4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <lustre_mds.h>
52 #include "mdd_internal.h"
54 static const struct lu_object_operations mdd_lu_obj_ops;
56 static int mdd_xattr_get(const struct lu_env *env,
57 struct md_object *obj, struct lu_buf *buf,
60 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
65 if (mdd_object_exists(obj) == 0)
68 rc = mdo_attr_get(env, obj, la);
69 if (unlikely(rc != 0)) {
71 obj->mod_flags |= DEAD_OBJ;
75 if (la->la_valid & LA_FLAGS &&
76 la->la_flags & LUSTRE_ORPHAN_FL)
77 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
82 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
84 struct mdd_thread_info *info;
86 lu_env_refill((struct lu_env *)env);
87 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
88 LASSERT(info != NULL);
92 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
96 buf = &mdd_env_info(env)->mti_buf[0];
102 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
103 const void *area, ssize_t len)
107 buf = &mdd_env_info(env)->mti_buf[0];
108 buf->lb_buf = (void *)area;
113 struct lu_object *mdd_object_alloc(const struct lu_env *env,
114 const struct lu_object_header *hdr,
117 struct mdd_object *mdd_obj;
119 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
120 if (mdd_obj != NULL) {
123 o = mdd2lu_obj(mdd_obj);
124 lu_object_init(o, NULL, d);
125 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
126 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
127 mdd_obj->mod_count = 0;
128 o->lo_ops = &mdd_lu_obj_ops;
135 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
136 const struct lu_object_conf *unused)
138 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
139 struct mdd_object *mdd_obj = lu2mdd_obj(o);
140 struct lu_object *below;
141 struct lu_device *under;
144 mdd_obj->mod_cltime = ktime_set(0, 0);
145 under = &d->mdd_child->dd_lu_dev;
146 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
148 RETURN(PTR_ERR(below));
150 lu_object_add(o, below);
155 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
159 if (lu_object_exists(o)) {
160 struct mdd_object *mdd_obj = lu2mdd_obj(o);
161 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
163 rc = mdd_la_get(env, mdd_obj, attr);
169 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
171 struct mdd_object *mdd = lu2mdd_obj(o);
174 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
177 static int mdd_object_print(const struct lu_env *env, void *cookie,
178 lu_printer_t p, const struct lu_object *o)
180 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
182 return (*p)(env, cookie,
183 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
184 mdd, mdd->mod_count, mdd->mod_valid,
185 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
188 static const struct lu_object_operations mdd_lu_obj_ops = {
189 .loo_object_init = mdd_object_init,
190 .loo_object_start = mdd_object_start,
191 .loo_object_free = mdd_object_free,
192 .loo_object_print = mdd_object_print,
195 struct mdd_object *mdd_object_find(const struct lu_env *env,
196 struct mdd_device *d,
197 const struct lu_fid *f)
199 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
203 * No permission check is needed.
205 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
208 struct mdd_object *mdd_obj = md2mdd_obj(obj);
213 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
214 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
215 ma->ma_attr.la_nlink = 0;
221 * No permission check is needed.
223 static int mdd_xattr_get(const struct lu_env *env,
224 struct md_object *obj, struct lu_buf *buf,
227 struct mdd_object *mdd_obj = md2mdd_obj(obj);
232 if (mdd_object_exists(mdd_obj) == 0) {
233 CERROR("%s: object "DFID" not found: rc = -2\n",
234 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
238 /* If the object has been destroyed, then do not get LMVEA, because
239 * it needs to load stripes from the iteration of the master object,
240 * and it will cause problem if master object has been destroyed, see
242 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
243 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
244 strcmp(name, XATTR_NAME_LMV) == 0))
247 /* If the object has been delete from the namespace, then
248 * get linkEA should return -ENOENT as well */
249 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
250 strcmp(name, XATTR_NAME_LINK) == 0))
253 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
254 rc = mdo_xattr_get(env, mdd_obj, buf, name);
255 mdd_read_unlock(env, mdd_obj);
261 * Permission check is done when open,
262 * no need check again.
264 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
267 struct mdd_object *mdd_obj = md2mdd_obj(obj);
268 struct dt_object *next;
273 if (mdd_object_exists(mdd_obj) == 0) {
274 CERROR("%s: object "DFID" not found: rc = -2\n",
275 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
279 next = mdd_object_child(mdd_obj);
280 LASSERT(next != NULL);
281 LASSERT(next->do_body_ops != NULL);
282 LASSERT(next->do_body_ops->dbo_read != NULL);
283 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
284 rc = dt_read(env, next, buf, &pos);
285 mdd_read_unlock(env, mdd_obj);
290 * No permission check is needed.
292 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
295 struct mdd_object *mdd_obj = md2mdd_obj(obj);
300 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
301 rc = mdo_xattr_list(env, mdd_obj, buf);
302 mdd_read_unlock(env, mdd_obj);
304 /* If the buffer is NULL then we are only here to get the
305 * length of the xattr name list. */
306 if (rc < 0 || buf->lb_buf == NULL)
310 * Filter out XATTR_NAME_LINK if this is an orphan object. See
313 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
314 char *end = (char *)buf->lb_buf + rc;
315 char *p = buf->lb_buf;
318 char *next = p + strlen(p) + 1;
320 if (strcmp(p, XATTR_NAME_LINK) == 0) {
322 memmove(p, next, end - next);
324 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
325 " of orphan "DFID"\n",
326 PFID(mdd_object_fid(mdd_obj)));
337 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
339 return mdo_invalidate(env, md2mdd_obj(obj));
342 int mdd_declare_create_object_internal(const struct lu_env *env,
343 struct mdd_object *p,
344 struct mdd_object *c,
345 struct lu_attr *attr,
346 struct thandle *handle,
347 const struct md_op_spec *spec,
348 struct dt_allocation_hint *hint)
350 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
351 const struct dt_index_features *feat = spec->sp_feat;
355 if (feat != &dt_directory_features && feat != NULL) {
356 dof->dof_type = DFT_INDEX;
357 dof->u.dof_idx.di_feat = feat;
359 dof->dof_type = dt_mode_to_dft(attr->la_mode);
360 if (dof->dof_type == DFT_REGULAR) {
361 dof->u.dof_reg.striped =
362 md_should_create(spec->sp_cr_flags);
363 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
364 dof->u.dof_reg.striped = 0;
365 /* is this replay? */
367 dof->u.dof_reg.striped = 0;
371 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
376 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
377 struct mdd_object *c, struct lu_attr *attr,
378 struct thandle *handle,
379 const struct md_op_spec *spec,
380 struct dt_allocation_hint *hint)
382 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
386 LASSERT(!mdd_object_exists(c));
388 rc = mdo_create_object(env, c, attr, hint, dof, handle);
393 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
394 const struct lu_attr *attr, struct thandle *handle,
400 rc = mdo_attr_set(env, obj, attr, handle);
401 #ifdef CONFIG_FS_POSIX_ACL
402 if (!rc && (attr->la_valid & LA_MODE) && needacl)
403 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
408 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
409 const struct lu_attr *oattr, struct lu_attr *attr,
410 struct thandle *handle)
415 LASSERT(attr->la_valid & LA_CTIME);
416 LASSERT(oattr != NULL);
418 /* Make sure the ctime is increased only, however, it's not strictly
419 * reliable at here because there is not guarantee to hold lock on
420 * object, so we just bypass some unnecessary cmtime setting first
421 * and OSD has to check it again. */
422 if (attr->la_ctime < oattr->la_ctime)
423 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
424 else if (attr->la_valid == LA_CTIME &&
425 attr->la_ctime == oattr->la_ctime)
426 attr->la_valid &= ~LA_CTIME;
428 if (attr->la_valid != 0)
429 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
434 * This gives the same functionality as the code between
435 * sys_chmod and inode_setattr
436 * chown_common and inode_setattr
437 * utimes and inode_setattr
438 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
440 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
441 const struct lu_attr *oattr, struct lu_attr *la,
442 const unsigned long flags)
451 /* Do not permit change file type */
452 if (la->la_valid & LA_TYPE)
455 /* They should not be processed by setattr */
456 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
459 LASSERT(oattr != NULL);
461 uc = lu_ucred_check(env);
465 if (la->la_valid == LA_CTIME) {
466 if (!(flags & MDS_PERM_BYPASS))
467 /* This is only for set ctime when rename's source is
469 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
471 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
472 la->la_valid &= ~LA_CTIME;
476 if (la->la_valid == LA_ATIME) {
477 /* This is an atime-only attribute update for close RPCs. */
478 if (la->la_atime < (oattr->la_atime +
479 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
480 la->la_valid &= ~LA_ATIME;
484 /* Check if flags change. */
485 if (la->la_valid & LA_FLAGS) {
486 unsigned int oldflags = oattr->la_flags &
487 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
488 unsigned int newflags = la->la_flags &
489 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
491 if ((uc->uc_fsuid != oattr->la_uid) &&
492 !md_capable(uc, CFS_CAP_FOWNER))
495 /* The IMMUTABLE and APPEND_ONLY flags can
496 * only be changed by the relevant capability. */
497 if ((oldflags ^ newflags) &&
498 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
501 if (!S_ISDIR(oattr->la_mode))
502 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
505 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
506 (la->la_valid & ~LA_FLAGS) &&
507 !(flags & MDS_PERM_BYPASS))
510 /* Check for setting the obj time. */
511 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
512 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
513 if ((uc->uc_fsuid != oattr->la_uid) &&
514 !md_capable(uc, CFS_CAP_FOWNER)) {
515 rc = mdd_permission_internal(env, obj, oattr,
522 if (la->la_valid & LA_KILL_SUID) {
523 la->la_valid &= ~LA_KILL_SUID;
524 if ((oattr->la_mode & S_ISUID) &&
525 !(la->la_valid & LA_MODE)) {
526 la->la_mode = oattr->la_mode;
527 la->la_valid |= LA_MODE;
529 la->la_mode &= ~S_ISUID;
532 if (la->la_valid & LA_KILL_SGID) {
533 la->la_valid &= ~LA_KILL_SGID;
534 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
535 (S_ISGID | S_IXGRP)) &&
536 !(la->la_valid & LA_MODE)) {
537 la->la_mode = oattr->la_mode;
538 la->la_valid |= LA_MODE;
540 la->la_mode &= ~S_ISGID;
543 /* Make sure a caller can chmod. */
544 if (la->la_valid & LA_MODE) {
545 if (!(flags & MDS_PERM_BYPASS) &&
546 (uc->uc_fsuid != oattr->la_uid) &&
547 !md_capable(uc, CFS_CAP_FOWNER))
550 if (la->la_mode == (umode_t) -1)
551 la->la_mode = oattr->la_mode;
553 la->la_mode = (la->la_mode & S_IALLUGO) |
554 (oattr->la_mode & ~S_IALLUGO);
556 /* Also check the setgid bit! */
557 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
558 la->la_gid : oattr->la_gid) &&
559 !md_capable(uc, CFS_CAP_FSETID))
560 la->la_mode &= ~S_ISGID;
562 la->la_mode = oattr->la_mode;
565 /* Make sure a caller can chown. */
566 if (la->la_valid & LA_UID) {
567 if (la->la_uid == (uid_t) -1)
568 la->la_uid = oattr->la_uid;
569 if (((uc->uc_fsuid != oattr->la_uid) ||
570 (la->la_uid != oattr->la_uid)) &&
571 !md_capable(uc, CFS_CAP_CHOWN))
574 /* If the user or group of a non-directory has been
575 * changed by a non-root user, remove the setuid bit.
576 * 19981026 David C Niemi <niemi@tux.org>
578 * Changed this to apply to all users, including root,
579 * to avoid some races. This is the behavior we had in
580 * 2.0. The check for non-root was definitely wrong
581 * for 2.2 anyway, as it should have been using
582 * CAP_FSETID rather than fsuid -- 19990830 SD. */
583 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
584 !S_ISDIR(oattr->la_mode)) {
585 la->la_mode &= ~S_ISUID;
586 la->la_valid |= LA_MODE;
590 /* Make sure caller can chgrp. */
591 if (la->la_valid & LA_GID) {
592 if (la->la_gid == (gid_t) -1)
593 la->la_gid = oattr->la_gid;
594 if (((uc->uc_fsuid != oattr->la_uid) ||
595 ((la->la_gid != oattr->la_gid) &&
596 !lustre_in_group_p(uc, la->la_gid))) &&
597 !md_capable(uc, CFS_CAP_CHOWN))
600 /* Likewise, if the user or group of a non-directory
601 * has been changed by a non-root user, remove the
602 * setgid bit UNLESS there is no group execute bit
603 * (this would be a file marked for mandatory
604 * locking). 19981026 David C Niemi <niemi@tux.org>
606 * Removed the fsuid check (see the comment above) --
608 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
609 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
610 la->la_mode &= ~S_ISGID;
611 la->la_valid |= LA_MODE;
615 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
616 if (!((flags & MDS_OWNEROVERRIDE) &&
617 (uc->uc_fsuid == oattr->la_uid)) &&
618 !(flags & MDS_PERM_BYPASS)) {
619 rc = mdd_permission_internal(env, obj, oattr,
626 if (la->la_valid & LA_CTIME) {
627 /* The pure setattr, it has the priority over what is
628 * already set, do not drop it if ctime is equal. */
629 if (la->la_ctime < oattr->la_ctime)
630 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
636 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
637 struct mdd_device *mdd,
638 enum changelog_rec_type type, int flags,
639 const struct lu_fid *fid,
640 struct thandle *handle)
642 const struct lu_ucred *uc = lu_ucred(env);
643 struct llog_changelog_rec *rec;
646 int xflags = CLFE_INVALID;
649 flags = (flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
650 if (uc != NULL && uc->uc_jobid[0] != '\0')
653 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
654 changelog_rec_offset(flags & CLF_SUPPORTED,
655 xflags & CLFE_SUPPORTED));
656 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
657 if (buf->lb_buf == NULL)
661 rec->cr_hdr.lrh_len = reclen;
662 rec->cr.cr_flags = flags;
663 rec->cr.cr_type = (__u32)type;
664 rec->cr.cr_tfid = *fid;
665 rec->cr.cr_namelen = 0;
667 if (flags & CLF_JOBID)
668 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
670 if (flags & CLF_EXTRA_FLAGS)
671 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
673 rc = mdd_changelog_store(env, mdd, rec, handle);
679 /** Store a data change changelog record
680 * If this fails, we must fail the whole transaction; we don't
681 * want the change to commit without the log entry.
682 * \param mdd_obj - mdd_object of change
683 * \param handle - transaction handle
685 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
686 enum changelog_rec_type type, int flags,
687 struct mdd_object *mdd_obj, struct thandle *handle)
691 LASSERT(mdd_obj != NULL);
692 LASSERT(handle != NULL);
695 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
697 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
700 if (mdd_is_volatile_obj(mdd_obj))
703 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
704 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
705 /* Don't need multiple updates in this log */
706 /* Don't check under lock - no big deal if we get an extra
711 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
712 mdo2fid(mdd_obj), handle);
714 mdd_obj->mod_cltime = ktime_get();
719 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
720 int flags, struct md_device *m, const struct lu_fid *fid)
722 struct thandle *handle;
723 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
728 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
730 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
733 LASSERT(fid != NULL);
735 handle = mdd_trans_create(env, mdd);
737 RETURN(PTR_ERR(handle));
739 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
743 rc = mdd_trans_start(env, mdd, handle);
747 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
751 rc = mdd_trans_stop(env, mdd, rc, handle);
757 * Save LMA extended attributes with data from \a ma.
759 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
760 * not, LMA EA will be first read from disk, modified and write back.
763 /* Precedence for choosing record type when multiple
764 * attributes change: setattr > mtime > ctime > atime
765 * (ctime changes when mtime does, plus chmod/chown.
766 * atime and ctime are independent.) */
767 static int mdd_attr_set_changelog(const struct lu_env *env,
768 struct md_object *obj, struct thandle *handle,
771 struct mdd_device *mdd = mdo2mdd(obj);
774 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
775 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
776 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
777 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
778 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
779 bits = bits & mdd->mdd_cl.mc_mask;
780 /* This is an implementation limit rather than a protocol limit */
781 CLASSERT(CL_LAST <= sizeof(int) * 8);
785 /* The record type is the lowest non-masked set bit */
788 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
789 return mdd_changelog_data_store(env, mdd, type, (int)valid,
790 md2mdd_obj(obj), handle);
793 static int mdd_declare_attr_set(const struct lu_env *env,
794 struct mdd_device *mdd,
795 struct mdd_object *obj,
796 const struct lu_attr *attr,
797 struct thandle *handle)
801 rc = mdo_declare_attr_set(env, obj, attr, handle);
805 #ifdef CONFIG_FS_POSIX_ACL
806 if (attr->la_valid & LA_MODE) {
807 mdd_read_lock(env, obj, MOR_TGT_CHILD);
808 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
809 XATTR_NAME_ACL_ACCESS);
810 mdd_read_unlock(env, obj);
811 if (rc == -EOPNOTSUPP || rc == -ENODATA)
817 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
818 rc = mdo_declare_xattr_set(env, obj, buf,
819 XATTR_NAME_ACL_ACCESS, 0,
827 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
834 * permission changes may require sync operation, to mitigate performance
835 * impact, only do this for dir and when permission is reduced.
837 * For regular files, version is updated with permission change (see VBR), async
838 * permission won't cause any issue, while missing permission change on
839 * directory may affect accessibility of other objects after recovery.
841 static inline bool permission_needs_sync(const struct lu_attr *old,
842 const struct lu_attr *new)
844 if (!S_ISDIR(old->la_mode))
847 if (new->la_valid & (LA_UID | LA_GID))
850 if (new->la_valid & LA_MODE &&
851 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
854 if ((new->la_valid & LA_MODE) &&
855 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
856 (old->la_mode & S_IRWXUGO))
862 /* set attr and LOV EA at once, return updated attr */
863 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
864 const struct md_attr *ma)
866 struct mdd_object *mdd_obj = md2mdd_obj(obj);
867 struct mdd_device *mdd = mdo2mdd(obj);
868 struct thandle *handle;
869 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
870 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
871 const struct lu_attr *la = &ma->ma_attr;
875 /* we do not use ->attr_set() for LOV/HSM EA any more */
876 LASSERT((ma->ma_valid & MA_LOV) == 0);
877 LASSERT((ma->ma_valid & MA_HSM) == 0);
879 rc = mdd_la_get(env, mdd_obj, attr);
883 *la_copy = ma->ma_attr;
884 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
888 /* no need to setattr anymore */
889 if (la_copy->la_valid == 0) {
890 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
891 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
892 PFID(mdo2fid(mdd_obj)), la->la_valid);
897 handle = mdd_trans_create(env, mdd);
899 RETURN(PTR_ERR(handle));
901 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
905 rc = mdd_trans_start(env, mdd, handle);
909 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
912 if (la->la_valid & (LA_MTIME | LA_CTIME))
913 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
914 la->la_mtime, la->la_ctime);
916 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
917 if (la_copy->la_valid & LA_FLAGS)
918 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
919 else if (la_copy->la_valid) /* setattr */
920 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
921 mdd_write_unlock(env, mdd_obj);
924 rc = mdd_attr_set_changelog(env, obj, handle,
930 rc = mdd_trans_stop(env, mdd, rc, handle);
935 static int mdd_xattr_sanity_check(const struct lu_env *env,
936 struct mdd_object *obj,
937 const struct lu_attr *attr,
940 struct lu_ucred *uc = lu_ucred_assert(env);
943 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
946 if (strncmp(XATTR_USER_PREFIX, name,
947 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
948 /* For sticky directories, only the owner and privileged user
949 * can write attributes. */
950 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
951 (uc->uc_fsuid != attr->la_uid) &&
952 !md_capable(uc, CFS_CAP_FOWNER))
955 if ((uc->uc_fsuid != attr->la_uid) &&
956 !md_capable(uc, CFS_CAP_FOWNER))
964 * Check if a string begins with a given prefix.
966 * \param str String to check
967 * \param prefix Substring to check at the beginning of \a str
968 * \return true/false whether the condition is verified.
970 static inline bool has_prefix(const char *str, const char *prefix)
972 return strncmp(prefix, str, strlen(prefix)) == 0;
976 * Indicate the kind of changelog to store (if any) for a xattr set/del.
978 * \param[in] xattr_name Full extended attribute name.
980 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
982 static enum changelog_rec_type
983 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
984 const char *xattr_name)
986 /* Layout changes systematically recorded */
987 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
988 strncmp(XATTR_LUSTRE_LOV, xattr_name,
989 strlen(XATTR_LUSTRE_LOV)) == 0)
992 /* HSM information changes systematically recorded */
993 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
996 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
997 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
998 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
1004 static int mdd_declare_xattr_set(const struct lu_env *env,
1005 struct mdd_device *mdd,
1006 struct mdd_object *obj,
1007 const struct lu_buf *buf,
1009 int fl, struct thandle *handle)
1013 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1017 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1018 return 0; /* no changelog to store */
1020 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1024 * Compare current and future data of HSM EA and add a changelog if needed.
1026 * Caller should have write-locked \param obj.
1028 * \param buf - Future HSM EA content.
1029 * \retval 0 if no changelog is needed or changelog was added properly.
1030 * \retval -ve errno if there was a problem
1032 static int mdd_hsm_update_locked(const struct lu_env *env,
1033 struct md_object *obj,
1034 const struct lu_buf *buf,
1035 struct thandle *handle, int *cl_flags)
1037 struct mdd_thread_info *info = mdd_env_info(env);
1038 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1039 struct lu_buf *current_buf;
1040 struct md_hsm *current_mh;
1041 struct md_hsm *new_mh;
1045 OBD_ALLOC_PTR(current_mh);
1046 if (current_mh == NULL)
1049 /* Read HSM attrs from disk */
1050 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1051 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1052 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1053 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1054 if (rc < 0 && rc != -ENODATA)
1056 else if (rc == -ENODATA)
1057 current_mh->mh_flags = 0;
1059 /* Map future HSM xattr */
1060 OBD_ALLOC_PTR(new_mh);
1062 GOTO(free, rc = -ENOMEM);
1063 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1067 /* Flags differ, set flags for the changelog that will be added */
1068 if (current_mh->mh_flags != new_mh->mh_flags) {
1069 hsm_set_cl_event(cl_flags, HE_STATE);
1070 if (new_mh->mh_flags & HS_DIRTY)
1071 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1074 OBD_FREE_PTR(new_mh);
1077 OBD_FREE_PTR(current_mh);
1081 static int mdd_declare_xattr_del(const struct lu_env *env,
1082 struct mdd_device *mdd,
1083 struct mdd_object *obj,
1085 struct thandle *handle);
1087 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1090 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1091 struct md_object *md_vic)
1093 struct mdd_device *mdd = mdo2mdd(md_obj);
1094 struct mdd_object *obj = md2mdd_obj(md_obj);
1095 struct mdd_object *vic = md2mdd_obj(md_vic);
1096 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1097 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1098 struct lov_mds_md *lmm;
1099 struct thandle *handle;
1103 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1104 if (rc == 0) /* same fid */
1107 handle = mdd_trans_create(env, mdd);
1109 RETURN(PTR_ERR(handle));
1112 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1113 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1115 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1116 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1119 /* get EA of victim file */
1120 memset(buf_vic, 0, sizeof(*buf_vic));
1121 rc = mdd_get_lov_ea(env, vic, buf_vic);
1128 /* parse the layout of victim file */
1129 lmm = buf_vic->lb_buf;
1130 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1131 GOTO(out, rc = -EINVAL);
1133 /* save EA of target file for restore */
1134 memset(buf, 0, sizeof(*buf));
1135 rc = mdd_get_lov_ea(env, obj, buf);
1139 /* Get rid of the layout from victim object */
1140 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1144 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1145 LU_XATTR_MERGE, handle);
1149 rc = mdd_trans_start(env, mdd, handle);
1153 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1158 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1159 if (rc) { /* wtf? */
1162 rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1163 LU_XATTR_REPLACE, handle);
1165 CERROR("%s: failed to rollback of layout of: "DFID
1166 ": %d, file state unknown\n",
1167 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1171 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1172 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1176 mdd_trans_stop(env, mdd, rc, handle);
1177 mdd_write_unlock(env, obj);
1178 mdd_write_unlock(env, vic);
1180 lu_buf_free(buf_vic);
1185 static int mdd_layout_merge_allowed(const struct lu_env *env,
1186 struct md_object *target,
1187 struct md_object *victim)
1189 struct mdd_object *o1 = md2mdd_obj(target);
1191 /* cannot extend directory's LOVEA */
1192 if (S_ISDIR(mdd_object_type(o1))) {
1193 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1194 mdd_obj_dev_name(o1));
1202 * The caller should guarantee to update the object ctime
1203 * after xattr_set if needed.
1205 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1206 const struct lu_buf *buf, const char *name,
1209 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1210 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1211 struct mdd_device *mdd = mdo2mdd(obj);
1212 struct thandle *handle;
1213 enum changelog_rec_type cl_type;
1218 rc = mdd_la_get(env, mdd_obj, attr);
1222 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1226 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 && fl == LU_XATTR_MERGE) {
1227 struct md_object *victim = buf->lb_buf;
1229 if (buf->lb_len != sizeof(victim))
1232 rc = mdd_layout_merge_allowed(env, obj, victim);
1236 /* merge layout of victim as a mirror of obj's. */
1237 rc = mdd_xattr_merge(env, obj, victim);
1241 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1242 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1243 struct posix_acl *acl;
1245 /* user may set empty ACL, which should be treated as removing
1247 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1250 RETURN(PTR_ERR(acl));
1252 rc = mdd_xattr_del(env, obj, name);
1255 posix_acl_release(acl);
1258 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1259 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1263 handle = mdd_trans_create(env, mdd);
1265 RETURN(PTR_ERR(handle));
1267 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1271 rc = mdd_trans_start(env, mdd, handle);
1275 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1277 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1278 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1280 mdd_write_unlock(env, mdd_obj);
1285 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1286 mdd_write_unlock(env, mdd_obj);
1290 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1294 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1299 return mdd_trans_stop(env, mdd, rc, handle);
1302 static int mdd_declare_xattr_del(const struct lu_env *env,
1303 struct mdd_device *mdd,
1304 struct mdd_object *obj,
1306 struct thandle *handle)
1310 rc = mdo_declare_xattr_del(env, obj, name, handle);
1314 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1315 return 0; /* no changelog to store */
1317 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1321 * The caller should guarantee to update the object ctime
1322 * after xattr_set if needed.
1324 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1327 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1328 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1329 struct mdd_device *mdd = mdo2mdd(obj);
1330 struct thandle *handle;
1334 rc = mdd_la_get(env, mdd_obj, attr);
1338 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1342 handle = mdd_trans_create(env, mdd);
1344 RETURN(PTR_ERR(handle));
1346 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1350 rc = mdd_trans_start(env, mdd, handle);
1354 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1355 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1356 mdd_write_unlock(env, mdd_obj);
1360 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1363 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1367 return mdd_trans_stop(env, mdd, rc, handle);
1371 * read lov EA of an object
1372 * return the lov EA in an allocated lu_buf
1374 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1375 struct lu_buf *lmm_buf)
1377 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1382 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1384 if (rc == -ERANGE) {
1385 /* mti_big_buf is allocated but is too small
1386 * we need to increase it */
1387 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1389 if (buf->lb_buf == NULL)
1390 GOTO(out, rc = -ENOMEM);
1401 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1402 /* mti_big_buf was not allocated, so we have to
1403 * allocate it based on the ea size */
1404 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1406 if (buf->lb_buf == NULL)
1407 GOTO(out, rc = -ENOMEM);
1411 lu_buf_alloc(lmm_buf, bufsize);
1412 if (lmm_buf->lb_buf == NULL)
1413 GOTO(out, rc = -ENOMEM);
1415 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1421 lu_buf_free(lmm_buf);
1425 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1426 struct mdd_object *o, struct lu_buf *buf,
1427 struct thandle *handle)
1429 struct hsm_attrs *attrs;
1435 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1440 attrs = buf->lb_buf;
1441 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1442 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1445 /* Add a changelog record for release. */
1446 hsm_set_cl_event(&flags, HE_RELEASE);
1447 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1453 * check if layout swapping between 2 objects is allowed
1455 * - only normal FIDs or non-system IGIFs
1456 * - same type of objects
1457 * - same owner/group (so quotas are still valid) unless this is from HSM
1460 static int mdd_layout_swap_allowed(const struct lu_env *env,
1461 struct mdd_object *o1,
1462 const struct lu_attr *attr1,
1463 struct mdd_object *o2,
1464 const struct lu_attr *attr2,
1467 const struct lu_fid *fid1, *fid2;
1473 if (!fid_is_norm(fid1) &&
1474 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1477 if (!fid_is_norm(fid2) &&
1478 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1481 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1482 if (S_ISDIR(mdd_object_type(o1)))
1484 if (S_ISREG(mdd_object_type(o1)))
1489 if (flags & SWAP_LAYOUTS_MDS_HSM)
1492 if ((attr1->la_uid != attr2->la_uid) ||
1493 (attr1->la_gid != attr2->la_gid))
1499 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1500 * look into the layout in MDD layer. */
1501 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1503 struct lov_comp_md_v1 *comp_v1;
1504 struct lov_mds_md *v1;
1508 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1509 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1510 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1516 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1517 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1520 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1521 off = le32_to_cpu(comp_v1->lcm_entries[i].
1523 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1528 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1529 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1540 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1542 return mdd_lmm_oi(lmm, oi, true);
1545 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1547 return mdd_lmm_oi(lmm, oi, false);
1550 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1552 struct lov_comp_md_v1 *comp_v1;
1554 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1555 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1557 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1559 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1560 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1561 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1562 __u16 tmp_gen = *gen;
1564 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1566 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1573 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1575 return mdd_lmm_gen(lmm, gen, true);
1578 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1580 return mdd_lmm_gen(lmm, gen, false);
1584 * swap layouts between 2 lustre objects
1586 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1587 struct md_object *obj2, __u64 flags)
1589 struct mdd_thread_info *info = mdd_env_info(env);
1590 struct mdd_object *fst_o = md2mdd_obj(obj1);
1591 struct mdd_object *snd_o = md2mdd_obj(obj2);
1592 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1593 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1594 struct mdd_device *mdd = mdo2mdd(obj1);
1595 struct lov_mds_md *fst_lmm, *snd_lmm;
1596 struct lu_buf *fst_buf = &info->mti_buf[0];
1597 struct lu_buf *snd_buf = &info->mti_buf[1];
1598 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1599 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1600 struct ost_id *saved_oi = NULL;
1601 struct thandle *handle;
1602 __u32 fst_gen, snd_gen, saved_gen;
1608 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1609 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1611 /* we have to sort the 2 obj, so locking will always
1612 * be in the same order, even in case of 2 concurrent swaps */
1613 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1614 if (rc == 0) /* same fid ? */
1620 rc = mdd_la_get(env, fst_o, fst_la);
1624 rc = mdd_la_get(env, snd_o, snd_la);
1628 /* check if layout swapping is allowed */
1629 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1633 handle = mdd_trans_create(env, mdd);
1635 RETURN(PTR_ERR(handle));
1637 /* objects are already sorted */
1638 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1639 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1641 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1642 if (rc < 0 && rc != -ENODATA)
1645 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1646 if (rc < 0 && rc != -ENODATA)
1649 /* swapping 2 non existant layouts is a success */
1650 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1653 /* to help inode migration between MDT, it is better to
1654 * start by the no layout file (if one), so we order the swap */
1655 if (snd_buf->lb_buf == NULL) {
1657 swap(fst_buf, snd_buf);
1660 fst_gen = snd_gen = 0;
1661 /* lmm and generation layout initialization */
1662 if (fst_buf->lb_buf != NULL) {
1663 fst_lmm = fst_buf->lb_buf;
1664 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1665 fst_fl = LU_XATTR_REPLACE;
1669 fst_fl = LU_XATTR_CREATE;
1672 snd_lmm = snd_buf->lb_buf;
1673 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1675 saved_gen = fst_gen;
1676 /* increase the generation layout numbers */
1681 * XXX The layout generation is used to generate component IDs for
1682 * the composite file, we have to do some special tweaks to make
1683 * sure the layout generation is always adequate for that job.
1686 /* Skip invalid generation number for composite layout */
1687 if ((snd_gen & LCME_ID_MASK) == 0)
1689 if ((fst_gen & LCME_ID_MASK) == 0)
1691 /* Make sure the generation is greater than all the component IDs */
1692 if (fst_gen < snd_gen)
1694 else if (fst_gen > snd_gen)
1697 /* set the file specific informations in lmm */
1698 if (fst_lmm != NULL) {
1699 saved_oi = &info->mti_oa.o_oi;
1700 mdd_get_lmm_oi(fst_lmm, saved_oi);
1701 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1702 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1703 mdd_set_lmm_oi(snd_lmm, saved_oi);
1705 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1706 cpu_to_le32(LOV_MAGIC_MAGIC))
1707 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1709 GOTO(stop, rc = -EPROTO);
1711 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1713 /* Prepare HSM attribute if it's required */
1714 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1715 const int buflen = sizeof(struct hsm_attrs);
1717 lu_buf_alloc(fst_hsm_buf, buflen);
1718 lu_buf_alloc(snd_hsm_buf, buflen);
1719 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1720 GOTO(stop, rc = -ENOMEM);
1722 /* Read HSM attribute */
1723 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1727 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1731 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1732 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1737 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1738 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1744 /* prepare transaction */
1745 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1750 if (fst_buf->lb_buf != NULL)
1751 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1752 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1755 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1760 rc = mdd_trans_start(env, mdd, handle);
1764 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1765 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1769 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1771 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1774 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1775 mdd_obj_dev_name(fst_o),
1776 PFID(mdo2fid(fst_o)), rc, rc2);
1781 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1785 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1788 if (fst_buf->lb_buf != NULL)
1789 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1790 LU_XATTR_REPLACE, handle);
1792 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1798 /* failure on second file, but first was done, so we have
1799 * to roll back first. */
1800 if (fst_buf->lb_buf != NULL) {
1801 mdd_set_lmm_oi(fst_lmm, saved_oi);
1802 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1803 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1804 LU_XATTR_REPLACE, handle);
1806 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1812 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1817 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1822 CERROR("%s: unable to roll back layout swap. FIDs: "
1823 DFID" and "DFID "error: %d/%d, steps: %d\n",
1824 mdd_obj_dev_name(fst_o),
1825 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1827 /* a solution to avoid journal commit is to panic,
1828 * but it has strong consequences so we use LBUG to
1829 * allow sysdamin to choose to panic or not
1836 /* Issue one changelog record per file */
1837 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1841 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1847 rc = mdd_trans_stop(env, mdd, rc, handle);
1849 mdd_write_unlock(env, snd_o);
1850 mdd_write_unlock(env, fst_o);
1852 lu_buf_free(fst_buf);
1853 lu_buf_free(snd_buf);
1854 lu_buf_free(fst_hsm_buf);
1855 lu_buf_free(snd_hsm_buf);
1859 static int mdd_declare_layout_change(const struct lu_env *env,
1860 struct mdd_device *mdd,
1861 struct mdd_object *obj,
1862 struct md_layout_change *mlc,
1863 struct thandle *handle)
1867 rc = mdo_declare_layout_change(env, obj, mlc, handle);
1871 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1874 /* For PFL, this is used to instantiate necessary component objects. */
1876 mdd_layout_instantiate_component(const struct lu_env *env,
1877 struct mdd_object *obj, struct md_layout_change *mlc,
1878 struct thandle *handle)
1880 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1884 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
1887 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1889 * It's possible that another layout write intent has already
1890 * instantiated our objects, so a -EALREADY returned, and we need to
1894 RETURN(rc == -EALREADY ? 0 : rc);
1896 rc = mdd_trans_start(env, mdd, handle);
1900 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1901 rc = mdo_layout_change(env, obj, mlc, handle);
1902 mdd_write_unlock(env, obj);
1906 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1911 * Change the FLR layout from RDONLY to WRITE_PENDING.
1913 * It picks the primary mirror, and bumps the layout version, and set
1914 * layout version xattr to OST objects in a sync tx. In order to facilitate
1915 * the handling of phantom writers from evicted clients, the clients carry
1916 * layout version of the file with write RPC, so that the OSTs can verify
1917 * if the write RPCs are legitimate, meaning not from evicted clients.
1920 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
1921 struct md_layout_change *mlc, struct thandle *handle)
1923 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1927 /* Verify acceptable operations */
1928 switch (mlc->mlc_opc) {
1929 case MD_LAYOUT_WRITE:
1931 case MD_LAYOUT_RESYNC:
1932 /* these are legal operations - this represents the case that
1933 * a few mirrors were missed in the last resync.
1934 * XXX: it will be supported later */
1935 case MD_LAYOUT_RESYNC_DONE:
1940 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1944 rc = mdd_declare_xattr_del(env, mdd, obj, XATTR_NAME_SOM, handle);
1948 /* record a changelog for data mover to consume */
1949 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1953 rc = mdd_trans_start(env, mdd, handle);
1957 /* it needs a sync tx to make FLR to work properly */
1958 handle->th_sync = 1;
1960 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1961 rc = mdo_layout_change(env, obj, mlc, handle);
1963 rc = mdo_xattr_del(env, obj, XATTR_NAME_SOM, handle);
1967 mdd_write_unlock(env, obj);
1971 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
1982 * Handle mirrored file state transition when it's in WRITE_PENDING.
1984 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
1985 * the file is in WRITE_PENDING state. If everything goes fine, the file's
1986 * layout version will be increased, and the file's state will be changed to
1990 mdd_layout_update_write_pending(const struct lu_env *env,
1991 struct mdd_object *obj, struct md_layout_change *mlc,
1992 struct thandle *handle)
1994 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1998 switch (mlc->mlc_opc) {
1999 case MD_LAYOUT_RESYNC:
2000 /* Upon receiving the resync request, it should
2001 * instantiate all stale components right away to get ready
2002 * for mirror copy. In order to avoid layout version change,
2003 * client should avoid sending LAYOUT_WRITE request at the
2006 case MD_LAYOUT_WRITE:
2007 /* legal race for concurrent write, the file state has been
2008 * changed by another client. */
2014 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2018 rc = mdd_trans_start(env, mdd, handle);
2022 /* it needs a sync tx to make FLR to work properly */
2023 handle->th_sync = 1;
2025 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2026 rc = mdo_layout_change(env, obj, mlc, handle);
2027 mdd_write_unlock(env, obj);
2038 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2040 * Only concurrent write and sync complete requests are possible when the
2041 * file is in SYNC_PENDING. For the latter request, it will pass in the
2042 * mirrors that have been synchronized, then the stale bit will be cleared
2043 * to make the file's state turn into RDONLY.
2044 * For concurrent write reqeust, it just needs to change the file's state
2045 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2046 * version because the version will be increased in the transition to
2047 * SYNC_PENDING later so that it can deny the write request from potential
2048 * evicted SYNC clients. */
2050 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2051 struct md_layout_change *mlc, struct thandle *handle)
2053 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2054 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2059 /* operation validation */
2060 switch (mlc->mlc_opc) {
2061 case MD_LAYOUT_RESYNC_DONE:
2062 /* resync complete. */
2063 case MD_LAYOUT_WRITE:
2064 /* concurrent write. */
2066 case MD_LAYOUT_RESYNC:
2067 /* resync again, most likely the previous run failed.
2068 * no-op if it's already in SYNC_PENDING state */
2074 if (mlc->mlc_som.lsa_valid & LSOM_FL_VALID) {
2075 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2076 if (rc && rc != -ENODATA)
2079 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2080 som_buf->lb_buf = &mlc->mlc_som;
2081 som_buf->lb_len = sizeof(mlc->mlc_som);
2084 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2088 /* record a changelog for the completion of resync */
2089 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2093 /* RESYNC_DONE has piggybacked size and blocks */
2095 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2096 XATTR_NAME_SOM, fl, handle);
2101 rc = mdd_trans_start(env, mdd, handle);
2105 /* it needs a sync tx to make FLR to work properly */
2106 handle->th_sync = 1;
2108 rc = mdo_layout_change(env, obj, mlc, handle);
2113 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2119 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2128 * Layout change callback for object.
2130 * This is only used by FLR for now. In the future, it can be exteneded to
2131 * handle all layout change.
2134 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2135 struct md_layout_change *mlc)
2137 struct mdd_object *obj = md2mdd_obj(o);
2138 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2139 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2140 struct lov_comp_md_v1 *lcm;
2141 struct thandle *handle;
2146 /* Verify acceptable operations */
2147 switch (mlc->mlc_opc) {
2148 case MD_LAYOUT_WRITE:
2149 case MD_LAYOUT_RESYNC:
2150 case MD_LAYOUT_RESYNC_DONE:
2156 handle = mdd_trans_create(env, mdd);
2158 RETURN(PTR_ERR(handle));
2160 rc = mdd_get_lov_ea(env, obj, buf);
2167 /* analyze the layout to make sure it's a FLR file */
2169 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2170 GOTO(out, rc = -EINVAL);
2172 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2174 /* please refer to HLD of FLR for state transition */
2175 switch (flr_state) {
2176 case LCM_FL_NOT_FLR:
2177 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2179 case LCM_FL_WRITE_PENDING:
2180 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2183 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2185 case LCM_FL_SYNC_PENDING:
2186 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2195 mdd_trans_stop(env, mdd, rc, handle);
2200 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2201 struct mdd_object *child, const struct lu_attr *attr,
2202 const struct md_op_spec *spec,
2203 struct dt_allocation_hint *hint)
2205 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2206 struct dt_object *nc = mdd_object_child(child);
2208 memset(hint, 0, sizeof(*hint));
2210 /* For striped directory, give striping EA to lod_ah_init, which will
2211 * decide the stripe_offset and stripe count by it. */
2212 if (S_ISDIR(attr->la_mode) &&
2213 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2214 hint->dah_eadata = spec->u.sp_ea.eadata;
2215 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2217 hint->dah_eadata = NULL;
2218 hint->dah_eadata_len = 0;
2221 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2222 hint->dah_eadata, hint->dah_eadata_len);
2223 /* @hint will be initialized by underlying device. */
2224 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2228 * do NOT or the MAY_*'s, you'll get the weakest
2230 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
2234 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2235 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2236 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2237 * owner can write to a file even if it is marked readonly to hide
2238 * its brokenness. (bug 5781) */
2239 if (flags & MDS_OPEN_OWNEROVERRIDE) {
2240 struct lu_ucred *uc = lu_ucred_check(env);
2242 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2246 if (flags & FMODE_READ)
2248 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
2250 if (flags & MDS_FMODE_EXEC)
2255 static int mdd_open_sanity_check(const struct lu_env *env,
2256 struct mdd_object *obj,
2257 const struct lu_attr *attr, int flag)
2263 if (mdd_is_dead_obj(obj))
2266 if (S_ISLNK(attr->la_mode))
2269 mode = accmode(env, attr, flag);
2271 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
2274 if (!(flag & MDS_OPEN_CREATED)) {
2275 rc = mdd_permission_internal(env, obj, attr, mode);
2280 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
2281 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
2282 flag &= ~MDS_OPEN_TRUNC;
2284 /* For writing append-only file must open it with append mode. */
2285 if (attr->la_flags & LUSTRE_APPEND_FL) {
2286 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
2288 if (flag & MDS_OPEN_TRUNC)
2295 static int mdd_open(const struct lu_env *env, struct md_object *obj,
2298 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2299 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2302 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2304 rc = mdd_la_get(env, mdd_obj, attr);
2308 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
2312 mdd_obj->mod_count++;
2315 mdd_write_unlock(env, mdd_obj);
2319 static int mdd_declare_close(const struct lu_env *env,
2320 struct mdd_object *obj,
2322 struct thandle *handle)
2326 rc = orph_declare_index_delete(env, obj, handle);
2330 return mdo_declare_destroy(env, obj, handle);
2334 * No permission check is needed.
2336 static int mdd_close(const struct lu_env *env, struct md_object *obj,
2337 struct md_attr *ma, int mode)
2339 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2340 struct mdd_device *mdd = mdo2mdd(obj);
2341 struct thandle *handle = NULL;
2344 bool blocked = false;
2347 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
2348 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2349 mdd_obj->mod_count--;
2350 mdd_write_unlock(env, mdd_obj);
2352 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
2353 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
2354 "list\n", PFID(mdd_object_fid(mdd_obj)));
2358 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
2359 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
2360 /* check without any lock */
2361 is_orphan = mdd_obj->mod_count == 1 &&
2362 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
2366 /* mdd_trans_create() maybe failed because of barrier_entry(),
2367 * under such case, the orphan MDT-object will be left in the
2368 * orphan list, and when the MDT remount next time, the unused
2369 * orphans will be destroyed automatically.
2371 * One exception: the former mdd_finish_unlink may failed to
2372 * add the orphan MDT-object to the orphan list, then if the
2373 * mdd_trans_create() failed because of barrier_entry(), the
2374 * MDT-object will become real orphan that is neither in the
2375 * namespace nor in the orphan list. Such bad case should be
2376 * very rare and will be handled by e2fsck/lfsck. */
2377 handle = mdd_trans_create(env, mdo2mdd(obj));
2378 if (IS_ERR(handle)) {
2379 rc = PTR_ERR(handle);
2380 if (rc != -EINPROGRESS)
2388 rc = mdd_declare_close(env, mdd_obj, ma, handle);
2392 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2396 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2402 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2403 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
2405 CERROR("Failed to get lu_attr of "DFID": %d\n",
2406 PFID(mdd_object_fid(mdd_obj)), rc);
2410 /* check again with lock */
2411 is_orphan = (mdd_obj->mod_count == 1) &&
2412 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
2413 ma->ma_attr.la_nlink == 0);
2415 if (is_orphan && !handle && !blocked) {
2416 mdd_write_unlock(env, mdd_obj);
2420 mdd_obj->mod_count--; /*release open count */
2422 if (!is_orphan || blocked)
2426 /* NB: Object maybe not in orphan list originally, it is rare case for
2427 * mdd_finish_unlink() failure, in that case, the object doesn't have
2428 * ORPHAN_OBJ flag */
2429 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2430 /* remove link to object from orphan index */
2431 LASSERT(handle != NULL);
2432 rc = __mdd_orphan_del(env, mdd_obj, handle);
2434 CERROR("%s: unable to delete "DFID" from orphan list: "
2435 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2436 PFID(mdd_object_fid(mdd_obj)), rc);
2437 /* If object was not deleted from orphan list, do not
2438 * destroy OSS objects, which will be done when next
2443 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2444 "list, OSS objects to be destroyed.\n",
2445 PFID(mdd_object_fid(mdd_obj)));
2448 rc = mdo_destroy(env, mdd_obj, handle);
2451 CERROR("%s: unable to delete "DFID" from orphan list: "
2452 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2453 PFID(mdd_object_fid(mdd_obj)), rc);
2458 mdd_write_unlock(env, mdd_obj);
2460 if (!rc && !blocked &&
2461 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2462 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2463 if (handle == NULL) {
2464 handle = mdd_trans_create(env, mdo2mdd(obj));
2466 GOTO(stop, rc = PTR_ERR(handle));
2468 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2473 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2478 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2483 if (handle != NULL && !IS_ERR(handle))
2484 rc = mdd_trans_stop(env, mdd, rc, handle);
2490 * Permission check is done when open,
2491 * no need check again.
2493 static int mdd_readpage_sanity_check(const struct lu_env *env,
2494 struct mdd_object *obj)
2496 struct dt_object *next = mdd_object_child(obj);
2500 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2508 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2509 size_t nob, const struct dt_it_ops *iops,
2510 struct dt_it *it, __u32 attr, void *arg)
2512 struct lu_dirpage *dp = &lp->lp_dir;
2516 struct lu_dirent *ent;
2517 struct lu_dirent *last = NULL;
2521 if (nob < sizeof(*dp))
2524 memset(area, 0, sizeof (*dp));
2525 area += sizeof (*dp);
2526 nob -= sizeof (*dp);
2533 len = iops->key_size(env, it);
2535 /* IAM iterator can return record with zero len. */
2539 hash = iops->store(env, it);
2540 if (unlikely(first)) {
2542 dp->ldp_hash_start = cpu_to_le64(hash);
2545 /* calculate max space required for lu_dirent */
2546 recsize = lu_dirent_calc_size(len, attr);
2548 if (nob >= recsize) {
2549 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2550 if (result == -ESTALE)
2555 /* osd might not able to pack all attributes,
2556 * so recheck rec length */
2557 recsize = le16_to_cpu(ent->lde_reclen);
2559 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2560 fid_le_to_cpu(&fid, &ent->lde_fid);
2561 if (fid_is_dot_lustre(&fid))
2565 result = (last != NULL) ? 0 :-EINVAL;
2569 ent = (void *)ent + recsize;
2573 result = iops->next(env, it);
2574 if (result == -ESTALE)
2576 } while (result == 0);
2579 dp->ldp_hash_end = cpu_to_le64(hash);
2581 if (last->lde_hash == dp->ldp_hash_end)
2582 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2583 last->lde_reclen = 0; /* end mark */
2586 /* end of directory */
2587 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2588 else if (result < 0)
2589 CWARN("build page failed: %d!\n", result);
2593 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2594 const struct lu_rdpg *rdpg)
2596 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2600 if (mdd_object_exists(mdd_obj) == 0) {
2601 CERROR("%s: object "DFID" not found: rc = -2\n",
2602 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2606 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2607 rc = mdd_readpage_sanity_check(env, mdd_obj);
2609 GOTO(out_unlock, rc);
2611 if (mdd_is_dead_obj(mdd_obj)) {
2613 struct lu_dirpage *dp;
2616 * According to POSIX, please do not return any entry to client:
2617 * even dot and dotdot should not be returned.
2619 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2620 PFID(mdd_object_fid(mdd_obj)));
2622 if (rdpg->rp_count <= 0)
2623 GOTO(out_unlock, rc = -EFAULT);
2624 LASSERT(rdpg->rp_pages != NULL);
2626 pg = rdpg->rp_pages[0];
2627 dp = (struct lu_dirpage *)kmap(pg);
2628 memset(dp, 0 , sizeof(struct lu_dirpage));
2629 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2630 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2631 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2633 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2636 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2637 mdd_dir_page_build, NULL);
2639 struct lu_dirpage *dp;
2641 dp = kmap(rdpg->rp_pages[0]);
2642 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2645 * No pages were processed, mark this for first page
2648 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2649 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2650 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2652 kunmap(rdpg->rp_pages[0]);
2655 GOTO(out_unlock, rc);
2657 mdd_read_unlock(env, mdd_obj);
2661 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2663 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2665 if (mdd_object_exists(mdd_obj) == 0) {
2668 CERROR("%s: object "DFID" not found: rc = %d\n",
2669 mdd_obj_dev_name(mdd_obj),
2670 PFID(mdd_object_fid(mdd_obj)), rc);
2673 return dt_object_sync(env, mdd_object_child(mdd_obj),
2677 static int mdd_object_lock(const struct lu_env *env,
2678 struct md_object *obj,
2679 struct lustre_handle *lh,
2680 struct ldlm_enqueue_info *einfo,
2681 union ldlm_policy_data *policy)
2683 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2684 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2688 static int mdd_object_unlock(const struct lu_env *env,
2689 struct md_object *obj,
2690 struct ldlm_enqueue_info *einfo,
2691 union ldlm_policy_data *policy)
2693 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2694 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2697 const struct md_object_operations mdd_obj_ops = {
2698 .moo_permission = mdd_permission,
2699 .moo_attr_get = mdd_attr_get,
2700 .moo_attr_set = mdd_attr_set,
2701 .moo_xattr_get = mdd_xattr_get,
2702 .moo_xattr_set = mdd_xattr_set,
2703 .moo_xattr_list = mdd_xattr_list,
2704 .moo_invalidate = mdd_invalidate,
2705 .moo_xattr_del = mdd_xattr_del,
2706 .moo_swap_layouts = mdd_swap_layouts,
2707 .moo_open = mdd_open,
2708 .moo_close = mdd_close,
2709 .moo_readpage = mdd_readpage,
2710 .moo_readlink = mdd_readlink,
2711 .moo_changelog = mdd_changelog,
2712 .moo_object_sync = mdd_object_sync,
2713 .moo_object_lock = mdd_object_lock,
2714 .moo_object_unlock = mdd_object_unlock,
2715 .moo_layout_change = mdd_layout_change,