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);
742 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
743 (la->la_valid & ~LA_FLAGS) &&
744 !(flags & MDS_PERM_BYPASS))
747 /* Check for setting the obj time. */
748 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
749 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
750 if ((uc->uc_fsuid != oattr->la_uid) &&
751 !md_capable(uc, CFS_CAP_FOWNER)) {
752 rc = mdd_permission_internal(env, obj, oattr,
759 if (la->la_valid & LA_KILL_SUID) {
760 la->la_valid &= ~LA_KILL_SUID;
761 if ((oattr->la_mode & S_ISUID) &&
762 !(la->la_valid & LA_MODE)) {
763 la->la_mode = oattr->la_mode;
764 la->la_valid |= LA_MODE;
766 la->la_mode &= ~S_ISUID;
769 if (la->la_valid & LA_KILL_SGID) {
770 la->la_valid &= ~LA_KILL_SGID;
771 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
772 (S_ISGID | S_IXGRP)) &&
773 !(la->la_valid & LA_MODE)) {
774 la->la_mode = oattr->la_mode;
775 la->la_valid |= LA_MODE;
777 la->la_mode &= ~S_ISGID;
780 /* Make sure a caller can chmod. */
781 if (la->la_valid & LA_MODE) {
782 if (!(flags & MDS_PERM_BYPASS) &&
783 (uc->uc_fsuid != oattr->la_uid) &&
784 !md_capable(uc, CFS_CAP_FOWNER))
787 if (la->la_mode == (umode_t) -1)
788 la->la_mode = oattr->la_mode;
790 la->la_mode = (la->la_mode & S_IALLUGO) |
791 (oattr->la_mode & ~S_IALLUGO);
793 /* Also check the setgid bit! */
794 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
795 la->la_gid : oattr->la_gid) &&
796 !md_capable(uc, CFS_CAP_FSETID))
797 la->la_mode &= ~S_ISGID;
799 la->la_mode = oattr->la_mode;
802 /* Make sure a caller can chown. */
803 if (la->la_valid & LA_UID) {
804 if (la->la_uid == (uid_t) -1)
805 la->la_uid = oattr->la_uid;
806 if (((uc->uc_fsuid != oattr->la_uid) ||
807 (la->la_uid != oattr->la_uid)) &&
808 !md_capable(uc, CFS_CAP_CHOWN))
811 /* If the user or group of a non-directory has been
812 * changed by a non-root user, remove the setuid bit.
813 * 19981026 David C Niemi <niemi@tux.org>
815 * Changed this to apply to all users, including root,
816 * to avoid some races. This is the behavior we had in
817 * 2.0. The check for non-root was definitely wrong
818 * for 2.2 anyway, as it should have been using
819 * CAP_FSETID rather than fsuid -- 19990830 SD. */
820 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
821 !S_ISDIR(oattr->la_mode)) {
822 la->la_mode &= ~S_ISUID;
823 la->la_valid |= LA_MODE;
827 /* Make sure caller can chgrp. */
828 if (la->la_valid & LA_GID) {
829 if (la->la_gid == (gid_t) -1)
830 la->la_gid = oattr->la_gid;
831 if (((uc->uc_fsuid != oattr->la_uid) ||
832 ((la->la_gid != oattr->la_gid) &&
833 !lustre_in_group_p(uc, la->la_gid))) &&
834 !md_capable(uc, CFS_CAP_CHOWN))
837 /* Likewise, if the user or group of a non-directory
838 * has been changed by a non-root user, remove the
839 * setgid bit UNLESS there is no group execute bit
840 * (this would be a file marked for mandatory
841 * locking). 19981026 David C Niemi <niemi@tux.org>
843 * Removed the fsuid check (see the comment above) --
845 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
846 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
847 la->la_mode &= ~S_ISGID;
848 la->la_valid |= LA_MODE;
852 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
853 if (!((flags & MDS_OWNEROVERRIDE) &&
854 (uc->uc_fsuid == oattr->la_uid)) &&
855 !(flags & MDS_PERM_BYPASS)) {
856 rc = mdd_permission_internal(env, obj, oattr,
863 if (la->la_valid & LA_CTIME) {
864 /* The pure setattr, it has the priority over what is
865 * already set, do not drop it if ctime is equal. */
866 if (la->la_ctime < oattr->la_ctime)
867 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
873 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
874 struct mdd_device *mdd,
875 enum changelog_rec_type type,
876 enum changelog_rec_flags clf_flags,
877 const struct lu_fid *fid,
878 const char *xattr_name,
879 struct thandle *handle)
881 const struct lu_ucred *uc = lu_ucred(env);
882 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
883 struct llog_changelog_rec *rec;
888 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
891 if (uc->uc_jobid[0] != '\0')
892 clf_flags |= CLF_JOBID;
893 xflags |= CLFE_UIDGID;
896 if (type == CL_OPEN || type == CL_DN_OPEN)
898 if (type == CL_SETXATTR || type == CL_GETXATTR)
899 xflags |= CLFE_XATTR;
901 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
902 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
903 xflags & CLFE_SUPPORTED));
904 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
905 if (buf->lb_buf == NULL)
909 rec->cr_hdr.lrh_len = reclen;
910 rec->cr.cr_flags = clf_flags;
911 rec->cr.cr_type = (__u32)type;
912 rec->cr.cr_tfid = *fid;
913 rec->cr.cr_namelen = 0;
915 if (clf_flags & CLF_JOBID)
916 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
918 if (clf_flags & CLF_EXTRA_FLAGS) {
919 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
920 if (xflags & CLFE_UIDGID)
921 mdd_changelog_rec_extra_uidgid(&rec->cr,
922 uc->uc_uid, uc->uc_gid);
923 if (xflags & CLFE_NID)
924 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
925 if (xflags & CLFE_OPEN)
926 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
927 if (xflags & CLFE_XATTR) {
928 if (xattr_name == NULL)
930 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
934 rc = mdd_changelog_store(env, mdd, rec, handle);
939 /** Store a data change changelog record
940 * If this fails, we must fail the whole transaction; we don't
941 * want the change to commit without the log entry.
942 * \param mdd_obj - mdd_object of change
943 * \param handle - transaction handle
945 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
946 enum changelog_rec_type type,
947 enum changelog_rec_flags clf_flags,
948 struct mdd_object *mdd_obj, struct thandle *handle)
952 LASSERT(mdd_obj != NULL);
953 LASSERT(handle != NULL);
955 if (!mdd_changelog_enabled(env, mdd, type))
958 if (mdd_is_volatile_obj(mdd_obj))
961 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
962 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
963 /* Don't need multiple updates in this log */
964 /* Don't check under lock - no big deal if we get an extra
969 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
970 mdd_object_fid(mdd_obj),
973 mdd_obj->mod_cltime = ktime_get();
978 int mdd_changelog_data_store_xattr(const struct lu_env *env,
979 struct mdd_device *mdd,
980 enum changelog_rec_type type,
981 enum changelog_rec_flags clf_flags,
982 struct mdd_object *mdd_obj,
983 const char *xattr_name,
984 struct thandle *handle)
988 LASSERT(mdd_obj != NULL);
989 LASSERT(handle != NULL);
991 if (!mdd_changelog_enabled(env, mdd, type))
994 if (mdd_is_volatile_obj(mdd_obj))
997 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
998 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
999 /* Don't need multiple updates in this log */
1000 /* Don't check under lock - no big deal if we get an extra
1006 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1007 mdd_object_fid(mdd_obj),
1008 xattr_name, handle);
1010 mdd_obj->mod_cltime = ktime_get();
1015 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
1016 * except for open flags have a dedicated record to store 32 bits of flags */
1017 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1018 enum changelog_rec_flags clf_flags,
1019 struct md_device *m, const struct lu_fid *fid)
1021 struct thandle *handle;
1022 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1026 LASSERT(fid != NULL);
1028 /* We'll check this again below, but we check now before we
1029 * start a transaction. */
1030 if (!mdd_changelog_enabled(env, mdd, type))
1033 handle = mdd_trans_create(env, mdd);
1035 RETURN(PTR_ERR(handle));
1037 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1041 rc = mdd_trans_start(env, mdd, handle);
1045 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1049 rc = mdd_trans_stop(env, mdd, rc, handle);
1055 * Save LMA extended attributes with data from \a ma.
1057 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1058 * not, LMA EA will be first read from disk, modified and write back.
1061 /* Precedence for choosing record type when multiple
1062 * attributes change: setattr > mtime > ctime > atime
1063 * (ctime changes when mtime does, plus chmod/chown.
1064 * atime and ctime are independent.) */
1065 static int mdd_attr_set_changelog(const struct lu_env *env,
1066 struct md_object *obj, struct thandle *handle,
1069 struct mdd_device *mdd = mdo2mdd(obj);
1072 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
1073 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
1074 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
1075 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
1076 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
1077 bits = bits & mdd->mdd_cl.mc_mask;
1078 /* This is an implementation limit rather than a protocol limit */
1079 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1083 /* The record type is the lowest non-masked set bit */
1086 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1087 return mdd_changelog_data_store(env, mdd, type, valid,
1088 md2mdd_obj(obj), handle);
1091 static int mdd_declare_attr_set(const struct lu_env *env,
1092 struct mdd_device *mdd,
1093 struct mdd_object *obj,
1094 const struct lu_attr *attr,
1095 struct thandle *handle)
1099 rc = mdo_declare_attr_set(env, obj, attr, handle);
1103 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1104 if (attr->la_valid & LA_MODE) {
1105 mdd_read_lock(env, obj, DT_TGT_CHILD);
1106 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1107 XATTR_NAME_ACL_ACCESS);
1108 mdd_read_unlock(env, obj);
1109 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1115 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1116 rc = mdo_declare_xattr_set(env, obj, buf,
1117 XATTR_NAME_ACL_ACCESS, 0,
1125 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1134 * permission changes may require sync operation, to mitigate performance
1135 * impact, only do this for dir and when permission is reduced.
1137 * For regular files, version is updated with permission change (see VBR), async
1138 * permission won't cause any issue, while missing permission change on
1139 * directory may affect accessibility of other objects after recovery.
1141 static inline bool permission_needs_sync(const struct lu_attr *old,
1142 const struct lu_attr *new)
1144 if (!S_ISDIR(old->la_mode))
1147 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1150 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1153 if (new->la_valid & LA_MODE) {
1154 /* turned on sticky bit */
1155 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1158 /* set-GID has no impact on what is allowed, not checked */
1160 /* turned off setuid bit, or one of rwx for someone */
1161 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1162 (old->la_mode & (0777 | S_ISUID)))
1169 static inline __u64 mdd_lmm_dom_size(void *buf)
1171 struct lov_mds_md *lmm = buf;
1172 struct lov_comp_md_v1 *comp_v1;
1173 struct lov_mds_md *v1;
1179 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1182 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1183 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1184 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1186 /* DoM entry is the first entry always */
1187 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1188 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1193 /* set attr and LOV EA at once, return updated attr */
1194 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1195 const struct md_attr *ma)
1197 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1198 struct mdd_device *mdd = mdo2mdd(obj);
1199 struct thandle *handle = NULL;
1200 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1201 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1202 const struct lu_attr *la = &ma->ma_attr;
1203 struct lu_ucred *uc;
1204 bool chrgrp_by_unprivileged_user = false;
1208 /* we do not use ->attr_set() for LOV/HSM EA any more */
1209 LASSERT((ma->ma_valid & MA_LOV) == 0);
1210 LASSERT((ma->ma_valid & MA_HSM) == 0);
1212 rc = mdd_la_get(env, mdd_obj, attr);
1216 *la_copy = ma->ma_attr;
1217 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1221 /* no need to setattr anymore */
1222 if (la_copy->la_valid == 0) {
1224 "%s: no valid attribute on "DFID", previous was %#llx\n",
1225 mdd_obj_dev_name(mdd_obj),
1226 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1231 /* If an unprivileged user changes group of some file,
1232 * the setattr operation will be processed synchronously to
1233 * honor the quota limit of the corresponding group. see LU-5152 */
1234 uc = lu_ucred_check(env);
1235 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1236 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1237 /* LU-10048: disable synchronous chgrp operation for it will
1238 * cause deadlock between MDT and OST.
1239 la_copy->la_valid |= LA_FLAGS;
1240 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1242 chrgrp_by_unprivileged_user = true;
1244 /* Flush the possible existing client setattr requests to OSTs
1245 * to keep the order with the current setattr operation that
1246 * will be sent directly to OSTs. see LU-5152 */
1247 /* LU-11303 disable sync as this is too heavyweight.
1248 * This should be replaced with a sync only for the object
1249 * being modified here, not the whole filesystem.
1250 rc = dt_sync(env, mdd->mdd_child);
1256 handle = mdd_trans_create(env, mdd);
1257 if (IS_ERR(handle)) {
1258 rc = PTR_ERR(handle);
1264 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1268 rc = mdd_trans_start(env, mdd, handle);
1272 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1273 permission_needs_sync(attr, la))
1274 handle->th_sync = 1;
1276 if (la->la_valid & (LA_MTIME | LA_CTIME))
1277 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1278 la->la_mtime, la->la_ctime);
1280 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1282 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1283 * otherwise following truncate will do nothing and truncated
1284 * data may be read again. This is a quick fix until LU-11033
1287 if (la_copy->la_valid & LA_SIZE) {
1288 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1290 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1294 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1295 la_copy->la_valid &= ~LA_SIZE;
1296 lu_buf_free(lov_buf);
1300 if (la_copy->la_valid) {
1301 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1303 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1304 /* rollback to the original gid */
1305 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1306 la_copy->la_gid = attr->la_gid;
1307 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1310 mdd_write_unlock(env, mdd_obj);
1314 rc = mdd_attr_set_changelog(env, obj, handle,
1318 rc = mdd_trans_stop(env, mdd, rc, handle);
1323 static int mdd_xattr_sanity_check(const struct lu_env *env,
1324 struct mdd_object *obj,
1325 const struct lu_attr *attr,
1328 struct lu_ucred *uc = lu_ucred_assert(env);
1331 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1334 if (strncmp(XATTR_USER_PREFIX, name,
1335 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1336 /* For sticky directories, only the owner and privileged user
1337 * can write attributes. */
1338 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1339 (uc->uc_fsuid != attr->la_uid) &&
1340 !md_capable(uc, CFS_CAP_FOWNER))
1342 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1343 (uc->uc_fsuid != attr->la_uid) &&
1344 !md_capable(uc, CFS_CAP_FOWNER)) {
1352 * Check if a string begins with a given prefix.
1354 * \param str String to check
1355 * \param prefix Substring to check at the beginning of \a str
1356 * \return true/false whether the condition is verified.
1358 static inline bool has_prefix(const char *str, const char *prefix)
1360 return strncmp(prefix, str, strlen(prefix)) == 0;
1364 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1366 * \param[in] xattr_name Full extended attribute name.
1368 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1370 static enum changelog_rec_type
1371 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1372 const char *xattr_name)
1374 /* Layout changes systematically recorded */
1375 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1376 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1377 allowed_lustre_lov(xattr_name))
1380 /* HSM information changes systematically recorded */
1381 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1384 /* Avoid logging SOM xattr for every file */
1385 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1388 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1389 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1390 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1391 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1392 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1398 static int mdd_declare_xattr_set(const struct lu_env *env,
1399 struct mdd_device *mdd,
1400 struct mdd_object *obj,
1401 const struct lu_buf *buf,
1403 int fl, struct thandle *handle)
1405 enum changelog_rec_type type;
1408 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1412 type = mdd_xattr_changelog_type(env, mdd, name);
1414 return 0; /* no changelog to store */
1416 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1420 * Compare current and future data of HSM EA and add a changelog if needed.
1422 * Caller should have write-locked \param obj.
1424 * \param buf - Future HSM EA content.
1425 * \retval 0 if no changelog is needed or changelog was added properly.
1426 * \retval -ve errno if there was a problem
1428 static int mdd_hsm_update_locked(const struct lu_env *env,
1429 struct md_object *obj,
1430 const struct lu_buf *buf,
1431 struct thandle *handle,
1432 enum changelog_rec_flags *clf_flags)
1434 struct mdd_thread_info *info = mdd_env_info(env);
1435 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1436 struct lu_buf *current_buf;
1437 struct md_hsm *current_mh;
1438 struct md_hsm *new_mh;
1442 OBD_ALLOC_PTR(current_mh);
1443 if (current_mh == NULL)
1446 /* Read HSM attrs from disk */
1447 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1449 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1451 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1452 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1453 if (rc < 0 && rc != -ENODATA)
1455 else if (rc == -ENODATA)
1456 current_mh->mh_flags = 0;
1458 /* Map future HSM xattr */
1459 OBD_ALLOC_PTR(new_mh);
1461 GOTO(free, rc = -ENOMEM);
1462 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1466 /* Flags differ, set flags for the changelog that will be added */
1467 if (current_mh->mh_flags != new_mh->mh_flags) {
1468 hsm_set_cl_event(clf_flags, HE_STATE);
1469 if (new_mh->mh_flags & HS_DIRTY)
1470 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1473 OBD_FREE_PTR(new_mh);
1476 OBD_FREE_PTR(current_mh);
1480 static int mdd_object_pfid_replace(const struct lu_env *env,
1481 struct mdd_object *o)
1483 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1484 struct thandle *handle;
1487 handle = mdd_trans_create(env, mdd);
1489 RETURN(PTR_ERR(handle));
1491 handle->th_complex = 1;
1493 /* it doesn't need to track the PFID update via llog, because LFSCK
1494 * will repair it even it goes wrong */
1495 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1500 rc = mdd_trans_start(env, mdd, handle);
1504 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1509 mdd_trans_stop(env, mdd, rc, handle);
1514 static int mdd_declare_xattr_del(const struct lu_env *env,
1515 struct mdd_device *mdd,
1516 struct mdd_object *obj,
1518 struct thandle *handle);
1520 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1523 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1524 struct md_object *md_vic)
1526 struct mdd_device *mdd = mdo2mdd(md_obj);
1527 struct mdd_object *obj = md2mdd_obj(md_obj);
1528 struct mdd_object *vic = md2mdd_obj(md_vic);
1529 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1530 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1531 struct lov_mds_md *lmm;
1532 struct thandle *handle;
1536 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1537 if (rc == 0) /* same fid */
1540 handle = mdd_trans_create(env, mdd);
1542 RETURN(PTR_ERR(handle));
1545 mdd_write_lock(env, obj, DT_TGT_CHILD);
1546 mdd_write_lock(env, vic, DT_TGT_CHILD);
1548 mdd_write_lock(env, vic, DT_TGT_CHILD);
1549 mdd_write_lock(env, obj, DT_TGT_CHILD);
1552 /* get EA of victim file */
1553 memset(buf_vic, 0, sizeof(*buf_vic));
1554 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1561 /* parse the layout of victim file */
1562 lmm = buf_vic->lb_buf;
1563 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1564 GOTO(out, rc = -EINVAL);
1566 /* save EA of target file for restore */
1567 memset(buf, 0, sizeof(*buf));
1568 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1572 /* Get rid of the layout from victim object */
1573 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1577 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1578 LU_XATTR_MERGE, handle);
1582 rc = mdd_trans_start(env, mdd, handle);
1586 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1591 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1593 GOTO(out_restore, rc);
1595 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1596 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1601 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1602 LU_XATTR_REPLACE, handle);
1604 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1605 mdd_obj_dev_name(obj),
1606 PFID(mdd_object_fid(obj)), rc2);
1610 mdd_trans_stop(env, mdd, rc, handle);
1611 mdd_write_unlock(env, obj);
1612 mdd_write_unlock(env, vic);
1614 lu_buf_free(buf_vic);
1617 (void) mdd_object_pfid_replace(env, obj);
1623 * Extract the mirror with specified mirror id, and store the splitted
1624 * mirror layout to @buf.
1626 * \param[in] comp_v1 mirrored layout
1627 * \param[in] mirror_id the mirror with mirror_id to be extracted
1628 * \param[out] buf store the layout excluding the extracted mirror,
1629 * caller free the buffer we allocated in this function
1630 * \param[out] buf_vic store the extracted layout, caller free the buffer
1631 * we allocated in this function
1633 * \retval 0 on success; < 0 if error happens
1635 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1636 struct lu_buf *buf, struct lu_buf *buf_vic)
1638 struct lov_comp_md_v1 *comp_rem;
1639 struct lov_comp_md_v1 *comp_vic;
1640 struct lov_comp_md_entry_v1 *entry;
1641 struct lov_comp_md_entry_v1 *entry_rem;
1642 struct lov_comp_md_entry_v1 *entry_vic;
1644 __u16 comp_cnt, count = 0;
1645 int lmm_size, lmm_size_vic = 0;
1647 int offset, offset_rem, offset_vic;
1649 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1650 /* comp_v1 should contains more than 1 mirror */
1651 if (mirror_cnt <= 1)
1653 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1654 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1656 for (i = 0; i < comp_cnt; i++) {
1657 entry = &comp_v1->lcm_entries[i];
1658 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1660 lmm_size_vic += sizeof(*entry);
1661 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1662 } else if (count > 0) {
1663 /* find the specified mirror */
1671 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1675 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1676 if (!buf_vic->lb_buf) {
1681 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1682 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1684 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1685 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1686 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1687 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1688 if (!comp_rem->lcm_mirror_count)
1689 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1691 memset(comp_vic, 0, sizeof(*comp_v1));
1692 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1693 comp_vic->lcm_mirror_count = 0;
1694 comp_vic->lcm_entry_count = cpu_to_le32(count);
1695 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1696 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1697 comp_vic->lcm_layout_gen = 0;
1699 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1700 offset_rem = sizeof(*comp_rem) +
1701 sizeof(*entry_rem) * (comp_cnt - count);
1702 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1703 for (i = j = k = 0; i < comp_cnt; i++) {
1704 struct lov_mds_md *lmm, *lmm_dst;
1707 entry = &comp_v1->lcm_entries[i];
1708 entry_vic = &comp_vic->lcm_entries[j];
1709 entry_rem = &comp_rem->lcm_entries[k];
1711 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1714 /* copy component entry */
1716 memcpy(entry_vic, entry, sizeof(*entry));
1717 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1718 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1721 memcpy(entry_rem, entry, sizeof(*entry));
1722 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1726 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1728 lmm_dst = (struct lov_mds_md *)
1729 ((char *)comp_vic + offset_vic);
1731 lmm_dst = (struct lov_mds_md *)
1732 ((char *)comp_rem + offset_rem);
1734 /* copy component entry blob */
1735 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1737 /* blob offset advance */
1738 offset += le32_to_cpu(entry->lcme_size);
1740 offset_vic += le32_to_cpu(entry->lcme_size);
1742 offset_rem += le32_to_cpu(entry->lcme_size);
1748 static int mdd_dom_data_truncate(const struct lu_env *env,
1749 struct mdd_device *mdd, struct mdd_object *mo);
1751 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1752 struct md_rejig_data *mrd)
1754 struct mdd_device *mdd = mdo2mdd(md_obj);
1755 struct mdd_object *obj = md2mdd_obj(md_obj);
1756 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1757 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1758 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1759 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1760 struct lov_comp_md_v1 *lcm;
1761 struct thandle *handle;
1763 bool dom_stripe = false;
1767 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1768 if (rc == 0) /* same fid */
1771 handle = mdd_trans_create(env, mdd);
1773 RETURN(PTR_ERR(handle));
1776 mdd_write_lock(env, obj, DT_TGT_CHILD);
1777 mdd_write_lock(env, vic, DT_TGT_CHILD);
1779 mdd_write_lock(env, vic, DT_TGT_CHILD);
1780 mdd_write_lock(env, obj, DT_TGT_CHILD);
1783 /* get EA of mirrored file */
1784 memset(buf_save, 0, sizeof(*buf));
1785 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1789 lcm = buf_save->lb_buf;
1790 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1791 GOTO(out, rc = -EINVAL);
1794 * Extract the mirror with specified mirror id, and store the splitted
1795 * mirror layout to the victim file.
1797 memset(buf, 0, sizeof(*buf));
1798 memset(buf_vic, 0, sizeof(*buf_vic));
1799 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1803 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1805 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1806 LU_XATTR_SPLIT, handle);
1809 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1810 LU_XATTR_SPLIT, handle);
1814 rc = mdd_trans_start(env, mdd, handle);
1818 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1823 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1826 GOTO(out_restore, rc);
1828 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1832 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1839 /* restore obj's layout */
1840 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1841 LU_XATTR_REPLACE, handle);
1843 CERROR("%s: failed rollback "DFID" layout: file state unkonwn: rc = %d\n",
1844 mdd_obj_dev_name(obj),
1845 PFID(mdd_object_fid(obj)), rc2);
1848 rc = mdd_trans_stop(env, mdd, rc, handle);
1850 /* Truncate local DOM data if all went well */
1851 if (!rc && dom_stripe)
1852 mdd_dom_data_truncate(env, mdd, obj);
1854 mdd_write_unlock(env, obj);
1855 mdd_write_unlock(env, vic);
1856 lu_buf_free(buf_save);
1858 lu_buf_free(buf_vic);
1861 (void) mdd_object_pfid_replace(env, obj);
1866 static int mdd_layout_merge_allowed(const struct lu_env *env,
1867 struct md_object *target,
1868 struct md_object *victim)
1870 struct mdd_object *o1 = md2mdd_obj(target);
1872 /* cannot extend directory's LOVEA */
1873 if (S_ISDIR(mdd_object_type(o1))) {
1874 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1875 mdd_obj_dev_name(o1));
1883 * The caller should guarantee to update the object ctime
1884 * after xattr_set if needed.
1886 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1887 const struct lu_buf *buf, const char *name,
1890 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1891 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1892 struct mdd_device *mdd = mdo2mdd(obj);
1893 struct thandle *handle;
1894 enum changelog_rec_type cl_type;
1895 enum changelog_rec_flags clf_flags = 0;
1899 rc = mdd_la_get(env, mdd_obj, attr);
1903 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1907 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1908 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1909 struct md_rejig_data *mrd = buf->lb_buf;
1910 struct md_object *victim = mrd->mrd_obj;
1912 if (buf->lb_len != sizeof(*mrd))
1915 rc = mdd_layout_merge_allowed(env, obj, victim);
1919 if (fl == LU_XATTR_MERGE)
1920 /* merge layout of victim as a mirror of obj's. */
1921 rc = mdd_xattr_merge(env, obj, victim);
1923 rc = mdd_xattr_split(env, obj, mrd);
1927 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1928 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1929 struct posix_acl *acl;
1931 /* user may set empty ACL, which should be treated as removing
1933 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1936 RETURN(PTR_ERR(acl));
1938 rc = mdd_xattr_del(env, obj, name);
1941 posix_acl_release(acl);
1944 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1945 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1949 handle = mdd_trans_create(env, mdd);
1951 RETURN(PTR_ERR(handle));
1953 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1957 rc = mdd_trans_start(env, mdd, handle);
1961 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1963 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1964 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1966 mdd_write_unlock(env, mdd_obj);
1971 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1972 mdd_write_unlock(env, mdd_obj);
1976 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1980 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1981 mdd_obj, name, handle);
1985 return mdd_trans_stop(env, mdd, rc, handle);
1988 static int mdd_declare_xattr_del(const struct lu_env *env,
1989 struct mdd_device *mdd,
1990 struct mdd_object *obj,
1992 struct thandle *handle)
1994 enum changelog_rec_type type;
1997 rc = mdo_declare_xattr_del(env, obj, name, handle);
2001 type = mdd_xattr_changelog_type(env, mdd, name);
2003 return 0; /* no changelog to store */
2005 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2009 * The caller should guarantee to update the object ctime
2010 * after xattr_set if needed.
2012 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2015 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2016 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2017 struct mdd_device *mdd = mdo2mdd(obj);
2018 struct thandle *handle;
2022 rc = mdd_la_get(env, mdd_obj, attr);
2026 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2030 handle = mdd_trans_create(env, mdd);
2032 RETURN(PTR_ERR(handle));
2034 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2038 rc = mdd_trans_start(env, mdd, handle);
2042 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2043 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2044 mdd_write_unlock(env, mdd_obj);
2048 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2051 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2056 return mdd_trans_stop(env, mdd, rc, handle);
2060 * read lov/lmv EA of an object
2061 * return the lov/lmv EA in an allocated lu_buf
2063 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2064 struct lu_buf *lmm_buf, const char *name)
2066 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2071 if (buf->lb_buf == NULL) {
2072 buf = lu_buf_check_and_alloc(buf, 4096);
2073 if (buf->lb_buf == NULL)
2078 rc = mdo_xattr_get(env, obj, buf, name);
2079 if (rc == -ERANGE) {
2080 /* mti_big_buf is allocated but is too small
2081 * we need to increase it */
2082 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2084 if (buf->lb_buf == NULL)
2087 } else if (rc < 0) {
2089 } else if (rc == 0) {
2093 lu_buf_alloc(lmm_buf, rc);
2094 if (lmm_buf->lb_buf == NULL)
2098 * we don't use lmm_buf directly, because we don't know xattr size, so
2099 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2101 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2106 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2107 struct mdd_object *o, struct lu_buf *buf,
2108 struct thandle *handle)
2110 struct hsm_attrs *attrs;
2111 enum hsm_states hsm_flags;
2112 enum changelog_rec_flags clf_flags = 0;
2116 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2121 attrs = buf->lb_buf;
2122 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2123 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2126 /* Add a changelog record for release. */
2127 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2128 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2129 clf_flags, o, handle);
2134 * check if layout swapping between 2 objects is allowed
2136 * - only normal FIDs or non-system IGIFs
2137 * - same type of objects
2138 * - same owner/group (so quotas are still valid) unless this is from HSM
2141 static int mdd_layout_swap_allowed(const struct lu_env *env,
2142 struct mdd_object *o1,
2143 const struct lu_attr *attr1,
2144 struct mdd_object *o2,
2145 const struct lu_attr *attr2,
2148 const struct lu_fid *fid1, *fid2;
2151 fid1 = mdd_object_fid(o1);
2152 fid2 = mdd_object_fid(o2);
2154 if (!fid_is_norm(fid1) &&
2155 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2158 if (!fid_is_norm(fid2) &&
2159 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2162 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2163 if (S_ISDIR(mdd_object_type(o1)))
2165 if (S_ISREG(mdd_object_type(o1)))
2170 if (flags & SWAP_LAYOUTS_MDS_HSM)
2173 if ((attr1->la_uid != attr2->la_uid) ||
2174 (attr1->la_gid != attr2->la_gid))
2180 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2181 * look into the layout in MDD layer. */
2182 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2184 struct lov_comp_md_v1 *comp_v1;
2185 struct lov_mds_md *v1;
2189 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2190 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2191 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2197 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2198 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2201 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2202 off = le32_to_cpu(comp_v1->lcm_entries[i].
2204 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2209 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2210 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2221 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2223 return mdd_lmm_oi(lmm, oi, true);
2226 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2228 return mdd_lmm_oi(lmm, oi, false);
2231 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2233 struct lov_comp_md_v1 *comp_v1;
2235 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2236 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2238 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2240 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2241 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2242 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2243 __u16 tmp_gen = *gen;
2245 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2247 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2254 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2256 return mdd_lmm_gen(lmm, gen, true);
2259 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2261 return mdd_lmm_gen(lmm, gen, false);
2264 static int mdd_dom_data_truncate(const struct lu_env *env,
2265 struct mdd_device *mdd, struct mdd_object *mo)
2268 struct dt_object *dom;
2271 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2273 GOTO(out, rc = -ENODATA);
2275 th = dt_trans_create(env, mdd->mdd_bottom);
2277 GOTO(out, rc = PTR_ERR(th));
2279 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2283 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2287 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2289 dt_trans_stop(env, mdd->mdd_bottom, th);
2291 /* Ignore failure but report the error */
2293 CERROR("%s: can't truncate DOM inode "DFID" data: rc = %d\n",
2294 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2299 * swap layouts between 2 lustre objects
2301 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2302 struct md_object *obj2, __u64 flags)
2304 struct mdd_thread_info *info = mdd_env_info(env);
2305 struct mdd_object *fst_o = md2mdd_obj(obj1);
2306 struct mdd_object *snd_o = md2mdd_obj(obj2);
2307 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2308 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2309 struct mdd_device *mdd = mdo2mdd(obj1);
2310 struct lov_mds_md *fst_lmm, *snd_lmm;
2311 struct lu_buf *fst_buf = &info->mti_buf[0];
2312 struct lu_buf *snd_buf = &info->mti_buf[1];
2313 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2314 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2315 struct ost_id *saved_oi = NULL;
2316 struct thandle *handle;
2317 struct mdd_object *dom_o = NULL;
2318 __u64 domsize_dom, domsize_vlt;
2319 __u32 fst_gen, snd_gen, saved_gen;
2325 BUILD_BUG_ON(ARRAY_SIZE(info->mti_buf) < 4);
2326 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2328 /* we have to sort the 2 obj, so locking will always
2329 * be in the same order, even in case of 2 concurrent swaps */
2330 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2331 if (rc == 0) /* same fid ? */
2337 rc = mdd_la_get(env, fst_o, fst_la);
2341 rc = mdd_la_get(env, snd_o, snd_la);
2345 /* check if layout swapping is allowed */
2346 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2350 handle = mdd_trans_create(env, mdd);
2352 RETURN(PTR_ERR(handle));
2354 /* objects are already sorted */
2355 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2356 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2358 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2359 if (rc < 0 && rc != -ENODATA)
2362 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2363 if (rc < 0 && rc != -ENODATA)
2366 /* check if file has DoM. DoM file can be migrated only to another
2367 * DoM layout with the same DoM component size or to an non-DOM
2368 * layout. After migration to OSTs layout, local MDT inode data
2369 * should be truncated.
2370 * Objects are sorted by FIDs, considering that original file's FID
2371 * is always smaller the snd_o is always original file we are migrating
2374 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2375 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2377 /* Only migration is supported for DoM files, not 'swap_layouts' so
2378 * target file must be volatile and orphan.
2380 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2381 dom_o = domsize_dom ? snd_o : NULL;
2382 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2383 swap(domsize_dom, domsize_vlt);
2384 dom_o = domsize_dom ? fst_o : NULL;
2385 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2386 /* 'lfs swap_layouts' case, neither file should have DoM */
2388 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2389 "use migration instead: rc = %d\n", rc);
2393 if (domsize_vlt > 0 && domsize_dom == 0) {
2396 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2397 mdd_obj_dev_name(snd_o),
2398 PFID(mdd_object_fid(snd_o)), rc);
2400 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2403 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2404 mdd_obj_dev_name(fst_o),
2405 PFID(mdd_object_fid(fst_o)), rc);
2407 } else if (domsize_vlt > 0) {
2408 /* Migration with the same DOM component size, no need to
2409 * truncate local data, it is still being used */
2413 /* swapping 2 non existant layouts is a success */
2414 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2417 /* to help inode migration between MDT, it is better to
2418 * start by the no layout file (if one), so we order the swap */
2419 if (snd_buf->lb_buf == NULL) {
2421 swap(fst_buf, snd_buf);
2424 fst_gen = snd_gen = 0;
2425 /* lmm and generation layout initialization */
2426 if (fst_buf->lb_buf != NULL) {
2427 fst_lmm = fst_buf->lb_buf;
2428 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2429 fst_fl = LU_XATTR_REPLACE;
2433 fst_fl = LU_XATTR_CREATE;
2436 snd_lmm = snd_buf->lb_buf;
2437 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2439 saved_gen = fst_gen;
2440 /* increase the generation layout numbers */
2445 * XXX The layout generation is used to generate component IDs for
2446 * the composite file, we have to do some special tweaks to make
2447 * sure the layout generation is always adequate for that job.
2450 /* Skip invalid generation number for composite layout */
2451 if ((snd_gen & LCME_ID_MASK) == 0)
2453 if ((fst_gen & LCME_ID_MASK) == 0)
2455 /* Make sure the generation is greater than all the component IDs */
2456 if (fst_gen < snd_gen)
2458 else if (fst_gen > snd_gen)
2461 /* set the file specific informations in lmm */
2462 if (fst_lmm != NULL) {
2463 saved_oi = &info->mti_oa.o_oi;
2464 mdd_get_lmm_oi(fst_lmm, saved_oi);
2465 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2466 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2467 mdd_set_lmm_oi(snd_lmm, saved_oi);
2469 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2470 cpu_to_le32(LOV_MAGIC_MAGIC))
2471 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2473 GOTO(stop, rc = -EPROTO);
2475 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2477 /* Prepare HSM attribute if it's required */
2478 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2479 const int buflen = sizeof(struct hsm_attrs);
2481 lu_buf_alloc(fst_hsm_buf, buflen);
2482 lu_buf_alloc(snd_hsm_buf, buflen);
2483 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2484 GOTO(stop, rc = -ENOMEM);
2486 /* Read HSM attribute */
2487 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2491 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2495 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2496 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2501 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2502 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2508 /* prepare transaction */
2509 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2514 if (fst_buf->lb_buf != NULL)
2515 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2516 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2519 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2524 rc = mdd_trans_start(env, mdd, handle);
2528 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2529 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2533 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2535 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2538 CERROR("%s: HSM error restoring "DFID": rc = %d/%d\n",
2539 mdd_obj_dev_name(fst_o),
2540 PFID(mdd_object_fid(fst_o)), rc, rc2);
2545 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2549 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2552 if (fst_buf->lb_buf != NULL)
2553 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2554 LU_XATTR_REPLACE, handle);
2556 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2559 GOTO(out_restore, rc);
2561 /* Issue one changelog record per file */
2562 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2566 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2575 /* failure on second file, but first was done, so we have
2576 * to roll back first. */
2577 if (fst_buf->lb_buf != NULL) {
2578 mdd_set_lmm_oi(fst_lmm, saved_oi);
2579 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2580 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2581 LU_XATTR_REPLACE, handle);
2583 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2589 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2594 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2599 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
2600 mdd_obj_dev_name(fst_o),
2601 PFID(mdd_object_fid(snd_o)),
2602 PFID(mdd_object_fid(fst_o)),
2604 /* a solution to avoid journal commit is to panic,
2605 * but it has strong consequences so we use LBUG to
2606 * allow sysdamin to choose to panic or not
2613 rc = mdd_trans_stop(env, mdd, rc, handle);
2615 /* Truncate local DOM data if all went well */
2617 mdd_dom_data_truncate(env, mdd, dom_o);
2619 mdd_write_unlock(env, snd_o);
2620 mdd_write_unlock(env, fst_o);
2622 lu_buf_free(fst_buf);
2623 lu_buf_free(snd_buf);
2624 lu_buf_free(fst_hsm_buf);
2625 lu_buf_free(snd_hsm_buf);
2628 (void) mdd_object_pfid_replace(env, fst_o);
2629 (void) mdd_object_pfid_replace(env, snd_o);
2634 static int mdd_declare_layout_change(const struct lu_env *env,
2635 struct mdd_device *mdd,
2636 struct mdd_object *obj,
2637 struct md_layout_change *mlc,
2638 struct thandle *handle)
2642 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2646 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2650 /* For PFL, this is used to instantiate necessary component objects. */
2652 mdd_layout_instantiate_component(const struct lu_env *env,
2653 struct mdd_object *obj, struct md_layout_change *mlc,
2654 struct thandle *handle)
2656 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2660 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2663 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2665 * It's possible that another layout write intent has already
2666 * instantiated our objects, so a -EALREADY returned, and we need to
2670 RETURN(rc == -EALREADY ? 0 : rc);
2672 rc = mdd_trans_start(env, mdd, handle);
2676 mdd_write_lock(env, obj, DT_TGT_CHILD);
2677 rc = mdo_layout_change(env, obj, mlc, handle);
2678 mdd_write_unlock(env, obj);
2682 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2687 * Change the FLR layout from RDONLY to WRITE_PENDING.
2689 * It picks the primary mirror, and bumps the layout version, and set
2690 * layout version xattr to OST objects in a sync tx. In order to facilitate
2691 * the handling of phantom writers from evicted clients, the clients carry
2692 * layout version of the file with write RPC, so that the OSTs can verify
2693 * if the write RPCs are legitimate, meaning not from evicted clients.
2696 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2697 struct md_layout_change *mlc, struct thandle *handle)
2699 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2700 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2701 struct lustre_som_attrs *som = &mlc->mlc_som;
2706 /* Verify acceptable operations */
2707 switch (mlc->mlc_opc) {
2708 case MD_LAYOUT_WRITE:
2709 case MD_LAYOUT_RESYNC:
2710 /* these are legal operations - this represents the case that
2711 * a few mirrors were missed in the last resync. */
2713 case MD_LAYOUT_RESYNC_DONE:
2718 som_buf->lb_buf = som;
2719 som_buf->lb_len = sizeof(*som);
2720 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2721 if (rc < 0 && rc != -ENODATA)
2725 lustre_som_swab(som);
2726 if (som->lsa_valid & SOM_FL_STRICT)
2727 fl = LU_XATTR_REPLACE;
2730 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2735 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2736 XATTR_NAME_SOM, fl, handle);
2741 /* record a changelog for data mover to consume */
2742 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2746 rc = mdd_trans_start(env, mdd, handle);
2750 /* it needs a sync tx to make FLR to work properly */
2751 handle->th_sync = 1;
2753 mdd_write_lock(env, obj, DT_TGT_CHILD);
2754 rc = mdo_layout_change(env, obj, mlc, handle);
2756 /* SOM state transition from STRICT to STALE */
2757 som->lsa_valid = SOM_FL_STALE;
2758 lustre_som_swab(som);
2759 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2762 mdd_write_unlock(env, obj);
2766 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2777 * Handle mirrored file state transition when it's in WRITE_PENDING.
2779 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2780 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2781 * layout version will be increased, and the file's state will be changed to
2785 mdd_layout_update_write_pending(const struct lu_env *env,
2786 struct mdd_object *obj, struct md_layout_change *mlc,
2787 struct thandle *handle)
2789 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2793 switch (mlc->mlc_opc) {
2794 case MD_LAYOUT_RESYNC:
2795 /* Upon receiving the resync request, it should
2796 * instantiate all stale components right away to get ready
2797 * for mirror copy. In order to avoid layout version change,
2798 * client should avoid sending LAYOUT_WRITE request at the
2801 case MD_LAYOUT_WRITE:
2802 /* legal race for concurrent write, the file state has been
2803 * changed by another client. */
2809 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2813 rc = mdd_trans_start(env, mdd, handle);
2817 /* it needs a sync tx to make FLR to work properly */
2818 handle->th_sync = 1;
2820 mdd_write_lock(env, obj, DT_TGT_CHILD);
2821 rc = mdo_layout_change(env, obj, mlc, handle);
2822 mdd_write_unlock(env, obj);
2833 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2835 * Only concurrent write and sync complete requests are possible when the
2836 * file is in SYNC_PENDING. For the latter request, it will pass in the
2837 * mirrors that have been synchronized, then the stale bit will be cleared
2838 * to make the file's state turn into RDONLY.
2839 * For concurrent write reqeust, it just needs to change the file's state
2840 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2841 * version because the version will be increased in the transition to
2842 * SYNC_PENDING later so that it can deny the write request from potential
2843 * evicted SYNC clients. */
2845 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2846 struct md_layout_change *mlc, struct thandle *handle)
2848 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2849 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2854 /* operation validation */
2855 switch (mlc->mlc_opc) {
2856 case MD_LAYOUT_RESYNC_DONE:
2857 /* resync complete. */
2858 case MD_LAYOUT_WRITE:
2859 /* concurrent write. */
2861 case MD_LAYOUT_RESYNC:
2862 /* resync again, most likely the previous run failed.
2863 * no-op if it's already in SYNC_PENDING state */
2869 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2870 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2871 if (rc < 0 && rc != -ENODATA)
2874 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2875 lustre_som_swab(&mlc->mlc_som);
2876 som_buf->lb_buf = &mlc->mlc_som;
2877 som_buf->lb_len = sizeof(mlc->mlc_som);
2880 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2884 /* record a changelog for the completion of resync */
2885 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2890 /* RESYNC_DONE has piggybacked size and blocks */
2892 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2893 XATTR_NAME_SOM, fl, handle);
2898 rc = mdd_trans_start(env, mdd, handle);
2902 /* it needs a sync tx to make FLR to work properly */
2903 handle->th_sync = 1;
2905 rc = mdo_layout_change(env, obj, mlc, handle);
2910 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2916 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2925 * Layout change callback for object.
2927 * This is only used by FLR for now. In the future, it can be exteneded to
2928 * handle all layout change.
2931 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2932 struct md_layout_change *mlc)
2934 struct mdd_object *obj = md2mdd_obj(o);
2935 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2936 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2937 struct lov_comp_md_v1 *lcm;
2938 struct thandle *handle;
2944 if (S_ISDIR(mdd_object_type(obj))) {
2945 switch (mlc->mlc_opc) {
2946 case MD_LAYOUT_SHRINK:
2947 rc = mdd_dir_layout_shrink(env, o, mlc);
2956 /* Verify acceptable operations */
2957 switch (mlc->mlc_opc) {
2958 case MD_LAYOUT_WRITE:
2959 case MD_LAYOUT_RESYNC:
2960 case MD_LAYOUT_RESYNC_DONE:
2966 handle = mdd_trans_create(env, mdd);
2968 RETURN(PTR_ERR(handle));
2970 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2977 /* analyze the layout to make sure it's a FLR file */
2979 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2980 GOTO(out, rc = -EINVAL);
2982 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2984 /* please refer to HLD of FLR for state transition */
2985 switch (flr_state) {
2987 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2989 case LCM_FL_WRITE_PENDING:
2990 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2993 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2995 case LCM_FL_SYNC_PENDING:
2996 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3005 mdd_trans_stop(env, mdd, rc, handle);
3010 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3011 struct mdd_object *child, const struct lu_attr *attr,
3012 const struct md_op_spec *spec,
3013 struct dt_allocation_hint *hint)
3015 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3016 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3017 struct dt_object *nc = mdd_object_child(child);
3019 memset(hint, 0, sizeof(*hint));
3021 /* For striped directory, give striping EA to lod_ah_init, which will
3022 * decide the stripe_offset and stripe count by it. */
3023 if (S_ISDIR(attr->la_mode) &&
3024 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3025 hint->dah_eadata = spec->u.sp_ea.eadata;
3026 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3028 hint->dah_eadata = NULL;
3029 hint->dah_eadata_len = 0;
3030 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
3031 if (mdd->mdd_append_stripe_count != 0 ||
3032 mdd->mdd_append_pool[0])
3034 "using O_APPEND file striping\n");
3035 if (mdd->mdd_append_stripe_count)
3036 hint->dah_append_stripes =
3037 mdd->mdd_append_stripe_count;
3038 if (mdd->mdd_append_pool[0])
3039 hint->dah_append_pool = mdd->mdd_append_pool;
3041 hint->dah_append_stripes = 0;
3045 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3046 hint->dah_eadata, hint->dah_eadata_len);
3047 /* @hint will be initialized by underlying device. */
3048 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3051 static int accmode(const struct lu_env *env, const struct lu_attr *la,
3054 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3055 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3056 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3057 * owner can write to a file even if it is marked readonly to hide
3058 * its brokenness. (bug 5781) */
3059 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3060 struct lu_ucred *uc = lu_ucred_check(env);
3062 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3066 return mds_accmode(open_flags);
3069 static int mdd_open_sanity_check(const struct lu_env *env,
3070 struct mdd_object *obj,
3071 const struct lu_attr *attr, u64 open_flags,
3077 /* EEXIST check, also opening of *open* orphans is allowed so we can
3078 * open-by-handle unlinked files
3080 if (mdd_is_dead_obj(obj) && !is_replay &&
3081 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3084 if (S_ISLNK(attr->la_mode))
3087 mode = accmode(env, attr, open_flags);
3089 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3092 if (!(open_flags & MDS_OPEN_CREATED)) {
3093 rc = mdd_permission_internal(env, obj, attr, mode);
3098 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3099 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3100 open_flags &= ~MDS_OPEN_TRUNC;
3102 /* For writing append-only file must open it with append mode. */
3103 if (attr->la_flags & LUSTRE_APPEND_FL) {
3104 if ((open_flags & MDS_FMODE_WRITE) &&
3105 !(open_flags & MDS_OPEN_APPEND))
3107 if (open_flags & MDS_OPEN_TRUNC)
3114 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3115 u64 open_flags, struct md_op_spec *spec)
3117 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3118 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3119 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3120 struct mdd_object_user *mou = NULL;
3121 const struct lu_ucred *uc = lu_ucred(env);
3122 struct mdd_device *mdd = mdo2mdd(obj);
3123 enum changelog_rec_type type = CL_OPEN;
3127 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3129 rc = mdd_la_get(env, mdd_obj, attr);
3133 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3135 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3140 mdd_obj->mod_count++;
3142 if (!mdd_changelog_enabled(env, mdd, type))
3146 /* look for existing opener in list under mdd_write_lock */
3147 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3152 /* add user to list */
3153 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3159 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3161 mdd_obj_user_free(mou);
3166 if (type == CL_DN_OPEN) {
3167 if (ktime_before(ktime_get(), mou->mou_deniednext))
3168 /* same user denied again same access within
3169 * time interval: do not record
3173 /* this user already denied, but some time ago:
3174 * update denied time
3176 mou->mou_deniednext =
3177 ktime_add(ktime_get(),
3178 ktime_set(mdd->mdd_cl.mc_deniednext,
3181 mou->mou_opencount++;
3182 /* same user opening file again with same flags:
3189 /* FYI, only the bottom 32 bits of open_flags are recorded */
3190 mdd_changelog(env, type, open_flags, md_dev, mdd_object_fid(mdd_obj));
3194 mdd_write_unlock(env, mdd_obj);
3198 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3199 struct md_attr *ma, struct thandle *handle)
3203 rc = mdd_orphan_declare_delete(env, obj, handle);
3207 return mdo_declare_destroy(env, obj, handle);
3211 * No permission check is needed.
3213 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3214 struct md_attr *ma, u64 open_flags)
3216 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3217 struct mdd_device *mdd = mdo2mdd(obj);
3218 struct thandle *handle = NULL;
3221 bool blocked = false;
3222 bool last_close_by_uid = false;
3223 const struct lu_ucred *uc = lu_ucred(env);
3226 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3227 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3228 mdd_obj->mod_count--;
3229 mdd_write_unlock(env, mdd_obj);
3231 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3232 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3233 "list\n", PFID(mdd_object_fid(mdd_obj)));
3237 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3238 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3239 /* check without any lock */
3240 is_orphan = mdd_obj->mod_count == 1 &&
3241 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3245 /* mdd_trans_create() maybe failed because of barrier_entry(),
3246 * under such case, the orphan MDT-object will be left in the
3247 * orphan list, and when the MDT remount next time, the unused
3248 * orphans will be destroyed automatically.
3250 * One exception: the former mdd_finish_unlink may failed to
3251 * add the orphan MDT-object to the orphan list, then if the
3252 * mdd_trans_create() failed because of barrier_entry(), the
3253 * MDT-object will become real orphan that is neither in the
3254 * namespace nor in the orphan list. Such bad case should be
3255 * very rare and will be handled by e2fsck/lfsck. */
3256 handle = mdd_trans_create(env, mdo2mdd(obj));
3257 if (IS_ERR(handle)) {
3258 rc = PTR_ERR(handle);
3259 if (rc != -EINPROGRESS)
3267 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3271 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3276 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3282 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3283 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3285 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3286 lu_dev_name(mdd2lu_dev(mdd)),
3287 PFID(mdd_object_fid(mdd_obj)), rc);
3291 /* check again with lock */
3292 is_orphan = (mdd_obj->mod_count == 1) &&
3293 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3294 ma->ma_attr.la_nlink == 0);
3296 if (is_orphan && !handle && !blocked) {
3297 mdd_write_unlock(env, mdd_obj);
3301 mdd_obj->mod_count--; /*release open count */
3303 /* under mdd write lock */
3304 /* If recording, see if we need to remove UID from list. uc is not
3305 * initialized if the client has been evicted. */
3306 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3307 struct mdd_object_user *mou;
3309 /* look for UID in list */
3310 /* If mou is NULL, it probably means logging was enabled after
3311 * the user had the file open. So the corresponding close
3312 * will not be logged.
3314 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3317 mou->mou_opencount--;
3318 if (mou->mou_opencount == 0) {
3319 mdd_obj_user_remove(mdd_obj, mou);
3320 last_close_by_uid = true;
3325 if (!is_orphan || blocked)
3329 /* NB: Object maybe not in orphan list originally, it is rare case for
3330 * mdd_finish_unlink() failure, in that case, the object doesn't have
3331 * ORPHAN_OBJ flag */
3332 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3333 /* remove link to object from orphan index */
3334 LASSERT(handle != NULL);
3335 rc = mdd_orphan_delete(env, mdd_obj, handle);
3337 CERROR("%s: unable to delete "DFID" from orphan list: "
3338 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3339 PFID(mdd_object_fid(mdd_obj)), rc);
3340 /* If object was not deleted from orphan list, do not
3341 * destroy OSS objects, which will be done when next
3346 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3347 "list, OSS objects to be destroyed.\n",
3348 PFID(mdd_object_fid(mdd_obj)));
3351 rc = mdo_destroy(env, mdd_obj, handle);
3354 CERROR("%s: unable to delete "DFID" from orphan list: "
3355 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3356 PFID(mdd_object_fid(mdd_obj)), rc);
3361 mdd_write_unlock(env, mdd_obj);
3363 if (rc != 0 || blocked ||
3364 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3367 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3368 * or if CL_OPEN was recorded and it's last close by user.
3369 * Changelogs mask may change between open and close operations, but
3370 * this is not a big deal if we have a CL_CLOSE entry with no matching
3371 * CL_OPEN. Plus Changelogs mask may not change often.
3373 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3374 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3375 MDS_OPEN_TRUNC))) ||
3376 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3377 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3378 if (handle == NULL) {
3379 handle = mdd_trans_create(env, mdo2mdd(obj));
3381 GOTO(stop, rc = PTR_ERR(handle));
3383 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3384 NULL, NULL, handle);
3388 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3393 /* FYI, only the bottom 32 bits of open_flags are recorded */
3394 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3399 if (handle != NULL && !IS_ERR(handle))
3400 rc = mdd_trans_stop(env, mdd, rc, handle);
3406 * Permission check is done when open,
3407 * no need check again.
3409 static int mdd_readpage_sanity_check(const struct lu_env *env,
3410 struct mdd_object *obj)
3412 struct dt_object *next = mdd_object_child(obj);
3416 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3424 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3425 size_t nob, const struct dt_it_ops *iops,
3426 struct dt_it *it, __u32 attr, void *arg)
3428 struct lu_dirpage *dp = &lp->lp_dir;
3432 struct lu_dirent *ent;
3433 struct lu_dirent *last = NULL;
3437 if (nob < sizeof(*dp))
3440 memset(area, 0, sizeof (*dp));
3441 area += sizeof (*dp);
3442 nob -= sizeof (*dp);
3449 len = iops->key_size(env, it);
3451 /* IAM iterator can return record with zero len. */
3455 hash = iops->store(env, it);
3456 if (unlikely(first)) {
3458 dp->ldp_hash_start = cpu_to_le64(hash);
3461 /* calculate max space required for lu_dirent */
3462 recsize = lu_dirent_calc_size(len, attr);
3464 if (nob >= recsize) {
3465 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3466 if (result == -ESTALE)
3471 /* osd might not able to pack all attributes,
3472 * so recheck rec length */
3473 recsize = le16_to_cpu(ent->lde_reclen);
3475 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3476 fid_le_to_cpu(&fid, &ent->lde_fid);
3477 if (fid_is_dot_lustre(&fid))
3481 result = (last != NULL) ? 0 :-EINVAL;
3485 ent = (void *)ent + recsize;
3489 result = iops->next(env, it);
3490 if (result == -ESTALE)
3492 } while (result == 0);
3495 dp->ldp_hash_end = cpu_to_le64(hash);
3497 if (last->lde_hash == dp->ldp_hash_end)
3498 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3499 last->lde_reclen = 0; /* end mark */
3502 /* end of directory */
3503 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3504 else if (result < 0)
3505 CWARN("build page failed: %d!\n", result);
3509 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3510 const struct lu_rdpg *rdpg)
3512 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3516 if (mdd_object_exists(mdd_obj) == 0) {
3517 CERROR("%s: object "DFID" not found: rc = -2\n",
3518 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3522 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3523 rc = mdd_readpage_sanity_check(env, mdd_obj);
3525 GOTO(out_unlock, rc);
3527 if (mdd_is_dead_obj(mdd_obj)) {
3529 struct lu_dirpage *dp;
3532 * According to POSIX, please do not return any entry to client:
3533 * even dot and dotdot should not be returned.
3535 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3536 PFID(mdd_object_fid(mdd_obj)));
3538 if (rdpg->rp_count <= 0)
3539 GOTO(out_unlock, rc = -EFAULT);
3540 LASSERT(rdpg->rp_pages != NULL);
3542 pg = rdpg->rp_pages[0];
3543 dp = (struct lu_dirpage *)kmap(pg);
3544 memset(dp, 0 , sizeof(struct lu_dirpage));
3545 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3546 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3547 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3549 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3552 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3553 mdd_dir_page_build, NULL);
3555 struct lu_dirpage *dp;
3557 dp = kmap(rdpg->rp_pages[0]);
3558 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3561 * No pages were processed, mark this for first page
3564 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3565 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3566 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3568 kunmap(rdpg->rp_pages[0]);
3571 GOTO(out_unlock, rc);
3573 mdd_read_unlock(env, mdd_obj);
3577 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3579 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3581 if (mdd_object_exists(mdd_obj) == 0) {
3584 CERROR("%s: object "DFID" not found: rc = %d\n",
3585 mdd_obj_dev_name(mdd_obj),
3586 PFID(mdd_object_fid(mdd_obj)), rc);
3589 return dt_object_sync(env, mdd_object_child(mdd_obj),
3593 static int mdd_object_lock(const struct lu_env *env,
3594 struct md_object *obj,
3595 struct lustre_handle *lh,
3596 struct ldlm_enqueue_info *einfo,
3597 union ldlm_policy_data *policy)
3599 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3600 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3604 static int mdd_object_unlock(const struct lu_env *env,
3605 struct md_object *obj,
3606 struct ldlm_enqueue_info *einfo,
3607 union ldlm_policy_data *policy)
3609 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3610 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3613 const struct md_object_operations mdd_obj_ops = {
3614 .moo_permission = mdd_permission,
3615 .moo_attr_get = mdd_attr_get,
3616 .moo_attr_set = mdd_attr_set,
3617 .moo_xattr_get = mdd_xattr_get,
3618 .moo_xattr_set = mdd_xattr_set,
3619 .moo_xattr_list = mdd_xattr_list,
3620 .moo_invalidate = mdd_invalidate,
3621 .moo_xattr_del = mdd_xattr_del,
3622 .moo_swap_layouts = mdd_swap_layouts,
3623 .moo_open = mdd_open,
3624 .moo_close = mdd_close,
3625 .moo_readpage = mdd_readpage,
3626 .moo_readlink = mdd_readlink,
3627 .moo_changelog = mdd_changelog,
3628 .moo_object_sync = mdd_object_sync,
3629 .moo_object_lock = mdd_object_lock,
3630 .moo_object_unlock = mdd_object_unlock,
3631 .moo_layout_change = mdd_layout_change,