1 /* -*- MODE: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mdd) routines
7 * Copyright (C) 2006 Cluster File Systems, Inc.
8 * Author: Wang Di <wangdi@clusterfs.com>
10 * This file is part of the Lustre file system, http://www.lustre.org
11 * Lustre is a trademark of Cluster File Systems, Inc.
13 * You may have signed or agreed to another license before downloading
14 * this software. If so, you are bound by the terms and conditions
15 * of that agreement, and the following does not apply to you. See the
16 * LICENSE file included with this distribution for more information.
18 * If you did not agree to a different license, then this copy of Lustre
19 * is open source software; you can redistribute it and/or modify it
20 * under the terms of version 2 of the GNU General Public License as
21 * published by the Free Software Foundation.
23 * In either case, Lustre is distributed in the hope that it will be
24 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
25 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
26 * license text for more details.
29 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_MDS
33 #include <linux/module.h>
34 #include <linux/jbd.h>
36 #include <obd_class.h>
37 #include <lustre_ver.h>
38 #include <obd_support.h>
39 #include <lprocfs_status.h>
41 #include <linux/ldiskfs_fs.h>
42 #include <lustre_mds.h>
43 #include <lustre/lustre_idl.h>
44 #include <lustre_fid.h>
46 #include "mdd_internal.h"
48 static const char dot[] = ".";
49 static const char dotdot[] = "..";
51 static struct lu_name lname_dotdot = {
56 static int __mdd_lookup(const struct lu_env *env, struct md_object *pobj,
57 const struct lu_name *lname, struct lu_fid* fid,
60 __mdd_lookup_locked(const struct lu_env *env, struct md_object *pobj,
61 const struct lu_name *lname, struct lu_fid* fid, int mask)
63 char *name = lname->ln_name;
64 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
65 struct dynlock_handle *dlh;
68 dlh = mdd_pdo_read_lock(env, mdd_obj, name);
69 if (unlikely(dlh == NULL))
71 rc = __mdd_lookup(env, pobj, lname, fid, mask);
72 mdd_pdo_read_unlock(env, mdd_obj, dlh);
77 static int mdd_lookup(const struct lu_env *env,
78 struct md_object *pobj, const struct lu_name *lname,
79 struct lu_fid* fid, struct md_op_spec *spec)
83 rc = __mdd_lookup_locked(env, pobj, lname, fid, MAY_EXEC);
88 static int mdd_parent_fid(const struct lu_env *env, struct mdd_object *obj,
91 return __mdd_lookup_locked(env, &obj->mod_obj, &lname_dotdot, fid, 0);
95 * For root fid use special function, whcih does not compare version component
96 * of fid. Vresion component is different for root fids on all MDTs.
98 static int mdd_is_root(struct mdd_device *mdd, const struct lu_fid *fid)
100 return fid_seq(&mdd->mdd_root_fid) == fid_seq(fid) &&
101 fid_oid(&mdd->mdd_root_fid) == fid_oid(fid);
105 * return 1: if lf is the fid of the ancestor of p1;
108 * return -EREMOTE: if remote object is found, in this
109 * case fid of remote object is saved to @pf;
111 * otherwise: values < 0, errors.
113 static int mdd_is_parent(const struct lu_env *env,
114 struct mdd_device *mdd,
115 struct mdd_object *p1,
116 const struct lu_fid *lf,
119 struct mdd_object *parent = NULL;
124 LASSERT(!lu_fid_eq(mdo2fid(p1), lf));
125 pfid = &mdd_env_info(env)->mti_fid;
127 /* Check for root first. */
128 if (mdd_is_root(mdd, mdo2fid(p1)))
132 /* this is done recursively, bypass capa for each obj */
133 mdd_set_capainfo(env, 4, p1, BYPASS_CAPA);
134 rc = mdd_parent_fid(env, p1, pfid);
137 if (mdd_is_root(mdd, pfid))
139 if (lu_fid_eq(pfid, lf))
142 mdd_object_put(env, parent);
143 parent = mdd_object_find(env, mdd, pfid);
145 /* cross-ref parent */
146 if (parent == NULL) {
149 GOTO(out, rc = -EREMOTE);
150 } else if (IS_ERR(parent))
151 GOTO(out, rc = PTR_ERR(parent));
156 if (parent && !IS_ERR(parent))
157 mdd_object_put(env, parent);
162 * No permission check is needed.
164 * returns 1: if fid is ancestor of @mo;
165 * returns 0: if fid is not a ancestor of @mo;
167 * returns EREMOTE if remote object is found, fid of remote object is saved to
170 * returns < 0: if error
172 static int mdd_is_subdir(const struct lu_env *env,
173 struct md_object *mo, const struct lu_fid *fid,
176 struct mdd_device *mdd = mdo2mdd(mo);
180 if (!S_ISDIR(mdd_object_type(md2mdd_obj(mo))))
183 rc = mdd_is_parent(env, mdd, md2mdd_obj(mo), fid, sfid);
187 } else if (rc == 1) {
188 /* found @fid is parent */
196 * Check that @dir contains no entries except (possibly) dot and dotdot.
201 * -ENOTDIR not a directory object
202 * -ENOTEMPTY not empty
206 static int mdd_dir_is_empty(const struct lu_env *env,
207 struct mdd_object *dir)
210 struct dt_object *obj;
211 struct dt_it_ops *iops;
215 obj = mdd_object_child(dir);
216 if (!dt_try_as_dir(env, obj))
219 iops = &obj->do_index_ops->dio_it;
220 it = iops->init(env, obj, 0, BYPASS_CAPA);
222 result = iops->get(env, it, (const void *)"");
225 for (result = 0, i = 0; result == 0 && i < 3; ++i)
226 result = iops->next(env, it);
229 else if (result == +1)
231 } else if (result == 0)
233 * Huh? Index contains no zero key?
244 static int __mdd_may_link(const struct lu_env *env, struct mdd_object *obj)
246 struct mdd_device *m = mdd_obj2mdd_dev(obj);
247 struct lu_attr *la = &mdd_env_info(env)->mti_la;
251 rc = mdd_la_get(env, obj, la, BYPASS_CAPA);
255 if (la->la_nlink >= m->mdd_dt_conf.ddp_max_nlink)
262 * Check whether it may create the cobj under the pobj.
265 int mdd_may_create(const struct lu_env *env, struct mdd_object *pobj,
266 struct mdd_object *cobj, int check_perm, int check_nlink)
271 if (cobj && mdd_object_exists(cobj))
274 if (mdd_is_dead_obj(pobj))
278 rc = mdd_permission_internal_locked(env, pobj, NULL,
279 MAY_WRITE | MAY_EXEC);
281 if (!rc && check_nlink)
282 rc = __mdd_may_link(env, pobj);
288 * Check whether can unlink from the pobj in the case of "cobj == NULL".
290 int mdd_may_unlink(const struct lu_env *env, struct mdd_object *pobj,
291 const struct md_attr *ma)
296 if (mdd_is_dead_obj(pobj))
299 if ((ma->ma_attr.la_valid & LA_FLAGS) &&
300 (ma->ma_attr.la_flags & (LUSTRE_APPEND_FL | LUSTRE_IMMUTABLE_FL)))
303 rc = mdd_permission_internal_locked(env, pobj, NULL,
304 MAY_WRITE | MAY_EXEC);
308 if (mdd_is_append(pobj))
315 * pobj == NULL is remote ops case, under such case, pobj's
316 * VTX feature has been checked already, no need check again.
318 static inline int mdd_is_sticky(const struct lu_env *env,
319 struct mdd_object *pobj,
320 struct mdd_object *cobj)
322 struct lu_attr *tmp_la = &mdd_env_info(env)->mti_la;
323 struct md_ucred *uc = md_ucred(env);
327 rc = mdd_la_get(env, pobj, tmp_la, BYPASS_CAPA);
331 if (!(tmp_la->la_mode & S_ISVTX) ||
332 (tmp_la->la_uid == uc->mu_fsuid))
336 rc = mdd_la_get(env, cobj, tmp_la, BYPASS_CAPA);
340 if (tmp_la->la_uid == uc->mu_fsuid)
343 return !mdd_capable(uc, CAP_FOWNER);
347 * Check whether it may delete the cobj from the pobj.
350 int mdd_may_delete(const struct lu_env *env, struct mdd_object *pobj,
351 struct mdd_object *cobj, struct md_attr *ma,
352 int check_perm, int check_empty)
358 if (!mdd_object_exists(cobj))
362 if (mdd_is_dead_obj(pobj))
366 rc = mdd_permission_internal_locked(env, pobj, NULL,
367 MAY_WRITE | MAY_EXEC);
372 if (mdd_is_append(pobj))
376 if (!(ma->ma_attr_flags & MDS_VTX_BYPASS) &&
377 mdd_is_sticky(env, pobj, cobj))
380 if (mdd_is_immutable(cobj) || mdd_is_append(cobj))
383 if ((ma->ma_attr.la_valid & LA_FLAGS) &&
384 (ma->ma_attr.la_flags & (LUSTRE_APPEND_FL | LUSTRE_IMMUTABLE_FL)))
387 if (S_ISDIR(ma->ma_attr.la_mode)) {
388 struct mdd_device *mdd = mdo2mdd(&cobj->mod_obj);
390 if (!S_ISDIR(mdd_object_type(cobj)))
393 if (lu_fid_eq(mdo2fid(cobj), &mdd->mdd_root_fid))
395 } else if (S_ISDIR(mdd_object_type(cobj)))
398 if (S_ISDIR(ma->ma_attr.la_mode) && check_empty)
399 rc = mdd_dir_is_empty(env, cobj);
406 * has mdd_write_lock on src already, but not on tgt yet
408 int mdd_link_sanity_check(const struct lu_env *env,
409 struct mdd_object *tgt_obj,
410 const struct lu_name *lname,
411 struct mdd_object *src_obj)
413 struct mdd_device *m = mdd_obj2mdd_dev(src_obj);
417 /* Local ops, no lookup before link, check filename length here. */
418 if (lname && (lname->ln_namelen > m->mdd_dt_conf.ddp_max_name_len))
419 RETURN(-ENAMETOOLONG);
421 if (mdd_is_immutable(src_obj) || mdd_is_append(src_obj))
424 if (S_ISDIR(mdd_object_type(src_obj)))
427 LASSERT(src_obj != tgt_obj);
429 rc = mdd_may_create(env, tgt_obj, NULL, 1, 0);
434 rc = __mdd_may_link(env, src_obj);
439 const struct dt_rec *__mdd_fid_rec(const struct lu_env *env,
440 const struct lu_fid *fid)
442 struct lu_fid_pack *pack = &mdd_env_info(env)->mti_pack;
444 fid_pack(pack, fid, &mdd_env_info(env)->mti_fid2);
445 return (const struct dt_rec *)pack;
449 /* insert named index, add reference if isdir */
450 static int __mdd_index_insert(const struct lu_env *env, struct mdd_object *pobj,
451 const struct lu_fid *lf, const char *name, int is_dir,
452 struct thandle *handle, struct lustre_capa *capa)
454 struct dt_object *next = mdd_object_child(pobj);
458 if (dt_try_as_dir(env, next)) {
459 rc = next->do_index_ops->dio_insert(env, next,
460 __mdd_fid_rec(env, lf),
461 (const struct dt_key *)name,
469 mdd_write_lock(env, pobj);
470 mdo_ref_add(env, pobj, handle);
471 mdd_write_unlock(env, pobj);
477 /* delete named index, drop reference if isdir */
478 static int __mdd_index_delete(const struct lu_env *env, struct mdd_object *pobj,
479 const char *name, int is_dir, struct thandle *handle,
480 struct lustre_capa *capa)
482 struct dt_object *next = mdd_object_child(pobj);
486 if (dt_try_as_dir(env, next)) {
487 rc = next->do_index_ops->dio_delete(env, next,
488 (struct dt_key *)name,
490 if (rc == 0 && is_dir) {
491 mdd_write_lock(env, pobj);
492 mdo_ref_del(env, pobj, handle);
493 mdd_write_unlock(env, pobj);
502 __mdd_index_insert_only(const struct lu_env *env, struct mdd_object *pobj,
503 const struct lu_fid *lf, const char *name,
504 struct thandle *handle, struct lustre_capa *capa)
506 struct dt_object *next = mdd_object_child(pobj);
510 if (dt_try_as_dir(env, next)) {
511 rc = next->do_index_ops->dio_insert(env, next,
512 __mdd_fid_rec(env, lf),
513 (const struct dt_key *)name,
521 static int mdd_link(const struct lu_env *env, struct md_object *tgt_obj,
522 struct md_object *src_obj, const struct lu_name *lname,
525 char *name = lname->ln_name;
526 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
527 struct mdd_object *mdd_tobj = md2mdd_obj(tgt_obj);
528 struct mdd_object *mdd_sobj = md2mdd_obj(src_obj);
529 struct mdd_device *mdd = mdo2mdd(src_obj);
530 struct dynlock_handle *dlh;
531 struct thandle *handle;
535 mdd_txn_param_build(env, mdd, MDD_TXN_LINK_OP);
536 handle = mdd_trans_start(env, mdd);
538 RETURN(PTR_ERR(handle));
540 dlh = mdd_pdo_write_lock(env, mdd_tobj, name);
542 GOTO(out_trans, rc = -ENOMEM);
543 mdd_write_lock(env, mdd_sobj);
545 rc = mdd_link_sanity_check(env, mdd_tobj, lname, mdd_sobj);
547 GOTO(out_unlock, rc);
549 rc = __mdd_index_insert_only(env, mdd_tobj, mdo2fid(mdd_sobj),
551 mdd_object_capa(env, mdd_tobj));
553 GOTO(out_unlock, rc);
555 mdo_ref_add(env, mdd_sobj, handle);
557 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
558 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
560 la->la_valid = LA_CTIME | LA_MTIME;
561 rc = mdd_attr_set_internal_locked(env, mdd_tobj, la, handle, 0);
563 GOTO(out_unlock, rc);
565 la->la_valid = LA_CTIME;
566 rc = mdd_attr_set_internal(env, mdd_sobj, la, handle, 0);
569 mdd_write_unlock(env, mdd_sobj);
570 mdd_pdo_write_unlock(env, mdd_tobj, dlh);
572 mdd_trans_stop(env, mdd, rc, handle);
576 /* caller should take a lock before calling */
577 int mdd_finish_unlink(const struct lu_env *env,
578 struct mdd_object *obj, struct md_attr *ma,
584 rc = mdd_iattr_get(env, obj, ma);
585 if (rc == 0 && ma->ma_attr.la_nlink == 0) {
586 /* add new orphan and the object
587 * will be deleted during the object_put() */
588 if (__mdd_orphan_add(env, obj, th) == 0)
589 obj->mod_flags |= ORPHAN_OBJ;
591 obj->mod_flags |= DEAD_OBJ;
592 if (obj->mod_count == 0)
593 rc = mdd_object_kill(env, obj, ma);
595 /* clear MA_LOV | MA_COOKIE, if we do not
596 * unlink it in case we get it somewhere */
597 ma->ma_valid &= ~(MA_LOV | MA_COOKIE);
599 ma->ma_valid &= ~(MA_LOV | MA_COOKIE);
606 * has mdd_write_lock on cobj already, but not on pobj yet
608 int mdd_unlink_sanity_check(const struct lu_env *env, struct mdd_object *pobj,
609 struct mdd_object *cobj, struct md_attr *ma)
614 rc = mdd_may_delete(env, pobj, cobj, ma, 1, 1);
619 static int mdd_unlink(const struct lu_env *env, struct md_object *pobj,
620 struct md_object *cobj, const struct lu_name *lname,
623 char *name = lname->ln_name;
624 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
625 struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
626 struct mdd_object *mdd_cobj = md2mdd_obj(cobj);
627 struct mdd_device *mdd = mdo2mdd(pobj);
628 struct dynlock_handle *dlh;
629 struct thandle *handle;
633 LASSERTF(mdd_object_exists(mdd_cobj) > 0, "FID is "DFID"\n",
634 PFID(mdd_object_fid(mdd_cobj)));
636 rc = mdd_log_txn_param_build(env, cobj, ma, MDD_TXN_UNLINK_OP);
640 handle = mdd_trans_start(env, mdd);
642 RETURN(PTR_ERR(handle));
645 dlh = mdd_pdo_write_lock(env, mdd_pobj, name);
647 GOTO(out_trans, rc = -ENOMEM);
648 mdd_write_lock(env, mdd_cobj);
650 is_dir = S_ISDIR(ma->ma_attr.la_mode);
651 rc = mdd_unlink_sanity_check(env, mdd_pobj, mdd_cobj, ma);
655 rc = __mdd_index_delete(env, mdd_pobj, name, is_dir, handle,
656 mdd_object_capa(env, mdd_pobj));
660 mdo_ref_del(env, mdd_cobj, handle);
663 mdo_ref_del(env, mdd_cobj, handle);
665 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
666 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
668 la->la_valid = LA_CTIME | LA_MTIME;
669 rc = mdd_attr_set_internal_locked(env, mdd_pobj, la, handle, 0);
673 la->la_valid = LA_CTIME;
674 rc = mdd_attr_set_internal(env, mdd_cobj, la, handle, 0);
678 rc = mdd_finish_unlink(env, mdd_cobj, ma, handle);
681 obd_set_info_async(mdd2obd_dev(mdd)->u.mds.mds_osc_exp,
682 strlen("unlinked"), "unlinked", 0,
686 mdd_write_unlock(env, mdd_cobj);
687 mdd_pdo_write_unlock(env, mdd_pobj, dlh);
689 mdd_trans_stop(env, mdd, rc, handle);
693 /* has not lock on pobj yet */
694 static int mdd_ni_sanity_check(const struct lu_env *env,
695 struct md_object *pobj,
696 const struct md_attr *ma)
698 struct mdd_object *obj = md2mdd_obj(pobj);
702 if (ma->ma_attr_flags & MDS_PERM_BYPASS)
705 rc = mdd_may_create(env, obj, NULL, 1, S_ISDIR(ma->ma_attr.la_mode));
713 static int mdd_name_insert(const struct lu_env *env,
714 struct md_object *pobj,
715 const struct lu_name *lname,
716 const struct lu_fid *fid,
717 const struct md_attr *ma)
719 char *name = lname->ln_name;
720 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
721 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
722 struct mdd_device *mdd = mdo2mdd(pobj);
723 struct dynlock_handle *dlh;
724 struct thandle *handle;
725 int is_dir = S_ISDIR(ma->ma_attr.la_mode);
729 mdd_txn_param_build(env, mdd, MDD_TXN_INDEX_INSERT_OP);
730 handle = mdd_trans_start(env, mdo2mdd(pobj));
732 RETURN(PTR_ERR(handle));
734 dlh = mdd_pdo_write_lock(env, mdd_obj, name);
736 GOTO(out_trans, rc = -ENOMEM);
738 rc = mdd_ni_sanity_check(env, pobj, ma);
740 GOTO(out_unlock, rc);
742 rc = __mdd_index_insert(env, mdd_obj, fid, name, is_dir,
743 handle, BYPASS_CAPA);
745 GOTO(out_unlock, rc);
748 * For some case, no need update obj's ctime (LA_CTIME is not set),
750 * For other cases, update obj's ctime (LA_CTIME is set),
753 if (ma->ma_attr.la_valid & LA_CTIME) {
754 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
755 la->la_valid = LA_CTIME | LA_MTIME;
756 rc = mdd_attr_set_internal_locked(env, mdd_obj, la, handle, 0);
760 mdd_pdo_write_unlock(env, mdd_obj, dlh);
762 mdd_trans_stop(env, mdo2mdd(pobj), rc, handle);
766 /* has not lock on pobj yet */
767 static int mdd_nr_sanity_check(const struct lu_env *env,
768 struct md_object *pobj,
769 const struct md_attr *ma)
771 struct mdd_object *obj = md2mdd_obj(pobj);
775 if (ma->ma_attr_flags & MDS_PERM_BYPASS)
778 rc = mdd_may_unlink(env, obj, ma);
786 static int mdd_name_remove(const struct lu_env *env,
787 struct md_object *pobj,
788 const struct lu_name *lname,
789 const struct md_attr *ma)
791 char *name = lname->ln_name;
792 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
793 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
794 struct mdd_device *mdd = mdo2mdd(pobj);
795 struct dynlock_handle *dlh;
796 struct thandle *handle;
797 int is_dir = S_ISDIR(ma->ma_attr.la_mode);
801 mdd_txn_param_build(env, mdd, MDD_TXN_INDEX_DELETE_OP);
802 handle = mdd_trans_start(env, mdd);
804 RETURN(PTR_ERR(handle));
806 dlh = mdd_pdo_write_lock(env, mdd_obj, name);
808 GOTO(out_trans, rc = -ENOMEM);
810 rc = mdd_nr_sanity_check(env, pobj, ma);
812 GOTO(out_unlock, rc);
814 rc = __mdd_index_delete(env, mdd_obj, name, is_dir,
815 handle, BYPASS_CAPA);
817 GOTO(out_unlock, rc);
820 * For some case, no need update obj's ctime (LA_CTIME is not set),
822 * For other cases, update obj's ctime (LA_CTIME is set),
825 if (ma->ma_attr.la_valid & LA_CTIME) {
826 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
827 la->la_valid = LA_CTIME | LA_MTIME;
828 rc = mdd_attr_set_internal_locked(env, mdd_obj, la, handle, 0);
832 mdd_pdo_write_unlock(env, mdd_obj, dlh);
834 mdd_trans_stop(env, mdd, rc, handle);
840 * has mdd_write_lock on tobj alreay, but not on tgt_pobj yet
842 static int mdd_rt_sanity_check(const struct lu_env *env,
843 struct mdd_object *tgt_pobj,
844 struct mdd_object *tobj,
850 if (unlikely(ma->ma_attr_flags & MDS_PERM_BYPASS))
853 /* XXX: for mdd_rename_tgt, "tobj == NULL" does not mean tobj not
854 * exist. In fact, tobj must exist, otherwise the call trace will be:
855 * mdt_reint_rename_tgt -> mdo_name_insert -> ... -> mdd_name_insert.
856 * When get here, tobj must be NOT NULL, the other case has been
857 * processed in cmr_rename_tgt before mdd_rename_tgt and enable
859 * So check may_delete, but not check nlink of tgt_pobj. */
861 rc = mdd_may_delete(env, tgt_pobj, tobj, ma, 1, 1);
866 static int mdd_rename_tgt(const struct lu_env *env,
867 struct md_object *pobj, struct md_object *tobj,
868 const struct lu_fid *lf, const struct lu_name *lname,
871 char *name = lname->ln_name;
872 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
873 struct mdd_object *mdd_tpobj = md2mdd_obj(pobj);
874 struct mdd_object *mdd_tobj = md2mdd_obj(tobj);
875 struct mdd_device *mdd = mdo2mdd(pobj);
876 struct dynlock_handle *dlh;
877 struct thandle *handle;
881 mdd_txn_param_build(env, mdd, MDD_TXN_RENAME_TGT_OP);
882 handle = mdd_trans_start(env, mdd);
884 RETURN(PTR_ERR(handle));
886 dlh = mdd_pdo_write_lock(env, mdd_tpobj, name);
888 GOTO(out_trans, rc = -ENOMEM);
890 mdd_write_lock(env, mdd_tobj);
892 rc = mdd_rt_sanity_check(env, mdd_tpobj, mdd_tobj, ma);
897 * If rename_tgt is called then we should just re-insert name with
898 * correct fid, no need to dec/inc parent nlink if obj is dir.
900 rc = __mdd_index_delete(env, mdd_tpobj, name, 0, handle, BYPASS_CAPA);
904 rc = __mdd_index_insert_only(env, mdd_tpobj, lf, name, handle,
909 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
910 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
912 la->la_valid = LA_CTIME | LA_MTIME;
913 rc = mdd_attr_set_internal_locked(env, mdd_tpobj, la, handle, 0);
918 * For tobj is remote case cmm layer has processed
919 * and pass NULL tobj to here. So when tobj is NOT NULL,
920 * it must be local one.
922 if (tobj && mdd_object_exists(mdd_tobj)) {
923 mdo_ref_del(env, mdd_tobj, handle);
925 /* Remove dot reference. */
926 if (S_ISDIR(ma->ma_attr.la_mode))
927 mdo_ref_del(env, mdd_tobj, handle);
929 la->la_valid = LA_CTIME;
930 rc = mdd_attr_set_internal(env, mdd_tobj, la, handle, 0);
934 rc = mdd_finish_unlink(env, mdd_tobj, ma, handle);
941 mdd_write_unlock(env, mdd_tobj);
942 mdd_pdo_write_unlock(env, mdd_tpobj, dlh);
944 mdd_trans_stop(env, mdd, rc, handle);
949 * The permission has been checked when obj created, no need check again.
951 static int mdd_cd_sanity_check(const struct lu_env *env,
952 struct mdd_object *obj)
957 if (!obj || mdd_is_dead_obj(obj))
964 static int mdd_create_data(const struct lu_env *env, struct md_object *pobj,
965 struct md_object *cobj, const struct md_op_spec *spec,
968 struct mdd_device *mdd = mdo2mdd(cobj);
969 struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
970 struct mdd_object *son = md2mdd_obj(cobj);
971 struct lu_attr *attr = &ma->ma_attr;
972 struct lov_mds_md *lmm = NULL;
974 struct thandle *handle;
978 rc = mdd_cd_sanity_check(env, son);
982 if (spec->sp_cr_flags & MDS_OPEN_DELAY_CREATE ||
983 !(spec->sp_cr_flags & FMODE_WRITE))
986 rc = mdd_lov_create(env, mdd, mdd_pobj, son, &lmm, &lmm_size,
991 mdd_txn_param_build(env, mdd, MDD_TXN_CREATE_DATA_OP);
992 handle = mdd_trans_start(env, mdd);
994 GOTO(out_free, rc = PTR_ERR(handle));
997 * XXX: Setting the lov ea is not locked but setting the attr is locked?
998 * Should this be fixed?
1001 /* Replay creates has objects already */
1003 if (spec->u.sp_ea.no_lov_create) {
1004 CDEBUG(D_INFO, "we already have lov ea\n");
1005 rc = mdd_lov_set_md(env, mdd_pobj, son,
1006 (struct lov_mds_md *)spec->u.sp_ea.eadata,
1007 spec->u.sp_ea.eadatalen, handle, 0);
1010 /* No need mdd_lsm_sanity_check here */
1011 rc = mdd_lov_set_md(env, mdd_pobj, son, lmm,
1012 lmm_size, handle, 0);
1015 rc = mdd_attr_get_internal_locked(env, son, ma);
1017 /* update lov_objid data, must be before transaction stop! */
1019 mdd_lov_objid_update(env, mdd);
1021 mdd_trans_stop(env, mdd, rc, handle);
1023 /* Finish mdd_lov_create() stuff. */
1024 mdd_lov_create_finish(env, mdd, lmm, lmm_size, spec);
1029 __mdd_lookup(const struct lu_env *env, struct md_object *pobj,
1030 const struct lu_name *lname, struct lu_fid* fid, int mask)
1032 char *name = lname->ln_name;
1033 const struct dt_key *key = (const struct dt_key *)name;
1034 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
1035 struct mdd_device *m = mdo2mdd(pobj);
1036 struct dt_object *dir = mdd_object_child(mdd_obj);
1037 struct lu_fid_pack *pack = &mdd_env_info(env)->mti_pack;
1041 if (unlikely(mdd_is_dead_obj(mdd_obj)))
1044 rc = mdd_object_exists(mdd_obj);
1045 if (unlikely(rc == 0))
1047 else if (unlikely(rc < 0)) {
1048 CERROR("Object "DFID" locates on remote server\n",
1049 PFID(mdo2fid(mdd_obj)));
1053 /* The common filename length check. */
1054 if (unlikely(lname->ln_namelen > m->mdd_dt_conf.ddp_max_name_len))
1055 RETURN(-ENAMETOOLONG);
1057 rc = mdd_permission_internal_locked(env, mdd_obj, NULL, mask);
1061 if (likely(S_ISDIR(mdd_object_type(mdd_obj)) &&
1062 dt_try_as_dir(env, dir))) {
1063 rc = dir->do_index_ops->dio_lookup(env, dir,
1064 (struct dt_rec *)pack, key,
1065 mdd_object_capa(env, mdd_obj));
1067 fid_unpack(pack, fid);
1074 int mdd_object_initialize(const struct lu_env *env, const struct lu_fid *pfid,
1075 struct mdd_object *child, struct md_attr *ma,
1076 struct thandle *handle)
1082 * Update attributes for child.
1085 * (1) the valid bits should be converted between Lustre and Linux;
1086 * (2) maybe, the child attributes should be set in OSD when creation.
1089 rc = mdd_attr_set_internal(env, child, &ma->ma_attr, handle, 0);
1093 if (S_ISDIR(ma->ma_attr.la_mode)) {
1094 /* Add "." and ".." for newly created dir */
1095 mdo_ref_add(env, child, handle);
1096 rc = __mdd_index_insert_only(env, child, mdo2fid(child),
1097 dot, handle, BYPASS_CAPA);
1099 rc = __mdd_index_insert_only(env, child, pfid,
1105 rc2 = __mdd_index_delete(env, child, dot, 0,
1106 handle, BYPASS_CAPA);
1108 CERROR("Failure to cleanup after dotdot"
1109 " creation: %d (%d)\n", rc2, rc);
1111 mdo_ref_del(env, child, handle);
1118 /* has not lock on pobj yet */
1119 static int mdd_create_sanity_check(const struct lu_env *env,
1120 struct md_object *pobj,
1121 const struct lu_name *lname,
1123 struct md_op_spec *spec)
1125 struct mdd_thread_info *info = mdd_env_info(env);
1126 struct lu_attr *la = &info->mti_la;
1127 struct lu_fid *fid = &info->mti_fid;
1128 struct mdd_object *obj = md2mdd_obj(pobj);
1129 struct mdd_device *m = mdo2mdd(pobj);
1130 int lookup = spec->sp_cr_lookup;
1135 if (mdd_is_dead_obj(obj))
1139 * In some cases this lookup is not needed - we know before if name
1140 * exists or not because MDT performs lookup for it.
1141 * name length check is done in lookup.
1145 * Check if the name already exist, though it will be checked in
1146 * _index_insert also, for avoiding rolling back if exists
1149 rc = __mdd_lookup_locked(env, pobj, lname, fid,
1150 MAY_WRITE | MAY_EXEC);
1152 RETURN(rc ? : -EEXIST);
1155 * Check WRITE permission for the parent.
1156 * EXEC permission have been checked
1157 * when lookup before create already.
1159 rc = mdd_permission_internal_locked(env, obj, NULL, MAY_WRITE);
1165 rc = mdd_la_get(env, obj, la, BYPASS_CAPA);
1169 if (la->la_mode & S_ISGID) {
1170 ma->ma_attr.la_gid = la->la_gid;
1171 if (S_ISDIR(ma->ma_attr.la_mode)) {
1172 ma->ma_attr.la_mode |= S_ISGID;
1173 ma->ma_attr.la_valid |= LA_MODE;
1177 switch (ma->ma_attr.la_mode & S_IFMT) {
1179 if (la->la_nlink >= m->mdd_dt_conf.ddp_max_nlink)
1185 unsigned int symlen = strlen(spec->u.sp_symname) + 1;
1187 if (symlen > (1 << m->mdd_dt_conf.ddp_block_shift))
1188 RETURN(-ENAMETOOLONG);
1207 * Create object and insert it into namespace.
1209 static int mdd_create(const struct lu_env *env,
1210 struct md_object *pobj,
1211 const struct lu_name *lname,
1212 struct md_object *child,
1213 struct md_op_spec *spec,
1216 char *name = lname->ln_name;
1217 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
1218 struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
1219 struct mdd_object *son = md2mdd_obj(child);
1220 struct mdd_device *mdd = mdo2mdd(pobj);
1221 struct lu_attr *attr = &ma->ma_attr;
1222 struct lov_mds_md *lmm = NULL;
1223 struct thandle *handle;
1224 int rc, created = 0, inserted = 0, lmm_size = 0;
1225 struct dynlock_handle *dlh;
1229 * Two operations have to be performed:
1231 * - allocation of new object (->do_create()), and
1233 * - insertion into parent index (->dio_insert()).
1235 * Due to locking, operation order is not important, when both are
1236 * successful, *but* error handling cases are quite different:
1238 * - if insertion is done first, and following object creation fails,
1239 * insertion has to be rolled back, but this operation might fail
1240 * also leaving us with dangling index entry.
1242 * - if creation is done first, is has to be undone if insertion
1243 * fails, leaving us with leaked space, which is neither good, nor
1246 * It seems that creation-first is simplest solution, but it is
1247 * sub-optimal in the frequent
1252 * case, because second mkdir is bound to create object, only to
1253 * destroy it immediately.
1255 * To avoid this follow local file systems that do double lookup:
1257 * 0. lookup -> -EEXIST (mdd_create_sanity_check())
1259 * 1. create (mdd_object_create_internal())
1261 * 2. insert (__mdd_index_insert(), lookup again)
1264 /* Sanity checks before big job. */
1265 rc = mdd_create_sanity_check(env, pobj, lname, ma, spec);
1270 * No RPC inside the transaction, so OST objects should be created at
1273 if (S_ISREG(attr->la_mode)) {
1274 rc = mdd_lov_create(env, mdd, mdd_pobj, son, &lmm, &lmm_size,
1280 mdd_txn_param_build(env, mdd, MDD_TXN_MKDIR_OP);
1281 handle = mdd_trans_start(env, mdd);
1283 GOTO(out_free, rc = PTR_ERR(handle));
1285 dlh = mdd_pdo_write_lock(env, mdd_pobj, name);
1287 GOTO(out_trans, rc = -ENOMEM);
1290 * XXX: Check that link can be added to the parent in mkdir case.
1293 mdd_write_lock(env, son);
1294 rc = mdd_object_create_internal(env, mdd_pobj, son, ma, handle);
1296 mdd_write_unlock(env, son);
1302 #ifdef CONFIG_FS_POSIX_ACL
1303 mdd_read_lock(env, mdd_pobj);
1304 rc = mdd_acl_init(env, mdd_pobj, son, &ma->ma_attr.la_mode, handle);
1305 mdd_read_unlock(env, mdd_pobj);
1307 mdd_write_unlock(env, son);
1310 ma->ma_attr.la_valid |= LA_MODE;
1314 rc = mdd_object_initialize(env, mdo2fid(mdd_pobj),
1316 mdd_write_unlock(env, son);
1319 * Object has no links, so it will be destroyed when last
1320 * reference is released. (XXX not now.)
1324 rc = __mdd_index_insert(env, mdd_pobj, mdo2fid(son),
1325 name, S_ISDIR(attr->la_mode), handle,
1326 mdd_object_capa(env, mdd_pobj));
1333 /* No need mdd_lsm_sanity_check here */
1334 rc = mdd_lov_set_md(env, mdd_pobj, son, lmm, lmm_size, handle, 0);
1336 CERROR("error on stripe info copy %d \n", rc);
1339 if (lmm && lmm_size > 0) {
1340 /* Set Lov here, do not get lmm again later */
1341 memcpy(ma->ma_lmm, lmm, lmm_size);
1342 ma->ma_lmm_size = lmm_size;
1343 ma->ma_valid |= MA_LOV;
1346 if (S_ISLNK(attr->la_mode)) {
1347 struct dt_object *dt = mdd_object_child(son);
1348 const char *target_name = spec->u.sp_symname;
1349 int sym_len = strlen(target_name);
1350 const struct lu_buf *buf;
1353 buf = mdd_buf_get_const(env, target_name, sym_len);
1354 rc = dt->do_body_ops->dbo_write(env, dt, buf, &pos, handle,
1355 mdd_object_capa(env, son));
1360 GOTO(cleanup, rc = -EFAULT);
1364 la->la_valid = LA_CTIME | LA_MTIME;
1365 rc = mdd_attr_set_internal_locked(env, mdd_pobj, la, handle, 0);
1369 /* Return attr back. */
1370 rc = mdd_attr_get_internal_locked(env, son, ma);
1373 if (rc && created) {
1377 rc2 = __mdd_index_delete(env, mdd_pobj, name,
1378 S_ISDIR(attr->la_mode),
1379 handle, BYPASS_CAPA);
1381 CERROR("error can not cleanup destroy %d\n",
1385 mdd_write_lock(env, son);
1386 mdo_ref_del(env, son, handle);
1387 mdd_write_unlock(env, son);
1391 /* update lov_objid data, must be before transaction stop! */
1393 mdd_lov_objid_update(env, mdd);
1395 mdd_pdo_write_unlock(env, mdd_pobj, dlh);
1397 mdd_trans_stop(env, mdd, rc, handle);
1399 /* finis lov_create stuff, free all temporary data */
1400 mdd_lov_create_finish(env, mdd, lmm, lmm_size, spec);
1405 * Get locks on parents in proper order
1406 * RETURN: < 0 - error, rename_order if successful
1414 static int mdd_rename_order(const struct lu_env *env,
1415 struct mdd_device *mdd,
1416 struct mdd_object *src_pobj,
1417 struct mdd_object *tgt_pobj)
1419 /* order of locking, 1 - tgt-src, 0 - src-tgt*/
1423 if (src_pobj == tgt_pobj)
1424 RETURN(MDD_RN_SAME);
1426 /* compared the parent child relationship of src_p&tgt_p */
1427 if (lu_fid_eq(&mdd->mdd_root_fid, mdo2fid(src_pobj))){
1429 } else if (lu_fid_eq(&mdd->mdd_root_fid, mdo2fid(tgt_pobj))) {
1432 rc = mdd_is_parent(env, mdd, src_pobj, mdo2fid(tgt_pobj), NULL);
1445 /* has not mdd_write{read}_lock on any obj yet. */
1446 static int mdd_rename_sanity_check(const struct lu_env *env,
1447 struct mdd_object *src_pobj,
1448 struct mdd_object *tgt_pobj,
1449 struct mdd_object *sobj,
1450 struct mdd_object *tobj,
1456 if (unlikely(ma->ma_attr_flags & MDS_PERM_BYPASS))
1459 /* XXX: when get here, sobj must NOT be NULL,
1460 * the other case has been processed in cml_rename
1461 * before mdd_rename and enable MDS_PERM_BYPASS. */
1463 rc = mdd_may_delete(env, src_pobj, sobj, ma, 1, 0);
1467 /* XXX: when get here, "tobj == NULL" means tobj must
1468 * NOT exist (neither on remote MDS, such case has been
1469 * processed in cml_rename before mdd_rename and enable
1471 * So check may_create, but not check may_unlink. */
1473 rc = mdd_may_create(env, tgt_pobj, NULL,
1474 (src_pobj != tgt_pobj), 0);
1476 rc = mdd_may_delete(env, tgt_pobj, tobj, ma,
1477 (src_pobj != tgt_pobj), 1);
1479 if (!rc && !tobj && (src_pobj != tgt_pobj) &&
1480 S_ISDIR(ma->ma_attr.la_mode))
1481 rc = __mdd_may_link(env, tgt_pobj);
1486 /* src object can be remote that is why we use only fid and type of object */
1487 static int mdd_rename(const struct lu_env *env,
1488 struct md_object *src_pobj, struct md_object *tgt_pobj,
1489 const struct lu_fid *lf, const struct lu_name *lsname,
1490 struct md_object *tobj, const struct lu_name *ltname,
1493 char *sname = lsname->ln_name;
1494 char *tname = ltname->ln_name;
1495 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
1496 struct mdd_object *mdd_spobj = md2mdd_obj(src_pobj);
1497 struct mdd_object *mdd_tpobj = md2mdd_obj(tgt_pobj);
1498 struct mdd_device *mdd = mdo2mdd(src_pobj);
1499 struct mdd_object *mdd_sobj = NULL;
1500 struct mdd_object *mdd_tobj = NULL;
1501 struct dynlock_handle *sdlh, *tdlh;
1502 struct thandle *handle;
1507 LASSERT(ma->ma_attr.la_mode & S_IFMT);
1508 is_dir = S_ISDIR(ma->ma_attr.la_mode);
1511 mdd_tobj = md2mdd_obj(tobj);
1513 mdd_txn_param_build(env, mdd, MDD_TXN_RENAME_OP);
1514 handle = mdd_trans_start(env, mdd);
1516 RETURN(PTR_ERR(handle));
1518 /* FIXME: Should consider tobj and sobj too in rename_lock. */
1519 rc = mdd_rename_order(env, mdd, mdd_spobj, mdd_tpobj);
1521 GOTO(cleanup_unlocked, rc);
1523 /* Get locks in determined order */
1524 if (rc == MDD_RN_SAME) {
1525 sdlh = mdd_pdo_write_lock(env, mdd_spobj, sname);
1526 /* check hashes to determine do we need one lock or two */
1527 if (mdd_name2hash(sname) != mdd_name2hash(tname))
1528 tdlh = mdd_pdo_write_lock(env, mdd_tpobj, tname);
1531 } else if (rc == MDD_RN_SRCTGT) {
1532 sdlh = mdd_pdo_write_lock(env, mdd_spobj, sname);
1533 tdlh = mdd_pdo_write_lock(env, mdd_tpobj, tname);
1535 tdlh = mdd_pdo_write_lock(env, mdd_tpobj, tname);
1536 sdlh = mdd_pdo_write_lock(env, mdd_spobj, sname);
1538 if (sdlh == NULL || tdlh == NULL)
1539 GOTO(cleanup, rc = -ENOMEM);
1541 mdd_sobj = mdd_object_find(env, mdd, lf);
1542 rc = mdd_rename_sanity_check(env, mdd_spobj, mdd_tpobj,
1543 mdd_sobj, mdd_tobj, ma);
1547 rc = __mdd_index_delete(env, mdd_spobj, sname, is_dir, handle,
1548 mdd_object_capa(env, mdd_spobj));
1553 * Here tobj can be remote one, so we do index_delete unconditionally
1554 * and -ENOENT is allowed.
1556 rc = __mdd_index_delete(env, mdd_tpobj, tname, is_dir, handle,
1557 mdd_object_capa(env, mdd_tpobj));
1558 if (rc != 0 && rc != -ENOENT)
1561 rc = __mdd_index_insert(env, mdd_tpobj, lf, tname, is_dir, handle,
1562 mdd_object_capa(env, mdd_tpobj));
1566 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
1567 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
1569 /* XXX: mdd_sobj must be local one if it is NOT NULL. */
1571 la->la_valid = LA_CTIME;
1572 rc = mdd_attr_set_internal_locked(env, mdd_sobj, la, handle, 0);
1578 * For tobj is remote case cmm layer has processed
1579 * and set tobj to NULL then. So when tobj is NOT NULL,
1580 * it must be local one.
1582 if (tobj && mdd_object_exists(mdd_tobj)) {
1583 mdd_write_lock(env, mdd_tobj);
1584 mdo_ref_del(env, mdd_tobj, handle);
1586 /* Remove dot reference. */
1588 mdo_ref_del(env, mdd_tobj, handle);
1590 la->la_valid = LA_CTIME;
1591 rc = mdd_attr_set_internal(env, mdd_tobj, la, handle, 0);
1595 rc = mdd_finish_unlink(env, mdd_tobj, ma, handle);
1596 mdd_write_unlock(env, mdd_tobj);
1601 la->la_valid = LA_CTIME | LA_MTIME;
1602 rc = mdd_attr_set_internal_locked(env, mdd_spobj, la, handle, 0);
1606 if (mdd_spobj != mdd_tpobj) {
1607 la->la_valid = LA_CTIME | LA_MTIME;
1608 rc = mdd_attr_set_internal_locked(env, mdd_tpobj, la,
1614 if (likely(tdlh) && sdlh != tdlh)
1615 mdd_pdo_write_unlock(env, mdd_tpobj, tdlh);
1617 mdd_pdo_write_unlock(env, mdd_spobj, sdlh);
1619 mdd_trans_stop(env, mdd, rc, handle);
1621 mdd_object_put(env, mdd_sobj);
1625 struct md_dir_operations mdd_dir_ops = {
1626 .mdo_is_subdir = mdd_is_subdir,
1627 .mdo_lookup = mdd_lookup,
1628 .mdo_create = mdd_create,
1629 .mdo_rename = mdd_rename,
1630 .mdo_link = mdd_link,
1631 .mdo_unlink = mdd_unlink,
1632 .mdo_name_insert = mdd_name_insert,
1633 .mdo_name_remove = mdd_name_remove,
1634 .mdo_rename_tgt = mdd_rename_tgt,
1635 .mdo_create_data = mdd_create_data