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);
256 * Subdir count limitation can be broken through.
258 if (la->la_nlink >= m->mdd_dt_conf.ddp_max_nlink &&
259 !S_ISDIR(la->la_mode))
266 * Check whether it may create the cobj under the pobj.
269 int mdd_may_create(const struct lu_env *env, struct mdd_object *pobj,
270 struct mdd_object *cobj, int check_perm, int check_nlink)
275 if (cobj && mdd_object_exists(cobj))
278 if (mdd_is_dead_obj(pobj))
282 rc = mdd_permission_internal_locked(env, pobj, NULL,
283 MAY_WRITE | MAY_EXEC);
285 if (!rc && check_nlink)
286 rc = __mdd_may_link(env, pobj);
292 * Check whether can unlink from the pobj in the case of "cobj == NULL".
294 int mdd_may_unlink(const struct lu_env *env, struct mdd_object *pobj,
295 const struct md_attr *ma)
300 if (mdd_is_dead_obj(pobj))
303 if ((ma->ma_attr.la_valid & LA_FLAGS) &&
304 (ma->ma_attr.la_flags & (LUSTRE_APPEND_FL | LUSTRE_IMMUTABLE_FL)))
307 rc = mdd_permission_internal_locked(env, pobj, NULL,
308 MAY_WRITE | MAY_EXEC);
312 if (mdd_is_append(pobj))
319 * pobj == NULL is remote ops case, under such case, pobj's
320 * VTX feature has been checked already, no need check again.
322 static inline int mdd_is_sticky(const struct lu_env *env,
323 struct mdd_object *pobj,
324 struct mdd_object *cobj)
326 struct lu_attr *tmp_la = &mdd_env_info(env)->mti_la;
327 struct md_ucred *uc = md_ucred(env);
331 rc = mdd_la_get(env, pobj, tmp_la, BYPASS_CAPA);
335 if (!(tmp_la->la_mode & S_ISVTX) ||
336 (tmp_la->la_uid == uc->mu_fsuid))
340 rc = mdd_la_get(env, cobj, tmp_la, BYPASS_CAPA);
344 if (tmp_la->la_uid == uc->mu_fsuid)
347 return !mdd_capable(uc, CAP_FOWNER);
351 * Check whether it may delete the cobj from the pobj.
354 int mdd_may_delete(const struct lu_env *env, struct mdd_object *pobj,
355 struct mdd_object *cobj, struct md_attr *ma,
356 int check_perm, int check_empty)
362 if (!mdd_object_exists(cobj))
366 if (mdd_is_dead_obj(pobj))
370 rc = mdd_permission_internal_locked(env, pobj, NULL,
371 MAY_WRITE | MAY_EXEC);
376 if (mdd_is_append(pobj))
380 if (!(ma->ma_attr_flags & MDS_VTX_BYPASS) &&
381 mdd_is_sticky(env, pobj, cobj))
384 if (mdd_is_immutable(cobj) || mdd_is_append(cobj))
387 if ((ma->ma_attr.la_valid & LA_FLAGS) &&
388 (ma->ma_attr.la_flags & (LUSTRE_APPEND_FL | LUSTRE_IMMUTABLE_FL)))
391 if (S_ISDIR(ma->ma_attr.la_mode)) {
392 struct mdd_device *mdd = mdo2mdd(&cobj->mod_obj);
394 if (!S_ISDIR(mdd_object_type(cobj)))
397 if (lu_fid_eq(mdo2fid(cobj), &mdd->mdd_root_fid))
399 } else if (S_ISDIR(mdd_object_type(cobj)))
402 if (S_ISDIR(ma->ma_attr.la_mode) && check_empty)
403 rc = mdd_dir_is_empty(env, cobj);
410 * has mdd_write_lock on src already, but not on tgt yet
412 int mdd_link_sanity_check(const struct lu_env *env,
413 struct mdd_object *tgt_obj,
414 const struct lu_name *lname,
415 struct mdd_object *src_obj)
417 struct mdd_device *m = mdd_obj2mdd_dev(src_obj);
421 /* Local ops, no lookup before link, check filename length here. */
422 if (lname && (lname->ln_namelen > m->mdd_dt_conf.ddp_max_name_len))
423 RETURN(-ENAMETOOLONG);
425 if (mdd_is_immutable(src_obj) || mdd_is_append(src_obj))
428 if (S_ISDIR(mdd_object_type(src_obj)))
431 LASSERT(src_obj != tgt_obj);
433 rc = mdd_may_create(env, tgt_obj, NULL, 1, 0);
438 rc = __mdd_may_link(env, src_obj);
443 const struct dt_rec *__mdd_fid_rec(const struct lu_env *env,
444 const struct lu_fid *fid)
446 struct lu_fid_pack *pack = &mdd_env_info(env)->mti_pack;
448 fid_pack(pack, fid, &mdd_env_info(env)->mti_fid2);
449 return (const struct dt_rec *)pack;
453 * If subdir count is up to ddp_max_nlink, then enable MNLINK_OBJ flag and
454 * assign i_nlink to 1 which means the i_nlink for subdir count is incredible
455 * (maybe too large to be represented). It is a trick to break through the
456 * "i_nlink" limitation for subdir count.
458 void __mdd_ref_add(const struct lu_env *env, struct mdd_object *obj,
459 struct thandle *handle)
461 struct lu_attr *tmp_la = &mdd_env_info(env)->mti_la;
462 struct mdd_device *m = mdd_obj2mdd_dev(obj);
464 if (!mdd_is_mnlink(obj)) {
465 if (S_ISDIR(mdd_object_type(obj))) {
466 if (mdd_la_get(env, obj, tmp_la, BYPASS_CAPA))
469 if (tmp_la->la_nlink >= m->mdd_dt_conf.ddp_max_nlink) {
470 obj->mod_flags |= MNLINK_OBJ;
471 tmp_la->la_nlink = 1;
472 tmp_la->la_valid = LA_NLINK;
473 mdd_attr_set_internal(env, obj, tmp_la, handle,
478 mdo_ref_add(env, obj, handle);
482 void __mdd_ref_del(const struct lu_env *env, struct mdd_object *obj,
483 struct thandle *handle, int is_dot)
485 if (!mdd_is_mnlink(obj) || is_dot)
486 mdo_ref_del(env, obj, handle);
489 /* insert named index, add reference if isdir */
490 static int __mdd_index_insert(const struct lu_env *env, struct mdd_object *pobj,
491 const struct lu_fid *lf, const char *name, int is_dir,
492 struct thandle *handle, struct lustre_capa *capa)
494 struct dt_object *next = mdd_object_child(pobj);
498 if (dt_try_as_dir(env, next)) {
499 rc = next->do_index_ops->dio_insert(env, next,
500 __mdd_fid_rec(env, lf),
501 (const struct dt_key *)name,
509 mdd_write_lock(env, pobj);
510 __mdd_ref_add(env, pobj, handle);
511 mdd_write_unlock(env, pobj);
517 /* delete named index, drop reference if isdir */
518 static int __mdd_index_delete(const struct lu_env *env, struct mdd_object *pobj,
519 const char *name, int is_dir, struct thandle *handle,
520 struct lustre_capa *capa)
522 struct dt_object *next = mdd_object_child(pobj);
526 if (dt_try_as_dir(env, next)) {
527 rc = next->do_index_ops->dio_delete(env, next,
528 (struct dt_key *)name,
530 if (rc == 0 && is_dir) {
533 if (name != NULL && name[0] == '.' && name[1] == 0)
535 mdd_write_lock(env, pobj);
536 __mdd_ref_del(env, pobj, handle, is_dot);
537 mdd_write_unlock(env, pobj);
546 __mdd_index_insert_only(const struct lu_env *env, struct mdd_object *pobj,
547 const struct lu_fid *lf, const char *name,
548 struct thandle *handle, struct lustre_capa *capa)
550 struct dt_object *next = mdd_object_child(pobj);
554 if (dt_try_as_dir(env, next)) {
555 rc = next->do_index_ops->dio_insert(env, next,
556 __mdd_fid_rec(env, lf),
557 (const struct dt_key *)name,
565 static int mdd_link(const struct lu_env *env, struct md_object *tgt_obj,
566 struct md_object *src_obj, const struct lu_name *lname,
569 char *name = lname->ln_name;
570 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
571 struct mdd_object *mdd_tobj = md2mdd_obj(tgt_obj);
572 struct mdd_object *mdd_sobj = md2mdd_obj(src_obj);
573 struct mdd_device *mdd = mdo2mdd(src_obj);
574 struct dynlock_handle *dlh;
575 struct thandle *handle;
579 mdd_txn_param_build(env, mdd, MDD_TXN_LINK_OP);
580 handle = mdd_trans_start(env, mdd);
582 RETURN(PTR_ERR(handle));
584 dlh = mdd_pdo_write_lock(env, mdd_tobj, name);
586 GOTO(out_trans, rc = -ENOMEM);
587 mdd_write_lock(env, mdd_sobj);
589 rc = mdd_link_sanity_check(env, mdd_tobj, lname, mdd_sobj);
591 GOTO(out_unlock, rc);
593 rc = __mdd_index_insert_only(env, mdd_tobj, mdo2fid(mdd_sobj),
595 mdd_object_capa(env, mdd_tobj));
597 GOTO(out_unlock, rc);
599 __mdd_ref_add(env, mdd_sobj, handle);
601 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
602 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
604 la->la_valid = LA_CTIME | LA_MTIME;
605 rc = mdd_attr_check_set_internal_locked(env, mdd_tobj, la, handle, 0);
607 GOTO(out_unlock, rc);
609 la->la_valid = LA_CTIME;
610 rc = mdd_attr_check_set_internal(env, mdd_sobj, la, handle, 0);
613 mdd_write_unlock(env, mdd_sobj);
614 mdd_pdo_write_unlock(env, mdd_tobj, dlh);
616 mdd_trans_stop(env, mdd, rc, handle);
620 /* caller should take a lock before calling */
621 int mdd_finish_unlink(const struct lu_env *env,
622 struct mdd_object *obj, struct md_attr *ma,
628 rc = mdd_iattr_get(env, obj, ma);
629 if (rc == 0 && ma->ma_attr.la_nlink == 0) {
630 /* add new orphan and the object
631 * will be deleted during the object_put() */
632 if (__mdd_orphan_add(env, obj, th) == 0)
633 obj->mod_flags |= ORPHAN_OBJ;
635 obj->mod_flags |= DEAD_OBJ;
636 if (obj->mod_count == 0)
637 rc = mdd_object_kill(env, obj, ma);
639 /* clear MA_LOV | MA_COOKIE, if we do not
640 * unlink it in case we get it somewhere */
641 ma->ma_valid &= ~(MA_LOV | MA_COOKIE);
643 ma->ma_valid &= ~(MA_LOV | MA_COOKIE);
650 * has mdd_write_lock on cobj already, but not on pobj yet
652 int mdd_unlink_sanity_check(const struct lu_env *env, struct mdd_object *pobj,
653 struct mdd_object *cobj, struct md_attr *ma)
658 rc = mdd_may_delete(env, pobj, cobj, ma, 1, 1);
663 static int mdd_unlink(const struct lu_env *env, struct md_object *pobj,
664 struct md_object *cobj, const struct lu_name *lname,
667 char *name = lname->ln_name;
668 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
669 struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
670 struct mdd_object *mdd_cobj = md2mdd_obj(cobj);
671 struct mdd_device *mdd = mdo2mdd(pobj);
672 struct dynlock_handle *dlh;
673 struct thandle *handle;
677 LASSERTF(mdd_object_exists(mdd_cobj) > 0, "FID is "DFID"\n",
678 PFID(mdd_object_fid(mdd_cobj)));
680 rc = mdd_log_txn_param_build(env, cobj, ma, MDD_TXN_UNLINK_OP);
684 handle = mdd_trans_start(env, mdd);
686 RETURN(PTR_ERR(handle));
689 dlh = mdd_pdo_write_lock(env, mdd_pobj, name);
691 GOTO(out_trans, rc = -ENOMEM);
692 mdd_write_lock(env, mdd_cobj);
694 is_dir = S_ISDIR(ma->ma_attr.la_mode);
695 rc = mdd_unlink_sanity_check(env, mdd_pobj, mdd_cobj, ma);
699 rc = __mdd_index_delete(env, mdd_pobj, name, is_dir, handle,
700 mdd_object_capa(env, mdd_pobj));
704 __mdd_ref_del(env, mdd_cobj, handle, 0);
707 __mdd_ref_del(env, mdd_cobj, handle, 1);
709 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
710 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
712 la->la_valid = LA_CTIME | LA_MTIME;
713 rc = mdd_attr_check_set_internal_locked(env, mdd_pobj, la, handle, 0);
717 la->la_valid = LA_CTIME;
718 rc = mdd_attr_check_set_internal(env, mdd_cobj, la, handle, 0);
722 rc = mdd_finish_unlink(env, mdd_cobj, ma, handle);
725 obd_set_info_async(mdd2obd_dev(mdd)->u.mds.mds_osc_exp,
726 sizeof(KEY_UNLINKED), KEY_UNLINKED, 0,
730 mdd_write_unlock(env, mdd_cobj);
731 mdd_pdo_write_unlock(env, mdd_pobj, dlh);
733 mdd_trans_stop(env, mdd, rc, handle);
737 /* has not lock on pobj yet */
738 static int mdd_ni_sanity_check(const struct lu_env *env,
739 struct md_object *pobj,
740 const struct md_attr *ma)
742 struct mdd_object *obj = md2mdd_obj(pobj);
746 if (ma->ma_attr_flags & MDS_PERM_BYPASS)
749 rc = mdd_may_create(env, obj, NULL, 1, S_ISDIR(ma->ma_attr.la_mode));
757 static int mdd_name_insert(const struct lu_env *env,
758 struct md_object *pobj,
759 const struct lu_name *lname,
760 const struct lu_fid *fid,
761 const struct md_attr *ma)
763 char *name = lname->ln_name;
764 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
765 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
766 struct mdd_device *mdd = mdo2mdd(pobj);
767 struct dynlock_handle *dlh;
768 struct thandle *handle;
769 int is_dir = S_ISDIR(ma->ma_attr.la_mode);
773 mdd_txn_param_build(env, mdd, MDD_TXN_INDEX_INSERT_OP);
774 handle = mdd_trans_start(env, mdo2mdd(pobj));
776 RETURN(PTR_ERR(handle));
778 dlh = mdd_pdo_write_lock(env, mdd_obj, name);
780 GOTO(out_trans, rc = -ENOMEM);
782 rc = mdd_ni_sanity_check(env, pobj, ma);
784 GOTO(out_unlock, rc);
786 rc = __mdd_index_insert(env, mdd_obj, fid, name, is_dir,
787 handle, BYPASS_CAPA);
789 GOTO(out_unlock, rc);
792 * For some case, no need update obj's ctime (LA_CTIME is not set),
794 * For other cases, update obj's ctime (LA_CTIME is set),
797 if (ma->ma_attr.la_valid & LA_CTIME) {
798 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
799 la->la_valid = LA_CTIME | LA_MTIME;
800 rc = mdd_attr_check_set_internal_locked(env, mdd_obj, la,
805 mdd_pdo_write_unlock(env, mdd_obj, dlh);
807 mdd_trans_stop(env, mdo2mdd(pobj), rc, handle);
811 /* has not lock on pobj yet */
812 static int mdd_nr_sanity_check(const struct lu_env *env,
813 struct md_object *pobj,
814 const struct md_attr *ma)
816 struct mdd_object *obj = md2mdd_obj(pobj);
820 if (ma->ma_attr_flags & MDS_PERM_BYPASS)
823 rc = mdd_may_unlink(env, obj, ma);
831 static int mdd_name_remove(const struct lu_env *env,
832 struct md_object *pobj,
833 const struct lu_name *lname,
834 const struct md_attr *ma)
836 char *name = lname->ln_name;
837 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
838 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
839 struct mdd_device *mdd = mdo2mdd(pobj);
840 struct dynlock_handle *dlh;
841 struct thandle *handle;
842 int is_dir = S_ISDIR(ma->ma_attr.la_mode);
846 mdd_txn_param_build(env, mdd, MDD_TXN_INDEX_DELETE_OP);
847 handle = mdd_trans_start(env, mdd);
849 RETURN(PTR_ERR(handle));
851 dlh = mdd_pdo_write_lock(env, mdd_obj, name);
853 GOTO(out_trans, rc = -ENOMEM);
855 rc = mdd_nr_sanity_check(env, pobj, ma);
857 GOTO(out_unlock, rc);
859 rc = __mdd_index_delete(env, mdd_obj, name, is_dir,
860 handle, BYPASS_CAPA);
862 GOTO(out_unlock, rc);
865 * For some case, no need update obj's ctime (LA_CTIME is not set),
867 * For other cases, update obj's ctime (LA_CTIME is set),
870 if (ma->ma_attr.la_valid & LA_CTIME) {
871 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
872 la->la_valid = LA_CTIME | LA_MTIME;
873 rc = mdd_attr_check_set_internal_locked(env, mdd_obj, la,
878 mdd_pdo_write_unlock(env, mdd_obj, dlh);
880 mdd_trans_stop(env, mdd, rc, handle);
886 * has mdd_write_lock on tobj alreay, but not on tgt_pobj yet
888 static int mdd_rt_sanity_check(const struct lu_env *env,
889 struct mdd_object *tgt_pobj,
890 struct mdd_object *tobj,
896 if (unlikely(ma->ma_attr_flags & MDS_PERM_BYPASS))
899 /* XXX: for mdd_rename_tgt, "tobj == NULL" does not mean tobj not
900 * exist. In fact, tobj must exist, otherwise the call trace will be:
901 * mdt_reint_rename_tgt -> mdo_name_insert -> ... -> mdd_name_insert.
902 * When get here, tobj must be NOT NULL, the other case has been
903 * processed in cmr_rename_tgt before mdd_rename_tgt and enable
905 * So check may_delete, but not check nlink of tgt_pobj. */
907 rc = mdd_may_delete(env, tgt_pobj, tobj, ma, 1, 1);
912 static int mdd_rename_tgt(const struct lu_env *env,
913 struct md_object *pobj, struct md_object *tobj,
914 const struct lu_fid *lf, const struct lu_name *lname,
917 char *name = lname->ln_name;
918 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
919 struct mdd_object *mdd_tpobj = md2mdd_obj(pobj);
920 struct mdd_object *mdd_tobj = md2mdd_obj(tobj);
921 struct mdd_device *mdd = mdo2mdd(pobj);
922 struct dynlock_handle *dlh;
923 struct thandle *handle;
927 mdd_txn_param_build(env, mdd, MDD_TXN_RENAME_TGT_OP);
928 handle = mdd_trans_start(env, mdd);
930 RETURN(PTR_ERR(handle));
932 dlh = mdd_pdo_write_lock(env, mdd_tpobj, name);
934 GOTO(out_trans, rc = -ENOMEM);
936 mdd_write_lock(env, mdd_tobj);
938 rc = mdd_rt_sanity_check(env, mdd_tpobj, mdd_tobj, ma);
943 * If rename_tgt is called then we should just re-insert name with
944 * correct fid, no need to dec/inc parent nlink if obj is dir.
946 rc = __mdd_index_delete(env, mdd_tpobj, name, 0, handle, BYPASS_CAPA);
950 rc = __mdd_index_insert_only(env, mdd_tpobj, lf, name, handle,
955 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
956 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
958 la->la_valid = LA_CTIME | LA_MTIME;
959 rc = mdd_attr_check_set_internal_locked(env, mdd_tpobj, la, handle, 0);
964 * For tobj is remote case cmm layer has processed
965 * and pass NULL tobj to here. So when tobj is NOT NULL,
966 * it must be local one.
968 if (tobj && mdd_object_exists(mdd_tobj)) {
969 __mdd_ref_del(env, mdd_tobj, handle, 0);
971 /* Remove dot reference. */
972 if (S_ISDIR(ma->ma_attr.la_mode))
973 __mdd_ref_del(env, mdd_tobj, handle, 1);
975 la->la_valid = LA_CTIME;
976 rc = mdd_attr_check_set_internal(env, mdd_tobj, la, handle, 0);
980 rc = mdd_finish_unlink(env, mdd_tobj, ma, handle);
987 mdd_write_unlock(env, mdd_tobj);
988 mdd_pdo_write_unlock(env, mdd_tpobj, dlh);
990 mdd_trans_stop(env, mdd, rc, handle);
995 * The permission has been checked when obj created, no need check again.
997 static int mdd_cd_sanity_check(const struct lu_env *env,
998 struct mdd_object *obj)
1003 if (!obj || mdd_is_dead_obj(obj))
1010 static int mdd_create_data(const struct lu_env *env, struct md_object *pobj,
1011 struct md_object *cobj, const struct md_op_spec *spec,
1014 struct mdd_device *mdd = mdo2mdd(cobj);
1015 struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
1016 struct mdd_object *son = md2mdd_obj(cobj);
1017 struct lu_attr *attr = &ma->ma_attr;
1018 struct lov_mds_md *lmm = NULL;
1020 struct thandle *handle;
1024 rc = mdd_cd_sanity_check(env, son);
1028 if (!md_should_create(spec->sp_cr_flags))
1031 rc = mdd_lov_create(env, mdd, mdd_pobj, son, &lmm, &lmm_size,
1036 mdd_txn_param_build(env, mdd, MDD_TXN_CREATE_DATA_OP);
1037 handle = mdd_trans_start(env, mdd);
1039 GOTO(out_free, rc = PTR_ERR(handle));
1042 * XXX: Setting the lov ea is not locked but setting the attr is locked?
1043 * Should this be fixed?
1046 /* Replay creates has objects already */
1048 if (spec->u.sp_ea.no_lov_create) {
1049 CDEBUG(D_INFO, "we already have lov ea\n");
1050 rc = mdd_lov_set_md(env, mdd_pobj, son,
1051 (struct lov_mds_md *)spec->u.sp_ea.eadata,
1052 spec->u.sp_ea.eadatalen, handle, 0);
1055 /* No need mdd_lsm_sanity_check here */
1056 rc = mdd_lov_set_md(env, mdd_pobj, son, lmm,
1057 lmm_size, handle, 0);
1060 rc = mdd_attr_get_internal_locked(env, son, ma);
1062 /* update lov_objid data, must be before transaction stop! */
1064 mdd_lov_objid_update(mdd, lmm);
1066 mdd_trans_stop(env, mdd, rc, handle);
1068 /* Finish mdd_lov_create() stuff. */
1069 mdd_lov_create_finish(env, mdd, lmm, lmm_size, spec);
1074 __mdd_lookup(const struct lu_env *env, struct md_object *pobj,
1075 const struct lu_name *lname, struct lu_fid* fid, int mask)
1077 char *name = lname->ln_name;
1078 const struct dt_key *key = (const struct dt_key *)name;
1079 struct mdd_object *mdd_obj = md2mdd_obj(pobj);
1080 struct mdd_device *m = mdo2mdd(pobj);
1081 struct dt_object *dir = mdd_object_child(mdd_obj);
1082 struct lu_fid_pack *pack = &mdd_env_info(env)->mti_pack;
1086 if (unlikely(mdd_is_dead_obj(mdd_obj)))
1089 rc = mdd_object_exists(mdd_obj);
1090 if (unlikely(rc == 0))
1092 else if (unlikely(rc < 0)) {
1093 CERROR("Object "DFID" locates on remote server\n",
1094 PFID(mdo2fid(mdd_obj)));
1098 /* The common filename length check. */
1099 if (unlikely(lname->ln_namelen > m->mdd_dt_conf.ddp_max_name_len))
1100 RETURN(-ENAMETOOLONG);
1102 rc = mdd_permission_internal_locked(env, mdd_obj, NULL, mask);
1106 if (likely(S_ISDIR(mdd_object_type(mdd_obj)) &&
1107 dt_try_as_dir(env, dir))) {
1108 rc = dir->do_index_ops->dio_lookup(env, dir,
1109 (struct dt_rec *)pack, key,
1110 mdd_object_capa(env, mdd_obj));
1112 rc = fid_unpack(pack, fid);
1119 int mdd_object_initialize(const struct lu_env *env, const struct lu_fid *pfid,
1120 struct mdd_object *child, struct md_attr *ma,
1121 struct thandle *handle)
1127 * Update attributes for child.
1130 * (1) the valid bits should be converted between Lustre and Linux;
1131 * (2) maybe, the child attributes should be set in OSD when creation.
1134 rc = mdd_attr_set_internal(env, child, &ma->ma_attr, handle, 0);
1138 if (S_ISDIR(ma->ma_attr.la_mode)) {
1139 /* Add "." and ".." for newly created dir */
1140 __mdd_ref_add(env, child, handle);
1141 rc = __mdd_index_insert_only(env, child, mdo2fid(child),
1142 dot, handle, BYPASS_CAPA);
1144 rc = __mdd_index_insert_only(env, child, pfid,
1150 rc2 = __mdd_index_delete(env, child, dot, 1,
1151 handle, BYPASS_CAPA);
1153 CERROR("Failure to cleanup after dotdot"
1154 " creation: %d (%d)\n", rc2, rc);
1161 /* has not lock on pobj yet */
1162 static int mdd_create_sanity_check(const struct lu_env *env,
1163 struct md_object *pobj,
1164 const struct lu_name *lname,
1166 struct md_op_spec *spec)
1168 struct mdd_thread_info *info = mdd_env_info(env);
1169 struct lu_attr *la = &info->mti_la;
1170 struct lu_fid *fid = &info->mti_fid;
1171 struct mdd_object *obj = md2mdd_obj(pobj);
1172 struct mdd_device *m = mdo2mdd(pobj);
1173 int lookup = spec->sp_cr_lookup;
1178 if (mdd_is_dead_obj(obj))
1182 * In some cases this lookup is not needed - we know before if name
1183 * exists or not because MDT performs lookup for it.
1184 * name length check is done in lookup.
1188 * Check if the name already exist, though it will be checked in
1189 * _index_insert also, for avoiding rolling back if exists
1192 rc = __mdd_lookup_locked(env, pobj, lname, fid,
1193 MAY_WRITE | MAY_EXEC);
1195 RETURN(rc ? : -EEXIST);
1198 * Check WRITE permission for the parent.
1199 * EXEC permission have been checked
1200 * when lookup before create already.
1202 rc = mdd_permission_internal_locked(env, obj, NULL, MAY_WRITE);
1208 rc = mdd_la_get(env, obj, la, BYPASS_CAPA);
1212 if (la->la_mode & S_ISGID) {
1213 ma->ma_attr.la_gid = la->la_gid;
1214 if (S_ISDIR(ma->ma_attr.la_mode)) {
1215 ma->ma_attr.la_mode |= S_ISGID;
1216 ma->ma_attr.la_valid |= LA_MODE;
1220 switch (ma->ma_attr.la_mode & S_IFMT) {
1222 unsigned int symlen = strlen(spec->u.sp_symname) + 1;
1224 if (symlen > (1 << m->mdd_dt_conf.ddp_block_shift))
1225 RETURN(-ENAMETOOLONG);
1245 * Create object and insert it into namespace.
1247 static int mdd_create(const struct lu_env *env,
1248 struct md_object *pobj,
1249 const struct lu_name *lname,
1250 struct md_object *child,
1251 struct md_op_spec *spec,
1254 char *name = lname->ln_name;
1255 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
1256 struct mdd_object *mdd_pobj = md2mdd_obj(pobj);
1257 struct mdd_object *son = md2mdd_obj(child);
1258 struct mdd_device *mdd = mdo2mdd(pobj);
1259 struct lu_attr *attr = &ma->ma_attr;
1260 struct lov_mds_md *lmm = NULL;
1261 struct thandle *handle;
1262 int rc, created = 0, initialized = 0, inserted = 0, lmm_size = 0;
1263 struct dynlock_handle *dlh;
1267 * Two operations have to be performed:
1269 * - allocation of new object (->do_create()), and
1271 * - insertion into parent index (->dio_insert()).
1273 * Due to locking, operation order is not important, when both are
1274 * successful, *but* error handling cases are quite different:
1276 * - if insertion is done first, and following object creation fails,
1277 * insertion has to be rolled back, but this operation might fail
1278 * also leaving us with dangling index entry.
1280 * - if creation is done first, is has to be undone if insertion
1281 * fails, leaving us with leaked space, which is neither good, nor
1284 * It seems that creation-first is simplest solution, but it is
1285 * sub-optimal in the frequent
1290 * case, because second mkdir is bound to create object, only to
1291 * destroy it immediately.
1293 * To avoid this follow local file systems that do double lookup:
1295 * 0. lookup -> -EEXIST (mdd_create_sanity_check())
1297 * 1. create (mdd_object_create_internal())
1299 * 2. insert (__mdd_index_insert(), lookup again)
1302 /* Sanity checks before big job. */
1303 rc = mdd_create_sanity_check(env, pobj, lname, ma, spec);
1308 * No RPC inside the transaction, so OST objects should be created at
1311 if (S_ISREG(attr->la_mode)) {
1312 rc = mdd_lov_create(env, mdd, mdd_pobj, son, &lmm, &lmm_size,
1318 mdd_txn_param_build(env, mdd, MDD_TXN_MKDIR_OP);
1319 handle = mdd_trans_start(env, mdd);
1321 GOTO(out_free, rc = PTR_ERR(handle));
1323 dlh = mdd_pdo_write_lock(env, mdd_pobj, name);
1325 GOTO(out_trans, rc = -ENOMEM);
1328 * XXX: Check that link can be added to the parent in mkdir case.
1331 mdd_write_lock(env, son);
1332 rc = mdd_object_create_internal(env, mdd_pobj, son, ma, handle);
1334 mdd_write_unlock(env, son);
1340 #ifdef CONFIG_FS_POSIX_ACL
1341 mdd_read_lock(env, mdd_pobj);
1342 rc = mdd_acl_init(env, mdd_pobj, son, &ma->ma_attr.la_mode, handle);
1343 mdd_read_unlock(env, mdd_pobj);
1345 mdd_write_unlock(env, son);
1348 ma->ma_attr.la_valid |= LA_MODE;
1352 rc = mdd_object_initialize(env, mdo2fid(mdd_pobj),
1354 mdd_write_unlock(env, son);
1357 * Object has no links, so it will be destroyed when last
1358 * reference is released. (XXX not now.)
1364 rc = __mdd_index_insert(env, mdd_pobj, mdo2fid(son),
1365 name, S_ISDIR(attr->la_mode), handle,
1366 mdd_object_capa(env, mdd_pobj));
1373 /* No need mdd_lsm_sanity_check here */
1374 rc = mdd_lov_set_md(env, mdd_pobj, son, lmm, lmm_size, handle, 0);
1376 CERROR("error on stripe info copy %d \n", rc);
1379 if (lmm && lmm_size > 0) {
1380 /* Set Lov here, do not get lmm again later */
1381 memcpy(ma->ma_lmm, lmm, lmm_size);
1382 ma->ma_lmm_size = lmm_size;
1383 ma->ma_valid |= MA_LOV;
1386 if (S_ISLNK(attr->la_mode)) {
1387 struct dt_object *dt = mdd_object_child(son);
1388 const char *target_name = spec->u.sp_symname;
1389 int sym_len = strlen(target_name);
1390 const struct lu_buf *buf;
1393 buf = mdd_buf_get_const(env, target_name, sym_len);
1394 rc = dt->do_body_ops->dbo_write(env, dt, buf, &pos, handle,
1395 mdd_object_capa(env, son));
1400 GOTO(cleanup, rc = -EFAULT);
1404 la->la_valid = LA_CTIME | LA_MTIME;
1405 rc = mdd_attr_check_set_internal_locked(env, mdd_pobj, la, handle, 0);
1409 /* Return attr back. */
1410 rc = mdd_attr_get_internal_locked(env, son, ma);
1413 if (rc && created) {
1417 rc2 = __mdd_index_delete(env, mdd_pobj, name,
1418 S_ISDIR(attr->la_mode),
1419 handle, BYPASS_CAPA);
1421 CERROR("error can not cleanup destroy %d\n",
1426 mdd_write_lock(env, son);
1427 __mdd_ref_del(env, son, handle, 0);
1428 if (initialized && S_ISDIR(attr->la_mode))
1429 __mdd_ref_del(env, son, handle, 1);
1430 mdd_write_unlock(env, son);
1434 /* update lov_objid data, must be before transaction stop! */
1436 mdd_lov_objid_update(mdd, lmm);
1438 mdd_pdo_write_unlock(env, mdd_pobj, dlh);
1440 mdd_trans_stop(env, mdd, rc, handle);
1442 /* finis lov_create stuff, free all temporary data */
1443 mdd_lov_create_finish(env, mdd, lmm, lmm_size, spec);
1448 * Get locks on parents in proper order
1449 * RETURN: < 0 - error, rename_order if successful
1457 static int mdd_rename_order(const struct lu_env *env,
1458 struct mdd_device *mdd,
1459 struct mdd_object *src_pobj,
1460 struct mdd_object *tgt_pobj)
1462 /* order of locking, 1 - tgt-src, 0 - src-tgt*/
1466 if (src_pobj == tgt_pobj)
1467 RETURN(MDD_RN_SAME);
1469 /* compared the parent child relationship of src_p&tgt_p */
1470 if (lu_fid_eq(&mdd->mdd_root_fid, mdo2fid(src_pobj))){
1472 } else if (lu_fid_eq(&mdd->mdd_root_fid, mdo2fid(tgt_pobj))) {
1475 rc = mdd_is_parent(env, mdd, src_pobj, mdo2fid(tgt_pobj), NULL);
1488 /* has not mdd_write{read}_lock on any obj yet. */
1489 static int mdd_rename_sanity_check(const struct lu_env *env,
1490 struct mdd_object *src_pobj,
1491 struct mdd_object *tgt_pobj,
1492 struct mdd_object *sobj,
1493 struct mdd_object *tobj,
1499 if (unlikely(ma->ma_attr_flags & MDS_PERM_BYPASS))
1502 /* XXX: when get here, sobj must NOT be NULL,
1503 * the other case has been processed in cml_rename
1504 * before mdd_rename and enable MDS_PERM_BYPASS. */
1506 rc = mdd_may_delete(env, src_pobj, sobj, ma, 1, 0);
1510 /* XXX: when get here, "tobj == NULL" means tobj must
1511 * NOT exist (neither on remote MDS, such case has been
1512 * processed in cml_rename before mdd_rename and enable
1514 * So check may_create, but not check may_unlink. */
1516 rc = mdd_may_create(env, tgt_pobj, NULL,
1517 (src_pobj != tgt_pobj), 0);
1519 rc = mdd_may_delete(env, tgt_pobj, tobj, ma,
1520 (src_pobj != tgt_pobj), 1);
1522 if (!rc && !tobj && (src_pobj != tgt_pobj) &&
1523 S_ISDIR(ma->ma_attr.la_mode))
1524 rc = __mdd_may_link(env, tgt_pobj);
1529 /* src object can be remote that is why we use only fid and type of object */
1530 static int mdd_rename(const struct lu_env *env,
1531 struct md_object *src_pobj, struct md_object *tgt_pobj,
1532 const struct lu_fid *lf, const struct lu_name *lsname,
1533 struct md_object *tobj, const struct lu_name *ltname,
1536 char *sname = lsname->ln_name;
1537 char *tname = ltname->ln_name;
1538 struct lu_attr *la = &mdd_env_info(env)->mti_la_for_fix;
1539 struct mdd_object *mdd_spobj = md2mdd_obj(src_pobj);
1540 struct mdd_object *mdd_tpobj = md2mdd_obj(tgt_pobj);
1541 struct mdd_device *mdd = mdo2mdd(src_pobj);
1542 struct mdd_object *mdd_sobj = NULL;
1543 struct mdd_object *mdd_tobj = NULL;
1544 struct dynlock_handle *sdlh, *tdlh;
1545 struct thandle *handle;
1550 LASSERT(ma->ma_attr.la_mode & S_IFMT);
1551 is_dir = S_ISDIR(ma->ma_attr.la_mode);
1554 mdd_tobj = md2mdd_obj(tobj);
1556 mdd_txn_param_build(env, mdd, MDD_TXN_RENAME_OP);
1557 handle = mdd_trans_start(env, mdd);
1559 RETURN(PTR_ERR(handle));
1561 /* FIXME: Should consider tobj and sobj too in rename_lock. */
1562 rc = mdd_rename_order(env, mdd, mdd_spobj, mdd_tpobj);
1564 GOTO(cleanup_unlocked, rc);
1566 /* Get locks in determined order */
1567 if (rc == MDD_RN_SAME) {
1568 sdlh = mdd_pdo_write_lock(env, mdd_spobj, sname);
1569 /* check hashes to determine do we need one lock or two */
1570 if (mdd_name2hash(sname) != mdd_name2hash(tname))
1571 tdlh = mdd_pdo_write_lock(env, mdd_tpobj, tname);
1574 } else if (rc == MDD_RN_SRCTGT) {
1575 sdlh = mdd_pdo_write_lock(env, mdd_spobj, sname);
1576 tdlh = mdd_pdo_write_lock(env, mdd_tpobj, tname);
1578 tdlh = mdd_pdo_write_lock(env, mdd_tpobj, tname);
1579 sdlh = mdd_pdo_write_lock(env, mdd_spobj, sname);
1581 if (sdlh == NULL || tdlh == NULL)
1582 GOTO(cleanup, rc = -ENOMEM);
1584 mdd_sobj = mdd_object_find(env, mdd, lf);
1585 rc = mdd_rename_sanity_check(env, mdd_spobj, mdd_tpobj,
1586 mdd_sobj, mdd_tobj, ma);
1590 rc = __mdd_index_delete(env, mdd_spobj, sname, is_dir, handle,
1591 mdd_object_capa(env, mdd_spobj));
1596 * Here tobj can be remote one, so we do index_delete unconditionally
1597 * and -ENOENT is allowed.
1599 rc = __mdd_index_delete(env, mdd_tpobj, tname, is_dir, handle,
1600 mdd_object_capa(env, mdd_tpobj));
1601 if (rc != 0 && rc != -ENOENT)
1604 rc = __mdd_index_insert(env, mdd_tpobj, lf, tname, is_dir, handle,
1605 mdd_object_capa(env, mdd_tpobj));
1609 LASSERT(ma->ma_attr.la_valid & LA_CTIME);
1610 la->la_ctime = la->la_mtime = ma->ma_attr.la_ctime;
1612 /* XXX: mdd_sobj must be local one if it is NOT NULL. */
1614 la->la_valid = LA_CTIME;
1615 rc = mdd_attr_check_set_internal_locked(env, mdd_sobj, la,
1622 * For tobj is remote case cmm layer has processed
1623 * and set tobj to NULL then. So when tobj is NOT NULL,
1624 * it must be local one.
1626 if (tobj && mdd_object_exists(mdd_tobj)) {
1627 mdd_write_lock(env, mdd_tobj);
1628 __mdd_ref_del(env, mdd_tobj, handle, 0);
1630 /* Remove dot reference. */
1632 __mdd_ref_del(env, mdd_tobj, handle, 1);
1634 la->la_valid = LA_CTIME;
1635 rc = mdd_attr_check_set_internal(env, mdd_tobj, la, handle, 0);
1639 rc = mdd_finish_unlink(env, mdd_tobj, ma, handle);
1640 mdd_write_unlock(env, mdd_tobj);
1645 la->la_valid = LA_CTIME | LA_MTIME;
1646 rc = mdd_attr_check_set_internal_locked(env, mdd_spobj, la, handle, 0);
1650 if (mdd_spobj != mdd_tpobj) {
1651 la->la_valid = LA_CTIME | LA_MTIME;
1652 rc = mdd_attr_check_set_internal_locked(env, mdd_tpobj, la,
1658 if (likely(tdlh) && sdlh != tdlh)
1659 mdd_pdo_write_unlock(env, mdd_tpobj, tdlh);
1661 mdd_pdo_write_unlock(env, mdd_spobj, sdlh);
1663 mdd_trans_stop(env, mdd, rc, handle);
1665 mdd_object_put(env, mdd_sobj);
1669 struct md_dir_operations mdd_dir_ops = {
1670 .mdo_is_subdir = mdd_is_subdir,
1671 .mdo_lookup = mdd_lookup,
1672 .mdo_create = mdd_create,
1673 .mdo_rename = mdd_rename,
1674 .mdo_link = mdd_link,
1675 .mdo_unlink = mdd_unlink,
1676 .mdo_name_insert = mdd_name_insert,
1677 .mdo_name_remove = mdd_name_remove,
1678 .mdo_rename_tgt = mdd_rename_tgt,
1679 .mdo_create_data = mdd_create_data