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(mdd_object_fid(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 mdd_object_fid(mdd_obj),
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 mdd_object_fid(mdd_obj),
972 mdd_obj->mod_cltime = ktime_get();
977 int mdd_changelog_data_store_xattr(const struct lu_env *env,
978 struct mdd_device *mdd,
979 enum changelog_rec_type type,
980 enum changelog_rec_flags clf_flags,
981 struct mdd_object *mdd_obj,
982 const char *xattr_name,
983 struct thandle *handle)
987 LASSERT(mdd_obj != NULL);
988 LASSERT(handle != NULL);
990 if (!mdd_changelog_enabled(env, mdd, type))
993 if (mdd_is_volatile_obj(mdd_obj))
996 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
997 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
998 /* Don't need multiple updates in this log */
999 /* Don't check under lock - no big deal if we get an extra
1005 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1006 mdd_object_fid(mdd_obj),
1007 xattr_name, handle);
1009 mdd_obj->mod_cltime = ktime_get();
1014 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
1015 * except for open flags have a dedicated record to store 32 bits of flags */
1016 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1017 enum changelog_rec_flags clf_flags,
1018 struct md_device *m, const struct lu_fid *fid)
1020 struct thandle *handle;
1021 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1025 LASSERT(fid != NULL);
1027 /* We'll check this again below, but we check now before we
1028 * start a transaction. */
1029 if (!mdd_changelog_enabled(env, mdd, type))
1032 handle = mdd_trans_create(env, mdd);
1034 RETURN(PTR_ERR(handle));
1036 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1040 rc = mdd_trans_start(env, mdd, handle);
1044 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1048 rc = mdd_trans_stop(env, mdd, rc, handle);
1054 * Save LMA extended attributes with data from \a ma.
1056 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1057 * not, LMA EA will be first read from disk, modified and write back.
1060 /* Precedence for choosing record type when multiple
1061 * attributes change: setattr > mtime > ctime > atime
1062 * (ctime changes when mtime does, plus chmod/chown.
1063 * atime and ctime are independent.) */
1064 static int mdd_attr_set_changelog(const struct lu_env *env,
1065 struct md_object *obj, struct thandle *handle,
1068 struct mdd_device *mdd = mdo2mdd(obj);
1071 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
1072 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
1073 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
1074 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
1075 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
1076 bits = bits & mdd->mdd_cl.mc_mask;
1077 /* This is an implementation limit rather than a protocol limit */
1078 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1082 /* The record type is the lowest non-masked set bit */
1085 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1086 return mdd_changelog_data_store(env, mdd, type, valid,
1087 md2mdd_obj(obj), handle);
1090 static int mdd_declare_attr_set(const struct lu_env *env,
1091 struct mdd_device *mdd,
1092 struct mdd_object *obj,
1093 const struct lu_attr *attr,
1094 struct thandle *handle)
1098 rc = mdo_declare_attr_set(env, obj, attr, handle);
1102 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1103 if (attr->la_valid & LA_MODE) {
1104 mdd_read_lock(env, obj, DT_TGT_CHILD);
1105 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1106 XATTR_NAME_ACL_ACCESS);
1107 mdd_read_unlock(env, obj);
1108 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1114 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1115 rc = mdo_declare_xattr_set(env, obj, buf,
1116 XATTR_NAME_ACL_ACCESS, 0,
1124 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1133 * permission changes may require sync operation, to mitigate performance
1134 * impact, only do this for dir and when permission is reduced.
1136 * For regular files, version is updated with permission change (see VBR), async
1137 * permission won't cause any issue, while missing permission change on
1138 * directory may affect accessibility of other objects after recovery.
1140 static inline bool permission_needs_sync(const struct lu_attr *old,
1141 const struct lu_attr *new)
1143 if (!S_ISDIR(old->la_mode))
1146 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1149 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1152 if (new->la_valid & LA_MODE) {
1153 /* turned on sticky bit */
1154 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1157 /* set-GID has no impact on what is allowed, not checked */
1159 /* turned off setuid bit, or one of rwx for someone */
1160 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1161 (old->la_mode & (0777 | S_ISUID)))
1168 static inline __u64 mdd_lmm_dom_size(void *buf)
1170 struct lov_mds_md *lmm = buf;
1171 struct lov_comp_md_v1 *comp_v1;
1172 struct lov_mds_md *v1;
1178 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1181 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1182 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1183 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1185 /* DoM entry is the first entry always */
1186 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1187 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1192 /* set attr and LOV EA at once, return updated attr */
1193 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1194 const struct md_attr *ma)
1196 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1197 struct mdd_device *mdd = mdo2mdd(obj);
1198 struct thandle *handle = NULL;
1199 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
1200 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1201 const struct lu_attr *la = &ma->ma_attr;
1202 struct lu_ucred *uc;
1203 bool chrgrp_by_unprivileged_user = false;
1207 /* we do not use ->attr_set() for LOV/HSM EA any more */
1208 LASSERT((ma->ma_valid & MA_LOV) == 0);
1209 LASSERT((ma->ma_valid & MA_HSM) == 0);
1211 rc = mdd_la_get(env, mdd_obj, attr);
1215 *la_copy = ma->ma_attr;
1216 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1220 /* no need to setattr anymore */
1221 if (la_copy->la_valid == 0) {
1223 "%s: no valid attribute on "DFID", previous was %#llx\n",
1224 mdd_obj_dev_name(mdd_obj),
1225 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1230 /* If an unprivileged user changes group of some file,
1231 * the setattr operation will be processed synchronously to
1232 * honor the quota limit of the corresponding group. see LU-5152 */
1233 uc = lu_ucred_check(env);
1234 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1235 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1236 /* LU-10048: disable synchronous chgrp operation for it will
1237 * cause deadlock between MDT and OST.
1238 la_copy->la_valid |= LA_FLAGS;
1239 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1241 chrgrp_by_unprivileged_user = true;
1243 /* Flush the possible existing client setattr requests to OSTs
1244 * to keep the order with the current setattr operation that
1245 * will be sent directly to OSTs. see LU-5152 */
1246 /* LU-11303 disable sync as this is too heavyweight.
1247 * This should be replaced with a sync only for the object
1248 * being modified here, not the whole filesystem.
1249 rc = dt_sync(env, mdd->mdd_child);
1255 handle = mdd_trans_create(env, mdd);
1256 if (IS_ERR(handle)) {
1257 rc = PTR_ERR(handle);
1263 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1267 rc = mdd_trans_start(env, mdd, handle);
1271 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1272 permission_needs_sync(attr, la))
1273 handle->th_sync = 1;
1275 if (la->la_valid & (LA_MTIME | LA_CTIME))
1276 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1277 la->la_mtime, la->la_ctime);
1279 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1281 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1282 * otherwise following truncate will do nothing and truncated
1283 * data may be read again. This is a quick fix until LU-11033
1286 if (la_copy->la_valid & LA_SIZE) {
1287 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1289 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1293 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1294 la_copy->la_valid &= ~LA_SIZE;
1295 lu_buf_free(lov_buf);
1299 if (la_copy->la_valid) {
1300 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1302 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1303 /* rollback to the original gid */
1304 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1305 la_copy->la_gid = attr->la_gid;
1306 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1309 mdd_write_unlock(env, mdd_obj);
1313 rc = mdd_attr_set_changelog(env, obj, handle,
1317 rc = mdd_trans_stop(env, mdd, rc, handle);
1322 static int mdd_xattr_sanity_check(const struct lu_env *env,
1323 struct mdd_object *obj,
1324 const struct lu_attr *attr,
1327 struct lu_ucred *uc = lu_ucred_assert(env);
1330 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1333 if (strncmp(XATTR_USER_PREFIX, name,
1334 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1335 /* For sticky directories, only the owner and privileged user
1336 * can write attributes. */
1337 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1338 (uc->uc_fsuid != attr->la_uid) &&
1339 !md_capable(uc, CFS_CAP_FOWNER))
1341 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1342 (uc->uc_fsuid != attr->la_uid) &&
1343 !md_capable(uc, CFS_CAP_FOWNER)) {
1351 * Check if a string begins with a given prefix.
1353 * \param str String to check
1354 * \param prefix Substring to check at the beginning of \a str
1355 * \return true/false whether the condition is verified.
1357 static inline bool has_prefix(const char *str, const char *prefix)
1359 return strncmp(prefix, str, strlen(prefix)) == 0;
1363 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1365 * \param[in] xattr_name Full extended attribute name.
1367 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1369 static enum changelog_rec_type
1370 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1371 const char *xattr_name)
1373 /* Layout changes systematically recorded */
1374 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1375 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1376 allowed_lustre_lov(xattr_name))
1379 /* HSM information changes systematically recorded */
1380 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1383 /* Avoid logging SOM xattr for every file */
1384 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1387 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1388 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1389 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1390 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1391 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1397 static int mdd_declare_xattr_set(const struct lu_env *env,
1398 struct mdd_device *mdd,
1399 struct mdd_object *obj,
1400 const struct lu_buf *buf,
1402 int fl, struct thandle *handle)
1404 enum changelog_rec_type type;
1407 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1411 type = mdd_xattr_changelog_type(env, mdd, name);
1413 return 0; /* no changelog to store */
1415 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1419 * Compare current and future data of HSM EA and add a changelog if needed.
1421 * Caller should have write-locked \param obj.
1423 * \param buf - Future HSM EA content.
1424 * \retval 0 if no changelog is needed or changelog was added properly.
1425 * \retval -ve errno if there was a problem
1427 static int mdd_hsm_update_locked(const struct lu_env *env,
1428 struct md_object *obj,
1429 const struct lu_buf *buf,
1430 struct thandle *handle,
1431 enum changelog_rec_flags *clf_flags)
1433 struct mdd_thread_info *info = mdd_env_info(env);
1434 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1435 struct lu_buf *current_buf;
1436 struct md_hsm *current_mh;
1437 struct md_hsm *new_mh;
1441 OBD_ALLOC_PTR(current_mh);
1442 if (current_mh == NULL)
1445 /* Read HSM attrs from disk */
1446 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1448 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1450 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1451 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1452 if (rc < 0 && rc != -ENODATA)
1454 else if (rc == -ENODATA)
1455 current_mh->mh_flags = 0;
1457 /* Map future HSM xattr */
1458 OBD_ALLOC_PTR(new_mh);
1460 GOTO(free, rc = -ENOMEM);
1461 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1465 /* Flags differ, set flags for the changelog that will be added */
1466 if (current_mh->mh_flags != new_mh->mh_flags) {
1467 hsm_set_cl_event(clf_flags, HE_STATE);
1468 if (new_mh->mh_flags & HS_DIRTY)
1469 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1472 OBD_FREE_PTR(new_mh);
1475 OBD_FREE_PTR(current_mh);
1479 static int mdd_object_pfid_replace(const struct lu_env *env,
1480 struct mdd_object *o)
1482 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1483 struct thandle *handle;
1486 handle = mdd_trans_create(env, mdd);
1488 RETURN(PTR_ERR(handle));
1490 handle->th_complex = 1;
1492 /* it doesn't need to track the PFID update via llog, because LFSCK
1493 * will repair it even it goes wrong */
1494 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1499 rc = mdd_trans_start(env, mdd, handle);
1503 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1508 mdd_trans_stop(env, mdd, rc, handle);
1513 static int mdd_declare_xattr_del(const struct lu_env *env,
1514 struct mdd_device *mdd,
1515 struct mdd_object *obj,
1517 struct thandle *handle);
1519 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1522 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1523 struct md_object *md_vic)
1525 struct mdd_device *mdd = mdo2mdd(md_obj);
1526 struct mdd_object *obj = md2mdd_obj(md_obj);
1527 struct mdd_object *vic = md2mdd_obj(md_vic);
1528 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1529 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1530 struct lov_mds_md *lmm;
1531 struct thandle *handle;
1535 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1536 if (rc == 0) /* same fid */
1539 handle = mdd_trans_create(env, mdd);
1541 RETURN(PTR_ERR(handle));
1544 mdd_write_lock(env, obj, DT_TGT_CHILD);
1545 mdd_write_lock(env, vic, DT_TGT_CHILD);
1547 mdd_write_lock(env, vic, DT_TGT_CHILD);
1548 mdd_write_lock(env, obj, DT_TGT_CHILD);
1551 /* get EA of victim file */
1552 memset(buf_vic, 0, sizeof(*buf_vic));
1553 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1560 /* parse the layout of victim file */
1561 lmm = buf_vic->lb_buf;
1562 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1563 GOTO(out, rc = -EINVAL);
1565 /* save EA of target file for restore */
1566 memset(buf, 0, sizeof(*buf));
1567 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1571 /* Get rid of the layout from victim object */
1572 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1576 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1577 LU_XATTR_MERGE, handle);
1581 rc = mdd_trans_start(env, mdd, handle);
1585 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1590 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1592 GOTO(out_restore, rc);
1594 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1595 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1600 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1601 LU_XATTR_REPLACE, handle);
1603 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1604 mdd_obj_dev_name(obj),
1605 PFID(mdd_object_fid(obj)), rc2);
1609 mdd_trans_stop(env, mdd, rc, handle);
1610 mdd_write_unlock(env, obj);
1611 mdd_write_unlock(env, vic);
1613 lu_buf_free(buf_vic);
1616 (void) mdd_object_pfid_replace(env, obj);
1622 * Extract the mirror with specified mirror id, and store the splitted
1623 * mirror layout to @buf.
1625 * \param[in] comp_v1 mirrored layout
1626 * \param[in] mirror_id the mirror with mirror_id to be extracted
1627 * \param[out] buf store the layout excluding the extracted mirror,
1628 * caller free the buffer we allocated in this function
1629 * \param[out] buf_vic store the extracted layout, caller free the buffer
1630 * we allocated in this function
1632 * \retval 0 on success; < 0 if error happens
1634 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1635 struct lu_buf *buf, struct lu_buf *buf_vic)
1637 struct lov_comp_md_v1 *comp_rem;
1638 struct lov_comp_md_v1 *comp_vic;
1639 struct lov_comp_md_entry_v1 *entry;
1640 struct lov_comp_md_entry_v1 *entry_rem;
1641 struct lov_comp_md_entry_v1 *entry_vic;
1643 __u16 comp_cnt, count = 0;
1644 int lmm_size, lmm_size_vic = 0;
1646 int offset, offset_rem, offset_vic;
1648 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1649 /* comp_v1 should contains more than 1 mirror */
1650 if (mirror_cnt <= 1)
1652 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1653 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1655 for (i = 0; i < comp_cnt; i++) {
1656 entry = &comp_v1->lcm_entries[i];
1657 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1659 lmm_size_vic += sizeof(*entry);
1660 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1661 } else if (count > 0) {
1662 /* find the specified mirror */
1670 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1674 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1675 if (!buf_vic->lb_buf) {
1680 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1681 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1683 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1684 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1685 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1686 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1687 if (!comp_rem->lcm_mirror_count)
1688 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1690 memset(comp_vic, 0, sizeof(*comp_v1));
1691 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1692 comp_vic->lcm_mirror_count = 0;
1693 comp_vic->lcm_entry_count = cpu_to_le32(count);
1694 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1695 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1696 comp_vic->lcm_layout_gen = 0;
1698 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1699 offset_rem = sizeof(*comp_rem) +
1700 sizeof(*entry_rem) * (comp_cnt - count);
1701 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1702 for (i = j = k = 0; i < comp_cnt; i++) {
1703 struct lov_mds_md *lmm, *lmm_dst;
1706 entry = &comp_v1->lcm_entries[i];
1707 entry_vic = &comp_vic->lcm_entries[j];
1708 entry_rem = &comp_rem->lcm_entries[k];
1710 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1713 /* copy component entry */
1715 memcpy(entry_vic, entry, sizeof(*entry));
1716 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1717 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1720 memcpy(entry_rem, entry, sizeof(*entry));
1721 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1725 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1727 lmm_dst = (struct lov_mds_md *)
1728 ((char *)comp_vic + offset_vic);
1730 lmm_dst = (struct lov_mds_md *)
1731 ((char *)comp_rem + offset_rem);
1733 /* copy component entry blob */
1734 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1736 /* blob offset advance */
1737 offset += le32_to_cpu(entry->lcme_size);
1739 offset_vic += le32_to_cpu(entry->lcme_size);
1741 offset_rem += le32_to_cpu(entry->lcme_size);
1747 static int mdd_dom_data_truncate(const struct lu_env *env,
1748 struct mdd_device *mdd, struct mdd_object *mo);
1750 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1751 struct md_rejig_data *mrd)
1753 struct mdd_device *mdd = mdo2mdd(md_obj);
1754 struct mdd_object *obj = md2mdd_obj(md_obj);
1755 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1756 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1757 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1758 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1759 struct lov_comp_md_v1 *lcm;
1760 struct thandle *handle;
1762 bool dom_stripe = false;
1766 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1767 if (rc == 0) /* same fid */
1770 handle = mdd_trans_create(env, mdd);
1772 RETURN(PTR_ERR(handle));
1775 mdd_write_lock(env, obj, DT_TGT_CHILD);
1776 mdd_write_lock(env, vic, DT_TGT_CHILD);
1778 mdd_write_lock(env, vic, DT_TGT_CHILD);
1779 mdd_write_lock(env, obj, DT_TGT_CHILD);
1782 /* get EA of mirrored file */
1783 memset(buf_save, 0, sizeof(*buf));
1784 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1788 lcm = buf_save->lb_buf;
1789 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1790 GOTO(out, rc = -EINVAL);
1793 * Extract the mirror with specified mirror id, and store the splitted
1794 * mirror layout to the victim file.
1796 memset(buf, 0, sizeof(*buf));
1797 memset(buf_vic, 0, sizeof(*buf_vic));
1798 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1802 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1804 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1805 LU_XATTR_SPLIT, handle);
1808 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1809 LU_XATTR_SPLIT, handle);
1813 rc = mdd_trans_start(env, mdd, handle);
1817 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1822 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1825 GOTO(out_restore, rc);
1827 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1831 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1838 /* restore obj's layout */
1839 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1840 LU_XATTR_REPLACE, handle);
1842 CERROR("%s: failed rollback "DFID" layout: file state unkonwn: rc = %d\n",
1843 mdd_obj_dev_name(obj),
1844 PFID(mdd_object_fid(obj)), rc2);
1847 rc = mdd_trans_stop(env, mdd, rc, handle);
1849 /* Truncate local DOM data if all went well */
1850 if (!rc && dom_stripe)
1851 mdd_dom_data_truncate(env, mdd, obj);
1853 mdd_write_unlock(env, obj);
1854 mdd_write_unlock(env, vic);
1855 lu_buf_free(buf_save);
1857 lu_buf_free(buf_vic);
1860 (void) mdd_object_pfid_replace(env, obj);
1865 static int mdd_layout_merge_allowed(const struct lu_env *env,
1866 struct md_object *target,
1867 struct md_object *victim)
1869 struct mdd_object *o1 = md2mdd_obj(target);
1871 /* cannot extend directory's LOVEA */
1872 if (S_ISDIR(mdd_object_type(o1))) {
1873 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1874 mdd_obj_dev_name(o1));
1882 * The caller should guarantee to update the object ctime
1883 * after xattr_set if needed.
1885 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1886 const struct lu_buf *buf, const char *name,
1889 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1890 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1891 struct mdd_device *mdd = mdo2mdd(obj);
1892 struct thandle *handle;
1893 enum changelog_rec_type cl_type;
1894 enum changelog_rec_flags clf_flags = 0;
1898 rc = mdd_la_get(env, mdd_obj, attr);
1902 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1906 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1907 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1908 struct md_rejig_data *mrd = buf->lb_buf;
1909 struct md_object *victim = mrd->mrd_obj;
1911 if (buf->lb_len != sizeof(*mrd))
1914 rc = mdd_layout_merge_allowed(env, obj, victim);
1918 if (fl == LU_XATTR_MERGE)
1919 /* merge layout of victim as a mirror of obj's. */
1920 rc = mdd_xattr_merge(env, obj, victim);
1922 rc = mdd_xattr_split(env, obj, mrd);
1926 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1927 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1928 struct posix_acl *acl;
1930 /* user may set empty ACL, which should be treated as removing
1932 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1935 RETURN(PTR_ERR(acl));
1937 rc = mdd_xattr_del(env, obj, name);
1940 posix_acl_release(acl);
1943 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1944 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1948 handle = mdd_trans_create(env, mdd);
1950 RETURN(PTR_ERR(handle));
1952 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1956 rc = mdd_trans_start(env, mdd, handle);
1960 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1962 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1963 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
1965 mdd_write_unlock(env, mdd_obj);
1970 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1971 mdd_write_unlock(env, mdd_obj);
1975 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1979 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
1980 mdd_obj, name, handle);
1984 return mdd_trans_stop(env, mdd, rc, handle);
1987 static int mdd_declare_xattr_del(const struct lu_env *env,
1988 struct mdd_device *mdd,
1989 struct mdd_object *obj,
1991 struct thandle *handle)
1993 enum changelog_rec_type type;
1996 rc = mdo_declare_xattr_del(env, obj, name, handle);
2000 type = mdd_xattr_changelog_type(env, mdd, name);
2002 return 0; /* no changelog to store */
2004 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2008 * The caller should guarantee to update the object ctime
2009 * after xattr_set if needed.
2011 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2014 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2015 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2016 struct mdd_device *mdd = mdo2mdd(obj);
2017 struct thandle *handle;
2021 rc = mdd_la_get(env, mdd_obj, attr);
2025 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2029 handle = mdd_trans_create(env, mdd);
2031 RETURN(PTR_ERR(handle));
2033 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2037 rc = mdd_trans_start(env, mdd, handle);
2041 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2042 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2043 mdd_write_unlock(env, mdd_obj);
2047 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2050 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2055 return mdd_trans_stop(env, mdd, rc, handle);
2059 * read lov/lmv EA of an object
2060 * return the lov/lmv EA in an allocated lu_buf
2062 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2063 struct lu_buf *lmm_buf, const char *name)
2065 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
2070 if (buf->lb_buf == NULL) {
2071 buf = lu_buf_check_and_alloc(buf, 4096);
2072 if (buf->lb_buf == NULL)
2077 rc = mdo_xattr_get(env, obj, buf, name);
2078 if (rc == -ERANGE) {
2079 /* mti_big_buf is allocated but is too small
2080 * we need to increase it */
2081 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
2083 if (buf->lb_buf == NULL)
2086 } else if (rc < 0) {
2088 } else if (rc == 0) {
2092 lu_buf_alloc(lmm_buf, rc);
2093 if (lmm_buf->lb_buf == NULL)
2097 * we don't use lmm_buf directly, because we don't know xattr size, so
2098 * by using mti_big_buf we can avoid calling mdo_xattr_get() twice.
2100 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2105 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2106 struct mdd_object *o, struct lu_buf *buf,
2107 struct thandle *handle)
2109 struct hsm_attrs *attrs;
2110 enum hsm_states hsm_flags;
2111 enum changelog_rec_flags clf_flags = 0;
2115 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2120 attrs = buf->lb_buf;
2121 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2122 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2125 /* Add a changelog record for release. */
2126 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2127 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2128 clf_flags, o, handle);
2133 * check if layout swapping between 2 objects is allowed
2135 * - only normal FIDs or non-system IGIFs
2136 * - same type of objects
2137 * - same owner/group (so quotas are still valid) unless this is from HSM
2140 static int mdd_layout_swap_allowed(const struct lu_env *env,
2141 struct mdd_object *o1,
2142 const struct lu_attr *attr1,
2143 struct mdd_object *o2,
2144 const struct lu_attr *attr2,
2147 const struct lu_fid *fid1, *fid2;
2150 fid1 = mdd_object_fid(o1);
2151 fid2 = mdd_object_fid(o2);
2153 if (!fid_is_norm(fid1) &&
2154 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2157 if (!fid_is_norm(fid2) &&
2158 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2161 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2162 if (S_ISDIR(mdd_object_type(o1)))
2164 if (S_ISREG(mdd_object_type(o1)))
2169 if (flags & SWAP_LAYOUTS_MDS_HSM)
2172 if ((attr1->la_uid != attr2->la_uid) ||
2173 (attr1->la_gid != attr2->la_gid))
2179 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2180 * look into the layout in MDD layer. */
2181 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2183 struct lov_comp_md_v1 *comp_v1;
2184 struct lov_mds_md *v1;
2188 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2189 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2190 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2196 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2197 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2200 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2201 off = le32_to_cpu(comp_v1->lcm_entries[i].
2203 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2208 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2209 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2220 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2222 return mdd_lmm_oi(lmm, oi, true);
2225 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2227 return mdd_lmm_oi(lmm, oi, false);
2230 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2232 struct lov_comp_md_v1 *comp_v1;
2234 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2235 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2237 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2239 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2240 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2241 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2242 __u16 tmp_gen = *gen;
2244 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2246 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2253 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2255 return mdd_lmm_gen(lmm, gen, true);
2258 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2260 return mdd_lmm_gen(lmm, gen, false);
2263 static int mdd_dom_data_truncate(const struct lu_env *env,
2264 struct mdd_device *mdd, struct mdd_object *mo)
2267 struct dt_object *dom;
2270 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2272 GOTO(out, rc = -ENODATA);
2274 th = dt_trans_create(env, mdd->mdd_bottom);
2276 GOTO(out, rc = PTR_ERR(th));
2278 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2282 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2286 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2288 dt_trans_stop(env, mdd->mdd_bottom, th);
2290 /* Ignore failure but report the error */
2292 CERROR("%s: can't truncate DOM inode "DFID" data: rc = %d\n",
2293 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2298 * swap layouts between 2 lustre objects
2300 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2301 struct md_object *obj2, __u64 flags)
2303 struct mdd_thread_info *info = mdd_env_info(env);
2304 struct mdd_object *fst_o = md2mdd_obj(obj1);
2305 struct mdd_object *snd_o = md2mdd_obj(obj2);
2306 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2307 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2308 struct mdd_device *mdd = mdo2mdd(obj1);
2309 struct lov_mds_md *fst_lmm, *snd_lmm;
2310 struct lu_buf *fst_buf = &info->mti_buf[0];
2311 struct lu_buf *snd_buf = &info->mti_buf[1];
2312 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
2313 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
2314 struct ost_id *saved_oi = NULL;
2315 struct thandle *handle;
2316 struct mdd_object *dom_o = NULL;
2317 __u64 domsize_dom, domsize_vlt;
2318 __u32 fst_gen, snd_gen, saved_gen;
2324 BUILD_BUG_ON(ARRAY_SIZE(info->mti_buf) < 4);
2325 memset(info->mti_buf, 0, sizeof(info->mti_buf));
2327 /* we have to sort the 2 obj, so locking will always
2328 * be in the same order, even in case of 2 concurrent swaps */
2329 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2330 if (rc == 0) /* same fid ? */
2336 rc = mdd_la_get(env, fst_o, fst_la);
2340 rc = mdd_la_get(env, snd_o, snd_la);
2344 /* check if layout swapping is allowed */
2345 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2349 handle = mdd_trans_create(env, mdd);
2351 RETURN(PTR_ERR(handle));
2353 /* objects are already sorted */
2354 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2355 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2357 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2358 if (rc < 0 && rc != -ENODATA)
2361 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2362 if (rc < 0 && rc != -ENODATA)
2365 /* check if file has DoM. DoM file can be migrated only to another
2366 * DoM layout with the same DoM component size or to an non-DOM
2367 * layout. After migration to OSTs layout, local MDT inode data
2368 * should be truncated.
2369 * Objects are sorted by FIDs, considering that original file's FID
2370 * is always smaller the snd_o is always original file we are migrating
2373 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2374 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2376 /* Only migration is supported for DoM files, not 'swap_layouts' so
2377 * target file must be volatile and orphan.
2379 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2380 dom_o = domsize_dom ? snd_o : NULL;
2381 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2382 swap(domsize_dom, domsize_vlt);
2383 dom_o = domsize_dom ? fst_o : NULL;
2384 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2385 /* 'lfs swap_layouts' case, neither file should have DoM */
2387 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2388 "use migration instead: rc = %d\n", rc);
2392 if (domsize_vlt > 0 && domsize_dom == 0) {
2395 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2396 mdd_obj_dev_name(snd_o),
2397 PFID(mdd_object_fid(snd_o)), rc);
2399 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2402 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2403 mdd_obj_dev_name(fst_o),
2404 PFID(mdd_object_fid(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: HSM error restoring "DFID": rc = %d/%d\n",
2538 mdd_obj_dev_name(fst_o),
2539 PFID(mdd_object_fid(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 of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
2599 mdd_obj_dev_name(fst_o),
2600 PFID(mdd_object_fid(snd_o)),
2601 PFID(mdd_object_fid(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;
2943 if (S_ISDIR(mdd_object_type(obj))) {
2944 switch (mlc->mlc_opc) {
2945 case MD_LAYOUT_SHRINK:
2946 rc = mdd_dir_layout_shrink(env, o, mlc);
2955 /* Verify acceptable operations */
2956 switch (mlc->mlc_opc) {
2957 case MD_LAYOUT_WRITE:
2958 case MD_LAYOUT_RESYNC:
2959 case MD_LAYOUT_RESYNC_DONE:
2965 handle = mdd_trans_create(env, mdd);
2967 RETURN(PTR_ERR(handle));
2969 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
2976 /* analyze the layout to make sure it's a FLR file */
2978 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2979 GOTO(out, rc = -EINVAL);
2981 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2983 /* please refer to HLD of FLR for state transition */
2984 switch (flr_state) {
2986 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2988 case LCM_FL_WRITE_PENDING:
2989 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2992 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2994 case LCM_FL_SYNC_PENDING:
2995 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3004 mdd_trans_stop(env, mdd, rc, handle);
3009 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3010 struct mdd_object *child, const struct lu_attr *attr,
3011 const struct md_op_spec *spec,
3012 struct dt_allocation_hint *hint)
3014 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3015 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3016 struct dt_object *nc = mdd_object_child(child);
3018 memset(hint, 0, sizeof(*hint));
3020 /* For striped directory, give striping EA to lod_ah_init, which will
3021 * decide the stripe_offset and stripe count by it. */
3022 if (S_ISDIR(attr->la_mode) &&
3023 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3024 hint->dah_eadata = spec->u.sp_ea.eadata;
3025 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3027 hint->dah_eadata = NULL;
3028 hint->dah_eadata_len = 0;
3029 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
3030 if (mdd->mdd_append_stripe_count != 0 ||
3031 mdd->mdd_append_pool[0])
3033 "using O_APPEND file striping\n");
3034 if (mdd->mdd_append_stripe_count)
3035 hint->dah_append_stripes =
3036 mdd->mdd_append_stripe_count;
3037 if (mdd->mdd_append_pool[0])
3038 hint->dah_append_pool = mdd->mdd_append_pool;
3040 hint->dah_append_stripes = 0;
3044 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3045 hint->dah_eadata, hint->dah_eadata_len);
3046 /* @hint will be initialized by underlying device. */
3047 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3050 static int accmode(const struct lu_env *env, const struct lu_attr *la,
3053 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3054 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3055 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3056 * owner can write to a file even if it is marked readonly to hide
3057 * its brokenness. (bug 5781) */
3058 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3059 struct lu_ucred *uc = lu_ucred_check(env);
3061 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3065 return mds_accmode(open_flags);
3068 static int mdd_open_sanity_check(const struct lu_env *env,
3069 struct mdd_object *obj,
3070 const struct lu_attr *attr, u64 open_flags,
3076 /* EEXIST check, also opening of *open* orphans is allowed so we can
3077 * open-by-handle unlinked files
3079 if (mdd_is_dead_obj(obj) && !is_replay &&
3080 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3083 if (S_ISLNK(attr->la_mode))
3086 mode = accmode(env, attr, open_flags);
3088 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
3091 if (!(open_flags & MDS_OPEN_CREATED)) {
3092 rc = mdd_permission_internal(env, obj, attr, mode);
3097 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3098 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3099 open_flags &= ~MDS_OPEN_TRUNC;
3101 /* For writing append-only file must open it with append mode. */
3102 if (attr->la_flags & LUSTRE_APPEND_FL) {
3103 if ((open_flags & MDS_FMODE_WRITE) &&
3104 !(open_flags & MDS_OPEN_APPEND))
3106 if (open_flags & MDS_OPEN_TRUNC)
3113 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3114 u64 open_flags, struct md_op_spec *spec)
3116 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3117 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3118 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3119 struct mdd_object_user *mou = NULL;
3120 const struct lu_ucred *uc = lu_ucred(env);
3121 struct mdd_device *mdd = mdo2mdd(obj);
3122 enum changelog_rec_type type = CL_OPEN;
3126 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3128 rc = mdd_la_get(env, mdd_obj, attr);
3132 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3134 if ((rc == -EACCES) && (mdd->mdd_cl.mc_mask & (1 << CL_DN_OPEN)))
3139 mdd_obj->mod_count++;
3141 if (!mdd_changelog_enabled(env, mdd, type))
3145 /* look for existing opener in list under mdd_write_lock */
3146 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3151 /* add user to list */
3152 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3158 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3160 mdd_obj_user_free(mou);
3165 if (type == CL_DN_OPEN) {
3166 if (ktime_before(ktime_get(), mou->mou_deniednext))
3167 /* same user denied again same access within
3168 * time interval: do not record
3172 /* this user already denied, but some time ago:
3173 * update denied time
3175 mou->mou_deniednext =
3176 ktime_add(ktime_get(),
3177 ktime_set(mdd->mdd_cl.mc_deniednext,
3180 mou->mou_opencount++;
3181 /* same user opening file again with same flags:
3188 /* FYI, only the bottom 32 bits of open_flags are recorded */
3189 mdd_changelog(env, type, open_flags, md_dev, mdd_object_fid(mdd_obj));
3193 mdd_write_unlock(env, mdd_obj);
3197 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3198 struct md_attr *ma, struct thandle *handle)
3202 rc = mdd_orphan_declare_delete(env, obj, handle);
3206 return mdo_declare_destroy(env, obj, handle);
3210 * No permission check is needed.
3212 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3213 struct md_attr *ma, u64 open_flags)
3215 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3216 struct mdd_device *mdd = mdo2mdd(obj);
3217 struct thandle *handle = NULL;
3220 bool blocked = false;
3221 bool last_close_by_uid = false;
3222 const struct lu_ucred *uc = lu_ucred(env);
3225 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3226 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3227 mdd_obj->mod_count--;
3228 mdd_write_unlock(env, mdd_obj);
3230 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3231 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3232 "list\n", PFID(mdd_object_fid(mdd_obj)));
3236 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3237 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3238 /* check without any lock */
3239 is_orphan = mdd_obj->mod_count == 1 &&
3240 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3244 /* mdd_trans_create() maybe failed because of barrier_entry(),
3245 * under such case, the orphan MDT-object will be left in the
3246 * orphan list, and when the MDT remount next time, the unused
3247 * orphans will be destroyed automatically.
3249 * One exception: the former mdd_finish_unlink may failed to
3250 * add the orphan MDT-object to the orphan list, then if the
3251 * mdd_trans_create() failed because of barrier_entry(), the
3252 * MDT-object will become real orphan that is neither in the
3253 * namespace nor in the orphan list. Such bad case should be
3254 * very rare and will be handled by e2fsck/lfsck. */
3255 handle = mdd_trans_create(env, mdo2mdd(obj));
3256 if (IS_ERR(handle)) {
3257 rc = PTR_ERR(handle);
3258 if (rc != -EINPROGRESS)
3266 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3270 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3275 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3281 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3282 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3284 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3285 lu_dev_name(mdd2lu_dev(mdd)),
3286 PFID(mdd_object_fid(mdd_obj)), rc);
3290 /* check again with lock */
3291 is_orphan = (mdd_obj->mod_count == 1) &&
3292 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3293 ma->ma_attr.la_nlink == 0);
3295 if (is_orphan && !handle && !blocked) {
3296 mdd_write_unlock(env, mdd_obj);
3300 mdd_obj->mod_count--; /*release open count */
3302 /* under mdd write lock */
3303 /* If recording, see if we need to remove UID from list. uc is not
3304 * initialized if the client has been evicted. */
3305 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3306 struct mdd_object_user *mou;
3308 /* look for UID in list */
3309 /* If mou is NULL, it probably means logging was enabled after
3310 * the user had the file open. So the corresponding close
3311 * will not be logged.
3313 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3316 mou->mou_opencount--;
3317 if (mou->mou_opencount == 0) {
3318 mdd_obj_user_remove(mdd_obj, mou);
3319 last_close_by_uid = true;
3324 if (!is_orphan || blocked)
3328 /* NB: Object maybe not in orphan list originally, it is rare case for
3329 * mdd_finish_unlink() failure, in that case, the object doesn't have
3330 * ORPHAN_OBJ flag */
3331 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3332 /* remove link to object from orphan index */
3333 LASSERT(handle != NULL);
3334 rc = mdd_orphan_delete(env, mdd_obj, handle);
3336 CERROR("%s: unable to delete "DFID" from orphan list: "
3337 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3338 PFID(mdd_object_fid(mdd_obj)), rc);
3339 /* If object was not deleted from orphan list, do not
3340 * destroy OSS objects, which will be done when next
3345 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3346 "list, OSS objects to be destroyed.\n",
3347 PFID(mdd_object_fid(mdd_obj)));
3350 rc = mdo_destroy(env, mdd_obj, handle);
3353 CERROR("%s: unable to delete "DFID" from orphan list: "
3354 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3355 PFID(mdd_object_fid(mdd_obj)), rc);
3360 mdd_write_unlock(env, mdd_obj);
3362 if (rc != 0 || blocked ||
3363 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3366 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3367 * or if CL_OPEN was recorded and it's last close by user.
3368 * Changelogs mask may change between open and close operations, but
3369 * this is not a big deal if we have a CL_CLOSE entry with no matching
3370 * CL_OPEN. Plus Changelogs mask may not change often.
3372 if (((!(mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) &&
3373 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3374 MDS_OPEN_TRUNC))) ||
3375 ((mdd->mdd_cl.mc_mask & (1 << CL_OPEN)) && last_close_by_uid)) &&
3376 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3377 if (handle == NULL) {
3378 handle = mdd_trans_create(env, mdo2mdd(obj));
3380 GOTO(stop, rc = PTR_ERR(handle));
3382 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3383 NULL, NULL, handle);
3387 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3392 /* FYI, only the bottom 32 bits of open_flags are recorded */
3393 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3398 if (handle != NULL && !IS_ERR(handle))
3399 rc = mdd_trans_stop(env, mdd, rc, handle);
3405 * Permission check is done when open,
3406 * no need check again.
3408 static int mdd_readpage_sanity_check(const struct lu_env *env,
3409 struct mdd_object *obj)
3411 struct dt_object *next = mdd_object_child(obj);
3415 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3423 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3424 size_t nob, const struct dt_it_ops *iops,
3425 struct dt_it *it, __u32 attr, void *arg)
3427 struct lu_dirpage *dp = &lp->lp_dir;
3431 struct lu_dirent *ent;
3432 struct lu_dirent *last = NULL;
3436 if (nob < sizeof(*dp))
3439 memset(area, 0, sizeof (*dp));
3440 area += sizeof (*dp);
3441 nob -= sizeof (*dp);
3448 len = iops->key_size(env, it);
3450 /* IAM iterator can return record with zero len. */
3454 hash = iops->store(env, it);
3455 if (unlikely(first)) {
3457 dp->ldp_hash_start = cpu_to_le64(hash);
3460 /* calculate max space required for lu_dirent */
3461 recsize = lu_dirent_calc_size(len, attr);
3463 if (nob >= recsize) {
3464 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3465 if (result == -ESTALE)
3470 /* osd might not able to pack all attributes,
3471 * so recheck rec length */
3472 recsize = le16_to_cpu(ent->lde_reclen);
3474 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3475 fid_le_to_cpu(&fid, &ent->lde_fid);
3476 if (fid_is_dot_lustre(&fid))
3480 result = (last != NULL) ? 0 :-EINVAL;
3484 ent = (void *)ent + recsize;
3488 result = iops->next(env, it);
3489 if (result == -ESTALE)
3491 } while (result == 0);
3494 dp->ldp_hash_end = cpu_to_le64(hash);
3496 if (last->lde_hash == dp->ldp_hash_end)
3497 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3498 last->lde_reclen = 0; /* end mark */
3501 /* end of directory */
3502 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3503 else if (result < 0)
3504 CWARN("build page failed: %d!\n", result);
3508 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3509 const struct lu_rdpg *rdpg)
3511 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3515 if (mdd_object_exists(mdd_obj) == 0) {
3516 CERROR("%s: object "DFID" not found: rc = -2\n",
3517 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3521 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3522 rc = mdd_readpage_sanity_check(env, mdd_obj);
3524 GOTO(out_unlock, rc);
3526 if (mdd_is_dead_obj(mdd_obj)) {
3528 struct lu_dirpage *dp;
3531 * According to POSIX, please do not return any entry to client:
3532 * even dot and dotdot should not be returned.
3534 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3535 PFID(mdd_object_fid(mdd_obj)));
3537 if (rdpg->rp_count <= 0)
3538 GOTO(out_unlock, rc = -EFAULT);
3539 LASSERT(rdpg->rp_pages != NULL);
3541 pg = rdpg->rp_pages[0];
3542 dp = (struct lu_dirpage *)kmap(pg);
3543 memset(dp, 0 , sizeof(struct lu_dirpage));
3544 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3545 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3546 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3548 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3551 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3552 mdd_dir_page_build, NULL);
3554 struct lu_dirpage *dp;
3556 dp = kmap(rdpg->rp_pages[0]);
3557 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3560 * No pages were processed, mark this for first page
3563 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3564 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3565 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3567 kunmap(rdpg->rp_pages[0]);
3570 GOTO(out_unlock, rc);
3572 mdd_read_unlock(env, mdd_obj);
3576 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3578 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3580 if (mdd_object_exists(mdd_obj) == 0) {
3583 CERROR("%s: object "DFID" not found: rc = %d\n",
3584 mdd_obj_dev_name(mdd_obj),
3585 PFID(mdd_object_fid(mdd_obj)), rc);
3588 return dt_object_sync(env, mdd_object_child(mdd_obj),
3592 static int mdd_object_lock(const struct lu_env *env,
3593 struct md_object *obj,
3594 struct lustre_handle *lh,
3595 struct ldlm_enqueue_info *einfo,
3596 union ldlm_policy_data *policy)
3598 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3599 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3603 static int mdd_object_unlock(const struct lu_env *env,
3604 struct md_object *obj,
3605 struct ldlm_enqueue_info *einfo,
3606 union ldlm_policy_data *policy)
3608 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3609 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3612 const struct md_object_operations mdd_obj_ops = {
3613 .moo_permission = mdd_permission,
3614 .moo_attr_get = mdd_attr_get,
3615 .moo_attr_set = mdd_attr_set,
3616 .moo_xattr_get = mdd_xattr_get,
3617 .moo_xattr_set = mdd_xattr_set,
3618 .moo_xattr_list = mdd_xattr_list,
3619 .moo_invalidate = mdd_invalidate,
3620 .moo_xattr_del = mdd_xattr_del,
3621 .moo_swap_layouts = mdd_swap_layouts,
3622 .moo_open = mdd_open,
3623 .moo_close = mdd_close,
3624 .moo_readpage = mdd_readpage,
3625 .moo_readlink = mdd_readlink,
3626 .moo_changelog = mdd_changelog,
3627 .moo_object_sync = mdd_object_sync,
3628 .moo_object_lock = mdd_object_lock,
3629 .moo_object_unlock = mdd_object_unlock,
3630 .moo_layout_change = mdd_layout_change,