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 xflags |= CLFE_UIDGID;
655 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
656 changelog_rec_offset(flags & CLF_SUPPORTED,
657 xflags & CLFE_SUPPORTED));
658 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
659 if (buf->lb_buf == NULL)
663 rec->cr_hdr.lrh_len = reclen;
664 rec->cr.cr_flags = flags;
665 rec->cr.cr_type = (__u32)type;
666 rec->cr.cr_tfid = *fid;
667 rec->cr.cr_namelen = 0;
669 if (flags & CLF_JOBID)
670 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
672 if (flags & CLF_EXTRA_FLAGS) {
673 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
674 if (xflags & CLFE_UIDGID)
675 mdd_changelog_rec_extra_uidgid(&rec->cr,
676 uc->uc_uid, uc->uc_gid);
679 rc = mdd_changelog_store(env, mdd, rec, handle);
685 /** Store a data change changelog record
686 * If this fails, we must fail the whole transaction; we don't
687 * want the change to commit without the log entry.
688 * \param mdd_obj - mdd_object of change
689 * \param handle - transaction handle
691 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
692 enum changelog_rec_type type, int flags,
693 struct mdd_object *mdd_obj, struct thandle *handle)
697 LASSERT(mdd_obj != NULL);
698 LASSERT(handle != NULL);
701 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
703 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
706 if (mdd_is_volatile_obj(mdd_obj))
709 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
710 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
711 /* Don't need multiple updates in this log */
712 /* Don't check under lock - no big deal if we get an extra
717 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
718 mdo2fid(mdd_obj), handle);
720 mdd_obj->mod_cltime = ktime_get();
725 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
726 int flags, struct md_device *m, const struct lu_fid *fid)
728 struct thandle *handle;
729 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
734 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
736 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
739 LASSERT(fid != NULL);
741 handle = mdd_trans_create(env, mdd);
743 RETURN(PTR_ERR(handle));
745 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
749 rc = mdd_trans_start(env, mdd, handle);
753 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
757 rc = mdd_trans_stop(env, mdd, rc, handle);
763 * Save LMA extended attributes with data from \a ma.
765 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
766 * not, LMA EA will be first read from disk, modified and write back.
769 /* Precedence for choosing record type when multiple
770 * attributes change: setattr > mtime > ctime > atime
771 * (ctime changes when mtime does, plus chmod/chown.
772 * atime and ctime are independent.) */
773 static int mdd_attr_set_changelog(const struct lu_env *env,
774 struct md_object *obj, struct thandle *handle,
777 struct mdd_device *mdd = mdo2mdd(obj);
780 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
781 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
782 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
783 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
784 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
785 bits = bits & mdd->mdd_cl.mc_mask;
786 /* This is an implementation limit rather than a protocol limit */
787 CLASSERT(CL_LAST <= sizeof(int) * 8);
791 /* The record type is the lowest non-masked set bit */
794 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
795 return mdd_changelog_data_store(env, mdd, type, (int)valid,
796 md2mdd_obj(obj), handle);
799 static int mdd_declare_attr_set(const struct lu_env *env,
800 struct mdd_device *mdd,
801 struct mdd_object *obj,
802 const struct lu_attr *attr,
803 struct thandle *handle)
807 rc = mdo_declare_attr_set(env, obj, attr, handle);
811 #ifdef CONFIG_FS_POSIX_ACL
812 if (attr->la_valid & LA_MODE) {
813 mdd_read_lock(env, obj, MOR_TGT_CHILD);
814 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
815 XATTR_NAME_ACL_ACCESS);
816 mdd_read_unlock(env, obj);
817 if (rc == -EOPNOTSUPP || rc == -ENODATA)
823 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
824 rc = mdo_declare_xattr_set(env, obj, buf,
825 XATTR_NAME_ACL_ACCESS, 0,
833 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
840 * permission changes may require sync operation, to mitigate performance
841 * impact, only do this for dir and when permission is reduced.
843 * For regular files, version is updated with permission change (see VBR), async
844 * permission won't cause any issue, while missing permission change on
845 * directory may affect accessibility of other objects after recovery.
847 static inline bool permission_needs_sync(const struct lu_attr *old,
848 const struct lu_attr *new)
850 if (!S_ISDIR(old->la_mode))
853 if (new->la_valid & (LA_UID | LA_GID))
856 if (new->la_valid & LA_MODE &&
857 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
860 if ((new->la_valid & LA_MODE) &&
861 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
862 (old->la_mode & S_IRWXUGO))
868 /* set attr and LOV EA at once, return updated attr */
869 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
870 const struct md_attr *ma)
872 struct mdd_object *mdd_obj = md2mdd_obj(obj);
873 struct mdd_device *mdd = mdo2mdd(obj);
874 struct thandle *handle;
875 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
876 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
877 const struct lu_attr *la = &ma->ma_attr;
881 /* we do not use ->attr_set() for LOV/HSM EA any more */
882 LASSERT((ma->ma_valid & MA_LOV) == 0);
883 LASSERT((ma->ma_valid & MA_HSM) == 0);
885 rc = mdd_la_get(env, mdd_obj, attr);
889 *la_copy = ma->ma_attr;
890 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
894 /* no need to setattr anymore */
895 if (la_copy->la_valid == 0) {
896 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
897 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
898 PFID(mdo2fid(mdd_obj)), la->la_valid);
903 handle = mdd_trans_create(env, mdd);
905 RETURN(PTR_ERR(handle));
907 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
911 rc = mdd_trans_start(env, mdd, handle);
915 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
918 if (la->la_valid & (LA_MTIME | LA_CTIME))
919 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
920 la->la_mtime, la->la_ctime);
922 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
923 if (la_copy->la_valid & LA_FLAGS)
924 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
925 else if (la_copy->la_valid) /* setattr */
926 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
927 mdd_write_unlock(env, mdd_obj);
930 rc = mdd_attr_set_changelog(env, obj, handle,
936 rc = mdd_trans_stop(env, mdd, rc, handle);
941 static int mdd_xattr_sanity_check(const struct lu_env *env,
942 struct mdd_object *obj,
943 const struct lu_attr *attr,
946 struct lu_ucred *uc = lu_ucred_assert(env);
949 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
952 if (strncmp(XATTR_USER_PREFIX, name,
953 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
954 /* For sticky directories, only the owner and privileged user
955 * can write attributes. */
956 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
957 (uc->uc_fsuid != attr->la_uid) &&
958 !md_capable(uc, CFS_CAP_FOWNER))
961 if ((uc->uc_fsuid != attr->la_uid) &&
962 !md_capable(uc, CFS_CAP_FOWNER))
970 * Check if a string begins with a given prefix.
972 * \param str String to check
973 * \param prefix Substring to check at the beginning of \a str
974 * \return true/false whether the condition is verified.
976 static inline bool has_prefix(const char *str, const char *prefix)
978 return strncmp(prefix, str, strlen(prefix)) == 0;
982 * Indicate the kind of changelog to store (if any) for a xattr set/del.
984 * \param[in] xattr_name Full extended attribute name.
986 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
988 static enum changelog_rec_type
989 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
990 const char *xattr_name)
992 /* Layout changes systematically recorded */
993 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
994 strncmp(XATTR_LUSTRE_LOV, xattr_name,
995 strlen(XATTR_LUSTRE_LOV)) == 0)
998 /* HSM information changes systematically recorded */
999 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1002 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1003 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1004 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
1010 static int mdd_declare_xattr_set(const struct lu_env *env,
1011 struct mdd_device *mdd,
1012 struct mdd_object *obj,
1013 const struct lu_buf *buf,
1015 int fl, struct thandle *handle)
1019 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1023 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1024 return 0; /* no changelog to store */
1026 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1030 * Compare current and future data of HSM EA and add a changelog if needed.
1032 * Caller should have write-locked \param obj.
1034 * \param buf - Future HSM EA content.
1035 * \retval 0 if no changelog is needed or changelog was added properly.
1036 * \retval -ve errno if there was a problem
1038 static int mdd_hsm_update_locked(const struct lu_env *env,
1039 struct md_object *obj,
1040 const struct lu_buf *buf,
1041 struct thandle *handle, int *cl_flags)
1043 struct mdd_thread_info *info = mdd_env_info(env);
1044 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1045 struct lu_buf *current_buf;
1046 struct md_hsm *current_mh;
1047 struct md_hsm *new_mh;
1051 OBD_ALLOC_PTR(current_mh);
1052 if (current_mh == NULL)
1055 /* Read HSM attrs from disk */
1056 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1057 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1058 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1059 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1060 if (rc < 0 && rc != -ENODATA)
1062 else if (rc == -ENODATA)
1063 current_mh->mh_flags = 0;
1065 /* Map future HSM xattr */
1066 OBD_ALLOC_PTR(new_mh);
1068 GOTO(free, rc = -ENOMEM);
1069 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1073 /* Flags differ, set flags for the changelog that will be added */
1074 if (current_mh->mh_flags != new_mh->mh_flags) {
1075 hsm_set_cl_event(cl_flags, HE_STATE);
1076 if (new_mh->mh_flags & HS_DIRTY)
1077 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1080 OBD_FREE_PTR(new_mh);
1083 OBD_FREE_PTR(current_mh);
1087 static int mdd_object_pfid_replace(const struct lu_env *env,
1088 struct mdd_object *o)
1090 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1091 struct thandle *handle;
1094 handle = mdd_trans_create(env, mdd);
1096 RETURN(PTR_ERR(handle));
1098 handle->th_complex = 1;
1100 /* it doesn't need to track the PFID update via llog, because LFSCK
1101 * will repair it even it goes wrong */
1102 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1107 rc = mdd_trans_start(env, mdd, handle);
1111 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1116 mdd_trans_stop(env, mdd, rc, handle);
1121 static int mdd_declare_xattr_del(const struct lu_env *env,
1122 struct mdd_device *mdd,
1123 struct mdd_object *obj,
1125 struct thandle *handle);
1127 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1130 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1131 struct md_object *md_vic)
1133 struct mdd_device *mdd = mdo2mdd(md_obj);
1134 struct mdd_object *obj = md2mdd_obj(md_obj);
1135 struct mdd_object *vic = md2mdd_obj(md_vic);
1136 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1137 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1138 struct lov_mds_md *lmm;
1139 struct thandle *handle;
1143 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1144 if (rc == 0) /* same fid */
1147 handle = mdd_trans_create(env, mdd);
1149 RETURN(PTR_ERR(handle));
1152 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1153 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1155 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1156 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1159 /* get EA of victim file */
1160 memset(buf_vic, 0, sizeof(*buf_vic));
1161 rc = mdd_get_lov_ea(env, vic, buf_vic);
1168 /* parse the layout of victim file */
1169 lmm = buf_vic->lb_buf;
1170 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1171 GOTO(out, rc = -EINVAL);
1173 /* save EA of target file for restore */
1174 memset(buf, 0, sizeof(*buf));
1175 rc = mdd_get_lov_ea(env, obj, buf);
1179 /* Get rid of the layout from victim object */
1180 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1184 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1185 LU_XATTR_MERGE, handle);
1189 rc = mdd_trans_start(env, mdd, handle);
1193 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1198 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1200 GOTO(out_restore, rc);
1202 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1203 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1208 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1209 LU_XATTR_REPLACE, handle);
1211 CERROR("%s: failed to rollback of layout of: "DFID
1212 ": %d, file state unknown\n",
1213 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1217 mdd_trans_stop(env, mdd, rc, handle);
1218 mdd_write_unlock(env, obj);
1219 mdd_write_unlock(env, vic);
1221 lu_buf_free(buf_vic);
1224 (void) mdd_object_pfid_replace(env, obj);
1229 static int mdd_layout_merge_allowed(const struct lu_env *env,
1230 struct md_object *target,
1231 struct md_object *victim)
1233 struct mdd_object *o1 = md2mdd_obj(target);
1235 /* cannot extend directory's LOVEA */
1236 if (S_ISDIR(mdd_object_type(o1))) {
1237 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1238 mdd_obj_dev_name(o1));
1246 * The caller should guarantee to update the object ctime
1247 * after xattr_set if needed.
1249 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1250 const struct lu_buf *buf, const char *name,
1253 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1254 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1255 struct mdd_device *mdd = mdo2mdd(obj);
1256 struct thandle *handle;
1257 enum changelog_rec_type cl_type;
1262 rc = mdd_la_get(env, mdd_obj, attr);
1266 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1270 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 && fl == LU_XATTR_MERGE) {
1271 struct md_object *victim = buf->lb_buf;
1273 if (buf->lb_len != sizeof(victim))
1276 rc = mdd_layout_merge_allowed(env, obj, victim);
1280 /* merge layout of victim as a mirror of obj's. */
1281 rc = mdd_xattr_merge(env, obj, victim);
1285 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1286 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1287 struct posix_acl *acl;
1289 /* user may set empty ACL, which should be treated as removing
1291 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1294 RETURN(PTR_ERR(acl));
1296 rc = mdd_xattr_del(env, obj, name);
1299 posix_acl_release(acl);
1302 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1303 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1307 handle = mdd_trans_create(env, mdd);
1309 RETURN(PTR_ERR(handle));
1311 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1315 rc = mdd_trans_start(env, mdd, handle);
1319 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1321 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1322 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1324 mdd_write_unlock(env, mdd_obj);
1329 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1330 mdd_write_unlock(env, mdd_obj);
1334 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1338 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1343 return mdd_trans_stop(env, mdd, rc, handle);
1346 static int mdd_declare_xattr_del(const struct lu_env *env,
1347 struct mdd_device *mdd,
1348 struct mdd_object *obj,
1350 struct thandle *handle)
1354 rc = mdo_declare_xattr_del(env, obj, name, handle);
1358 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1359 return 0; /* no changelog to store */
1361 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1365 * The caller should guarantee to update the object ctime
1366 * after xattr_set if needed.
1368 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1371 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1372 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1373 struct mdd_device *mdd = mdo2mdd(obj);
1374 struct thandle *handle;
1378 rc = mdd_la_get(env, mdd_obj, attr);
1382 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1386 handle = mdd_trans_create(env, mdd);
1388 RETURN(PTR_ERR(handle));
1390 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1394 rc = mdd_trans_start(env, mdd, handle);
1398 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1399 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1400 mdd_write_unlock(env, mdd_obj);
1404 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1407 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1411 return mdd_trans_stop(env, mdd, rc, handle);
1415 * read lov EA of an object
1416 * return the lov EA in an allocated lu_buf
1418 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1419 struct lu_buf *lmm_buf)
1421 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1426 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1428 if (rc == -ERANGE) {
1429 /* mti_big_buf is allocated but is too small
1430 * we need to increase it */
1431 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1433 if (buf->lb_buf == NULL)
1434 GOTO(out, rc = -ENOMEM);
1445 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1446 /* mti_big_buf was not allocated, so we have to
1447 * allocate it based on the ea size */
1448 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1450 if (buf->lb_buf == NULL)
1451 GOTO(out, rc = -ENOMEM);
1455 lu_buf_alloc(lmm_buf, bufsize);
1456 if (lmm_buf->lb_buf == NULL)
1457 GOTO(out, rc = -ENOMEM);
1459 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1465 lu_buf_free(lmm_buf);
1469 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1470 struct mdd_object *o, struct lu_buf *buf,
1471 struct thandle *handle)
1473 struct hsm_attrs *attrs;
1479 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1484 attrs = buf->lb_buf;
1485 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1486 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1489 /* Add a changelog record for release. */
1490 hsm_set_cl_event(&flags, HE_RELEASE);
1491 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1497 * check if layout swapping between 2 objects is allowed
1499 * - only normal FIDs or non-system IGIFs
1500 * - same type of objects
1501 * - same owner/group (so quotas are still valid) unless this is from HSM
1504 static int mdd_layout_swap_allowed(const struct lu_env *env,
1505 struct mdd_object *o1,
1506 const struct lu_attr *attr1,
1507 struct mdd_object *o2,
1508 const struct lu_attr *attr2,
1511 const struct lu_fid *fid1, *fid2;
1517 if (!fid_is_norm(fid1) &&
1518 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1521 if (!fid_is_norm(fid2) &&
1522 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1525 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1526 if (S_ISDIR(mdd_object_type(o1)))
1528 if (S_ISREG(mdd_object_type(o1)))
1533 if (flags & SWAP_LAYOUTS_MDS_HSM)
1536 if ((attr1->la_uid != attr2->la_uid) ||
1537 (attr1->la_gid != attr2->la_gid))
1543 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1544 * look into the layout in MDD layer. */
1545 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1547 struct lov_comp_md_v1 *comp_v1;
1548 struct lov_mds_md *v1;
1552 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1553 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1554 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1560 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1561 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1564 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1565 off = le32_to_cpu(comp_v1->lcm_entries[i].
1567 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1572 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1573 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1584 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1586 return mdd_lmm_oi(lmm, oi, true);
1589 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1591 return mdd_lmm_oi(lmm, oi, false);
1594 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1596 struct lov_comp_md_v1 *comp_v1;
1598 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1599 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1601 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1603 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1604 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1605 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1606 __u16 tmp_gen = *gen;
1608 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1610 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1617 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1619 return mdd_lmm_gen(lmm, gen, true);
1622 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1624 return mdd_lmm_gen(lmm, gen, false);
1628 * swap layouts between 2 lustre objects
1630 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1631 struct md_object *obj2, __u64 flags)
1633 struct mdd_thread_info *info = mdd_env_info(env);
1634 struct mdd_object *fst_o = md2mdd_obj(obj1);
1635 struct mdd_object *snd_o = md2mdd_obj(obj2);
1636 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1637 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1638 struct mdd_device *mdd = mdo2mdd(obj1);
1639 struct lov_mds_md *fst_lmm, *snd_lmm;
1640 struct lu_buf *fst_buf = &info->mti_buf[0];
1641 struct lu_buf *snd_buf = &info->mti_buf[1];
1642 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1643 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1644 struct ost_id *saved_oi = NULL;
1645 struct thandle *handle;
1646 __u32 fst_gen, snd_gen, saved_gen;
1652 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1653 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1655 /* we have to sort the 2 obj, so locking will always
1656 * be in the same order, even in case of 2 concurrent swaps */
1657 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1658 if (rc == 0) /* same fid ? */
1664 rc = mdd_la_get(env, fst_o, fst_la);
1668 rc = mdd_la_get(env, snd_o, snd_la);
1672 /* check if layout swapping is allowed */
1673 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1677 handle = mdd_trans_create(env, mdd);
1679 RETURN(PTR_ERR(handle));
1681 /* objects are already sorted */
1682 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1683 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1685 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1686 if (rc < 0 && rc != -ENODATA)
1689 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1690 if (rc < 0 && rc != -ENODATA)
1693 /* swapping 2 non existant layouts is a success */
1694 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1697 /* to help inode migration between MDT, it is better to
1698 * start by the no layout file (if one), so we order the swap */
1699 if (snd_buf->lb_buf == NULL) {
1701 swap(fst_buf, snd_buf);
1704 fst_gen = snd_gen = 0;
1705 /* lmm and generation layout initialization */
1706 if (fst_buf->lb_buf != NULL) {
1707 fst_lmm = fst_buf->lb_buf;
1708 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1709 fst_fl = LU_XATTR_REPLACE;
1713 fst_fl = LU_XATTR_CREATE;
1716 snd_lmm = snd_buf->lb_buf;
1717 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1719 saved_gen = fst_gen;
1720 /* increase the generation layout numbers */
1725 * XXX The layout generation is used to generate component IDs for
1726 * the composite file, we have to do some special tweaks to make
1727 * sure the layout generation is always adequate for that job.
1730 /* Skip invalid generation number for composite layout */
1731 if ((snd_gen & LCME_ID_MASK) == 0)
1733 if ((fst_gen & LCME_ID_MASK) == 0)
1735 /* Make sure the generation is greater than all the component IDs */
1736 if (fst_gen < snd_gen)
1738 else if (fst_gen > snd_gen)
1741 /* set the file specific informations in lmm */
1742 if (fst_lmm != NULL) {
1743 saved_oi = &info->mti_oa.o_oi;
1744 mdd_get_lmm_oi(fst_lmm, saved_oi);
1745 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1746 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1747 mdd_set_lmm_oi(snd_lmm, saved_oi);
1749 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1750 cpu_to_le32(LOV_MAGIC_MAGIC))
1751 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1753 GOTO(stop, rc = -EPROTO);
1755 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1757 /* Prepare HSM attribute if it's required */
1758 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1759 const int buflen = sizeof(struct hsm_attrs);
1761 lu_buf_alloc(fst_hsm_buf, buflen);
1762 lu_buf_alloc(snd_hsm_buf, buflen);
1763 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1764 GOTO(stop, rc = -ENOMEM);
1766 /* Read HSM attribute */
1767 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1771 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1775 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1776 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1781 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1782 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1788 /* prepare transaction */
1789 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1794 if (fst_buf->lb_buf != NULL)
1795 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1796 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1799 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1804 rc = mdd_trans_start(env, mdd, handle);
1808 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1809 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1813 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1815 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1818 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1819 mdd_obj_dev_name(fst_o),
1820 PFID(mdo2fid(fst_o)), rc, rc2);
1825 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1829 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1832 if (fst_buf->lb_buf != NULL)
1833 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1834 LU_XATTR_REPLACE, handle);
1836 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1839 GOTO(out_restore, rc);
1841 /* Issue one changelog record per file */
1842 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1846 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1855 /* failure on second file, but first was done, so we have
1856 * to roll back first. */
1857 if (fst_buf->lb_buf != NULL) {
1858 mdd_set_lmm_oi(fst_lmm, saved_oi);
1859 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1860 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1861 LU_XATTR_REPLACE, handle);
1863 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1869 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1874 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1879 CERROR("%s: unable to roll back layout swap. FIDs: "
1880 DFID" and "DFID "error: %d/%d, steps: %d\n",
1881 mdd_obj_dev_name(fst_o),
1882 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1884 /* a solution to avoid journal commit is to panic,
1885 * but it has strong consequences so we use LBUG to
1886 * allow sysdamin to choose to panic or not
1893 rc = mdd_trans_stop(env, mdd, rc, handle);
1895 mdd_write_unlock(env, snd_o);
1896 mdd_write_unlock(env, fst_o);
1898 lu_buf_free(fst_buf);
1899 lu_buf_free(snd_buf);
1900 lu_buf_free(fst_hsm_buf);
1901 lu_buf_free(snd_hsm_buf);
1904 (void) mdd_object_pfid_replace(env, fst_o);
1905 (void) mdd_object_pfid_replace(env, snd_o);
1910 static int mdd_declare_layout_change(const struct lu_env *env,
1911 struct mdd_device *mdd,
1912 struct mdd_object *obj,
1913 struct md_layout_change *mlc,
1914 struct thandle *handle)
1918 rc = mdo_declare_layout_change(env, obj, mlc, handle);
1922 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1925 /* For PFL, this is used to instantiate necessary component objects. */
1927 mdd_layout_instantiate_component(const struct lu_env *env,
1928 struct mdd_object *obj, struct md_layout_change *mlc,
1929 struct thandle *handle)
1931 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1935 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
1938 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1940 * It's possible that another layout write intent has already
1941 * instantiated our objects, so a -EALREADY returned, and we need to
1945 RETURN(rc == -EALREADY ? 0 : rc);
1947 rc = mdd_trans_start(env, mdd, handle);
1951 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1952 rc = mdo_layout_change(env, obj, mlc, handle);
1953 mdd_write_unlock(env, obj);
1957 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1962 * Change the FLR layout from RDONLY to WRITE_PENDING.
1964 * It picks the primary mirror, and bumps the layout version, and set
1965 * layout version xattr to OST objects in a sync tx. In order to facilitate
1966 * the handling of phantom writers from evicted clients, the clients carry
1967 * layout version of the file with write RPC, so that the OSTs can verify
1968 * if the write RPCs are legitimate, meaning not from evicted clients.
1971 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
1972 struct md_layout_change *mlc, struct thandle *handle)
1974 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1978 /* Verify acceptable operations */
1979 switch (mlc->mlc_opc) {
1980 case MD_LAYOUT_WRITE:
1982 case MD_LAYOUT_RESYNC:
1983 /* these are legal operations - this represents the case that
1984 * a few mirrors were missed in the last resync.
1985 * XXX: it will be supported later */
1986 case MD_LAYOUT_RESYNC_DONE:
1991 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1995 rc = mdd_declare_xattr_del(env, mdd, obj, XATTR_NAME_SOM, handle);
1999 /* record a changelog for data mover to consume */
2000 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2004 rc = mdd_trans_start(env, mdd, handle);
2008 /* it needs a sync tx to make FLR to work properly */
2009 handle->th_sync = 1;
2011 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2012 rc = mdo_layout_change(env, obj, mlc, handle);
2014 rc = mdo_xattr_del(env, obj, XATTR_NAME_SOM, handle);
2018 mdd_write_unlock(env, obj);
2022 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2033 * Handle mirrored file state transition when it's in WRITE_PENDING.
2035 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2036 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2037 * layout version will be increased, and the file's state will be changed to
2041 mdd_layout_update_write_pending(const struct lu_env *env,
2042 struct mdd_object *obj, struct md_layout_change *mlc,
2043 struct thandle *handle)
2045 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2049 switch (mlc->mlc_opc) {
2050 case MD_LAYOUT_RESYNC:
2051 /* Upon receiving the resync request, it should
2052 * instantiate all stale components right away to get ready
2053 * for mirror copy. In order to avoid layout version change,
2054 * client should avoid sending LAYOUT_WRITE request at the
2057 case MD_LAYOUT_WRITE:
2058 /* legal race for concurrent write, the file state has been
2059 * changed by another client. */
2065 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2069 rc = mdd_trans_start(env, mdd, handle);
2073 /* it needs a sync tx to make FLR to work properly */
2074 handle->th_sync = 1;
2076 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2077 rc = mdo_layout_change(env, obj, mlc, handle);
2078 mdd_write_unlock(env, obj);
2089 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2091 * Only concurrent write and sync complete requests are possible when the
2092 * file is in SYNC_PENDING. For the latter request, it will pass in the
2093 * mirrors that have been synchronized, then the stale bit will be cleared
2094 * to make the file's state turn into RDONLY.
2095 * For concurrent write reqeust, it just needs to change the file's state
2096 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2097 * version because the version will be increased in the transition to
2098 * SYNC_PENDING later so that it can deny the write request from potential
2099 * evicted SYNC clients. */
2101 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2102 struct md_layout_change *mlc, struct thandle *handle)
2104 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2105 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2110 /* operation validation */
2111 switch (mlc->mlc_opc) {
2112 case MD_LAYOUT_RESYNC_DONE:
2113 /* resync complete. */
2114 case MD_LAYOUT_WRITE:
2115 /* concurrent write. */
2117 case MD_LAYOUT_RESYNC:
2118 /* resync again, most likely the previous run failed.
2119 * no-op if it's already in SYNC_PENDING state */
2125 if (mlc->mlc_som.lsa_valid & LSOM_FL_VALID) {
2126 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2127 if (rc && rc != -ENODATA)
2130 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2131 som_buf->lb_buf = &mlc->mlc_som;
2132 som_buf->lb_len = sizeof(mlc->mlc_som);
2135 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2139 /* record a changelog for the completion of resync */
2140 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2144 /* RESYNC_DONE has piggybacked size and blocks */
2146 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2147 XATTR_NAME_SOM, fl, handle);
2152 rc = mdd_trans_start(env, mdd, handle);
2156 /* it needs a sync tx to make FLR to work properly */
2157 handle->th_sync = 1;
2159 rc = mdo_layout_change(env, obj, mlc, handle);
2164 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2170 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2179 * Layout change callback for object.
2181 * This is only used by FLR for now. In the future, it can be exteneded to
2182 * handle all layout change.
2185 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2186 struct md_layout_change *mlc)
2188 struct mdd_object *obj = md2mdd_obj(o);
2189 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2190 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2191 struct lov_comp_md_v1 *lcm;
2192 struct thandle *handle;
2197 /* Verify acceptable operations */
2198 switch (mlc->mlc_opc) {
2199 case MD_LAYOUT_WRITE:
2200 case MD_LAYOUT_RESYNC:
2201 case MD_LAYOUT_RESYNC_DONE:
2207 handle = mdd_trans_create(env, mdd);
2209 RETURN(PTR_ERR(handle));
2211 rc = mdd_get_lov_ea(env, obj, buf);
2218 /* analyze the layout to make sure it's a FLR file */
2220 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2221 GOTO(out, rc = -EINVAL);
2223 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2225 /* please refer to HLD of FLR for state transition */
2226 switch (flr_state) {
2227 case LCM_FL_NOT_FLR:
2228 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2230 case LCM_FL_WRITE_PENDING:
2231 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2234 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2236 case LCM_FL_SYNC_PENDING:
2237 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2246 mdd_trans_stop(env, mdd, rc, handle);
2251 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2252 struct mdd_object *child, const struct lu_attr *attr,
2253 const struct md_op_spec *spec,
2254 struct dt_allocation_hint *hint)
2256 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2257 struct dt_object *nc = mdd_object_child(child);
2259 memset(hint, 0, sizeof(*hint));
2261 /* For striped directory, give striping EA to lod_ah_init, which will
2262 * decide the stripe_offset and stripe count by it. */
2263 if (S_ISDIR(attr->la_mode) &&
2264 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2265 hint->dah_eadata = spec->u.sp_ea.eadata;
2266 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2268 hint->dah_eadata = NULL;
2269 hint->dah_eadata_len = 0;
2272 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2273 hint->dah_eadata, hint->dah_eadata_len);
2274 /* @hint will be initialized by underlying device. */
2275 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2279 * do NOT or the MAY_*'s, you'll get the weakest
2281 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
2285 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2286 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2287 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2288 * owner can write to a file even if it is marked readonly to hide
2289 * its brokenness. (bug 5781) */
2290 if (flags & MDS_OPEN_OWNEROVERRIDE) {
2291 struct lu_ucred *uc = lu_ucred_check(env);
2293 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2297 if (flags & FMODE_READ)
2299 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
2301 if (flags & MDS_FMODE_EXEC)
2306 static int mdd_open_sanity_check(const struct lu_env *env,
2307 struct mdd_object *obj,
2308 const struct lu_attr *attr, int flag)
2314 if (mdd_is_dead_obj(obj))
2317 if (S_ISLNK(attr->la_mode))
2320 mode = accmode(env, attr, flag);
2322 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
2325 if (!(flag & MDS_OPEN_CREATED)) {
2326 rc = mdd_permission_internal(env, obj, attr, mode);
2331 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
2332 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
2333 flag &= ~MDS_OPEN_TRUNC;
2335 /* For writing append-only file must open it with append mode. */
2336 if (attr->la_flags & LUSTRE_APPEND_FL) {
2337 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
2339 if (flag & MDS_OPEN_TRUNC)
2346 static int mdd_open(const struct lu_env *env, struct md_object *obj,
2349 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2350 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2353 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2355 rc = mdd_la_get(env, mdd_obj, attr);
2359 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
2363 mdd_obj->mod_count++;
2366 mdd_write_unlock(env, mdd_obj);
2370 static int mdd_declare_close(const struct lu_env *env,
2371 struct mdd_object *obj,
2373 struct thandle *handle)
2377 rc = orph_declare_index_delete(env, obj, handle);
2381 return mdo_declare_destroy(env, obj, handle);
2385 * No permission check is needed.
2387 static int mdd_close(const struct lu_env *env, struct md_object *obj,
2388 struct md_attr *ma, int mode)
2390 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2391 struct mdd_device *mdd = mdo2mdd(obj);
2392 struct thandle *handle = NULL;
2395 bool blocked = false;
2398 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
2399 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2400 mdd_obj->mod_count--;
2401 mdd_write_unlock(env, mdd_obj);
2403 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
2404 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
2405 "list\n", PFID(mdd_object_fid(mdd_obj)));
2409 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
2410 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
2411 /* check without any lock */
2412 is_orphan = mdd_obj->mod_count == 1 &&
2413 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
2417 /* mdd_trans_create() maybe failed because of barrier_entry(),
2418 * under such case, the orphan MDT-object will be left in the
2419 * orphan list, and when the MDT remount next time, the unused
2420 * orphans will be destroyed automatically.
2422 * One exception: the former mdd_finish_unlink may failed to
2423 * add the orphan MDT-object to the orphan list, then if the
2424 * mdd_trans_create() failed because of barrier_entry(), the
2425 * MDT-object will become real orphan that is neither in the
2426 * namespace nor in the orphan list. Such bad case should be
2427 * very rare and will be handled by e2fsck/lfsck. */
2428 handle = mdd_trans_create(env, mdo2mdd(obj));
2429 if (IS_ERR(handle)) {
2430 rc = PTR_ERR(handle);
2431 if (rc != -EINPROGRESS)
2439 rc = mdd_declare_close(env, mdd_obj, ma, handle);
2443 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2447 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2453 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2454 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
2456 CERROR("Failed to get lu_attr of "DFID": %d\n",
2457 PFID(mdd_object_fid(mdd_obj)), rc);
2461 /* check again with lock */
2462 is_orphan = (mdd_obj->mod_count == 1) &&
2463 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
2464 ma->ma_attr.la_nlink == 0);
2466 if (is_orphan && !handle && !blocked) {
2467 mdd_write_unlock(env, mdd_obj);
2471 mdd_obj->mod_count--; /*release open count */
2473 if (!is_orphan || blocked)
2477 /* NB: Object maybe not in orphan list originally, it is rare case for
2478 * mdd_finish_unlink() failure, in that case, the object doesn't have
2479 * ORPHAN_OBJ flag */
2480 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2481 /* remove link to object from orphan index */
2482 LASSERT(handle != NULL);
2483 rc = __mdd_orphan_del(env, mdd_obj, handle);
2485 CERROR("%s: unable to delete "DFID" from orphan list: "
2486 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2487 PFID(mdd_object_fid(mdd_obj)), rc);
2488 /* If object was not deleted from orphan list, do not
2489 * destroy OSS objects, which will be done when next
2494 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2495 "list, OSS objects to be destroyed.\n",
2496 PFID(mdd_object_fid(mdd_obj)));
2499 rc = mdo_destroy(env, mdd_obj, handle);
2502 CERROR("%s: unable to delete "DFID" from orphan list: "
2503 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2504 PFID(mdd_object_fid(mdd_obj)), rc);
2509 mdd_write_unlock(env, mdd_obj);
2511 if (!rc && !blocked &&
2512 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2513 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2514 if (handle == NULL) {
2515 handle = mdd_trans_create(env, mdo2mdd(obj));
2517 GOTO(stop, rc = PTR_ERR(handle));
2519 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2524 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2529 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2534 if (handle != NULL && !IS_ERR(handle))
2535 rc = mdd_trans_stop(env, mdd, rc, handle);
2541 * Permission check is done when open,
2542 * no need check again.
2544 static int mdd_readpage_sanity_check(const struct lu_env *env,
2545 struct mdd_object *obj)
2547 struct dt_object *next = mdd_object_child(obj);
2551 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2559 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2560 size_t nob, const struct dt_it_ops *iops,
2561 struct dt_it *it, __u32 attr, void *arg)
2563 struct lu_dirpage *dp = &lp->lp_dir;
2567 struct lu_dirent *ent;
2568 struct lu_dirent *last = NULL;
2572 if (nob < sizeof(*dp))
2575 memset(area, 0, sizeof (*dp));
2576 area += sizeof (*dp);
2577 nob -= sizeof (*dp);
2584 len = iops->key_size(env, it);
2586 /* IAM iterator can return record with zero len. */
2590 hash = iops->store(env, it);
2591 if (unlikely(first)) {
2593 dp->ldp_hash_start = cpu_to_le64(hash);
2596 /* calculate max space required for lu_dirent */
2597 recsize = lu_dirent_calc_size(len, attr);
2599 if (nob >= recsize) {
2600 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2601 if (result == -ESTALE)
2606 /* osd might not able to pack all attributes,
2607 * so recheck rec length */
2608 recsize = le16_to_cpu(ent->lde_reclen);
2610 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2611 fid_le_to_cpu(&fid, &ent->lde_fid);
2612 if (fid_is_dot_lustre(&fid))
2616 result = (last != NULL) ? 0 :-EINVAL;
2620 ent = (void *)ent + recsize;
2624 result = iops->next(env, it);
2625 if (result == -ESTALE)
2627 } while (result == 0);
2630 dp->ldp_hash_end = cpu_to_le64(hash);
2632 if (last->lde_hash == dp->ldp_hash_end)
2633 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2634 last->lde_reclen = 0; /* end mark */
2637 /* end of directory */
2638 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2639 else if (result < 0)
2640 CWARN("build page failed: %d!\n", result);
2644 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2645 const struct lu_rdpg *rdpg)
2647 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2651 if (mdd_object_exists(mdd_obj) == 0) {
2652 CERROR("%s: object "DFID" not found: rc = -2\n",
2653 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2657 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2658 rc = mdd_readpage_sanity_check(env, mdd_obj);
2660 GOTO(out_unlock, rc);
2662 if (mdd_is_dead_obj(mdd_obj)) {
2664 struct lu_dirpage *dp;
2667 * According to POSIX, please do not return any entry to client:
2668 * even dot and dotdot should not be returned.
2670 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2671 PFID(mdd_object_fid(mdd_obj)));
2673 if (rdpg->rp_count <= 0)
2674 GOTO(out_unlock, rc = -EFAULT);
2675 LASSERT(rdpg->rp_pages != NULL);
2677 pg = rdpg->rp_pages[0];
2678 dp = (struct lu_dirpage *)kmap(pg);
2679 memset(dp, 0 , sizeof(struct lu_dirpage));
2680 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2681 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2682 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2684 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2687 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2688 mdd_dir_page_build, NULL);
2690 struct lu_dirpage *dp;
2692 dp = kmap(rdpg->rp_pages[0]);
2693 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2696 * No pages were processed, mark this for first page
2699 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2700 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2701 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2703 kunmap(rdpg->rp_pages[0]);
2706 GOTO(out_unlock, rc);
2708 mdd_read_unlock(env, mdd_obj);
2712 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2714 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2716 if (mdd_object_exists(mdd_obj) == 0) {
2719 CERROR("%s: object "DFID" not found: rc = %d\n",
2720 mdd_obj_dev_name(mdd_obj),
2721 PFID(mdd_object_fid(mdd_obj)), rc);
2724 return dt_object_sync(env, mdd_object_child(mdd_obj),
2728 static int mdd_object_lock(const struct lu_env *env,
2729 struct md_object *obj,
2730 struct lustre_handle *lh,
2731 struct ldlm_enqueue_info *einfo,
2732 union ldlm_policy_data *policy)
2734 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2735 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2739 static int mdd_object_unlock(const struct lu_env *env,
2740 struct md_object *obj,
2741 struct ldlm_enqueue_info *einfo,
2742 union ldlm_policy_data *policy)
2744 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2745 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2748 const struct md_object_operations mdd_obj_ops = {
2749 .moo_permission = mdd_permission,
2750 .moo_attr_get = mdd_attr_get,
2751 .moo_attr_set = mdd_attr_set,
2752 .moo_xattr_get = mdd_xattr_get,
2753 .moo_xattr_set = mdd_xattr_set,
2754 .moo_xattr_list = mdd_xattr_list,
2755 .moo_invalidate = mdd_invalidate,
2756 .moo_xattr_del = mdd_xattr_del,
2757 .moo_swap_layouts = mdd_swap_layouts,
2758 .moo_open = mdd_open,
2759 .moo_close = mdd_close,
2760 .moo_readpage = mdd_readpage,
2761 .moo_readlink = mdd_readlink,
2762 .moo_changelog = mdd_changelog,
2763 .moo_object_sync = mdd_object_sync,
2764 .moo_object_lock = mdd_object_lock,
2765 .moo_object_unlock = mdd_object_unlock,
2766 .moo_layout_change = mdd_layout_change,