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 struct lu_fid *pfid,
76 const char *xattr_name,
77 struct thandle *handle);
79 static inline bool has_prefix(const char *str, const char *prefix);
82 static u32 flags_helper(u64 open_flags)
86 if (open_flags & MDS_FMODE_EXEC) {
87 open_mode = MDS_FMODE_EXEC;
89 if (open_flags & MDS_FMODE_READ)
90 open_mode = MDS_FMODE_READ;
92 (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
93 open_mode |= MDS_FMODE_WRITE;
99 /** Allocate/init a user and its sub-structures.
104 * \retval mou [OUT] success valid structure
107 static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags,
108 uid_t uid, gid_t gid)
110 struct mdd_object_user *mou;
114 OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem);
116 RETURN(ERR_PTR(-ENOMEM));
118 mou->mou_open_flags = open_flags;
119 mou->mou_uidgid = ((__u64)uid << 32) | gid;
120 mou->mou_opencount = 0;
121 mou->mou_deniednext = ktime_set(0, 0);
127 * Free a user and its sub-structures.
129 * \param mou [IN] user to be freed.
131 static void mdd_obj_user_free(struct mdd_object_user *mou)
133 OBD_SLAB_FREE_PTR(mou, mdd_object_kmem);
137 * Find if UID/GID already has this file open
139 * Caller should have write-locked \param mdd_obj.
140 * \param mdd_obj [IN] mdd_obj
141 * \param uid [IN] client uid
142 * \param gid [IN] client gid
143 * \retval user pointer or NULL if not found
146 struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj,
147 uid_t uid, gid_t gid,
150 struct mdd_object_user *mou;
155 uidgid = ((__u64)uid << 32) | gid;
156 list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) {
157 if (mou->mou_uidgid == uidgid &&
158 flags_helper(mou->mou_open_flags) ==
159 flags_helper(open_flags))
166 * Add a user to the list of openers for this file
168 * Caller should have write-locked \param mdd_obj.
169 * \param mdd_obj [IN] mdd_obj
170 * \param mou [IN] user
172 * \retval -ve failure
174 static int mdd_obj_user_add(struct mdd_object *mdd_obj,
175 struct mdd_object_user *mou,
178 struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj);
179 struct mdd_object_user *tmp;
180 __u32 uid = mou->mou_uidgid >> 32;
181 __u32 gid = mou->mou_uidgid & ((1UL << 32) - 1);
184 tmp = mdd_obj_user_find(mdd_obj, uid, gid, mou->mou_open_flags);
188 list_add_tail(&mou->mou_list, &mdd_obj->mod_users);
191 /* next 'access denied' notification cannot happen before
194 mou->mou_deniednext =
195 ktime_add(ktime_get(),
196 ktime_set(mdd->mdd_cl.mc_deniednext, 0));
198 mou->mou_opencount++;
203 * Remove UID from the list
205 * Caller should have write-locked \param mdd_obj.
206 * \param mdd_obj [IN] mdd_obj
207 * \param uid [IN] user
208 * \retval -ve failure
210 static int mdd_obj_user_remove(struct mdd_object *mdd_obj,
211 struct mdd_object_user *mou)
218 list_del_init(&mou->mou_list);
220 mdd_obj_user_free(mou);
224 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
229 if (mdd_object_exists(obj) == 0)
232 rc = mdo_attr_get(env, obj, la);
233 if (unlikely(rc != 0)) {
235 obj->mod_flags |= DEAD_OBJ;
239 if (la->la_valid & LA_FLAGS && la->la_flags & LUSTRE_ORPHAN_FL)
240 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
245 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
247 return lu_env_info(env, &mdd_thread_key);
250 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
254 buf = &mdd_env_info(env)->mti_buf[0];
260 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
261 const void *area, ssize_t len)
265 buf = &mdd_env_info(env)->mti_buf[0];
266 buf->lb_buf = (void *)area;
271 struct lu_object *mdd_object_alloc(const struct lu_env *env,
272 const struct lu_object_header *hdr,
275 struct mdd_object *mdd_obj;
278 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
282 o = mdd2lu_obj(mdd_obj);
283 lu_object_init(o, NULL, d);
284 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
285 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
286 mdd_obj->mod_count = 0;
287 o->lo_ops = &mdd_lu_obj_ops;
288 INIT_LIST_HEAD(&mdd_obj->mod_users);
293 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
294 const struct lu_object_conf *unused)
296 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
297 struct mdd_object *mdd_obj = lu2mdd_obj(o);
298 struct lu_object *below;
299 struct lu_device *under;
302 mdd_obj->mod_cltime = ktime_set(0, 0);
303 under = &d->mdd_child->dd_lu_dev;
304 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
306 RETURN(PTR_ERR(below));
308 lu_object_add(o, below);
313 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
317 if (lu_object_exists(o)) {
318 struct mdd_object *mdd_obj = lu2mdd_obj(o);
319 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
321 rc = mdd_la_get(env, mdd_obj, attr);
327 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
329 struct mdd_object *mdd = lu2mdd_obj(o);
330 struct mdd_object_user *mou, *tmp2;
333 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
334 list_del(&mou->mou_list);
335 mdd_obj_user_free(mou);
339 /* mdd doesn't contain an lu_object_header, so don't need call_rcu */
340 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
343 static int mdd_object_print(const struct lu_env *env, void *cookie,
344 lu_printer_t p, const struct lu_object *o)
346 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
348 return (*p)(env, cookie,
349 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
350 mdd, mdd->mod_count, mdd->mod_valid,
351 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
354 static const struct lu_object_operations mdd_lu_obj_ops = {
355 .loo_object_init = mdd_object_init,
356 .loo_object_start = mdd_object_start,
357 .loo_object_free = mdd_object_free,
358 .loo_object_print = mdd_object_print,
361 struct mdd_object *mdd_object_find(const struct lu_env *env,
362 struct mdd_device *d,
363 const struct lu_fid *f)
365 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
369 * No permission check is needed.
371 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
374 struct mdd_object *mdd_obj = md2mdd_obj(obj);
379 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
380 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
381 ma->ma_attr.la_nlink = 0;
387 * No permission check is needed.
389 static int mdd_xattr_get(const struct lu_env *env,
390 struct md_object *obj, struct lu_buf *buf,
393 struct mdd_object *mdd_obj = md2mdd_obj(obj);
394 struct mdd_device *mdd;
399 if (mdd_object_exists(mdd_obj) == 0) {
400 CERROR("%s: object "DFID" not found: rc = -2\n",
401 mdd_obj_dev_name(mdd_obj),
402 PFID(mdd_object_fid(mdd_obj)));
406 /* If the object has been destroyed, then do not get LMVEA, because
407 * it needs to load stripes from the iteration of the master object,
408 * and it will cause problem if master object has been destroyed, see
410 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
411 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
412 strcmp(name, XATTR_NAME_LMV) == 0))
415 /* If the object has been delete from the namespace, then
416 * get linkEA should return -ENOENT as well */
417 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
418 strcmp(name, XATTR_NAME_LINK) == 0))
421 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
422 rc = mdo_xattr_get(env, mdd_obj, buf, name);
423 mdd_read_unlock(env, mdd_obj);
427 /* record only getting user xattrs and acls */
428 if (rc >= 0 && buf->lb_buf &&
429 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
430 (has_prefix(name, XATTR_USER_PREFIX) ||
431 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
432 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
433 struct thandle *handle;
436 LASSERT(mdd_object_fid(mdd_obj) != NULL);
438 handle = mdd_trans_create(env, mdd);
440 RETURN(PTR_ERR(handle));
442 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
447 rc2 = mdd_trans_start(env, mdd, handle);
451 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
452 mdd_object_fid(mdd_obj),
456 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
465 * Permission check is done when open,
466 * no need check again.
468 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
471 struct mdd_object *mdd_obj = md2mdd_obj(obj);
472 struct dt_object *next;
477 if (mdd_object_exists(mdd_obj) == 0) {
478 CERROR("%s: object "DFID" not found: rc = -2\n",
479 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
483 next = mdd_object_child(mdd_obj);
484 LASSERT(next != NULL);
485 LASSERT(next->do_body_ops != NULL);
486 LASSERT(next->do_body_ops->dbo_read != NULL);
487 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
488 rc = dt_read(env, next, buf, &pos);
489 mdd_read_unlock(env, mdd_obj);
494 * No permission check is needed.
496 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
499 struct mdd_object *mdd_obj = md2mdd_obj(obj);
504 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
505 rc = mdo_xattr_list(env, mdd_obj, buf);
506 mdd_read_unlock(env, mdd_obj);
508 /* If the buffer is NULL then we are only here to get the
509 * length of the xattr name list. */
510 if (rc < 0 || buf->lb_buf == NULL)
514 * Filter out XATTR_NAME_LINK if this is an orphan object. See
517 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
518 char *end = (char *)buf->lb_buf + rc;
519 char *p = buf->lb_buf;
522 char *next = p + strlen(p) + 1;
524 if (strcmp(p, XATTR_NAME_LINK) == 0) {
526 memmove(p, next, end - next);
528 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
529 " of orphan "DFID"\n",
530 PFID(mdd_object_fid(mdd_obj)));
541 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
543 return mdo_invalidate(env, md2mdd_obj(obj));
546 int mdd_declare_create_object_internal(const struct lu_env *env,
547 struct mdd_object *p,
548 struct mdd_object *c,
549 struct lu_attr *attr,
550 struct thandle *handle,
551 const struct md_op_spec *spec,
552 struct dt_allocation_hint *hint)
554 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
555 const struct dt_index_features *feat = spec->sp_feat;
559 if (feat != &dt_directory_features && feat != NULL) {
560 dof->dof_type = DFT_INDEX;
561 dof->u.dof_idx.di_feat = feat;
563 dof->dof_type = dt_mode_to_dft(attr->la_mode);
564 if (dof->dof_type == DFT_REGULAR) {
565 dof->u.dof_reg.striped =
566 md_should_create(spec->sp_cr_flags);
567 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
568 dof->u.dof_reg.striped = 0;
569 /* is this replay? */
571 dof->u.dof_reg.striped = 0;
575 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
580 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
581 struct mdd_object *c, struct lu_attr *attr,
582 struct thandle *handle,
583 const struct md_op_spec *spec,
584 struct dt_allocation_hint *hint)
586 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
590 LASSERT(!mdd_object_exists(c));
592 rc = mdo_create_object(env, c, attr, hint, dof, handle);
597 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
598 const struct lu_attr *attr, struct thandle *handle,
604 rc = mdo_attr_set(env, obj, attr, handle);
605 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
606 if (!rc && (attr->la_valid & LA_MODE) && needacl)
607 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
612 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
613 const struct lu_attr *oattr, struct lu_attr *attr,
614 struct thandle *handle)
619 LASSERT(attr->la_valid & LA_CTIME);
620 LASSERT(oattr != NULL);
622 /* Make sure the ctime is increased only, however, it's not strictly
623 * reliable at here because there is not guarantee to hold lock on
624 * object, so we just bypass some unnecessary cmtime setting first
625 * and OSD has to check it again. */
626 if (attr->la_ctime < oattr->la_ctime)
627 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
628 else if (attr->la_valid == LA_CTIME &&
629 attr->la_ctime == oattr->la_ctime)
630 attr->la_valid &= ~LA_CTIME;
632 if (attr->la_valid != 0)
633 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
638 static bool is_project_state_change(const struct lu_attr *oattr,
641 if (la->la_valid & LA_PROJID &&
642 oattr->la_projid != la->la_projid)
645 if ((la->la_valid & LA_FLAGS) &&
646 (la->la_flags & LUSTRE_PROJINHERIT_FL) !=
647 (oattr->la_flags & LUSTRE_PROJINHERIT_FL))
654 * This gives the same functionality as the code between
655 * sys_chmod and inode_setattr
656 * chown_common and inode_setattr
657 * utimes and inode_setattr
658 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
660 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
661 const struct lu_attr *oattr, struct lu_attr *la,
662 const struct md_attr *ma)
666 const unsigned long flags = ma->ma_attr_flags;
673 /* Do not permit change file type */
674 if (la->la_valid & LA_TYPE)
677 /* They should not be processed by setattr */
678 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
681 LASSERT(oattr != NULL);
683 uc = lu_ucred_check(env);
687 if (is_project_state_change(oattr, la)) {
688 if (!md_capable(uc, CFS_CAP_SYS_RESOURCE) &&
689 !lustre_in_group_p(uc, ma->ma_enable_chprojid_gid) &&
690 !(ma->ma_enable_chprojid_gid == -1 &&
691 mdd_permission_internal(env, obj, oattr, MAY_WRITE)))
695 if (la->la_valid == LA_CTIME) {
696 if (!(flags & MDS_PERM_BYPASS))
697 /* This is only for set ctime when rename's source is
699 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
701 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
702 la->la_valid &= ~LA_CTIME;
706 if (flags & MDS_CLOSE_UPDATE_TIMES &&
707 la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
708 /* This is an atime/mtime/ctime attribute update for
711 if (la->la_valid & LA_ATIME &&
712 la->la_atime <= (oattr->la_atime +
713 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
714 la->la_valid &= ~LA_ATIME;
715 if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime)
716 la->la_valid &= ~LA_CTIME;
717 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
718 la->la_valid &= ~LA_MTIME;
722 /* Check if flags change. */
723 if (la->la_valid & LA_FLAGS) {
724 unsigned int oldflags = oattr->la_flags &
725 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
726 unsigned int newflags = la->la_flags &
727 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
729 if ((uc->uc_fsuid != oattr->la_uid) &&
730 !md_capable(uc, CFS_CAP_FOWNER))
733 /* The IMMUTABLE and APPEND_ONLY flags can
734 * only be changed by the relevant capability. */
735 if ((oldflags ^ newflags) &&
736 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
739 if (!S_ISDIR(oattr->la_mode)) {
740 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
741 } else if (la->la_flags & LUSTRE_ENCRYPT_FL) {
742 /* when trying to add encryption flag on dir,
743 * make sure it is empty
745 rc = mdd_dir_is_empty(env, obj);
752 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
753 (la->la_valid & ~LA_FLAGS) &&
754 !(flags & MDS_PERM_BYPASS))
757 /* Check for setting the obj time. */
758 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
759 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
760 if ((uc->uc_fsuid != oattr->la_uid) &&
761 !md_capable(uc, CFS_CAP_FOWNER)) {
762 rc = mdd_permission_internal(env, obj, oattr,
769 if (la->la_valid & LA_KILL_SUID) {
770 la->la_valid &= ~LA_KILL_SUID;
771 if ((oattr->la_mode & S_ISUID) &&
772 !(la->la_valid & LA_MODE)) {
773 la->la_mode = oattr->la_mode;
774 la->la_valid |= LA_MODE;
776 la->la_mode &= ~S_ISUID;
779 if (la->la_valid & LA_KILL_SGID) {
780 la->la_valid &= ~LA_KILL_SGID;
781 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
782 (S_ISGID | S_IXGRP)) &&
783 !(la->la_valid & LA_MODE)) {
784 la->la_mode = oattr->la_mode;
785 la->la_valid |= LA_MODE;
787 la->la_mode &= ~S_ISGID;
790 /* Make sure a caller can chmod. */
791 if (la->la_valid & LA_MODE) {
792 if (!(flags & MDS_PERM_BYPASS) &&
793 (uc->uc_fsuid != oattr->la_uid) &&
794 !md_capable(uc, CFS_CAP_FOWNER))
797 if (la->la_mode == (umode_t) -1)
798 la->la_mode = oattr->la_mode;
800 la->la_mode = (la->la_mode & S_IALLUGO) |
801 (oattr->la_mode & ~S_IALLUGO);
803 /* Also check the setgid bit! */
804 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
805 la->la_gid : oattr->la_gid) &&
806 !md_capable(uc, CFS_CAP_FSETID))
807 la->la_mode &= ~S_ISGID;
809 la->la_mode = oattr->la_mode;
812 /* Make sure a caller can chown. */
813 if (la->la_valid & LA_UID) {
814 if (la->la_uid == (uid_t) -1)
815 la->la_uid = oattr->la_uid;
816 if (((uc->uc_fsuid != oattr->la_uid) ||
817 (la->la_uid != oattr->la_uid)) &&
818 !md_capable(uc, CFS_CAP_CHOWN))
821 /* If the user or group of a non-directory has been
822 * changed by a non-root user, remove the setuid bit.
823 * 19981026 David C Niemi <niemi@tux.org>
825 * Changed this to apply to all users, including root,
826 * to avoid some races. This is the behavior we had in
827 * 2.0. The check for non-root was definitely wrong
828 * for 2.2 anyway, as it should have been using
829 * CAP_FSETID rather than fsuid -- 19990830 SD. */
830 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
831 !S_ISDIR(oattr->la_mode)) {
832 la->la_mode &= ~S_ISUID;
833 la->la_valid |= LA_MODE;
837 /* Make sure caller can chgrp. */
838 if (la->la_valid & LA_GID) {
839 if (la->la_gid == (gid_t) -1)
840 la->la_gid = oattr->la_gid;
841 if (((uc->uc_fsuid != oattr->la_uid) ||
842 ((la->la_gid != oattr->la_gid) &&
843 !lustre_in_group_p(uc, la->la_gid))) &&
844 !md_capable(uc, CFS_CAP_CHOWN))
847 /* Likewise, if the user or group of a non-directory
848 * has been changed by a non-root user, remove the
849 * setgid bit UNLESS there is no group execute bit
850 * (this would be a file marked for mandatory
851 * locking). 19981026 David C Niemi <niemi@tux.org>
853 * Removed the fsuid check (see the comment above) --
855 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
856 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
857 la->la_mode &= ~S_ISGID;
858 la->la_valid |= LA_MODE;
862 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
863 if (!((flags & MDS_OWNEROVERRIDE) &&
864 (uc->uc_fsuid == oattr->la_uid)) &&
865 !(flags & MDS_PERM_BYPASS)) {
866 rc = mdd_permission_internal(env, obj, oattr,
873 if (la->la_valid & LA_CTIME) {
874 /* The pure setattr, it has the priority over what is
875 * already set, do not drop it if ctime is equal. */
876 if (la->la_ctime < oattr->la_ctime)
877 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
883 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
884 struct mdd_device *mdd,
885 enum changelog_rec_type type,
886 enum changelog_rec_flags clf_flags,
887 const struct lu_fid *fid,
888 const struct lu_fid *pfid,
889 const char *xattr_name,
890 struct thandle *handle)
892 const struct lu_ucred *uc = lu_ucred(env);
893 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
894 struct llog_changelog_rec *rec;
899 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
902 if (uc->uc_jobid[0] != '\0')
903 clf_flags |= CLF_JOBID;
904 xflags |= CLFE_UIDGID;
907 if (type == CL_OPEN || type == CL_DN_OPEN)
909 if (type == CL_SETXATTR || type == CL_GETXATTR)
910 xflags |= CLFE_XATTR;
912 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
913 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
914 xflags & CLFE_SUPPORTED));
915 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
916 if (buf->lb_buf == NULL)
920 rec->cr_hdr.lrh_len = reclen;
921 rec->cr.cr_flags = clf_flags;
922 rec->cr.cr_type = (__u32)type;
923 rec->cr.cr_tfid = *fid;
925 rec->cr.cr_pfid = *pfid;
926 rec->cr.cr_namelen = 0;
928 if (clf_flags & CLF_JOBID)
929 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
931 if (clf_flags & CLF_EXTRA_FLAGS) {
932 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
933 if (xflags & CLFE_UIDGID)
934 mdd_changelog_rec_extra_uidgid(&rec->cr,
935 uc->uc_uid, uc->uc_gid);
936 if (xflags & CLFE_NID)
937 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
938 if (xflags & CLFE_OPEN)
939 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
940 if (xflags & CLFE_XATTR) {
941 if (xattr_name == NULL)
943 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
947 rc = mdd_changelog_store(env, mdd, rec, handle);
952 /** Store a data change changelog record
953 * If this fails, we must fail the whole transaction; we don't
954 * want the change to commit without the log entry.
955 * \param mdd_obj - mdd_object of change
956 * \param handle - transaction handle
957 * \param pfid - parent FID for CL_MTIME changelogs
959 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
960 enum changelog_rec_type type,
961 enum changelog_rec_flags clf_flags,
962 struct mdd_object *mdd_obj, struct thandle *handle,
963 const struct lu_fid *pfid)
967 LASSERT(mdd_obj != NULL);
968 LASSERT(handle != NULL);
970 if (!mdd_changelog_enabled(env, mdd, type))
973 if (mdd_is_volatile_obj(mdd_obj))
976 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
977 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
978 /* Don't need multiple updates in this log */
979 /* Don't check under lock - no big deal if we get an extra
984 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
985 mdd_object_fid(mdd_obj), pfid,
988 mdd_obj->mod_cltime = ktime_get();
993 int mdd_changelog_data_store_xattr(const struct lu_env *env,
994 struct mdd_device *mdd,
995 enum changelog_rec_type type,
996 enum changelog_rec_flags clf_flags,
997 struct mdd_object *mdd_obj,
998 const char *xattr_name,
999 struct thandle *handle)
1003 LASSERT(mdd_obj != NULL);
1004 LASSERT(handle != NULL);
1006 if (!mdd_changelog_enabled(env, mdd, type))
1009 if (mdd_is_volatile_obj(mdd_obj))
1012 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1013 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1014 /* Don't need multiple updates in this log */
1015 /* Don't check under lock - no big deal if we get an extra
1021 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1022 mdd_object_fid(mdd_obj), NULL,
1023 xattr_name, handle);
1025 mdd_obj->mod_cltime = ktime_get();
1030 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
1031 * except for open flags have a dedicated record to store 32 bits of flags */
1032 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1033 enum changelog_rec_flags clf_flags,
1034 struct md_device *m, const struct lu_fid *fid)
1036 struct thandle *handle;
1037 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1041 LASSERT(fid != NULL);
1043 /* We'll check this again below, but we check now before we
1044 * start a transaction. */
1045 if (!mdd_changelog_enabled(env, mdd, type))
1048 handle = mdd_trans_create(env, mdd);
1050 RETURN(PTR_ERR(handle));
1052 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1056 rc = mdd_trans_start(env, mdd, handle);
1060 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1061 fid, NULL, NULL, handle);
1064 rc = mdd_trans_stop(env, mdd, rc, handle);
1070 * Save LMA extended attributes with data from \a ma.
1072 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1073 * not, LMA EA will be first read from disk, modified and write back.
1076 /* Precedence for choosing record type when multiple
1077 * attributes change: setattr > mtime > ctime > atime
1078 * (ctime changes when mtime does, plus chmod/chown.
1079 * atime and ctime are independent.) */
1080 static int mdd_attr_set_changelog(const struct lu_env *env,
1081 struct md_object *obj, struct thandle *handle,
1082 const struct lu_fid *pfid, __u64 valid)
1084 struct mdd_device *mdd = mdo2mdd(obj);
1087 bits = (valid & LA_SIZE) ? BIT(CL_TRUNC) : 0;
1088 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? BIT(CL_SETATTR) : 0;
1089 bits |= (valid & LA_MTIME) ? BIT(CL_MTIME) : 0;
1090 bits |= (valid & LA_CTIME) ? BIT(CL_CTIME) : 0;
1091 bits |= (valid & LA_ATIME) ? BIT(CL_ATIME) : 0;
1092 bits = bits & mdd->mdd_cl.mc_mask;
1093 /* This is an implementation limit rather than a protocol limit */
1094 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1098 /* The record type is the lowest non-masked set bit */
1101 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1102 return mdd_changelog_data_store(env, mdd, type, valid, md2mdd_obj(obj),
1106 static int mdd_declare_attr_set(const struct lu_env *env,
1107 struct mdd_device *mdd,
1108 struct mdd_object *obj,
1109 const struct lu_attr *attr,
1110 struct thandle *handle)
1114 rc = mdo_declare_attr_set(env, obj, attr, handle);
1118 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1119 if (attr->la_valid & LA_MODE) {
1120 mdd_read_lock(env, obj, DT_TGT_CHILD);
1121 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1122 XATTR_NAME_ACL_ACCESS);
1123 mdd_read_unlock(env, obj);
1124 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1130 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1131 rc = mdo_declare_xattr_set(env, obj, buf,
1132 XATTR_NAME_ACL_ACCESS, 0,
1140 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1149 * permission changes may require sync operation, to mitigate performance
1150 * impact, only do this for dir and when permission is reduced.
1152 * For regular files, version is updated with permission change (see VBR), async
1153 * permission won't cause any issue, while missing permission change on
1154 * directory may affect accessibility of other objects after recovery.
1156 static inline bool permission_needs_sync(const struct lu_attr *old,
1157 const struct lu_attr *new)
1159 if (!S_ISDIR(old->la_mode))
1162 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1165 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1168 if (new->la_valid & LA_MODE) {
1169 /* turned on sticky bit */
1170 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1173 /* set-GID has no impact on what is allowed, not checked */
1175 /* turned off setuid bit, or one of rwx for someone */
1176 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1177 (old->la_mode & (0777 | S_ISUID)))
1184 static inline __u64 mdd_lmm_dom_size(void *buf)
1186 struct lov_mds_md *lmm = buf;
1187 struct lov_comp_md_v1 *comp_v1;
1188 struct lov_mds_md *v1;
1194 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1197 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1198 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1199 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1201 /* DoM entry is the first entry always */
1202 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1203 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1208 /* set attr and LOV EA at once, return updated attr */
1209 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1210 const struct md_attr *ma)
1212 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1213 struct mdd_device *mdd = mdo2mdd(obj);
1214 struct thandle *handle = NULL;
1215 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1216 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1217 const struct lu_attr *la = &ma->ma_attr;
1218 struct lu_ucred *uc;
1219 bool chrgrp_by_unprivileged_user = false;
1223 /* we do not use ->attr_set() for LOV/HSM EA any more */
1224 LASSERT((ma->ma_valid & MA_LOV) == 0);
1225 LASSERT((ma->ma_valid & MA_HSM) == 0);
1227 rc = mdd_la_get(env, mdd_obj, attr);
1231 *la_copy = ma->ma_attr;
1232 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1236 /* no need to setattr anymore */
1237 if (la_copy->la_valid == 0) {
1239 "%s: no valid attribute on "DFID", previous was %#llx\n",
1240 mdd_obj_dev_name(mdd_obj),
1241 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1246 /* If an unprivileged user changes group of some file,
1247 * the setattr operation will be processed synchronously to
1248 * honor the quota limit of the corresponding group. see LU-5152 */
1249 uc = lu_ucred_check(env);
1250 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1251 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1252 /* LU-10048: disable synchronous chgrp operation for it will
1253 * cause deadlock between MDT and OST.
1254 la_copy->la_valid |= LA_FLAGS;
1255 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1257 chrgrp_by_unprivileged_user = true;
1259 /* Flush the possible existing client setattr requests to OSTs
1260 * to keep the order with the current setattr operation that
1261 * will be sent directly to OSTs. see LU-5152 */
1262 /* LU-11303 disable sync as this is too heavyweight.
1263 * This should be replaced with a sync only for the object
1264 * being modified here, not the whole filesystem.
1265 rc = dt_sync(env, mdd->mdd_child);
1271 handle = mdd_trans_create(env, mdd);
1272 if (IS_ERR(handle)) {
1273 rc = PTR_ERR(handle);
1279 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1283 rc = mdd_trans_start(env, mdd, handle);
1287 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1288 permission_needs_sync(attr, la))
1289 handle->th_sync = 1;
1291 if (la->la_valid & (LA_MTIME | LA_CTIME))
1292 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1293 la->la_mtime, la->la_ctime);
1295 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1297 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1298 * otherwise following truncate will do nothing and truncated
1299 * data may be read again. This is a quick fix until LU-11033
1302 if (la_copy->la_valid & LA_SIZE) {
1303 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1305 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1309 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1310 la_copy->la_valid &= ~LA_SIZE;
1311 lu_buf_free(lov_buf);
1315 if (la_copy->la_valid) {
1316 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1318 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1319 /* rollback to the original gid */
1320 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1321 la_copy->la_gid = attr->la_gid;
1322 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1325 mdd_write_unlock(env, mdd_obj);
1329 rc = mdd_attr_set_changelog(env, obj, handle, &ma->ma_pfid,
1333 rc = mdd_trans_stop(env, mdd, rc, handle);
1338 static int mdd_xattr_sanity_check(const struct lu_env *env,
1339 struct mdd_object *obj,
1340 const struct lu_attr *attr,
1343 struct lu_ucred *uc = lu_ucred_assert(env);
1346 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1349 if (strncmp(XATTR_USER_PREFIX, name,
1350 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1351 /* For sticky directories, only the owner and privileged user
1352 * can write attributes. */
1353 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1354 (uc->uc_fsuid != attr->la_uid) &&
1355 !md_capable(uc, CFS_CAP_FOWNER))
1357 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1358 (uc->uc_fsuid != attr->la_uid) &&
1359 !md_capable(uc, CFS_CAP_FOWNER)) {
1367 * Check if a string begins with a given prefix.
1369 * \param str String to check
1370 * \param prefix Substring to check at the beginning of \a str
1371 * \return true/false whether the condition is verified.
1373 static inline bool has_prefix(const char *str, const char *prefix)
1375 return strncmp(prefix, str, strlen(prefix)) == 0;
1379 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1381 * \param[in] xattr_name Full extended attribute name.
1383 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1385 static enum changelog_rec_type
1386 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1387 const char *xattr_name)
1389 /* Layout changes systematically recorded */
1390 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1391 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1392 allowed_lustre_lov(xattr_name))
1395 /* HSM information changes systematically recorded */
1396 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1399 /* Avoid logging SOM xattr for every file */
1400 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1403 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1404 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1405 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1406 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1407 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1413 static int mdd_declare_xattr_set(const struct lu_env *env,
1414 struct mdd_device *mdd,
1415 struct mdd_object *obj,
1416 const struct lu_buf *buf,
1418 int fl, struct thandle *handle)
1420 enum changelog_rec_type type;
1423 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1427 type = mdd_xattr_changelog_type(env, mdd, name);
1429 return 0; /* no changelog to store */
1431 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1435 * Compare current and future data of HSM EA and add a changelog if needed.
1437 * Caller should have write-locked \param obj.
1439 * \param buf - Future HSM EA content.
1440 * \retval 0 if no changelog is needed or changelog was added properly.
1441 * \retval -ve errno if there was a problem
1443 static int mdd_hsm_update_locked(const struct lu_env *env,
1444 struct md_object *obj,
1445 const struct lu_buf *buf,
1446 struct thandle *handle,
1447 enum changelog_rec_flags *clf_flags)
1449 struct mdd_thread_info *info = mdd_env_info(env);
1450 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1451 struct lu_buf *current_buf;
1452 struct md_hsm *current_mh;
1453 struct md_hsm *new_mh;
1457 OBD_ALLOC_PTR(current_mh);
1458 if (current_mh == NULL)
1461 /* Read HSM attrs from disk */
1462 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1464 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1466 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1467 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1468 if (rc < 0 && rc != -ENODATA)
1470 else if (rc == -ENODATA)
1471 current_mh->mh_flags = 0;
1473 /* Map future HSM xattr */
1474 OBD_ALLOC_PTR(new_mh);
1476 GOTO(free, rc = -ENOMEM);
1477 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1481 /* Flags differ, set flags for the changelog that will be added */
1482 if (current_mh->mh_flags != new_mh->mh_flags) {
1483 hsm_set_cl_event(clf_flags, HE_STATE);
1484 if (new_mh->mh_flags & HS_DIRTY)
1485 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1488 OBD_FREE_PTR(new_mh);
1491 OBD_FREE_PTR(current_mh);
1495 static int mdd_object_pfid_replace(const struct lu_env *env,
1496 struct mdd_object *o)
1498 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1499 struct thandle *handle;
1502 handle = mdd_trans_create(env, mdd);
1504 RETURN(PTR_ERR(handle));
1506 handle->th_complex = 1;
1508 /* it doesn't need to track the PFID update via llog, because LFSCK
1509 * will repair it even it goes wrong */
1510 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1515 rc = mdd_trans_start(env, mdd, handle);
1519 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1524 mdd_trans_stop(env, mdd, rc, handle);
1529 static int mdd_declare_xattr_del(const struct lu_env *env,
1530 struct mdd_device *mdd,
1531 struct mdd_object *obj,
1533 struct thandle *handle);
1535 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1538 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1539 struct md_object *md_vic)
1541 struct mdd_device *mdd = mdo2mdd(md_obj);
1542 struct mdd_object *obj = md2mdd_obj(md_obj);
1543 struct mdd_object *vic = md2mdd_obj(md_vic);
1544 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1545 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1546 struct lov_mds_md *lmm;
1547 struct thandle *handle;
1551 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1552 if (rc == 0) /* same fid */
1555 handle = mdd_trans_create(env, mdd);
1557 RETURN(PTR_ERR(handle));
1560 mdd_write_lock(env, obj, DT_TGT_CHILD);
1561 mdd_write_lock(env, vic, DT_TGT_CHILD);
1563 mdd_write_lock(env, vic, DT_TGT_CHILD);
1564 mdd_write_lock(env, obj, DT_TGT_CHILD);
1567 /* get EA of victim file */
1568 memset(buf_vic, 0, sizeof(*buf_vic));
1569 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1576 /* parse the layout of victim file */
1577 lmm = buf_vic->lb_buf;
1578 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1579 GOTO(out, rc = -EINVAL);
1581 /* save EA of target file for restore */
1582 memset(buf, 0, sizeof(*buf));
1583 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1587 /* Get rid of the layout from victim object */
1588 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1592 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1593 LU_XATTR_MERGE, handle);
1597 rc = mdd_trans_start(env, mdd, handle);
1601 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1606 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1608 GOTO(out_restore, rc);
1610 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1612 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle,
1618 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1619 LU_XATTR_REPLACE, handle);
1621 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1622 mdd_obj_dev_name(obj),
1623 PFID(mdd_object_fid(obj)), rc2);
1627 mdd_trans_stop(env, mdd, rc, handle);
1628 mdd_write_unlock(env, obj);
1629 mdd_write_unlock(env, vic);
1631 lu_buf_free(buf_vic);
1634 (void) mdd_object_pfid_replace(env, obj);
1640 * Extract the mirror with specified mirror id, and store the splitted
1641 * mirror layout to @buf.
1643 * \param[in] comp_v1 mirrored layout
1644 * \param[in] mirror_id the mirror with mirror_id to be extracted
1645 * \param[out] buf store the layout excluding the extracted mirror,
1646 * caller free the buffer we allocated in this function
1647 * \param[out] buf_vic store the extracted layout, caller free the buffer
1648 * we allocated in this function
1650 * \retval 0 on success; < 0 if error happens
1652 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1653 struct lu_buf *buf, struct lu_buf *buf_vic)
1655 struct lov_comp_md_v1 *comp_rem;
1656 struct lov_comp_md_v1 *comp_vic;
1657 struct lov_comp_md_entry_v1 *entry;
1658 struct lov_comp_md_entry_v1 *entry_rem;
1659 struct lov_comp_md_entry_v1 *entry_vic;
1661 __u16 comp_cnt, count = 0;
1662 int lmm_size, lmm_size_vic = 0;
1664 int offset, offset_rem, offset_vic;
1666 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1667 /* comp_v1 should contains more than 1 mirror */
1668 if (mirror_cnt <= 1)
1670 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1671 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1673 for (i = 0; i < comp_cnt; i++) {
1674 entry = &comp_v1->lcm_entries[i];
1675 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1677 lmm_size_vic += sizeof(*entry);
1678 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1679 } else if (count > 0) {
1680 /* find the specified mirror */
1688 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1692 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1693 if (!buf_vic->lb_buf) {
1698 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1699 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1701 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1702 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1703 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1704 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1705 if (!comp_rem->lcm_mirror_count)
1706 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1708 memset(comp_vic, 0, sizeof(*comp_v1));
1709 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1710 comp_vic->lcm_mirror_count = 0;
1711 comp_vic->lcm_entry_count = cpu_to_le32(count);
1712 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1713 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1714 comp_vic->lcm_layout_gen = 0;
1716 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1717 offset_rem = sizeof(*comp_rem) +
1718 sizeof(*entry_rem) * (comp_cnt - count);
1719 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1720 for (i = j = k = 0; i < comp_cnt; i++) {
1721 struct lov_mds_md *lmm, *lmm_dst;
1724 entry = &comp_v1->lcm_entries[i];
1725 entry_vic = &comp_vic->lcm_entries[j];
1726 entry_rem = &comp_rem->lcm_entries[k];
1728 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1731 /* copy component entry */
1733 memcpy(entry_vic, entry, sizeof(*entry));
1734 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1735 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1738 memcpy(entry_rem, entry, sizeof(*entry));
1739 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1743 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1745 lmm_dst = (struct lov_mds_md *)
1746 ((char *)comp_vic + offset_vic);
1748 lmm_dst = (struct lov_mds_md *)
1749 ((char *)comp_rem + offset_rem);
1751 /* copy component entry blob */
1752 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1754 /* blob offset advance */
1755 offset += le32_to_cpu(entry->lcme_size);
1757 offset_vic += le32_to_cpu(entry->lcme_size);
1759 offset_rem += le32_to_cpu(entry->lcme_size);
1765 static int mdd_dom_data_truncate(const struct lu_env *env,
1766 struct mdd_device *mdd, struct mdd_object *mo);
1768 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1769 struct md_rejig_data *mrd)
1771 struct mdd_device *mdd = mdo2mdd(md_obj);
1772 struct mdd_object *obj = md2mdd_obj(md_obj);
1773 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1774 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1775 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1776 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1777 struct lov_comp_md_v1 *lcm;
1778 struct thandle *handle;
1780 bool dom_stripe = false;
1784 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1785 if (rc == 0) /* same fid */
1788 handle = mdd_trans_create(env, mdd);
1790 RETURN(PTR_ERR(handle));
1793 mdd_write_lock(env, obj, DT_TGT_CHILD);
1794 mdd_write_lock(env, vic, DT_TGT_CHILD);
1796 mdd_write_lock(env, vic, DT_TGT_CHILD);
1797 mdd_write_lock(env, obj, DT_TGT_CHILD);
1800 /* get EA of mirrored file */
1801 memset(buf_save, 0, sizeof(*buf));
1802 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1806 lcm = buf_save->lb_buf;
1807 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1808 GOTO(out, rc = -EINVAL);
1811 * Extract the mirror with specified mirror id, and store the splitted
1812 * mirror layout to the victim file.
1814 memset(buf, 0, sizeof(*buf));
1815 memset(buf_vic, 0, sizeof(*buf_vic));
1816 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1820 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1822 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1823 LU_XATTR_SPLIT, handle);
1826 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1827 LU_XATTR_SPLIT, handle);
1831 rc = mdd_trans_start(env, mdd, handle);
1835 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1840 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1843 GOTO(out_restore, rc);
1845 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1850 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle,
1858 /* restore obj's layout */
1859 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1860 LU_XATTR_REPLACE, handle);
1862 CERROR("%s: failed rollback "DFID" layout: file state unkonwn: rc = %d\n",
1863 mdd_obj_dev_name(obj),
1864 PFID(mdd_object_fid(obj)), rc2);
1867 rc = mdd_trans_stop(env, mdd, rc, handle);
1869 /* Truncate local DOM data if all went well */
1870 if (!rc && dom_stripe)
1871 mdd_dom_data_truncate(env, mdd, obj);
1873 mdd_write_unlock(env, obj);
1874 mdd_write_unlock(env, vic);
1875 lu_buf_free(buf_save);
1877 lu_buf_free(buf_vic);
1880 (void) mdd_object_pfid_replace(env, obj);
1885 static int mdd_layout_merge_allowed(const struct lu_env *env,
1886 struct md_object *target,
1887 struct md_object *victim)
1889 struct mdd_object *o1 = md2mdd_obj(target);
1891 /* cannot extend directory's LOVEA */
1892 if (S_ISDIR(mdd_object_type(o1))) {
1893 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1894 mdd_obj_dev_name(o1));
1902 * The caller should guarantee to update the object ctime
1903 * after xattr_set if needed.
1905 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1906 const struct lu_buf *buf, const char *name,
1909 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1910 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1911 struct mdd_device *mdd = mdo2mdd(obj);
1912 struct thandle *handle;
1913 enum changelog_rec_type cl_type;
1914 enum changelog_rec_flags clf_flags = 0;
1918 rc = mdd_la_get(env, mdd_obj, attr);
1922 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1926 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1927 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1928 struct md_rejig_data *mrd = buf->lb_buf;
1929 struct md_object *victim = mrd->mrd_obj;
1931 if (buf->lb_len != sizeof(*mrd))
1934 rc = mdd_layout_merge_allowed(env, obj, victim);
1938 if (fl == LU_XATTR_MERGE)
1939 /* merge layout of victim as a mirror of obj's. */
1940 rc = mdd_xattr_merge(env, obj, victim);
1942 rc = mdd_xattr_split(env, obj, mrd);
1946 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1947 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1948 struct posix_acl *acl;
1950 /* user may set empty ACL, which should be treated as removing
1952 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1955 RETURN(PTR_ERR(acl));
1957 rc = mdd_xattr_del(env, obj, name);
1960 posix_acl_release(acl);
1963 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1964 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1968 handle = mdd_trans_create(env, mdd);
1970 RETURN(PTR_ERR(handle));
1972 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1976 rc = mdd_trans_start(env, mdd, handle);
1980 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1982 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1983 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1985 mdd_write_unlock(env, mdd_obj);
1990 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1991 mdd_write_unlock(env, mdd_obj);
1995 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1999 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
2000 mdd_obj, name, handle);
2004 return mdd_trans_stop(env, mdd, rc, handle);
2007 static int mdd_declare_xattr_del(const struct lu_env *env,
2008 struct mdd_device *mdd,
2009 struct mdd_object *obj,
2011 struct thandle *handle)
2013 enum changelog_rec_type type;
2016 rc = mdo_declare_xattr_del(env, obj, name, handle);
2020 type = mdd_xattr_changelog_type(env, mdd, name);
2022 return 0; /* no changelog to store */
2024 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2028 * The caller should guarantee to update the object ctime
2029 * after xattr_set if needed.
2031 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2034 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2035 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2036 struct mdd_device *mdd = mdo2mdd(obj);
2037 struct thandle *handle;
2041 rc = mdd_la_get(env, mdd_obj, attr);
2045 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2049 handle = mdd_trans_create(env, mdd);
2051 RETURN(PTR_ERR(handle));
2053 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2057 rc = mdd_trans_start(env, mdd, handle);
2061 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2062 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2063 mdd_write_unlock(env, mdd_obj);
2067 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2070 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2075 return mdd_trans_stop(env, mdd, rc, handle);
2079 * read lov/lmv EA of an object
2080 * return the lov/lmv EA in an allocated lu_buf
2082 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2083 struct lu_buf *lmm_buf, const char *name)
2085 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2090 if (buf->lb_buf == NULL) {
2091 buf = lu_buf_check_and_alloc(buf, 4096);
2092 if (buf->lb_buf == NULL)
2097 rc = mdo_xattr_get(env, obj, buf, name);
2098 if (rc == -ERANGE) {
2099 /* mti_big_buf is allocated but is too small
2100 * we need to increase it */
2101 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2103 if (buf->lb_buf == NULL)
2106 } else if (rc < 0) {
2108 } else if (rc == 0) {
2112 lu_buf_alloc(lmm_buf, rc);
2113 if (lmm_buf->lb_buf == NULL)
2117 * we don't use lmm_buf directly, because we don't know xattr size, so
2118 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2120 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2125 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2126 struct mdd_object *o, struct lu_buf *buf,
2127 struct thandle *handle)
2129 struct hsm_attrs *attrs;
2130 enum hsm_states hsm_flags;
2131 enum changelog_rec_flags clf_flags = 0;
2135 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2140 attrs = buf->lb_buf;
2141 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2142 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2145 /* Add a changelog record for release. */
2146 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2147 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2148 clf_flags, o, handle, NULL);
2153 * check if layout swapping between 2 objects is allowed
2155 * - only normal FIDs or non-system IGIFs
2156 * - same type of objects
2157 * - same owner/group (so quotas are still valid) unless this is from HSM
2160 static int mdd_layout_swap_allowed(const struct lu_env *env,
2161 struct mdd_object *o1,
2162 const struct lu_attr *attr1,
2163 struct mdd_object *o2,
2164 const struct lu_attr *attr2,
2167 const struct lu_fid *fid1, *fid2;
2170 fid1 = mdd_object_fid(o1);
2171 fid2 = mdd_object_fid(o2);
2173 if (!fid_is_norm(fid1) &&
2174 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2177 if (!fid_is_norm(fid2) &&
2178 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2181 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2182 if (S_ISDIR(mdd_object_type(o1)))
2184 if (S_ISREG(mdd_object_type(o1)))
2189 if (flags & SWAP_LAYOUTS_MDS_HSM)
2192 if ((attr1->la_uid != attr2->la_uid) ||
2193 (attr1->la_gid != attr2->la_gid))
2199 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2200 * look into the layout in MDD layer. */
2201 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2203 struct lov_comp_md_v1 *comp_v1;
2204 struct lov_mds_md *v1;
2208 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2209 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2210 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2216 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2217 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2220 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2221 off = le32_to_cpu(comp_v1->lcm_entries[i].
2223 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2228 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2229 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2240 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2242 return mdd_lmm_oi(lmm, oi, true);
2245 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2247 return mdd_lmm_oi(lmm, oi, false);
2250 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2252 struct lov_comp_md_v1 *comp_v1;
2254 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2255 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2257 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2259 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2260 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2261 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2262 __u16 tmp_gen = *gen;
2264 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2266 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2273 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2275 return mdd_lmm_gen(lmm, gen, true);
2278 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2280 return mdd_lmm_gen(lmm, gen, false);
2283 static int mdd_dom_data_truncate(const struct lu_env *env,
2284 struct mdd_device *mdd, struct mdd_object *mo)
2287 struct dt_object *dom;
2290 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2292 GOTO(out, rc = -ENODATA);
2294 th = dt_trans_create(env, mdd->mdd_bottom);
2296 GOTO(out, rc = PTR_ERR(th));
2298 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2302 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2306 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2308 dt_trans_stop(env, mdd->mdd_bottom, th);
2310 /* Ignore failure but report the error */
2312 CERROR("%s: can't truncate DOM inode "DFID" data: rc = %d\n",
2313 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2318 * swap layouts between 2 lustre objects
2320 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2321 struct md_object *obj2, __u64 flags)
2323 struct mdd_thread_info *info = mdd_env_info(env);
2324 struct mdd_object *fst_o = md2mdd_obj(obj1);
2325 struct mdd_object *snd_o = md2mdd_obj(obj2);
2326 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2327 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2328 struct mdd_device *mdd = mdo2mdd(obj1);
2329 struct lov_mds_md *fst_lmm, *snd_lmm;
2330 struct lu_buf *fst_buf = &info->mti_buf[0];
2331 struct lu_buf *snd_buf = &info->mti_buf[1];
2332 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2333 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2334 struct ost_id *saved_oi = NULL;
2335 struct thandle *handle;
2336 struct mdd_object *dom_o = NULL;
2337 __u64 domsize_dom, domsize_vlt;
2338 __u32 fst_gen, snd_gen, saved_gen;
2344 BUILD_BUG_ON(ARRAY_SIZE(info->mti_buf) < 4);
2345 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2347 /* we have to sort the 2 obj, so locking will always
2348 * be in the same order, even in case of 2 concurrent swaps */
2349 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2350 if (rc == 0) /* same fid ? */
2356 rc = mdd_la_get(env, fst_o, fst_la);
2360 rc = mdd_la_get(env, snd_o, snd_la);
2364 /* check if layout swapping is allowed */
2365 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2369 handle = mdd_trans_create(env, mdd);
2371 RETURN(PTR_ERR(handle));
2373 /* objects are already sorted */
2374 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2375 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2377 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2378 if (rc < 0 && rc != -ENODATA)
2381 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2382 if (rc < 0 && rc != -ENODATA)
2385 /* check if file has DoM. DoM file can be migrated only to another
2386 * DoM layout with the same DoM component size or to an non-DOM
2387 * layout. After migration to OSTs layout, local MDT inode data
2388 * should be truncated.
2389 * Objects are sorted by FIDs, considering that original file's FID
2390 * is always smaller the snd_o is always original file we are migrating
2393 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2394 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2396 /* Only migration is supported for DoM files, not 'swap_layouts' so
2397 * target file must be volatile and orphan.
2399 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2400 dom_o = domsize_dom ? snd_o : NULL;
2401 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2402 swap(domsize_dom, domsize_vlt);
2403 dom_o = domsize_dom ? fst_o : NULL;
2404 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2405 /* 'lfs swap_layouts' case, neither file should have DoM */
2407 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2408 "use migration instead: rc = %d\n", rc);
2412 if (domsize_vlt > 0 && domsize_dom == 0) {
2415 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2416 mdd_obj_dev_name(snd_o),
2417 PFID(mdd_object_fid(snd_o)), rc);
2419 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2422 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2423 mdd_obj_dev_name(fst_o),
2424 PFID(mdd_object_fid(fst_o)), rc);
2426 } else if (domsize_vlt > 0) {
2427 /* Migration with the same DOM component size, no need to
2428 * truncate local data, it is still being used */
2432 /* swapping 2 non existant layouts is a success */
2433 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2436 /* to help inode migration between MDT, it is better to
2437 * start by the no layout file (if one), so we order the swap */
2438 if (snd_buf->lb_buf == NULL) {
2440 swap(fst_buf, snd_buf);
2443 fst_gen = snd_gen = 0;
2444 /* lmm and generation layout initialization */
2445 if (fst_buf->lb_buf != NULL) {
2446 fst_lmm = fst_buf->lb_buf;
2447 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2448 fst_fl = LU_XATTR_REPLACE;
2452 fst_fl = LU_XATTR_CREATE;
2455 snd_lmm = snd_buf->lb_buf;
2456 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2458 saved_gen = fst_gen;
2459 /* increase the generation layout numbers */
2464 * XXX The layout generation is used to generate component IDs for
2465 * the composite file, we have to do some special tweaks to make
2466 * sure the layout generation is always adequate for that job.
2469 /* Skip invalid generation number for composite layout */
2470 if ((snd_gen & LCME_ID_MASK) == 0)
2472 if ((fst_gen & LCME_ID_MASK) == 0)
2474 /* Make sure the generation is greater than all the component IDs */
2475 if (fst_gen < snd_gen)
2477 else if (fst_gen > snd_gen)
2480 /* set the file specific informations in lmm */
2481 if (fst_lmm != NULL) {
2482 struct ost_id temp_oi;
2484 saved_oi = &info->mti_oa.o_oi;
2485 mdd_get_lmm_oi(fst_lmm, saved_oi);
2486 mdd_get_lmm_oi(snd_lmm, &temp_oi);
2487 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2488 mdd_set_lmm_oi(fst_lmm, &temp_oi);
2489 mdd_set_lmm_oi(snd_lmm, saved_oi);
2491 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2492 cpu_to_le32(LOV_MAGIC_MAGIC))
2493 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2495 GOTO(stop, rc = -EPROTO);
2497 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2499 /* Prepare HSM attribute if it's required */
2500 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2501 const int buflen = sizeof(struct hsm_attrs);
2503 lu_buf_alloc(fst_hsm_buf, buflen);
2504 lu_buf_alloc(snd_hsm_buf, buflen);
2505 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2506 GOTO(stop, rc = -ENOMEM);
2508 /* Read HSM attribute */
2509 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2513 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2517 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2518 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2523 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2524 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2530 /* prepare transaction */
2531 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2536 if (fst_buf->lb_buf != NULL)
2537 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2538 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2541 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2546 rc = mdd_trans_start(env, mdd, handle);
2550 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2551 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2555 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2557 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2560 CERROR("%s: HSM error restoring "DFID": rc = %d/%d\n",
2561 mdd_obj_dev_name(fst_o),
2562 PFID(mdd_object_fid(fst_o)), rc, rc2);
2567 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2571 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2574 if (fst_buf->lb_buf != NULL)
2575 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2576 LU_XATTR_REPLACE, handle);
2578 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2581 GOTO(out_restore, rc);
2583 /* Issue one changelog record per file */
2584 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle,
2589 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle,
2599 /* failure on second file, but first was done, so we have
2600 * to roll back first. */
2601 if (fst_buf->lb_buf != NULL) {
2602 mdd_set_lmm_oi(fst_lmm, saved_oi);
2603 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2604 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2605 LU_XATTR_REPLACE, handle);
2607 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2613 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2618 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2623 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
2624 mdd_obj_dev_name(fst_o),
2625 PFID(mdd_object_fid(snd_o)),
2626 PFID(mdd_object_fid(fst_o)),
2628 /* a solution to avoid journal commit is to panic,
2629 * but it has strong consequences so we use LBUG to
2630 * allow sysdamin to choose to panic or not
2637 rc = mdd_trans_stop(env, mdd, rc, handle);
2639 /* Truncate local DOM data if all went well */
2641 mdd_dom_data_truncate(env, mdd, dom_o);
2643 mdd_write_unlock(env, snd_o);
2644 mdd_write_unlock(env, fst_o);
2646 lu_buf_free(fst_buf);
2647 lu_buf_free(snd_buf);
2648 lu_buf_free(fst_hsm_buf);
2649 lu_buf_free(snd_hsm_buf);
2652 (void) mdd_object_pfid_replace(env, fst_o);
2653 (void) mdd_object_pfid_replace(env, snd_o);
2658 static int mdd_declare_layout_change(const struct lu_env *env,
2659 struct mdd_device *mdd,
2660 struct mdd_object *obj,
2661 struct md_layout_change *mlc,
2662 struct thandle *handle)
2666 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2670 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2674 /* For PFL, this is used to instantiate necessary component objects. */
2676 mdd_layout_instantiate_component(const struct lu_env *env,
2677 struct mdd_object *obj, struct md_layout_change *mlc,
2678 struct thandle *handle)
2680 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2684 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2687 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2689 * It's possible that another layout write intent has already
2690 * instantiated our objects, so a -EALREADY returned, and we need to
2694 RETURN(rc == -EALREADY ? 0 : rc);
2696 rc = mdd_trans_start(env, mdd, handle);
2700 mdd_write_lock(env, obj, DT_TGT_CHILD);
2701 rc = mdo_layout_change(env, obj, mlc, handle);
2702 mdd_write_unlock(env, obj);
2706 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
2712 * Change the FLR layout from RDONLY to WRITE_PENDING.
2714 * It picks the primary mirror, and bumps the layout version, and set
2715 * layout version xattr to OST objects in a sync tx. In order to facilitate
2716 * the handling of phantom writers from evicted clients, the clients carry
2717 * layout version of the file with write RPC, so that the OSTs can verify
2718 * if the write RPCs are legitimate, meaning not from evicted clients.
2721 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2722 struct md_layout_change *mlc, struct thandle *handle)
2724 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2725 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2726 struct lustre_som_attrs *som = &mlc->mlc_som;
2731 /* Verify acceptable operations */
2732 switch (mlc->mlc_opc) {
2733 case MD_LAYOUT_WRITE:
2734 case MD_LAYOUT_RESYNC:
2735 /* these are legal operations - this represents the case that
2736 * a few mirrors were missed in the last resync. */
2738 case MD_LAYOUT_RESYNC_DONE:
2743 som_buf->lb_buf = som;
2744 som_buf->lb_len = sizeof(*som);
2745 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2746 if (rc < 0 && rc != -ENODATA)
2750 lustre_som_swab(som);
2751 if (som->lsa_valid & SOM_FL_STRICT)
2752 fl = LU_XATTR_REPLACE;
2755 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2760 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2761 XATTR_NAME_SOM, fl, handle);
2766 /* record a changelog for data mover to consume */
2767 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2771 rc = mdd_trans_start(env, mdd, handle);
2775 /* it needs a sync tx to make FLR to work properly */
2776 handle->th_sync = 1;
2778 mdd_write_lock(env, obj, DT_TGT_CHILD);
2779 rc = mdo_layout_change(env, obj, mlc, handle);
2781 /* SOM state transition from STRICT to STALE */
2782 som->lsa_valid = SOM_FL_STALE;
2783 lustre_som_swab(som);
2784 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2787 mdd_write_unlock(env, obj);
2791 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle, NULL);
2802 * Handle mirrored file state transition when it's in WRITE_PENDING.
2804 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2805 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2806 * layout version will be increased, and the file's state will be changed to
2810 mdd_layout_update_write_pending(const struct lu_env *env,
2811 struct mdd_object *obj, struct md_layout_change *mlc,
2812 struct thandle *handle)
2814 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2818 switch (mlc->mlc_opc) {
2819 case MD_LAYOUT_RESYNC:
2820 /* Upon receiving the resync request, it should
2821 * instantiate all stale components right away to get ready
2822 * for mirror copy. In order to avoid layout version change,
2823 * client should avoid sending LAYOUT_WRITE request at the
2826 case MD_LAYOUT_WRITE:
2827 /* legal race for concurrent write, the file state has been
2828 * changed by another client. */
2834 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2838 rc = mdd_trans_start(env, mdd, handle);
2842 /* it needs a sync tx to make FLR to work properly */
2843 handle->th_sync = 1;
2845 mdd_write_lock(env, obj, DT_TGT_CHILD);
2846 rc = mdo_layout_change(env, obj, mlc, handle);
2847 mdd_write_unlock(env, obj);
2858 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2860 * Only concurrent write and sync complete requests are possible when the
2861 * file is in SYNC_PENDING. For the latter request, it will pass in the
2862 * mirrors that have been synchronized, then the stale bit will be cleared
2863 * to make the file's state turn into RDONLY.
2864 * For concurrent write reqeust, it just needs to change the file's state
2865 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2866 * version because the version will be increased in the transition to
2867 * SYNC_PENDING later so that it can deny the write request from potential
2868 * evicted SYNC clients. */
2870 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2871 struct md_layout_change *mlc, struct thandle *handle)
2873 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2874 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2879 /* operation validation */
2880 switch (mlc->mlc_opc) {
2881 case MD_LAYOUT_RESYNC_DONE:
2882 /* resync complete. */
2883 case MD_LAYOUT_WRITE:
2884 /* concurrent write. */
2886 case MD_LAYOUT_RESYNC:
2887 /* resync again, most likely the previous run failed.
2888 * no-op if it's already in SYNC_PENDING state */
2894 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2895 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2896 if (rc < 0 && rc != -ENODATA)
2899 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2900 lustre_som_swab(&mlc->mlc_som);
2901 som_buf->lb_buf = &mlc->mlc_som;
2902 som_buf->lb_len = sizeof(mlc->mlc_som);
2905 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2909 /* record a changelog for the completion of resync */
2910 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2915 /* RESYNC_DONE has piggybacked size and blocks */
2917 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2918 XATTR_NAME_SOM, fl, handle);
2923 rc = mdd_trans_start(env, mdd, handle);
2927 /* it needs a sync tx to make FLR to work properly */
2928 handle->th_sync = 1;
2930 rc = mdo_layout_change(env, obj, mlc, handle);
2935 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2941 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle,
2951 * Layout change callback for object.
2953 * This is only used by FLR for now. In the future, it can be exteneded to
2954 * handle all layout change.
2957 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2958 struct md_layout_change *mlc)
2960 struct mdd_object *obj = md2mdd_obj(o);
2961 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2962 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2963 struct lov_comp_md_v1 *lcm;
2964 struct thandle *handle;
2970 if (S_ISDIR(mdd_object_type(obj))) {
2971 switch (mlc->mlc_opc) {
2972 case MD_LAYOUT_SHRINK:
2973 rc = mdd_dir_layout_shrink(env, o, mlc);
2975 case MD_LAYOUT_SPLIT:
2976 rc = mdd_dir_layout_split(env, o, mlc);
2985 /* Verify acceptable operations */
2986 switch (mlc->mlc_opc) {
2987 case MD_LAYOUT_WRITE:
2988 case MD_LAYOUT_RESYNC:
2989 case MD_LAYOUT_RESYNC_DONE:
2995 handle = mdd_trans_create(env, mdd);
2997 RETURN(PTR_ERR(handle));
2999 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
3006 /* analyze the layout to make sure it's a FLR file */
3008 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
3009 GOTO(out, rc = -EINVAL);
3011 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
3013 /* please refer to HLD of FLR for state transition */
3014 switch (flr_state) {
3016 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
3018 case LCM_FL_WRITE_PENDING:
3019 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
3022 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
3024 case LCM_FL_SYNC_PENDING:
3025 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3034 mdd_trans_stop(env, mdd, rc, handle);
3039 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3040 struct mdd_object *child, const struct lu_attr *attr,
3041 const struct md_op_spec *spec,
3042 struct dt_allocation_hint *hint)
3044 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3045 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3046 struct dt_object *nc = mdd_object_child(child);
3048 memset(hint, 0, sizeof(*hint));
3050 /* For striped directory, give striping EA to lod_ah_init, which will
3051 * decide the stripe_offset and stripe count by it. */
3052 if (S_ISDIR(attr->la_mode) &&
3053 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3054 hint->dah_eadata = spec->u.sp_ea.eadata;
3055 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3057 hint->dah_eadata = NULL;
3058 hint->dah_eadata_len = 0;
3059 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
3060 if (mdd->mdd_append_stripe_count != 0 ||
3061 mdd->mdd_append_pool[0])
3063 "using O_APPEND file striping\n");
3064 if (mdd->mdd_append_stripe_count)
3065 hint->dah_append_stripes =
3066 mdd->mdd_append_stripe_count;
3067 if (mdd->mdd_append_pool[0])
3068 hint->dah_append_pool = mdd->mdd_append_pool;
3070 hint->dah_append_stripes = 0;
3074 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3075 hint->dah_eadata, hint->dah_eadata_len);
3076 /* @hint will be initialized by underlying device. */
3077 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3080 static int accmode(const struct lu_env *env, const struct lu_attr *la,
3083 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3084 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3085 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3086 * owner can write to a file even if it is marked readonly to hide
3087 * its brokenness. (bug 5781) */
3088 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3089 struct lu_ucred *uc = lu_ucred_check(env);
3091 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3095 return mds_accmode(open_flags);
3098 static int mdd_open_sanity_check(const struct lu_env *env,
3099 struct mdd_object *obj,
3100 const struct lu_attr *attr, u64 open_flags,
3106 /* EEXIST check, also opening of *open* orphans is allowed so we can
3107 * open-by-handle unlinked files
3109 if (mdd_is_dead_obj(obj) && !is_replay &&
3110 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3113 if (S_ISLNK(attr->la_mode))
3116 mode = accmode(env, attr, open_flags);
3118 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3121 if (!(open_flags & MDS_OPEN_CREATED)) {
3122 rc = mdd_permission_internal(env, obj, attr, mode);
3127 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3128 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3129 open_flags &= ~MDS_OPEN_TRUNC;
3131 /* For writing append-only file must open it with append mode. */
3132 if (attr->la_flags & LUSTRE_APPEND_FL) {
3133 if ((open_flags & MDS_FMODE_WRITE) &&
3134 !(open_flags & MDS_OPEN_APPEND))
3136 if (open_flags & MDS_OPEN_TRUNC)
3143 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3144 u64 open_flags, struct md_op_spec *spec)
3146 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3147 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3148 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3149 struct mdd_object_user *mou = NULL;
3150 const struct lu_ucred *uc = lu_ucred(env);
3151 struct mdd_device *mdd = mdo2mdd(obj);
3152 enum changelog_rec_type type = CL_OPEN;
3156 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3158 rc = mdd_la_get(env, mdd_obj, attr);
3162 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3164 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & BIT(CL_DN_OPEN)))
3169 mdd_obj->mod_count++;
3171 if (!mdd_changelog_enabled(env, mdd, type))
3175 /* look for existing opener in list under mdd_write_lock */
3176 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3181 /* add user to list */
3182 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3188 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3190 mdd_obj_user_free(mou);
3195 if (type == CL_DN_OPEN) {
3196 if (ktime_before(ktime_get(), mou->mou_deniednext))
3197 /* same user denied again same access within
3198 * time interval: do not record
3202 /* this user already denied, but some time ago:
3203 * update denied time
3205 mou->mou_deniednext =
3206 ktime_add(ktime_get(),
3207 ktime_set(mdd->mdd_cl.mc_deniednext,
3210 mou->mou_opencount++;
3211 /* same user opening file again with same flags:
3218 /* FYI, only the bottom 32 bits of open_flags are recorded */
3219 mdd_changelog(env, type, open_flags, md_dev, mdd_object_fid(mdd_obj));
3223 mdd_write_unlock(env, mdd_obj);
3227 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3228 struct md_attr *ma, struct thandle *handle)
3232 rc = mdd_orphan_declare_delete(env, obj, handle);
3236 return mdo_declare_destroy(env, obj, handle);
3240 * No permission check is needed.
3242 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3243 struct md_attr *ma, u64 open_flags)
3245 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3246 struct mdd_device *mdd = mdo2mdd(obj);
3247 struct thandle *handle = NULL;
3250 bool blocked = false;
3251 bool last_close_by_uid = false;
3252 const struct lu_ucred *uc = lu_ucred(env);
3255 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3256 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3257 mdd_obj->mod_count--;
3258 mdd_write_unlock(env, mdd_obj);
3260 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3261 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3262 "list\n", PFID(mdd_object_fid(mdd_obj)));
3266 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3267 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3268 /* check without any lock */
3269 is_orphan = mdd_obj->mod_count == 1 &&
3270 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3274 /* mdd_trans_create() maybe failed because of barrier_entry(),
3275 * under such case, the orphan MDT-object will be left in the
3276 * orphan list, and when the MDT remount next time, the unused
3277 * orphans will be destroyed automatically.
3279 * One exception: the former mdd_finish_unlink may failed to
3280 * add the orphan MDT-object to the orphan list, then if the
3281 * mdd_trans_create() failed because of barrier_entry(), the
3282 * MDT-object will become real orphan that is neither in the
3283 * namespace nor in the orphan list. Such bad case should be
3284 * very rare and will be handled by e2fsck/lfsck. */
3285 handle = mdd_trans_create(env, mdo2mdd(obj));
3286 if (IS_ERR(handle)) {
3287 rc = PTR_ERR(handle);
3288 if (rc != -EINPROGRESS)
3296 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3300 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3305 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3311 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3312 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3314 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3315 lu_dev_name(mdd2lu_dev(mdd)),
3316 PFID(mdd_object_fid(mdd_obj)), rc);
3320 /* check again with lock */
3321 is_orphan = (mdd_obj->mod_count == 1) &&
3322 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3323 ma->ma_attr.la_nlink == 0);
3325 if (is_orphan && !handle && !blocked) {
3326 mdd_write_unlock(env, mdd_obj);
3330 mdd_obj->mod_count--; /*release open count */
3332 /* under mdd write lock */
3333 /* If recording, see if we need to remove UID from list. uc is not
3334 * initialized if the client has been evicted. */
3335 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3336 struct mdd_object_user *mou;
3338 /* look for UID in list */
3339 /* If mou is NULL, it probably means logging was enabled after
3340 * the user had the file open. So the corresponding close
3341 * will not be logged.
3343 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3346 mou->mou_opencount--;
3347 if (mou->mou_opencount == 0) {
3348 mdd_obj_user_remove(mdd_obj, mou);
3349 last_close_by_uid = true;
3354 if (!is_orphan || blocked)
3358 /* NB: Object maybe not in orphan list originally, it is rare case for
3359 * mdd_finish_unlink() failure, in that case, the object doesn't have
3360 * ORPHAN_OBJ flag */
3361 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3362 /* remove link to object from orphan index */
3363 LASSERT(handle != NULL);
3364 rc = mdd_orphan_delete(env, mdd_obj, handle);
3366 CERROR("%s: unable to delete "DFID" from orphan list: "
3367 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3368 PFID(mdd_object_fid(mdd_obj)), rc);
3369 /* If object was not deleted from orphan list, do not
3370 * destroy OSS objects, which will be done when next
3375 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3376 "list, OSS objects to be destroyed.\n",
3377 PFID(mdd_object_fid(mdd_obj)));
3380 rc = mdo_destroy(env, mdd_obj, handle);
3383 CERROR("%s: unable to delete "DFID" from orphan list: "
3384 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3385 PFID(mdd_object_fid(mdd_obj)), rc);
3390 mdd_write_unlock(env, mdd_obj);
3392 if (rc != 0 || blocked ||
3393 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3396 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3397 * or if CL_OPEN was recorded and it's last close by user.
3398 * Changelogs mask may change between open and close operations, but
3399 * this is not a big deal if we have a CL_CLOSE entry with no matching
3400 * CL_OPEN. Plus Changelogs mask may not change often.
3402 if (((!(mdd->mdd_cl.mc_mask & BIT(CL_OPEN)) &&
3403 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3404 MDS_OPEN_TRUNC))) ||
3405 ((mdd->mdd_cl.mc_mask & BIT(CL_OPEN)) && last_close_by_uid)) &&
3406 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3407 if (handle == NULL) {
3408 handle = mdd_trans_create(env, mdo2mdd(obj));
3410 GOTO(stop, rc = PTR_ERR(handle));
3412 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3413 NULL, NULL, handle);
3417 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3422 /* FYI, only the bottom 32 bits of open_flags are recorded */
3423 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3424 mdd_obj, handle, NULL);
3428 if (handle != NULL && !IS_ERR(handle))
3429 rc = mdd_trans_stop(env, mdd, rc, handle);
3435 * Permission check is done when open,
3436 * no need check again.
3438 static int mdd_readpage_sanity_check(const struct lu_env *env,
3439 struct mdd_object *obj)
3441 struct dt_object *next = mdd_object_child(obj);
3445 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3453 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3454 size_t nob, const struct dt_it_ops *iops,
3455 struct dt_it *it, __u32 attr, void *arg)
3457 struct lu_dirpage *dp = &lp->lp_dir;
3461 struct lu_dirent *ent;
3462 struct lu_dirent *last = NULL;
3466 if (nob < sizeof(*dp))
3469 memset(area, 0, sizeof (*dp));
3470 area += sizeof (*dp);
3471 nob -= sizeof (*dp);
3478 len = iops->key_size(env, it);
3480 /* IAM iterator can return record with zero len. */
3484 hash = iops->store(env, it);
3485 if (unlikely(first)) {
3487 dp->ldp_hash_start = cpu_to_le64(hash);
3490 /* calculate max space required for lu_dirent */
3491 recsize = lu_dirent_calc_size(len, attr);
3493 if (nob >= recsize) {
3494 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3495 if (result == -ESTALE)
3500 /* osd might not able to pack all attributes,
3501 * so recheck rec length */
3502 recsize = le16_to_cpu(ent->lde_reclen);
3504 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3505 fid_le_to_cpu(&fid, &ent->lde_fid);
3506 if (fid_is_dot_lustre(&fid))
3510 result = (last != NULL) ? 0 :-EINVAL;
3514 ent = (void *)ent + recsize;
3518 result = iops->next(env, it);
3519 if (result == -ESTALE)
3521 } while (result == 0);
3524 dp->ldp_hash_end = cpu_to_le64(hash);
3526 if (last->lde_hash == dp->ldp_hash_end)
3527 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3528 last->lde_reclen = 0; /* end mark */
3531 /* end of directory */
3532 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3533 else if (result < 0)
3534 CWARN("build page failed: %d!\n", result);
3538 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3539 const struct lu_rdpg *rdpg)
3541 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3545 if (mdd_object_exists(mdd_obj) == 0) {
3546 CERROR("%s: object "DFID" not found: rc = -2\n",
3547 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3551 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3552 rc = mdd_readpage_sanity_check(env, mdd_obj);
3554 GOTO(out_unlock, rc);
3556 if (mdd_is_dead_obj(mdd_obj)) {
3558 struct lu_dirpage *dp;
3561 * According to POSIX, please do not return any entry to client:
3562 * even dot and dotdot should not be returned.
3564 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3565 PFID(mdd_object_fid(mdd_obj)));
3567 if (rdpg->rp_count <= 0)
3568 GOTO(out_unlock, rc = -EFAULT);
3569 LASSERT(rdpg->rp_pages != NULL);
3571 pg = rdpg->rp_pages[0];
3572 dp = (struct lu_dirpage *)kmap(pg);
3573 memset(dp, 0 , sizeof(struct lu_dirpage));
3574 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3575 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3576 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3578 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3581 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3582 mdd_dir_page_build, NULL);
3584 struct lu_dirpage *dp;
3586 dp = kmap(rdpg->rp_pages[0]);
3587 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3590 * No pages were processed, mark this for first page
3593 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3594 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3595 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3597 kunmap(rdpg->rp_pages[0]);
3600 GOTO(out_unlock, rc);
3602 mdd_read_unlock(env, mdd_obj);
3606 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3608 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3610 if (mdd_object_exists(mdd_obj) == 0) {
3613 CERROR("%s: object "DFID" not found: rc = %d\n",
3614 mdd_obj_dev_name(mdd_obj),
3615 PFID(mdd_object_fid(mdd_obj)), rc);
3618 return dt_object_sync(env, mdd_object_child(mdd_obj),
3622 static int mdd_object_lock(const struct lu_env *env,
3623 struct md_object *obj,
3624 struct lustre_handle *lh,
3625 struct ldlm_enqueue_info *einfo,
3626 union ldlm_policy_data *policy)
3628 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3629 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3633 static int mdd_object_unlock(const struct lu_env *env,
3634 struct md_object *obj,
3635 struct ldlm_enqueue_info *einfo,
3636 union ldlm_policy_data *policy)
3638 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3639 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3642 const struct md_object_operations mdd_obj_ops = {
3643 .moo_permission = mdd_permission,
3644 .moo_attr_get = mdd_attr_get,
3645 .moo_attr_set = mdd_attr_set,
3646 .moo_xattr_get = mdd_xattr_get,
3647 .moo_xattr_set = mdd_xattr_set,
3648 .moo_xattr_list = mdd_xattr_list,
3649 .moo_invalidate = mdd_invalidate,
3650 .moo_xattr_del = mdd_xattr_del,
3651 .moo_swap_layouts = mdd_swap_layouts,
3652 .moo_open = mdd_open,
3653 .moo_close = mdd_close,
3654 .moo_readpage = mdd_readpage,
3655 .moo_readlink = mdd_readlink,
3656 .moo_changelog = mdd_changelog,
3657 .moo_object_sync = mdd_object_sync,
3658 .moo_object_lock = mdd_object_lock,
3659 .moo_object_unlock = mdd_object_unlock,
3660 .moo_layout_change = mdd_layout_change,