* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2014, Intel Corporation.
+ * Copyright (c) 2011, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
if (unlikely(mdd_is_dead_obj(mdd_obj)))
RETURN(-ESTALE);
+ if (!mdd_object_exists(mdd_obj))
+ RETURN(-ESTALE);
+
if (mdd_object_remote(mdd_obj)) {
CDEBUG(D_INFO, "%s: Object "DFID" locates on remote server\n",
mdd2obd_dev(m)->obd_name, PFID(mdo2fid(mdd_obj)));
- } else if (!mdd_object_exists(mdd_obj)) {
- RETURN(-ESTALE);
}
rc = mdd_permission_internal_locked(env, mdd_obj, pattr, mask,
parent = mdd_object_find(env, mdd, pfid);
if (IS_ERR(parent))
GOTO(out, rc = PTR_ERR(parent));
+
+ if (!mdd_object_exists(parent))
+ GOTO(out, rc = -EINVAL);
+
p1 = parent;
}
EXIT;
const struct lu_attr *pattr, struct mdd_object *cobj,
bool check_perm)
{
- struct mdd_thread_info *info = mdd_env_info(env);
- struct lu_buf *xbuf;
int rc = 0;
ENTRY;
if (mdd_is_dead_obj(pobj))
RETURN(-ENOENT);
- /* If the parent is a sub-stripe, check whether it is dead */
- xbuf = mdd_buf_get(env, info->mti_key, sizeof(info->mti_key));
- rc = mdo_xattr_get(env, pobj, xbuf, XATTR_NAME_LMV);
- if (unlikely(rc > 0)) {
- struct lmv_mds_md_v1 *lmv1 = xbuf->lb_buf;
-
- if (le32_to_cpu(lmv1->lmv_magic) == LMV_MAGIC_STRIPE &&
- le32_to_cpu(lmv1->lmv_hash_type) & LMV_HASH_FLAG_DEAD)
- RETURN(-ESTALE);
- }
- rc = 0;
-
if (check_perm)
rc = mdd_permission_internal_locked(env, pobj, pattr,
MAY_WRITE | MAY_EXEC,
int mdd_links_write(const struct lu_env *env, struct mdd_object *mdd_obj,
struct linkea_data *ldata, struct thandle *handle)
{
- const struct lu_buf *buf = mdd_buf_get_const(env, ldata->ld_buf->lb_buf,
- ldata->ld_leh->leh_len);
+ const struct lu_buf *buf;
int rc;
+ if (ldata == NULL || ldata->ld_buf == NULL ||
+ ldata->ld_leh == NULL)
+ return 0;
+
+ buf = mdd_buf_get_const(env, ldata->ld_buf->lb_buf,
+ ldata->ld_leh->leh_len);
if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_LINKEA))
return 0;
return rc;
}
-static int mdd_mark_dead_object(const struct lu_env *env,
+static int mdd_mark_orphan_object(const struct lu_env *env,
struct mdd_object *obj, struct thandle *handle,
bool declare)
{
struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
int rc;
- if (!declare)
- obj->mod_flags |= DEAD_OBJ;
-
if (!S_ISDIR(mdd_object_type(obj)))
return 0;
attr->la_valid = LA_FLAGS;
- attr->la_flags = LUSTRE_SLAVE_DEAD_FL;
+ attr->la_flags = LUSTRE_ORPHAN_FL;
if (declare)
rc = mdo_declare_attr_set(env, obj, attr, handle);
{
int rc;
- rc = mdd_mark_dead_object(env, obj, handle, true);
+ /* Sigh, we do not know if the unlink object will become orphan in
+ * declare phase, but fortunately the flags here does not matter
+ * in current declare implementation */
+ rc = mdd_mark_orphan_object(env, obj, handle, true);
if (rc != 0)
return rc;
- rc = orph_declare_index_insert(env, obj, mdd_object_type(obj), handle);
+ rc = mdo_declare_destroy(env, obj, handle);
if (rc != 0)
return rc;
- rc = mdo_declare_destroy(env, obj, handle);
+ rc = orph_declare_index_insert(env, obj, mdd_object_type(obj), handle);
if (rc != 0)
return rc;
struct thandle *th)
{
int rc = 0;
- int is_dir = S_ISDIR(ma->ma_attr.la_mode);
- ENTRY;
+ int is_dir = S_ISDIR(ma->ma_attr.la_mode);
+ ENTRY;
- LASSERT(mdd_write_locked(env, obj) != 0);
+ LASSERT(mdd_write_locked(env, obj) != 0);
if (ma->ma_attr.la_nlink == 0 || is_dir) {
- rc = mdd_mark_dead_object(env, obj, th, false);
- if (rc != 0)
- RETURN(rc);
-
- /* add new orphan and the object
- * will be deleted during mdd_close() */
- if (obj->mod_count) {
- rc = __mdd_orphan_add(env, obj, th);
- if (rc == 0)
- CDEBUG(D_HA, "Object "DFID" is inserted into "
- "orphan list, open count = %d\n",
- PFID(mdd_object_fid(obj)),
- obj->mod_count);
- else
- CERROR("Object "DFID" fail to be an orphan, "
- "open count = %d, maybe cause failed "
- "open replay\n",
- PFID(mdd_object_fid(obj)),
- obj->mod_count);
- } else {
+ /* add new orphan and the object
+ * will be deleted during mdd_close() */
+ obj->mod_flags |= DEAD_OBJ;
+ if (obj->mod_count) {
+ rc = __mdd_orphan_add(env, obj, th);
+ if (rc == 0)
+ CDEBUG(D_HA, "Object "DFID" is inserted into "
+ "orphan list, open count = %d\n",
+ PFID(mdd_object_fid(obj)),
+ obj->mod_count);
+ else
+ CERROR("Object "DFID" fail to be an orphan, "
+ "open count = %d, maybe cause failed "
+ "open replay\n",
+ PFID(mdd_object_fid(obj)),
+ obj->mod_count);
+
+ /* mark object as an orphan here, not
+ * before __mdd_orphan_add() as racing
+ * mdd_la_get() may propagate ORPHAN_OBJ
+ * causing the asserition */
+ rc = mdd_mark_orphan_object(env, obj, th, false);
+ } else {
rc = mdo_destroy(env, obj, th);
- }
+ }
} else if (!is_dir) {
/* old files may not have link ea; ignore errors */
mdd_links_del(env, obj, mdo2fid(pobj), lname, th);
if (rc < 0)
RETURN(false);
- ma->ma_valid = MA_HSM;
+ ma->ma_valid |= MA_HSM;
}
if (ma->ma_hsm.mh_flags & HS_EXISTS)
RETURN(true);
struct mdd_object *mdd_cobj = NULL;
struct mdd_device *mdd = mdo2mdd(pobj);
struct thandle *handle;
- int rc, is_dir = 0;
+ int rc, is_dir = 0, cl_flags = 0;
ENTRY;
/* cobj == NULL means only delete name entry */
RETURN(rc);
is_dir = S_ISDIR(cattr->la_mode);
+ /* search for an existing archive.
+ * we should check ahead as the object
+ * can be destroyed in this transaction */
+ if (mdd_hsm_archive_exists(env, mdd_cobj, ma))
+ cl_flags |= CLF_UNLINK_HSM_EXISTS;
}
rc = mdd_unlink_sanity_check(env, mdd_pobj, pattr, mdd_cobj, cattr);
ma->ma_attr = *cattr;
ma->ma_valid |= MA_INODE;
rc = mdd_finish_unlink(env, mdd_cobj, ma, mdd_pobj, lname, handle);
+ if (rc != 0)
+ GOTO(cleanup, rc);
/* fetch updated nlink */
- if (rc == 0)
- rc = mdd_la_get(env, mdd_cobj, cattr);
+ rc = mdd_la_get(env, mdd_cobj, cattr);
+ /* if object is removed then we can't get its attrs,
+ * use last get */
+ if (rc == -ENOENT) {
+ cattr->la_nlink = 0;
+ rc = 0;
+ }
- /* if object is removed then we can't get its attrs, use last get */
if (cattr->la_nlink == 0) {
ma->ma_attr = *cattr;
ma->ma_valid |= MA_INODE;
}
+
EXIT;
cleanup:
if (likely(mdd_cobj != NULL))
mdd_write_unlock(env, mdd_cobj);
if (rc == 0) {
- int cl_flags = 0;
-
- if (cattr->la_nlink == 0) {
+ if (cattr->la_nlink == 0)
cl_flags |= CLF_UNLINK_LAST;
- /* search for an existing archive */
- if (mdd_hsm_archive_exists(env, mdd_cobj, ma))
- cl_flags |= CLF_UNLINK_HSM_EXISTS;
- }
+ else
+ cl_flags &= ~CLF_UNLINK_HSM_EXISTS;
rc = mdd_changelog_ns_store(env, mdd,
is_dir ? CL_RMDIR : CL_UNLINK, cl_flags,
int rc;
ENTRY;
- /* EEXIST check */
- if (mdd_is_dead_obj(obj))
- RETURN(-ENOENT);
+ /* EEXIST check */
+ if (mdd_is_dead_obj(obj))
+ RETURN(-ENOENT);
/*
* In some cases this lookup is not needed - we know before if name
RETURN(rc);
}
+static int mdd_index_delete(const struct lu_env *env,
+ struct mdd_object *mdd_pobj,
+ struct lu_attr *cattr,
+ const struct lu_name *lname)
+{
+ struct mdd_device *mdd = mdo2mdd(&mdd_pobj->mod_obj);
+ struct thandle *handle;
+ int rc;
+ ENTRY;
+
+ handle = mdd_trans_create(env, mdd);
+ if (IS_ERR(handle))
+ RETURN(PTR_ERR(handle));
+
+ rc = mdo_declare_index_delete(env, mdd_pobj, lname->ln_name,
+ handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+
+ if (S_ISDIR(cattr->la_mode)) {
+ rc = mdo_declare_ref_del(env, mdd_pobj, handle);
+ if (rc != 0)
+ GOTO(stop, rc);
+ }
+
+ /* Since this will only be used in the error handler path,
+ * Let's set the thandle to be local and not mess the transno */
+ handle->th_local = 1;
+ rc = mdd_trans_start(env, mdd, handle);
+ if (rc)
+ GOTO(stop, rc);
+
+ rc = __mdd_index_delete(env, mdd_pobj, lname->ln_name,
+ S_ISDIR(cattr->la_mode), handle);
+ if (rc)
+ GOTO(stop, rc);
+stop:
+ mdd_trans_stop(env, mdd, rc, handle);
+ RETURN(rc);
+}
+
/*
* Create object and insert it into namespace.
*/
NULL, handle);
out_stop:
rc2 = mdd_trans_stop(env, mdd, rc, handle);
- if (rc == 0)
+ if (rc == 0) {
+ /* If creation fails, it is most likely due to the remote update
+ * failure, because local transaction will mostly succeed at
+ * this stage. There is no easy way to rollback all of previous
+ * updates, so let's remove the object from namespace, and
+ * LFSCK should handle the orphan object. */
+ if (rc2 < 0 && !mdd_object_remote(mdd_pobj))
+ mdd_index_delete(env, mdd_pobj, attr, lname);
rc = rc2;
+ }
out_free:
if (is_vmalloc_addr(ldata->ld_buf))
/* if we vmalloced a large buffer drop it */
rc = mdd_la_get(env, mdd_tobj, tattr);
if (rc)
GOTO(out_pending, rc);
+ /* search for an existing archive.
+ * we should check ahead as the object
+ * can be destroyed in this transaction */
+ if (mdd_hsm_archive_exists(env, mdd_tobj, ma))
+ cl_flags |= CLF_RENAME_LAST_EXISTS;
}
rc = mdd_la_get(env, mdd_tpobj, tpattr);
/* fetch updated nlink */
rc = mdd_la_get(env, mdd_tobj, tattr);
- if (rc != 0) {
+ if (rc == -ENOENT) {
+ /* the object got removed, let's
+ * return the latest known attributes */
+ tattr->la_nlink = 0;
+ rc = 0;
+ } else if (rc != 0) {
CERROR("%s: Failed to get nlink for tobj "
DFID": rc = %d\n",
mdd2obd_dev(mdd)->obd_name,
ma->ma_attr = *tattr;
ma->ma_valid |= MA_INODE;
- if (tattr->la_nlink == 0) {
+ if (tattr->la_nlink == 0)
cl_flags |= CLF_RENAME_LAST;
- if (mdd_hsm_archive_exists(env, mdd_tobj, ma))
- cl_flags |= CLF_RENAME_LAST_EXISTS;
- }
+ else
+ cl_flags &= ~CLF_RENAME_LAST_EXISTS;
}
la->la_valid = LA_CTIME | LA_MTIME;
**/
static int mdd_linkea_update_child_internal(const struct lu_env *env,
struct mdd_object *parent,
+ struct mdd_object *newparent,
struct mdd_object *child,
const char *name, int namelen,
struct thandle *handle,
&lname, &fid);
if (strncmp(lname.ln_name, name, namelen) != 0 ||
- lu_fid_eq(&fid, mdd_object_fid(parent))) {
+ !lu_fid_eq(&fid, mdd_object_fid(parent))) {
ldata.ld_lee = (struct link_ea_entry *)
((char *)ldata.ld_lee +
ldata.ld_reclen);
CDEBUG(D_INFO, "%s: update "DFID" with %.*s:"DFID"\n",
mdd2obd_dev(mdd)->obd_name, PFID(mdd_object_fid(child)),
lname.ln_namelen, lname.ln_name,
- PFID(mdd_object_fid(parent)));
+ PFID(mdd_object_fid(newparent)));
/* update to the new parent fid */
linkea_entry_pack(ldata.ld_lee, &lname,
- mdd_object_fid(parent));
+ mdd_object_fid(newparent));
if (declare)
rc = mdd_declare_links_add(env, child, handle, &ldata,
MLAO_IGNORE);
static int mdd_linkea_declare_update_child(const struct lu_env *env,
struct mdd_object *parent,
+ struct mdd_object *newparent,
struct mdd_object *child,
const char *name, int namelen,
struct thandle *handle)
{
- return mdd_linkea_update_child_internal(env, parent, child, name,
+ return mdd_linkea_update_child_internal(env, parent, newparent,
+ child, name,
namelen, handle, true);
}
static int mdd_linkea_update_child(const struct lu_env *env,
struct mdd_object *parent,
+ struct mdd_object *newparent,
struct mdd_object *child,
const char *name, int namelen,
struct thandle *handle)
{
- return mdd_linkea_update_child_internal(env, parent, child, name,
+ return mdd_linkea_update_child_internal(env, parent, newparent,
+ child, name,
namelen, handle, false);
}
buf, 0, handle);
if (rc != 0)
return rc;
- } else if (S_ISDIR(la->la_mode)) {
+ } else if (S_ISDIR(la->la_mode) && ldata != NULL) {
rc = mdd_declare_links_add(env, mdd_tobj, handle, ldata,
MLAO_IGNORE);
if (rc != 0)
}
} else if (S_ISDIR(la->la_mode)) {
rc = mdd_links_read(env, mdd_sobj, ldata);
- if (rc < 0 && rc != -ENODATA)
+ if (rc == -ENODATA) {
+ /* ignore the non-linkEA error */
+ ldata = NULL;
+ rc = 0;
+ }
+ if (rc < 0)
RETURN(rc);
}
if (rc != 0)
GOTO(stop_trans, rc);
+ /* don't set nlink from the original object */
+ la->la_valid &= ~LA_NLINK;
+
/* create the target object */
rc = mdd_object_create(env, mdd_pobj, mdd_tobj, la, spec, NULL, NULL,
hint, handle);
if (rc != 0)
GOTO(stop_trans, rc);
- if (S_ISDIR(la->la_mode)) {
+ if (S_ISDIR(la->la_mode) && ldata != NULL) {
rc = mdd_links_write(env, mdd_tobj, ldata, handle);
if (rc != 0)
GOTO(stop_trans, rc);
handle = mdd_trans_create(env, mdd);
if (IS_ERR(handle))
- GOTO(out, rc = PTR_ERR(handle));
+ GOTO(out_put, rc = PTR_ERR(handle));
/* Note: this transaction is part of migration, and it is not
* the last step of migration, so we set th_local = 1 to avoid
GOTO(out_put, rc);
}
- rc = mdd_linkea_declare_update_child(env, mdd_tobj,
+ rc = mdd_linkea_declare_update_child(env, mdd_sobj,mdd_tobj,
child, name,
strlen(name),
handle);
GOTO(out_put, rc);
}
- rc = mdd_linkea_update_child(env, mdd_tobj, child, name,
+ rc = mdd_linkea_update_child(env, mdd_sobj, mdd_tobj,
+ child, name,
strlen(name), handle);
out_put:
lname, ldata, handle);
if (rc != 0)
GOTO(stop_trans, rc);
+
+ /* linkea update might decrease the source object
+ * nlink, let's get the attr again after ref_del */
+ rc = mdd_la_get(env, mdd_sobj, so_attr);
+ if (rc != 0)
+ GOTO(stop_trans, rc);
}
if (S_ISREG(so_attr->la_mode)) {
mdd_write_lock(env, mdd_sobj, MOR_TGT_CHILD);
- /* Increase mod_count to add the source object to the orphan list,
- * so if other clients still send RPC to the old object, then these
- * objects can help the request to find the new object, see
- * mdt_reint_open() */
- mdd_sobj->mod_count++;
- rc = mdd_finish_unlink(env, mdd_sobj, ma, mdd_pobj, lname, handle);
- mdd_sobj->mod_count--;
+ mdd_sobj->mod_flags |= DEAD_OBJ;
+ rc = mdd_mark_orphan_object(env, mdd_sobj, handle, false);
+ if (rc != 0)
+ GOTO(out_unlock, rc);
+
+ rc = __mdd_orphan_add(env, mdd_sobj, handle);
if (rc != 0)
GOTO(out_unlock, rc);
RETURN(rc);
}
+static int mdd_fld_lookup(const struct lu_env *env, struct mdd_device *mdd,
+ const struct lu_fid *fid, __u32 *mdt_index)
+{
+ struct lu_seq_range *range = &mdd_env_info(env)->mti_range;
+ struct seq_server_site *ss;
+ int rc;
+
+ ss = mdd->mdd_md_dev.md_lu_dev.ld_site->ld_seq_site;
+
+ range->lsr_flags = LU_SEQ_RANGE_MDT;
+ rc = fld_server_lookup(env, ss->ss_server_fld, fid->f_seq, range);
+ if (rc != 0)
+ return rc;
+
+ *mdt_index = range->lsr_index;
+
+ return 0;
+}
/**
* Check whether we should migrate the file/dir
* return val
{
struct mdd_thread_info *info = mdd_env_info(env);
struct linkea_data *ldata = &info->mti_link_data;
+ struct mdd_device *mdd = mdo2mdd(&pobj->mod_obj);
int mgr_easize;
struct lu_buf *mgr_buf;
int count;
int rc;
-
+ __u64 mdt_index;
ENTRY;
mgr_easize = lmv_mds_md_size(2, LMV_MAGIC_V1);
RETURN(rc);
}
+ mdt_index = mdd->mdd_md_dev.md_lu_dev.ld_site->ld_seq_site->ss_node_id;
/* If there are still links locally, then the file will not be
* migrated. */
LASSERT(ldata->ld_leh != NULL);
ldata->ld_lee = (struct link_ea_entry *)(ldata->ld_leh + 1);
for (count = 0; count < ldata->ld_leh->leh_reccount; count++) {
- struct mdd_device *mdd = mdo2mdd(&sobj->mod_obj);
- struct mdd_object *lpobj;
struct lu_name lname;
struct lu_fid fid;
+ __u32 parent_mdt_index;
linkea_entry_unpack(ldata->ld_lee, &ldata->ld_reclen,
&lname, &fid);
ldata->ld_lee = (struct link_ea_entry *)((char *)ldata->ld_lee +
ldata->ld_reclen);
- lpobj = mdd_object_find(env, mdd, &fid);
- if (IS_ERR(lpobj)) {
- CWARN("%s: cannot find obj "DFID": rc = %ld\n",
- mdd2obd_dev(mdd)->obd_name, PFID(&fid),
- PTR_ERR(lpobj));
- continue;
- }
- if (!mdd_object_exists(lpobj) || mdd_object_remote(lpobj)) {
- CDEBUG(D_INFO, DFID"%.*s: is on remote MDT.\n",
- PFID(&fid), lname.ln_namelen, lname.ln_name);
- mdd_object_put(env, lpobj);
+ rc = mdd_fld_lookup(env, mdd, &fid, &parent_mdt_index);
+ if (rc != 0)
+ RETURN(rc);
+
+ /* Migrate the object only if none of its parents are on the
+ * current MDT. */
+ if (parent_mdt_index != mdt_index)
continue;
- }
CDEBUG(D_INFO, DFID"still has local entry %.*s "DFID"\n",
PFID(mdd_object_fid(sobj)), lname.ln_namelen,
lname.ln_name, PFID(&fid));
- mdd_object_put(env, lpobj);
rc = 1;
break;
}
struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
struct mdd_device *mdd = mdo2mdd(pobj);
struct mdd_object *mdd_sobj = md2mdd_obj(sobj);
- struct mdd_object *mdd_tobj = NULL;
+ struct mdd_object *mdd_tobj = md2mdd_obj(tobj);
struct lu_attr *so_attr = MDD_ENV_VAR(env, cattr);
struct lu_attr *pattr = MDD_ENV_VAR(env, pattr);
int rc;
/* step 1: Check whether the orphan object has been created, and create
* orphan object on the remote MDT if needed */
- mdd_tobj = md2mdd_obj(tobj);
if (!mdd_object_exists(mdd_tobj)) {
rc = mdd_migrate_create(env, mdd_pobj, mdd_sobj, mdd_tobj,
lname, so_attr);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff