#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_opts.mo_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) &&
+ !mdt_object_obf(o))
*version = dt_version_get(info->mti_env, mdt_obj2dt(o));
else
*version = ENOENT_VERSION;
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);
+ }
+ 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)
return rc;
}
+/**
+ * Check HSM flags and add HS_DIRTY flag if relevant.
+ *
+ * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
+ * and is not RELEASED.
+ */
+int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
+ struct md_attr *ma)
+{
+ int rc;
+ ENTRY;
+
+ /* If the file was modified, add the dirty flag */
+ ma->ma_need = MA_HSM;
+ 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);
+ 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))) {
+
+ ma->ma_hsm.mh_flags |= HS_DIRTY;
+ rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
+ if (rc) {
+ CERROR("file attribute change error for "DFID": %d\n",
+ PFID(lu_object_fid(&mo->mot_obj.mo_lu)), rc);
+ RETURN(rc);
+ }
+ }
+
+ RETURN(rc);
+}
+
static int mdt_reint_setattr(struct mdt_thread_info *info,
struct mdt_lock_handle *lhc)
{
} else
LBUG();
+ /* If file data is modified, add the dirty flag */
+ if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
+ rc = mdt_add_dirty_flag(info, mo, ma);
+
ma->ma_need = MA_INODE;
ma->ma_valid = 0;
rc = mdt_attr_get_complex(info, mo, ma);
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_opts.mo_oss_capa &&
+ exp_connect_flags(info->mti_exp) & OBD_CONNECT_OSS_CAPA &&
+ S_ISREG(lu_object_attr(&mo->mot_obj.mo_lu)) &&
+ (ma->ma_attr.la_valid & LA_SIZE) && !som_au) {
struct lustre_capa *capa;
capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
RETURN(err_serious(-ENOENT));
/*
- * 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));
+ 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);
- 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)
+ GOTO(unlock_parent, rc);
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);
+ 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 (!mdt_is_dne_client(req->rq_export))
+ /* Return -EIO for old client */
+ GOTO(unlock_parent, 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",
+ 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(out_unlock_parent, rc);
- }
+ 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);
EXIT;
mdt_object_unlock_put(info, mc, child_lh, rc);
-out_unlock_parent:
- mdt_object_unlock_put(info, mp, parent_lh, rc);
+unlock_parent:
+ mdt_object_unlock(info, mp, parent_lh, rc);
+put_parent:
+ mdt_object_put(info->mti_env, mp);
out:
return rc;
}
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) {
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, 4, 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);
}
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);
- }
+ 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 (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));
+ 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_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) {
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,
git://git.whamcloud.com / fs / lustre-release.git / blobdiff