* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2012, Intel Corporation.
+ * Copyright (c) 2011, 2013, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#define DEBUG_SUBSYSTEM S_MDS
#include "mdt_internal.h"
-#include "lu_time.h"
static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
struct md_attr *ma)
if (repbody->valid & OBD_MD_FLMDSCAPA)
RETURN(rc);
- if (rc == 0 && info->mti_mdt->mdt_opts.mo_mds_capa &&
- info->mti_exp->exp_connect_flags & OBD_CONNECT_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;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
struct mdt_object *o, __u64 *version)
{
LASSERT(o);
- LASSERT(mdt_object_exists(o) >= 0);
- if (mdt_object_exists(o) > 0 && !mdt_object_obf(o))
+ if (mdt_object_exists(o) && !mdt_object_remote(o) &&
+ !fid_is_obf(mdt_object_fid(o)))
*version = dt_version_get(info->mti_env, mdt_obj2dt(o));
else
*version = ENOENT_VERSION;
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);
+ if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
+ RETURN(-EPERM);
+
+ 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);
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);
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)) {
+ if (uc->uc_gid !=
+ mdt->mdt_enable_remote_dir_gid &&
+ mdt->mdt_enable_remote_dir_gid != -1) {
+ CERROR("%s: Creating remote dir is only"
+ " permitted for administrator or"
+ " set mdt_enable_remote_dir_gid:"
+ " rc = %d\n",
+ mdt_obd_name(mdt), -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",
+ mdt_obd_name(mdt));
+ GOTO(out_put_child, rc = -EPERM);
+ }
+ if (!mdt_is_dne_client(mdt_info_req(info)->rq_export)) {
+ /* Return -EIO for old client */
+ GOTO(out_put_child, rc = -EIO);
+ }
+
+ }
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
/* capa for cross-ref will be stored here */
int rc;
ENTRY;
- /* attr shouldn't be set on remote object */
- LASSERT(mdt_object_exists(mo) >= 0);
+ /* 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);
- if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
- lockpart |= MDS_INODELOCK_LOOKUP;
+ /* 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)
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 (info->mti_dlm_req)
ldlm_request_cancel(req, info->mti_dlm_req, 0);
- repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
+ repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
if (IS_ERR(mo))
GOTO(out, rc = PTR_ERR(mo));
- if (mdt_object_obf(mo))
- GOTO(out_put, rc = -EPERM);
-
/* start a log jounal handle if needed */
if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM)) {
if ((ma->ma_attr.la_valid & LA_SIZE) ||
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 &&
- info->mti_exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
- S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
- (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
+ 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)) &&
+ (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
struct lustre_capa *capa;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
break;
default:
CERROR("%s: Unsupported mode %o\n",
- mdt2obd_dev(info->mti_mdt)->obd_name,
+ mdt_obd_name(info->mti_mdt),
info->mti_attr.ma_attr.la_mode);
RETURN(err_serious(-EOPNOTSUPP));
}
struct mdt_lock_handle *child_lh;
struct lu_name *lname;
int rc;
- ENTRY;
+ int no_name = 0;
+ ENTRY;
DEBUG_REQ(D_INODE, req, "unlink "DFID"/%s", PFID(rr->rr_fid1),
rr->rr_name);
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: lock the parent.
+ * step 1: Found the parent.
*/
- parent_lh = &info->mti_lh[MDT_LH_PARENT];
- mdt_lock_pdo_init(parent_lh, LCK_PW, rr->rr_name,
- rr->rr_namelen);
+ mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
+ if (IS_ERR(mp)) {
+ rc = PTR_ERR(mp);
+ GOTO(out, rc);
+ }
- mp = mdt_object_find_lock(info, rr->rr_fid1, parent_lh,
- MDS_INODELOCK_UPDATE);
- if (IS_ERR(mp))
- GOTO(out, rc = PTR_ERR(mp));
+ 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);
- if (mdt_object_obf(mp))
- GOTO(out_unlock_parent, rc = -EPERM);
+ rc = mdt_version_get_check_save(info, mp, 0);
+ if (rc)
+ GOTO(unlock_parent, rc);
+ }
- rc = mdt_version_get_check_save(info, mp, 0);
- if (rc)
- GOTO(out_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);
- /* step 2: find & lock the child */
- lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
- /* 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(out_unlock_parent, rc);
+ /* 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));
- /* 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(out_unlock_parent, rc = PTR_ERR(mc));
child_lh = &info->mti_lh[MDT_LH_CHILD];
mdt_lock_reg_init(child_lh, LCK_EX);
- 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(out_unlock_parent, rc);
- }
+ 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);
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);
+
+ rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
+ mdt_object_child(mc), lname, ma, no_name);
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 (rc == 0)
+ mdt_handle_last_unlink(info, mc, ma);
if (ma->ma_valid & MA_INODE) {
switch (ma->ma_attr.la_mode & S_IFMT) {
}
EXIT;
-
- mdt_object_unlock_put(info, mc, child_lh, rc);
-out_unlock_parent:
- mdt_object_unlock_put(info, mp, parent_lh, rc);
+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;
}
if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
RETURN(-EPERM);
+ if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
+ fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(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,
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);
if (IS_ERR(ms))
GOTO(out_unlock_parent, rc = PTR_ERR(ms));
- rc = mdt_object_lock(info, ms, lhs, MDS_INODELOCK_UPDATE,
- MDT_CROSS_LOCK);
+ if (mdt_object_remote(ms)) {
+ mdt_object_put(info->mti_env, ms);
+ 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 |
+ MDS_INODELOCK_XATTR, MDT_CROSS_LOCK);
if (rc != 0) {
mdt_object_put(info->mti_env, ms);
GOTO(out_unlock_parent, 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,
struct lustre_handle *lh)
{
- struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
- ldlm_policy_data_t *policy = &info->mti_policy;
- struct ldlm_res_id *res_id = &info->mti_res_id;
- __u64 flags = 0;
- struct md_site *ms;
- int rc;
- ENTRY;
-
- ms = mdt_md_site(info->mti_mdt);
- fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
-
- memset(policy, 0, sizeof *policy);
- policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
-
- if (ms->ms_control_exp == NULL) {
- flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
-
- /*
- * Current node is controller, that is mdt0, where we should
- * take BFL lock.
- */
- rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
- LCK_EX, &flags, ldlm_blocking_ast,
- ldlm_completion_ast, NULL, NULL, 0,
- LVB_T_NONE,
- &info->mti_exp->exp_handle.h_cookie,
- lh);
- } else {
- struct ldlm_enqueue_info einfo = { LDLM_IBITS, LCK_EX,
- ldlm_blocking_ast, ldlm_completion_ast, NULL, NULL, NULL };
- /*
- * This is the case mdt0 is remote node, issue DLM lock like
- * other clients.
- */
- rc = ldlm_cli_enqueue(ms->ms_control_exp, NULL, &einfo, res_id,
- policy, &flags, NULL, 0, LVB_T_NONE, lh,
- 0);
- }
+ struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
+ ldlm_policy_data_t *policy = &info->mti_policy;
+ struct ldlm_res_id *res_id = &info->mti_res_id;
+ __u64 flags = 0;
+ int rc;
+ ENTRY;
+ fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
+
+ memset(policy, 0, sizeof *policy);
+ policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
+#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
+ */
+ flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
+ /*
+ * Current node is controller, that is mdt0, where we should
+ * take BFL lock.
+ */
+ rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
+ LCK_EX, &flags, ldlm_blocking_ast,
+ ldlm_completion_ast, NULL, NULL, 0,
+ LVB_T_NONE,
+ &info->mti_exp->exp_handle.h_cookie,
+ lh);
+#else
+#warning "Local rename lock is invalid for DNE phase II."
+#endif
RETURN(rc);
}
* 2 - src child; 3 - tgt child.
* Update on disk version of src child.
*/
+/**
+ * For DNE phase I, only these renames are allowed
+ * mv src_p/src_c tgt_p/tgt_c
+ * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
+ * 2. src_p and tgt_p are same directory, and tgt_c does not
+ * exists. In this case, all of modification will happen
+ * in the MDT where ithesource parent is, only one remote
+ * update is needed, i.e. set c_time/m_time on the child.
+ * And tgt_c will be still in the same MDT as the original
+ * src_c.
+ */
static int mdt_reint_rename(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
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);
- }
+ if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1) ||
+ fid_is_obf(rr->rr_fid2) || fid_is_dot_lustre(rr->rr_fid2))
+ RETURN(-EPERM);
+
+ 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];
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);
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);
- rc = mdt_object_exists(mtgtdir);
- if (rc == 0) {
- GOTO(out_put_target, rc = -ESTALE);
- } else if (rc > 0) {
- /* 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);
- }
- }
+ 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);
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_obf(mold)) {
- mdt_object_put(info->mti_env, mold);
+ if (fid_is_obf(old_fid) || fid_is_dot_lustre(old_fid))
GOTO(out_unlock_target, rc = -EPERM);
- }
+
+ mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
+ if (IS_ERR(mold))
+ GOTO(out_unlock_target, rc = PTR_ERR(mold));
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);
lu_fid_eq(new_fid, rr->rr_fid2))
GOTO(out_unlock_old, rc = -EINVAL);
+ if (fid_is_obf(new_fid) || fid_is_dot_lustre(new_fid))
+ GOTO(out_unlock_old, rc = -EPERM);
+
+ if (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_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)) {
+ if (mdt_object_remote(mnew)) {
mdt_object_put(info->mti_env, mnew);
- GOTO(out_unlock_old, rc = -EPERM);
+ 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);
+ /* 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 */
out_unlock_source:
mdt_object_unlock_put(info, msrcdir, lh_srcdirp, rc);
out_rename_lock:
- mdt_rename_unlock(&rename_lh);
- return rc;
+ if (lustre_handle_is_used(&rename_lh))
+ mdt_rename_unlock(&rename_lh);
+ return rc;
}
typedef int (*mdt_reinter)(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc);
static mdt_reinter reinters[REINT_MAX] = {
- [REINT_SETATTR] = mdt_reint_setattr,
- [REINT_CREATE] = mdt_reint_create,
- [REINT_LINK] = mdt_reint_link,
- [REINT_UNLINK] = mdt_reint_unlink,
- [REINT_RENAME] = mdt_reint_rename,
- [REINT_OPEN] = mdt_reint_open,
- [REINT_SETXATTR] = mdt_reint_setxattr
+ [REINT_SETATTR] = mdt_reint_setattr,
+ [REINT_CREATE] = mdt_reint_create,
+ [REINT_LINK] = mdt_reint_link,
+ [REINT_UNLINK] = mdt_reint_unlink,
+ [REINT_RENAME] = mdt_reint_rename,
+ [REINT_OPEN] = mdt_reint_open,
+ [REINT_SETXATTR] = mdt_reint_setxattr,
+ [REINT_RMENTRY] = mdt_reint_unlink
};
int mdt_reint_rec(struct mdt_thread_info *info,