if (repbody->valid & OBD_MD_FLMDSCAPA)
RETURN(rc);
- if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa &&
+ if (rc == 0 && info->mti_mdt->mdt_lut.lut_mds_capa &&
exp_connect_flags(info->mti_exp) & OBD_CONNECT_MDS_CAPA) {
struct lustre_capa *capa;
OBD_FAIL_MDS_REINT_CREATE_WRITE);
/* Version of child will be updated on disk. */
- info->mti_mos = child;
+ tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
rc = mdt_version_get_check_save(info, child, 2);
if (rc)
GOTO(out_put_child, rc);
/* VBR: update version if attr changed are important for recovery */
if (do_vbr) {
/* update on-disk version of changed object */
- info->mti_mos = mo;
+ tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
rc = mdt_version_get_check_save(info, mo, 0);
if (rc)
GOTO(out_unlock, rc);
rc = mdt_attr_get_complex(info, mo, ma);
if (rc) {
CERROR("file attribute read error for "DFID": %d.\n",
- PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
+ PFID(mdt_object_fid(mo)), rc);
RETURN(rc);
}
/* If an up2date copy exists in the backend, add dirty flag */
if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
&& !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
+ struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_CHILD];
ma->ma_hsm.mh_flags |= HS_DIRTY;
+
+ mdt_lock_reg_init(lh, LCK_PW);
+ rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR,
+ MDT_LOCAL_LOCK);
+ if (rc != 0)
+ RETURN(rc);
+
rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
- if (rc) {
+ if (rc)
CERROR("file attribute change error for "DFID": %d\n",
- PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
- RETURN(rc);
- }
+ PFID(mdt_object_fid(mo)), rc);
+ mdt_object_unlock(info, mo, lh, rc);
}
RETURN(rc);
if (rc)
GOTO(out_put, rc);
- mfd = mdt_mfd_new();
+ mfd = mdt_mfd_new(med);
if (mfd == NULL) {
mdt_write_put(mo);
GOTO(out_put, rc = -ENOMEM);
LASSERT(info->mti_ioepoch);
spin_lock(&med->med_open_lock);
- mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
+ mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
+ req_is_replay(req));
if (mfd == NULL) {
spin_unlock(&med->med_open_lock);
CDEBUG(D_INODE, "no handle for file close: "
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
rc = mdt_attr_get_complex(info, mo, ma);
- if (rc != 0)
- GOTO(out_put, rc);
+ if (rc != 0)
+ GOTO(out_put, rc);
- mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
+ mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
- if (info->mti_mdt->mdt_opts.mo_oss_capa &&
+ if (info->mti_mdt->mdt_lut.lut_oss_capa &&
exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
- S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
+ S_ISREG(lu_object_attr(&mo->mot_obj)) &&
(ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
struct lustre_capa *capa;
repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
GOTO(unlock_child, rc = -EREMOTE);
} else if (info->mti_spec.sp_rm_entry) {
- CERROR("%s: lfs rmdir should not be used on local dir %s\n",
- mdt_obd_name(info->mti_mdt),
- (char *)rr->rr_name);
- GOTO(put_child, rc = -EPERM);
+ 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);
}
- rc = mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_FULL,
- MDT_CROSS_LOCK);
- if (rc != 0) {
+ /* 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);
+ /*
+ * 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)
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));
+ CERROR("%s: source inode "DFID" on remote MDT from "DFID"\n",
+ mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1),
+ PFID(rr->rr_fid2));
GOTO(out_unlock_parent, rc = -EXDEV);
}
- rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
- MDT_CROSS_LOCK);
+ rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE |
+ MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, ms);
GOTO(out_unlock_parent, rc);
mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
OBD_FAIL_MDS_REINT_LINK_WRITE);
- info->mti_mos = ms;
+ tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
rc = mdt_version_get_check_save(info, ms, 1);
if (rc)
GOTO(out_unlock_child, rc);
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))) {
+ if (rc != -ENOENT) {
+ CDEBUG(D_INFO, "link target %.*s existed!\n",
+ rr->rr_namelen, (char *)rr->rr_name);
+ GOTO(out_unlock_child, rc = -EEXIST);
+ }
info->mti_ver[2] = ENOENT_VERSION;
mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
}
* (lh->mlh_pdo_hash) in parallel directory lock.
*/
static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
- struct mdt_lock_handle *lh,
- struct mdt_object *obj, __u64 ibits)
+ struct mdt_lock_handle *lh,
+ struct mdt_object *obj, __u64 ibits)
{
- struct ldlm_res_id *res_id = &info->mti_res_id;
- struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
- ldlm_policy_data_t *policy = &info->mti_policy;
- int rc;
+ struct ldlm_res_id *res = &info->mti_res_id;
+ struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
+ ldlm_policy_data_t *policy = &info->mti_policy;
+ int rc;
- /*
- * Finish res_id initializing by name hash marking part of
- * directory which is taking modification.
- */
- LASSERT(lh->mlh_pdo_hash != 0);
- fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res_id);
- memset(policy, 0, sizeof(*policy));
- policy->l_inodebits.bits = ibits;
- /*
- * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
- * going to be sent to client. If it is - mdt_intent_policy() path will
- * fix it up and turn FL_LOCAL flag off.
- */
- rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
- res_id, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
- &info->mti_exp->exp_handle.h_cookie);
- return rc;
+ /*
+ * Finish res_id initializing by name hash marking part of
+ * directory which is taking modification.
+ */
+ LASSERT(lh->mlh_pdo_hash != 0);
+ fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
+ memset(policy, 0, sizeof(*policy));
+ policy->l_inodebits.bits = ibits;
+ /*
+ * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
+ * going to be sent to client. If it is - mdt_intent_policy() path will
+ * fix it up and turn FL_LOCAL flag off.
+ */
+ rc = mdt_fid_lock(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
+ res, LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB,
+ &info->mti_exp->exp_handle.h_cookie);
+ return rc;
}
static int mdt_rename_lock(struct mdt_thread_info *info,
memset(policy, 0, sizeof *policy);
policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
-#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 4, 53, 0)
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 5, 53, 0)
/* In phase I, we will not do cross-rename, so local BFL lock would
* be enough
*/
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);
+ rc = mdt_object_lock(info, mold, lh_oldp, MDS_INODELOCK_LOOKUP |
+ MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, mold);
GOTO(out_unlock_target, rc);
}
- info->mti_mos = mold;
+ tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
/* save version after locking */
mdt_version_get_save(info, mold, 2);
mdt_set_capainfo(info, 2, old_fid, BYPASS_CAPA);
GOTO(out_unlock_old, rc = -EXDEV);
}
- rc = mdt_object_lock(info, mnew, lh_newp,
- MDS_INODELOCK_FULL, MDT_CROSS_LOCK);
+ /* We used to acquire MDS_INODELOCK_FULL here but we
+ * can't do this now because a running HSM restore on
+ * the rename onto victim will hold the layout
+ * lock. See LU-4002. */
+ rc = mdt_object_lock(info, mnew, lh_newp,
+ MDS_INODELOCK_LOOKUP |
+ MDS_INODELOCK_UPDATE,
+ MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, mnew);
GOTO(out_unlock_old, rc);
} else if (rc != -EREMOTE && rc != -ENOENT) {
GOTO(out_unlock_old, rc);
} else {
- mdt_enoent_version_save(info, 3);
+ /* If mnew does not exist and mold are remote directory,
+ * it only allows rename if they are under same directory */
+ if (mtgtdir != msrcdir && mdt_object_remote(mold)) {
+ CDEBUG(D_INFO, "Src child "DFID" is on another MDT\n",
+ PFID(old_fid));
+ GOTO(out_unlock_old, rc = -EXDEV);
+ }
+ mdt_enoent_version_save(info, 3);
}
/* step 5: rename it */
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);
- /* Check if @dst is subdir of @src. */
- rc = mdt_rename_sanity(info, old_fid);
- if (rc)
- GOTO(out_unlock_new, rc);
+ if (mnew != NULL)
+ mutex_lock(&mnew->mot_lov_mutex);
- 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);
+ 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) {
+ if (mnew != NULL)
+ mutex_unlock(&mnew->mot_lov_mutex);
+
+ /* 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);
+ if (mnew)
+ mdt_handle_last_unlink(info, mnew, ma);
mdt_rename_counter_tally(info, info->mti_mdt, req,
msrcdir, mtgtdir);