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/
31 * lustre/mdd/mdd_object.c
33 * Lustre Metadata Server (mdd) routines
35 * Author: Wang Di <wangdi@clusterfs.com>
38 #define DEBUG_SUBSYSTEM S_MDS
40 #include <linux/module.h>
42 #include <obd_class.h>
43 #include <obd_support.h>
44 #include <lprocfs_status.h>
45 /* fid_be_cpu(), fid_cpu_to_be(). */
46 #include <lustre_fid.h>
47 #include <lustre_idmap.h>
48 #include <uapi/linux/lustre/lustre_param.h>
49 #include <lustre_mds.h>
51 #include "mdd_internal.h"
53 static const struct lu_object_operations mdd_lu_obj_ops;
55 struct mdd_object_user {
56 struct list_head mou_list; /**< linked off mod_users */
57 u64 mou_open_flags; /**< open mode by client */
58 __u64 mou_uidgid; /**< uid_gid on client */
59 int mou_opencount; /**< # opened */
60 ktime_t mou_deniednext; /**< time of next access denied
65 static int mdd_xattr_get(const struct lu_env *env,
66 struct md_object *obj, struct lu_buf *buf,
69 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
70 struct mdd_device *mdd,
71 enum changelog_rec_type type,
72 enum changelog_rec_flags clf_flags,
73 const struct lu_fid *fid,
74 const struct lu_fid *pfid,
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)->mdi_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)->mdi_buf[0];
265 buf->lb_buf = (void *)area;
270 struct lu_object *mdd_object_alloc(const struct lu_env *env,
271 const struct lu_object_header *hdr,
274 struct mdd_object *mdd_obj;
277 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
281 o = mdd2lu_obj(mdd_obj);
282 lu_object_init(o, NULL, d);
283 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
284 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
285 mdd_obj->mod_count = 0;
286 o->lo_ops = &mdd_lu_obj_ops;
287 INIT_LIST_HEAD(&mdd_obj->mod_users);
292 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
293 const struct lu_object_conf *unused)
295 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
296 struct mdd_object *mdd_obj = lu2mdd_obj(o);
297 struct lu_object *below;
298 struct lu_device *under;
301 mdd_obj->mod_cltime = ktime_set(0, 0);
302 under = &d->mdd_child->dd_lu_dev;
303 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
305 RETURN(PTR_ERR(below));
307 lu_object_add(o, below);
312 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
316 if (lu_object_exists(o)) {
317 struct mdd_object *mdd_obj = lu2mdd_obj(o);
318 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
320 rc = mdd_la_get(env, mdd_obj, attr);
326 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
328 struct mdd_object *mdd = lu2mdd_obj(o);
329 struct mdd_object_user *mou, *tmp2;
332 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
333 list_del(&mou->mou_list);
334 mdd_obj_user_free(mou);
338 /* mdd doesn't contain an lu_object_header, so don't need call_rcu */
339 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
342 static int mdd_object_print(const struct lu_env *env, void *cookie,
343 lu_printer_t p, const struct lu_object *o)
345 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
347 return (*p)(env, cookie,
348 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
349 mdd, mdd->mod_count, mdd->mod_valid,
350 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
353 static const struct lu_object_operations mdd_lu_obj_ops = {
354 .loo_object_init = mdd_object_init,
355 .loo_object_start = mdd_object_start,
356 .loo_object_free = mdd_object_free,
357 .loo_object_print = mdd_object_print,
360 struct mdd_object *mdd_object_find(const struct lu_env *env,
361 struct mdd_device *d,
362 const struct lu_fid *f)
364 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
368 * No permission check is needed.
370 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
373 struct mdd_object *mdd_obj = md2mdd_obj(obj);
378 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
379 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
380 ma->ma_attr.la_nlink = 0;
386 * No permission check is needed.
388 static int mdd_xattr_get(const struct lu_env *env,
389 struct md_object *obj, struct lu_buf *buf,
392 struct mdd_object *mdd_obj = md2mdd_obj(obj);
393 struct mdd_device *mdd;
398 if (mdd_object_exists(mdd_obj) == 0) {
399 CERROR("%s: object "DFID" not found: rc = -2\n",
400 mdd_obj_dev_name(mdd_obj),
401 PFID(mdd_object_fid(mdd_obj)));
405 /* If the object has been destroyed, then do not get LMVEA, because
406 * it needs to load stripes from the iteration of the master object,
407 * and it will cause problem if master object has been destroyed, see
409 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
410 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
411 strcmp(name, XATTR_NAME_LMV) == 0))
414 /* If the object has been delete from the namespace, then
415 * get linkEA should return -ENOENT as well */
416 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
417 strcmp(name, XATTR_NAME_LINK) == 0))
420 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
421 rc = mdo_xattr_get(env, mdd_obj, buf, name);
422 mdd_read_unlock(env, mdd_obj);
426 /* record only getting user xattrs and acls */
427 if (rc >= 0 && buf->lb_buf &&
428 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
429 (has_prefix(name, XATTR_USER_PREFIX) ||
430 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
431 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
432 struct thandle *handle;
435 LASSERT(mdd_object_fid(mdd_obj) != NULL);
437 handle = mdd_trans_create(env, mdd);
439 RETURN(PTR_ERR(handle));
441 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
446 rc2 = mdd_trans_start(env, mdd, handle);
450 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
451 mdd_object_fid(mdd_obj),
455 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
464 * Permission check is done when open,
465 * no need check again.
467 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
470 struct mdd_object *mdd_obj = md2mdd_obj(obj);
471 struct dt_object *next;
476 if (mdd_object_exists(mdd_obj) == 0) {
477 CERROR("%s: object "DFID" not found: rc = -2\n",
478 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
482 next = mdd_object_child(mdd_obj);
483 LASSERT(next != NULL);
484 LASSERT(next->do_body_ops != NULL);
485 LASSERT(next->do_body_ops->dbo_read != NULL);
486 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
487 rc = dt_read(env, next, buf, &pos);
488 mdd_read_unlock(env, mdd_obj);
493 * No permission check is needed.
495 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
498 struct mdd_object *mdd_obj = md2mdd_obj(obj);
503 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
504 rc = mdo_xattr_list(env, mdd_obj, buf);
505 mdd_read_unlock(env, mdd_obj);
507 /* If the buffer is NULL then we are only here to get the
508 * length of the xattr name list. */
509 if (rc < 0 || buf->lb_buf == NULL)
513 * Filter out XATTR_NAME_LINK if this is an orphan object. See
516 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
517 char *end = (char *)buf->lb_buf + rc;
518 char *p = buf->lb_buf;
521 char *next = p + strlen(p) + 1;
523 if (strcmp(p, XATTR_NAME_LINK) == 0) {
525 memmove(p, next, end - next);
527 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
528 " of orphan "DFID"\n",
529 PFID(mdd_object_fid(mdd_obj)));
540 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
542 return mdo_invalidate(env, md2mdd_obj(obj));
545 int mdd_declare_create_object_internal(const struct lu_env *env,
546 struct mdd_object *p,
547 struct mdd_object *c,
548 struct lu_attr *attr,
549 struct thandle *handle,
550 const struct md_op_spec *spec,
551 struct dt_allocation_hint *hint)
553 struct dt_object_format *dof = &mdd_env_info(env)->mdi_dof;
554 const struct dt_index_features *feat = spec->sp_feat;
558 if (feat != &dt_directory_features && feat != NULL) {
559 dof->dof_type = DFT_INDEX;
560 dof->u.dof_idx.di_feat = feat;
562 dof->dof_type = dt_mode_to_dft(attr->la_mode);
563 if (dof->dof_type == DFT_REGULAR) {
564 dof->u.dof_reg.striped =
565 md_should_create(spec->sp_cr_flags);
566 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
567 dof->u.dof_reg.striped = 0;
568 /* is this replay? */
570 dof->u.dof_reg.striped = 0;
574 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
579 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
580 struct mdd_object *c, struct lu_attr *attr,
581 struct thandle *handle,
582 const struct md_op_spec *spec,
583 struct dt_allocation_hint *hint)
585 struct dt_object_format *dof = &mdd_env_info(env)->mdi_dof;
589 LASSERT(!mdd_object_exists(c));
591 rc = mdo_create_object(env, c, attr, hint, dof, handle);
596 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
597 const struct lu_attr *attr, struct thandle *handle,
603 rc = mdo_attr_set(env, obj, attr, handle);
604 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
605 if (!rc && (attr->la_valid & LA_MODE) && needacl)
606 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
611 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
612 const struct lu_attr *oattr, struct lu_attr *attr,
613 struct thandle *handle)
618 LASSERT(attr->la_valid & LA_CTIME);
619 LASSERT(oattr != NULL);
621 /* Make sure the ctime is increased only, however, it's not strictly
622 * reliable at here because there is not guarantee to hold lock on
623 * object, so we just bypass some unnecessary cmtime setting first
624 * and OSD has to check it again. */
625 if (attr->la_ctime < oattr->la_ctime)
626 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
627 else if (attr->la_valid == LA_CTIME &&
628 attr->la_ctime == oattr->la_ctime)
629 attr->la_valid &= ~LA_CTIME;
631 if (attr->la_valid != 0)
632 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
637 static bool is_project_state_change(const struct lu_attr *oattr,
640 if (la->la_valid & LA_PROJID &&
641 oattr->la_projid != la->la_projid)
644 if ((la->la_valid & LA_FLAGS) &&
645 (la->la_flags & LUSTRE_PROJINHERIT_FL) !=
646 (oattr->la_flags & LUSTRE_PROJINHERIT_FL))
653 * This gives the same functionality as the code between
654 * sys_chmod and inode_setattr
655 * chown_common and inode_setattr
656 * utimes and inode_setattr
657 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
659 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
660 const struct lu_attr *oattr, struct lu_attr *la,
661 const struct md_attr *ma)
665 const unsigned long flags = ma->ma_attr_flags;
672 /* Do not permit change file type */
673 if (la->la_valid & LA_TYPE)
676 /* They should not be processed by setattr */
677 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
680 LASSERT(oattr != NULL);
682 uc = lu_ucred_check(env);
686 if (is_project_state_change(oattr, la)) {
687 if (!md_capable(uc, CAP_SYS_RESOURCE) &&
688 !lustre_in_group_p(uc, ma->ma_enable_chprojid_gid) &&
689 !(ma->ma_enable_chprojid_gid == -1 &&
690 mdd_permission_internal(env, obj, oattr, MAY_WRITE)))
694 if (la->la_valid == LA_CTIME) {
695 if (!(flags & MDS_PERM_BYPASS))
696 /* This is only for set ctime when rename's source is
698 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
700 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
701 la->la_valid &= ~LA_CTIME;
705 if (flags & MDS_CLOSE_UPDATE_TIMES &&
706 la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
707 /* This is an atime/mtime/ctime attribute update for
710 if (la->la_valid & LA_ATIME &&
711 la->la_atime <= (oattr->la_atime +
712 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
713 la->la_valid &= ~LA_ATIME;
714 if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime)
715 la->la_valid &= ~LA_CTIME;
716 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
717 la->la_valid &= ~LA_MTIME;
721 /* Check if flags change. */
722 if (la->la_valid & LA_FLAGS) {
723 unsigned int oldflags = oattr->la_flags &
724 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
725 unsigned int newflags = la->la_flags &
726 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
728 if ((uc->uc_fsuid != oattr->la_uid) &&
729 !md_capable(uc, CAP_FOWNER))
732 /* The IMMUTABLE and APPEND_ONLY flags can
733 * only be changed by the relevant capability. */
734 if ((oldflags ^ newflags) &&
735 !md_capable(uc, CAP_LINUX_IMMUTABLE))
738 if (!S_ISDIR(oattr->la_mode)) {
739 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
740 } else if (la->la_flags & LUSTRE_ENCRYPT_FL) {
741 /* when trying to add encryption flag on dir,
742 * make sure it is empty
744 rc = mdd_dir_is_empty(env, obj);
751 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
752 (la->la_valid & ~LA_FLAGS) &&
753 !(flags & MDS_PERM_BYPASS))
756 /* Check for setting the obj time. */
757 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
758 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
759 if ((uc->uc_fsuid != oattr->la_uid) &&
760 !md_capable(uc, CAP_FOWNER)) {
761 rc = mdd_permission_internal(env, obj, oattr,
768 if (la->la_valid & LA_KILL_SUID) {
769 la->la_valid &= ~LA_KILL_SUID;
770 if ((oattr->la_mode & S_ISUID) &&
771 !(la->la_valid & LA_MODE)) {
772 la->la_mode = oattr->la_mode;
773 la->la_valid |= LA_MODE;
775 la->la_mode &= ~S_ISUID;
778 if (la->la_valid & LA_KILL_SGID) {
779 la->la_valid &= ~LA_KILL_SGID;
780 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
781 (S_ISGID | S_IXGRP)) &&
782 !(la->la_valid & LA_MODE)) {
783 la->la_mode = oattr->la_mode;
784 la->la_valid |= LA_MODE;
786 la->la_mode &= ~S_ISGID;
789 /* Make sure a caller can chmod. */
790 if (la->la_valid & LA_MODE) {
791 if (!(flags & MDS_PERM_BYPASS) &&
792 (uc->uc_fsuid != oattr->la_uid) &&
793 !md_capable(uc, CAP_FOWNER))
796 if (la->la_mode == (umode_t) -1)
797 la->la_mode = oattr->la_mode;
799 la->la_mode = (la->la_mode & S_IALLUGO) |
800 (oattr->la_mode & ~S_IALLUGO);
802 /* Also check the setgid bit! */
803 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
804 la->la_gid : oattr->la_gid) &&
805 !md_capable(uc, CAP_FSETID))
806 la->la_mode &= ~S_ISGID;
808 la->la_mode = oattr->la_mode;
811 /* Make sure a caller can chown. */
812 if (la->la_valid & LA_UID) {
813 if (la->la_uid == (uid_t) -1)
814 la->la_uid = oattr->la_uid;
815 if (((uc->uc_fsuid != oattr->la_uid) ||
816 (la->la_uid != oattr->la_uid)) &&
817 !md_capable(uc, CAP_CHOWN))
820 /* If the user or group of a non-directory has been
821 * changed by a non-root user, remove the setuid bit.
822 * 19981026 David C Niemi <niemi@tux.org>
824 * Changed this to apply to all users, including root,
825 * to avoid some races. This is the behavior we had in
826 * 2.0. The check for non-root was definitely wrong
827 * for 2.2 anyway, as it should have been using
828 * CAP_FSETID rather than fsuid -- 19990830 SD. */
829 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
830 !S_ISDIR(oattr->la_mode)) {
831 la->la_mode &= ~S_ISUID;
832 la->la_valid |= LA_MODE;
836 /* Make sure caller can chgrp. */
837 if (la->la_valid & LA_GID) {
838 if (la->la_gid == (gid_t) -1)
839 la->la_gid = oattr->la_gid;
840 if (((uc->uc_fsuid != oattr->la_uid) ||
841 ((la->la_gid != oattr->la_gid) &&
842 !lustre_in_group_p(uc, la->la_gid))) &&
843 !md_capable(uc, CAP_CHOWN))
846 /* Likewise, if the user or group of a non-directory
847 * has been changed by a non-root user, remove the
848 * setgid bit UNLESS there is no group execute bit
849 * (this would be a file marked for mandatory
850 * locking). 19981026 David C Niemi <niemi@tux.org>
852 * Removed the fsuid check (see the comment above) --
854 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
855 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
856 la->la_mode &= ~S_ISGID;
857 la->la_valid |= LA_MODE;
861 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
862 if (!((flags & MDS_OWNEROVERRIDE) &&
863 (uc->uc_fsuid == oattr->la_uid)) &&
864 !(flags & MDS_PERM_BYPASS)) {
865 rc = mdd_permission_internal(env, obj, oattr,
872 if (la->la_valid & LA_CTIME) {
874 * The pure setattr, it has the priority over what is
875 * already set, do not drop it if ctime is equal.
877 if (la->la_ctime < oattr->la_ctime)
878 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
884 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
885 struct mdd_device *mdd,
886 enum changelog_rec_type type,
887 enum changelog_rec_flags clf_flags,
888 const struct lu_fid *fid,
889 const struct lu_fid *pfid,
890 const char *xattr_name,
891 struct thandle *handle)
893 const struct lu_ucred *uc = lu_ucred(env);
894 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
895 struct llog_changelog_rec *rec;
900 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
903 if (uc->uc_jobid[0] != '\0')
904 clf_flags |= CLF_JOBID;
905 xflags |= CLFE_UIDGID;
908 if (type == CL_OPEN || type == CL_DN_OPEN)
910 if (type == CL_SETXATTR || type == CL_GETXATTR)
911 xflags |= CLFE_XATTR;
913 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
914 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
915 xflags & CLFE_SUPPORTED));
916 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_chlg_buf, reclen);
917 if (buf->lb_buf == NULL)
921 rec->cr_hdr.lrh_len = reclen;
922 rec->cr.cr_flags = clf_flags;
923 rec->cr.cr_type = (__u32)type;
924 rec->cr.cr_tfid = *fid;
926 rec->cr.cr_pfid = *pfid;
927 rec->cr.cr_namelen = 0;
929 if (clf_flags & CLF_JOBID)
930 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
932 if (clf_flags & CLF_EXTRA_FLAGS) {
933 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
934 if (xflags & CLFE_UIDGID)
935 mdd_changelog_rec_extra_uidgid(&rec->cr,
936 uc->uc_uid, uc->uc_gid);
937 if (xflags & CLFE_NID)
938 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
939 if (xflags & CLFE_OPEN)
940 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
941 if (xflags & CLFE_XATTR) {
942 if (xattr_name == NULL)
944 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
948 rc = mdd_changelog_store(env, mdd, rec, handle);
953 /** Store a data change changelog record
954 * If this fails, we must fail the whole transaction; we don't
955 * want the change to commit without the log entry.
956 * \param mdd_obj - mdd_object of change
957 * \param handle - transaction handle
958 * \param pfid - parent FID for CL_MTIME changelogs
960 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
961 enum changelog_rec_type type,
962 enum changelog_rec_flags clf_flags,
963 struct mdd_object *mdd_obj, struct thandle *handle,
964 const struct lu_fid *pfid)
968 LASSERT(mdd_obj != NULL);
969 LASSERT(handle != NULL);
971 if (!mdd_changelog_enabled(env, mdd, type))
974 if (mdd_is_volatile_obj(mdd_obj))
977 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
978 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
979 /* Don't need multiple updates in this log */
980 /* Don't check under lock - no big deal if we get an extra
985 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
986 mdd_object_fid(mdd_obj), pfid,
989 mdd_obj->mod_cltime = ktime_get();
994 int mdd_changelog_data_store_xattr(const struct lu_env *env,
995 struct mdd_device *mdd,
996 enum changelog_rec_type type,
997 enum changelog_rec_flags clf_flags,
998 struct mdd_object *mdd_obj,
999 const char *xattr_name,
1000 struct thandle *handle)
1004 LASSERT(mdd_obj != NULL);
1005 LASSERT(handle != NULL);
1007 if (!mdd_changelog_enabled(env, mdd, type))
1010 if (mdd_is_volatile_obj(mdd_obj))
1013 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1014 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1015 /* Don't need multiple updates in this log */
1016 /* Don't check under lock - no big deal if we get an extra
1022 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1023 mdd_object_fid(mdd_obj), NULL,
1024 xattr_name, handle);
1026 mdd_obj->mod_cltime = ktime_get();
1031 /* only the bottom CLF_FLAGSHIFT bits of @flags are stored in the record,
1032 * except for open flags have a dedicated record to store 32 bits of flags */
1033 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1034 enum changelog_rec_flags clf_flags,
1035 struct md_device *m, const struct lu_fid *fid)
1037 struct thandle *handle;
1038 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1042 LASSERT(fid != NULL);
1044 /* We'll check this again below, but we check now before we
1045 * start a transaction. */
1046 if (!mdd_changelog_enabled(env, mdd, type))
1049 handle = mdd_trans_create(env, mdd);
1051 RETURN(PTR_ERR(handle));
1053 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1057 rc = mdd_trans_start(env, mdd, handle);
1061 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1062 fid, NULL, NULL, handle);
1065 rc = mdd_trans_stop(env, mdd, rc, handle);
1071 * Save LMA extended attributes with data from \a ma.
1073 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1074 * not, LMA EA will be first read from disk, modified and write back.
1077 /* Precedence for choosing record type when multiple
1078 * attributes change: setattr > mtime > ctime > atime
1079 * (ctime changes when mtime does, plus chmod/chown.
1080 * atime and ctime are independent.) */
1081 static int mdd_attr_set_changelog(const struct lu_env *env,
1082 struct md_object *obj, struct thandle *handle,
1083 const struct lu_fid *pfid, __u64 valid)
1085 struct mdd_device *mdd = mdo2mdd(obj);
1088 bits = (valid & LA_SIZE) ? BIT(CL_TRUNC) : 0;
1089 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? BIT(CL_SETATTR) : 0;
1090 bits |= (valid & LA_MTIME) ? BIT(CL_MTIME) : 0;
1091 bits |= (valid & LA_CTIME) ? BIT(CL_CTIME) : 0;
1092 bits |= (valid & LA_ATIME) ? BIT(CL_ATIME) : 0;
1093 bits = bits & mdd->mdd_cl.mc_current_mask;
1094 /* This is an implementation limit rather than a protocol limit */
1095 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1099 /* The record type is the lowest non-masked set bit */
1102 /* XXX: we only store the low CLF_FLAGMASK bits of la_valid */
1103 return mdd_changelog_data_store(env, mdd, type, valid, md2mdd_obj(obj),
1107 static int mdd_declare_attr_set(const struct lu_env *env,
1108 struct mdd_device *mdd,
1109 struct mdd_object *obj,
1110 const struct lu_attr *attr,
1111 struct thandle *handle)
1115 rc = mdo_declare_attr_set(env, obj, attr, handle);
1119 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1120 if (attr->la_valid & LA_MODE) {
1121 mdd_read_lock(env, obj, DT_TGT_CHILD);
1122 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1123 XATTR_NAME_ACL_ACCESS);
1124 mdd_read_unlock(env, obj);
1125 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1131 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1132 rc = mdo_declare_xattr_set(env, obj, buf,
1133 XATTR_NAME_ACL_ACCESS, 0,
1141 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1150 * permission changes may require sync operation, to mitigate performance
1151 * impact, only do this for dir and when permission is reduced.
1153 * For regular files, version is updated with permission change (see VBR), async
1154 * permission won't cause any issue, while missing permission change on
1155 * directory may affect accessibility of other objects after recovery.
1157 static inline bool permission_needs_sync(const struct lu_attr *old,
1158 const struct lu_attr *new)
1160 if (!S_ISDIR(old->la_mode))
1163 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1166 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1169 if (new->la_valid & LA_MODE) {
1170 /* turned on sticky bit */
1171 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1174 /* set-GID has no impact on what is allowed, not checked */
1176 /* turned off setuid bit, or one of rwx for someone */
1177 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1178 (old->la_mode & (0777 | S_ISUID)))
1185 static inline __u64 mdd_lmm_dom_size(void *buf)
1187 struct lov_mds_md *lmm = buf;
1188 struct lov_comp_md_v1 *comp_v1;
1189 struct lov_mds_md *v1;
1195 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1198 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1199 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1200 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1202 /* DoM entry is the first entry always */
1203 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) == LOV_PATTERN_MDT)
1204 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1209 /* set attr and LOV EA at once, return updated attr */
1210 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1211 const struct md_attr *ma)
1213 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1214 struct mdd_device *mdd = mdo2mdd(obj);
1215 struct thandle *handle = NULL;
1216 struct lu_attr *la_copy = &mdd_env_info(env)->mdi_la_for_fix;
1217 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1218 const struct lu_attr *la = &ma->ma_attr;
1219 struct lu_ucred *uc;
1220 bool chrgrp_by_unprivileged_user = false;
1224 /* we do not use ->attr_set() for LOV/HSM EA any more */
1225 LASSERT((ma->ma_valid & MA_LOV) == 0);
1226 LASSERT((ma->ma_valid & MA_HSM) == 0);
1228 rc = mdd_la_get(env, mdd_obj, attr);
1232 *la_copy = ma->ma_attr;
1233 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1237 /* no need to setattr anymore */
1238 if (la_copy->la_valid == 0) {
1240 "%s: no valid attribute on "DFID", previous was %#llx\n",
1241 mdd_obj_dev_name(mdd_obj),
1242 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1247 /* If an unprivileged user changes group of some file,
1248 * the setattr operation will be processed synchronously to
1249 * honor the quota limit of the corresponding group. see LU-5152 */
1250 uc = lu_ucred_check(env);
1251 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1252 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1253 /* LU-10048: disable synchronous chgrp operation for it will
1254 * cause deadlock between MDT and OST.
1255 la_copy->la_valid |= LA_FLAGS;
1256 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
1258 chrgrp_by_unprivileged_user = true;
1260 /* Flush the possible existing client setattr requests to OSTs
1261 * to keep the order with the current setattr operation that
1262 * will be sent directly to OSTs. see LU-5152 */
1263 /* LU-11303 disable sync as this is too heavyweight.
1264 * This should be replaced with a sync only for the object
1265 * being modified here, not the whole filesystem.
1266 rc = dt_sync(env, mdd->mdd_child);
1272 handle = mdd_trans_create(env, mdd);
1273 if (IS_ERR(handle)) {
1274 rc = PTR_ERR(handle);
1280 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1284 rc = mdd_trans_start(env, mdd, handle);
1288 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1289 permission_needs_sync(attr, la))
1290 handle->th_sync = 1;
1292 if (la->la_valid & (LA_MTIME | LA_CTIME))
1293 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1294 la->la_mtime, la->la_ctime);
1296 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1298 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1299 * otherwise following truncate will do nothing and truncated
1300 * data may be read again. This is a quick fix until LU-11033
1303 if (la_copy->la_valid & LA_SIZE) {
1304 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1306 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1310 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1311 la_copy->la_valid &= ~LA_SIZE;
1312 lu_buf_free(lov_buf);
1316 if (la_copy->la_valid) {
1317 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1319 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1320 /* rollback to the original gid */
1321 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1322 la_copy->la_gid = attr->la_gid;
1323 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1326 mdd_write_unlock(env, mdd_obj);
1330 rc = mdd_attr_set_changelog(env, obj, handle, &ma->ma_pfid,
1334 rc = mdd_trans_stop(env, mdd, rc, handle);
1339 static int mdd_xattr_sanity_check(const struct lu_env *env,
1340 struct mdd_object *obj,
1341 const struct lu_attr *attr,
1344 struct lu_ucred *uc = lu_ucred_assert(env);
1347 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1350 if (strncmp(XATTR_USER_PREFIX, name,
1351 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1352 /* For sticky directories, only the owner and privileged user
1353 * can write attributes. */
1354 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1355 (uc->uc_fsuid != attr->la_uid) &&
1356 !md_capable(uc, CAP_FOWNER))
1358 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1359 (uc->uc_fsuid != attr->la_uid) &&
1360 !md_capable(uc, CAP_FOWNER)) {
1368 * Check if a string begins with a given prefix.
1370 * \param str String to check
1371 * \param prefix Substring to check at the beginning of \a str
1372 * \return true/false whether the condition is verified.
1374 static inline bool has_prefix(const char *str, const char *prefix)
1376 return strncmp(prefix, str, strlen(prefix)) == 0;
1380 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1382 * \param[in] xattr_name Full extended attribute name.
1384 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1386 static enum changelog_rec_type
1387 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1388 const char *xattr_name)
1390 /* Layout changes systematically recorded */
1391 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1392 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1393 allowed_lustre_lov(xattr_name))
1396 /* HSM information changes systematically recorded */
1397 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1400 /* Avoid logging SOM xattr for every file */
1401 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1404 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1405 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1406 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1407 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1408 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1414 static int mdd_declare_xattr_set(const struct lu_env *env,
1415 struct mdd_device *mdd,
1416 struct mdd_object *obj,
1417 const struct lu_buf *buf,
1419 int fl, struct thandle *handle)
1421 enum changelog_rec_type type;
1424 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1428 type = mdd_xattr_changelog_type(env, mdd, name);
1430 return 0; /* no changelog to store */
1432 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1436 * Compare current and future data of HSM EA and add a changelog if needed.
1438 * Caller should have write-locked \param obj.
1440 * \param buf - Future HSM EA content.
1441 * \retval 0 if no changelog is needed or changelog was added properly.
1442 * \retval -ve errno if there was a problem
1444 static int mdd_hsm_update_locked(const struct lu_env *env,
1445 struct md_object *obj,
1446 const struct lu_buf *buf,
1447 struct thandle *handle,
1448 enum changelog_rec_flags *clf_flags)
1450 struct mdd_thread_info *info = mdd_env_info(env);
1451 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1452 struct lu_buf *current_buf;
1453 struct md_hsm *current_mh;
1454 struct md_hsm *new_mh;
1458 OBD_ALLOC_PTR(current_mh);
1459 if (current_mh == NULL)
1462 /* Read HSM attrs from disk */
1463 current_buf = lu_buf_check_and_alloc(&info->mdi_xattr_buf,
1465 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1467 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1468 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1469 if (rc < 0 && rc != -ENODATA)
1471 else if (rc == -ENODATA)
1472 current_mh->mh_flags = 0;
1474 /* Map future HSM xattr */
1475 OBD_ALLOC_PTR(new_mh);
1477 GOTO(free, rc = -ENOMEM);
1478 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1482 /* Flags differ, set flags for the changelog that will be added */
1483 if (current_mh->mh_flags != new_mh->mh_flags) {
1484 hsm_set_cl_event(clf_flags, HE_STATE);
1485 if (new_mh->mh_flags & HS_DIRTY)
1486 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1489 OBD_FREE_PTR(new_mh);
1492 OBD_FREE_PTR(current_mh);
1496 static int mdd_object_pfid_replace(const struct lu_env *env,
1497 struct mdd_object *o)
1499 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1500 struct thandle *handle;
1503 handle = mdd_trans_create(env, mdd);
1505 RETURN(PTR_ERR(handle));
1507 handle->th_complex = 1;
1509 /* it doesn't need to track the PFID update via llog, because LFSCK
1510 * will repair it even it goes wrong */
1511 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1516 rc = mdd_trans_start(env, mdd, handle);
1520 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1525 mdd_trans_stop(env, mdd, rc, handle);
1530 static int mdd_declare_xattr_del(const struct lu_env *env,
1531 struct mdd_device *mdd,
1532 struct mdd_object *obj,
1534 struct thandle *handle);
1536 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1539 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1540 struct md_object *md_vic)
1542 struct mdd_device *mdd = mdo2mdd(md_obj);
1543 struct mdd_object *obj = md2mdd_obj(md_obj);
1544 struct mdd_object *vic = md2mdd_obj(md_vic);
1545 struct lu_buf *buf = &mdd_env_info(env)->mdi_buf[0];
1546 struct lu_buf *buf_vic = &mdd_env_info(env)->mdi_buf[1];
1547 struct lov_mds_md *lmm;
1548 struct thandle *handle;
1552 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1553 if (rc == 0) /* same fid */
1556 handle = mdd_trans_create(env, mdd);
1558 RETURN(PTR_ERR(handle));
1561 mdd_write_lock(env, obj, DT_TGT_CHILD);
1562 mdd_write_lock(env, vic, DT_TGT_CHILD);
1564 mdd_write_lock(env, vic, DT_TGT_CHILD);
1565 mdd_write_lock(env, obj, DT_TGT_CHILD);
1568 /* get EA of victim file */
1569 memset(buf_vic, 0, sizeof(*buf_vic));
1570 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1577 /* parse the layout of victim file */
1578 lmm = buf_vic->lb_buf;
1579 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1580 GOTO(out, rc = -EINVAL);
1582 /* save EA of target file for restore */
1583 memset(buf, 0, sizeof(*buf));
1584 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1588 /* Get rid of the layout from victim object */
1589 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1593 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_NAME_LOV,
1594 LU_XATTR_MERGE, handle);
1598 rc = mdd_trans_start(env, mdd, handle);
1602 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV, LU_XATTR_MERGE,
1607 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1609 GOTO(out_restore, rc);
1611 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1613 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle,
1619 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1620 LU_XATTR_REPLACE, handle);
1622 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1623 mdd_obj_dev_name(obj),
1624 PFID(mdd_object_fid(obj)), rc2);
1628 mdd_trans_stop(env, mdd, rc, handle);
1629 mdd_write_unlock(env, obj);
1630 mdd_write_unlock(env, vic);
1632 lu_buf_free(buf_vic);
1635 (void) mdd_object_pfid_replace(env, obj);
1641 * Extract the mirror with specified mirror id, and store the splitted
1642 * mirror layout to @buf.
1644 * \param[in] comp_v1 mirrored layout
1645 * \param[in] mirror_id the mirror with mirror_id to be extracted
1646 * \param[out] buf store the layout excluding the extracted mirror,
1647 * caller free the buffer we allocated in this function
1648 * \param[out] buf_vic store the extracted layout, caller free the buffer
1649 * we allocated in this function
1651 * \retval 0 on success; < 0 if error happens
1653 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1654 struct lu_buf *buf, struct lu_buf *buf_vic)
1656 struct lov_comp_md_v1 *comp_rem;
1657 struct lov_comp_md_v1 *comp_vic;
1658 struct lov_comp_md_entry_v1 *entry;
1659 struct lov_comp_md_entry_v1 *entry_rem;
1660 struct lov_comp_md_entry_v1 *entry_vic;
1662 __u16 comp_cnt, count = 0;
1663 int lmm_size, lmm_size_vic = 0;
1665 int offset, offset_rem, offset_vic;
1667 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1668 /* comp_v1 should contains more than 1 mirror */
1669 if (mirror_cnt <= 1)
1671 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1672 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1674 for (i = 0; i < comp_cnt; i++) {
1675 entry = &comp_v1->lcm_entries[i];
1676 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1678 lmm_size_vic += sizeof(*entry);
1679 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1680 } else if (count > 0) {
1681 /* find the specified mirror */
1689 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1693 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1694 if (!buf_vic->lb_buf) {
1699 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1700 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1702 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1703 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1704 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1705 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1706 if (!comp_rem->lcm_mirror_count)
1707 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1709 memset(comp_vic, 0, sizeof(*comp_v1));
1710 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1711 comp_vic->lcm_mirror_count = 0;
1712 comp_vic->lcm_entry_count = cpu_to_le32(count);
1713 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1714 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1715 comp_vic->lcm_layout_gen = 0;
1717 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1718 offset_rem = sizeof(*comp_rem) +
1719 sizeof(*entry_rem) * (comp_cnt - count);
1720 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1721 for (i = j = k = 0; i < comp_cnt; i++) {
1722 struct lov_mds_md *lmm, *lmm_dst;
1725 entry = &comp_v1->lcm_entries[i];
1726 entry_vic = &comp_vic->lcm_entries[j];
1727 entry_rem = &comp_rem->lcm_entries[k];
1729 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1732 /* copy component entry */
1734 memcpy(entry_vic, entry, sizeof(*entry));
1735 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1736 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1739 memcpy(entry_rem, entry, sizeof(*entry));
1740 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1744 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1746 lmm_dst = (struct lov_mds_md *)
1747 ((char *)comp_vic + offset_vic);
1749 lmm_dst = (struct lov_mds_md *)
1750 ((char *)comp_rem + offset_rem);
1752 /* copy component entry blob */
1753 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1755 /* blob offset advance */
1756 offset += le32_to_cpu(entry->lcme_size);
1758 offset_vic += le32_to_cpu(entry->lcme_size);
1760 offset_rem += le32_to_cpu(entry->lcme_size);
1766 static int mdd_dom_data_truncate(const struct lu_env *env,
1767 struct mdd_device *mdd, struct mdd_object *mo);
1769 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1770 struct md_rejig_data *mrd)
1772 struct mdd_device *mdd = mdo2mdd(md_obj);
1773 struct mdd_object *obj = md2mdd_obj(md_obj);
1774 struct mdd_object *vic = NULL;
1775 struct lu_buf *buf = &mdd_env_info(env)->mdi_buf[0];
1776 struct lu_buf *buf_save = &mdd_env_info(env)->mdi_buf[1];
1777 struct lu_buf *buf_vic = &mdd_env_info(env)->mdi_buf[2];
1778 struct lov_comp_md_v1 *lcm;
1779 struct thandle *handle;
1781 bool dom_stripe = false;
1786 * NULL @mrd_obj means mirror deleting, and use NULL vic to indicate
1790 vic = md2mdd_obj(mrd->mrd_obj);
1792 handle = mdd_trans_create(env, mdd);
1794 RETURN(PTR_ERR(handle));
1796 /* get EA of mirrored file */
1797 memset(buf_save, 0, sizeof(*buf));
1798 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1802 lcm = buf_save->lb_buf;
1803 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1804 GOTO(stop, rc = -EINVAL);
1807 * Extract the mirror with specified mirror id, and store the splitted
1808 * mirror layout to the victim buffer.
1810 memset(buf, 0, sizeof(*buf));
1811 memset(buf_vic, 0, sizeof(*buf_vic));
1812 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1816 * @buf stores layout w/o the specified mirror, @buf_vic stores the
1820 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1824 * non delete mirror split
1826 * declare obj set remaining layout in @buf, will set obj's
1829 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1830 LU_XATTR_SPLIT, handle);
1834 /* declare vic set splitted layout in @buf_vic */
1835 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic,
1836 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1842 * declare delete mirror objects in @buf_vic, will change obj's
1845 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic,
1846 XATTR_NAME_LOV, LU_XATTR_PURGE,
1851 /* declare obj set remaining layout in @buf */
1852 rc = mdd_declare_xattr_set(env, mdd, obj, buf,
1853 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1859 rc = mdd_trans_start(env, mdd, handle);
1864 /* don't use the same file to save the splitted mirror */
1865 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1867 GOTO(stop, rc = -EPERM);
1870 mdd_write_lock(env, obj, DT_TGT_CHILD);
1871 mdd_write_lock(env, vic, DT_TGT_CHILD);
1873 mdd_write_lock(env, vic, DT_TGT_CHILD);
1874 mdd_write_lock(env, obj, DT_TGT_CHILD);
1877 mdd_write_lock(env, obj, DT_TGT_CHILD);
1880 /* set obj's layout in @buf */
1881 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_SPLIT,
1887 /* set vic's layout in @buf_vic */
1888 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV,
1889 LU_XATTR_CREATE, handle);
1891 GOTO(out_restore, rc);
1893 /* delete mirror objects */
1894 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV,
1895 LU_XATTR_PURGE, handle);
1897 GOTO(out_restore, rc);
1900 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1903 GOTO(out_restore, rc);
1906 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic,
1909 GOTO(out_restore, rc);
1914 /* restore obj's in-memory and on-disk layout */
1915 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1916 LU_XATTR_REPLACE, handle);
1918 CERROR("%s: failed rollback "DFID
1919 " layout: file state unknown: rc = %d\n",
1920 mdd_obj_dev_name(obj),
1921 PFID(mdd_object_fid(obj)), rc);
1925 mdd_write_unlock(env, obj);
1927 mdd_write_unlock(env, vic);
1929 rc = mdd_trans_stop(env, mdd, rc, handle);
1931 /* Truncate local DOM data if all went well */
1932 if (!rc && dom_stripe)
1933 mdd_dom_data_truncate(env, mdd, obj);
1935 lu_buf_free(buf_save);
1937 lu_buf_free(buf_vic);
1940 (void) mdd_object_pfid_replace(env, obj);
1945 static int mdd_layout_merge_allowed(const struct lu_env *env,
1946 struct md_object *target,
1947 struct md_object *victim)
1949 struct mdd_object *o1 = md2mdd_obj(target);
1951 /* cannot extend directory's LOVEA */
1952 if (S_ISDIR(mdd_object_type(o1))) {
1953 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1954 mdd_obj_dev_name(o1));
1962 * The caller should guarantee to update the object ctime
1963 * after xattr_set if needed.
1965 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1966 const struct lu_buf *buf, const char *name,
1969 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1970 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1971 struct mdd_device *mdd = mdo2mdd(obj);
1972 struct thandle *handle;
1973 enum changelog_rec_type cl_type;
1974 enum changelog_rec_flags clf_flags = 0;
1978 rc = mdd_la_get(env, mdd_obj, attr);
1982 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1986 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1987 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1988 struct md_rejig_data *mrd = buf->lb_buf;
1989 struct md_object *victim = mrd->mrd_obj;
1991 if (buf->lb_len != sizeof(*mrd))
1995 if (fl == LU_XATTR_MERGE) {
1996 rc = mdd_layout_merge_allowed(env, obj, victim);
1999 /* merge layout of victim as a mirror of obj's. */
2000 rc = mdd_xattr_merge(env, obj, victim);
2002 rc = mdd_xattr_split(env, obj, mrd);
2007 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
2008 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
2009 struct posix_acl *acl;
2011 /* user may set empty ACL, which should be treated as removing
2013 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
2016 RETURN(PTR_ERR(acl));
2018 rc = mdd_xattr_del(env, obj, name);
2021 posix_acl_release(acl);
2024 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
2025 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
2029 handle = mdd_trans_create(env, mdd);
2031 RETURN(PTR_ERR(handle));
2033 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
2037 rc = mdd_trans_start(env, mdd, handle);
2041 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2043 if (strcmp(XATTR_NAME_HSM, name) == 0) {
2044 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
2046 mdd_write_unlock(env, mdd_obj);
2051 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
2052 mdd_write_unlock(env, mdd_obj);
2056 cl_type = mdd_xattr_changelog_type(env, mdd, name);
2060 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
2061 mdd_obj, name, handle);
2065 return mdd_trans_stop(env, mdd, rc, handle);
2068 static int mdd_declare_xattr_del(const struct lu_env *env,
2069 struct mdd_device *mdd,
2070 struct mdd_object *obj,
2072 struct thandle *handle)
2074 enum changelog_rec_type type;
2077 rc = mdo_declare_xattr_del(env, obj, name, handle);
2081 type = mdd_xattr_changelog_type(env, mdd, name);
2083 return 0; /* no changelog to store */
2085 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2089 * The caller should guarantee to update the object ctime
2090 * after xattr_set if needed.
2092 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2095 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2096 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2097 struct mdd_device *mdd = mdo2mdd(obj);
2098 struct thandle *handle;
2102 rc = mdd_la_get(env, mdd_obj, attr);
2106 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2110 handle = mdd_trans_create(env, mdd);
2112 RETURN(PTR_ERR(handle));
2114 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2118 rc = mdd_trans_start(env, mdd, handle);
2122 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2123 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2124 mdd_write_unlock(env, mdd_obj);
2128 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2131 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2136 return mdd_trans_stop(env, mdd, rc, handle);
2140 * read lov/lmv EA of an object
2141 * return the lov/lmv EA in an allocated lu_buf
2143 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2144 struct lu_buf *lmm_buf, const char *name)
2146 struct lu_buf *buf = &mdd_env_info(env)->mdi_big_buf;
2151 if (buf->lb_buf == NULL) {
2152 buf = lu_buf_check_and_alloc(buf, 4096);
2153 if (buf->lb_buf == NULL)
2158 rc = mdo_xattr_get(env, obj, buf, name);
2159 if (rc == -ERANGE) {
2160 /* mdi_big_buf is allocated but is too small
2161 * we need to increase it */
2162 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_big_buf,
2164 if (buf->lb_buf == NULL)
2167 } else if (rc < 0) {
2169 } else if (rc == 0) {
2173 lu_buf_alloc(lmm_buf, rc);
2174 if (lmm_buf->lb_buf == NULL)
2178 * we don't use lmm_buf directly, because we don't know xattr size, so
2179 * by using mdi_big_buf we can avoid calling mdo_xattr_get() twice.
2181 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2186 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2187 struct mdd_object *o, struct lu_buf *buf,
2188 struct thandle *handle)
2190 struct hsm_attrs *attrs;
2191 enum hsm_states hsm_flags;
2192 enum changelog_rec_flags clf_flags = 0;
2196 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2201 attrs = buf->lb_buf;
2202 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2203 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2206 /* Add a changelog record for release. */
2207 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2208 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2209 clf_flags, o, handle, NULL);
2214 * check if layout swapping between 2 objects is allowed
2216 * - only normal FIDs or non-system IGIFs
2217 * - same type of objects
2218 * - same owner/group (so quotas are still valid) unless this is from HSM
2221 static int mdd_layout_swap_allowed(const struct lu_env *env,
2222 struct mdd_object *o1,
2223 const struct lu_attr *attr1,
2224 struct mdd_object *o2,
2225 const struct lu_attr *attr2,
2228 const struct lu_fid *fid1, *fid2;
2231 fid1 = mdd_object_fid(o1);
2232 fid2 = mdd_object_fid(o2);
2234 if (!fid_is_norm(fid1) &&
2235 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2238 if (!fid_is_norm(fid2) &&
2239 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2242 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2243 if (S_ISDIR(mdd_object_type(o1)))
2245 if (S_ISREG(mdd_object_type(o1)))
2250 if (flags & SWAP_LAYOUTS_MDS_HSM)
2253 if ((attr1->la_uid != attr2->la_uid) ||
2254 (attr1->la_gid != attr2->la_gid))
2260 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2261 * look into the layout in MDD layer. */
2262 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2264 struct lov_comp_md_v1 *comp_v1;
2265 struct lov_mds_md *v1;
2269 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2270 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2271 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2277 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2278 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2281 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2282 off = le32_to_cpu(comp_v1->lcm_entries[i].
2284 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2289 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2290 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2301 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2303 return mdd_lmm_oi(lmm, oi, true);
2306 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2308 return mdd_lmm_oi(lmm, oi, false);
2311 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2313 struct lov_comp_md_v1 *comp_v1;
2315 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2316 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2318 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2320 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2321 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2322 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2323 __u16 tmp_gen = *gen;
2325 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2327 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2334 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2336 return mdd_lmm_gen(lmm, gen, true);
2339 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2341 return mdd_lmm_gen(lmm, gen, false);
2344 static int mdd_dom_data_truncate(const struct lu_env *env,
2345 struct mdd_device *mdd, struct mdd_object *mo)
2348 struct dt_object *dom;
2351 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2353 GOTO(out, rc = -ENODATA);
2355 th = dt_trans_create(env, mdd->mdd_bottom);
2357 GOTO(out, rc = PTR_ERR(th));
2359 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2363 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2367 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2369 dt_trans_stop(env, mdd->mdd_bottom, th);
2371 /* Ignore failure but report the error */
2373 CERROR("%s: can't truncate DOM inode "DFID" data: rc = %d\n",
2374 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2379 * swap layouts between 2 lustre objects
2381 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2382 struct md_object *obj2, __u64 flags)
2384 struct mdd_thread_info *info = mdd_env_info(env);
2385 struct mdd_object *fst_o = md2mdd_obj(obj1);
2386 struct mdd_object *snd_o = md2mdd_obj(obj2);
2387 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2388 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2389 struct mdd_device *mdd = mdo2mdd(obj1);
2390 struct lov_mds_md *fst_lmm, *snd_lmm;
2391 struct lu_buf *fst_buf = &info->mdi_buf[0];
2392 struct lu_buf *snd_buf = &info->mdi_buf[1];
2393 struct lu_buf *fst_hsm_buf = &info->mdi_buf[2];
2394 struct lu_buf *snd_hsm_buf = &info->mdi_buf[3];
2395 struct ost_id *saved_oi = NULL;
2396 struct thandle *handle;
2397 struct mdd_object *dom_o = NULL;
2398 __u64 domsize_dom, domsize_vlt;
2399 __u32 fst_gen, snd_gen, saved_gen;
2405 BUILD_BUG_ON(ARRAY_SIZE(info->mdi_buf) < 4);
2406 memset(info->mdi_buf, 0, sizeof(info->mdi_buf));
2408 /* we have to sort the 2 obj, so locking will always
2409 * be in the same order, even in case of 2 concurrent swaps */
2410 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2411 if (rc == 0) /* same fid ? */
2417 rc = mdd_la_get(env, fst_o, fst_la);
2421 rc = mdd_la_get(env, snd_o, snd_la);
2425 /* check if layout swapping is allowed */
2426 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2430 handle = mdd_trans_create(env, mdd);
2432 RETURN(PTR_ERR(handle));
2434 /* objects are already sorted */
2435 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2436 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2438 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2439 if (rc < 0 && rc != -ENODATA)
2442 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2443 if (rc < 0 && rc != -ENODATA)
2446 /* check if file has DoM. DoM file can be migrated only to another
2447 * DoM layout with the same DoM component size or to an non-DOM
2448 * layout. After migration to OSTs layout, local MDT inode data
2449 * should be truncated.
2450 * Objects are sorted by FIDs, considering that original file's FID
2451 * is always smaller the snd_o is always original file we are migrating
2454 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2455 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2457 /* Only migration is supported for DoM files, not 'swap_layouts' so
2458 * target file must be volatile and orphan.
2460 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2461 dom_o = domsize_dom ? snd_o : NULL;
2462 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2463 swap(domsize_dom, domsize_vlt);
2464 dom_o = domsize_dom ? fst_o : NULL;
2465 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2466 /* 'lfs swap_layouts' case, neither file should have DoM */
2468 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2469 "use migration instead: rc = %d\n", rc);
2473 if (domsize_vlt > 0 && domsize_dom == 0) {
2476 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2477 mdd_obj_dev_name(snd_o),
2478 PFID(mdd_object_fid(snd_o)), rc);
2480 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2483 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2484 mdd_obj_dev_name(fst_o),
2485 PFID(mdd_object_fid(fst_o)), rc);
2487 } else if (domsize_vlt > 0) {
2488 /* Migration with the same DOM component size, no need to
2489 * truncate local data, it is still being used */
2493 /* swapping 2 non existant layouts is a success */
2494 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2497 /* to help inode migration between MDT, it is better to
2498 * start by the no layout file (if one), so we order the swap */
2499 if (snd_buf->lb_buf == NULL) {
2501 swap(fst_buf, snd_buf);
2504 fst_gen = snd_gen = 0;
2505 /* lmm and generation layout initialization */
2506 if (fst_buf->lb_buf != NULL) {
2507 fst_lmm = fst_buf->lb_buf;
2508 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2509 fst_fl = LU_XATTR_REPLACE;
2513 fst_fl = LU_XATTR_CREATE;
2516 snd_lmm = snd_buf->lb_buf;
2517 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2519 saved_gen = fst_gen;
2520 /* increase the generation layout numbers */
2525 * XXX The layout generation is used to generate component IDs for
2526 * the composite file, we have to do some special tweaks to make
2527 * sure the layout generation is always adequate for that job.
2530 /* Skip invalid generation number for composite layout */
2531 if ((snd_gen & LCME_ID_MASK) == 0)
2533 if ((fst_gen & LCME_ID_MASK) == 0)
2535 /* Make sure the generation is greater than all the component IDs */
2536 if (fst_gen < snd_gen)
2538 else if (fst_gen > snd_gen)
2541 /* set the file specific informations in lmm */
2542 if (fst_lmm != NULL) {
2543 struct ost_id temp_oi;
2545 saved_oi = &info->mdi_oa.o_oi;
2546 mdd_get_lmm_oi(fst_lmm, saved_oi);
2547 mdd_get_lmm_oi(snd_lmm, &temp_oi);
2548 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2549 mdd_set_lmm_oi(fst_lmm, &temp_oi);
2550 mdd_set_lmm_oi(snd_lmm, saved_oi);
2552 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2553 cpu_to_le32(LOV_MAGIC_MAGIC))
2554 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2556 GOTO(stop, rc = -EPROTO);
2558 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2560 /* Prepare HSM attribute if it's required */
2561 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2562 const int buflen = sizeof(struct hsm_attrs);
2564 lu_buf_alloc(fst_hsm_buf, buflen);
2565 lu_buf_alloc(snd_hsm_buf, buflen);
2566 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2567 GOTO(stop, rc = -ENOMEM);
2569 /* Read HSM attribute */
2570 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2574 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2578 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2579 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2584 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2585 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2591 /* prepare transaction */
2592 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2597 if (fst_buf->lb_buf != NULL)
2598 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2599 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2602 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2607 rc = mdd_trans_start(env, mdd, handle);
2611 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2612 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2616 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2618 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2621 CERROR("%s: HSM error restoring "DFID": rc = %d/%d\n",
2622 mdd_obj_dev_name(fst_o),
2623 PFID(mdd_object_fid(fst_o)), rc, rc2);
2628 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2632 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2635 if (fst_buf->lb_buf != NULL)
2636 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2637 LU_XATTR_REPLACE, handle);
2639 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2642 GOTO(out_restore, rc);
2644 /* Issue one changelog record per file */
2645 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle,
2650 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle,
2660 /* failure on second file, but first was done, so we have
2661 * to roll back first. */
2662 if (fst_buf->lb_buf != NULL) {
2663 mdd_set_lmm_oi(fst_lmm, saved_oi);
2664 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2665 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2666 LU_XATTR_REPLACE, handle);
2668 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2674 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2679 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2684 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
2685 mdd_obj_dev_name(fst_o),
2686 PFID(mdd_object_fid(snd_o)),
2687 PFID(mdd_object_fid(fst_o)),
2689 /* a solution to avoid journal commit is to panic,
2690 * but it has strong consequences so we use LBUG to
2691 * allow sysdamin to choose to panic or not
2698 rc = mdd_trans_stop(env, mdd, rc, handle);
2700 /* Truncate local DOM data if all went well */
2702 mdd_dom_data_truncate(env, mdd, dom_o);
2704 mdd_write_unlock(env, snd_o);
2705 mdd_write_unlock(env, fst_o);
2707 lu_buf_free(fst_buf);
2708 lu_buf_free(snd_buf);
2709 lu_buf_free(fst_hsm_buf);
2710 lu_buf_free(snd_hsm_buf);
2713 (void) mdd_object_pfid_replace(env, fst_o);
2714 (void) mdd_object_pfid_replace(env, snd_o);
2719 static int mdd_declare_layout_change(const struct lu_env *env,
2720 struct mdd_device *mdd,
2721 struct mdd_object *obj,
2722 struct md_layout_change *mlc,
2723 struct thandle *handle)
2727 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2731 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2735 /* For PFL, this is used to instantiate necessary component objects. */
2737 mdd_layout_instantiate_component(const struct lu_env *env,
2738 struct mdd_object *obj, struct md_layout_change *mlc,
2739 struct thandle *handle)
2741 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2745 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2748 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2750 * It's possible that another layout write intent has already
2751 * instantiated our objects, so a -EALREADY returned, and we need to
2755 RETURN(rc == -EALREADY ? 0 : rc);
2757 rc = mdd_trans_start(env, mdd, handle);
2761 mdd_write_lock(env, obj, DT_TGT_CHILD);
2762 rc = mdo_layout_change(env, obj, mlc, handle);
2763 mdd_write_unlock(env, obj);
2767 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
2773 * Change the FLR layout from RDONLY to WRITE_PENDING.
2775 * It picks the primary mirror, and bumps the layout version, and set
2776 * layout version xattr to OST objects in a sync tx. In order to facilitate
2777 * the handling of phantom writers from evicted clients, the clients carry
2778 * layout version of the file with write RPC, so that the OSTs can verify
2779 * if the write RPCs are legitimate, meaning not from evicted clients.
2782 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2783 struct md_layout_change *mlc, struct thandle *handle)
2785 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2786 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
2787 struct lustre_som_attrs *som = &mlc->mlc_som;
2792 /* Verify acceptable operations */
2793 switch (mlc->mlc_opc) {
2794 case MD_LAYOUT_WRITE:
2795 case MD_LAYOUT_RESYNC:
2796 /* these are legal operations - this represents the case that
2797 * a few mirrors were missed in the last resync. */
2799 case MD_LAYOUT_RESYNC_DONE:
2804 som_buf->lb_buf = som;
2805 som_buf->lb_len = sizeof(*som);
2806 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2807 if (rc < 0 && rc != -ENODATA)
2811 lustre_som_swab(som);
2812 if (som->lsa_valid & SOM_FL_STRICT)
2813 fl = LU_XATTR_REPLACE;
2815 if (mlc->mlc_opc == MD_LAYOUT_WRITE &&
2816 mlc->mlc_intent->li_extent.e_end > som->lsa_size) {
2817 som->lsa_size = mlc->mlc_intent->li_extent.e_end + 1;
2818 fl = LU_XATTR_REPLACE;
2822 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2827 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2828 XATTR_NAME_SOM, fl, handle);
2833 /* record a changelog for data mover to consume */
2834 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2838 rc = mdd_trans_start(env, mdd, handle);
2842 /* it needs a sync tx to make FLR to work properly */
2843 handle->th_sync = 1;
2845 mdd_write_lock(env, obj, DT_TGT_CHILD);
2846 rc = mdo_layout_change(env, obj, mlc, handle);
2848 /* SOM state transition from STRICT to STALE */
2849 som->lsa_valid = SOM_FL_STALE;
2850 lustre_som_swab(som);
2851 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2854 mdd_write_unlock(env, obj);
2858 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle, NULL);
2869 * Handle mirrored file state transition when it's in WRITE_PENDING.
2871 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2872 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2873 * layout version will be increased, and the file's state will be changed to
2877 mdd_layout_update_write_pending(const struct lu_env *env,
2878 struct mdd_object *obj, struct md_layout_change *mlc,
2879 struct thandle *handle)
2881 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2882 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
2883 struct lustre_som_attrs *som = &mlc->mlc_som;
2888 switch (mlc->mlc_opc) {
2889 case MD_LAYOUT_RESYNC:
2890 /* Upon receiving the resync request, it should
2891 * instantiate all stale components right away to get ready
2892 * for mirror copy. In order to avoid layout version change,
2893 * client should avoid sending LAYOUT_WRITE request at the
2896 case MD_LAYOUT_WRITE:
2898 * legal race for concurrent write, the file state has been
2899 * changed by another client. Or a jump over file size and
2902 som_buf->lb_buf = som;
2903 som_buf->lb_len = sizeof(*som);
2904 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2905 if (rc < 0 && rc != -ENODATA)
2909 lustre_som_swab(som);
2910 if (mlc->mlc_intent->li_extent.e_end > som->lsa_size) {
2912 mlc->mlc_intent->li_extent.e_end + 1;
2913 fl = LU_XATTR_REPLACE;
2921 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2926 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2927 XATTR_NAME_SOM, fl, handle);
2932 rc = mdd_trans_start(env, mdd, handle);
2936 /* it needs a sync tx to make FLR to work properly */
2937 handle->th_sync = 1;
2939 mdd_write_lock(env, obj, DT_TGT_CHILD);
2940 rc = mdo_layout_change(env, obj, mlc, handle);
2942 som->lsa_valid = SOM_FL_STALE;
2943 lustre_som_swab(som);
2944 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2947 mdd_write_unlock(env, obj);
2958 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2960 * Only concurrent write and sync complete requests are possible when the
2961 * file is in SYNC_PENDING. For the latter request, it will pass in the
2962 * mirrors that have been synchronized, then the stale bit will be cleared
2963 * to make the file's state turn into RDONLY.
2964 * For concurrent write reqeust, it just needs to change the file's state
2965 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2966 * version because the version will be increased in the transition to
2967 * SYNC_PENDING later so that it can deny the write request from potential
2968 * evicted SYNC clients. */
2970 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2971 struct md_layout_change *mlc, struct thandle *handle)
2973 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2974 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
2979 /* operation validation */
2980 switch (mlc->mlc_opc) {
2981 case MD_LAYOUT_RESYNC_DONE:
2982 /* resync complete. */
2983 case MD_LAYOUT_WRITE:
2984 /* concurrent write. */
2986 case MD_LAYOUT_RESYNC:
2987 /* resync again, most likely the previous run failed.
2988 * no-op if it's already in SYNC_PENDING state */
2994 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
2995 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2996 if (rc < 0 && rc != -ENODATA)
2999 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
3000 lustre_som_swab(&mlc->mlc_som);
3001 som_buf->lb_buf = &mlc->mlc_som;
3002 som_buf->lb_len = sizeof(mlc->mlc_som);
3005 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3009 /* record a changelog for the completion of resync */
3010 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
3015 /* RESYNC_DONE has piggybacked size and blocks */
3017 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3018 XATTR_NAME_SOM, fl, handle);
3023 rc = mdd_trans_start(env, mdd, handle);
3027 /* it needs a sync tx to make FLR to work properly */
3028 handle->th_sync = 1;
3030 rc = mdo_layout_change(env, obj, mlc, handle);
3035 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3041 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle,
3051 * Layout change callback for object.
3053 * This is only used by FLR for now. In the future, it can be exteneded to
3054 * handle all layout change.
3057 mdd_layout_change(const struct lu_env *env, struct md_object *o,
3058 struct md_layout_change *mlc)
3060 struct mdd_object *obj = md2mdd_obj(o);
3061 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3062 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
3063 struct lov_comp_md_v1 *lcm;
3064 struct thandle *handle;
3070 if (S_ISDIR(mdd_object_type(obj))) {
3071 switch (mlc->mlc_opc) {
3072 case MD_LAYOUT_SHRINK:
3073 rc = mdd_dir_layout_shrink(env, o, mlc);
3075 case MD_LAYOUT_SPLIT:
3076 rc = mdd_dir_layout_split(env, o, mlc);
3085 /* Verify acceptable operations */
3086 switch (mlc->mlc_opc) {
3087 case MD_LAYOUT_WRITE:
3088 case MD_LAYOUT_RESYNC:
3089 case MD_LAYOUT_RESYNC_DONE:
3095 handle = mdd_trans_create(env, mdd);
3097 RETURN(PTR_ERR(handle));
3099 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
3106 /* analyze the layout to make sure it's a FLR file */
3108 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
3109 GOTO(out, rc = -EINVAL);
3111 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
3113 /* please refer to HLD of FLR for state transition */
3114 switch (flr_state) {
3116 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
3118 case LCM_FL_WRITE_PENDING:
3119 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
3122 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
3124 case LCM_FL_SYNC_PENDING:
3125 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3134 mdd_trans_stop(env, mdd, rc, handle);
3139 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3140 struct mdd_object *child, const struct lu_attr *attr,
3141 const struct md_op_spec *spec,
3142 struct dt_allocation_hint *hint)
3144 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3145 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3146 struct dt_object *nc = mdd_object_child(child);
3148 memset(hint, 0, sizeof(*hint));
3150 /* For striped directory, give striping EA to lod_ah_init, which will
3151 * decide the stripe_offset and stripe count by it. */
3152 if (S_ISDIR(attr->la_mode) &&
3153 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3154 hint->dah_eadata = spec->u.sp_ea.eadata;
3155 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3157 hint->dah_eadata = NULL;
3158 hint->dah_eadata_len = 0;
3159 if (spec->sp_cr_flags & MDS_OPEN_APPEND) {
3160 if (mdd->mdd_append_stripe_count != 0 ||
3161 mdd->mdd_append_pool[0])
3163 "using O_APPEND file striping\n");
3164 if (mdd->mdd_append_stripe_count)
3165 hint->dah_append_stripes =
3166 mdd->mdd_append_stripe_count;
3167 if (mdd->mdd_append_pool[0])
3168 hint->dah_append_pool = mdd->mdd_append_pool;
3170 hint->dah_append_stripes = 0;
3174 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3175 hint->dah_eadata, hint->dah_eadata_len);
3176 /* @hint will be initialized by underlying device. */
3177 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3180 static int mdd_accmode(const struct lu_env *env, const struct lu_attr *la,
3183 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3184 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3185 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3186 * owner can write to a file even if it is marked readonly to hide
3187 * its brokenness. (bug 5781) */
3188 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3189 struct lu_ucred *uc = lu_ucred_check(env);
3191 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3195 return mds_accmode(open_flags);
3198 static int mdd_open_sanity_check(const struct lu_env *env,
3199 struct mdd_object *obj,
3200 const struct lu_attr *attr, u64 open_flags,
3203 unsigned int may_mask;
3207 /* EEXIST check, also opening of *open* orphans is allowed so we can
3208 * open-by-handle unlinked files
3210 if (mdd_is_dead_obj(obj) && !is_replay &&
3211 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3214 if (S_ISLNK(attr->la_mode))
3217 may_mask = mdd_accmode(env, attr, open_flags);
3219 if (S_ISDIR(attr->la_mode) && (may_mask & MAY_WRITE))
3222 if (!(open_flags & MDS_OPEN_CREATED)) {
3223 rc = mdd_permission_internal(env, obj, attr, may_mask);
3228 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3229 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3230 open_flags &= ~MDS_OPEN_TRUNC;
3232 /* For writing append-only file must open it with append mode. */
3233 if (attr->la_flags & LUSTRE_APPEND_FL) {
3234 if ((open_flags & MDS_FMODE_WRITE) &&
3235 !(open_flags & MDS_OPEN_APPEND))
3237 if (open_flags & MDS_OPEN_TRUNC)
3244 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3245 u64 open_flags, struct md_op_spec *spec)
3247 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3248 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3249 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3250 struct mdd_object_user *mou = NULL;
3251 const struct lu_ucred *uc = lu_ucred(env);
3252 struct mdd_device *mdd = mdo2mdd(obj);
3253 enum changelog_rec_type type = CL_OPEN;
3257 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3259 rc = mdd_la_get(env, mdd_obj, attr);
3263 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3265 if ((rc == -EACCES) && (mdd->mdd_cl.mc_current_mask & BIT(CL_DN_OPEN)))
3270 mdd_obj->mod_count++;
3272 if (!mdd_changelog_enabled(env, mdd, type))
3276 /* look for existing opener in list under mdd_write_lock */
3277 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3282 /* add user to list */
3283 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3289 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3291 mdd_obj_user_free(mou);
3296 if (type == CL_DN_OPEN) {
3297 if (ktime_before(ktime_get(), mou->mou_deniednext))
3298 /* same user denied again same access within
3299 * time interval: do not record
3303 /* this user already denied, but some time ago:
3304 * update denied time
3306 mou->mou_deniednext =
3307 ktime_add(ktime_get(),
3308 ktime_set(mdd->mdd_cl.mc_deniednext,
3311 mou->mou_opencount++;
3312 /* same user opening file again with same flags:
3319 /* FYI, only the bottom 32 bits of open_flags are recorded */
3320 mdd_changelog(env, type, open_flags, md_dev, mdd_object_fid(mdd_obj));
3324 mdd_write_unlock(env, mdd_obj);
3328 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3329 struct md_attr *ma, struct thandle *handle)
3333 rc = mdd_orphan_declare_delete(env, obj, handle);
3337 return mdo_declare_destroy(env, obj, handle);
3341 * No permission check is needed.
3343 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3344 struct md_attr *ma, u64 open_flags)
3346 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3347 struct mdd_device *mdd = mdo2mdd(obj);
3348 struct thandle *handle = NULL;
3351 bool blocked = false;
3352 bool last_close_by_uid = false;
3353 const struct lu_ucred *uc = lu_ucred(env);
3356 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3357 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3358 mdd_obj->mod_count--;
3359 mdd_write_unlock(env, mdd_obj);
3361 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3362 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3363 "list\n", PFID(mdd_object_fid(mdd_obj)));
3367 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3368 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3369 /* check without any lock */
3370 is_orphan = mdd_obj->mod_count == 1 &&
3371 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3375 /* mdd_trans_create() maybe failed because of barrier_entry(),
3376 * under such case, the orphan MDT-object will be left in the
3377 * orphan list, and when the MDT remount next time, the unused
3378 * orphans will be destroyed automatically.
3380 * One exception: the former mdd_finish_unlink may failed to
3381 * add the orphan MDT-object to the orphan list, then if the
3382 * mdd_trans_create() failed because of barrier_entry(), the
3383 * MDT-object will become real orphan that is neither in the
3384 * namespace nor in the orphan list. Such bad case should be
3385 * very rare and will be handled by e2fsck/lfsck. */
3386 handle = mdd_trans_create(env, mdo2mdd(obj));
3387 if (IS_ERR(handle)) {
3388 rc = PTR_ERR(handle);
3389 if (rc != -EINPROGRESS)
3397 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3401 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3406 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3412 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3413 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3415 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3416 lu_dev_name(mdd2lu_dev(mdd)),
3417 PFID(mdd_object_fid(mdd_obj)), rc);
3421 /* check again with lock */
3422 is_orphan = (mdd_obj->mod_count == 1) &&
3423 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3424 ma->ma_attr.la_nlink == 0);
3426 if (is_orphan && !handle && !blocked) {
3427 mdd_write_unlock(env, mdd_obj);
3431 mdd_obj->mod_count--; /*release open count */
3433 /* under mdd write lock */
3434 /* If recording, see if we need to remove UID from list. uc is not
3435 * initialized if the client has been evicted. */
3436 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3437 struct mdd_object_user *mou;
3439 /* look for UID in list */
3440 /* If mou is NULL, it probably means logging was enabled after
3441 * the user had the file open. So the corresponding close
3442 * will not be logged.
3444 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3447 mou->mou_opencount--;
3448 if (mou->mou_opencount == 0) {
3449 mdd_obj_user_remove(mdd_obj, mou);
3450 last_close_by_uid = true;
3455 if (!is_orphan || blocked)
3459 /* NB: Object maybe not in orphan list originally, it is rare case for
3460 * mdd_finish_unlink() failure, in that case, the object doesn't have
3461 * ORPHAN_OBJ flag */
3462 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3463 /* remove link to object from orphan index */
3464 LASSERT(handle != NULL);
3465 rc = mdd_orphan_delete(env, mdd_obj, handle);
3467 CERROR("%s: unable to delete "DFID" from orphan list: "
3468 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3469 PFID(mdd_object_fid(mdd_obj)), rc);
3470 /* If object was not deleted from orphan list, do not
3471 * destroy OSS objects, which will be done when next
3476 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3477 "list, OSS objects to be destroyed.\n",
3478 PFID(mdd_object_fid(mdd_obj)));
3481 rc = mdo_destroy(env, mdd_obj, handle);
3484 CERROR("%s: unable to delete "DFID" from orphan list: "
3485 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3486 PFID(mdd_object_fid(mdd_obj)), rc);
3491 mdd_write_unlock(env, mdd_obj);
3493 if (rc != 0 || blocked ||
3494 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3497 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3498 * or if CL_OPEN was recorded and it's last close by user.
3499 * Changelogs mask may change between open and close operations, but
3500 * this is not a big deal if we have a CL_CLOSE entry with no matching
3501 * CL_OPEN. Plus Changelogs mask may not change often.
3503 if (((!(mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3504 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3505 MDS_OPEN_TRUNC))) ||
3506 ((mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3507 last_close_by_uid)) &&
3508 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3509 if (handle == NULL) {
3510 handle = mdd_trans_create(env, mdo2mdd(obj));
3512 GOTO(stop, rc = PTR_ERR(handle));
3514 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3515 NULL, NULL, handle);
3519 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3524 /* FYI, only the bottom 32 bits of open_flags are recorded */
3525 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3526 mdd_obj, handle, NULL);
3530 if (handle != NULL && !IS_ERR(handle))
3531 rc = mdd_trans_stop(env, mdd, rc, handle);
3537 * Permission check is done when open,
3538 * no need check again.
3540 static int mdd_readpage_sanity_check(const struct lu_env *env,
3541 struct mdd_object *obj)
3543 struct dt_object *next = mdd_object_child(obj);
3547 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3555 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
3556 size_t nob, const struct dt_it_ops *iops,
3557 struct dt_it *it, __u32 attr, void *arg)
3559 struct lu_dirpage *dp = &lp->lp_dir;
3563 struct lu_dirent *ent;
3564 struct lu_dirent *last = NULL;
3568 if (nob < sizeof(*dp))
3571 memset(area, 0, sizeof (*dp));
3572 area += sizeof (*dp);
3573 nob -= sizeof (*dp);
3580 len = iops->key_size(env, it);
3582 /* IAM iterator can return record with zero len. */
3586 hash = iops->store(env, it);
3587 if (unlikely(first)) {
3589 dp->ldp_hash_start = cpu_to_le64(hash);
3592 /* calculate max space required for lu_dirent */
3593 recsize = lu_dirent_calc_size(len, attr);
3595 if (nob >= recsize) {
3596 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3597 if (result == -ESTALE)
3602 /* osd might not able to pack all attributes,
3603 * so recheck rec length */
3604 recsize = le16_to_cpu(ent->lde_reclen);
3606 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3607 fid_le_to_cpu(&fid, &ent->lde_fid);
3608 if (fid_is_dot_lustre(&fid))
3612 result = (last != NULL) ? 0 :-EINVAL;
3616 ent = (void *)ent + recsize;
3620 result = iops->next(env, it);
3621 if (result == -ESTALE)
3623 } while (result == 0);
3626 dp->ldp_hash_end = cpu_to_le64(hash);
3628 if (last->lde_hash == dp->ldp_hash_end)
3629 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3630 last->lde_reclen = 0; /* end mark */
3633 /* end of directory */
3634 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3635 else if (result < 0)
3636 CWARN("build page failed: %d!\n", result);
3640 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3641 const struct lu_rdpg *rdpg)
3643 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3647 if (mdd_object_exists(mdd_obj) == 0) {
3648 CERROR("%s: object "DFID" not found: rc = -2\n",
3649 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3653 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3654 rc = mdd_readpage_sanity_check(env, mdd_obj);
3656 GOTO(out_unlock, rc);
3658 if (mdd_is_dead_obj(mdd_obj)) {
3660 struct lu_dirpage *dp;
3663 * According to POSIX, please do not return any entry to client:
3664 * even dot and dotdot should not be returned.
3666 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3667 PFID(mdd_object_fid(mdd_obj)));
3669 if (rdpg->rp_count <= 0)
3670 GOTO(out_unlock, rc = -EFAULT);
3671 LASSERT(rdpg->rp_pages != NULL);
3673 pg = rdpg->rp_pages[0];
3674 dp = (struct lu_dirpage *)kmap(pg);
3675 memset(dp, 0 , sizeof(struct lu_dirpage));
3676 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3677 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3678 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3680 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3683 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3684 mdd_dir_page_build, NULL);
3686 struct lu_dirpage *dp;
3688 dp = kmap(rdpg->rp_pages[0]);
3689 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3692 * No pages were processed, mark this for first page
3695 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3696 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3697 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3699 kunmap(rdpg->rp_pages[0]);
3702 GOTO(out_unlock, rc);
3704 mdd_read_unlock(env, mdd_obj);
3708 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3710 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3712 if (mdd_object_exists(mdd_obj) == 0) {
3715 CERROR("%s: object "DFID" not found: rc = %d\n",
3716 mdd_obj_dev_name(mdd_obj),
3717 PFID(mdd_object_fid(mdd_obj)), rc);
3720 return dt_object_sync(env, mdd_object_child(mdd_obj),
3724 static int mdd_object_lock(const struct lu_env *env,
3725 struct md_object *obj,
3726 struct lustre_handle *lh,
3727 struct ldlm_enqueue_info *einfo,
3728 union ldlm_policy_data *policy)
3730 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3731 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3735 static int mdd_object_unlock(const struct lu_env *env,
3736 struct md_object *obj,
3737 struct ldlm_enqueue_info *einfo,
3738 union ldlm_policy_data *policy)
3740 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3741 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3744 const struct md_object_operations mdd_obj_ops = {
3745 .moo_permission = mdd_permission,
3746 .moo_attr_get = mdd_attr_get,
3747 .moo_attr_set = mdd_attr_set,
3748 .moo_xattr_get = mdd_xattr_get,
3749 .moo_xattr_set = mdd_xattr_set,
3750 .moo_xattr_list = mdd_xattr_list,
3751 .moo_invalidate = mdd_invalidate,
3752 .moo_xattr_del = mdd_xattr_del,
3753 .moo_swap_layouts = mdd_swap_layouts,
3754 .moo_open = mdd_open,
3755 .moo_close = mdd_close,
3756 .moo_readpage = mdd_readpage,
3757 .moo_readlink = mdd_readlink,
3758 .moo_changelog = mdd_changelog,
3759 .moo_object_sync = mdd_object_sync,
3760 .moo_object_lock = mdd_object_lock,
3761 .moo_object_unlock = mdd_object_unlock,
3762 .moo_layout_change = mdd_layout_change,