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 int flags, const struct lu_fid *fid,
74 const char *xattr_name,
75 struct thandle *handle);
77 static inline bool has_prefix(const char *str, const char *prefix);
80 static u32 flags_helper(u64 open_flags)
84 if (open_flags & MDS_FMODE_EXEC) {
85 open_mode = MDS_FMODE_EXEC;
87 if (open_flags & MDS_FMODE_READ)
88 open_mode = MDS_FMODE_READ;
90 (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
91 open_mode |= MDS_FMODE_WRITE;
97 /** Allocate/init a user and its sub-structures.
102 * \retval mou [OUT] success valid structure
105 static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags,
106 uid_t uid, gid_t gid)
108 struct mdd_object_user *mou;
112 OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem);
114 RETURN(ERR_PTR(-ENOMEM));
116 mou->mou_open_flags = open_flags;
117 mou->mou_uidgid = ((__u64)uid << 32) | gid;
118 mou->mou_opencount = 0;
119 mou->mou_deniednext = ktime_set(0, 0);
125 * Free a user and its sub-structures.
127 * \param mou [IN] user to be freed.
129 static void mdd_obj_user_free(struct mdd_object_user *mou)
131 OBD_SLAB_FREE_PTR(mou, mdd_object_kmem);
135 * Find if UID/GID already has this file open
137 * Caller should have write-locked \param mdd_obj.
138 * \param mdd_obj [IN] mdd_obj
139 * \param uid [IN] client uid
140 * \param gid [IN] client gid
141 * \retval user pointer or NULL if not found
144 struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj,
145 uid_t uid, gid_t gid,
148 struct mdd_object_user *mou;
153 uidgid = ((__u64)uid << 32) | gid;
154 list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) {
155 if (mou->mou_uidgid == uidgid &&
156 flags_helper(mou->mou_open_flags) ==
157 flags_helper(open_flags))
164 * Add a user to the list of openers for this file
166 * Caller should have write-locked \param mdd_obj.
167 * \param mdd_obj [IN] mdd_obj
168 * \param mou [IN] user
170 * \retval -ve failure
172 static int mdd_obj_user_add(struct mdd_object *mdd_obj,
173 struct mdd_object_user *mou,
176 struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj);
177 struct mdd_object_user *tmp;
178 __u32 uid = mou->mou_uidgid >> 32;
179 __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 &&
239 la->la_flags & LUSTRE_ORPHAN_FL)
240 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
245 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
247 struct mdd_thread_info *info;
249 lu_env_refill((struct lu_env *)env);
250 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
251 LASSERT(info != NULL);
255 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
259 buf = &mdd_env_info(env)->mti_buf[0];
265 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
266 const void *area, ssize_t len)
270 buf = &mdd_env_info(env)->mti_buf[0];
271 buf->lb_buf = (void *)area;
276 struct lu_object *mdd_object_alloc(const struct lu_env *env,
277 const struct lu_object_header *hdr,
280 struct mdd_object *mdd_obj;
283 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
287 o = mdd2lu_obj(mdd_obj);
288 lu_object_init(o, NULL, d);
289 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
290 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
291 mdd_obj->mod_count = 0;
292 o->lo_ops = &mdd_lu_obj_ops;
293 INIT_LIST_HEAD(&mdd_obj->mod_users);
298 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
299 const struct lu_object_conf *unused)
301 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
302 struct mdd_object *mdd_obj = lu2mdd_obj(o);
303 struct lu_object *below;
304 struct lu_device *under;
307 mdd_obj->mod_cltime = ktime_set(0, 0);
308 under = &d->mdd_child->dd_lu_dev;
309 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
311 RETURN(PTR_ERR(below));
313 lu_object_add(o, below);
318 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
322 if (lu_object_exists(o)) {
323 struct mdd_object *mdd_obj = lu2mdd_obj(o);
324 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
326 rc = mdd_la_get(env, mdd_obj, attr);
332 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
334 struct mdd_object *mdd = lu2mdd_obj(o);
335 struct mdd_object_user *mou, *tmp2;
338 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
339 list_del(&mou->mou_list);
340 mdd_obj_user_free(mou);
344 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
347 static int mdd_object_print(const struct lu_env *env, void *cookie,
348 lu_printer_t p, const struct lu_object *o)
350 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
352 return (*p)(env, cookie,
353 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
354 mdd, mdd->mod_count, mdd->mod_valid,
355 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
358 static const struct lu_object_operations mdd_lu_obj_ops = {
359 .loo_object_init = mdd_object_init,
360 .loo_object_start = mdd_object_start,
361 .loo_object_free = mdd_object_free,
362 .loo_object_print = mdd_object_print,
365 struct mdd_object *mdd_object_find(const struct lu_env *env,
366 struct mdd_device *d,
367 const struct lu_fid *f)
369 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
373 * No permission check is needed.
375 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
378 struct mdd_object *mdd_obj = md2mdd_obj(obj);
383 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
384 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
385 ma->ma_attr.la_nlink = 0;
391 * No permission check is needed.
393 static int mdd_xattr_get(const struct lu_env *env,
394 struct md_object *obj, struct lu_buf *buf,
397 struct mdd_object *mdd_obj = md2mdd_obj(obj);
398 struct mdd_device *mdd;
403 if (mdd_object_exists(mdd_obj) == 0) {
404 CERROR("%s: object "DFID" not found: rc = -2\n",
405 mdd_obj_dev_name(mdd_obj),
406 PFID(mdd_object_fid(mdd_obj)));
410 /* If the object has been destroyed, then do not get LMVEA, because
411 * it needs to load stripes from the iteration of the master object,
412 * and it will cause problem if master object has been destroyed, see
414 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
415 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
416 strcmp(name, XATTR_NAME_LMV) == 0))
419 /* If the object has been delete from the namespace, then
420 * get linkEA should return -ENOENT as well */
421 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
422 strcmp(name, XATTR_NAME_LINK) == 0))
425 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
426 rc = mdo_xattr_get(env, mdd_obj, buf, name);
427 mdd_read_unlock(env, mdd_obj);
431 /* record only getting user xattrs and acls */
433 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
434 (has_prefix(name, XATTR_USER_PREFIX) ||
435 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
436 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
437 struct thandle *handle;
440 LASSERT(mdo2fid(mdd_obj) != NULL);
442 handle = mdd_trans_create(env, mdd);
444 RETURN(PTR_ERR(handle));
446 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
451 rc2 = mdd_trans_start(env, mdd, handle);
455 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
456 mdo2fid(mdd_obj), name,
460 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
469 * Permission check is done when open,
470 * no need check again.
472 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
475 struct mdd_object *mdd_obj = md2mdd_obj(obj);
476 struct dt_object *next;
481 if (mdd_object_exists(mdd_obj) == 0) {
482 CERROR("%s: object "DFID" not found: rc = -2\n",
483 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
487 next = mdd_object_child(mdd_obj);
488 LASSERT(next != NULL);
489 LASSERT(next->do_body_ops != NULL);
490 LASSERT(next->do_body_ops->dbo_read != NULL);
491 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
492 rc = dt_read(env, next, buf, &pos);
493 mdd_read_unlock(env, mdd_obj);
498 * No permission check is needed.
500 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
503 struct mdd_object *mdd_obj = md2mdd_obj(obj);
508 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
509 rc = mdo_xattr_list(env, mdd_obj, buf);
510 mdd_read_unlock(env, mdd_obj);
512 /* If the buffer is NULL then we are only here to get the
513 * length of the xattr name list. */
514 if (rc < 0 || buf->lb_buf == NULL)
518 * Filter out XATTR_NAME_LINK if this is an orphan object. See
521 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
522 char *end = (char *)buf->lb_buf + rc;
523 char *p = buf->lb_buf;
526 char *next = p + strlen(p) + 1;
528 if (strcmp(p, XATTR_NAME_LINK) == 0) {
530 memmove(p, next, end - next);
532 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
533 " of orphan "DFID"\n",
534 PFID(mdd_object_fid(mdd_obj)));
545 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
547 return mdo_invalidate(env, md2mdd_obj(obj));
550 int mdd_declare_create_object_internal(const struct lu_env *env,
551 struct mdd_object *p,
552 struct mdd_object *c,
553 struct lu_attr *attr,
554 struct thandle *handle,
555 const struct md_op_spec *spec,
556 struct dt_allocation_hint *hint)
558 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
559 const struct dt_index_features *feat = spec->sp_feat;
563 if (feat != &dt_directory_features && feat != NULL) {
564 dof->dof_type = DFT_INDEX;
565 dof->u.dof_idx.di_feat = feat;
567 dof->dof_type = dt_mode_to_dft(attr->la_mode);
568 if (dof->dof_type == DFT_REGULAR) {
569 dof->u.dof_reg.striped =
570 md_should_create(spec->sp_cr_flags);
571 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
572 dof->u.dof_reg.striped = 0;
573 /* is this replay? */
575 dof->u.dof_reg.striped = 0;
579 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
584 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
585 struct mdd_object *c, struct lu_attr *attr,
586 struct thandle *handle,
587 const struct md_op_spec *spec,
588 struct dt_allocation_hint *hint)
590 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
594 LASSERT(!mdd_object_exists(c));
596 rc = mdo_create_object(env, c, attr, hint, dof, handle);
601 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
602 const struct lu_attr *attr, struct thandle *handle,
608 rc = mdo_attr_set(env, obj, attr, handle);
609 #ifdef CONFIG_FS_POSIX_ACL
610 if (!rc && (attr->la_valid & LA_MODE) && needacl)
611 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
616 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
617 const struct lu_attr *oattr, struct lu_attr *attr,
618 struct thandle *handle)
623 LASSERT(attr->la_valid & LA_CTIME);
624 LASSERT(oattr != NULL);
626 /* Make sure the ctime is increased only, however, it's not strictly
627 * reliable at here because there is not guarantee to hold lock on
628 * object, so we just bypass some unnecessary cmtime setting first
629 * and OSD has to check it again. */
630 if (attr->la_ctime < oattr->la_ctime)
631 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
632 else if (attr->la_valid == LA_CTIME &&
633 attr->la_ctime == oattr->la_ctime)
634 attr->la_valid &= ~LA_CTIME;
636 if (attr->la_valid != 0)
637 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
642 * This gives the same functionality as the code between
643 * sys_chmod and inode_setattr
644 * chown_common and inode_setattr
645 * utimes and inode_setattr
646 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
648 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
649 const struct lu_attr *oattr, struct lu_attr *la,
650 const unsigned long flags)
659 /* Do not permit change file type */
660 if (la->la_valid & LA_TYPE)
663 /* They should not be processed by setattr */
664 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
667 LASSERT(oattr != NULL);
669 uc = lu_ucred_check(env);
673 if (la->la_valid == LA_CTIME) {
674 if (!(flags & MDS_PERM_BYPASS))
675 /* This is only for set ctime when rename's source is
677 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
679 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
680 la->la_valid &= ~LA_CTIME;
684 if (la->la_valid == LA_ATIME) {
685 /* This is an atime-only attribute update for close RPCs. */
686 if (la->la_atime < (oattr->la_atime +
687 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
688 la->la_valid &= ~LA_ATIME;
692 /* Check if flags change. */
693 if (la->la_valid & LA_FLAGS) {
694 unsigned int oldflags = oattr->la_flags &
695 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
696 unsigned int newflags = la->la_flags &
697 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
699 if ((uc->uc_fsuid != oattr->la_uid) &&
700 !md_capable(uc, CFS_CAP_FOWNER))
703 /* The IMMUTABLE and APPEND_ONLY flags can
704 * only be changed by the relevant capability. */
705 if ((oldflags ^ newflags) &&
706 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
709 if (!S_ISDIR(oattr->la_mode))
710 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
713 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
714 (la->la_valid & ~LA_FLAGS) &&
715 !(flags & MDS_PERM_BYPASS))
718 /* Check for setting the obj time. */
719 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
720 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
721 if ((uc->uc_fsuid != oattr->la_uid) &&
722 !md_capable(uc, CFS_CAP_FOWNER)) {
723 rc = mdd_permission_internal(env, obj, oattr,
730 if (la->la_valid & LA_KILL_SUID) {
731 la->la_valid &= ~LA_KILL_SUID;
732 if ((oattr->la_mode & S_ISUID) &&
733 !(la->la_valid & LA_MODE)) {
734 la->la_mode = oattr->la_mode;
735 la->la_valid |= LA_MODE;
737 la->la_mode &= ~S_ISUID;
740 if (la->la_valid & LA_KILL_SGID) {
741 la->la_valid &= ~LA_KILL_SGID;
742 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
743 (S_ISGID | S_IXGRP)) &&
744 !(la->la_valid & LA_MODE)) {
745 la->la_mode = oattr->la_mode;
746 la->la_valid |= LA_MODE;
748 la->la_mode &= ~S_ISGID;
751 /* Make sure a caller can chmod. */
752 if (la->la_valid & LA_MODE) {
753 if (!(flags & MDS_PERM_BYPASS) &&
754 (uc->uc_fsuid != oattr->la_uid) &&
755 !md_capable(uc, CFS_CAP_FOWNER))
758 if (la->la_mode == (umode_t) -1)
759 la->la_mode = oattr->la_mode;
761 la->la_mode = (la->la_mode & S_IALLUGO) |
762 (oattr->la_mode & ~S_IALLUGO);
764 /* Also check the setgid bit! */
765 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
766 la->la_gid : oattr->la_gid) &&
767 !md_capable(uc, CFS_CAP_FSETID))
768 la->la_mode &= ~S_ISGID;
770 la->la_mode = oattr->la_mode;
773 /* Make sure a caller can chown. */
774 if (la->la_valid & LA_UID) {
775 if (la->la_uid == (uid_t) -1)
776 la->la_uid = oattr->la_uid;
777 if (((uc->uc_fsuid != oattr->la_uid) ||
778 (la->la_uid != oattr->la_uid)) &&
779 !md_capable(uc, CFS_CAP_CHOWN))
782 /* If the user or group of a non-directory has been
783 * changed by a non-root user, remove the setuid bit.
784 * 19981026 David C Niemi <niemi@tux.org>
786 * Changed this to apply to all users, including root,
787 * to avoid some races. This is the behavior we had in
788 * 2.0. The check for non-root was definitely wrong
789 * for 2.2 anyway, as it should have been using
790 * CAP_FSETID rather than fsuid -- 19990830 SD. */
791 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
792 !S_ISDIR(oattr->la_mode)) {
793 la->la_mode &= ~S_ISUID;
794 la->la_valid |= LA_MODE;
798 /* Make sure caller can chgrp. */
799 if (la->la_valid & LA_GID) {
800 if (la->la_gid == (gid_t) -1)
801 la->la_gid = oattr->la_gid;
802 if (((uc->uc_fsuid != oattr->la_uid) ||
803 ((la->la_gid != oattr->la_gid) &&
804 !lustre_in_group_p(uc, la->la_gid))) &&
805 !md_capable(uc, CFS_CAP_CHOWN))
808 /* Likewise, if the user or group of a non-directory
809 * has been changed by a non-root user, remove the
810 * setgid bit UNLESS there is no group execute bit
811 * (this would be a file marked for mandatory
812 * locking). 19981026 David C Niemi <niemi@tux.org>
814 * Removed the fsuid check (see the comment above) --
816 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
817 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
818 la->la_mode &= ~S_ISGID;
819 la->la_valid |= LA_MODE;
823 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
824 if (!((flags & MDS_OWNEROVERRIDE) &&
825 (uc->uc_fsuid == oattr->la_uid)) &&
826 !(flags & MDS_PERM_BYPASS)) {
827 rc = mdd_permission_internal(env, obj, oattr,
834 if (la->la_valid & LA_CTIME) {
835 /* The pure setattr, it has the priority over what is
836 * already set, do not drop it if ctime is equal. */
837 if (la->la_ctime < oattr->la_ctime)
838 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
844 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
845 struct mdd_device *mdd,
846 enum changelog_rec_type type,
847 int flags, const struct lu_fid *fid,
848 const char *xattr_name,
849 struct thandle *handle)
851 const struct lu_ucred *uc = lu_ucred(env);
852 struct llog_changelog_rec *rec;
855 int xflags = CLFE_INVALID;
858 flags = (flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
861 if (uc->uc_jobid[0] != '\0')
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(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 = flags;
881 rec->cr.cr_type = (__u32)type;
882 rec->cr.cr_tfid = *fid;
883 rec->cr.cr_namelen = 0;
885 if (flags & CLF_JOBID)
886 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
888 if (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, 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, int flags,
917 struct mdd_object *mdd_obj, struct thandle *handle)
921 LASSERT(mdd_obj != NULL);
922 LASSERT(handle != NULL);
924 if (!mdd_changelog_enabled(env, mdd, type))
927 if (mdd_is_volatile_obj(mdd_obj))
930 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
931 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
932 /* Don't need multiple updates in this log */
933 /* Don't check under lock - no big deal if we get an extra
938 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
939 mdo2fid(mdd_obj), NULL, handle);
941 mdd_obj->mod_cltime = ktime_get();
946 int mdd_changelog_data_store_xattr(const struct lu_env *env,
947 struct mdd_device *mdd,
948 enum changelog_rec_type type,
949 int flags, struct mdd_object *mdd_obj,
950 const char *xattr_name,
951 struct thandle *handle)
955 LASSERT(mdd_obj != NULL);
956 LASSERT(handle != NULL);
958 if (!mdd_changelog_enabled(env, mdd, type))
961 if (mdd_is_volatile_obj(mdd_obj))
964 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
965 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
966 /* Don't need multiple updates in this log */
967 /* Don't check under lock - no big deal if we get an extra
973 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
974 mdo2fid(mdd_obj), xattr_name,
977 mdd_obj->mod_cltime = ktime_get();
982 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
983 int flags, struct md_device *m, const struct lu_fid *fid)
985 struct thandle *handle;
986 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
990 LASSERT(fid != NULL);
992 /* We'll check this again below, but we check now before we
993 * start a transaction. */
994 if (!mdd_changelog_enabled(env, mdd, type))
997 handle = mdd_trans_create(env, mdd);
999 RETURN(PTR_ERR(handle));
1001 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1005 rc = mdd_trans_start(env, mdd, handle);
1009 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
1013 rc = mdd_trans_stop(env, mdd, rc, handle);
1019 * Save LMA extended attributes with data from \a ma.
1021 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1022 * not, LMA EA will be first read from disk, modified and write back.
1025 /* Precedence for choosing record type when multiple
1026 * attributes change: setattr > mtime > ctime > atime
1027 * (ctime changes when mtime does, plus chmod/chown.
1028 * atime and ctime are independent.) */
1029 static int mdd_attr_set_changelog(const struct lu_env *env,
1030 struct md_object *obj, struct thandle *handle,
1033 struct mdd_device *mdd = mdo2mdd(obj);
1036 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
1037 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
1038 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
1039 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
1040 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
1041 bits = bits & mdd->mdd_cl.mc_mask;
1042 /* This is an implementation limit rather than a protocol limit */
1043 CLASSERT(CL_LAST <= sizeof(int) * 8);
1047 /* The record type is the lowest non-masked set bit */
1050 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
1051 return mdd_changelog_data_store(env, mdd, type, (int)valid,
1052 md2mdd_obj(obj), handle);
1055 static int mdd_declare_attr_set(const struct lu_env *env,
1056 struct mdd_device *mdd,
1057 struct mdd_object *obj,
1058 const struct lu_attr *attr,
1059 struct thandle *handle)
1063 rc = mdo_declare_attr_set(env, obj, attr, handle);
1067 #ifdef CONFIG_FS_POSIX_ACL
1068 if (attr->la_valid & LA_MODE) {
1069 mdd_read_lock(env, obj, MOR_TGT_CHILD);
1070 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1071 XATTR_NAME_ACL_ACCESS);
1072 mdd_read_unlock(env, obj);
1073 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1079 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1080 rc = mdo_declare_xattr_set(env, obj, buf,
1081 XATTR_NAME_ACL_ACCESS, 0,
1089 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1097 * permission changes may require sync operation, to mitigate performance
1098 * impact, only do this for dir and when permission is reduced.
1100 * For regular files, version is updated with permission change (see VBR), async
1101 * permission won't cause any issue, while missing permission change on
1102 * directory may affect accessibility of other objects after recovery.
1104 static inline bool permission_needs_sync(const struct lu_attr *old,
1105 const struct lu_attr *new)
1107 if (!S_ISDIR(old->la_mode))
1110 if (new->la_valid & (LA_UID | LA_GID))
1113 if (new->la_valid & LA_MODE &&
1114 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
1117 if ((new->la_valid & LA_MODE) &&
1118 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
1119 (old->la_mode & S_IRWXUGO))
1125 static inline __u64 mdd_lmm_dom_size(void *buf)
1127 struct lov_mds_md *lmm = buf;
1128 struct lov_comp_md_v1 *comp_v1;
1129 struct lov_mds_md *v1;
1135 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1138 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1139 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1140 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1142 /* DoM entry is the first entry always */
1143 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1144 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1149 /* set attr and LOV EA at once, return updated attr */
1150 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1151 const struct md_attr *ma)
1153 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1154 struct mdd_device *mdd = mdo2mdd(obj);
1155 struct thandle *handle = NULL;
1156 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1157 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1158 const struct lu_attr *la = &ma->ma_attr;
1159 struct lu_ucred *uc;
1163 /* we do not use ->attr_set() for LOV/HSM EA any more */
1164 LASSERT((ma->ma_valid & MA_LOV) == 0);
1165 LASSERT((ma->ma_valid & MA_HSM) == 0);
1167 rc = mdd_la_get(env, mdd_obj, attr);
1171 *la_copy = ma->ma_attr;
1172 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
1176 /* no need to setattr anymore */
1177 if (la_copy->la_valid == 0) {
1178 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
1179 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
1180 PFID(mdo2fid(mdd_obj)), la->la_valid);
1185 /* If an unprivileged user changes group of some file,
1186 * the setattr operation will be processed synchronously to
1187 * honor the quota limit of the corresponding group. see LU-5152 */
1188 uc = lu_ucred_check(env);
1189 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1190 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1191 la_copy->la_valid |= LA_FLAGS;
1192 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1194 /* Flush the possible existing sync requests to OSTs to
1195 * keep the order of sync for the current setattr operation
1196 * will be sent directly to OSTs. see LU-5152 */
1197 rc = dt_sync(env, mdd->mdd_child);
1202 handle = mdd_trans_create(env, mdd);
1203 if (IS_ERR(handle)) {
1204 rc = PTR_ERR(handle);
1210 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1214 rc = mdd_trans_start(env, mdd, handle);
1218 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
1219 handle->th_sync = 1;
1221 if (la->la_valid & (LA_MTIME | LA_CTIME))
1222 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1223 la->la_mtime, la->la_ctime);
1225 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1227 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1228 * otherwise following truncate will do nothing and truncated
1229 * data may be read again. This is a quick fix until LU-11033
1232 if (la_copy->la_valid & LA_SIZE) {
1233 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1235 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1239 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1240 la_copy->la_valid &= ~LA_SIZE;
1241 lu_buf_free(lov_buf);
1245 if (la_copy->la_valid) {
1246 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1248 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1249 /* rollback to the original gid */
1250 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1251 la_copy->la_gid = attr->la_gid;
1252 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1255 mdd_write_unlock(env, mdd_obj);
1259 rc = mdd_attr_set_changelog(env, obj, handle,
1263 rc = mdd_trans_stop(env, mdd, rc, handle);
1268 static int mdd_xattr_sanity_check(const struct lu_env *env,
1269 struct mdd_object *obj,
1270 const struct lu_attr *attr,
1273 struct lu_ucred *uc = lu_ucred_assert(env);
1276 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1279 if (strncmp(XATTR_USER_PREFIX, name,
1280 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1281 /* For sticky directories, only the owner and privileged user
1282 * can write attributes. */
1283 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1284 (uc->uc_fsuid != attr->la_uid) &&
1285 !md_capable(uc, CFS_CAP_FOWNER))
1287 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1288 (uc->uc_fsuid != attr->la_uid) &&
1289 !md_capable(uc, CFS_CAP_FOWNER)) {
1297 * Check if a string begins with a given prefix.
1299 * \param str String to check
1300 * \param prefix Substring to check at the beginning of \a str
1301 * \return true/false whether the condition is verified.
1303 static inline bool has_prefix(const char *str, const char *prefix)
1305 return strncmp(prefix, str, strlen(prefix)) == 0;
1309 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1311 * \param[in] xattr_name Full extended attribute name.
1313 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1315 static enum changelog_rec_type
1316 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1317 const char *xattr_name)
1319 /* Layout changes systematically recorded */
1320 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1321 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1322 strlen(XATTR_LUSTRE_LOV)) == 0)
1325 /* HSM information changes systematically recorded */
1326 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1329 /* Avoid logging SOM xattr for every file */
1330 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1333 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1334 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1335 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1336 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1337 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1343 static int mdd_declare_xattr_set(const struct lu_env *env,
1344 struct mdd_device *mdd,
1345 struct mdd_object *obj,
1346 const struct lu_buf *buf,
1348 int fl, struct thandle *handle)
1350 enum changelog_rec_type type;
1353 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1357 type = mdd_xattr_changelog_type(env, mdd, name);
1359 return 0; /* no changelog to store */
1361 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1365 * Compare current and future data of HSM EA and add a changelog if needed.
1367 * Caller should have write-locked \param obj.
1369 * \param buf - Future HSM EA content.
1370 * \retval 0 if no changelog is needed or changelog was added properly.
1371 * \retval -ve errno if there was a problem
1373 static int mdd_hsm_update_locked(const struct lu_env *env,
1374 struct md_object *obj,
1375 const struct lu_buf *buf,
1376 struct thandle *handle, int *cl_flags)
1378 struct mdd_thread_info *info = mdd_env_info(env);
1379 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1380 struct lu_buf *current_buf;
1381 struct md_hsm *current_mh;
1382 struct md_hsm *new_mh;
1386 OBD_ALLOC_PTR(current_mh);
1387 if (current_mh == NULL)
1390 /* Read HSM attrs from disk */
1391 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1392 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1393 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1394 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1395 if (rc < 0 && rc != -ENODATA)
1397 else if (rc == -ENODATA)
1398 current_mh->mh_flags = 0;
1400 /* Map future HSM xattr */
1401 OBD_ALLOC_PTR(new_mh);
1403 GOTO(free, rc = -ENOMEM);
1404 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1408 /* Flags differ, set flags for the changelog that will be added */
1409 if (current_mh->mh_flags != new_mh->mh_flags) {
1410 hsm_set_cl_event(cl_flags, HE_STATE);
1411 if (new_mh->mh_flags & HS_DIRTY)
1412 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1415 OBD_FREE_PTR(new_mh);
1418 OBD_FREE_PTR(current_mh);
1422 static int mdd_object_pfid_replace(const struct lu_env *env,
1423 struct mdd_object *o)
1425 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1426 struct thandle *handle;
1429 handle = mdd_trans_create(env, mdd);
1431 RETURN(PTR_ERR(handle));
1433 handle->th_complex = 1;
1435 /* it doesn't need to track the PFID update via llog, because LFSCK
1436 * will repair it even it goes wrong */
1437 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1442 rc = mdd_trans_start(env, mdd, handle);
1446 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1451 mdd_trans_stop(env, mdd, rc, handle);
1456 static int mdd_declare_xattr_del(const struct lu_env *env,
1457 struct mdd_device *mdd,
1458 struct mdd_object *obj,
1460 struct thandle *handle);
1462 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1465 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1466 struct md_object *md_vic)
1468 struct mdd_device *mdd = mdo2mdd(md_obj);
1469 struct mdd_object *obj = md2mdd_obj(md_obj);
1470 struct mdd_object *vic = md2mdd_obj(md_vic);
1471 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1472 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1473 struct lov_mds_md *lmm;
1474 struct thandle *handle;
1478 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1479 if (rc == 0) /* same fid */
1482 handle = mdd_trans_create(env, mdd);
1484 RETURN(PTR_ERR(handle));
1487 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1488 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1490 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1491 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1494 /* get EA of victim file */
1495 memset(buf_vic, 0, sizeof(*buf_vic));
1496 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1503 /* parse the layout of victim file */
1504 lmm = buf_vic->lb_buf;
1505 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1506 GOTO(out, rc = -EINVAL);
1508 /* save EA of target file for restore */
1509 memset(buf, 0, sizeof(*buf));
1510 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1514 /* Get rid of the layout from victim object */
1515 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1519 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1520 LU_XATTR_MERGE, handle);
1524 rc = mdd_trans_start(env, mdd, handle);
1528 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1533 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1535 GOTO(out_restore, rc);
1537 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1538 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1543 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1544 LU_XATTR_REPLACE, handle);
1546 CERROR("%s: failed to rollback of layout of: "DFID
1547 ": %d, file state unknown\n",
1548 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1552 mdd_trans_stop(env, mdd, rc, handle);
1553 mdd_write_unlock(env, obj);
1554 mdd_write_unlock(env, vic);
1556 lu_buf_free(buf_vic);
1559 (void) mdd_object_pfid_replace(env, obj);
1565 * Extract the mirror with specified mirror id, and store the splitted
1566 * mirror layout to @buf.
1568 * \param[in] comp_v1 mirrored layout
1569 * \param[in] mirror_id the mirror with mirror_id to be extracted
1570 * \param[out] buf store the layout excluding the extracted mirror,
1571 * caller free the buffer we allocated in this function
1572 * \param[out] buf_vic store the extracted layout, caller free the buffer
1573 * we allocated in this function
1575 * \retval 0 on success; < 0 if error happens
1577 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1578 struct lu_buf *buf, struct lu_buf *buf_vic)
1580 struct lov_comp_md_v1 *comp_rem;
1581 struct lov_comp_md_v1 *comp_vic;
1582 struct lov_comp_md_entry_v1 *entry;
1583 struct lov_comp_md_entry_v1 *entry_rem;
1584 struct lov_comp_md_entry_v1 *entry_vic;
1586 __u16 comp_cnt, count = 0;
1587 int lmm_size, lmm_size_vic = 0;
1589 int offset, offset_rem, offset_vic;
1591 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1592 /* comp_v1 should contains more than 1 mirror */
1593 if (mirror_cnt <= 1)
1595 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1596 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1598 for (i = 0; i < comp_cnt; i++) {
1599 entry = &comp_v1->lcm_entries[i];
1600 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1602 lmm_size_vic += sizeof(*entry);
1603 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1604 } else if (count > 0) {
1605 /* find the specified mirror */
1613 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1617 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1618 if (!buf_vic->lb_buf) {
1623 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1624 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1626 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1627 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1628 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1629 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1630 if (!comp_rem->lcm_mirror_count)
1631 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1633 memset(comp_vic, 0, sizeof(*comp_v1));
1634 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1635 comp_vic->lcm_mirror_count = 0;
1636 comp_vic->lcm_entry_count = cpu_to_le32(count);
1637 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1638 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1639 comp_vic->lcm_layout_gen = 0;
1641 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1642 offset_rem = sizeof(*comp_rem) +
1643 sizeof(*entry_rem) * (comp_cnt - count);
1644 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1645 for (i = j = k = 0; i < comp_cnt; i++) {
1646 struct lov_mds_md *lmm, *lmm_dst;
1649 entry = &comp_v1->lcm_entries[i];
1650 entry_vic = &comp_vic->lcm_entries[j];
1651 entry_rem = &comp_rem->lcm_entries[k];
1653 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1656 /* copy component entry */
1658 memcpy(entry_vic, entry, sizeof(*entry));
1659 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1660 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1663 memcpy(entry_rem, entry, sizeof(*entry));
1664 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1668 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1670 lmm_dst = (struct lov_mds_md *)
1671 ((char *)comp_vic + offset_vic);
1673 lmm_dst = (struct lov_mds_md *)
1674 ((char *)comp_rem + offset_rem);
1676 /* copy component entry blob */
1677 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1679 /* blob offset advance */
1680 offset += le32_to_cpu(entry->lcme_size);
1682 offset_vic += le32_to_cpu(entry->lcme_size);
1684 offset_rem += le32_to_cpu(entry->lcme_size);
1690 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1691 struct md_rejig_data *mrd)
1693 struct mdd_device *mdd = mdo2mdd(md_obj);
1694 struct mdd_object *obj = md2mdd_obj(md_obj);
1695 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1696 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1697 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1698 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1699 struct lov_comp_md_v1 *lcm;
1700 struct thandle *handle;
1704 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1705 if (rc == 0) /* same fid */
1708 handle = mdd_trans_create(env, mdd);
1710 RETURN(PTR_ERR(handle));
1713 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1714 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1716 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1717 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1720 /* get EA of mirrored file */
1721 memset(buf_save, 0, sizeof(*buf));
1722 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1726 lcm = buf_save->lb_buf;
1727 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1728 GOTO(out, rc = -EINVAL);
1731 * Extract the mirror with specified mirror id, and store the splitted
1732 * mirror layout to the victim file.
1734 memset(buf, 0, sizeof(*buf));
1735 memset(buf_vic, 0, sizeof(*buf_vic));
1736 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1740 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1741 LU_XATTR_SPLIT, handle);
1744 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1745 LU_XATTR_SPLIT, handle);
1749 rc = mdd_trans_start(env, mdd, handle);
1753 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1758 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1761 GOTO(out_restore, rc);
1763 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1767 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1774 /* restore obj's layout */
1775 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1776 LU_XATTR_REPLACE, handle);
1778 CERROR("%s: failed to rollback of layout of: "DFID
1779 ": %d, file state unkonwn.\n",
1780 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1783 mdd_trans_stop(env, mdd, rc, handle);
1784 mdd_write_unlock(env, obj);
1785 mdd_write_unlock(env, vic);
1786 lu_buf_free(buf_save);
1788 lu_buf_free(buf_vic);
1791 (void) mdd_object_pfid_replace(env, obj);
1796 static int mdd_layout_merge_allowed(const struct lu_env *env,
1797 struct md_object *target,
1798 struct md_object *victim)
1800 struct mdd_object *o1 = md2mdd_obj(target);
1802 /* cannot extend directory's LOVEA */
1803 if (S_ISDIR(mdd_object_type(o1))) {
1804 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1805 mdd_obj_dev_name(o1));
1813 * The caller should guarantee to update the object ctime
1814 * after xattr_set if needed.
1816 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1817 const struct lu_buf *buf, const char *name,
1820 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1821 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1822 struct mdd_device *mdd = mdo2mdd(obj);
1823 struct thandle *handle;
1824 enum changelog_rec_type cl_type;
1829 rc = mdd_la_get(env, mdd_obj, attr);
1833 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1837 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1838 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1839 struct md_rejig_data *mrd = buf->lb_buf;
1840 struct md_object *victim = mrd->mrd_obj;
1842 if (buf->lb_len != sizeof(*mrd))
1845 rc = mdd_layout_merge_allowed(env, obj, victim);
1849 if (fl == LU_XATTR_MERGE)
1850 /* merge layout of victim as a mirror of obj's. */
1851 rc = mdd_xattr_merge(env, obj, victim);
1853 rc = mdd_xattr_split(env, obj, mrd);
1857 if (strcmp(name, XATTR_NAME_LMV) == 0) {
1858 rc = mdd_dir_layout_shrink(env, obj, buf);
1862 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1863 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1864 struct posix_acl *acl;
1866 /* user may set empty ACL, which should be treated as removing
1868 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1871 RETURN(PTR_ERR(acl));
1873 rc = mdd_xattr_del(env, obj, name);
1876 posix_acl_release(acl);
1879 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1880 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1884 handle = mdd_trans_create(env, mdd);
1886 RETURN(PTR_ERR(handle));
1888 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1892 rc = mdd_trans_start(env, mdd, handle);
1896 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1898 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1899 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1901 mdd_write_unlock(env, mdd_obj);
1906 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1907 mdd_write_unlock(env, mdd_obj);
1911 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1915 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, cl_flags,
1916 mdd_obj, name, handle);
1920 return mdd_trans_stop(env, mdd, rc, handle);
1923 static int mdd_declare_xattr_del(const struct lu_env *env,
1924 struct mdd_device *mdd,
1925 struct mdd_object *obj,
1927 struct thandle *handle)
1929 enum changelog_rec_type type;
1932 rc = mdo_declare_xattr_del(env, obj, name, handle);
1936 type = mdd_xattr_changelog_type(env, mdd, name);
1938 return 0; /* no changelog to store */
1940 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1944 * The caller should guarantee to update the object ctime
1945 * after xattr_set if needed.
1947 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1950 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1951 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1952 struct mdd_device *mdd = mdo2mdd(obj);
1953 struct thandle *handle;
1957 rc = mdd_la_get(env, mdd_obj, attr);
1961 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1965 handle = mdd_trans_create(env, mdd);
1967 RETURN(PTR_ERR(handle));
1969 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1973 rc = mdd_trans_start(env, mdd, handle);
1977 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1978 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1979 mdd_write_unlock(env, mdd_obj);
1983 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1986 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
1991 return mdd_trans_stop(env, mdd, rc, handle);
1995 * read lov/lmv EA of an object
1996 * return the lov/lmv EA in an allocated lu_buf
1998 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
1999 struct lu_buf *lmm_buf, const char *name)
2001 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2006 if (buf->lb_buf == NULL) {
2007 buf = lu_buf_check_and_alloc(buf, 4096);
2008 if (buf->lb_buf == NULL)
2013 rc = mdo_xattr_get(env, obj, buf, name);
2014 if (rc == -ERANGE) {
2015 /* mti_big_buf is allocated but is too small
2016 * we need to increase it */
2017 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2019 if (buf->lb_buf == NULL)
2022 } else if (rc < 0) {
2024 } else if (rc == 0) {
2028 lu_buf_alloc(lmm_buf, rc);
2029 if (lmm_buf->lb_buf == NULL)
2033 * we don't use lmm_buf directly, because we don't know xattr size, so
2034 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2036 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2041 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2042 struct mdd_object *o, struct lu_buf *buf,
2043 struct thandle *handle)
2045 struct hsm_attrs *attrs;
2051 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2056 attrs = buf->lb_buf;
2057 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2058 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2061 /* Add a changelog record for release. */
2062 hsm_set_cl_event(&flags, HE_RELEASE);
2063 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2069 * check if layout swapping between 2 objects is allowed
2071 * - only normal FIDs or non-system IGIFs
2072 * - same type of objects
2073 * - same owner/group (so quotas are still valid) unless this is from HSM
2076 static int mdd_layout_swap_allowed(const struct lu_env *env,
2077 struct mdd_object *o1,
2078 const struct lu_attr *attr1,
2079 struct mdd_object *o2,
2080 const struct lu_attr *attr2,
2083 const struct lu_fid *fid1, *fid2;
2089 if (!fid_is_norm(fid1) &&
2090 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2093 if (!fid_is_norm(fid2) &&
2094 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2097 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2098 if (S_ISDIR(mdd_object_type(o1)))
2100 if (S_ISREG(mdd_object_type(o1)))
2105 if (flags & SWAP_LAYOUTS_MDS_HSM)
2108 if ((attr1->la_uid != attr2->la_uid) ||
2109 (attr1->la_gid != attr2->la_gid))
2115 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2116 * look into the layout in MDD layer. */
2117 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2119 struct lov_comp_md_v1 *comp_v1;
2120 struct lov_mds_md *v1;
2124 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2125 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2126 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2132 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2133 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2136 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2137 off = le32_to_cpu(comp_v1->lcm_entries[i].
2139 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2144 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2145 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2156 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2158 return mdd_lmm_oi(lmm, oi, true);
2161 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2163 return mdd_lmm_oi(lmm, oi, false);
2166 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2168 struct lov_comp_md_v1 *comp_v1;
2170 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2171 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2173 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2175 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2176 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2177 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2178 __u16 tmp_gen = *gen;
2180 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2182 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2189 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2191 return mdd_lmm_gen(lmm, gen, true);
2194 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2196 return mdd_lmm_gen(lmm, gen, false);
2199 static int mdd_dom_data_truncate(const struct lu_env *env,
2200 struct mdd_device *mdd, struct mdd_object *mo)
2203 struct dt_object *dom;
2206 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2208 GOTO(out, rc = -ENODATA);
2210 th = dt_trans_create(env, mdd->mdd_bottom);
2212 GOTO(out, rc = PTR_ERR(th));
2214 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2218 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2222 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2224 dt_trans_stop(env, mdd->mdd_bottom, th);
2226 /* Ignore failure but report the error */
2228 CERROR("%s: "DFID" can't truncate DOM inode data, rc = %d\n",
2229 mdd_obj_dev_name(mo), PFID(mdo2fid(mo)), rc);
2234 * swap layouts between 2 lustre objects
2236 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2237 struct md_object *obj2, __u64 flags)
2239 struct mdd_thread_info *info = mdd_env_info(env);
2240 struct mdd_object *fst_o = md2mdd_obj(obj1);
2241 struct mdd_object *snd_o = md2mdd_obj(obj2);
2242 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2243 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2244 struct mdd_device *mdd = mdo2mdd(obj1);
2245 struct lov_mds_md *fst_lmm, *snd_lmm;
2246 struct lu_buf *fst_buf = &info->mti_buf[0];
2247 struct lu_buf *snd_buf = &info->mti_buf[1];
2248 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2249 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2250 struct ost_id *saved_oi = NULL;
2251 struct thandle *handle;
2252 struct mdd_object *dom_o = NULL;
2253 __u64 domsize_dom, domsize_vlt;
2254 __u32 fst_gen, snd_gen, saved_gen;
2260 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
2261 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2263 /* we have to sort the 2 obj, so locking will always
2264 * be in the same order, even in case of 2 concurrent swaps */
2265 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
2266 if (rc == 0) /* same fid ? */
2272 rc = mdd_la_get(env, fst_o, fst_la);
2276 rc = mdd_la_get(env, snd_o, snd_la);
2280 /* check if layout swapping is allowed */
2281 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2285 handle = mdd_trans_create(env, mdd);
2287 RETURN(PTR_ERR(handle));
2289 /* objects are already sorted */
2290 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
2291 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
2293 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2294 if (rc < 0 && rc != -ENODATA)
2297 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2298 if (rc < 0 && rc != -ENODATA)
2301 /* check if file has DoM. DoM file can be migrated only to another
2302 * DoM layout with the same DoM component size or to an non-DOM
2303 * layout. After migration to OSTs layout, local MDT inode data
2304 * should be truncated.
2305 * Objects are sorted by FIDs, considering that original file's FID
2306 * is always smaller the snd_o is always original file we are migrating
2309 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2310 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2312 /* Only migration is supported for DoM files, not 'swap_layouts' so
2313 * target file must be volatile and orphan.
2315 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2316 dom_o = domsize_dom ? snd_o : NULL;
2317 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2318 swap(domsize_dom, domsize_vlt);
2319 dom_o = domsize_dom ? fst_o : NULL;
2320 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2321 /* 'lfs swap_layouts' case, neither file should have DoM */
2323 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2324 "use migration instead: rc = %d\n", rc);
2328 if (domsize_vlt > 0 && domsize_dom == 0) {
2330 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": OST to DOM "
2331 "migration is not supported: rc = %d\n",
2332 PFID(mdo2fid(snd_o)), rc);
2334 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2336 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": new layout "
2337 "must have the same DoM component size: rc = %d\n",
2338 PFID(mdo2fid(fst_o)), rc);
2340 } else if (domsize_vlt > 0) {
2341 /* Migration with the same DOM component size, no need to
2342 * truncate local data, it is still being used */
2346 /* swapping 2 non existant layouts is a success */
2347 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2350 /* to help inode migration between MDT, it is better to
2351 * start by the no layout file (if one), so we order the swap */
2352 if (snd_buf->lb_buf == NULL) {
2354 swap(fst_buf, snd_buf);
2357 fst_gen = snd_gen = 0;
2358 /* lmm and generation layout initialization */
2359 if (fst_buf->lb_buf != NULL) {
2360 fst_lmm = fst_buf->lb_buf;
2361 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2362 fst_fl = LU_XATTR_REPLACE;
2366 fst_fl = LU_XATTR_CREATE;
2369 snd_lmm = snd_buf->lb_buf;
2370 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2372 saved_gen = fst_gen;
2373 /* increase the generation layout numbers */
2378 * XXX The layout generation is used to generate component IDs for
2379 * the composite file, we have to do some special tweaks to make
2380 * sure the layout generation is always adequate for that job.
2383 /* Skip invalid generation number for composite layout */
2384 if ((snd_gen & LCME_ID_MASK) == 0)
2386 if ((fst_gen & LCME_ID_MASK) == 0)
2388 /* Make sure the generation is greater than all the component IDs */
2389 if (fst_gen < snd_gen)
2391 else if (fst_gen > snd_gen)
2394 /* set the file specific informations in lmm */
2395 if (fst_lmm != NULL) {
2396 saved_oi = &info->mti_oa.o_oi;
2397 mdd_get_lmm_oi(fst_lmm, saved_oi);
2398 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2399 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2400 mdd_set_lmm_oi(snd_lmm, saved_oi);
2402 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2403 cpu_to_le32(LOV_MAGIC_MAGIC))
2404 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2406 GOTO(stop, rc = -EPROTO);
2408 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2410 /* Prepare HSM attribute if it's required */
2411 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2412 const int buflen = sizeof(struct hsm_attrs);
2414 lu_buf_alloc(fst_hsm_buf, buflen);
2415 lu_buf_alloc(snd_hsm_buf, buflen);
2416 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2417 GOTO(stop, rc = -ENOMEM);
2419 /* Read HSM attribute */
2420 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2424 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2428 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2429 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2434 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2435 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2441 /* prepare transaction */
2442 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2447 if (fst_buf->lb_buf != NULL)
2448 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2449 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2452 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2457 rc = mdd_trans_start(env, mdd, handle);
2461 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2462 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2466 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2468 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2471 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
2472 mdd_obj_dev_name(fst_o),
2473 PFID(mdo2fid(fst_o)), rc, rc2);
2478 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2482 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2485 if (fst_buf->lb_buf != NULL)
2486 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2487 LU_XATTR_REPLACE, handle);
2489 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2492 GOTO(out_restore, rc);
2494 /* Issue one changelog record per file */
2495 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2499 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2508 /* failure on second file, but first was done, so we have
2509 * to roll back first. */
2510 if (fst_buf->lb_buf != NULL) {
2511 mdd_set_lmm_oi(fst_lmm, saved_oi);
2512 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2513 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2514 LU_XATTR_REPLACE, handle);
2516 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2522 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2527 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2532 CERROR("%s: unable to roll back layout swap. FIDs: "
2533 DFID" and "DFID "error: %d/%d, steps: %d\n",
2534 mdd_obj_dev_name(fst_o),
2535 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
2537 /* a solution to avoid journal commit is to panic,
2538 * but it has strong consequences so we use LBUG to
2539 * allow sysdamin to choose to panic or not
2546 rc = mdd_trans_stop(env, mdd, rc, handle);
2548 /* Truncate local DOM data if all went well */
2550 mdd_dom_data_truncate(env, mdd, dom_o);
2552 mdd_write_unlock(env, snd_o);
2553 mdd_write_unlock(env, fst_o);
2555 lu_buf_free(fst_buf);
2556 lu_buf_free(snd_buf);
2557 lu_buf_free(fst_hsm_buf);
2558 lu_buf_free(snd_hsm_buf);
2561 (void) mdd_object_pfid_replace(env, fst_o);
2562 (void) mdd_object_pfid_replace(env, snd_o);
2567 static int mdd_declare_layout_change(const struct lu_env *env,
2568 struct mdd_device *mdd,
2569 struct mdd_object *obj,
2570 struct md_layout_change *mlc,
2571 struct thandle *handle)
2575 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2579 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2583 /* For PFL, this is used to instantiate necessary component objects. */
2585 mdd_layout_instantiate_component(const struct lu_env *env,
2586 struct mdd_object *obj, struct md_layout_change *mlc,
2587 struct thandle *handle)
2589 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2593 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2596 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2598 * It's possible that another layout write intent has already
2599 * instantiated our objects, so a -EALREADY returned, and we need to
2603 RETURN(rc == -EALREADY ? 0 : rc);
2605 rc = mdd_trans_start(env, mdd, handle);
2609 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2610 rc = mdo_layout_change(env, obj, mlc, handle);
2611 mdd_write_unlock(env, obj);
2615 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2620 * Change the FLR layout from RDONLY to WRITE_PENDING.
2622 * It picks the primary mirror, and bumps the layout version, and set
2623 * layout version xattr to OST objects in a sync tx. In order to facilitate
2624 * the handling of phantom writers from evicted clients, the clients carry
2625 * layout version of the file with write RPC, so that the OSTs can verify
2626 * if the write RPCs are legitimate, meaning not from evicted clients.
2629 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2630 struct md_layout_change *mlc, struct thandle *handle)
2632 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2633 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2634 struct lustre_som_attrs *som = &mlc->mlc_som;
2639 /* Verify acceptable operations */
2640 switch (mlc->mlc_opc) {
2641 case MD_LAYOUT_WRITE:
2642 case MD_LAYOUT_RESYNC:
2643 /* these are legal operations - this represents the case that
2644 * a few mirrors were missed in the last resync. */
2646 case MD_LAYOUT_RESYNC_DONE:
2651 som_buf->lb_buf = som;
2652 som_buf->lb_len = sizeof(*som);
2653 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2654 if (rc < 0 && rc != -ENODATA)
2658 lustre_som_swab(som);
2659 if (som->lsa_valid & SOM_FL_STRICT)
2660 fl = LU_XATTR_REPLACE;
2663 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2668 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2669 XATTR_NAME_SOM, fl, handle);
2674 /* record a changelog for data mover to consume */
2675 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2679 rc = mdd_trans_start(env, mdd, handle);
2683 /* it needs a sync tx to make FLR to work properly */
2684 handle->th_sync = 1;
2686 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2687 rc = mdo_layout_change(env, obj, mlc, handle);
2689 /* SOM state transition from STRICT to STALE */
2690 som->lsa_valid = SOM_FL_STALE;
2691 lustre_som_swab(som);
2692 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2695 mdd_write_unlock(env, obj);
2699 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2710 * Handle mirrored file state transition when it's in WRITE_PENDING.
2712 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2713 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2714 * layout version will be increased, and the file's state will be changed to
2718 mdd_layout_update_write_pending(const struct lu_env *env,
2719 struct mdd_object *obj, struct md_layout_change *mlc,
2720 struct thandle *handle)
2722 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2726 switch (mlc->mlc_opc) {
2727 case MD_LAYOUT_RESYNC:
2728 /* Upon receiving the resync request, it should
2729 * instantiate all stale components right away to get ready
2730 * for mirror copy. In order to avoid layout version change,
2731 * client should avoid sending LAYOUT_WRITE request at the
2734 case MD_LAYOUT_WRITE:
2735 /* legal race for concurrent write, the file state has been
2736 * changed by another client. */
2742 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2746 rc = mdd_trans_start(env, mdd, handle);
2750 /* it needs a sync tx to make FLR to work properly */
2751 handle->th_sync = 1;
2753 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2754 rc = mdo_layout_change(env, obj, mlc, handle);
2755 mdd_write_unlock(env, obj);
2766 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2768 * Only concurrent write and sync complete requests are possible when the
2769 * file is in SYNC_PENDING. For the latter request, it will pass in the
2770 * mirrors that have been synchronized, then the stale bit will be cleared
2771 * to make the file's state turn into RDONLY.
2772 * For concurrent write reqeust, it just needs to change the file's state
2773 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2774 * version because the version will be increased in the transition to
2775 * SYNC_PENDING later so that it can deny the write request from potential
2776 * evicted SYNC clients. */
2778 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2779 struct md_layout_change *mlc, struct thandle *handle)
2781 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2782 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2787 /* operation validation */
2788 switch (mlc->mlc_opc) {
2789 case MD_LAYOUT_RESYNC_DONE:
2790 /* resync complete. */
2791 case MD_LAYOUT_WRITE:
2792 /* concurrent write. */
2794 case MD_LAYOUT_RESYNC:
2795 /* resync again, most likely the previous run failed.
2796 * no-op if it's already in SYNC_PENDING state */
2802 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2803 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2804 if (rc < 0 && rc != -ENODATA)
2807 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2808 lustre_som_swab(&mlc->mlc_som);
2809 som_buf->lb_buf = &mlc->mlc_som;
2810 som_buf->lb_len = sizeof(mlc->mlc_som);
2813 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2817 /* record a changelog for the completion of resync */
2818 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2823 /* RESYNC_DONE has piggybacked size and blocks */
2825 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2826 XATTR_NAME_SOM, fl, handle);
2831 rc = mdd_trans_start(env, mdd, handle);
2835 /* it needs a sync tx to make FLR to work properly */
2836 handle->th_sync = 1;
2838 rc = mdo_layout_change(env, obj, mlc, handle);
2843 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2849 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2858 * Layout change callback for object.
2860 * This is only used by FLR for now. In the future, it can be exteneded to
2861 * handle all layout change.
2864 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2865 struct md_layout_change *mlc)
2867 struct mdd_object *obj = md2mdd_obj(o);
2868 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2869 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2870 struct lov_comp_md_v1 *lcm;
2871 struct thandle *handle;
2876 /* Verify acceptable operations */
2877 switch (mlc->mlc_opc) {
2878 case MD_LAYOUT_WRITE:
2879 case MD_LAYOUT_RESYNC:
2880 case MD_LAYOUT_RESYNC_DONE:
2886 handle = mdd_trans_create(env, mdd);
2888 RETURN(PTR_ERR(handle));
2890 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2897 /* analyze the layout to make sure it's a FLR file */
2899 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2900 GOTO(out, rc = -EINVAL);
2902 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2904 /* please refer to HLD of FLR for state transition */
2905 switch (flr_state) {
2907 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2909 case LCM_FL_WRITE_PENDING:
2910 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2913 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2915 case LCM_FL_SYNC_PENDING:
2916 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2925 mdd_trans_stop(env, mdd, rc, handle);
2930 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2931 struct mdd_object *child, const struct lu_attr *attr,
2932 const struct md_op_spec *spec,
2933 struct dt_allocation_hint *hint)
2935 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2936 struct dt_object *nc = mdd_object_child(child);
2938 memset(hint, 0, sizeof(*hint));
2940 /* For striped directory, give striping EA to lod_ah_init, which will
2941 * decide the stripe_offset and stripe count by it. */
2942 if (S_ISDIR(attr->la_mode) &&
2943 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2944 hint->dah_eadata = spec->u.sp_ea.eadata;
2945 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2947 hint->dah_eadata = NULL;
2948 hint->dah_eadata_len = 0;
2951 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2952 hint->dah_eadata, hint->dah_eadata_len);
2953 /* @hint will be initialized by underlying device. */
2954 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2957 static int accmode(const struct lu_env *env, const struct lu_attr *la,
2960 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2961 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2962 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2963 * owner can write to a file even if it is marked readonly to hide
2964 * its brokenness. (bug 5781) */
2965 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
2966 struct lu_ucred *uc = lu_ucred_check(env);
2968 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2972 return mds_accmode(open_flags);
2975 static int mdd_open_sanity_check(const struct lu_env *env,
2976 struct mdd_object *obj,
2977 const struct lu_attr *attr, u64 open_flags)
2982 /* EEXIST check, also opening of *open* orphans is allowed so we can
2983 * open-by-handle unlinked files
2985 if (mdd_is_dead_obj(obj) &&
2986 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
2989 if (S_ISLNK(attr->la_mode))
2992 mode = accmode(env, attr, open_flags);
2994 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
2997 if (!(open_flags & MDS_OPEN_CREATED)) {
2998 rc = mdd_permission_internal(env, obj, attr, mode);
3003 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3004 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3005 open_flags &= ~MDS_OPEN_TRUNC;
3007 /* For writing append-only file must open it with append mode. */
3008 if (attr->la_flags & LUSTRE_APPEND_FL) {
3009 if ((open_flags & MDS_FMODE_WRITE) &&
3010 !(open_flags & MDS_OPEN_APPEND))
3012 if (open_flags & MDS_OPEN_TRUNC)
3019 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3022 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3023 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3024 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3025 struct mdd_object_user *mou = NULL;
3026 const struct lu_ucred *uc = lu_ucred(env);
3027 struct mdd_device *mdd = mdo2mdd(obj);
3028 enum changelog_rec_type type = CL_OPEN;
3031 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3033 rc = mdd_la_get(env, mdd_obj, attr);
3037 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags);
3038 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3043 mdd_obj->mod_count++;
3045 if (!mdd_changelog_enabled(env, mdd, type))
3049 /* look for existing opener in list under mdd_write_lock */
3050 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3055 /* add user to list */
3056 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3062 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3064 mdd_obj_user_free(mou);
3069 if (type == CL_DN_OPEN) {
3070 if (ktime_before(ktime_get(), mou->mou_deniednext))
3071 /* same user denied again same access within
3072 * time interval: do not record
3076 /* this user already denied, but some time ago:
3077 * update denied time
3079 mou->mou_deniednext =
3080 ktime_add(ktime_get(),
3081 ktime_set(mdd->mdd_cl.mc_deniednext,
3084 mou->mou_opencount++;
3085 /* same user opening file again with same flags:
3092 mdd_changelog(env, type, open_flags, md_dev, mdo2fid(mdd_obj));
3096 mdd_write_unlock(env, mdd_obj);
3100 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3101 struct md_attr *ma, struct thandle *handle)
3105 rc = mdd_orphan_declare_delete(env, obj, handle);
3109 return mdo_declare_destroy(env, obj, handle);
3113 * No permission check is needed.
3115 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3116 struct md_attr *ma, u64 open_flags)
3118 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3119 struct mdd_device *mdd = mdo2mdd(obj);
3120 struct thandle *handle = NULL;
3123 bool blocked = false;
3124 int last_close_by_uid = 0;
3125 const struct lu_ucred *uc = lu_ucred(env);
3128 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3129 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3130 mdd_obj->mod_count--;
3131 mdd_write_unlock(env, mdd_obj);
3133 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3134 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3135 "list\n", PFID(mdd_object_fid(mdd_obj)));
3139 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3140 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3141 /* check without any lock */
3142 is_orphan = mdd_obj->mod_count == 1 &&
3143 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3147 /* mdd_trans_create() maybe failed because of barrier_entry(),
3148 * under such case, the orphan MDT-object will be left in the
3149 * orphan list, and when the MDT remount next time, the unused
3150 * orphans will be destroyed automatically.
3152 * One exception: the former mdd_finish_unlink may failed to
3153 * add the orphan MDT-object to the orphan list, then if the
3154 * mdd_trans_create() failed because of barrier_entry(), the
3155 * MDT-object will become real orphan that is neither in the
3156 * namespace nor in the orphan list. Such bad case should be
3157 * very rare and will be handled by e2fsck/lfsck. */
3158 handle = mdd_trans_create(env, mdo2mdd(obj));
3159 if (IS_ERR(handle)) {
3160 rc = PTR_ERR(handle);
3161 if (rc != -EINPROGRESS)
3169 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3173 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3178 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3184 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3185 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3187 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3188 lu_dev_name(mdd2lu_dev(mdd)),
3189 PFID(mdd_object_fid(mdd_obj)), rc);
3193 /* check again with lock */
3194 is_orphan = (mdd_obj->mod_count == 1) &&
3195 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3196 ma->ma_attr.la_nlink == 0);
3198 if (is_orphan && !handle && !blocked) {
3199 mdd_write_unlock(env, mdd_obj);
3203 mdd_obj->mod_count--; /*release open count */
3205 /* under mdd write lock */
3206 /* If recording, see if we need to remove UID from list */
3207 if (mdd_changelog_enabled(env, mdd, CL_OPEN)) {
3208 struct mdd_object_user *mou;
3210 /* look for UID in list */
3211 /* If mou is NULL, it probably means logging was enabled after
3212 * the user had the file open. So the corresponding close
3213 * will not be logged.
3215 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3218 mou->mou_opencount--;
3219 if (mou->mou_opencount == 0) {
3220 mdd_obj_user_remove(mdd_obj, mou);
3221 last_close_by_uid = 1;
3226 if (!is_orphan || blocked)
3230 /* NB: Object maybe not in orphan list originally, it is rare case for
3231 * mdd_finish_unlink() failure, in that case, the object doesn't have
3232 * ORPHAN_OBJ flag */
3233 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3234 /* remove link to object from orphan index */
3235 LASSERT(handle != NULL);
3236 rc = mdd_orphan_delete(env, mdd_obj, handle);
3238 CERROR("%s: unable to delete "DFID" from orphan list: "
3239 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3240 PFID(mdd_object_fid(mdd_obj)), rc);
3241 /* If object was not deleted from orphan list, do not
3242 * destroy OSS objects, which will be done when next
3247 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3248 "list, OSS objects to be destroyed.\n",
3249 PFID(mdd_object_fid(mdd_obj)));
3252 rc = mdo_destroy(env, mdd_obj, handle);
3255 CERROR("%s: unable to delete "DFID" from orphan list: "
3256 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3257 PFID(mdd_object_fid(mdd_obj)), rc);
3262 mdd_write_unlock(env, mdd_obj);
3264 if (rc != 0 || blocked ||
3265 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3268 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3269 * or if CL_OPEN was recorded and it's last close by user.
3270 * Changelogs mask may change between open and close operations, but
3271 * this is not a big deal if we have a CL_CLOSE entry with no matching
3272 * CL_OPEN. Plus Changelogs mask may not change often.
3274 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3275 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3276 MDS_OPEN_TRUNC))) ||
3277 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3278 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3279 if (handle == NULL) {
3280 handle = mdd_trans_create(env, mdo2mdd(obj));
3282 GOTO(stop, rc = PTR_ERR(handle));
3284 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3285 NULL, NULL, handle);
3289 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3294 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3299 if (handle != NULL && !IS_ERR(handle))
3300 rc = mdd_trans_stop(env, mdd, rc, handle);
3306 * Permission check is done when open,
3307 * no need check again.
3309 static int mdd_readpage_sanity_check(const struct lu_env *env,
3310 struct mdd_object *obj)
3312 struct dt_object *next = mdd_object_child(obj);
3316 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3324 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3325 size_t nob, const struct dt_it_ops *iops,
3326 struct dt_it *it, __u32 attr, void *arg)
3328 struct lu_dirpage *dp = &lp->lp_dir;
3332 struct lu_dirent *ent;
3333 struct lu_dirent *last = NULL;
3337 if (nob < sizeof(*dp))
3340 memset(area, 0, sizeof (*dp));
3341 area += sizeof (*dp);
3342 nob -= sizeof (*dp);
3349 len = iops->key_size(env, it);
3351 /* IAM iterator can return record with zero len. */
3355 hash = iops->store(env, it);
3356 if (unlikely(first)) {
3358 dp->ldp_hash_start = cpu_to_le64(hash);
3361 /* calculate max space required for lu_dirent */
3362 recsize = lu_dirent_calc_size(len, attr);
3364 if (nob >= recsize) {
3365 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3366 if (result == -ESTALE)
3371 /* osd might not able to pack all attributes,
3372 * so recheck rec length */
3373 recsize = le16_to_cpu(ent->lde_reclen);
3375 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3376 fid_le_to_cpu(&fid, &ent->lde_fid);
3377 if (fid_is_dot_lustre(&fid))
3381 result = (last != NULL) ? 0 :-EINVAL;
3385 ent = (void *)ent + recsize;
3389 result = iops->next(env, it);
3390 if (result == -ESTALE)
3392 } while (result == 0);
3395 dp->ldp_hash_end = cpu_to_le64(hash);
3397 if (last->lde_hash == dp->ldp_hash_end)
3398 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3399 last->lde_reclen = 0; /* end mark */
3402 /* end of directory */
3403 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3404 else if (result < 0)
3405 CWARN("build page failed: %d!\n", result);
3409 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3410 const struct lu_rdpg *rdpg)
3412 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3416 if (mdd_object_exists(mdd_obj) == 0) {
3417 CERROR("%s: object "DFID" not found: rc = -2\n",
3418 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3422 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
3423 rc = mdd_readpage_sanity_check(env, mdd_obj);
3425 GOTO(out_unlock, rc);
3427 if (mdd_is_dead_obj(mdd_obj)) {
3429 struct lu_dirpage *dp;
3432 * According to POSIX, please do not return any entry to client:
3433 * even dot and dotdot should not be returned.
3435 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3436 PFID(mdd_object_fid(mdd_obj)));
3438 if (rdpg->rp_count <= 0)
3439 GOTO(out_unlock, rc = -EFAULT);
3440 LASSERT(rdpg->rp_pages != NULL);
3442 pg = rdpg->rp_pages[0];
3443 dp = (struct lu_dirpage *)kmap(pg);
3444 memset(dp, 0 , sizeof(struct lu_dirpage));
3445 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3446 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3447 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3449 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3452 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3453 mdd_dir_page_build, NULL);
3455 struct lu_dirpage *dp;
3457 dp = kmap(rdpg->rp_pages[0]);
3458 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3461 * No pages were processed, mark this for first page
3464 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3465 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3466 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3468 kunmap(rdpg->rp_pages[0]);
3471 GOTO(out_unlock, rc);
3473 mdd_read_unlock(env, mdd_obj);
3477 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3479 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3481 if (mdd_object_exists(mdd_obj) == 0) {
3484 CERROR("%s: object "DFID" not found: rc = %d\n",
3485 mdd_obj_dev_name(mdd_obj),
3486 PFID(mdd_object_fid(mdd_obj)), rc);
3489 return dt_object_sync(env, mdd_object_child(mdd_obj),
3493 static int mdd_object_lock(const struct lu_env *env,
3494 struct md_object *obj,
3495 struct lustre_handle *lh,
3496 struct ldlm_enqueue_info *einfo,
3497 union ldlm_policy_data *policy)
3499 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3500 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3504 static int mdd_object_unlock(const struct lu_env *env,
3505 struct md_object *obj,
3506 struct ldlm_enqueue_info *einfo,
3507 union ldlm_policy_data *policy)
3509 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3510 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3513 const struct md_object_operations mdd_obj_ops = {
3514 .moo_permission = mdd_permission,
3515 .moo_attr_get = mdd_attr_get,
3516 .moo_attr_set = mdd_attr_set,
3517 .moo_xattr_get = mdd_xattr_get,
3518 .moo_xattr_set = mdd_xattr_set,
3519 .moo_xattr_list = mdd_xattr_list,
3520 .moo_invalidate = mdd_invalidate,
3521 .moo_xattr_del = mdd_xattr_del,
3522 .moo_swap_layouts = mdd_swap_layouts,
3523 .moo_open = mdd_open,
3524 .moo_close = mdd_close,
3525 .moo_readpage = mdd_readpage,
3526 .moo_readlink = mdd_readlink,
3527 .moo_changelog = mdd_changelog,
3528 .moo_object_sync = mdd_object_sync,
3529 .moo_object_lock = mdd_object_lock,
3530 .moo_object_unlock = mdd_object_unlock,
3531 .moo_layout_change = mdd_layout_change,