return rc;
}
+static int lfsck_lmv_set(const struct lu_env *env,
+ struct lfsck_instance *lfsck,
+ struct dt_object *obj,
+ struct lmv_mds_md_v1 *lmv)
+{
+ struct dt_device *dev = lfsck->li_next;
+ struct thandle *th = NULL;
+ struct lu_buf buf = { lmv, sizeof(*lmv) };
+ int rc;
+
+ ENTRY;
+
+ th = dt_trans_create(env, dev);
+ if (IS_ERR(th))
+ RETURN(PTR_ERR(th));
+
+ rc = dt_declare_xattr_set(env, obj, &buf, XATTR_NAME_LMV".set", 0, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start_local(env, dev, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_xattr_set(env, obj, &buf, XATTR_NAME_LMV".set", 0, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ EXIT;
+stop:
+ dt_trans_stop(env, dev, th);
+
+ return rc;
+}
+
+static int lfsck_lmv_delete(const struct lu_env *env,
+ struct lfsck_instance *lfsck,
+ struct dt_object *obj)
+{
+ struct dt_device *dev = lfsck->li_next;
+ struct thandle *th = NULL;
+ int rc;
+
+ ENTRY;
+
+ th = dt_trans_create(env, dev);
+ if (IS_ERR(th))
+ RETURN(PTR_ERR(th));
+
+ rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LMV, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = dt_trans_start_local(env, dev, th);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ rc = dt_xattr_del(env, obj, XATTR_NAME_LMV, th);
+ if (rc)
+ GOTO(stop, rc);
+
+ EXIT;
+stop:
+ dt_trans_stop(env, dev, th);
+
+ return rc;
+}
+
+static inline int lfsck_object_is_shard(const struct lu_env *env,
+ struct lfsck_instance *lfsck,
+ struct dt_object *obj,
+ const struct lu_name *lname)
+{
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lmv_mds_md_v1 *lmv = &info->lti_lmv;
+ int rc;
+
+ rc = lfsck_shard_name_to_index(env, lname->ln_name, lname->ln_namelen,
+ lfsck_object_type(obj),
+ lfsck_dto2fid(obj));
+ if (rc < 0)
+ return 0;
+
+ rc = lfsck_read_stripe_lmv(env, lfsck, obj, lmv);
+ if (rc == -ENODATA)
+ return 0;
+
+ if (!rc && lmv->lmv_magic == LMV_MAGIC_STRIPE)
+ return 1;
+
+ return rc;
+}
+
/**
* Add the specified name entry back to namespace.
*
* it is quite possible that the name entry is lost. Then the LFSCK
* should add the name entry back to the namespace.
*
+ * If \a child is shard, which means \a parent is a striped directory,
+ * if \a parent has LMV, we need to delete it before insertion because
+ * now parent's striping is broken and can't be parsed correctly.
+ *
* \param[in] env pointer to the thread context
* \param[in] com pointer to the lfsck component
* \param[in] parent pointer to the directory under which the name entry
* will be inserted into
* \param[in] child pointer to the object referenced by the name entry
* that to be inserted into the parent
- * \param[in] name the name for the child in the parent directory
+ * \param[in] lname the name for the child in the parent directory
*
* \retval positive number for repaired cases
* \retval 0 if nothing to be repaired
struct lfsck_component *com,
struct dt_object *parent,
struct dt_object *child,
- const char *name)
+ const struct lu_name *lname)
{
- struct lfsck_thread_info *info = lfsck_env_info(env);
- struct lu_attr *la = &info->lti_la;
- struct dt_insert_rec *rec = &info->lti_dt_rec;
- struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_attr *la = &info->lti_la;
+ struct dt_insert_rec *rec = &info->lti_dt_rec;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
/* The child and its name may be on different MDTs. */
- const struct lu_fid *pfid = lfsck_dto2fid(parent);
- const struct lu_fid *cfid = lfsck_dto2fid(child);
- struct dt_device *dev = lfsck->li_next;
- struct thandle *th = NULL;
- struct lfsck_lock_handle *llh = &info->lti_llh;
- int rc = 0;
+ const struct lu_fid *pfid = lfsck_dto2fid(parent);
+ const struct lu_fid *cfid = lfsck_dto2fid(child);
+ struct dt_device *dev = lfsck->li_next;
+ struct thandle *th = NULL;
+ struct lfsck_lock_handle *llh = &info->lti_llh;
+ struct lmv_mds_md_v1 *lmv = &info->lti_lmv;
+ struct lu_buf buf = { lmv, sizeof(*lmv) };
+ /* whether parent's LMV is deleted before insertion */
+ bool parent_lmv_deleted = false;
+ /* whether parent's LMV is missing */
+ bool parent_lmv_lost = false;
+ int rc = 0;
+
ENTRY;
/* @parent/@child may be based on lfsck->li_bottom,
if (IS_ERR(parent))
GOTO(log, rc = PTR_ERR(parent));
- if (unlikely(!dt_try_as_dir(env, parent)))
- GOTO(log, rc = -ENOTDIR);
-
child = lfsck_object_locate(dev, child);
if (IS_ERR(child))
GOTO(log, rc = PTR_ERR(child));
if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
GOTO(log, rc = 1);
- rc = lfsck_lock(env, lfsck, parent, name, llh,
- MDS_INODELOCK_UPDATE, LCK_PW);
- if (rc != 0)
+ rc = lfsck_lock(env, lfsck, parent, lname->ln_name, llh,
+ MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
+ MDS_INODELOCK_XATTR, LCK_EX);
+ if (rc)
GOTO(log, rc);
+ rc = lfsck_object_is_shard(env, lfsck, child, lname);
+ if (rc < 0)
+ GOTO(unlock, rc);
+
+ if (rc == 1) {
+ rc = lfsck_read_stripe_lmv(env, lfsck, parent, lmv);
+ if (!rc) {
+ /*
+ * To add a shard, we need to convert parent to a
+ * plain directory by deleting its LMV, and after
+ * insertion set it back.
+ */
+ rc = lfsck_lmv_delete(env, lfsck, parent);
+ if (rc)
+ GOTO(unlock, rc);
+ parent_lmv_deleted = true;
+ lmv->lmv_layout_version++;
+ lfsck_lmv_header_cpu_to_le(lmv, lmv);
+ } else if (rc == -ENODATA) {
+ struct lu_seq_range *range = &info->lti_range;
+ struct seq_server_site *ss = lfsck_dev_site(lfsck);
+
+ rc = lfsck_read_stripe_lmv(env, lfsck, child, lmv);
+ if (rc)
+ GOTO(unlock, rc);
+
+ fld_range_set_mdt(range);
+ rc = fld_server_lookup(env, ss->ss_server_fld,
+ fid_seq(lfsck_dto2fid(parent)), range);
+ if (rc)
+ GOTO(unlock, rc);
+
+ parent_lmv_lost = true;
+ lmv->lmv_magic = LMV_MAGIC;
+ lmv->lmv_master_mdt_index = range->lsr_index;
+ lmv->lmv_layout_version++;
+ lfsck_lmv_header_cpu_to_le(lmv, lmv);
+ } else {
+ GOTO(unlock, rc);
+ }
+ }
+
+ if (unlikely(!dt_try_as_dir(env, parent)))
+ GOTO(unlock, rc = -ENOTDIR);
+
th = dt_trans_create(env, dev);
if (IS_ERR(th))
GOTO(unlock, rc = PTR_ERR(th));
rec->rec_type = lfsck_object_type(child) & S_IFMT;
rec->rec_fid = cfid;
rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
- (const struct dt_key *)name, th);
+ (const struct dt_key *)lname->ln_name, th);
if (rc != 0)
GOTO(stop, rc);
GOTO(stop, rc);
}
- memset(la, 0, sizeof(*la));
+ if (parent_lmv_lost) {
+ rc = dt_declare_xattr_set(env, parent, &buf,
+ XATTR_NAME_LMV".set", 0, th);
+ if (rc)
+ GOTO(stop, rc);
+ }
+
la->la_ctime = ktime_get_real_seconds();
la->la_valid = LA_CTIME;
rc = dt_declare_attr_set(env, parent, la, th);
GOTO(stop, rc);
rc = dt_insert(env, parent, (const struct dt_rec *)rec,
- (const struct dt_key *)name, th);
+ (const struct dt_key *)lname->ln_name, th);
if (rc != 0)
GOTO(stop, rc);
GOTO(stop, rc);
}
- la->la_ctime = ktime_get_real_seconds();
+ if (parent_lmv_lost) {
+ rc = dt_xattr_set(env, parent, &buf, XATTR_NAME_LMV".set", 0,
+ th);
+ if (rc)
+ GOTO(stop, rc);
+ }
+
rc = dt_attr_set(env, parent, la, th);
if (rc != 0)
GOTO(stop, rc);
dt_trans_stop(env, dev, th);
unlock:
+ if (parent_lmv_deleted)
+ lfsck_lmv_set(env, lfsck, parent, lmv);
+
lfsck_unlock(llh);
log:
CDEBUG(D_LFSCK, "%s: namespace LFSCK insert object "DFID" with "
"the name %s and type %o to the parent "DFID": rc = %d\n",
- lfsck_lfsck2name(lfsck), PFID(cfid), name,
+ lfsck_lfsck2name(lfsck), PFID(cfid), lname->ln_name,
lfsck_object_type(child) & S_IFMT, PFID(pfid), rc);
if (rc != 0) {
lost_parent:
lmv = &info->lti_lmv;
- rc = lfsck_read_stripe_lmv(env, child, lmv);
+ rc = lfsck_read_stripe_lmv(env, lfsck, child, lmv);
if (rc != 0 && rc != -ENODATA)
GOTO(out, rc);
if (rc >= 0) {
/* Add the missing name entry to the parent. */
rc = lfsck_namespace_insert_normal(env, com, parent,
- child, cname->ln_name);
+ child, cname);
if (unlikely(rc == -EEXIST)) {
/* Unfortunately, someone reused the name
* under the parent by race. So we have
}
lfsck_ibits_unlock(lh, LCK_EX);
- rc = lfsck_namespace_check_name(env, parent, child, cname);
+ rc = lfsck_namespace_check_name(env, lfsck, parent, child,
+ cname);
if (rc == -ENOENT)
goto lost_parent;
/* Add the missing name entry back to the namespace. */
rc = lfsck_namespace_insert_normal(env, com, parent, child,
- cname->ln_name);
+ cname);
if (unlikely(rc == -ESTALE))
/* It may happen when the remote object has been
* removed, but the local MDT is not aware of that. */
/* Add the missing name entry to the parent. */
rc = lfsck_namespace_insert_normal(env, com,
- parent, child, cname->ln_name);
+ parent, child, cname);
if (unlikely(rc == -EEXIST))
/* Unfortunately, someone reused the
* name under the parent by race. So we
GOTO(out, rc = 0);
}
- rc = lfsck_namespace_check_name(env, parent, child, cname);
+ rc = lfsck_namespace_check_name(env, lfsck, parent, child,
+ cname);
if (rc == -ENOENT)
goto lost_parent;
/* Add the missing name entry back to the namespace. */
rc = lfsck_namespace_insert_normal(env, com, parent, child,
- cname->ln_name);
+ cname);
if (unlikely(rc == -ESTALE))
/* It may happen when the remote object has been
* removed, but the local MDT is not aware of that. */
struct dt_object *child,
struct lfsck_namespace_req *lnr)
{
- struct lfsck_thread_info *info = lfsck_env_info(env);
- struct lu_attr *la = &info->lti_la;
- struct dt_allocation_hint *hint = &info->lti_hint;
- struct dt_object_format *dof = &info->lti_dof;
- struct dt_insert_rec *rec = &info->lti_dt_rec;
- struct lmv_mds_md_v1 *lmv2 = &info->lti_lmv2;
- const struct lu_name *cname;
- const struct lu_fid *pfid = lfsck_dto2fid(parent);
- const struct lu_fid *cfid = lfsck_dto2fid(child);
- struct linkea_data ldata = { NULL };
- struct lfsck_lock_handle *llh = &info->lti_llh;
- struct lu_buf linkea_buf;
- struct lu_buf lmv_buf;
- struct lfsck_instance *lfsck = com->lc_lfsck;
- struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
- struct dt_device *dev = lfsck->li_next;
- struct thandle *th = NULL;
- int rc = 0;
- __u16 type = lnr->lnr_type;
- bool create;
+ struct lfsck_thread_info *info = lfsck_env_info(env);
+ struct lu_attr *la = &info->lti_la;
+ struct dt_allocation_hint *hint = &info->lti_hint;
+ struct dt_object_format *dof = &info->lti_dof;
+ struct dt_insert_rec *rec = &info->lti_dt_rec;
+ struct lmv_mds_md_v1 *lmv2 = &info->lti_lmv2;
+ const struct lu_name *cname;
+ const struct lu_fid *pfid = lfsck_dto2fid(parent);
+ const struct lu_fid *cfid = lfsck_dto2fid(child);
+ struct linkea_data ldata = { NULL };
+ struct lfsck_lock_handle *llh = &info->lti_llh;
+ struct lustre_handle rlh = { 0 };
+ struct lustre_handle clh = { 0 };
+ struct lu_buf linkea_buf;
+ struct lu_buf lmv_buf;
+ struct lfsck_instance *lfsck = com->lc_lfsck;
+ struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
+ struct dt_device *dev = lfsck->li_next;
+ struct thandle *th = NULL;
+ int rc = 0;
+ __u16 type = lnr->lnr_type;
+ bool create;
ENTRY;
cname = lfsck_name_get_const(env, lnr->lnr_name, lnr->lnr_namelen);
GOTO(log, rc);
rc = lfsck_lock(env, lfsck, parent, lnr->lnr_name, llh,
- MDS_INODELOCK_UPDATE, LCK_PR);
+ MDS_INODELOCK_UPDATE, LCK_PW);
if (rc != 0)
GOTO(log, rc);
if (rc != 0)
GOTO(log, rc);
+ if (dt_object_remote(child)) {
+ rc = lfsck_remote_lookup_lock(env, lfsck, parent, child, &rlh,
+ LCK_EX);
+ if (rc != 0)
+ GOTO(log, rc);
+ }
+
+ rc = lfsck_ibits_lock(env, lfsck, child, &clh,
+ MDS_INODELOCK_UPDATE | MDS_INODELOCK_LOOKUP |
+ MDS_INODELOCK_XATTR, LCK_EX);
+ if (rc != 0)
+ GOTO(unlock_remote_lookup, rc);
+
/* Set the ctime as zero, then others can know it is created for
* repairing dangling name entry by LFSCK. And if the LFSCK made
* wrong decision and the real MDT-object has been found later,
* then the LFSCK has chance to fix the incosistency properly. */
memset(la, 0, sizeof(*la));
- la->la_mode = (type & S_IFMT) | 0600;
- la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
- LA_ATIME | LA_MTIME | LA_CTIME;
+ if (S_ISDIR(type))
+ la->la_mode = (type & S_IFMT) | 0700;
+ else
+ la->la_mode = (type & S_IFMT) | 0600;
+ la->la_valid = LA_TYPE | LA_MODE | LA_CTIME;
- child->do_ops->do_ah_init(env, hint, parent, child,
- la->la_mode & S_IFMT);
+ /*
+ * if it's directory, skip do_ah_init() to create a plain directory
+ * because it may have shards already, which will be inserted back
+ * later, besides, it may be remote, and creating stripe directory
+ * remotely is not supported.
+ */
+ if (S_ISREG(type))
+ child->do_ops->do_ah_init(env, hint, parent, child, type);
memset(dof, 0, sizeof(*dof));
dof->dof_type = dt_mode_to_dft(type);
th = dt_trans_create(env, dev);
if (IS_ERR(th))
- GOTO(log, rc = PTR_ERR(th));
+ GOTO(unlock_child, rc = PTR_ERR(th));
/* 1a. create child. */
rc = dt_declare_create(env, child, la, hint, dof, th);
lfsck_lmv_header_cpu_to_le(lmv2, lmv2);
lfsck_buf_init(&lmv_buf, lmv2, sizeof(*lmv2));
rc = dt_declare_xattr_set(env, child, &lmv_buf,
- XATTR_NAME_LMV, 0, th);
+ XATTR_NAME_LMV".set", 0, th);
if (rc != 0)
GOTO(stop, rc);
}
if (rc != 0)
GOTO(stop, rc);
+ /* 7a. if child is remote, delete and insert to generate local agent */
+ if (dt_object_remote(child)) {
+ rc = dt_declare_delete(env, parent,
+ (const struct dt_key *)lnr->lnr_name,
+ th);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
+ (const struct dt_key *)lnr->lnr_name,
+ th);
+ if (rc)
+ GOTO(stop, rc);
+ }
+
rc = dt_trans_start_local(env, dev, th);
if (rc != 0)
GOTO(stop, rc = (rc == -EEXIST ? 1 : rc));
/* 5b. generate slave LMV EA. */
if (lnr->lnr_lmv != NULL && lnr->lnr_lmv->ll_lmv_master) {
- rc = dt_xattr_set(env, child, &lmv_buf, XATTR_NAME_LMV,
- 0, th);
+ rc = dt_xattr_set(env, child, &lmv_buf,
+ XATTR_NAME_LMV".set", 0, th);
if (rc != 0)
GOTO(unlock, rc);
}
/* 6b. insert linkEA for child. */
rc = dt_xattr_set(env, child, &linkea_buf,
XATTR_NAME_LINK, 0, th);
+ if (rc)
+ GOTO(unlock, rc);
+
+ /* 7b. if child is remote, delete and insert to generate local agent */
+ if (dt_object_remote(child)) {
+ rc = dt_delete(env, parent,
+ (const struct dt_key *)lnr->lnr_name, th);
+ if (rc)
+ GOTO(unlock, rc);
+
+ rec->rec_type = type;
+ rec->rec_fid = cfid;
+ rc = dt_insert(env, parent, (const struct dt_rec *)rec,
+ (const struct dt_key *)lnr->lnr_name, th);
+ if (rc)
+ GOTO(unlock, rc);
+ }
GOTO(unlock, rc);
stop:
dt_trans_stop(env, dev, th);
+unlock_child:
+ lfsck_ibits_unlock(&clh, LCK_EX);
+unlock_remote_lookup:
+ if (dt_object_remote(child))
+ lfsck_ibits_unlock(&rlh, LCK_EX);
log:
lfsck_unlock(llh);
CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant found dangling "