4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <lustre_mds.h>
52 #include "mdd_internal.h"
54 static const struct lu_object_operations mdd_lu_obj_ops;
56 struct mdd_object_user {
57 struct list_head mou_list; /**< linked off mod_users */
58 u64 mou_open_flags; /**< open mode by client */
59 __u64 mou_uidgid; /**< uid_gid on client */
60 int mou_opencount; /**< # opened */
61 ktime_t mou_deniednext; /**< time of next access denied
66 static int mdd_xattr_get(const struct lu_env *env,
67 struct md_object *obj, struct lu_buf *buf,
70 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
71 struct mdd_device *mdd,
72 enum changelog_rec_type type,
73 enum changelog_rec_flags clf_flags,
74 const struct lu_fid *fid,
75 const char *xattr_name,
76 struct thandle *handle);
78 static inline bool has_prefix(const char *str, const char *prefix);
81 static u32 flags_helper(u64 open_flags)
85 if (open_flags & MDS_FMODE_EXEC) {
86 open_mode = MDS_FMODE_EXEC;
88 if (open_flags & MDS_FMODE_READ)
89 open_mode = MDS_FMODE_READ;
91 (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
92 open_mode |= MDS_FMODE_WRITE;
98 /** Allocate/init a user and its sub-structures.
103 * \retval mou [OUT] success valid structure
106 static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags,
107 uid_t uid, gid_t gid)
109 struct mdd_object_user *mou;
113 OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem);
115 RETURN(ERR_PTR(-ENOMEM));
117 mou->mou_open_flags = open_flags;
118 mou->mou_uidgid = ((__u64)uid << 32) | gid;
119 mou->mou_opencount = 0;
120 mou->mou_deniednext = ktime_set(0, 0);
126 * Free a user and its sub-structures.
128 * \param mou [IN] user to be freed.
130 static void mdd_obj_user_free(struct mdd_object_user *mou)
132 OBD_SLAB_FREE_PTR(mou, mdd_object_kmem);
136 * Find if UID/GID already has this file open
138 * Caller should have write-locked \param mdd_obj.
139 * \param mdd_obj [IN] mdd_obj
140 * \param uid [IN] client uid
141 * \param gid [IN] client gid
142 * \retval user pointer or NULL if not found
145 struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj,
146 uid_t uid, gid_t gid,
149 struct mdd_object_user *mou;
154 uidgid = ((__u64)uid << 32) | gid;
155 list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) {
156 if (mou->mou_uidgid == uidgid &&
157 flags_helper(mou->mou_open_flags) ==
158 flags_helper(open_flags))
165 * Add a user to the list of openers for this file
167 * Caller should have write-locked \param mdd_obj.
168 * \param mdd_obj [IN] mdd_obj
169 * \param mou [IN] user
171 * \retval -ve failure
173 static int mdd_obj_user_add(struct mdd_object *mdd_obj,
174 struct mdd_object_user *mou,
177 struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj);
178 struct mdd_object_user *tmp;
179 __u32 uid = mou->mou_uidgid >> 32;
180 __u32 gid = mou->mou_uidgid & ((1UL << 32) - 1);
183 tmp = mdd_obj_user_find(mdd_obj, uid, gid, mou->mou_open_flags);
187 list_add_tail(&mou->mou_list, &mdd_obj->mod_users);
190 /* next 'access denied' notification cannot happen before
193 mou->mou_deniednext =
194 ktime_add(ktime_get(),
195 ktime_set(mdd->mdd_cl.mc_deniednext, 0));
197 mou->mou_opencount++;
202 * Remove UID from the list
204 * Caller should have write-locked \param mdd_obj.
205 * \param mdd_obj [IN] mdd_obj
206 * \param uid [IN] user
207 * \retval -ve failure
209 static int mdd_obj_user_remove(struct mdd_object *mdd_obj,
210 struct mdd_object_user *mou)
217 list_del_init(&mou->mou_list);
219 mdd_obj_user_free(mou);
223 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
228 if (mdd_object_exists(obj) == 0)
231 rc = mdo_attr_get(env, obj, la);
232 if (unlikely(rc != 0)) {
234 obj->mod_flags |= DEAD_OBJ;
238 if (la->la_valid & LA_FLAGS && la->la_flags & LUSTRE_ORPHAN_FL)
239 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
244 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
246 struct mdd_thread_info *info;
248 lu_env_refill((struct lu_env *)env);
249 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
250 LASSERT(info != NULL);
254 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
258 buf = &mdd_env_info(env)->mti_buf[0];
264 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
265 const void *area, ssize_t len)
269 buf = &mdd_env_info(env)->mti_buf[0];
270 buf->lb_buf = (void *)area;
275 struct lu_object *mdd_object_alloc(const struct lu_env *env,
276 const struct lu_object_header *hdr,
279 struct mdd_object *mdd_obj;
282 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
286 o = mdd2lu_obj(mdd_obj);
287 lu_object_init(o, NULL, d);
288 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
289 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
290 mdd_obj->mod_count = 0;
291 o->lo_ops = &mdd_lu_obj_ops;
292 INIT_LIST_HEAD(&mdd_obj->mod_users);
297 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
298 const struct lu_object_conf *unused)
300 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
301 struct mdd_object *mdd_obj = lu2mdd_obj(o);
302 struct lu_object *below;
303 struct lu_device *under;
306 mdd_obj->mod_cltime = ktime_set(0, 0);
307 under = &d->mdd_child->dd_lu_dev;
308 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
310 RETURN(PTR_ERR(below));
312 lu_object_add(o, below);
317 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
321 if (lu_object_exists(o)) {
322 struct mdd_object *mdd_obj = lu2mdd_obj(o);
323 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
325 rc = mdd_la_get(env, mdd_obj, attr);
331 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
333 struct mdd_object *mdd = lu2mdd_obj(o);
334 struct mdd_object_user *mou, *tmp2;
337 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
338 list_del(&mou->mou_list);
339 mdd_obj_user_free(mou);
343 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
346 static int mdd_object_print(const struct lu_env *env, void *cookie,
347 lu_printer_t p, const struct lu_object *o)
349 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
351 return (*p)(env, cookie,
352 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
353 mdd, mdd->mod_count, mdd->mod_valid,
354 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
357 static const struct lu_object_operations mdd_lu_obj_ops = {
358 .loo_object_init = mdd_object_init,
359 .loo_object_start = mdd_object_start,
360 .loo_object_free = mdd_object_free,
361 .loo_object_print = mdd_object_print,
364 struct mdd_object *mdd_object_find(const struct lu_env *env,
365 struct mdd_device *d,
366 const struct lu_fid *f)
368 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
372 * No permission check is needed.
374 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
377 struct mdd_object *mdd_obj = md2mdd_obj(obj);
382 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
383 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
384 ma->ma_attr.la_nlink = 0;
390 * No permission check is needed.
392 static int mdd_xattr_get(const struct lu_env *env,
393 struct md_object *obj, struct lu_buf *buf,
396 struct mdd_object *mdd_obj = md2mdd_obj(obj);
397 struct mdd_device *mdd;
402 if (mdd_object_exists(mdd_obj) == 0) {
403 CERROR("%s: object "DFID" not found: rc = -2\n",
404 mdd_obj_dev_name(mdd_obj),
405 PFID(mdd_object_fid(mdd_obj)));
409 /* If the object has been destroyed, then do not get LMVEA, because
410 * it needs to load stripes from the iteration of the master object,
411 * and it will cause problem if master object has been destroyed, see
413 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
414 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
415 strcmp(name, XATTR_NAME_LMV) == 0))
418 /* If the object has been delete from the namespace, then
419 * get linkEA should return -ENOENT as well */
420 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
421 strcmp(name, XATTR_NAME_LINK) == 0))
424 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
425 rc = mdo_xattr_get(env, mdd_obj, buf, name);
426 mdd_read_unlock(env, mdd_obj);
430 /* record only getting user xattrs and acls */
431 if (rc >= 0 && buf->lb_buf &&
432 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
433 (has_prefix(name, XATTR_USER_PREFIX) ||
434 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
435 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
436 struct thandle *handle;
439 LASSERT(mdo2fid(mdd_obj) != NULL);
441 handle = mdd_trans_create(env, mdd);
443 RETURN(PTR_ERR(handle));
445 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
450 rc2 = mdd_trans_start(env, mdd, handle);
454 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
455 mdo2fid(mdd_obj), name,
459 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
468 * Permission check is done when open,
469 * no need check again.
471 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
474 struct mdd_object *mdd_obj = md2mdd_obj(obj);
475 struct dt_object *next;
480 if (mdd_object_exists(mdd_obj) == 0) {
481 CERROR("%s: object "DFID" not found: rc = -2\n",
482 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
486 next = mdd_object_child(mdd_obj);
487 LASSERT(next != NULL);
488 LASSERT(next->do_body_ops != NULL);
489 LASSERT(next->do_body_ops->dbo_read != NULL);
490 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
491 rc = dt_read(env, next, buf, &pos);
492 mdd_read_unlock(env, mdd_obj);
497 * No permission check is needed.
499 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
502 struct mdd_object *mdd_obj = md2mdd_obj(obj);
507 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
508 rc = mdo_xattr_list(env, mdd_obj, buf);
509 mdd_read_unlock(env, mdd_obj);
511 /* If the buffer is NULL then we are only here to get the
512 * length of the xattr name list. */
513 if (rc < 0 || buf->lb_buf == NULL)
517 * Filter out XATTR_NAME_LINK if this is an orphan object. See
520 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
521 char *end = (char *)buf->lb_buf + rc;
522 char *p = buf->lb_buf;
525 char *next = p + strlen(p) + 1;
527 if (strcmp(p, XATTR_NAME_LINK) == 0) {
529 memmove(p, next, end - next);
531 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
532 " of orphan "DFID"\n",
533 PFID(mdd_object_fid(mdd_obj)));
544 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
546 return mdo_invalidate(env, md2mdd_obj(obj));
549 int mdd_declare_create_object_internal(const struct lu_env *env,
550 struct mdd_object *p,
551 struct mdd_object *c,
552 struct lu_attr *attr,
553 struct thandle *handle,
554 const struct md_op_spec *spec,
555 struct dt_allocation_hint *hint)
557 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
558 const struct dt_index_features *feat = spec->sp_feat;
562 if (feat != &dt_directory_features && feat != NULL) {
563 dof->dof_type = DFT_INDEX;
564 dof->u.dof_idx.di_feat = feat;
566 dof->dof_type = dt_mode_to_dft(attr->la_mode);
567 if (dof->dof_type == DFT_REGULAR) {
568 dof->u.dof_reg.striped =
569 md_should_create(spec->sp_cr_flags);
570 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
571 dof->u.dof_reg.striped = 0;
572 /* is this replay? */
574 dof->u.dof_reg.striped = 0;
578 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
583 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
584 struct mdd_object *c, struct lu_attr *attr,
585 struct thandle *handle,
586 const struct md_op_spec *spec,
587 struct dt_allocation_hint *hint)
589 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
593 LASSERT(!mdd_object_exists(c));
595 rc = mdo_create_object(env, c, attr, hint, dof, handle);
600 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
601 const struct lu_attr *attr, struct thandle *handle,
607 rc = mdo_attr_set(env, obj, attr, handle);
608 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
609 if (!rc && (attr->la_valid & LA_MODE) && needacl)
610 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
615 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
616 const struct lu_attr *oattr, struct lu_attr *attr,
617 struct thandle *handle)
622 LASSERT(attr->la_valid & LA_CTIME);
623 LASSERT(oattr != NULL);
625 /* Make sure the ctime is increased only, however, it's not strictly
626 * reliable at here because there is not guarantee to hold lock on
627 * object, so we just bypass some unnecessary cmtime setting first
628 * and OSD has to check it again. */
629 if (attr->la_ctime < oattr->la_ctime)
630 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
631 else if (attr->la_valid == LA_CTIME &&
632 attr->la_ctime == oattr->la_ctime)
633 attr->la_valid &= ~LA_CTIME;
635 if (attr->la_valid != 0)
636 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
641 * This gives the same functionality as the code between
642 * sys_chmod and inode_setattr
643 * chown_common and inode_setattr
644 * utimes and inode_setattr
645 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
647 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
648 const struct lu_attr *oattr, struct lu_attr *la,
649 const unsigned long flags)
658 /* Do not permit change file type */
659 if (la->la_valid & LA_TYPE)
662 /* They should not be processed by setattr */
663 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
666 LASSERT(oattr != NULL);
668 uc = lu_ucred_check(env);
672 if (la->la_valid == LA_CTIME) {
673 if (!(flags & MDS_PERM_BYPASS))
674 /* This is only for set ctime when rename's source is
676 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
678 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
679 la->la_valid &= ~LA_CTIME;
683 if (flags & MDS_CLOSE_UPDATE_TIMES &&
684 la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
685 /* This is an atime/mtime/ctime attribute update for
688 if (la->la_valid & LA_ATIME &&
689 la->la_atime <= (oattr->la_atime +
690 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
691 la->la_valid &= ~LA_ATIME;
692 if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime)
693 la->la_valid &= ~LA_CTIME;
694 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
695 la->la_valid &= ~LA_MTIME;
699 /* Check if flags change. */
700 if (la->la_valid & LA_FLAGS) {
701 unsigned int oldflags = oattr->la_flags &
702 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
703 unsigned int newflags = la->la_flags &
704 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
706 if ((uc->uc_fsuid != oattr->la_uid) &&
707 !md_capable(uc, CFS_CAP_FOWNER))
710 /* The IMMUTABLE and APPEND_ONLY flags can
711 * only be changed by the relevant capability. */
712 if ((oldflags ^ newflags) &&
713 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
716 if (!S_ISDIR(oattr->la_mode))
717 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
720 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
721 (la->la_valid & ~LA_FLAGS) &&
722 !(flags & MDS_PERM_BYPASS))
725 /* Check for setting the obj time. */
726 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
727 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
728 if ((uc->uc_fsuid != oattr->la_uid) &&
729 !md_capable(uc, CFS_CAP_FOWNER)) {
730 rc = mdd_permission_internal(env, obj, oattr,
737 if (la->la_valid & LA_KILL_SUID) {
738 la->la_valid &= ~LA_KILL_SUID;
739 if ((oattr->la_mode & S_ISUID) &&
740 !(la->la_valid & LA_MODE)) {
741 la->la_mode = oattr->la_mode;
742 la->la_valid |= LA_MODE;
744 la->la_mode &= ~S_ISUID;
747 if (la->la_valid & LA_KILL_SGID) {
748 la->la_valid &= ~LA_KILL_SGID;
749 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
750 (S_ISGID | S_IXGRP)) &&
751 !(la->la_valid & LA_MODE)) {
752 la->la_mode = oattr->la_mode;
753 la->la_valid |= LA_MODE;
755 la->la_mode &= ~S_ISGID;
758 /* Make sure a caller can chmod. */
759 if (la->la_valid & LA_MODE) {
760 if (!(flags & MDS_PERM_BYPASS) &&
761 (uc->uc_fsuid != oattr->la_uid) &&
762 !md_capable(uc, CFS_CAP_FOWNER))
765 if (la->la_mode == (umode_t) -1)
766 la->la_mode = oattr->la_mode;
768 la->la_mode = (la->la_mode & S_IALLUGO) |
769 (oattr->la_mode & ~S_IALLUGO);
771 /* Also check the setgid bit! */
772 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
773 la->la_gid : oattr->la_gid) &&
774 !md_capable(uc, CFS_CAP_FSETID))
775 la->la_mode &= ~S_ISGID;
777 la->la_mode = oattr->la_mode;
780 /* Make sure a caller can chown. */
781 if (la->la_valid & LA_UID) {
782 if (la->la_uid == (uid_t) -1)
783 la->la_uid = oattr->la_uid;
784 if (((uc->uc_fsuid != oattr->la_uid) ||
785 (la->la_uid != oattr->la_uid)) &&
786 !md_capable(uc, CFS_CAP_CHOWN))
789 /* If the user or group of a non-directory has been
790 * changed by a non-root user, remove the setuid bit.
791 * 19981026 David C Niemi <niemi@tux.org>
793 * Changed this to apply to all users, including root,
794 * to avoid some races. This is the behavior we had in
795 * 2.0. The check for non-root was definitely wrong
796 * for 2.2 anyway, as it should have been using
797 * CAP_FSETID rather than fsuid -- 19990830 SD. */
798 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
799 !S_ISDIR(oattr->la_mode)) {
800 la->la_mode &= ~S_ISUID;
801 la->la_valid |= LA_MODE;
805 /* Make sure caller can chgrp. */
806 if (la->la_valid & LA_GID) {
807 if (la->la_gid == (gid_t) -1)
808 la->la_gid = oattr->la_gid;
809 if (((uc->uc_fsuid != oattr->la_uid) ||
810 ((la->la_gid != oattr->la_gid) &&
811 !lustre_in_group_p(uc, la->la_gid))) &&
812 !md_capable(uc, CFS_CAP_CHOWN))
815 /* Likewise, if the user or group of a non-directory
816 * has been changed by a non-root user, remove the
817 * setgid bit UNLESS there is no group execute bit
818 * (this would be a file marked for mandatory
819 * locking). 19981026 David C Niemi <niemi@tux.org>
821 * Removed the fsuid check (see the comment above) --
823 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
824 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
825 la->la_mode &= ~S_ISGID;
826 la->la_valid |= LA_MODE;
830 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
831 if (!((flags & MDS_OWNEROVERRIDE) &&
832 (uc->uc_fsuid == oattr->la_uid)) &&
833 !(flags & MDS_PERM_BYPASS)) {
834 rc = mdd_permission_internal(env, obj, oattr,
841 if (la->la_valid & LA_CTIME) {
842 /* The pure setattr, it has the priority over what is
843 * already set, do not drop it if ctime is equal. */
844 if (la->la_ctime < oattr->la_ctime)
845 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
851 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
852 struct mdd_device *mdd,
853 enum changelog_rec_type type,
854 enum changelog_rec_flags clf_flags,
855 const struct lu_fid *fid,
856 const char *xattr_name,
857 struct thandle *handle)
859 const struct lu_ucred *uc = lu_ucred(env);
860 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
861 struct llog_changelog_rec *rec;
866 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
869 if (uc->uc_jobid[0] != '\0')
870 clf_flags |= CLF_JOBID;
871 xflags |= CLFE_UIDGID;
874 if (type == CL_OPEN || type == CL_DN_OPEN)
876 if (type == CL_SETXATTR || type == CL_GETXATTR)
877 xflags |= CLFE_XATTR;
879 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
880 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
881 xflags & CLFE_SUPPORTED));
882 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
883 if (buf->lb_buf == NULL)
887 rec->cr_hdr.lrh_len = reclen;
888 rec->cr.cr_flags = clf_flags;
889 rec->cr.cr_type = (__u32)type;
890 rec->cr.cr_tfid = *fid;
891 rec->cr.cr_namelen = 0;
893 if (clf_flags & CLF_JOBID)
894 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
896 if (clf_flags & CLF_EXTRA_FLAGS) {
897 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
898 if (xflags & CLFE_UIDGID)
899 mdd_changelog_rec_extra_uidgid(&rec->cr,
900 uc->uc_uid, uc->uc_gid);
901 if (xflags & CLFE_NID)
902 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
903 if (xflags & CLFE_OPEN)
904 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
905 if (xflags & CLFE_XATTR) {
906 if (xattr_name == NULL)
908 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
912 rc = mdd_changelog_store(env, mdd, rec, handle);
917 /** Store a data change changelog record
918 * If this fails, we must fail the whole transaction; we don't
919 * want the change to commit without the log entry.
920 * \param mdd_obj - mdd_object of change
921 * \param handle - transaction handle
923 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
924 enum changelog_rec_type type,
925 enum changelog_rec_flags clf_flags,
926 struct mdd_object *mdd_obj, struct thandle *handle)
930 LASSERT(mdd_obj != NULL);
931 LASSERT(handle != NULL);
933 if (!mdd_changelog_enabled(env, mdd, type))
936 if (mdd_is_volatile_obj(mdd_obj))
939 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
940 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
941 /* Don't need multiple updates in this log */
942 /* Don't check under lock - no big deal if we get an extra
947 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
948 mdo2fid(mdd_obj), NULL, handle);
950 mdd_obj->mod_cltime = ktime_get();
955 int mdd_changelog_data_store_xattr(const struct lu_env *env,
956 struct mdd_device *mdd,
957 enum changelog_rec_type type,
958 enum changelog_rec_flags clf_flags,
959 struct mdd_object *mdd_obj,
960 const char *xattr_name,
961 struct thandle *handle)
965 LASSERT(mdd_obj != NULL);
966 LASSERT(handle != NULL);
968 if (!mdd_changelog_enabled(env, mdd, type))
971 if (mdd_is_volatile_obj(mdd_obj))
974 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
975 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
976 /* Don't need multiple updates in this log */
977 /* Don't check under lock - no big deal if we get an extra
983 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
984 mdo2fid(mdd_obj), xattr_name,
987 mdd_obj->mod_cltime = ktime_get();
992 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
993 * except for open flags have a dedicated record to store 32 bits of flags */
994 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
995 enum changelog_rec_flags clf_flags,
996 struct md_device *m, const struct lu_fid *fid)
998 struct thandle *handle;
999 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1003 LASSERT(fid != NULL);
1005 /* We'll check this again below, but we check now before we
1006 * start a transaction. */
1007 if (!mdd_changelog_enabled(env, mdd, type))
1010 handle = mdd_trans_create(env, mdd);
1012 RETURN(PTR_ERR(handle));
1014 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1018 rc = mdd_trans_start(env, mdd, handle);
1022 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1026 rc = mdd_trans_stop(env, mdd, rc, handle);
1032 * Save LMA extended attributes with data from \a ma.
1034 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1035 * not, LMA EA will be first read from disk, modified and write back.
1038 /* Precedence for choosing record type when multiple
1039 * attributes change: setattr > mtime > ctime > atime
1040 * (ctime changes when mtime does, plus chmod/chown.
1041 * atime and ctime are independent.) */
1042 static int mdd_attr_set_changelog(const struct lu_env *env,
1043 struct md_object *obj, struct thandle *handle,
1046 struct mdd_device *mdd = mdo2mdd(obj);
1049 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
1050 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
1051 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
1052 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
1053 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
1054 bits = bits & mdd->mdd_cl.mc_mask;
1055 /* This is an implementation limit rather than a protocol limit */
1056 CLASSERT(CL_LAST <= sizeof(int) * 8);
1060 /* The record type is the lowest non-masked set bit */
1063 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1064 return mdd_changelog_data_store(env, mdd, type, valid,
1065 md2mdd_obj(obj), handle);
1068 static int mdd_declare_attr_set(const struct lu_env *env,
1069 struct mdd_device *mdd,
1070 struct mdd_object *obj,
1071 const struct lu_attr *attr,
1072 struct thandle *handle)
1076 rc = mdo_declare_attr_set(env, obj, attr, handle);
1080 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1081 if (attr->la_valid & LA_MODE) {
1082 mdd_read_lock(env, obj, DT_TGT_CHILD);
1083 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1084 XATTR_NAME_ACL_ACCESS);
1085 mdd_read_unlock(env, obj);
1086 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1092 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1093 rc = mdo_declare_xattr_set(env, obj, buf,
1094 XATTR_NAME_ACL_ACCESS, 0,
1102 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1111 * permission changes may require sync operation, to mitigate performance
1112 * impact, only do this for dir and when permission is reduced.
1114 * For regular files, version is updated with permission change (see VBR), async
1115 * permission won't cause any issue, while missing permission change on
1116 * directory may affect accessibility of other objects after recovery.
1118 static inline bool permission_needs_sync(const struct lu_attr *old,
1119 const struct lu_attr *new)
1121 if (!S_ISDIR(old->la_mode))
1124 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1127 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1130 if (new->la_valid & LA_MODE) {
1131 /* turned on sticky bit */
1132 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1135 /* set-GID has no impact on what is allowed, not checked */
1137 /* turned off setuid bit, or one of rwx for someone */
1138 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1139 (old->la_mode & (0777 | S_ISUID)))
1146 static inline __u64 mdd_lmm_dom_size(void *buf)
1148 struct lov_mds_md *lmm = buf;
1149 struct lov_comp_md_v1 *comp_v1;
1150 struct lov_mds_md *v1;
1156 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1159 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1160 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1161 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1163 /* DoM entry is the first entry always */
1164 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1165 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1170 /* set attr and LOV EA at once, return updated attr */
1171 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1172 const struct md_attr *ma)
1174 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1175 struct mdd_device *mdd = mdo2mdd(obj);
1176 struct thandle *handle = NULL;
1177 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1178 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1179 const struct lu_attr *la = &ma->ma_attr;
1180 struct lu_ucred *uc;
1181 bool chrgrp_by_unprivileged_user = false;
1185 /* we do not use ->attr_set() for LOV/HSM EA any more */
1186 LASSERT((ma->ma_valid & MA_LOV) == 0);
1187 LASSERT((ma->ma_valid & MA_HSM) == 0);
1189 rc = mdd_la_get(env, mdd_obj, attr);
1193 *la_copy = ma->ma_attr;
1194 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
1198 /* no need to setattr anymore */
1199 if (la_copy->la_valid == 0) {
1200 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
1201 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
1202 PFID(mdo2fid(mdd_obj)), la->la_valid);
1207 /* If an unprivileged user changes group of some file,
1208 * the setattr operation will be processed synchronously to
1209 * honor the quota limit of the corresponding group. see LU-5152 */
1210 uc = lu_ucred_check(env);
1211 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1212 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1213 /* LU-10048: disable synchronous chgrp operation for it will
1214 * cause deadlock between MDT and OST.
1215 la_copy->la_valid |= LA_FLAGS;
1216 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1218 chrgrp_by_unprivileged_user = true;
1220 /* Flush the possible existing client setattr requests to OSTs
1221 * to keep the order with the current setattr operation that
1222 * will be sent directly to OSTs. see LU-5152 */
1223 /* LU-11303 disable sync as this is too heavyweight.
1224 * This should be replaced with a sync only for the object
1225 * being modified here, not the whole filesystem.
1226 rc = dt_sync(env, mdd->mdd_child);
1232 handle = mdd_trans_create(env, mdd);
1233 if (IS_ERR(handle)) {
1234 rc = PTR_ERR(handle);
1240 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1244 rc = mdd_trans_start(env, mdd, handle);
1248 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1249 permission_needs_sync(attr, la))
1250 handle->th_sync = 1;
1252 if (la->la_valid & (LA_MTIME | LA_CTIME))
1253 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1254 la->la_mtime, la->la_ctime);
1256 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1258 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1259 * otherwise following truncate will do nothing and truncated
1260 * data may be read again. This is a quick fix until LU-11033
1263 if (la_copy->la_valid & LA_SIZE) {
1264 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1266 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1270 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1271 la_copy->la_valid &= ~LA_SIZE;
1272 lu_buf_free(lov_buf);
1276 if (la_copy->la_valid) {
1277 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1279 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1280 /* rollback to the original gid */
1281 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1282 la_copy->la_gid = attr->la_gid;
1283 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1286 mdd_write_unlock(env, mdd_obj);
1290 rc = mdd_attr_set_changelog(env, obj, handle,
1294 rc = mdd_trans_stop(env, mdd, rc, handle);
1299 static int mdd_xattr_sanity_check(const struct lu_env *env,
1300 struct mdd_object *obj,
1301 const struct lu_attr *attr,
1304 struct lu_ucred *uc = lu_ucred_assert(env);
1307 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1310 if (strncmp(XATTR_USER_PREFIX, name,
1311 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1312 /* For sticky directories, only the owner and privileged user
1313 * can write attributes. */
1314 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1315 (uc->uc_fsuid != attr->la_uid) &&
1316 !md_capable(uc, CFS_CAP_FOWNER))
1318 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1319 (uc->uc_fsuid != attr->la_uid) &&
1320 !md_capable(uc, CFS_CAP_FOWNER)) {
1328 * Check if a string begins with a given prefix.
1330 * \param str String to check
1331 * \param prefix Substring to check at the beginning of \a str
1332 * \return true/false whether the condition is verified.
1334 static inline bool has_prefix(const char *str, const char *prefix)
1336 return strncmp(prefix, str, strlen(prefix)) == 0;
1340 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1342 * \param[in] xattr_name Full extended attribute name.
1344 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1346 static enum changelog_rec_type
1347 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1348 const char *xattr_name)
1350 /* Layout changes systematically recorded */
1351 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1352 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1353 strlen(XATTR_LUSTRE_LOV)) == 0)
1356 /* HSM information changes systematically recorded */
1357 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1360 /* Avoid logging SOM xattr for every file */
1361 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1364 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1365 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1366 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1367 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1368 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1374 static int mdd_declare_xattr_set(const struct lu_env *env,
1375 struct mdd_device *mdd,
1376 struct mdd_object *obj,
1377 const struct lu_buf *buf,
1379 int fl, struct thandle *handle)
1381 enum changelog_rec_type type;
1384 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1388 type = mdd_xattr_changelog_type(env, mdd, name);
1390 return 0; /* no changelog to store */
1392 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1396 * Compare current and future data of HSM EA and add a changelog if needed.
1398 * Caller should have write-locked \param obj.
1400 * \param buf - Future HSM EA content.
1401 * \retval 0 if no changelog is needed or changelog was added properly.
1402 * \retval -ve errno if there was a problem
1404 static int mdd_hsm_update_locked(const struct lu_env *env,
1405 struct md_object *obj,
1406 const struct lu_buf *buf,
1407 struct thandle *handle,
1408 enum changelog_rec_flags *clf_flags)
1410 struct mdd_thread_info *info = mdd_env_info(env);
1411 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1412 struct lu_buf *current_buf;
1413 struct md_hsm *current_mh;
1414 struct md_hsm *new_mh;
1418 OBD_ALLOC_PTR(current_mh);
1419 if (current_mh == NULL)
1422 /* Read HSM attrs from disk */
1423 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1424 MIN(mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1426 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1427 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1428 if (rc < 0 && rc != -ENODATA)
1430 else if (rc == -ENODATA)
1431 current_mh->mh_flags = 0;
1433 /* Map future HSM xattr */
1434 OBD_ALLOC_PTR(new_mh);
1436 GOTO(free, rc = -ENOMEM);
1437 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1441 /* Flags differ, set flags for the changelog that will be added */
1442 if (current_mh->mh_flags != new_mh->mh_flags) {
1443 hsm_set_cl_event(clf_flags, HE_STATE);
1444 if (new_mh->mh_flags & HS_DIRTY)
1445 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1448 OBD_FREE_PTR(new_mh);
1451 OBD_FREE_PTR(current_mh);
1455 static int mdd_object_pfid_replace(const struct lu_env *env,
1456 struct mdd_object *o)
1458 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1459 struct thandle *handle;
1462 handle = mdd_trans_create(env, mdd);
1464 RETURN(PTR_ERR(handle));
1466 handle->th_complex = 1;
1468 /* it doesn't need to track the PFID update via llog, because LFSCK
1469 * will repair it even it goes wrong */
1470 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1475 rc = mdd_trans_start(env, mdd, handle);
1479 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1484 mdd_trans_stop(env, mdd, rc, handle);
1489 static int mdd_declare_xattr_del(const struct lu_env *env,
1490 struct mdd_device *mdd,
1491 struct mdd_object *obj,
1493 struct thandle *handle);
1495 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1498 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1499 struct md_object *md_vic)
1501 struct mdd_device *mdd = mdo2mdd(md_obj);
1502 struct mdd_object *obj = md2mdd_obj(md_obj);
1503 struct mdd_object *vic = md2mdd_obj(md_vic);
1504 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1505 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1506 struct lov_mds_md *lmm;
1507 struct thandle *handle;
1511 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1512 if (rc == 0) /* same fid */
1515 handle = mdd_trans_create(env, mdd);
1517 RETURN(PTR_ERR(handle));
1520 mdd_write_lock(env, obj, DT_TGT_CHILD);
1521 mdd_write_lock(env, vic, DT_TGT_CHILD);
1523 mdd_write_lock(env, vic, DT_TGT_CHILD);
1524 mdd_write_lock(env, obj, DT_TGT_CHILD);
1527 /* get EA of victim file */
1528 memset(buf_vic, 0, sizeof(*buf_vic));
1529 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1536 /* parse the layout of victim file */
1537 lmm = buf_vic->lb_buf;
1538 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1539 GOTO(out, rc = -EINVAL);
1541 /* save EA of target file for restore */
1542 memset(buf, 0, sizeof(*buf));
1543 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1547 /* Get rid of the layout from victim object */
1548 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1552 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1553 LU_XATTR_MERGE, handle);
1557 rc = mdd_trans_start(env, mdd, handle);
1561 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1566 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1568 GOTO(out_restore, rc);
1570 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1571 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1576 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1577 LU_XATTR_REPLACE, handle);
1579 CERROR("%s: failed to rollback of layout of: "DFID
1580 ": %d, file state unknown\n",
1581 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1585 mdd_trans_stop(env, mdd, rc, handle);
1586 mdd_write_unlock(env, obj);
1587 mdd_write_unlock(env, vic);
1589 lu_buf_free(buf_vic);
1592 (void) mdd_object_pfid_replace(env, obj);
1598 * Extract the mirror with specified mirror id, and store the splitted
1599 * mirror layout to @buf.
1601 * \param[in] comp_v1 mirrored layout
1602 * \param[in] mirror_id the mirror with mirror_id to be extracted
1603 * \param[out] buf store the layout excluding the extracted mirror,
1604 * caller free the buffer we allocated in this function
1605 * \param[out] buf_vic store the extracted layout, caller free the buffer
1606 * we allocated in this function
1608 * \retval 0 on success; < 0 if error happens
1610 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1611 struct lu_buf *buf, struct lu_buf *buf_vic)
1613 struct lov_comp_md_v1 *comp_rem;
1614 struct lov_comp_md_v1 *comp_vic;
1615 struct lov_comp_md_entry_v1 *entry;
1616 struct lov_comp_md_entry_v1 *entry_rem;
1617 struct lov_comp_md_entry_v1 *entry_vic;
1619 __u16 comp_cnt, count = 0;
1620 int lmm_size, lmm_size_vic = 0;
1622 int offset, offset_rem, offset_vic;
1624 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1625 /* comp_v1 should contains more than 1 mirror */
1626 if (mirror_cnt <= 1)
1628 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1629 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1631 for (i = 0; i < comp_cnt; i++) {
1632 entry = &comp_v1->lcm_entries[i];
1633 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1635 lmm_size_vic += sizeof(*entry);
1636 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1637 } else if (count > 0) {
1638 /* find the specified mirror */
1646 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1650 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1651 if (!buf_vic->lb_buf) {
1656 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1657 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1659 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1660 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1661 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1662 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1663 if (!comp_rem->lcm_mirror_count)
1664 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1666 memset(comp_vic, 0, sizeof(*comp_v1));
1667 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1668 comp_vic->lcm_mirror_count = 0;
1669 comp_vic->lcm_entry_count = cpu_to_le32(count);
1670 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1671 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1672 comp_vic->lcm_layout_gen = 0;
1674 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1675 offset_rem = sizeof(*comp_rem) +
1676 sizeof(*entry_rem) * (comp_cnt - count);
1677 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1678 for (i = j = k = 0; i < comp_cnt; i++) {
1679 struct lov_mds_md *lmm, *lmm_dst;
1682 entry = &comp_v1->lcm_entries[i];
1683 entry_vic = &comp_vic->lcm_entries[j];
1684 entry_rem = &comp_rem->lcm_entries[k];
1686 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1689 /* copy component entry */
1691 memcpy(entry_vic, entry, sizeof(*entry));
1692 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1693 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1696 memcpy(entry_rem, entry, sizeof(*entry));
1697 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1701 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1703 lmm_dst = (struct lov_mds_md *)
1704 ((char *)comp_vic + offset_vic);
1706 lmm_dst = (struct lov_mds_md *)
1707 ((char *)comp_rem + offset_rem);
1709 /* copy component entry blob */
1710 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1712 /* blob offset advance */
1713 offset += le32_to_cpu(entry->lcme_size);
1715 offset_vic += le32_to_cpu(entry->lcme_size);
1717 offset_rem += le32_to_cpu(entry->lcme_size);
1723 static int mdd_dom_data_truncate(const struct lu_env *env,
1724 struct mdd_device *mdd, struct mdd_object *mo);
1726 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1727 struct md_rejig_data *mrd)
1729 struct mdd_device *mdd = mdo2mdd(md_obj);
1730 struct mdd_object *obj = md2mdd_obj(md_obj);
1731 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1732 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1733 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1734 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1735 struct lov_comp_md_v1 *lcm;
1736 struct thandle *handle;
1738 bool dom_stripe = false;
1742 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1743 if (rc == 0) /* same fid */
1746 handle = mdd_trans_create(env, mdd);
1748 RETURN(PTR_ERR(handle));
1751 mdd_write_lock(env, obj, DT_TGT_CHILD);
1752 mdd_write_lock(env, vic, DT_TGT_CHILD);
1754 mdd_write_lock(env, vic, DT_TGT_CHILD);
1755 mdd_write_lock(env, obj, DT_TGT_CHILD);
1758 /* get EA of mirrored file */
1759 memset(buf_save, 0, sizeof(*buf));
1760 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1764 lcm = buf_save->lb_buf;
1765 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1766 GOTO(out, rc = -EINVAL);
1769 * Extract the mirror with specified mirror id, and store the splitted
1770 * mirror layout to the victim file.
1772 memset(buf, 0, sizeof(*buf));
1773 memset(buf_vic, 0, sizeof(*buf_vic));
1774 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1778 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1780 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1781 LU_XATTR_SPLIT, handle);
1784 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1785 LU_XATTR_SPLIT, handle);
1789 rc = mdd_trans_start(env, mdd, handle);
1793 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1798 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1801 GOTO(out_restore, rc);
1803 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1807 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1814 /* restore obj's layout */
1815 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1816 LU_XATTR_REPLACE, handle);
1818 CERROR("%s: failed to rollback of layout of: "DFID
1819 ": %d, file state unkonwn.\n",
1820 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1823 rc = mdd_trans_stop(env, mdd, rc, handle);
1825 /* Truncate local DOM data if all went well */
1826 if (!rc && dom_stripe)
1827 mdd_dom_data_truncate(env, mdd, obj);
1829 mdd_write_unlock(env, obj);
1830 mdd_write_unlock(env, vic);
1831 lu_buf_free(buf_save);
1833 lu_buf_free(buf_vic);
1836 (void) mdd_object_pfid_replace(env, obj);
1841 static int mdd_layout_merge_allowed(const struct lu_env *env,
1842 struct md_object *target,
1843 struct md_object *victim)
1845 struct mdd_object *o1 = md2mdd_obj(target);
1847 /* cannot extend directory's LOVEA */
1848 if (S_ISDIR(mdd_object_type(o1))) {
1849 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1850 mdd_obj_dev_name(o1));
1858 * The caller should guarantee to update the object ctime
1859 * after xattr_set if needed.
1861 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1862 const struct lu_buf *buf, const char *name,
1865 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1866 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1867 struct mdd_device *mdd = mdo2mdd(obj);
1868 struct thandle *handle;
1869 enum changelog_rec_type cl_type;
1870 enum changelog_rec_flags clf_flags = 0;
1874 rc = mdd_la_get(env, mdd_obj, attr);
1878 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1882 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1883 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1884 struct md_rejig_data *mrd = buf->lb_buf;
1885 struct md_object *victim = mrd->mrd_obj;
1887 if (buf->lb_len != sizeof(*mrd))
1890 rc = mdd_layout_merge_allowed(env, obj, victim);
1894 if (fl == LU_XATTR_MERGE)
1895 /* merge layout of victim as a mirror of obj's. */
1896 rc = mdd_xattr_merge(env, obj, victim);
1898 rc = mdd_xattr_split(env, obj, mrd);
1902 if (strcmp(name, XATTR_NAME_LMV) == 0) {
1903 rc = mdd_dir_layout_shrink(env, obj, buf);
1907 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1908 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1909 struct posix_acl *acl;
1911 /* user may set empty ACL, which should be treated as removing
1913 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1916 RETURN(PTR_ERR(acl));
1918 rc = mdd_xattr_del(env, obj, name);
1921 posix_acl_release(acl);
1924 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1925 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1929 handle = mdd_trans_create(env, mdd);
1931 RETURN(PTR_ERR(handle));
1933 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1937 rc = mdd_trans_start(env, mdd, handle);
1941 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1943 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1944 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1946 mdd_write_unlock(env, mdd_obj);
1951 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1952 mdd_write_unlock(env, mdd_obj);
1956 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1960 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1961 mdd_obj, name, handle);
1965 return mdd_trans_stop(env, mdd, rc, handle);
1968 static int mdd_declare_xattr_del(const struct lu_env *env,
1969 struct mdd_device *mdd,
1970 struct mdd_object *obj,
1972 struct thandle *handle)
1974 enum changelog_rec_type type;
1977 rc = mdo_declare_xattr_del(env, obj, name, handle);
1981 type = mdd_xattr_changelog_type(env, mdd, name);
1983 return 0; /* no changelog to store */
1985 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1989 * The caller should guarantee to update the object ctime
1990 * after xattr_set if needed.
1992 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1995 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1996 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1997 struct mdd_device *mdd = mdo2mdd(obj);
1998 struct thandle *handle;
2002 rc = mdd_la_get(env, mdd_obj, attr);
2006 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2010 handle = mdd_trans_create(env, mdd);
2012 RETURN(PTR_ERR(handle));
2014 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2018 rc = mdd_trans_start(env, mdd, handle);
2022 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2023 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2024 mdd_write_unlock(env, mdd_obj);
2028 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2031 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2036 return mdd_trans_stop(env, mdd, rc, handle);
2040 * read lov/lmv EA of an object
2041 * return the lov/lmv EA in an allocated lu_buf
2043 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2044 struct lu_buf *lmm_buf, const char *name)
2046 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2051 if (buf->lb_buf == NULL) {
2052 buf = lu_buf_check_and_alloc(buf, 4096);
2053 if (buf->lb_buf == NULL)
2058 rc = mdo_xattr_get(env, obj, buf, name);
2059 if (rc == -ERANGE) {
2060 /* mti_big_buf is allocated but is too small
2061 * we need to increase it */
2062 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2064 if (buf->lb_buf == NULL)
2067 } else if (rc < 0) {
2069 } else if (rc == 0) {
2073 lu_buf_alloc(lmm_buf, rc);
2074 if (lmm_buf->lb_buf == NULL)
2078 * we don't use lmm_buf directly, because we don't know xattr size, so
2079 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2081 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2086 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2087 struct mdd_object *o, struct lu_buf *buf,
2088 struct thandle *handle)
2090 struct hsm_attrs *attrs;
2091 enum hsm_states hsm_flags;
2092 enum changelog_rec_flags clf_flags = 0;
2096 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2101 attrs = buf->lb_buf;
2102 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2103 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2106 /* Add a changelog record for release. */
2107 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2108 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2109 clf_flags, o, handle);
2114 * check if layout swapping between 2 objects is allowed
2116 * - only normal FIDs or non-system IGIFs
2117 * - same type of objects
2118 * - same owner/group (so quotas are still valid) unless this is from HSM
2121 static int mdd_layout_swap_allowed(const struct lu_env *env,
2122 struct mdd_object *o1,
2123 const struct lu_attr *attr1,
2124 struct mdd_object *o2,
2125 const struct lu_attr *attr2,
2128 const struct lu_fid *fid1, *fid2;
2134 if (!fid_is_norm(fid1) &&
2135 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2138 if (!fid_is_norm(fid2) &&
2139 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2142 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2143 if (S_ISDIR(mdd_object_type(o1)))
2145 if (S_ISREG(mdd_object_type(o1)))
2150 if (flags & SWAP_LAYOUTS_MDS_HSM)
2153 if ((attr1->la_uid != attr2->la_uid) ||
2154 (attr1->la_gid != attr2->la_gid))
2160 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2161 * look into the layout in MDD layer. */
2162 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2164 struct lov_comp_md_v1 *comp_v1;
2165 struct lov_mds_md *v1;
2169 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2170 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2171 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2177 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2178 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2181 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2182 off = le32_to_cpu(comp_v1->lcm_entries[i].
2184 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2189 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2190 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2201 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2203 return mdd_lmm_oi(lmm, oi, true);
2206 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2208 return mdd_lmm_oi(lmm, oi, false);
2211 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2213 struct lov_comp_md_v1 *comp_v1;
2215 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2216 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2218 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2220 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2221 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2222 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2223 __u16 tmp_gen = *gen;
2225 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2227 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2234 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2236 return mdd_lmm_gen(lmm, gen, true);
2239 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2241 return mdd_lmm_gen(lmm, gen, false);
2244 static int mdd_dom_data_truncate(const struct lu_env *env,
2245 struct mdd_device *mdd, struct mdd_object *mo)
2248 struct dt_object *dom;
2251 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2253 GOTO(out, rc = -ENODATA);
2255 th = dt_trans_create(env, mdd->mdd_bottom);
2257 GOTO(out, rc = PTR_ERR(th));
2259 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2263 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2267 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2269 dt_trans_stop(env, mdd->mdd_bottom, th);
2271 /* Ignore failure but report the error */
2273 CERROR("%s: "DFID" can't truncate DOM inode data, rc = %d\n",
2274 mdd_obj_dev_name(mo), PFID(mdo2fid(mo)), rc);
2279 * swap layouts between 2 lustre objects
2281 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2282 struct md_object *obj2, __u64 flags)
2284 struct mdd_thread_info *info = mdd_env_info(env);
2285 struct mdd_object *fst_o = md2mdd_obj(obj1);
2286 struct mdd_object *snd_o = md2mdd_obj(obj2);
2287 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2288 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2289 struct mdd_device *mdd = mdo2mdd(obj1);
2290 struct lov_mds_md *fst_lmm, *snd_lmm;
2291 struct lu_buf *fst_buf = &info->mti_buf[0];
2292 struct lu_buf *snd_buf = &info->mti_buf[1];
2293 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2294 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2295 struct ost_id *saved_oi = NULL;
2296 struct thandle *handle;
2297 struct mdd_object *dom_o = NULL;
2298 __u64 domsize_dom, domsize_vlt;
2299 __u32 fst_gen, snd_gen, saved_gen;
2305 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
2306 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2308 /* we have to sort the 2 obj, so locking will always
2309 * be in the same order, even in case of 2 concurrent swaps */
2310 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
2311 if (rc == 0) /* same fid ? */
2317 rc = mdd_la_get(env, fst_o, fst_la);
2321 rc = mdd_la_get(env, snd_o, snd_la);
2325 /* check if layout swapping is allowed */
2326 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2330 handle = mdd_trans_create(env, mdd);
2332 RETURN(PTR_ERR(handle));
2334 /* objects are already sorted */
2335 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2336 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2338 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2339 if (rc < 0 && rc != -ENODATA)
2342 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2343 if (rc < 0 && rc != -ENODATA)
2346 /* check if file has DoM. DoM file can be migrated only to another
2347 * DoM layout with the same DoM component size or to an non-DOM
2348 * layout. After migration to OSTs layout, local MDT inode data
2349 * should be truncated.
2350 * Objects are sorted by FIDs, considering that original file's FID
2351 * is always smaller the snd_o is always original file we are migrating
2354 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2355 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2357 /* Only migration is supported for DoM files, not 'swap_layouts' so
2358 * target file must be volatile and orphan.
2360 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2361 dom_o = domsize_dom ? snd_o : NULL;
2362 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2363 swap(domsize_dom, domsize_vlt);
2364 dom_o = domsize_dom ? fst_o : NULL;
2365 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2366 /* 'lfs swap_layouts' case, neither file should have DoM */
2368 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2369 "use migration instead: rc = %d\n", rc);
2373 if (domsize_vlt > 0 && domsize_dom == 0) {
2375 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": OST to DOM "
2376 "migration is not supported: rc = %d\n",
2377 PFID(mdo2fid(snd_o)), rc);
2379 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2381 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": new layout "
2382 "must have the same DoM component size: rc = %d\n",
2383 PFID(mdo2fid(fst_o)), rc);
2385 } else if (domsize_vlt > 0) {
2386 /* Migration with the same DOM component size, no need to
2387 * truncate local data, it is still being used */
2391 /* swapping 2 non existant layouts is a success */
2392 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2395 /* to help inode migration between MDT, it is better to
2396 * start by the no layout file (if one), so we order the swap */
2397 if (snd_buf->lb_buf == NULL) {
2399 swap(fst_buf, snd_buf);
2402 fst_gen = snd_gen = 0;
2403 /* lmm and generation layout initialization */
2404 if (fst_buf->lb_buf != NULL) {
2405 fst_lmm = fst_buf->lb_buf;
2406 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2407 fst_fl = LU_XATTR_REPLACE;
2411 fst_fl = LU_XATTR_CREATE;
2414 snd_lmm = snd_buf->lb_buf;
2415 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2417 saved_gen = fst_gen;
2418 /* increase the generation layout numbers */
2423 * XXX The layout generation is used to generate component IDs for
2424 * the composite file, we have to do some special tweaks to make
2425 * sure the layout generation is always adequate for that job.
2428 /* Skip invalid generation number for composite layout */
2429 if ((snd_gen & LCME_ID_MASK) == 0)
2431 if ((fst_gen & LCME_ID_MASK) == 0)
2433 /* Make sure the generation is greater than all the component IDs */
2434 if (fst_gen < snd_gen)
2436 else if (fst_gen > snd_gen)
2439 /* set the file specific informations in lmm */
2440 if (fst_lmm != NULL) {
2441 saved_oi = &info->mti_oa.o_oi;
2442 mdd_get_lmm_oi(fst_lmm, saved_oi);
2443 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2444 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2445 mdd_set_lmm_oi(snd_lmm, saved_oi);
2447 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2448 cpu_to_le32(LOV_MAGIC_MAGIC))
2449 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2451 GOTO(stop, rc = -EPROTO);
2453 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2455 /* Prepare HSM attribute if it's required */
2456 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2457 const int buflen = sizeof(struct hsm_attrs);
2459 lu_buf_alloc(fst_hsm_buf, buflen);
2460 lu_buf_alloc(snd_hsm_buf, buflen);
2461 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2462 GOTO(stop, rc = -ENOMEM);
2464 /* Read HSM attribute */
2465 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2469 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2473 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2474 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2479 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2480 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2486 /* prepare transaction */
2487 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2492 if (fst_buf->lb_buf != NULL)
2493 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2494 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2497 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2502 rc = mdd_trans_start(env, mdd, handle);
2506 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2507 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2511 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2513 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2516 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
2517 mdd_obj_dev_name(fst_o),
2518 PFID(mdo2fid(fst_o)), rc, rc2);
2523 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2527 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2530 if (fst_buf->lb_buf != NULL)
2531 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2532 LU_XATTR_REPLACE, handle);
2534 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2537 GOTO(out_restore, rc);
2539 /* Issue one changelog record per file */
2540 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2544 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2553 /* failure on second file, but first was done, so we have
2554 * to roll back first. */
2555 if (fst_buf->lb_buf != NULL) {
2556 mdd_set_lmm_oi(fst_lmm, saved_oi);
2557 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2558 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2559 LU_XATTR_REPLACE, handle);
2561 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2567 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2572 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2577 CERROR("%s: unable to roll back layout swap. FIDs: "
2578 DFID" and "DFID "error: %d/%d, steps: %d\n",
2579 mdd_obj_dev_name(fst_o),
2580 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
2582 /* a solution to avoid journal commit is to panic,
2583 * but it has strong consequences so we use LBUG to
2584 * allow sysdamin to choose to panic or not
2591 rc = mdd_trans_stop(env, mdd, rc, handle);
2593 /* Truncate local DOM data if all went well */
2595 mdd_dom_data_truncate(env, mdd, dom_o);
2597 mdd_write_unlock(env, snd_o);
2598 mdd_write_unlock(env, fst_o);
2600 lu_buf_free(fst_buf);
2601 lu_buf_free(snd_buf);
2602 lu_buf_free(fst_hsm_buf);
2603 lu_buf_free(snd_hsm_buf);
2606 (void) mdd_object_pfid_replace(env, fst_o);
2607 (void) mdd_object_pfid_replace(env, snd_o);
2612 static int mdd_declare_layout_change(const struct lu_env *env,
2613 struct mdd_device *mdd,
2614 struct mdd_object *obj,
2615 struct md_layout_change *mlc,
2616 struct thandle *handle)
2620 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2624 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2628 /* For PFL, this is used to instantiate necessary component objects. */
2630 mdd_layout_instantiate_component(const struct lu_env *env,
2631 struct mdd_object *obj, struct md_layout_change *mlc,
2632 struct thandle *handle)
2634 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2638 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2641 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2643 * It's possible that another layout write intent has already
2644 * instantiated our objects, so a -EALREADY returned, and we need to
2648 RETURN(rc == -EALREADY ? 0 : rc);
2650 rc = mdd_trans_start(env, mdd, handle);
2654 mdd_write_lock(env, obj, DT_TGT_CHILD);
2655 rc = mdo_layout_change(env, obj, mlc, handle);
2656 mdd_write_unlock(env, obj);
2660 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2665 * Change the FLR layout from RDONLY to WRITE_PENDING.
2667 * It picks the primary mirror, and bumps the layout version, and set
2668 * layout version xattr to OST objects in a sync tx. In order to facilitate
2669 * the handling of phantom writers from evicted clients, the clients carry
2670 * layout version of the file with write RPC, so that the OSTs can verify
2671 * if the write RPCs are legitimate, meaning not from evicted clients.
2674 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2675 struct md_layout_change *mlc, struct thandle *handle)
2677 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2678 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2679 struct lustre_som_attrs *som = &mlc->mlc_som;
2684 /* Verify acceptable operations */
2685 switch (mlc->mlc_opc) {
2686 case MD_LAYOUT_WRITE:
2687 case MD_LAYOUT_RESYNC:
2688 /* these are legal operations - this represents the case that
2689 * a few mirrors were missed in the last resync. */
2691 case MD_LAYOUT_RESYNC_DONE:
2696 som_buf->lb_buf = som;
2697 som_buf->lb_len = sizeof(*som);
2698 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2699 if (rc < 0 && rc != -ENODATA)
2703 lustre_som_swab(som);
2704 if (som->lsa_valid & SOM_FL_STRICT)
2705 fl = LU_XATTR_REPLACE;
2708 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2713 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2714 XATTR_NAME_SOM, fl, handle);
2719 /* record a changelog for data mover to consume */
2720 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2724 rc = mdd_trans_start(env, mdd, handle);
2728 /* it needs a sync tx to make FLR to work properly */
2729 handle->th_sync = 1;
2731 mdd_write_lock(env, obj, DT_TGT_CHILD);
2732 rc = mdo_layout_change(env, obj, mlc, handle);
2734 /* SOM state transition from STRICT to STALE */
2735 som->lsa_valid = SOM_FL_STALE;
2736 lustre_som_swab(som);
2737 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2740 mdd_write_unlock(env, obj);
2744 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2755 * Handle mirrored file state transition when it's in WRITE_PENDING.
2757 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2758 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2759 * layout version will be increased, and the file's state will be changed to
2763 mdd_layout_update_write_pending(const struct lu_env *env,
2764 struct mdd_object *obj, struct md_layout_change *mlc,
2765 struct thandle *handle)
2767 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2771 switch (mlc->mlc_opc) {
2772 case MD_LAYOUT_RESYNC:
2773 /* Upon receiving the resync request, it should
2774 * instantiate all stale components right away to get ready
2775 * for mirror copy. In order to avoid layout version change,
2776 * client should avoid sending LAYOUT_WRITE request at the
2779 case MD_LAYOUT_WRITE:
2780 /* legal race for concurrent write, the file state has been
2781 * changed by another client. */
2787 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2791 rc = mdd_trans_start(env, mdd, handle);
2795 /* it needs a sync tx to make FLR to work properly */
2796 handle->th_sync = 1;
2798 mdd_write_lock(env, obj, DT_TGT_CHILD);
2799 rc = mdo_layout_change(env, obj, mlc, handle);
2800 mdd_write_unlock(env, obj);
2811 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2813 * Only concurrent write and sync complete requests are possible when the
2814 * file is in SYNC_PENDING. For the latter request, it will pass in the
2815 * mirrors that have been synchronized, then the stale bit will be cleared
2816 * to make the file's state turn into RDONLY.
2817 * For concurrent write reqeust, it just needs to change the file's state
2818 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2819 * version because the version will be increased in the transition to
2820 * SYNC_PENDING later so that it can deny the write request from potential
2821 * evicted SYNC clients. */
2823 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2824 struct md_layout_change *mlc, struct thandle *handle)
2826 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2827 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2832 /* operation validation */
2833 switch (mlc->mlc_opc) {
2834 case MD_LAYOUT_RESYNC_DONE:
2835 /* resync complete. */
2836 case MD_LAYOUT_WRITE:
2837 /* concurrent write. */
2839 case MD_LAYOUT_RESYNC:
2840 /* resync again, most likely the previous run failed.
2841 * no-op if it's already in SYNC_PENDING state */
2847 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2848 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2849 if (rc < 0 && rc != -ENODATA)
2852 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2853 lustre_som_swab(&mlc->mlc_som);
2854 som_buf->lb_buf = &mlc->mlc_som;
2855 som_buf->lb_len = sizeof(mlc->mlc_som);
2858 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2862 /* record a changelog for the completion of resync */
2863 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2868 /* RESYNC_DONE has piggybacked size and blocks */
2870 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2871 XATTR_NAME_SOM, fl, handle);
2876 rc = mdd_trans_start(env, mdd, handle);
2880 /* it needs a sync tx to make FLR to work properly */
2881 handle->th_sync = 1;
2883 rc = mdo_layout_change(env, obj, mlc, handle);
2888 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2894 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2903 * Layout change callback for object.
2905 * This is only used by FLR for now. In the future, it can be exteneded to
2906 * handle all layout change.
2909 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2910 struct md_layout_change *mlc)
2912 struct mdd_object *obj = md2mdd_obj(o);
2913 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2914 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2915 struct lov_comp_md_v1 *lcm;
2916 struct thandle *handle;
2921 /* Verify acceptable operations */
2922 switch (mlc->mlc_opc) {
2923 case MD_LAYOUT_WRITE:
2924 case MD_LAYOUT_RESYNC:
2925 case MD_LAYOUT_RESYNC_DONE:
2931 handle = mdd_trans_create(env, mdd);
2933 RETURN(PTR_ERR(handle));
2935 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2942 /* analyze the layout to make sure it's a FLR file */
2944 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2945 GOTO(out, rc = -EINVAL);
2947 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2949 /* please refer to HLD of FLR for state transition */
2950 switch (flr_state) {
2952 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2954 case LCM_FL_WRITE_PENDING:
2955 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2958 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2960 case LCM_FL_SYNC_PENDING:
2961 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2970 mdd_trans_stop(env, mdd, rc, handle);
2975 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2976 struct mdd_object *child, const struct lu_attr *attr,
2977 const struct md_op_spec *spec,
2978 struct dt_allocation_hint *hint)
2980 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2981 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
2982 struct dt_object *nc = mdd_object_child(child);
2984 memset(hint, 0, sizeof(*hint));
2986 /* For striped directory, give striping EA to lod_ah_init, which will
2987 * decide the stripe_offset and stripe count by it. */
2988 if (S_ISDIR(attr->la_mode) &&
2989 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2990 hint->dah_eadata = spec->u.sp_ea.eadata;
2991 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2993 hint->dah_eadata = NULL;
2994 hint->dah_eadata_len = 0;
2995 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
2996 if (mdd->mdd_append_stripe_count != 0 ||
2997 mdd->mdd_append_pool[0])
2999 "using O_APPEND file striping\n");
3000 if (mdd->mdd_append_stripe_count)
3001 hint->dah_append_stripes =
3002 mdd->mdd_append_stripe_count;
3003 if (mdd->mdd_append_pool[0])
3004 hint->dah_append_pool = mdd->mdd_append_pool;
3006 hint->dah_append_stripes = 0;
3010 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3011 hint->dah_eadata, hint->dah_eadata_len);
3012 /* @hint will be initialized by underlying device. */
3013 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3016 static int accmode(const struct lu_env *env, const struct lu_attr *la,
3019 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3020 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3021 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3022 * owner can write to a file even if it is marked readonly to hide
3023 * its brokenness. (bug 5781) */
3024 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3025 struct lu_ucred *uc = lu_ucred_check(env);
3027 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3031 return mds_accmode(open_flags);
3034 static int mdd_open_sanity_check(const struct lu_env *env,
3035 struct mdd_object *obj,
3036 const struct lu_attr *attr, u64 open_flags,
3042 /* EEXIST check, also opening of *open* orphans is allowed so we can
3043 * open-by-handle unlinked files
3045 if (mdd_is_dead_obj(obj) && !is_replay &&
3046 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3049 if (S_ISLNK(attr->la_mode))
3052 mode = accmode(env, attr, open_flags);
3054 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3057 if (!(open_flags & MDS_OPEN_CREATED)) {
3058 rc = mdd_permission_internal(env, obj, attr, mode);
3063 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3064 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3065 open_flags &= ~MDS_OPEN_TRUNC;
3067 /* For writing append-only file must open it with append mode. */
3068 if (attr->la_flags & LUSTRE_APPEND_FL) {
3069 if ((open_flags & MDS_FMODE_WRITE) &&
3070 !(open_flags & MDS_OPEN_APPEND))
3072 if (open_flags & MDS_OPEN_TRUNC)
3079 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3080 u64 open_flags, struct md_op_spec *spec)
3082 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3083 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3084 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3085 struct mdd_object_user *mou = NULL;
3086 const struct lu_ucred *uc = lu_ucred(env);
3087 struct mdd_device *mdd = mdo2mdd(obj);
3088 enum changelog_rec_type type = CL_OPEN;
3092 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3094 rc = mdd_la_get(env, mdd_obj, attr);
3098 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3100 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3105 mdd_obj->mod_count++;
3107 if (!mdd_changelog_enabled(env, mdd, type))
3111 /* look for existing opener in list under mdd_write_lock */
3112 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3117 /* add user to list */
3118 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3124 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3126 mdd_obj_user_free(mou);
3131 if (type == CL_DN_OPEN) {
3132 if (ktime_before(ktime_get(), mou->mou_deniednext))
3133 /* same user denied again same access within
3134 * time interval: do not record
3138 /* this user already denied, but some time ago:
3139 * update denied time
3141 mou->mou_deniednext =
3142 ktime_add(ktime_get(),
3143 ktime_set(mdd->mdd_cl.mc_deniednext,
3146 mou->mou_opencount++;
3147 /* same user opening file again with same flags:
3154 /* FYI, only the bottom 32 bits of open_flags are recorded */
3155 mdd_changelog(env, type, open_flags, md_dev, mdo2fid(mdd_obj));
3159 mdd_write_unlock(env, mdd_obj);
3163 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3164 struct md_attr *ma, struct thandle *handle)
3168 rc = mdd_orphan_declare_delete(env, obj, handle);
3172 return mdo_declare_destroy(env, obj, handle);
3176 * No permission check is needed.
3178 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3179 struct md_attr *ma, u64 open_flags)
3181 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3182 struct mdd_device *mdd = mdo2mdd(obj);
3183 struct thandle *handle = NULL;
3186 bool blocked = false;
3187 bool last_close_by_uid = false;
3188 const struct lu_ucred *uc = lu_ucred(env);
3191 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3192 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3193 mdd_obj->mod_count--;
3194 mdd_write_unlock(env, mdd_obj);
3196 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3197 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3198 "list\n", PFID(mdd_object_fid(mdd_obj)));
3202 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3203 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3204 /* check without any lock */
3205 is_orphan = mdd_obj->mod_count == 1 &&
3206 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3210 /* mdd_trans_create() maybe failed because of barrier_entry(),
3211 * under such case, the orphan MDT-object will be left in the
3212 * orphan list, and when the MDT remount next time, the unused
3213 * orphans will be destroyed automatically.
3215 * One exception: the former mdd_finish_unlink may failed to
3216 * add the orphan MDT-object to the orphan list, then if the
3217 * mdd_trans_create() failed because of barrier_entry(), the
3218 * MDT-object will become real orphan that is neither in the
3219 * namespace nor in the orphan list. Such bad case should be
3220 * very rare and will be handled by e2fsck/lfsck. */
3221 handle = mdd_trans_create(env, mdo2mdd(obj));
3222 if (IS_ERR(handle)) {
3223 rc = PTR_ERR(handle);
3224 if (rc != -EINPROGRESS)
3232 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3236 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3241 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3247 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3248 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3250 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3251 lu_dev_name(mdd2lu_dev(mdd)),
3252 PFID(mdd_object_fid(mdd_obj)), rc);
3256 /* check again with lock */
3257 is_orphan = (mdd_obj->mod_count == 1) &&
3258 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3259 ma->ma_attr.la_nlink == 0);
3261 if (is_orphan && !handle && !blocked) {
3262 mdd_write_unlock(env, mdd_obj);
3266 mdd_obj->mod_count--; /*release open count */
3268 /* under mdd write lock */
3269 /* If recording, see if we need to remove UID from list. uc is not
3270 * initialized if the client has been evicted. */
3271 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3272 struct mdd_object_user *mou;
3274 /* look for UID in list */
3275 /* If mou is NULL, it probably means logging was enabled after
3276 * the user had the file open. So the corresponding close
3277 * will not be logged.
3279 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3282 mou->mou_opencount--;
3283 if (mou->mou_opencount == 0) {
3284 mdd_obj_user_remove(mdd_obj, mou);
3285 last_close_by_uid = true;
3290 if (!is_orphan || blocked)
3294 /* NB: Object maybe not in orphan list originally, it is rare case for
3295 * mdd_finish_unlink() failure, in that case, the object doesn't have
3296 * ORPHAN_OBJ flag */
3297 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3298 /* remove link to object from orphan index */
3299 LASSERT(handle != NULL);
3300 rc = mdd_orphan_delete(env, mdd_obj, handle);
3302 CERROR("%s: unable to delete "DFID" from orphan list: "
3303 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3304 PFID(mdd_object_fid(mdd_obj)), rc);
3305 /* If object was not deleted from orphan list, do not
3306 * destroy OSS objects, which will be done when next
3311 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3312 "list, OSS objects to be destroyed.\n",
3313 PFID(mdd_object_fid(mdd_obj)));
3316 rc = mdo_destroy(env, mdd_obj, handle);
3319 CERROR("%s: unable to delete "DFID" from orphan list: "
3320 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3321 PFID(mdd_object_fid(mdd_obj)), rc);
3326 mdd_write_unlock(env, mdd_obj);
3328 if (rc != 0 || blocked ||
3329 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3332 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3333 * or if CL_OPEN was recorded and it's last close by user.
3334 * Changelogs mask may change between open and close operations, but
3335 * this is not a big deal if we have a CL_CLOSE entry with no matching
3336 * CL_OPEN. Plus Changelogs mask may not change often.
3338 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3339 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3340 MDS_OPEN_TRUNC))) ||
3341 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3342 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3343 if (handle == NULL) {
3344 handle = mdd_trans_create(env, mdo2mdd(obj));
3346 GOTO(stop, rc = PTR_ERR(handle));
3348 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3349 NULL, NULL, handle);
3353 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3358 /* FYI, only the bottom 32 bits of open_flags are recorded */
3359 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3364 if (handle != NULL && !IS_ERR(handle))
3365 rc = mdd_trans_stop(env, mdd, rc, handle);
3371 * Permission check is done when open,
3372 * no need check again.
3374 static int mdd_readpage_sanity_check(const struct lu_env *env,
3375 struct mdd_object *obj)
3377 struct dt_object *next = mdd_object_child(obj);
3381 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3389 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3390 size_t nob, const struct dt_it_ops *iops,
3391 struct dt_it *it, __u32 attr, void *arg)
3393 struct lu_dirpage *dp = &lp->lp_dir;
3397 struct lu_dirent *ent;
3398 struct lu_dirent *last = NULL;
3402 if (nob < sizeof(*dp))
3405 memset(area, 0, sizeof (*dp));
3406 area += sizeof (*dp);
3407 nob -= sizeof (*dp);
3414 len = iops->key_size(env, it);
3416 /* IAM iterator can return record with zero len. */
3420 hash = iops->store(env, it);
3421 if (unlikely(first)) {
3423 dp->ldp_hash_start = cpu_to_le64(hash);
3426 /* calculate max space required for lu_dirent */
3427 recsize = lu_dirent_calc_size(len, attr);
3429 if (nob >= recsize) {
3430 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3431 if (result == -ESTALE)
3436 /* osd might not able to pack all attributes,
3437 * so recheck rec length */
3438 recsize = le16_to_cpu(ent->lde_reclen);
3440 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3441 fid_le_to_cpu(&fid, &ent->lde_fid);
3442 if (fid_is_dot_lustre(&fid))
3446 result = (last != NULL) ? 0 :-EINVAL;
3450 ent = (void *)ent + recsize;
3454 result = iops->next(env, it);
3455 if (result == -ESTALE)
3457 } while (result == 0);
3460 dp->ldp_hash_end = cpu_to_le64(hash);
3462 if (last->lde_hash == dp->ldp_hash_end)
3463 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3464 last->lde_reclen = 0; /* end mark */
3467 /* end of directory */
3468 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3469 else if (result < 0)
3470 CWARN("build page failed: %d!\n", result);
3474 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3475 const struct lu_rdpg *rdpg)
3477 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3481 if (mdd_object_exists(mdd_obj) == 0) {
3482 CERROR("%s: object "DFID" not found: rc = -2\n",
3483 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3487 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3488 rc = mdd_readpage_sanity_check(env, mdd_obj);
3490 GOTO(out_unlock, rc);
3492 if (mdd_is_dead_obj(mdd_obj)) {
3494 struct lu_dirpage *dp;
3497 * According to POSIX, please do not return any entry to client:
3498 * even dot and dotdot should not be returned.
3500 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3501 PFID(mdd_object_fid(mdd_obj)));
3503 if (rdpg->rp_count <= 0)
3504 GOTO(out_unlock, rc = -EFAULT);
3505 LASSERT(rdpg->rp_pages != NULL);
3507 pg = rdpg->rp_pages[0];
3508 dp = (struct lu_dirpage *)kmap(pg);
3509 memset(dp, 0 , sizeof(struct lu_dirpage));
3510 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3511 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3512 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3514 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3517 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3518 mdd_dir_page_build, NULL);
3520 struct lu_dirpage *dp;
3522 dp = kmap(rdpg->rp_pages[0]);
3523 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3526 * No pages were processed, mark this for first page
3529 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3530 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3531 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3533 kunmap(rdpg->rp_pages[0]);
3536 GOTO(out_unlock, rc);
3538 mdd_read_unlock(env, mdd_obj);
3542 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3544 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3546 if (mdd_object_exists(mdd_obj) == 0) {
3549 CERROR("%s: object "DFID" not found: rc = %d\n",
3550 mdd_obj_dev_name(mdd_obj),
3551 PFID(mdd_object_fid(mdd_obj)), rc);
3554 return dt_object_sync(env, mdd_object_child(mdd_obj),
3558 static int mdd_object_lock(const struct lu_env *env,
3559 struct md_object *obj,
3560 struct lustre_handle *lh,
3561 struct ldlm_enqueue_info *einfo,
3562 union ldlm_policy_data *policy)
3564 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3565 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3569 static int mdd_object_unlock(const struct lu_env *env,
3570 struct md_object *obj,
3571 struct ldlm_enqueue_info *einfo,
3572 union ldlm_policy_data *policy)
3574 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3575 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3578 const struct md_object_operations mdd_obj_ops = {
3579 .moo_permission = mdd_permission,
3580 .moo_attr_get = mdd_attr_get,
3581 .moo_attr_set = mdd_attr_set,
3582 .moo_xattr_get = mdd_xattr_get,
3583 .moo_xattr_set = mdd_xattr_set,
3584 .moo_xattr_list = mdd_xattr_list,
3585 .moo_invalidate = mdd_invalidate,
3586 .moo_xattr_del = mdd_xattr_del,
3587 .moo_swap_layouts = mdd_swap_layouts,
3588 .moo_open = mdd_open,
3589 .moo_close = mdd_close,
3590 .moo_readpage = mdd_readpage,
3591 .moo_readlink = mdd_readlink,
3592 .moo_changelog = mdd_changelog,
3593 .moo_object_sync = mdd_object_sync,
3594 .moo_object_lock = mdd_object_lock,
3595 .moo_object_unlock = mdd_object_unlock,
3596 .moo_layout_change = mdd_layout_change,