4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <lustre_mds.h>
52 #include "mdd_internal.h"
54 static const struct lu_object_operations mdd_lu_obj_ops;
56 struct mdd_object_user {
57 struct list_head mou_list; /**< linked off mod_users */
58 u64 mou_open_flags; /**< open mode by client */
59 __u64 mou_uidgid; /**< uid_gid on client */
60 int mou_opencount; /**< # opened */
61 ktime_t mou_deniednext; /**< time of next access denied
66 static int mdd_xattr_get(const struct lu_env *env,
67 struct md_object *obj, struct lu_buf *buf,
70 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
71 struct mdd_device *mdd,
72 enum changelog_rec_type type,
73 enum changelog_rec_flags clf_flags,
74 const struct lu_fid *fid,
75 const char *xattr_name,
76 struct thandle *handle);
78 static inline bool has_prefix(const char *str, const char *prefix);
81 static u32 flags_helper(u64 open_flags)
85 if (open_flags & MDS_FMODE_EXEC) {
86 open_mode = MDS_FMODE_EXEC;
88 if (open_flags & MDS_FMODE_READ)
89 open_mode = MDS_FMODE_READ;
91 (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
92 open_mode |= MDS_FMODE_WRITE;
98 /** Allocate/init a user and its sub-structures.
103 * \retval mou [OUT] success valid structure
106 static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags,
107 uid_t uid, gid_t gid)
109 struct mdd_object_user *mou;
113 OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem);
115 RETURN(ERR_PTR(-ENOMEM));
117 mou->mou_open_flags = open_flags;
118 mou->mou_uidgid = ((__u64)uid << 32) | gid;
119 mou->mou_opencount = 0;
120 mou->mou_deniednext = ktime_set(0, 0);
126 * Free a user and its sub-structures.
128 * \param mou [IN] user to be freed.
130 static void mdd_obj_user_free(struct mdd_object_user *mou)
132 OBD_SLAB_FREE_PTR(mou, mdd_object_kmem);
136 * Find if UID/GID already has this file open
138 * Caller should have write-locked \param mdd_obj.
139 * \param mdd_obj [IN] mdd_obj
140 * \param uid [IN] client uid
141 * \param gid [IN] client gid
142 * \retval user pointer or NULL if not found
145 struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj,
146 uid_t uid, gid_t gid,
149 struct mdd_object_user *mou;
154 uidgid = ((__u64)uid << 32) | gid;
155 list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) {
156 if (mou->mou_uidgid == uidgid &&
157 flags_helper(mou->mou_open_flags) ==
158 flags_helper(open_flags))
165 * Add a user to the list of openers for this file
167 * Caller should have write-locked \param mdd_obj.
168 * \param mdd_obj [IN] mdd_obj
169 * \param mou [IN] user
171 * \retval -ve failure
173 static int mdd_obj_user_add(struct mdd_object *mdd_obj,
174 struct mdd_object_user *mou,
177 struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj);
178 struct mdd_object_user *tmp;
179 __u32 uid = mou->mou_uidgid >> 32;
180 __u32 gid = mou->mou_uidgid & ((1UL << 32) - 1);
183 tmp = mdd_obj_user_find(mdd_obj, uid, gid, mou->mou_open_flags);
187 list_add_tail(&mou->mou_list, &mdd_obj->mod_users);
190 /* next 'access denied' notification cannot happen before
193 mou->mou_deniednext =
194 ktime_add(ktime_get(),
195 ktime_set(mdd->mdd_cl.mc_deniednext, 0));
197 mou->mou_opencount++;
202 * Remove UID from the list
204 * Caller should have write-locked \param mdd_obj.
205 * \param mdd_obj [IN] mdd_obj
206 * \param uid [IN] user
207 * \retval -ve failure
209 static int mdd_obj_user_remove(struct mdd_object *mdd_obj,
210 struct mdd_object_user *mou)
217 list_del_init(&mou->mou_list);
219 mdd_obj_user_free(mou);
223 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
228 if (mdd_object_exists(obj) == 0)
231 rc = mdo_attr_get(env, obj, la);
232 if (unlikely(rc != 0)) {
234 obj->mod_flags |= DEAD_OBJ;
238 if (la->la_valid & LA_FLAGS && la->la_flags & LUSTRE_ORPHAN_FL)
239 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
244 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
246 return lu_env_info(env, &mdd_thread_key);
249 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
253 buf = &mdd_env_info(env)->mti_buf[0];
259 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
260 const void *area, ssize_t len)
264 buf = &mdd_env_info(env)->mti_buf[0];
265 buf->lb_buf = (void *)area;
270 struct lu_object *mdd_object_alloc(const struct lu_env *env,
271 const struct lu_object_header *hdr,
274 struct mdd_object *mdd_obj;
277 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
281 o = mdd2lu_obj(mdd_obj);
282 lu_object_init(o, NULL, d);
283 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
284 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
285 mdd_obj->mod_count = 0;
286 o->lo_ops = &mdd_lu_obj_ops;
287 INIT_LIST_HEAD(&mdd_obj->mod_users);
292 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
293 const struct lu_object_conf *unused)
295 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
296 struct mdd_object *mdd_obj = lu2mdd_obj(o);
297 struct lu_object *below;
298 struct lu_device *under;
301 mdd_obj->mod_cltime = ktime_set(0, 0);
302 under = &d->mdd_child->dd_lu_dev;
303 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
305 RETURN(PTR_ERR(below));
307 lu_object_add(o, below);
312 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
316 if (lu_object_exists(o)) {
317 struct mdd_object *mdd_obj = lu2mdd_obj(o);
318 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
320 rc = mdd_la_get(env, mdd_obj, attr);
326 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
328 struct mdd_object *mdd = lu2mdd_obj(o);
329 struct mdd_object_user *mou, *tmp2;
332 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
333 list_del(&mou->mou_list);
334 mdd_obj_user_free(mou);
338 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
341 static int mdd_object_print(const struct lu_env *env, void *cookie,
342 lu_printer_t p, const struct lu_object *o)
344 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
346 return (*p)(env, cookie,
347 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
348 mdd, mdd->mod_count, mdd->mod_valid,
349 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
352 static const struct lu_object_operations mdd_lu_obj_ops = {
353 .loo_object_init = mdd_object_init,
354 .loo_object_start = mdd_object_start,
355 .loo_object_free = mdd_object_free,
356 .loo_object_print = mdd_object_print,
359 struct mdd_object *mdd_object_find(const struct lu_env *env,
360 struct mdd_device *d,
361 const struct lu_fid *f)
363 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
367 * No permission check is needed.
369 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
372 struct mdd_object *mdd_obj = md2mdd_obj(obj);
377 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
378 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
379 ma->ma_attr.la_nlink = 0;
385 * No permission check is needed.
387 static int mdd_xattr_get(const struct lu_env *env,
388 struct md_object *obj, struct lu_buf *buf,
391 struct mdd_object *mdd_obj = md2mdd_obj(obj);
392 struct mdd_device *mdd;
397 if (mdd_object_exists(mdd_obj) == 0) {
398 CERROR("%s: object "DFID" not found: rc = -2\n",
399 mdd_obj_dev_name(mdd_obj),
400 PFID(mdd_object_fid(mdd_obj)));
404 /* If the object has been destroyed, then do not get LMVEA, because
405 * it needs to load stripes from the iteration of the master object,
406 * and it will cause problem if master object has been destroyed, see
408 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
409 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
410 strcmp(name, XATTR_NAME_LMV) == 0))
413 /* If the object has been delete from the namespace, then
414 * get linkEA should return -ENOENT as well */
415 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
416 strcmp(name, XATTR_NAME_LINK) == 0))
419 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
420 rc = mdo_xattr_get(env, mdd_obj, buf, name);
421 mdd_read_unlock(env, mdd_obj);
425 /* record only getting user xattrs and acls */
426 if (rc >= 0 && buf->lb_buf &&
427 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
428 (has_prefix(name, XATTR_USER_PREFIX) ||
429 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
430 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
431 struct thandle *handle;
434 LASSERT(mdo2fid(mdd_obj) != NULL);
436 handle = mdd_trans_create(env, mdd);
438 RETURN(PTR_ERR(handle));
440 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
445 rc2 = mdd_trans_start(env, mdd, handle);
449 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
450 mdo2fid(mdd_obj), name,
454 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
463 * Permission check is done when open,
464 * no need check again.
466 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
469 struct mdd_object *mdd_obj = md2mdd_obj(obj);
470 struct dt_object *next;
475 if (mdd_object_exists(mdd_obj) == 0) {
476 CERROR("%s: object "DFID" not found: rc = -2\n",
477 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
481 next = mdd_object_child(mdd_obj);
482 LASSERT(next != NULL);
483 LASSERT(next->do_body_ops != NULL);
484 LASSERT(next->do_body_ops->dbo_read != NULL);
485 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
486 rc = dt_read(env, next, buf, &pos);
487 mdd_read_unlock(env, mdd_obj);
492 * No permission check is needed.
494 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
497 struct mdd_object *mdd_obj = md2mdd_obj(obj);
502 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
503 rc = mdo_xattr_list(env, mdd_obj, buf);
504 mdd_read_unlock(env, mdd_obj);
506 /* If the buffer is NULL then we are only here to get the
507 * length of the xattr name list. */
508 if (rc < 0 || buf->lb_buf == NULL)
512 * Filter out XATTR_NAME_LINK if this is an orphan object. See
515 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
516 char *end = (char *)buf->lb_buf + rc;
517 char *p = buf->lb_buf;
520 char *next = p + strlen(p) + 1;
522 if (strcmp(p, XATTR_NAME_LINK) == 0) {
524 memmove(p, next, end - next);
526 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
527 " of orphan "DFID"\n",
528 PFID(mdd_object_fid(mdd_obj)));
539 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
541 return mdo_invalidate(env, md2mdd_obj(obj));
544 int mdd_declare_create_object_internal(const struct lu_env *env,
545 struct mdd_object *p,
546 struct mdd_object *c,
547 struct lu_attr *attr,
548 struct thandle *handle,
549 const struct md_op_spec *spec,
550 struct dt_allocation_hint *hint)
552 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
553 const struct dt_index_features *feat = spec->sp_feat;
557 if (feat != &dt_directory_features && feat != NULL) {
558 dof->dof_type = DFT_INDEX;
559 dof->u.dof_idx.di_feat = feat;
561 dof->dof_type = dt_mode_to_dft(attr->la_mode);
562 if (dof->dof_type == DFT_REGULAR) {
563 dof->u.dof_reg.striped =
564 md_should_create(spec->sp_cr_flags);
565 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
566 dof->u.dof_reg.striped = 0;
567 /* is this replay? */
569 dof->u.dof_reg.striped = 0;
573 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
578 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
579 struct mdd_object *c, struct lu_attr *attr,
580 struct thandle *handle,
581 const struct md_op_spec *spec,
582 struct dt_allocation_hint *hint)
584 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
588 LASSERT(!mdd_object_exists(c));
590 rc = mdo_create_object(env, c, attr, hint, dof, handle);
595 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
596 const struct lu_attr *attr, struct thandle *handle,
602 rc = mdo_attr_set(env, obj, attr, handle);
603 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
604 if (!rc && (attr->la_valid & LA_MODE) && needacl)
605 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
610 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
611 const struct lu_attr *oattr, struct lu_attr *attr,
612 struct thandle *handle)
617 LASSERT(attr->la_valid & LA_CTIME);
618 LASSERT(oattr != NULL);
620 /* Make sure the ctime is increased only, however, it's not strictly
621 * reliable at here because there is not guarantee to hold lock on
622 * object, so we just bypass some unnecessary cmtime setting first
623 * and OSD has to check it again. */
624 if (attr->la_ctime < oattr->la_ctime)
625 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
626 else if (attr->la_valid == LA_CTIME &&
627 attr->la_ctime == oattr->la_ctime)
628 attr->la_valid &= ~LA_CTIME;
630 if (attr->la_valid != 0)
631 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
636 static bool is_project_state_change(const struct lu_attr *oattr,
639 if (la->la_valid & LA_PROJID &&
640 oattr->la_projid != la->la_projid)
643 if ((la->la_valid & LA_FLAGS) &&
644 (la->la_flags & LUSTRE_PROJINHERIT_FL) !=
645 (oattr->la_flags & LUSTRE_PROJINHERIT_FL))
652 * This gives the same functionality as the code between
653 * sys_chmod and inode_setattr
654 * chown_common and inode_setattr
655 * utimes and inode_setattr
656 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
658 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
659 const struct lu_attr *oattr, struct lu_attr *la,
660 const struct md_attr *ma)
664 const unsigned long flags = ma->ma_attr_flags;
671 /* Do not permit change file type */
672 if (la->la_valid & LA_TYPE)
675 /* They should not be processed by setattr */
676 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
679 LASSERT(oattr != NULL);
681 uc = lu_ucred_check(env);
685 if (is_project_state_change(oattr, la)) {
686 if (!md_capable(uc, CFS_CAP_SYS_RESOURCE) &&
687 !lustre_in_group_p(uc, ma->ma_enable_chprojid_gid) &&
688 !(ma->ma_enable_chprojid_gid == -1 &&
689 mdd_permission_internal(env, obj, oattr, MAY_WRITE)))
693 if (la->la_valid == LA_CTIME) {
694 if (!(flags & MDS_PERM_BYPASS))
695 /* This is only for set ctime when rename's source is
697 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
699 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
700 la->la_valid &= ~LA_CTIME;
704 if (flags & MDS_CLOSE_UPDATE_TIMES &&
705 la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
706 /* This is an atime/mtime/ctime attribute update for
709 if (la->la_valid & LA_ATIME &&
710 la->la_atime <= (oattr->la_atime +
711 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
712 la->la_valid &= ~LA_ATIME;
713 if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime)
714 la->la_valid &= ~LA_CTIME;
715 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
716 la->la_valid &= ~LA_MTIME;
720 /* Check if flags change. */
721 if (la->la_valid & LA_FLAGS) {
722 unsigned int oldflags = oattr->la_flags &
723 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
724 unsigned int newflags = la->la_flags &
725 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
727 if ((uc->uc_fsuid != oattr->la_uid) &&
728 !md_capable(uc, CFS_CAP_FOWNER))
731 /* The IMMUTABLE and APPEND_ONLY flags can
732 * only be changed by the relevant capability. */
733 if ((oldflags ^ newflags) &&
734 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
737 if (!S_ISDIR(oattr->la_mode))
738 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
741 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
742 (la->la_valid & ~LA_FLAGS) &&
743 !(flags & MDS_PERM_BYPASS))
746 /* Check for setting the obj time. */
747 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
748 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
749 if ((uc->uc_fsuid != oattr->la_uid) &&
750 !md_capable(uc, CFS_CAP_FOWNER)) {
751 rc = mdd_permission_internal(env, obj, oattr,
758 if (la->la_valid & LA_KILL_SUID) {
759 la->la_valid &= ~LA_KILL_SUID;
760 if ((oattr->la_mode & S_ISUID) &&
761 !(la->la_valid & LA_MODE)) {
762 la->la_mode = oattr->la_mode;
763 la->la_valid |= LA_MODE;
765 la->la_mode &= ~S_ISUID;
768 if (la->la_valid & LA_KILL_SGID) {
769 la->la_valid &= ~LA_KILL_SGID;
770 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
771 (S_ISGID | S_IXGRP)) &&
772 !(la->la_valid & LA_MODE)) {
773 la->la_mode = oattr->la_mode;
774 la->la_valid |= LA_MODE;
776 la->la_mode &= ~S_ISGID;
779 /* Make sure a caller can chmod. */
780 if (la->la_valid & LA_MODE) {
781 if (!(flags & MDS_PERM_BYPASS) &&
782 (uc->uc_fsuid != oattr->la_uid) &&
783 !md_capable(uc, CFS_CAP_FOWNER))
786 if (la->la_mode == (umode_t) -1)
787 la->la_mode = oattr->la_mode;
789 la->la_mode = (la->la_mode & S_IALLUGO) |
790 (oattr->la_mode & ~S_IALLUGO);
792 /* Also check the setgid bit! */
793 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
794 la->la_gid : oattr->la_gid) &&
795 !md_capable(uc, CFS_CAP_FSETID))
796 la->la_mode &= ~S_ISGID;
798 la->la_mode = oattr->la_mode;
801 /* Make sure a caller can chown. */
802 if (la->la_valid & LA_UID) {
803 if (la->la_uid == (uid_t) -1)
804 la->la_uid = oattr->la_uid;
805 if (((uc->uc_fsuid != oattr->la_uid) ||
806 (la->la_uid != oattr->la_uid)) &&
807 !md_capable(uc, CFS_CAP_CHOWN))
810 /* If the user or group of a non-directory has been
811 * changed by a non-root user, remove the setuid bit.
812 * 19981026 David C Niemi <niemi@tux.org>
814 * Changed this to apply to all users, including root,
815 * to avoid some races. This is the behavior we had in
816 * 2.0. The check for non-root was definitely wrong
817 * for 2.2 anyway, as it should have been using
818 * CAP_FSETID rather than fsuid -- 19990830 SD. */
819 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
820 !S_ISDIR(oattr->la_mode)) {
821 la->la_mode &= ~S_ISUID;
822 la->la_valid |= LA_MODE;
826 /* Make sure caller can chgrp. */
827 if (la->la_valid & LA_GID) {
828 if (la->la_gid == (gid_t) -1)
829 la->la_gid = oattr->la_gid;
830 if (((uc->uc_fsuid != oattr->la_uid) ||
831 ((la->la_gid != oattr->la_gid) &&
832 !lustre_in_group_p(uc, la->la_gid))) &&
833 !md_capable(uc, CFS_CAP_CHOWN))
836 /* Likewise, if the user or group of a non-directory
837 * has been changed by a non-root user, remove the
838 * setgid bit UNLESS there is no group execute bit
839 * (this would be a file marked for mandatory
840 * locking). 19981026 David C Niemi <niemi@tux.org>
842 * Removed the fsuid check (see the comment above) --
844 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
845 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
846 la->la_mode &= ~S_ISGID;
847 la->la_valid |= LA_MODE;
851 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
852 if (!((flags & MDS_OWNEROVERRIDE) &&
853 (uc->uc_fsuid == oattr->la_uid)) &&
854 !(flags & MDS_PERM_BYPASS)) {
855 rc = mdd_permission_internal(env, obj, oattr,
862 if (la->la_valid & LA_CTIME) {
863 /* The pure setattr, it has the priority over what is
864 * already set, do not drop it if ctime is equal. */
865 if (la->la_ctime < oattr->la_ctime)
866 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
872 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
873 struct mdd_device *mdd,
874 enum changelog_rec_type type,
875 enum changelog_rec_flags clf_flags,
876 const struct lu_fid *fid,
877 const char *xattr_name,
878 struct thandle *handle)
880 const struct lu_ucred *uc = lu_ucred(env);
881 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
882 struct llog_changelog_rec *rec;
887 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
890 if (uc->uc_jobid[0] != '\0')
891 clf_flags |= CLF_JOBID;
892 xflags |= CLFE_UIDGID;
895 if (type == CL_OPEN || type == CL_DN_OPEN)
897 if (type == CL_SETXATTR || type == CL_GETXATTR)
898 xflags |= CLFE_XATTR;
900 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
901 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
902 xflags & CLFE_SUPPORTED));
903 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
904 if (buf->lb_buf == NULL)
908 rec->cr_hdr.lrh_len = reclen;
909 rec->cr.cr_flags = clf_flags;
910 rec->cr.cr_type = (__u32)type;
911 rec->cr.cr_tfid = *fid;
912 rec->cr.cr_namelen = 0;
914 if (clf_flags & CLF_JOBID)
915 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
917 if (clf_flags & CLF_EXTRA_FLAGS) {
918 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
919 if (xflags & CLFE_UIDGID)
920 mdd_changelog_rec_extra_uidgid(&rec->cr,
921 uc->uc_uid, uc->uc_gid);
922 if (xflags & CLFE_NID)
923 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
924 if (xflags & CLFE_OPEN)
925 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
926 if (xflags & CLFE_XATTR) {
927 if (xattr_name == NULL)
929 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
933 rc = mdd_changelog_store(env, mdd, rec, handle);
938 /** Store a data change changelog record
939 * If this fails, we must fail the whole transaction; we don't
940 * want the change to commit without the log entry.
941 * \param mdd_obj - mdd_object of change
942 * \param handle - transaction handle
944 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
945 enum changelog_rec_type type,
946 enum changelog_rec_flags clf_flags,
947 struct mdd_object *mdd_obj, struct thandle *handle)
951 LASSERT(mdd_obj != NULL);
952 LASSERT(handle != NULL);
954 if (!mdd_changelog_enabled(env, mdd, type))
957 if (mdd_is_volatile_obj(mdd_obj))
960 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
961 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
962 /* Don't need multiple updates in this log */
963 /* Don't check under lock - no big deal if we get an extra
968 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
969 mdo2fid(mdd_obj), NULL, handle);
971 mdd_obj->mod_cltime = ktime_get();
976 int mdd_changelog_data_store_xattr(const struct lu_env *env,
977 struct mdd_device *mdd,
978 enum changelog_rec_type type,
979 enum changelog_rec_flags clf_flags,
980 struct mdd_object *mdd_obj,
981 const char *xattr_name,
982 struct thandle *handle)
986 LASSERT(mdd_obj != NULL);
987 LASSERT(handle != NULL);
989 if (!mdd_changelog_enabled(env, mdd, type))
992 if (mdd_is_volatile_obj(mdd_obj))
995 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
996 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
997 /* Don't need multiple updates in this log */
998 /* Don't check under lock - no big deal if we get an extra
1004 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1005 mdo2fid(mdd_obj), xattr_name,
1008 mdd_obj->mod_cltime = ktime_get();
1013 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
1014 * except for open flags have a dedicated record to store 32 bits of flags */
1015 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1016 enum changelog_rec_flags clf_flags,
1017 struct md_device *m, const struct lu_fid *fid)
1019 struct thandle *handle;
1020 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1024 LASSERT(fid != NULL);
1026 /* We'll check this again below, but we check now before we
1027 * start a transaction. */
1028 if (!mdd_changelog_enabled(env, mdd, type))
1031 handle = mdd_trans_create(env, mdd);
1033 RETURN(PTR_ERR(handle));
1035 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1039 rc = mdd_trans_start(env, mdd, handle);
1043 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1047 rc = mdd_trans_stop(env, mdd, rc, handle);
1053 * Save LMA extended attributes with data from \a ma.
1055 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1056 * not, LMA EA will be first read from disk, modified and write back.
1059 /* Precedence for choosing record type when multiple
1060 * attributes change: setattr > mtime > ctime > atime
1061 * (ctime changes when mtime does, plus chmod/chown.
1062 * atime and ctime are independent.) */
1063 static int mdd_attr_set_changelog(const struct lu_env *env,
1064 struct md_object *obj, struct thandle *handle,
1067 struct mdd_device *mdd = mdo2mdd(obj);
1070 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
1071 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
1072 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
1073 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
1074 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
1075 bits = bits & mdd->mdd_cl.mc_mask;
1076 /* This is an implementation limit rather than a protocol limit */
1077 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1081 /* The record type is the lowest non-masked set bit */
1084 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1085 return mdd_changelog_data_store(env, mdd, type, valid,
1086 md2mdd_obj(obj), handle);
1089 static int mdd_declare_attr_set(const struct lu_env *env,
1090 struct mdd_device *mdd,
1091 struct mdd_object *obj,
1092 const struct lu_attr *attr,
1093 struct thandle *handle)
1097 rc = mdo_declare_attr_set(env, obj, attr, handle);
1101 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1102 if (attr->la_valid & LA_MODE) {
1103 mdd_read_lock(env, obj, DT_TGT_CHILD);
1104 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1105 XATTR_NAME_ACL_ACCESS);
1106 mdd_read_unlock(env, obj);
1107 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1113 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1114 rc = mdo_declare_xattr_set(env, obj, buf,
1115 XATTR_NAME_ACL_ACCESS, 0,
1123 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1132 * permission changes may require sync operation, to mitigate performance
1133 * impact, only do this for dir and when permission is reduced.
1135 * For regular files, version is updated with permission change (see VBR), async
1136 * permission won't cause any issue, while missing permission change on
1137 * directory may affect accessibility of other objects after recovery.
1139 static inline bool permission_needs_sync(const struct lu_attr *old,
1140 const struct lu_attr *new)
1142 if (!S_ISDIR(old->la_mode))
1145 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1148 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1151 if (new->la_valid & LA_MODE) {
1152 /* turned on sticky bit */
1153 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1156 /* set-GID has no impact on what is allowed, not checked */
1158 /* turned off setuid bit, or one of rwx for someone */
1159 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1160 (old->la_mode & (0777 | S_ISUID)))
1167 static inline __u64 mdd_lmm_dom_size(void *buf)
1169 struct lov_mds_md *lmm = buf;
1170 struct lov_comp_md_v1 *comp_v1;
1171 struct lov_mds_md *v1;
1177 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1180 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1181 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1182 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1184 /* DoM entry is the first entry always */
1185 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1186 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1191 /* set attr and LOV EA at once, return updated attr */
1192 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1193 const struct md_attr *ma)
1195 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1196 struct mdd_device *mdd = mdo2mdd(obj);
1197 struct thandle *handle = NULL;
1198 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1199 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1200 const struct lu_attr *la = &ma->ma_attr;
1201 struct lu_ucred *uc;
1202 bool chrgrp_by_unprivileged_user = false;
1206 /* we do not use ->attr_set() for LOV/HSM EA any more */
1207 LASSERT((ma->ma_valid & MA_LOV) == 0);
1208 LASSERT((ma->ma_valid & MA_HSM) == 0);
1210 rc = mdd_la_get(env, mdd_obj, attr);
1214 *la_copy = ma->ma_attr;
1215 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1219 /* no need to setattr anymore */
1220 if (la_copy->la_valid == 0) {
1221 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
1222 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
1223 PFID(mdo2fid(mdd_obj)), la->la_valid);
1228 /* If an unprivileged user changes group of some file,
1229 * the setattr operation will be processed synchronously to
1230 * honor the quota limit of the corresponding group. see LU-5152 */
1231 uc = lu_ucred_check(env);
1232 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1233 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1234 /* LU-10048: disable synchronous chgrp operation for it will
1235 * cause deadlock between MDT and OST.
1236 la_copy->la_valid |= LA_FLAGS;
1237 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1239 chrgrp_by_unprivileged_user = true;
1241 /* Flush the possible existing client setattr requests to OSTs
1242 * to keep the order with the current setattr operation that
1243 * will be sent directly to OSTs. see LU-5152 */
1244 /* LU-11303 disable sync as this is too heavyweight.
1245 * This should be replaced with a sync only for the object
1246 * being modified here, not the whole filesystem.
1247 rc = dt_sync(env, mdd->mdd_child);
1253 handle = mdd_trans_create(env, mdd);
1254 if (IS_ERR(handle)) {
1255 rc = PTR_ERR(handle);
1261 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1265 rc = mdd_trans_start(env, mdd, handle);
1269 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1270 permission_needs_sync(attr, la))
1271 handle->th_sync = 1;
1273 if (la->la_valid & (LA_MTIME | LA_CTIME))
1274 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1275 la->la_mtime, la->la_ctime);
1277 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1279 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1280 * otherwise following truncate will do nothing and truncated
1281 * data may be read again. This is a quick fix until LU-11033
1284 if (la_copy->la_valid & LA_SIZE) {
1285 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1287 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1291 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1292 la_copy->la_valid &= ~LA_SIZE;
1293 lu_buf_free(lov_buf);
1297 if (la_copy->la_valid) {
1298 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1300 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1301 /* rollback to the original gid */
1302 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1303 la_copy->la_gid = attr->la_gid;
1304 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1307 mdd_write_unlock(env, mdd_obj);
1311 rc = mdd_attr_set_changelog(env, obj, handle,
1315 rc = mdd_trans_stop(env, mdd, rc, handle);
1320 static int mdd_xattr_sanity_check(const struct lu_env *env,
1321 struct mdd_object *obj,
1322 const struct lu_attr *attr,
1325 struct lu_ucred *uc = lu_ucred_assert(env);
1328 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1331 if (strncmp(XATTR_USER_PREFIX, name,
1332 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1333 /* For sticky directories, only the owner and privileged user
1334 * can write attributes. */
1335 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1336 (uc->uc_fsuid != attr->la_uid) &&
1337 !md_capable(uc, CFS_CAP_FOWNER))
1339 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1340 (uc->uc_fsuid != attr->la_uid) &&
1341 !md_capable(uc, CFS_CAP_FOWNER)) {
1349 * Check if a string begins with a given prefix.
1351 * \param str String to check
1352 * \param prefix Substring to check at the beginning of \a str
1353 * \return true/false whether the condition is verified.
1355 static inline bool has_prefix(const char *str, const char *prefix)
1357 return strncmp(prefix, str, strlen(prefix)) == 0;
1361 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1363 * \param[in] xattr_name Full extended attribute name.
1365 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1367 static enum changelog_rec_type
1368 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1369 const char *xattr_name)
1371 /* Layout changes systematically recorded */
1372 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1373 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1374 strlen(XATTR_LUSTRE_LOV)) == 0)
1377 /* HSM information changes systematically recorded */
1378 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1381 /* Avoid logging SOM xattr for every file */
1382 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1385 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1386 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1387 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1388 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1389 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1395 static int mdd_declare_xattr_set(const struct lu_env *env,
1396 struct mdd_device *mdd,
1397 struct mdd_object *obj,
1398 const struct lu_buf *buf,
1400 int fl, struct thandle *handle)
1402 enum changelog_rec_type type;
1405 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1409 type = mdd_xattr_changelog_type(env, mdd, name);
1411 return 0; /* no changelog to store */
1413 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1417 * Compare current and future data of HSM EA and add a changelog if needed.
1419 * Caller should have write-locked \param obj.
1421 * \param buf - Future HSM EA content.
1422 * \retval 0 if no changelog is needed or changelog was added properly.
1423 * \retval -ve errno if there was a problem
1425 static int mdd_hsm_update_locked(const struct lu_env *env,
1426 struct md_object *obj,
1427 const struct lu_buf *buf,
1428 struct thandle *handle,
1429 enum changelog_rec_flags *clf_flags)
1431 struct mdd_thread_info *info = mdd_env_info(env);
1432 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1433 struct lu_buf *current_buf;
1434 struct md_hsm *current_mh;
1435 struct md_hsm *new_mh;
1439 OBD_ALLOC_PTR(current_mh);
1440 if (current_mh == NULL)
1443 /* Read HSM attrs from disk */
1444 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1445 MIN(mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1447 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1448 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1449 if (rc < 0 && rc != -ENODATA)
1451 else if (rc == -ENODATA)
1452 current_mh->mh_flags = 0;
1454 /* Map future HSM xattr */
1455 OBD_ALLOC_PTR(new_mh);
1457 GOTO(free, rc = -ENOMEM);
1458 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1462 /* Flags differ, set flags for the changelog that will be added */
1463 if (current_mh->mh_flags != new_mh->mh_flags) {
1464 hsm_set_cl_event(clf_flags, HE_STATE);
1465 if (new_mh->mh_flags & HS_DIRTY)
1466 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1469 OBD_FREE_PTR(new_mh);
1472 OBD_FREE_PTR(current_mh);
1476 static int mdd_object_pfid_replace(const struct lu_env *env,
1477 struct mdd_object *o)
1479 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1480 struct thandle *handle;
1483 handle = mdd_trans_create(env, mdd);
1485 RETURN(PTR_ERR(handle));
1487 handle->th_complex = 1;
1489 /* it doesn't need to track the PFID update via llog, because LFSCK
1490 * will repair it even it goes wrong */
1491 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1496 rc = mdd_trans_start(env, mdd, handle);
1500 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1505 mdd_trans_stop(env, mdd, rc, handle);
1510 static int mdd_declare_xattr_del(const struct lu_env *env,
1511 struct mdd_device *mdd,
1512 struct mdd_object *obj,
1514 struct thandle *handle);
1516 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1519 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1520 struct md_object *md_vic)
1522 struct mdd_device *mdd = mdo2mdd(md_obj);
1523 struct mdd_object *obj = md2mdd_obj(md_obj);
1524 struct mdd_object *vic = md2mdd_obj(md_vic);
1525 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1526 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1527 struct lov_mds_md *lmm;
1528 struct thandle *handle;
1532 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1533 if (rc == 0) /* same fid */
1536 handle = mdd_trans_create(env, mdd);
1538 RETURN(PTR_ERR(handle));
1541 mdd_write_lock(env, obj, DT_TGT_CHILD);
1542 mdd_write_lock(env, vic, DT_TGT_CHILD);
1544 mdd_write_lock(env, vic, DT_TGT_CHILD);
1545 mdd_write_lock(env, obj, DT_TGT_CHILD);
1548 /* get EA of victim file */
1549 memset(buf_vic, 0, sizeof(*buf_vic));
1550 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1557 /* parse the layout of victim file */
1558 lmm = buf_vic->lb_buf;
1559 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1560 GOTO(out, rc = -EINVAL);
1562 /* save EA of target file for restore */
1563 memset(buf, 0, sizeof(*buf));
1564 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1568 /* Get rid of the layout from victim object */
1569 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1573 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1574 LU_XATTR_MERGE, handle);
1578 rc = mdd_trans_start(env, mdd, handle);
1582 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1587 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1589 GOTO(out_restore, rc);
1591 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1592 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1597 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1598 LU_XATTR_REPLACE, handle);
1600 CERROR("%s: failed to rollback of layout of: "DFID
1601 ": %d, file state unknown\n",
1602 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1606 mdd_trans_stop(env, mdd, rc, handle);
1607 mdd_write_unlock(env, obj);
1608 mdd_write_unlock(env, vic);
1610 lu_buf_free(buf_vic);
1613 (void) mdd_object_pfid_replace(env, obj);
1619 * Extract the mirror with specified mirror id, and store the splitted
1620 * mirror layout to @buf.
1622 * \param[in] comp_v1 mirrored layout
1623 * \param[in] mirror_id the mirror with mirror_id to be extracted
1624 * \param[out] buf store the layout excluding the extracted mirror,
1625 * caller free the buffer we allocated in this function
1626 * \param[out] buf_vic store the extracted layout, caller free the buffer
1627 * we allocated in this function
1629 * \retval 0 on success; < 0 if error happens
1631 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1632 struct lu_buf *buf, struct lu_buf *buf_vic)
1634 struct lov_comp_md_v1 *comp_rem;
1635 struct lov_comp_md_v1 *comp_vic;
1636 struct lov_comp_md_entry_v1 *entry;
1637 struct lov_comp_md_entry_v1 *entry_rem;
1638 struct lov_comp_md_entry_v1 *entry_vic;
1640 __u16 comp_cnt, count = 0;
1641 int lmm_size, lmm_size_vic = 0;
1643 int offset, offset_rem, offset_vic;
1645 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1646 /* comp_v1 should contains more than 1 mirror */
1647 if (mirror_cnt <= 1)
1649 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1650 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1652 for (i = 0; i < comp_cnt; i++) {
1653 entry = &comp_v1->lcm_entries[i];
1654 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1656 lmm_size_vic += sizeof(*entry);
1657 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1658 } else if (count > 0) {
1659 /* find the specified mirror */
1667 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1671 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1672 if (!buf_vic->lb_buf) {
1677 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1678 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1680 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1681 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1682 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1683 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1684 if (!comp_rem->lcm_mirror_count)
1685 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1687 memset(comp_vic, 0, sizeof(*comp_v1));
1688 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1689 comp_vic->lcm_mirror_count = 0;
1690 comp_vic->lcm_entry_count = cpu_to_le32(count);
1691 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1692 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1693 comp_vic->lcm_layout_gen = 0;
1695 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1696 offset_rem = sizeof(*comp_rem) +
1697 sizeof(*entry_rem) * (comp_cnt - count);
1698 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1699 for (i = j = k = 0; i < comp_cnt; i++) {
1700 struct lov_mds_md *lmm, *lmm_dst;
1703 entry = &comp_v1->lcm_entries[i];
1704 entry_vic = &comp_vic->lcm_entries[j];
1705 entry_rem = &comp_rem->lcm_entries[k];
1707 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1710 /* copy component entry */
1712 memcpy(entry_vic, entry, sizeof(*entry));
1713 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1714 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1717 memcpy(entry_rem, entry, sizeof(*entry));
1718 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1722 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1724 lmm_dst = (struct lov_mds_md *)
1725 ((char *)comp_vic + offset_vic);
1727 lmm_dst = (struct lov_mds_md *)
1728 ((char *)comp_rem + offset_rem);
1730 /* copy component entry blob */
1731 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1733 /* blob offset advance */
1734 offset += le32_to_cpu(entry->lcme_size);
1736 offset_vic += le32_to_cpu(entry->lcme_size);
1738 offset_rem += le32_to_cpu(entry->lcme_size);
1744 static int mdd_dom_data_truncate(const struct lu_env *env,
1745 struct mdd_device *mdd, struct mdd_object *mo);
1747 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1748 struct md_rejig_data *mrd)
1750 struct mdd_device *mdd = mdo2mdd(md_obj);
1751 struct mdd_object *obj = md2mdd_obj(md_obj);
1752 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1753 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1754 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1755 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1756 struct lov_comp_md_v1 *lcm;
1757 struct thandle *handle;
1759 bool dom_stripe = false;
1763 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1764 if (rc == 0) /* same fid */
1767 handle = mdd_trans_create(env, mdd);
1769 RETURN(PTR_ERR(handle));
1772 mdd_write_lock(env, obj, DT_TGT_CHILD);
1773 mdd_write_lock(env, vic, DT_TGT_CHILD);
1775 mdd_write_lock(env, vic, DT_TGT_CHILD);
1776 mdd_write_lock(env, obj, DT_TGT_CHILD);
1779 /* get EA of mirrored file */
1780 memset(buf_save, 0, sizeof(*buf));
1781 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1785 lcm = buf_save->lb_buf;
1786 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1787 GOTO(out, rc = -EINVAL);
1790 * Extract the mirror with specified mirror id, and store the splitted
1791 * mirror layout to the victim file.
1793 memset(buf, 0, sizeof(*buf));
1794 memset(buf_vic, 0, sizeof(*buf_vic));
1795 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1799 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1801 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1802 LU_XATTR_SPLIT, handle);
1805 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1806 LU_XATTR_SPLIT, handle);
1810 rc = mdd_trans_start(env, mdd, handle);
1814 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1819 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1822 GOTO(out_restore, rc);
1824 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1828 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1835 /* restore obj's layout */
1836 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1837 LU_XATTR_REPLACE, handle);
1839 CERROR("%s: failed to rollback of layout of: "DFID
1840 ": %d, file state unkonwn.\n",
1841 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1844 rc = mdd_trans_stop(env, mdd, rc, handle);
1846 /* Truncate local DOM data if all went well */
1847 if (!rc && dom_stripe)
1848 mdd_dom_data_truncate(env, mdd, obj);
1850 mdd_write_unlock(env, obj);
1851 mdd_write_unlock(env, vic);
1852 lu_buf_free(buf_save);
1854 lu_buf_free(buf_vic);
1857 (void) mdd_object_pfid_replace(env, obj);
1862 static int mdd_layout_merge_allowed(const struct lu_env *env,
1863 struct md_object *target,
1864 struct md_object *victim)
1866 struct mdd_object *o1 = md2mdd_obj(target);
1868 /* cannot extend directory's LOVEA */
1869 if (S_ISDIR(mdd_object_type(o1))) {
1870 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1871 mdd_obj_dev_name(o1));
1879 * The caller should guarantee to update the object ctime
1880 * after xattr_set if needed.
1882 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1883 const struct lu_buf *buf, const char *name,
1886 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1887 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1888 struct mdd_device *mdd = mdo2mdd(obj);
1889 struct thandle *handle;
1890 enum changelog_rec_type cl_type;
1891 enum changelog_rec_flags clf_flags = 0;
1895 rc = mdd_la_get(env, mdd_obj, attr);
1899 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1903 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1904 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1905 struct md_rejig_data *mrd = buf->lb_buf;
1906 struct md_object *victim = mrd->mrd_obj;
1908 if (buf->lb_len != sizeof(*mrd))
1911 rc = mdd_layout_merge_allowed(env, obj, victim);
1915 if (fl == LU_XATTR_MERGE)
1916 /* merge layout of victim as a mirror of obj's. */
1917 rc = mdd_xattr_merge(env, obj, victim);
1919 rc = mdd_xattr_split(env, obj, mrd);
1923 if (strcmp(name, XATTR_NAME_LMV) == 0) {
1924 rc = mdd_dir_layout_shrink(env, obj, buf);
1928 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1929 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1930 struct posix_acl *acl;
1932 /* user may set empty ACL, which should be treated as removing
1934 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1937 RETURN(PTR_ERR(acl));
1939 rc = mdd_xattr_del(env, obj, name);
1942 posix_acl_release(acl);
1945 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1946 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1950 handle = mdd_trans_create(env, mdd);
1952 RETURN(PTR_ERR(handle));
1954 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1958 rc = mdd_trans_start(env, mdd, handle);
1962 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1964 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1965 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1967 mdd_write_unlock(env, mdd_obj);
1972 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1973 mdd_write_unlock(env, mdd_obj);
1977 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1981 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1982 mdd_obj, name, handle);
1986 return mdd_trans_stop(env, mdd, rc, handle);
1989 static int mdd_declare_xattr_del(const struct lu_env *env,
1990 struct mdd_device *mdd,
1991 struct mdd_object *obj,
1993 struct thandle *handle)
1995 enum changelog_rec_type type;
1998 rc = mdo_declare_xattr_del(env, obj, name, handle);
2002 type = mdd_xattr_changelog_type(env, mdd, name);
2004 return 0; /* no changelog to store */
2006 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2010 * The caller should guarantee to update the object ctime
2011 * after xattr_set if needed.
2013 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2016 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2017 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2018 struct mdd_device *mdd = mdo2mdd(obj);
2019 struct thandle *handle;
2023 rc = mdd_la_get(env, mdd_obj, attr);
2027 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2031 handle = mdd_trans_create(env, mdd);
2033 RETURN(PTR_ERR(handle));
2035 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2039 rc = mdd_trans_start(env, mdd, handle);
2043 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2044 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2045 mdd_write_unlock(env, mdd_obj);
2049 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2052 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2057 return mdd_trans_stop(env, mdd, rc, handle);
2061 * read lov/lmv EA of an object
2062 * return the lov/lmv EA in an allocated lu_buf
2064 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2065 struct lu_buf *lmm_buf, const char *name)
2067 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2072 if (buf->lb_buf == NULL) {
2073 buf = lu_buf_check_and_alloc(buf, 4096);
2074 if (buf->lb_buf == NULL)
2079 rc = mdo_xattr_get(env, obj, buf, name);
2080 if (rc == -ERANGE) {
2081 /* mti_big_buf is allocated but is too small
2082 * we need to increase it */
2083 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2085 if (buf->lb_buf == NULL)
2088 } else if (rc < 0) {
2090 } else if (rc == 0) {
2094 lu_buf_alloc(lmm_buf, rc);
2095 if (lmm_buf->lb_buf == NULL)
2099 * we don't use lmm_buf directly, because we don't know xattr size, so
2100 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2102 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2107 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2108 struct mdd_object *o, struct lu_buf *buf,
2109 struct thandle *handle)
2111 struct hsm_attrs *attrs;
2112 enum hsm_states hsm_flags;
2113 enum changelog_rec_flags clf_flags = 0;
2117 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2122 attrs = buf->lb_buf;
2123 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2124 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2127 /* Add a changelog record for release. */
2128 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2129 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2130 clf_flags, o, handle);
2135 * check if layout swapping between 2 objects is allowed
2137 * - only normal FIDs or non-system IGIFs
2138 * - same type of objects
2139 * - same owner/group (so quotas are still valid) unless this is from HSM
2142 static int mdd_layout_swap_allowed(const struct lu_env *env,
2143 struct mdd_object *o1,
2144 const struct lu_attr *attr1,
2145 struct mdd_object *o2,
2146 const struct lu_attr *attr2,
2149 const struct lu_fid *fid1, *fid2;
2155 if (!fid_is_norm(fid1) &&
2156 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2159 if (!fid_is_norm(fid2) &&
2160 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2163 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2164 if (S_ISDIR(mdd_object_type(o1)))
2166 if (S_ISREG(mdd_object_type(o1)))
2171 if (flags & SWAP_LAYOUTS_MDS_HSM)
2174 if ((attr1->la_uid != attr2->la_uid) ||
2175 (attr1->la_gid != attr2->la_gid))
2181 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2182 * look into the layout in MDD layer. */
2183 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2185 struct lov_comp_md_v1 *comp_v1;
2186 struct lov_mds_md *v1;
2190 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2191 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2192 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2198 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2199 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2202 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2203 off = le32_to_cpu(comp_v1->lcm_entries[i].
2205 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2210 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2211 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2222 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2224 return mdd_lmm_oi(lmm, oi, true);
2227 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2229 return mdd_lmm_oi(lmm, oi, false);
2232 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2234 struct lov_comp_md_v1 *comp_v1;
2236 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2237 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2239 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2241 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2242 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2243 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2244 __u16 tmp_gen = *gen;
2246 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2248 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2255 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2257 return mdd_lmm_gen(lmm, gen, true);
2260 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2262 return mdd_lmm_gen(lmm, gen, false);
2265 static int mdd_dom_data_truncate(const struct lu_env *env,
2266 struct mdd_device *mdd, struct mdd_object *mo)
2269 struct dt_object *dom;
2272 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2274 GOTO(out, rc = -ENODATA);
2276 th = dt_trans_create(env, mdd->mdd_bottom);
2278 GOTO(out, rc = PTR_ERR(th));
2280 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2284 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2288 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2290 dt_trans_stop(env, mdd->mdd_bottom, th);
2292 /* Ignore failure but report the error */
2294 CERROR("%s: "DFID" can't truncate DOM inode data, rc = %d\n",
2295 mdd_obj_dev_name(mo), PFID(mdo2fid(mo)), rc);
2300 * swap layouts between 2 lustre objects
2302 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2303 struct md_object *obj2, __u64 flags)
2305 struct mdd_thread_info *info = mdd_env_info(env);
2306 struct mdd_object *fst_o = md2mdd_obj(obj1);
2307 struct mdd_object *snd_o = md2mdd_obj(obj2);
2308 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2309 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2310 struct mdd_device *mdd = mdo2mdd(obj1);
2311 struct lov_mds_md *fst_lmm, *snd_lmm;
2312 struct lu_buf *fst_buf = &info->mti_buf[0];
2313 struct lu_buf *snd_buf = &info->mti_buf[1];
2314 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2315 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2316 struct ost_id *saved_oi = NULL;
2317 struct thandle *handle;
2318 struct mdd_object *dom_o = NULL;
2319 __u64 domsize_dom, domsize_vlt;
2320 __u32 fst_gen, snd_gen, saved_gen;
2326 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
2327 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2329 /* we have to sort the 2 obj, so locking will always
2330 * be in the same order, even in case of 2 concurrent swaps */
2331 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
2332 if (rc == 0) /* same fid ? */
2338 rc = mdd_la_get(env, fst_o, fst_la);
2342 rc = mdd_la_get(env, snd_o, snd_la);
2346 /* check if layout swapping is allowed */
2347 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2351 handle = mdd_trans_create(env, mdd);
2353 RETURN(PTR_ERR(handle));
2355 /* objects are already sorted */
2356 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2357 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2359 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2360 if (rc < 0 && rc != -ENODATA)
2363 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2364 if (rc < 0 && rc != -ENODATA)
2367 /* check if file has DoM. DoM file can be migrated only to another
2368 * DoM layout with the same DoM component size or to an non-DOM
2369 * layout. After migration to OSTs layout, local MDT inode data
2370 * should be truncated.
2371 * Objects are sorted by FIDs, considering that original file's FID
2372 * is always smaller the snd_o is always original file we are migrating
2375 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2376 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2378 /* Only migration is supported for DoM files, not 'swap_layouts' so
2379 * target file must be volatile and orphan.
2381 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2382 dom_o = domsize_dom ? snd_o : NULL;
2383 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2384 swap(domsize_dom, domsize_vlt);
2385 dom_o = domsize_dom ? fst_o : NULL;
2386 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2387 /* 'lfs swap_layouts' case, neither file should have DoM */
2389 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2390 "use migration instead: rc = %d\n", rc);
2394 if (domsize_vlt > 0 && domsize_dom == 0) {
2396 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": OST to DOM "
2397 "migration is not supported: rc = %d\n",
2398 PFID(mdo2fid(snd_o)), rc);
2400 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2402 CDEBUG(D_LAYOUT, "cannot swap layout for "DFID": new layout "
2403 "must have the same DoM component size: rc = %d\n",
2404 PFID(mdo2fid(fst_o)), rc);
2406 } else if (domsize_vlt > 0) {
2407 /* Migration with the same DOM component size, no need to
2408 * truncate local data, it is still being used */
2412 /* swapping 2 non existant layouts is a success */
2413 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2416 /* to help inode migration between MDT, it is better to
2417 * start by the no layout file (if one), so we order the swap */
2418 if (snd_buf->lb_buf == NULL) {
2420 swap(fst_buf, snd_buf);
2423 fst_gen = snd_gen = 0;
2424 /* lmm and generation layout initialization */
2425 if (fst_buf->lb_buf != NULL) {
2426 fst_lmm = fst_buf->lb_buf;
2427 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2428 fst_fl = LU_XATTR_REPLACE;
2432 fst_fl = LU_XATTR_CREATE;
2435 snd_lmm = snd_buf->lb_buf;
2436 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2438 saved_gen = fst_gen;
2439 /* increase the generation layout numbers */
2444 * XXX The layout generation is used to generate component IDs for
2445 * the composite file, we have to do some special tweaks to make
2446 * sure the layout generation is always adequate for that job.
2449 /* Skip invalid generation number for composite layout */
2450 if ((snd_gen & LCME_ID_MASK) == 0)
2452 if ((fst_gen & LCME_ID_MASK) == 0)
2454 /* Make sure the generation is greater than all the component IDs */
2455 if (fst_gen < snd_gen)
2457 else if (fst_gen > snd_gen)
2460 /* set the file specific informations in lmm */
2461 if (fst_lmm != NULL) {
2462 saved_oi = &info->mti_oa.o_oi;
2463 mdd_get_lmm_oi(fst_lmm, saved_oi);
2464 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2465 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2466 mdd_set_lmm_oi(snd_lmm, saved_oi);
2468 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2469 cpu_to_le32(LOV_MAGIC_MAGIC))
2470 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2472 GOTO(stop, rc = -EPROTO);
2474 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2476 /* Prepare HSM attribute if it's required */
2477 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2478 const int buflen = sizeof(struct hsm_attrs);
2480 lu_buf_alloc(fst_hsm_buf, buflen);
2481 lu_buf_alloc(snd_hsm_buf, buflen);
2482 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2483 GOTO(stop, rc = -ENOMEM);
2485 /* Read HSM attribute */
2486 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2490 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2494 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2495 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2500 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2501 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2507 /* prepare transaction */
2508 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2513 if (fst_buf->lb_buf != NULL)
2514 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2515 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2518 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2523 rc = mdd_trans_start(env, mdd, handle);
2527 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2528 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2532 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2534 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2537 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
2538 mdd_obj_dev_name(fst_o),
2539 PFID(mdo2fid(fst_o)), rc, rc2);
2544 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2548 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2551 if (fst_buf->lb_buf != NULL)
2552 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2553 LU_XATTR_REPLACE, handle);
2555 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2558 GOTO(out_restore, rc);
2560 /* Issue one changelog record per file */
2561 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2565 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2574 /* failure on second file, but first was done, so we have
2575 * to roll back first. */
2576 if (fst_buf->lb_buf != NULL) {
2577 mdd_set_lmm_oi(fst_lmm, saved_oi);
2578 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2579 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2580 LU_XATTR_REPLACE, handle);
2582 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2588 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2593 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2598 CERROR("%s: unable to roll back layout swap. FIDs: "
2599 DFID" and "DFID "error: %d/%d, steps: %d\n",
2600 mdd_obj_dev_name(fst_o),
2601 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
2603 /* a solution to avoid journal commit is to panic,
2604 * but it has strong consequences so we use LBUG to
2605 * allow sysdamin to choose to panic or not
2612 rc = mdd_trans_stop(env, mdd, rc, handle);
2614 /* Truncate local DOM data if all went well */
2616 mdd_dom_data_truncate(env, mdd, dom_o);
2618 mdd_write_unlock(env, snd_o);
2619 mdd_write_unlock(env, fst_o);
2621 lu_buf_free(fst_buf);
2622 lu_buf_free(snd_buf);
2623 lu_buf_free(fst_hsm_buf);
2624 lu_buf_free(snd_hsm_buf);
2627 (void) mdd_object_pfid_replace(env, fst_o);
2628 (void) mdd_object_pfid_replace(env, snd_o);
2633 static int mdd_declare_layout_change(const struct lu_env *env,
2634 struct mdd_device *mdd,
2635 struct mdd_object *obj,
2636 struct md_layout_change *mlc,
2637 struct thandle *handle)
2641 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2645 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2649 /* For PFL, this is used to instantiate necessary component objects. */
2651 mdd_layout_instantiate_component(const struct lu_env *env,
2652 struct mdd_object *obj, struct md_layout_change *mlc,
2653 struct thandle *handle)
2655 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2659 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2662 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2664 * It's possible that another layout write intent has already
2665 * instantiated our objects, so a -EALREADY returned, and we need to
2669 RETURN(rc == -EALREADY ? 0 : rc);
2671 rc = mdd_trans_start(env, mdd, handle);
2675 mdd_write_lock(env, obj, DT_TGT_CHILD);
2676 rc = mdo_layout_change(env, obj, mlc, handle);
2677 mdd_write_unlock(env, obj);
2681 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2686 * Change the FLR layout from RDONLY to WRITE_PENDING.
2688 * It picks the primary mirror, and bumps the layout version, and set
2689 * layout version xattr to OST objects in a sync tx. In order to facilitate
2690 * the handling of phantom writers from evicted clients, the clients carry
2691 * layout version of the file with write RPC, so that the OSTs can verify
2692 * if the write RPCs are legitimate, meaning not from evicted clients.
2695 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2696 struct md_layout_change *mlc, struct thandle *handle)
2698 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2699 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2700 struct lustre_som_attrs *som = &mlc->mlc_som;
2705 /* Verify acceptable operations */
2706 switch (mlc->mlc_opc) {
2707 case MD_LAYOUT_WRITE:
2708 case MD_LAYOUT_RESYNC:
2709 /* these are legal operations - this represents the case that
2710 * a few mirrors were missed in the last resync. */
2712 case MD_LAYOUT_RESYNC_DONE:
2717 som_buf->lb_buf = som;
2718 som_buf->lb_len = sizeof(*som);
2719 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2720 if (rc < 0 && rc != -ENODATA)
2724 lustre_som_swab(som);
2725 if (som->lsa_valid & SOM_FL_STRICT)
2726 fl = LU_XATTR_REPLACE;
2729 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2734 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2735 XATTR_NAME_SOM, fl, handle);
2740 /* record a changelog for data mover to consume */
2741 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2745 rc = mdd_trans_start(env, mdd, handle);
2749 /* it needs a sync tx to make FLR to work properly */
2750 handle->th_sync = 1;
2752 mdd_write_lock(env, obj, DT_TGT_CHILD);
2753 rc = mdo_layout_change(env, obj, mlc, handle);
2755 /* SOM state transition from STRICT to STALE */
2756 som->lsa_valid = SOM_FL_STALE;
2757 lustre_som_swab(som);
2758 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2761 mdd_write_unlock(env, obj);
2765 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2776 * Handle mirrored file state transition when it's in WRITE_PENDING.
2778 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2779 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2780 * layout version will be increased, and the file's state will be changed to
2784 mdd_layout_update_write_pending(const struct lu_env *env,
2785 struct mdd_object *obj, struct md_layout_change *mlc,
2786 struct thandle *handle)
2788 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2792 switch (mlc->mlc_opc) {
2793 case MD_LAYOUT_RESYNC:
2794 /* Upon receiving the resync request, it should
2795 * instantiate all stale components right away to get ready
2796 * for mirror copy. In order to avoid layout version change,
2797 * client should avoid sending LAYOUT_WRITE request at the
2800 case MD_LAYOUT_WRITE:
2801 /* legal race for concurrent write, the file state has been
2802 * changed by another client. */
2808 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2812 rc = mdd_trans_start(env, mdd, handle);
2816 /* it needs a sync tx to make FLR to work properly */
2817 handle->th_sync = 1;
2819 mdd_write_lock(env, obj, DT_TGT_CHILD);
2820 rc = mdo_layout_change(env, obj, mlc, handle);
2821 mdd_write_unlock(env, obj);
2832 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2834 * Only concurrent write and sync complete requests are possible when the
2835 * file is in SYNC_PENDING. For the latter request, it will pass in the
2836 * mirrors that have been synchronized, then the stale bit will be cleared
2837 * to make the file's state turn into RDONLY.
2838 * For concurrent write reqeust, it just needs to change the file's state
2839 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2840 * version because the version will be increased in the transition to
2841 * SYNC_PENDING later so that it can deny the write request from potential
2842 * evicted SYNC clients. */
2844 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2845 struct md_layout_change *mlc, struct thandle *handle)
2847 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2848 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2853 /* operation validation */
2854 switch (mlc->mlc_opc) {
2855 case MD_LAYOUT_RESYNC_DONE:
2856 /* resync complete. */
2857 case MD_LAYOUT_WRITE:
2858 /* concurrent write. */
2860 case MD_LAYOUT_RESYNC:
2861 /* resync again, most likely the previous run failed.
2862 * no-op if it's already in SYNC_PENDING state */
2868 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2869 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2870 if (rc < 0 && rc != -ENODATA)
2873 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2874 lustre_som_swab(&mlc->mlc_som);
2875 som_buf->lb_buf = &mlc->mlc_som;
2876 som_buf->lb_len = sizeof(mlc->mlc_som);
2879 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2883 /* record a changelog for the completion of resync */
2884 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
2889 /* RESYNC_DONE has piggybacked size and blocks */
2891 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2892 XATTR_NAME_SOM, fl, handle);
2897 rc = mdd_trans_start(env, mdd, handle);
2901 /* it needs a sync tx to make FLR to work properly */
2902 handle->th_sync = 1;
2904 rc = mdo_layout_change(env, obj, mlc, handle);
2909 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2915 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2924 * Layout change callback for object.
2926 * This is only used by FLR for now. In the future, it can be exteneded to
2927 * handle all layout change.
2930 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2931 struct md_layout_change *mlc)
2933 struct mdd_object *obj = md2mdd_obj(o);
2934 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2935 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2936 struct lov_comp_md_v1 *lcm;
2937 struct thandle *handle;
2942 /* Verify acceptable operations */
2943 switch (mlc->mlc_opc) {
2944 case MD_LAYOUT_WRITE:
2945 case MD_LAYOUT_RESYNC:
2946 case MD_LAYOUT_RESYNC_DONE:
2952 handle = mdd_trans_create(env, mdd);
2954 RETURN(PTR_ERR(handle));
2956 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2963 /* analyze the layout to make sure it's a FLR file */
2965 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2966 GOTO(out, rc = -EINVAL);
2968 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2970 /* please refer to HLD of FLR for state transition */
2971 switch (flr_state) {
2973 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2975 case LCM_FL_WRITE_PENDING:
2976 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2979 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2981 case LCM_FL_SYNC_PENDING:
2982 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2991 mdd_trans_stop(env, mdd, rc, handle);
2996 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2997 struct mdd_object *child, const struct lu_attr *attr,
2998 const struct md_op_spec *spec,
2999 struct dt_allocation_hint *hint)
3001 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3002 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3003 struct dt_object *nc = mdd_object_child(child);
3005 memset(hint, 0, sizeof(*hint));
3007 /* For striped directory, give striping EA to lod_ah_init, which will
3008 * decide the stripe_offset and stripe count by it. */
3009 if (S_ISDIR(attr->la_mode) &&
3010 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3011 hint->dah_eadata = spec->u.sp_ea.eadata;
3012 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3014 hint->dah_eadata = NULL;
3015 hint->dah_eadata_len = 0;
3016 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
3017 if (mdd->mdd_append_stripe_count != 0 ||
3018 mdd->mdd_append_pool[0])
3020 "using O_APPEND file striping\n");
3021 if (mdd->mdd_append_stripe_count)
3022 hint->dah_append_stripes =
3023 mdd->mdd_append_stripe_count;
3024 if (mdd->mdd_append_pool[0])
3025 hint->dah_append_pool = mdd->mdd_append_pool;
3027 hint->dah_append_stripes = 0;
3031 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3032 hint->dah_eadata, hint->dah_eadata_len);
3033 /* @hint will be initialized by underlying device. */
3034 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3037 static int accmode(const struct lu_env *env, const struct lu_attr *la,
3040 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3041 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3042 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3043 * owner can write to a file even if it is marked readonly to hide
3044 * its brokenness. (bug 5781) */
3045 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3046 struct lu_ucred *uc = lu_ucred_check(env);
3048 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3052 return mds_accmode(open_flags);
3055 static int mdd_open_sanity_check(const struct lu_env *env,
3056 struct mdd_object *obj,
3057 const struct lu_attr *attr, u64 open_flags,
3063 /* EEXIST check, also opening of *open* orphans is allowed so we can
3064 * open-by-handle unlinked files
3066 if (mdd_is_dead_obj(obj) && !is_replay &&
3067 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3070 if (S_ISLNK(attr->la_mode))
3073 mode = accmode(env, attr, open_flags);
3075 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3078 if (!(open_flags & MDS_OPEN_CREATED)) {
3079 rc = mdd_permission_internal(env, obj, attr, mode);
3084 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3085 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3086 open_flags &= ~MDS_OPEN_TRUNC;
3088 /* For writing append-only file must open it with append mode. */
3089 if (attr->la_flags & LUSTRE_APPEND_FL) {
3090 if ((open_flags & MDS_FMODE_WRITE) &&
3091 !(open_flags & MDS_OPEN_APPEND))
3093 if (open_flags & MDS_OPEN_TRUNC)
3100 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3101 u64 open_flags, struct md_op_spec *spec)
3103 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3104 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3105 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3106 struct mdd_object_user *mou = NULL;
3107 const struct lu_ucred *uc = lu_ucred(env);
3108 struct mdd_device *mdd = mdo2mdd(obj);
3109 enum changelog_rec_type type = CL_OPEN;
3113 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3115 rc = mdd_la_get(env, mdd_obj, attr);
3119 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3121 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3126 mdd_obj->mod_count++;
3128 if (!mdd_changelog_enabled(env, mdd, type))
3132 /* look for existing opener in list under mdd_write_lock */
3133 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3138 /* add user to list */
3139 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3145 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3147 mdd_obj_user_free(mou);
3152 if (type == CL_DN_OPEN) {
3153 if (ktime_before(ktime_get(), mou->mou_deniednext))
3154 /* same user denied again same access within
3155 * time interval: do not record
3159 /* this user already denied, but some time ago:
3160 * update denied time
3162 mou->mou_deniednext =
3163 ktime_add(ktime_get(),
3164 ktime_set(mdd->mdd_cl.mc_deniednext,
3167 mou->mou_opencount++;
3168 /* same user opening file again with same flags:
3175 /* FYI, only the bottom 32 bits of open_flags are recorded */
3176 mdd_changelog(env, type, open_flags, md_dev, mdo2fid(mdd_obj));
3180 mdd_write_unlock(env, mdd_obj);
3184 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3185 struct md_attr *ma, struct thandle *handle)
3189 rc = mdd_orphan_declare_delete(env, obj, handle);
3193 return mdo_declare_destroy(env, obj, handle);
3197 * No permission check is needed.
3199 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3200 struct md_attr *ma, u64 open_flags)
3202 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3203 struct mdd_device *mdd = mdo2mdd(obj);
3204 struct thandle *handle = NULL;
3207 bool blocked = false;
3208 bool last_close_by_uid = false;
3209 const struct lu_ucred *uc = lu_ucred(env);
3212 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3213 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3214 mdd_obj->mod_count--;
3215 mdd_write_unlock(env, mdd_obj);
3217 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3218 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3219 "list\n", PFID(mdd_object_fid(mdd_obj)));
3223 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3224 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3225 /* check without any lock */
3226 is_orphan = mdd_obj->mod_count == 1 &&
3227 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3231 /* mdd_trans_create() maybe failed because of barrier_entry(),
3232 * under such case, the orphan MDT-object will be left in the
3233 * orphan list, and when the MDT remount next time, the unused
3234 * orphans will be destroyed automatically.
3236 * One exception: the former mdd_finish_unlink may failed to
3237 * add the orphan MDT-object to the orphan list, then if the
3238 * mdd_trans_create() failed because of barrier_entry(), the
3239 * MDT-object will become real orphan that is neither in the
3240 * namespace nor in the orphan list. Such bad case should be
3241 * very rare and will be handled by e2fsck/lfsck. */
3242 handle = mdd_trans_create(env, mdo2mdd(obj));
3243 if (IS_ERR(handle)) {
3244 rc = PTR_ERR(handle);
3245 if (rc != -EINPROGRESS)
3253 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3257 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3262 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3268 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3269 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3271 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3272 lu_dev_name(mdd2lu_dev(mdd)),
3273 PFID(mdd_object_fid(mdd_obj)), rc);
3277 /* check again with lock */
3278 is_orphan = (mdd_obj->mod_count == 1) &&
3279 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3280 ma->ma_attr.la_nlink == 0);
3282 if (is_orphan && !handle && !blocked) {
3283 mdd_write_unlock(env, mdd_obj);
3287 mdd_obj->mod_count--; /*release open count */
3289 /* under mdd write lock */
3290 /* If recording, see if we need to remove UID from list. uc is not
3291 * initialized if the client has been evicted. */
3292 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3293 struct mdd_object_user *mou;
3295 /* look for UID in list */
3296 /* If mou is NULL, it probably means logging was enabled after
3297 * the user had the file open. So the corresponding close
3298 * will not be logged.
3300 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3303 mou->mou_opencount--;
3304 if (mou->mou_opencount == 0) {
3305 mdd_obj_user_remove(mdd_obj, mou);
3306 last_close_by_uid = true;
3311 if (!is_orphan || blocked)
3315 /* NB: Object maybe not in orphan list originally, it is rare case for
3316 * mdd_finish_unlink() failure, in that case, the object doesn't have
3317 * ORPHAN_OBJ flag */
3318 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3319 /* remove link to object from orphan index */
3320 LASSERT(handle != NULL);
3321 rc = mdd_orphan_delete(env, mdd_obj, handle);
3323 CERROR("%s: unable to delete "DFID" from orphan list: "
3324 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3325 PFID(mdd_object_fid(mdd_obj)), rc);
3326 /* If object was not deleted from orphan list, do not
3327 * destroy OSS objects, which will be done when next
3332 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3333 "list, OSS objects to be destroyed.\n",
3334 PFID(mdd_object_fid(mdd_obj)));
3337 rc = mdo_destroy(env, mdd_obj, handle);
3340 CERROR("%s: unable to delete "DFID" from orphan list: "
3341 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3342 PFID(mdd_object_fid(mdd_obj)), rc);
3347 mdd_write_unlock(env, mdd_obj);
3349 if (rc != 0 || blocked ||
3350 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3353 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3354 * or if CL_OPEN was recorded and it's last close by user.
3355 * Changelogs mask may change between open and close operations, but
3356 * this is not a big deal if we have a CL_CLOSE entry with no matching
3357 * CL_OPEN. Plus Changelogs mask may not change often.
3359 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3360 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3361 MDS_OPEN_TRUNC))) ||
3362 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3363 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3364 if (handle == NULL) {
3365 handle = mdd_trans_create(env, mdo2mdd(obj));
3367 GOTO(stop, rc = PTR_ERR(handle));
3369 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3370 NULL, NULL, handle);
3374 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3379 /* FYI, only the bottom 32 bits of open_flags are recorded */
3380 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3385 if (handle != NULL && !IS_ERR(handle))
3386 rc = mdd_trans_stop(env, mdd, rc, handle);
3392 * Permission check is done when open,
3393 * no need check again.
3395 static int mdd_readpage_sanity_check(const struct lu_env *env,
3396 struct mdd_object *obj)
3398 struct dt_object *next = mdd_object_child(obj);
3402 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3410 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3411 size_t nob, const struct dt_it_ops *iops,
3412 struct dt_it *it, __u32 attr, void *arg)
3414 struct lu_dirpage *dp = &lp->lp_dir;
3418 struct lu_dirent *ent;
3419 struct lu_dirent *last = NULL;
3423 if (nob < sizeof(*dp))
3426 memset(area, 0, sizeof (*dp));
3427 area += sizeof (*dp);
3428 nob -= sizeof (*dp);
3435 len = iops->key_size(env, it);
3437 /* IAM iterator can return record with zero len. */
3441 hash = iops->store(env, it);
3442 if (unlikely(first)) {
3444 dp->ldp_hash_start = cpu_to_le64(hash);
3447 /* calculate max space required for lu_dirent */
3448 recsize = lu_dirent_calc_size(len, attr);
3450 if (nob >= recsize) {
3451 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3452 if (result == -ESTALE)
3457 /* osd might not able to pack all attributes,
3458 * so recheck rec length */
3459 recsize = le16_to_cpu(ent->lde_reclen);
3461 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3462 fid_le_to_cpu(&fid, &ent->lde_fid);
3463 if (fid_is_dot_lustre(&fid))
3467 result = (last != NULL) ? 0 :-EINVAL;
3471 ent = (void *)ent + recsize;
3475 result = iops->next(env, it);
3476 if (result == -ESTALE)
3478 } while (result == 0);
3481 dp->ldp_hash_end = cpu_to_le64(hash);
3483 if (last->lde_hash == dp->ldp_hash_end)
3484 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3485 last->lde_reclen = 0; /* end mark */
3488 /* end of directory */
3489 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3490 else if (result < 0)
3491 CWARN("build page failed: %d!\n", result);
3495 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3496 const struct lu_rdpg *rdpg)
3498 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3502 if (mdd_object_exists(mdd_obj) == 0) {
3503 CERROR("%s: object "DFID" not found: rc = -2\n",
3504 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3508 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3509 rc = mdd_readpage_sanity_check(env, mdd_obj);
3511 GOTO(out_unlock, rc);
3513 if (mdd_is_dead_obj(mdd_obj)) {
3515 struct lu_dirpage *dp;
3518 * According to POSIX, please do not return any entry to client:
3519 * even dot and dotdot should not be returned.
3521 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3522 PFID(mdd_object_fid(mdd_obj)));
3524 if (rdpg->rp_count <= 0)
3525 GOTO(out_unlock, rc = -EFAULT);
3526 LASSERT(rdpg->rp_pages != NULL);
3528 pg = rdpg->rp_pages[0];
3529 dp = (struct lu_dirpage *)kmap(pg);
3530 memset(dp, 0 , sizeof(struct lu_dirpage));
3531 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3532 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3533 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3535 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3538 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3539 mdd_dir_page_build, NULL);
3541 struct lu_dirpage *dp;
3543 dp = kmap(rdpg->rp_pages[0]);
3544 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3547 * No pages were processed, mark this for first page
3550 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3551 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3552 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3554 kunmap(rdpg->rp_pages[0]);
3557 GOTO(out_unlock, rc);
3559 mdd_read_unlock(env, mdd_obj);
3563 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3565 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3567 if (mdd_object_exists(mdd_obj) == 0) {
3570 CERROR("%s: object "DFID" not found: rc = %d\n",
3571 mdd_obj_dev_name(mdd_obj),
3572 PFID(mdd_object_fid(mdd_obj)), rc);
3575 return dt_object_sync(env, mdd_object_child(mdd_obj),
3579 static int mdd_object_lock(const struct lu_env *env,
3580 struct md_object *obj,
3581 struct lustre_handle *lh,
3582 struct ldlm_enqueue_info *einfo,
3583 union ldlm_policy_data *policy)
3585 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3586 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3590 static int mdd_object_unlock(const struct lu_env *env,
3591 struct md_object *obj,
3592 struct ldlm_enqueue_info *einfo,
3593 union ldlm_policy_data *policy)
3595 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3596 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3599 const struct md_object_operations mdd_obj_ops = {
3600 .moo_permission = mdd_permission,
3601 .moo_attr_get = mdd_attr_get,
3602 .moo_attr_set = mdd_attr_set,
3603 .moo_xattr_get = mdd_xattr_get,
3604 .moo_xattr_set = mdd_xattr_set,
3605 .moo_xattr_list = mdd_xattr_list,
3606 .moo_invalidate = mdd_invalidate,
3607 .moo_xattr_del = mdd_xattr_del,
3608 .moo_swap_layouts = mdd_swap_layouts,
3609 .moo_open = mdd_open,
3610 .moo_close = mdd_close,
3611 .moo_readpage = mdd_readpage,
3612 .moo_readlink = mdd_readlink,
3613 .moo_changelog = mdd_changelog,
3614 .moo_object_sync = mdd_object_sync,
3615 .moo_object_lock = mdd_object_lock,
3616 .moo_object_unlock = mdd_object_unlock,
3617 .moo_layout_change = mdd_layout_change,