void mdd_read_lock(const struct lu_env *env, struct mdd_object *obj,
enum mdd_object_role role)
{
struct dt_object *next = mdd_object_child(obj);
next->do_ops->do_read_lock(env, next, role);
}
/*
* To enable DNE functionality we need FID of /ROOT directory
* (which is / as seen by the clients) to belong to MDT0 and
* not to FID_SEQ_LOCAL_FILE or some other local sequence,
* which can be used by any node, so can't be part of FLDB.
*
* Pre-production code was using FID_SEQ_LOCAL_FILE for /ROOT
* making few existing setups incompatibile with DNE. This
* applies to ZFS-based setups only as ldiskfs-based setups
* are still using IGIF to identify /ROOT.
*
* The intention of this code is to fix on-disk state to use
* FID_SEQ_ROOT for /ROOT:
* - "." and ".." references in /ROOT itself and it`s subdirectories
* - LinkEA in all the objects listed in /ROOT
*
* Given only ZFS is affected where "." and ".." are not stored, we need to:
* - delete "." and ".." from /ROOT and its subdirectories
* - rename references in LinkEA in all the objects listed in /ROOT
*
* This code is subject for removal in 2.5
*/
static int mdd_convert_remove_dots(const struct lu_env *env,
struct mdd_device *mdd,
struct mdd_object *o)
{
struct thandle *th;
const struct dt_key *dot = (const struct dt_key *)".";
const struct dt_key *dotdot = (const struct dt_key *)"..";
int rc;
if (dt_try_as_dir(env, mdd_object_child(o)) == 0)
RETURN(-ENOTDIR);
/* remove "."/".." and do not insert them back - not stored in ZFS */
th = dt_trans_create(env, mdd->mdd_child);
if (IS_ERR(th))
RETURN(PTR_ERR(th));
rc = dt_declare_delete(env, mdd_object_child(o), dot, th);
if (rc)
GOTO(out, rc);
rc = dt_declare_delete(env, mdd_object_child(o), dotdot, th);
if (rc)
GOTO(out, rc);
rc = dt_trans_start_local(env, mdd->mdd_child, th);
if (rc)
GOTO(out, rc);
/* ignore non-existing "."/".." - we stored them on disk for
* pre-production systems, but this is not how regular ZFS works */
rc = dt_delete(env, mdd_object_child(o), dot, th, BYPASS_CAPA);
if (rc == -ENOENT)
rc = 0;
if (rc)
GOTO(out, rc);
rc = dt_delete(env, mdd_object_child(o), dotdot, th, BYPASS_CAPA);
if (rc == -ENOENT)
rc = 0;
if (rc)
GOTO(out, rc);
out:
dt_trans_stop(env, mdd->mdd_child, th);
RETURN(rc);
}
static int orph_index_insert(const struct lu_env *env,
struct mdd_object *obj,
__u32 op,
struct thandle *th)
{
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
struct dt_object *dor = mdd->mdd_orphans;
const struct lu_fid *lf_dor = lu_object_fid(&dor->do_lu);
struct dt_object *next = mdd_object_child(obj);
int rc;
ENTRY;
LASSERT(mdd_write_locked(env, obj) != 0);
LASSERT(!(obj->mod_flags & ORPHAN_OBJ));
mdd_orphan_write_lock(env, mdd);
rc = mdd_orphan_insert_obj(env, mdd, obj, op, th);
if (rc)
GOTO(out, rc);
mdo_ref_add(env, obj, th);
if (!S_ISDIR(mdd_object_type(obj)))
goto out;
mdo_ref_add(env, obj, th);
mdd_orphan_ref_add(env, mdd, th);
/* try best to fixup directory, dont return errors
* from here */
if (!dt_try_as_dir(env, next))
goto out;
next->do_index_ops->dio_delete(env, next,
(const struct dt_key *)dotdot,
th, BYPASS_CAPA);
next->do_index_ops->dio_insert(env, next,
(struct dt_rec *)lf_dor,
(const struct dt_key *)dotdot,
th, BYPASS_CAPA, 1);
out:
if (rc == 0)
obj->mod_flags |= ORPHAN_OBJ;
mdd_orphan_write_unlock(env, mdd);
RETURN(rc);
}
/**
* add an orphan \a obj to the orphan index.
* \param obj file or directory.
* \param th transaction for index insert.
*
* \pre obj nlink == 0 && obj->mod_count != 0
*
* \retval 0 success
* \retval -ve index operation error.
*/
int mdd_orphan_insert(const struct lu_env *env, struct mdd_object *obj,
struct thandle *th)
{
struct mdd_device *mdd = mdo2mdd(&obj->mod_obj);
struct dt_object *dor = mdd->mdd_orphans;
const struct lu_fid *lf_dor = lu_object_fid(&dor->do_lu);
struct dt_object *next = mdd_object_child(obj);
struct dt_insert_rec *rec = &mdd_env_info(env)->mti_dt_rec;
int rc;
ENTRY;
LASSERT(mdd_write_locked(env, obj) != 0);
LASSERT(!(obj->mod_flags & ORPHAN_OBJ));
dt_write_lock(env, mdd->mdd_orphans, MOR_TGT_ORPHAN);
rc = mdd_orphan_insert_obj(env, mdd, obj, th);
if (rc)
GOTO(out, rc);
mdo_ref_add(env, obj, th);
if (!S_ISDIR(mdd_object_type(obj)))
GOTO(out, rc = 0);
mdo_ref_add(env, obj, th);
dt_ref_add(env, mdd->mdd_orphans, th);
/* try best to fixup directory, do not return errors from here */
if (!dt_try_as_dir(env, next))
GOTO(out, rc = 0);
dt_delete(env, next, (const struct dt_key *)dotdot, th);
rec->rec_fid = lf_dor;
rec->rec_type = S_IFDIR;
dt_insert(env, next, (const struct dt_rec *)rec,
(const struct dt_key *)dotdot, th);
out:
if (rc == 0)
obj->mod_flags |= ORPHAN_OBJ;
dt_write_unlock(env, mdd->mdd_orphans);
RETURN(rc);
}
int mdd_compat_fixes(const struct lu_env *env, struct mdd_device *mdd)
{
struct mdd_thread_info *info = mdd_env_info(env);
struct mdd_object *root;
struct dt_object *o;
struct lustre_mdt_attrs *lma;
struct lu_buf buf;
int rc;
ENTRY;
/* IGIF FIDS are valid for old 1.8 and 2.[123] ROOT and are kept.
* Normal FIDs used by Xyratex 1.8->2.1 upgrade tool are also kept. */
if (fid_is_igif(&mdd->mdd_root_fid) || fid_is_norm(&mdd->mdd_root_fid))
RETURN(0);
/*
* FID is supposed to be FID_SEQ_ROOT for:
* - new ldiskfs fs
* - new ZFS fs
* - old ZFS fs, by now processed with osd_convert_root_to_new_seq()
*/
if (fid_seq(&mdd->mdd_root_fid) != FID_SEQ_ROOT) {
CERROR("%s: wrong FID "DFID" is used for /ROOT\n",
mdd2obd_dev(mdd)->obd_name,
PFID(&mdd->mdd_root_fid));
RETURN(-EINVAL);
}
root = mdd_object_find(env, mdd, &mdd->mdd_root_fid);
if (IS_ERR(root))
RETURN(PTR_ERR(root));
o = mdd_object_child(root);
CDEBUG(D_OTHER, "/ROOT = "DFID"\n", PFID(&mdd->mdd_root_fid));
if (dt_try_as_dir(env, o) == 0) {
CERROR("%s: not a directory\n", mdd2obd_dev(mdd)->obd_name);
GOTO(out, rc = -ENOTDIR);
}
lma = (struct lustre_mdt_attrs *)&info->mti_xattr_buf;
CLASSERT(sizeof(info->mti_xattr_buf) >= LMA_OLD_SIZE);
buf.lb_len = LMA_OLD_SIZE;
buf.lb_buf = lma;
rc = mdo_xattr_get(env, root, &buf, XATTR_NAME_LMA, BYPASS_CAPA);
if (rc < 0 && rc != -ENODATA) {
CERROR("%s: can't fetch LMA: rc = %d\n",
mdd2obd_dev(mdd)->obd_name, rc);
GOTO(out, rc);
}
lustre_lma_swab(lma);
if (lu_fid_eq(&lma->lma_self_fid, &mdd->mdd_root_fid)) {
/* /ROOT has been converted already
* or was correct from the beginning */
CDEBUG(D_OTHER, "%s: converted already\n",
mdd2obd_dev(mdd)->obd_name);
GOTO(out, rc = 0);
}
/* this is supposed to happen only on pre-production ZFS backend */
if (strcmp(mdd->mdd_bottom->dd_lu_dev.ld_type->ldt_name,
LUSTRE_OSD_ZFS_NAME) != 0) {
CERROR("%s: "DFID" is used on ldiskfs?!\n",
mdd2obd_dev(mdd)->obd_name, PFID(&mdd->mdd_root_fid));
GOTO(out, rc = -ENOTSUPP);
}
LCONSOLE_INFO("%s: FID of /ROOT has been changed. "
"Please remount the clients.\n",
mdd2obd_dev(mdd)->obd_name);
/* Fill FLDB first */
rc = mdd_fill_fldb(env, mdd);
if (rc)
GOTO(out, rc);
/* remove ./.. from /ROOT */
rc = mdd_convert_remove_dots(env, mdd, root);
if (rc)
GOTO(out, rc);
/* go over the directory, fix all the objects */
rc = mdd_fix_children(env, mdd, o);
if (rc)
GOTO(out, rc);
/* Update LMA on /ROOT. Done for simplicity in MDD, not in osd-zfs.
* Correct LMA will imply the whole directory has been coverted
* successfully, otherwise it will be retried on next mount. */
rc = mdd_convert_lma(env, mdd, root);
out:
mdd_object_put(env, root);
RETURN(rc);
}
int mdd_write_locked(const struct lu_env *env, struct mdd_object *obj)
{
struct dt_object *next = mdd_object_child(obj);
return next->do_ops->do_write_locked(env, next);
}
void mdd_read_unlock(const struct lu_env *env, struct mdd_object *obj)
{
struct dt_object *next = mdd_object_child(obj);
next->do_ops->do_read_unlock(env, next);
}
