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 /* This is an atime/mtime/ctime update for close RPCs. */
726 if (!(la->la_valid & LA_CTIME) ||
727 (la->la_ctime <= oattr->la_ctime))
728 la->la_valid &= ~(LA_MTIME | LA_CTIME);
730 if (la->la_valid & LA_ATIME &&
731 (la->la_atime <= obj->mod_atime_set ||
732 la->la_atime <= (oattr->la_atime +
733 mdd_obj2mdd_dev(obj)->mdd_atime_diff)))
734 la->la_valid &= ~LA_ATIME;
736 obj->mod_atime_set = la->la_atime;
737 if (la->la_valid & LA_MTIME && la->la_mtime <= oattr->la_mtime)
738 la->la_valid &= ~LA_MTIME;
740 } else if ((la->la_valid & LA_ATIME) && (la->la_valid & LA_CTIME)) {
741 /* save the time when atime was changed, in case this is
742 * set-in-past, to not lose it later on close. */
743 obj->mod_atime_set = la->la_ctime;
746 /* Check if flags change. */
747 if (la->la_valid & LA_FLAGS) {
748 unsigned int oldflags = oattr->la_flags &
749 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
750 unsigned int newflags = la->la_flags &
751 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
753 if ((uc->uc_fsuid != oattr->la_uid) &&
754 !cap_raised(uc->uc_cap, CAP_FOWNER))
757 /* The IMMUTABLE and APPEND_ONLY flags can
758 * only be changed by the relevant capability.
760 if ((oldflags ^ newflags) &&
761 !cap_raised(uc->uc_cap, CAP_LINUX_IMMUTABLE))
764 if (!S_ISDIR(oattr->la_mode)) {
765 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
766 } else if (la->la_flags & LUSTRE_ENCRYPT_FL) {
767 /* when trying to add encryption flag on dir,
768 * make sure it is empty
770 rc = mdd_dir_is_empty(env, obj);
777 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
778 (la->la_valid & ~LA_FLAGS) &&
779 !(flags & MDS_PERM_BYPASS))
782 /* Check for setting the obj time. */
783 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
784 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
785 if ((uc->uc_fsuid != oattr->la_uid) &&
786 !cap_raised(uc->uc_cap, CAP_FOWNER)) {
787 rc = mdd_permission_internal(env, obj, oattr,
794 if (la->la_valid & LA_KILL_SUID) {
795 la->la_valid &= ~LA_KILL_SUID;
796 if ((oattr->la_mode & S_ISUID) &&
797 !(la->la_valid & LA_MODE)) {
798 la->la_mode = oattr->la_mode;
799 la->la_valid |= LA_MODE;
801 la->la_mode &= ~S_ISUID;
804 if (la->la_valid & LA_KILL_SGID) {
805 la->la_valid &= ~LA_KILL_SGID;
806 if (((oattr->la_mode & (S_ISGID | 0010)) ==
808 !(la->la_valid & LA_MODE)) {
809 la->la_mode = oattr->la_mode;
810 la->la_valid |= LA_MODE;
812 la->la_mode &= ~S_ISGID;
815 /* Make sure a caller can chmod. */
816 if (la->la_valid & LA_MODE) {
817 /* we want rbac roles to have precedence over any other
818 * permission or capability checks
820 if (!uc->uc_rbac_file_perms ||
821 (!(flags & MDS_PERM_BYPASS) &&
822 (uc->uc_fsuid != oattr->la_uid) &&
823 !cap_raised(uc->uc_cap, CAP_FOWNER)))
826 if (la->la_mode == (umode_t) -1)
827 la->la_mode = oattr->la_mode;
829 la->la_mode = (la->la_mode & S_IALLUGO) |
830 (oattr->la_mode & ~S_IALLUGO);
832 /* Also check the setgid bit! */
833 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
834 la->la_gid : oattr->la_gid) &&
835 !cap_raised(uc->uc_cap, CAP_FSETID))
836 la->la_mode &= ~S_ISGID;
838 la->la_mode = oattr->la_mode;
841 /* Make sure a caller can chown. */
842 if (la->la_valid & LA_UID) {
843 if (la->la_uid == (uid_t) -1)
844 la->la_uid = oattr->la_uid;
845 /* we want rbac roles to have precedence over any other
846 * permission or capability checks
848 if (!uc->uc_rbac_file_perms ||
849 ((uc->uc_fsuid != oattr->la_uid ||
850 la->la_uid != oattr->la_uid) &&
851 !cap_raised(uc->uc_cap, CAP_CHOWN)))
854 /* If the user or group of a non-directory has been
855 * changed by a non-root user, remove the setuid bit.
856 * 19981026 David C Niemi <niemi@tux.org>
858 * Changed this to apply to all users, including root,
859 * to avoid some races. This is the behavior we had in
860 * 2.0. The check for non-root was definitely wrong
861 * for 2.2 anyway, as it should have been using
862 * CAP_FSETID rather than fsuid -- 19990830 SD.
864 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
865 !S_ISDIR(oattr->la_mode)) {
866 la->la_mode &= ~S_ISUID;
867 la->la_valid |= LA_MODE;
871 /* Make sure caller can chgrp. */
872 if (la->la_valid & LA_GID) {
873 if (la->la_gid == (gid_t) -1)
874 la->la_gid = oattr->la_gid;
875 /* we want rbac roles to have precedence over any other
876 * permission or capability checks
878 if (!uc->uc_rbac_file_perms ||
879 ((uc->uc_fsuid != oattr->la_uid ||
880 (la->la_gid != oattr->la_gid &&
881 !lustre_in_group_p(uc, la->la_gid))) &&
882 !cap_raised(uc->uc_cap, CAP_CHOWN)))
885 /* Likewise, if the user or group of a non-directory
886 * has been changed by a non-root user, remove the
887 * setgid bit UNLESS there is no group execute bit
888 * (this would be a file marked for mandatory
889 * locking). 19981026 David C Niemi <niemi@tux.org>
891 * Removed the fsuid check (see the comment above) --
894 if (((oattr->la_mode & (S_ISGID | 0010)) ==
895 (S_ISGID | 0010)) && !S_ISDIR(oattr->la_mode)) {
896 la->la_mode &= ~S_ISGID;
897 la->la_valid |= LA_MODE;
901 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
902 if (!((flags & MDS_OWNEROVERRIDE) &&
903 (uc->uc_fsuid == oattr->la_uid)) &&
904 !(flags & MDS_PERM_BYPASS)) {
905 rc = mdd_permission_internal(env, obj, oattr,
912 if (la->la_valid & LA_CTIME) {
914 * The pure setattr, it has the priority over what is
915 * already set, do not drop it if ctime is equal.
917 if (la->la_ctime < oattr->la_ctime)
918 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
924 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
925 struct mdd_device *mdd,
926 enum changelog_rec_type type,
927 enum changelog_rec_flags clf_flags,
928 const struct lu_fid *fid,
929 const struct lu_fid *pfid,
930 const char *xattr_name,
931 struct thandle *handle)
933 const struct lu_ucred *uc = lu_ucred(env);
934 enum changelog_rec_extra_flags xflags = CLFE_INVALID;
935 struct llog_changelog_rec *rec;
940 clf_flags = (clf_flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
943 if (uc->uc_jobid[0] != '\0')
944 clf_flags |= CLF_JOBID;
945 xflags |= CLFE_UIDGID;
948 if (type == CL_OPEN || type == CL_DN_OPEN)
950 if (type == CL_SETXATTR || type == CL_GETXATTR)
951 xflags |= CLFE_XATTR;
953 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
954 changelog_rec_offset(clf_flags & CLF_SUPPORTED,
955 xflags & CLFE_SUPPORTED));
956 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_chlg_buf, reclen);
957 if (buf->lb_buf == NULL)
961 rec->cr_hdr.lrh_len = reclen;
962 rec->cr.cr_flags = clf_flags;
963 rec->cr.cr_type = (__u32)type;
964 rec->cr.cr_tfid = *fid;
966 rec->cr.cr_pfid = *pfid;
967 rec->cr.cr_namelen = 0;
969 if (clf_flags & CLF_JOBID)
970 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
972 if (clf_flags & CLF_EXTRA_FLAGS) {
973 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
974 if (xflags & CLFE_UIDGID)
975 mdd_changelog_rec_extra_uidgid(&rec->cr,
976 uc->uc_uid, uc->uc_gid);
977 if (xflags & CLFE_NID)
978 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
979 if (xflags & CLFE_OPEN)
980 mdd_changelog_rec_extra_omode(&rec->cr, clf_flags);
981 if (xflags & CLFE_XATTR) {
982 if (xattr_name == NULL)
984 mdd_changelog_rec_extra_xattr(&rec->cr, xattr_name);
988 rc = mdd_changelog_store(env, mdd, rec, handle);
993 /** Store a data change changelog record
994 * If this fails, we must fail the whole transaction; we don't
995 * want the change to commit without the log entry.
996 * \param mdd_obj - mdd_object of change
997 * \param handle - transaction handle
998 * \param pfid - parent FID for CL_MTIME changelogs
1000 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
1001 enum changelog_rec_type type,
1002 enum changelog_rec_flags clf_flags,
1003 struct mdd_object *mdd_obj, struct thandle *handle,
1004 const struct lu_fid *pfid)
1008 LASSERT(mdd_obj != NULL);
1009 LASSERT(handle != NULL);
1011 if (!mdd_changelog_enabled(env, mdd, type))
1014 if (mdd_is_volatile_obj(mdd_obj))
1017 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1018 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1019 /* Don't need multiple updates in this log */
1020 /* Don't check under lock - no big deal if we get extra entry */
1024 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1025 mdd_object_fid(mdd_obj), pfid,
1028 mdd_obj->mod_cltime = ktime_get();
1033 int mdd_changelog_data_store_xattr(const struct lu_env *env,
1034 struct mdd_device *mdd,
1035 enum changelog_rec_type type,
1036 enum changelog_rec_flags clf_flags,
1037 struct mdd_object *mdd_obj,
1038 const char *xattr_name,
1039 struct thandle *handle)
1043 LASSERT(mdd_obj != NULL);
1044 LASSERT(handle != NULL);
1046 if (!mdd_changelog_enabled(env, mdd, type))
1049 if (mdd_is_volatile_obj(mdd_obj))
1052 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
1053 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
1054 /* Don't need multiple updates in this log */
1055 /* Don't check under lock - no big deal if we get an extra
1061 rc = mdd_changelog_data_store_by_fid(env, mdd, type, clf_flags,
1062 mdd_object_fid(mdd_obj), NULL,
1063 xattr_name, handle);
1065 mdd_obj->mod_cltime = ktime_get();
1070 /* Only the bottom CLF_FLAGMASK bits of @flags are stored in the record,
1071 * except for open flags have a dedicated record to store 32 bits of flags.
1073 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
1074 enum changelog_rec_flags clf_flags,
1075 struct md_device *m, const struct lu_fid *fid)
1077 struct thandle *handle;
1078 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
1083 LASSERT(fid != NULL);
1085 /* We'll check this again below, but we check now before we
1086 * start a transaction.
1088 if (!mdd_changelog_enabled(env, mdd, type))
1091 handle = mdd_trans_create(env, mdd);
1093 RETURN(PTR_ERR(handle));
1095 rc = mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1099 rc = mdd_trans_start(env, mdd, handle);
1103 /* note we only store the low CLF_FLAGMASK bits of la_valid */
1104 rc = mdd_changelog_data_store_by_fid(env, mdd, type,
1105 clf_flags & CLF_FLAGMASK,
1106 fid, NULL, NULL, handle);
1109 rc = mdd_trans_stop(env, mdd, rc, handle);
1115 * Save LMA extended attributes with data from \a ma.
1117 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
1118 * not, LMA EA will be first read from disk, modified and write back.
1121 /* Precedence for choosing record type when multiple
1122 * attributes change: setattr > mtime > ctime > atime
1123 * (ctime changes when mtime does, plus chmod/chown.
1124 * atime and ctime are independent.)
1126 static int mdd_attr_set_changelog(const struct lu_env *env,
1127 struct md_object *obj, struct thandle *handle,
1128 const struct lu_fid *pfid, __u64 valid)
1130 struct mdd_device *mdd = mdo2mdd(obj);
1133 bits = (valid & LA_SIZE) ? BIT(CL_TRUNC) : 0;
1134 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? BIT(CL_SETATTR) : 0;
1135 bits |= (valid & LA_MTIME) ? BIT(CL_MTIME) : 0;
1136 bits |= (valid & LA_CTIME) ? BIT(CL_CTIME) : 0;
1137 bits |= (valid & LA_ATIME) ? BIT(CL_ATIME) : 0;
1138 bits = bits & mdd->mdd_cl.mc_current_mask;
1139 /* This is an implementation limit rather than a protocol limit */
1140 BUILD_BUG_ON(CL_LAST > sizeof(int) * 8);
1144 /* The record type is the lowest non-masked set bit */
1147 /* only the low CLF_FLAGMASK bits of la_valid are stored */
1148 return mdd_changelog_data_store(env, mdd, type, valid & CLF_FLAGMASK,
1149 md2mdd_obj(obj), handle, pfid);
1152 static int mdd_declare_attr_set(const struct lu_env *env,
1153 struct mdd_device *mdd,
1154 struct mdd_object *obj,
1155 const struct lu_attr *attr,
1156 struct thandle *handle)
1160 rc = mdo_declare_attr_set(env, obj, attr, handle);
1164 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
1165 if (attr->la_valid & LA_MODE) {
1166 mdd_read_lock(env, obj, DT_TGT_CHILD);
1167 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
1168 XATTR_NAME_ACL_ACCESS);
1169 mdd_read_unlock(env, obj);
1170 if (rc == -EOPNOTSUPP || rc == -ENODATA)
1176 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
1178 rc = mdo_declare_xattr_set(env, obj, buf,
1179 XATTR_NAME_ACL_ACCESS, 0,
1187 rc = mdd_declare_changelog_store(env, mdd, CL_SETXATTR, NULL, NULL,
1196 * permission changes may require sync operation, to mitigate performance
1197 * impact, only do this for dir and when permission is reduced.
1199 * For regular files, version is updated with permission change (see VBR), async
1200 * permission won't cause any issue, while missing permission change on
1201 * directory may affect accessibility of other objects after recovery.
1203 static inline bool permission_needs_sync(const struct lu_attr *old,
1204 const struct lu_attr *new)
1206 if (!S_ISDIR(old->la_mode))
1209 if (new->la_valid & LA_UID && old->la_uid != new->la_uid)
1212 if (new->la_valid & LA_GID && old->la_gid != new->la_gid)
1215 if (new->la_valid & LA_MODE) {
1216 /* turned on sticky bit */
1217 if (!(old->la_mode & S_ISVTX) && (new->la_mode & S_ISVTX))
1220 /* set-GID has no impact on what is allowed, not checked */
1222 /* turned off setuid bit, or one of rwx for someone */
1223 if (((new->la_mode & old->la_mode) & (0777 | S_ISUID)) !=
1224 (old->la_mode & (0777 | S_ISUID)))
1231 static inline __u64 mdd_lmm_dom_size(void *buf)
1233 struct lov_mds_md *lmm = buf;
1234 struct lov_comp_md_v1 *comp_v1;
1235 struct lov_mds_md *v1;
1241 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1244 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1245 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1246 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1248 /* DoM entry is the first entry always */
1249 if (lov_pattern(le32_to_cpu(v1->lmm_pattern)) & LOV_PATTERN_MDT)
1250 return le64_to_cpu(comp_v1->lcm_entries[0].lcme_extent.e_end);
1255 /* set attr and LOV EA at once, return updated attr */
1256 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
1257 const struct md_attr *ma)
1259 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1260 struct mdd_device *mdd = mdo2mdd(obj);
1261 struct thandle *handle = NULL;
1262 struct lu_attr *la_copy = &mdd_env_info(env)->mdi_la_for_fix;
1263 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1264 const struct lu_attr *la = &ma->ma_attr;
1265 struct lu_ucred *uc;
1266 struct lquota_id_info qi;
1267 bool quota_reserved = false;
1268 bool chrgrp_by_unprivileged_user = false;
1273 /* we do not use ->attr_set() for LOV/HSM EA any more */
1274 LASSERT((ma->ma_valid & MA_LOV) == 0);
1275 LASSERT((ma->ma_valid & MA_HSM) == 0);
1277 rc = mdd_la_get(env, mdd_obj, attr);
1281 *la_copy = ma->ma_attr;
1282 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1283 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma);
1284 mdd_write_unlock(env, mdd_obj);
1288 /* no need to setattr anymore */
1289 if (la_copy->la_valid == 0) {
1291 "%s: no valid attribute on "DFID", previous was %#llx\n",
1292 mdd_obj_dev_name(mdd_obj),
1293 PFID(mdd_object_fid(mdd_obj)), la->la_valid);
1298 /* If an unprivileged user changes group of some file,
1299 * the setattr operation will be processed synchronously to
1300 * honor the quota limit of the corresponding group. see LU-5152
1302 uc = lu_ucred_check(env);
1303 memset(&qi, 0, sizeof(qi));
1304 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
1305 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
1306 CDEBUG(D_QUOTA, "%s: reserve quota for changing group: gid=%u size=%llu\n",
1307 mdd2obd_dev(mdd)->obd_name, la->la_gid, la->la_size);
1309 if (la->la_valid & LA_BLOCKS)
1310 qi.lqi_space = la->la_blocks << 9;
1311 else if (la->la_valid & LA_SIZE)
1312 qi.lqi_space = la->la_size;
1313 /* use local attr gotten above */
1314 else if (attr->la_valid & LA_BLOCKS)
1315 qi.lqi_space = attr->la_blocks << 9;
1316 else if (attr->la_valid & LA_SIZE)
1317 qi.lqi_space = attr->la_size;
1319 if (qi.lqi_space > 0) {
1320 qi.lqi_id.qid_gid = la->la_gid;
1321 qi.lqi_type = GRPQUOTA;
1322 qi.lqi_space = toqb(qi.lqi_space);
1323 qi.lqi_is_blk = true;
1324 rc = dt_reserve_or_free_quota(env, mdd->mdd_bottom,
1328 CDEBUG(D_QUOTA, "%s: failed to reserve quota for gid %d size %llu\n",
1329 mdd2obd_dev(mdd)->obd_name,
1330 la->la_gid, qi.lqi_space);
1335 quota_reserved = true;
1336 la_copy->la_valid |= LA_FLAGS;
1339 chrgrp_by_unprivileged_user = true;
1341 /* Flush the possible existing client setattr requests to OSTs
1342 * to keep the order with the current setattr operation that
1343 * will be sent directly to OSTs. see LU-5152
1345 * LU-11303 disable sync as this is too heavyweight.
1346 * This should be replaced with a sync only for the object
1347 * being modified here, not the whole filesystem.
1348 rc = dt_sync(env, mdd->mdd_child);
1354 handle = mdd_trans_create(env, mdd);
1355 if (IS_ERR(handle)) {
1356 rc = PTR_ERR(handle);
1362 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
1366 rc = mdd_trans_start(env, mdd, handle);
1370 if (!chrgrp_by_unprivileged_user && mdd->mdd_sync_permission &&
1371 permission_needs_sync(attr, la))
1372 handle->th_sync = 1;
1374 if (la->la_valid & (LA_MTIME | LA_CTIME))
1375 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
1376 la->la_mtime, la->la_ctime);
1378 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
1380 /* LU-10509: setattr of LA_SIZE should be skipped case of DOM,
1381 * otherwise following truncate will do nothing and truncated
1382 * data may be read again. This is a quick fix until LU-11033
1385 if (la_copy->la_valid & LA_SIZE) {
1386 struct lu_buf *lov_buf = mdd_buf_get(env, NULL, 0);
1388 rc = mdd_stripe_get(env, mdd_obj, lov_buf, XATTR_NAME_LOV);
1392 if (mdd_lmm_dom_size(lov_buf->lb_buf) > 0)
1393 la_copy->la_valid &= ~LA_SIZE;
1394 lu_buf_free(lov_buf);
1398 if (la_copy->la_valid) {
1399 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1401 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
1402 /* rollback to the original gid */
1403 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
1404 la_copy->la_gid = attr->la_gid;
1405 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
1408 mdd_write_unlock(env, mdd_obj);
1412 rc = mdd_attr_set_changelog(env, obj, handle, &ma->ma_pfid,
1415 if (rc == 0 && quota_reserved) {
1416 struct thandle *sub_th;
1418 sub_th = thandle_get_sub_by_dt(env, handle, mdd->mdd_bottom);
1419 if (unlikely(IS_ERR(sub_th))) {
1421 dt_reserve_or_free_quota(env, mdd->mdd_bottom, &qi);
1423 sub_th->th_reserved_quota = qi;
1425 } else if (quota_reserved) {
1427 dt_reserve_or_free_quota(env, mdd->mdd_bottom, &qi);
1431 rc = mdd_trans_stop(env, mdd, rc, handle);
1436 static int mdd_xattr_sanity_check(const struct lu_env *env,
1437 struct mdd_object *obj,
1438 const struct lu_attr *attr,
1441 struct lu_ucred *uc = lu_ucred_assert(env);
1445 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
1448 if (strncmp(XATTR_USER_PREFIX, name,
1449 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
1450 /* For sticky directories, only the owner and privileged user
1451 * can write attributes.
1453 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
1454 (uc->uc_fsuid != attr->la_uid) &&
1455 !cap_raised(uc->uc_cap, CAP_FOWNER))
1457 } else if (strcmp(name, XATTR_NAME_SOM) != 0 &&
1458 (uc->uc_fsuid != attr->la_uid) &&
1459 !cap_raised(uc->uc_cap, CAP_FOWNER)) {
1467 * Check if a string begins with a given prefix.
1469 * \param str String to check
1470 * \param prefix Substring to check at the beginning of \a str
1471 * \return true/false whether the condition is verified.
1473 static inline bool has_prefix(const char *str, const char *prefix)
1475 return strncmp(prefix, str, strlen(prefix)) == 0;
1479 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1481 * \param[in] xattr_name Full extended attribute name.
1483 * \return type of changelog to use, or CL_NONE if no changelog is to be emitted
1485 static enum changelog_rec_type
1486 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1487 const char *xattr_name)
1489 /* Layout changes systematically recorded */
1490 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1491 strcmp(XATTR_LUSTRE_LOV, xattr_name) == 0 ||
1492 allowed_lustre_lov(xattr_name))
1495 /* HSM information changes systematically recorded */
1496 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1499 /* Avoid logging SOM xattr for every file */
1500 if (strcmp(XATTR_NAME_SOM, xattr_name) == 0)
1503 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1504 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1505 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1506 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1507 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1513 static int mdd_declare_xattr_set(const struct lu_env *env,
1514 struct mdd_device *mdd,
1515 struct mdd_object *obj,
1516 const struct lu_buf *buf,
1518 int fl, struct thandle *handle)
1520 enum changelog_rec_type type;
1523 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1527 type = mdd_xattr_changelog_type(env, mdd, name);
1529 return 0; /* no changelog to store */
1531 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
1535 * Compare current and future data of HSM EA and add a changelog if needed.
1537 * Caller should have write-locked \param obj.
1539 * \param buf - Future HSM EA content.
1540 * \retval 0 if no changelog is needed or changelog was added properly.
1541 * \retval -ve errno if there was a problem
1543 static int mdd_hsm_update_locked(const struct lu_env *env,
1544 struct md_object *obj,
1545 const struct lu_buf *buf,
1546 struct thandle *handle,
1547 enum changelog_rec_flags *clf_flags)
1549 struct mdd_thread_info *info = mdd_env_info(env);
1550 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1551 struct lu_buf *current_buf;
1552 struct md_hsm *current_mh;
1553 struct md_hsm *new_mh;
1557 OBD_ALLOC_PTR(current_mh);
1558 if (current_mh == NULL)
1561 /* Read HSM attrs from disk */
1562 current_buf = lu_buf_check_and_alloc(&info->mdi_xattr_buf,
1564 mdd_obj2mdd_dev(mdd_obj)->mdd_dt_conf.ddp_max_ea_size,
1566 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1567 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1568 if (rc < 0 && rc != -ENODATA)
1570 else if (rc == -ENODATA)
1571 current_mh->mh_flags = 0;
1573 /* Map future HSM xattr */
1574 OBD_ALLOC_PTR(new_mh);
1576 GOTO(free, rc = -ENOMEM);
1577 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1581 /* Flags differ, set flags for the changelog that will be added */
1582 if (current_mh->mh_flags != new_mh->mh_flags) {
1583 hsm_set_cl_event(clf_flags, HE_STATE);
1584 if (new_mh->mh_flags & HS_DIRTY)
1585 hsm_set_cl_flags(clf_flags, CLF_HSM_DIRTY);
1588 OBD_FREE_PTR(new_mh);
1591 OBD_FREE_PTR(current_mh);
1595 static int mdd_object_pfid_replace(const struct lu_env *env,
1596 struct mdd_object *o)
1598 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1599 struct thandle *handle;
1602 handle = mdd_trans_create(env, mdd);
1604 RETURN(PTR_ERR(handle));
1606 handle->th_complex = 1;
1608 /* it doesn't need to track the PFID update via llog, because LFSCK
1609 * will repair it even it goes wrong
1611 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1616 rc = mdd_trans_start(env, mdd, handle);
1620 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1625 mdd_trans_stop(env, mdd, rc, handle);
1630 static int mdd_declare_xattr_del(const struct lu_env *env,
1631 struct mdd_device *mdd,
1632 struct mdd_object *obj,
1634 struct thandle *handle);
1636 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1639 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1640 struct md_object *md_vic)
1642 struct mdd_device *mdd = mdo2mdd(md_obj);
1643 struct mdd_object *obj = md2mdd_obj(md_obj);
1644 struct mdd_object *vic = md2mdd_obj(md_vic);
1646 struct lu_buf *buf_vic;
1647 struct lov_mds_md *lmm;
1648 struct thandle *handle;
1649 struct lu_attr *cattr = MDD_ENV_VAR(env, cattr);
1650 struct lu_attr *tattr = MDD_ENV_VAR(env, tattr);
1651 bool is_same_projid;
1657 rc = mdd_la_get(env, obj, cattr);
1661 rc = mdd_la_get(env, vic, tattr);
1665 is_same_projid = cattr->la_projid == tattr->la_projid;
1667 if (lu_fid_cmp(mdd_object_fid(obj), mdd_object_fid(vic)) == 0)
1670 handle = mdd_trans_create(env, mdd);
1672 RETURN(PTR_ERR(handle));
1674 buf = &mdd_env_info(env)->mdi_buf[0];
1675 buf_vic = &mdd_env_info(env)->mdi_buf[1];
1677 if (!is_same_projid) {
1678 rc = mdd_declare_attr_set(env, mdd, vic, cattr, handle);
1683 /* get EA of victim file */
1684 memset(buf_vic, 0, sizeof(*buf_vic));
1685 rc = mdd_stripe_get(env, vic, buf_vic, XATTR_NAME_LOV);
1692 /* parse the layout of victim file */
1693 lmm = buf_vic->lb_buf;
1694 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1695 GOTO(stop, rc = -EINVAL);
1697 /* save EA of target file for restore */
1698 memset(buf, 0, sizeof(*buf));
1699 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
1703 /* Get rid of the layout from victim object */
1704 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1708 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_NAME_LOV,
1709 LU_XATTR_MERGE, handle);
1713 rc = mdd_trans_start(env, mdd, handle);
1717 rc = mdd_write_lock_two_objects(env, obj, vic);
1721 if (!is_same_projid) {
1722 cattr->la_valid = LA_PROJID;
1723 rc = mdd_attr_set_internal(env, vic, cattr, handle, 0);
1728 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV, LU_XATTR_MERGE,
1733 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1735 GOTO(out_restore, rc);
1737 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
1739 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle,
1744 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1745 LU_XATTR_REPLACE, handle);
1747 CERROR("%s: failed rollback of "DFID" layout: file state unknown: rc = %d\n",
1748 mdd_obj_dev_name(obj),
1749 PFID(mdd_object_fid(obj)), rc2);
1753 mdd_write_unlock_two_objects(env, obj, vic);
1755 mdd_trans_stop(env, mdd, rc, handle);
1757 lu_buf_free(buf_vic);
1760 * -EAGAIN means transaction execution phase detect the layout
1761 * has been changed by others.
1764 (void) mdd_object_pfid_replace(env, obj);
1765 else if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
1766 GOTO(retry, retried);
1772 * Extract the mirror with specified mirror id, and store the splitted
1773 * mirror layout to @buf.
1775 * \param[in] comp_v1 mirrored layout
1776 * \param[in] mirror_id the mirror with mirror_id to be extracted
1777 * \param[out] buf store the layout excluding the extracted mirror,
1778 * caller free the buffer we allocated in this function
1779 * \param[out] buf_vic store the extracted layout, caller free the buffer
1780 * we allocated in this function
1782 * \retval 0 on success; < 0 if error happens
1784 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1785 struct lu_buf *buf, struct lu_buf *buf_vic)
1787 struct lov_comp_md_v1 *comp_rem;
1788 struct lov_comp_md_v1 *comp_vic;
1789 struct lov_comp_md_entry_v1 *entry;
1790 struct lov_comp_md_entry_v1 *entry_rem;
1791 struct lov_comp_md_entry_v1 *entry_vic;
1793 __u16 comp_cnt, count = 0;
1794 int lmm_size, lmm_size_vic = 0;
1796 int offset, offset_rem, offset_vic;
1798 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1799 /* comp_v1 should contains more than 1 mirror */
1800 if (mirror_cnt <= 1)
1802 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1803 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1805 for (i = 0; i < comp_cnt; i++) {
1806 entry = &comp_v1->lcm_entries[i];
1807 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1809 lmm_size_vic += sizeof(*entry);
1810 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1811 } else if (count > 0) {
1812 /* find the specified mirror */
1820 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1824 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1825 if (!buf_vic->lb_buf) {
1830 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1831 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1833 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1834 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1835 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1836 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1837 if (!comp_rem->lcm_mirror_count)
1838 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1840 memset(comp_vic, 0, sizeof(*comp_v1));
1841 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1842 comp_vic->lcm_mirror_count = 0;
1843 comp_vic->lcm_entry_count = cpu_to_le32(count);
1844 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1845 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1846 comp_vic->lcm_layout_gen = 0;
1848 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1849 offset_rem = sizeof(*comp_rem) +
1850 sizeof(*entry_rem) * (comp_cnt - count);
1851 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1852 for (i = j = k = 0; i < comp_cnt; i++) {
1853 struct lov_mds_md *lmm, *lmm_dst;
1856 entry = &comp_v1->lcm_entries[i];
1857 entry_vic = &comp_vic->lcm_entries[j];
1858 entry_rem = &comp_rem->lcm_entries[k];
1860 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1863 /* copy component entry */
1865 memcpy(entry_vic, entry, sizeof(*entry));
1866 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1867 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1870 memcpy(entry_rem, entry, sizeof(*entry));
1871 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1875 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1877 lmm_dst = (struct lov_mds_md *)
1878 ((char *)comp_vic + offset_vic);
1880 lmm_dst = (struct lov_mds_md *)
1881 ((char *)comp_rem + offset_rem);
1883 /* copy component entry blob */
1884 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1886 /* blob offset advance */
1887 offset += le32_to_cpu(entry->lcme_size);
1889 offset_vic += le32_to_cpu(entry->lcme_size);
1891 offset_rem += le32_to_cpu(entry->lcme_size);
1897 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1898 struct md_rejig_data *mrd)
1900 struct mdd_device *mdd = mdo2mdd(md_obj);
1901 struct mdd_object *obj = md2mdd_obj(md_obj);
1902 struct mdd_object *vic = NULL;
1904 struct lu_buf *buf_save;
1905 struct lu_buf *buf_vic;
1906 struct lov_comp_md_v1 *lcm;
1907 struct thandle *handle;
1909 bool dom_stripe = false;
1915 * NULL @mrd_obj means mirror deleting, and use NULL vic to indicate
1919 vic = md2mdd_obj(mrd->mrd_obj);
1922 buf = &mdd_env_info(env)->mdi_buf[0];
1923 buf_save = &mdd_env_info(env)->mdi_buf[1];
1924 buf_vic = &mdd_env_info(env)->mdi_buf[2];
1926 handle = mdd_trans_create(env, mdd);
1928 RETURN(PTR_ERR(handle));
1930 /* get EA of mirrored file */
1931 memset(buf_save, 0, sizeof(*buf));
1932 rc = mdd_stripe_get(env, obj, buf_save, XATTR_NAME_LOV);
1936 lcm = buf_save->lb_buf;
1937 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1938 GOTO(stop, rc = -EINVAL);
1941 * Extract the mirror with specified mirror id, and store the splitted
1942 * mirror layout to the victim buffer.
1944 memset(buf, 0, sizeof(*buf));
1945 memset(buf_vic, 0, sizeof(*buf_vic));
1946 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1950 * @buf stores layout w/o the specified mirror, @buf_vic stores the
1954 dom_stripe = mdd_lmm_dom_size(buf_vic->lb_buf) > 0;
1958 * non delete mirror split
1960 * declare obj set remaining layout in @buf, will set obj's
1963 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1964 LU_XATTR_SPLIT, handle);
1968 /* declare vic set splitted layout in @buf_vic */
1969 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic,
1970 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1976 * declare delete mirror objects in @buf_vic, will change obj's
1979 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic,
1980 XATTR_NAME_LOV, LU_XATTR_PURGE,
1985 /* declare obj set remaining layout in @buf */
1986 rc = mdd_declare_xattr_set(env, mdd, obj, buf,
1987 XATTR_NAME_LOV, LU_XATTR_SPLIT,
1993 rc = mdd_trans_start(env, mdd, handle);
1997 rc = mdd_write_lock_two_objects(env, obj, vic);
2001 /* set obj's layout in @buf */
2002 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_SPLIT,
2008 /* set vic's layout in @buf_vic */
2009 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV,
2010 LU_XATTR_CREATE, handle);
2012 GOTO(out_restore, rc);
2014 /* delete mirror objects */
2015 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_NAME_LOV,
2016 LU_XATTR_PURGE, handle);
2018 GOTO(out_restore, rc);
2021 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
2024 GOTO(out_restore, rc);
2027 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic,
2030 GOTO(out_restore, rc);
2035 /* restore obj's in-memory and on-disk layout */
2036 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
2037 LU_XATTR_REPLACE, handle);
2039 CERROR("%s: failed rollback "DFID" layout: file state unknown: rc = %d\n",
2040 mdd_obj_dev_name(obj),
2041 PFID(mdd_object_fid(obj)), rc);
2045 mdd_write_unlock_two_objects(env, obj, vic);
2047 rc = mdd_trans_stop(env, mdd, rc, handle);
2049 /* Truncate local DOM data if all went well */
2050 if (!rc && dom_stripe)
2051 mdd_dom_fixup(env, mdd, obj, NULL);
2053 lu_buf_free(buf_save);
2055 lu_buf_free(buf_vic);
2058 * -EAGAIN means transaction execution phase detect the layout
2059 * has been changed by others.
2062 (void) mdd_object_pfid_replace(env, obj);
2063 else if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
2064 GOTO(retry, retried);
2069 static int mdd_layout_merge_allowed(const struct lu_env *env,
2070 struct md_object *target,
2071 struct md_object *victim)
2073 struct mdd_object *o1 = md2mdd_obj(target);
2076 /* cannot extend directory's LOVEA */
2077 if (S_ISDIR(mdd_object_type(o1))) {
2079 CERROR("%s: Don't extend directory's LOVEA, just set it: rc = %d\n",
2080 mdd_obj_dev_name(o1), rc);
2088 * The caller should guarantee to update the object ctime
2089 * after xattr_set if needed.
2091 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
2092 const struct lu_buf *buf, const char *name,
2095 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2096 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2097 struct mdd_device *mdd = mdo2mdd(obj);
2098 struct thandle *handle;
2099 enum changelog_rec_type cl_type;
2100 enum changelog_rec_flags clf_flags = 0;
2107 rc = mdd_la_get(env, mdd_obj, attr);
2111 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2115 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
2116 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
2117 struct md_rejig_data *mrd = buf->lb_buf;
2118 struct md_object *victim = mrd->mrd_obj;
2120 if (buf->lb_len != sizeof(*mrd))
2124 if (fl == LU_XATTR_MERGE) {
2125 rc = mdd_layout_merge_allowed(env, obj, victim);
2128 /* merge layout of victim as a mirror of obj's. */
2129 rc = mdd_xattr_merge(env, obj, victim);
2131 rc = mdd_xattr_split(env, obj, mrd);
2136 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
2137 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
2138 struct posix_acl *acl;
2140 /* user set empty ACL, this should be treated as removing ACL */
2141 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
2144 RETURN(PTR_ERR(acl));
2146 rc = mdd_xattr_del(env, obj, name);
2149 posix_acl_release(acl);
2152 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
2153 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
2157 handle = mdd_trans_create(env, mdd);
2159 RETURN(PTR_ERR(handle));
2161 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
2165 rc = mdd_trans_start(env, mdd, handle);
2169 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2171 if (strcmp(XATTR_NAME_HSM, name) == 0) {
2172 rc = mdd_hsm_update_locked(env, obj, buf, handle, &clf_flags);
2174 GOTO(out_unlock, rc);
2177 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
2179 mdd_write_unlock(env, mdd_obj);
2183 cl_type = mdd_xattr_changelog_type(env, mdd, name);
2184 if (cl_type != CL_NONE)
2185 rc = mdd_changelog_data_store_xattr(env, mdd, cl_type,
2186 clf_flags, mdd_obj, name,
2190 rc = mdd_trans_stop(env, mdd, rc, handle);
2193 * -EAGAIN means transaction execution phase detect the layout
2194 * has been changed by others.
2196 if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
2197 GOTO(retry, retried);
2202 static int mdd_declare_xattr_del(const struct lu_env *env,
2203 struct mdd_device *mdd,
2204 struct mdd_object *obj,
2206 struct thandle *handle)
2208 enum changelog_rec_type type;
2211 rc = mdo_declare_xattr_del(env, obj, name, handle);
2215 type = mdd_xattr_changelog_type(env, mdd, name);
2217 return 0; /* no changelog to store */
2219 return mdd_declare_changelog_store(env, mdd, type, NULL, NULL, handle);
2223 * The caller should guarantee to update the object ctime
2224 * after xattr_set if needed.
2226 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
2229 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2230 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2231 struct mdd_device *mdd = mdo2mdd(obj);
2232 struct thandle *handle;
2237 rc = mdd_la_get(env, mdd_obj, attr);
2241 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
2245 handle = mdd_trans_create(env, mdd);
2247 RETURN(PTR_ERR(handle));
2249 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
2253 rc = mdd_trans_start(env, mdd, handle);
2257 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
2258 rc = mdo_xattr_del(env, mdd_obj, name, handle);
2259 mdd_write_unlock(env, mdd_obj);
2263 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
2266 rc = mdd_changelog_data_store_xattr(env, mdd, CL_SETXATTR, 0, mdd_obj,
2271 return mdd_trans_stop(env, mdd, rc, handle);
2275 * read lov/lmv EA of an object
2276 * return the lov/lmv EA in an allocated lu_buf
2278 int mdd_stripe_get(const struct lu_env *env, struct mdd_object *obj,
2279 struct lu_buf *lmm_buf, const char *name)
2281 struct lu_buf *buf = &mdd_env_info(env)->mdi_big_buf;
2286 if (buf->lb_buf == NULL) {
2287 buf = lu_buf_check_and_alloc(buf, 4096);
2288 if (buf->lb_buf == NULL)
2293 rc = mdo_xattr_get(env, obj, buf, name);
2294 if (rc == -ERANGE) {
2295 /* mdi_big_buf is allocated but too small need to increase it */
2296 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mdi_big_buf,
2298 if (buf->lb_buf == NULL)
2301 } else if (rc < 0) {
2303 } else if (rc == 0) {
2307 lu_buf_alloc(lmm_buf, rc);
2308 if (lmm_buf->lb_buf == NULL)
2312 * we don't use lmm_buf directly, because we don't know xattr size, so
2313 * by using mdi_big_buf we can avoid calling mdo_xattr_get() twice.
2315 memcpy(lmm_buf->lb_buf, buf->lb_buf, rc);
2320 static int emit_changelog_after_swap_layout(const struct lu_env *env,
2321 struct thandle *handle,
2322 struct mdd_object *o,
2323 struct lu_buf *hsm_buf)
2325 enum changelog_rec_flags flags = 0;
2326 enum changelog_rec_type type;
2327 enum hsm_states hsm_flags;
2328 struct hsm_attrs *attrs;
2331 attrs = hsm_buf->lb_buf;
2332 hsm_flags = le32_to_cpu(attrs->hsm_flags);
2334 if ((hsm_flags & HS_RELEASED) && !mdd_is_dead_obj(o)) {
2335 hsm_set_cl_event(&flags, HE_RELEASE);
2341 return mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), type,
2342 flags, o, handle, NULL);
2346 * check if layout swapping between 2 objects is allowed
2348 * - only normal FIDs or non-system IGIFs
2349 * - same type of objects
2350 * - same owner/group (so quotas are still valid) unless this is from HSM
2353 static int mdd_layout_swap_allowed(const struct lu_env *env,
2354 struct mdd_object *o1,
2355 const struct lu_attr *attr1,
2356 struct mdd_object *o2,
2357 const struct lu_attr *attr2,
2360 const struct lu_fid *fid1, *fid2;
2364 fid1 = mdd_object_fid(o1);
2365 fid2 = mdd_object_fid(o2);
2367 if (!fid_is_norm(fid1) &&
2368 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
2371 if (!fid_is_norm(fid2) &&
2372 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
2375 if (mdd_object_type(o1) != mdd_object_type(o2)) {
2376 if (S_ISDIR(mdd_object_type(o1)))
2378 if (S_ISREG(mdd_object_type(o1)))
2383 /* Do not check uid/gid for release since the orphan is created as root
2385 if (flags & SWAP_LAYOUTS_MDS_RELEASE)
2388 if ((attr1->la_uid != attr2->la_uid) ||
2389 (attr1->la_gid != attr2->la_gid))
2395 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
2396 * look into the layout in MDD layer.
2398 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
2400 struct lov_comp_md_v1 *comp_v1;
2401 struct lov_mds_md *v1;
2405 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2406 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2407 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
2415 off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
2416 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
2417 if (le32_to_cpu(v1->lmm_magic) != LOV_MAGIC_FOREIGN) {
2425 comp_v1->lcm_entries[i].lcme_offset);
2426 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2428 if (le32_to_cpu(v1->lmm_magic) ==
2435 for (i = 0; i < le32_to_cpu(ent_count); i++) {
2436 off = le32_to_cpu(comp_v1->lcm_entries[i].
2438 v1 = (struct lov_mds_md *)((char *)comp_v1 +
2440 if (le32_to_cpu(v1->lmm_magic) !=
2445 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2446 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2457 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2459 return mdd_lmm_oi(lmm, oi, true);
2462 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
2464 return mdd_lmm_oi(lmm, oi, false);
2467 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
2469 struct lov_comp_md_v1 *comp_v1;
2471 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
2472 comp_v1 = (struct lov_comp_md_v1 *)lmm;
2474 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
2476 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
2477 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
2478 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
2479 __u16 tmp_gen = *gen;
2482 *gen = le16_to_cpu(lmm->lmm_layout_gen);
2484 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
2491 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2493 return mdd_lmm_gen(lmm, gen, true);
2496 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
2498 return mdd_lmm_gen(lmm, gen, false);
2501 int mdd_dom_fixup(const struct lu_env *env, struct mdd_device *mdd,
2502 struct mdd_object *mo, struct mdd_object *vo)
2504 struct dt_object *dom, *vlt;
2505 dt_obj_version_t dv = 0;
2512 vlt = dt_object_locate(mdd_object_child(vo), mdd->mdd_bottom);
2514 GOTO(out, rc = -ENOENT);
2515 dv = dt_data_version_get(env, vlt);
2517 GOTO(out, rc = -ENODATA);
2520 dom = dt_object_locate(mdd_object_child(mo), mdd->mdd_bottom);
2522 th = dt_trans_create(env, mdd->mdd_bottom);
2524 GOTO(out, rc = PTR_ERR(th));
2527 rc = dt_declare_data_version_set(env, dom, th);
2531 rc = dt_declare_data_version_del(env, dom, th);
2534 rc = dt_declare_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2539 rc = dt_trans_start_local(env, mdd->mdd_bottom, th);
2544 dt_data_version_set(env, dom, dv, th);
2546 dt_data_version_del(env, dom, th);
2547 rc = dt_punch(env, dom, 0, OBD_OBJECT_EOF, th);
2551 dt_trans_stop(env, mdd->mdd_bottom, th);
2553 /* Ignore failure but report the error */
2555 CERROR("%s: can't manage DOM file "DFID" data: rc = %d\n",
2556 mdd_obj_dev_name(mo), PFID(mdd_object_fid(mo)), rc);
2560 /* Swap Layout HSM object information: every information needed to decide how
2561 * to update the HSM xattr flag during a layout swap.
2563 struct sl_hsm_object_info {
2564 struct mdd_object *o; /* the object whose HSM attribute to check */
2565 struct lu_buf *hsm_buf; /* buffer initialized with the content of the
2568 int xattr_flags; /* 0: do not update HSM xattr
2569 * LU_XATTR_REPLACE or LU_XATTR_CREATE: update
2570 * the xattr with this flag
2572 __u64 dv; /* dataversion of the object */
2575 /* Read the HSM xattr and store its content into \p hsm_buf. */
2576 static int fetch_hsm_xattr(const struct lu_env *env, struct mdd_object *o,
2577 struct lu_buf *hsm_buf)
2581 lu_buf_alloc(hsm_buf, sizeof(struct hsm_attrs));
2582 if (hsm_buf->lb_buf == NULL)
2585 rc = mdo_xattr_get(env, o, hsm_buf, XATTR_NAME_HSM);
2592 /* return true if HSM xattr needs update */
2593 static bool swap_hsm_set_dirty(struct lu_buf *out_buf, struct lu_buf *src_buf)
2595 struct md_hsm src_hsm;
2597 lustre_buf2hsm(src_buf->lb_buf, src_buf->lb_len, &src_hsm);
2598 if (!(src_hsm.mh_flags & (HS_ARCHIVED | HS_EXISTS)) ||
2599 src_hsm.mh_flags & HS_DIRTY)
2600 /* do not update HSM attr of non archived or dirty files */
2603 src_hsm.mh_flags |= HS_DIRTY;
2604 lustre_hsm2buf(out_buf->lb_buf, &src_hsm);
2609 /* return 1 HSM xattr needs update
2610 * 0 do not update HSM xattr
2611 * -EPERM data version mismatch, no swap
2613 static int swap_hsm_update_version(struct lu_buf *out_buf, __u64 out_dv,
2614 struct lu_buf *src_buf, __u64 src_dv)
2616 struct md_hsm src_hsm;
2618 /* Put lb_len as a second argument since we know that src_buf is a valid
2621 lustre_buf2hsm(src_buf->lb_buf, src_buf->lb_len, &src_hsm);
2623 if (!(src_hsm.mh_flags & (HS_ARCHIVED | HS_EXISTS)) ||
2624 src_hsm.mh_flags & HS_DIRTY)
2627 /* migration with old client -> set the dirty flag */
2628 if (!src_dv || !out_dv) {
2629 src_hsm.mh_flags |= HS_DIRTY;
2633 if (src_hsm.mh_arch_ver != src_dv) {
2635 "HSM archive version and previous data version mismatch (arch_ver=%llu, prev_ver=%llu, new_ver=%llu)\n",
2636 src_hsm.mh_arch_ver, src_dv, out_dv);
2640 src_hsm.mh_arch_ver = out_dv;
2642 lustre_hsm2buf(out_buf->lb_buf, &src_hsm);
2647 /* Allow HSM xattr swap only with volatile/orphan files.
2648 * This is used by HSM release/restore.
2650 static inline bool swap_hsm_xattr_allowed(bool fst_has_hsm, bool snd_has_hsm,
2651 unsigned long o_fst_flag,
2652 unsigned long o_snd_flag)
2654 return (fst_has_hsm && snd_has_hsm &&
2655 ((o_fst_flag | o_snd_flag) & (ORPHAN_OBJ | VOLATILE_OBJ)));
2658 /* Read and update the data version of both objects if necessary. If they have
2659 * to be updated, declare the update.
2660 * \p xattr_flags will be updated if necessary to be used by mdo_xattr_set
2662 * \p dv contains the data version of the file before and after the migration.
2663 * It can be NULL when swapping layouts in other contexts (HSM or
2666 static int swap_layouts_prepare_hsm_attr(const struct lu_env *env,
2667 struct mdd_device *mdd,
2668 struct thandle *handle,
2669 struct sl_hsm_object_info *fst,
2670 struct sl_hsm_object_info *snd)
2672 unsigned long o_fst_fl;
2673 unsigned long o_snd_fl;
2677 fst->xattr_flags = 0;
2678 snd->xattr_flags = 0;
2680 rc = fetch_hsm_xattr(env, fst->o, fst->hsm_buf);
2681 if (rc != -ENODATA && rc != 0)
2684 rc2 = fetch_hsm_xattr(env, snd->o, snd->hsm_buf);
2685 if (rc2 != -ENODATA && rc2 != 0)
2688 /* if nothing to swap, not an error */
2692 /* swap if the first object have no HSM xattr */
2698 o_fst_fl = fst->o->mod_flags;
2699 o_snd_fl = snd->o->mod_flags;
2701 if ((o_snd_fl & VOLATILE_OBJ) && rc2) {
2702 /* migration of fst */
2703 rc = swap_hsm_update_version(snd->hsm_buf, snd->dv,
2704 fst->hsm_buf, fst->dv);
2710 fst->xattr_flags = LU_XATTR_REPLACE;
2712 } else if (swap_hsm_xattr_allowed(!rc, !rc2, o_fst_fl, o_snd_fl)) {
2713 /* HSM release/restore -> HSM xattr swap */
2714 fst->xattr_flags = LU_XATTR_REPLACE;
2715 snd->xattr_flags = LU_XATTR_REPLACE;
2717 } else if (!rc && !rc2) {
2718 /* swap on 2 archived files is not supported (no rollback) */
2722 /* swap layout with HSM fst and non-HSM snd */
2723 if (!swap_hsm_set_dirty(snd->hsm_buf, fst->hsm_buf))
2726 fst->xattr_flags = LU_XATTR_REPLACE;
2729 if (fst->xattr_flags) {
2730 rc = mdd_declare_xattr_set(env, mdd, fst->o, snd->hsm_buf,
2731 XATTR_NAME_HSM, fst->xattr_flags,
2737 if (snd->xattr_flags) {
2738 rc = mdd_declare_xattr_set(env, mdd, snd->o, fst->hsm_buf,
2739 XATTR_NAME_HSM, snd->xattr_flags,
2748 /* swap layouts between 2 lustre objects */
2749 static int mdd_swap_layouts(const struct lu_env *env,
2750 struct md_object *obj1, struct md_object *obj2,
2751 __u64 dv1, __u64 dv2, __u64 flags)
2753 struct mdd_thread_info *info = mdd_env_info(env);
2754 struct mdd_object *fst_o = md2mdd_obj(obj1);
2755 struct mdd_object *snd_o = md2mdd_obj(obj2);
2756 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
2757 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
2758 struct mdd_device *mdd = mdo2mdd(obj1);
2759 struct lov_mds_md *fst_lmm, *snd_lmm;
2760 struct sl_hsm_object_info fst_info;
2761 struct sl_hsm_object_info snd_info;
2762 struct mdd_object *vlt_o = NULL;
2763 struct mdd_object *dom_o = NULL;
2764 struct ost_id *saved_oi = NULL;
2765 struct lu_buf *fst_hsm_buf;
2766 struct lu_buf *snd_hsm_buf;
2767 struct lu_buf *fst_buf;
2768 struct lu_buf *snd_buf;
2769 struct thandle *handle;
2782 BUILD_BUG_ON(ARRAY_SIZE(info->mdi_buf) < 4);
2784 fst_buf = &info->mdi_buf[0];
2785 snd_buf = &info->mdi_buf[1];
2786 fst_hsm_buf = &info->mdi_buf[2];
2787 snd_hsm_buf = &info->mdi_buf[3];
2789 memset(info->mdi_buf, 0, sizeof(info->mdi_buf));
2791 /* we have to sort the 2 obj, so locking will always
2792 * be in the same order, even in case of 2 concurrent swaps
2794 rc = lu_fid_cmp(mdd_object_fid(fst_o), mdd_object_fid(snd_o));
2795 if (rc == 0) /* same fid ? */
2803 rc = mdd_la_get(env, fst_o, fst_la);
2807 rc = mdd_la_get(env, snd_o, snd_la);
2811 /* check if layout swapping is allowed */
2812 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
2816 handle = mdd_trans_create(env, mdd);
2818 RETURN(PTR_ERR(handle));
2820 rc = mdd_stripe_get(env, fst_o, fst_buf, XATTR_NAME_LOV);
2821 if (rc < 0 && rc != -ENODATA)
2824 rc = mdd_stripe_get(env, snd_o, snd_buf, XATTR_NAME_LOV);
2825 if (rc < 0 && rc != -ENODATA)
2828 /* check if file has DoM. DoM file can be migrated only to another
2829 * DoM layout with the same DoM component size or to an non-DOM
2830 * layout. After migration to OSTs layout, local MDT inode data
2831 * should be truncated.
2832 * Objects are sorted by FIDs, considering that original file's FID
2833 * is always smaller the snd_o is always original file we are migrating
2836 domsize_dom = mdd_lmm_dom_size(snd_buf->lb_buf);
2837 domsize_vlt = mdd_lmm_dom_size(fst_buf->lb_buf);
2839 /* Only migration is supported for DoM files, not 'swap_layouts' so
2840 * target file must be volatile and orphan.
2842 if (fst_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2843 vlt_o = domsize_vlt ? fst_o : NULL;
2844 dom_o = domsize_dom ? snd_o : NULL;
2845 } else if (snd_o->mod_flags & (ORPHAN_OBJ | VOLATILE_OBJ)) {
2846 swap(domsize_dom, domsize_vlt);
2847 vlt_o = domsize_vlt ? snd_o : NULL;
2848 dom_o = domsize_dom ? fst_o : NULL;
2849 } else if (domsize_dom > 0 || domsize_vlt > 0) {
2850 /* 'lfs swap_layouts' case, neither file should have DoM */
2852 CDEBUG(D_LAYOUT, "cannot swap layouts with DOM component, use migration instead: rc = %d\n",
2857 if (domsize_vlt > 0 && domsize_dom == 0) {
2860 "%s: cannot swap "DFID" layout: OST to DOM migration not supported: rc = %d\n",
2861 mdd_obj_dev_name(snd_o),
2862 PFID(mdd_object_fid(snd_o)), rc);
2864 } else if (domsize_vlt > 0 && domsize_dom != domsize_vlt) {
2867 "%s: cannot swap "DFID" layout: new layout must have same DoM component size: rc = %d\n",
2868 mdd_obj_dev_name(fst_o),
2869 PFID(mdd_object_fid(fst_o)), rc);
2873 /* swapping 2 non existant layouts is a success */
2874 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
2877 /* to help inode migration between MDT, it is better to
2878 * start by the no layout file (if one), so we order the swap
2880 if (snd_buf->lb_buf == NULL) {
2882 swap(fst_buf, snd_buf);
2886 fst_gen = snd_gen = 0;
2887 /* lmm and generation layout initialization */
2888 if (fst_buf->lb_buf != NULL) {
2889 fst_lmm = fst_buf->lb_buf;
2890 mdd_get_lmm_gen(fst_lmm, &fst_gen);
2891 fst_fl = LU_XATTR_REPLACE;
2895 fst_fl = LU_XATTR_CREATE;
2898 snd_lmm = snd_buf->lb_buf;
2899 mdd_get_lmm_gen(snd_lmm, &snd_gen);
2901 saved_gen = fst_gen;
2902 /* increase the generation layout numbers */
2907 * XXX The layout generation is used to generate component IDs for
2908 * the composite file, we have to do some special tweaks to make
2909 * sure the layout generation is always adequate for that job.
2912 /* Skip invalid generation number for composite layout */
2913 if ((snd_gen & LCME_ID_MASK) == 0)
2915 if ((fst_gen & LCME_ID_MASK) == 0)
2917 /* Make sure the generation is greater than all the component IDs */
2918 if (fst_gen < snd_gen)
2920 else if (fst_gen > snd_gen)
2923 /* set the file specific informations in lmm */
2924 if (fst_lmm != NULL) {
2925 struct ost_id temp_oi;
2927 saved_oi = &info->mdi_oa.o_oi;
2928 mdd_get_lmm_oi(fst_lmm, saved_oi);
2929 mdd_get_lmm_oi(snd_lmm, &temp_oi);
2930 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2931 mdd_set_lmm_oi(fst_lmm, &temp_oi);
2932 mdd_set_lmm_oi(snd_lmm, saved_oi);
2934 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2935 cpu_to_le32(LOV_MAGIC_MAGIC))
2936 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2938 GOTO(stop, rc = -EPROTO);
2940 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2943 fst_info.hsm_buf = fst_hsm_buf;
2947 snd_info.hsm_buf = snd_hsm_buf;
2949 rc = swap_layouts_prepare_hsm_attr(env, mdd, handle, &fst_info,
2954 /* prepare transaction */
2955 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2960 if (fst_buf->lb_buf != NULL)
2961 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2962 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2965 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2970 rc = mdd_trans_start(env, mdd, handle);
2974 rc = mdd_write_lock_two_objects(env, fst_o, snd_o);
2978 if (!mdd_object_exists(fst_o))
2979 GOTO(unlock, rc = -ENOENT);
2980 if (!mdd_object_exists(snd_o))
2981 GOTO(unlock, rc = -ENOENT);
2983 if (fst_info.xattr_flags) {
2984 rc = mdo_xattr_set(env, fst_o, snd_info.hsm_buf,
2985 XATTR_NAME_HSM, fst_info.xattr_flags,
2991 if (snd_info.xattr_flags) {
2992 rc = mdo_xattr_set(env, snd_o, fst_info.hsm_buf,
2993 XATTR_NAME_HSM, snd_info.xattr_flags,
2996 GOTO(out_restore_hsm_fst, rc);
2999 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
3001 GOTO(out_restore_hsm, rc);
3003 if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
3006 if (fst_buf->lb_buf != NULL)
3007 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
3008 LU_XATTR_REPLACE, handle);
3010 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
3013 GOTO(out_restore, rc);
3015 /* Issue one changelog record per file */
3016 rc = emit_changelog_after_swap_layout(env, handle, fst_o, fst_hsm_buf);
3020 rc = emit_changelog_after_swap_layout(env, handle, snd_o, snd_hsm_buf);
3029 /* failure on second file, but first was done, so we have
3030 * to roll back first.
3032 if (fst_buf->lb_buf != NULL) {
3033 mdd_set_lmm_oi(fst_lmm, saved_oi);
3034 mdd_set_lmm_gen(fst_lmm, &saved_gen);
3035 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
3036 LU_XATTR_REPLACE, handle);
3038 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
3044 if (snd_info.xattr_flags) {
3045 /* roll back swap HSM */
3047 rc2 = mdo_xattr_set(env, snd_o, snd_hsm_buf,
3048 XATTR_NAME_HSM, LU_XATTR_REPLACE,
3054 out_restore_hsm_fst:
3055 if (fst_info.xattr_flags) {
3057 rc2 = mdo_xattr_set(env, fst_o, fst_hsm_buf,
3058 XATTR_NAME_HSM, LU_XATTR_REPLACE,
3065 CERROR("%s: unable to roll back layout swap of "DFID" and "DFID", steps: %d: rc = %d/%d\n",
3066 mdd_obj_dev_name(fst_o),
3067 PFID(mdd_object_fid(snd_o)),
3068 PFID(mdd_object_fid(fst_o)),
3070 /* a solution to avoid journal commit is to panic,
3071 * but it has strong consequences so we use LBUG to
3072 * allow sysdamin to choose to panic or not
3079 mdd_write_unlock_two_objects(env, fst_o, snd_o);
3082 rc = mdd_trans_stop(env, mdd, rc, handle);
3084 /* Truncate local DOM data if all went well, except migration case
3085 * with the same DOM component size. In that case a local data is
3086 * still in use and shouldn't be deleted.
3089 mdd_dom_fixup(env, mdd, dom_o, vlt_o);
3091 lu_buf_free(fst_buf);
3092 lu_buf_free(snd_buf);
3093 lu_buf_free(fst_hsm_buf);
3094 lu_buf_free(snd_hsm_buf);
3097 (void) mdd_object_pfid_replace(env, fst_o);
3098 (void) mdd_object_pfid_replace(env, snd_o);
3099 } else if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED) {
3101 * -EAGAIN means transaction execution phase detect the layout
3102 * has been changed by others.
3104 GOTO(retry, retried);
3110 static int mdd_declare_layout_change(const struct lu_env *env,
3111 struct mdd_device *mdd,
3112 struct mdd_object *obj,
3113 struct md_layout_change *mlc,
3114 struct thandle *handle)
3118 rc = mdo_declare_layout_change(env, obj, mlc, handle);
3122 return mdd_declare_changelog_store(env, mdd, CL_LAYOUT, NULL, NULL,
3126 /* For PFL, this is used to instantiate necessary component objects. */
3128 mdd_layout_instantiate_component(const struct lu_env *env,
3129 struct mdd_object *obj, struct md_layout_change *mlc,
3130 struct thandle *handle)
3132 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3137 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
3140 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3142 * It's possible that another layout write intent has already
3143 * instantiated our objects, so a -EALREADY returned, and we need to
3147 RETURN(rc == -EALREADY ? 0 : rc);
3149 rc = mdd_trans_start(env, mdd, handle);
3153 mdd_write_lock(env, obj, DT_TGT_CHILD);
3154 rc = mdo_layout_change(env, obj, mlc, handle);
3155 mdd_write_unlock(env, obj);
3159 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
3165 * Change the FLR layout from RDONLY to WRITE_PENDING.
3167 * It picks the primary mirror, and bumps the layout version, and set
3168 * layout version xattr to OST objects in a sync tx. In order to facilitate
3169 * the handling of phantom writers from evicted clients, the clients carry
3170 * layout version of the file with write RPC, so that the OSTs can verify
3171 * if the write RPCs are legitimate, meaning not from evicted clients.
3174 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
3175 struct md_layout_change *mlc, struct thandle *handle)
3177 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3178 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3179 struct lustre_som_attrs *som = &mlc->mlc_som;
3185 /* Verify acceptable operations */
3186 switch (mlc->mlc_opc) {
3187 case MD_LAYOUT_WRITE:
3188 case MD_LAYOUT_RESYNC:
3189 /* these are legal operations - this represents the case that
3190 * a few mirrors were missed in the last resync.
3193 case MD_LAYOUT_RESYNC_DONE:
3198 som_buf->lb_buf = som;
3199 som_buf->lb_len = sizeof(*som);
3200 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
3201 if (rc < 0 && rc != -ENODATA)
3205 lustre_som_swab(som);
3206 if (som->lsa_valid & SOM_FL_STRICT)
3207 fl = LU_XATTR_REPLACE;
3209 if (mlc->mlc_opc == MD_LAYOUT_WRITE &&
3210 mlc->mlc_intent->lai_extent.e_end > som->lsa_size) {
3211 som->lsa_size = mlc->mlc_intent->lai_extent.e_end + 1;
3212 fl = LU_XATTR_REPLACE;
3216 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3221 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3222 XATTR_NAME_SOM, fl, handle);
3227 /* record a changelog for data mover to consume */
3228 rc = mdd_declare_changelog_store(env, mdd, CL_FLRW, NULL, NULL, handle);
3232 rc = mdd_trans_start(env, mdd, handle);
3236 /* it needs a sync tx to make FLR to work properly */
3237 handle->th_sync = 1;
3239 mdd_write_lock(env, obj, DT_TGT_CHILD);
3240 rc = mdo_layout_change(env, obj, mlc, handle);
3242 /* SOM state transition from STRICT to STALE */
3243 som->lsa_valid = SOM_FL_STALE;
3244 lustre_som_swab(som);
3245 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3248 mdd_write_unlock(env, obj);
3252 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle, NULL);
3259 * Handle mirrored file state transition when it's in WRITE_PENDING.
3261 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
3262 * the file is in WRITE_PENDING state. If everything goes fine, the file's
3263 * layout version will be increased, and the file's state will be changed to
3267 mdd_layout_update_write_pending(const struct lu_env *env,
3268 struct mdd_object *obj, struct md_layout_change *mlc,
3269 struct thandle *handle)
3271 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3272 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3273 struct lustre_som_attrs *som = &mlc->mlc_som;
3279 switch (mlc->mlc_opc) {
3280 case MD_LAYOUT_RESYNC:
3281 /* Upon receiving the resync request, it should
3282 * instantiate all stale components right away to get ready
3283 * for mirror copy. In order to avoid layout version change,
3284 * client should avoid sending LAYOUT_WRITE request at the
3288 case MD_LAYOUT_WRITE:
3290 * legal race for concurrent write, the file state has been
3291 * changed by another client. Or a jump over file size and
3294 som_buf->lb_buf = som;
3295 som_buf->lb_len = sizeof(*som);
3296 rc = mdo_xattr_get(env, obj, som_buf, XATTR_NAME_SOM);
3297 if (rc < 0 && rc != -ENODATA)
3301 lustre_som_swab(som);
3302 if (mlc->mlc_intent->lai_extent.e_end > som->lsa_size) {
3304 mlc->mlc_intent->lai_extent.e_end + 1;
3305 fl = LU_XATTR_REPLACE;
3313 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3318 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3319 XATTR_NAME_SOM, fl, handle);
3324 rc = mdd_trans_start(env, mdd, handle);
3328 /* it needs a sync tx to make FLR to work properly */
3329 handle->th_sync = 1;
3331 mdd_write_lock(env, obj, DT_TGT_CHILD);
3332 rc = mdo_layout_change(env, obj, mlc, handle);
3334 som->lsa_valid = SOM_FL_STALE;
3335 lustre_som_swab(som);
3336 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3339 mdd_write_unlock(env, obj);
3346 * Handle the requests when a FLR file's state is in SYNC_PENDING.
3348 * Only concurrent write and sync complete requests are possible when the
3349 * file is in SYNC_PENDING. For the latter request, it will pass in the
3350 * mirrors that have been synchronized, then the stale bit will be cleared
3351 * to make the file's state turn into RDONLY.
3352 * For concurrent write reqeust, it just needs to change the file's state
3353 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
3354 * version because the version will be increased in the transition to
3355 * SYNC_PENDING later so that it can deny the write request from potential
3356 * evicted SYNC clients.
3359 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
3360 struct md_layout_change *mlc, struct thandle *handle)
3362 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3363 struct lu_buf *som_buf = &mdd_env_info(env)->mdi_buf[1];
3369 /* operation validation */
3370 switch (mlc->mlc_opc) {
3371 case MD_LAYOUT_RESYNC_DONE:
3372 /* resync complete. */
3373 case MD_LAYOUT_WRITE:
3374 /* concurrent write. */
3376 case MD_LAYOUT_RESYNC:
3377 /* resync again, most likely the previous run failed.
3378 * no-op if it's already in SYNC_PENDING state
3385 if (mlc->mlc_som.lsa_valid & SOM_FL_STRICT) {
3386 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
3387 if (rc < 0 && rc != -ENODATA)
3390 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
3391 lustre_som_swab(&mlc->mlc_som);
3392 som_buf->lb_buf = &mlc->mlc_som;
3393 som_buf->lb_len = sizeof(mlc->mlc_som);
3396 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3400 /* record a changelog for the completion of resync */
3401 rc = mdd_declare_changelog_store(env, mdd, CL_RESYNC, NULL, NULL,
3406 /* RESYNC_DONE has piggybacked size and blocks */
3408 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
3409 XATTR_NAME_SOM, fl, handle);
3414 rc = mdd_trans_start(env, mdd, handle);
3418 /* it needs a sync tx to make FLR to work properly */
3419 handle->th_sync = 1;
3421 mdd_write_lock(env, obj, DT_TGT_CHILD);
3422 rc = mdo_layout_change(env, obj, mlc, handle);
3423 mdd_write_unlock(env, obj);
3428 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
3434 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle,
3440 /* Update the layout for PCC-RO. */
3442 mdd_layout_update_pccro(const struct lu_env *env, struct md_object *o,
3443 struct md_layout_change *mlc)
3445 struct mdd_object *obj = md2mdd_obj(o);
3446 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3447 struct thandle *handle;
3452 handle = mdd_trans_create(env, mdd);
3454 RETURN(PTR_ERR(handle));
3456 /* TODO: Set SOM strict correct when the file is PCC-RO cached. */
3457 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
3459 * It is possible that another layout write intent has already
3460 * set/cleared read-only flag on the object, so as to return
3461 * -EALREADY, and we need to do nothing in this case.
3464 GOTO(out, rc == -EALREADY ? rc = 0 : rc);
3466 rc = mdd_trans_start(env, mdd, handle);
3470 mdd_write_lock(env, obj, DT_TGT_CHILD);
3471 rc = mdo_layout_change(env, obj, mlc, handle);
3472 mdd_write_unlock(env, obj);
3476 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle,
3479 mdd_trans_stop(env, mdd, rc, handle);
3485 * Layout change callback for object.
3487 * This is only used by FLR and PCC-RO for now. In the future, it can be
3488 * exteneded to handle all layout change.
3491 mdd_layout_change(const struct lu_env *env, struct md_object *o,
3492 struct md_layout_change *mlc)
3494 struct mdd_object *obj = md2mdd_obj(o);
3495 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
3496 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
3497 struct lov_comp_md_v1 *lcm;
3498 struct thandle *handle;
3505 if (S_ISDIR(mdd_object_type(obj))) {
3506 switch (mlc->mlc_opc) {
3507 case MD_LAYOUT_SHRINK:
3508 rc = mdd_dir_layout_shrink(env, o, mlc);
3510 case MD_LAYOUT_SPLIT:
3511 rc = mdd_dir_layout_split(env, o, mlc);
3520 /* Verify acceptable operations */
3521 switch (mlc->mlc_opc) {
3522 case MD_LAYOUT_WRITE: {
3523 struct layout_intent *intent = mlc->mlc_intent;
3525 if (intent->lai_opc == LAYOUT_INTENT_PCCRO_SET ||
3526 intent->lai_opc == LAYOUT_INTENT_PCCRO_CLEAR)
3527 RETURN(mdd_layout_update_pccro(env, o, mlc));
3529 case MD_LAYOUT_RESYNC:
3530 case MD_LAYOUT_RESYNC_DONE:
3537 handle = mdd_trans_create(env, mdd);
3539 RETURN(PTR_ERR(handle));
3541 rc = mdd_stripe_get(env, obj, buf, XATTR_NAME_LOV);
3548 /* analyze the layout to make sure it's a FLR file */
3550 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
3551 GOTO(out, rc = -EINVAL);
3553 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
3555 /* please refer to HLD of FLR for state transition */
3556 switch (flr_state) {
3558 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
3560 case LCM_FL_WRITE_PENDING:
3561 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
3564 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
3566 case LCM_FL_SYNC_PENDING:
3567 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
3576 mdd_trans_stop(env, mdd, rc, handle);
3580 * -EAGAIN means transaction execution phase detect the layout
3581 * has been changed by others.
3583 if (rc == -EAGAIN && retried++ < MAX_TRANS_RETRIED)
3584 GOTO(retry, retried);
3589 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
3590 struct mdd_object *child, const struct lu_attr *attr,
3591 const struct md_op_spec *spec,
3592 struct dt_allocation_hint *hint)
3594 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
3595 struct mdd_device *mdd = mdd_obj2mdd_dev(child);
3596 struct dt_object *nc = mdd_object_child(child);
3598 memset(hint, 0, sizeof(*hint));
3600 /* For striped directory, give striping EA to lod_ah_init, which will
3601 * decide the stripe_offset and stripe count by it.
3603 if (S_ISDIR(attr->la_mode) && spec) {
3604 if (unlikely(spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
3605 hint->dah_eadata = spec->u.sp_ea.eadata;
3606 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
3607 if (spec->sp_cr_flags & MDS_OPEN_DEFAULT_LMV)
3608 hint->dah_eadata_is_dmv = 1;
3610 hint->dah_dmv_imp_inherit = spec->sp_dmv_imp_inherit;
3611 } else if (S_ISREG(attr->la_mode) &&
3612 spec->sp_cr_flags & MDS_OPEN_APPEND) {
3613 hint->dah_append_stripe_count = mdd->mdd_append_stripe_count;
3614 hint->dah_append_pool = mdd->mdd_append_pool;
3617 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
3618 hint->dah_eadata, hint->dah_eadata_len);
3619 /* @hint will be initialized by underlying device. */
3620 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
3623 static int mdd_accmode(const struct lu_env *env, const struct lu_attr *la,
3626 /* Sadly, NFSD reopens a file repeatedly during operation, so the
3627 * "acc_mode = 0" allowance for newly-created files isn't honoured.
3628 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
3629 * owner can write to a file even if it is marked readonly to hide
3630 * its brokenness. (bug 5781)
3632 if (open_flags & MDS_OPEN_OWNEROVERRIDE) {
3633 struct lu_ucred *uc = lu_ucred_check(env);
3635 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
3639 return mds_accmode(open_flags);
3642 static int mdd_open_sanity_check(const struct lu_env *env,
3643 struct mdd_object *obj,
3644 const struct lu_attr *attr, u64 open_flags,
3647 unsigned int may_mask;
3652 /* EEXIST check, also opening of *open* orphans is allowed so we can
3653 * open-by-handle unlinked files
3655 if (mdd_is_dead_obj(obj) && !is_replay &&
3656 likely(!(mdd_is_orphan_obj(obj) && obj->mod_count > 0)))
3659 if (S_ISLNK(attr->la_mode))
3662 may_mask = mdd_accmode(env, attr, open_flags);
3664 if (S_ISDIR(attr->la_mode) && (may_mask & MAY_WRITE))
3667 if (!(open_flags & MDS_OPEN_CREATED)) {
3668 rc = mdd_permission_internal(env, obj, attr, may_mask);
3673 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
3674 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
3675 open_flags &= ~MDS_OPEN_TRUNC;
3677 /* For writing append-only file must open it with append mode. */
3678 if (attr->la_flags & LUSTRE_APPEND_FL) {
3679 if ((open_flags & MDS_FMODE_WRITE) &&
3680 !(open_flags & MDS_OPEN_APPEND))
3682 if (open_flags & MDS_OPEN_TRUNC)
3689 static int mdd_open(const struct lu_env *env, struct md_object *obj,
3690 u64 open_flags, struct md_op_spec *spec)
3692 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3693 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
3694 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
3695 struct mdd_object_user *mou = NULL;
3696 const struct lu_ucred *uc = lu_ucred(env);
3697 struct mdd_device *mdd = mdo2mdd(obj);
3698 enum changelog_rec_type type = CL_OPEN;
3703 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3705 rc = mdd_la_get(env, mdd_obj, attr);
3709 rc = mdd_open_sanity_check(env, mdd_obj, attr, open_flags,
3711 if ((rc == -EACCES) && (mdd->mdd_cl.mc_current_mask & BIT(CL_DN_OPEN)))
3716 mdd_obj->mod_count++;
3718 if (!mdd_changelog_enabled(env, mdd, type))
3722 /* look for existing opener in list under mdd_write_lock */
3723 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid, open_flags);
3728 /* add user to list */
3729 mou = mdd_obj_user_alloc(open_flags, uc->uc_uid, uc->uc_gid);
3735 rc2 = mdd_obj_user_add(mdd_obj, mou, type == CL_DN_OPEN);
3737 mdd_obj_user_free(mou);
3742 if (type == CL_DN_OPEN) {
3743 if (ktime_before(ktime_get(), mou->mou_deniednext))
3744 /* same user denied again same access within
3745 * time interval: do not record
3749 /* this user already denied, but some time ago:
3750 * update denied time
3752 mou->mou_deniednext =
3753 ktime_add(ktime_get(),
3754 ktime_set(mdd->mdd_cl.mc_deniednext,
3757 mou->mou_opencount++;
3758 /* same user opening file again with same flags:
3765 /* we can't hold object lock over transaction start
3766 * and we don't actually need the object to be locked
3768 mdd_write_unlock(env, mdd_obj);
3770 /* FYI, only the bottom 32 bits of open_flags are recorded */
3771 /* only the low CLF_FLAGMASK bits of la_valid are stored*/
3772 mdd_changelog(env, type, open_flags & CLF_FLAGMASK, md_dev,
3773 mdd_object_fid(mdd_obj));
3778 mdd_write_unlock(env, mdd_obj);
3783 static int mdd_declare_close(const struct lu_env *env, struct mdd_object *obj,
3784 struct md_attr *ma, struct thandle *handle)
3788 rc = mdd_orphan_declare_delete(env, obj, handle);
3792 return mdo_declare_destroy(env, obj, handle);
3795 /* No permission check is needed. */
3796 static int mdd_close(const struct lu_env *env, struct md_object *obj,
3797 struct md_attr *ma, u64 open_flags)
3799 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3800 struct mdd_device *mdd = mdo2mdd(obj);
3801 struct thandle *handle = NULL;
3804 bool blocked = false;
3805 bool last_close_by_uid = false;
3806 const struct lu_ucred *uc = lu_ucred(env);
3810 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
3811 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3812 mdd_obj->mod_count--;
3813 mdd_write_unlock(env, mdd_obj);
3815 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
3816 CDEBUG(D_HA, "Object "DFID" is retained in orphan list\n",
3817 PFID(mdd_object_fid(mdd_obj)));
3821 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
3822 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ).
3824 /* check without any lock */
3825 is_orphan = mdd_obj->mod_count == 1 &&
3826 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
3830 /* mdd_trans_create() maybe failed because of barrier_entry(),
3831 * under such case, the orphan MDT-object will be left in the
3832 * orphan list, and when the MDT remount next time, the unused
3833 * orphans will be destroyed automatically.
3835 * One exception: the former mdd_finish_unlink may failed to
3836 * add the orphan MDT-object to the orphan list, then if the
3837 * mdd_trans_create() failed because of barrier_entry(), the
3838 * MDT-object will become real orphan that is neither in the
3839 * namespace nor in the orphan list. Such bad case should be
3840 * very rare and will be handled by e2fsck/lfsck.
3842 handle = mdd_trans_create(env, mdo2mdd(obj));
3843 if (IS_ERR(handle)) {
3844 rc = PTR_ERR(handle);
3845 if (rc != -EINPROGRESS)
3853 rc = mdd_declare_close(env, mdd_obj, ma, handle);
3857 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE, NULL, NULL,
3862 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3868 mdd_write_lock(env, mdd_obj, DT_TGT_CHILD);
3869 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
3871 CERROR("%s: failed to get lu_attr of "DFID": rc = %d\n",
3872 lu_dev_name(mdd2lu_dev(mdd)),
3873 PFID(mdd_object_fid(mdd_obj)), rc);
3877 /* check again with lock */
3878 is_orphan = (mdd_obj->mod_count == 1) &&
3879 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
3880 ma->ma_attr.la_nlink == 0);
3882 if (is_orphan && !handle && !blocked) {
3883 mdd_write_unlock(env, mdd_obj);
3887 mdd_obj->mod_count--; /*release open count */
3889 /* under mdd write lock */
3890 /* If recording, see if we need to remove UID from list. uc is not
3891 * initialized if the client has been evicted.
3893 if (mdd_changelog_enabled(env, mdd, CL_OPEN) && uc) {
3894 struct mdd_object_user *mou;
3896 /* look for UID in list */
3897 /* If mou is NULL, it probably means logging was enabled after
3898 * the user had the file open. So the corresponding close
3899 * will not be logged.
3901 mou = mdd_obj_user_find(mdd_obj, uc->uc_uid, uc->uc_gid,
3904 mou->mou_opencount--;
3905 if (mou->mou_opencount == 0) {
3906 mdd_obj_user_remove(mdd_obj, mou);
3907 last_close_by_uid = true;
3912 if (!is_orphan || blocked)
3916 /* NB: Object maybe not in orphan list originally, it is rare case for
3917 * mdd_finish_unlink() failure, in that case, the object doesn't have
3920 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
3921 /* remove link to object from orphan index */
3922 LASSERT(handle != NULL);
3923 rc = mdd_orphan_delete(env, mdd_obj, handle);
3925 CERROR("%s: unable to delete "DFID" from orphan list: rc = %d\n",
3926 lu_dev_name(mdd2lu_dev(mdd)),
3927 PFID(mdd_object_fid(mdd_obj)), rc);
3928 /* If object was not deleted from orphan list, do not
3929 * destroy OSS objects, which will be done when next
3935 CDEBUG(D_HA, "Object "DFID" is deleted from orphan list, OSS objects to be destroyed.\n",
3936 PFID(mdd_object_fid(mdd_obj)));
3939 rc = mdo_destroy(env, mdd_obj, handle);
3942 CERROR("%s: unable to delete "DFID" from orphan list: rc = %d\n",
3943 lu_dev_name(mdd2lu_dev(mdd)),
3944 PFID(mdd_object_fid(mdd_obj)), rc);
3949 mdd_write_unlock(env, mdd_obj);
3951 if (rc != 0 || blocked ||
3952 !mdd_changelog_enabled(env, mdd, CL_CLOSE))
3955 /* Record CL_CLOSE in changelog only if file was opened in write mode,
3956 * or if CL_OPEN was recorded and it's last close by user.
3957 * Changelogs mask may change between open and close operations, but
3958 * this is not a big deal if we have a CL_CLOSE entry with no matching
3959 * CL_OPEN. Plus Changelogs mask may not change often.
3961 if (((!(mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3962 (open_flags & (MDS_FMODE_WRITE | MDS_OPEN_APPEND |
3963 MDS_OPEN_TRUNC))) ||
3964 ((mdd->mdd_cl.mc_current_mask & BIT(CL_OPEN)) &&
3965 last_close_by_uid)) &&
3966 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
3967 if (handle == NULL) {
3968 handle = mdd_trans_create(env, mdo2mdd(obj));
3970 GOTO(stop, rc = PTR_ERR(handle));
3972 rc = mdd_declare_changelog_store(env, mdd, CL_CLOSE,
3973 NULL, NULL, handle);
3977 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
3982 /* FYI, only the bottom 32 bits of open_flags are recorded */
3983 mdd_changelog_data_store(env, mdd, CL_CLOSE, open_flags,
3984 mdd_obj, handle, NULL);
3988 if (handle != NULL && !IS_ERR(handle))
3989 rc = mdd_trans_stop(env, mdd, rc, handle);
3994 /* Permission check is done when open, no need check again. */
3995 static int mdd_readpage_sanity_check(const struct lu_env *env,
3996 struct mdd_object *obj)
3998 if (!dt_try_as_dir(env, mdd_object_child(obj), true))
4004 static int mdd_dir_page_build(const struct lu_env *env, struct dt_object *obj,
4005 union lu_page *lp, size_t bytes,
4006 const struct dt_it_ops *iops,
4007 struct dt_it *it, __u32 attr, void *arg)
4009 struct lu_dirpage *dp = &lp->lp_dir;
4012 struct lu_dirent *ent;
4013 struct lu_dirent *last = NULL;
4018 if (bytes < sizeof(*dp))
4019 GOTO(out_err, result = -EOVERFLOW);
4021 memset(area, 0, sizeof(*dp));
4022 area += sizeof(*dp);
4023 bytes -= sizeof(*dp);
4030 len = iops->key_size(env, it);
4032 /* IAM iterator can return record with zero len. */
4036 hash = iops->store(env, it);
4037 if (unlikely(first)) {
4039 dp->ldp_hash_start = cpu_to_le64(hash);
4042 /* calculate max space required for lu_dirent */
4043 recsize = lu_dirent_calc_size(len, attr);
4045 if (bytes >= recsize &&
4046 !CFS_FAIL_CHECK(OBD_FAIL_MDS_DIR_PAGE_WALK)) {
4047 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
4048 if (result == -ESTALE)
4053 /* OSD might not able to pack all attributes, so
4054 * recheck record length had room to store FID
4056 recsize = le16_to_cpu(ent->lde_reclen);
4058 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
4059 fid_le_to_cpu(&fid, &ent->lde_fid);
4060 if (fid_is_dot_lustre(&fid))
4061 GOTO(next, recsize);
4064 result = (last != NULL) ? 0 : -EBADSLT;
4068 ent = (void *)ent + recsize;
4072 result = iops->next(env, it);
4073 if (result == -ESTALE)
4075 } while (result == 0);
4078 dp->ldp_hash_end = cpu_to_le64(hash);
4080 if (last->lde_hash == dp->ldp_hash_end)
4081 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
4082 last->lde_reclen = 0; /* end mark */
4086 /* end of directory */
4087 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
4088 else if (result < 0)
4089 CWARN("%s: build page failed for "DFID": rc = %d\n",
4090 lu_dev_name(obj->do_lu.lo_dev),
4091 PFID(lu_object_fid(&obj->do_lu)), result);
4096 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
4097 const struct lu_rdpg *rdpg)
4099 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4104 if (mdd_object_exists(mdd_obj) == 0) {
4106 CERROR("%s: object "DFID" not found: rc = %d\n",
4107 mdd_obj_dev_name(mdd_obj),
4108 PFID(mdd_object_fid(mdd_obj)), rc);
4112 mdd_read_lock(env, mdd_obj, DT_TGT_CHILD);
4113 rc = mdd_readpage_sanity_check(env, mdd_obj);
4115 GOTO(out_unlock, rc);
4117 if (mdd_is_dead_obj(mdd_obj)) {
4119 struct lu_dirpage *dp;
4122 * According to POSIX, please do not return any entry to client:
4123 * even dot and dotdot should not be returned.
4125 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
4126 PFID(mdd_object_fid(mdd_obj)));
4128 if (rdpg->rp_count <= 0)
4129 GOTO(out_unlock, rc = -EFAULT);
4130 LASSERT(rdpg->rp_pages != NULL);
4132 pg = rdpg->rp_pages[0];
4133 dp = (struct lu_dirpage *)kmap(pg);
4134 memset(dp, 0, sizeof(struct lu_dirpage));
4135 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
4136 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
4137 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
4139 GOTO(out_unlock, rc = LU_PAGE_SIZE);
4142 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
4143 mdd_dir_page_build, NULL);
4145 struct lu_dirpage *dp;
4147 dp = kmap(rdpg->rp_pages[0]);
4148 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
4151 * No pages were processed, mark this for first page
4154 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
4155 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
4156 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
4158 kunmap(rdpg->rp_pages[0]);
4161 GOTO(out_unlock, rc);
4163 mdd_read_unlock(env, mdd_obj);
4167 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
4169 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4171 if (mdd_object_exists(mdd_obj) == 0) {
4174 CERROR("%s: object "DFID" not found: rc = %d\n",
4175 mdd_obj_dev_name(mdd_obj),
4176 PFID(mdd_object_fid(mdd_obj)), rc);
4179 return dt_object_sync(env, mdd_object_child(mdd_obj),
4183 static int mdd_object_lock(const struct lu_env *env,
4184 struct md_object *obj,
4185 struct lustre_handle *lh,
4186 struct ldlm_enqueue_info *einfo,
4187 union ldlm_policy_data *policy)
4189 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4191 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
4195 static int mdd_object_unlock(const struct lu_env *env,
4196 struct md_object *obj,
4197 struct ldlm_enqueue_info *einfo,
4198 union ldlm_policy_data *policy)
4200 struct mdd_object *mdd_obj = md2mdd_obj(obj);
4202 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
4205 const struct md_object_operations mdd_obj_ops = {
4206 .moo_permission = mdd_permission,
4207 .moo_attr_get = mdd_attr_get,
4208 .moo_attr_set = mdd_attr_set,
4209 .moo_xattr_get = mdd_xattr_get,
4210 .moo_xattr_set = mdd_xattr_set,
4211 .moo_xattr_list = mdd_xattr_list,
4212 .moo_invalidate = mdd_invalidate,
4213 .moo_xattr_del = mdd_xattr_del,
4214 .moo_swap_layouts = mdd_swap_layouts,
4215 .moo_open = mdd_open,
4216 .moo_close = mdd_close,
4217 .moo_readpage = mdd_readpage,
4218 .moo_readlink = mdd_readlink,
4219 .moo_changelog = mdd_changelog,
4220 .moo_object_sync = mdd_object_sync,
4221 .moo_object_lock = mdd_object_lock,
4222 .moo_object_unlock = mdd_object_unlock,
4223 .moo_layout_change = mdd_layout_change,