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 /* time of next access denied notificaiton */
61 ktime_t mou_deniednext;
64 static int mdd_xattr_get(const struct lu_env *env,
65 struct md_object *obj, struct lu_buf *buf,
68 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
69 struct mdd_device *mdd,
70 enum changelog_rec_type type,
71 enum changelog_rec_flags clf_flags,
72 const struct lu_fid *fid,
73 const struct lu_fid *pfid,
74 const char *xattr_name,
75 struct thandle *handle);
77 static inline bool has_prefix(const char *str, const char *prefix);
80 static u32 flags_helper(u64 open_flags)
84 if (open_flags & MDS_FMODE_EXEC) {
85 open_mode = MDS_FMODE_EXEC;
87 if (open_flags & MDS_FMODE_READ)
88 open_mode = MDS_FMODE_READ;
90 (MDS_FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
91 open_mode |= MDS_FMODE_WRITE;
97 /** Allocate/init a user and its sub-structures.
102 * \retval mou [OUT] success valid structure
105 static struct mdd_object_user *mdd_obj_user_alloc(u64 open_flags,
106 uid_t uid, gid_t gid)
108 struct mdd_object_user *mou;
112 OBD_SLAB_ALLOC_PTR(mou, mdd_object_kmem);
114 RETURN(ERR_PTR(-ENOMEM));
116 mou->mou_open_flags = open_flags;
117 mou->mou_uidgid = ((__u64)uid << 32) | gid;
118 mou->mou_opencount = 0;
119 mou->mou_deniednext = ktime_set(0, 0);
125 * Free a user and its sub-structures.
127 * \param mou [IN] user to be freed.
129 static void mdd_obj_user_free(struct mdd_object_user *mou)
131 OBD_SLAB_FREE_PTR(mou, mdd_object_kmem);
135 * Find if UID/GID already has this file open
137 * Caller should have write-locked \param mdd_obj.
138 * \param mdd_obj [IN] mdd_obj
139 * \param uid [IN] client uid
140 * \param gid [IN] client gid
141 * \retval user pointer or NULL if not found
144 struct mdd_object_user *mdd_obj_user_find(struct mdd_object *mdd_obj,
145 uid_t uid, gid_t gid,
148 struct mdd_object_user *mou;
153 uidgid = ((__u64)uid << 32) | gid;
154 list_for_each_entry(mou, &mdd_obj->mod_users, mou_list) {
155 if (mou->mou_uidgid == uidgid &&
156 flags_helper(mou->mou_open_flags) ==
157 flags_helper(open_flags))
164 * Add a user to the list of openers for this file
166 * Caller should have write-locked \param mdd_obj.
167 * \param mdd_obj [IN] mdd_obj
168 * \param mou [IN] user
170 * \retval -ve failure
172 static int mdd_obj_user_add(struct mdd_object *mdd_obj,
173 struct mdd_object_user *mou,
176 struct mdd_device *mdd = mdd_obj2mdd_dev(mdd_obj);
177 struct mdd_object_user *tmp;
178 __u32 uid = mou->mou_uidgid >> 32;
179 __u32 gid = mou->mou_uidgid & ((1UL << 32) - 1);
182 tmp = mdd_obj_user_find(mdd_obj, uid, gid, mou->mou_open_flags);
186 list_add_tail(&mou->mou_list, &mdd_obj->mod_users);
189 /* next 'access denied' notification cannot happen before
192 mou->mou_deniednext =
193 ktime_add(ktime_get(),
194 ktime_set(mdd->mdd_cl.mc_deniednext, 0));
196 mou->mou_opencount++;
201 * Remove UID from the list
203 * Caller should have write-locked \param mdd_obj.
204 * \param mdd_obj [IN] mdd_obj
205 * \param uid [IN] user
206 * \retval -ve failure
208 static int mdd_obj_user_remove(struct mdd_object *mdd_obj,
209 struct mdd_object_user *mou)
216 list_del_init(&mou->mou_list);
218 mdd_obj_user_free(mou);
222 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
227 if (mdd_object_exists(obj) == 0)
230 rc = mdo_attr_get(env, obj, la);
231 if (unlikely(rc != 0)) {
233 obj->mod_flags |= DEAD_OBJ;
237 if (la->la_valid & LA_FLAGS && la->la_flags & LUSTRE_ORPHAN_FL)
238 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
243 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
245 return lu_env_info(env, &mdd_thread_key);
248 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
252 buf = &mdd_env_info(env)->mdi_buf[0];
258 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
259 const void *area, ssize_t len)
263 buf = &mdd_env_info(env)->mdi_buf[0];
264 buf->lb_buf = (void *)area;
269 struct lu_object *mdd_object_alloc(const struct lu_env *env,
270 const struct lu_object_header *hdr,
273 struct mdd_device *mdd = lu2mdd_dev(d);
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 obd_obt_init(mdd2obd_dev(mdd));
287 o->lo_ops = &mdd_lu_obj_ops;
288 INIT_LIST_HEAD(&mdd_obj->mod_users);
293 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
294 const struct lu_object_conf *unused)
296 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
297 struct mdd_object *mdd_obj = lu2mdd_obj(o);
298 struct lu_object *below;
299 struct lu_device *under;
303 mdd_obj->mod_cltime = ktime_set(0, 0);
304 under = &d->mdd_child->dd_lu_dev;
305 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
307 RETURN(PTR_ERR(below));
309 lu_object_add(o, below);
314 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
318 if (lu_object_exists(o)) {
319 struct mdd_object *mdd_obj = lu2mdd_obj(o);
320 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
322 rc = mdd_la_get(env, mdd_obj, attr);
328 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
330 struct mdd_object *mdd = lu2mdd_obj(o);
331 struct mdd_object_user *mou, *tmp2;
334 list_for_each_entry_safe(mou, tmp2, &mdd->mod_users, mou_list) {
335 list_del(&mou->mou_list);
336 mdd_obj_user_free(mou);
340 /* mdd doesn't contain an lu_object_header, so don't need call_rcu */
341 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
344 static int mdd_object_print(const struct lu_env *env, void *cookie,
345 lu_printer_t p, const struct lu_object *o)
347 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
349 return (*p)(env, cookie,
350 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
351 mdd, mdd->mod_count, mdd->mod_valid,
352 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
355 static const struct lu_object_operations mdd_lu_obj_ops = {
356 .loo_object_init = mdd_object_init,
357 .loo_object_start = mdd_object_start,
358 .loo_object_free = mdd_object_free,
359 .loo_object_print = mdd_object_print,
362 struct mdd_object *mdd_object_find(const struct lu_env *env,
363 struct mdd_device *d,
364 const struct lu_fid *f)
366 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
370 * No permission check is needed.
372 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
375 struct mdd_object *mdd_obj = md2mdd_obj(obj);
380 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
381 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
382 ma->ma_attr.la_nlink = 0;
388 * No permission check is needed.
390 static int mdd_xattr_get(const struct lu_env *env,
391 struct md_object *obj, struct lu_buf *buf,
394 struct mdd_object *mdd_obj = md2mdd_obj(obj);
395 struct mdd_device *mdd;
400 if (mdd_object_exists(mdd_obj) == 0) {
402 CERROR("%s: object "DFID" not found: rc = %d\n",
403 mdd_obj_dev_name(mdd_obj),
404 PFID(mdd_object_fid(mdd_obj)), rc);
408 /* If the object has been destroyed, then do not get LMVEA, because
409 * it needs to load stripes from the iteration of the master object,
410 * and it will cause problem if master object has been destroyed, see
413 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
414 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
415 strcmp(name, XATTR_NAME_LMV) == 0))
418 /* If the object has been delete from the namespace, then
419 * get linkEA should return -ENOENT as well
421 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
422 strcmp(name, XATTR_NAME_LINK) == 0))
425 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
426 rc = mdo_xattr_get(env, mdd_obj, buf, name);
427 mdd_read_unlock(env, mdd_obj);
431 /* record only getting user xattrs and acls */
432 if (rc >= 0 && buf->lb_buf &&
433 mdd_changelog_enabled(env, mdd, CL_GETXATTR) &&
434 (has_prefix(name, XATTR_USER_PREFIX) ||
435 has_prefix(name, XATTR_NAME_POSIX_ACL_ACCESS) ||
436 has_prefix(name, XATTR_NAME_POSIX_ACL_DEFAULT))) {
437 struct thandle *handle;
440 LASSERT(mdd_object_fid(mdd_obj) != NULL);
442 handle = mdd_trans_create(env, mdd);
444 RETURN(PTR_ERR(handle));
446 rc2 = mdd_declare_changelog_store(env, mdd, CL_GETXATTR, NULL,
451 rc2 = mdd_trans_start(env, mdd, handle);
455 rc2 = mdd_changelog_data_store_by_fid(env, mdd, CL_GETXATTR, 0,
456 mdd_object_fid(mdd_obj),
460 rc2 = mdd_trans_stop(env, mdd, rc2, handle);
468 /* Permission check is done when open, no need check again. */
469 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
472 struct mdd_object *mdd_obj = md2mdd_obj(obj);
473 struct dt_object *next;
479 if (mdd_object_exists(mdd_obj) == 0) {
481 CERROR("%s: object "DFID" not found: rc = %d\n",
482 mdd_obj_dev_name(mdd_obj),
483 PFID(mdd_object_fid(mdd_obj)), rc);
487 next = mdd_object_child(mdd_obj);
488 LASSERT(next != NULL);
489 LASSERT(next->do_body_ops != NULL);
490 LASSERT(next->do_body_ops->dbo_read != NULL);
491 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
492 rc = dt_read(env, next, buf, &pos);
493 mdd_read_unlock(env, mdd_obj);
497 /* No permission check is needed. */
498 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
501 struct mdd_object *mdd_obj = md2mdd_obj(obj);
506 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
507 rc = mdo_xattr_list(env, mdd_obj, buf);
508 mdd_read_unlock(env, mdd_obj);
510 /* If the buffer is NULL then we are only here to get the
511 * length of the xattr name list.
513 if (rc < 0 || buf->lb_buf == NULL)
517 * Filter out XATTR_NAME_LINK if this is an orphan object. See
520 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
521 char *end = (char *)buf->lb_buf + rc;
522 char *p = buf->lb_buf;
525 char *next = p + strlen(p) + 1;
527 if (strcmp(p, XATTR_NAME_LINK) == 0) {
529 memmove(p, next, end - next);
531 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
532 " of orphan "DFID"\n",
533 PFID(mdd_object_fid(mdd_obj)));
544 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
546 return mdo_invalidate(env, md2mdd_obj(obj));
549 int mdd_declare_create_object_internal(const struct lu_env *env,
550 struct mdd_object *p,
551 struct mdd_object *c,
552 struct lu_attr *attr,
553 struct thandle *handle,
554 const struct md_op_spec *spec,
555 struct dt_allocation_hint *hint)
557 struct dt_object_format *dof = &mdd_env_info(env)->mdi_dof;
558 const struct dt_index_features *feat = spec->sp_feat;
563 if (feat != &dt_directory_features && feat != NULL) {
564 dof->dof_type = DFT_INDEX;
565 dof->u.dof_idx.di_feat = feat;
567 dof->dof_type = dt_mode_to_dft(attr->la_mode);
568 if (dof->dof_type == DFT_REGULAR) {
569 dof->u.dof_reg.striped =
570 md_should_create(spec->sp_cr_flags);
571 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
572 dof->u.dof_reg.striped = 0;
573 /* is this replay? */
575 dof->u.dof_reg.striped = 0;
579 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
584 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
585 struct mdd_object *c, struct lu_attr *attr,
586 struct thandle *handle,
587 const struct md_op_spec *spec,
588 struct dt_allocation_hint *hint)
590 struct dt_object_format *dof = &mdd_env_info(env)->mdi_dof;
595 LASSERT(!mdd_object_exists(c));
597 rc = mdo_create_object(env, c, attr, hint, dof, handle);
602 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
603 const struct lu_attr *attr, struct thandle *handle,
610 rc = mdo_attr_set(env, obj, attr, handle);
611 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
612 if (!rc && (attr->la_valid & LA_MODE) && needacl)
613 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
618 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
619 const struct lu_attr *oattr, struct lu_attr *attr,
620 struct thandle *handle)
626 LASSERT(attr->la_valid & LA_CTIME);
627 LASSERT(oattr != NULL);
629 /* Make sure the ctime is increased only, however, it's not strictly
630 * reliable at here because there is not guarantee to hold lock on
631 * object, so we just bypass some unnecessary cmtime setting first
632 * and OSD has to check it again.
634 if (attr->la_ctime < oattr->la_ctime)
635 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
636 else if (attr->la_valid == LA_CTIME &&
637 attr->la_ctime == oattr->la_ctime)
638 attr->la_valid &= ~LA_CTIME;
640 if (attr->la_valid != 0)
641 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
646 static bool is_project_state_change(const struct lu_attr *oattr,
649 if (la->la_valid & LA_PROJID &&
650 oattr->la_projid != la->la_projid)
653 if ((la->la_valid & LA_FLAGS) &&
654 (la->la_flags & LUSTRE_PROJINHERIT_FL) !=
655 (oattr->la_flags & LUSTRE_PROJINHERIT_FL))
662 * This gives the same functionality as the code between
663 * sys_chmod and inode_setattr
664 * chown_common and inode_setattr
665 * utimes and inode_setattr
666 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
668 * @param[in] oattr Original, current object attributes.
669 * @param[in,out] la New attributes to be evaluated and potentially
670 * modified depending on oattr content and other rules.
671 * @param[in] ma md_attr to be evaluated for that object.
673 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
674 const struct lu_attr *oattr, struct lu_attr *la,
675 const struct md_attr *ma)
679 const unsigned long flags = ma->ma_attr_flags;
686 /* Do not permit change file type */
687 if (la->la_valid & LA_TYPE)
690 /* They should not be processed by setattr */
691 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
694 LASSERT(oattr != NULL);
696 uc = lu_ucred_check(env);
700 if (is_project_state_change(oattr, la)) {
701 /* we want rbac roles to have precedence over any other
702 * permission or capability checks
704 if (!uc->uc_rbac_file_perms ||
705 (!cap_raised(uc->uc_cap, CAP_SYS_RESOURCE) &&
706 !lustre_in_group_p(uc, ma->ma_enable_chprojid_gid) &&
707 !(ma->ma_enable_chprojid_gid == -1 &&
708 mdd_permission_internal(env, obj, oattr, MAY_WRITE))))
712 if (la->la_valid == LA_CTIME) {
713 if (!(flags & MDS_PERM_BYPASS))
714 /* This is only for set ctime when rename's source is
717 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
719 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
720 la->la_valid &= ~LA_CTIME;
724 if (flags & MDS_CLOSE_UPDATE_TIMES &&
725 la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
726 /* This is an atime/mtime/ctime attribute update for
729 if (la->la_valid & LA_ATIME &&
730 la->la_atime <= (oattr->la_atime +
731 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
732 la->la_valid &= ~LA_ATIME;
733 if (la->la_valid & LA_CTIME && la->la_ctime <= oattr->la_ctime)
734 la->la_valid &= ~LA_CTIME;
735 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
736 la->la_valid &= ~LA_MTIME;
740 /* Check if flags change. */
741 if (la->la_valid & LA_FLAGS) {
742 unsigned int oldflags = oattr->la_flags &
743 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
744 unsigned int newflags = la->la_flags &
745 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
747 if ((uc->uc_fsuid != oattr->la_uid) &&
748 !cap_raised(uc->uc_cap, CAP_FOWNER))
751 /* The IMMUTABLE and APPEND_ONLY flags can
752 * only be changed by the relevant capability.
754 if ((oldflags ^ newflags) &&
755 !cap_raised(uc->uc_cap, CAP_LINUX_IMMUTABLE))
758 if (!S_ISDIR(oattr->la_mode)) {
759 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
760 } else if (la->la_flags & LUSTRE_ENCRYPT_FL) {
761 /* when trying to add encryption flag on dir,
762 * make sure it is empty
764 rc = mdd_dir_is_empty(env, obj);
771 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
772 (la->la_valid & ~LA_FLAGS) &&
773 !(flags & MDS_PERM_BYPASS))
776 /* Check for setting the obj time. */
777 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
778 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
779 if ((uc->uc_fsuid != oattr->la_uid) &&
780 !cap_raised(uc->uc_cap, CAP_FOWNER)) {
781 rc = mdd_permission_internal(env, obj, oattr,
788 if (la->la_valid & LA_KILL_SUID) {
789 la->la_valid &= ~LA_KILL_SUID;
790 if ((oattr->la_mode & S_ISUID) &&
791 !(la->la_valid & LA_MODE)) {
792 la->la_mode = oattr->la_mode;
793 la->la_valid |= LA_MODE;
795 la->la_mode &= ~S_ISUID;
798 if (la->la_valid & LA_KILL_SGID) {
799 la->la_valid &= ~LA_KILL_SGID;
800 if (((oattr->la_mode & (S_ISGID | 0010)) ==
802 !(la->la_valid & LA_MODE)) {
803 la->la_mode = oattr->la_mode;
804 la->la_valid |= LA_MODE;
806 la->la_mode &= ~S_ISGID;
809 /* Make sure a caller can chmod. */
810 if (la->la_valid & LA_MODE) {
811 /* we want rbac roles to have precedence over any other
812 * permission or capability checks
814 if (!uc->uc_rbac_file_perms ||
815 (!(flags & MDS_PERM_BYPASS) &&
816 (uc->uc_fsuid != oattr->la_uid) &&
817 !cap_raised(uc->uc_cap, CAP_FOWNER)))
820 if (la->la_mode == (umode_t) -1)
821 la->la_mode = oattr->la_mode;
823 la->la_mode = (la->la_mode & S_IALLUGO) |
824 (oattr->la_mode & ~S_IALLUGO);
826 /* Also check the setgid bit! */
827 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
828 la->la_gid : oattr->la_gid) &&
829 !cap_raised(uc->uc_cap, CAP_FSETID))
830 la->la_mode &= ~S_ISGID;
832 la->la_mode = oattr->la_mode;
835 /* Make sure a caller can chown. */
836 if (la->la_valid & LA_UID) {
837 if (la->la_uid == (uid_t) -1)
838 la->la_uid = oattr->la_uid;
839 /* we want rbac roles to have precedence over any other
840 * permission or capability checks
842 if (!uc->uc_rbac_file_perms ||
843 ((uc->uc_fsuid != oattr->la_uid ||
844 la->la_uid != oattr->la_uid) &&
845 !cap_raised(uc->uc_cap, CAP_CHOWN)))
848 /* If the user or group of a non-directory has been
849 * changed by a non-root user, remove the setuid bit.
850 * 19981026 David C Niemi <niemi@tux.org>
852 * Changed this to apply to all users, including root,
853 * to avoid some races. This is the behavior we had in
854 * 2.0. The check for non-root was definitely wrong
855 * for 2.2 anyway, as it should have been using
856 * CAP_FSETID rather than fsuid -- 19990830 SD.
858 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
859 !S_ISDIR(oattr->la_mode)) {
860 la->la_mode &= ~S_ISUID;
861 la->la_valid |= LA_MODE;
865 /* Make sure caller can chgrp. */
866 if (la->la_valid & LA_GID) {
867 if (la->la_gid == (gid_t) -1)
868 la->la_gid = oattr->la_gid;
869 /* we want rbac roles to have precedence over any other
870 * permission or capability checks
872 if (!uc->uc_rbac_file_perms ||
873 ((uc->uc_fsuid != oattr->la_uid ||
874 (la->la_gid != oattr->la_gid &&
875 !lustre_in_group_p(uc, la->la_gid))) &&
876 !cap_raised(uc->uc_cap, CAP_CHOWN)))
879 /* Likewise, if the user or group of a non-directory
880 * has been changed by a non-root user, remove the
881 * setgid bit UNLESS there is no group execute bit
882 * (this would be a file marked for mandatory
883 * locking). 19981026 David C Niemi <niemi@tux.org>
885 * Removed the fsuid check (see the comment above) --
888 if (((oattr->la_mode & (S_ISGID | 0010)) ==
889 (S_ISGID | 0010)) && !S_ISDIR(oattr->la_mode)) {
890 la->la_mode &= ~S_ISGID;
891 la->la_valid |= LA_MODE;
895 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
896 if (!((flags & MDS_OWNEROVERRIDE) &&
897 (uc->uc_fsuid == oattr->la_uid)) &&
898 !(flags & MDS_PERM_BYPASS)) {
899 rc = mdd_permission_internal(env, obj, oattr,
906 if (la->la_valid & LA_CTIME) {
908 * The pure setattr, it has the priority over what is
909 * already set, do not drop it if ctime is equal.
911 if (la->la_ctime < oattr->la_ctime)
912 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
918 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
919 struct mdd_device *mdd,
920 enum changelog_rec_type type,
921 enum changelog_rec_flags clf_flags,
922 const struct lu_fid *fid,
923 const struct lu_fid *pfid,
924 const char *xattr_name,
925 struct thandle *handle)
927 const struct lu_ucred *uc = lu_ucred(env);
928 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
929 struct llog_changelog_rec *rec;
934 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
937 if (uc->uc_jobid[0] != '\0')
938 clf_flags |= CLF_JOBID;
939 xflags |= CLFE_UIDGID;
942 if (type == CL_OPEN || type == CL_DN_OPEN)
944 if (type == CL_SETXATTR || type == CL_GETXATTR)
945 xflags |= CLFE_XATTR;
947 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
948 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
949 xflags & CLFE_SUPPORTED));
950 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_chlg_buf, reclen);
951 if (buf->lb_buf == NULL)
955 rec->cr_hdr.lrh_len = reclen;
956 rec->cr.cr_flags = clf_flags;
957 rec->cr.cr_type = (__u32)type;
958 rec->cr.cr_tfid = *fid;
960 rec->cr.cr_pfid = *pfid;
961 rec->cr.cr_namelen = 0;
963 if (clf_flags & CLF_JOBID)
964 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
966 if (clf_flags & CLF_EXTRA_FLAGS) {
967 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
968 if (xflags & CLFE_UIDGID)
969 mdd_changelog_rec_extra_uidgid(&rec->cr,
970 uc->uc_uid, uc->uc_gid);
971 if (xflags & CLFE_NID)
972 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
973 if (xflags & CLFE_OPEN)
974 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
975 if (xflags & CLFE_XATTR) {
976 if (xattr_name == NULL)
978 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
982 rc = mdd_changelog_store(env, mdd, rec, handle);
987 /** Store a data change changelog record
988 * If this fails, we must fail the whole transaction; we don't
989 * want the change to commit without the log entry.
990 * \param mdd_obj - mdd_object of change
991 * \param handle - transaction handle
992 * \param pfid - parent FID for CL_MTIME changelogs
994 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
995 enum changelog_rec_type type,
996 enum changelog_rec_flags clf_flags,
997 struct mdd_object *mdd_obj, struct thandle *handle,
998 const struct lu_fid *pfid)
1002 LASSERT(mdd_obj != NULL);
1003 LASSERT(handle != NULL);
1005 if (!mdd_changelog_enabled(env, mdd, type))
1008 if (mdd_is_volatile_obj(mdd_obj))
1011 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1012 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1013 /* Don't need multiple updates in this log */
1014 /* Don't check under lock - no big deal if we get extra entry */
1018 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1019 mdd_object_fid(mdd_obj), pfid,
1022 mdd_obj->mod_cltime = ktime_get();
1027 int mdd_changelog_data_store_xattr(const struct lu_env *env,
1028 struct mdd_device *mdd,
1029 enum changelog_rec_type type,
1030 enum changelog_rec_flags clf_flags,
1031 struct mdd_object *mdd_obj,
1032 const char *xattr_name,
1033 struct thandle *handle)
1037 LASSERT(mdd_obj != NULL);
1038 LASSERT(handle != NULL);
1040 if (!mdd_changelog_enabled(env, mdd, type))
1043 if (mdd_is_volatile_obj(mdd_obj))
1046 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1047 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1048 /* Don't need multiple updates in this log */
1049 /* Don't check under lock - no big deal if we get an extra
1055 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1056 mdd_object_fid(mdd_obj), NULL,
1057 xattr_name, handle);
1059 mdd_obj->mod_cltime = ktime_get();
1064 /* Only the bottom CLF_FLAGMASK bits of @flags are stored in the record,
1065 * except for open flags have a dedicated record to store 32 bits of flags.
1067 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1068 enum changelog_rec_flags clf_flags,
1069 struct md_device *m, const struct lu_fid *fid)
1071 struct thandle *handle;
1072 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1077 LASSERT(fid != NULL);
1079 /* We'll check this again below, but we check now before we
1080 * start a transaction.
1082 if (!mdd_changelog_enabled(env, mdd, type))
1085 handle = mdd_trans_create(env, mdd);
1087 RETURN(PTR_ERR(handle));
1089 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1093 rc = mdd_trans_start(env, mdd, handle);
1097 /* note we only store the low CLF_FLAGMASK bits of la_valid */
1098 rc = mdd_changelog_data_store_by_fid(env, mdd, type,
1099 clf_flags & CLF_FLAGMASK,
1100 fid, NULL, NULL, handle);
1103 rc = mdd_trans_stop(env, mdd, rc, handle);
1109 * Save LMA extended attributes with data from \a ma.
1111 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1112 * not, LMA EA will be first read from disk, modified and write back.
1115 /* Precedence for choosing record type when multiple
1116 * attributes change: setattr > mtime > ctime > atime
1117 * (ctime changes when mtime does, plus chmod/chown.
1118 * atime and ctime are independent.)
1120 static int mdd_attr_set_changelog(const struct lu_env *env,
1121 struct md_object *obj, struct thandle *handle,
1122 const struct lu_fid *pfid, __u64 valid)
1124 struct mdd_device *mdd = mdo2mdd(obj);
1127 bits = (valid & LA_SIZE) ? BIT(CL_TRUNC) : 0;
1128 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? BIT(CL_SETATTR) : 0;
1129 bits |= (valid & LA_MTIME) ? BIT(CL_MTIME) : 0;
1130 bits |= (valid & LA_CTIME) ? BIT(CL_CTIME) : 0;
1131 bits |= (valid & LA_ATIME) ? BIT(CL_ATIME) : 0;
1132 bits = bits & mdd->mdd_cl.mc_current_mask;
1133 /* This is an implementation limit rather than a protocol limit */
1134 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1138 /* The record type is the lowest non-masked set bit */
1141 /* only the low CLF_FLAGMASK bits of la_valid are stored */
1142 return mdd_changelog_data_store(env, mdd, type, valid & CLF_FLAGMASK,
1143 md2mdd_obj(obj), handle, pfid);
1146 static int mdd_declare_attr_set(const struct lu_env *env,
1147 struct mdd_device *mdd,
1148 struct mdd_object *obj,
1149 const struct lu_attr *attr,
1150 struct thandle *handle)
1154 rc = mdo_declare_attr_set(env, obj, attr, handle);
1158 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1159 if (attr->la_valid & LA_MODE) {
1160 mdd_read_lock(env, obj, DT_TGT_CHILD);
1161 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1162 XATTR_NAME_ACL_ACCESS);
1163 mdd_read_unlock(env, obj);
1164 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1170 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1172 rc = mdo_declare_xattr_set(env, obj, buf,
1173 XATTR_NAME_ACL_ACCESS, 0,
1181 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1190 * permission changes may require sync operation, to mitigate performance
1191 * impact, only do this for dir and when permission is reduced.
1193 * For regular files, version is updated with permission change (see VBR), async
1194 * permission won't cause any issue, while missing permission change on
1195 * directory may affect accessibility of other objects after recovery.
1197 static inline bool permission_needs_sync(const struct lu_attr *old,
1198 const struct lu_attr *new)
1200 if (!S_ISDIR(old->la_mode))
1203 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1206 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1209 if (new->la_valid & LA_MODE) {
1210 /* turned on sticky bit */
1211 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1214 /* set-GID has no impact on what is allowed, not checked */
1216 /* turned off setuid bit, or one of rwx for someone */
1217 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1218 (old->la_mode & (0777 | S_ISUID)))
1225 static inline __u64 mdd_lmm_dom_size(void *buf)
1227 struct lov_mds_md *lmm = buf;
1228 struct lov_comp_md_v1 *comp_v1;
1229 struct lov_mds_md *v1;
1235 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1238 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1239 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1240 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1242 /* DoM entry is the first entry always */
1243 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) & LOV_PATTERN_MDT)
1244 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1249 /* set attr and LOV EA at once, return updated attr */
1250 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1251 const struct md_attr *ma)
1253 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1254 struct mdd_device *mdd = mdo2mdd(obj);
1255 struct thandle *handle = NULL;
1256 struct lu_attr *la_copy = &mdd_env_info(env)->mdi_la_for_fix;
1257 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1258 const struct lu_attr *la = &ma->ma_attr;
1259 struct lu_ucred *uc;
1260 struct lquota_id_info qi;
1261 bool quota_reserved = false;
1262 bool chrgrp_by_unprivileged_user = false;
1267 /* we do not use ->attr_set() for LOV/HSM EA any more */
1268 LASSERT((ma->ma_valid & MA_LOV) == 0);
1269 LASSERT((ma->ma_valid & MA_HSM) == 0);
1271 rc = mdd_la_get(env, mdd_obj, attr);
1275 *la_copy = ma->ma_attr;
1276 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1280 /* no need to setattr anymore */
1281 if (la_copy->la_valid == 0) {
1283 "%s: no valid attribute on "DFID", previous was %#llx\n",
1284 mdd_obj_dev_name(mdd_obj),
1285 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1290 /* If an unprivileged user changes group of some file,
1291 * the setattr operation will be processed synchronously to
1292 * honor the quota limit of the corresponding group. see LU-5152
1294 uc = lu_ucred_check(env);
1295 memset(&qi, 0, sizeof(qi));
1296 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1297 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1298 CDEBUG(D_QUOTA, "%s: reserve quota for changing group: gid=%u size=%llu\n",
1299 mdd2obd_dev(mdd)->obd_name, la->la_gid, la->la_size);
1301 if (la->la_valid & LA_BLOCKS)
1302 qi.lqi_space = la->la_blocks << 9;
1303 else if (la->la_valid & LA_SIZE)
1304 qi.lqi_space = la->la_size;
1305 /* use local attr gotten above */
1306 else if (attr->la_valid & LA_BLOCKS)
1307 qi.lqi_space = attr->la_blocks << 9;
1308 else if (attr->la_valid & LA_SIZE)
1309 qi.lqi_space = attr->la_size;
1311 if (qi.lqi_space > 0) {
1312 qi.lqi_id.qid_gid = la->la_gid;
1313 qi.lqi_type = GRPQUOTA;
1314 qi.lqi_space = toqb(qi.lqi_space);
1315 qi.lqi_is_blk = true;
1316 rc = dt_reserve_or_free_quota(env, mdd->mdd_bottom,
1320 CDEBUG(D_QUOTA, "%s: failed to reserve quota for gid %d size %llu\n",
1321 mdd2obd_dev(mdd)->obd_name,
1322 la->la_gid, qi.lqi_space);
1327 quota_reserved = true;
1328 la_copy->la_valid |= LA_FLAGS;
1331 chrgrp_by_unprivileged_user = true;
1333 /* Flush the possible existing client setattr requests to OSTs
1334 * to keep the order with the current setattr operation that
1335 * will be sent directly to OSTs. see LU-5152
1337 * LU-11303 disable sync as this is too heavyweight.
1338 * This should be replaced with a sync only for the object
1339 * being modified here, not the whole filesystem.
1340 rc = dt_sync(env, mdd->mdd_child);
1346 handle = mdd_trans_create(env, mdd);
1347 if (IS_ERR(handle)) {
1348 rc = PTR_ERR(handle);
1354 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1358 rc = mdd_trans_start(env, mdd, handle);
1362 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1363 permission_needs_sync(attr, la))
1364 handle->th_sync = 1;
1366 if (la->la_valid & (LA_MTIME | LA_CTIME))
1367 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1368 la->la_mtime, la->la_ctime);
1370 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1372 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1373 * otherwise following truncate will do nothing and truncated
1374 * data may be read again. This is a quick fix until LU-11033
1377 if (la_copy->la_valid & LA_SIZE) {
1378 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1380 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1384 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1385 la_copy->la_valid &= ~LA_SIZE;
1386 lu_buf_free(lov_buf);
1390 if (la_copy->la_valid) {
1391 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1393 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1394 /* rollback to the original gid */
1395 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1396 la_copy->la_gid = attr->la_gid;
1397 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1400 mdd_write_unlock(env, mdd_obj);
1404 rc = mdd_attr_set_changelog(env, obj, handle, &ma->ma_pfid,
1407 if (rc == 0 && quota_reserved) {
1408 struct thandle *sub_th;
1410 sub_th = thandle_get_sub_by_dt(env, handle, mdd->mdd_bottom);
1411 if (unlikely(IS_ERR(sub_th))) {
1413 dt_reserve_or_free_quota(env, mdd->mdd_bottom, &qi);
1415 sub_th->th_reserved_quota = qi;
1417 } else if (quota_reserved) {
1419 dt_reserve_or_free_quota(env, mdd->mdd_bottom, &qi);
1423 rc = mdd_trans_stop(env, mdd, rc, handle);
1428 static int mdd_xattr_sanity_check(const struct lu_env *env,
1429 struct mdd_object *obj,
1430 const struct lu_attr *attr,
1433 struct lu_ucred *uc = lu_ucred_assert(env);
1437 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1440 if (strncmp(XATTR_USER_PREFIX, name,
1441 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1442 /* For sticky directories, only the owner and privileged user
1443 * can write attributes.
1445 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1446 (uc->uc_fsuid != attr->la_uid) &&
1447 !cap_raised(uc->uc_cap, CAP_FOWNER))
1449 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1450 (uc->uc_fsuid != attr->la_uid) &&
1451 !cap_raised(uc->uc_cap, CAP_FOWNER)) {
1459 * Check if a string begins with a given prefix.
1461 * \param str String to check
1462 * \param prefix Substring to check at the beginning of \a str
1463 * \return true/false whether the condition is verified.
1465 static inline bool has_prefix(const char *str, const char *prefix)
1467 return strncmp(prefix, str, strlen(prefix)) == 0;
1471 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1473 * \param[in] xattr_name Full extended attribute name.
1475 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1477 static enum changelog_rec_type
1478 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1479 const char *xattr_name)
1481 /* Layout changes systematically recorded */
1482 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1483 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1484 allowed_lustre_lov(xattr_name))
1487 /* HSM information changes systematically recorded */
1488 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1491 /* Avoid logging SOM xattr for every file */
1492 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1495 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1496 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1497 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1498 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1499 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1505 static int mdd_declare_xattr_set(const struct lu_env *env,
1506 struct mdd_device *mdd,
1507 struct mdd_object *obj,
1508 const struct lu_buf *buf,
1510 int fl, struct thandle *handle)
1512 enum changelog_rec_type type;
1515 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1519 type = mdd_xattr_changelog_type(env, mdd, name);
1521 return 0; /* no changelog to store */
1523 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1527 * Compare current and future data of HSM EA and add a changelog if needed.
1529 * Caller should have write-locked \param obj.
1531 * \param buf - Future HSM EA content.
1532 * \retval 0 if no changelog is needed or changelog was added properly.
1533 * \retval -ve errno if there was a problem
1535 static int mdd_hsm_update_locked(const struct lu_env *env,
1536 struct md_object *obj,
1537 const struct lu_buf *buf,
1538 struct thandle *handle,
1539 enum changelog_rec_flags *clf_flags)
1541 struct mdd_thread_info *info = mdd_env_info(env);
1542 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1543 struct lu_buf *current_buf;
1544 struct md_hsm *current_mh;
1545 struct md_hsm *new_mh;
1549 OBD_ALLOC_PTR(current_mh);
1550 if (current_mh == NULL)
1553 /* Read HSM attrs from disk */
1554 current_buf = lu_buf_check_and_alloc(&info->mdi_xattr_buf,
1556 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1558 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1559 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1560 if (rc < 0 && rc != -ENODATA)
1562 else if (rc == -ENODATA)
1563 current_mh->mh_flags = 0;
1565 /* Map future HSM xattr */
1566 OBD_ALLOC_PTR(new_mh);
1568 GOTO(free, rc = -ENOMEM);
1569 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1573 /* Flags differ, set flags for the changelog that will be added */
1574 if (current_mh->mh_flags != new_mh->mh_flags) {
1575 hsm_set_cl_event(clf_flags, HE_STATE);
1576 if (new_mh->mh_flags & HS_DIRTY)
1577 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1580 OBD_FREE_PTR(new_mh);
1583 OBD_FREE_PTR(current_mh);
1587 static int mdd_object_pfid_replace(const struct lu_env *env,
1588 struct mdd_object *o)
1590 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1591 struct thandle *handle;
1594 handle = mdd_trans_create(env, mdd);
1596 RETURN(PTR_ERR(handle));
1598 handle->th_complex = 1;
1600 /* it doesn't need to track the PFID update via llog, because LFSCK
1601 * will repair it even it goes wrong
1603 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1608 rc = mdd_trans_start(env, mdd, handle);
1612 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1617 mdd_trans_stop(env, mdd, rc, handle);
1622 static int mdd_declare_xattr_del(const struct lu_env *env,
1623 struct mdd_device *mdd,
1624 struct mdd_object *obj,
1626 struct thandle *handle);
1628 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1631 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1632 struct md_object *md_vic)
1634 struct mdd_device *mdd = mdo2mdd(md_obj);
1635 struct mdd_object *obj = md2mdd_obj(md_obj);
1636 struct mdd_object *vic = md2mdd_obj(md_vic);
1638 struct lu_buf *buf_vic;
1639 struct lov_mds_md *lmm;
1640 struct thandle *handle;
1641 struct lu_attr *cattr = MDD_ENV_VAR(env, cattr);
1642 struct lu_attr *tattr = MDD_ENV_VAR(env, tattr);
1643 bool is_same_projid;
1649 rc = mdd_la_get(env, obj, cattr);
1653 rc = mdd_la_get(env, vic, tattr);
1657 is_same_projid = cattr->la_projid == tattr->la_projid;
1659 if (lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic)) == 0)
1662 handle = mdd_trans_create(env, mdd);
1664 RETURN(PTR_ERR(handle));
1666 buf = &mdd_env_info(env)->mdi_buf[0];
1667 buf_vic = &mdd_env_info(env)->mdi_buf[1];
1669 if (!is_same_projid) {
1670 rc = mdd_declare_attr_set(env, mdd, vic, cattr, handle);
1675 /* get EA of victim file */
1676 memset(buf_vic, 0, sizeof(*buf_vic));
1677 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1684 /* parse the layout of victim file */
1685 lmm = buf_vic->lb_buf;
1686 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1687 GOTO(stop, rc = -EINVAL);
1689 /* save EA of target file for restore */
1690 memset(buf, 0, sizeof(*buf));
1691 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1695 /* Get rid of the layout from victim object */
1696 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1700 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_NAME_LOV,
1701 LU_XATTR_MERGE, handle);
1705 rc = mdd_trans_start(env, mdd, handle);
1709 rc = mdd_write_lock_two_objects(env, obj, vic);
1713 if (!is_same_projid) {
1714 cattr->la_valid = LA_PROJID;
1715 rc = mdd_attr_set_internal(env, vic, cattr, handle, 0);
1720 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV, LU_XATTR_MERGE,
1725 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1727 GOTO(out_restore, rc);
1729 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1731 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle,
1736 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1737 LU_XATTR_REPLACE, handle);
1739 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1740 mdd_obj_dev_name(obj),
1741 PFID(mdd_object_fid(obj)), rc2);
1745 mdd_write_unlock_two_objects(env, obj, vic);
1747 mdd_trans_stop(env, mdd, rc, handle);
1749 lu_buf_free(buf_vic);
1752 * -EAGAIN means transaction execution phase detect the layout
1753 * has been changed by others.
1756 (void) mdd_object_pfid_replace(env, obj);
1757 else if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
1758 GOTO(retry, retried);
1764 * Extract the mirror with specified mirror id, and store the splitted
1765 * mirror layout to @buf.
1767 * \param[in] comp_v1 mirrored layout
1768 * \param[in] mirror_id the mirror with mirror_id to be extracted
1769 * \param[out] buf store the layout excluding the extracted mirror,
1770 * caller free the buffer we allocated in this function
1771 * \param[out] buf_vic store the extracted layout, caller free the buffer
1772 * we allocated in this function
1774 * \retval 0 on success; < 0 if error happens
1776 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1777 struct lu_buf *buf, struct lu_buf *buf_vic)
1779 struct lov_comp_md_v1 *comp_rem;
1780 struct lov_comp_md_v1 *comp_vic;
1781 struct lov_comp_md_entry_v1 *entry;
1782 struct lov_comp_md_entry_v1 *entry_rem;
1783 struct lov_comp_md_entry_v1 *entry_vic;
1785 __u16 comp_cnt, count = 0;
1786 int lmm_size, lmm_size_vic = 0;
1788 int offset, offset_rem, offset_vic;
1790 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1791 /* comp_v1 should contains more than 1 mirror */
1792 if (mirror_cnt <= 1)
1794 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1795 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1797 for (i = 0; i < comp_cnt; i++) {
1798 entry = &comp_v1->lcm_entries[i];
1799 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1801 lmm_size_vic += sizeof(*entry);
1802 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1803 } else if (count > 0) {
1804 /* find the specified mirror */
1812 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1816 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1817 if (!buf_vic->lb_buf) {
1822 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1823 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1825 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1826 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1827 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1828 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1829 if (!comp_rem->lcm_mirror_count)
1830 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1832 memset(comp_vic, 0, sizeof(*comp_v1));
1833 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1834 comp_vic->lcm_mirror_count = 0;
1835 comp_vic->lcm_entry_count = cpu_to_le32(count);
1836 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1837 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1838 comp_vic->lcm_layout_gen = 0;
1840 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1841 offset_rem = sizeof(*comp_rem) +
1842 sizeof(*entry_rem) * (comp_cnt - count);
1843 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1844 for (i = j = k = 0; i < comp_cnt; i++) {
1845 struct lov_mds_md *lmm, *lmm_dst;
1848 entry = &comp_v1->lcm_entries[i];
1849 entry_vic = &comp_vic->lcm_entries[j];
1850 entry_rem = &comp_rem->lcm_entries[k];
1852 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1855 /* copy component entry */
1857 memcpy(entry_vic, entry, sizeof(*entry));
1858 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1859 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1862 memcpy(entry_rem, entry, sizeof(*entry));
1863 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1867 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1869 lmm_dst = (struct lov_mds_md *)
1870 ((char *)comp_vic + offset_vic);
1872 lmm_dst = (struct lov_mds_md *)
1873 ((char *)comp_rem + offset_rem);
1875 /* copy component entry blob */
1876 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1878 /* blob offset advance */
1879 offset += le32_to_cpu(entry->lcme_size);
1881 offset_vic += le32_to_cpu(entry->lcme_size);
1883 offset_rem += le32_to_cpu(entry->lcme_size);
1889 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1890 struct md_rejig_data *mrd)
1892 struct mdd_device *mdd = mdo2mdd(md_obj);
1893 struct mdd_object *obj = md2mdd_obj(md_obj);
1894 struct mdd_object *vic = NULL;
1896 struct lu_buf *buf_save;
1897 struct lu_buf *buf_vic;
1898 struct lov_comp_md_v1 *lcm;
1899 struct thandle *handle;
1901 bool dom_stripe = false;
1907 * NULL @mrd_obj means mirror deleting, and use NULL vic to indicate
1911 vic = md2mdd_obj(mrd->mrd_obj);
1914 buf = &mdd_env_info(env)->mdi_buf[0];
1915 buf_save = &mdd_env_info(env)->mdi_buf[1];
1916 buf_vic = &mdd_env_info(env)->mdi_buf[2];
1918 handle = mdd_trans_create(env, mdd);
1920 RETURN(PTR_ERR(handle));
1922 /* get EA of mirrored file */
1923 memset(buf_save, 0, sizeof(*buf));
1924 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1928 lcm = buf_save->lb_buf;
1929 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1930 GOTO(stop, rc = -EINVAL);
1933 * Extract the mirror with specified mirror id, and store the splitted
1934 * mirror layout to the victim buffer.
1936 memset(buf, 0, sizeof(*buf));
1937 memset(buf_vic, 0, sizeof(*buf_vic));
1938 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1942 * @buf stores layout w/o the specified mirror, @buf_vic stores the
1946 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1950 * non delete mirror split
1952 * declare obj set remaining layout in @buf, will set obj's
1955 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1956 LU_XATTR_SPLIT, handle);
1960 /* declare vic set splitted layout in @buf_vic */
1961 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic,
1962 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1968 * declare delete mirror objects in @buf_vic, will change obj's
1971 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic,
1972 XATTR_NAME_LOV, LU_XATTR_PURGE,
1977 /* declare obj set remaining layout in @buf */
1978 rc = mdd_declare_xattr_set(env, mdd, obj, buf,
1979 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1985 rc = mdd_trans_start(env, mdd, handle);
1989 rc = mdd_write_lock_two_objects(env, obj, vic);
1993 /* set obj's layout in @buf */
1994 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_SPLIT,
2000 /* set vic's layout in @buf_vic */
2001 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV,
2002 LU_XATTR_CREATE, handle);
2004 GOTO(out_restore, rc);
2006 /* delete mirror objects */
2007 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV,
2008 LU_XATTR_PURGE, handle);
2010 GOTO(out_restore, rc);
2013 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
2016 GOTO(out_restore, rc);
2019 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic,
2022 GOTO(out_restore, rc);
2027 /* restore obj's in-memory and on-disk layout */
2028 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
2029 LU_XATTR_REPLACE, handle);
2031 CERROR("%s: failed rollback "DFID" layout: file state unknown: rc = %d\n",
2032 mdd_obj_dev_name(obj),
2033 PFID(mdd_object_fid(obj)), rc);
2037 mdd_write_unlock_two_objects(env, obj, vic);
2039 rc = mdd_trans_stop(env, mdd, rc, handle);
2041 /* Truncate local DOM data if all went well */
2042 if (!rc && dom_stripe)
2043 mdd_dom_fixup(env, mdd, obj, NULL);
2045 lu_buf_free(buf_save);
2047 lu_buf_free(buf_vic);
2050 * -EAGAIN means transaction execution phase detect the layout
2051 * has been changed by others.
2054 (void) mdd_object_pfid_replace(env, obj);
2055 else if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
2056 GOTO(retry, retried);
2061 static int mdd_layout_merge_allowed(const struct lu_env *env,
2062 struct md_object *target,
2063 struct md_object *victim)
2065 struct mdd_object *o1 = md2mdd_obj(target);
2068 /* cannot extend directory's LOVEA */
2069 if (S_ISDIR(mdd_object_type(o1))) {
2071 CERROR("%s: Don't extend directory's LOVEA, just set it: rc = %d\n",
2072 mdd_obj_dev_name(o1), rc);
2080 * The caller should guarantee to update the object ctime
2081 * after xattr_set if needed.
2083 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
2084 const struct lu_buf *buf, const char *name,
2087 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2088 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2089 struct mdd_device *mdd = mdo2mdd(obj);
2090 struct thandle *handle;
2091 enum changelog_rec_type cl_type;
2092 enum changelog_rec_flags clf_flags = 0;
2099 rc = mdd_la_get(env, mdd_obj, attr);
2103 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2107 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
2108 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
2109 struct md_rejig_data *mrd = buf->lb_buf;
2110 struct md_object *victim = mrd->mrd_obj;
2112 if (buf->lb_len != sizeof(*mrd))
2116 if (fl == LU_XATTR_MERGE) {
2117 rc = mdd_layout_merge_allowed(env, obj, victim);
2120 /* merge layout of victim as a mirror of obj's. */
2121 rc = mdd_xattr_merge(env, obj, victim);
2123 rc = mdd_xattr_split(env, obj, mrd);
2128 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
2129 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
2130 struct posix_acl *acl;
2132 /* user set empty ACL, this should be treated as removing ACL */
2133 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
2136 RETURN(PTR_ERR(acl));
2138 rc = mdd_xattr_del(env, obj, name);
2141 posix_acl_release(acl);
2144 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
2145 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
2149 handle = mdd_trans_create(env, mdd);
2151 RETURN(PTR_ERR(handle));
2153 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
2157 rc = mdd_trans_start(env, mdd, handle);
2161 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2163 if (strcmp(XATTR_NAME_HSM, name) == 0) {
2164 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
2166 GOTO(out_unlock, rc);
2169 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
2171 mdd_write_unlock(env, mdd_obj);
2175 cl_type = mdd_xattr_changelog_type(env, mdd, name);
2176 if (cl_type != CL_NONE)
2177 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type,
2178 clf_flags, mdd_obj, name,
2182 rc = mdd_trans_stop(env, mdd, rc, handle);
2185 * -EAGAIN means transaction execution phase detect the layout
2186 * has been changed by others.
2188 if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
2189 GOTO(retry, retried);
2194 static int mdd_declare_xattr_del(const struct lu_env *env,
2195 struct mdd_device *mdd,
2196 struct mdd_object *obj,
2198 struct thandle *handle)
2200 enum changelog_rec_type type;
2203 rc = mdo_declare_xattr_del(env, obj, name, handle);
2207 type = mdd_xattr_changelog_type(env, mdd, name);
2209 return 0; /* no changelog to store */
2211 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2215 * The caller should guarantee to update the object ctime
2216 * after xattr_set if needed.
2218 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2221 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2222 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2223 struct mdd_device *mdd = mdo2mdd(obj);
2224 struct thandle *handle;
2229 rc = mdd_la_get(env, mdd_obj, attr);
2233 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2237 handle = mdd_trans_create(env, mdd);
2239 RETURN(PTR_ERR(handle));
2241 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2245 rc = mdd_trans_start(env, mdd, handle);
2249 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2250 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2251 mdd_write_unlock(env, mdd_obj);
2255 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2258 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2263 return mdd_trans_stop(env, mdd, rc, handle);
2267 * read lov/lmv EA of an object
2268 * return the lov/lmv EA in an allocated lu_buf
2270 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2271 struct lu_buf *lmm_buf, const char *name)
2273 struct lu_buf *buf = &mdd_env_info(env)->mdi_big_buf;
2278 if (buf->lb_buf == NULL) {
2279 buf = lu_buf_check_and_alloc(buf, 4096);
2280 if (buf->lb_buf == NULL)
2285 rc = mdo_xattr_get(env, obj, buf, name);
2286 if (rc == -ERANGE) {
2287 /* mdi_big_buf is allocated but too small need to increase it */
2288 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_big_buf,
2290 if (buf->lb_buf == NULL)
2293 } else if (rc < 0) {
2295 } else if (rc == 0) {
2299 lu_buf_alloc(lmm_buf, rc);
2300 if (lmm_buf->lb_buf == NULL)
2304 * we don't use lmm_buf directly, because we don't know xattr size, so
2305 * by using mdi_big_buf we can avoid calling mdo_xattr_get() twice.
2307 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2312 static int emit_changelog_after_swap_layout(const struct lu_env *env,
2313 struct thandle *handle,
2314 struct mdd_object *o,
2315 struct lu_buf *hsm_buf)
2317 enum changelog_rec_flags flags = 0;
2318 enum changelog_rec_type type;
2319 enum hsm_states hsm_flags;
2320 struct hsm_attrs *attrs;
2323 attrs = hsm_buf->lb_buf;
2324 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2326 if ((hsm_flags & HS_RELEASED) && !mdd_is_dead_obj(o)) {
2327 hsm_set_cl_event(&flags, HE_RELEASE);
2333 return mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), type,
2334 flags, o, handle, NULL);
2338 * check if layout swapping between 2 objects is allowed
2340 * - only normal FIDs or non-system IGIFs
2341 * - same type of objects
2342 * - same owner/group (so quotas are still valid) unless this is from HSM
2345 static int mdd_layout_swap_allowed(const struct lu_env *env,
2346 struct mdd_object *o1,
2347 const struct lu_attr *attr1,
2348 struct mdd_object *o2,
2349 const struct lu_attr *attr2,
2352 const struct lu_fid *fid1, *fid2;
2356 fid1 = mdd_object_fid(o1);
2357 fid2 = mdd_object_fid(o2);
2359 if (!fid_is_norm(fid1) &&
2360 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2363 if (!fid_is_norm(fid2) &&
2364 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2367 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2368 if (S_ISDIR(mdd_object_type(o1)))
2370 if (S_ISREG(mdd_object_type(o1)))
2375 /* Do not check uid/gid for release since the orphan is created as root
2377 if (flags & SWAP_LAYOUTS_MDS_RELEASE)
2380 if ((attr1->la_uid != attr2->la_uid) ||
2381 (attr1->la_gid != attr2->la_gid))
2387 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2388 * look into the layout in MDD layer.
2390 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2392 struct lov_comp_md_v1 *comp_v1;
2393 struct lov_mds_md *v1;
2397 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2398 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2399 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2407 off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
2408 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2409 if (le32_to_cpu(v1->lmm_magic) != LOV_MAGIC_FOREIGN) {
2417 comp_v1->lcm_entries[i].lcme_offset);
2418 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2420 if (le32_to_cpu(v1->lmm_magic) ==
2427 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2428 off = le32_to_cpu(comp_v1->lcm_entries[i].
2430 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2432 if (le32_to_cpu(v1->lmm_magic) !=
2437 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2438 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2449 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2451 return mdd_lmm_oi(lmm, oi, true);
2454 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2456 return mdd_lmm_oi(lmm, oi, false);
2459 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2461 struct lov_comp_md_v1 *comp_v1;
2463 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2464 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2466 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2468 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2469 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2470 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2471 __u16 tmp_gen = *gen;
2474 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2476 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2483 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2485 return mdd_lmm_gen(lmm, gen, true);
2488 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2490 return mdd_lmm_gen(lmm, gen, false);
2493 int mdd_dom_fixup(const struct lu_env *env, struct mdd_device *mdd,
2494 struct mdd_object *mo, struct mdd_object *vo)
2496 struct dt_object *dom, *vlt;
2497 dt_obj_version_t dv = 0;
2504 vlt = dt_object_locate(mdd_object_child(vo), mdd->mdd_bottom);
2506 GOTO(out, rc = -ENOENT);
2507 dv = dt_data_version_get(env, vlt);
2509 GOTO(out, rc = -ENODATA);
2512 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2514 th = dt_trans_create(env, mdd->mdd_bottom);
2516 GOTO(out, rc = PTR_ERR(th));
2519 rc = dt_declare_data_version_set(env, dom, th);
2523 rc = dt_declare_data_version_del(env, dom, th);
2526 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2531 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2536 dt_data_version_set(env, dom, dv, th);
2538 dt_data_version_del(env, dom, th);
2539 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2543 dt_trans_stop(env, mdd->mdd_bottom, th);
2545 /* Ignore failure but report the error */
2547 CERROR("%s: can't manage DOM file "DFID" data: rc = %d\n",
2548 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2552 /* Swap Layout HSM object information: every information needed to decide how
2553 * to update the HSM xattr flag during a layout swap.
2555 struct sl_hsm_object_info {
2556 struct mdd_object *o; /* the object whose HSM attribute to check */
2557 struct lu_buf *hsm_buf; /* buffer initialized with the content of the
2560 int xattr_flags; /* 0: do not update HSM xattr
2561 * LU_XATTR_REPLACE or LU_XATTR_CREATE: update
2562 * the xattr with this flag
2564 __u64 dv; /* dataversion of the object */
2567 /* Read the HSM xattr and store its content into \p hsm_buf. */
2568 static int fetch_hsm_xattr(const struct lu_env *env, struct mdd_object *o,
2569 struct lu_buf *hsm_buf)
2573 lu_buf_alloc(hsm_buf, sizeof(struct hsm_attrs));
2574 if (hsm_buf->lb_buf == NULL)
2577 rc = mdo_xattr_get(env, o, hsm_buf, XATTR_NAME_HSM);
2584 /* return true if HSM xattr needs update */
2585 static bool swap_hsm_set_dirty(struct lu_buf *out_buf, struct lu_buf *src_buf)
2587 struct md_hsm src_hsm;
2589 lustre_buf2hsm(src_buf->lb_buf, src_buf->lb_len, &src_hsm);
2590 if (!(src_hsm.mh_flags & (HS_ARCHIVED | HS_EXISTS)) ||
2591 src_hsm.mh_flags & HS_DIRTY)
2592 /* do not update HSM attr of non archived or dirty files */
2595 src_hsm.mh_flags |= HS_DIRTY;
2596 lustre_hsm2buf(out_buf->lb_buf, &src_hsm);
2601 /* return 1 HSM xattr needs update
2602 * 0 do not update HSM xattr
2603 * -EPERM data version mismatch, no swap
2605 static int swap_hsm_update_version(struct lu_buf *out_buf, __u64 out_dv,
2606 struct lu_buf *src_buf, __u64 src_dv)
2608 struct md_hsm src_hsm;
2610 /* Put lb_len as a second argument since we know that src_buf is a valid
2613 lustre_buf2hsm(src_buf->lb_buf, src_buf->lb_len, &src_hsm);
2615 if (!(src_hsm.mh_flags & (HS_ARCHIVED | HS_EXISTS)) ||
2616 src_hsm.mh_flags & HS_DIRTY)
2619 /* migration with old client -> set the dirty flag */
2620 if (!src_dv || !out_dv) {
2621 src_hsm.mh_flags |= HS_DIRTY;
2625 if (src_hsm.mh_arch_ver != src_dv) {
2627 "HSM archive version and previous data version mismatch (arch_ver=%llu, prev_ver=%llu, new_ver=%llu)\n",
2628 src_hsm.mh_arch_ver, src_dv, out_dv);
2632 src_hsm.mh_arch_ver = out_dv;
2634 lustre_hsm2buf(out_buf->lb_buf, &src_hsm);
2639 /* Allow HSM xattr swap only with volatile/orphan files.
2640 * This is used by HSM release/restore.
2642 static inline bool swap_hsm_xattr_allowed(bool fst_has_hsm, bool snd_has_hsm,
2643 unsigned long o_fst_flag,
2644 unsigned long o_snd_flag)
2646 return (fst_has_hsm && snd_has_hsm &&
2647 ((o_fst_flag | o_snd_flag) & (ORPHAN_OBJ | VOLATILE_OBJ)));
2650 /* Read and update the data version of both objects if necessary. If they have
2651 * to be updated, declare the update.
2652 * \p xattr_flags will be updated if necessary to be used by mdo_xattr_set
2654 * \p dv contains the data version of the file before and after the migration.
2655 * It can be NULL when swapping layouts in other contexts (HSM or
2658 static int swap_layouts_prepare_hsm_attr(const struct lu_env *env,
2659 struct mdd_device *mdd,
2660 struct thandle *handle,
2661 struct sl_hsm_object_info *fst,
2662 struct sl_hsm_object_info *snd)
2664 unsigned long o_fst_fl;
2665 unsigned long o_snd_fl;
2669 fst->xattr_flags = 0;
2670 snd->xattr_flags = 0;
2672 rc = fetch_hsm_xattr(env, fst->o, fst->hsm_buf);
2673 if (rc != -ENODATA && rc != 0)
2676 rc2 = fetch_hsm_xattr(env, snd->o, snd->hsm_buf);
2677 if (rc2 != -ENODATA && rc2 != 0)
2680 /* if nothing to swap, not an error */
2684 /* swap if the first object have no HSM xattr */
2690 o_fst_fl = fst->o->mod_flags;
2691 o_snd_fl = snd->o->mod_flags;
2693 if ((o_snd_fl & VOLATILE_OBJ) && rc2) {
2694 /* migration of fst */
2695 rc = swap_hsm_update_version(snd->hsm_buf, snd->dv,
2696 fst->hsm_buf, fst->dv);
2702 fst->xattr_flags = LU_XATTR_REPLACE;
2704 } else if (swap_hsm_xattr_allowed(!rc, !rc2, o_fst_fl, o_snd_fl)) {
2705 /* HSM release/restore -> HSM xattr swap */
2706 fst->xattr_flags = LU_XATTR_REPLACE;
2707 snd->xattr_flags = LU_XATTR_REPLACE;
2709 } else if (!rc && !rc2) {
2710 /* swap on 2 archived files is not supported (no rollback) */
2714 /* swap layout with HSM fst and non-HSM snd */
2715 if (!swap_hsm_set_dirty(snd->hsm_buf, fst->hsm_buf))
2718 fst->xattr_flags = LU_XATTR_REPLACE;
2721 if (fst->xattr_flags) {
2722 rc = mdd_declare_xattr_set(env, mdd, fst->o, snd->hsm_buf,
2723 XATTR_NAME_HSM, fst->xattr_flags,
2729 if (snd->xattr_flags) {
2730 rc = mdd_declare_xattr_set(env, mdd, snd->o, fst->hsm_buf,
2731 XATTR_NAME_HSM, snd->xattr_flags,
2740 /* swap layouts between 2 lustre objects */
2741 static int mdd_swap_layouts(const struct lu_env *env,
2742 struct md_object *obj1, struct md_object *obj2,
2743 __u64 dv1, __u64 dv2, __u64 flags)
2745 struct mdd_thread_info *info = mdd_env_info(env);
2746 struct mdd_object *fst_o = md2mdd_obj(obj1);
2747 struct mdd_object *snd_o = md2mdd_obj(obj2);
2748 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2749 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2750 struct mdd_device *mdd = mdo2mdd(obj1);
2751 struct lov_mds_md *fst_lmm, *snd_lmm;
2752 struct sl_hsm_object_info fst_info;
2753 struct sl_hsm_object_info snd_info;
2754 struct mdd_object *vlt_o = NULL;
2755 struct mdd_object *dom_o = NULL;
2756 struct ost_id *saved_oi = NULL;
2757 struct lu_buf *fst_hsm_buf;
2758 struct lu_buf *snd_hsm_buf;
2759 struct lu_buf *fst_buf;
2760 struct lu_buf *snd_buf;
2761 struct thandle *handle;
2774 BUILD_BUG_ON(ARRAY_SIZE(info->mdi_buf) < 4);
2776 fst_buf = &info->mdi_buf[0];
2777 snd_buf = &info->mdi_buf[1];
2778 fst_hsm_buf = &info->mdi_buf[2];
2779 snd_hsm_buf = &info->mdi_buf[3];
2781 memset(info->mdi_buf, 0, sizeof(info->mdi_buf));
2783 /* we have to sort the 2 obj, so locking will always
2784 * be in the same order, even in case of 2 concurrent swaps
2786 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2787 if (rc == 0) /* same fid ? */
2795 rc = mdd_la_get(env, fst_o, fst_la);
2799 rc = mdd_la_get(env, snd_o, snd_la);
2803 /* check if layout swapping is allowed */
2804 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2808 handle = mdd_trans_create(env, mdd);
2810 RETURN(PTR_ERR(handle));
2812 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2813 if (rc < 0 && rc != -ENODATA)
2816 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2817 if (rc < 0 && rc != -ENODATA)
2820 /* check if file has DoM. DoM file can be migrated only to another
2821 * DoM layout with the same DoM component size or to an non-DOM
2822 * layout. After migration to OSTs layout, local MDT inode data
2823 * should be truncated.
2824 * Objects are sorted by FIDs, considering that original file's FID
2825 * is always smaller the snd_o is always original file we are migrating
2828 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2829 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2831 /* Only migration is supported for DoM files, not 'swap_layouts' so
2832 * target file must be volatile and orphan.
2834 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2835 vlt_o = domsize_vlt ? fst_o : NULL;
2836 dom_o = domsize_dom ? snd_o : NULL;
2837 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2838 swap(domsize_dom, domsize_vlt);
2839 vlt_o = domsize_vlt ? snd_o : NULL;
2840 dom_o = domsize_dom ? fst_o : NULL;
2841 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2842 /* 'lfs swap_layouts' case, neither file should have DoM */
2844 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, use migration instead: rc = %d\n",
2849 if (domsize_vlt > 0 && domsize_dom == 0) {
2852 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2853 mdd_obj_dev_name(snd_o),
2854 PFID(mdd_object_fid(snd_o)), rc);
2856 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2859 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2860 mdd_obj_dev_name(fst_o),
2861 PFID(mdd_object_fid(fst_o)), rc);
2865 /* swapping 2 non existant layouts is a success */
2866 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2869 /* to help inode migration between MDT, it is better to
2870 * start by the no layout file (if one), so we order the swap
2872 if (snd_buf->lb_buf == NULL) {
2874 swap(fst_buf, snd_buf);
2878 fst_gen = snd_gen = 0;
2879 /* lmm and generation layout initialization */
2880 if (fst_buf->lb_buf != NULL) {
2881 fst_lmm = fst_buf->lb_buf;
2882 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2883 fst_fl = LU_XATTR_REPLACE;
2887 fst_fl = LU_XATTR_CREATE;
2890 snd_lmm = snd_buf->lb_buf;
2891 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2893 saved_gen = fst_gen;
2894 /* increase the generation layout numbers */
2899 * XXX The layout generation is used to generate component IDs for
2900 * the composite file, we have to do some special tweaks to make
2901 * sure the layout generation is always adequate for that job.
2904 /* Skip invalid generation number for composite layout */
2905 if ((snd_gen & LCME_ID_MASK) == 0)
2907 if ((fst_gen & LCME_ID_MASK) == 0)
2909 /* Make sure the generation is greater than all the component IDs */
2910 if (fst_gen < snd_gen)
2912 else if (fst_gen > snd_gen)
2915 /* set the file specific informations in lmm */
2916 if (fst_lmm != NULL) {
2917 struct ost_id temp_oi;
2919 saved_oi = &info->mdi_oa.o_oi;
2920 mdd_get_lmm_oi(fst_lmm, saved_oi);
2921 mdd_get_lmm_oi(snd_lmm, &temp_oi);
2922 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2923 mdd_set_lmm_oi(fst_lmm, &temp_oi);
2924 mdd_set_lmm_oi(snd_lmm, saved_oi);
2926 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2927 cpu_to_le32(LOV_MAGIC_MAGIC))
2928 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2930 GOTO(stop, rc = -EPROTO);
2932 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2935 fst_info.hsm_buf = fst_hsm_buf;
2939 snd_info.hsm_buf = snd_hsm_buf;
2941 rc = swap_layouts_prepare_hsm_attr(env, mdd, handle, &fst_info,
2946 /* prepare transaction */
2947 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2952 if (fst_buf->lb_buf != NULL)
2953 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2954 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2957 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2962 rc = mdd_trans_start(env, mdd, handle);
2966 rc = mdd_write_lock_two_objects(env, fst_o, snd_o);
2970 if (!mdd_object_exists(fst_o))
2971 GOTO(unlock, rc = -ENOENT);
2972 if (!mdd_object_exists(snd_o))
2973 GOTO(unlock, rc = -ENOENT);
2975 if (fst_info.xattr_flags) {
2976 rc = mdo_xattr_set(env, fst_o, snd_info.hsm_buf,
2977 XATTR_NAME_HSM, fst_info.xattr_flags,
2983 if (snd_info.xattr_flags) {
2984 rc = mdo_xattr_set(env, snd_o, fst_info.hsm_buf,
2985 XATTR_NAME_HSM, snd_info.xattr_flags,
2988 GOTO(out_restore_hsm_fst, rc);
2991 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2993 GOTO(out_restore_hsm, rc);
2995 if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2998 if (fst_buf->lb_buf != NULL)
2999 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
3000 LU_XATTR_REPLACE, handle);
3002 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
3005 GOTO(out_restore, rc);
3007 /* Issue one changelog record per file */
3008 rc = emit_changelog_after_swap_layout(env, handle, fst_o, fst_hsm_buf);
3012 rc = emit_changelog_after_swap_layout(env, handle, snd_o, snd_hsm_buf);
3021 /* failure on second file, but first was done, so we have
3022 * to roll back first.
3024 if (fst_buf->lb_buf != NULL) {
3025 mdd_set_lmm_oi(fst_lmm, saved_oi);
3026 mdd_set_lmm_gen(fst_lmm, &saved_gen);
3027 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
3028 LU_XATTR_REPLACE, handle);
3030 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
3036 if (snd_info.xattr_flags) {
3037 /* roll back swap HSM */
3039 rc2 = mdo_xattr_set(env, snd_o, snd_hsm_buf,
3040 XATTR_NAME_HSM, LU_XATTR_REPLACE,
3046 out_restore_hsm_fst:
3047 if (fst_info.xattr_flags) {
3049 rc2 = mdo_xattr_set(env, fst_o, fst_hsm_buf,
3050 XATTR_NAME_HSM, LU_XATTR_REPLACE,
3057 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
3058 mdd_obj_dev_name(fst_o),
3059 PFID(mdd_object_fid(snd_o)),
3060 PFID(mdd_object_fid(fst_o)),
3062 /* a solution to avoid journal commit is to panic,
3063 * but it has strong consequences so we use LBUG to
3064 * allow sysdamin to choose to panic or not
3071 mdd_write_unlock_two_objects(env, fst_o, snd_o);
3074 rc = mdd_trans_stop(env, mdd, rc, handle);
3076 /* Truncate local DOM data if all went well, except migration case
3077 * with the same DOM component size. In that case a local data is
3078 * still in use and shouldn't be deleted.
3081 mdd_dom_fixup(env, mdd, dom_o, vlt_o);
3083 lu_buf_free(fst_buf);
3084 lu_buf_free(snd_buf);
3085 lu_buf_free(fst_hsm_buf);
3086 lu_buf_free(snd_hsm_buf);
3089 (void) mdd_object_pfid_replace(env, fst_o);
3090 (void) mdd_object_pfid_replace(env, snd_o);
3091 } else if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED) {
3093 * -EAGAIN means transaction execution phase detect the layout
3094 * has been changed by others.
3096 GOTO(retry, retried);
3102 static int mdd_declare_layout_change(const struct lu_env *env,
3103 struct mdd_device *mdd,
3104 struct mdd_object *obj,
3105 struct md_layout_change *mlc,
3106 struct thandle *handle)
3110 rc = mdo_declare_layout_change(env, obj, mlc, handle);
3114 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
3118 /* For PFL, this is used to instantiate necessary component objects. */
3120 mdd_layout_instantiate_component(const struct lu_env *env,
3121 struct mdd_object *obj, struct md_layout_change *mlc,
3122 struct thandle *handle)
3124 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3129 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
3132 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3134 * It's possible that another layout write intent has already
3135 * instantiated our objects, so a -EALREADY returned, and we need to
3139 RETURN(rc == -EALREADY ? 0 : rc);
3141 rc = mdd_trans_start(env, mdd, handle);
3145 mdd_write_lock(env, obj, DT_TGT_CHILD);
3146 rc = mdo_layout_change(env, obj, mlc, handle);
3147 mdd_write_unlock(env, obj);
3151 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
3157 * Change the FLR layout from RDONLY to WRITE_PENDING.
3159 * It picks the primary mirror, and bumps the layout version, and set
3160 * layout version xattr to OST objects in a sync tx. In order to facilitate
3161 * the handling of phantom writers from evicted clients, the clients carry
3162 * layout version of the file with write RPC, so that the OSTs can verify
3163 * if the write RPCs are legitimate, meaning not from evicted clients.
3166 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
3167 struct md_layout_change *mlc, struct thandle *handle)
3169 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3170 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3171 struct lustre_som_attrs *som = &mlc->mlc_som;
3177 /* Verify acceptable operations */
3178 switch (mlc->mlc_opc) {
3179 case MD_LAYOUT_WRITE:
3180 case MD_LAYOUT_RESYNC:
3181 /* these are legal operations - this represents the case that
3182 * a few mirrors were missed in the last resync.
3185 case MD_LAYOUT_RESYNC_DONE:
3190 som_buf->lb_buf = som;
3191 som_buf->lb_len = sizeof(*som);
3192 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
3193 if (rc < 0 && rc != -ENODATA)
3197 lustre_som_swab(som);
3198 if (som->lsa_valid & SOM_FL_STRICT)
3199 fl = LU_XATTR_REPLACE;
3201 if (mlc->mlc_opc == MD_LAYOUT_WRITE &&
3202 mlc->mlc_intent->lai_extent.e_end > som->lsa_size) {
3203 som->lsa_size = mlc->mlc_intent->lai_extent.e_end + 1;
3204 fl = LU_XATTR_REPLACE;
3208 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3213 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3214 XATTR_NAME_SOM, fl, handle);
3219 /* record a changelog for data mover to consume */
3220 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
3224 rc = mdd_trans_start(env, mdd, handle);
3228 /* it needs a sync tx to make FLR to work properly */
3229 handle->th_sync = 1;
3231 mdd_write_lock(env, obj, DT_TGT_CHILD);
3232 rc = mdo_layout_change(env, obj, mlc, handle);
3234 /* SOM state transition from STRICT to STALE */
3235 som->lsa_valid = SOM_FL_STALE;
3236 lustre_som_swab(som);
3237 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3240 mdd_write_unlock(env, obj);
3244 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle, NULL);
3251 * Handle mirrored file state transition when it's in WRITE_PENDING.
3253 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
3254 * the file is in WRITE_PENDING state. If everything goes fine, the file's
3255 * layout version will be increased, and the file's state will be changed to
3259 mdd_layout_update_write_pending(const struct lu_env *env,
3260 struct mdd_object *obj, struct md_layout_change *mlc,
3261 struct thandle *handle)
3263 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3264 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3265 struct lustre_som_attrs *som = &mlc->mlc_som;
3271 switch (mlc->mlc_opc) {
3272 case MD_LAYOUT_RESYNC:
3273 /* Upon receiving the resync request, it should
3274 * instantiate all stale components right away to get ready
3275 * for mirror copy. In order to avoid layout version change,
3276 * client should avoid sending LAYOUT_WRITE request at the
3280 case MD_LAYOUT_WRITE:
3282 * legal race for concurrent write, the file state has been
3283 * changed by another client. Or a jump over file size and
3286 som_buf->lb_buf = som;
3287 som_buf->lb_len = sizeof(*som);
3288 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
3289 if (rc < 0 && rc != -ENODATA)
3293 lustre_som_swab(som);
3294 if (mlc->mlc_intent->lai_extent.e_end > som->lsa_size) {
3296 mlc->mlc_intent->lai_extent.e_end + 1;
3297 fl = LU_XATTR_REPLACE;
3305 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3310 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3311 XATTR_NAME_SOM, fl, handle);
3316 rc = mdd_trans_start(env, mdd, handle);
3320 /* it needs a sync tx to make FLR to work properly */
3321 handle->th_sync = 1;
3323 mdd_write_lock(env, obj, DT_TGT_CHILD);
3324 rc = mdo_layout_change(env, obj, mlc, handle);
3326 som->lsa_valid = SOM_FL_STALE;
3327 lustre_som_swab(som);
3328 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3331 mdd_write_unlock(env, obj);
3338 * Handle the requests when a FLR file's state is in SYNC_PENDING.
3340 * Only concurrent write and sync complete requests are possible when the
3341 * file is in SYNC_PENDING. For the latter request, it will pass in the
3342 * mirrors that have been synchronized, then the stale bit will be cleared
3343 * to make the file's state turn into RDONLY.
3344 * For concurrent write reqeust, it just needs to change the file's state
3345 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
3346 * version because the version will be increased in the transition to
3347 * SYNC_PENDING later so that it can deny the write request from potential
3348 * evicted SYNC clients.
3351 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
3352 struct md_layout_change *mlc, struct thandle *handle)
3354 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3355 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3361 /* operation validation */
3362 switch (mlc->mlc_opc) {
3363 case MD_LAYOUT_RESYNC_DONE:
3364 /* resync complete. */
3365 case MD_LAYOUT_WRITE:
3366 /* concurrent write. */
3368 case MD_LAYOUT_RESYNC:
3369 /* resync again, most likely the previous run failed.
3370 * no-op if it's already in SYNC_PENDING state
3377 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
3378 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
3379 if (rc < 0 && rc != -ENODATA)
3382 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
3383 lustre_som_swab(&mlc->mlc_som);
3384 som_buf->lb_buf = &mlc->mlc_som;
3385 som_buf->lb_len = sizeof(mlc->mlc_som);
3388 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3392 /* record a changelog for the completion of resync */
3393 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
3398 /* RESYNC_DONE has piggybacked size and blocks */
3400 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3401 XATTR_NAME_SOM, fl, handle);
3406 rc = mdd_trans_start(env, mdd, handle);
3410 /* it needs a sync tx to make FLR to work properly */
3411 handle->th_sync = 1;
3413 mdd_write_lock(env, obj, DT_TGT_CHILD);
3414 rc = mdo_layout_change(env, obj, mlc, handle);
3415 mdd_write_unlock(env, obj);
3420 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3426 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle,
3432 /* Update the layout for PCC-RO. */
3434 mdd_layout_update_pccro(const struct lu_env *env, struct md_object *o,
3435 struct md_layout_change *mlc)
3437 struct mdd_object *obj = md2mdd_obj(o);
3438 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3439 struct thandle *handle;
3444 handle = mdd_trans_create(env, mdd);
3446 RETURN(PTR_ERR(handle));
3448 /* TODO: Set SOM strict correct when the file is PCC-RO cached. */
3449 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3451 * It is possible that another layout write intent has already
3452 * set/cleared read-only flag on the object, so as to return
3453 * -EALREADY, and we need to do nothing in this case.
3456 GOTO(out, rc == -EALREADY ? rc = 0 : rc);
3458 rc = mdd_trans_start(env, mdd, handle);
3462 mdd_write_lock(env, obj, DT_TGT_CHILD);
3463 rc = mdo_layout_change(env, obj, mlc, handle);
3464 mdd_write_unlock(env, obj);
3468 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
3471 mdd_trans_stop(env, mdd, rc, handle);
3477 * Layout change callback for object.
3479 * This is only used by FLR and PCC-RO for now. In the future, it can be
3480 * exteneded to handle all layout change.
3483 mdd_layout_change(const struct lu_env *env, struct md_object *o,
3484 struct md_layout_change *mlc)
3486 struct mdd_object *obj = md2mdd_obj(o);
3487 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3488 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
3489 struct lov_comp_md_v1 *lcm;
3490 struct thandle *handle;
3497 if (S_ISDIR(mdd_object_type(obj))) {
3498 switch (mlc->mlc_opc) {
3499 case MD_LAYOUT_SHRINK:
3500 rc = mdd_dir_layout_shrink(env, o, mlc);
3502 case MD_LAYOUT_SPLIT:
3503 rc = mdd_dir_layout_split(env, o, mlc);
3512 /* Verify acceptable operations */
3513 switch (mlc->mlc_opc) {
3514 case MD_LAYOUT_WRITE: {
3515 struct layout_intent *intent = mlc->mlc_intent;
3517 if (intent->lai_opc == LAYOUT_INTENT_PCCRO_SET ||
3518 intent->lai_opc == LAYOUT_INTENT_PCCRO_CLEAR)
3519 RETURN(mdd_layout_update_pccro(env, o, mlc));
3521 case MD_LAYOUT_RESYNC:
3522 case MD_LAYOUT_RESYNC_DONE:
3529 handle = mdd_trans_create(env, mdd);
3531 RETURN(PTR_ERR(handle));
3533 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
3540 /* analyze the layout to make sure it's a FLR file */
3542 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
3543 GOTO(out, rc = -EINVAL);
3545 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
3547 /* please refer to HLD of FLR for state transition */
3548 switch (flr_state) {
3550 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
3552 case LCM_FL_WRITE_PENDING:
3553 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
3556 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
3558 case LCM_FL_SYNC_PENDING:
3559 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3568 mdd_trans_stop(env, mdd, rc, handle);
3572 * -EAGAIN means transaction execution phase detect the layout
3573 * has been changed by others.
3575 if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
3576 GOTO(retry, retried);
3581 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3582 struct mdd_object *child, const struct lu_attr *attr,
3583 const struct md_op_spec *spec,
3584 struct dt_allocation_hint *hint)
3586 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3587 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3588 struct dt_object *nc = mdd_object_child(child);
3590 memset(hint, 0, sizeof(*hint));
3592 /* For striped directory, give striping EA to lod_ah_init, which will
3593 * decide the stripe_offset and stripe count by it.
3595 if (S_ISDIR(attr->la_mode) && spec) {
3596 if (unlikely(spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3597 hint->dah_eadata = spec->u.sp_ea.eadata;
3598 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3599 if (spec->sp_cr_flags & MDS_OPEN_DEFAULT_LMV)
3600 hint->dah_eadata_is_dmv = 1;
3602 hint->dah_dmv_imp_inherit = spec->sp_dmv_imp_inherit;
3603 } else if (S_ISREG(attr->la_mode) &&
3604 spec->sp_cr_flags & MDS_OPEN_APPEND) {
3605 hint->dah_append_stripe_count = mdd->mdd_append_stripe_count;
3606 hint->dah_append_pool = mdd->mdd_append_pool;
3609 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3610 hint->dah_eadata, hint->dah_eadata_len);
3611 /* @hint will be initialized by underlying device. */
3612 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3615 static int mdd_accmode(const struct lu_env *env, const struct lu_attr *la,
3618 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3619 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3620 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3621 * owner can write to a file even if it is marked readonly to hide
3622 * its brokenness. (bug 5781)
3624 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3625 struct lu_ucred *uc = lu_ucred_check(env);
3627 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3631 return mds_accmode(open_flags);
3634 static int mdd_open_sanity_check(const struct lu_env *env,
3635 struct mdd_object *obj,
3636 const struct lu_attr *attr, u64 open_flags,
3639 unsigned int may_mask;
3644 /* EEXIST check, also opening of *open* orphans is allowed so we can
3645 * open-by-handle unlinked files
3647 if (mdd_is_dead_obj(obj) && !is_replay &&
3648 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3651 if (S_ISLNK(attr->la_mode))
3654 may_mask = mdd_accmode(env, attr, open_flags);
3656 if (S_ISDIR(attr->la_mode) && (may_mask & MAY_WRITE))
3659 if (!(open_flags & MDS_OPEN_CREATED)) {
3660 rc = mdd_permission_internal(env, obj, attr, may_mask);
3665 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3666 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3667 open_flags &= ~MDS_OPEN_TRUNC;
3669 /* For writing append-only file must open it with append mode. */
3670 if (attr->la_flags & LUSTRE_APPEND_FL) {
3671 if ((open_flags & MDS_FMODE_WRITE) &&
3672 !(open_flags & MDS_OPEN_APPEND))
3674 if (open_flags & MDS_OPEN_TRUNC)
3681 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3682 u64 open_flags, struct md_op_spec *spec)
3684 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3685 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3686 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3687 struct mdd_object_user *mou = NULL;
3688 const struct lu_ucred *uc = lu_ucred(env);
3689 struct mdd_device *mdd = mdo2mdd(obj);
3690 enum changelog_rec_type type = CL_OPEN;
3695 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3697 rc = mdd_la_get(env, mdd_obj, attr);
3701 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3703 if ((rc == -EACCES) && (mdd->mdd_cl.mc_current_mask & BIT(CL_DN_OPEN)))
3708 mdd_obj->mod_count++;
3710 if (!mdd_changelog_enabled(env, mdd, type))
3714 /* look for existing opener in list under mdd_write_lock */
3715 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3720 /* add user to list */
3721 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3727 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3729 mdd_obj_user_free(mou);
3734 if (type == CL_DN_OPEN) {
3735 if (ktime_before(ktime_get(), mou->mou_deniednext))
3736 /* same user denied again same access within
3737 * time interval: do not record
3741 /* this user already denied, but some time ago:
3742 * update denied time
3744 mou->mou_deniednext =
3745 ktime_add(ktime_get(),
3746 ktime_set(mdd->mdd_cl.mc_deniednext,
3749 mou->mou_opencount++;
3750 /* same user opening file again with same flags:
3757 /* we can't hold object lock over transaction start
3758 * and we don't actually need the object to be locked
3760 mdd_write_unlock(env, mdd_obj);
3762 /* FYI, only the bottom 32 bits of open_flags are recorded */
3763 /* only the low CLF_FLAGMASK bits of la_valid are stored*/
3764 mdd_changelog(env, type, open_flags & CLF_FLAGMASK, md_dev,
3765 mdd_object_fid(mdd_obj));
3770 mdd_write_unlock(env, mdd_obj);
3775 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3776 struct md_attr *ma, struct thandle *handle)
3780 rc = mdd_orphan_declare_delete(env, obj, handle);
3784 return mdo_declare_destroy(env, obj, handle);
3787 /* No permission check is needed. */
3788 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3789 struct md_attr *ma, u64 open_flags)
3791 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3792 struct mdd_device *mdd = mdo2mdd(obj);
3793 struct thandle *handle = NULL;
3796 bool blocked = false;
3797 bool last_close_by_uid = false;
3798 const struct lu_ucred *uc = lu_ucred(env);
3802 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3803 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3804 mdd_obj->mod_count--;
3805 mdd_write_unlock(env, mdd_obj);
3807 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3808 CDEBUG(D_HA, "Object "DFID" is retained in orphan list\n",
3809 PFID(mdd_object_fid(mdd_obj)));
3813 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3814 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ).
3816 /* check without any lock */
3817 is_orphan = mdd_obj->mod_count == 1 &&
3818 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3822 /* mdd_trans_create() maybe failed because of barrier_entry(),
3823 * under such case, the orphan MDT-object will be left in the
3824 * orphan list, and when the MDT remount next time, the unused
3825 * orphans will be destroyed automatically.
3827 * One exception: the former mdd_finish_unlink may failed to
3828 * add the orphan MDT-object to the orphan list, then if the
3829 * mdd_trans_create() failed because of barrier_entry(), the
3830 * MDT-object will become real orphan that is neither in the
3831 * namespace nor in the orphan list. Such bad case should be
3832 * very rare and will be handled by e2fsck/lfsck.
3834 handle = mdd_trans_create(env, mdo2mdd(obj));
3835 if (IS_ERR(handle)) {
3836 rc = PTR_ERR(handle);
3837 if (rc != -EINPROGRESS)
3845 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3849 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3854 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3860 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3861 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3863 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3864 lu_dev_name(mdd2lu_dev(mdd)),
3865 PFID(mdd_object_fid(mdd_obj)), rc);
3869 /* check again with lock */
3870 is_orphan = (mdd_obj->mod_count == 1) &&
3871 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3872 ma->ma_attr.la_nlink == 0);
3874 if (is_orphan && !handle && !blocked) {
3875 mdd_write_unlock(env, mdd_obj);
3879 mdd_obj->mod_count--; /*release open count */
3881 /* under mdd write lock */
3882 /* If recording, see if we need to remove UID from list. uc is not
3883 * initialized if the client has been evicted.
3885 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3886 struct mdd_object_user *mou;
3888 /* look for UID in list */
3889 /* If mou is NULL, it probably means logging was enabled after
3890 * the user had the file open. So the corresponding close
3891 * will not be logged.
3893 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3896 mou->mou_opencount--;
3897 if (mou->mou_opencount == 0) {
3898 mdd_obj_user_remove(mdd_obj, mou);
3899 last_close_by_uid = true;
3904 if (!is_orphan || blocked)
3908 /* NB: Object maybe not in orphan list originally, it is rare case for
3909 * mdd_finish_unlink() failure, in that case, the object doesn't have
3912 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3913 /* remove link to object from orphan index */
3914 LASSERT(handle != NULL);
3915 rc = mdd_orphan_delete(env, mdd_obj, handle);
3917 CERROR("%s: unable to delete "DFID" from orphan list: rc = %d\n",
3918 lu_dev_name(mdd2lu_dev(mdd)),
3919 PFID(mdd_object_fid(mdd_obj)), rc);
3920 /* If object was not deleted from orphan list, do not
3921 * destroy OSS objects, which will be done when next
3927 CDEBUG(D_HA, "Object "DFID" is deleted from orphan list, OSS objects to be destroyed.\n",
3928 PFID(mdd_object_fid(mdd_obj)));
3931 rc = mdo_destroy(env, mdd_obj, handle);
3934 CERROR("%s: unable to delete "DFID" from orphan list: rc = %d\n",
3935 lu_dev_name(mdd2lu_dev(mdd)),
3936 PFID(mdd_object_fid(mdd_obj)), rc);
3941 mdd_write_unlock(env, mdd_obj);
3943 if (rc != 0 || blocked ||
3944 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3947 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3948 * or if CL_OPEN was recorded and it's last close by user.
3949 * Changelogs mask may change between open and close operations, but
3950 * this is not a big deal if we have a CL_CLOSE entry with no matching
3951 * CL_OPEN. Plus Changelogs mask may not change often.
3953 if (((!(mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3954 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3955 MDS_OPEN_TRUNC))) ||
3956 ((mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3957 last_close_by_uid)) &&
3958 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3959 if (handle == NULL) {
3960 handle = mdd_trans_create(env, mdo2mdd(obj));
3962 GOTO(stop, rc = PTR_ERR(handle));
3964 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3965 NULL, NULL, handle);
3969 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3974 /* FYI, only the bottom 32 bits of open_flags are recorded */
3975 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3976 mdd_obj, handle, NULL);
3980 if (handle != NULL && !IS_ERR(handle))
3981 rc = mdd_trans_stop(env, mdd, rc, handle);
3986 /* Permission check is done when open, no need check again. */
3987 static int mdd_readpage_sanity_check(const struct lu_env *env,
3988 struct mdd_object *obj)
3990 if (!dt_try_as_dir(env, mdd_object_child(obj), true))
3996 static int mdd_dir_page_build(const struct lu_env *env, struct dt_object *obj,
3997 union lu_page *lp, size_t bytes,
3998 const struct dt_it_ops *iops,
3999 struct dt_it *it, __u32 attr, void *arg)
4001 struct lu_dirpage *dp = &lp->lp_dir;
4004 struct lu_dirent *ent;
4005 struct lu_dirent *last = NULL;
4010 if (bytes < sizeof(*dp))
4011 GOTO(out_err, result = -EOVERFLOW);
4013 memset(area, 0, sizeof(*dp));
4014 area += sizeof(*dp);
4015 bytes -= sizeof(*dp);
4022 len = iops->key_size(env, it);
4024 /* IAM iterator can return record with zero len. */
4028 hash = iops->store(env, it);
4029 if (unlikely(first)) {
4031 dp->ldp_hash_start = cpu_to_le64(hash);
4034 /* calculate max space required for lu_dirent */
4035 recsize = lu_dirent_calc_size(len, attr);
4037 if (bytes >= recsize &&
4038 !CFS_FAIL_CHECK(OBD_FAIL_MDS_DIR_PAGE_WALK)) {
4039 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
4040 if (result == -ESTALE)
4045 /* OSD might not able to pack all attributes, so
4046 * recheck record length had room to store FID
4048 recsize = le16_to_cpu(ent->lde_reclen);
4050 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
4051 fid_le_to_cpu(&fid, &ent->lde_fid);
4052 if (fid_is_dot_lustre(&fid))
4053 GOTO(next, recsize);
4056 result = (last != NULL) ? 0 : -EBADSLT;
4060 ent = (void *)ent + recsize;
4064 result = iops->next(env, it);
4065 if (result == -ESTALE)
4067 } while (result == 0);
4070 dp->ldp_hash_end = cpu_to_le64(hash);
4072 if (last->lde_hash == dp->ldp_hash_end)
4073 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
4074 last->lde_reclen = 0; /* end mark */
4078 /* end of directory */
4079 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
4080 else if (result < 0)
4081 CWARN("%s: build page failed for "DFID": rc = %d\n",
4082 lu_dev_name(obj->do_lu.lo_dev),
4083 PFID(lu_object_fid(&obj->do_lu)), result);
4088 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
4089 const struct lu_rdpg *rdpg)
4091 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4096 if (mdd_object_exists(mdd_obj) == 0) {
4098 CERROR("%s: object "DFID" not found: rc = %d\n",
4099 mdd_obj_dev_name(mdd_obj),
4100 PFID(mdd_object_fid(mdd_obj)), rc);
4104 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
4105 rc = mdd_readpage_sanity_check(env, mdd_obj);
4107 GOTO(out_unlock, rc);
4109 if (mdd_is_dead_obj(mdd_obj)) {
4111 struct lu_dirpage *dp;
4114 * According to POSIX, please do not return any entry to client:
4115 * even dot and dotdot should not be returned.
4117 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
4118 PFID(mdd_object_fid(mdd_obj)));
4120 if (rdpg->rp_count <= 0)
4121 GOTO(out_unlock, rc = -EFAULT);
4122 LASSERT(rdpg->rp_pages != NULL);
4124 pg = rdpg->rp_pages[0];
4125 dp = (struct lu_dirpage *)kmap(pg);
4126 memset(dp, 0, sizeof(struct lu_dirpage));
4127 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
4128 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
4129 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
4131 GOTO(out_unlock, rc = LU_PAGE_SIZE);
4134 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
4135 mdd_dir_page_build, NULL);
4137 struct lu_dirpage *dp;
4139 dp = kmap(rdpg->rp_pages[0]);
4140 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
4143 * No pages were processed, mark this for first page
4146 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
4147 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
4148 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
4150 kunmap(rdpg->rp_pages[0]);
4153 GOTO(out_unlock, rc);
4155 mdd_read_unlock(env, mdd_obj);
4159 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
4161 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4163 if (mdd_object_exists(mdd_obj) == 0) {
4166 CERROR("%s: object "DFID" not found: rc = %d\n",
4167 mdd_obj_dev_name(mdd_obj),
4168 PFID(mdd_object_fid(mdd_obj)), rc);
4171 return dt_object_sync(env, mdd_object_child(mdd_obj),
4175 static int mdd_object_lock(const struct lu_env *env,
4176 struct md_object *obj,
4177 struct lustre_handle *lh,
4178 struct ldlm_enqueue_info *einfo,
4179 union ldlm_policy_data *policy)
4181 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4183 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
4187 static int mdd_object_unlock(const struct lu_env *env,
4188 struct md_object *obj,
4189 struct ldlm_enqueue_info *einfo,
4190 union ldlm_policy_data *policy)
4192 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4194 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
4197 const struct md_object_operations mdd_obj_ops = {
4198 .moo_permission = mdd_permission,
4199 .moo_attr_get = mdd_attr_get,
4200 .moo_attr_set = mdd_attr_set,
4201 .moo_xattr_get = mdd_xattr_get,
4202 .moo_xattr_set = mdd_xattr_set,
4203 .moo_xattr_list = mdd_xattr_list,
4204 .moo_invalidate = mdd_invalidate,
4205 .moo_xattr_del = mdd_xattr_del,
4206 .moo_swap_layouts = mdd_swap_layouts,
4207 .moo_open = mdd_open,
4208 .moo_close = mdd_close,
4209 .moo_readpage = mdd_readpage,
4210 .moo_readlink = mdd_readlink,
4211 .moo_changelog = mdd_changelog,
4212 .moo_object_sync = mdd_object_sync,
4213 .moo_object_lock = mdd_object_lock,
4214 .moo_object_unlock = mdd_object_unlock,
4215 .moo_layout_change = mdd_layout_change,