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;
657 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
658 changelog_rec_offset(flags & CLF_SUPPORTED,
659 xflags & CLFE_SUPPORTED));
660 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
661 if (buf->lb_buf == NULL)
665 rec->cr_hdr.lrh_len = reclen;
666 rec->cr.cr_flags = flags;
667 rec->cr.cr_type = (__u32)type;
668 rec->cr.cr_tfid = *fid;
669 rec->cr.cr_namelen = 0;
671 if (flags & CLF_JOBID)
672 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
674 if (flags & CLF_EXTRA_FLAGS) {
675 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
676 if (xflags & CLFE_UIDGID)
677 mdd_changelog_rec_extra_uidgid(&rec->cr,
678 uc->uc_uid, uc->uc_gid);
681 rc = mdd_changelog_store(env, mdd, rec, handle);
687 /** Store a data change changelog record
688 * If this fails, we must fail the whole transaction; we don't
689 * want the change to commit without the log entry.
690 * \param mdd_obj - mdd_object of change
691 * \param handle - transaction handle
693 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
694 enum changelog_rec_type type, int flags,
695 struct mdd_object *mdd_obj, struct thandle *handle)
699 LASSERT(mdd_obj != NULL);
700 LASSERT(handle != NULL);
703 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
705 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
708 if (mdd_is_volatile_obj(mdd_obj))
711 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
712 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
713 /* Don't need multiple updates in this log */
714 /* Don't check under lock - no big deal if we get an extra
719 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
720 mdo2fid(mdd_obj), handle);
722 mdd_obj->mod_cltime = ktime_get();
727 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
728 int flags, struct md_device *m, const struct lu_fid *fid)
730 struct thandle *handle;
731 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
736 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
738 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
741 LASSERT(fid != NULL);
743 handle = mdd_trans_create(env, mdd);
745 RETURN(PTR_ERR(handle));
747 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
751 rc = mdd_trans_start(env, mdd, handle);
755 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
759 rc = mdd_trans_stop(env, mdd, rc, handle);
765 * Save LMA extended attributes with data from \a ma.
767 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
768 * not, LMA EA will be first read from disk, modified and write back.
771 /* Precedence for choosing record type when multiple
772 * attributes change: setattr > mtime > ctime > atime
773 * (ctime changes when mtime does, plus chmod/chown.
774 * atime and ctime are independent.) */
775 static int mdd_attr_set_changelog(const struct lu_env *env,
776 struct md_object *obj, struct thandle *handle,
779 struct mdd_device *mdd = mdo2mdd(obj);
782 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
783 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
784 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
785 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
786 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
787 bits = bits & mdd->mdd_cl.mc_mask;
788 /* This is an implementation limit rather than a protocol limit */
789 CLASSERT(CL_LAST <= sizeof(int) * 8);
793 /* The record type is the lowest non-masked set bit */
796 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
797 return mdd_changelog_data_store(env, mdd, type, (int)valid,
798 md2mdd_obj(obj), handle);
801 static int mdd_declare_attr_set(const struct lu_env *env,
802 struct mdd_device *mdd,
803 struct mdd_object *obj,
804 const struct lu_attr *attr,
805 struct thandle *handle)
809 rc = mdo_declare_attr_set(env, obj, attr, handle);
813 #ifdef CONFIG_FS_POSIX_ACL
814 if (attr->la_valid & LA_MODE) {
815 mdd_read_lock(env, obj, MOR_TGT_CHILD);
816 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
817 XATTR_NAME_ACL_ACCESS);
818 mdd_read_unlock(env, obj);
819 if (rc == -EOPNOTSUPP || rc == -ENODATA)
825 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
826 rc = mdo_declare_xattr_set(env, obj, buf,
827 XATTR_NAME_ACL_ACCESS, 0,
835 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
842 * permission changes may require sync operation, to mitigate performance
843 * impact, only do this for dir and when permission is reduced.
845 * For regular files, version is updated with permission change (see VBR), async
846 * permission won't cause any issue, while missing permission change on
847 * directory may affect accessibility of other objects after recovery.
849 static inline bool permission_needs_sync(const struct lu_attr *old,
850 const struct lu_attr *new)
852 if (!S_ISDIR(old->la_mode))
855 if (new->la_valid & (LA_UID | LA_GID))
858 if (new->la_valid & LA_MODE &&
859 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
862 if ((new->la_valid & LA_MODE) &&
863 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
864 (old->la_mode & S_IRWXUGO))
870 /* set attr and LOV EA at once, return updated attr */
871 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
872 const struct md_attr *ma)
874 struct mdd_object *mdd_obj = md2mdd_obj(obj);
875 struct mdd_device *mdd = mdo2mdd(obj);
876 struct thandle *handle;
877 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
878 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
879 const struct lu_attr *la = &ma->ma_attr;
883 /* we do not use ->attr_set() for LOV/HSM EA any more */
884 LASSERT((ma->ma_valid & MA_LOV) == 0);
885 LASSERT((ma->ma_valid & MA_HSM) == 0);
887 rc = mdd_la_get(env, mdd_obj, attr);
891 *la_copy = ma->ma_attr;
892 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
896 /* no need to setattr anymore */
897 if (la_copy->la_valid == 0) {
898 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
899 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
900 PFID(mdo2fid(mdd_obj)), la->la_valid);
905 handle = mdd_trans_create(env, mdd);
907 RETURN(PTR_ERR(handle));
909 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
913 rc = mdd_trans_start(env, mdd, handle);
917 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
920 if (la->la_valid & (LA_MTIME | LA_CTIME))
921 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
922 la->la_mtime, la->la_ctime);
924 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
925 if (la_copy->la_valid & LA_FLAGS)
926 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
927 else if (la_copy->la_valid) /* setattr */
928 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
929 mdd_write_unlock(env, mdd_obj);
932 rc = mdd_attr_set_changelog(env, obj, handle,
938 rc = mdd_trans_stop(env, mdd, rc, handle);
943 static int mdd_xattr_sanity_check(const struct lu_env *env,
944 struct mdd_object *obj,
945 const struct lu_attr *attr,
948 struct lu_ucred *uc = lu_ucred_assert(env);
951 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
954 if (strncmp(XATTR_USER_PREFIX, name,
955 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
956 /* For sticky directories, only the owner and privileged user
957 * can write attributes. */
958 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
959 (uc->uc_fsuid != attr->la_uid) &&
960 !md_capable(uc, CFS_CAP_FOWNER))
963 if ((uc->uc_fsuid != attr->la_uid) &&
964 !md_capable(uc, CFS_CAP_FOWNER))
972 * Check if a string begins with a given prefix.
974 * \param str String to check
975 * \param prefix Substring to check at the beginning of \a str
976 * \return true/false whether the condition is verified.
978 static inline bool has_prefix(const char *str, const char *prefix)
980 return strncmp(prefix, str, strlen(prefix)) == 0;
984 * Indicate the kind of changelog to store (if any) for a xattr set/del.
986 * \param[in] xattr_name Full extended attribute name.
988 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
990 static enum changelog_rec_type
991 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
992 const char *xattr_name)
994 /* Layout changes systematically recorded */
995 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
996 strncmp(XATTR_LUSTRE_LOV, xattr_name,
997 strlen(XATTR_LUSTRE_LOV)) == 0)
1000 /* HSM information changes systematically recorded */
1001 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1004 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1005 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1006 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
1012 static int mdd_declare_xattr_set(const struct lu_env *env,
1013 struct mdd_device *mdd,
1014 struct mdd_object *obj,
1015 const struct lu_buf *buf,
1017 int fl, struct thandle *handle)
1021 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1025 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1026 return 0; /* no changelog to store */
1028 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1032 * Compare current and future data of HSM EA and add a changelog if needed.
1034 * Caller should have write-locked \param obj.
1036 * \param buf - Future HSM EA content.
1037 * \retval 0 if no changelog is needed or changelog was added properly.
1038 * \retval -ve errno if there was a problem
1040 static int mdd_hsm_update_locked(const struct lu_env *env,
1041 struct md_object *obj,
1042 const struct lu_buf *buf,
1043 struct thandle *handle, int *cl_flags)
1045 struct mdd_thread_info *info = mdd_env_info(env);
1046 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1047 struct lu_buf *current_buf;
1048 struct md_hsm *current_mh;
1049 struct md_hsm *new_mh;
1053 OBD_ALLOC_PTR(current_mh);
1054 if (current_mh == NULL)
1057 /* Read HSM attrs from disk */
1058 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1059 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1060 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1061 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1062 if (rc < 0 && rc != -ENODATA)
1064 else if (rc == -ENODATA)
1065 current_mh->mh_flags = 0;
1067 /* Map future HSM xattr */
1068 OBD_ALLOC_PTR(new_mh);
1070 GOTO(free, rc = -ENOMEM);
1071 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1075 /* Flags differ, set flags for the changelog that will be added */
1076 if (current_mh->mh_flags != new_mh->mh_flags) {
1077 hsm_set_cl_event(cl_flags, HE_STATE);
1078 if (new_mh->mh_flags & HS_DIRTY)
1079 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1082 OBD_FREE_PTR(new_mh);
1085 OBD_FREE_PTR(current_mh);
1089 static int mdd_object_pfid_replace(const struct lu_env *env,
1090 struct mdd_object *o)
1092 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1093 struct thandle *handle;
1096 handle = mdd_trans_create(env, mdd);
1098 RETURN(PTR_ERR(handle));
1100 handle->th_complex = 1;
1102 /* it doesn't need to track the PFID update via llog, because LFSCK
1103 * will repair it even it goes wrong */
1104 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1109 rc = mdd_trans_start(env, mdd, handle);
1113 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1118 mdd_trans_stop(env, mdd, rc, handle);
1123 static int mdd_declare_xattr_del(const struct lu_env *env,
1124 struct mdd_device *mdd,
1125 struct mdd_object *obj,
1127 struct thandle *handle);
1129 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1132 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1133 struct md_object *md_vic)
1135 struct mdd_device *mdd = mdo2mdd(md_obj);
1136 struct mdd_object *obj = md2mdd_obj(md_obj);
1137 struct mdd_object *vic = md2mdd_obj(md_vic);
1138 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1139 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1140 struct lov_mds_md *lmm;
1141 struct thandle *handle;
1145 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1146 if (rc == 0) /* same fid */
1149 handle = mdd_trans_create(env, mdd);
1151 RETURN(PTR_ERR(handle));
1154 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1155 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1157 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1158 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1161 /* get EA of victim file */
1162 memset(buf_vic, 0, sizeof(*buf_vic));
1163 rc = mdd_get_lov_ea(env, vic, buf_vic);
1170 /* parse the layout of victim file */
1171 lmm = buf_vic->lb_buf;
1172 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1173 GOTO(out, rc = -EINVAL);
1175 /* save EA of target file for restore */
1176 memset(buf, 0, sizeof(*buf));
1177 rc = mdd_get_lov_ea(env, obj, buf);
1181 /* Get rid of the layout from victim object */
1182 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1186 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1187 LU_XATTR_MERGE, handle);
1191 rc = mdd_trans_start(env, mdd, handle);
1195 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1200 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1202 GOTO(out_restore, rc);
1204 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1205 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1210 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1211 LU_XATTR_REPLACE, handle);
1213 CERROR("%s: failed to rollback of layout of: "DFID
1214 ": %d, file state unknown\n",
1215 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1219 mdd_trans_stop(env, mdd, rc, handle);
1220 mdd_write_unlock(env, obj);
1221 mdd_write_unlock(env, vic);
1223 lu_buf_free(buf_vic);
1226 (void) mdd_object_pfid_replace(env, obj);
1231 static int mdd_layout_merge_allowed(const struct lu_env *env,
1232 struct md_object *target,
1233 struct md_object *victim)
1235 struct mdd_object *o1 = md2mdd_obj(target);
1237 /* cannot extend directory's LOVEA */
1238 if (S_ISDIR(mdd_object_type(o1))) {
1239 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1240 mdd_obj_dev_name(o1));
1248 * The caller should guarantee to update the object ctime
1249 * after xattr_set if needed.
1251 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1252 const struct lu_buf *buf, const char *name,
1255 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1256 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1257 struct mdd_device *mdd = mdo2mdd(obj);
1258 struct thandle *handle;
1259 enum changelog_rec_type cl_type;
1264 rc = mdd_la_get(env, mdd_obj, attr);
1268 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1272 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 && fl == LU_XATTR_MERGE) {
1273 struct md_object *victim = buf->lb_buf;
1275 if (buf->lb_len != sizeof(victim))
1278 rc = mdd_layout_merge_allowed(env, obj, victim);
1282 /* merge layout of victim as a mirror of obj's. */
1283 rc = mdd_xattr_merge(env, obj, victim);
1287 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1288 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1289 struct posix_acl *acl;
1291 /* user may set empty ACL, which should be treated as removing
1293 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1296 RETURN(PTR_ERR(acl));
1298 rc = mdd_xattr_del(env, obj, name);
1301 posix_acl_release(acl);
1304 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1305 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1309 handle = mdd_trans_create(env, mdd);
1311 RETURN(PTR_ERR(handle));
1313 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1317 rc = mdd_trans_start(env, mdd, handle);
1321 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1323 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1324 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1326 mdd_write_unlock(env, mdd_obj);
1331 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1332 mdd_write_unlock(env, mdd_obj);
1336 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1340 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1345 return mdd_trans_stop(env, mdd, rc, handle);
1348 static int mdd_declare_xattr_del(const struct lu_env *env,
1349 struct mdd_device *mdd,
1350 struct mdd_object *obj,
1352 struct thandle *handle)
1356 rc = mdo_declare_xattr_del(env, obj, name, handle);
1360 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1361 return 0; /* no changelog to store */
1363 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1367 * The caller should guarantee to update the object ctime
1368 * after xattr_set if needed.
1370 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1373 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1374 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1375 struct mdd_device *mdd = mdo2mdd(obj);
1376 struct thandle *handle;
1380 rc = mdd_la_get(env, mdd_obj, attr);
1384 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1388 handle = mdd_trans_create(env, mdd);
1390 RETURN(PTR_ERR(handle));
1392 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1396 rc = mdd_trans_start(env, mdd, handle);
1400 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1401 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1402 mdd_write_unlock(env, mdd_obj);
1406 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1409 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1413 return mdd_trans_stop(env, mdd, rc, handle);
1417 * read lov EA of an object
1418 * return the lov EA in an allocated lu_buf
1420 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1421 struct lu_buf *lmm_buf)
1423 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1428 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1430 if (rc == -ERANGE) {
1431 /* mti_big_buf is allocated but is too small
1432 * we need to increase it */
1433 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1435 if (buf->lb_buf == NULL)
1436 GOTO(out, rc = -ENOMEM);
1447 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1448 /* mti_big_buf was not allocated, so we have to
1449 * allocate it based on the ea size */
1450 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1452 if (buf->lb_buf == NULL)
1453 GOTO(out, rc = -ENOMEM);
1457 lu_buf_alloc(lmm_buf, bufsize);
1458 if (lmm_buf->lb_buf == NULL)
1459 GOTO(out, rc = -ENOMEM);
1461 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1467 lu_buf_free(lmm_buf);
1471 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1472 struct mdd_object *o, struct lu_buf *buf,
1473 struct thandle *handle)
1475 struct hsm_attrs *attrs;
1481 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1486 attrs = buf->lb_buf;
1487 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1488 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1491 /* Add a changelog record for release. */
1492 hsm_set_cl_event(&flags, HE_RELEASE);
1493 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1499 * check if layout swapping between 2 objects is allowed
1501 * - only normal FIDs or non-system IGIFs
1502 * - same type of objects
1503 * - same owner/group (so quotas are still valid) unless this is from HSM
1506 static int mdd_layout_swap_allowed(const struct lu_env *env,
1507 struct mdd_object *o1,
1508 const struct lu_attr *attr1,
1509 struct mdd_object *o2,
1510 const struct lu_attr *attr2,
1513 const struct lu_fid *fid1, *fid2;
1519 if (!fid_is_norm(fid1) &&
1520 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1523 if (!fid_is_norm(fid2) &&
1524 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1527 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1528 if (S_ISDIR(mdd_object_type(o1)))
1530 if (S_ISREG(mdd_object_type(o1)))
1535 if (flags & SWAP_LAYOUTS_MDS_HSM)
1538 if ((attr1->la_uid != attr2->la_uid) ||
1539 (attr1->la_gid != attr2->la_gid))
1545 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1546 * look into the layout in MDD layer. */
1547 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1549 struct lov_comp_md_v1 *comp_v1;
1550 struct lov_mds_md *v1;
1554 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1555 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1556 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1562 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1563 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1566 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1567 off = le32_to_cpu(comp_v1->lcm_entries[i].
1569 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1574 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1575 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1586 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1588 return mdd_lmm_oi(lmm, oi, true);
1591 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1593 return mdd_lmm_oi(lmm, oi, false);
1596 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1598 struct lov_comp_md_v1 *comp_v1;
1600 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1601 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1603 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1605 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1606 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1607 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1608 __u16 tmp_gen = *gen;
1610 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1612 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1619 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1621 return mdd_lmm_gen(lmm, gen, true);
1624 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1626 return mdd_lmm_gen(lmm, gen, false);
1630 * swap layouts between 2 lustre objects
1632 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1633 struct md_object *obj2, __u64 flags)
1635 struct mdd_thread_info *info = mdd_env_info(env);
1636 struct mdd_object *fst_o = md2mdd_obj(obj1);
1637 struct mdd_object *snd_o = md2mdd_obj(obj2);
1638 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1639 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1640 struct mdd_device *mdd = mdo2mdd(obj1);
1641 struct lov_mds_md *fst_lmm, *snd_lmm;
1642 struct lu_buf *fst_buf = &info->mti_buf[0];
1643 struct lu_buf *snd_buf = &info->mti_buf[1];
1644 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1645 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1646 struct ost_id *saved_oi = NULL;
1647 struct thandle *handle;
1648 __u32 fst_gen, snd_gen, saved_gen;
1654 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1655 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1657 /* we have to sort the 2 obj, so locking will always
1658 * be in the same order, even in case of 2 concurrent swaps */
1659 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1660 if (rc == 0) /* same fid ? */
1666 rc = mdd_la_get(env, fst_o, fst_la);
1670 rc = mdd_la_get(env, snd_o, snd_la);
1674 /* check if layout swapping is allowed */
1675 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1679 handle = mdd_trans_create(env, mdd);
1681 RETURN(PTR_ERR(handle));
1683 /* objects are already sorted */
1684 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1685 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1687 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1688 if (rc < 0 && rc != -ENODATA)
1691 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1692 if (rc < 0 && rc != -ENODATA)
1695 /* swapping 2 non existant layouts is a success */
1696 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1699 /* to help inode migration between MDT, it is better to
1700 * start by the no layout file (if one), so we order the swap */
1701 if (snd_buf->lb_buf == NULL) {
1703 swap(fst_buf, snd_buf);
1706 fst_gen = snd_gen = 0;
1707 /* lmm and generation layout initialization */
1708 if (fst_buf->lb_buf != NULL) {
1709 fst_lmm = fst_buf->lb_buf;
1710 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1711 fst_fl = LU_XATTR_REPLACE;
1715 fst_fl = LU_XATTR_CREATE;
1718 snd_lmm = snd_buf->lb_buf;
1719 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1721 saved_gen = fst_gen;
1722 /* increase the generation layout numbers */
1727 * XXX The layout generation is used to generate component IDs for
1728 * the composite file, we have to do some special tweaks to make
1729 * sure the layout generation is always adequate for that job.
1732 /* Skip invalid generation number for composite layout */
1733 if ((snd_gen & LCME_ID_MASK) == 0)
1735 if ((fst_gen & LCME_ID_MASK) == 0)
1737 /* Make sure the generation is greater than all the component IDs */
1738 if (fst_gen < snd_gen)
1740 else if (fst_gen > snd_gen)
1743 /* set the file specific informations in lmm */
1744 if (fst_lmm != NULL) {
1745 saved_oi = &info->mti_oa.o_oi;
1746 mdd_get_lmm_oi(fst_lmm, saved_oi);
1747 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1748 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1749 mdd_set_lmm_oi(snd_lmm, saved_oi);
1751 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1752 cpu_to_le32(LOV_MAGIC_MAGIC))
1753 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1755 GOTO(stop, rc = -EPROTO);
1757 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1759 /* Prepare HSM attribute if it's required */
1760 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1761 const int buflen = sizeof(struct hsm_attrs);
1763 lu_buf_alloc(fst_hsm_buf, buflen);
1764 lu_buf_alloc(snd_hsm_buf, buflen);
1765 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1766 GOTO(stop, rc = -ENOMEM);
1768 /* Read HSM attribute */
1769 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1773 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1777 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1778 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1783 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1784 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1790 /* prepare transaction */
1791 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1796 if (fst_buf->lb_buf != NULL)
1797 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1798 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1801 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1806 rc = mdd_trans_start(env, mdd, handle);
1810 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1811 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1815 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1817 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1820 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1821 mdd_obj_dev_name(fst_o),
1822 PFID(mdo2fid(fst_o)), rc, rc2);
1827 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1831 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1834 if (fst_buf->lb_buf != NULL)
1835 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1836 LU_XATTR_REPLACE, handle);
1838 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1841 GOTO(out_restore, rc);
1843 /* Issue one changelog record per file */
1844 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1848 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1857 /* failure on second file, but first was done, so we have
1858 * to roll back first. */
1859 if (fst_buf->lb_buf != NULL) {
1860 mdd_set_lmm_oi(fst_lmm, saved_oi);
1861 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1862 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1863 LU_XATTR_REPLACE, handle);
1865 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1871 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1876 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1881 CERROR("%s: unable to roll back layout swap. FIDs: "
1882 DFID" and "DFID "error: %d/%d, steps: %d\n",
1883 mdd_obj_dev_name(fst_o),
1884 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1886 /* a solution to avoid journal commit is to panic,
1887 * but it has strong consequences so we use LBUG to
1888 * allow sysdamin to choose to panic or not
1895 rc = mdd_trans_stop(env, mdd, rc, handle);
1897 mdd_write_unlock(env, snd_o);
1898 mdd_write_unlock(env, fst_o);
1900 lu_buf_free(fst_buf);
1901 lu_buf_free(snd_buf);
1902 lu_buf_free(fst_hsm_buf);
1903 lu_buf_free(snd_hsm_buf);
1906 (void) mdd_object_pfid_replace(env, fst_o);
1907 (void) mdd_object_pfid_replace(env, snd_o);
1912 static int mdd_declare_layout_change(const struct lu_env *env,
1913 struct mdd_device *mdd,
1914 struct mdd_object *obj,
1915 struct md_layout_change *mlc,
1916 struct thandle *handle)
1920 rc = mdo_declare_layout_change(env, obj, mlc, handle);
1924 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1927 /* For PFL, this is used to instantiate necessary component objects. */
1929 mdd_layout_instantiate_component(const struct lu_env *env,
1930 struct mdd_object *obj, struct md_layout_change *mlc,
1931 struct thandle *handle)
1933 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1937 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
1940 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1942 * It's possible that another layout write intent has already
1943 * instantiated our objects, so a -EALREADY returned, and we need to
1947 RETURN(rc == -EALREADY ? 0 : rc);
1949 rc = mdd_trans_start(env, mdd, handle);
1953 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1954 rc = mdo_layout_change(env, obj, mlc, handle);
1955 mdd_write_unlock(env, obj);
1959 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1964 * Change the FLR layout from RDONLY to WRITE_PENDING.
1966 * It picks the primary mirror, and bumps the layout version, and set
1967 * layout version xattr to OST objects in a sync tx. In order to facilitate
1968 * the handling of phantom writers from evicted clients, the clients carry
1969 * layout version of the file with write RPC, so that the OSTs can verify
1970 * if the write RPCs are legitimate, meaning not from evicted clients.
1973 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
1974 struct md_layout_change *mlc, struct thandle *handle)
1976 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
1980 /* Verify acceptable operations */
1981 switch (mlc->mlc_opc) {
1982 case MD_LAYOUT_WRITE:
1984 case MD_LAYOUT_RESYNC:
1985 /* these are legal operations - this represents the case that
1986 * a few mirrors were missed in the last resync.
1987 * XXX: it will be supported later */
1988 case MD_LAYOUT_RESYNC_DONE:
1993 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
1997 rc = mdd_declare_xattr_del(env, mdd, obj, XATTR_NAME_SOM, handle);
2001 /* record a changelog for data mover to consume */
2002 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2006 rc = mdd_trans_start(env, mdd, handle);
2010 /* it needs a sync tx to make FLR to work properly */
2011 handle->th_sync = 1;
2013 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2014 rc = mdo_layout_change(env, obj, mlc, handle);
2016 rc = mdo_xattr_del(env, obj, XATTR_NAME_SOM, handle);
2020 mdd_write_unlock(env, obj);
2024 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2035 * Handle mirrored file state transition when it's in WRITE_PENDING.
2037 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2038 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2039 * layout version will be increased, and the file's state will be changed to
2043 mdd_layout_update_write_pending(const struct lu_env *env,
2044 struct mdd_object *obj, struct md_layout_change *mlc,
2045 struct thandle *handle)
2047 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2051 switch (mlc->mlc_opc) {
2052 case MD_LAYOUT_RESYNC:
2053 /* Upon receiving the resync request, it should
2054 * instantiate all stale components right away to get ready
2055 * for mirror copy. In order to avoid layout version change,
2056 * client should avoid sending LAYOUT_WRITE request at the
2059 case MD_LAYOUT_WRITE:
2060 /* legal race for concurrent write, the file state has been
2061 * changed by another client. */
2067 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2071 rc = mdd_trans_start(env, mdd, handle);
2075 /* it needs a sync tx to make FLR to work properly */
2076 handle->th_sync = 1;
2078 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2079 rc = mdo_layout_change(env, obj, mlc, handle);
2080 mdd_write_unlock(env, obj);
2091 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2093 * Only concurrent write and sync complete requests are possible when the
2094 * file is in SYNC_PENDING. For the latter request, it will pass in the
2095 * mirrors that have been synchronized, then the stale bit will be cleared
2096 * to make the file's state turn into RDONLY.
2097 * For concurrent write reqeust, it just needs to change the file's state
2098 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2099 * version because the version will be increased in the transition to
2100 * SYNC_PENDING later so that it can deny the write request from potential
2101 * evicted SYNC clients. */
2103 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2104 struct md_layout_change *mlc, struct thandle *handle)
2106 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2107 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2112 /* operation validation */
2113 switch (mlc->mlc_opc) {
2114 case MD_LAYOUT_RESYNC_DONE:
2115 /* resync complete. */
2116 case MD_LAYOUT_WRITE:
2117 /* concurrent write. */
2119 case MD_LAYOUT_RESYNC:
2120 /* resync again, most likely the previous run failed.
2121 * no-op if it's already in SYNC_PENDING state */
2127 if (mlc->mlc_som.lsa_valid & LSOM_FL_VALID) {
2128 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2129 if (rc && rc != -ENODATA)
2132 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2133 som_buf->lb_buf = &mlc->mlc_som;
2134 som_buf->lb_len = sizeof(mlc->mlc_som);
2137 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2141 /* record a changelog for the completion of resync */
2142 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2146 /* RESYNC_DONE has piggybacked size and blocks */
2148 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2149 XATTR_NAME_SOM, fl, handle);
2154 rc = mdd_trans_start(env, mdd, handle);
2158 /* it needs a sync tx to make FLR to work properly */
2159 handle->th_sync = 1;
2161 rc = mdo_layout_change(env, obj, mlc, handle);
2166 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2172 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2181 * Layout change callback for object.
2183 * This is only used by FLR for now. In the future, it can be exteneded to
2184 * handle all layout change.
2187 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2188 struct md_layout_change *mlc)
2190 struct mdd_object *obj = md2mdd_obj(o);
2191 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2192 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2193 struct lov_comp_md_v1 *lcm;
2194 struct thandle *handle;
2199 /* Verify acceptable operations */
2200 switch (mlc->mlc_opc) {
2201 case MD_LAYOUT_WRITE:
2202 case MD_LAYOUT_RESYNC:
2203 case MD_LAYOUT_RESYNC_DONE:
2209 handle = mdd_trans_create(env, mdd);
2211 RETURN(PTR_ERR(handle));
2213 rc = mdd_get_lov_ea(env, obj, buf);
2220 /* analyze the layout to make sure it's a FLR file */
2222 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2223 GOTO(out, rc = -EINVAL);
2225 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2227 /* please refer to HLD of FLR for state transition */
2228 switch (flr_state) {
2229 case LCM_FL_NOT_FLR:
2230 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2232 case LCM_FL_WRITE_PENDING:
2233 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2236 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2238 case LCM_FL_SYNC_PENDING:
2239 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2248 mdd_trans_stop(env, mdd, rc, handle);
2253 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2254 struct mdd_object *child, const struct lu_attr *attr,
2255 const struct md_op_spec *spec,
2256 struct dt_allocation_hint *hint)
2258 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2259 struct dt_object *nc = mdd_object_child(child);
2261 memset(hint, 0, sizeof(*hint));
2263 /* For striped directory, give striping EA to lod_ah_init, which will
2264 * decide the stripe_offset and stripe count by it. */
2265 if (S_ISDIR(attr->la_mode) &&
2266 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2267 hint->dah_eadata = spec->u.sp_ea.eadata;
2268 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2270 hint->dah_eadata = NULL;
2271 hint->dah_eadata_len = 0;
2274 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2275 hint->dah_eadata, hint->dah_eadata_len);
2276 /* @hint will be initialized by underlying device. */
2277 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2281 * do NOT or the MAY_*'s, you'll get the weakest
2283 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
2287 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2288 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2289 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2290 * owner can write to a file even if it is marked readonly to hide
2291 * its brokenness. (bug 5781) */
2292 if (flags & MDS_OPEN_OWNEROVERRIDE) {
2293 struct lu_ucred *uc = lu_ucred_check(env);
2295 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2299 if (flags & FMODE_READ)
2301 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
2303 if (flags & MDS_FMODE_EXEC)
2308 static int mdd_open_sanity_check(const struct lu_env *env,
2309 struct mdd_object *obj,
2310 const struct lu_attr *attr, int flag)
2316 if (mdd_is_dead_obj(obj))
2319 if (S_ISLNK(attr->la_mode))
2322 mode = accmode(env, attr, flag);
2324 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
2327 if (!(flag & MDS_OPEN_CREATED)) {
2328 rc = mdd_permission_internal(env, obj, attr, mode);
2333 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
2334 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
2335 flag &= ~MDS_OPEN_TRUNC;
2337 /* For writing append-only file must open it with append mode. */
2338 if (attr->la_flags & LUSTRE_APPEND_FL) {
2339 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
2341 if (flag & MDS_OPEN_TRUNC)
2348 static int mdd_open(const struct lu_env *env, struct md_object *obj,
2351 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2352 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2355 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2357 rc = mdd_la_get(env, mdd_obj, attr);
2361 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
2365 mdd_obj->mod_count++;
2368 mdd_write_unlock(env, mdd_obj);
2372 static int mdd_declare_close(const struct lu_env *env,
2373 struct mdd_object *obj,
2375 struct thandle *handle)
2379 rc = orph_declare_index_delete(env, obj, handle);
2383 return mdo_declare_destroy(env, obj, handle);
2387 * No permission check is needed.
2389 static int mdd_close(const struct lu_env *env, struct md_object *obj,
2390 struct md_attr *ma, int mode)
2392 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2393 struct mdd_device *mdd = mdo2mdd(obj);
2394 struct thandle *handle = NULL;
2397 bool blocked = false;
2400 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
2401 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2402 mdd_obj->mod_count--;
2403 mdd_write_unlock(env, mdd_obj);
2405 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
2406 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
2407 "list\n", PFID(mdd_object_fid(mdd_obj)));
2411 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
2412 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
2413 /* check without any lock */
2414 is_orphan = mdd_obj->mod_count == 1 &&
2415 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
2419 /* mdd_trans_create() maybe failed because of barrier_entry(),
2420 * under such case, the orphan MDT-object will be left in the
2421 * orphan list, and when the MDT remount next time, the unused
2422 * orphans will be destroyed automatically.
2424 * One exception: the former mdd_finish_unlink may failed to
2425 * add the orphan MDT-object to the orphan list, then if the
2426 * mdd_trans_create() failed because of barrier_entry(), the
2427 * MDT-object will become real orphan that is neither in the
2428 * namespace nor in the orphan list. Such bad case should be
2429 * very rare and will be handled by e2fsck/lfsck. */
2430 handle = mdd_trans_create(env, mdo2mdd(obj));
2431 if (IS_ERR(handle)) {
2432 rc = PTR_ERR(handle);
2433 if (rc != -EINPROGRESS)
2441 rc = mdd_declare_close(env, mdd_obj, ma, handle);
2445 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2449 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2455 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2456 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
2458 CERROR("Failed to get lu_attr of "DFID": %d\n",
2459 PFID(mdd_object_fid(mdd_obj)), rc);
2463 /* check again with lock */
2464 is_orphan = (mdd_obj->mod_count == 1) &&
2465 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
2466 ma->ma_attr.la_nlink == 0);
2468 if (is_orphan && !handle && !blocked) {
2469 mdd_write_unlock(env, mdd_obj);
2473 mdd_obj->mod_count--; /*release open count */
2475 if (!is_orphan || blocked)
2479 /* NB: Object maybe not in orphan list originally, it is rare case for
2480 * mdd_finish_unlink() failure, in that case, the object doesn't have
2481 * ORPHAN_OBJ flag */
2482 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2483 /* remove link to object from orphan index */
2484 LASSERT(handle != NULL);
2485 rc = __mdd_orphan_del(env, mdd_obj, handle);
2487 CERROR("%s: unable to delete "DFID" from orphan list: "
2488 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2489 PFID(mdd_object_fid(mdd_obj)), rc);
2490 /* If object was not deleted from orphan list, do not
2491 * destroy OSS objects, which will be done when next
2496 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2497 "list, OSS objects to be destroyed.\n",
2498 PFID(mdd_object_fid(mdd_obj)));
2501 rc = mdo_destroy(env, mdd_obj, handle);
2504 CERROR("%s: unable to delete "DFID" from orphan list: "
2505 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2506 PFID(mdd_object_fid(mdd_obj)), rc);
2511 mdd_write_unlock(env, mdd_obj);
2513 if (!rc && !blocked &&
2514 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2515 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2516 if (handle == NULL) {
2517 handle = mdd_trans_create(env, mdo2mdd(obj));
2519 GOTO(stop, rc = PTR_ERR(handle));
2521 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2526 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2531 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2536 if (handle != NULL && !IS_ERR(handle))
2537 rc = mdd_trans_stop(env, mdd, rc, handle);
2543 * Permission check is done when open,
2544 * no need check again.
2546 static int mdd_readpage_sanity_check(const struct lu_env *env,
2547 struct mdd_object *obj)
2549 struct dt_object *next = mdd_object_child(obj);
2553 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2561 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2562 size_t nob, const struct dt_it_ops *iops,
2563 struct dt_it *it, __u32 attr, void *arg)
2565 struct lu_dirpage *dp = &lp->lp_dir;
2569 struct lu_dirent *ent;
2570 struct lu_dirent *last = NULL;
2574 if (nob < sizeof(*dp))
2577 memset(area, 0, sizeof (*dp));
2578 area += sizeof (*dp);
2579 nob -= sizeof (*dp);
2586 len = iops->key_size(env, it);
2588 /* IAM iterator can return record with zero len. */
2592 hash = iops->store(env, it);
2593 if (unlikely(first)) {
2595 dp->ldp_hash_start = cpu_to_le64(hash);
2598 /* calculate max space required for lu_dirent */
2599 recsize = lu_dirent_calc_size(len, attr);
2601 if (nob >= recsize) {
2602 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2603 if (result == -ESTALE)
2608 /* osd might not able to pack all attributes,
2609 * so recheck rec length */
2610 recsize = le16_to_cpu(ent->lde_reclen);
2612 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2613 fid_le_to_cpu(&fid, &ent->lde_fid);
2614 if (fid_is_dot_lustre(&fid))
2618 result = (last != NULL) ? 0 :-EINVAL;
2622 ent = (void *)ent + recsize;
2626 result = iops->next(env, it);
2627 if (result == -ESTALE)
2629 } while (result == 0);
2632 dp->ldp_hash_end = cpu_to_le64(hash);
2634 if (last->lde_hash == dp->ldp_hash_end)
2635 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2636 last->lde_reclen = 0; /* end mark */
2639 /* end of directory */
2640 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2641 else if (result < 0)
2642 CWARN("build page failed: %d!\n", result);
2646 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2647 const struct lu_rdpg *rdpg)
2649 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2653 if (mdd_object_exists(mdd_obj) == 0) {
2654 CERROR("%s: object "DFID" not found: rc = -2\n",
2655 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2659 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2660 rc = mdd_readpage_sanity_check(env, mdd_obj);
2662 GOTO(out_unlock, rc);
2664 if (mdd_is_dead_obj(mdd_obj)) {
2666 struct lu_dirpage *dp;
2669 * According to POSIX, please do not return any entry to client:
2670 * even dot and dotdot should not be returned.
2672 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2673 PFID(mdd_object_fid(mdd_obj)));
2675 if (rdpg->rp_count <= 0)
2676 GOTO(out_unlock, rc = -EFAULT);
2677 LASSERT(rdpg->rp_pages != NULL);
2679 pg = rdpg->rp_pages[0];
2680 dp = (struct lu_dirpage *)kmap(pg);
2681 memset(dp, 0 , sizeof(struct lu_dirpage));
2682 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2683 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2684 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2686 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2689 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2690 mdd_dir_page_build, NULL);
2692 struct lu_dirpage *dp;
2694 dp = kmap(rdpg->rp_pages[0]);
2695 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2698 * No pages were processed, mark this for first page
2701 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2702 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2703 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2705 kunmap(rdpg->rp_pages[0]);
2708 GOTO(out_unlock, rc);
2710 mdd_read_unlock(env, mdd_obj);
2714 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2716 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2718 if (mdd_object_exists(mdd_obj) == 0) {
2721 CERROR("%s: object "DFID" not found: rc = %d\n",
2722 mdd_obj_dev_name(mdd_obj),
2723 PFID(mdd_object_fid(mdd_obj)), rc);
2726 return dt_object_sync(env, mdd_object_child(mdd_obj),
2730 static int mdd_object_lock(const struct lu_env *env,
2731 struct md_object *obj,
2732 struct lustre_handle *lh,
2733 struct ldlm_enqueue_info *einfo,
2734 union ldlm_policy_data *policy)
2736 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2737 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2741 static int mdd_object_unlock(const struct lu_env *env,
2742 struct md_object *obj,
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_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2750 const struct md_object_operations mdd_obj_ops = {
2751 .moo_permission = mdd_permission,
2752 .moo_attr_get = mdd_attr_get,
2753 .moo_attr_set = mdd_attr_set,
2754 .moo_xattr_get = mdd_xattr_get,
2755 .moo_xattr_set = mdd_xattr_set,
2756 .moo_xattr_list = mdd_xattr_list,
2757 .moo_invalidate = mdd_invalidate,
2758 .moo_xattr_del = mdd_xattr_del,
2759 .moo_swap_layouts = mdd_swap_layouts,
2760 .moo_open = mdd_open,
2761 .moo_close = mdd_close,
2762 .moo_readpage = mdd_readpage,
2763 .moo_readlink = mdd_readlink,
2764 .moo_changelog = mdd_changelog,
2765 .moo_object_sync = mdd_object_sync,
2766 .moo_object_lock = mdd_object_lock,
2767 .moo_object_unlock = mdd_object_unlock,
2768 .moo_layout_change = mdd_layout_change,