int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
struct md_attr *ma, int flags)
{
struct mdt_lock_handle *lh;
int do_vbr = ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID|LA_FLAGS);
__u64 lockpart = MDS_INODELOCK_UPDATE;
int rc;
ENTRY;
/* attr shouldn't be set on remote object */
LASSERT(!mdt_object_remote(mo));
lh = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_reg_init(lh, LCK_PW);
/* Even though the new MDT will grant PERM lock to the old
* client, but the old client will almost ignore that during
* So it needs to revoke both LOOKUP and PERM lock here, so
* both new and old client can cancel the dcache */
if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
rc = mdt_object_lock(info, mo, lh, lockpart, MDT_LOCAL_LOCK);
if (rc != 0)
RETURN(rc);
if (mdt_object_exists(mo) == 0)
GOTO(out_unlock, rc = -ENOENT);
/* all attrs are packed into mti_attr in unpack_setattr */
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_SETATTR_WRITE);
/* This is only for set ctime when rename's source is on remote MDS. */
if (unlikely(ma->ma_attr.la_valid == LA_CTIME))
ma->ma_attr_flags |= MDS_VTX_BYPASS;
/* VBR: update version if attr changed are important for recovery */
if (do_vbr) {
/* update on-disk version of changed object */
info->mti_mos = mo;
rc = mdt_version_get_check_save(info, mo, 0);
if (rc)
GOTO(out_unlock, rc);
}
/* all attrs are packed into mti_attr in unpack_setattr */
rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
if (rc != 0)
GOTO(out_unlock, rc);
EXIT;
out_unlock:
mdt_object_unlock(info, mo, lh, rc);
return rc;
}
/**
* Get version of object by fid.
*
* Return real version or ENOENT_VERSION if object doesn't exist
*/
static void mdt_obj_version_get(struct mdt_thread_info *info,
struct mdt_object *o, __u64 *version)
{
LASSERT(o);
if (mdt_object_exists(o) && !mdt_object_remote(o) &&
!mdt_object_obf(o))
*version = dt_version_get(info->mti_env, mdt_obj2dt(o));
else
*version = ENOENT_VERSION;
CDEBUG(D_INODE, "FID "DFID" version is "LPX64"\n",
PFID(mdt_object_fid(o)), *version);
}
static int mdt_lvbo_fill(struct ldlm_lock *lock, void *lvb, int lvblen)
{
struct lu_env env;
struct mdt_thread_info *info;
struct mdt_device *mdt;
struct lu_fid *fid;
struct mdt_object *obj = NULL;
struct md_object *child = NULL;
int rc;
ENTRY;
mdt = ldlm_lock_to_ns(lock)->ns_lvbp;
if (IS_LQUOTA_RES(lock->l_resource)) {
if (mdt->mdt_qmt_dev == NULL)
RETURN(0);
/* call lvbo fill function of quota master */
rc = qmt_hdls.qmth_lvbo_fill(mdt->mdt_qmt_dev, lock, lvb,
lvblen);
RETURN(rc);
}
if (!ldlm_has_layout(lock))
RETURN(0);
/* layout lock will be granted to client, fill in lvb with layout */
/* XXX create an env to talk to mdt stack. We should get this env from
* ptlrpc_thread->t_env. */
rc = lu_env_init(&env, LCT_MD_THREAD);
/* Likely ENOMEM */
if (rc)
RETURN(rc);
info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
/* Likely ENOMEM */
if (info == NULL)
GOTO(out, rc = -ENOMEM);
memset(info, 0, sizeof *info);
info->mti_env = &env;
info->mti_exp = lock->l_export;
info->mti_mdt = mdt;
/* XXX get fid by resource id. why don't include fid in ldlm_resource */
fid = &info->mti_tmp_fid2;
fid_extract_from_res_name(fid, &lock->l_resource->lr_name);
obj = mdt_object_find(&env, info->mti_mdt, fid);
if (IS_ERR(obj))
GOTO(out, rc = PTR_ERR(obj));
if (!mdt_object_exists(obj) || mdt_object_remote(obj))
GOTO(out, rc = -ENOENT);
child = mdt_object_child(obj);
/* get the length of lsm */
rc = mo_xattr_get(&env, child, &LU_BUF_NULL, XATTR_NAME_LOV);
if (rc < 0)
GOTO(out, rc);
if (rc > 0) {
struct lu_buf *lmm = NULL;
if (lvblen < rc) {
CERROR("%s: expected %d actual %d.\n",
mdt_obd_name(mdt), rc, lvblen);
GOTO(out, rc = -ERANGE);
}
lmm = &info->mti_buf;
lmm->lb_buf = lvb;
lmm->lb_len = rc;
rc = mo_xattr_get(&env, child, lmm, XATTR_NAME_LOV);
if (rc < 0)
GOTO(out, rc);
}
out:
if (obj != NULL && !IS_ERR(obj))
mdt_object_put(&env, obj);
lu_env_fini(&env);
RETURN(rc < 0 ? 0 : rc);
}
/* shrink dir layout after migration */
static int mdt_dir_layout_shrink(struct mdt_thread_info *info)
{
const struct lu_env *env = info->mti_env;
struct mdt_device *mdt = info->mti_mdt;
struct lu_ucred *uc = mdt_ucred(info);
struct mdt_reint_record *rr = &info->mti_rr;
struct lmv_user_md *lmu = rr->rr_eadata;
__u32 lum_stripe_count = lmu->lum_stripe_count;
struct lu_buf *buf = &info->mti_buf;
struct lmv_mds_md_v1 *lmv;
struct md_attr *ma = &info->mti_attr;
struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
struct mdt_object *pobj = NULL;
struct mdt_object *obj;
struct mdt_lock_handle *lhp = NULL;
struct mdt_lock_handle *lhc;
int rc;
ENTRY;
if (!mdt->mdt_enable_dir_migration)
RETURN(-EPERM);
if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
mdt->mdt_enable_remote_dir_gid != -1)
RETURN(-EPERM);
/* mti_big_lmm is used to save LMV, but it may be uninitialized. */
if (unlikely(!info->mti_big_lmm)) {
info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
if (!info->mti_big_lmm)
RETURN(-ENOMEM);
}
obj = mdt_object_find(env, mdt, rr->rr_fid1);
if (IS_ERR(obj))
RETURN(PTR_ERR(obj));
/* get parent from PFID */
rc = mdt_attr_get_pfid(info, obj, &ma->ma_pfid);
if (rc)
GOTO(put_obj, rc);
pobj = mdt_object_find(env, mdt, &ma->ma_pfid);
if (IS_ERR(pobj))
GOTO(put_obj, rc = PTR_ERR(pobj));
/* revoke object remote LOOKUP lock */
if (mdt_object_remote(pobj)) {
rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
if (rc)
GOTO(put_pobj, rc);
}
/*
* lock parent if dir will be shrunk to 1 stripe, because dir will be
* converted to normal directory, as will change dir fid and update
* namespace of parent.
*/
lhp = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_reg_init(lhp, LCK_PW);
if (le32_to_cpu(lmu->lum_stripe_count) < 2) {
rc = mdt_reint_object_lock(info, pobj, lhp,
MDS_INODELOCK_UPDATE, true);
if (rc)
GOTO(put_pobj, rc);
}
/* lock object */
lhc = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(lhc, LCK_EX);
rc = mdt_reint_striped_lock(info, obj, lhc, MDS_INODELOCK_FULL, einfo,
true);
if (rc)
GOTO(unlock_pobj, rc);
ma->ma_lmv = info->mti_big_lmm;
ma->ma_lmv_size = info->mti_big_lmmsize;
ma->ma_valid = 0;
rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
if (rc)
GOTO(unlock_obj, rc);
/* user may run 'lfs migrate' multiple times, so it's shrunk already */
if (!(ma->ma_valid & MA_LMV))
GOTO(unlock_obj, rc = -EALREADY);
lmv = &ma->ma_lmv->lmv_md_v1;
/* ditto */
if (!(le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION))
GOTO(unlock_obj, rc = -EALREADY);
lum_stripe_count = lmu->lum_stripe_count;
if (!lum_stripe_count)
lum_stripe_count = cpu_to_le32(1);
if (lmv->lmv_migrate_offset != lum_stripe_count) {
CERROR("%s: "DFID" migrate mdt count mismatch %u != %u\n",
mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
lmv->lmv_migrate_offset, lmu->lum_stripe_count);
GOTO(unlock_obj, rc = -EINVAL);
}
if (lmv->lmv_master_mdt_index != lmu->lum_stripe_offset) {
CERROR("%s: "DFID" migrate mdt index mismatch %u != %u\n",
mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
lmv->lmv_master_mdt_index, lmu->lum_stripe_offset);
GOTO(unlock_obj, rc = -EINVAL);
}
if (lum_stripe_count > 1 &&
(lmv->lmv_hash_type & cpu_to_le32(LMV_HASH_TYPE_MASK)) !=
lmu->lum_hash_type) {
CERROR("%s: "DFID" migrate mdt hash mismatch %u != %u\n",
mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
lmv->lmv_hash_type, lmu->lum_hash_type);
GOTO(unlock_obj, rc = -EINVAL);
}
buf->lb_buf = rr->rr_eadata;
buf->lb_len = rr->rr_eadatalen;
rc = mo_xattr_set(env, mdt_object_child(obj), buf, XATTR_NAME_LMV, 0);
GOTO(unlock_obj, rc);
unlock_obj:
mdt_reint_striped_unlock(info, obj, lhc, einfo, rc);
unlock_pobj:
mdt_object_unlock(info, pobj, lhp, rc);
put_pobj:
mdt_object_put(env, pobj);
put_obj:
mdt_object_put(env, obj);
return rc;
}
/*
* VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
* name.
*/
static int mdt_reint_link(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct md_attr *ma = &info->mti_attr;
struct mdt_object *ms;
struct mdt_object *mp;
struct mdt_lock_handle *lhs;
struct mdt_lock_handle *lhp;
struct lu_name *lname;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/%s",
PFID(rr->rr_fid1), PFID(rr->rr_fid2), rr->rr_name);
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
RETURN(err_serious(-ENOENT));
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
/* Invalid case so return error immediately instead of
* processing it */
if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
RETURN(-EPERM);
/* step 1: find & lock the target parent dir */
lhp = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_pdo_init(lhp, LCK_PW, rr->rr_name,
rr->rr_namelen);
mp = mdt_object_find_lock(info, rr->rr_fid2, lhp,
MDS_INODELOCK_UPDATE);
if (IS_ERR(mp))
RETURN(PTR_ERR(mp));
if (mdt_object_obf(mp))
GOTO(out_unlock_parent, rc = -EPERM);
rc = mdt_version_get_check_save(info, mp, 0);
if (rc)
GOTO(out_unlock_parent, rc);
/* step 2: find & lock the source */
lhs = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(lhs, LCK_EX);
ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(ms))
GOTO(out_unlock_parent, rc = PTR_ERR(ms));
if (mdt_object_remote(ms)) {
mdt_object_put(info->mti_env, ms);
CERROR("Target directory "DFID" is on another MDT\n",
PFID(rr->rr_fid1));
GOTO(out_unlock_parent, rc = -EXDEV);
}
rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, ms);
GOTO(out_unlock_parent, rc);
}
/* step 3: link it */
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_LINK_WRITE);
info->mti_mos = ms;
rc = mdt_version_get_check_save(info, ms, 1);
if (rc)
GOTO(out_unlock_child, rc);
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
/** check target version by name during replay */
rc = mdt_lookup_version_check(info, mp, lname, &info->mti_tmp_fid1, 2);
if (rc != 0 && rc != -ENOENT)
GOTO(out_unlock_child, rc);
/* save version of file name for replay, it must be ENOENT here */
if (!req_is_replay(mdt_info_req(info))) {
info->mti_ver[2] = ENOENT_VERSION;
mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
}
rc = mdo_link(info->mti_env, mdt_object_child(mp),
mdt_object_child(ms), lname, ma);
if (rc == 0)
mdt_counter_incr(req, LPROC_MDT_LINK);
EXIT;
out_unlock_child:
mdt_object_unlock_put(info, ms, lhs, rc);
out_unlock_parent:
mdt_object_unlock_put(info, mp, lhp, rc);
return rc;
}
/*
* VBR: save parent version in reply and child version getting by its name.
* Version of child is getting and checking during its lookup. If
*/
static int mdt_reint_unlink(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct md_attr *ma = &info->mti_attr;
struct lu_fid *child_fid = &info->mti_tmp_fid1;
struct mdt_object *mp;
struct mdt_object *mc;
struct mdt_lock_handle *parent_lh;
struct mdt_lock_handle *child_lh;
struct lu_name *lname;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
rr->rr_name);
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
RETURN(err_serious(-ENOENT));
/*
* step 1: Found the parent.
*/
mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(mp)) {
rc = PTR_ERR(mp);
GOTO(out, rc);
}
if (mdt_object_obf(mp))
GOTO(put_parent, rc = -EPERM);
parent_lh = &info->mti_lh[MDT_LH_PARENT];
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
if (mdt_object_remote(mp)) {
mdt_lock_reg_init(parent_lh, LCK_EX);
rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
parent_lh->mlh_rreg_mode,
MDS_INODELOCK_UPDATE);
if (rc != ELDLM_OK)
GOTO(put_parent, rc);
} else {
mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
rr->rr_namelen);
rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
MDT_LOCAL_LOCK);
if (rc)
GOTO(put_parent, rc);
rc = mdt_version_get_check_save(info, mp, 0);
if (rc)
GOTO(unlock_parent, rc);
}
/* step 2: find & lock the child */
/* lookup child object along with version checking */
fid_zero(child_fid);
rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
if (rc != 0)
GOTO(unlock_parent, rc);
mdt_reint_init_ma(info, ma);
/* We will lock the child regardless it is local or remote. No harm. */
mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
if (IS_ERR(mc))
GOTO(unlock_parent, rc = PTR_ERR(mc));
child_lh = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(child_lh, LCK_EX);
if (mdt_object_remote(mc)) {
struct mdt_body *repbody;
if (!fid_is_zero(rr->rr_fid2)) {
CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n",
mdt2obd_dev(info->mti_mdt)->obd_name,
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
GOTO(unlock_parent, rc = -ENOENT);
}
CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
mdt2obd_dev(info->mti_mdt)->obd_name,
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
if (info->mti_spec.sp_rm_entry) {
struct lu_ucred *uc = mdt_ucred(info);
if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
CERROR("%s: unlink remote entry is only "
"permitted for administrator: rc = %d\n",
mdt2obd_dev(info->mti_mdt)->obd_name,
-EPERM);
GOTO(unlock_parent, rc = -EPERM);
}
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
NULL, lname, ma);
mdt_object_put(info->mti_env, mc);
GOTO(unlock_parent, rc);
}
/* Revoke the LOOKUP lock of the remote object granted by
* this MDT. Since the unlink will happen on another MDT,
* it will release the LOOKUP lock right away. Then What
* would happen if another client try to grab the LOOKUP
* lock at the same time with unlink XXX */
mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
MDT_CROSS_LOCK);
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
LASSERT(repbody != NULL);
repbody->fid1 = *mdt_object_fid(mc);
repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
mdt_object_unlock_put(info, mc, child_lh, rc);
GOTO(unlock_parent, rc = -EREMOTE);
} else if (info->mti_spec.sp_rm_entry) {
CERROR("%s: lfs rmdir should not be used on local dir %s\n",
mdt2obd_dev(info->mti_mdt)->obd_name,
(char *)rr->rr_name);
mdt_object_put(info->mti_env, mc);
GOTO(unlock_parent, rc = -EPERM);
}
rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, mc);
GOTO(unlock_parent, rc);
}
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_UNLINK_WRITE);
/* save version when object is locked */
mdt_version_get_save(info, mc, 1);
/*
* Now we can only make sure we need MA_INODE, in mdd layer, will check
* whether need MA_LOV and MA_COOKIE.
*/
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
mdt_object_child(mc), lname, ma);
if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
rc = mdt_attr_get_complex(info, mc, ma);
if (rc == 0)
mdt_handle_last_unlink(info, mc, ma);
if (ma->ma_valid & MA_INODE) {
switch (ma->ma_attr.la_mode & S_IFMT) {
case S_IFDIR:
mdt_counter_incr(req, LPROC_MDT_RMDIR);
break;
case S_IFREG:
case S_IFLNK:
case S_IFCHR:
case S_IFBLK:
case S_IFIFO:
case S_IFSOCK:
mdt_counter_incr(req, LPROC_MDT_UNLINK);
break;
default:
LASSERTF(0, "bad file type %o unlinking\n",
ma->ma_attr.la_mode);
}
}
EXIT;
mdt_object_unlock_put(info, mc, child_lh, rc);
unlock_parent:
mdt_object_unlock(info, mp, parent_lh, rc);
put_parent:
mdt_object_put(info->mti_env, mp);
out:
return rc;
}
/*
* VBR: we save three versions in reply:
* 0 - parent. Check that parent version is the same during replay.
* 1 - name. Version of 'name' if file exists with the same name or
* ENOENT_VERSION, it is needed because file may appear due to missed replays.
* 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
* check.
*/
static int mdt_md_create(struct mdt_thread_info *info)
{
struct mdt_device *mdt = info->mti_mdt;
struct mdt_object *parent;
struct mdt_object *child;
struct mdt_lock_handle *lh;
struct mdt_body *repbody;
struct md_attr *ma = &info->mti_attr;
struct mdt_reint_record *rr = &info->mti_rr;
struct lu_name *lname;
int rc;
ENTRY;
DEBUG_REQ(D_INODE, mdt_info_req(info), "Create (%s->"DFID") in "DFID,
rr->rr_name, PFID(rr->rr_fid2), PFID(rr->rr_fid1));
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
lh = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_pdo_init(lh, LCK_PW, rr->rr_name, rr->rr_namelen);
parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
MDS_INODELOCK_UPDATE);
if (IS_ERR(parent))
RETURN(PTR_ERR(parent));
if (mdt_object_obf(parent))
GOTO(out_put_parent, rc = -EPERM);
rc = mdt_version_get_check_save(info, parent, 0);
if (rc)
GOTO(out_put_parent, rc);
/*
* Check child name version during replay.
* During create replay a file may exist with same name.
*/
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
rc = mdt_lookup_version_check(info, parent, lname,
&info->mti_tmp_fid1, 1);
if (rc == 0)
GOTO(out_put_parent, rc = -EEXIST);
/* -ENOENT is expected here */
if (rc != -ENOENT)
GOTO(out_put_parent, rc);
/* save version of file name for replay, it must be ENOENT here */
mdt_enoent_version_save(info, 1);
child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
if (likely(!IS_ERR(child))) {
struct md_object *next = mdt_object_child(parent);
if (mdt_object_remote(child)) {
struct seq_server_site *ss;
struct lu_ucred *uc = mdt_ucred(info);
if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
CERROR("%s: Creating remote dir is only "
"permitted for administrator: rc = %d\n",
mdt2obd_dev(mdt)->obd_name, -EPERM);
GOTO(out_put_child, rc = -EPERM);
}
ss = mdt_seq_site(mdt);
if (ss->ss_node_id != 0 &&
mdt->mdt_enable_remote_dir == 0) {
CERROR("%s: remote dir is only permitted on"
" MDT0 or set_param"
" mdt.*.enable_remote_dir=1\n",
mdt2obd_dev(mdt)->obd_name);
GOTO(out_put_child, rc = -EPERM);
}
}
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
/* capa for cross-ref will be stored here */
ma->ma_capa = req_capsule_server_get(info->mti_pill,
&RMF_CAPA1);
LASSERT(ma->ma_capa);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_CREATE_WRITE);
/* Version of child will be updated on disk. */
info->mti_mos = child;
rc = mdt_version_get_check_save(info, child, 2);
if (rc)
GOTO(out_put_child, rc);
/* Let lower layer know current lock mode. */
info->mti_spec.sp_cr_mode =
mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
/*
* Do not perform lookup sanity check. We know that name does
* not exist.
*/
info->mti_spec.sp_cr_lookup = 0;
info->mti_spec.sp_feat = &dt_directory_features;
rc = mdo_create(info->mti_env, next, lname,
mdt_object_child(child),
&info->mti_spec, ma);
if (rc == 0)
rc = mdt_attr_get_complex(info, child, ma);
if (rc == 0) {
/* Return fid & attr to client. */
if (ma->ma_valid & MA_INODE)
mdt_pack_attr2body(info, repbody, &ma->ma_attr,
mdt_object_fid(child));
}
out_put_child:
mdt_object_put(info->mti_env, child);
} else {
rc = PTR_ERR(child);
}
mdt_create_pack_capa(info, rc, child, repbody);
out_put_parent:
mdt_object_unlock_put(info, parent, lh, rc);
RETURN(rc);
}
/*
* VBR: rename versions in reply: 0 - src parent; 1 - tgt parent;
* 2 - src child; 3 - tgt child.
* Update on disk version of src child.
*/
static int mdt_reint_rename(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
struct mdt_reint_record *rr = &info->mti_rr;
struct md_attr *ma = &info->mti_attr;
struct ptlrpc_request *req = mdt_info_req(info);
struct mdt_object *msrcdir;
struct mdt_object *mtgtdir;
struct mdt_object *mold;
struct mdt_object *mnew = NULL;
struct mdt_lock_handle *lh_srcdirp;
struct mdt_lock_handle *lh_tgtdirp;
struct mdt_lock_handle *lh_oldp;
struct mdt_lock_handle *lh_newp;
struct lu_fid *old_fid = &info->mti_tmp_fid1;
struct lu_fid *new_fid = &info->mti_tmp_fid2;
struct lustre_handle rename_lh = { 0 };
struct lu_name slname = { 0 };
struct lu_name *lname;
int rc;
ENTRY;
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
DEBUG_REQ(D_INODE, req, "rename "DFID"/%s to "DFID"/%s",
PFID(rr->rr_fid1), rr->rr_name,
PFID(rr->rr_fid2), rr->rr_tgt);
rc = mdt_rename_lock(info, &rename_lh);
if (rc) {
CERROR("Can't lock FS for rename, rc %d\n", rc);
RETURN(rc);
}
lh_newp = &info->mti_lh[MDT_LH_NEW];
/* step 1: lock the source dir. */
lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
mdt_lock_pdo_init(lh_srcdirp, LCK_PW, rr->rr_name,
rr->rr_namelen);
msrcdir = mdt_object_find_lock(info, rr->rr_fid1, lh_srcdirp,
MDS_INODELOCK_UPDATE);
if (IS_ERR(msrcdir))
GOTO(out_rename_lock, rc = PTR_ERR(msrcdir));
if (mdt_object_obf(msrcdir))
GOTO(out_unlock_source, rc = -EPERM);
rc = mdt_version_get_check_save(info, msrcdir, 0);
if (rc)
GOTO(out_unlock_source, rc);
/* step 2: find & lock the target dir. */
lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, rr->rr_tgt,
rr->rr_tgtlen);
if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
mdt_object_get(info->mti_env, msrcdir);
mtgtdir = msrcdir;
if (lh_tgtdirp->mlh_pdo_hash != lh_srcdirp->mlh_pdo_hash) {
rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
MDS_INODELOCK_UPDATE);
if (rc)
GOTO(out_unlock_source, rc);
OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
}
} else {
mtgtdir = mdt_object_find(info->mti_env, info->mti_mdt,
rr->rr_fid2);
if (IS_ERR(mtgtdir))
GOTO(out_unlock_source, rc = PTR_ERR(mtgtdir));
if (mdt_object_obf(mtgtdir))
GOTO(out_put_target, rc = -EPERM);
/* check early, the real version will be saved after locking */
rc = mdt_version_get_check(info, mtgtdir, 1);
if (rc)
GOTO(out_put_target, rc);
if (unlikely(mdt_object_remote(mtgtdir))) {
CDEBUG(D_INFO, "Source dir "DFID" target dir "DFID
"on different MDTs\n", PFID(rr->rr_fid1),
PFID(rr->rr_fid2));
GOTO(out_put_target, rc = -EXDEV);
} else {
if (likely(mdt_object_exists(mtgtdir))) {
/* we lock the target dir if it is local */
rc = mdt_object_lock(info, mtgtdir, lh_tgtdirp,
MDS_INODELOCK_UPDATE,
MDT_LOCAL_LOCK);
if (rc != 0)
GOTO(out_put_target, rc);
/* get and save correct version after locking */
mdt_version_get_save(info, mtgtdir, 1);
} else {
GOTO(out_put_target, rc = -ESTALE);
}
}
}
/* step 3: find & lock the old object. */
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
mdt_name_copy(&slname, lname);
fid_zero(old_fid);
rc = mdt_lookup_version_check(info, msrcdir, &slname, old_fid, 2);
if (rc != 0)
GOTO(out_unlock_target, rc);
if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
GOTO(out_unlock_target, rc = -EINVAL);
mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
if (IS_ERR(mold))
GOTO(out_unlock_target, rc = PTR_ERR(mold));
if (mdt_object_remote(mold)) {
mdt_object_put(info->mti_env, mold);
CDEBUG(D_INFO, "Source child "DFID" is on another MDT\n",
PFID(old_fid));
GOTO(out_unlock_target, rc = -EXDEV);
}
if (mdt_object_obf(mold)) {
mdt_object_put(info->mti_env, mold);
GOTO(out_unlock_target, rc = -EPERM);
}
lh_oldp = &info->mti_lh[MDT_LH_OLD];
mdt_lock_reg_init(lh_oldp, LCK_EX);
rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP,
MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, mold);
GOTO(out_unlock_target, rc);
}
info->mti_mos = mold;
/* save version after locking */
mdt_version_get_save(info, mold, 2);
mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
/* step 4: find & lock the new object. */
/* new target object may not exist now */
lname = mdt_name(info->mti_env, (char *)rr->rr_tgt, rr->rr_tgtlen);
/* lookup with version checking */
fid_zero(new_fid);
rc = mdt_lookup_version_check(info, mtgtdir, lname, new_fid, 3);
if (rc == 0) {
/* the new_fid should have been filled at this moment */
if (lu_fid_eq(old_fid, new_fid))
GOTO(out_unlock_old, rc);
if (lu_fid_eq(new_fid, rr->rr_fid1) ||
lu_fid_eq(new_fid, rr->rr_fid2))
GOTO(out_unlock_old, rc = -EINVAL);
mdt_lock_reg_init(lh_newp, LCK_EX);
mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
if (IS_ERR(mnew))
GOTO(out_unlock_old, rc = PTR_ERR(mnew));
if (mdt_object_obf(mnew)) {
mdt_object_put(info->mti_env, mnew);
GOTO(out_unlock_old, rc = -EPERM);
}
if (mdt_object_remote(mnew)) {
mdt_object_put(info->mti_env, mnew);
CDEBUG(D_INFO, "src child "DFID" is on another MDT\n",
PFID(new_fid));
GOTO(out_unlock_old, rc = -EXDEV);
}
rc = mdt_object_lock(info, mnew, lh_newp,
MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, mnew);
GOTO(out_unlock_old, rc);
}
/* get and save version after locking */
mdt_version_get_save(info, mnew, 3);
mdt_set_capainfo(info, 3, new_fid, BYPASS_CAPA);
} else if (rc != -EREMOTE && rc != -ENOENT) {
GOTO(out_unlock_old, rc);
} else {
mdt_enoent_version_save(info, 3);
}
/* step 5: rename it */
mdt_reint_init_ma(info, ma);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_RENAME_WRITE);
/* Check if @dst is subdir of @src. */
rc = mdt_rename_sanity(info, old_fid);
if (rc)
GOTO(out_unlock_new, rc);
rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
mdt_object_child(mtgtdir), old_fid, &slname,
(mnew ? mdt_object_child(mnew) : NULL),
lname, ma);
/* handle last link of tgt object */
if (rc == 0) {
mdt_counter_incr(req, LPROC_MDT_RENAME);
if (mnew)
mdt_handle_last_unlink(info, mnew, ma);
mdt_rename_counter_tally(info, info->mti_mdt, req,
msrcdir, mtgtdir);
}
EXIT;
out_unlock_new:
if (mnew)
mdt_object_unlock_put(info, mnew, lh_newp, rc);
out_unlock_old:
mdt_object_unlock_put(info, mold, lh_oldp, rc);
out_unlock_target:
mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
out_put_target:
mdt_object_put(info->mti_env, mtgtdir);
out_unlock_source:
mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
out_rename_lock:
if (lustre_handle_is_used(&rename_lh))
mdt_rename_unlock(&rename_lh);
return rc;
}
/*
* VBR: save parent version in reply and child version getting by its name.
* Version of child is getting and checking during its lookup. If
*/
static int mdt_reint_unlink(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
struct mdt_reint_record *rr = &info->mti_rr;
struct ptlrpc_request *req = mdt_info_req(info);
struct md_attr *ma = &info->mti_attr;
struct lu_fid *child_fid = &info->mti_tmp_fid1;
struct mdt_object *mp;
struct mdt_object *mc;
struct mdt_lock_handle *parent_lh;
struct mdt_lock_handle *child_lh;
struct lu_name *lname;
int rc;
int no_name = 0;
ENTRY;
DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
rr->rr_name);
if (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
RETURN(err_serious(-ENOENT));
if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
RETURN(-EPERM);
/*
* step 1: Found the parent.
*/
mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(mp)) {
rc = PTR_ERR(mp);
GOTO(out, rc);
}
parent_lh = &info->mti_lh[MDT_LH_PARENT];
lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
if (mdt_object_remote(mp)) {
mdt_lock_reg_init(parent_lh, LCK_EX);
rc = mdt_remote_object_lock(info, mp, &parent_lh->mlh_rreg_lh,
parent_lh->mlh_rreg_mode,
MDS_INODELOCK_UPDATE);
if (rc != ELDLM_OK)
GOTO(put_parent, rc);
} else {
mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
rr->rr_namelen);
rc = mdt_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
MDT_LOCAL_LOCK);
if (rc)
GOTO(put_parent, rc);
rc = mdt_version_get_check_save(info, mp, 0);
if (rc)
GOTO(unlock_parent, rc);
}
/* step 2: find & lock the child */
/* lookup child object along with version checking */
fid_zero(child_fid);
rc = mdt_lookup_version_check(info, mp, lname, child_fid, 1);
if (rc != 0) {
/* Name might not be able to find during resend of
* remote unlink, considering following case.
* dir_A is a remote directory, the name entry of
* dir_A is on MDT0, the directory is on MDT1,
*
* 1. client sends unlink req to MDT1.
* 2. MDT1 sends name delete update to MDT0.
* 3. name entry is being deleted in MDT0 synchronously.
* 4. MDT1 is restarted.
* 5. client resends unlink req to MDT1. So it can not
* find the name entry on MDT0 anymore.
* In this case, MDT1 only needs to destory the local
* directory.
* */
if (mdt_object_remote(mp) && rc == -ENOENT &&
!fid_is_zero(rr->rr_fid2) &&
lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
no_name = 1;
*child_fid = *rr->rr_fid2;
} else {
GOTO(unlock_parent, rc);
}
}
if (fid_is_obf(child_fid) || fid_is_dot_lustre(child_fid))
GOTO(unlock_parent, rc = -EPERM);
mdt_reint_init_ma(info, ma);
/* We will lock the child regardless it is local or remote. No harm. */
mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
if (IS_ERR(mc))
GOTO(unlock_parent, rc = PTR_ERR(mc));
child_lh = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(child_lh, LCK_EX);
if (mdt_object_remote(mc)) {
struct mdt_body *repbody;
if (!fid_is_zero(rr->rr_fid2)) {
CDEBUG(D_INFO, "%s: name %s can not find "DFID"\n",
mdt_obd_name(info->mti_mdt),
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
GOTO(put_child, rc = -ENOENT);
}
CDEBUG(D_INFO, "%s: name %s: "DFID" is another MDT\n",
mdt_obd_name(info->mti_mdt),
(char *)rr->rr_name, PFID(mdt_object_fid(mc)));
if (!mdt_is_dne_client(req->rq_export))
/* Return -EIO for old client */
GOTO(put_child, rc = -EIO);
if (info->mti_spec.sp_rm_entry) {
struct lu_ucred *uc = mdt_ucred(info);
if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
CERROR("%s: unlink remote entry is only "
"permitted for administrator: rc = %d\n",
mdt_obd_name(info->mti_mdt),
-EPERM);
GOTO(put_child, rc = -EPERM);
}
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
NULL, lname, ma, no_name);
GOTO(put_child, rc);
}
/* Revoke the LOOKUP lock of the remote object granted by
* this MDT. Since the unlink will happen on another MDT,
* it will release the LOOKUP lock right away. Then What
* would happen if another client try to grab the LOOKUP
* lock at the same time with unlink XXX */
mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP,
MDT_CROSS_LOCK);
repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
LASSERT(repbody != NULL);
repbody->fid1 = *mdt_object_fid(mc);
repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
GOTO(unlock_child, rc = -EREMOTE);
} else if (info->mti_spec.sp_rm_entry) {
rc = -EPERM;
CDEBUG(D_INFO, "%s: no rm_entry on local dir '%s': rc = %d\n",
mdt_obd_name(info->mti_mdt), (char *)rr->rr_name, rc);
GOTO(put_child, rc);
}
/* We used to acquire MDS_INODELOCK_FULL here but we can't do
* this now because a running HSM restore on the child (unlink
* victim) will hold the layout lock. See LU-4002. */
rc = mdt_object_lock(info, mc, child_lh,
MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE,
MDT_CROSS_LOCK);
if (rc != 0)
GOTO(put_child, rc);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_UNLINK_WRITE);
/* save version when object is locked */
mdt_version_get_save(info, mc, 1);
/*
* Now we can only make sure we need MA_INODE, in mdd layer, will check
* whether need MA_LOV and MA_COOKIE.
*/
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
mutex_lock(&mc->mot_lov_mutex);
rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
mdt_object_child(mc), lname, ma, no_name);
mutex_unlock(&mc->mot_lov_mutex);
if (rc == 0 && !lu_object_is_dying(&mc->mot_header))
rc = mdt_attr_get_complex(info, mc, ma);
if (rc == 0)
mdt_handle_last_unlink(info, mc, ma);
if (ma->ma_valid & MA_INODE) {
switch (ma->ma_attr.la_mode & S_IFMT) {
case S_IFDIR:
mdt_counter_incr(req, LPROC_MDT_RMDIR);
break;
case S_IFREG:
case S_IFLNK:
case S_IFCHR:
case S_IFBLK:
case S_IFIFO:
case S_IFSOCK:
mdt_counter_incr(req, LPROC_MDT_UNLINK);
break;
default:
LASSERTF(0, "bad file type %o unlinking\n",
ma->ma_attr.la_mode);
}
}
EXIT;
unlock_child:
mdt_object_unlock(info, mc, child_lh, rc);
put_child:
mdt_object_put(info->mti_env, mc);
unlock_parent:
mdt_object_unlock(info, mp, parent_lh, rc);
put_parent:
mdt_object_put(info->mti_env, mp);
out:
return rc;
}