Whamcloud - gitweb
LU-4019 ofd: setattr don't udpate lvbo with object referenced
[fs/lustre-release.git] / lustre / ofd / ofd_obd.c
index be76426..8bcbd06 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, Intel Corporation.
+ * Copyright (c) 2012, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 
 #define DEBUG_SUBSYSTEM S_FILTER
 
+#include <lustre/lustre_idl.h>
 #include "ofd_internal.h"
 #include <obd_cksum.h>
 #include <lustre_quota.h>
+#include <lustre_lfsck.h>
 
 static int ofd_export_stats_init(struct ofd_device *ofd,
                                 struct obd_export *exp, void *client_nid)
 {
        struct obd_device       *obd = ofd_obd(ofd);
        struct nid_stat         *stats;
-       int                      num_stats, i;
+       int                      num_stats;
        int                      rc, newnid = 0;
 
        ENTRY;
 
+       LASSERT(obd->obd_uses_nid_stats);
+
        if (obd_uuid_equals(&exp->exp_client_uuid, &obd->obd_uuid))
                /* Self-export gets no proc entry */
                RETURN(0);
@@ -75,21 +79,7 @@ static int ofd_export_stats_init(struct ofd_device *ofd,
        stats = exp->exp_nid_stats;
        LASSERT(stats != NULL);
 
-       OBD_ALLOC(stats->nid_brw_stats, sizeof(struct brw_stats));
-       if (stats->nid_brw_stats == NULL)
-               GOTO(clean, rc = -ENOMEM);
-
-       for (i = 0; i < BRW_LAST; i++)
-               spin_lock_init(&stats->nid_brw_stats->hist[i].oh_lock);
-
-       rc = lprocfs_seq_create(stats->nid_proc, "brw_stats", 0644,
-                               &ofd_per_nid_stats_fops, stats);
-       if (rc)
-               CWARN("Error adding the brw_stats file\n");
-
-       num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
-                    LPROC_OFD_LAST - 1;
-
+       num_stats = NUM_OBD_STATS + LPROC_OFD_LAST;
        stats->nid_stats = lprocfs_alloc_stats(num_stats,
                                               LPROCFS_STATS_FLAG_NOPERCPU);
        if (stats->nid_stats == NULL)
@@ -119,7 +109,8 @@ clean:
 
 static int ofd_parse_connect_data(const struct lu_env *env,
                                  struct obd_export *exp,
-                                 struct obd_connect_data *data)
+                                 struct obd_connect_data *data,
+                                 bool new_connection)
 {
        struct ofd_device                *ofd = ofd_exp(exp);
        struct filter_export_data        *fed = &exp->exp_filter_data;
@@ -146,7 +137,6 @@ static int ofd_parse_connect_data(const struct lu_env *env,
        fed->fed_group = data->ocd_group;
 
        data->ocd_connect_flags &= OST_CONNECT_SUPPORTED;
-       exp->exp_connect_data = *data;
        data->ocd_version = LUSTRE_VERSION_CODE;
 
        /* Kindly make sure the SKIP_ORPHAN flag is from MDS. */
@@ -165,8 +155,9 @@ static int ofd_parse_connect_data(const struct lu_env *env,
                data->ocd_grant_extent = ofd->ofd_dt_conf.ddp_grant_frag >> 10;
        }
 
-       if (exp_connect_flags(exp) & OBD_CONNECT_GRANT)
-               data->ocd_grant = ofd_grant_connect(env, exp, data->ocd_grant);
+       if (data->ocd_connect_flags & OBD_CONNECT_GRANT)
+               data->ocd_grant = ofd_grant_connect(env, exp, data->ocd_grant,
+                                                   new_connection);
 
        if (data->ocd_connect_flags & OBD_CONNECT_INDEX) {
                struct lr_server_data *lsd = &ofd->ofd_lut.lut_lsd;
@@ -237,8 +228,8 @@ static int ofd_parse_connect_data(const struct lu_env *env,
        if (data->ocd_connect_flags & OBD_CONNECT_MAXBYTES)
                data->ocd_maxbytes = ofd->ofd_dt_conf.ddp_maxbytes;
 
-       if (data->ocd_connect_flags & OBD_CONNECT_PINGLESS) {
-               if (suppress_pings) {
+       if (OCD_HAS_FLAG(data, PINGLESS)) {
+               if (ptlrpc_pinger_suppress_pings()) {
                        spin_lock(&exp->exp_obd->obd_dev_lock);
                        list_del_init(&exp->exp_obd_chain_timed);
                        spin_unlock(&exp->exp_obd->obd_dev_lock);
@@ -271,7 +262,7 @@ static int ofd_obd_reconnect(const struct lu_env *env, struct obd_export *exp,
        }
 
        ofd_info_init(env, exp);
-       rc = ofd_parse_connect_data(env, exp, data);
+       rc = ofd_parse_connect_data(env, exp, data, false);
        if (rc == 0)
                ofd_export_stats_init(ofd, exp, localdata);
 
@@ -308,7 +299,7 @@ static int ofd_obd_connect(const struct lu_env *env, struct obd_export **_exp,
 
        ofd_info_init(env, exp);
 
-       rc = ofd_parse_connect_data(env, exp, data);
+       rc = ofd_parse_connect_data(env, exp, data, true);
        if (rc)
                GOTO(out, rc);
 
@@ -324,7 +315,7 @@ static int ofd_obd_connect(const struct lu_env *env, struct obd_export **_exp,
        }
 
        CDEBUG(D_HA, "%s: get connection from MDS %d\n", obd->obd_name,
-              data->ocd_group);
+              data ? data->ocd_group : -1);
 
 out:
        if (rc != 0) {
@@ -539,7 +530,7 @@ static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
                        struct lov_stripe_md *lsm)
 {
        struct ofd_device       *ofd;
-       int                      rc = 0;
+       int                     rc = 0;
 
        ENTRY;
 
@@ -548,13 +539,21 @@ static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
                RETURN(-EINVAL);
        }
 
+       /* Because ofd_get_info might be called from
+        * handle_request_in as well(see LU-3239), where env might
+        * not be initilaized correctly, and le_ses might be in
+        * an un-initialized state, so only refill le_ctx here */
+       rc = lu_env_refill((struct lu_env *)env);
+       if (rc != 0)
+               RETURN(rc);
+
        ofd = ofd_exp(exp);
 
        if (KEY_IS(KEY_BLOCKSIZE)) {
                __u32 *blocksize = val;
                if (blocksize) {
                        if (*vallen < sizeof(*blocksize))
-                               RETURN(-EOVERFLOW);
+                               GOTO(out, rc = -EOVERFLOW);
                        *blocksize = 1 << ofd->ofd_dt_conf.ddp_block_shift;
                }
                *vallen = sizeof(*blocksize);
@@ -562,7 +561,7 @@ static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
                __u32 *blocksize_bits = val;
                if (blocksize_bits) {
                        if (*vallen < sizeof(*blocksize_bits))
-                               RETURN(-EOVERFLOW);
+                               GOTO(out, rc = -EOVERFLOW);
                        *blocksize_bits = ofd->ofd_dt_conf.ddp_block_shift;
                }
                *vallen = sizeof(*blocksize_bits);
@@ -572,44 +571,43 @@ static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
 
                if (val == NULL) {
                        *vallen = sizeof(obd_id);
-                       RETURN(0);
+                       GOTO(out, rc = 0);
                }
                ofd_info_init(env, exp);
                oseq = ofd_seq_load(env, ofd,
                                    (obd_seq)exp->exp_filter_data.fed_group);
-               LASSERT(oseq != NULL);
+               LASSERT(!IS_ERR(oseq));
                if (last_id) {
                        if (*vallen < sizeof(*last_id)) {
                                ofd_seq_put(env, oseq);
-                               RETURN(-EOVERFLOW);
+                               GOTO(out, rc = -EOVERFLOW);
                        }
                        *last_id = ofd_seq_last_oid(oseq);
                }
                ofd_seq_put(env, oseq);
                *vallen = sizeof(*last_id);
        } else if (KEY_IS(KEY_FIEMAP)) {
-               struct ofd_thread_info          *info;
                struct ofd_device               *ofd = ofd_exp(exp);
                struct ofd_object               *fo;
                struct ll_fiemap_info_key       *fm_key = key;
+               struct lu_fid                    fid;
 
                if (val == NULL) {
                        *vallen = fiemap_count_to_size(
                                               fm_key->fiemap.fm_extent_count);
-                       RETURN(0);
+                       GOTO(out, rc = 0);
                }
 
-               info = ofd_info_init(env, exp);
-
-               fid_ostid_unpack(&info->fti_fid, &fm_key->oa.o_oi, 0);
-
+               rc = ostid_to_fid(&fid, &fm_key->oa.o_oi, 0);
+               if (rc != 0)
+                       GOTO(out, rc);
                CDEBUG(D_INODE, "get FIEMAP of object "DFID"\n",
-                      PFID(&info->fti_fid));
+                      PFID(&fid));
 
-               fo = ofd_object_find(env, ofd, &info->fti_fid);
+               fo = ofd_object_find(env, ofd, &fid);
                if (IS_ERR(fo)) {
                        CERROR("%s: error finding object "DFID"\n",
-                              exp->exp_obd->obd_name, PFID(&info->fti_fid));
+                              exp->exp_obd->obd_name, PFID(&fid));
                        rc = PTR_ERR(fo);
                } else {
                        struct ll_user_fiemap *fiemap = val;
@@ -630,42 +628,43 @@ static int ofd_get_info(const struct lu_env *env, struct obd_export *exp,
                *((__u32 *) val) = ofd->ofd_sync_lock_cancel;
                *vallen = sizeof(__u32);
        } else if (KEY_IS(KEY_LAST_FID)) {
-               struct lu_env      env;
-               struct ofd_device *ofd = ofd_exp(exp);
-               struct ofd_seq    *oseq;
-               struct lu_fid     *last_fid = val;
-               int             rc;
+               struct ofd_device       *ofd = ofd_exp(exp);
+               struct ofd_seq          *oseq;
+               struct lu_fid           *fid = val;
+               int                     rc;
 
-               if (last_fid == NULL) {
+               if (fid == NULL) {
                        *vallen = sizeof(struct lu_fid);
-                       RETURN(0);
+                       GOTO(out, rc = 0);
                }
 
-               if (*vallen < sizeof(*last_fid))
-                       RETURN(-EOVERFLOW);
+               if (*vallen < sizeof(*fid))
+                       GOTO(out, rc = -EOVERFLOW);
 
-               rc = lu_env_init(&env, LCT_DT_THREAD);
-               if (rc != 0)
-                       RETURN(rc);
-               ofd_info_init(&env, exp);
-               fid_le_to_cpu(last_fid, last_fid);
-               oseq = ofd_seq_load(&env, ofd, fid_seq(last_fid));
+               ofd_info_init(env, exp);
+
+               fid_le_to_cpu(fid, fid);
+
+               oseq = ofd_seq_load(env, ofd,
+                                   ostid_seq((struct ost_id *)fid));
                if (IS_ERR(oseq))
-                       GOTO(out_fid, rc = PTR_ERR(oseq));
+                       GOTO(out, rc = PTR_ERR(oseq));
 
-               last_fid->f_seq = oseq->os_seq;
-               last_fid->f_oid = oseq->os_last_oid;
-               fid_cpu_to_le(last_fid, last_fid);
+               rc = ostid_to_fid(fid, &oseq->os_oi,
+                            ofd->ofd_lut.lut_lsd.lsd_osd_index);
+               if (rc != 0)
+                       GOTO(out_put, rc);
 
-               *vallen = sizeof(*last_fid);
-               ofd_seq_put(&env, oseq);
-out_fid:
-               lu_env_fini(&env);
+               CDEBUG(D_HA, "%s: LAST FID is "DFID"\n", ofd_name(ofd),
+                      PFID(fid));
+               *vallen = sizeof(*fid);
+out_put:
+               ofd_seq_put(env, oseq);
        } else {
                CERROR("Not supported key %s\n", (char*)key);
                rc = -EOPNOTSUPP;
        }
-
+out:
        RETURN(rc);
 }
 
@@ -839,10 +838,12 @@ int ofd_setattr(const struct lu_env *env, struct obd_export *exp,
        info = ofd_info_init(env, exp);
        ofd_oti2info(info, oti);
 
-       fid_ostid_unpack(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
-       ofd_build_resid(&info->fti_fid, &info->fti_resid);
+       rc = ostid_to_fid(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
+       if (rc != 0)
+               RETURN(rc);
 
-       rc = ofd_auth_capa(exp, &info->fti_fid, oa->o_seq,
+       ost_fid_build_resid(&info->fti_fid, &info->fti_resid);
+       rc = ofd_auth_capa(exp, &info->fti_fid, ostid_seq(&oa->o_oi),
                           oinfo_capa(oinfo), CAPA_OPC_META_WRITE);
        if (rc)
                GOTO(out, rc);
@@ -886,18 +887,6 @@ int ofd_setattr(const struct lu_env *env, struct obd_export *exp,
        if (rc)
                GOTO(out_unlock, rc);
 
-       res = ldlm_resource_get(ns, NULL, &info->fti_resid, LDLM_EXTENT, 0);
-       if (res != NULL) {
-               ldlm_res_lvbo_update(res, NULL, 0);
-               ldlm_resource_putref(res);
-       }
-
-       oinfo->oi_oa->o_valid = OBD_MD_FLID;
-
-       /* Quota release needs uid/gid info */
-       rc = ofd_attr_get(env, fo, &info->fti_attr);
-       obdo_from_la(oinfo->oi_oa, &info->fti_attr,
-                    OFD_VALID_FLAGS | LA_UID | LA_GID);
        ofd_info2oti(info, oti);
 
        ofd_counter_incr(exp, LPROC_OFD_STATS_SETATTR, oti->oti_jobid, 1);
@@ -905,6 +894,19 @@ int ofd_setattr(const struct lu_env *env, struct obd_export *exp,
 out_unlock:
        ofd_object_put(env, fo);
 out:
+       if (rc == 0) {
+               /* we do not call this before to avoid lu_object_find() in
+                *  ->lvbo_update() holding another reference on the object.
+                * otherwise concurrent destroy can make the object unavailable
+                * for 2nd lu_object_find() waiting for the first reference
+                * to go... deadlock! */
+               res = ldlm_resource_get(ns, NULL, &info->fti_resid, LDLM_EXTENT, 0);
+               if (res != NULL) {
+                       ldlm_res_lvbo_update(res, NULL, 0);
+                       ldlm_resource_putref(res);
+               }
+       }
+
        return rc;
 }
 
@@ -925,15 +927,17 @@ static int ofd_punch(const struct lu_env *env, struct obd_export *exp,
        info = ofd_info_init(env, exp);
        ofd_oti2info(info, oti);
 
-       fid_ostid_unpack(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
-       ofd_build_resid(&info->fti_fid, &info->fti_resid);
+       rc = ostid_to_fid(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
+       if (rc != 0)
+               RETURN(rc);
+       ost_fid_build_resid(&info->fti_fid, &info->fti_resid);
 
        CDEBUG(D_INODE, "calling punch for object "DFID", valid = "LPX64
               ", start = "LPD64", end = "LPD64"\n", PFID(&info->fti_fid),
               oinfo->oi_oa->o_valid, oinfo->oi_policy.l_extent.start,
               oinfo->oi_policy.l_extent.end);
 
-       rc = ofd_auth_capa(exp, &info->fti_fid, oinfo->oi_oa->o_seq,
+       rc = ofd_auth_capa(exp, &info->fti_fid, ostid_seq(&oinfo->oi_oa->o_oi),
                           oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
        if (rc)
                GOTO(out_env, rc);
@@ -998,7 +1002,8 @@ static int ofd_destroy_by_fid(const struct lu_env *env,
 {
        struct ofd_thread_info  *info = ofd_info(env);
        struct lustre_handle     lockh;
-       __u64                    flags = LDLM_AST_DISCARD_DATA, rc = 0;
+       __u64                    flags = LDLM_FL_AST_DISCARD_DATA;
+       __u64                    rc = 0;
        ldlm_policy_data_t       policy = {
                                        .l_extent = { 0, OBD_OBJECT_EOF }
                                 };
@@ -1009,10 +1014,12 @@ static int ofd_destroy_by_fid(const struct lu_env *env,
        fo = ofd_object_find(env, ofd, fid);
        if (IS_ERR(fo))
                RETURN(PTR_ERR(fo));
+       if (!ofd_object_exists(fo))
+               GOTO(out, rc = -ENOENT);
 
        /* Tell the clients that the object is gone now and that they should
         * throw away any cached pages. */
-       ofd_build_resid(fid, &info->fti_resid);
+       ost_fid_build_resid(fid, &info->fti_resid);
        rc = ldlm_cli_enqueue_local(ofd->ofd_namespace, &info->fti_resid,
                                    LDLM_EXTENT, &policy, LCK_PW, &flags,
                                    ldlm_blocking_ast, ldlm_completion_ast,
@@ -1025,7 +1032,8 @@ static int ofd_destroy_by_fid(const struct lu_env *env,
        LASSERT(fo != NULL);
 
        rc = ofd_object_destroy(env, fo, orphan);
-
+       EXIT;
+out:
        ofd_object_put(env, fo);
        RETURN(rc);
 }
@@ -1046,7 +1054,7 @@ int ofd_destroy(const struct lu_env *env, struct obd_export *exp,
        ofd_oti2info(info, oti);
 
        if (!(oa->o_valid & OBD_MD_FLGROUP))
-               oa->o_seq = 0;
+               ostid_set_seq_mdt0(&oa->o_oi);
 
        /* check that o_misc makes sense */
        if (oa->o_valid & OBD_MD_FLOBJCOUNT)
@@ -1062,10 +1070,18 @@ int ofd_destroy(const struct lu_env *env, struct obd_export *exp,
         */
        if (info->fti_transno == 0) /* not replay */
                info->fti_mult_trans = 1;
+
+       CDEBUG(D_HA, "%s: Destroy object "DOSTID" count %d\n", ofd_name(ofd),
+              POSTID(&oa->o_oi), count);
        while (count > 0) {
                int lrc;
 
-               fid_ostid_unpack(&info->fti_fid, &oa->o_oi, 0);
+               lrc = ostid_to_fid(&info->fti_fid, &oa->o_oi, 0);
+               if (lrc != 0) {
+                       if (rc == 0)
+                               rc = lrc;
+                       GOTO(out, rc);
+               }
                lrc = ofd_destroy_by_fid(env, ofd, &info->fti_fid, 0);
                if (lrc == -ENOENT) {
                        CDEBUG(D_INODE,
@@ -1081,7 +1097,7 @@ int ofd_destroy(const struct lu_env *env, struct obd_export *exp,
                        rc = lrc;
                }
                count--;
-               oa->o_id++;
+               ostid_inc_id(&oa->o_oi);
        }
 
        /* if we have transaction then there were some deletions, we don't
@@ -1101,6 +1117,7 @@ int ofd_destroy(const struct lu_env *env, struct obd_export *exp,
                info->fti_transno = 0;
        }
        ofd_info2oti(info, oti);
+out:
        RETURN(rc);
 }
 
@@ -1113,14 +1130,15 @@ static int ofd_orphans_destroy(const struct lu_env *env,
        int                      skip_orphan;
        int                      rc = 0;
        struct ost_id            oi = oa->o_oi;
+       __u64                    end_id = ostid_id(&oa->o_oi);
        struct ofd_seq          *oseq;
 
        ENTRY;
 
-       oseq = ofd_seq_get(ofd, oa->o_seq);
+       oseq = ofd_seq_get(ofd, ostid_seq(&oa->o_oi));
        if (oseq == NULL) {
-               CERROR("%s: Can not find seq for "LPU64":"LPU64"\n",
-                      ofd_name(ofd), oa->o_seq, oa->o_id);
+               CERROR("%s: Can not find seq for "DOSTID"\n",
+                      ofd_name(ofd), POSTID(&oa->o_oi));
                RETURN(-EINVAL);
        }
 
@@ -1128,34 +1146,38 @@ static int ofd_orphans_destroy(const struct lu_env *env,
        skip_orphan = !!(exp_connect_flags(exp) & OBD_CONNECT_SKIP_ORPHAN);
 
        last = ofd_seq_last_oid(oseq);
-       LCONSOLE_INFO("%s: deleting orphan objects from "LPX64":"LPU64
-                     " to "LPU64"\n", ofd_name(ofd), oa->o_seq,
-                     oa->o_id + 1, last);
+       LCONSOLE(D_INFO, "%s: deleting orphan objects from "DOSTID
+                " to "DOSTID"\n", ofd_name(ofd), ostid_seq(&oa->o_oi),
+                end_id + 1, ostid_seq(&oa->o_oi), last);
 
-       for (oi.oi_id = last; oi.oi_id > oa->o_id; oi.oi_id--) {
-               fid_ostid_unpack(&info->fti_fid, &oi, 0);
+       for (ostid_set_id(&oi, last); ostid_id(&oi) > end_id;
+                         ostid_dec_id(&oi)) {
+               rc = ostid_to_fid(&info->fti_fid, &oi, 0);
+               if (rc != 0)
+                       GOTO(out_put, rc);
                rc = ofd_destroy_by_fid(env, ofd, &info->fti_fid, 1);
                if (rc && rc != -ENOENT) /* this is pretty fatal... */
-                       CEMERG("error destroying precreated id "LPU64": %d\n",
-                              oi.oi_id, rc);
+                       CEMERG("%s: error destroying precreated id "DOSTID
+                              ": rc = %d\n", ofd_name(ofd), POSTID(&oi), rc);
                if (!skip_orphan) {
-                       ofd_seq_last_oid_set(oseq, oi.oi_id - 1);
+                       ofd_seq_last_oid_set(oseq, ostid_id(&oi) - 1);
                        /* update last_id on disk periodically so that if we
                         * restart * we don't need to re-scan all of the just
                         * deleted objects. */
-                       if ((oi.oi_id & 511) == 0)
+                       if ((ostid_id(&oi) & 511) == 0)
                                ofd_seq_last_oid_write(env, ofd, oseq);
                }
        }
-       CDEBUG(D_HA, "%s: after destroy: set last_objids["LPU64"] = "LPU64"\n",
-              ofd_obd(ofd)->obd_name, oa->o_seq, oa->o_id);
+       CDEBUG(D_HA, "%s: after destroy: set last_id to "DOSTID"\n",
+              ofd_obd(ofd)->obd_name, POSTID(&oa->o_oi));
        if (!skip_orphan) {
                rc = ofd_seq_last_oid_write(env, ofd, oseq);
        } else {
                /* don't reuse orphan object, return last used objid */
-               oa->o_id = last;
+               ostid_set_id(&oa->o_oi, last);
                rc = 0;
        }
+out_put:
        ofd_seq_put(env, oseq);
        RETURN(rc);
 }
@@ -1166,7 +1188,7 @@ int ofd_create(const struct lu_env *env, struct obd_export *exp,
 {
        struct ofd_device       *ofd = ofd_exp(exp);
        struct ofd_thread_info  *info;
-       obd_seq                 seq = oa->o_seq;
+       obd_seq                 seq = ostid_seq(&oa->o_oi);
        struct ofd_seq          *oseq;
        int                     rc = 0, diff;
        int                     sync_trans = 0;
@@ -1176,24 +1198,24 @@ int ofd_create(const struct lu_env *env, struct obd_export *exp,
        info = ofd_info_init(env, exp);
        ofd_oti2info(info, oti);
 
-       LASSERT(oa->o_seq >= FID_SEQ_OST_MDT0);
+       LASSERT(ostid_seq(&oa->o_oi) >= FID_SEQ_OST_MDT0);
        LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 
-       CDEBUG(D_INFO, "ofd_create(oa->o_seq="LPU64",oa->o_id="LPU64")\n",
-              seq, oa->o_id);
+       CDEBUG(D_INFO, "ofd_create("DOSTID")\n", POSTID(&oa->o_oi));
 
        oseq = ofd_seq_load(env, ofd, seq);
-       if (oseq == NULL) {
-               CERROR("%s: Can't find oseq "LPX64"\n", ofd_name(ofd), seq);
+       if (IS_ERR(oseq)) {
+               CERROR("%s: Can't find FID Sequence "LPX64": rc = %ld\n",
+                      ofd_name(ofd), seq, PTR_ERR(oseq));
                RETURN(-EINVAL);
        }
 
        if ((oa->o_valid & OBD_MD_FLFLAGS) &&
            (oa->o_flags & OBD_FL_RECREATE_OBJS)) {
                if (!ofd_obd(ofd)->obd_recovering ||
-                   oa->o_id > ofd_seq_last_oid(oseq)) {
-                       CERROR("recreate objid "LPU64" > last id "LPU64"\n",
-                                       oa->o_id, ofd_seq_last_oid(oseq));
+                   ostid_id(&oa->o_oi) > ofd_seq_last_oid(oseq)) {
+                       CERROR("recreate objid "DOSTID" > last id "LPU64"\n",
+                              POSTID(&oa->o_oi), ofd_seq_last_oid(oseq));
                        GOTO(out_nolock, rc = -EINVAL);
                }
                /* do nothing because we create objects during first write */
@@ -1212,11 +1234,13 @@ int ofd_create(const struct lu_env *env, struct obd_export *exp,
                oseq->os_destroys_in_progress = 1;
                mutex_lock(&oseq->os_create_lock);
                if (!oseq->os_destroys_in_progress) {
-                       CERROR("%s:["LPU64"] destroys_in_progress already cleared\n",
-                              exp->exp_obd->obd_name, oa->o_seq);
+                       CERROR("%s:["LPU64"] destroys_in_progress already"
+                              " cleared\n", exp->exp_obd->obd_name,
+                              ostid_seq(&oa->o_oi));
+                       ostid_set_id(&oa->o_oi, ofd_seq_last_oid(oseq));
                        GOTO(out, rc = 0);
                }
-               diff = oa->o_id - ofd_seq_last_oid(oseq);
+               diff = ostid_id(&oa->o_oi) - ofd_seq_last_oid(oseq);
                CDEBUG(D_HA, "ofd_last_id() = "LPU64" -> diff = %d\n",
                        ofd_seq_last_oid(oseq), diff);
                if (-diff > OST_MAX_PRECREATE) {
@@ -1238,24 +1262,25 @@ int ofd_create(const struct lu_env *env, struct obd_export *exp,
                }
                /* only precreate if seq is 0, IDIF or normal and also o_id
                 * must be specfied */
-               if ((!fid_seq_is_mdt(oa->o_seq) &&
-                    !fid_seq_is_norm(oa->o_seq) &&
-                    !fid_seq_is_idif(oa->o_seq)) || oa->o_id == 0) {
+               if ((!fid_seq_is_mdt(ostid_seq(&oa->o_oi)) &&
+                    !fid_seq_is_norm(ostid_seq(&oa->o_oi)) &&
+                    !fid_seq_is_idif(ostid_seq(&oa->o_oi))) ||
+                               ostid_id(&oa->o_oi) == 0) {
                        diff = 1; /* shouldn't we create this right now? */
                } else {
-                       diff = oa->o_id - ofd_seq_last_oid(oseq);
+                       diff = ostid_id(&oa->o_oi) - ofd_seq_last_oid(oseq);
                        /* Do sync create if the seq is about to used up */
-                       if (fid_seq_is_idif(oa->o_seq) ||
-                           fid_seq_is_mdt0(oa->o_seq)) {
-                               if (unlikely(oa->o_id >= IDIF_MAX_OID - 1))
+                       if (fid_seq_is_idif(ostid_seq(&oa->o_oi)) ||
+                           fid_seq_is_mdt0(ostid_seq(&oa->o_oi))) {
+                               if (unlikely(ostid_id(&oa->o_oi) >= IDIF_MAX_OID - 1))
                                        sync_trans = 1;
-                       } else if (fid_seq_is_norm(oa->o_seq)) {
-                               if (unlikely(oa->o_id >=
+                       } else if (fid_seq_is_norm(ostid_seq(&oa->o_oi))) {
+                               if (unlikely(ostid_id(&oa->o_oi) >=
                                             LUSTRE_DATA_SEQ_MAX_WIDTH - 1))
                                        sync_trans = 1;
                        } else {
-                               CERROR("%s : invalid o_seq "LPX64": rc = %d\n",
-                                      ofd_name(ofd), oa->o_seq, -EINVAL);
+                               CERROR("%s : invalid o_seq "DOSTID": rc = %d\n",
+                                      ofd_name(ofd), POSTID(&oa->o_oi), -EINVAL);
                                GOTO(out, rc = -EINVAL);
                        }
                }
@@ -1273,20 +1298,36 @@ int ofd_create(const struct lu_env *env, struct obd_export *exp,
                                              ofd_obd(ofd)->obd_self_export,
                                              &diff);
                        if (rc) {
-                               CDEBUG(D_HA, "%s: failed to acquire grant space"
-                                      "for precreate (%d)\n",
-                                      ofd_name(ofd), diff);
+                               CDEBUG(D_HA, "%s: failed to acquire grant "
+                                      "space for precreate (%d): rc = %d\n",
+                                      ofd_name(ofd), diff, rc);
                                diff = 0;
                        }
                }
 
+               /* This can happen if a new OST is formatted and installed
+                * in place of an old one at the same index.  Instead of
+                * precreating potentially millions of deleted old objects
+                * (possibly filling the OST), only precreate the last batch.
+                * LFSCK will eventually clean up any orphans. LU-14 */
+               if (diff > 5 * OST_MAX_PRECREATE) {
+                       diff = OST_MAX_PRECREATE / 2;
+                       LCONSOLE_WARN("%s: precreate FID "DOSTID" is over %u "
+                                     "larger than the LAST_ID "DOSTID", only "
+                                     "precreating the last %u objects.\n",
+                                     ofd_name(ofd), POSTID(&oa->o_oi),
+                                     5 * OST_MAX_PRECREATE,
+                                     POSTID(&oseq->os_oi), diff);
+                       ofd_seq_last_oid_set(oseq, ostid_id(&oa->o_oi) - diff);
+               }
+
                while (diff > 0) {
                        next_id = ofd_seq_last_oid(oseq) + 1;
                        count = ofd_precreate_batch(ofd, diff);
 
-                       CDEBUG(D_HA, "%s: reserve %d objects in group "LPU64
+                       CDEBUG(D_HA, "%s: reserve %d objects in group "LPX64
                               " at "LPU64"\n", ofd_obd(ofd)->obd_name,
-                              count, oa->o_seq, next_id);
+                              count, ostid_seq(&oa->o_oi), next_id);
 
                        if (cfs_time_after(jiffies, enough_time)) {
                                LCONSOLE_WARN("%s: Slow creates, %d/%d objects"
@@ -1311,9 +1352,10 @@ int ofd_create(const struct lu_env *env, struct obd_export *exp,
                         * them, even if last creation failed */
                        rc = 0;
                else
-                       CERROR("unable to precreate: %d\n", rc);
+                       CERROR("%s: unable to precreate: rc = %d\n",
+                              ofd_name(ofd), rc);
 
-               oa->o_id = ofd_seq_last_oid(oseq);
+               ostid_set_id(&oa->o_oi, ofd_seq_last_oid(oseq));
                oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 
                if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
@@ -1329,7 +1371,7 @@ out_nolock:
        if (rc == 0 && ea != NULL) {
                struct lov_stripe_md *lsm = *ea;
 
-               lsm->lsm_object_id = oa->o_id;
+               lsm->lsm_oi = oa->o_oi;
        }
        ofd_seq_put(env, oseq);
        RETURN(rc);
@@ -1341,15 +1383,16 @@ int ofd_getattr(const struct lu_env *env, struct obd_export *exp,
        struct ofd_device       *ofd = ofd_exp(exp);
        struct ofd_thread_info  *info;
        struct ofd_object       *fo;
-       __u64                    curr_version;
        int                      rc = 0;
 
        ENTRY;
 
        info = ofd_info_init(env, exp);
 
-       fid_ostid_unpack(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
-       rc = ofd_auth_capa(exp, &info->fti_fid, oinfo->oi_oa->o_seq,
+       rc = ostid_to_fid(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
+       if (rc != 0)
+               GOTO(out, rc);
+       rc = ofd_auth_capa(exp, &info->fti_fid, ostid_seq(&oinfo->oi_oa->o_oi),
                           oinfo_capa(oinfo), CAPA_OPC_META_READ);
        if (rc)
                GOTO(out, rc);
@@ -1357,19 +1400,27 @@ int ofd_getattr(const struct lu_env *env, struct obd_export *exp,
        fo = ofd_object_find(env, ofd, &info->fti_fid);
        if (IS_ERR(fo))
                GOTO(out, rc = PTR_ERR(fo));
+       if (!ofd_object_exists(fo))
+               GOTO(out_put, rc = -ENOENT);
+
        LASSERT(fo != NULL);
        rc = ofd_attr_get(env, fo, &info->fti_attr);
        oinfo->oi_oa->o_valid = OBD_MD_FLID;
-       if (rc == 0)
+       if (rc == 0) {
+               __u64 curr_version;
+
                obdo_from_la(oinfo->oi_oa, &info->fti_attr,
                             OFD_VALID_FLAGS | LA_UID | LA_GID);
 
-       /* Store object version in reply */
-       curr_version = dt_version_get(env, ofd_object_child(fo));
-       if ((__s64)curr_version != -EOPNOTSUPP) {
-               oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
-               oinfo->oi_oa->o_data_version = curr_version;
+               /* Store object version in reply */
+               curr_version = dt_version_get(env, ofd_object_child(fo));
+               if ((__s64)curr_version != -EOPNOTSUPP) {
+                       oinfo->oi_oa->o_valid |= OBD_MD_FLDATAVERSION;
+                       oinfo->oi_oa->o_data_version = curr_version;
+               }
        }
+
+out_put:
        ofd_object_put(env, fo);
 out:
        RETURN(rc);
@@ -1393,9 +1444,11 @@ static int ofd_sync(const struct lu_env *env, struct obd_export *exp,
        }
 
        info = ofd_info_init(env, exp);
-       fid_ostid_unpack(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
+       rc = ostid_to_fid(&info->fti_fid, &oinfo->oi_oa->o_oi, 0);
+       if (rc != 0)
+               GOTO(out, rc);
 
-       rc = ofd_auth_capa(exp, &info->fti_fid, oinfo->oi_oa->o_seq,
+       rc = ofd_auth_capa(exp, &info->fti_fid, ostid_seq(&oinfo->oi_oa->o_oi),
                           oinfo_capa(oinfo), CAPA_OPC_OSS_TRUNC);
        if (rc)
                GOTO(out, rc);
@@ -1430,6 +1483,63 @@ out:
        return rc;
 }
 
+static int ofd_ioc_get_obj_version(const struct lu_env *env,
+                                  struct ofd_device *ofd, void *karg)
+{
+       struct obd_ioctl_data *data = karg;
+       struct lu_fid          fid;
+       struct ofd_object     *fo;
+       dt_obj_version_t       version;
+       int                    rc = 0;
+
+       ENTRY;
+
+       if (data->ioc_inlbuf2 == NULL || data->ioc_inllen2 != sizeof(version))
+               GOTO(out, rc = -EINVAL);
+
+       if (data->ioc_inlbuf1 != NULL && data->ioc_inllen1 == sizeof(fid)) {
+               fid = *(struct lu_fid *)data->ioc_inlbuf1;
+       } else if (data->ioc_inlbuf3 != NULL &&
+                  data->ioc_inllen3 == sizeof(__u64) &&
+                  data->ioc_inlbuf4 != NULL &&
+                  data->ioc_inllen4 == sizeof(__u64)) {
+               struct ost_id ostid;
+
+               ostid_set_seq(&ostid, *(__u64 *)data->ioc_inlbuf4);
+               ostid_set_id(&ostid, *(__u64 *)data->ioc_inlbuf3);
+               rc = ostid_to_fid(&fid, &ostid, 0);
+               if (rc != 0)
+                       GOTO(out, rc);
+       } else {
+               GOTO(out, rc = -EINVAL);
+       }
+
+       if (!fid_is_sane(&fid))
+               GOTO(out, rc = -EINVAL);
+
+       fo = ofd_object_find(env, ofd, &fid);
+       if (IS_ERR(fo))
+               GOTO(out, rc = PTR_ERR(fo));
+
+       if (!ofd_object_exists(fo))
+               GOTO(out_fo, rc = -ENOENT);
+
+       if (lu_object_remote(&fo->ofo_obj.do_lu))
+               GOTO(out_fo, rc = -EREMOTE);
+
+       version = dt_version_get(env, ofd_object_child(fo));
+       if (version == 0)
+               GOTO(out_fo, rc = -EIO);
+
+       *(dt_obj_version_t *)data->ioc_inlbuf2 = version;
+
+       EXIT;
+out_fo:
+       ofd_object_put(env, fo);
+out:
+       return rc;
+}
+
 int ofd_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                  void *karg, void *uarg)
 {
@@ -1441,7 +1551,7 @@ int ofd_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
        ENTRY;
 
        CDEBUG(D_IOCTL, "handling ioctl cmd %#x\n", cmd);
-       rc = lu_env_init(&env, LCT_LOCAL);
+       rc = lu_env_init(&env, LCT_DT_THREAD);
        if (rc)
                RETURN(rc);
 
@@ -1459,6 +1569,27 @@ int ofd_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
                if (rc == 0)
                        rc = dt_ro(&env, ofd->ofd_osd);
                break;
+       case OBD_IOC_START_LFSCK: {
+               struct obd_ioctl_data *data = karg;
+               struct lfsck_start_param lsp;
+
+               if (unlikely(data == NULL)) {
+                       rc = -EINVAL;
+                       break;
+               }
+
+               lsp.lsp_start = (struct lfsck_start *)(data->ioc_inlbuf1);
+               lsp.lsp_namespace = ofd->ofd_namespace;
+               rc = lfsck_start(&env, ofd->ofd_osd, &lsp);
+               break;
+       }
+       case OBD_IOC_STOP_LFSCK: {
+               rc = lfsck_stop(&env, ofd->ofd_osd, false);
+               break;
+       }
+       case OBD_IOC_GET_OBJ_VERSION:
+               rc = ofd_ioc_get_obj_version(&env, ofd, karg);
+               break;
        default:
                CERROR("%s: not supported cmd = %d\n", obd->obd_name, cmd);
                rc = -ENOTTY;
@@ -1514,11 +1645,11 @@ static int ofd_health_check(const struct lu_env *nul, struct obd_device *obd)
                GOTO(out, rc = -EROFS);
 
 #ifdef USE_HEALTH_CHECK_WRITE
-       OBD_ALLOC(info->fti_buf.lb_buf, CFS_PAGE_SIZE);
+       OBD_ALLOC(info->fti_buf.lb_buf, PAGE_CACHE_SIZE);
        if (info->fti_buf.lb_buf == NULL)
                GOTO(out, rc = -ENOMEM);
 
-       info->fti_buf.lb_len = CFS_PAGE_SIZE;
+       info->fti_buf.lb_len = PAGE_CACHE_SIZE;
        info->fti_off = 0;
 
        th = dt_trans_create(&env, ofd->ofd_osd);
@@ -1537,7 +1668,7 @@ static int ofd_health_check(const struct lu_env *nul, struct obd_device *obd)
        }
        dt_trans_stop(&env, ofd->ofd_osd, th);
 
-       OBD_FREE(info->fti_buf.lb_buf, CFS_PAGE_SIZE);
+       OBD_FREE(info->fti_buf.lb_buf, PAGE_CACHE_SIZE);
 
        CDEBUG(D_INFO, "write 1 page synchronously for checking io rc %d\n",rc);
 #endif