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,
1102 * permission changes may require sync operation, to mitigate performance
1103 * impact, only do this for dir and when permission is reduced.
1105 * For regular files, version is updated with permission change (see VBR), async
1106 * permission won't cause any issue, while missing permission change on
1107 * directory may affect accessibility of other objects after recovery.
1109 static inline bool permission_needs_sync(const struct lu_attr *old,
1110 const struct lu_attr *new)
1112 if (!S_ISDIR(old->la_mode))
1115 if (new->la_valid & (LA_UID | LA_GID))
1118 if (new->la_valid & LA_MODE &&
1119 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
1122 if ((new->la_valid & LA_MODE) &&
1123 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
1124 (old->la_mode & S_IRWXUGO))
1130 static inline __u64 mdd_lmm_dom_size(void *buf)
1132 struct lov_mds_md *lmm = buf;
1133 struct lov_comp_md_v1 *comp_v1;
1134 struct lov_mds_md *v1;
1140 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1143 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1144 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1145 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1147 /* DoM entry is the first entry always */
1148 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1149 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1154 /* set attr and LOV EA at once, return updated attr */
1155 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1156 const struct md_attr *ma)
1158 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1159 struct mdd_device *mdd = mdo2mdd(obj);
1160 struct thandle *handle = NULL;
1161 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1162 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1163 const struct lu_attr *la = &ma->ma_attr;
1164 struct lu_ucred *uc;
1165 bool chrgrp_by_unprivileged_user = false;
1169 /* we do not use ->attr_set() for LOV/HSM EA any more */
1170 LASSERT((ma->ma_valid & MA_LOV) == 0);
1171 LASSERT((ma->ma_valid & MA_HSM) == 0);
1173 rc = mdd_la_get(env, mdd_obj, attr);
1177 *la_copy = ma->ma_attr;
1178 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
1182 /* no need to setattr anymore */
1183 if (la_copy->la_valid == 0) {
1184 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
1185 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
1186 PFID(mdo2fid(mdd_obj)), la->la_valid);
1191 /* If an unprivileged user changes group of some file,
1192 * the setattr operation will be processed synchronously to
1193 * honor the quota limit of the corresponding group. see LU-5152 */
1194 uc = lu_ucred_check(env);
1195 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1196 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1197 /* LU-10048: disable synchronous chgrp operation for it will
1198 * cause deadlock between MDT and OST.
1199 la_copy->la_valid |= LA_FLAGS;
1200 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1202 chrgrp_by_unprivileged_user = true;
1204 /* Flush the possible existing client setattr requests to OSTs
1205 * to keep the order with the current setattr operation that
1206 * will be sent directly to OSTs. see LU-5152 */
1207 /* LU-11303 disable sync as this is too heavyweight.
1208 * This should be replaced with a sync only for the object
1209 * being modified here, not the whole filesystem.
1210 rc = dt_sync(env, mdd->mdd_child);
1216 handle = mdd_trans_create(env, mdd);
1217 if (IS_ERR(handle)) {
1218 rc = PTR_ERR(handle);
1224 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1228 rc = mdd_trans_start(env, mdd, handle);
1232 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1233 permission_needs_sync(attr, la))
1234 handle->th_sync = 1;
1236 if (la->la_valid & (LA_MTIME | LA_CTIME))
1237 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1238 la->la_mtime, la->la_ctime);
1240 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1242 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1243 * otherwise following truncate will do nothing and truncated
1244 * data may be read again. This is a quick fix until LU-11033
1247 if (la_copy->la_valid & LA_SIZE) {
1248 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1250 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1254 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1255 la_copy->la_valid &= ~LA_SIZE;
1256 lu_buf_free(lov_buf);
1260 if (la_copy->la_valid) {
1261 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1263 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1264 /* rollback to the original gid */
1265 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1266 la_copy->la_gid = attr->la_gid;
1267 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1270 mdd_write_unlock(env, mdd_obj);
1274 rc = mdd_attr_set_changelog(env, obj, handle,
1278 rc = mdd_trans_stop(env, mdd, rc, handle);
1283 static int mdd_xattr_sanity_check(const struct lu_env *env,
1284 struct mdd_object *obj,
1285 const struct lu_attr *attr,
1288 struct lu_ucred *uc = lu_ucred_assert(env);
1291 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1294 if (strncmp(XATTR_USER_PREFIX, name,
1295 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1296 /* For sticky directories, only the owner and privileged user
1297 * can write attributes. */
1298 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1299 (uc->uc_fsuid != attr->la_uid) &&
1300 !md_capable(uc, CFS_CAP_FOWNER))
1302 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1303 (uc->uc_fsuid != attr->la_uid) &&
1304 !md_capable(uc, CFS_CAP_FOWNER)) {
1312 * Check if a string begins with a given prefix.
1314 * \param str String to check
1315 * \param prefix Substring to check at the beginning of \a str
1316 * \return true/false whether the condition is verified.
1318 static inline bool has_prefix(const char *str, const char *prefix)
1320 return strncmp(prefix, str, strlen(prefix)) == 0;
1324 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1326 * \param[in] xattr_name Full extended attribute name.
1328 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1330 static enum changelog_rec_type
1331 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1332 const char *xattr_name)
1334 /* Layout changes systematically recorded */
1335 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1336 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1337 strlen(XATTR_LUSTRE_LOV)) == 0)
1340 /* HSM information changes systematically recorded */
1341 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1344 /* Avoid logging SOM xattr for every file */
1345 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1348 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1349 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1350 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1351 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1352 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1358 static int mdd_declare_xattr_set(const struct lu_env *env,
1359 struct mdd_device *mdd,
1360 struct mdd_object *obj,
1361 const struct lu_buf *buf,
1363 int fl, struct thandle *handle)
1365 enum changelog_rec_type type;
1368 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1372 type = mdd_xattr_changelog_type(env, mdd, name);
1374 return 0; /* no changelog to store */
1376 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1380 * Compare current and future data of HSM EA and add a changelog if needed.
1382 * Caller should have write-locked \param obj.
1384 * \param buf - Future HSM EA content.
1385 * \retval 0 if no changelog is needed or changelog was added properly.
1386 * \retval -ve errno if there was a problem
1388 static int mdd_hsm_update_locked(const struct lu_env *env,
1389 struct md_object *obj,
1390 const struct lu_buf *buf,
1391 struct thandle *handle,
1392 enum changelog_rec_flags *clf_flags)
1394 struct mdd_thread_info *info = mdd_env_info(env);
1395 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1396 struct lu_buf *current_buf;
1397 struct md_hsm *current_mh;
1398 struct md_hsm *new_mh;
1402 OBD_ALLOC_PTR(current_mh);
1403 if (current_mh == NULL)
1406 /* Read HSM attrs from disk */
1407 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1408 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1409 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1410 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1411 if (rc < 0 && rc != -ENODATA)
1413 else if (rc == -ENODATA)
1414 current_mh->mh_flags = 0;
1416 /* Map future HSM xattr */
1417 OBD_ALLOC_PTR(new_mh);
1419 GOTO(free, rc = -ENOMEM);
1420 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1424 /* Flags differ, set flags for the changelog that will be added */
1425 if (current_mh->mh_flags != new_mh->mh_flags) {
1426 hsm_set_cl_event(clf_flags, HE_STATE);
1427 if (new_mh->mh_flags & HS_DIRTY)
1428 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1431 OBD_FREE_PTR(new_mh);
1434 OBD_FREE_PTR(current_mh);
1438 static int mdd_object_pfid_replace(const struct lu_env *env,
1439 struct mdd_object *o)
1441 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1442 struct thandle *handle;
1445 handle = mdd_trans_create(env, mdd);
1447 RETURN(PTR_ERR(handle));
1449 handle->th_complex = 1;
1451 /* it doesn't need to track the PFID update via llog, because LFSCK
1452 * will repair it even it goes wrong */
1453 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1458 rc = mdd_trans_start(env, mdd, handle);
1462 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1467 mdd_trans_stop(env, mdd, rc, handle);
1472 static int mdd_declare_xattr_del(const struct lu_env *env,
1473 struct mdd_device *mdd,
1474 struct mdd_object *obj,
1476 struct thandle *handle);
1478 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1481 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1482 struct md_object *md_vic)
1484 struct mdd_device *mdd = mdo2mdd(md_obj);
1485 struct mdd_object *obj = md2mdd_obj(md_obj);
1486 struct mdd_object *vic = md2mdd_obj(md_vic);
1487 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1488 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1489 struct lov_mds_md *lmm;
1490 struct thandle *handle;
1494 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1495 if (rc == 0) /* same fid */
1498 handle = mdd_trans_create(env, mdd);
1500 RETURN(PTR_ERR(handle));
1503 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1504 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1506 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1507 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1510 /* get EA of victim file */
1511 memset(buf_vic, 0, sizeof(*buf_vic));
1512 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1519 /* parse the layout of victim file */
1520 lmm = buf_vic->lb_buf;
1521 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1522 GOTO(out, rc = -EINVAL);
1524 /* save EA of target file for restore */
1525 memset(buf, 0, sizeof(*buf));
1526 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1530 /* Get rid of the layout from victim object */
1531 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1535 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1536 LU_XATTR_MERGE, handle);
1540 rc = mdd_trans_start(env, mdd, handle);
1544 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1549 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1551 GOTO(out_restore, rc);
1553 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1554 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1559 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1560 LU_XATTR_REPLACE, handle);
1562 CERROR("%s: failed to rollback of layout of: "DFID
1563 ": %d, file state unknown\n",
1564 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1568 mdd_trans_stop(env, mdd, rc, handle);
1569 mdd_write_unlock(env, obj);
1570 mdd_write_unlock(env, vic);
1572 lu_buf_free(buf_vic);
1575 (void) mdd_object_pfid_replace(env, obj);
1581 * Extract the mirror with specified mirror id, and store the splitted
1582 * mirror layout to @buf.
1584 * \param[in] comp_v1 mirrored layout
1585 * \param[in] mirror_id the mirror with mirror_id to be extracted
1586 * \param[out] buf store the layout excluding the extracted mirror,
1587 * caller free the buffer we allocated in this function
1588 * \param[out] buf_vic store the extracted layout, caller free the buffer
1589 * we allocated in this function
1591 * \retval 0 on success; < 0 if error happens
1593 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1594 struct lu_buf *buf, struct lu_buf *buf_vic)
1596 struct lov_comp_md_v1 *comp_rem;
1597 struct lov_comp_md_v1 *comp_vic;
1598 struct lov_comp_md_entry_v1 *entry;
1599 struct lov_comp_md_entry_v1 *entry_rem;
1600 struct lov_comp_md_entry_v1 *entry_vic;
1602 __u16 comp_cnt, count = 0;
1603 int lmm_size, lmm_size_vic = 0;
1605 int offset, offset_rem, offset_vic;
1607 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1608 /* comp_v1 should contains more than 1 mirror */
1609 if (mirror_cnt <= 1)
1611 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1612 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1614 for (i = 0; i < comp_cnt; i++) {
1615 entry = &comp_v1->lcm_entries[i];
1616 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1618 lmm_size_vic += sizeof(*entry);
1619 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1620 } else if (count > 0) {
1621 /* find the specified mirror */
1629 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1633 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1634 if (!buf_vic->lb_buf) {
1639 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1640 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1642 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1643 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1644 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1645 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1646 if (!comp_rem->lcm_mirror_count)
1647 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1649 memset(comp_vic, 0, sizeof(*comp_v1));
1650 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1651 comp_vic->lcm_mirror_count = 0;
1652 comp_vic->lcm_entry_count = cpu_to_le32(count);
1653 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1654 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1655 comp_vic->lcm_layout_gen = 0;
1657 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1658 offset_rem = sizeof(*comp_rem) +
1659 sizeof(*entry_rem) * (comp_cnt - count);
1660 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1661 for (i = j = k = 0; i < comp_cnt; i++) {
1662 struct lov_mds_md *lmm, *lmm_dst;
1665 entry = &comp_v1->lcm_entries[i];
1666 entry_vic = &comp_vic->lcm_entries[j];
1667 entry_rem = &comp_rem->lcm_entries[k];
1669 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1672 /* copy component entry */
1674 memcpy(entry_vic, entry, sizeof(*entry));
1675 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1676 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1679 memcpy(entry_rem, entry, sizeof(*entry));
1680 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1684 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1686 lmm_dst = (struct lov_mds_md *)
1687 ((char *)comp_vic + offset_vic);
1689 lmm_dst = (struct lov_mds_md *)
1690 ((char *)comp_rem + offset_rem);
1692 /* copy component entry blob */
1693 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1695 /* blob offset advance */
1696 offset += le32_to_cpu(entry->lcme_size);
1698 offset_vic += le32_to_cpu(entry->lcme_size);
1700 offset_rem += le32_to_cpu(entry->lcme_size);
1706 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1707 struct md_rejig_data *mrd)
1709 struct mdd_device *mdd = mdo2mdd(md_obj);
1710 struct mdd_object *obj = md2mdd_obj(md_obj);
1711 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1712 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1713 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1714 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1715 struct lov_comp_md_v1 *lcm;
1716 struct thandle *handle;
1720 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1721 if (rc == 0) /* same fid */
1724 handle = mdd_trans_create(env, mdd);
1726 RETURN(PTR_ERR(handle));
1729 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1730 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1732 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1733 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1736 /* get EA of mirrored file */
1737 memset(buf_save, 0, sizeof(*buf));
1738 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1742 lcm = buf_save->lb_buf;
1743 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1744 GOTO(out, rc = -EINVAL);
1747 * Extract the mirror with specified mirror id, and store the splitted
1748 * mirror layout to the victim file.
1750 memset(buf, 0, sizeof(*buf));
1751 memset(buf_vic, 0, sizeof(*buf_vic));
1752 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1756 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1757 LU_XATTR_SPLIT, handle);
1760 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1761 LU_XATTR_SPLIT, handle);
1765 rc = mdd_trans_start(env, mdd, handle);
1769 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1774 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1777 GOTO(out_restore, rc);
1779 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1783 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1790 /* restore obj's layout */
1791 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1792 LU_XATTR_REPLACE, handle);
1794 CERROR("%s: failed to rollback of layout of: "DFID
1795 ": %d, file state unkonwn.\n",
1796 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1799 mdd_trans_stop(env, mdd, rc, handle);
1800 mdd_write_unlock(env, obj);
1801 mdd_write_unlock(env, vic);
1802 lu_buf_free(buf_save);
1804 lu_buf_free(buf_vic);
1807 (void) mdd_object_pfid_replace(env, obj);
1812 static int mdd_layout_merge_allowed(const struct lu_env *env,
1813 struct md_object *target,
1814 struct md_object *victim)
1816 struct mdd_object *o1 = md2mdd_obj(target);
1818 /* cannot extend directory's LOVEA */
1819 if (S_ISDIR(mdd_object_type(o1))) {
1820 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1821 mdd_obj_dev_name(o1));
1829 * The caller should guarantee to update the object ctime
1830 * after xattr_set if needed.
1832 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1833 const struct lu_buf *buf, const char *name,
1836 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1837 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1838 struct mdd_device *mdd = mdo2mdd(obj);
1839 struct thandle *handle;
1840 enum changelog_rec_type cl_type;
1841 enum changelog_rec_flags clf_flags = 0;
1845 rc = mdd_la_get(env, mdd_obj, attr);
1849 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1853 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1854 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1855 struct md_rejig_data *mrd = buf->lb_buf;
1856 struct md_object *victim = mrd->mrd_obj;
1858 if (buf->lb_len != sizeof(*mrd))
1861 rc = mdd_layout_merge_allowed(env, obj, victim);
1865 if (fl == LU_XATTR_MERGE)
1866 /* merge layout of victim as a mirror of obj's. */
1867 rc = mdd_xattr_merge(env, obj, victim);
1869 rc = mdd_xattr_split(env, obj, mrd);
1873 if (strcmp(name, XATTR_NAME_LMV) == 0) {
1874 rc = mdd_dir_layout_shrink(env, obj, buf);
1878 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1879 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1880 struct posix_acl *acl;
1882 /* user may set empty ACL, which should be treated as removing
1884 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1887 RETURN(PTR_ERR(acl));
1889 rc = mdd_xattr_del(env, obj, name);
1892 posix_acl_release(acl);
1895 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1896 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1900 handle = mdd_trans_create(env, mdd);
1902 RETURN(PTR_ERR(handle));
1904 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1908 rc = mdd_trans_start(env, mdd, handle);
1912 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1914 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1915 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1917 mdd_write_unlock(env, mdd_obj);
1922 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1923 mdd_write_unlock(env, mdd_obj);
1927 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1931 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1932 mdd_obj, name, handle);
1936 return mdd_trans_stop(env, mdd, rc, handle);
1939 static int mdd_declare_xattr_del(const struct lu_env *env,
1940 struct mdd_device *mdd,
1941 struct mdd_object *obj,
1943 struct thandle *handle)
1945 enum changelog_rec_type type;
1948 rc = mdo_declare_xattr_del(env, obj, name, handle);
1952 type = mdd_xattr_changelog_type(env, mdd, name);
1954 return 0; /* no changelog to store */
1956 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1960 * The caller should guarantee to update the object ctime
1961 * after xattr_set if needed.
1963 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1966 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1967 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1968 struct mdd_device *mdd = mdo2mdd(obj);
1969 struct thandle *handle;
1973 rc = mdd_la_get(env, mdd_obj, attr);
1977 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1981 handle = mdd_trans_create(env, mdd);
1983 RETURN(PTR_ERR(handle));
1985 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1989 rc = mdd_trans_start(env, mdd, handle);
1993 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1994 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1995 mdd_write_unlock(env, mdd_obj);
1999 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2002 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2007 return mdd_trans_stop(env, mdd, rc, handle);
2011 * read lov/lmv EA of an object
2012 * return the lov/lmv EA in an allocated lu_buf
2014 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2015 struct lu_buf *lmm_buf, const char *name)
2017 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2022 if (buf->lb_buf == NULL) {
2023 buf = lu_buf_check_and_alloc(buf, 4096);
2024 if (buf->lb_buf == NULL)
2029 rc = mdo_xattr_get(env, obj, buf, name);
2030 if (rc == -ERANGE) {
2031 /* mti_big_buf is allocated but is too small
2032 * we need to increase it */
2033 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2035 if (buf->lb_buf == NULL)
2038 } else if (rc < 0) {
2040 } else if (rc == 0) {
2044 lu_buf_alloc(lmm_buf, rc);
2045 if (lmm_buf->lb_buf == NULL)
2049 * we don't use lmm_buf directly, because we don't know xattr size, so
2050 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2052 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2057 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2058 struct mdd_object *o, struct lu_buf *buf,
2059 struct thandle *handle)
2061 struct hsm_attrs *attrs;
2062 enum hsm_states hsm_flags;
2063 enum changelog_rec_flags clf_flags = 0;
2067 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2072 attrs = buf->lb_buf;
2073 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2074 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2077 /* Add a changelog record for release. */
2078 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2079 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2080 clf_flags, o, handle);
2085 * check if layout swapping between 2 objects is allowed
2087 * - only normal FIDs or non-system IGIFs
2088 * - same type of objects
2089 * - same owner/group (so quotas are still valid) unless this is from HSM
2092 static int mdd_layout_swap_allowed(const struct lu_env *env,
2093 struct mdd_object *o1,
2094 const struct lu_attr *attr1,
2095 struct mdd_object *o2,
2096 const struct lu_attr *attr2,
2099 const struct lu_fid *fid1, *fid2;
2105 if (!fid_is_norm(fid1) &&
2106 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2109 if (!fid_is_norm(fid2) &&
2110 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2113 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2114 if (S_ISDIR(mdd_object_type(o1)))
2116 if (S_ISREG(mdd_object_type(o1)))
2121 if (flags & SWAP_LAYOUTS_MDS_HSM)
2124 if ((attr1->la_uid != attr2->la_uid) ||
2125 (attr1->la_gid != attr2->la_gid))
2131 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2132 * look into the layout in MDD layer. */
2133 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2135 struct lov_comp_md_v1 *comp_v1;
2136 struct lov_mds_md *v1;
2140 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2141 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2142 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2148 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2149 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2152 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2153 off = le32_to_cpu(comp_v1->lcm_entries[i].
2155 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2160 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2161 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2172 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2174 return mdd_lmm_oi(lmm, oi, true);
2177 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2179 return mdd_lmm_oi(lmm, oi, false);
2182 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2184 struct lov_comp_md_v1 *comp_v1;
2186 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2187 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2189 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2191 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2192 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2193 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2194 __u16 tmp_gen = *gen;
2196 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2198 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2205 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2207 return mdd_lmm_gen(lmm, gen, true);
2210 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2212 return mdd_lmm_gen(lmm, gen, false);
2215 static int mdd_dom_data_truncate(const struct lu_env *env,
2216 struct mdd_device *mdd, struct mdd_object *mo)
2219 struct dt_object *dom;
2222 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2224 GOTO(out, rc = -ENODATA);
2226 th = dt_trans_create(env, mdd->mdd_bottom);
2228 GOTO(out, rc = PTR_ERR(th));
2230 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2234 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2238 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2240 dt_trans_stop(env, mdd->mdd_bottom, th);
2242 /* Ignore failure but report the error */
2244 CERROR("%s: "DFID" can't truncate DOM inode data, rc = %d\n",
2245 mdd_obj_dev_name(mo), PFID(mdo2fid(mo)), rc);
2250 * swap layouts between 2 lustre objects
2252 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2253 struct md_object *obj2, __u64 flags)
2255 struct mdd_thread_info *info = mdd_env_info(env);
2256 struct mdd_object *fst_o = md2mdd_obj(obj1);
2257 struct mdd_object *snd_o = md2mdd_obj(obj2);
2258 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2259 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2260 struct mdd_device *mdd = mdo2mdd(obj1);
2261 struct lov_mds_md *fst_lmm, *snd_lmm;
2262 struct lu_buf *fst_buf = &info->mti_buf[0];
2263 struct lu_buf *snd_buf = &info->mti_buf[1];
2264 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2265 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2266 struct ost_id *saved_oi = NULL;
2267 struct thandle *handle;
2268 struct mdd_object *dom_o = NULL;
2269 __u64 domsize_dom, domsize_vlt;
2270 __u32 fst_gen, snd_gen, saved_gen;
2276 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
2277 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2279 /* we have to sort the 2 obj, so locking will always
2280 * be in the same order, even in case of 2 concurrent swaps */
2281 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
2282 if (rc == 0) /* same fid ? */
2288 rc = mdd_la_get(env, fst_o, fst_la);
2292 rc = mdd_la_get(env, snd_o, snd_la);
2296 /* check if layout swapping is allowed */
2297 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2301 handle = mdd_trans_create(env, mdd);
2303 RETURN(PTR_ERR(handle));
2305 /* objects are already sorted */
2306 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
2307 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
2309 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2310 if (rc < 0 && rc != -ENODATA)
2313 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2314 if (rc < 0 && rc != -ENODATA)
2317 /* check if file has DoM. DoM file can be migrated only to another
2318 * DoM layout with the same DoM component size or to an non-DOM
2319 * layout. After migration to OSTs layout, local MDT inode data
2320 * should be truncated.
2321 * Objects are sorted by FIDs, considering that original file's FID
2322 * is always smaller the snd_o is always original file we are migrating
2325 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2326 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2328 /* Only migration is supported for DoM files, not 'swap_layouts' so
2329 * target file must be volatile and orphan.
2331 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2332 dom_o = domsize_dom ? snd_o : NULL;
2333 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2334 swap(domsize_dom, domsize_vlt);
2335 dom_o = domsize_dom ? fst_o : NULL;
2336 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2337 /* 'lfs swap_layouts' case, neither file should have DoM */
2339 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2340 "use migration instead: rc = %d\n", rc);
2344 if (domsize_vlt > 0 && domsize_dom == 0) {
2346 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": OST to DOM "
2347 "migration is not supported: rc = %d\n",
2348 PFID(mdo2fid(snd_o)), rc);
2350 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2352 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": new layout "
2353 "must have the same DoM component size: rc = %d\n",
2354 PFID(mdo2fid(fst_o)), rc);
2356 } else if (domsize_vlt > 0) {
2357 /* Migration with the same DOM component size, no need to
2358 * truncate local data, it is still being used */
2362 /* swapping 2 non existant layouts is a success */
2363 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2366 /* to help inode migration between MDT, it is better to
2367 * start by the no layout file (if one), so we order the swap */
2368 if (snd_buf->lb_buf == NULL) {
2370 swap(fst_buf, snd_buf);
2373 fst_gen = snd_gen = 0;
2374 /* lmm and generation layout initialization */
2375 if (fst_buf->lb_buf != NULL) {
2376 fst_lmm = fst_buf->lb_buf;
2377 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2378 fst_fl = LU_XATTR_REPLACE;
2382 fst_fl = LU_XATTR_CREATE;
2385 snd_lmm = snd_buf->lb_buf;
2386 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2388 saved_gen = fst_gen;
2389 /* increase the generation layout numbers */
2394 * XXX The layout generation is used to generate component IDs for
2395 * the composite file, we have to do some special tweaks to make
2396 * sure the layout generation is always adequate for that job.
2399 /* Skip invalid generation number for composite layout */
2400 if ((snd_gen & LCME_ID_MASK) == 0)
2402 if ((fst_gen & LCME_ID_MASK) == 0)
2404 /* Make sure the generation is greater than all the component IDs */
2405 if (fst_gen < snd_gen)
2407 else if (fst_gen > snd_gen)
2410 /* set the file specific informations in lmm */
2411 if (fst_lmm != NULL) {
2412 saved_oi = &info->mti_oa.o_oi;
2413 mdd_get_lmm_oi(fst_lmm, saved_oi);
2414 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2415 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2416 mdd_set_lmm_oi(snd_lmm, saved_oi);
2418 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2419 cpu_to_le32(LOV_MAGIC_MAGIC))
2420 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2422 GOTO(stop, rc = -EPROTO);
2424 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2426 /* Prepare HSM attribute if it's required */
2427 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2428 const int buflen = sizeof(struct hsm_attrs);
2430 lu_buf_alloc(fst_hsm_buf, buflen);
2431 lu_buf_alloc(snd_hsm_buf, buflen);
2432 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2433 GOTO(stop, rc = -ENOMEM);
2435 /* Read HSM attribute */
2436 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2440 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2444 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2445 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2450 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2451 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2457 /* prepare transaction */
2458 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2463 if (fst_buf->lb_buf != NULL)
2464 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2465 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2468 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2473 rc = mdd_trans_start(env, mdd, handle);
2477 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2478 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2482 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2484 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2487 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
2488 mdd_obj_dev_name(fst_o),
2489 PFID(mdo2fid(fst_o)), rc, rc2);
2494 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2498 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2501 if (fst_buf->lb_buf != NULL)
2502 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2503 LU_XATTR_REPLACE, handle);
2505 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2508 GOTO(out_restore, rc);
2510 /* Issue one changelog record per file */
2511 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2515 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2524 /* failure on second file, but first was done, so we have
2525 * to roll back first. */
2526 if (fst_buf->lb_buf != NULL) {
2527 mdd_set_lmm_oi(fst_lmm, saved_oi);
2528 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2529 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2530 LU_XATTR_REPLACE, handle);
2532 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2538 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2543 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2548 CERROR("%s: unable to roll back layout swap. FIDs: "
2549 DFID" and "DFID "error: %d/%d, steps: %d\n",
2550 mdd_obj_dev_name(fst_o),
2551 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
2553 /* a solution to avoid journal commit is to panic,
2554 * but it has strong consequences so we use LBUG to
2555 * allow sysdamin to choose to panic or not
2562 rc = mdd_trans_stop(env, mdd, rc, handle);
2564 /* Truncate local DOM data if all went well */
2566 mdd_dom_data_truncate(env, mdd, dom_o);
2568 mdd_write_unlock(env, snd_o);
2569 mdd_write_unlock(env, fst_o);
2571 lu_buf_free(fst_buf);
2572 lu_buf_free(snd_buf);
2573 lu_buf_free(fst_hsm_buf);
2574 lu_buf_free(snd_hsm_buf);
2577 (void) mdd_object_pfid_replace(env, fst_o);
2578 (void) mdd_object_pfid_replace(env, snd_o);
2583 static int mdd_declare_layout_change(const struct lu_env *env,
2584 struct mdd_device *mdd,
2585 struct mdd_object *obj,
2586 struct md_layout_change *mlc,
2587 struct thandle *handle)
2591 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2595 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2599 /* For PFL, this is used to instantiate necessary component objects. */
2601 mdd_layout_instantiate_component(const struct lu_env *env,
2602 struct mdd_object *obj, struct md_layout_change *mlc,
2603 struct thandle *handle)
2605 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2609 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2612 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2614 * It's possible that another layout write intent has already
2615 * instantiated our objects, so a -EALREADY returned, and we need to
2619 RETURN(rc == -EALREADY ? 0 : rc);
2621 rc = mdd_trans_start(env, mdd, handle);
2625 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2626 rc = mdo_layout_change(env, obj, mlc, handle);
2627 mdd_write_unlock(env, obj);
2631 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2636 * Change the FLR layout from RDONLY to WRITE_PENDING.
2638 * It picks the primary mirror, and bumps the layout version, and set
2639 * layout version xattr to OST objects in a sync tx. In order to facilitate
2640 * the handling of phantom writers from evicted clients, the clients carry
2641 * layout version of the file with write RPC, so that the OSTs can verify
2642 * if the write RPCs are legitimate, meaning not from evicted clients.
2645 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2646 struct md_layout_change *mlc, struct thandle *handle)
2648 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2649 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2650 struct lustre_som_attrs *som = &mlc->mlc_som;
2655 /* Verify acceptable operations */
2656 switch (mlc->mlc_opc) {
2657 case MD_LAYOUT_WRITE:
2658 case MD_LAYOUT_RESYNC:
2659 /* these are legal operations - this represents the case that
2660 * a few mirrors were missed in the last resync. */
2662 case MD_LAYOUT_RESYNC_DONE:
2667 som_buf->lb_buf = som;
2668 som_buf->lb_len = sizeof(*som);
2669 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2670 if (rc < 0 && rc != -ENODATA)
2674 lustre_som_swab(som);
2675 if (som->lsa_valid & SOM_FL_STRICT)
2676 fl = LU_XATTR_REPLACE;
2679 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2684 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2685 XATTR_NAME_SOM, fl, handle);
2690 /* record a changelog for data mover to consume */
2691 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2695 rc = mdd_trans_start(env, mdd, handle);
2699 /* it needs a sync tx to make FLR to work properly */
2700 handle->th_sync = 1;
2702 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2703 rc = mdo_layout_change(env, obj, mlc, handle);
2705 /* SOM state transition from STRICT to STALE */
2706 som->lsa_valid = SOM_FL_STALE;
2707 lustre_som_swab(som);
2708 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2711 mdd_write_unlock(env, obj);
2715 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2726 * Handle mirrored file state transition when it's in WRITE_PENDING.
2728 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2729 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2730 * layout version will be increased, and the file's state will be changed to
2734 mdd_layout_update_write_pending(const struct lu_env *env,
2735 struct mdd_object *obj, struct md_layout_change *mlc,
2736 struct thandle *handle)
2738 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2742 switch (mlc->mlc_opc) {
2743 case MD_LAYOUT_RESYNC:
2744 /* Upon receiving the resync request, it should
2745 * instantiate all stale components right away to get ready
2746 * for mirror copy. In order to avoid layout version change,
2747 * client should avoid sending LAYOUT_WRITE request at the
2750 case MD_LAYOUT_WRITE:
2751 /* legal race for concurrent write, the file state has been
2752 * changed by another client. */
2758 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2762 rc = mdd_trans_start(env, mdd, handle);
2766 /* it needs a sync tx to make FLR to work properly */
2767 handle->th_sync = 1;
2769 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2770 rc = mdo_layout_change(env, obj, mlc, handle);
2771 mdd_write_unlock(env, obj);
2782 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2784 * Only concurrent write and sync complete requests are possible when the
2785 * file is in SYNC_PENDING. For the latter request, it will pass in the
2786 * mirrors that have been synchronized, then the stale bit will be cleared
2787 * to make the file's state turn into RDONLY.
2788 * For concurrent write reqeust, it just needs to change the file's state
2789 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2790 * version because the version will be increased in the transition to
2791 * SYNC_PENDING later so that it can deny the write request from potential
2792 * evicted SYNC clients. */
2794 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2795 struct md_layout_change *mlc, struct thandle *handle)
2797 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2798 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2803 /* operation validation */
2804 switch (mlc->mlc_opc) {
2805 case MD_LAYOUT_RESYNC_DONE:
2806 /* resync complete. */
2807 case MD_LAYOUT_WRITE:
2808 /* concurrent write. */
2810 case MD_LAYOUT_RESYNC:
2811 /* resync again, most likely the previous run failed.
2812 * no-op if it's already in SYNC_PENDING state */
2818 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2819 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2820 if (rc < 0 && rc != -ENODATA)
2823 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2824 lustre_som_swab(&mlc->mlc_som);
2825 som_buf->lb_buf = &mlc->mlc_som;
2826 som_buf->lb_len = sizeof(mlc->mlc_som);
2829 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2833 /* record a changelog for the completion of resync */
2834 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2839 /* RESYNC_DONE has piggybacked size and blocks */
2841 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2842 XATTR_NAME_SOM, fl, handle);
2847 rc = mdd_trans_start(env, mdd, handle);
2851 /* it needs a sync tx to make FLR to work properly */
2852 handle->th_sync = 1;
2854 rc = mdo_layout_change(env, obj, mlc, handle);
2859 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2865 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2874 * Layout change callback for object.
2876 * This is only used by FLR for now. In the future, it can be exteneded to
2877 * handle all layout change.
2880 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2881 struct md_layout_change *mlc)
2883 struct mdd_object *obj = md2mdd_obj(o);
2884 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2885 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2886 struct lov_comp_md_v1 *lcm;
2887 struct thandle *handle;
2892 /* Verify acceptable operations */
2893 switch (mlc->mlc_opc) {
2894 case MD_LAYOUT_WRITE:
2895 case MD_LAYOUT_RESYNC:
2896 case MD_LAYOUT_RESYNC_DONE:
2902 handle = mdd_trans_create(env, mdd);
2904 RETURN(PTR_ERR(handle));
2906 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2913 /* analyze the layout to make sure it's a FLR file */
2915 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2916 GOTO(out, rc = -EINVAL);
2918 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2920 /* please refer to HLD of FLR for state transition */
2921 switch (flr_state) {
2923 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2925 case LCM_FL_WRITE_PENDING:
2926 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2929 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2931 case LCM_FL_SYNC_PENDING:
2932 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2941 mdd_trans_stop(env, mdd, rc, handle);
2946 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2947 struct mdd_object *child, const struct lu_attr *attr,
2948 const struct md_op_spec *spec,
2949 struct dt_allocation_hint *hint)
2951 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2952 struct dt_object *nc = mdd_object_child(child);
2954 memset(hint, 0, sizeof(*hint));
2956 /* For striped directory, give striping EA to lod_ah_init, which will
2957 * decide the stripe_offset and stripe count by it. */
2958 if (S_ISDIR(attr->la_mode) &&
2959 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2960 hint->dah_eadata = spec->u.sp_ea.eadata;
2961 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2963 hint->dah_eadata = NULL;
2964 hint->dah_eadata_len = 0;
2967 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2968 hint->dah_eadata, hint->dah_eadata_len);
2969 /* @hint will be initialized by underlying device. */
2970 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2973 static int accmode(const struct lu_env *env, const struct lu_attr *la,
2976 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2977 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2978 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2979 * owner can write to a file even if it is marked readonly to hide
2980 * its brokenness. (bug 5781) */
2981 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
2982 struct lu_ucred *uc = lu_ucred_check(env);
2984 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2988 return mds_accmode(open_flags);
2991 static int mdd_open_sanity_check(const struct lu_env *env,
2992 struct mdd_object *obj,
2993 const struct lu_attr *attr, u64 open_flags)
2998 /* EEXIST check, also opening of *open* orphans is allowed so we can
2999 * open-by-handle unlinked files
3001 if (mdd_is_dead_obj(obj) &&
3002 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3005 if (S_ISLNK(attr->la_mode))
3008 mode = accmode(env, attr, open_flags);
3010 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3013 if (!(open_flags & MDS_OPEN_CREATED)) {
3014 rc = mdd_permission_internal(env, obj, attr, mode);
3019 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3020 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3021 open_flags &= ~MDS_OPEN_TRUNC;
3023 /* For writing append-only file must open it with append mode. */
3024 if (attr->la_flags & LUSTRE_APPEND_FL) {
3025 if ((open_flags & MDS_FMODE_WRITE) &&
3026 !(open_flags & MDS_OPEN_APPEND))
3028 if (open_flags & MDS_OPEN_TRUNC)
3035 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3038 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3039 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3040 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3041 struct mdd_object_user *mou = NULL;
3042 const struct lu_ucred *uc = lu_ucred(env);
3043 struct mdd_device *mdd = mdo2mdd(obj);
3044 enum changelog_rec_type type = CL_OPEN;
3047 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3049 rc = mdd_la_get(env, mdd_obj, attr);
3053 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags);
3054 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3059 mdd_obj->mod_count++;
3061 if (!mdd_changelog_enabled(env, mdd, type))
3065 /* look for existing opener in list under mdd_write_lock */
3066 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3071 /* add user to list */
3072 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3078 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3080 mdd_obj_user_free(mou);
3085 if (type == CL_DN_OPEN) {
3086 if (ktime_before(ktime_get(), mou->mou_deniednext))
3087 /* same user denied again same access within
3088 * time interval: do not record
3092 /* this user already denied, but some time ago:
3093 * update denied time
3095 mou->mou_deniednext =
3096 ktime_add(ktime_get(),
3097 ktime_set(mdd->mdd_cl.mc_deniednext,
3100 mou->mou_opencount++;
3101 /* same user opening file again with same flags:
3108 /* FYI, only the bottom 32 bits of open_flags are recorded */
3109 mdd_changelog(env, type, open_flags, md_dev, mdo2fid(mdd_obj));
3113 mdd_write_unlock(env, mdd_obj);
3117 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3118 struct md_attr *ma, struct thandle *handle)
3122 rc = mdd_orphan_declare_delete(env, obj, handle);
3126 return mdo_declare_destroy(env, obj, handle);
3130 * No permission check is needed.
3132 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3133 struct md_attr *ma, u64 open_flags)
3135 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3136 struct mdd_device *mdd = mdo2mdd(obj);
3137 struct thandle *handle = NULL;
3140 bool blocked = false;
3141 bool last_close_by_uid = false;
3142 const struct lu_ucred *uc = lu_ucred(env);
3145 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3146 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3147 mdd_obj->mod_count--;
3148 mdd_write_unlock(env, mdd_obj);
3150 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3151 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3152 "list\n", PFID(mdd_object_fid(mdd_obj)));
3156 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3157 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3158 /* check without any lock */
3159 is_orphan = mdd_obj->mod_count == 1 &&
3160 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3164 /* mdd_trans_create() maybe failed because of barrier_entry(),
3165 * under such case, the orphan MDT-object will be left in the
3166 * orphan list, and when the MDT remount next time, the unused
3167 * orphans will be destroyed automatically.
3169 * One exception: the former mdd_finish_unlink may failed to
3170 * add the orphan MDT-object to the orphan list, then if the
3171 * mdd_trans_create() failed because of barrier_entry(), the
3172 * MDT-object will become real orphan that is neither in the
3173 * namespace nor in the orphan list. Such bad case should be
3174 * very rare and will be handled by e2fsck/lfsck. */
3175 handle = mdd_trans_create(env, mdo2mdd(obj));
3176 if (IS_ERR(handle)) {
3177 rc = PTR_ERR(handle);
3178 if (rc != -EINPROGRESS)
3186 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3190 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3195 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3201 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
3202 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3204 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3205 lu_dev_name(mdd2lu_dev(mdd)),
3206 PFID(mdd_object_fid(mdd_obj)), rc);
3210 /* check again with lock */
3211 is_orphan = (mdd_obj->mod_count == 1) &&
3212 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3213 ma->ma_attr.la_nlink == 0);
3215 if (is_orphan && !handle && !blocked) {
3216 mdd_write_unlock(env, mdd_obj);
3220 mdd_obj->mod_count--; /*release open count */
3222 /* under mdd write lock */
3223 /* If recording, see if we need to remove UID from list. uc is not
3224 * initialized if the client has been evicted. */
3225 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3226 struct mdd_object_user *mou;
3228 /* look for UID in list */
3229 /* If mou is NULL, it probably means logging was enabled after
3230 * the user had the file open. So the corresponding close
3231 * will not be logged.
3233 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3236 mou->mou_opencount--;
3237 if (mou->mou_opencount == 0) {
3238 mdd_obj_user_remove(mdd_obj, mou);
3239 last_close_by_uid = true;
3244 if (!is_orphan || blocked)
3248 /* NB: Object maybe not in orphan list originally, it is rare case for
3249 * mdd_finish_unlink() failure, in that case, the object doesn't have
3250 * ORPHAN_OBJ flag */
3251 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3252 /* remove link to object from orphan index */
3253 LASSERT(handle != NULL);
3254 rc = mdd_orphan_delete(env, mdd_obj, handle);
3256 CERROR("%s: unable to delete "DFID" from orphan list: "
3257 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3258 PFID(mdd_object_fid(mdd_obj)), rc);
3259 /* If object was not deleted from orphan list, do not
3260 * destroy OSS objects, which will be done when next
3265 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3266 "list, OSS objects to be destroyed.\n",
3267 PFID(mdd_object_fid(mdd_obj)));
3270 rc = mdo_destroy(env, mdd_obj, handle);
3273 CERROR("%s: unable to delete "DFID" from orphan list: "
3274 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3275 PFID(mdd_object_fid(mdd_obj)), rc);
3280 mdd_write_unlock(env, mdd_obj);
3282 if (rc != 0 || blocked ||
3283 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3286 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3287 * or if CL_OPEN was recorded and it's last close by user.
3288 * Changelogs mask may change between open and close operations, but
3289 * this is not a big deal if we have a CL_CLOSE entry with no matching
3290 * CL_OPEN. Plus Changelogs mask may not change often.
3292 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3293 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3294 MDS_OPEN_TRUNC))) ||
3295 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3296 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3297 if (handle == NULL) {
3298 handle = mdd_trans_create(env, mdo2mdd(obj));
3300 GOTO(stop, rc = PTR_ERR(handle));
3302 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3303 NULL, NULL, handle);
3307 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3312 /* FYI, only the bottom 32 bits of open_flags are recorded */
3313 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3318 if (handle != NULL && !IS_ERR(handle))
3319 rc = mdd_trans_stop(env, mdd, rc, handle);
3325 * Permission check is done when open,
3326 * no need check again.
3328 static int mdd_readpage_sanity_check(const struct lu_env *env,
3329 struct mdd_object *obj)
3331 struct dt_object *next = mdd_object_child(obj);
3335 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3343 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3344 size_t nob, const struct dt_it_ops *iops,
3345 struct dt_it *it, __u32 attr, void *arg)
3347 struct lu_dirpage *dp = &lp->lp_dir;
3351 struct lu_dirent *ent;
3352 struct lu_dirent *last = NULL;
3356 if (nob < sizeof(*dp))
3359 memset(area, 0, sizeof (*dp));
3360 area += sizeof (*dp);
3361 nob -= sizeof (*dp);
3368 len = iops->key_size(env, it);
3370 /* IAM iterator can return record with zero len. */
3374 hash = iops->store(env, it);
3375 if (unlikely(first)) {
3377 dp->ldp_hash_start = cpu_to_le64(hash);
3380 /* calculate max space required for lu_dirent */
3381 recsize = lu_dirent_calc_size(len, attr);
3383 if (nob >= recsize) {
3384 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3385 if (result == -ESTALE)
3390 /* osd might not able to pack all attributes,
3391 * so recheck rec length */
3392 recsize = le16_to_cpu(ent->lde_reclen);
3394 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3395 fid_le_to_cpu(&fid, &ent->lde_fid);
3396 if (fid_is_dot_lustre(&fid))
3400 result = (last != NULL) ? 0 :-EINVAL;
3404 ent = (void *)ent + recsize;
3408 result = iops->next(env, it);
3409 if (result == -ESTALE)
3411 } while (result == 0);
3414 dp->ldp_hash_end = cpu_to_le64(hash);
3416 if (last->lde_hash == dp->ldp_hash_end)
3417 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3418 last->lde_reclen = 0; /* end mark */
3421 /* end of directory */
3422 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3423 else if (result < 0)
3424 CWARN("build page failed: %d!\n", result);
3428 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3429 const struct lu_rdpg *rdpg)
3431 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3435 if (mdd_object_exists(mdd_obj) == 0) {
3436 CERROR("%s: object "DFID" not found: rc = -2\n",
3437 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3441 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
3442 rc = mdd_readpage_sanity_check(env, mdd_obj);
3444 GOTO(out_unlock, rc);
3446 if (mdd_is_dead_obj(mdd_obj)) {
3448 struct lu_dirpage *dp;
3451 * According to POSIX, please do not return any entry to client:
3452 * even dot and dotdot should not be returned.
3454 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3455 PFID(mdd_object_fid(mdd_obj)));
3457 if (rdpg->rp_count <= 0)
3458 GOTO(out_unlock, rc = -EFAULT);
3459 LASSERT(rdpg->rp_pages != NULL);
3461 pg = rdpg->rp_pages[0];
3462 dp = (struct lu_dirpage *)kmap(pg);
3463 memset(dp, 0 , sizeof(struct lu_dirpage));
3464 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3465 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3466 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3468 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3471 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3472 mdd_dir_page_build, NULL);
3474 struct lu_dirpage *dp;
3476 dp = kmap(rdpg->rp_pages[0]);
3477 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3480 * No pages were processed, mark this for first page
3483 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3484 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3485 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3487 kunmap(rdpg->rp_pages[0]);
3490 GOTO(out_unlock, rc);
3492 mdd_read_unlock(env, mdd_obj);
3496 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3498 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3500 if (mdd_object_exists(mdd_obj) == 0) {
3503 CERROR("%s: object "DFID" not found: rc = %d\n",
3504 mdd_obj_dev_name(mdd_obj),
3505 PFID(mdd_object_fid(mdd_obj)), rc);
3508 return dt_object_sync(env, mdd_object_child(mdd_obj),
3512 static int mdd_object_lock(const struct lu_env *env,
3513 struct md_object *obj,
3514 struct lustre_handle *lh,
3515 struct ldlm_enqueue_info *einfo,
3516 union ldlm_policy_data *policy)
3518 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3519 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3523 static int mdd_object_unlock(const struct lu_env *env,
3524 struct md_object *obj,
3525 struct ldlm_enqueue_info *einfo,
3526 union ldlm_policy_data *policy)
3528 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3529 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3532 const struct md_object_operations mdd_obj_ops = {
3533 .moo_permission = mdd_permission,
3534 .moo_attr_get = mdd_attr_get,
3535 .moo_attr_set = mdd_attr_set,
3536 .moo_xattr_get = mdd_xattr_get,
3537 .moo_xattr_set = mdd_xattr_set,
3538 .moo_xattr_list = mdd_xattr_list,
3539 .moo_invalidate = mdd_invalidate,
3540 .moo_xattr_del = mdd_xattr_del,
3541 .moo_swap_layouts = mdd_swap_layouts,
3542 .moo_open = mdd_open,
3543 .moo_close = mdd_close,
3544 .moo_readpage = mdd_readpage,
3545 .moo_readlink = mdd_readlink,
3546 .moo_changelog = mdd_changelog,
3547 .moo_object_sync = mdd_object_sync,
3548 .moo_object_lock = mdd_object_lock,
3549 .moo_object_unlock = mdd_object_unlock,
3550 .moo_layout_change = mdd_layout_change,