+ ltd->ltd_layout_gen = llmd->llmd_touch_gen;
+ list_move_tail(<d->ltd_layout_phase_list,
+ &llmd->llmd_mdt_phase1_list);
+ atomic_inc(<d->ltd_ref);
+ laia->laia_ltd = ltd;
+ spin_unlock(<ds->ltd_lock);
+ rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
+ lfsck_layout_master_async_interpret,
+ laia, LFSCK_NOTIFY);
+ if (rc != 0) {
+ CDEBUG(D_LFSCK, "%s: layout LFSCK fail to "
+ "notify MDT %x for phase1_done: "
+ "rc = %d\n", lfsck_lfsck2name(lfsck),
+ ltd->ltd_index, rc);
+ lfsck_tgt_put(ltd);
+ }
+ spin_lock(<ds->ltd_lock);
+ }
+ spin_unlock(<ds->ltd_lock);
+ break;
+ default:
+ CDEBUG(D_LFSCK, "%s: layout LFSCK unexpected event: rc = %d\n",
+ lfsck_lfsck2name(lfsck), lr->lr_event);
+ rc = -EINVAL;
+ break;
+ }
+
+ rc = ptlrpc_set_wait(set);
+ ptlrpc_set_destroy(set);
+
+ RETURN(rc);
+}
+
+static int lfsck_layout_double_scan_result(const struct lu_env *env,
+ struct lfsck_component *com,
+ int rc)
+{
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct lfsck_layout *lo = com->lc_file_ram;
+ struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
+
+ down_write(&com->lc_sem);
+ lo->ll_run_time_phase2 += cfs_duration_sec(cfs_time_current() +
+ HALF_SEC - lfsck->li_time_last_checkpoint);
+ lo->ll_time_last_checkpoint = cfs_time_current_sec();
+ lo->ll_objs_checked_phase2 += com->lc_new_checked;
+
+ if (rc > 0) {
+ com->lc_journal = 0;
+ if (lo->ll_flags & LF_INCOMPLETE)
+ lo->ll_status = LS_PARTIAL;
+ else
+ lo->ll_status = LS_COMPLETED;
+ if (!(bk->lb_param & LPF_DRYRUN))
+ lo->ll_flags &= ~(LF_SCANNED_ONCE | LF_INCONSISTENT);
+ lo->ll_time_last_complete = lo->ll_time_last_checkpoint;
+ lo->ll_success_count++;
+ } else if (rc == 0) {
+ lo->ll_status = lfsck->li_status;
+ if (lo->ll_status == 0)
+ lo->ll_status = LS_STOPPED;
+ } else {
+ lo->ll_status = LS_FAILED;
+ }
+
+ rc = lfsck_layout_store(env, com);
+ up_write(&com->lc_sem);
+
+ return rc;
+}
+
+static int lfsck_layout_lock(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct dt_object *obj,
+ struct lustre_handle *lh, __u64 bits)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ ldlm_policy_data_t *policy = &info->lti_policy;
+ struct ldlm_res_id *resid = &info->lti_resid;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ __u64 flags = LDLM_FL_ATOMIC_CB;
+ int rc;
+
+ LASSERT(lfsck->li_namespace != NULL);
+
+ memset(policy, 0, sizeof(*policy));
+ policy->l_inodebits.bits = bits;
+ fid_build_reg_res_name(lfsck_dto2fid(obj), resid);
+ rc = ldlm_cli_enqueue_local(lfsck->li_namespace, resid, LDLM_IBITS,
+ policy, LCK_EX, &flags, ldlm_blocking_ast,
+ ldlm_completion_ast, NULL, NULL, 0,
+ LVB_T_NONE, NULL, lh);
+ if (rc == ELDLM_OK) {
+ rc = 0;
+ } else {
+ memset(lh, 0, sizeof(*lh));
+ rc = -EIO;
+ }
+
+ return rc;
+}
+
+static void lfsck_layout_unlock(struct lustre_handle *lh)
+{
+ if (lustre_handle_is_used(lh)) {
+ ldlm_lock_decref(lh, LCK_EX);
+ memset(lh, 0, sizeof(*lh));
+ }
+}
+
+static int lfsck_layout_trans_stop(const struct lu_env *env,
+ struct dt_device *dev,
+ struct thandle *handle, int result)
+{
+ int rc;
+
+ handle->th_result = result;
+ rc = dt_trans_stop(env, dev, handle);
+ if (rc > 0)
+ rc = 0;
+ else if (rc == 0)
+ rc = 1;
+
+ return rc;
+}
+
+/**
+ * Get the system default stripe size.
+ *
+ * \param[in] env pointer to the thread context
+ * \param[in] lfsck pointer to the lfsck instance
+ * \param[out] size pointer to the default stripe size
+ *
+ * \retval 0 for success
+ * \retval negative error number on failure
+ */
+static int lfsck_layout_get_def_stripesize(const struct lu_env *env,
+ struct lfsck_instance *lfsck,
+ __u32 *size)
+{
+ struct lov_user_md *lum = &lfsck_env_info(env)->lti_lum;
+ struct dt_object *root;
+ int rc;
+
+ root = dt_locate(env, lfsck->li_next, &lfsck->li_local_root_fid);
+ if (IS_ERR(root))
+ return PTR_ERR(root);
+
+ /* Get the default stripe size via xattr_get on the backend root. */
+ rc = dt_xattr_get(env, root, lfsck_buf_get(env, lum, sizeof(*lum)),
+ XATTR_NAME_LOV, BYPASS_CAPA);
+ if (rc > 0) {
+ /* The lum->lmm_stripe_size is LE mode. The *size also
+ * should be LE mode. So it is unnecessary to convert. */
+ *size = lum->lmm_stripe_size;
+ rc = 0;
+ } else if (unlikely(rc == 0)) {
+ rc = -EINVAL;
+ }
+
+ lfsck_object_put(env, root);
+
+ return rc;
+}
+
+/**
+ * \retval +1: repaired
+ * \retval 0: did nothing
+ * \retval -ve: on error
+ */
+static int lfsck_layout_refill_lovea(const struct lu_env *env,
+ struct thandle *handle,
+ struct dt_object *parent,
+ struct lu_fid *cfid,
+ struct lu_buf *buf,
+ struct lov_ost_data_v1 *slot,
+ int fl, __u32 ost_idx)
+{
+ struct ost_id *oi = &lfsck_env_info(env)->lti_oi;
+ struct lov_mds_md_v1 *lmm = buf->lb_buf;
+ int rc;
+
+ fid_to_ostid(cfid, oi);
+ ostid_cpu_to_le(oi, &slot->l_ost_oi);
+ slot->l_ost_gen = cpu_to_le32(0);
+ slot->l_ost_idx = cpu_to_le32(ost_idx);
+
+ if (le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE) {
+ struct lov_ost_data_v1 *objs;
+ int i;
+ __u16 count;
+
+ count = le16_to_cpu(lmm->lmm_stripe_count);
+ if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1)
+ objs = &lmm->lmm_objects[0];
+ else
+ objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
+ for (i = 0; i < count; i++, objs++) {
+ if (objs != slot && lovea_slot_is_dummy(objs))
+ break;
+ }
+
+ /* If the @slot is the last dummy slot to be refilled,
+ * then drop LOV_PATTERN_F_HOLE from lmm::lmm_pattern. */
+ if (i == count)
+ lmm->lmm_pattern &= ~cpu_to_le32(LOV_PATTERN_F_HOLE);
+ }
+
+ rc = dt_xattr_set(env, parent, buf, XATTR_NAME_LOV, fl, handle,
+ BYPASS_CAPA);
+ if (rc == 0)
+ rc = 1;
+
+ return rc;
+}
+
+/**
+ * \retval +1: repaired
+ * \retval 0: did nothing
+ * \retval -ve: on error
+ */
+static int lfsck_layout_extend_lovea(const struct lu_env *env,
+ struct lfsck_instance *lfsck,
+ struct thandle *handle,
+ struct dt_object *parent,
+ struct lu_fid *cfid,
+ struct lu_buf *buf, int fl,
+ __u32 ost_idx, __u32 ea_off, bool reset)
+{
+ struct lov_mds_md_v1 *lmm = buf->lb_buf;
+ struct lov_ost_data_v1 *objs;
+ int rc;
+ __u16 count;
+ bool hole = false;
+ ENTRY;
+
+ if (fl == LU_XATTR_CREATE || reset) {
+ __u32 pattern = LOV_PATTERN_RAID0;
+
+ count = ea_off + 1;
+ LASSERT(buf->lb_len == lov_mds_md_size(count, LOV_MAGIC_V1));
+
+ if (ea_off != 0 || reset) {
+ pattern |= LOV_PATTERN_F_HOLE;
+ hole = true;
+ }
+
+ memset(lmm, 0, buf->lb_len);
+ lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
+ lmm->lmm_pattern = cpu_to_le32(pattern);
+ fid_to_lmm_oi(lfsck_dto2fid(parent), &lmm->lmm_oi);
+ lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
+
+ rc = lfsck_layout_get_def_stripesize(env, lfsck,
+ &lmm->lmm_stripe_size);
+ if (rc != 0)
+ RETURN(rc);
+
+ objs = &lmm->lmm_objects[ea_off];
+ } else {
+ __u32 magic = le32_to_cpu(lmm->lmm_magic);
+ int gap;
+
+ count = le16_to_cpu(lmm->lmm_stripe_count);
+ if (magic == LOV_MAGIC_V1)
+ objs = &lmm->lmm_objects[count];
+ else
+ objs = &((struct lov_mds_md_v3 *)lmm)->
+ lmm_objects[count];
+
+ gap = ea_off - count;
+ if (gap >= 0)
+ count = ea_off + 1;
+ LASSERT(buf->lb_len == lov_mds_md_size(count, magic));
+
+ if (gap > 0) {
+ memset(objs, 0, gap * sizeof(*objs));
+ lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_HOLE);
+ hole = true;
+ }
+
+ lmm->lmm_layout_gen =
+ cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
+ objs += gap;
+ }
+
+ lmm->lmm_stripe_count = cpu_to_le16(count);
+ rc = lfsck_layout_refill_lovea(env, handle, parent, cfid, buf, objs,
+ fl, ost_idx);
+
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant extend layout EA for "
+ DFID": parent "DFID", OST-index %u, stripe-index %u, fl %d, "
+ "reset %s, %s LOV EA hole: rc = %d\n",
+ lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
+ ost_idx, ea_off, fl, reset ? "yes" : "no",
+ hole ? "with" : "without", rc);
+
+ RETURN(rc);
+}
+
+/**
+ * \retval +1: repaired
+ * \retval 0: did nothing
+ * \retval -ve: on error
+ */
+static int lfsck_layout_update_pfid(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct dt_object *parent,
+ struct lu_fid *cfid,
+ struct dt_device *cdev, __u32 ea_off)
+{
+ struct filter_fid *pfid = &lfsck_env_info(env)->lti_new_pfid;
+ struct dt_object *child;
+ struct thandle *handle;
+ const struct lu_fid *tfid = lu_object_fid(&parent->do_lu);
+ struct lu_buf *buf;
+ int rc = 0;
+ ENTRY;
+
+ child = lfsck_object_find_by_dev(env, cdev, cfid);
+ if (IS_ERR(child))
+ RETURN(PTR_ERR(child));
+
+ handle = dt_trans_create(env, cdev);
+ if (IS_ERR(handle))
+ GOTO(out, rc = PTR_ERR(handle));
+
+ pfid->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
+ pfid->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
+ /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
+ * MDT-object's FID::f_ver, instead it is the OST-object index in its
+ * parent MDT-object's layout EA. */
+ pfid->ff_parent.f_stripe_idx = cpu_to_le32(ea_off);
+ buf = lfsck_buf_get(env, pfid, sizeof(struct filter_fid));
+
+ rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start(env, cdev, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle,
+ BYPASS_CAPA);
+
+ GOTO(stop, rc = (rc == 0 ? 1 : rc));
+
+stop:
+ dt_trans_stop(env, cdev, handle);
+
+out:
+ lu_object_put(env, &child->do_lu);
+
+ return rc;
+}
+
+/**
+ * This function will create the MDT-object with the given (partial) LOV EA.
+ *
+ * Under some data corruption cases, the MDT-object of the file may be lost,
+ * but its OST-objects, or some of them are there. The layout LFSCK needs to
+ * re-create the MDT-object with the orphan OST-object(s) information.
+ *
+ * On the other hand, the LFSCK may has created some OST-object for repairing
+ * dangling LOV EA reference, but as the LFSCK processing, it may find that
+ * the old OST-object is there and should replace the former new created OST
+ * object. Unfortunately, some others have modified such newly created object.
+ * To keep the data (both new and old), the LFSCK will create MDT-object with
+ * new FID to reference the original OST-object.
+ *
+ * \param[in] env pointer to the thread context
+ * \param[in] com pointer to the lfsck component
+ * \param[in] ltd pointer to target device descriptor
+ * \param[in] rec pointer to the record for the orphan OST-object
+ * \param[in] cfid pointer to FID for the orphan OST-object
+ * \param[in] infix additional information, such as the FID for original
+ * MDT-object and the stripe offset in the LOV EA
+ * \param[in] type the type for describing why the orphan MDT-object is
+ * created. The rules are as following:
+ *
+ * type "C": Multiple OST-objects claim the same MDT-object and the
+ * same slot in the layout EA. Then the LFSCK will create
+ * new MDT-object(s) to hold the conflict OST-object(s).
+ *
+ * type "N": The orphan OST-object does not know which one was the
+ * real parent MDT-object, so the LFSCK uses new FID for
+ * its parent MDT-object.
+ *
+ * type "R": The orphan OST-object knows its parent MDT-object FID,
+ * but does not know the position (the file name) in the
+ * namespace.
+ *
+ * The orphan name will be like:
+ * ${FID}-${infix}-${type}-${conflict_version}
+ *
+ * \param[in] ea_off the stripe offset in the LOV EA
+ *
+ * \retval positive on repaired something
+ * \retval 0 if needs to repair nothing
+ * \retval negative error number on failure
+ */
+static int lfsck_layout_recreate_parent(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_tgt_desc *ltd,
+ struct lu_orphan_rec *rec,
+ struct lu_fid *cfid,
+ const char *infix,
+ const char *type,
+ __u32 ea_off)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct dt_insert_rec *dtrec = &info->lti_dt_rec;
+ char *name = info->lti_key;
+ struct lu_attr *la = &info->lti_la;
+ struct dt_object_format *dof = &info->lti_dof;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct lu_fid *pfid = &rec->lor_fid;
+ struct lu_fid *tfid = &info->lti_fid3;
+ struct dt_device *next = lfsck->li_next;
+ struct dt_object *pobj = NULL;
+ struct dt_object *cobj = NULL;
+ struct thandle *th = NULL;
+ struct lu_buf *pbuf = NULL;
+ struct lu_buf *ea_buf = &info->lti_big_buf;
+ struct lustre_handle lh = { 0 };
+ struct linkea_data ldata = { 0 };
+ struct lu_buf linkea_buf;
+ const struct lu_name *pname;
+ int buflen = ea_buf->lb_len;
+ int idx = 0;
+ int rc = 0;
+ ENTRY;
+
+ /* Create .lustre/lost+found/MDTxxxx when needed. */
+ if (unlikely(lfsck->li_lpf_obj == NULL)) {
+ rc = lfsck_create_lpf(env, lfsck);
+ if (rc != 0)
+ GOTO(log, rc);
+ }
+
+ if (fid_is_zero(pfid)) {
+ struct filter_fid *ff = &info->lti_new_pfid;
+
+ rc = lfsck_fid_alloc(env, lfsck, pfid, false);
+ if (rc != 0)
+ RETURN(rc);
+
+ ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
+ ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
+ /* Currently, the filter_fid::ff_parent::f_ver is not the
+ * real parent MDT-object's FID::f_ver, instead it is the
+ * OST-object index in its parent MDT-object's layout EA. */
+ ff->ff_parent.f_stripe_idx = cpu_to_le32(ea_off);
+ pbuf = lfsck_buf_get(env, ff, sizeof(struct filter_fid));
+ cobj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
+ if (IS_ERR(cobj))
+ GOTO(log, rc = PTR_ERR(cobj));
+ }
+
+ pobj = lfsck_object_find_by_dev(env, lfsck->li_bottom, pfid);
+ if (IS_ERR(pobj))
+ GOTO(put, rc = PTR_ERR(pobj));
+
+ LASSERT(infix != NULL);
+ LASSERT(type != NULL);
+
+ do {
+ snprintf(name, NAME_MAX, DFID"%s-%s-%d", PFID(pfid), infix,
+ type, idx++);
+ rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
+ (const struct dt_key *)name, BYPASS_CAPA);
+ if (rc != 0 && rc != -ENOENT)
+ GOTO(put, rc);
+ } while (rc == 0);
+
+ rc = linkea_data_new(&ldata,
+ &lfsck_env_info(env)->lti_linkea_buf);
+ if (rc != 0)
+ GOTO(put, rc);
+
+ pname = lfsck_name_get_const(env, name, strlen(name));
+ rc = linkea_add_buf(&ldata, pname, lfsck_dto2fid(lfsck->li_lpf_obj));
+ if (rc != 0)
+ GOTO(put, rc);
+
+ memset(la, 0, sizeof(*la));
+ la->la_uid = rec->lor_uid;
+ la->la_gid = rec->lor_gid;
+ la->la_mode = S_IFREG | S_IRUSR;
+ la->la_valid = LA_MODE | LA_UID | LA_GID;
+
+ memset(dof, 0, sizeof(*dof));
+ dof->dof_type = dt_mode_to_dft(S_IFREG);
+
+ rc = lov_mds_md_size(ea_off + 1, LOV_MAGIC_V1);
+ if (buflen < rc) {
+ lu_buf_realloc(ea_buf, rc);
+ buflen = ea_buf->lb_len;
+ if (ea_buf->lb_buf == NULL)
+ GOTO(put, rc = -ENOMEM);
+ } else {
+ ea_buf->lb_len = rc;
+ }
+
+ /* Hold update lock on the .lustre/lost+found/MDTxxxx/.
+ *
+ * XXX: Currently, we do not grab the PDO lock as normal create cases,
+ * because creating MDT-object for orphan OST-object is rare, we
+ * do not much care about the performance. It can be improved in
+ * the future when needed. */
+ rc = lfsck_layout_lock(env, com, lfsck->li_lpf_obj, &lh,
+ MDS_INODELOCK_UPDATE);
+ if (rc != 0)
+ GOTO(put, rc);
+
+ th = dt_trans_create(env, next);
+ if (IS_ERR(th))
+ GOTO(unlock, rc = PTR_ERR(th));
+
+ /* 1a. Update OST-object's parent information remotely.
+ *
+ * If other subsequent modifications failed, then next LFSCK scanning
+ * will process the OST-object as orphan again with known parent FID. */
+ if (cobj != NULL) {
+ rc = dt_declare_xattr_set(env, cobj, pbuf, XATTR_NAME_FID, 0, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+ }
+
+ /* 2a. Create the MDT-object locally. */
+ rc = dt_declare_create(env, pobj, la, NULL, dof, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ /* 3a. Add layout EA for the MDT-object. */
+ rc = dt_declare_xattr_set(env, pobj, ea_buf, XATTR_NAME_LOV,
+ LU_XATTR_CREATE, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ /* 4a. Insert the MDT-object to .lustre/lost+found/MDTxxxx/ */
+ dtrec->rec_fid = pfid;
+ dtrec->rec_type = S_IFREG;
+ rc = dt_declare_insert(env, lfsck->li_lpf_obj,
+ (const struct dt_rec *)dtrec,
+ (const struct dt_key *)name, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ /* 5a. insert linkEA for parent. */
+ linkea_buf.lb_buf = ldata.ld_buf->lb_buf;
+ linkea_buf.lb_len = ldata.ld_leh->leh_len;
+ rc = dt_declare_xattr_set(env, pobj, &linkea_buf,
+ XATTR_NAME_LINK, 0, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start(env, next, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ /* 1b. Update OST-object's parent information remotely. */
+ if (cobj != NULL) {
+ rc = dt_xattr_set(env, cobj, pbuf, XATTR_NAME_FID, 0, th,
+ BYPASS_CAPA);
+ if (rc != 0)
+ GOTO(stop, rc);
+ }
+
+ dt_write_lock(env, pobj, 0);
+ /* 2b. Create the MDT-object locally. */
+ rc = dt_create(env, pobj, la, NULL, dof, th);
+ if (rc == 0)
+ /* 3b. Add layout EA for the MDT-object. */
+ rc = lfsck_layout_extend_lovea(env, lfsck, th, pobj, cfid,
+ ea_buf, LU_XATTR_CREATE,
+ ltd->ltd_index, ea_off, false);
+ dt_write_unlock(env, pobj);
+ if (rc < 0)
+ GOTO(stop, rc);
+
+ /* 4b. Insert the MDT-object to .lustre/lost+found/MDTxxxx/ */
+ rc = dt_insert(env, lfsck->li_lpf_obj, (const struct dt_rec *)dtrec,
+ (const struct dt_key *)name, th, BYPASS_CAPA, 1);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ /* 5b. insert linkEA for parent. */
+ rc = dt_xattr_set(env, pobj, &linkea_buf,
+ XATTR_NAME_LINK, 0, th, BYPASS_CAPA);
+
+ GOTO(stop, rc);
+
+stop:
+ dt_trans_stop(env, next, th);
+
+unlock:
+ lfsck_layout_unlock(&lh);
+
+put:
+ if (cobj != NULL && !IS_ERR(cobj))
+ lu_object_put(env, &cobj->do_lu);
+ if (pobj != NULL && !IS_ERR(pobj))
+ lu_object_put(env, &pobj->do_lu);
+ ea_buf->lb_len = buflen;
+
+log:
+ if (rc < 0)
+ CDEBUG(D_LFSCK, "%s layout LFSCK assistant failed to "
+ "recreate the lost MDT-object: parent "DFID
+ ", child "DFID", OST-index %u, stripe-index %u, "
+ "infix %s, type %s: rc = %d\n",
+ lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
+ ltd->ltd_index, ea_off, infix, type, rc);
+
+ return rc >= 0 ? 1 : rc;
+}
+
+static int lfsck_layout_master_conditional_destroy(const struct lu_env *env,
+ struct lfsck_component *com,
+ const struct lu_fid *fid,
+ __u32 index)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lfsck_request *lr = &info->lti_lr;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct lfsck_tgt_desc *ltd;
+ struct ptlrpc_request *req;
+ struct lfsck_request *tmp;
+ struct obd_export *exp;
+ int rc = 0;
+ ENTRY;
+
+ ltd = lfsck_tgt_get(&lfsck->li_ost_descs, index);
+ if (unlikely(ltd == NULL))
+ RETURN(-ENXIO);
+
+ exp = ltd->ltd_exp;
+ if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
+ GOTO(put, rc = -EOPNOTSUPP);
+
+ req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
+ if (req == NULL)
+ GOTO(put, rc = -ENOMEM);
+
+ rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
+ if (rc != 0) {
+ ptlrpc_request_free(req);
+
+ GOTO(put, rc);
+ }
+
+ memset(lr, 0, sizeof(*lr));
+ lr->lr_event = LE_CONDITIONAL_DESTROY;
+ lr->lr_active = LFSCK_TYPE_LAYOUT;
+ lr->lr_fid = *fid;
+
+ tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
+ *tmp = *lr;
+ ptlrpc_request_set_replen(req);
+
+ rc = ptlrpc_queue_wait(req);
+ ptlrpc_req_finished(req);
+
+ GOTO(put, rc);
+
+put:
+ lfsck_tgt_put(ltd);
+
+ return rc;
+}
+
+static int lfsck_layout_slave_conditional_destroy(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_request *lr)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_attr *la = &info->lti_la;
+ ldlm_policy_data_t *policy = &info->lti_policy;
+ struct ldlm_res_id *resid = &info->lti_resid;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct dt_device *dev = lfsck->li_bottom;
+ struct lu_fid *fid = &lr->lr_fid;
+ struct dt_object *obj;
+ struct thandle *th = NULL;
+ struct lustre_handle lh = { 0 };
+ __u64 flags = 0;
+ int rc = 0;
+ ENTRY;
+
+ obj = lfsck_object_find_by_dev(env, dev, fid);
+ if (IS_ERR(obj))
+ RETURN(PTR_ERR(obj));
+
+ dt_read_lock(env, obj, 0);
+ if (dt_object_exists(obj) == 0) {
+ dt_read_unlock(env, obj);
+
+ GOTO(put, rc = -ENOENT);
+ }
+
+ /* Get obj's attr without lock firstly. */
+ rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
+ dt_read_unlock(env, obj);
+ if (rc != 0)
+ GOTO(put, rc);
+
+ if (likely(la->la_ctime != 0 || la->la_mode & S_ISUID))
+ GOTO(put, rc = -ETXTBSY);
+
+ /* Acquire extent lock on [0, EOF] to sync with all possible written. */
+ LASSERT(lfsck->li_namespace != NULL);
+
+ memset(policy, 0, sizeof(*policy));
+ policy->l_extent.end = OBD_OBJECT_EOF;
+ ost_fid_build_resid(fid, resid);
+ rc = ldlm_cli_enqueue_local(lfsck->li_namespace, resid, LDLM_EXTENT,
+ policy, LCK_EX, &flags, ldlm_blocking_ast,
+ ldlm_completion_ast, NULL, NULL, 0,
+ LVB_T_NONE, NULL, &lh);
+ if (rc != ELDLM_OK)
+ GOTO(put, rc = -EIO);
+
+ dt_write_lock(env, obj, 0);
+ /* Get obj's attr within lock again. */
+ rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
+ if (rc != 0)
+ GOTO(unlock, rc);
+
+ if (la->la_ctime != 0)
+ GOTO(unlock, rc = -ETXTBSY);
+
+ th = dt_trans_create(env, dev);
+ if (IS_ERR(th))
+ GOTO(unlock, rc = PTR_ERR(th));
+
+ rc = dt_declare_ref_del(env, obj, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_declare_destroy(env, obj, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start_local(env, dev, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_ref_del(env, obj, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_destroy(env, obj, th);
+ if (rc == 0)
+ CDEBUG(D_LFSCK, "%s: layout LFSCK destroyed the empty "
+ "OST-object "DFID" that was created for reparing "
+ "dangling referenced case. But the original missed "
+ "OST-object is found now.\n",
+ lfsck_lfsck2name(lfsck), PFID(fid));
+
+ GOTO(stop, rc);
+
+stop:
+ dt_trans_stop(env, dev, th);
+
+unlock:
+ dt_write_unlock(env, obj);
+ ldlm_lock_decref(&lh, LCK_EX);
+
+put:
+ lu_object_put(env, &obj->do_lu);
+
+ return rc;
+}
+
+/**
+ * Some OST-object has occupied the specified layout EA slot.
+ * Such OST-object may be generated by the LFSCK when repair
+ * dangling referenced MDT-object, which can be indicated by
+ * attr::la_ctime == 0 but without S_ISUID in la_mode. If it
+ * is true and such OST-object has not been modified yet, we
+ * will replace it with the orphan OST-object; otherwise the
+ * LFSCK will create new MDT-object to reference the orphan.
+ *
+ * \retval +1: repaired
+ * \retval 0: did nothing
+ * \retval -ve: on error
+ */
+static int lfsck_layout_conflict_create(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_tgt_desc *ltd,
+ struct lu_orphan_rec *rec,
+ struct dt_object *parent,
+ struct lu_fid *cfid,
+ struct lu_buf *ea_buf,
+ struct lov_ost_data_v1 *slot,
+ __u32 ea_off, __u32 ori_len)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_fid *cfid2 = &info->lti_fid2;
+ struct ost_id *oi = &info->lti_oi;
+ char *infix = info->lti_tmpbuf;
+ struct lov_mds_md_v1 *lmm = ea_buf->lb_buf;
+ struct dt_device *dev = com->lc_lfsck->li_bottom;
+ struct thandle *th = NULL;
+ struct lustre_handle lh = { 0 };
+ __u32 ost_idx2 = le32_to_cpu(slot->l_ost_idx);
+ int rc = 0;
+ ENTRY;
+
+ ostid_le_to_cpu(&slot->l_ost_oi, oi);
+ ostid_to_fid(cfid2, oi, ost_idx2);
+
+ /* Hold layout lock on the parent to prevent others to access. */
+ rc = lfsck_layout_lock(env, com, parent, &lh,
+ MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR);
+ if (rc != 0)
+ GOTO(out, rc);
+
+ rc = lfsck_layout_master_conditional_destroy(env, com, cfid2, ost_idx2);
+
+ /* If the conflict OST-obejct is not created for fixing dangling
+ * referenced MDT-object in former LFSCK check/repair, or it has
+ * been modified by others, then we cannot destroy it. Re-create
+ * a new MDT-object for the orphan OST-object. */
+ if (rc == -ETXTBSY) {
+ /* No need the layout lock on the original parent. */
+ lfsck_layout_unlock(&lh);
+ ea_buf->lb_len = ori_len;
+
+ fid_zero(&rec->lor_fid);
+ snprintf(infix, LFSCK_TMPBUF_LEN, "-"DFID"-%x",
+ PFID(lu_object_fid(&parent->do_lu)), ea_off);
+ rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
+ infix, "C", ea_off);
+
+ RETURN(rc);
+ }
+
+ if (rc != 0 && rc != -ENOENT)
+ GOTO(unlock, rc);
+
+ th = dt_trans_create(env, dev);
+ if (IS_ERR(th))
+ GOTO(unlock, rc = PTR_ERR(th));
+
+ rc = dt_declare_xattr_set(env, parent, ea_buf, XATTR_NAME_LOV,
+ LU_XATTR_REPLACE, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start_local(env, dev, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ dt_write_lock(env, parent, 0);
+ lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
+ rc = lfsck_layout_refill_lovea(env, th, parent, cfid, ea_buf, slot,
+ LU_XATTR_REPLACE, ltd->ltd_index);
+ dt_write_unlock(env, parent);
+
+ GOTO(stop, rc);
+
+stop:
+ dt_trans_stop(env, dev, th);
+
+unlock:
+ lfsck_layout_unlock(&lh);
+
+out:
+ ea_buf->lb_len = ori_len;
+
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant replaced the conflict "
+ "OST-object "DFID" on the OST %x with the orphan "DFID" on "
+ "the OST %x: parent "DFID", stripe-index %u: rc = %d\n",
+ lfsck_lfsck2name(com->lc_lfsck), PFID(cfid2), ost_idx2,
+ PFID(cfid), ltd->ltd_index, PFID(lfsck_dto2fid(parent)),
+ ea_off, rc);
+
+ return rc >= 0 ? 1 : rc;
+}
+
+/**
+ * \retval +1: repaired
+ * \retval 0: did nothing
+ * \retval -ve: on error
+ */
+static int lfsck_layout_recreate_lovea(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_tgt_desc *ltd,
+ struct lu_orphan_rec *rec,
+ struct dt_object *parent,
+ struct lu_fid *cfid,
+ __u32 ost_idx, __u32 ea_off)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_buf *buf = &info->lti_big_buf;
+ struct lu_fid *fid = &info->lti_fid2;
+ struct ost_id *oi = &info->lti_oi;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct dt_device *dt = lfsck->li_bottom;
+ struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
+ struct thandle *handle = NULL;
+ size_t buflen = buf->lb_len;
+ size_t lovea_size;
+ struct lov_mds_md_v1 *lmm;
+ struct lov_ost_data_v1 *objs;
+ struct lustre_handle lh = { 0 };
+ __u32 magic;
+ int fl = 0;
+ int rc = 0;
+ int rc1;
+ int i;
+ __u16 count;
+ bool locked = false;
+ ENTRY;
+
+ rc = lfsck_layout_lock(env, com, parent, &lh,
+ MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR);
+ if (rc != 0) {
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant failed to recreate "
+ "LOV EA for "DFID": parent "DFID", OST-index %u, "
+ "stripe-index %u: rc = %d\n",
+ lfsck_lfsck2name(lfsck), PFID(cfid),
+ PFID(lfsck_dto2fid(parent)), ost_idx, ea_off, rc);
+
+ RETURN(rc);
+ }
+
+again:
+ if (locked) {
+ dt_write_unlock(env, parent);
+ locked = false;
+ }
+
+ if (handle != NULL) {
+ dt_trans_stop(env, dt, handle);
+ handle = NULL;
+ }
+
+ if (rc < 0)
+ GOTO(unlock_layout, rc);
+
+ lovea_size = rc;
+ if (buf->lb_len < lovea_size) {
+ lu_buf_realloc(buf, lovea_size);
+ buflen = buf->lb_len;
+ if (buf->lb_buf == NULL)
+ GOTO(unlock_layout, rc = -ENOMEM);
+ }
+
+ if (!(bk->lb_param & LPF_DRYRUN)) {
+ handle = dt_trans_create(env, dt);
+ if (IS_ERR(handle))
+ GOTO(unlock_layout, rc = PTR_ERR(handle));
+
+ rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
+ fl, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start_local(env, dt, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+ }
+
+ dt_write_lock(env, parent, 0);
+ locked = true;
+ rc = dt_xattr_get(env, parent, buf, XATTR_NAME_LOV, BYPASS_CAPA);
+ if (rc == -ERANGE) {
+ rc = dt_xattr_get(env, parent, &LU_BUF_NULL, XATTR_NAME_LOV,
+ BYPASS_CAPA);
+ LASSERT(rc != 0);
+ goto again;
+ } else if (rc == -ENODATA || rc == 0) {
+ lovea_size = lov_mds_md_size(ea_off + 1, LOV_MAGIC_V1);
+ /* If the declared is not big enough, re-try. */
+ if (buf->lb_len < lovea_size) {
+ rc = lovea_size;
+ goto again;
+ }
+ fl = LU_XATTR_CREATE;
+ } else if (rc < 0) {
+ GOTO(unlock_parent, rc);
+ } else if (unlikely(buf->lb_len == 0)) {
+ goto again;
+ } else {
+ fl = LU_XATTR_REPLACE;
+ lovea_size = rc;
+ }
+
+ if (fl == LU_XATTR_CREATE) {
+ if (bk->lb_param & LPF_DRYRUN)
+ GOTO(unlock_parent, rc = 1);
+
+ LASSERT(buf->lb_len >= lovea_size);
+
+ buf->lb_len = lovea_size;
+ rc = lfsck_layout_extend_lovea(env, lfsck, handle, parent, cfid,
+ buf, fl, ost_idx, ea_off, false);
+
+ GOTO(unlock_parent, rc);
+ }
+
+ lmm = buf->lb_buf;
+ rc1 = lfsck_layout_verify_header(lmm);
+
+ /* If the LOV EA crashed, the rebuild it. */
+ if (rc1 == -EINVAL) {
+ if (bk->lb_param & LPF_DRYRUN)
+ GOTO(unlock_parent, rc = 1);
+
+ LASSERT(buf->lb_len >= lovea_size);
+
+ buf->lb_len = lovea_size;
+ memset(lmm, 0, buf->lb_len);
+ rc = lfsck_layout_extend_lovea(env, lfsck, handle, parent, cfid,
+ buf, fl, ost_idx, ea_off, true);
+
+ GOTO(unlock_parent, rc);
+ }
+
+ /* For other unknown magic/pattern, keep the current LOV EA. */
+ if (rc1 != 0)
+ GOTO(unlock_parent, rc = rc1);
+
+ /* Currently, we only support LOV_MAGIC_V1/LOV_MAGIC_V3 which has
+ * been verified in lfsck_layout_verify_header() already. If some
+ * new magic introduced in the future, then layout LFSCK needs to
+ * be updated also. */
+ magic = le32_to_cpu(lmm->lmm_magic);
+ if (magic == LOV_MAGIC_V1) {
+ objs = &lmm->lmm_objects[0];
+ } else {
+ LASSERT(magic == LOV_MAGIC_V3);
+ objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
+ }
+
+ count = le16_to_cpu(lmm->lmm_stripe_count);
+ if (count == 0)
+ GOTO(unlock_parent, rc = -EINVAL);
+ LASSERT(count > 0);
+
+ /* Exceed the current end of MDT-object layout EA. Then extend it. */
+ if (count <= ea_off) {
+ if (bk->lb_param & LPF_DRYRUN)
+ GOTO(unlock_parent, rc = 1);
+
+ lovea_size = lov_mds_md_size(ea_off + 1, magic);
+ /* If the declared is not big enough, re-try. */
+ if (buf->lb_len < lovea_size) {
+ rc = lovea_size;
+ goto again;
+ }
+ buf->lb_len = lovea_size;
+ rc = lfsck_layout_extend_lovea(env, lfsck, handle, parent, cfid,
+ buf, fl, ost_idx, ea_off, false);
+
+ GOTO(unlock_parent, rc);
+ }
+
+ LASSERTF(rc > 0, "invalid rc = %d\n", rc);
+
+ buf->lb_len = lovea_size;
+ for (i = 0; i < count; i++, objs++) {
+ /* The MDT-object was created via lfsck_layout_recover_create()
+ * by others before, and we fill the dummy layout EA. */
+ if (lovea_slot_is_dummy(objs)) {
+ if (i != ea_off)
+ continue;
+
+ if (bk->lb_param & LPF_DRYRUN)
+ GOTO(unlock_parent, rc = 1);
+
+ lmm->lmm_layout_gen =
+ cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
+ rc = lfsck_layout_refill_lovea(env, handle, parent,
+ cfid, buf, objs, fl,
+ ost_idx);
+
+ CDEBUG(D_LFSCK, "%s layout LFSCK assistant fill "
+ "dummy layout slot for "DFID": parent "DFID
+ ", OST-index %u, stripe-index %u: rc = %d\n",
+ lfsck_lfsck2name(lfsck), PFID(cfid),
+ PFID(lfsck_dto2fid(parent)), ost_idx, i, rc);
+
+ GOTO(unlock_parent, rc);
+ }
+
+ ostid_le_to_cpu(&objs->l_ost_oi, oi);
+ ostid_to_fid(fid, oi, le32_to_cpu(objs->l_ost_idx));
+ /* It should be rare case, the slot is there, but the LFSCK
+ * does not handle it during the first-phase cycle scanning. */
+ if (unlikely(lu_fid_eq(fid, cfid))) {
+ if (i == ea_off) {
+ GOTO(unlock_parent, rc = 0);
+ } else {
+ /* Rare case that the OST-object index
+ * does not match the parent MDT-object
+ * layout EA. We trust the later one. */
+ if (bk->lb_param & LPF_DRYRUN)
+ GOTO(unlock_parent, rc = 1);
+
+ dt_write_unlock(env, parent);
+ if (handle != NULL)
+ dt_trans_stop(env, dt, handle);
+ lfsck_layout_unlock(&lh);
+ buf->lb_len = buflen;
+ rc = lfsck_layout_update_pfid(env, com, parent,
+ cfid, ltd->ltd_tgt, i);
+
+ CDEBUG(D_LFSCK, "%s layout LFSCK assistant "
+ "updated OST-object's pfid for "DFID
+ ": parent "DFID", OST-index %u, "
+ "stripe-index %u: rc = %d\n",
+ lfsck_lfsck2name(lfsck), PFID(cfid),
+ PFID(lfsck_dto2fid(parent)),
+ ltd->ltd_index, i, rc);
+
+ RETURN(rc);
+ }
+ }
+ }
+
+ /* The MDT-object exists, but related layout EA slot is occupied
+ * by others. */
+ if (bk->lb_param & LPF_DRYRUN)
+ GOTO(unlock_parent, rc = 1);
+
+ dt_write_unlock(env, parent);
+ if (handle != NULL)
+ dt_trans_stop(env, dt, handle);
+ lfsck_layout_unlock(&lh);
+ if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1)
+ objs = &lmm->lmm_objects[ea_off];
+ else
+ objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[ea_off];
+ rc = lfsck_layout_conflict_create(env, com, ltd, rec, parent, cfid,
+ buf, objs, ea_off, buflen);
+
+ RETURN(rc);
+
+unlock_parent:
+ if (locked)
+ dt_write_unlock(env, parent);
+
+stop:
+ if (handle != NULL)
+ dt_trans_stop(env, dt, handle);
+
+unlock_layout:
+ lfsck_layout_unlock(&lh);
+ buf->lb_len = buflen;
+
+ return rc;
+}
+
+static int lfsck_layout_scan_orphan_one(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_tgt_desc *ltd,
+ struct lu_orphan_rec *rec,
+ struct lu_fid *cfid)
+{
+ struct lfsck_layout *lo = com->lc_file_ram;
+ struct lu_fid *pfid = &rec->lor_fid;
+ struct dt_object *parent = NULL;
+ __u32 ea_off = pfid->f_stripe_idx;
+ int rc = 0;
+ ENTRY;
+
+ if (!fid_is_sane(cfid))
+ GOTO(out, rc = -EINVAL);
+
+ if (fid_is_zero(pfid)) {
+ rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
+ "", "N", ea_off);
+ GOTO(out, rc);
+ }
+
+ pfid->f_ver = 0;
+ if (!fid_is_sane(pfid))
+ GOTO(out, rc = -EINVAL);
+
+ parent = lfsck_object_find_by_dev(env, com->lc_lfsck->li_bottom, pfid);
+ if (IS_ERR(parent))
+ GOTO(out, rc = PTR_ERR(parent));
+
+ if (unlikely(dt_object_remote(parent) != 0))
+ GOTO(put, rc = -EXDEV);
+
+ if (dt_object_exists(parent) == 0) {
+ lu_object_put(env, &parent->do_lu);
+ rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
+ "", "R", ea_off);
+ GOTO(out, rc);
+ }
+
+ if (!S_ISREG(lu_object_attr(&parent->do_lu)))
+ GOTO(put, rc = -EISDIR);
+
+ rc = lfsck_layout_recreate_lovea(env, com, ltd, rec, parent, cfid,
+ ltd->ltd_index, ea_off);
+
+ GOTO(put, rc);
+
+put:
+ if (rc <= 0)
+ lu_object_put(env, &parent->do_lu);
+ else
+ /* The layout EA is changed, need to be reloaded next time. */
+ lu_object_put_nocache(env, &parent->do_lu);
+
+out:
+ down_write(&com->lc_sem);
+ com->lc_new_scanned++;
+ com->lc_new_checked++;
+ if (rc > 0) {
+ lo->ll_objs_repaired[LLIT_ORPHAN - 1]++;
+ rc = 0;
+ } else if (rc < 0) {
+ lo->ll_objs_failed_phase2++;
+ }
+ up_write(&com->lc_sem);
+
+ return rc;
+}
+
+static int lfsck_layout_scan_orphan(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_tgt_desc *ltd)
+{
+ struct lfsck_layout *lo = com->lc_file_ram;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct ost_id *oi = &info->lti_oi;
+ struct lu_fid *fid = &info->lti_fid;
+ struct dt_object *obj;
+ const struct dt_it_ops *iops;
+ struct dt_it *di;
+ int rc = 0;
+ ENTRY;
+
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant starts the orphan "
+ "scanning for OST%04x\n",
+ lfsck_lfsck2name(lfsck), ltd->ltd_index);
+
+ ostid_set_seq(oi, FID_SEQ_IDIF);
+ ostid_set_id(oi, 0);
+ ostid_to_fid(fid, oi, ltd->ltd_index);
+ obj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, fid);
+ if (unlikely(IS_ERR(obj)))
+ GOTO(log, rc = PTR_ERR(obj));
+
+ rc = obj->do_ops->do_index_try(env, obj, &dt_lfsck_orphan_features);
+ if (rc != 0)
+ GOTO(put, rc);
+
+ iops = &obj->do_index_ops->dio_it;
+ di = iops->init(env, obj, 0, BYPASS_CAPA);
+ if (IS_ERR(di))
+ GOTO(put, rc = PTR_ERR(di));
+
+ rc = iops->load(env, di, 0);
+ if (rc == -ESRCH) {
+ /* -ESRCH means that the orphan OST-objects rbtree has been
+ * cleanup because of the OSS server restart or other errors. */
+ lo->ll_flags |= LF_INCOMPLETE;
+ GOTO(fini, rc);
+ }
+
+ if (rc == 0)
+ rc = iops->next(env, di);
+ else if (rc > 0)
+ rc = 0;
+
+ if (rc < 0)
+ GOTO(fini, rc);
+
+ if (rc > 0)
+ GOTO(fini, rc = 0);
+
+ do {
+ struct dt_key *key;
+ struct lu_orphan_rec *rec = &info->lti_rec;
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY3) &&
+ cfs_fail_val > 0) {
+ struct ptlrpc_thread *thread = &lfsck->li_thread;
+ struct l_wait_info lwi;
+
+ lwi = LWI_TIMEOUT(cfs_time_seconds(cfs_fail_val),
+ NULL, NULL);
+ l_wait_event(thread->t_ctl_waitq,
+ !thread_is_running(thread),
+ &lwi);
+ }
+
+ key = iops->key(env, di);
+ com->lc_fid_latest_scanned_phase2 = *(struct lu_fid *)key;
+ rc = iops->rec(env, di, (struct dt_rec *)rec, 0);
+ if (rc == 0)
+ rc = lfsck_layout_scan_orphan_one(env, com, ltd, rec,
+ &com->lc_fid_latest_scanned_phase2);
+ if (rc != 0 && bk->lb_param & LPF_FAILOUT)
+ GOTO(fini, rc);
+
+ lfsck_control_speed_by_self(com);
+ do {
+ rc = iops->next(env, di);
+ } while (rc < 0 && !(bk->lb_param & LPF_FAILOUT));
+ } while (rc == 0);
+
+ GOTO(fini, rc);
+
+fini:
+ iops->put(env, di);
+ iops->fini(env, di);
+put:
+ lu_object_put(env, &obj->do_lu);
+
+log:
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant finished the orphan "
+ "scanning for OST%04x: rc = %d\n",
+ lfsck_lfsck2name(lfsck), ltd->ltd_index, rc);
+
+ return rc > 0 ? 0 : rc;
+}
+
+/* For the MDT-object with dangling reference, we need to repare the
+ * inconsistency according to the LFSCK sponsor's requirement:
+ *
+ * 1) Keep the inconsistency there and report the inconsistency case,
+ * then give the chance to the application to find related issues,
+ * and the users can make the decision about how to handle it with
+ * more human knownledge. (by default)
+ *
+ * 2) Re-create the missed OST-object with the FID/owner information. */
+static int lfsck_layout_repair_dangling(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_layout_req *llr,
+ const struct lu_attr *pla)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct filter_fid *pfid = &info->lti_new_pfid;
+ struct dt_allocation_hint *hint = &info->lti_hint;
+ struct lu_attr *cla = &info->lti_la2;
+ struct dt_object *parent = llr->llr_parent->llo_obj;
+ struct dt_object *child = llr->llr_child;
+ struct dt_device *dev = lfsck_obj2dt_dev(child);
+ const struct lu_fid *tfid = lu_object_fid(&parent->do_lu);
+ struct thandle *handle;
+ struct lu_buf *buf;
+ struct lustre_handle lh = { 0 };
+ int rc;
+ bool create;
+ ENTRY;
+
+ if (com->lc_lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ)
+ create = true;
+ else
+ create = false;
+
+ if (!create)
+ GOTO(log, rc = 1);
+
+ memset(cla, 0, sizeof(*cla));
+ cla->la_uid = pla->la_uid;
+ cla->la_gid = pla->la_gid;
+ cla->la_mode = S_IFREG | 0666;
+ cla->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
+ LA_ATIME | LA_MTIME | LA_CTIME;
+
+ rc = lfsck_layout_lock(env, com, parent, &lh,
+ MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR);
+ if (rc != 0)
+ GOTO(log, rc);
+
+ handle = dt_trans_create(env, dev);
+ if (IS_ERR(handle))
+ GOTO(unlock1, rc = PTR_ERR(handle));
+
+ hint->dah_parent = NULL;
+ hint->dah_mode = 0;
+ pfid->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
+ pfid->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
+ /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
+ * MDT-object's FID::f_ver, instead it is the OST-object index in its
+ * parent MDT-object's layout EA. */
+ pfid->ff_parent.f_stripe_idx = cpu_to_le32(llr->llr_lov_idx);
+ buf = lfsck_buf_get(env, pfid, sizeof(struct filter_fid));
+
+ rc = dt_declare_create(env, child, cla, hint, NULL, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID,
+ LU_XATTR_CREATE, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start(env, dev, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ dt_read_lock(env, parent, 0);
+ if (unlikely(lu_object_is_dying(parent->do_lu.lo_header)))
+ GOTO(unlock2, rc = 1);
+
+ rc = dt_create(env, child, cla, hint, NULL, handle);
+ if (rc != 0)
+ GOTO(unlock2, rc);
+
+ rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, LU_XATTR_CREATE,
+ handle, BYPASS_CAPA);
+
+ GOTO(unlock2, rc);
+
+unlock2:
+ dt_read_unlock(env, parent);
+
+stop:
+ rc = lfsck_layout_trans_stop(env, dev, handle, rc);
+
+unlock1:
+ lfsck_layout_unlock(&lh);
+
+log:
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found dangling "
+ "reference for: parent "DFID", child "DFID", OST-index %u, "
+ "stripe-index %u, owner %u/%u. %s: rc = %d\n",
+ lfsck_lfsck2name(com->lc_lfsck), PFID(lfsck_dto2fid(parent)),
+ PFID(lfsck_dto2fid(child)), llr->llr_ost_idx,
+ llr->llr_lov_idx, pla->la_uid, pla->la_gid,
+ create ? "Create the lost OST-object as required" :
+ "Keep the MDT-object there by default", rc);
+
+ return rc;
+}
+
+/* If the OST-object does not recognize the MDT-object as its parent, and
+ * there is no other MDT-object claims as its parent, then just trust the
+ * given MDT-object as its parent. So update the OST-object filter_fid. */
+static int lfsck_layout_repair_unmatched_pair(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_layout_req *llr,
+ const struct lu_attr *pla)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct filter_fid *pfid = &info->lti_new_pfid;
+ struct lu_attr *tla = &info->lti_la3;
+ struct dt_object *parent = llr->llr_parent->llo_obj;
+ struct dt_object *child = llr->llr_child;
+ struct dt_device *dev = lfsck_obj2dt_dev(child);
+ const struct lu_fid *tfid = lu_object_fid(&parent->do_lu);
+ struct thandle *handle;
+ struct lu_buf *buf;
+ struct lustre_handle lh = { 0 };
+ int rc;
+ ENTRY;
+
+ rc = lfsck_layout_lock(env, com, parent, &lh,
+ MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR);
+ if (rc != 0)
+ GOTO(log, rc);
+
+ handle = dt_trans_create(env, dev);
+ if (IS_ERR(handle))
+ GOTO(unlock1, rc = PTR_ERR(handle));
+
+ pfid->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
+ pfid->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
+ /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
+ * MDT-object's FID::f_ver, instead it is the OST-object index in its
+ * parent MDT-object's layout EA. */
+ pfid->ff_parent.f_stripe_idx = cpu_to_le32(llr->llr_lov_idx);
+ buf = lfsck_buf_get(env, pfid, sizeof(struct filter_fid));
+
+ rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ tla->la_valid = LA_UID | LA_GID;
+ tla->la_uid = pla->la_uid;
+ tla->la_gid = pla->la_gid;
+ rc = dt_declare_attr_set(env, child, tla, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start(env, dev, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ dt_write_lock(env, parent, 0);
+ if (unlikely(lu_object_is_dying(parent->do_lu.lo_header)))
+ GOTO(unlock2, rc = 1);
+
+ rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle,
+ BYPASS_CAPA);
+ if (rc != 0)
+ GOTO(unlock2, rc);
+
+ /* Get the latest parent's owner. */
+ rc = dt_attr_get(env, parent, tla, BYPASS_CAPA);
+ if (rc != 0)
+ GOTO(unlock2, rc);
+
+ tla->la_valid = LA_UID | LA_GID;
+ rc = dt_attr_set(env, child, tla, handle, BYPASS_CAPA);
+
+ GOTO(unlock2, rc);
+
+unlock2:
+ dt_write_unlock(env, parent);
+
+stop:
+ rc = lfsck_layout_trans_stop(env, dev, handle, rc);
+
+unlock1:
+ lfsck_layout_unlock(&lh);
+
+log:
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired unmatched "
+ "MDT-OST pair for: parent "DFID", child "DFID", OST-index %u, "
+ "stripe-index %u, owner %u/%u: rc = %d\n",
+ lfsck_lfsck2name(com->lc_lfsck), PFID(lfsck_dto2fid(parent)),
+ PFID(lfsck_dto2fid(child)), llr->llr_ost_idx, llr->llr_lov_idx,
+ pla->la_uid, pla->la_gid, rc);
+
+ return rc;
+}
+
+/* If there are more than one MDT-objects claim as the OST-object's parent,
+ * and the OST-object only recognizes one of them, then we need to generate
+ * new OST-object(s) with new fid(s) for the non-recognized MDT-object(s). */
+static int lfsck_layout_repair_multiple_references(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_layout_req *llr,
+ struct lu_attr *la,
+ struct lu_buf *buf)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct dt_allocation_hint *hint = &info->lti_hint;
+ struct dt_object_format *dof = &info->lti_dof;
+ struct dt_device *pdev = com->lc_lfsck->li_next;
+ struct ost_id *oi = &info->lti_oi;
+ struct dt_object *parent = llr->llr_parent->llo_obj;
+ struct dt_device *cdev = lfsck_obj2dt_dev(llr->llr_child);
+ struct dt_object *child = NULL;
+ struct lu_device *d = &cdev->dd_lu_dev;
+ struct lu_object *o = NULL;
+ struct thandle *handle;
+ struct lov_mds_md_v1 *lmm;
+ struct lov_ost_data_v1 *objs;
+ struct lustre_handle lh = { 0 };
+ __u32 magic;
+ int rc;
+ ENTRY;
+
+ rc = lfsck_layout_lock(env, com, parent, &lh,
+ MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR);
+ if (rc != 0)
+ GOTO(log, rc);
+
+ handle = dt_trans_create(env, pdev);
+ if (IS_ERR(handle))
+ GOTO(unlock1, rc = PTR_ERR(handle));
+
+ o = lu_object_anon(env, d, NULL);
+ if (IS_ERR(o))
+ GOTO(stop, rc = PTR_ERR(o));
+
+ child = container_of(o, struct dt_object, do_lu);
+ o = lu_object_locate(o->lo_header, d->ld_type);
+ if (unlikely(o == NULL))
+ GOTO(stop, rc = -EINVAL);
+
+ child = container_of(o, struct dt_object, do_lu);
+ la->la_valid = LA_UID | LA_GID;
+ hint->dah_parent = NULL;
+ hint->dah_mode = 0;
+ dof->dof_type = DFT_REGULAR;
+ rc = dt_declare_create(env, child, la, NULL, NULL, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
+ LU_XATTR_REPLACE, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start(env, pdev, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ dt_write_lock(env, parent, 0);
+ if (unlikely(lu_object_is_dying(parent->do_lu.lo_header)))
+ GOTO(unlock2, rc = 0);
+
+ rc = dt_xattr_get(env, parent, buf, XATTR_NAME_LOV, BYPASS_CAPA);
+ if (unlikely(rc == 0 || rc == -ENODATA || rc == -ERANGE))
+ GOTO(unlock2, rc = 0);
+
+ lmm = buf->lb_buf;
+ /* Someone change layout during the LFSCK, no need to repair then. */
+ if (le16_to_cpu(lmm->lmm_layout_gen) != llr->llr_parent->llo_gen)
+ GOTO(unlock2, rc = 0);
+
+ rc = dt_create(env, child, la, hint, dof, handle);
+ if (rc != 0)
+ GOTO(unlock2, rc);
+
+ /* Currently, we only support LOV_MAGIC_V1/LOV_MAGIC_V3 which has
+ * been verified in lfsck_layout_verify_header() already. If some
+ * new magic introduced in the future, then layout LFSCK needs to
+ * be updated also. */
+ magic = le32_to_cpu(lmm->lmm_magic);
+ if (magic == LOV_MAGIC_V1) {
+ objs = &lmm->lmm_objects[0];
+ } else {
+ LASSERT(magic == LOV_MAGIC_V3);
+ objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
+ }
+
+ lmm->lmm_layout_gen = cpu_to_le16(llr->llr_parent->llo_gen + 1);
+ fid_to_ostid(lu_object_fid(&child->do_lu), oi);
+ ostid_cpu_to_le(oi, &objs[llr->llr_lov_idx].l_ost_oi);
+ objs[llr->llr_lov_idx].l_ost_gen = cpu_to_le32(0);
+ objs[llr->llr_lov_idx].l_ost_idx = cpu_to_le32(llr->llr_ost_idx);
+ rc = dt_xattr_set(env, parent, buf, XATTR_NAME_LOV,
+ LU_XATTR_REPLACE, handle, BYPASS_CAPA);
+
+ GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
+
+unlock2:
+ dt_write_unlock(env, parent);
+
+stop:
+ if (child != NULL)
+ lu_object_put(env, &child->do_lu);
+
+ dt_trans_stop(env, pdev, handle);
+
+unlock1:
+ lfsck_layout_unlock(&lh);
+
+log:
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired multiple "
+ "references for: parent "DFID", OST-index %u, stripe-index %u, "
+ "owner %u/%u: rc = %d\n",
+ lfsck_lfsck2name(com->lc_lfsck), PFID(lfsck_dto2fid(parent)),
+ llr->llr_ost_idx, llr->llr_lov_idx, la->la_uid, la->la_gid, rc);
+
+ return rc;
+}
+
+/* If the MDT-object and the OST-object have different owner information,
+ * then trust the MDT-object, because the normal chown/chgrp handle order
+ * is from MDT to OST, and it is possible that some chown/chgrp operation
+ * is partly done. */
+static int lfsck_layout_repair_owner(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct lfsck_layout_req *llr,
+ struct lu_attr *pla)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_attr *tla = &info->lti_la3;
+ struct dt_object *parent = llr->llr_parent->llo_obj;
+ struct dt_object *child = llr->llr_child;
+ struct dt_device *dev = lfsck_obj2dt_dev(child);
+ struct thandle *handle;
+ int rc;
+ ENTRY;
+
+ handle = dt_trans_create(env, dev);
+ if (IS_ERR(handle))
+ GOTO(log, rc = PTR_ERR(handle));
+
+ tla->la_uid = pla->la_uid;
+ tla->la_gid = pla->la_gid;
+ tla->la_valid = LA_UID | LA_GID;
+ rc = dt_declare_attr_set(env, child, tla, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start(env, dev, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ /* Use the dt_object lock to serialize with destroy and attr_set. */
+ dt_read_lock(env, parent, 0);
+ if (unlikely(lu_object_is_dying(parent->do_lu.lo_header)))
+ GOTO(unlock, rc = 1);
+
+ /* Get the latest parent's owner. */
+ rc = dt_attr_get(env, parent, tla, BYPASS_CAPA);
+ if (rc != 0)
+ GOTO(unlock, rc);
+
+ /* Some others chown/chgrp during the LFSCK, needs to do nothing. */
+ if (unlikely(tla->la_uid != pla->la_uid ||
+ tla->la_gid != pla->la_gid))
+ GOTO(unlock, rc = 1);
+
+ tla->la_valid = LA_UID | LA_GID;
+ rc = dt_attr_set(env, child, tla, handle, BYPASS_CAPA);
+
+ GOTO(unlock, rc);
+
+unlock:
+ dt_read_unlock(env, parent);
+
+stop:
+ rc = lfsck_layout_trans_stop(env, dev, handle, rc);
+
+log:
+ CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired inconsistent "
+ "file owner for: parent "DFID", child "DFID", OST-index %u, "
+ "stripe-index %u, owner %u/%u: rc = %d\n",
+ lfsck_lfsck2name(com->lc_lfsck), PFID(lfsck_dto2fid(parent)),
+ PFID(lfsck_dto2fid(child)), llr->llr_ost_idx, llr->llr_lov_idx,
+ pla->la_uid, pla->la_gid, rc);
+
+ return rc;
+}
+
+/* Check whether the OST-object correctly back points to the
+ * MDT-object (@parent) via the XATTR_NAME_FID xattr (@pfid). */
+static int lfsck_layout_check_parent(const struct lu_env *env,
+ struct lfsck_component *com,
+ struct dt_object *parent,
+ const struct lu_fid *pfid,
+ const struct lu_fid *cfid,
+ const struct lu_attr *pla,
+ const struct lu_attr *cla,
+ struct lfsck_layout_req *llr,
+ struct lu_buf *lov_ea, __u32 idx)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_buf *buf = &info->lti_big_buf;
+ struct dt_object *tobj;
+ struct lov_mds_md_v1 *lmm;
+ struct lov_ost_data_v1 *objs;
+ int rc;
+ int i;
+ __u32 magic;
+ __u16 count;
+ ENTRY;
+
+ if (fid_is_zero(pfid)) {
+ /* client never wrote. */
+ if (cla->la_size == 0 && cla->la_blocks == 0) {
+ if (unlikely(cla->la_uid != pla->la_uid ||
+ cla->la_gid != pla->la_gid))
+ RETURN (LLIT_INCONSISTENT_OWNER);
+
+ RETURN(0);
+ }
+
+ RETURN(LLIT_UNMATCHED_PAIR);
+ }
+
+ if (unlikely(!fid_is_sane(pfid)))
+ RETURN(LLIT_UNMATCHED_PAIR);
+
+ if (lu_fid_eq(pfid, lu_object_fid(&parent->do_lu))) {
+ if (llr->llr_lov_idx == idx)
+ RETURN(0);
+
+ RETURN(LLIT_UNMATCHED_PAIR);
+ }
+
+ tobj = lfsck_object_find(env, com->lc_lfsck, pfid);
+ if (tobj == NULL)
+ RETURN(LLIT_UNMATCHED_PAIR);
+
+ if (IS_ERR(tobj))
+ RETURN(PTR_ERR(tobj));
+
+ if (!dt_object_exists(tobj))
+ GOTO(out, rc = LLIT_UNMATCHED_PAIR);
+
+ /* Load the tobj's layout EA, in spite of it is a local MDT-object or
+ * remote one on another MDT. Then check whether the given OST-object
+ * is in such layout. If yes, it is multiple referenced, otherwise it
+ * is unmatched referenced case. */
+ rc = lfsck_layout_get_lovea(env, tobj, buf, NULL);
+ if (rc == 0)
+ GOTO(out, rc = LLIT_UNMATCHED_PAIR);
+
+ if (rc < 0)
+ GOTO(out, rc);
+
+ lmm = buf->lb_buf;
+ magic = le32_to_cpu(lmm->lmm_magic);
+ if (magic == LOV_MAGIC_V1) {
+ objs = &lmm->lmm_objects[0];
+ } else {
+ LASSERT(magic == LOV_MAGIC_V3);
+ objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
+ }
+
+ count = le16_to_cpu(lmm->lmm_stripe_count);
+ for (i = 0; i < count; i++, objs++) {
+ struct lu_fid *tfid = &info->lti_fid2;
+ struct ost_id *oi = &info->lti_oi;
+
+ if (lovea_slot_is_dummy(objs))
+ continue;
+
+ ostid_le_to_cpu(&objs->l_ost_oi, oi);
+ ostid_to_fid(tfid, oi, le32_to_cpu(objs->l_ost_idx));
+ if (lu_fid_eq(cfid, tfid)) {
+ *lov_ea = *buf;
+
+ GOTO(out, rc = LLIT_MULTIPLE_REFERENCED);