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 return lu_env_info(env, &mdd_thread_key);
249 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
253 buf = &mdd_env_info(env)->mti_buf[0];
259 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
260 const void *area, ssize_t len)
264 buf = &mdd_env_info(env)->mti_buf[0];
265 buf->lb_buf = (void *)area;
270 struct lu_object *mdd_object_alloc(const struct lu_env *env,
271 const struct lu_object_header *hdr,
274 struct mdd_object *mdd_obj;
277 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
281 o = mdd2lu_obj(mdd_obj);
282 lu_object_init(o, NULL, d);
283 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
284 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
285 mdd_obj->mod_count = 0;
286 o->lo_ops = &mdd_lu_obj_ops;
287 INIT_LIST_HEAD(&mdd_obj->mod_users);
292 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
293 const struct lu_object_conf *unused)
295 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
296 struct mdd_object *mdd_obj = lu2mdd_obj(o);
297 struct lu_object *below;
298 struct lu_device *under;
301 mdd_obj->mod_cltime = ktime_set(0, 0);
302 under = &d->mdd_child->dd_lu_dev;
303 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
305 RETURN(PTR_ERR(below));
307 lu_object_add(o, below);
312 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
316 if (lu_object_exists(o)) {
317 struct mdd_object *mdd_obj = lu2mdd_obj(o);
318 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
320 rc = mdd_la_get(env, mdd_obj, attr);
326 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
328 struct mdd_object *mdd = lu2mdd_obj(o);
329 struct mdd_object_user *mou, *tmp2;
332 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
333 list_del(&mou->mou_list);
334 mdd_obj_user_free(mou);
338 /* mdd doesn't contain an lu_object_header, so don't need call_rcu */
339 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
342 static int mdd_object_print(const struct lu_env *env, void *cookie,
343 lu_printer_t p, const struct lu_object *o)
345 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
347 return (*p)(env, cookie,
348 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
349 mdd, mdd->mod_count, mdd->mod_valid,
350 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
353 static const struct lu_object_operations mdd_lu_obj_ops = {
354 .loo_object_init = mdd_object_init,
355 .loo_object_start = mdd_object_start,
356 .loo_object_free = mdd_object_free,
357 .loo_object_print = mdd_object_print,
360 struct mdd_object *mdd_object_find(const struct lu_env *env,
361 struct mdd_device *d,
362 const struct lu_fid *f)
364 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
368 * No permission check is needed.
370 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
373 struct mdd_object *mdd_obj = md2mdd_obj(obj);
378 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
379 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
380 ma->ma_attr.la_nlink = 0;
386 * No permission check is needed.
388 static int mdd_xattr_get(const struct lu_env *env,
389 struct md_object *obj, struct lu_buf *buf,
392 struct mdd_object *mdd_obj = md2mdd_obj(obj);
393 struct mdd_device *mdd;
398 if (mdd_object_exists(mdd_obj) == 0) {
399 CERROR("%s: object "DFID" not found: rc = -2\n",
400 mdd_obj_dev_name(mdd_obj),
401 PFID(mdd_object_fid(mdd_obj)));
405 /* If the object has been destroyed, then do not get LMVEA, because
406 * it needs to load stripes from the iteration of the master object,
407 * and it will cause problem if master object has been destroyed, see
409 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
410 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
411 strcmp(name, XATTR_NAME_LMV) == 0))
414 /* If the object has been delete from the namespace, then
415 * get linkEA should return -ENOENT as well */
416 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
417 strcmp(name, XATTR_NAME_LINK) == 0))
420 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
421 rc = mdo_xattr_get(env, mdd_obj, buf, name);
422 mdd_read_unlock(env, mdd_obj);
426 /* record only getting user xattrs and acls */
427 if (rc >= 0 && buf->lb_buf &&
428 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
429 (has_prefix(name, XATTR_USER_PREFIX) ||
430 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
431 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
432 struct thandle *handle;
435 LASSERT(mdd_object_fid(mdd_obj) != NULL);
437 handle = mdd_trans_create(env, mdd);
439 RETURN(PTR_ERR(handle));
441 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
446 rc2 = mdd_trans_start(env, mdd, handle);
450 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
451 mdd_object_fid(mdd_obj),
455 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
464 * Permission check is done when open,
465 * no need check again.
467 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
470 struct mdd_object *mdd_obj = md2mdd_obj(obj);
471 struct dt_object *next;
476 if (mdd_object_exists(mdd_obj) == 0) {
477 CERROR("%s: object "DFID" not found: rc = -2\n",
478 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
482 next = mdd_object_child(mdd_obj);
483 LASSERT(next != NULL);
484 LASSERT(next->do_body_ops != NULL);
485 LASSERT(next->do_body_ops->dbo_read != NULL);
486 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
487 rc = dt_read(env, next, buf, &pos);
488 mdd_read_unlock(env, mdd_obj);
493 * No permission check is needed.
495 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
498 struct mdd_object *mdd_obj = md2mdd_obj(obj);
503 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
504 rc = mdo_xattr_list(env, mdd_obj, buf);
505 mdd_read_unlock(env, mdd_obj);
507 /* If the buffer is NULL then we are only here to get the
508 * length of the xattr name list. */
509 if (rc < 0 || buf->lb_buf == NULL)
513 * Filter out XATTR_NAME_LINK if this is an orphan object. See
516 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
517 char *end = (char *)buf->lb_buf + rc;
518 char *p = buf->lb_buf;
521 char *next = p + strlen(p) + 1;
523 if (strcmp(p, XATTR_NAME_LINK) == 0) {
525 memmove(p, next, end - next);
527 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
528 " of orphan "DFID"\n",
529 PFID(mdd_object_fid(mdd_obj)));
540 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
542 return mdo_invalidate(env, md2mdd_obj(obj));
545 int mdd_declare_create_object_internal(const struct lu_env *env,
546 struct mdd_object *p,
547 struct mdd_object *c,
548 struct lu_attr *attr,
549 struct thandle *handle,
550 const struct md_op_spec *spec,
551 struct dt_allocation_hint *hint)
553 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
554 const struct dt_index_features *feat = spec->sp_feat;
558 if (feat != &dt_directory_features && feat != NULL) {
559 dof->dof_type = DFT_INDEX;
560 dof->u.dof_idx.di_feat = feat;
562 dof->dof_type = dt_mode_to_dft(attr->la_mode);
563 if (dof->dof_type == DFT_REGULAR) {
564 dof->u.dof_reg.striped =
565 md_should_create(spec->sp_cr_flags);
566 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
567 dof->u.dof_reg.striped = 0;
568 /* is this replay? */
570 dof->u.dof_reg.striped = 0;
574 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
579 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
580 struct mdd_object *c, struct lu_attr *attr,
581 struct thandle *handle,
582 const struct md_op_spec *spec,
583 struct dt_allocation_hint *hint)
585 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
589 LASSERT(!mdd_object_exists(c));
591 rc = mdo_create_object(env, c, attr, hint, dof, handle);
596 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
597 const struct lu_attr *attr, struct thandle *handle,
603 rc = mdo_attr_set(env, obj, attr, handle);
604 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
605 if (!rc && (attr->la_valid & LA_MODE) && needacl)
606 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
611 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
612 const struct lu_attr *oattr, struct lu_attr *attr,
613 struct thandle *handle)
618 LASSERT(attr->la_valid & LA_CTIME);
619 LASSERT(oattr != NULL);
621 /* Make sure the ctime is increased only, however, it's not strictly
622 * reliable at here because there is not guarantee to hold lock on
623 * object, so we just bypass some unnecessary cmtime setting first
624 * and OSD has to check it again. */
625 if (attr->la_ctime < oattr->la_ctime)
626 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
627 else if (attr->la_valid == LA_CTIME &&
628 attr->la_ctime == oattr->la_ctime)
629 attr->la_valid &= ~LA_CTIME;
631 if (attr->la_valid != 0)
632 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
637 static bool is_project_state_change(const struct lu_attr *oattr,
640 if (la->la_valid & LA_PROJID &&
641 oattr->la_projid != la->la_projid)
644 if ((la->la_valid & LA_FLAGS) &&
645 (la->la_flags & LUSTRE_PROJINHERIT_FL) !=
646 (oattr->la_flags & LUSTRE_PROJINHERIT_FL))
653 * This gives the same functionality as the code between
654 * sys_chmod and inode_setattr
655 * chown_common and inode_setattr
656 * utimes and inode_setattr
657 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
659 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
660 const struct lu_attr *oattr, struct lu_attr *la,
661 const struct md_attr *ma)
665 const unsigned long flags = ma->ma_attr_flags;
672 /* Do not permit change file type */
673 if (la->la_valid & LA_TYPE)
676 /* They should not be processed by setattr */
677 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
680 LASSERT(oattr != NULL);
682 uc = lu_ucred_check(env);
686 if (is_project_state_change(oattr, la)) {
687 if (!md_capable(uc, CFS_CAP_SYS_RESOURCE) &&
688 !lustre_in_group_p(uc, ma->ma_enable_chprojid_gid) &&
689 !(ma->ma_enable_chprojid_gid == -1 &&
690 mdd_permission_internal(env, obj, oattr, MAY_WRITE)))
694 if (la->la_valid == LA_CTIME) {
695 if (!(flags & MDS_PERM_BYPASS))
696 /* This is only for set ctime when rename's source is
698 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
700 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
701 la->la_valid &= ~LA_CTIME;
705 if (flags & MDS_CLOSE_UPDATE_TIMES &&
706 la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
707 /* This is an atime/mtime/ctime attribute update for
710 if (la->la_valid & LA_ATIME &&
711 la->la_atime <= (oattr->la_atime +
712 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
713 la->la_valid &= ~LA_ATIME;
714 if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime)
715 la->la_valid &= ~LA_CTIME;
716 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
717 la->la_valid &= ~LA_MTIME;
721 /* Check if flags change. */
722 if (la->la_valid & LA_FLAGS) {
723 unsigned int oldflags = oattr->la_flags &
724 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
725 unsigned int newflags = la->la_flags &
726 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
728 if ((uc->uc_fsuid != oattr->la_uid) &&
729 !md_capable(uc, CFS_CAP_FOWNER))
732 /* The IMMUTABLE and APPEND_ONLY flags can
733 * only be changed by the relevant capability. */
734 if ((oldflags ^ newflags) &&
735 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
738 if (!S_ISDIR(oattr->la_mode)) {
739 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
740 } else if (la->la_flags & LUSTRE_ENCRYPT_FL) {
741 /* when trying to add encryption flag on dir,
742 * make sure it is empty
744 rc = mdd_dir_is_empty(env, obj);
751 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
752 (la->la_valid & ~LA_FLAGS) &&
753 !(flags & MDS_PERM_BYPASS))
756 /* Check for setting the obj time. */
757 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
758 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
759 if ((uc->uc_fsuid != oattr->la_uid) &&
760 !md_capable(uc, CFS_CAP_FOWNER)) {
761 rc = mdd_permission_internal(env, obj, oattr,
768 if (la->la_valid & LA_KILL_SUID) {
769 la->la_valid &= ~LA_KILL_SUID;
770 if ((oattr->la_mode & S_ISUID) &&
771 !(la->la_valid & LA_MODE)) {
772 la->la_mode = oattr->la_mode;
773 la->la_valid |= LA_MODE;
775 la->la_mode &= ~S_ISUID;
778 if (la->la_valid & LA_KILL_SGID) {
779 la->la_valid &= ~LA_KILL_SGID;
780 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
781 (S_ISGID | S_IXGRP)) &&
782 !(la->la_valid & LA_MODE)) {
783 la->la_mode = oattr->la_mode;
784 la->la_valid |= LA_MODE;
786 la->la_mode &= ~S_ISGID;
789 /* Make sure a caller can chmod. */
790 if (la->la_valid & LA_MODE) {
791 if (!(flags & MDS_PERM_BYPASS) &&
792 (uc->uc_fsuid != oattr->la_uid) &&
793 !md_capable(uc, CFS_CAP_FOWNER))
796 if (la->la_mode == (umode_t) -1)
797 la->la_mode = oattr->la_mode;
799 la->la_mode = (la->la_mode & S_IALLUGO) |
800 (oattr->la_mode & ~S_IALLUGO);
802 /* Also check the setgid bit! */
803 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
804 la->la_gid : oattr->la_gid) &&
805 !md_capable(uc, CFS_CAP_FSETID))
806 la->la_mode &= ~S_ISGID;
808 la->la_mode = oattr->la_mode;
811 /* Make sure a caller can chown. */
812 if (la->la_valid & LA_UID) {
813 if (la->la_uid == (uid_t) -1)
814 la->la_uid = oattr->la_uid;
815 if (((uc->uc_fsuid != oattr->la_uid) ||
816 (la->la_uid != oattr->la_uid)) &&
817 !md_capable(uc, CFS_CAP_CHOWN))
820 /* If the user or group of a non-directory has been
821 * changed by a non-root user, remove the setuid bit.
822 * 19981026 David C Niemi <niemi@tux.org>
824 * Changed this to apply to all users, including root,
825 * to avoid some races. This is the behavior we had in
826 * 2.0. The check for non-root was definitely wrong
827 * for 2.2 anyway, as it should have been using
828 * CAP_FSETID rather than fsuid -- 19990830 SD. */
829 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
830 !S_ISDIR(oattr->la_mode)) {
831 la->la_mode &= ~S_ISUID;
832 la->la_valid |= LA_MODE;
836 /* Make sure caller can chgrp. */
837 if (la->la_valid & LA_GID) {
838 if (la->la_gid == (gid_t) -1)
839 la->la_gid = oattr->la_gid;
840 if (((uc->uc_fsuid != oattr->la_uid) ||
841 ((la->la_gid != oattr->la_gid) &&
842 !lustre_in_group_p(uc, la->la_gid))) &&
843 !md_capable(uc, CFS_CAP_CHOWN))
846 /* Likewise, if the user or group of a non-directory
847 * has been changed by a non-root user, remove the
848 * setgid bit UNLESS there is no group execute bit
849 * (this would be a file marked for mandatory
850 * locking). 19981026 David C Niemi <niemi@tux.org>
852 * Removed the fsuid check (see the comment above) --
854 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
855 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
856 la->la_mode &= ~S_ISGID;
857 la->la_valid |= LA_MODE;
861 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
862 if (!((flags & MDS_OWNEROVERRIDE) &&
863 (uc->uc_fsuid == oattr->la_uid)) &&
864 !(flags & MDS_PERM_BYPASS)) {
865 rc = mdd_permission_internal(env, obj, oattr,
872 if (la->la_valid & LA_CTIME) {
873 /* The pure setattr, it has the priority over what is
874 * already set, do not drop it if ctime is equal. */
875 if (la->la_ctime < oattr->la_ctime)
876 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
882 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
883 struct mdd_device *mdd,
884 enum changelog_rec_type type,
885 enum changelog_rec_flags clf_flags,
886 const struct lu_fid *fid,
887 const char *xattr_name,
888 struct thandle *handle)
890 const struct lu_ucred *uc = lu_ucred(env);
891 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
892 struct llog_changelog_rec *rec;
897 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
900 if (uc->uc_jobid[0] != '\0')
901 clf_flags |= CLF_JOBID;
902 xflags |= CLFE_UIDGID;
905 if (type == CL_OPEN || type == CL_DN_OPEN)
907 if (type == CL_SETXATTR || type == CL_GETXATTR)
908 xflags |= CLFE_XATTR;
910 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
911 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
912 xflags & CLFE_SUPPORTED));
913 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
914 if (buf->lb_buf == NULL)
918 rec->cr_hdr.lrh_len = reclen;
919 rec->cr.cr_flags = clf_flags;
920 rec->cr.cr_type = (__u32)type;
921 rec->cr.cr_tfid = *fid;
922 rec->cr.cr_namelen = 0;
924 if (clf_flags & CLF_JOBID)
925 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
927 if (clf_flags & CLF_EXTRA_FLAGS) {
928 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
929 if (xflags & CLFE_UIDGID)
930 mdd_changelog_rec_extra_uidgid(&rec->cr,
931 uc->uc_uid, uc->uc_gid);
932 if (xflags & CLFE_NID)
933 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
934 if (xflags & CLFE_OPEN)
935 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
936 if (xflags & CLFE_XATTR) {
937 if (xattr_name == NULL)
939 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
943 rc = mdd_changelog_store(env, mdd, rec, handle);
948 /** Store a data change changelog record
949 * If this fails, we must fail the whole transaction; we don't
950 * want the change to commit without the log entry.
951 * \param mdd_obj - mdd_object of change
952 * \param handle - transaction handle
954 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
955 enum changelog_rec_type type,
956 enum changelog_rec_flags clf_flags,
957 struct mdd_object *mdd_obj, struct thandle *handle)
961 LASSERT(mdd_obj != NULL);
962 LASSERT(handle != NULL);
964 if (!mdd_changelog_enabled(env, mdd, type))
967 if (mdd_is_volatile_obj(mdd_obj))
970 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
971 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
972 /* Don't need multiple updates in this log */
973 /* Don't check under lock - no big deal if we get an extra
978 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
979 mdd_object_fid(mdd_obj),
982 mdd_obj->mod_cltime = ktime_get();
987 int mdd_changelog_data_store_xattr(const struct lu_env *env,
988 struct mdd_device *mdd,
989 enum changelog_rec_type type,
990 enum changelog_rec_flags clf_flags,
991 struct mdd_object *mdd_obj,
992 const char *xattr_name,
993 struct thandle *handle)
997 LASSERT(mdd_obj != NULL);
998 LASSERT(handle != NULL);
1000 if (!mdd_changelog_enabled(env, mdd, type))
1003 if (mdd_is_volatile_obj(mdd_obj))
1006 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1007 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1008 /* Don't need multiple updates in this log */
1009 /* Don't check under lock - no big deal if we get an extra
1015 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1016 mdd_object_fid(mdd_obj),
1017 xattr_name, handle);
1019 mdd_obj->mod_cltime = ktime_get();
1024 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
1025 * except for open flags have a dedicated record to store 32 bits of flags */
1026 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1027 enum changelog_rec_flags clf_flags,
1028 struct md_device *m, const struct lu_fid *fid)
1030 struct thandle *handle;
1031 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1035 LASSERT(fid != NULL);
1037 /* We'll check this again below, but we check now before we
1038 * start a transaction. */
1039 if (!mdd_changelog_enabled(env, mdd, type))
1042 handle = mdd_trans_create(env, mdd);
1044 RETURN(PTR_ERR(handle));
1046 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1050 rc = mdd_trans_start(env, mdd, handle);
1054 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1058 rc = mdd_trans_stop(env, mdd, rc, handle);
1064 * Save LMA extended attributes with data from \a ma.
1066 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1067 * not, LMA EA will be first read from disk, modified and write back.
1070 /* Precedence for choosing record type when multiple
1071 * attributes change: setattr > mtime > ctime > atime
1072 * (ctime changes when mtime does, plus chmod/chown.
1073 * atime and ctime are independent.) */
1074 static int mdd_attr_set_changelog(const struct lu_env *env,
1075 struct md_object *obj, struct thandle *handle,
1078 struct mdd_device *mdd = mdo2mdd(obj);
1081 bits = (valid & LA_SIZE) ? BIT(CL_TRUNC) : 0;
1082 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? BIT(CL_SETATTR) : 0;
1083 bits |= (valid & LA_MTIME) ? BIT(CL_MTIME) : 0;
1084 bits |= (valid & LA_CTIME) ? BIT(CL_CTIME) : 0;
1085 bits |= (valid & LA_ATIME) ? BIT(CL_ATIME) : 0;
1086 bits = bits & mdd->mdd_cl.mc_mask;
1087 /* This is an implementation limit rather than a protocol limit */
1088 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1092 /* The record type is the lowest non-masked set bit */
1095 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1096 return mdd_changelog_data_store(env, mdd, type, valid,
1097 md2mdd_obj(obj), handle);
1100 static int mdd_declare_attr_set(const struct lu_env *env,
1101 struct mdd_device *mdd,
1102 struct mdd_object *obj,
1103 const struct lu_attr *attr,
1104 struct thandle *handle)
1108 rc = mdo_declare_attr_set(env, obj, attr, handle);
1112 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1113 if (attr->la_valid & LA_MODE) {
1114 mdd_read_lock(env, obj, DT_TGT_CHILD);
1115 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1116 XATTR_NAME_ACL_ACCESS);
1117 mdd_read_unlock(env, obj);
1118 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1124 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1125 rc = mdo_declare_xattr_set(env, obj, buf,
1126 XATTR_NAME_ACL_ACCESS, 0,
1134 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1143 * permission changes may require sync operation, to mitigate performance
1144 * impact, only do this for dir and when permission is reduced.
1146 * For regular files, version is updated with permission change (see VBR), async
1147 * permission won't cause any issue, while missing permission change on
1148 * directory may affect accessibility of other objects after recovery.
1150 static inline bool permission_needs_sync(const struct lu_attr *old,
1151 const struct lu_attr *new)
1153 if (!S_ISDIR(old->la_mode))
1156 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1159 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1162 if (new->la_valid & LA_MODE) {
1163 /* turned on sticky bit */
1164 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1167 /* set-GID has no impact on what is allowed, not checked */
1169 /* turned off setuid bit, or one of rwx for someone */
1170 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1171 (old->la_mode & (0777 | S_ISUID)))
1178 static inline __u64 mdd_lmm_dom_size(void *buf)
1180 struct lov_mds_md *lmm = buf;
1181 struct lov_comp_md_v1 *comp_v1;
1182 struct lov_mds_md *v1;
1188 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1191 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1192 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1193 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1195 /* DoM entry is the first entry always */
1196 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1197 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1202 /* set attr and LOV EA at once, return updated attr */
1203 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1204 const struct md_attr *ma)
1206 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1207 struct mdd_device *mdd = mdo2mdd(obj);
1208 struct thandle *handle = NULL;
1209 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1210 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1211 const struct lu_attr *la = &ma->ma_attr;
1212 struct lu_ucred *uc;
1213 bool chrgrp_by_unprivileged_user = false;
1217 /* we do not use ->attr_set() for LOV/HSM EA any more */
1218 LASSERT((ma->ma_valid & MA_LOV) == 0);
1219 LASSERT((ma->ma_valid & MA_HSM) == 0);
1221 rc = mdd_la_get(env, mdd_obj, attr);
1225 *la_copy = ma->ma_attr;
1226 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1230 /* no need to setattr anymore */
1231 if (la_copy->la_valid == 0) {
1233 "%s: no valid attribute on "DFID", previous was %#llx\n",
1234 mdd_obj_dev_name(mdd_obj),
1235 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1240 /* If an unprivileged user changes group of some file,
1241 * the setattr operation will be processed synchronously to
1242 * honor the quota limit of the corresponding group. see LU-5152 */
1243 uc = lu_ucred_check(env);
1244 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1245 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1246 /* LU-10048: disable synchronous chgrp operation for it will
1247 * cause deadlock between MDT and OST.
1248 la_copy->la_valid |= LA_FLAGS;
1249 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1251 chrgrp_by_unprivileged_user = true;
1253 /* Flush the possible existing client setattr requests to OSTs
1254 * to keep the order with the current setattr operation that
1255 * will be sent directly to OSTs. see LU-5152 */
1256 /* LU-11303 disable sync as this is too heavyweight.
1257 * This should be replaced with a sync only for the object
1258 * being modified here, not the whole filesystem.
1259 rc = dt_sync(env, mdd->mdd_child);
1265 handle = mdd_trans_create(env, mdd);
1266 if (IS_ERR(handle)) {
1267 rc = PTR_ERR(handle);
1273 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1277 rc = mdd_trans_start(env, mdd, handle);
1281 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1282 permission_needs_sync(attr, la))
1283 handle->th_sync = 1;
1285 if (la->la_valid & (LA_MTIME | LA_CTIME))
1286 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1287 la->la_mtime, la->la_ctime);
1289 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1291 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1292 * otherwise following truncate will do nothing and truncated
1293 * data may be read again. This is a quick fix until LU-11033
1296 if (la_copy->la_valid & LA_SIZE) {
1297 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1299 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1303 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1304 la_copy->la_valid &= ~LA_SIZE;
1305 lu_buf_free(lov_buf);
1309 if (la_copy->la_valid) {
1310 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1312 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1313 /* rollback to the original gid */
1314 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1315 la_copy->la_gid = attr->la_gid;
1316 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1319 mdd_write_unlock(env, mdd_obj);
1323 rc = mdd_attr_set_changelog(env, obj, handle,
1327 rc = mdd_trans_stop(env, mdd, rc, handle);
1332 static int mdd_xattr_sanity_check(const struct lu_env *env,
1333 struct mdd_object *obj,
1334 const struct lu_attr *attr,
1337 struct lu_ucred *uc = lu_ucred_assert(env);
1340 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1343 if (strncmp(XATTR_USER_PREFIX, name,
1344 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1345 /* For sticky directories, only the owner and privileged user
1346 * can write attributes. */
1347 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1348 (uc->uc_fsuid != attr->la_uid) &&
1349 !md_capable(uc, CFS_CAP_FOWNER))
1351 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1352 (uc->uc_fsuid != attr->la_uid) &&
1353 !md_capable(uc, CFS_CAP_FOWNER)) {
1361 * Check if a string begins with a given prefix.
1363 * \param str String to check
1364 * \param prefix Substring to check at the beginning of \a str
1365 * \return true/false whether the condition is verified.
1367 static inline bool has_prefix(const char *str, const char *prefix)
1369 return strncmp(prefix, str, strlen(prefix)) == 0;
1373 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1375 * \param[in] xattr_name Full extended attribute name.
1377 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1379 static enum changelog_rec_type
1380 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1381 const char *xattr_name)
1383 /* Layout changes systematically recorded */
1384 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1385 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1386 allowed_lustre_lov(xattr_name))
1389 /* HSM information changes systematically recorded */
1390 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1393 /* Avoid logging SOM xattr for every file */
1394 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1397 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1398 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1399 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1400 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1401 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1407 static int mdd_declare_xattr_set(const struct lu_env *env,
1408 struct mdd_device *mdd,
1409 struct mdd_object *obj,
1410 const struct lu_buf *buf,
1412 int fl, struct thandle *handle)
1414 enum changelog_rec_type type;
1417 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1421 type = mdd_xattr_changelog_type(env, mdd, name);
1423 return 0; /* no changelog to store */
1425 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1429 * Compare current and future data of HSM EA and add a changelog if needed.
1431 * Caller should have write-locked \param obj.
1433 * \param buf - Future HSM EA content.
1434 * \retval 0 if no changelog is needed or changelog was added properly.
1435 * \retval -ve errno if there was a problem
1437 static int mdd_hsm_update_locked(const struct lu_env *env,
1438 struct md_object *obj,
1439 const struct lu_buf *buf,
1440 struct thandle *handle,
1441 enum changelog_rec_flags *clf_flags)
1443 struct mdd_thread_info *info = mdd_env_info(env);
1444 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1445 struct lu_buf *current_buf;
1446 struct md_hsm *current_mh;
1447 struct md_hsm *new_mh;
1451 OBD_ALLOC_PTR(current_mh);
1452 if (current_mh == NULL)
1455 /* Read HSM attrs from disk */
1456 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1458 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1460 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1461 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1462 if (rc < 0 && rc != -ENODATA)
1464 else if (rc == -ENODATA)
1465 current_mh->mh_flags = 0;
1467 /* Map future HSM xattr */
1468 OBD_ALLOC_PTR(new_mh);
1470 GOTO(free, rc = -ENOMEM);
1471 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1475 /* Flags differ, set flags for the changelog that will be added */
1476 if (current_mh->mh_flags != new_mh->mh_flags) {
1477 hsm_set_cl_event(clf_flags, HE_STATE);
1478 if (new_mh->mh_flags & HS_DIRTY)
1479 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1482 OBD_FREE_PTR(new_mh);
1485 OBD_FREE_PTR(current_mh);
1489 static int mdd_object_pfid_replace(const struct lu_env *env,
1490 struct mdd_object *o)
1492 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1493 struct thandle *handle;
1496 handle = mdd_trans_create(env, mdd);
1498 RETURN(PTR_ERR(handle));
1500 handle->th_complex = 1;
1502 /* it doesn't need to track the PFID update via llog, because LFSCK
1503 * will repair it even it goes wrong */
1504 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1509 rc = mdd_trans_start(env, mdd, handle);
1513 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1518 mdd_trans_stop(env, mdd, rc, handle);
1523 static int mdd_declare_xattr_del(const struct lu_env *env,
1524 struct mdd_device *mdd,
1525 struct mdd_object *obj,
1527 struct thandle *handle);
1529 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1532 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1533 struct md_object *md_vic)
1535 struct mdd_device *mdd = mdo2mdd(md_obj);
1536 struct mdd_object *obj = md2mdd_obj(md_obj);
1537 struct mdd_object *vic = md2mdd_obj(md_vic);
1538 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1539 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1540 struct lov_mds_md *lmm;
1541 struct thandle *handle;
1545 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1546 if (rc == 0) /* same fid */
1549 handle = mdd_trans_create(env, mdd);
1551 RETURN(PTR_ERR(handle));
1554 mdd_write_lock(env, obj, DT_TGT_CHILD);
1555 mdd_write_lock(env, vic, DT_TGT_CHILD);
1557 mdd_write_lock(env, vic, DT_TGT_CHILD);
1558 mdd_write_lock(env, obj, DT_TGT_CHILD);
1561 /* get EA of victim file */
1562 memset(buf_vic, 0, sizeof(*buf_vic));
1563 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1570 /* parse the layout of victim file */
1571 lmm = buf_vic->lb_buf;
1572 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1573 GOTO(out, rc = -EINVAL);
1575 /* save EA of target file for restore */
1576 memset(buf, 0, sizeof(*buf));
1577 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1581 /* Get rid of the layout from victim object */
1582 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1586 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1587 LU_XATTR_MERGE, handle);
1591 rc = mdd_trans_start(env, mdd, handle);
1595 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1600 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1602 GOTO(out_restore, rc);
1604 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1605 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1610 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1611 LU_XATTR_REPLACE, handle);
1613 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1614 mdd_obj_dev_name(obj),
1615 PFID(mdd_object_fid(obj)), rc2);
1619 mdd_trans_stop(env, mdd, rc, handle);
1620 mdd_write_unlock(env, obj);
1621 mdd_write_unlock(env, vic);
1623 lu_buf_free(buf_vic);
1626 (void) mdd_object_pfid_replace(env, obj);
1632 * Extract the mirror with specified mirror id, and store the splitted
1633 * mirror layout to @buf.
1635 * \param[in] comp_v1 mirrored layout
1636 * \param[in] mirror_id the mirror with mirror_id to be extracted
1637 * \param[out] buf store the layout excluding the extracted mirror,
1638 * caller free the buffer we allocated in this function
1639 * \param[out] buf_vic store the extracted layout, caller free the buffer
1640 * we allocated in this function
1642 * \retval 0 on success; < 0 if error happens
1644 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1645 struct lu_buf *buf, struct lu_buf *buf_vic)
1647 struct lov_comp_md_v1 *comp_rem;
1648 struct lov_comp_md_v1 *comp_vic;
1649 struct lov_comp_md_entry_v1 *entry;
1650 struct lov_comp_md_entry_v1 *entry_rem;
1651 struct lov_comp_md_entry_v1 *entry_vic;
1653 __u16 comp_cnt, count = 0;
1654 int lmm_size, lmm_size_vic = 0;
1656 int offset, offset_rem, offset_vic;
1658 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1659 /* comp_v1 should contains more than 1 mirror */
1660 if (mirror_cnt <= 1)
1662 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1663 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1665 for (i = 0; i < comp_cnt; i++) {
1666 entry = &comp_v1->lcm_entries[i];
1667 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1669 lmm_size_vic += sizeof(*entry);
1670 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1671 } else if (count > 0) {
1672 /* find the specified mirror */
1680 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1684 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1685 if (!buf_vic->lb_buf) {
1690 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1691 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1693 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1694 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1695 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1696 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1697 if (!comp_rem->lcm_mirror_count)
1698 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1700 memset(comp_vic, 0, sizeof(*comp_v1));
1701 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1702 comp_vic->lcm_mirror_count = 0;
1703 comp_vic->lcm_entry_count = cpu_to_le32(count);
1704 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1705 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1706 comp_vic->lcm_layout_gen = 0;
1708 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1709 offset_rem = sizeof(*comp_rem) +
1710 sizeof(*entry_rem) * (comp_cnt - count);
1711 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1712 for (i = j = k = 0; i < comp_cnt; i++) {
1713 struct lov_mds_md *lmm, *lmm_dst;
1716 entry = &comp_v1->lcm_entries[i];
1717 entry_vic = &comp_vic->lcm_entries[j];
1718 entry_rem = &comp_rem->lcm_entries[k];
1720 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1723 /* copy component entry */
1725 memcpy(entry_vic, entry, sizeof(*entry));
1726 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1727 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1730 memcpy(entry_rem, entry, sizeof(*entry));
1731 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1735 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1737 lmm_dst = (struct lov_mds_md *)
1738 ((char *)comp_vic + offset_vic);
1740 lmm_dst = (struct lov_mds_md *)
1741 ((char *)comp_rem + offset_rem);
1743 /* copy component entry blob */
1744 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1746 /* blob offset advance */
1747 offset += le32_to_cpu(entry->lcme_size);
1749 offset_vic += le32_to_cpu(entry->lcme_size);
1751 offset_rem += le32_to_cpu(entry->lcme_size);
1757 static int mdd_dom_data_truncate(const struct lu_env *env,
1758 struct mdd_device *mdd, struct mdd_object *mo);
1760 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1761 struct md_rejig_data *mrd)
1763 struct mdd_device *mdd = mdo2mdd(md_obj);
1764 struct mdd_object *obj = md2mdd_obj(md_obj);
1765 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1766 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1767 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1768 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1769 struct lov_comp_md_v1 *lcm;
1770 struct thandle *handle;
1772 bool dom_stripe = false;
1776 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1777 if (rc == 0) /* same fid */
1780 handle = mdd_trans_create(env, mdd);
1782 RETURN(PTR_ERR(handle));
1785 mdd_write_lock(env, obj, DT_TGT_CHILD);
1786 mdd_write_lock(env, vic, DT_TGT_CHILD);
1788 mdd_write_lock(env, vic, DT_TGT_CHILD);
1789 mdd_write_lock(env, obj, DT_TGT_CHILD);
1792 /* get EA of mirrored file */
1793 memset(buf_save, 0, sizeof(*buf));
1794 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1798 lcm = buf_save->lb_buf;
1799 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1800 GOTO(out, rc = -EINVAL);
1803 * Extract the mirror with specified mirror id, and store the splitted
1804 * mirror layout to the victim file.
1806 memset(buf, 0, sizeof(*buf));
1807 memset(buf_vic, 0, sizeof(*buf_vic));
1808 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1812 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1814 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1815 LU_XATTR_SPLIT, handle);
1818 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1819 LU_XATTR_SPLIT, handle);
1823 rc = mdd_trans_start(env, mdd, handle);
1827 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1832 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1835 GOTO(out_restore, rc);
1837 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1841 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1848 /* restore obj's layout */
1849 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1850 LU_XATTR_REPLACE, handle);
1852 CERROR("%s: failed rollback "DFID" layout: file state unkonwn: rc = %d\n",
1853 mdd_obj_dev_name(obj),
1854 PFID(mdd_object_fid(obj)), rc2);
1857 rc = mdd_trans_stop(env, mdd, rc, handle);
1859 /* Truncate local DOM data if all went well */
1860 if (!rc && dom_stripe)
1861 mdd_dom_data_truncate(env, mdd, obj);
1863 mdd_write_unlock(env, obj);
1864 mdd_write_unlock(env, vic);
1865 lu_buf_free(buf_save);
1867 lu_buf_free(buf_vic);
1870 (void) mdd_object_pfid_replace(env, obj);
1875 static int mdd_layout_merge_allowed(const struct lu_env *env,
1876 struct md_object *target,
1877 struct md_object *victim)
1879 struct mdd_object *o1 = md2mdd_obj(target);
1881 /* cannot extend directory's LOVEA */
1882 if (S_ISDIR(mdd_object_type(o1))) {
1883 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1884 mdd_obj_dev_name(o1));
1892 * The caller should guarantee to update the object ctime
1893 * after xattr_set if needed.
1895 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1896 const struct lu_buf *buf, const char *name,
1899 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1900 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1901 struct mdd_device *mdd = mdo2mdd(obj);
1902 struct thandle *handle;
1903 enum changelog_rec_type cl_type;
1904 enum changelog_rec_flags clf_flags = 0;
1908 rc = mdd_la_get(env, mdd_obj, attr);
1912 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1916 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1917 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1918 struct md_rejig_data *mrd = buf->lb_buf;
1919 struct md_object *victim = mrd->mrd_obj;
1921 if (buf->lb_len != sizeof(*mrd))
1924 rc = mdd_layout_merge_allowed(env, obj, victim);
1928 if (fl == LU_XATTR_MERGE)
1929 /* merge layout of victim as a mirror of obj's. */
1930 rc = mdd_xattr_merge(env, obj, victim);
1932 rc = mdd_xattr_split(env, obj, mrd);
1936 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1937 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1938 struct posix_acl *acl;
1940 /* user may set empty ACL, which should be treated as removing
1942 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1945 RETURN(PTR_ERR(acl));
1947 rc = mdd_xattr_del(env, obj, name);
1950 posix_acl_release(acl);
1953 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1954 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1958 handle = mdd_trans_create(env, mdd);
1960 RETURN(PTR_ERR(handle));
1962 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1966 rc = mdd_trans_start(env, mdd, handle);
1970 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1972 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1973 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1975 mdd_write_unlock(env, mdd_obj);
1980 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1981 mdd_write_unlock(env, mdd_obj);
1985 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1989 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1990 mdd_obj, name, handle);
1994 return mdd_trans_stop(env, mdd, rc, handle);
1997 static int mdd_declare_xattr_del(const struct lu_env *env,
1998 struct mdd_device *mdd,
1999 struct mdd_object *obj,
2001 struct thandle *handle)
2003 enum changelog_rec_type type;
2006 rc = mdo_declare_xattr_del(env, obj, name, handle);
2010 type = mdd_xattr_changelog_type(env, mdd, name);
2012 return 0; /* no changelog to store */
2014 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2018 * The caller should guarantee to update the object ctime
2019 * after xattr_set if needed.
2021 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2024 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2025 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2026 struct mdd_device *mdd = mdo2mdd(obj);
2027 struct thandle *handle;
2031 rc = mdd_la_get(env, mdd_obj, attr);
2035 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2039 handle = mdd_trans_create(env, mdd);
2041 RETURN(PTR_ERR(handle));
2043 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2047 rc = mdd_trans_start(env, mdd, handle);
2051 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2052 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2053 mdd_write_unlock(env, mdd_obj);
2057 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2060 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2065 return mdd_trans_stop(env, mdd, rc, handle);
2069 * read lov/lmv EA of an object
2070 * return the lov/lmv EA in an allocated lu_buf
2072 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2073 struct lu_buf *lmm_buf, const char *name)
2075 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2080 if (buf->lb_buf == NULL) {
2081 buf = lu_buf_check_and_alloc(buf, 4096);
2082 if (buf->lb_buf == NULL)
2087 rc = mdo_xattr_get(env, obj, buf, name);
2088 if (rc == -ERANGE) {
2089 /* mti_big_buf is allocated but is too small
2090 * we need to increase it */
2091 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2093 if (buf->lb_buf == NULL)
2096 } else if (rc < 0) {
2098 } else if (rc == 0) {
2102 lu_buf_alloc(lmm_buf, rc);
2103 if (lmm_buf->lb_buf == NULL)
2107 * we don't use lmm_buf directly, because we don't know xattr size, so
2108 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2110 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2115 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2116 struct mdd_object *o, struct lu_buf *buf,
2117 struct thandle *handle)
2119 struct hsm_attrs *attrs;
2120 enum hsm_states hsm_flags;
2121 enum changelog_rec_flags clf_flags = 0;
2125 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2130 attrs = buf->lb_buf;
2131 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2132 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2135 /* Add a changelog record for release. */
2136 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2137 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2138 clf_flags, o, handle);
2143 * check if layout swapping between 2 objects is allowed
2145 * - only normal FIDs or non-system IGIFs
2146 * - same type of objects
2147 * - same owner/group (so quotas are still valid) unless this is from HSM
2150 static int mdd_layout_swap_allowed(const struct lu_env *env,
2151 struct mdd_object *o1,
2152 const struct lu_attr *attr1,
2153 struct mdd_object *o2,
2154 const struct lu_attr *attr2,
2157 const struct lu_fid *fid1, *fid2;
2160 fid1 = mdd_object_fid(o1);
2161 fid2 = mdd_object_fid(o2);
2163 if (!fid_is_norm(fid1) &&
2164 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2167 if (!fid_is_norm(fid2) &&
2168 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2171 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2172 if (S_ISDIR(mdd_object_type(o1)))
2174 if (S_ISREG(mdd_object_type(o1)))
2179 if (flags & SWAP_LAYOUTS_MDS_HSM)
2182 if ((attr1->la_uid != attr2->la_uid) ||
2183 (attr1->la_gid != attr2->la_gid))
2189 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2190 * look into the layout in MDD layer. */
2191 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2193 struct lov_comp_md_v1 *comp_v1;
2194 struct lov_mds_md *v1;
2198 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2199 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2200 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2206 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2207 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2210 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2211 off = le32_to_cpu(comp_v1->lcm_entries[i].
2213 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2218 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2219 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2230 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2232 return mdd_lmm_oi(lmm, oi, true);
2235 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2237 return mdd_lmm_oi(lmm, oi, false);
2240 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2242 struct lov_comp_md_v1 *comp_v1;
2244 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2245 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2247 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2249 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2250 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2251 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2252 __u16 tmp_gen = *gen;
2254 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2256 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2263 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2265 return mdd_lmm_gen(lmm, gen, true);
2268 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2270 return mdd_lmm_gen(lmm, gen, false);
2273 static int mdd_dom_data_truncate(const struct lu_env *env,
2274 struct mdd_device *mdd, struct mdd_object *mo)
2277 struct dt_object *dom;
2280 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2282 GOTO(out, rc = -ENODATA);
2284 th = dt_trans_create(env, mdd->mdd_bottom);
2286 GOTO(out, rc = PTR_ERR(th));
2288 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2292 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2296 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2298 dt_trans_stop(env, mdd->mdd_bottom, th);
2300 /* Ignore failure but report the error */
2302 CERROR("%s: can't truncate DOM inode "DFID" data: rc = %d\n",
2303 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2308 * swap layouts between 2 lustre objects
2310 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2311 struct md_object *obj2, __u64 flags)
2313 struct mdd_thread_info *info = mdd_env_info(env);
2314 struct mdd_object *fst_o = md2mdd_obj(obj1);
2315 struct mdd_object *snd_o = md2mdd_obj(obj2);
2316 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2317 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2318 struct mdd_device *mdd = mdo2mdd(obj1);
2319 struct lov_mds_md *fst_lmm, *snd_lmm;
2320 struct lu_buf *fst_buf = &info->mti_buf[0];
2321 struct lu_buf *snd_buf = &info->mti_buf[1];
2322 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2323 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2324 struct ost_id *saved_oi = NULL;
2325 struct thandle *handle;
2326 struct mdd_object *dom_o = NULL;
2327 __u64 domsize_dom, domsize_vlt;
2328 __u32 fst_gen, snd_gen, saved_gen;
2334 BUILD_BUG_ON(ARRAY_SIZE(info->mti_buf) < 4);
2335 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2337 /* we have to sort the 2 obj, so locking will always
2338 * be in the same order, even in case of 2 concurrent swaps */
2339 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2340 if (rc == 0) /* same fid ? */
2346 rc = mdd_la_get(env, fst_o, fst_la);
2350 rc = mdd_la_get(env, snd_o, snd_la);
2354 /* check if layout swapping is allowed */
2355 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2359 handle = mdd_trans_create(env, mdd);
2361 RETURN(PTR_ERR(handle));
2363 /* objects are already sorted */
2364 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2365 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2367 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2368 if (rc < 0 && rc != -ENODATA)
2371 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2372 if (rc < 0 && rc != -ENODATA)
2375 /* check if file has DoM. DoM file can be migrated only to another
2376 * DoM layout with the same DoM component size or to an non-DOM
2377 * layout. After migration to OSTs layout, local MDT inode data
2378 * should be truncated.
2379 * Objects are sorted by FIDs, considering that original file's FID
2380 * is always smaller the snd_o is always original file we are migrating
2383 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2384 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2386 /* Only migration is supported for DoM files, not 'swap_layouts' so
2387 * target file must be volatile and orphan.
2389 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2390 dom_o = domsize_dom ? snd_o : NULL;
2391 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2392 swap(domsize_dom, domsize_vlt);
2393 dom_o = domsize_dom ? fst_o : NULL;
2394 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2395 /* 'lfs swap_layouts' case, neither file should have DoM */
2397 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2398 "use migration instead: rc = %d\n", rc);
2402 if (domsize_vlt > 0 && domsize_dom == 0) {
2405 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2406 mdd_obj_dev_name(snd_o),
2407 PFID(mdd_object_fid(snd_o)), rc);
2409 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2412 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2413 mdd_obj_dev_name(fst_o),
2414 PFID(mdd_object_fid(fst_o)), rc);
2416 } else if (domsize_vlt > 0) {
2417 /* Migration with the same DOM component size, no need to
2418 * truncate local data, it is still being used */
2422 /* swapping 2 non existant layouts is a success */
2423 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2426 /* to help inode migration between MDT, it is better to
2427 * start by the no layout file (if one), so we order the swap */
2428 if (snd_buf->lb_buf == NULL) {
2430 swap(fst_buf, snd_buf);
2433 fst_gen = snd_gen = 0;
2434 /* lmm and generation layout initialization */
2435 if (fst_buf->lb_buf != NULL) {
2436 fst_lmm = fst_buf->lb_buf;
2437 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2438 fst_fl = LU_XATTR_REPLACE;
2442 fst_fl = LU_XATTR_CREATE;
2445 snd_lmm = snd_buf->lb_buf;
2446 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2448 saved_gen = fst_gen;
2449 /* increase the generation layout numbers */
2454 * XXX The layout generation is used to generate component IDs for
2455 * the composite file, we have to do some special tweaks to make
2456 * sure the layout generation is always adequate for that job.
2459 /* Skip invalid generation number for composite layout */
2460 if ((snd_gen & LCME_ID_MASK) == 0)
2462 if ((fst_gen & LCME_ID_MASK) == 0)
2464 /* Make sure the generation is greater than all the component IDs */
2465 if (fst_gen < snd_gen)
2467 else if (fst_gen > snd_gen)
2470 /* set the file specific informations in lmm */
2471 if (fst_lmm != NULL) {
2472 struct ost_id temp_oi;
2474 saved_oi = &info->mti_oa.o_oi;
2475 mdd_get_lmm_oi(fst_lmm, saved_oi);
2476 mdd_get_lmm_oi(snd_lmm, &temp_oi);
2477 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2478 mdd_set_lmm_oi(fst_lmm, &temp_oi);
2479 mdd_set_lmm_oi(snd_lmm, saved_oi);
2481 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2482 cpu_to_le32(LOV_MAGIC_MAGIC))
2483 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2485 GOTO(stop, rc = -EPROTO);
2487 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2489 /* Prepare HSM attribute if it's required */
2490 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2491 const int buflen = sizeof(struct hsm_attrs);
2493 lu_buf_alloc(fst_hsm_buf, buflen);
2494 lu_buf_alloc(snd_hsm_buf, buflen);
2495 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2496 GOTO(stop, rc = -ENOMEM);
2498 /* Read HSM attribute */
2499 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2503 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2507 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2508 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2513 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2514 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2520 /* prepare transaction */
2521 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2526 if (fst_buf->lb_buf != NULL)
2527 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2528 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2531 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2536 rc = mdd_trans_start(env, mdd, handle);
2540 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2541 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2545 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2547 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2550 CERROR("%s: HSM error restoring "DFID": rc = %d/%d\n",
2551 mdd_obj_dev_name(fst_o),
2552 PFID(mdd_object_fid(fst_o)), rc, rc2);
2557 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2561 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2564 if (fst_buf->lb_buf != NULL)
2565 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2566 LU_XATTR_REPLACE, handle);
2568 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2571 GOTO(out_restore, rc);
2573 /* Issue one changelog record per file */
2574 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2578 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2587 /* failure on second file, but first was done, so we have
2588 * to roll back first. */
2589 if (fst_buf->lb_buf != NULL) {
2590 mdd_set_lmm_oi(fst_lmm, saved_oi);
2591 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2592 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2593 LU_XATTR_REPLACE, handle);
2595 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2601 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2606 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2611 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
2612 mdd_obj_dev_name(fst_o),
2613 PFID(mdd_object_fid(snd_o)),
2614 PFID(mdd_object_fid(fst_o)),
2616 /* a solution to avoid journal commit is to panic,
2617 * but it has strong consequences so we use LBUG to
2618 * allow sysdamin to choose to panic or not
2625 rc = mdd_trans_stop(env, mdd, rc, handle);
2627 /* Truncate local DOM data if all went well */
2629 mdd_dom_data_truncate(env, mdd, dom_o);
2631 mdd_write_unlock(env, snd_o);
2632 mdd_write_unlock(env, fst_o);
2634 lu_buf_free(fst_buf);
2635 lu_buf_free(snd_buf);
2636 lu_buf_free(fst_hsm_buf);
2637 lu_buf_free(snd_hsm_buf);
2640 (void) mdd_object_pfid_replace(env, fst_o);
2641 (void) mdd_object_pfid_replace(env, snd_o);
2646 static int mdd_declare_layout_change(const struct lu_env *env,
2647 struct mdd_device *mdd,
2648 struct mdd_object *obj,
2649 struct md_layout_change *mlc,
2650 struct thandle *handle)
2654 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2658 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2662 /* For PFL, this is used to instantiate necessary component objects. */
2664 mdd_layout_instantiate_component(const struct lu_env *env,
2665 struct mdd_object *obj, struct md_layout_change *mlc,
2666 struct thandle *handle)
2668 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2672 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2675 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2677 * It's possible that another layout write intent has already
2678 * instantiated our objects, so a -EALREADY returned, and we need to
2682 RETURN(rc == -EALREADY ? 0 : rc);
2684 rc = mdd_trans_start(env, mdd, handle);
2688 mdd_write_lock(env, obj, DT_TGT_CHILD);
2689 rc = mdo_layout_change(env, obj, mlc, handle);
2690 mdd_write_unlock(env, obj);
2694 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2699 * Change the FLR layout from RDONLY to WRITE_PENDING.
2701 * It picks the primary mirror, and bumps the layout version, and set
2702 * layout version xattr to OST objects in a sync tx. In order to facilitate
2703 * the handling of phantom writers from evicted clients, the clients carry
2704 * layout version of the file with write RPC, so that the OSTs can verify
2705 * if the write RPCs are legitimate, meaning not from evicted clients.
2708 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2709 struct md_layout_change *mlc, struct thandle *handle)
2711 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2712 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2713 struct lustre_som_attrs *som = &mlc->mlc_som;
2718 /* Verify acceptable operations */
2719 switch (mlc->mlc_opc) {
2720 case MD_LAYOUT_WRITE:
2721 case MD_LAYOUT_RESYNC:
2722 /* these are legal operations - this represents the case that
2723 * a few mirrors were missed in the last resync. */
2725 case MD_LAYOUT_RESYNC_DONE:
2730 som_buf->lb_buf = som;
2731 som_buf->lb_len = sizeof(*som);
2732 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2733 if (rc < 0 && rc != -ENODATA)
2737 lustre_som_swab(som);
2738 if (som->lsa_valid & SOM_FL_STRICT)
2739 fl = LU_XATTR_REPLACE;
2742 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2747 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2748 XATTR_NAME_SOM, fl, handle);
2753 /* record a changelog for data mover to consume */
2754 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2758 rc = mdd_trans_start(env, mdd, handle);
2762 /* it needs a sync tx to make FLR to work properly */
2763 handle->th_sync = 1;
2765 mdd_write_lock(env, obj, DT_TGT_CHILD);
2766 rc = mdo_layout_change(env, obj, mlc, handle);
2768 /* SOM state transition from STRICT to STALE */
2769 som->lsa_valid = SOM_FL_STALE;
2770 lustre_som_swab(som);
2771 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2774 mdd_write_unlock(env, obj);
2778 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2789 * Handle mirrored file state transition when it's in WRITE_PENDING.
2791 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2792 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2793 * layout version will be increased, and the file's state will be changed to
2797 mdd_layout_update_write_pending(const struct lu_env *env,
2798 struct mdd_object *obj, struct md_layout_change *mlc,
2799 struct thandle *handle)
2801 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2805 switch (mlc->mlc_opc) {
2806 case MD_LAYOUT_RESYNC:
2807 /* Upon receiving the resync request, it should
2808 * instantiate all stale components right away to get ready
2809 * for mirror copy. In order to avoid layout version change,
2810 * client should avoid sending LAYOUT_WRITE request at the
2813 case MD_LAYOUT_WRITE:
2814 /* legal race for concurrent write, the file state has been
2815 * changed by another client. */
2821 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2825 rc = mdd_trans_start(env, mdd, handle);
2829 /* it needs a sync tx to make FLR to work properly */
2830 handle->th_sync = 1;
2832 mdd_write_lock(env, obj, DT_TGT_CHILD);
2833 rc = mdo_layout_change(env, obj, mlc, handle);
2834 mdd_write_unlock(env, obj);
2845 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2847 * Only concurrent write and sync complete requests are possible when the
2848 * file is in SYNC_PENDING. For the latter request, it will pass in the
2849 * mirrors that have been synchronized, then the stale bit will be cleared
2850 * to make the file's state turn into RDONLY.
2851 * For concurrent write reqeust, it just needs to change the file's state
2852 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2853 * version because the version will be increased in the transition to
2854 * SYNC_PENDING later so that it can deny the write request from potential
2855 * evicted SYNC clients. */
2857 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2858 struct md_layout_change *mlc, struct thandle *handle)
2860 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2861 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2866 /* operation validation */
2867 switch (mlc->mlc_opc) {
2868 case MD_LAYOUT_RESYNC_DONE:
2869 /* resync complete. */
2870 case MD_LAYOUT_WRITE:
2871 /* concurrent write. */
2873 case MD_LAYOUT_RESYNC:
2874 /* resync again, most likely the previous run failed.
2875 * no-op if it's already in SYNC_PENDING state */
2881 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2882 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2883 if (rc < 0 && rc != -ENODATA)
2886 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2887 lustre_som_swab(&mlc->mlc_som);
2888 som_buf->lb_buf = &mlc->mlc_som;
2889 som_buf->lb_len = sizeof(mlc->mlc_som);
2892 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2896 /* record a changelog for the completion of resync */
2897 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2902 /* RESYNC_DONE has piggybacked size and blocks */
2904 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2905 XATTR_NAME_SOM, fl, handle);
2910 rc = mdd_trans_start(env, mdd, handle);
2914 /* it needs a sync tx to make FLR to work properly */
2915 handle->th_sync = 1;
2917 rc = mdo_layout_change(env, obj, mlc, handle);
2922 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2928 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2937 * Layout change callback for object.
2939 * This is only used by FLR for now. In the future, it can be exteneded to
2940 * handle all layout change.
2943 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2944 struct md_layout_change *mlc)
2946 struct mdd_object *obj = md2mdd_obj(o);
2947 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2948 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2949 struct lov_comp_md_v1 *lcm;
2950 struct thandle *handle;
2956 if (S_ISDIR(mdd_object_type(obj))) {
2957 switch (mlc->mlc_opc) {
2958 case MD_LAYOUT_SHRINK:
2959 rc = mdd_dir_layout_shrink(env, o, mlc);
2961 case MD_LAYOUT_SPLIT:
2962 rc = mdd_dir_layout_split(env, o, mlc);
2971 /* Verify acceptable operations */
2972 switch (mlc->mlc_opc) {
2973 case MD_LAYOUT_WRITE:
2974 case MD_LAYOUT_RESYNC:
2975 case MD_LAYOUT_RESYNC_DONE:
2981 handle = mdd_trans_create(env, mdd);
2983 RETURN(PTR_ERR(handle));
2985 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2992 /* analyze the layout to make sure it's a FLR file */
2994 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2995 GOTO(out, rc = -EINVAL);
2997 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2999 /* please refer to HLD of FLR for state transition */
3000 switch (flr_state) {
3002 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
3004 case LCM_FL_WRITE_PENDING:
3005 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
3008 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
3010 case LCM_FL_SYNC_PENDING:
3011 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3020 mdd_trans_stop(env, mdd, rc, handle);
3025 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3026 struct mdd_object *child, const struct lu_attr *attr,
3027 const struct md_op_spec *spec,
3028 struct dt_allocation_hint *hint)
3030 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3031 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3032 struct dt_object *nc = mdd_object_child(child);
3034 memset(hint, 0, sizeof(*hint));
3036 /* For striped directory, give striping EA to lod_ah_init, which will
3037 * decide the stripe_offset and stripe count by it. */
3038 if (S_ISDIR(attr->la_mode) &&
3039 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3040 hint->dah_eadata = spec->u.sp_ea.eadata;
3041 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3043 hint->dah_eadata = NULL;
3044 hint->dah_eadata_len = 0;
3045 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
3046 if (mdd->mdd_append_stripe_count != 0 ||
3047 mdd->mdd_append_pool[0])
3049 "using O_APPEND file striping\n");
3050 if (mdd->mdd_append_stripe_count)
3051 hint->dah_append_stripes =
3052 mdd->mdd_append_stripe_count;
3053 if (mdd->mdd_append_pool[0])
3054 hint->dah_append_pool = mdd->mdd_append_pool;
3056 hint->dah_append_stripes = 0;
3060 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3061 hint->dah_eadata, hint->dah_eadata_len);
3062 /* @hint will be initialized by underlying device. */
3063 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3066 static int accmode(const struct lu_env *env, const struct lu_attr *la,
3069 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3070 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3071 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3072 * owner can write to a file even if it is marked readonly to hide
3073 * its brokenness. (bug 5781) */
3074 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3075 struct lu_ucred *uc = lu_ucred_check(env);
3077 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3081 return mds_accmode(open_flags);
3084 static int mdd_open_sanity_check(const struct lu_env *env,
3085 struct mdd_object *obj,
3086 const struct lu_attr *attr, u64 open_flags,
3092 /* EEXIST check, also opening of *open* orphans is allowed so we can
3093 * open-by-handle unlinked files
3095 if (mdd_is_dead_obj(obj) && !is_replay &&
3096 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3099 if (S_ISLNK(attr->la_mode))
3102 mode = accmode(env, attr, open_flags);
3104 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3107 if (!(open_flags & MDS_OPEN_CREATED)) {
3108 rc = mdd_permission_internal(env, obj, attr, mode);
3113 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3114 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3115 open_flags &= ~MDS_OPEN_TRUNC;
3117 /* For writing append-only file must open it with append mode. */
3118 if (attr->la_flags & LUSTRE_APPEND_FL) {
3119 if ((open_flags & MDS_FMODE_WRITE) &&
3120 !(open_flags & MDS_OPEN_APPEND))
3122 if (open_flags & MDS_OPEN_TRUNC)
3129 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3130 u64 open_flags, struct md_op_spec *spec)
3132 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3133 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3134 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3135 struct mdd_object_user *mou = NULL;
3136 const struct lu_ucred *uc = lu_ucred(env);
3137 struct mdd_device *mdd = mdo2mdd(obj);
3138 enum changelog_rec_type type = CL_OPEN;
3142 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3144 rc = mdd_la_get(env, mdd_obj, attr);
3148 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3150 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & BIT(CL_DN_OPEN)))
3155 mdd_obj->mod_count++;
3157 if (!mdd_changelog_enabled(env, mdd, type))
3161 /* look for existing opener in list under mdd_write_lock */
3162 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3167 /* add user to list */
3168 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3174 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3176 mdd_obj_user_free(mou);
3181 if (type == CL_DN_OPEN) {
3182 if (ktime_before(ktime_get(), mou->mou_deniednext))
3183 /* same user denied again same access within
3184 * time interval: do not record
3188 /* this user already denied, but some time ago:
3189 * update denied time
3191 mou->mou_deniednext =
3192 ktime_add(ktime_get(),
3193 ktime_set(mdd->mdd_cl.mc_deniednext,
3196 mou->mou_opencount++;
3197 /* same user opening file again with same flags:
3204 /* FYI, only the bottom 32 bits of open_flags are recorded */
3205 mdd_changelog(env, type, open_flags, md_dev, mdd_object_fid(mdd_obj));
3209 mdd_write_unlock(env, mdd_obj);
3213 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3214 struct md_attr *ma, struct thandle *handle)
3218 rc = mdd_orphan_declare_delete(env, obj, handle);
3222 return mdo_declare_destroy(env, obj, handle);
3226 * No permission check is needed.
3228 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3229 struct md_attr *ma, u64 open_flags)
3231 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3232 struct mdd_device *mdd = mdo2mdd(obj);
3233 struct thandle *handle = NULL;
3236 bool blocked = false;
3237 bool last_close_by_uid = false;
3238 const struct lu_ucred *uc = lu_ucred(env);
3241 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3242 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3243 mdd_obj->mod_count--;
3244 mdd_write_unlock(env, mdd_obj);
3246 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3247 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3248 "list\n", PFID(mdd_object_fid(mdd_obj)));
3252 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3253 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3254 /* check without any lock */
3255 is_orphan = mdd_obj->mod_count == 1 &&
3256 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3260 /* mdd_trans_create() maybe failed because of barrier_entry(),
3261 * under such case, the orphan MDT-object will be left in the
3262 * orphan list, and when the MDT remount next time, the unused
3263 * orphans will be destroyed automatically.
3265 * One exception: the former mdd_finish_unlink may failed to
3266 * add the orphan MDT-object to the orphan list, then if the
3267 * mdd_trans_create() failed because of barrier_entry(), the
3268 * MDT-object will become real orphan that is neither in the
3269 * namespace nor in the orphan list. Such bad case should be
3270 * very rare and will be handled by e2fsck/lfsck. */
3271 handle = mdd_trans_create(env, mdo2mdd(obj));
3272 if (IS_ERR(handle)) {
3273 rc = PTR_ERR(handle);
3274 if (rc != -EINPROGRESS)
3282 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3286 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3291 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3297 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3298 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3300 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3301 lu_dev_name(mdd2lu_dev(mdd)),
3302 PFID(mdd_object_fid(mdd_obj)), rc);
3306 /* check again with lock */
3307 is_orphan = (mdd_obj->mod_count == 1) &&
3308 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3309 ma->ma_attr.la_nlink == 0);
3311 if (is_orphan && !handle && !blocked) {
3312 mdd_write_unlock(env, mdd_obj);
3316 mdd_obj->mod_count--; /*release open count */
3318 /* under mdd write lock */
3319 /* If recording, see if we need to remove UID from list. uc is not
3320 * initialized if the client has been evicted. */
3321 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3322 struct mdd_object_user *mou;
3324 /* look for UID in list */
3325 /* If mou is NULL, it probably means logging was enabled after
3326 * the user had the file open. So the corresponding close
3327 * will not be logged.
3329 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3332 mou->mou_opencount--;
3333 if (mou->mou_opencount == 0) {
3334 mdd_obj_user_remove(mdd_obj, mou);
3335 last_close_by_uid = true;
3340 if (!is_orphan || blocked)
3344 /* NB: Object maybe not in orphan list originally, it is rare case for
3345 * mdd_finish_unlink() failure, in that case, the object doesn't have
3346 * ORPHAN_OBJ flag */
3347 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3348 /* remove link to object from orphan index */
3349 LASSERT(handle != NULL);
3350 rc = mdd_orphan_delete(env, mdd_obj, handle);
3352 CERROR("%s: unable to delete "DFID" from orphan list: "
3353 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3354 PFID(mdd_object_fid(mdd_obj)), rc);
3355 /* If object was not deleted from orphan list, do not
3356 * destroy OSS objects, which will be done when next
3361 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3362 "list, OSS objects to be destroyed.\n",
3363 PFID(mdd_object_fid(mdd_obj)));
3366 rc = mdo_destroy(env, mdd_obj, handle);
3369 CERROR("%s: unable to delete "DFID" from orphan list: "
3370 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3371 PFID(mdd_object_fid(mdd_obj)), rc);
3376 mdd_write_unlock(env, mdd_obj);
3378 if (rc != 0 || blocked ||
3379 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3382 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3383 * or if CL_OPEN was recorded and it's last close by user.
3384 * Changelogs mask may change between open and close operations, but
3385 * this is not a big deal if we have a CL_CLOSE entry with no matching
3386 * CL_OPEN. Plus Changelogs mask may not change often.
3388 if (((!(mdd->mdd_cl.mc_mask & BIT(CL_OPEN)) &&
3389 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3390 MDS_OPEN_TRUNC))) ||
3391 ((mdd->mdd_cl.mc_mask & BIT(CL_OPEN)) && last_close_by_uid)) &&
3392 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3393 if (handle == NULL) {
3394 handle = mdd_trans_create(env, mdo2mdd(obj));
3396 GOTO(stop, rc = PTR_ERR(handle));
3398 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3399 NULL, NULL, handle);
3403 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3408 /* FYI, only the bottom 32 bits of open_flags are recorded */
3409 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3414 if (handle != NULL && !IS_ERR(handle))
3415 rc = mdd_trans_stop(env, mdd, rc, handle);
3421 * Permission check is done when open,
3422 * no need check again.
3424 static int mdd_readpage_sanity_check(const struct lu_env *env,
3425 struct mdd_object *obj)
3427 struct dt_object *next = mdd_object_child(obj);
3431 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3439 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3440 size_t nob, const struct dt_it_ops *iops,
3441 struct dt_it *it, __u32 attr, void *arg)
3443 struct lu_dirpage *dp = &lp->lp_dir;
3447 struct lu_dirent *ent;
3448 struct lu_dirent *last = NULL;
3452 if (nob < sizeof(*dp))
3455 memset(area, 0, sizeof (*dp));
3456 area += sizeof (*dp);
3457 nob -= sizeof (*dp);
3464 len = iops->key_size(env, it);
3466 /* IAM iterator can return record with zero len. */
3470 hash = iops->store(env, it);
3471 if (unlikely(first)) {
3473 dp->ldp_hash_start = cpu_to_le64(hash);
3476 /* calculate max space required for lu_dirent */
3477 recsize = lu_dirent_calc_size(len, attr);
3479 if (nob >= recsize) {
3480 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3481 if (result == -ESTALE)
3486 /* osd might not able to pack all attributes,
3487 * so recheck rec length */
3488 recsize = le16_to_cpu(ent->lde_reclen);
3490 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3491 fid_le_to_cpu(&fid, &ent->lde_fid);
3492 if (fid_is_dot_lustre(&fid))
3496 result = (last != NULL) ? 0 :-EINVAL;
3500 ent = (void *)ent + recsize;
3504 result = iops->next(env, it);
3505 if (result == -ESTALE)
3507 } while (result == 0);
3510 dp->ldp_hash_end = cpu_to_le64(hash);
3512 if (last->lde_hash == dp->ldp_hash_end)
3513 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3514 last->lde_reclen = 0; /* end mark */
3517 /* end of directory */
3518 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3519 else if (result < 0)
3520 CWARN("build page failed: %d!\n", result);
3524 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3525 const struct lu_rdpg *rdpg)
3527 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3531 if (mdd_object_exists(mdd_obj) == 0) {
3532 CERROR("%s: object "DFID" not found: rc = -2\n",
3533 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3537 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3538 rc = mdd_readpage_sanity_check(env, mdd_obj);
3540 GOTO(out_unlock, rc);
3542 if (mdd_is_dead_obj(mdd_obj)) {
3544 struct lu_dirpage *dp;
3547 * According to POSIX, please do not return any entry to client:
3548 * even dot and dotdot should not be returned.
3550 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3551 PFID(mdd_object_fid(mdd_obj)));
3553 if (rdpg->rp_count <= 0)
3554 GOTO(out_unlock, rc = -EFAULT);
3555 LASSERT(rdpg->rp_pages != NULL);
3557 pg = rdpg->rp_pages[0];
3558 dp = (struct lu_dirpage *)kmap(pg);
3559 memset(dp, 0 , sizeof(struct lu_dirpage));
3560 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3561 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3562 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3564 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3567 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3568 mdd_dir_page_build, NULL);
3570 struct lu_dirpage *dp;
3572 dp = kmap(rdpg->rp_pages[0]);
3573 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3576 * No pages were processed, mark this for first page
3579 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3580 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3581 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3583 kunmap(rdpg->rp_pages[0]);
3586 GOTO(out_unlock, rc);
3588 mdd_read_unlock(env, mdd_obj);
3592 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3594 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3596 if (mdd_object_exists(mdd_obj) == 0) {
3599 CERROR("%s: object "DFID" not found: rc = %d\n",
3600 mdd_obj_dev_name(mdd_obj),
3601 PFID(mdd_object_fid(mdd_obj)), rc);
3604 return dt_object_sync(env, mdd_object_child(mdd_obj),
3608 static int mdd_object_lock(const struct lu_env *env,
3609 struct md_object *obj,
3610 struct lustre_handle *lh,
3611 struct ldlm_enqueue_info *einfo,
3612 union ldlm_policy_data *policy)
3614 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3615 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3619 static int mdd_object_unlock(const struct lu_env *env,
3620 struct md_object *obj,
3621 struct ldlm_enqueue_info *einfo,
3622 union ldlm_policy_data *policy)
3624 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3625 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3628 const struct md_object_operations mdd_obj_ops = {
3629 .moo_permission = mdd_permission,
3630 .moo_attr_get = mdd_attr_get,
3631 .moo_attr_set = mdd_attr_set,
3632 .moo_xattr_get = mdd_xattr_get,
3633 .moo_xattr_set = mdd_xattr_set,
3634 .moo_xattr_list = mdd_xattr_list,
3635 .moo_invalidate = mdd_invalidate,
3636 .moo_xattr_del = mdd_xattr_del,
3637 .moo_swap_layouts = mdd_swap_layouts,
3638 .moo_open = mdd_open,
3639 .moo_close = mdd_close,
3640 .moo_readpage = mdd_readpage,
3641 .moo_readlink = mdd_readlink,
3642 .moo_changelog = mdd_changelog,
3643 .moo_object_sync = mdd_object_sync,
3644 .moo_object_lock = mdd_object_lock,
3645 .moo_object_unlock = mdd_object_unlock,
3646 .moo_layout_change = mdd_layout_change,