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 (!cap_raised(uc->uc_cap, 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 !cap_raised(uc->uc_cap, CAP_FOWNER))
732 /* The IMMUTABLE and APPEND_ONLY flags can
733 * only be changed by the relevant capability. */
734 if ((oldflags ^ newflags) &&
735 !cap_raised(uc->uc_cap, 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 !cap_raised(uc->uc_cap, 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 !cap_raised(uc->uc_cap, 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 !cap_raised(uc->uc_cap, 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 !cap_raised(uc->uc_cap, 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 !cap_raised(uc->uc_cap, 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 struct lquota_id_info qi;
1221 bool quota_reserved = false;
1222 bool chrgrp_by_unprivileged_user = false;
1226 /* we do not use ->attr_set() for LOV/HSM EA any more */
1227 LASSERT((ma->ma_valid & MA_LOV) == 0);
1228 LASSERT((ma->ma_valid & MA_HSM) == 0);
1230 rc = mdd_la_get(env, mdd_obj, attr);
1234 *la_copy = ma->ma_attr;
1235 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1239 /* no need to setattr anymore */
1240 if (la_copy->la_valid == 0) {
1242 "%s: no valid attribute on "DFID", previous was %#llx\n",
1243 mdd_obj_dev_name(mdd_obj),
1244 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1249 /* If an unprivileged user changes group of some file,
1250 * the setattr operation will be processed synchronously to
1251 * honor the quota limit of the corresponding group. see LU-5152 */
1252 uc = lu_ucred_check(env);
1253 memset(&qi, 0, sizeof(qi));
1254 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1255 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1256 CDEBUG(D_QUOTA, "%s: reserve quota for changing group: gid=%u size=%llu\n",
1257 mdd2obd_dev(mdd)->obd_name, la->la_gid, la->la_size);
1259 if (la->la_valid & LA_BLOCKS)
1260 qi.lqi_space = la->la_blocks << 9;
1261 else if (la->la_valid & LA_SIZE)
1262 qi.lqi_space = la->la_size;
1263 /* use local attr gotten above */
1264 else if (attr->la_valid & LA_BLOCKS)
1265 qi.lqi_space = attr->la_blocks << 9;
1266 else if (attr->la_valid & LA_SIZE)
1267 qi.lqi_space = attr->la_size;
1269 if (qi.lqi_space > 0) {
1270 qi.lqi_id.qid_gid = la->la_gid;
1271 qi.lqi_type = GRPQUOTA;
1272 qi.lqi_space = toqb(qi.lqi_space);
1273 qi.lqi_is_blk = true;
1274 rc = dt_reserve_or_free_quota(env, mdd->mdd_bottom,
1278 CDEBUG(D_QUOTA, "%s: failed to reserve quota for gid %d size %llu\n",
1279 mdd2obd_dev(mdd)->obd_name,
1280 la->la_gid, qi.lqi_space);
1285 quota_reserved = true;
1286 la_copy->la_valid |= LA_FLAGS;
1289 chrgrp_by_unprivileged_user = true;
1291 /* Flush the possible existing client setattr requests to OSTs
1292 * to keep the order with the current setattr operation that
1293 * will be sent directly to OSTs. see LU-5152 */
1294 /* LU-11303 disable sync as this is too heavyweight.
1295 * This should be replaced with a sync only for the object
1296 * being modified here, not the whole filesystem.
1297 rc = dt_sync(env, mdd->mdd_child);
1303 handle = mdd_trans_create(env, mdd);
1304 if (IS_ERR(handle)) {
1305 rc = PTR_ERR(handle);
1311 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1315 rc = mdd_trans_start(env, mdd, handle);
1319 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1320 permission_needs_sync(attr, la))
1321 handle->th_sync = 1;
1323 if (la->la_valid & (LA_MTIME | LA_CTIME))
1324 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1325 la->la_mtime, la->la_ctime);
1327 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1329 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1330 * otherwise following truncate will do nothing and truncated
1331 * data may be read again. This is a quick fix until LU-11033
1334 if (la_copy->la_valid & LA_SIZE) {
1335 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1337 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1341 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1342 la_copy->la_valid &= ~LA_SIZE;
1343 lu_buf_free(lov_buf);
1347 if (la_copy->la_valid) {
1348 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1350 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1351 /* rollback to the original gid */
1352 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1353 la_copy->la_gid = attr->la_gid;
1354 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1357 mdd_write_unlock(env, mdd_obj);
1361 rc = mdd_attr_set_changelog(env, obj, handle, &ma->ma_pfid,
1364 if (rc == 0 && quota_reserved) {
1365 struct thandle *sub_th;
1367 sub_th = thandle_get_sub_by_dt(env, handle, mdd->mdd_bottom);
1368 if (unlikely(IS_ERR(sub_th))) {
1370 dt_reserve_or_free_quota(env, mdd->mdd_bottom, &qi);
1372 sub_th->th_reserved_quota = qi;
1374 } else if (quota_reserved) {
1376 dt_reserve_or_free_quota(env, mdd->mdd_bottom, &qi);
1380 rc = mdd_trans_stop(env, mdd, rc, handle);
1385 static int mdd_xattr_sanity_check(const struct lu_env *env,
1386 struct mdd_object *obj,
1387 const struct lu_attr *attr,
1390 struct lu_ucred *uc = lu_ucred_assert(env);
1393 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1396 if (strncmp(XATTR_USER_PREFIX, name,
1397 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1398 /* For sticky directories, only the owner and privileged user
1399 * can write attributes. */
1400 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1401 (uc->uc_fsuid != attr->la_uid) &&
1402 !cap_raised(uc->uc_cap, CAP_FOWNER))
1404 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1405 (uc->uc_fsuid != attr->la_uid) &&
1406 !cap_raised(uc->uc_cap, CAP_FOWNER)) {
1414 * Check if a string begins with a given prefix.
1416 * \param str String to check
1417 * \param prefix Substring to check at the beginning of \a str
1418 * \return true/false whether the condition is verified.
1420 static inline bool has_prefix(const char *str, const char *prefix)
1422 return strncmp(prefix, str, strlen(prefix)) == 0;
1426 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1428 * \param[in] xattr_name Full extended attribute name.
1430 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1432 static enum changelog_rec_type
1433 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1434 const char *xattr_name)
1436 /* Layout changes systematically recorded */
1437 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1438 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1439 allowed_lustre_lov(xattr_name))
1442 /* HSM information changes systematically recorded */
1443 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1446 /* Avoid logging SOM xattr for every file */
1447 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1450 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1451 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1452 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1453 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1454 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1460 static int mdd_declare_xattr_set(const struct lu_env *env,
1461 struct mdd_device *mdd,
1462 struct mdd_object *obj,
1463 const struct lu_buf *buf,
1465 int fl, struct thandle *handle)
1467 enum changelog_rec_type type;
1470 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1474 type = mdd_xattr_changelog_type(env, mdd, name);
1476 return 0; /* no changelog to store */
1478 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1482 * Compare current and future data of HSM EA and add a changelog if needed.
1484 * Caller should have write-locked \param obj.
1486 * \param buf - Future HSM EA content.
1487 * \retval 0 if no changelog is needed or changelog was added properly.
1488 * \retval -ve errno if there was a problem
1490 static int mdd_hsm_update_locked(const struct lu_env *env,
1491 struct md_object *obj,
1492 const struct lu_buf *buf,
1493 struct thandle *handle,
1494 enum changelog_rec_flags *clf_flags)
1496 struct mdd_thread_info *info = mdd_env_info(env);
1497 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1498 struct lu_buf *current_buf;
1499 struct md_hsm *current_mh;
1500 struct md_hsm *new_mh;
1504 OBD_ALLOC_PTR(current_mh);
1505 if (current_mh == NULL)
1508 /* Read HSM attrs from disk */
1509 current_buf = lu_buf_check_and_alloc(&info->mdi_xattr_buf,
1511 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1513 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1514 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1515 if (rc < 0 && rc != -ENODATA)
1517 else if (rc == -ENODATA)
1518 current_mh->mh_flags = 0;
1520 /* Map future HSM xattr */
1521 OBD_ALLOC_PTR(new_mh);
1523 GOTO(free, rc = -ENOMEM);
1524 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1528 /* Flags differ, set flags for the changelog that will be added */
1529 if (current_mh->mh_flags != new_mh->mh_flags) {
1530 hsm_set_cl_event(clf_flags, HE_STATE);
1531 if (new_mh->mh_flags & HS_DIRTY)
1532 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1535 OBD_FREE_PTR(new_mh);
1538 OBD_FREE_PTR(current_mh);
1542 static int mdd_object_pfid_replace(const struct lu_env *env,
1543 struct mdd_object *o)
1545 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1546 struct thandle *handle;
1549 handle = mdd_trans_create(env, mdd);
1551 RETURN(PTR_ERR(handle));
1553 handle->th_complex = 1;
1555 /* it doesn't need to track the PFID update via llog, because LFSCK
1556 * will repair it even it goes wrong */
1557 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1562 rc = mdd_trans_start(env, mdd, handle);
1566 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1571 mdd_trans_stop(env, mdd, rc, handle);
1576 static int mdd_declare_xattr_del(const struct lu_env *env,
1577 struct mdd_device *mdd,
1578 struct mdd_object *obj,
1580 struct thandle *handle);
1582 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1585 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1586 struct md_object *md_vic)
1588 struct mdd_device *mdd = mdo2mdd(md_obj);
1589 struct mdd_object *obj = md2mdd_obj(md_obj);
1590 struct mdd_object *vic = md2mdd_obj(md_vic);
1591 struct lu_buf *buf = &mdd_env_info(env)->mdi_buf[0];
1592 struct lu_buf *buf_vic = &mdd_env_info(env)->mdi_buf[1];
1593 struct lov_mds_md *lmm;
1594 struct thandle *handle;
1598 lock_order = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1599 if (lock_order == 0) /* same fid */
1602 handle = mdd_trans_create(env, mdd);
1604 RETURN(PTR_ERR(handle));
1606 /* get EA of victim file */
1607 memset(buf_vic, 0, sizeof(*buf_vic));
1608 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1615 /* parse the layout of victim file */
1616 lmm = buf_vic->lb_buf;
1617 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1618 GOTO(stop, rc = -EINVAL);
1620 /* save EA of target file for restore */
1621 memset(buf, 0, sizeof(*buf));
1622 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1626 /* Get rid of the layout from victim object */
1627 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1631 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_NAME_LOV,
1632 LU_XATTR_MERGE, handle);
1636 rc = mdd_trans_start(env, mdd, handle);
1640 if (lock_order > 0) {
1641 mdd_write_lock(env, obj, DT_TGT_CHILD);
1642 mdd_write_lock(env, vic, DT_TGT_CHILD);
1644 mdd_write_lock(env, vic, DT_TGT_CHILD);
1645 mdd_write_lock(env, obj, DT_TGT_CHILD);
1648 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV, LU_XATTR_MERGE,
1653 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1655 GOTO(out_restore, rc);
1657 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1659 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle,
1665 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1666 LU_XATTR_REPLACE, handle);
1668 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1669 mdd_obj_dev_name(obj),
1670 PFID(mdd_object_fid(obj)), rc2);
1674 mdd_write_unlock(env, obj);
1675 mdd_write_unlock(env, vic);
1677 mdd_trans_stop(env, mdd, rc, handle);
1679 lu_buf_free(buf_vic);
1682 (void) mdd_object_pfid_replace(env, obj);
1688 * Extract the mirror with specified mirror id, and store the splitted
1689 * mirror layout to @buf.
1691 * \param[in] comp_v1 mirrored layout
1692 * \param[in] mirror_id the mirror with mirror_id to be extracted
1693 * \param[out] buf store the layout excluding the extracted mirror,
1694 * caller free the buffer we allocated in this function
1695 * \param[out] buf_vic store the extracted layout, caller free the buffer
1696 * we allocated in this function
1698 * \retval 0 on success; < 0 if error happens
1700 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1701 struct lu_buf *buf, struct lu_buf *buf_vic)
1703 struct lov_comp_md_v1 *comp_rem;
1704 struct lov_comp_md_v1 *comp_vic;
1705 struct lov_comp_md_entry_v1 *entry;
1706 struct lov_comp_md_entry_v1 *entry_rem;
1707 struct lov_comp_md_entry_v1 *entry_vic;
1709 __u16 comp_cnt, count = 0;
1710 int lmm_size, lmm_size_vic = 0;
1712 int offset, offset_rem, offset_vic;
1714 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1715 /* comp_v1 should contains more than 1 mirror */
1716 if (mirror_cnt <= 1)
1718 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1719 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1721 for (i = 0; i < comp_cnt; i++) {
1722 entry = &comp_v1->lcm_entries[i];
1723 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1725 lmm_size_vic += sizeof(*entry);
1726 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1727 } else if (count > 0) {
1728 /* find the specified mirror */
1736 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1740 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1741 if (!buf_vic->lb_buf) {
1746 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1747 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1749 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1750 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1751 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1752 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1753 if (!comp_rem->lcm_mirror_count)
1754 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1756 memset(comp_vic, 0, sizeof(*comp_v1));
1757 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1758 comp_vic->lcm_mirror_count = 0;
1759 comp_vic->lcm_entry_count = cpu_to_le32(count);
1760 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1761 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1762 comp_vic->lcm_layout_gen = 0;
1764 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1765 offset_rem = sizeof(*comp_rem) +
1766 sizeof(*entry_rem) * (comp_cnt - count);
1767 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1768 for (i = j = k = 0; i < comp_cnt; i++) {
1769 struct lov_mds_md *lmm, *lmm_dst;
1772 entry = &comp_v1->lcm_entries[i];
1773 entry_vic = &comp_vic->lcm_entries[j];
1774 entry_rem = &comp_rem->lcm_entries[k];
1776 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1779 /* copy component entry */
1781 memcpy(entry_vic, entry, sizeof(*entry));
1782 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1783 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1786 memcpy(entry_rem, entry, sizeof(*entry));
1787 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1791 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1793 lmm_dst = (struct lov_mds_md *)
1794 ((char *)comp_vic + offset_vic);
1796 lmm_dst = (struct lov_mds_md *)
1797 ((char *)comp_rem + offset_rem);
1799 /* copy component entry blob */
1800 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1802 /* blob offset advance */
1803 offset += le32_to_cpu(entry->lcme_size);
1805 offset_vic += le32_to_cpu(entry->lcme_size);
1807 offset_rem += le32_to_cpu(entry->lcme_size);
1813 static int mdd_dom_data_truncate(const struct lu_env *env,
1814 struct mdd_device *mdd, struct mdd_object *mo);
1816 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1817 struct md_rejig_data *mrd)
1819 struct mdd_device *mdd = mdo2mdd(md_obj);
1820 struct mdd_object *obj = md2mdd_obj(md_obj);
1821 struct mdd_object *vic = NULL;
1822 struct lu_buf *buf = &mdd_env_info(env)->mdi_buf[0];
1823 struct lu_buf *buf_save = &mdd_env_info(env)->mdi_buf[1];
1824 struct lu_buf *buf_vic = &mdd_env_info(env)->mdi_buf[2];
1825 struct lov_comp_md_v1 *lcm;
1826 struct thandle *handle;
1828 bool dom_stripe = false;
1833 * NULL @mrd_obj means mirror deleting, and use NULL vic to indicate
1837 vic = md2mdd_obj(mrd->mrd_obj);
1839 handle = mdd_trans_create(env, mdd);
1841 RETURN(PTR_ERR(handle));
1843 /* get EA of mirrored file */
1844 memset(buf_save, 0, sizeof(*buf));
1845 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1849 lcm = buf_save->lb_buf;
1850 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1851 GOTO(stop, rc = -EINVAL);
1854 * Extract the mirror with specified mirror id, and store the splitted
1855 * mirror layout to the victim buffer.
1857 memset(buf, 0, sizeof(*buf));
1858 memset(buf_vic, 0, sizeof(*buf_vic));
1859 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1863 * @buf stores layout w/o the specified mirror, @buf_vic stores the
1867 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1871 * non delete mirror split
1873 * declare obj set remaining layout in @buf, will set obj's
1876 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1877 LU_XATTR_SPLIT, handle);
1881 /* declare vic set splitted layout in @buf_vic */
1882 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic,
1883 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1889 * declare delete mirror objects in @buf_vic, will change obj's
1892 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic,
1893 XATTR_NAME_LOV, LU_XATTR_PURGE,
1898 /* declare obj set remaining layout in @buf */
1899 rc = mdd_declare_xattr_set(env, mdd, obj, buf,
1900 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1906 rc = mdd_trans_start(env, mdd, handle);
1911 /* don't use the same file to save the splitted mirror */
1912 rc = lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic));
1914 GOTO(stop, rc = -EPERM);
1917 mdd_write_lock(env, obj, DT_TGT_CHILD);
1918 mdd_write_lock(env, vic, DT_TGT_CHILD);
1920 mdd_write_lock(env, vic, DT_TGT_CHILD);
1921 mdd_write_lock(env, obj, DT_TGT_CHILD);
1924 mdd_write_lock(env, obj, DT_TGT_CHILD);
1927 /* set obj's layout in @buf */
1928 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_SPLIT,
1934 /* set vic's layout in @buf_vic */
1935 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV,
1936 LU_XATTR_CREATE, handle);
1938 GOTO(out_restore, rc);
1940 /* delete mirror objects */
1941 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV,
1942 LU_XATTR_PURGE, handle);
1944 GOTO(out_restore, rc);
1947 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1950 GOTO(out_restore, rc);
1953 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic,
1956 GOTO(out_restore, rc);
1961 /* restore obj's in-memory and on-disk layout */
1962 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1963 LU_XATTR_REPLACE, handle);
1965 CERROR("%s: failed rollback "DFID
1966 " layout: file state unknown: rc = %d\n",
1967 mdd_obj_dev_name(obj),
1968 PFID(mdd_object_fid(obj)), rc);
1972 mdd_write_unlock(env, obj);
1974 mdd_write_unlock(env, vic);
1976 rc = mdd_trans_stop(env, mdd, rc, handle);
1978 /* Truncate local DOM data if all went well */
1979 if (!rc && dom_stripe)
1980 mdd_dom_data_truncate(env, mdd, obj);
1982 lu_buf_free(buf_save);
1984 lu_buf_free(buf_vic);
1987 (void) mdd_object_pfid_replace(env, obj);
1992 static int mdd_layout_merge_allowed(const struct lu_env *env,
1993 struct md_object *target,
1994 struct md_object *victim)
1996 struct mdd_object *o1 = md2mdd_obj(target);
1998 /* cannot extend directory's LOVEA */
1999 if (S_ISDIR(mdd_object_type(o1))) {
2000 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
2001 mdd_obj_dev_name(o1));
2009 * The caller should guarantee to update the object ctime
2010 * after xattr_set if needed.
2012 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
2013 const struct lu_buf *buf, const char *name,
2016 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2017 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2018 struct mdd_device *mdd = mdo2mdd(obj);
2019 struct thandle *handle;
2020 enum changelog_rec_type cl_type;
2021 enum changelog_rec_flags clf_flags = 0;
2025 rc = mdd_la_get(env, mdd_obj, attr);
2029 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2033 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
2034 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
2035 struct md_rejig_data *mrd = buf->lb_buf;
2036 struct md_object *victim = mrd->mrd_obj;
2038 if (buf->lb_len != sizeof(*mrd))
2042 if (fl == LU_XATTR_MERGE) {
2043 rc = mdd_layout_merge_allowed(env, obj, victim);
2046 /* merge layout of victim as a mirror of obj's. */
2047 rc = mdd_xattr_merge(env, obj, victim);
2049 rc = mdd_xattr_split(env, obj, mrd);
2054 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
2055 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
2056 struct posix_acl *acl;
2058 /* user may set empty ACL, which should be treated as removing
2060 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
2063 RETURN(PTR_ERR(acl));
2065 rc = mdd_xattr_del(env, obj, name);
2068 posix_acl_release(acl);
2071 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
2072 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
2076 handle = mdd_trans_create(env, mdd);
2078 RETURN(PTR_ERR(handle));
2080 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
2084 rc = mdd_trans_start(env, mdd, handle);
2088 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2090 if (strcmp(XATTR_NAME_HSM, name) == 0) {
2091 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
2093 mdd_write_unlock(env, mdd_obj);
2098 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
2099 mdd_write_unlock(env, mdd_obj);
2103 cl_type = mdd_xattr_changelog_type(env, mdd, name);
2107 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type, clf_flags,
2108 mdd_obj, name, handle);
2112 return mdd_trans_stop(env, mdd, rc, handle);
2115 static int mdd_declare_xattr_del(const struct lu_env *env,
2116 struct mdd_device *mdd,
2117 struct mdd_object *obj,
2119 struct thandle *handle)
2121 enum changelog_rec_type type;
2124 rc = mdo_declare_xattr_del(env, obj, name, handle);
2128 type = mdd_xattr_changelog_type(env, mdd, name);
2130 return 0; /* no changelog to store */
2132 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2136 * The caller should guarantee to update the object ctime
2137 * after xattr_set if needed.
2139 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2142 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2143 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2144 struct mdd_device *mdd = mdo2mdd(obj);
2145 struct thandle *handle;
2149 rc = mdd_la_get(env, mdd_obj, attr);
2153 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2157 handle = mdd_trans_create(env, mdd);
2159 RETURN(PTR_ERR(handle));
2161 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2165 rc = mdd_trans_start(env, mdd, handle);
2169 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2170 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2171 mdd_write_unlock(env, mdd_obj);
2175 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2178 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2183 return mdd_trans_stop(env, mdd, rc, handle);
2187 * read lov/lmv EA of an object
2188 * return the lov/lmv EA in an allocated lu_buf
2190 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2191 struct lu_buf *lmm_buf, const char *name)
2193 struct lu_buf *buf = &mdd_env_info(env)->mdi_big_buf;
2198 if (buf->lb_buf == NULL) {
2199 buf = lu_buf_check_and_alloc(buf, 4096);
2200 if (buf->lb_buf == NULL)
2205 rc = mdo_xattr_get(env, obj, buf, name);
2206 if (rc == -ERANGE) {
2207 /* mdi_big_buf is allocated but is too small
2208 * we need to increase it */
2209 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_big_buf,
2211 if (buf->lb_buf == NULL)
2214 } else if (rc < 0) {
2216 } else if (rc == 0) {
2220 lu_buf_alloc(lmm_buf, rc);
2221 if (lmm_buf->lb_buf == NULL)
2225 * we don't use lmm_buf directly, because we don't know xattr size, so
2226 * by using mdi_big_buf we can avoid calling mdo_xattr_get() twice.
2228 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2233 static int mdd_xattr_hsm_replace(const struct lu_env *env,
2234 struct mdd_object *o, struct lu_buf *buf,
2235 struct thandle *handle)
2237 struct hsm_attrs *attrs;
2238 enum hsm_states hsm_flags;
2239 enum changelog_rec_flags clf_flags = 0;
2243 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
2248 attrs = buf->lb_buf;
2249 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2250 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
2253 /* Add a changelog record for release. */
2254 hsm_set_cl_event(&clf_flags, HE_RELEASE);
2255 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
2256 clf_flags, o, handle, NULL);
2261 * check if layout swapping between 2 objects is allowed
2263 * - only normal FIDs or non-system IGIFs
2264 * - same type of objects
2265 * - same owner/group (so quotas are still valid) unless this is from HSM
2268 static int mdd_layout_swap_allowed(const struct lu_env *env,
2269 struct mdd_object *o1,
2270 const struct lu_attr *attr1,
2271 struct mdd_object *o2,
2272 const struct lu_attr *attr2,
2275 const struct lu_fid *fid1, *fid2;
2278 fid1 = mdd_object_fid(o1);
2279 fid2 = mdd_object_fid(o2);
2281 if (!fid_is_norm(fid1) &&
2282 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2285 if (!fid_is_norm(fid2) &&
2286 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2289 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2290 if (S_ISDIR(mdd_object_type(o1)))
2292 if (S_ISREG(mdd_object_type(o1)))
2297 if (flags & SWAP_LAYOUTS_MDS_HSM)
2300 if ((attr1->la_uid != attr2->la_uid) ||
2301 (attr1->la_gid != attr2->la_gid))
2307 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2308 * look into the layout in MDD layer. */
2309 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2311 struct lov_comp_md_v1 *comp_v1;
2312 struct lov_mds_md *v1;
2316 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2317 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2318 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2324 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
2325 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2328 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2329 off = le32_to_cpu(comp_v1->lcm_entries[i].
2331 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2336 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2337 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2348 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2350 return mdd_lmm_oi(lmm, oi, true);
2353 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2355 return mdd_lmm_oi(lmm, oi, false);
2358 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2360 struct lov_comp_md_v1 *comp_v1;
2362 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2363 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2365 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2367 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2368 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2369 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2370 __u16 tmp_gen = *gen;
2372 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2374 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2381 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2383 return mdd_lmm_gen(lmm, gen, true);
2386 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2388 return mdd_lmm_gen(lmm, gen, false);
2391 static int mdd_dom_data_truncate(const struct lu_env *env,
2392 struct mdd_device *mdd, struct mdd_object *mo)
2395 struct dt_object *dom;
2398 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2400 GOTO(out, rc = -ENODATA);
2402 th = dt_trans_create(env, mdd->mdd_bottom);
2404 GOTO(out, rc = PTR_ERR(th));
2406 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2410 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2414 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2416 dt_trans_stop(env, mdd->mdd_bottom, th);
2418 /* Ignore failure but report the error */
2420 CERROR("%s: can't truncate DOM inode "DFID" data: rc = %d\n",
2421 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2426 * swap layouts between 2 lustre objects
2428 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
2429 struct md_object *obj2, __u64 flags)
2431 struct mdd_thread_info *info = mdd_env_info(env);
2432 struct mdd_object *fst_o = md2mdd_obj(obj1);
2433 struct mdd_object *snd_o = md2mdd_obj(obj2);
2434 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2435 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2436 struct mdd_device *mdd = mdo2mdd(obj1);
2437 struct lov_mds_md *fst_lmm, *snd_lmm;
2438 struct lu_buf *fst_buf = &info->mdi_buf[0];
2439 struct lu_buf *snd_buf = &info->mdi_buf[1];
2440 struct lu_buf *fst_hsm_buf = &info->mdi_buf[2];
2441 struct lu_buf *snd_hsm_buf = &info->mdi_buf[3];
2442 struct ost_id *saved_oi = NULL;
2443 struct thandle *handle;
2444 struct mdd_object *dom_o = NULL;
2445 __u64 domsize_dom, domsize_vlt;
2446 __u32 fst_gen, snd_gen, saved_gen;
2452 BUILD_BUG_ON(ARRAY_SIZE(info->mdi_buf) < 4);
2453 memset(info->mdi_buf, 0, sizeof(info->mdi_buf));
2455 /* we have to sort the 2 obj, so locking will always
2456 * be in the same order, even in case of 2 concurrent swaps */
2457 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2458 if (rc == 0) /* same fid ? */
2464 rc = mdd_la_get(env, fst_o, fst_la);
2468 rc = mdd_la_get(env, snd_o, snd_la);
2472 /* check if layout swapping is allowed */
2473 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2477 handle = mdd_trans_create(env, mdd);
2479 RETURN(PTR_ERR(handle));
2481 /* objects are already sorted */
2482 mdd_write_lock(env, fst_o, DT_TGT_CHILD);
2483 mdd_write_lock(env, snd_o, DT_TGT_CHILD);
2485 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2486 if (rc < 0 && rc != -ENODATA)
2489 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2490 if (rc < 0 && rc != -ENODATA)
2493 /* check if file has DoM. DoM file can be migrated only to another
2494 * DoM layout with the same DoM component size or to an non-DOM
2495 * layout. After migration to OSTs layout, local MDT inode data
2496 * should be truncated.
2497 * Objects are sorted by FIDs, considering that original file's FID
2498 * is always smaller the snd_o is always original file we are migrating
2501 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2502 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2504 /* Only migration is supported for DoM files, not 'swap_layouts' so
2505 * target file must be volatile and orphan.
2507 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2508 dom_o = domsize_dom ? snd_o : NULL;
2509 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2510 swap(domsize_dom, domsize_vlt);
2511 dom_o = domsize_dom ? fst_o : NULL;
2512 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2513 /* 'lfs swap_layouts' case, neither file should have DoM */
2515 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, "
2516 "use migration instead: rc = %d\n", rc);
2520 if (domsize_vlt > 0 && domsize_dom == 0) {
2523 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2524 mdd_obj_dev_name(snd_o),
2525 PFID(mdd_object_fid(snd_o)), rc);
2527 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2530 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2531 mdd_obj_dev_name(fst_o),
2532 PFID(mdd_object_fid(fst_o)), rc);
2534 } else if (domsize_vlt > 0) {
2535 /* Migration with the same DOM component size, no need to
2536 * truncate local data, it is still being used */
2540 /* swapping 2 non existant layouts is a success */
2541 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2544 /* to help inode migration between MDT, it is better to
2545 * start by the no layout file (if one), so we order the swap */
2546 if (snd_buf->lb_buf == NULL) {
2548 swap(fst_buf, snd_buf);
2551 fst_gen = snd_gen = 0;
2552 /* lmm and generation layout initialization */
2553 if (fst_buf->lb_buf != NULL) {
2554 fst_lmm = fst_buf->lb_buf;
2555 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2556 fst_fl = LU_XATTR_REPLACE;
2560 fst_fl = LU_XATTR_CREATE;
2563 snd_lmm = snd_buf->lb_buf;
2564 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2566 saved_gen = fst_gen;
2567 /* increase the generation layout numbers */
2572 * XXX The layout generation is used to generate component IDs for
2573 * the composite file, we have to do some special tweaks to make
2574 * sure the layout generation is always adequate for that job.
2577 /* Skip invalid generation number for composite layout */
2578 if ((snd_gen & LCME_ID_MASK) == 0)
2580 if ((fst_gen & LCME_ID_MASK) == 0)
2582 /* Make sure the generation is greater than all the component IDs */
2583 if (fst_gen < snd_gen)
2585 else if (fst_gen > snd_gen)
2588 /* set the file specific informations in lmm */
2589 if (fst_lmm != NULL) {
2590 struct ost_id temp_oi;
2592 saved_oi = &info->mdi_oa.o_oi;
2593 mdd_get_lmm_oi(fst_lmm, saved_oi);
2594 mdd_get_lmm_oi(snd_lmm, &temp_oi);
2595 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2596 mdd_set_lmm_oi(fst_lmm, &temp_oi);
2597 mdd_set_lmm_oi(snd_lmm, saved_oi);
2599 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2600 cpu_to_le32(LOV_MAGIC_MAGIC))
2601 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2603 GOTO(stop, rc = -EPROTO);
2605 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2607 /* Prepare HSM attribute if it's required */
2608 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2609 const int buflen = sizeof(struct hsm_attrs);
2611 lu_buf_alloc(fst_hsm_buf, buflen);
2612 lu_buf_alloc(snd_hsm_buf, buflen);
2613 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2614 GOTO(stop, rc = -ENOMEM);
2616 /* Read HSM attribute */
2617 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2621 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2625 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2626 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2631 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2632 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2638 /* prepare transaction */
2639 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2644 if (fst_buf->lb_buf != NULL)
2645 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2646 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2649 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2654 rc = mdd_trans_start(env, mdd, handle);
2658 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2659 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2663 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2665 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2668 CERROR("%s: HSM error restoring "DFID": rc = %d/%d\n",
2669 mdd_obj_dev_name(fst_o),
2670 PFID(mdd_object_fid(fst_o)), rc, rc2);
2675 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2679 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2682 if (fst_buf->lb_buf != NULL)
2683 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2684 LU_XATTR_REPLACE, handle);
2686 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2689 GOTO(out_restore, rc);
2691 /* Issue one changelog record per file */
2692 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle,
2697 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle,
2707 /* failure on second file, but first was done, so we have
2708 * to roll back first. */
2709 if (fst_buf->lb_buf != NULL) {
2710 mdd_set_lmm_oi(fst_lmm, saved_oi);
2711 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2712 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2713 LU_XATTR_REPLACE, handle);
2715 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2720 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2722 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2728 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf,
2735 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
2736 mdd_obj_dev_name(fst_o),
2737 PFID(mdd_object_fid(snd_o)),
2738 PFID(mdd_object_fid(fst_o)),
2740 /* a solution to avoid journal commit is to panic,
2741 * but it has strong consequences so we use LBUG to
2742 * allow sysdamin to choose to panic or not
2749 rc = mdd_trans_stop(env, mdd, rc, handle);
2751 /* Truncate local DOM data if all went well */
2753 mdd_dom_data_truncate(env, mdd, dom_o);
2755 mdd_write_unlock(env, snd_o);
2756 mdd_write_unlock(env, fst_o);
2758 lu_buf_free(fst_buf);
2759 lu_buf_free(snd_buf);
2760 lu_buf_free(fst_hsm_buf);
2761 lu_buf_free(snd_hsm_buf);
2764 (void) mdd_object_pfid_replace(env, fst_o);
2765 (void) mdd_object_pfid_replace(env, snd_o);
2770 static int mdd_declare_layout_change(const struct lu_env *env,
2771 struct mdd_device *mdd,
2772 struct mdd_object *obj,
2773 struct md_layout_change *mlc,
2774 struct thandle *handle)
2778 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2782 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
2786 /* For PFL, this is used to instantiate necessary component objects. */
2788 mdd_layout_instantiate_component(const struct lu_env *env,
2789 struct mdd_object *obj, struct md_layout_change *mlc,
2790 struct thandle *handle)
2792 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2796 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2799 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2801 * It's possible that another layout write intent has already
2802 * instantiated our objects, so a -EALREADY returned, and we need to
2806 RETURN(rc == -EALREADY ? 0 : rc);
2808 rc = mdd_trans_start(env, mdd, handle);
2812 mdd_write_lock(env, obj, DT_TGT_CHILD);
2813 rc = mdo_layout_change(env, obj, mlc, handle);
2814 mdd_write_unlock(env, obj);
2818 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
2824 * Change the FLR layout from RDONLY to WRITE_PENDING.
2826 * It picks the primary mirror, and bumps the layout version, and set
2827 * layout version xattr to OST objects in a sync tx. In order to facilitate
2828 * the handling of phantom writers from evicted clients, the clients carry
2829 * layout version of the file with write RPC, so that the OSTs can verify
2830 * if the write RPCs are legitimate, meaning not from evicted clients.
2833 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2834 struct md_layout_change *mlc, struct thandle *handle)
2836 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2837 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
2838 struct lustre_som_attrs *som = &mlc->mlc_som;
2843 /* Verify acceptable operations */
2844 switch (mlc->mlc_opc) {
2845 case MD_LAYOUT_WRITE:
2846 case MD_LAYOUT_RESYNC:
2847 /* these are legal operations - this represents the case that
2848 * a few mirrors were missed in the last resync. */
2850 case MD_LAYOUT_RESYNC_DONE:
2855 som_buf->lb_buf = som;
2856 som_buf->lb_len = sizeof(*som);
2857 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2858 if (rc < 0 && rc != -ENODATA)
2862 lustre_som_swab(som);
2863 if (som->lsa_valid & SOM_FL_STRICT)
2864 fl = LU_XATTR_REPLACE;
2866 if (mlc->mlc_opc == MD_LAYOUT_WRITE &&
2867 mlc->mlc_intent->li_extent.e_end > som->lsa_size) {
2868 som->lsa_size = mlc->mlc_intent->li_extent.e_end + 1;
2869 fl = LU_XATTR_REPLACE;
2873 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2878 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2879 XATTR_NAME_SOM, fl, handle);
2884 /* record a changelog for data mover to consume */
2885 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
2889 rc = mdd_trans_start(env, mdd, handle);
2893 /* it needs a sync tx to make FLR to work properly */
2894 handle->th_sync = 1;
2896 mdd_write_lock(env, obj, DT_TGT_CHILD);
2897 rc = mdo_layout_change(env, obj, mlc, handle);
2899 /* SOM state transition from STRICT to STALE */
2900 som->lsa_valid = SOM_FL_STALE;
2901 lustre_som_swab(som);
2902 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2905 mdd_write_unlock(env, obj);
2909 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle, NULL);
2920 * Handle mirrored file state transition when it's in WRITE_PENDING.
2922 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2923 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2924 * layout version will be increased, and the file's state will be changed to
2928 mdd_layout_update_write_pending(const struct lu_env *env,
2929 struct mdd_object *obj, struct md_layout_change *mlc,
2930 struct thandle *handle)
2932 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2933 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
2934 struct lustre_som_attrs *som = &mlc->mlc_som;
2939 switch (mlc->mlc_opc) {
2940 case MD_LAYOUT_RESYNC:
2941 /* Upon receiving the resync request, it should
2942 * instantiate all stale components right away to get ready
2943 * for mirror copy. In order to avoid layout version change,
2944 * client should avoid sending LAYOUT_WRITE request at the
2947 case MD_LAYOUT_WRITE:
2949 * legal race for concurrent write, the file state has been
2950 * changed by another client. Or a jump over file size and
2953 som_buf->lb_buf = som;
2954 som_buf->lb_len = sizeof(*som);
2955 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
2956 if (rc < 0 && rc != -ENODATA)
2960 lustre_som_swab(som);
2961 if (mlc->mlc_intent->li_extent.e_end > som->lsa_size) {
2963 mlc->mlc_intent->li_extent.e_end + 1;
2964 fl = LU_XATTR_REPLACE;
2972 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2977 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2978 XATTR_NAME_SOM, fl, handle);
2983 rc = mdd_trans_start(env, mdd, handle);
2987 /* it needs a sync tx to make FLR to work properly */
2988 handle->th_sync = 1;
2990 mdd_write_lock(env, obj, DT_TGT_CHILD);
2991 rc = mdo_layout_change(env, obj, mlc, handle);
2993 som->lsa_valid = SOM_FL_STALE;
2994 lustre_som_swab(som);
2995 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2998 mdd_write_unlock(env, obj);
3009 * Handle the requests when a FLR file's state is in SYNC_PENDING.
3011 * Only concurrent write and sync complete requests are possible when the
3012 * file is in SYNC_PENDING. For the latter request, it will pass in the
3013 * mirrors that have been synchronized, then the stale bit will be cleared
3014 * to make the file's state turn into RDONLY.
3015 * For concurrent write reqeust, it just needs to change the file's state
3016 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
3017 * version because the version will be increased in the transition to
3018 * SYNC_PENDING later so that it can deny the write request from potential
3019 * evicted SYNC clients. */
3021 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
3022 struct md_layout_change *mlc, struct thandle *handle)
3024 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3025 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3030 /* operation validation */
3031 switch (mlc->mlc_opc) {
3032 case MD_LAYOUT_RESYNC_DONE:
3033 /* resync complete. */
3034 case MD_LAYOUT_WRITE:
3035 /* concurrent write. */
3037 case MD_LAYOUT_RESYNC:
3038 /* resync again, most likely the previous run failed.
3039 * no-op if it's already in SYNC_PENDING state */
3045 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
3046 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
3047 if (rc < 0 && rc != -ENODATA)
3050 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
3051 lustre_som_swab(&mlc->mlc_som);
3052 som_buf->lb_buf = &mlc->mlc_som;
3053 som_buf->lb_len = sizeof(mlc->mlc_som);
3056 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3060 /* record a changelog for the completion of resync */
3061 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
3066 /* RESYNC_DONE has piggybacked size and blocks */
3068 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3069 XATTR_NAME_SOM, fl, handle);
3074 rc = mdd_trans_start(env, mdd, handle);
3078 /* it needs a sync tx to make FLR to work properly */
3079 handle->th_sync = 1;
3081 rc = mdo_layout_change(env, obj, mlc, handle);
3086 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3092 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle,
3102 * Layout change callback for object.
3104 * This is only used by FLR for now. In the future, it can be exteneded to
3105 * handle all layout change.
3108 mdd_layout_change(const struct lu_env *env, struct md_object *o,
3109 struct md_layout_change *mlc)
3111 struct mdd_object *obj = md2mdd_obj(o);
3112 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3113 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
3114 struct lov_comp_md_v1 *lcm;
3115 struct thandle *handle;
3121 if (S_ISDIR(mdd_object_type(obj))) {
3122 switch (mlc->mlc_opc) {
3123 case MD_LAYOUT_SHRINK:
3124 rc = mdd_dir_layout_shrink(env, o, mlc);
3126 case MD_LAYOUT_SPLIT:
3127 rc = mdd_dir_layout_split(env, o, mlc);
3136 /* Verify acceptable operations */
3137 switch (mlc->mlc_opc) {
3138 case MD_LAYOUT_WRITE:
3139 case MD_LAYOUT_RESYNC:
3140 case MD_LAYOUT_RESYNC_DONE:
3146 handle = mdd_trans_create(env, mdd);
3148 RETURN(PTR_ERR(handle));
3150 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
3157 /* analyze the layout to make sure it's a FLR file */
3159 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
3160 GOTO(out, rc = -EINVAL);
3162 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
3164 /* please refer to HLD of FLR for state transition */
3165 switch (flr_state) {
3167 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
3169 case LCM_FL_WRITE_PENDING:
3170 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
3173 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
3175 case LCM_FL_SYNC_PENDING:
3176 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3185 mdd_trans_stop(env, mdd, rc, handle);
3190 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3191 struct mdd_object *child, const struct lu_attr *attr,
3192 const struct md_op_spec *spec,
3193 struct dt_allocation_hint *hint)
3195 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3196 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3197 struct dt_object *nc = mdd_object_child(child);
3199 memset(hint, 0, sizeof(*hint));
3201 /* For striped directory, give striping EA to lod_ah_init, which will
3202 * decide the stripe_offset and stripe count by it. */
3203 if (S_ISDIR(attr->la_mode) &&
3204 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3205 hint->dah_eadata = spec->u.sp_ea.eadata;
3206 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3207 } else if (S_ISREG(attr->la_mode) &&
3208 spec->sp_cr_flags & MDS_OPEN_APPEND) {
3209 hint->dah_append_stripe_count = mdd->mdd_append_stripe_count;
3210 hint->dah_append_pool = mdd->mdd_append_pool;
3213 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3214 hint->dah_eadata, hint->dah_eadata_len);
3215 /* @hint will be initialized by underlying device. */
3216 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3219 static int mdd_accmode(const struct lu_env *env, const struct lu_attr *la,
3222 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3223 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3224 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3225 * owner can write to a file even if it is marked readonly to hide
3226 * its brokenness. (bug 5781) */
3227 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3228 struct lu_ucred *uc = lu_ucred_check(env);
3230 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3234 return mds_accmode(open_flags);
3237 static int mdd_open_sanity_check(const struct lu_env *env,
3238 struct mdd_object *obj,
3239 const struct lu_attr *attr, u64 open_flags,
3242 unsigned int may_mask;
3246 /* EEXIST check, also opening of *open* orphans is allowed so we can
3247 * open-by-handle unlinked files
3249 if (mdd_is_dead_obj(obj) && !is_replay &&
3250 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3253 if (S_ISLNK(attr->la_mode))
3256 may_mask = mdd_accmode(env, attr, open_flags);
3258 if (S_ISDIR(attr->la_mode) && (may_mask & MAY_WRITE))
3261 if (!(open_flags & MDS_OPEN_CREATED)) {
3262 rc = mdd_permission_internal(env, obj, attr, may_mask);
3267 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3268 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3269 open_flags &= ~MDS_OPEN_TRUNC;
3271 /* For writing append-only file must open it with append mode. */
3272 if (attr->la_flags & LUSTRE_APPEND_FL) {
3273 if ((open_flags & MDS_FMODE_WRITE) &&
3274 !(open_flags & MDS_OPEN_APPEND))
3276 if (open_flags & MDS_OPEN_TRUNC)
3283 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3284 u64 open_flags, struct md_op_spec *spec)
3286 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3287 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3288 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3289 struct mdd_object_user *mou = NULL;
3290 const struct lu_ucred *uc = lu_ucred(env);
3291 struct mdd_device *mdd = mdo2mdd(obj);
3292 enum changelog_rec_type type = CL_OPEN;
3296 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3298 rc = mdd_la_get(env, mdd_obj, attr);
3302 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3304 if ((rc == -EACCES) && (mdd->mdd_cl.mc_current_mask & BIT(CL_DN_OPEN)))
3309 mdd_obj->mod_count++;
3311 if (!mdd_changelog_enabled(env, mdd, type))
3315 /* look for existing opener in list under mdd_write_lock */
3316 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3321 /* add user to list */
3322 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3328 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3330 mdd_obj_user_free(mou);
3335 if (type == CL_DN_OPEN) {
3336 if (ktime_before(ktime_get(), mou->mou_deniednext))
3337 /* same user denied again same access within
3338 * time interval: do not record
3342 /* this user already denied, but some time ago:
3343 * update denied time
3345 mou->mou_deniednext =
3346 ktime_add(ktime_get(),
3347 ktime_set(mdd->mdd_cl.mc_deniednext,
3350 mou->mou_opencount++;
3351 /* same user opening file again with same flags:
3358 /* FYI, only the bottom 32 bits of open_flags are recorded */
3359 mdd_changelog(env, type, open_flags, md_dev, mdd_object_fid(mdd_obj));
3363 mdd_write_unlock(env, mdd_obj);
3367 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3368 struct md_attr *ma, struct thandle *handle)
3372 rc = mdd_orphan_declare_delete(env, obj, handle);
3376 return mdo_declare_destroy(env, obj, handle);
3380 * No permission check is needed.
3382 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3383 struct md_attr *ma, u64 open_flags)
3385 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3386 struct mdd_device *mdd = mdo2mdd(obj);
3387 struct thandle *handle = NULL;
3390 bool blocked = false;
3391 bool last_close_by_uid = false;
3392 const struct lu_ucred *uc = lu_ucred(env);
3395 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3396 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3397 mdd_obj->mod_count--;
3398 mdd_write_unlock(env, mdd_obj);
3400 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3401 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
3402 "list\n", PFID(mdd_object_fid(mdd_obj)));
3406 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3407 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
3408 /* check without any lock */
3409 is_orphan = mdd_obj->mod_count == 1 &&
3410 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3414 /* mdd_trans_create() maybe failed because of barrier_entry(),
3415 * under such case, the orphan MDT-object will be left in the
3416 * orphan list, and when the MDT remount next time, the unused
3417 * orphans will be destroyed automatically.
3419 * One exception: the former mdd_finish_unlink may failed to
3420 * add the orphan MDT-object to the orphan list, then if the
3421 * mdd_trans_create() failed because of barrier_entry(), the
3422 * MDT-object will become real orphan that is neither in the
3423 * namespace nor in the orphan list. Such bad case should be
3424 * very rare and will be handled by e2fsck/lfsck. */
3425 handle = mdd_trans_create(env, mdo2mdd(obj));
3426 if (IS_ERR(handle)) {
3427 rc = PTR_ERR(handle);
3428 if (rc != -EINPROGRESS)
3436 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3440 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3445 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3451 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3452 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3454 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3455 lu_dev_name(mdd2lu_dev(mdd)),
3456 PFID(mdd_object_fid(mdd_obj)), rc);
3460 /* check again with lock */
3461 is_orphan = (mdd_obj->mod_count == 1) &&
3462 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3463 ma->ma_attr.la_nlink == 0);
3465 if (is_orphan && !handle && !blocked) {
3466 mdd_write_unlock(env, mdd_obj);
3470 mdd_obj->mod_count--; /*release open count */
3472 /* under mdd write lock */
3473 /* If recording, see if we need to remove UID from list. uc is not
3474 * initialized if the client has been evicted. */
3475 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3476 struct mdd_object_user *mou;
3478 /* look for UID in list */
3479 /* If mou is NULL, it probably means logging was enabled after
3480 * the user had the file open. So the corresponding close
3481 * will not be logged.
3483 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3486 mou->mou_opencount--;
3487 if (mou->mou_opencount == 0) {
3488 mdd_obj_user_remove(mdd_obj, mou);
3489 last_close_by_uid = true;
3494 if (!is_orphan || blocked)
3498 /* NB: Object maybe not in orphan list originally, it is rare case for
3499 * mdd_finish_unlink() failure, in that case, the object doesn't have
3500 * ORPHAN_OBJ flag */
3501 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3502 /* remove link to object from orphan index */
3503 LASSERT(handle != NULL);
3504 rc = mdd_orphan_delete(env, mdd_obj, handle);
3506 CERROR("%s: unable to delete "DFID" from orphan list: "
3507 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3508 PFID(mdd_object_fid(mdd_obj)), rc);
3509 /* If object was not deleted from orphan list, do not
3510 * destroy OSS objects, which will be done when next
3515 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
3516 "list, OSS objects to be destroyed.\n",
3517 PFID(mdd_object_fid(mdd_obj)));
3520 rc = mdo_destroy(env, mdd_obj, handle);
3523 CERROR("%s: unable to delete "DFID" from orphan list: "
3524 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
3525 PFID(mdd_object_fid(mdd_obj)), rc);
3530 mdd_write_unlock(env, mdd_obj);
3532 if (rc != 0 || blocked ||
3533 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3536 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3537 * or if CL_OPEN was recorded and it's last close by user.
3538 * Changelogs mask may change between open and close operations, but
3539 * this is not a big deal if we have a CL_CLOSE entry with no matching
3540 * CL_OPEN. Plus Changelogs mask may not change often.
3542 if (((!(mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3543 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3544 MDS_OPEN_TRUNC))) ||
3545 ((mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3546 last_close_by_uid)) &&
3547 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3548 if (handle == NULL) {
3549 handle = mdd_trans_create(env, mdo2mdd(obj));
3551 GOTO(stop, rc = PTR_ERR(handle));
3553 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3554 NULL, NULL, handle);
3558 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3563 /* FYI, only the bottom 32 bits of open_flags are recorded */
3564 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3565 mdd_obj, handle, NULL);
3569 if (handle != NULL && !IS_ERR(handle))
3570 rc = mdd_trans_stop(env, mdd, rc, handle);
3576 * Permission check is done when open,
3577 * no need check again.
3579 static int mdd_readpage_sanity_check(const struct lu_env *env,
3580 struct mdd_object *obj)
3582 struct dt_object *next = mdd_object_child(obj);
3586 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
3594 static int mdd_dir_page_build(const struct lu_env *env, struct dt_object *obj,
3595 union lu_page *lp, size_t bytes,
3596 const struct dt_it_ops *iops,
3597 struct dt_it *it, __u32 attr, void *arg)
3599 struct lu_dirpage *dp = &lp->lp_dir;
3602 struct lu_dirent *ent;
3603 struct lu_dirent *last = NULL;
3608 if (bytes < sizeof(*dp))
3609 GOTO(out_err, result = -EOVERFLOW);
3611 memset(area, 0, sizeof(*dp));
3612 area += sizeof(*dp);
3613 bytes -= sizeof(*dp);
3620 len = iops->key_size(env, it);
3622 /* IAM iterator can return record with zero len. */
3626 hash = iops->store(env, it);
3627 if (unlikely(first)) {
3629 dp->ldp_hash_start = cpu_to_le64(hash);
3632 /* calculate max space required for lu_dirent */
3633 recsize = lu_dirent_calc_size(len, attr);
3635 if (bytes >= recsize &&
3636 !OBD_FAIL_CHECK(OBD_FAIL_MDS_DIR_PAGE_WALK)) {
3637 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
3638 if (result == -ESTALE)
3643 /* OSD might not able to pack all attributes, so
3644 * recheck record length had room to store FID
3646 recsize = le16_to_cpu(ent->lde_reclen);
3648 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
3649 fid_le_to_cpu(&fid, &ent->lde_fid);
3650 if (fid_is_dot_lustre(&fid))
3651 GOTO(next, recsize);
3654 result = (last != NULL) ? 0 : -EBADSLT;
3658 ent = (void *)ent + recsize;
3662 result = iops->next(env, it);
3663 if (result == -ESTALE)
3665 } while (result == 0);
3668 dp->ldp_hash_end = cpu_to_le64(hash);
3670 if (last->lde_hash == dp->ldp_hash_end)
3671 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
3672 last->lde_reclen = 0; /* end mark */
3676 /* end of directory */
3677 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3678 else if (result < 0)
3679 CWARN("%s: build page failed for "DFID": rc = %d\n",
3680 lu_dev_name(obj->do_lu.lo_dev),
3681 PFID(lu_object_fid(&obj->do_lu)), result);
3686 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
3687 const struct lu_rdpg *rdpg)
3689 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3693 if (mdd_object_exists(mdd_obj) == 0) {
3694 CERROR("%s: object "DFID" not found: rc = -2\n",
3695 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
3699 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
3700 rc = mdd_readpage_sanity_check(env, mdd_obj);
3702 GOTO(out_unlock, rc);
3704 if (mdd_is_dead_obj(mdd_obj)) {
3706 struct lu_dirpage *dp;
3709 * According to POSIX, please do not return any entry to client:
3710 * even dot and dotdot should not be returned.
3712 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
3713 PFID(mdd_object_fid(mdd_obj)));
3715 if (rdpg->rp_count <= 0)
3716 GOTO(out_unlock, rc = -EFAULT);
3717 LASSERT(rdpg->rp_pages != NULL);
3719 pg = rdpg->rp_pages[0];
3720 dp = (struct lu_dirpage *)kmap(pg);
3721 memset(dp, 0 , sizeof(struct lu_dirpage));
3722 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3723 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3724 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3726 GOTO(out_unlock, rc = LU_PAGE_SIZE);
3729 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
3730 mdd_dir_page_build, NULL);
3732 struct lu_dirpage *dp;
3734 dp = kmap(rdpg->rp_pages[0]);
3735 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
3738 * No pages were processed, mark this for first page
3741 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
3742 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
3743 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
3745 kunmap(rdpg->rp_pages[0]);
3748 GOTO(out_unlock, rc);
3750 mdd_read_unlock(env, mdd_obj);
3754 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
3756 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3758 if (mdd_object_exists(mdd_obj) == 0) {
3761 CERROR("%s: object "DFID" not found: rc = %d\n",
3762 mdd_obj_dev_name(mdd_obj),
3763 PFID(mdd_object_fid(mdd_obj)), rc);
3766 return dt_object_sync(env, mdd_object_child(mdd_obj),
3770 static int mdd_object_lock(const struct lu_env *env,
3771 struct md_object *obj,
3772 struct lustre_handle *lh,
3773 struct ldlm_enqueue_info *einfo,
3774 union ldlm_policy_data *policy)
3776 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3777 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3781 static int mdd_object_unlock(const struct lu_env *env,
3782 struct md_object *obj,
3783 struct ldlm_enqueue_info *einfo,
3784 union ldlm_policy_data *policy)
3786 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3787 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3790 const struct md_object_operations mdd_obj_ops = {
3791 .moo_permission = mdd_permission,
3792 .moo_attr_get = mdd_attr_get,
3793 .moo_attr_set = mdd_attr_set,
3794 .moo_xattr_get = mdd_xattr_get,
3795 .moo_xattr_set = mdd_xattr_set,
3796 .moo_xattr_list = mdd_xattr_list,
3797 .moo_invalidate = mdd_invalidate,
3798 .moo_xattr_del = mdd_xattr_del,
3799 .moo_swap_layouts = mdd_swap_layouts,
3800 .moo_open = mdd_open,
3801 .moo_close = mdd_close,
3802 .moo_readpage = mdd_readpage,
3803 .moo_readlink = mdd_readlink,
3804 .moo_changelog = mdd_changelog,
3805 .moo_object_sync = mdd_object_sync,
3806 .moo_object_lock = mdd_object_lock,
3807 .moo_object_unlock = mdd_object_unlock,
3808 .moo_layout_change = mdd_layout_change,