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 struct mdd_object_user {
57 struct list_head mou_list; /**< linked off mod_users */
58 u64 mou_open_flags; /**< open mode by client */
59 __u64 mou_uidgid; /**< uid_gid on client */
60 int mou_opencount; /**< # opened */
61 ktime_t mou_deniednext; /**< time of next access denied
66 static int mdd_xattr_get(const struct lu_env *env,
67 struct md_object *obj, struct lu_buf *buf,
70 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
71 struct mdd_device *mdd,
72 enum changelog_rec_type type,
73 enum changelog_rec_flags clf_flags,
74 const struct lu_fid *fid,
75 const char *xattr_name,
76 struct thandle *handle);
78 static inline bool has_prefix(const char *str, const char *prefix);
81 static u32 flags_helper(u64 open_flags)
85 if (open_flags & MDS_FMODE_EXEC) {
86 open_mode = MDS_FMODE_EXEC;
88 if (open_flags & MDS_FMODE_READ)
89 open_mode = MDS_FMODE_READ;
91 (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
92 open_mode |= MDS_FMODE_WRITE;
98 /** Allocate/init a user and its sub-structures.
103 * \retval mou [OUT] success valid structure
106 static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags,
107 uid_t uid, gid_t gid)
109 struct mdd_object_user *mou;
113 OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem);
115 RETURN(ERR_PTR(-ENOMEM));
117 mou->mou_open_flags = open_flags;
118 mou->mou_uidgid = ((__u64)uid << 32) | gid;
119 mou->mou_opencount = 0;
120 mou->mou_deniednext = ktime_set(0, 0);
126 * Free a user and its sub-structures.
128 * \param mou [IN] user to be freed.
130 static void mdd_obj_user_free(struct mdd_object_user *mou)
132 OBD_SLAB_FREE_PTR(mou, mdd_object_kmem);
136 * Find if UID/GID already has this file open
138 * Caller should have write-locked \param mdd_obj.
139 * \param mdd_obj [IN] mdd_obj
140 * \param uid [IN] client uid
141 * \param gid [IN] client gid
142 * \retval user pointer or NULL if not found
145 struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj,
146 uid_t uid, gid_t gid,
149 struct mdd_object_user *mou;
154 uidgid = ((__u64)uid << 32) | gid;
155 list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) {
156 if (mou->mou_uidgid == uidgid &&
157 flags_helper(mou->mou_open_flags) ==
158 flags_helper(open_flags))
165 * Add a user to the list of openers for this file
167 * Caller should have write-locked \param mdd_obj.
168 * \param mdd_obj [IN] mdd_obj
169 * \param mou [IN] user
171 * \retval -ve failure
173 static int mdd_obj_user_add(struct mdd_object *mdd_obj,
174 struct mdd_object_user *mou,
177 struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj);
178 struct mdd_object_user *tmp;
179 __u32 uid = mou->mou_uidgid >> 32;
180 __u32 gid = mou->mou_uidgid & ((1UL << 32) - 1);
183 tmp = mdd_obj_user_find(mdd_obj, uid, gid, mou->mou_open_flags);
187 list_add_tail(&mou->mou_list, &mdd_obj->mod_users);
190 /* next 'access denied' notification cannot happen before
193 mou->mou_deniednext =
194 ktime_add(ktime_get(),
195 ktime_set(mdd->mdd_cl.mc_deniednext, 0));
197 mou->mou_opencount++;
202 * Remove UID from the list
204 * Caller should have write-locked \param mdd_obj.
205 * \param mdd_obj [IN] mdd_obj
206 * \param uid [IN] user
207 * \retval -ve failure
209 static int mdd_obj_user_remove(struct mdd_object *mdd_obj,
210 struct mdd_object_user *mou)
217 list_del_init(&mou->mou_list);
219 mdd_obj_user_free(mou);
223 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
228 if (mdd_object_exists(obj) == 0)
231 rc = mdo_attr_get(env, obj, la);
232 if (unlikely(rc != 0)) {
234 obj->mod_flags |= DEAD_OBJ;
238 if (la->la_valid & LA_FLAGS && la->la_flags & LUSTRE_ORPHAN_FL)
239 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
244 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
246 struct mdd_thread_info *info;
248 lu_env_refill((struct lu_env *)env);
249 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
250 LASSERT(info != NULL);
254 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
258 buf = &mdd_env_info(env)->mti_buf[0];
264 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
265 const void *area, ssize_t len)
269 buf = &mdd_env_info(env)->mti_buf[0];
270 buf->lb_buf = (void *)area;
275 struct lu_object *mdd_object_alloc(const struct lu_env *env,
276 const struct lu_object_header *hdr,
279 struct mdd_object *mdd_obj;
282 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
286 o = mdd2lu_obj(mdd_obj);
287 lu_object_init(o, NULL, d);
288 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
289 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
290 mdd_obj->mod_count = 0;
291 o->lo_ops = &mdd_lu_obj_ops;
292 INIT_LIST_HEAD(&mdd_obj->mod_users);
297 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
298 const struct lu_object_conf *unused)
300 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
301 struct mdd_object *mdd_obj = lu2mdd_obj(o);
302 struct lu_object *below;
303 struct lu_device *under;
306 mdd_obj->mod_cltime = ktime_set(0, 0);
307 under = &d->mdd_child->dd_lu_dev;
308 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
310 RETURN(PTR_ERR(below));
312 lu_object_add(o, below);
317 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
321 if (lu_object_exists(o)) {
322 struct mdd_object *mdd_obj = lu2mdd_obj(o);
323 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
325 rc = mdd_la_get(env, mdd_obj, attr);
331 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
333 struct mdd_object *mdd = lu2mdd_obj(o);
334 struct mdd_object_user *mou, *tmp2;
337 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
338 list_del(&mou->mou_list);
339 mdd_obj_user_free(mou);
343 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
346 static int mdd_object_print(const struct lu_env *env, void *cookie,
347 lu_printer_t p, const struct lu_object *o)
349 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
351 return (*p)(env, cookie,
352 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
353 mdd, mdd->mod_count, mdd->mod_valid,
354 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
357 static const struct lu_object_operations mdd_lu_obj_ops = {
358 .loo_object_init = mdd_object_init,
359 .loo_object_start = mdd_object_start,
360 .loo_object_free = mdd_object_free,
361 .loo_object_print = mdd_object_print,
364 struct mdd_object *mdd_object_find(const struct lu_env *env,
365 struct mdd_device *d,
366 const struct lu_fid *f)
368 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
372 * No permission check is needed.
374 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
377 struct mdd_object *mdd_obj = md2mdd_obj(obj);
382 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
383 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
384 ma->ma_attr.la_nlink = 0;
390 * No permission check is needed.
392 static int mdd_xattr_get(const struct lu_env *env,
393 struct md_object *obj, struct lu_buf *buf,
396 struct mdd_object *mdd_obj = md2mdd_obj(obj);
397 struct mdd_device *mdd;
402 if (mdd_object_exists(mdd_obj) == 0) {
403 CERROR("%s: object "DFID" not found: rc = -2\n",
404 mdd_obj_dev_name(mdd_obj),
405 PFID(mdd_object_fid(mdd_obj)));
409 /* If the object has been destroyed, then do not get LMVEA, because
410 * it needs to load stripes from the iteration of the master object,
411 * and it will cause problem if master object has been destroyed, see
413 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
414 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
415 strcmp(name, XATTR_NAME_LMV) == 0))
418 /* If the object has been delete from the namespace, then
419 * get linkEA should return -ENOENT as well */
420 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
421 strcmp(name, XATTR_NAME_LINK) == 0))
424 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
425 rc = mdo_xattr_get(env, mdd_obj, buf, name);
426 mdd_read_unlock(env, mdd_obj);
430 /* record only getting user xattrs and acls */
432 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
433 (has_prefix(name, XATTR_USER_PREFIX) ||
434 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
435 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
436 struct thandle *handle;
439 LASSERT(mdo2fid(mdd_obj) != NULL);
441 handle = mdd_trans_create(env, mdd);
443 RETURN(PTR_ERR(handle));
445 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
450 rc2 = mdd_trans_start(env, mdd, handle);
454 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
455 mdo2fid(mdd_obj), name,
459 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
468 * Permission check is done when open,
469 * no need check again.
471 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
474 struct mdd_object *mdd_obj = md2mdd_obj(obj);
475 struct dt_object *next;
480 if (mdd_object_exists(mdd_obj) == 0) {
481 CERROR("%s: object "DFID" not found: rc = -2\n",
482 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
486 next = mdd_object_child(mdd_obj);
487 LASSERT(next != NULL);
488 LASSERT(next->do_body_ops != NULL);
489 LASSERT(next->do_body_ops->dbo_read != NULL);
490 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
491 rc = dt_read(env, next, buf, &pos);
492 mdd_read_unlock(env, mdd_obj);
497 * No permission check is needed.
499 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
502 struct mdd_object *mdd_obj = md2mdd_obj(obj);
507 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
508 rc = mdo_xattr_list(env, mdd_obj, buf);
509 mdd_read_unlock(env, mdd_obj);
511 /* If the buffer is NULL then we are only here to get the
512 * length of the xattr name list. */
513 if (rc < 0 || buf->lb_buf == NULL)
517 * Filter out XATTR_NAME_LINK if this is an orphan object. See
520 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
521 char *end = (char *)buf->lb_buf + rc;
522 char *p = buf->lb_buf;
525 char *next = p + strlen(p) + 1;
527 if (strcmp(p, XATTR_NAME_LINK) == 0) {
529 memmove(p, next, end - next);
531 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
532 " of orphan "DFID"\n",
533 PFID(mdd_object_fid(mdd_obj)));
544 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
546 return mdo_invalidate(env, md2mdd_obj(obj));
549 int mdd_declare_create_object_internal(const struct lu_env *env,
550 struct mdd_object *p,
551 struct mdd_object *c,
552 struct lu_attr *attr,
553 struct thandle *handle,
554 const struct md_op_spec *spec,
555 struct dt_allocation_hint *hint)
557 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
558 const struct dt_index_features *feat = spec->sp_feat;
562 if (feat != &dt_directory_features && feat != NULL) {
563 dof->dof_type = DFT_INDEX;
564 dof->u.dof_idx.di_feat = feat;
566 dof->dof_type = dt_mode_to_dft(attr->la_mode);
567 if (dof->dof_type == DFT_REGULAR) {
568 dof->u.dof_reg.striped =
569 md_should_create(spec->sp_cr_flags);
570 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
571 dof->u.dof_reg.striped = 0;
572 /* is this replay? */
574 dof->u.dof_reg.striped = 0;
578 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
583 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
584 struct mdd_object *c, struct lu_attr *attr,
585 struct thandle *handle,
586 const struct md_op_spec *spec,
587 struct dt_allocation_hint *hint)
589 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
593 LASSERT(!mdd_object_exists(c));
595 rc = mdo_create_object(env, c, attr, hint, dof, handle);
600 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
601 const struct lu_attr *attr, struct thandle *handle,
607 rc = mdo_attr_set(env, obj, attr, handle);
608 #ifdef CONFIG_FS_POSIX_ACL
609 if (!rc && (attr->la_valid & LA_MODE) && needacl)
610 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
615 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
616 const struct lu_attr *oattr, struct lu_attr *attr,
617 struct thandle *handle)
622 LASSERT(attr->la_valid & LA_CTIME);
623 LASSERT(oattr != NULL);
625 /* Make sure the ctime is increased only, however, it's not strictly
626 * reliable at here because there is not guarantee to hold lock on
627 * object, so we just bypass some unnecessary cmtime setting first
628 * and OSD has to check it again. */
629 if (attr->la_ctime < oattr->la_ctime)
630 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
631 else if (attr->la_valid == LA_CTIME &&
632 attr->la_ctime == oattr->la_ctime)
633 attr->la_valid &= ~LA_CTIME;
635 if (attr->la_valid != 0)
636 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
641 * This gives the same functionality as the code between
642 * sys_chmod and inode_setattr
643 * chown_common and inode_setattr
644 * utimes and inode_setattr
645 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
647 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
648 const struct lu_attr *oattr, struct lu_attr *la,
649 const unsigned long flags)
658 /* Do not permit change file type */
659 if (la->la_valid & LA_TYPE)
662 /* They should not be processed by setattr */
663 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
666 LASSERT(oattr != NULL);
668 uc = lu_ucred_check(env);
672 if (la->la_valid == LA_CTIME) {
673 if (!(flags & MDS_PERM_BYPASS))
674 /* This is only for set ctime when rename's source is
676 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
678 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
679 la->la_valid &= ~LA_CTIME;
683 if (la->la_valid == LA_ATIME) {
684 /* This is an atime-only attribute update for close RPCs. */
685 if (la->la_atime < (oattr->la_atime +
686 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
687 la->la_valid &= ~LA_ATIME;
691 /* Check if flags change. */
692 if (la->la_valid & LA_FLAGS) {
693 unsigned int oldflags = oattr->la_flags &
694 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
695 unsigned int newflags = la->la_flags &
696 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
698 if ((uc->uc_fsuid != oattr->la_uid) &&
699 !md_capable(uc, CFS_CAP_FOWNER))
702 /* The IMMUTABLE and APPEND_ONLY flags can
703 * only be changed by the relevant capability. */
704 if ((oldflags ^ newflags) &&
705 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
708 if (!S_ISDIR(oattr->la_mode))
709 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
712 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
713 (la->la_valid & ~LA_FLAGS) &&
714 !(flags & MDS_PERM_BYPASS))
717 /* Check for setting the obj time. */
718 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
719 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
720 if ((uc->uc_fsuid != oattr->la_uid) &&
721 !md_capable(uc, CFS_CAP_FOWNER)) {
722 rc = mdd_permission_internal(env, obj, oattr,
729 if (la->la_valid & LA_KILL_SUID) {
730 la->la_valid &= ~LA_KILL_SUID;
731 if ((oattr->la_mode & S_ISUID) &&
732 !(la->la_valid & LA_MODE)) {
733 la->la_mode = oattr->la_mode;
734 la->la_valid |= LA_MODE;
736 la->la_mode &= ~S_ISUID;
739 if (la->la_valid & LA_KILL_SGID) {
740 la->la_valid &= ~LA_KILL_SGID;
741 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
742 (S_ISGID | S_IXGRP)) &&
743 !(la->la_valid & LA_MODE)) {
744 la->la_mode = oattr->la_mode;
745 la->la_valid |= LA_MODE;
747 la->la_mode &= ~S_ISGID;
750 /* Make sure a caller can chmod. */
751 if (la->la_valid & LA_MODE) {
752 if (!(flags & MDS_PERM_BYPASS) &&
753 (uc->uc_fsuid != oattr->la_uid) &&
754 !md_capable(uc, CFS_CAP_FOWNER))
757 if (la->la_mode == (umode_t) -1)
758 la->la_mode = oattr->la_mode;
760 la->la_mode = (la->la_mode & S_IALLUGO) |
761 (oattr->la_mode & ~S_IALLUGO);
763 /* Also check the setgid bit! */
764 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
765 la->la_gid : oattr->la_gid) &&
766 !md_capable(uc, CFS_CAP_FSETID))
767 la->la_mode &= ~S_ISGID;
769 la->la_mode = oattr->la_mode;
772 /* Make sure a caller can chown. */
773 if (la->la_valid & LA_UID) {
774 if (la->la_uid == (uid_t) -1)
775 la->la_uid = oattr->la_uid;
776 if (((uc->uc_fsuid != oattr->la_uid) ||
777 (la->la_uid != oattr->la_uid)) &&
778 !md_capable(uc, CFS_CAP_CHOWN))
781 /* If the user or group of a non-directory has been
782 * changed by a non-root user, remove the setuid bit.
783 * 19981026 David C Niemi <niemi@tux.org>
785 * Changed this to apply to all users, including root,
786 * to avoid some races. This is the behavior we had in
787 * 2.0. The check for non-root was definitely wrong
788 * for 2.2 anyway, as it should have been using
789 * CAP_FSETID rather than fsuid -- 19990830 SD. */
790 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
791 !S_ISDIR(oattr->la_mode)) {
792 la->la_mode &= ~S_ISUID;
793 la->la_valid |= LA_MODE;
797 /* Make sure caller can chgrp. */
798 if (la->la_valid & LA_GID) {
799 if (la->la_gid == (gid_t) -1)
800 la->la_gid = oattr->la_gid;
801 if (((uc->uc_fsuid != oattr->la_uid) ||
802 ((la->la_gid != oattr->la_gid) &&
803 !lustre_in_group_p(uc, la->la_gid))) &&
804 !md_capable(uc, CFS_CAP_CHOWN))
807 /* Likewise, if the user or group of a non-directory
808 * has been changed by a non-root user, remove the
809 * setgid bit UNLESS there is no group execute bit
810 * (this would be a file marked for mandatory
811 * locking). 19981026 David C Niemi <niemi@tux.org>
813 * Removed the fsuid check (see the comment above) --
815 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
816 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
817 la->la_mode &= ~S_ISGID;
818 la->la_valid |= LA_MODE;
822 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
823 if (!((flags & MDS_OWNEROVERRIDE) &&
824 (uc->uc_fsuid == oattr->la_uid)) &&
825 !(flags & MDS_PERM_BYPASS)) {
826 rc = mdd_permission_internal(env, obj, oattr,
833 if (la->la_valid & LA_CTIME) {
834 /* The pure setattr, it has the priority over what is
835 * already set, do not drop it if ctime is equal. */
836 if (la->la_ctime < oattr->la_ctime)
837 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
843 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
844 struct mdd_device *mdd,
845 enum changelog_rec_type type,
846 enum changelog_rec_flags clf_flags,
847 const struct lu_fid *fid,
848 const char *xattr_name,
849 struct thandle *handle)
851 const struct lu_ucred *uc = lu_ucred(env);
852 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
853 struct llog_changelog_rec *rec;
858 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
861 if (uc->uc_jobid[0] != '\0')
862 clf_flags |= CLF_JOBID;
863 xflags |= CLFE_UIDGID;
866 if (type == CL_OPEN || type == CL_DN_OPEN)
868 if (type == CL_SETXATTR || type == CL_GETXATTR)
869 xflags |= CLFE_XATTR;
871 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
872 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
873 xflags & CLFE_SUPPORTED));
874 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
875 if (buf->lb_buf == NULL)
879 rec->cr_hdr.lrh_len = reclen;
880 rec->cr.cr_flags = clf_flags;
881 rec->cr.cr_type = (__u32)type;
882 rec->cr.cr_tfid = *fid;
883 rec->cr.cr_namelen = 0;
885 if (clf_flags & CLF_JOBID)
886 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
888 if (clf_flags & CLF_EXTRA_FLAGS) {
889 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
890 if (xflags & CLFE_UIDGID)
891 mdd_changelog_rec_extra_uidgid(&rec->cr,
892 uc->uc_uid, uc->uc_gid);
893 if (xflags & CLFE_NID)
894 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
895 if (xflags & CLFE_OPEN)
896 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
897 if (xflags & CLFE_XATTR) {
898 if (xattr_name == NULL)
900 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
904 rc = mdd_changelog_store(env, mdd, rec, handle);
909 /** Store a data change changelog record
910 * If this fails, we must fail the whole transaction; we don't
911 * want the change to commit without the log entry.
912 * \param mdd_obj - mdd_object of change
913 * \param handle - transaction handle
915 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
916 enum changelog_rec_type type,
917 enum changelog_rec_flags clf_flags,
918 struct mdd_object *mdd_obj, struct thandle *handle)
922 LASSERT(mdd_obj != NULL);
923 LASSERT(handle != NULL);
925 if (!mdd_changelog_enabled(env, mdd, type))
928 if (mdd_is_volatile_obj(mdd_obj))
931 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
932 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
933 /* Don't need multiple updates in this log */
934 /* Don't check under lock - no big deal if we get an extra
939 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
940 mdo2fid(mdd_obj), NULL, handle);
942 mdd_obj->mod_cltime = ktime_get();
947 int mdd_changelog_data_store_xattr(const struct lu_env *env,
948 struct mdd_device *mdd,
949 enum changelog_rec_type type,
950 enum changelog_rec_flags clf_flags,
951 struct mdd_object *mdd_obj,
952 const char *xattr_name,
953 struct thandle *handle)
957 LASSERT(mdd_obj != NULL);
958 LASSERT(handle != NULL);
960 if (!mdd_changelog_enabled(env, mdd, type))
963 if (mdd_is_volatile_obj(mdd_obj))
966 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
967 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
968 /* Don't need multiple updates in this log */
969 /* Don't check under lock - no big deal if we get an extra
975 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
976 mdo2fid(mdd_obj), xattr_name,
979 mdd_obj->mod_cltime = ktime_get();
984 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
985 * except for open flags have a dedicated record to store 32 bits of flags */
986 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
987 enum changelog_rec_flags clf_flags,
988 struct md_device *m, const struct lu_fid *fid)
990 struct thandle *handle;
991 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
995 LASSERT(fid != NULL);
997 /* We'll check this again below, but we check now before we
998 * start a transaction. */
999 if (!mdd_changelog_enabled(env, mdd, type))
1002 handle = mdd_trans_create(env, mdd);
1004 RETURN(PTR_ERR(handle));
1006 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1010 rc = mdd_trans_start(env, mdd, handle);
1014 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1018 rc = mdd_trans_stop(env, mdd, rc, handle);
1024 * Save LMA extended attributes with data from \a ma.
1026 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1027 * not, LMA EA will be first read from disk, modified and write back.
1030 /* Precedence for choosing record type when multiple
1031 * attributes change: setattr > mtime > ctime > atime
1032 * (ctime changes when mtime does, plus chmod/chown.
1033 * atime and ctime are independent.) */
1034 static int mdd_attr_set_changelog(const struct lu_env *env,
1035 struct md_object *obj, struct thandle *handle,
1038 struct mdd_device *mdd = mdo2mdd(obj);
1041 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
1042 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
1043 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
1044 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
1045 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
1046 bits = bits & mdd->mdd_cl.mc_mask;
1047 /* This is an implementation limit rather than a protocol limit */
1048 CLASSERT(CL_LAST <= sizeof(int) * 8);
1052 /* The record type is the lowest non-masked set bit */
1055 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1056 return mdd_changelog_data_store(env, mdd, type, valid,
1057 md2mdd_obj(obj), handle);
1060 static int mdd_declare_attr_set(const struct lu_env *env,
1061 struct mdd_device *mdd,
1062 struct mdd_object *obj,
1063 const struct lu_attr *attr,
1064 struct thandle *handle)
1068 rc = mdo_declare_attr_set(env, obj, attr, handle);
1072 #ifdef CONFIG_FS_POSIX_ACL
1073 if (attr->la_valid & LA_MODE) {
1074 mdd_read_lock(env, obj, MOR_TGT_CHILD);
1075 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1076 XATTR_NAME_ACL_ACCESS);
1077 mdd_read_unlock(env, obj);
1078 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1084 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1085 rc = mdo_declare_xattr_set(env, obj, buf,
1086 XATTR_NAME_ACL_ACCESS, 0,
1094 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1103 * permission changes may require sync operation, to mitigate performance
1104 * impact, only do this for dir and when permission is reduced.
1106 * For regular files, version is updated with permission change (see VBR), async
1107 * permission won't cause any issue, while missing permission change on
1108 * directory may affect accessibility of other objects after recovery.
1110 static inline bool permission_needs_sync(const struct lu_attr *old,
1111 const struct lu_attr *new)
1113 if (!S_ISDIR(old->la_mode))
1116 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1119 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1122 if (new->la_valid & LA_MODE) {
1123 /* turned on sticky bit */
1124 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1127 /* set-GID has no impact on what is allowed, not checked */
1129 /* turned off setuid bit, or one of rwx for someone */
1130 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1131 (old->la_mode & (0777 | S_ISUID)))
1138 static inline __u64 mdd_lmm_dom_size(void *buf)
1140 struct lov_mds_md *lmm = buf;
1141 struct lov_comp_md_v1 *comp_v1;
1142 struct lov_mds_md *v1;
1148 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1151 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1152 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1153 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1155 /* DoM entry is the first entry always */
1156 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1157 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1162 /* set attr and LOV EA at once, return updated attr */
1163 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1164 const struct md_attr *ma)
1166 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1167 struct mdd_device *mdd = mdo2mdd(obj);
1168 struct thandle *handle = NULL;
1169 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1170 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1171 const struct lu_attr *la = &ma->ma_attr;
1172 struct lu_ucred *uc;
1173 bool chrgrp_by_unprivileged_user = false;
1177 /* we do not use ->attr_set() for LOV/HSM EA any more */
1178 LASSERT((ma->ma_valid & MA_LOV) == 0);
1179 LASSERT((ma->ma_valid & MA_HSM) == 0);
1181 rc = mdd_la_get(env, mdd_obj, attr);
1185 *la_copy = ma->ma_attr;
1186 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
1190 /* no need to setattr anymore */
1191 if (la_copy->la_valid == 0) {
1192 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
1193 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
1194 PFID(mdo2fid(mdd_obj)), la->la_valid);
1199 /* If an unprivileged user changes group of some file,
1200 * the setattr operation will be processed synchronously to
1201 * honor the quota limit of the corresponding group. see LU-5152 */
1202 uc = lu_ucred_check(env);
1203 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1204 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1205 /* LU-10048: disable synchronous chgrp operation for it will
1206 * cause deadlock between MDT and OST.
1207 la_copy->la_valid |= LA_FLAGS;
1208 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1210 chrgrp_by_unprivileged_user = true;
1212 /* Flush the possible existing client setattr requests to OSTs
1213 * to keep the order with the current setattr operation that
1214 * will be sent directly to OSTs. see LU-5152 */
1215 /* LU-11303 disable sync as this is too heavyweight.
1216 * This should be replaced with a sync only for the object
1217 * being modified here, not the whole filesystem.
1218 rc = dt_sync(env, mdd->mdd_child);
1224 handle = mdd_trans_create(env, mdd);
1225 if (IS_ERR(handle)) {
1226 rc = PTR_ERR(handle);
1232 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1236 rc = mdd_trans_start(env, mdd, handle);
1240 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1241 permission_needs_sync(attr, la))
1242 handle->th_sync = 1;
1244 if (la->la_valid & (LA_MTIME | LA_CTIME))
1245 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1246 la->la_mtime, la->la_ctime);
1248 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1250 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1251 * otherwise following truncate will do nothing and truncated
1252 * data may be read again. This is a quick fix until LU-11033
1255 if (la_copy->la_valid & LA_SIZE) {
1256 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1258 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1262 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1263 la_copy->la_valid &= ~LA_SIZE;
1264 lu_buf_free(lov_buf);
1268 if (la_copy->la_valid) {
1269 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1271 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1272 /* rollback to the original gid */
1273 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1274 la_copy->la_gid = attr->la_gid;
1275 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1278 mdd_write_unlock(env, mdd_obj);
1282 rc = mdd_attr_set_changelog(env, obj, handle,
1286 rc = mdd_trans_stop(env, mdd, rc, handle);
1291 static int mdd_xattr_sanity_check(const struct lu_env *env,
1292 struct mdd_object *obj,
1293 const struct lu_attr *attr,
1296 struct lu_ucred *uc = lu_ucred_assert(env);
1299 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1302 if (strncmp(XATTR_USER_PREFIX, name,
1303 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1304 /* For sticky directories, only the owner and privileged user
1305 * can write attributes. */
1306 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1307 (uc->uc_fsuid != attr->la_uid) &&
1308 !md_capable(uc, CFS_CAP_FOWNER))
1310 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1311 (uc->uc_fsuid != attr->la_uid) &&
1312 !md_capable(uc, CFS_CAP_FOWNER)) {
1320 * Check if a string begins with a given prefix.
1322 * \param str String to check
1323 * \param prefix Substring to check at the beginning of \a str
1324 * \return true/false whether the condition is verified.
1326 static inline bool has_prefix(const char *str, const char *prefix)
1328 return strncmp(prefix, str, strlen(prefix)) == 0;
1332 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1334 * \param[in] xattr_name Full extended attribute name.
1336 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1338 static enum changelog_rec_type
1339 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1340 const char *xattr_name)
1342 /* Layout changes systematically recorded */
1343 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1344 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1345 strlen(XATTR_LUSTRE_LOV)) == 0)
1348 /* HSM information changes systematically recorded */
1349 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1352 /* Avoid logging SOM xattr for every file */
1353 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1356 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1357 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1358 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1359 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1360 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1366 static int mdd_declare_xattr_set(const struct lu_env *env,
1367 struct mdd_device *mdd,
1368 struct mdd_object *obj,
1369 const struct lu_buf *buf,
1371 int fl, struct thandle *handle)
1373 enum changelog_rec_type type;
1376 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1380 type = mdd_xattr_changelog_type(env, mdd, name);
1382 return 0; /* no changelog to store */
1384 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1388 * Compare current and future data of HSM EA and add a changelog if needed.
1390 * Caller should have write-locked \param obj.
1392 * \param buf - Future HSM EA content.
1393 * \retval 0 if no changelog is needed or changelog was added properly.
1394 * \retval -ve errno if there was a problem
1396 static int mdd_hsm_update_locked(const struct lu_env *env,
1397 struct md_object *obj,
1398 const struct lu_buf *buf,
1399 struct thandle *handle,
1400 enum changelog_rec_flags *clf_flags)
1402 struct mdd_thread_info *info = mdd_env_info(env);
1403 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1404 struct lu_buf *current_buf;
1405 struct md_hsm *current_mh;
1406 struct md_hsm *new_mh;
1410 OBD_ALLOC_PTR(current_mh);
1411 if (current_mh == NULL)
1414 /* Read HSM attrs from disk */
1415 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1416 MIN(mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1418 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1419 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1420 if (rc < 0 && rc != -ENODATA)
1422 else if (rc == -ENODATA)
1423 current_mh->mh_flags = 0;
1425 /* Map future HSM xattr */
1426 OBD_ALLOC_PTR(new_mh);
1428 GOTO(free, rc = -ENOMEM);
1429 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1433 /* Flags differ, set flags for the changelog that will be added */
1434 if (current_mh->mh_flags != new_mh->mh_flags) {
1435 hsm_set_cl_event(clf_flags, HE_STATE);
1436 if (new_mh->mh_flags & HS_DIRTY)
1437 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1440 OBD_FREE_PTR(new_mh);
1443 OBD_FREE_PTR(current_mh);
1447 static int mdd_object_pfid_replace(const struct lu_env *env,
1448 struct mdd_object *o)
1450 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1451 struct thandle *handle;
1454 handle = mdd_trans_create(env, mdd);
1456 RETURN(PTR_ERR(handle));
1458 handle->th_complex = 1;
1460 /* it doesn't need to track the PFID update via llog, because LFSCK
1461 * will repair it even it goes wrong */
1462 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1467 rc = mdd_trans_start(env, mdd, handle);
1471 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1476 mdd_trans_stop(env, mdd, rc, handle);
1481 static int mdd_declare_xattr_del(const struct lu_env *env,
1482 struct mdd_device *mdd,
1483 struct mdd_object *obj,
1485 struct thandle *handle);
1487 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1490 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1491 struct md_object *md_vic)
1493 struct mdd_device *mdd = mdo2mdd(md_obj);
1494 struct mdd_object *obj = md2mdd_obj(md_obj);
1495 struct mdd_object *vic = md2mdd_obj(md_vic);
1496 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1497 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1498 struct lov_mds_md *lmm;
1499 struct thandle *handle;
1503 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1504 if (rc == 0) /* same fid */
1507 handle = mdd_trans_create(env, mdd);
1509 RETURN(PTR_ERR(handle));
1512 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1513 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1515 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1516 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1519 /* get EA of victim file */
1520 memset(buf_vic, 0, sizeof(*buf_vic));
1521 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1528 /* parse the layout of victim file */
1529 lmm = buf_vic->lb_buf;
1530 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1531 GOTO(out, rc = -EINVAL);
1533 /* save EA of target file for restore */
1534 memset(buf, 0, sizeof(*buf));
1535 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1539 /* Get rid of the layout from victim object */
1540 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1544 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1545 LU_XATTR_MERGE, handle);
1549 rc = mdd_trans_start(env, mdd, handle);
1553 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1558 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1560 GOTO(out_restore, rc);
1562 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1563 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1568 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1569 LU_XATTR_REPLACE, handle);
1571 CERROR("%s: failed to rollback of layout of: "DFID
1572 ": %d, file state unknown\n",
1573 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1577 mdd_trans_stop(env, mdd, rc, handle);
1578 mdd_write_unlock(env, obj);
1579 mdd_write_unlock(env, vic);
1581 lu_buf_free(buf_vic);
1584 (void) mdd_object_pfid_replace(env, obj);
1590 * Extract the mirror with specified mirror id, and store the splitted
1591 * mirror layout to @buf.
1593 * \param[in] comp_v1 mirrored layout
1594 * \param[in] mirror_id the mirror with mirror_id to be extracted
1595 * \param[out] buf store the layout excluding the extracted mirror,
1596 * caller free the buffer we allocated in this function
1597 * \param[out] buf_vic store the extracted layout, caller free the buffer
1598 * we allocated in this function
1600 * \retval 0 on success; < 0 if error happens
1602 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1603 struct lu_buf *buf, struct lu_buf *buf_vic)
1605 struct lov_comp_md_v1 *comp_rem;
1606 struct lov_comp_md_v1 *comp_vic;
1607 struct lov_comp_md_entry_v1 *entry;
1608 struct lov_comp_md_entry_v1 *entry_rem;
1609 struct lov_comp_md_entry_v1 *entry_vic;
1611 __u16 comp_cnt, count = 0;
1612 int lmm_size, lmm_size_vic = 0;
1614 int offset, offset_rem, offset_vic;
1616 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1617 /* comp_v1 should contains more than 1 mirror */
1618 if (mirror_cnt <= 1)
1620 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1621 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1623 for (i = 0; i < comp_cnt; i++) {
1624 entry = &comp_v1->lcm_entries[i];
1625 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1627 lmm_size_vic += sizeof(*entry);
1628 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1629 } else if (count > 0) {
1630 /* find the specified mirror */
1638 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1642 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1643 if (!buf_vic->lb_buf) {
1648 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1649 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1651 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1652 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1653 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1654 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1655 if (!comp_rem->lcm_mirror_count)
1656 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1658 memset(comp_vic, 0, sizeof(*comp_v1));
1659 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1660 comp_vic->lcm_mirror_count = 0;
1661 comp_vic->lcm_entry_count = cpu_to_le32(count);
1662 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1663 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1664 comp_vic->lcm_layout_gen = 0;
1666 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1667 offset_rem = sizeof(*comp_rem) +
1668 sizeof(*entry_rem) * (comp_cnt - count);
1669 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1670 for (i = j = k = 0; i < comp_cnt; i++) {
1671 struct lov_mds_md *lmm, *lmm_dst;
1674 entry = &comp_v1->lcm_entries[i];
1675 entry_vic = &comp_vic->lcm_entries[j];
1676 entry_rem = &comp_rem->lcm_entries[k];
1678 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1681 /* copy component entry */
1683 memcpy(entry_vic, entry, sizeof(*entry));
1684 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1685 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1688 memcpy(entry_rem, entry, sizeof(*entry));
1689 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1693 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1695 lmm_dst = (struct lov_mds_md *)
1696 ((char *)comp_vic + offset_vic);
1698 lmm_dst = (struct lov_mds_md *)
1699 ((char *)comp_rem + offset_rem);
1701 /* copy component entry blob */
1702 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1704 /* blob offset advance */
1705 offset += le32_to_cpu(entry->lcme_size);
1707 offset_vic += le32_to_cpu(entry->lcme_size);
1709 offset_rem += le32_to_cpu(entry->lcme_size);
1715 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1716 struct md_rejig_data *mrd)
1718 struct mdd_device *mdd = mdo2mdd(md_obj);
1719 struct mdd_object *obj = md2mdd_obj(md_obj);
1720 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1721 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1722 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1723 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1724 struct lov_comp_md_v1 *lcm;
1725 struct thandle *handle;
1729 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1730 if (rc == 0) /* same fid */
1733 handle = mdd_trans_create(env, mdd);
1735 RETURN(PTR_ERR(handle));
1738 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1739 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1741 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1742 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1745 /* get EA of mirrored file */
1746 memset(buf_save, 0, sizeof(*buf));
1747 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1751 lcm = buf_save->lb_buf;
1752 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1753 GOTO(out, rc = -EINVAL);
1756 * Extract the mirror with specified mirror id, and store the splitted
1757 * mirror layout to the victim file.
1759 memset(buf, 0, sizeof(*buf));
1760 memset(buf_vic, 0, sizeof(*buf_vic));
1761 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1765 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1766 LU_XATTR_SPLIT, handle);
1769 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1770 LU_XATTR_SPLIT, handle);
1774 rc = mdd_trans_start(env, mdd, handle);
1778 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1783 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1786 GOTO(out_restore, rc);
1788 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1792 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1799 /* restore obj's layout */
1800 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1801 LU_XATTR_REPLACE, handle);
1803 CERROR("%s: failed to rollback of layout of: "DFID
1804 ": %d, file state unkonwn.\n",
1805 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1808 mdd_trans_stop(env, mdd, rc, handle);
1809 mdd_write_unlock(env, obj);
1810 mdd_write_unlock(env, vic);
1811 lu_buf_free(buf_save);
1813 lu_buf_free(buf_vic);
1816 (void) mdd_object_pfid_replace(env, obj);
1821 static int mdd_layout_merge_allowed(const struct lu_env *env,
1822 struct md_object *target,
1823 struct md_object *victim)
1825 struct mdd_object *o1 = md2mdd_obj(target);
1827 /* cannot extend directory's LOVEA */
1828 if (S_ISDIR(mdd_object_type(o1))) {
1829 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1830 mdd_obj_dev_name(o1));
1838 * The caller should guarantee to update the object ctime
1839 * after xattr_set if needed.
1841 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1842 const struct lu_buf *buf, const char *name,
1845 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1846 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1847 struct mdd_device *mdd = mdo2mdd(obj);
1848 struct thandle *handle;
1849 enum changelog_rec_type cl_type;
1850 enum changelog_rec_flags clf_flags = 0;
1854 rc = mdd_la_get(env, mdd_obj, attr);
1858 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1862 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1863 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1864 struct md_rejig_data *mrd = buf->lb_buf;
1865 struct md_object *victim = mrd->mrd_obj;
1867 if (buf->lb_len != sizeof(*mrd))
1870 rc = mdd_layout_merge_allowed(env, obj, victim);
1874 if (fl == LU_XATTR_MERGE)
1875 /* merge layout of victim as a mirror of obj's. */
1876 rc = mdd_xattr_merge(env, obj, victim);
1878 rc = mdd_xattr_split(env, obj, mrd);
1882 if (strcmp(name, XATTR_NAME_LMV) == 0) {
1883 rc = mdd_dir_layout_shrink(env, obj, buf);
1887 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1888 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1889 struct posix_acl *acl;
1891 /* user may set empty ACL, which should be treated as removing
1893 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1896 RETURN(PTR_ERR(acl));
1898 rc = mdd_xattr_del(env, obj, name);
1901 posix_acl_release(acl);
1904 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1905 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1909 handle = mdd_trans_create(env, mdd);
1911 RETURN(PTR_ERR(handle));
1913 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1917 rc = mdd_trans_start(env, mdd, handle);
1921 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1923 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1924 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1926 mdd_write_unlock(env, mdd_obj);
1931 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1932 mdd_write_unlock(env, mdd_obj);
1936 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1940 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1941 mdd_obj, name, handle);
1945 return mdd_trans_stop(env, mdd, rc, handle);
1948 static int mdd_declare_xattr_del(const struct lu_env *env,
1949 struct mdd_device *mdd,
1950 struct mdd_object *obj,
1952 struct thandle *handle)
1954 enum changelog_rec_type type;
1957 rc = mdo_declare_xattr_del(env, obj, name, handle);
1961 type = mdd_xattr_changelog_type(env, mdd, name);
1963 return 0; /* no changelog to store */
1965 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1969 * The caller should guarantee to update the object ctime
1970 * after xattr_set if needed.
1972 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1975 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1976 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1977 struct mdd_device *mdd = mdo2mdd(obj);
1978 struct thandle *handle;
1982 rc = mdd_la_get(env, mdd_obj, attr);
1986 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1990 handle = mdd_trans_create(env, mdd);
1992 RETURN(PTR_ERR(handle));
1994 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1998 rc = mdd_trans_start(env, mdd, handle);
2002 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2003 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2004 mdd_write_unlock(env, mdd_obj);
2008 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2011 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2016 return mdd_trans_stop(env, mdd, rc, handle);
2020 * read lov/lmv EA of an object
2021 * return the lov/lmv EA in an allocated lu_buf
2023 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2024 struct lu_buf *lmm_buf, const char *name)
2026 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2031 if (buf->lb_buf == NULL) {
2032 buf = lu_buf_check_and_alloc(buf, 4096);
2033 if (buf->lb_buf == NULL)
2038 rc = mdo_xattr_get(env, obj, buf, name);
2039 if (rc == -ERANGE) {
2040 /* mti_big_buf is allocated but is too small
2041 * we need to increase it */
2042 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2044 if (buf->lb_buf == NULL)
2047 } else if (rc < 0) {
2049 } else if (rc == 0) {
2053 lu_buf_alloc(lmm_buf, rc);
2054 if (lmm_buf->lb_buf == NULL)
2058 * we don't use lmm_buf directly, because we don't know xattr size, so
2059 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2061 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2066 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2067 struct mdd_object *o, struct lu_buf *buf,
2068 struct thandle *handle)
2070 struct hsm_attrs *attrs;
2071 enum hsm_states hsm_flags;
2072 enum changelog_rec_flags clf_flags = 0;
2076 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2081 attrs = buf->lb_buf;
2082 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2083 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2086 /* Add a changelog record for release. */
2087 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2088 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2089 clf_flags, o, handle);
2094 * check if layout swapping between 2 objects is allowed
2096 * - only normal FIDs or non-system IGIFs
2097 * - same type of objects
2098 * - same owner/group (so quotas are still valid) unless this is from HSM
2101 static int mdd_layout_swap_allowed(const struct lu_env *env,
2102 struct mdd_object *o1,
2103 const struct lu_attr *attr1,
2104 struct mdd_object *o2,
2105 const struct lu_attr *attr2,
2108 const struct lu_fid *fid1, *fid2;
2114 if (!fid_is_norm(fid1) &&
2115 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2118 if (!fid_is_norm(fid2) &&
2119 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2122 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2123 if (S_ISDIR(mdd_object_type(o1)))
2125 if (S_ISREG(mdd_object_type(o1)))
2130 if (flags & SWAP_LAYOUTS_MDS_HSM)
2133 if ((attr1->la_uid != attr2->la_uid) ||
2134 (attr1->la_gid != attr2->la_gid))
2140 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2141 * look into the layout in MDD layer. */
2142 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2144 struct lov_comp_md_v1 *comp_v1;
2145 struct lov_mds_md *v1;
2149 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2150 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2151 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2157 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2158 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2161 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2162 off = le32_to_cpu(comp_v1->lcm_entries[i].
2164 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2169 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2170 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2181 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2183 return mdd_lmm_oi(lmm, oi, true);
2186 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2188 return mdd_lmm_oi(lmm, oi, false);
2191 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2193 struct lov_comp_md_v1 *comp_v1;
2195 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2196 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2198 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2200 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2201 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2202 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2203 __u16 tmp_gen = *gen;
2205 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2207 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2214 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2216 return mdd_lmm_gen(lmm, gen, true);
2219 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2221 return mdd_lmm_gen(lmm, gen, false);
2224 static int mdd_dom_data_truncate(const struct lu_env *env,
2225 struct mdd_device *mdd, struct mdd_object *mo)
2228 struct dt_object *dom;
2231 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2233 GOTO(out, rc = -ENODATA);
2235 th = dt_trans_create(env, mdd->mdd_bottom);
2237 GOTO(out, rc = PTR_ERR(th));
2239 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2243 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2247 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2249 dt_trans_stop(env, mdd->mdd_bottom, th);
2251 /* Ignore failure but report the error */
2253 CERROR("%s: "DFID" can't truncate DOM inode data, rc = %d\n",
2254 mdd_obj_dev_name(mo), PFID(mdo2fid(mo)), rc);
2259 * swap layouts between 2 lustre objects
2261 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2262 struct md_object *obj2, __u64 flags)
2264 struct mdd_thread_info *info = mdd_env_info(env);
2265 struct mdd_object *fst_o = md2mdd_obj(obj1);
2266 struct mdd_object *snd_o = md2mdd_obj(obj2);
2267 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2268 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2269 struct mdd_device *mdd = mdo2mdd(obj1);
2270 struct lov_mds_md *fst_lmm, *snd_lmm;
2271 struct lu_buf *fst_buf = &info->mti_buf[0];
2272 struct lu_buf *snd_buf = &info->mti_buf[1];
2273 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2274 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2275 struct ost_id *saved_oi = NULL;
2276 struct thandle *handle;
2277 struct mdd_object *dom_o = NULL;
2278 __u64 domsize_dom, domsize_vlt;
2279 __u32 fst_gen, snd_gen, saved_gen;
2285 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
2286 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2288 /* we have to sort the 2 obj, so locking will always
2289 * be in the same order, even in case of 2 concurrent swaps */
2290 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
2291 if (rc == 0) /* same fid ? */
2297 rc = mdd_la_get(env, fst_o, fst_la);
2301 rc = mdd_la_get(env, snd_o, snd_la);
2305 /* check if layout swapping is allowed */
2306 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2310 handle = mdd_trans_create(env, mdd);
2312 RETURN(PTR_ERR(handle));
2314 /* objects are already sorted */
2315 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
2316 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
2318 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2319 if (rc < 0 && rc != -ENODATA)
2322 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2323 if (rc < 0 && rc != -ENODATA)
2326 /* check if file has DoM. DoM file can be migrated only to another
2327 * DoM layout with the same DoM component size or to an non-DOM
2328 * layout. After migration to OSTs layout, local MDT inode data
2329 * should be truncated.
2330 * Objects are sorted by FIDs, considering that original file's FID
2331 * is always smaller the snd_o is always original file we are migrating
2334 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2335 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2337 /* Only migration is supported for DoM files, not 'swap_layouts' so
2338 * target file must be volatile and orphan.
2340 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2341 dom_o = domsize_dom ? snd_o : NULL;
2342 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2343 swap(domsize_dom, domsize_vlt);
2344 dom_o = domsize_dom ? fst_o : NULL;
2345 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2346 /* 'lfs swap_layouts' case, neither file should have DoM */
2348 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2349 "use migration instead: rc = %d\n", rc);
2353 if (domsize_vlt > 0 && domsize_dom == 0) {
2355 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": OST to DOM "
2356 "migration is not supported: rc = %d\n",
2357 PFID(mdo2fid(snd_o)), rc);
2359 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2361 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": new layout "
2362 "must have the same DoM component size: rc = %d\n",
2363 PFID(mdo2fid(fst_o)), rc);
2365 } else if (domsize_vlt > 0) {
2366 /* Migration with the same DOM component size, no need to
2367 * truncate local data, it is still being used */
2371 /* swapping 2 non existant layouts is a success */
2372 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2375 /* to help inode migration between MDT, it is better to
2376 * start by the no layout file (if one), so we order the swap */
2377 if (snd_buf->lb_buf == NULL) {
2379 swap(fst_buf, snd_buf);
2382 fst_gen = snd_gen = 0;
2383 /* lmm and generation layout initialization */
2384 if (fst_buf->lb_buf != NULL) {
2385 fst_lmm = fst_buf->lb_buf;
2386 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2387 fst_fl = LU_XATTR_REPLACE;
2391 fst_fl = LU_XATTR_CREATE;
2394 snd_lmm = snd_buf->lb_buf;
2395 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2397 saved_gen = fst_gen;
2398 /* increase the generation layout numbers */
2403 * XXX The layout generation is used to generate component IDs for
2404 * the composite file, we have to do some special tweaks to make
2405 * sure the layout generation is always adequate for that job.
2408 /* Skip invalid generation number for composite layout */
2409 if ((snd_gen & LCME_ID_MASK) == 0)
2411 if ((fst_gen & LCME_ID_MASK) == 0)
2413 /* Make sure the generation is greater than all the component IDs */
2414 if (fst_gen < snd_gen)
2416 else if (fst_gen > snd_gen)
2419 /* set the file specific informations in lmm */
2420 if (fst_lmm != NULL) {
2421 saved_oi = &info->mti_oa.o_oi;
2422 mdd_get_lmm_oi(fst_lmm, saved_oi);
2423 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2424 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2425 mdd_set_lmm_oi(snd_lmm, saved_oi);
2427 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2428 cpu_to_le32(LOV_MAGIC_MAGIC))
2429 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2431 GOTO(stop, rc = -EPROTO);
2433 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2435 /* Prepare HSM attribute if it's required */
2436 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2437 const int buflen = sizeof(struct hsm_attrs);
2439 lu_buf_alloc(fst_hsm_buf, buflen);
2440 lu_buf_alloc(snd_hsm_buf, buflen);
2441 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2442 GOTO(stop, rc = -ENOMEM);
2444 /* Read HSM attribute */
2445 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2449 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2453 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2454 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2459 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2460 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2466 /* prepare transaction */
2467 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2472 if (fst_buf->lb_buf != NULL)
2473 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2474 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2477 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2482 rc = mdd_trans_start(env, mdd, handle);
2486 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2487 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2491 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2493 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2496 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
2497 mdd_obj_dev_name(fst_o),
2498 PFID(mdo2fid(fst_o)), rc, rc2);
2503 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2507 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2510 if (fst_buf->lb_buf != NULL)
2511 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2512 LU_XATTR_REPLACE, handle);
2514 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2517 GOTO(out_restore, rc);
2519 /* Issue one changelog record per file */
2520 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2524 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2533 /* failure on second file, but first was done, so we have
2534 * to roll back first. */
2535 if (fst_buf->lb_buf != NULL) {
2536 mdd_set_lmm_oi(fst_lmm, saved_oi);
2537 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2538 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2539 LU_XATTR_REPLACE, handle);
2541 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2547 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2552 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2557 CERROR("%s: unable to roll back layout swap. FIDs: "
2558 DFID" and "DFID "error: %d/%d, steps: %d\n",
2559 mdd_obj_dev_name(fst_o),
2560 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
2562 /* a solution to avoid journal commit is to panic,
2563 * but it has strong consequences so we use LBUG to
2564 * allow sysdamin to choose to panic or not
2571 rc = mdd_trans_stop(env, mdd, rc, handle);
2573 /* Truncate local DOM data if all went well */
2575 mdd_dom_data_truncate(env, mdd, dom_o);
2577 mdd_write_unlock(env, snd_o);
2578 mdd_write_unlock(env, fst_o);
2580 lu_buf_free(fst_buf);
2581 lu_buf_free(snd_buf);
2582 lu_buf_free(fst_hsm_buf);
2583 lu_buf_free(snd_hsm_buf);
2586 (void) mdd_object_pfid_replace(env, fst_o);
2587 (void) mdd_object_pfid_replace(env, snd_o);
2592 static int mdd_declare_layout_change(const struct lu_env *env,
2593 struct mdd_device *mdd,
2594 struct mdd_object *obj,
2595 struct md_layout_change *mlc,
2596 struct thandle *handle)
2600 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2604 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2608 /* For PFL, this is used to instantiate necessary component objects. */
2610 mdd_layout_instantiate_component(const struct lu_env *env,
2611 struct mdd_object *obj, struct md_layout_change *mlc,
2612 struct thandle *handle)
2614 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2618 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2621 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2623 * It's possible that another layout write intent has already
2624 * instantiated our objects, so a -EALREADY returned, and we need to
2628 RETURN(rc == -EALREADY ? 0 : rc);
2630 rc = mdd_trans_start(env, mdd, handle);
2634 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2635 rc = mdo_layout_change(env, obj, mlc, handle);
2636 mdd_write_unlock(env, obj);
2640 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2645 * Change the FLR layout from RDONLY to WRITE_PENDING.
2647 * It picks the primary mirror, and bumps the layout version, and set
2648 * layout version xattr to OST objects in a sync tx. In order to facilitate
2649 * the handling of phantom writers from evicted clients, the clients carry
2650 * layout version of the file with write RPC, so that the OSTs can verify
2651 * if the write RPCs are legitimate, meaning not from evicted clients.
2654 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2655 struct md_layout_change *mlc, struct thandle *handle)
2657 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2658 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2659 struct lustre_som_attrs *som = &mlc->mlc_som;
2664 /* Verify acceptable operations */
2665 switch (mlc->mlc_opc) {
2666 case MD_LAYOUT_WRITE:
2667 case MD_LAYOUT_RESYNC:
2668 /* these are legal operations - this represents the case that
2669 * a few mirrors were missed in the last resync. */
2671 case MD_LAYOUT_RESYNC_DONE:
2676 som_buf->lb_buf = som;
2677 som_buf->lb_len = sizeof(*som);
2678 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2679 if (rc < 0 && rc != -ENODATA)
2683 lustre_som_swab(som);
2684 if (som->lsa_valid & SOM_FL_STRICT)
2685 fl = LU_XATTR_REPLACE;
2688 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2693 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2694 XATTR_NAME_SOM, fl, handle);
2699 /* record a changelog for data mover to consume */
2700 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2704 rc = mdd_trans_start(env, mdd, handle);
2708 /* it needs a sync tx to make FLR to work properly */
2709 handle->th_sync = 1;
2711 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2712 rc = mdo_layout_change(env, obj, mlc, handle);
2714 /* SOM state transition from STRICT to STALE */
2715 som->lsa_valid = SOM_FL_STALE;
2716 lustre_som_swab(som);
2717 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2720 mdd_write_unlock(env, obj);
2724 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2735 * Handle mirrored file state transition when it's in WRITE_PENDING.
2737 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2738 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2739 * layout version will be increased, and the file's state will be changed to
2743 mdd_layout_update_write_pending(const struct lu_env *env,
2744 struct mdd_object *obj, struct md_layout_change *mlc,
2745 struct thandle *handle)
2747 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2751 switch (mlc->mlc_opc) {
2752 case MD_LAYOUT_RESYNC:
2753 /* Upon receiving the resync request, it should
2754 * instantiate all stale components right away to get ready
2755 * for mirror copy. In order to avoid layout version change,
2756 * client should avoid sending LAYOUT_WRITE request at the
2759 case MD_LAYOUT_WRITE:
2760 /* legal race for concurrent write, the file state has been
2761 * changed by another client. */
2767 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2771 rc = mdd_trans_start(env, mdd, handle);
2775 /* it needs a sync tx to make FLR to work properly */
2776 handle->th_sync = 1;
2778 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2779 rc = mdo_layout_change(env, obj, mlc, handle);
2780 mdd_write_unlock(env, obj);
2791 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2793 * Only concurrent write and sync complete requests are possible when the
2794 * file is in SYNC_PENDING. For the latter request, it will pass in the
2795 * mirrors that have been synchronized, then the stale bit will be cleared
2796 * to make the file's state turn into RDONLY.
2797 * For concurrent write reqeust, it just needs to change the file's state
2798 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2799 * version because the version will be increased in the transition to
2800 * SYNC_PENDING later so that it can deny the write request from potential
2801 * evicted SYNC clients. */
2803 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2804 struct md_layout_change *mlc, struct thandle *handle)
2806 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2807 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2812 /* operation validation */
2813 switch (mlc->mlc_opc) {
2814 case MD_LAYOUT_RESYNC_DONE:
2815 /* resync complete. */
2816 case MD_LAYOUT_WRITE:
2817 /* concurrent write. */
2819 case MD_LAYOUT_RESYNC:
2820 /* resync again, most likely the previous run failed.
2821 * no-op if it's already in SYNC_PENDING state */
2827 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2828 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2829 if (rc < 0 && rc != -ENODATA)
2832 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2833 lustre_som_swab(&mlc->mlc_som);
2834 som_buf->lb_buf = &mlc->mlc_som;
2835 som_buf->lb_len = sizeof(mlc->mlc_som);
2838 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2842 /* record a changelog for the completion of resync */
2843 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2848 /* RESYNC_DONE has piggybacked size and blocks */
2850 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2851 XATTR_NAME_SOM, fl, handle);
2856 rc = mdd_trans_start(env, mdd, handle);
2860 /* it needs a sync tx to make FLR to work properly */
2861 handle->th_sync = 1;
2863 rc = mdo_layout_change(env, obj, mlc, handle);
2868 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2874 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2883 * Layout change callback for object.
2885 * This is only used by FLR for now. In the future, it can be exteneded to
2886 * handle all layout change.
2889 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2890 struct md_layout_change *mlc)
2892 struct mdd_object *obj = md2mdd_obj(o);
2893 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2894 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2895 struct lov_comp_md_v1 *lcm;
2896 struct thandle *handle;
2901 /* Verify acceptable operations */
2902 switch (mlc->mlc_opc) {
2903 case MD_LAYOUT_WRITE:
2904 case MD_LAYOUT_RESYNC:
2905 case MD_LAYOUT_RESYNC_DONE:
2911 handle = mdd_trans_create(env, mdd);
2913 RETURN(PTR_ERR(handle));
2915 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2922 /* analyze the layout to make sure it's a FLR file */
2924 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2925 GOTO(out, rc = -EINVAL);
2927 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2929 /* please refer to HLD of FLR for state transition */
2930 switch (flr_state) {
2932 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2934 case LCM_FL_WRITE_PENDING:
2935 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2938 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2940 case LCM_FL_SYNC_PENDING:
2941 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2950 mdd_trans_stop(env, mdd, rc, handle);
2955 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2956 struct mdd_object *child, const struct lu_attr *attr,
2957 const struct md_op_spec *spec,
2958 struct dt_allocation_hint *hint)
2960 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2961 struct dt_object *nc = mdd_object_child(child);
2963 memset(hint, 0, sizeof(*hint));
2965 /* For striped directory, give striping EA to lod_ah_init, which will
2966 * decide the stripe_offset and stripe count by it. */
2967 if (S_ISDIR(attr->la_mode) &&
2968 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2969 hint->dah_eadata = spec->u.sp_ea.eadata;
2970 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2972 hint->dah_eadata = NULL;
2973 hint->dah_eadata_len = 0;
2976 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2977 hint->dah_eadata, hint->dah_eadata_len);
2978 /* @hint will be initialized by underlying device. */
2979 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2982 static int accmode(const struct lu_env *env, const struct lu_attr *la,
2985 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2986 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2987 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2988 * owner can write to a file even if it is marked readonly to hide
2989 * its brokenness. (bug 5781) */
2990 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
2991 struct lu_ucred *uc = lu_ucred_check(env);
2993 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2997 return mds_accmode(open_flags);
3000 static int mdd_open_sanity_check(const struct lu_env *env,
3001 struct mdd_object *obj,
3002 const struct lu_attr *attr, u64 open_flags)
3007 /* EEXIST check, also opening of *open* orphans is allowed so we can
3008 * open-by-handle unlinked files
3010 if (mdd_is_dead_obj(obj) &&
3011 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3014 if (S_ISLNK(attr->la_mode))
3017 mode = accmode(env, attr, open_flags);
3019 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3022 if (!(open_flags & MDS_OPEN_CREATED)) {
3023 rc = mdd_permission_internal(env, obj, attr, mode);
3028 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3029 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3030 open_flags &= ~MDS_OPEN_TRUNC;
3032 /* For writing append-only file must open it with append mode. */
3033 if (attr->la_flags & LUSTRE_APPEND_FL) {
3034 if ((open_flags & MDS_FMODE_WRITE) &&
3035 !(open_flags & MDS_OPEN_APPEND))
3037 if (open_flags & MDS_OPEN_TRUNC)
3044 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3047 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3048 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3049 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3050 struct mdd_object_user *mou = NULL;
3051 const struct lu_ucred *uc = lu_ucred(env);
3052 struct mdd_device *mdd = mdo2mdd(obj);
3053 enum changelog_rec_type type = CL_OPEN;
3056 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3058 rc = mdd_la_get(env, mdd_obj, attr);
3062 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags);
3063 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3068 mdd_obj->mod_count++;
3070 if (!mdd_changelog_enabled(env, mdd, type))
3074 /* look for existing opener in list under mdd_write_lock */
3075 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3080 /* add user to list */
3081 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3087 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3089 mdd_obj_user_free(mou);
3094 if (type == CL_DN_OPEN) {
3095 if (ktime_before(ktime_get(), mou->mou_deniednext))
3096 /* same user denied again same access within
3097 * time interval: do not record
3101 /* this user already denied, but some time ago:
3102 * update denied time
3104 mou->mou_deniednext =
3105 ktime_add(ktime_get(),
3106 ktime_set(mdd->mdd_cl.mc_deniednext,
3109 mou->mou_opencount++;
3110 /* same user opening file again with same flags:
3117 /* FYI, only the bottom 32 bits of open_flags are recorded */
3118 mdd_changelog(env, type, open_flags, md_dev, mdo2fid(mdd_obj));
3122 mdd_write_unlock(env, mdd_obj);
3126 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3127 struct md_attr *ma, struct thandle *handle)
3131 rc = mdd_orphan_declare_delete(env, obj, handle);
3135 return mdo_declare_destroy(env, obj, handle);
3139 * No permission check is needed.
3141 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3142 struct md_attr *ma, u64 open_flags)
3144 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3145 struct mdd_device *mdd = mdo2mdd(obj);
3146 struct thandle *handle = NULL;
3149 bool blocked = false;
3150 bool last_close_by_uid = false;
3151 const struct lu_ucred *uc = lu_ucred(env);
3154 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3155 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3156 mdd_obj->mod_count--;
3157 mdd_write_unlock(env, mdd_obj);
3159 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3160 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3161 "list\n", PFID(mdd_object_fid(mdd_obj)));
3165 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3166 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3167 /* check without any lock */
3168 is_orphan = mdd_obj->mod_count == 1 &&
3169 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3173 /* mdd_trans_create() maybe failed because of barrier_entry(),
3174 * under such case, the orphan MDT-object will be left in the
3175 * orphan list, and when the MDT remount next time, the unused
3176 * orphans will be destroyed automatically.
3178 * One exception: the former mdd_finish_unlink may failed to
3179 * add the orphan MDT-object to the orphan list, then if the
3180 * mdd_trans_create() failed because of barrier_entry(), the
3181 * MDT-object will become real orphan that is neither in the
3182 * namespace nor in the orphan list. Such bad case should be
3183 * very rare and will be handled by e2fsck/lfsck. */
3184 handle = mdd_trans_create(env, mdo2mdd(obj));
3185 if (IS_ERR(handle)) {
3186 rc = PTR_ERR(handle);
3187 if (rc != -EINPROGRESS)
3195 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3199 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3204 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3210 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3211 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3213 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3214 lu_dev_name(mdd2lu_dev(mdd)),
3215 PFID(mdd_object_fid(mdd_obj)), rc);
3219 /* check again with lock */
3220 is_orphan = (mdd_obj->mod_count == 1) &&
3221 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3222 ma->ma_attr.la_nlink == 0);
3224 if (is_orphan && !handle && !blocked) {
3225 mdd_write_unlock(env, mdd_obj);
3229 mdd_obj->mod_count--; /*release open count */
3231 /* under mdd write lock */
3232 /* If recording, see if we need to remove UID from list. uc is not
3233 * initialized if the client has been evicted. */
3234 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3235 struct mdd_object_user *mou;
3237 /* look for UID in list */
3238 /* If mou is NULL, it probably means logging was enabled after
3239 * the user had the file open. So the corresponding close
3240 * will not be logged.
3242 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3245 mou->mou_opencount--;
3246 if (mou->mou_opencount == 0) {
3247 mdd_obj_user_remove(mdd_obj, mou);
3248 last_close_by_uid = true;
3253 if (!is_orphan || blocked)
3257 /* NB: Object maybe not in orphan list originally, it is rare case for
3258 * mdd_finish_unlink() failure, in that case, the object doesn't have
3259 * ORPHAN_OBJ flag */
3260 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3261 /* remove link to object from orphan index */
3262 LASSERT(handle != NULL);
3263 rc = mdd_orphan_delete(env, mdd_obj, handle);
3265 CERROR("%s: unable to delete "DFID" from orphan list: "
3266 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3267 PFID(mdd_object_fid(mdd_obj)), rc);
3268 /* If object was not deleted from orphan list, do not
3269 * destroy OSS objects, which will be done when next
3274 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3275 "list, OSS objects to be destroyed.\n",
3276 PFID(mdd_object_fid(mdd_obj)));
3279 rc = mdo_destroy(env, mdd_obj, handle);
3282 CERROR("%s: unable to delete "DFID" from orphan list: "
3283 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3284 PFID(mdd_object_fid(mdd_obj)), rc);
3289 mdd_write_unlock(env, mdd_obj);
3291 if (rc != 0 || blocked ||
3292 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3295 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3296 * or if CL_OPEN was recorded and it's last close by user.
3297 * Changelogs mask may change between open and close operations, but
3298 * this is not a big deal if we have a CL_CLOSE entry with no matching
3299 * CL_OPEN. Plus Changelogs mask may not change often.
3301 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3302 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3303 MDS_OPEN_TRUNC))) ||
3304 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3305 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3306 if (handle == NULL) {
3307 handle = mdd_trans_create(env, mdo2mdd(obj));
3309 GOTO(stop, rc = PTR_ERR(handle));
3311 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3312 NULL, NULL, handle);
3316 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3321 /* FYI, only the bottom 32 bits of open_flags are recorded */
3322 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3327 if (handle != NULL && !IS_ERR(handle))
3328 rc = mdd_trans_stop(env, mdd, rc, handle);
3334 * Permission check is done when open,
3335 * no need check again.
3337 static int mdd_readpage_sanity_check(const struct lu_env *env,
3338 struct mdd_object *obj)
3340 struct dt_object *next = mdd_object_child(obj);
3344 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3352 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3353 size_t nob, const struct dt_it_ops *iops,
3354 struct dt_it *it, __u32 attr, void *arg)
3356 struct lu_dirpage *dp = &lp->lp_dir;
3360 struct lu_dirent *ent;
3361 struct lu_dirent *last = NULL;
3365 if (nob < sizeof(*dp))
3368 memset(area, 0, sizeof (*dp));
3369 area += sizeof (*dp);
3370 nob -= sizeof (*dp);
3377 len = iops->key_size(env, it);
3379 /* IAM iterator can return record with zero len. */
3383 hash = iops->store(env, it);
3384 if (unlikely(first)) {
3386 dp->ldp_hash_start = cpu_to_le64(hash);
3389 /* calculate max space required for lu_dirent */
3390 recsize = lu_dirent_calc_size(len, attr);
3392 if (nob >= recsize) {
3393 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3394 if (result == -ESTALE)
3399 /* osd might not able to pack all attributes,
3400 * so recheck rec length */
3401 recsize = le16_to_cpu(ent->lde_reclen);
3403 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3404 fid_le_to_cpu(&fid, &ent->lde_fid);
3405 if (fid_is_dot_lustre(&fid))
3409 result = (last != NULL) ? 0 :-EINVAL;
3413 ent = (void *)ent + recsize;
3417 result = iops->next(env, it);
3418 if (result == -ESTALE)
3420 } while (result == 0);
3423 dp->ldp_hash_end = cpu_to_le64(hash);
3425 if (last->lde_hash == dp->ldp_hash_end)
3426 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3427 last->lde_reclen = 0; /* end mark */
3430 /* end of directory */
3431 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3432 else if (result < 0)
3433 CWARN("build page failed: %d!\n", result);
3437 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3438 const struct lu_rdpg *rdpg)
3440 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3444 if (mdd_object_exists(mdd_obj) == 0) {
3445 CERROR("%s: object "DFID" not found: rc = -2\n",
3446 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3450 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
3451 rc = mdd_readpage_sanity_check(env, mdd_obj);
3453 GOTO(out_unlock, rc);
3455 if (mdd_is_dead_obj(mdd_obj)) {
3457 struct lu_dirpage *dp;
3460 * According to POSIX, please do not return any entry to client:
3461 * even dot and dotdot should not be returned.
3463 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3464 PFID(mdd_object_fid(mdd_obj)));
3466 if (rdpg->rp_count <= 0)
3467 GOTO(out_unlock, rc = -EFAULT);
3468 LASSERT(rdpg->rp_pages != NULL);
3470 pg = rdpg->rp_pages[0];
3471 dp = (struct lu_dirpage *)kmap(pg);
3472 memset(dp, 0 , sizeof(struct lu_dirpage));
3473 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3474 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3475 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3477 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3480 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3481 mdd_dir_page_build, NULL);
3483 struct lu_dirpage *dp;
3485 dp = kmap(rdpg->rp_pages[0]);
3486 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3489 * No pages were processed, mark this for first page
3492 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3493 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3494 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3496 kunmap(rdpg->rp_pages[0]);
3499 GOTO(out_unlock, rc);
3501 mdd_read_unlock(env, mdd_obj);
3505 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3507 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3509 if (mdd_object_exists(mdd_obj) == 0) {
3512 CERROR("%s: object "DFID" not found: rc = %d\n",
3513 mdd_obj_dev_name(mdd_obj),
3514 PFID(mdd_object_fid(mdd_obj)), rc);
3517 return dt_object_sync(env, mdd_object_child(mdd_obj),
3521 static int mdd_object_lock(const struct lu_env *env,
3522 struct md_object *obj,
3523 struct lustre_handle *lh,
3524 struct ldlm_enqueue_info *einfo,
3525 union ldlm_policy_data *policy)
3527 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3528 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3532 static int mdd_object_unlock(const struct lu_env *env,
3533 struct md_object *obj,
3534 struct ldlm_enqueue_info *einfo,
3535 union ldlm_policy_data *policy)
3537 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3538 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3541 const struct md_object_operations mdd_obj_ops = {
3542 .moo_permission = mdd_permission,
3543 .moo_attr_get = mdd_attr_get,
3544 .moo_attr_set = mdd_attr_set,
3545 .moo_xattr_get = mdd_xattr_get,
3546 .moo_xattr_set = mdd_xattr_set,
3547 .moo_xattr_list = mdd_xattr_list,
3548 .moo_invalidate = mdd_invalidate,
3549 .moo_xattr_del = mdd_xattr_del,
3550 .moo_swap_layouts = mdd_swap_layouts,
3551 .moo_open = mdd_open,
3552 .moo_close = mdd_close,
3553 .moo_readpage = mdd_readpage,
3554 .moo_readlink = mdd_readlink,
3555 .moo_changelog = mdd_changelog,
3556 .moo_object_sync = mdd_object_sync,
3557 .moo_object_lock = mdd_object_lock,
3558 .moo_object_unlock = mdd_object_unlock,
3559 .moo_layout_change = mdd_layout_change,