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, 2017, 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;
652 if (uc->uc_jobid[0] != '\0')
654 xflags |= CLFE_UIDGID;
660 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
661 changelog_rec_offset(flags & CLF_SUPPORTED,
662 xflags & CLFE_SUPPORTED));
663 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
664 if (buf->lb_buf == NULL)
668 rec->cr_hdr.lrh_len = reclen;
669 rec->cr.cr_flags = flags;
670 rec->cr.cr_type = (__u32)type;
671 rec->cr.cr_tfid = *fid;
672 rec->cr.cr_namelen = 0;
674 if (flags & CLF_JOBID)
675 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
677 if (flags & CLF_EXTRA_FLAGS) {
678 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
679 if (xflags & CLFE_UIDGID)
680 mdd_changelog_rec_extra_uidgid(&rec->cr,
681 uc->uc_uid, uc->uc_gid);
682 if (xflags & CLFE_NID)
683 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
684 if (xflags & CLFE_OPEN)
685 mdd_changelog_rec_extra_omode(&rec->cr, flags);
688 rc = mdd_changelog_store(env, mdd, rec, handle);
693 /** Store a data change changelog record
694 * If this fails, we must fail the whole transaction; we don't
695 * want the change to commit without the log entry.
696 * \param mdd_obj - mdd_object of change
697 * \param handle - transaction handle
699 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
700 enum changelog_rec_type type, int flags,
701 struct mdd_object *mdd_obj, struct thandle *handle)
705 LASSERT(mdd_obj != NULL);
706 LASSERT(handle != NULL);
709 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
711 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
714 if (mdd_is_volatile_obj(mdd_obj))
717 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
718 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
719 /* Don't need multiple updates in this log */
720 /* Don't check under lock - no big deal if we get an extra
725 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
726 mdo2fid(mdd_obj), handle);
728 mdd_obj->mod_cltime = ktime_get();
733 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
734 int flags, struct md_device *m, const struct lu_fid *fid)
736 struct thandle *handle;
737 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
742 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
744 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
747 LASSERT(fid != NULL);
749 handle = mdd_trans_create(env, mdd);
751 RETURN(PTR_ERR(handle));
753 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
757 rc = mdd_trans_start(env, mdd, handle);
761 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
765 rc = mdd_trans_stop(env, mdd, rc, handle);
771 * Save LMA extended attributes with data from \a ma.
773 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
774 * not, LMA EA will be first read from disk, modified and write back.
777 /* Precedence for choosing record type when multiple
778 * attributes change: setattr > mtime > ctime > atime
779 * (ctime changes when mtime does, plus chmod/chown.
780 * atime and ctime are independent.) */
781 static int mdd_attr_set_changelog(const struct lu_env *env,
782 struct md_object *obj, struct thandle *handle,
785 struct mdd_device *mdd = mdo2mdd(obj);
788 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
789 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
790 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
791 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
792 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
793 bits = bits & mdd->mdd_cl.mc_mask;
794 /* This is an implementation limit rather than a protocol limit */
795 CLASSERT(CL_LAST <= sizeof(int) * 8);
799 /* The record type is the lowest non-masked set bit */
802 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
803 return mdd_changelog_data_store(env, mdd, type, (int)valid,
804 md2mdd_obj(obj), handle);
807 static int mdd_declare_attr_set(const struct lu_env *env,
808 struct mdd_device *mdd,
809 struct mdd_object *obj,
810 const struct lu_attr *attr,
811 struct thandle *handle)
815 rc = mdo_declare_attr_set(env, obj, attr, handle);
819 #ifdef CONFIG_FS_POSIX_ACL
820 if (attr->la_valid & LA_MODE) {
821 mdd_read_lock(env, obj, MOR_TGT_CHILD);
822 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
823 XATTR_NAME_ACL_ACCESS);
824 mdd_read_unlock(env, obj);
825 if (rc == -EOPNOTSUPP || rc == -ENODATA)
831 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
832 rc = mdo_declare_xattr_set(env, obj, buf,
833 XATTR_NAME_ACL_ACCESS, 0,
841 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
848 * permission changes may require sync operation, to mitigate performance
849 * impact, only do this for dir and when permission is reduced.
851 * For regular files, version is updated with permission change (see VBR), async
852 * permission won't cause any issue, while missing permission change on
853 * directory may affect accessibility of other objects after recovery.
855 static inline bool permission_needs_sync(const struct lu_attr *old,
856 const struct lu_attr *new)
858 if (!S_ISDIR(old->la_mode))
861 if (new->la_valid & (LA_UID | LA_GID))
864 if (new->la_valid & LA_MODE &&
865 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
868 if ((new->la_valid & LA_MODE) &&
869 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
870 (old->la_mode & S_IRWXUGO))
876 /* set attr and LOV EA at once, return updated attr */
877 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
878 const struct md_attr *ma)
880 struct mdd_object *mdd_obj = md2mdd_obj(obj);
881 struct mdd_device *mdd = mdo2mdd(obj);
882 struct thandle *handle;
883 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
884 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
885 const struct lu_attr *la = &ma->ma_attr;
889 /* we do not use ->attr_set() for LOV/HSM EA any more */
890 LASSERT((ma->ma_valid & MA_LOV) == 0);
891 LASSERT((ma->ma_valid & MA_HSM) == 0);
893 rc = mdd_la_get(env, mdd_obj, attr);
897 *la_copy = ma->ma_attr;
898 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
902 /* no need to setattr anymore */
903 if (la_copy->la_valid == 0) {
904 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
905 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
906 PFID(mdo2fid(mdd_obj)), la->la_valid);
911 handle = mdd_trans_create(env, mdd);
913 RETURN(PTR_ERR(handle));
915 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
919 rc = mdd_trans_start(env, mdd, handle);
923 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
926 if (la->la_valid & (LA_MTIME | LA_CTIME))
927 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
928 la->la_mtime, la->la_ctime);
930 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
931 if (la_copy->la_valid & LA_FLAGS)
932 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
933 else if (la_copy->la_valid) /* setattr */
934 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
935 mdd_write_unlock(env, mdd_obj);
938 rc = mdd_attr_set_changelog(env, obj, handle,
944 rc = mdd_trans_stop(env, mdd, rc, handle);
949 static int mdd_xattr_sanity_check(const struct lu_env *env,
950 struct mdd_object *obj,
951 const struct lu_attr *attr,
954 struct lu_ucred *uc = lu_ucred_assert(env);
957 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
960 if (strncmp(XATTR_USER_PREFIX, name,
961 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
962 /* For sticky directories, only the owner and privileged user
963 * can write attributes. */
964 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
965 (uc->uc_fsuid != attr->la_uid) &&
966 !md_capable(uc, CFS_CAP_FOWNER))
969 if ((uc->uc_fsuid != attr->la_uid) &&
970 !md_capable(uc, CFS_CAP_FOWNER))
978 * Check if a string begins with a given prefix.
980 * \param str String to check
981 * \param prefix Substring to check at the beginning of \a str
982 * \return true/false whether the condition is verified.
984 static inline bool has_prefix(const char *str, const char *prefix)
986 return strncmp(prefix, str, strlen(prefix)) == 0;
990 * Indicate the kind of changelog to store (if any) for a xattr set/del.
992 * \param[in] xattr_name Full extended attribute name.
994 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
996 static enum changelog_rec_type
997 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
998 const char *xattr_name)
1000 /* Layout changes systematically recorded */
1001 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1002 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1003 strlen(XATTR_LUSTRE_LOV)) == 0)
1006 /* HSM information changes systematically recorded */
1007 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1010 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1011 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1012 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
1018 static int mdd_declare_xattr_set(const struct lu_env *env,
1019 struct mdd_device *mdd,
1020 struct mdd_object *obj,
1021 const struct lu_buf *buf,
1023 int fl, struct thandle *handle)
1027 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1031 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1032 return 0; /* no changelog to store */
1034 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1038 * Compare current and future data of HSM EA and add a changelog if needed.
1040 * Caller should have write-locked \param obj.
1042 * \param buf - Future HSM EA content.
1043 * \retval 0 if no changelog is needed or changelog was added properly.
1044 * \retval -ve errno if there was a problem
1046 static int mdd_hsm_update_locked(const struct lu_env *env,
1047 struct md_object *obj,
1048 const struct lu_buf *buf,
1049 struct thandle *handle, int *cl_flags)
1051 struct mdd_thread_info *info = mdd_env_info(env);
1052 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1053 struct lu_buf *current_buf;
1054 struct md_hsm *current_mh;
1055 struct md_hsm *new_mh;
1059 OBD_ALLOC_PTR(current_mh);
1060 if (current_mh == NULL)
1063 /* Read HSM attrs from disk */
1064 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1065 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1066 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1067 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1068 if (rc < 0 && rc != -ENODATA)
1070 else if (rc == -ENODATA)
1071 current_mh->mh_flags = 0;
1073 /* Map future HSM xattr */
1074 OBD_ALLOC_PTR(new_mh);
1076 GOTO(free, rc = -ENOMEM);
1077 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1081 /* Flags differ, set flags for the changelog that will be added */
1082 if (current_mh->mh_flags != new_mh->mh_flags) {
1083 hsm_set_cl_event(cl_flags, HE_STATE);
1084 if (new_mh->mh_flags & HS_DIRTY)
1085 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1088 OBD_FREE_PTR(new_mh);
1091 OBD_FREE_PTR(current_mh);
1095 static int mdd_object_pfid_replace(const struct lu_env *env,
1096 struct mdd_object *o)
1098 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1099 struct thandle *handle;
1102 handle = mdd_trans_create(env, mdd);
1104 RETURN(PTR_ERR(handle));
1106 handle->th_complex = 1;
1108 /* it doesn't need to track the PFID update via llog, because LFSCK
1109 * will repair it even it goes wrong */
1110 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1115 rc = mdd_trans_start(env, mdd, handle);
1119 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1124 mdd_trans_stop(env, mdd, rc, handle);
1129 static int mdd_declare_xattr_del(const struct lu_env *env,
1130 struct mdd_device *mdd,
1131 struct mdd_object *obj,
1133 struct thandle *handle);
1135 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1138 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1139 struct md_object *md_vic)
1141 struct mdd_device *mdd = mdo2mdd(md_obj);
1142 struct mdd_object *obj = md2mdd_obj(md_obj);
1143 struct mdd_object *vic = md2mdd_obj(md_vic);
1144 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1145 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1146 struct lov_mds_md *lmm;
1147 struct thandle *handle;
1151 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1152 if (rc == 0) /* same fid */
1155 handle = mdd_trans_create(env, mdd);
1157 RETURN(PTR_ERR(handle));
1160 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1161 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1163 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1164 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1167 /* get EA of victim file */
1168 memset(buf_vic, 0, sizeof(*buf_vic));
1169 rc = mdd_get_lov_ea(env, vic, buf_vic);
1176 /* parse the layout of victim file */
1177 lmm = buf_vic->lb_buf;
1178 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1179 GOTO(out, rc = -EINVAL);
1181 /* save EA of target file for restore */
1182 memset(buf, 0, sizeof(*buf));
1183 rc = mdd_get_lov_ea(env, obj, buf);
1187 /* Get rid of the layout from victim object */
1188 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1192 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1193 LU_XATTR_MERGE, handle);
1197 rc = mdd_trans_start(env, mdd, handle);
1201 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1206 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1208 GOTO(out_restore, rc);
1210 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1211 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1216 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1217 LU_XATTR_REPLACE, handle);
1219 CERROR("%s: failed to rollback of layout of: "DFID
1220 ": %d, file state unknown\n",
1221 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1225 mdd_trans_stop(env, mdd, rc, handle);
1226 mdd_write_unlock(env, obj);
1227 mdd_write_unlock(env, vic);
1229 lu_buf_free(buf_vic);
1232 (void) mdd_object_pfid_replace(env, obj);
1237 static int mdd_layout_merge_allowed(const struct lu_env *env,
1238 struct md_object *target,
1239 struct md_object *victim)
1241 struct mdd_object *o1 = md2mdd_obj(target);
1243 /* cannot extend directory's LOVEA */
1244 if (S_ISDIR(mdd_object_type(o1))) {
1245 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1246 mdd_obj_dev_name(o1));
1254 * The caller should guarantee to update the object ctime
1255 * after xattr_set if needed.
1257 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1258 const struct lu_buf *buf, const char *name,
1261 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1262 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1263 struct mdd_device *mdd = mdo2mdd(obj);
1264 struct thandle *handle;
1265 enum changelog_rec_type cl_type;
1270 rc = mdd_la_get(env, mdd_obj, attr);
1274 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1278 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 && fl == LU_XATTR_MERGE) {
1279 struct md_object *victim = buf->lb_buf;
1281 if (buf->lb_len != sizeof(victim))
1284 rc = mdd_layout_merge_allowed(env, obj, victim);
1288 /* merge layout of victim as a mirror of obj's. */
1289 rc = mdd_xattr_merge(env, obj, victim);
1293 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1294 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1295 struct posix_acl *acl;
1297 /* user may set empty ACL, which should be treated as removing
1299 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1302 RETURN(PTR_ERR(acl));
1304 rc = mdd_xattr_del(env, obj, name);
1307 posix_acl_release(acl);
1310 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1311 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1315 handle = mdd_trans_create(env, mdd);
1317 RETURN(PTR_ERR(handle));
1319 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1323 rc = mdd_trans_start(env, mdd, handle);
1327 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1329 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1330 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1332 mdd_write_unlock(env, mdd_obj);
1337 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1338 mdd_write_unlock(env, mdd_obj);
1342 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1346 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1351 return mdd_trans_stop(env, mdd, rc, handle);
1354 static int mdd_declare_xattr_del(const struct lu_env *env,
1355 struct mdd_device *mdd,
1356 struct mdd_object *obj,
1358 struct thandle *handle)
1362 rc = mdo_declare_xattr_del(env, obj, name, handle);
1366 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1367 return 0; /* no changelog to store */
1369 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1373 * The caller should guarantee to update the object ctime
1374 * after xattr_set if needed.
1376 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1379 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1380 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1381 struct mdd_device *mdd = mdo2mdd(obj);
1382 struct thandle *handle;
1386 rc = mdd_la_get(env, mdd_obj, attr);
1390 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1394 handle = mdd_trans_create(env, mdd);
1396 RETURN(PTR_ERR(handle));
1398 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1402 rc = mdd_trans_start(env, mdd, handle);
1406 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1407 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1408 mdd_write_unlock(env, mdd_obj);
1412 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1415 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1419 return mdd_trans_stop(env, mdd, rc, handle);
1423 * read lov EA of an object
1424 * return the lov EA in an allocated lu_buf
1426 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1427 struct lu_buf *lmm_buf)
1429 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1434 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1436 if (rc == -ERANGE) {
1437 /* mti_big_buf is allocated but is too small
1438 * we need to increase it */
1439 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1441 if (buf->lb_buf == NULL)
1442 GOTO(out, rc = -ENOMEM);
1453 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1454 /* mti_big_buf was not allocated, so we have to
1455 * allocate it based on the ea size */
1456 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1458 if (buf->lb_buf == NULL)
1459 GOTO(out, rc = -ENOMEM);
1463 lu_buf_alloc(lmm_buf, bufsize);
1464 if (lmm_buf->lb_buf == NULL)
1465 GOTO(out, rc = -ENOMEM);
1467 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1473 lu_buf_free(lmm_buf);
1477 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1478 struct mdd_object *o, struct lu_buf *buf,
1479 struct thandle *handle)
1481 struct hsm_attrs *attrs;
1487 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1492 attrs = buf->lb_buf;
1493 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1494 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1497 /* Add a changelog record for release. */
1498 hsm_set_cl_event(&flags, HE_RELEASE);
1499 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1505 * check if layout swapping between 2 objects is allowed
1507 * - only normal FIDs or non-system IGIFs
1508 * - same type of objects
1509 * - same owner/group (so quotas are still valid) unless this is from HSM
1512 static int mdd_layout_swap_allowed(const struct lu_env *env,
1513 struct mdd_object *o1,
1514 const struct lu_attr *attr1,
1515 struct mdd_object *o2,
1516 const struct lu_attr *attr2,
1519 const struct lu_fid *fid1, *fid2;
1525 if (!fid_is_norm(fid1) &&
1526 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1529 if (!fid_is_norm(fid2) &&
1530 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1533 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1534 if (S_ISDIR(mdd_object_type(o1)))
1536 if (S_ISREG(mdd_object_type(o1)))
1541 if (flags & SWAP_LAYOUTS_MDS_HSM)
1544 if ((attr1->la_uid != attr2->la_uid) ||
1545 (attr1->la_gid != attr2->la_gid))
1551 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1552 * look into the layout in MDD layer. */
1553 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1555 struct lov_comp_md_v1 *comp_v1;
1556 struct lov_mds_md *v1;
1560 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1561 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1562 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1568 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1569 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1572 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1573 off = le32_to_cpu(comp_v1->lcm_entries[i].
1575 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1580 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1581 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1592 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1594 return mdd_lmm_oi(lmm, oi, true);
1597 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1599 return mdd_lmm_oi(lmm, oi, false);
1602 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1604 struct lov_comp_md_v1 *comp_v1;
1606 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1607 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1609 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1611 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1612 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1613 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1614 __u16 tmp_gen = *gen;
1616 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1618 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1625 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1627 return mdd_lmm_gen(lmm, gen, true);
1630 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1632 return mdd_lmm_gen(lmm, gen, false);
1636 * swap layouts between 2 lustre objects
1638 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1639 struct md_object *obj2, __u64 flags)
1641 struct mdd_thread_info *info = mdd_env_info(env);
1642 struct mdd_object *fst_o = md2mdd_obj(obj1);
1643 struct mdd_object *snd_o = md2mdd_obj(obj2);
1644 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1645 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1646 struct mdd_device *mdd = mdo2mdd(obj1);
1647 struct lov_mds_md *fst_lmm, *snd_lmm;
1648 struct lu_buf *fst_buf = &info->mti_buf[0];
1649 struct lu_buf *snd_buf = &info->mti_buf[1];
1650 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1651 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1652 struct ost_id *saved_oi = NULL;
1653 struct thandle *handle;
1654 __u32 fst_gen, snd_gen, saved_gen;
1660 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1661 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1663 /* we have to sort the 2 obj, so locking will always
1664 * be in the same order, even in case of 2 concurrent swaps */
1665 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1666 if (rc == 0) /* same fid ? */
1672 rc = mdd_la_get(env, fst_o, fst_la);
1676 rc = mdd_la_get(env, snd_o, snd_la);
1680 /* check if layout swapping is allowed */
1681 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1685 handle = mdd_trans_create(env, mdd);
1687 RETURN(PTR_ERR(handle));
1689 /* objects are already sorted */
1690 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1691 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1693 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1694 if (rc < 0 && rc != -ENODATA)
1697 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1698 if (rc < 0 && rc != -ENODATA)
1701 /* swapping 2 non existant layouts is a success */
1702 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1705 /* to help inode migration between MDT, it is better to
1706 * start by the no layout file (if one), so we order the swap */
1707 if (snd_buf->lb_buf == NULL) {
1709 swap(fst_buf, snd_buf);
1712 fst_gen = snd_gen = 0;
1713 /* lmm and generation layout initialization */
1714 if (fst_buf->lb_buf != NULL) {
1715 fst_lmm = fst_buf->lb_buf;
1716 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1717 fst_fl = LU_XATTR_REPLACE;
1721 fst_fl = LU_XATTR_CREATE;
1724 snd_lmm = snd_buf->lb_buf;
1725 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1727 saved_gen = fst_gen;
1728 /* increase the generation layout numbers */
1733 * XXX The layout generation is used to generate component IDs for
1734 * the composite file, we have to do some special tweaks to make
1735 * sure the layout generation is always adequate for that job.
1738 /* Skip invalid generation number for composite layout */
1739 if ((snd_gen & LCME_ID_MASK) == 0)
1741 if ((fst_gen & LCME_ID_MASK) == 0)
1743 /* Make sure the generation is greater than all the component IDs */
1744 if (fst_gen < snd_gen)
1746 else if (fst_gen > snd_gen)
1749 /* set the file specific informations in lmm */
1750 if (fst_lmm != NULL) {
1751 saved_oi = &info->mti_oa.o_oi;
1752 mdd_get_lmm_oi(fst_lmm, saved_oi);
1753 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1754 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1755 mdd_set_lmm_oi(snd_lmm, saved_oi);
1757 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1758 cpu_to_le32(LOV_MAGIC_MAGIC))
1759 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1761 GOTO(stop, rc = -EPROTO);
1763 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1765 /* Prepare HSM attribute if it's required */
1766 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1767 const int buflen = sizeof(struct hsm_attrs);
1769 lu_buf_alloc(fst_hsm_buf, buflen);
1770 lu_buf_alloc(snd_hsm_buf, buflen);
1771 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1772 GOTO(stop, rc = -ENOMEM);
1774 /* Read HSM attribute */
1775 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1779 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1783 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1784 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1789 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1790 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1796 /* prepare transaction */
1797 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1802 if (fst_buf->lb_buf != NULL)
1803 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1804 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1807 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1812 rc = mdd_trans_start(env, mdd, handle);
1816 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1817 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1821 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1823 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1826 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1827 mdd_obj_dev_name(fst_o),
1828 PFID(mdo2fid(fst_o)), rc, rc2);
1833 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1837 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1840 if (fst_buf->lb_buf != NULL)
1841 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1842 LU_XATTR_REPLACE, handle);
1844 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1847 GOTO(out_restore, rc);
1849 /* Issue one changelog record per file */
1850 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1854 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1863 /* failure on second file, but first was done, so we have
1864 * to roll back first. */
1865 if (fst_buf->lb_buf != NULL) {
1866 mdd_set_lmm_oi(fst_lmm, saved_oi);
1867 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1868 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1869 LU_XATTR_REPLACE, handle);
1871 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1877 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1882 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1887 CERROR("%s: unable to roll back layout swap. FIDs: "
1888 DFID" and "DFID "error: %d/%d, steps: %d\n",
1889 mdd_obj_dev_name(fst_o),
1890 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1892 /* a solution to avoid journal commit is to panic,
1893 * but it has strong consequences so we use LBUG to
1894 * allow sysdamin to choose to panic or not
1901 rc = mdd_trans_stop(env, mdd, rc, handle);
1903 mdd_write_unlock(env, snd_o);
1904 mdd_write_unlock(env, fst_o);
1906 lu_buf_free(fst_buf);
1907 lu_buf_free(snd_buf);
1908 lu_buf_free(fst_hsm_buf);
1909 lu_buf_free(snd_hsm_buf);
1912 (void) mdd_object_pfid_replace(env, fst_o);
1913 (void) mdd_object_pfid_replace(env, snd_o);
1918 static int mdd_declare_layout_change(const struct lu_env *env,
1919 struct mdd_device *mdd,
1920 struct mdd_object *obj,
1921 struct md_layout_change *mlc,
1922 struct thandle *handle)
1926 rc = mdo_declare_layout_change(env, obj, mlc, handle);
1930 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1933 /* For PFL, this is used to instantiate necessary component objects. */
1935 mdd_layout_instantiate_component(const struct lu_env *env,
1936 struct mdd_object *obj, struct md_layout_change *mlc,
1937 struct thandle *handle)
1939 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1943 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
1946 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1948 * It's possible that another layout write intent has already
1949 * instantiated our objects, so a -EALREADY returned, and we need to
1953 RETURN(rc == -EALREADY ? 0 : rc);
1955 rc = mdd_trans_start(env, mdd, handle);
1959 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1960 rc = mdo_layout_change(env, obj, mlc, handle);
1961 mdd_write_unlock(env, obj);
1965 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1970 * Change the FLR layout from RDONLY to WRITE_PENDING.
1972 * It picks the primary mirror, and bumps the layout version, and set
1973 * layout version xattr to OST objects in a sync tx. In order to facilitate
1974 * the handling of phantom writers from evicted clients, the clients carry
1975 * layout version of the file with write RPC, so that the OSTs can verify
1976 * if the write RPCs are legitimate, meaning not from evicted clients.
1979 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
1980 struct md_layout_change *mlc, struct thandle *handle)
1982 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1986 /* Verify acceptable operations */
1987 switch (mlc->mlc_opc) {
1988 case MD_LAYOUT_WRITE:
1990 case MD_LAYOUT_RESYNC:
1991 /* these are legal operations - this represents the case that
1992 * a few mirrors were missed in the last resync.
1993 * XXX: it will be supported later */
1994 case MD_LAYOUT_RESYNC_DONE:
1999 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2003 rc = mdd_declare_xattr_del(env, mdd, obj, XATTR_NAME_SOM, handle);
2007 /* record a changelog for data mover to consume */
2008 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2012 rc = mdd_trans_start(env, mdd, handle);
2016 /* it needs a sync tx to make FLR to work properly */
2017 handle->th_sync = 1;
2019 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2020 rc = mdo_layout_change(env, obj, mlc, handle);
2022 rc = mdo_xattr_del(env, obj, XATTR_NAME_SOM, handle);
2026 mdd_write_unlock(env, obj);
2030 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2041 * Handle mirrored file state transition when it's in WRITE_PENDING.
2043 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2044 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2045 * layout version will be increased, and the file's state will be changed to
2049 mdd_layout_update_write_pending(const struct lu_env *env,
2050 struct mdd_object *obj, struct md_layout_change *mlc,
2051 struct thandle *handle)
2053 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2057 switch (mlc->mlc_opc) {
2058 case MD_LAYOUT_RESYNC:
2059 /* Upon receiving the resync request, it should
2060 * instantiate all stale components right away to get ready
2061 * for mirror copy. In order to avoid layout version change,
2062 * client should avoid sending LAYOUT_WRITE request at the
2065 case MD_LAYOUT_WRITE:
2066 /* legal race for concurrent write, the file state has been
2067 * changed by another client. */
2073 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2077 rc = mdd_trans_start(env, mdd, handle);
2081 /* it needs a sync tx to make FLR to work properly */
2082 handle->th_sync = 1;
2084 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2085 rc = mdo_layout_change(env, obj, mlc, handle);
2086 mdd_write_unlock(env, obj);
2097 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2099 * Only concurrent write and sync complete requests are possible when the
2100 * file is in SYNC_PENDING. For the latter request, it will pass in the
2101 * mirrors that have been synchronized, then the stale bit will be cleared
2102 * to make the file's state turn into RDONLY.
2103 * For concurrent write reqeust, it just needs to change the file's state
2104 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2105 * version because the version will be increased in the transition to
2106 * SYNC_PENDING later so that it can deny the write request from potential
2107 * evicted SYNC clients. */
2109 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2110 struct md_layout_change *mlc, struct thandle *handle)
2112 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2113 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2118 /* operation validation */
2119 switch (mlc->mlc_opc) {
2120 case MD_LAYOUT_RESYNC_DONE:
2121 /* resync complete. */
2122 case MD_LAYOUT_WRITE:
2123 /* concurrent write. */
2125 case MD_LAYOUT_RESYNC:
2126 /* resync again, most likely the previous run failed.
2127 * no-op if it's already in SYNC_PENDING state */
2133 if (mlc->mlc_som.lsa_valid & LSOM_FL_VALID) {
2134 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2135 if (rc && rc != -ENODATA)
2138 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2139 som_buf->lb_buf = &mlc->mlc_som;
2140 som_buf->lb_len = sizeof(mlc->mlc_som);
2143 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2147 /* record a changelog for the completion of resync */
2148 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2152 /* RESYNC_DONE has piggybacked size and blocks */
2154 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2155 XATTR_NAME_SOM, fl, handle);
2160 rc = mdd_trans_start(env, mdd, handle);
2164 /* it needs a sync tx to make FLR to work properly */
2165 handle->th_sync = 1;
2167 rc = mdo_layout_change(env, obj, mlc, handle);
2172 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2178 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2187 * Layout change callback for object.
2189 * This is only used by FLR for now. In the future, it can be exteneded to
2190 * handle all layout change.
2193 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2194 struct md_layout_change *mlc)
2196 struct mdd_object *obj = md2mdd_obj(o);
2197 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2198 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2199 struct lov_comp_md_v1 *lcm;
2200 struct thandle *handle;
2205 /* Verify acceptable operations */
2206 switch (mlc->mlc_opc) {
2207 case MD_LAYOUT_WRITE:
2208 case MD_LAYOUT_RESYNC:
2209 case MD_LAYOUT_RESYNC_DONE:
2215 handle = mdd_trans_create(env, mdd);
2217 RETURN(PTR_ERR(handle));
2219 rc = mdd_get_lov_ea(env, obj, buf);
2226 /* analyze the layout to make sure it's a FLR file */
2228 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2229 GOTO(out, rc = -EINVAL);
2231 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2233 /* please refer to HLD of FLR for state transition */
2234 switch (flr_state) {
2235 case LCM_FL_NOT_FLR:
2236 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2238 case LCM_FL_WRITE_PENDING:
2239 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2242 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2244 case LCM_FL_SYNC_PENDING:
2245 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2254 mdd_trans_stop(env, mdd, rc, handle);
2259 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2260 struct mdd_object *child, const struct lu_attr *attr,
2261 const struct md_op_spec *spec,
2262 struct dt_allocation_hint *hint)
2264 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2265 struct dt_object *nc = mdd_object_child(child);
2267 memset(hint, 0, sizeof(*hint));
2269 /* For striped directory, give striping EA to lod_ah_init, which will
2270 * decide the stripe_offset and stripe count by it. */
2271 if (S_ISDIR(attr->la_mode) &&
2272 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2273 hint->dah_eadata = spec->u.sp_ea.eadata;
2274 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2276 hint->dah_eadata = NULL;
2277 hint->dah_eadata_len = 0;
2280 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2281 hint->dah_eadata, hint->dah_eadata_len);
2282 /* @hint will be initialized by underlying device. */
2283 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2287 * do NOT or the MAY_*'s, you'll get the weakest
2289 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
2293 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2294 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2295 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2296 * owner can write to a file even if it is marked readonly to hide
2297 * its brokenness. (bug 5781) */
2298 if (flags & MDS_OPEN_OWNEROVERRIDE) {
2299 struct lu_ucred *uc = lu_ucred_check(env);
2301 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2305 if (flags & FMODE_READ)
2307 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
2309 if (flags & MDS_FMODE_EXEC)
2314 static int mdd_open_sanity_check(const struct lu_env *env,
2315 struct mdd_object *obj,
2316 const struct lu_attr *attr, int flag)
2322 if (mdd_is_dead_obj(obj))
2325 if (S_ISLNK(attr->la_mode))
2328 mode = accmode(env, attr, flag);
2330 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
2333 if (!(flag & MDS_OPEN_CREATED)) {
2334 rc = mdd_permission_internal(env, obj, attr, mode);
2339 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
2340 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
2341 flag &= ~MDS_OPEN_TRUNC;
2343 /* For writing append-only file must open it with append mode. */
2344 if (attr->la_flags & LUSTRE_APPEND_FL) {
2345 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
2347 if (flag & MDS_OPEN_TRUNC)
2354 static int mdd_open(const struct lu_env *env, struct md_object *obj,
2357 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2358 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
2359 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2362 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2364 rc = mdd_la_get(env, mdd_obj, attr);
2368 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
2372 mdd_obj->mod_count++;
2374 mdd_changelog(env, CL_OPEN, flags, md_dev, mdo2fid(mdd_obj));
2378 mdd_write_unlock(env, mdd_obj);
2382 static int mdd_declare_close(const struct lu_env *env,
2383 struct mdd_object *obj,
2385 struct thandle *handle)
2389 rc = orph_declare_index_delete(env, obj, handle);
2393 return mdo_declare_destroy(env, obj, handle);
2397 * No permission check is needed.
2399 static int mdd_close(const struct lu_env *env, struct md_object *obj,
2400 struct md_attr *ma, int mode)
2402 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2403 struct mdd_device *mdd = mdo2mdd(obj);
2404 struct thandle *handle = NULL;
2407 bool blocked = false;
2410 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
2411 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2412 mdd_obj->mod_count--;
2413 mdd_write_unlock(env, mdd_obj);
2415 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
2416 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
2417 "list\n", PFID(mdd_object_fid(mdd_obj)));
2421 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
2422 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
2423 /* check without any lock */
2424 is_orphan = mdd_obj->mod_count == 1 &&
2425 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
2429 /* mdd_trans_create() maybe failed because of barrier_entry(),
2430 * under such case, the orphan MDT-object will be left in the
2431 * orphan list, and when the MDT remount next time, the unused
2432 * orphans will be destroyed automatically.
2434 * One exception: the former mdd_finish_unlink may failed to
2435 * add the orphan MDT-object to the orphan list, then if the
2436 * mdd_trans_create() failed because of barrier_entry(), the
2437 * MDT-object will become real orphan that is neither in the
2438 * namespace nor in the orphan list. Such bad case should be
2439 * very rare and will be handled by e2fsck/lfsck. */
2440 handle = mdd_trans_create(env, mdo2mdd(obj));
2441 if (IS_ERR(handle)) {
2442 rc = PTR_ERR(handle);
2443 if (rc != -EINPROGRESS)
2451 rc = mdd_declare_close(env, mdd_obj, ma, handle);
2455 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2459 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2465 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2466 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
2468 CERROR("Failed to get lu_attr of "DFID": %d\n",
2469 PFID(mdd_object_fid(mdd_obj)), rc);
2473 /* check again with lock */
2474 is_orphan = (mdd_obj->mod_count == 1) &&
2475 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
2476 ma->ma_attr.la_nlink == 0);
2478 if (is_orphan && !handle && !blocked) {
2479 mdd_write_unlock(env, mdd_obj);
2483 mdd_obj->mod_count--; /*release open count */
2485 if (!is_orphan || blocked)
2489 /* NB: Object maybe not in orphan list originally, it is rare case for
2490 * mdd_finish_unlink() failure, in that case, the object doesn't have
2491 * ORPHAN_OBJ flag */
2492 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2493 /* remove link to object from orphan index */
2494 LASSERT(handle != NULL);
2495 rc = __mdd_orphan_del(env, mdd_obj, handle);
2497 CERROR("%s: unable to delete "DFID" from orphan list: "
2498 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2499 PFID(mdd_object_fid(mdd_obj)), rc);
2500 /* If object was not deleted from orphan list, do not
2501 * destroy OSS objects, which will be done when next
2506 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2507 "list, OSS objects to be destroyed.\n",
2508 PFID(mdd_object_fid(mdd_obj)));
2511 rc = mdo_destroy(env, mdd_obj, handle);
2514 CERROR("%s: unable to delete "DFID" from orphan list: "
2515 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2516 PFID(mdd_object_fid(mdd_obj)), rc);
2521 mdd_write_unlock(env, mdd_obj);
2523 if (!rc && !blocked &&
2524 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2525 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2526 if (handle == NULL) {
2527 handle = mdd_trans_create(env, mdo2mdd(obj));
2529 GOTO(stop, rc = PTR_ERR(handle));
2531 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2536 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2541 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2546 if (handle != NULL && !IS_ERR(handle))
2547 rc = mdd_trans_stop(env, mdd, rc, handle);
2553 * Permission check is done when open,
2554 * no need check again.
2556 static int mdd_readpage_sanity_check(const struct lu_env *env,
2557 struct mdd_object *obj)
2559 struct dt_object *next = mdd_object_child(obj);
2563 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2571 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2572 size_t nob, const struct dt_it_ops *iops,
2573 struct dt_it *it, __u32 attr, void *arg)
2575 struct lu_dirpage *dp = &lp->lp_dir;
2579 struct lu_dirent *ent;
2580 struct lu_dirent *last = NULL;
2584 if (nob < sizeof(*dp))
2587 memset(area, 0, sizeof (*dp));
2588 area += sizeof (*dp);
2589 nob -= sizeof (*dp);
2596 len = iops->key_size(env, it);
2598 /* IAM iterator can return record with zero len. */
2602 hash = iops->store(env, it);
2603 if (unlikely(first)) {
2605 dp->ldp_hash_start = cpu_to_le64(hash);
2608 /* calculate max space required for lu_dirent */
2609 recsize = lu_dirent_calc_size(len, attr);
2611 if (nob >= recsize) {
2612 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2613 if (result == -ESTALE)
2618 /* osd might not able to pack all attributes,
2619 * so recheck rec length */
2620 recsize = le16_to_cpu(ent->lde_reclen);
2622 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2623 fid_le_to_cpu(&fid, &ent->lde_fid);
2624 if (fid_is_dot_lustre(&fid))
2628 result = (last != NULL) ? 0 :-EINVAL;
2632 ent = (void *)ent + recsize;
2636 result = iops->next(env, it);
2637 if (result == -ESTALE)
2639 } while (result == 0);
2642 dp->ldp_hash_end = cpu_to_le64(hash);
2644 if (last->lde_hash == dp->ldp_hash_end)
2645 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2646 last->lde_reclen = 0; /* end mark */
2649 /* end of directory */
2650 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2651 else if (result < 0)
2652 CWARN("build page failed: %d!\n", result);
2656 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2657 const struct lu_rdpg *rdpg)
2659 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2663 if (mdd_object_exists(mdd_obj) == 0) {
2664 CERROR("%s: object "DFID" not found: rc = -2\n",
2665 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2669 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2670 rc = mdd_readpage_sanity_check(env, mdd_obj);
2672 GOTO(out_unlock, rc);
2674 if (mdd_is_dead_obj(mdd_obj)) {
2676 struct lu_dirpage *dp;
2679 * According to POSIX, please do not return any entry to client:
2680 * even dot and dotdot should not be returned.
2682 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2683 PFID(mdd_object_fid(mdd_obj)));
2685 if (rdpg->rp_count <= 0)
2686 GOTO(out_unlock, rc = -EFAULT);
2687 LASSERT(rdpg->rp_pages != NULL);
2689 pg = rdpg->rp_pages[0];
2690 dp = (struct lu_dirpage *)kmap(pg);
2691 memset(dp, 0 , sizeof(struct lu_dirpage));
2692 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2693 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2694 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2696 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2699 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2700 mdd_dir_page_build, NULL);
2702 struct lu_dirpage *dp;
2704 dp = kmap(rdpg->rp_pages[0]);
2705 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2708 * No pages were processed, mark this for first page
2711 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2712 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2713 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2715 kunmap(rdpg->rp_pages[0]);
2718 GOTO(out_unlock, rc);
2720 mdd_read_unlock(env, mdd_obj);
2724 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2726 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2728 if (mdd_object_exists(mdd_obj) == 0) {
2731 CERROR("%s: object "DFID" not found: rc = %d\n",
2732 mdd_obj_dev_name(mdd_obj),
2733 PFID(mdd_object_fid(mdd_obj)), rc);
2736 return dt_object_sync(env, mdd_object_child(mdd_obj),
2740 static int mdd_object_lock(const struct lu_env *env,
2741 struct md_object *obj,
2742 struct lustre_handle *lh,
2743 struct ldlm_enqueue_info *einfo,
2744 union ldlm_policy_data *policy)
2746 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2747 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2751 static int mdd_object_unlock(const struct lu_env *env,
2752 struct md_object *obj,
2753 struct ldlm_enqueue_info *einfo,
2754 union ldlm_policy_data *policy)
2756 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2757 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2760 const struct md_object_operations mdd_obj_ops = {
2761 .moo_permission = mdd_permission,
2762 .moo_attr_get = mdd_attr_get,
2763 .moo_attr_set = mdd_attr_set,
2764 .moo_xattr_get = mdd_xattr_get,
2765 .moo_xattr_set = mdd_xattr_set,
2766 .moo_xattr_list = mdd_xattr_list,
2767 .moo_invalidate = mdd_invalidate,
2768 .moo_xattr_del = mdd_xattr_del,
2769 .moo_swap_layouts = mdd_swap_layouts,
2770 .moo_open = mdd_open,
2771 .moo_close = mdd_close,
2772 .moo_readpage = mdd_readpage,
2773 .moo_readlink = mdd_readlink,
2774 .moo_changelog = mdd_changelog,
2775 .moo_object_sync = mdd_object_sync,
2776 .moo_object_lock = mdd_object_lock,
2777 .moo_object_unlock = mdd_object_unlock,
2778 .moo_layout_change = mdd_layout_change,