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, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <lustre_mds.h>
52 #include "mdd_internal.h"
54 static const struct lu_object_operations mdd_lu_obj_ops;
56 static int mdd_xattr_get(const struct lu_env *env,
57 struct md_object *obj, struct lu_buf *buf,
60 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
65 if (mdd_object_exists(obj) == 0)
68 rc = mdo_attr_get(env, obj, la);
69 if (unlikely(rc != 0)) {
71 obj->mod_flags |= DEAD_OBJ;
75 if (la->la_valid & LA_FLAGS &&
76 la->la_flags & LUSTRE_ORPHAN_FL)
77 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
82 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
84 struct mdd_thread_info *info;
86 lu_env_refill((struct lu_env *)env);
87 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
88 LASSERT(info != NULL);
92 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
96 buf = &mdd_env_info(env)->mti_buf[0];
102 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
103 const void *area, ssize_t len)
107 buf = &mdd_env_info(env)->mti_buf[0];
108 buf->lb_buf = (void *)area;
113 struct lu_object *mdd_object_alloc(const struct lu_env *env,
114 const struct lu_object_header *hdr,
117 struct mdd_object *mdd_obj;
119 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
120 if (mdd_obj != NULL) {
123 o = mdd2lu_obj(mdd_obj);
124 lu_object_init(o, NULL, d);
125 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
126 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
127 mdd_obj->mod_count = 0;
128 o->lo_ops = &mdd_lu_obj_ops;
135 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
136 const struct lu_object_conf *unused)
138 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
139 struct mdd_object *mdd_obj = lu2mdd_obj(o);
140 struct lu_object *below;
141 struct lu_device *under;
144 mdd_obj->mod_cltime = ktime_set(0, 0);
145 under = &d->mdd_child->dd_lu_dev;
146 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
148 RETURN(PTR_ERR(below));
150 lu_object_add(o, below);
155 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
159 if (lu_object_exists(o)) {
160 struct mdd_object *mdd_obj = lu2mdd_obj(o);
161 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
163 rc = mdd_la_get(env, mdd_obj, attr);
169 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
171 struct mdd_object *mdd = lu2mdd_obj(o);
174 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
177 static int mdd_object_print(const struct lu_env *env, void *cookie,
178 lu_printer_t p, const struct lu_object *o)
180 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
182 return (*p)(env, cookie,
183 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
184 mdd, mdd->mod_count, mdd->mod_valid,
185 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
188 static const struct lu_object_operations mdd_lu_obj_ops = {
189 .loo_object_init = mdd_object_init,
190 .loo_object_start = mdd_object_start,
191 .loo_object_free = mdd_object_free,
192 .loo_object_print = mdd_object_print,
195 struct mdd_object *mdd_object_find(const struct lu_env *env,
196 struct mdd_device *d,
197 const struct lu_fid *f)
199 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
203 * No permission check is needed.
205 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
208 struct mdd_object *mdd_obj = md2mdd_obj(obj);
213 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
214 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
215 ma->ma_attr.la_nlink = 0;
221 * No permission check is needed.
223 static int mdd_xattr_get(const struct lu_env *env,
224 struct md_object *obj, struct lu_buf *buf,
227 struct mdd_object *mdd_obj = md2mdd_obj(obj);
232 if (mdd_object_exists(mdd_obj) == 0) {
233 CERROR("%s: object "DFID" not found: rc = -2\n",
234 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
238 /* If the object has been destroyed, then do not get LMVEA, because
239 * it needs to load stripes from the iteration of the master object,
240 * and it will cause problem if master object has been destroyed, see
242 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
243 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
244 strcmp(name, XATTR_NAME_LMV) == 0))
247 /* If the object has been delete from the namespace, then
248 * get linkEA should return -ENOENT as well */
249 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
250 strcmp(name, XATTR_NAME_LINK) == 0))
253 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
254 rc = mdo_xattr_get(env, mdd_obj, buf, name);
255 mdd_read_unlock(env, mdd_obj);
261 * Permission check is done when open,
262 * no need check again.
264 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
267 struct mdd_object *mdd_obj = md2mdd_obj(obj);
268 struct dt_object *next;
273 if (mdd_object_exists(mdd_obj) == 0) {
274 CERROR("%s: object "DFID" not found: rc = -2\n",
275 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
279 next = mdd_object_child(mdd_obj);
280 LASSERT(next != NULL);
281 LASSERT(next->do_body_ops != NULL);
282 LASSERT(next->do_body_ops->dbo_read != NULL);
283 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
284 rc = dt_read(env, next, buf, &pos);
285 mdd_read_unlock(env, mdd_obj);
290 * No permission check is needed.
292 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
295 struct mdd_object *mdd_obj = md2mdd_obj(obj);
300 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
301 rc = mdo_xattr_list(env, mdd_obj, buf);
302 mdd_read_unlock(env, mdd_obj);
304 /* If the buffer is NULL then we are only here to get the
305 * length of the xattr name list. */
306 if (rc < 0 || buf->lb_buf == NULL)
310 * Filter out XATTR_NAME_LINK if this is an orphan object. See
313 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
314 char *end = (char *)buf->lb_buf + rc;
315 char *p = buf->lb_buf;
318 char *next = p + strlen(p) + 1;
320 if (strcmp(p, XATTR_NAME_LINK) == 0) {
322 memmove(p, next, end - next);
324 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
325 " of orphan "DFID"\n",
326 PFID(mdd_object_fid(mdd_obj)));
337 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
339 return mdo_invalidate(env, md2mdd_obj(obj));
342 int mdd_declare_create_object_internal(const struct lu_env *env,
343 struct mdd_object *p,
344 struct mdd_object *c,
345 struct lu_attr *attr,
346 struct thandle *handle,
347 const struct md_op_spec *spec,
348 struct dt_allocation_hint *hint)
350 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
351 const struct dt_index_features *feat = spec->sp_feat;
355 if (feat != &dt_directory_features && feat != NULL) {
356 dof->dof_type = DFT_INDEX;
357 dof->u.dof_idx.di_feat = feat;
359 dof->dof_type = dt_mode_to_dft(attr->la_mode);
360 if (dof->dof_type == DFT_REGULAR) {
361 dof->u.dof_reg.striped =
362 md_should_create(spec->sp_cr_flags);
363 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
364 dof->u.dof_reg.striped = 0;
365 /* is this replay? */
367 dof->u.dof_reg.striped = 0;
371 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
376 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
377 struct mdd_object *c, struct lu_attr *attr,
378 struct thandle *handle,
379 const struct md_op_spec *spec,
380 struct dt_allocation_hint *hint)
382 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
386 LASSERT(!mdd_object_exists(c));
388 rc = mdo_create_object(env, c, attr, hint, dof, handle);
393 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
394 const struct lu_attr *attr, struct thandle *handle,
400 rc = mdo_attr_set(env, obj, attr, handle);
401 #ifdef CONFIG_FS_POSIX_ACL
402 if (!rc && (attr->la_valid & LA_MODE) && needacl)
403 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
408 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
409 const struct lu_attr *oattr, struct lu_attr *attr,
410 struct thandle *handle)
415 LASSERT(attr->la_valid & LA_CTIME);
416 LASSERT(oattr != NULL);
418 /* Make sure the ctime is increased only, however, it's not strictly
419 * reliable at here because there is not guarantee to hold lock on
420 * object, so we just bypass some unnecessary cmtime setting first
421 * and OSD has to check it again. */
422 if (attr->la_ctime < oattr->la_ctime)
423 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
424 else if (attr->la_valid == LA_CTIME &&
425 attr->la_ctime == oattr->la_ctime)
426 attr->la_valid &= ~LA_CTIME;
428 if (attr->la_valid != 0)
429 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
434 * This gives the same functionality as the code between
435 * sys_chmod and inode_setattr
436 * chown_common and inode_setattr
437 * utimes and inode_setattr
438 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
440 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
441 const struct lu_attr *oattr, struct lu_attr *la,
442 const unsigned long flags)
451 /* Do not permit change file type */
452 if (la->la_valid & LA_TYPE)
455 /* They should not be processed by setattr */
456 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
459 LASSERT(oattr != NULL);
461 uc = lu_ucred_check(env);
465 if (la->la_valid == LA_CTIME) {
466 if (!(flags & MDS_PERM_BYPASS))
467 /* This is only for set ctime when rename's source is
469 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
471 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
472 la->la_valid &= ~LA_CTIME;
476 if (la->la_valid == LA_ATIME) {
477 /* This is an atime-only attribute update for close RPCs. */
478 if (la->la_atime < (oattr->la_atime +
479 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
480 la->la_valid &= ~LA_ATIME;
484 /* Check if flags change. */
485 if (la->la_valid & LA_FLAGS) {
486 unsigned int oldflags = oattr->la_flags &
487 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
488 unsigned int newflags = la->la_flags &
489 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
491 if ((uc->uc_fsuid != oattr->la_uid) &&
492 !md_capable(uc, CFS_CAP_FOWNER))
495 /* The IMMUTABLE and APPEND_ONLY flags can
496 * only be changed by the relevant capability. */
497 if ((oldflags ^ newflags) &&
498 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
501 if (!S_ISDIR(oattr->la_mode))
502 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
505 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
506 (la->la_valid & ~LA_FLAGS) &&
507 !(flags & MDS_PERM_BYPASS))
510 /* Check for setting the obj time. */
511 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
512 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
513 if ((uc->uc_fsuid != oattr->la_uid) &&
514 !md_capable(uc, CFS_CAP_FOWNER)) {
515 rc = mdd_permission_internal(env, obj, oattr,
522 if (la->la_valid & LA_KILL_SUID) {
523 la->la_valid &= ~LA_KILL_SUID;
524 if ((oattr->la_mode & S_ISUID) &&
525 !(la->la_valid & LA_MODE)) {
526 la->la_mode = oattr->la_mode;
527 la->la_valid |= LA_MODE;
529 la->la_mode &= ~S_ISUID;
532 if (la->la_valid & LA_KILL_SGID) {
533 la->la_valid &= ~LA_KILL_SGID;
534 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
535 (S_ISGID | S_IXGRP)) &&
536 !(la->la_valid & LA_MODE)) {
537 la->la_mode = oattr->la_mode;
538 la->la_valid |= LA_MODE;
540 la->la_mode &= ~S_ISGID;
543 /* Make sure a caller can chmod. */
544 if (la->la_valid & LA_MODE) {
545 if (!(flags & MDS_PERM_BYPASS) &&
546 (uc->uc_fsuid != oattr->la_uid) &&
547 !md_capable(uc, CFS_CAP_FOWNER))
550 if (la->la_mode == (umode_t) -1)
551 la->la_mode = oattr->la_mode;
553 la->la_mode = (la->la_mode & S_IALLUGO) |
554 (oattr->la_mode & ~S_IALLUGO);
556 /* Also check the setgid bit! */
557 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
558 la->la_gid : oattr->la_gid) &&
559 !md_capable(uc, CFS_CAP_FSETID))
560 la->la_mode &= ~S_ISGID;
562 la->la_mode = oattr->la_mode;
565 /* Make sure a caller can chown. */
566 if (la->la_valid & LA_UID) {
567 if (la->la_uid == (uid_t) -1)
568 la->la_uid = oattr->la_uid;
569 if (((uc->uc_fsuid != oattr->la_uid) ||
570 (la->la_uid != oattr->la_uid)) &&
571 !md_capable(uc, CFS_CAP_CHOWN))
574 /* If the user or group of a non-directory has been
575 * changed by a non-root user, remove the setuid bit.
576 * 19981026 David C Niemi <niemi@tux.org>
578 * Changed this to apply to all users, including root,
579 * to avoid some races. This is the behavior we had in
580 * 2.0. The check for non-root was definitely wrong
581 * for 2.2 anyway, as it should have been using
582 * CAP_FSETID rather than fsuid -- 19990830 SD. */
583 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
584 !S_ISDIR(oattr->la_mode)) {
585 la->la_mode &= ~S_ISUID;
586 la->la_valid |= LA_MODE;
590 /* Make sure caller can chgrp. */
591 if (la->la_valid & LA_GID) {
592 if (la->la_gid == (gid_t) -1)
593 la->la_gid = oattr->la_gid;
594 if (((uc->uc_fsuid != oattr->la_uid) ||
595 ((la->la_gid != oattr->la_gid) &&
596 !lustre_in_group_p(uc, la->la_gid))) &&
597 !md_capable(uc, CFS_CAP_CHOWN))
600 /* Likewise, if the user or group of a non-directory
601 * has been changed by a non-root user, remove the
602 * setgid bit UNLESS there is no group execute bit
603 * (this would be a file marked for mandatory
604 * locking). 19981026 David C Niemi <niemi@tux.org>
606 * Removed the fsuid check (see the comment above) --
608 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
609 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
610 la->la_mode &= ~S_ISGID;
611 la->la_valid |= LA_MODE;
615 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
616 if (!((flags & MDS_OWNEROVERRIDE) &&
617 (uc->uc_fsuid == oattr->la_uid)) &&
618 !(flags & MDS_PERM_BYPASS)) {
619 rc = mdd_permission_internal(env, obj, oattr,
626 if (la->la_valid & LA_CTIME) {
627 /* The pure setattr, it has the priority over what is
628 * already set, do not drop it if ctime is equal. */
629 if (la->la_ctime < oattr->la_ctime)
630 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
636 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
637 struct mdd_device *mdd,
638 enum changelog_rec_type type, int flags,
639 const struct lu_fid *fid,
640 struct thandle *handle)
642 const struct lu_ucred *uc = lu_ucred(env);
643 struct llog_changelog_rec *rec;
648 flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
649 if (uc != NULL && uc->uc_jobid[0] != '\0')
652 reclen = llog_data_len(changelog_rec_offset(flags & CLF_SUPPORTED));
653 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
654 if (buf->lb_buf == NULL)
658 rec->cr.cr_flags = flags;
659 rec->cr.cr_type = (__u32)type;
660 rec->cr.cr_tfid = *fid;
661 rec->cr.cr_namelen = 0;
663 if (flags & CLF_JOBID)
664 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
666 rc = mdd_changelog_store(env, mdd, rec, handle);
672 /** Store a data change changelog record
673 * If this fails, we must fail the whole transaction; we don't
674 * want the change to commit without the log entry.
675 * \param mdd_obj - mdd_object of change
676 * \param handle - transaction handle
678 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
679 enum changelog_rec_type type, int flags,
680 struct mdd_object *mdd_obj, struct thandle *handle)
684 LASSERT(mdd_obj != NULL);
685 LASSERT(handle != NULL);
688 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
690 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
693 if (mdd_is_volatile_obj(mdd_obj))
696 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
697 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
698 /* Don't need multiple updates in this log */
699 /* Don't check under lock - no big deal if we get an extra
704 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
705 mdo2fid(mdd_obj), handle);
707 mdd_obj->mod_cltime = ktime_get();
712 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
713 int flags, struct md_device *m, const struct lu_fid *fid)
715 struct thandle *handle;
716 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
721 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
723 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
726 LASSERT(fid != NULL);
728 handle = mdd_trans_create(env, mdd);
730 RETURN(PTR_ERR(handle));
732 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
736 rc = mdd_trans_start(env, mdd, handle);
740 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
744 rc = mdd_trans_stop(env, mdd, rc, handle);
750 * Save LMA extended attributes with data from \a ma.
752 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
753 * not, LMA EA will be first read from disk, modified and write back.
756 /* Precedence for choosing record type when multiple
757 * attributes change: setattr > mtime > ctime > atime
758 * (ctime changes when mtime does, plus chmod/chown.
759 * atime and ctime are independent.) */
760 static int mdd_attr_set_changelog(const struct lu_env *env,
761 struct md_object *obj, struct thandle *handle,
764 struct mdd_device *mdd = mdo2mdd(obj);
767 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
768 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
769 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
770 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
771 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
772 bits = bits & mdd->mdd_cl.mc_mask;
773 /* This is an implementation limit rather than a protocol limit */
774 CLASSERT(CL_LAST <= sizeof(int) * 8);
778 /* The record type is the lowest non-masked set bit */
781 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
782 return mdd_changelog_data_store(env, mdd, type, (int)valid,
783 md2mdd_obj(obj), handle);
786 static int mdd_declare_attr_set(const struct lu_env *env,
787 struct mdd_device *mdd,
788 struct mdd_object *obj,
789 const struct lu_attr *attr,
790 struct thandle *handle)
794 rc = mdo_declare_attr_set(env, obj, attr, handle);
798 #ifdef CONFIG_FS_POSIX_ACL
799 if (attr->la_valid & LA_MODE) {
800 mdd_read_lock(env, obj, MOR_TGT_CHILD);
801 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
802 XATTR_NAME_ACL_ACCESS);
803 mdd_read_unlock(env, obj);
804 if (rc == -EOPNOTSUPP || rc == -ENODATA)
810 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
811 rc = mdo_declare_xattr_set(env, obj, buf,
812 XATTR_NAME_ACL_ACCESS, 0,
820 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
827 * permission changes may require sync operation, to mitigate performance
828 * impact, only do this for dir and when permission is reduced.
830 * For regular files, version is updated with permission change (see VBR), async
831 * permission won't cause any issue, while missing permission change on
832 * directory may affect accessibility of other objects after recovery.
834 static inline bool permission_needs_sync(const struct lu_attr *old,
835 const struct lu_attr *new)
837 if (!S_ISDIR(old->la_mode))
840 if (new->la_valid & (LA_UID | LA_GID))
843 if (new->la_valid & LA_MODE &&
844 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
847 if ((new->la_valid & LA_MODE) &&
848 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
849 (old->la_mode & S_IRWXUGO))
855 /* set attr and LOV EA at once, return updated attr */
856 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
857 const struct md_attr *ma)
859 struct mdd_object *mdd_obj = md2mdd_obj(obj);
860 struct mdd_device *mdd = mdo2mdd(obj);
861 struct thandle *handle;
862 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
863 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
864 const struct lu_attr *la = &ma->ma_attr;
868 /* we do not use ->attr_set() for LOV/HSM EA any more */
869 LASSERT((ma->ma_valid & MA_LOV) == 0);
870 LASSERT((ma->ma_valid & MA_HSM) == 0);
872 rc = mdd_la_get(env, mdd_obj, attr);
876 *la_copy = ma->ma_attr;
877 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
881 /* no need to setattr anymore */
882 if (la_copy->la_valid == 0) {
883 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
884 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
885 PFID(mdo2fid(mdd_obj)), la->la_valid);
890 handle = mdd_trans_create(env, mdd);
892 RETURN(PTR_ERR(handle));
894 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
898 rc = mdd_trans_start(env, mdd, handle);
902 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
905 if (la->la_valid & (LA_MTIME | LA_CTIME))
906 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
907 la->la_mtime, la->la_ctime);
909 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
910 if (la_copy->la_valid & LA_FLAGS)
911 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
912 else if (la_copy->la_valid) /* setattr */
913 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
914 mdd_write_unlock(env, mdd_obj);
917 rc = mdd_attr_set_changelog(env, obj, handle,
923 rc = mdd_trans_stop(env, mdd, rc, handle);
928 static int mdd_xattr_sanity_check(const struct lu_env *env,
929 struct mdd_object *obj,
930 const struct lu_attr *attr,
933 struct lu_ucred *uc = lu_ucred_assert(env);
936 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
939 if (strncmp(XATTR_USER_PREFIX, name,
940 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
941 /* For sticky directories, only the owner and privileged user
942 * can write attributes. */
943 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
944 (uc->uc_fsuid != attr->la_uid) &&
945 !md_capable(uc, CFS_CAP_FOWNER))
948 if ((uc->uc_fsuid != attr->la_uid) &&
949 !md_capable(uc, CFS_CAP_FOWNER))
957 * Check if a string begins with a given prefix.
959 * \param str String to check
960 * \param prefix Substring to check at the beginning of \a str
961 * \return true/false whether the condition is verified.
963 static inline bool has_prefix(const char *str, const char *prefix)
965 return strncmp(prefix, str, strlen(prefix)) == 0;
969 * Indicate the kind of changelog to store (if any) for a xattr set/del.
971 * \param[in] xattr_name Full extended attribute name.
973 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
975 static enum changelog_rec_type
976 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
977 const char *xattr_name)
979 /* Layout changes systematically recorded */
980 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
981 strncmp(XATTR_LUSTRE_LOV, xattr_name,
982 strlen(XATTR_LUSTRE_LOV)) == 0)
985 /* HSM information changes systematically recorded */
986 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
989 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
990 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
991 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
997 static int mdd_declare_xattr_set(const struct lu_env *env,
998 struct mdd_device *mdd,
999 struct mdd_object *obj,
1000 const struct lu_buf *buf,
1002 int fl, struct thandle *handle)
1006 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1010 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1011 return 0; /* no changelog to store */
1013 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1017 * Compare current and future data of HSM EA and add a changelog if needed.
1019 * Caller should have write-locked \param obj.
1021 * \param buf - Future HSM EA content.
1022 * \retval 0 if no changelog is needed or changelog was added properly.
1023 * \retval -ve errno if there was a problem
1025 static int mdd_hsm_update_locked(const struct lu_env *env,
1026 struct md_object *obj,
1027 const struct lu_buf *buf,
1028 struct thandle *handle, int *cl_flags)
1030 struct mdd_thread_info *info = mdd_env_info(env);
1031 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1032 struct lu_buf *current_buf;
1033 struct md_hsm *current_mh;
1034 struct md_hsm *new_mh;
1038 OBD_ALLOC_PTR(current_mh);
1039 if (current_mh == NULL)
1042 /* Read HSM attrs from disk */
1043 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1044 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1045 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1046 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1047 if (rc < 0 && rc != -ENODATA)
1049 else if (rc == -ENODATA)
1050 current_mh->mh_flags = 0;
1052 /* Map future HSM xattr */
1053 OBD_ALLOC_PTR(new_mh);
1055 GOTO(free, rc = -ENOMEM);
1056 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1060 /* Flags differ, set flags for the changelog that will be added */
1061 if (current_mh->mh_flags != new_mh->mh_flags) {
1062 hsm_set_cl_event(cl_flags, HE_STATE);
1063 if (new_mh->mh_flags & HS_DIRTY)
1064 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1067 OBD_FREE_PTR(new_mh);
1070 OBD_FREE_PTR(current_mh);
1074 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1078 * The caller should guarantee to update the object ctime
1079 * after xattr_set if needed.
1081 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1082 const struct lu_buf *buf, const char *name,
1085 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1086 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1087 struct mdd_device *mdd = mdo2mdd(obj);
1088 struct thandle *handle;
1089 enum changelog_rec_type cl_type;
1094 rc = mdd_la_get(env, mdd_obj, attr);
1098 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1102 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1103 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1104 struct posix_acl *acl;
1106 /* user may set empty ACL, which should be treated as removing
1108 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1111 RETURN(PTR_ERR(acl));
1113 rc = mdd_xattr_del(env, obj, name);
1116 posix_acl_release(acl);
1119 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1120 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1124 handle = mdd_trans_create(env, mdd);
1126 RETURN(PTR_ERR(handle));
1128 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1132 rc = mdd_trans_start(env, mdd, handle);
1136 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1138 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1139 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1141 mdd_write_unlock(env, mdd_obj);
1146 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1147 mdd_write_unlock(env, mdd_obj);
1151 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1155 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1160 return mdd_trans_stop(env, mdd, rc, handle);
1163 static int mdd_declare_xattr_del(const struct lu_env *env,
1164 struct mdd_device *mdd,
1165 struct mdd_object *obj,
1167 struct thandle *handle)
1171 rc = mdo_declare_xattr_del(env, obj, name, handle);
1175 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1176 return 0; /* no changelog to store */
1178 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1182 * The caller should guarantee to update the object ctime
1183 * after xattr_set if needed.
1185 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1188 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1189 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1190 struct mdd_device *mdd = mdo2mdd(obj);
1191 struct thandle *handle;
1195 rc = mdd_la_get(env, mdd_obj, attr);
1199 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1203 handle = mdd_trans_create(env, mdd);
1205 RETURN(PTR_ERR(handle));
1207 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1211 rc = mdd_trans_start(env, mdd, handle);
1215 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1216 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1217 mdd_write_unlock(env, mdd_obj);
1221 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1224 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1228 return mdd_trans_stop(env, mdd, rc, handle);
1232 * read lov EA of an object
1233 * return the lov EA in an allocated lu_buf
1235 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1236 struct lu_buf *lmm_buf)
1238 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1243 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1245 if (rc == -ERANGE) {
1246 /* mti_big_buf is allocated but is too small
1247 * we need to increase it */
1248 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1250 if (buf->lb_buf == NULL)
1251 GOTO(out, rc = -ENOMEM);
1262 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1263 /* mti_big_buf was not allocated, so we have to
1264 * allocate it based on the ea size */
1265 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1267 if (buf->lb_buf == NULL)
1268 GOTO(out, rc = -ENOMEM);
1272 lu_buf_alloc(lmm_buf, bufsize);
1273 if (lmm_buf->lb_buf == NULL)
1274 GOTO(out, rc = -ENOMEM);
1276 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1282 lu_buf_free(lmm_buf);
1286 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1287 struct mdd_object *o, struct lu_buf *buf,
1288 struct thandle *handle)
1290 struct hsm_attrs *attrs;
1296 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1301 attrs = buf->lb_buf;
1302 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1303 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1306 /* Add a changelog record for release. */
1307 hsm_set_cl_event(&flags, HE_RELEASE);
1308 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1314 * check if layout swapping between 2 objects is allowed
1316 * - only normal FIDs or non-system IGIFs
1317 * - same type of objects
1318 * - same owner/group (so quotas are still valid) unless this is from HSM
1321 static int mdd_layout_swap_allowed(const struct lu_env *env,
1322 struct mdd_object *o1,
1323 const struct lu_attr *attr1,
1324 struct mdd_object *o2,
1325 const struct lu_attr *attr2,
1328 const struct lu_fid *fid1, *fid2;
1334 if (!fid_is_norm(fid1) &&
1335 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1338 if (!fid_is_norm(fid2) &&
1339 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1342 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1343 if (S_ISDIR(mdd_object_type(o1)))
1345 if (S_ISREG(mdd_object_type(o1)))
1350 if (flags & SWAP_LAYOUTS_MDS_HSM)
1353 if ((attr1->la_uid != attr2->la_uid) ||
1354 (attr1->la_gid != attr2->la_gid))
1360 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1361 * look into the layout in MDD layer. */
1362 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1364 struct lov_comp_md_v1 *comp_v1;
1365 struct lov_mds_md *v1;
1369 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1370 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1371 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1377 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1378 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1381 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1382 off = le32_to_cpu(comp_v1->lcm_entries[i].
1384 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1389 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1390 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1401 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1403 return mdd_lmm_oi(lmm, oi, true);
1406 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1408 return mdd_lmm_oi(lmm, oi, false);
1411 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1413 struct lov_comp_md_v1 *comp_v1;
1415 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1416 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1418 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1420 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1421 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1422 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1423 __u16 tmp_gen = *gen;
1425 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1427 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1434 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1436 return mdd_lmm_gen(lmm, gen, true);
1439 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1441 return mdd_lmm_gen(lmm, gen, false);
1445 * swap layouts between 2 lustre objects
1447 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1448 struct md_object *obj2, __u64 flags)
1450 struct mdd_thread_info *info = mdd_env_info(env);
1451 struct mdd_object *fst_o = md2mdd_obj(obj1);
1452 struct mdd_object *snd_o = md2mdd_obj(obj2);
1453 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1454 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1455 struct mdd_device *mdd = mdo2mdd(obj1);
1456 struct lov_mds_md *fst_lmm, *snd_lmm;
1457 struct lu_buf *fst_buf = &info->mti_buf[0];
1458 struct lu_buf *snd_buf = &info->mti_buf[1];
1459 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1460 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1461 struct ost_id *saved_oi = NULL;
1462 struct thandle *handle;
1463 __u32 fst_gen, snd_gen, saved_gen;
1469 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1470 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1472 /* we have to sort the 2 obj, so locking will always
1473 * be in the same order, even in case of 2 concurrent swaps */
1474 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1475 if (rc == 0) /* same fid ? */
1481 rc = mdd_la_get(env, fst_o, fst_la);
1485 rc = mdd_la_get(env, snd_o, snd_la);
1489 /* check if layout swapping is allowed */
1490 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1494 handle = mdd_trans_create(env, mdd);
1496 RETURN(PTR_ERR(handle));
1498 /* objects are already sorted */
1499 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1500 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1502 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1503 if (rc < 0 && rc != -ENODATA)
1506 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1507 if (rc < 0 && rc != -ENODATA)
1510 /* swapping 2 non existant layouts is a success */
1511 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1514 /* to help inode migration between MDT, it is better to
1515 * start by the no layout file (if one), so we order the swap */
1516 if (snd_buf->lb_buf == NULL) {
1518 swap(fst_buf, snd_buf);
1521 fst_gen = snd_gen = 0;
1522 /* lmm and generation layout initialization */
1523 if (fst_buf->lb_buf != NULL) {
1524 fst_lmm = fst_buf->lb_buf;
1525 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1526 fst_fl = LU_XATTR_REPLACE;
1530 fst_fl = LU_XATTR_CREATE;
1533 snd_lmm = snd_buf->lb_buf;
1534 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1536 saved_gen = fst_gen;
1537 /* increase the generation layout numbers */
1542 * XXX The layout generation is used to generate component IDs for
1543 * the composite file, we have to do some special tweaks to make
1544 * sure the layout generation is always adequate for that job.
1547 /* Skip invalid generation number for composite layout */
1548 if ((snd_gen & LCME_ID_MASK) == 0)
1550 if ((fst_gen & LCME_ID_MASK) == 0)
1552 /* Make sure the generation is greater than all the component IDs */
1553 if (fst_gen < snd_gen)
1555 else if (fst_gen > snd_gen)
1558 /* set the file specific informations in lmm */
1559 if (fst_lmm != NULL) {
1560 saved_oi = &info->mti_oa.o_oi;
1561 mdd_get_lmm_oi(fst_lmm, saved_oi);
1562 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1563 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1564 mdd_set_lmm_oi(snd_lmm, saved_oi);
1566 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1567 cpu_to_le32(LOV_MAGIC_MAGIC))
1568 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1570 GOTO(stop, rc = -EPROTO);
1572 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1574 /* Prepare HSM attribute if it's required */
1575 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1576 const int buflen = sizeof(struct hsm_attrs);
1578 lu_buf_alloc(fst_hsm_buf, buflen);
1579 lu_buf_alloc(snd_hsm_buf, buflen);
1580 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1581 GOTO(stop, rc = -ENOMEM);
1583 /* Read HSM attribute */
1584 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1588 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1592 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1593 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1598 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1599 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1605 /* prepare transaction */
1606 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1611 if (fst_buf->lb_buf != NULL)
1612 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1613 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1616 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1621 rc = mdd_trans_start(env, mdd, handle);
1625 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1626 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1630 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1632 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1635 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1636 mdd_obj_dev_name(fst_o),
1637 PFID(mdo2fid(fst_o)), rc, rc2);
1642 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1646 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1649 if (fst_buf->lb_buf != NULL)
1650 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1651 LU_XATTR_REPLACE, handle);
1653 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1659 /* failure on second file, but first was done, so we have
1660 * to roll back first. */
1661 if (fst_buf->lb_buf != NULL) {
1662 mdd_set_lmm_oi(fst_lmm, saved_oi);
1663 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1664 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1665 LU_XATTR_REPLACE, handle);
1667 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1673 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1678 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1683 CERROR("%s: unable to roll back layout swap. FIDs: "
1684 DFID" and "DFID "error: %d/%d, steps: %d\n",
1685 mdd_obj_dev_name(fst_o),
1686 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1688 /* a solution to avoid journal commit is to panic,
1689 * but it has strong consequences so we use LBUG to
1690 * allow sysdamin to choose to panic or not
1697 /* Issue one changelog record per file */
1698 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1702 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1708 rc = mdd_trans_stop(env, mdd, rc, handle);
1710 mdd_write_unlock(env, snd_o);
1711 mdd_write_unlock(env, fst_o);
1713 lu_buf_free(fst_buf);
1714 lu_buf_free(snd_buf);
1715 lu_buf_free(fst_hsm_buf);
1716 lu_buf_free(snd_hsm_buf);
1720 static int mdd_declare_layout_change(const struct lu_env *env,
1721 struct mdd_device *mdd,
1722 struct mdd_object *obj,
1723 struct layout_intent *layout,
1724 const struct lu_buf *buf,
1725 struct thandle *handle)
1729 rc = mdo_declare_layout_change(env, obj, layout, buf, handle);
1733 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1736 /* For PFL, this is used to instantiate necessary component objects. */
1737 int mdd_layout_change(const struct lu_env *env, struct md_object *obj,
1738 struct layout_intent *layout, const struct lu_buf *buf)
1740 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1741 struct mdd_device *mdd = mdo2mdd(obj);
1742 struct thandle *handle;
1746 handle = mdd_trans_create(env, mdd);
1748 RETURN(PTR_ERR(handle));
1750 rc = mdd_declare_layout_change(env, mdd, mdd_obj, layout, buf, handle);
1752 * It's possible that another layout write intent has already
1753 * instantiated our objects, so a -EALREADY returned, and we need to
1757 GOTO(stop, rc = (rc == -EALREADY) ? 0 : rc);
1759 rc = mdd_trans_start(env, mdd, handle);
1763 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1764 rc = mdo_layout_change(env, mdd_obj, layout, buf, handle);
1765 mdd_write_unlock(env, mdd_obj);
1769 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, handle);
1771 RETURN(mdd_trans_stop(env, mdd, rc, handle));
1774 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
1775 struct mdd_object *child, const struct lu_attr *attr,
1776 const struct md_op_spec *spec,
1777 struct dt_allocation_hint *hint)
1779 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
1780 struct dt_object *nc = mdd_object_child(child);
1782 memset(hint, 0, sizeof(*hint));
1784 /* For striped directory, give striping EA to lod_ah_init, which will
1785 * decide the stripe_offset and stripe count by it. */
1786 if (S_ISDIR(attr->la_mode) &&
1787 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
1788 hint->dah_eadata = spec->u.sp_ea.eadata;
1789 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
1791 hint->dah_eadata = NULL;
1792 hint->dah_eadata_len = 0;
1795 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
1796 hint->dah_eadata, hint->dah_eadata_len);
1797 /* @hint will be initialized by underlying device. */
1798 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
1802 * do NOT or the MAY_*'s, you'll get the weakest
1804 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
1808 /* Sadly, NFSD reopens a file repeatedly during operation, so the
1809 * "acc_mode = 0" allowance for newly-created files isn't honoured.
1810 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
1811 * owner can write to a file even if it is marked readonly to hide
1812 * its brokenness. (bug 5781) */
1813 if (flags & MDS_OPEN_OWNEROVERRIDE) {
1814 struct lu_ucred *uc = lu_ucred_check(env);
1816 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
1820 if (flags & FMODE_READ)
1822 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
1824 if (flags & MDS_FMODE_EXEC)
1829 static int mdd_open_sanity_check(const struct lu_env *env,
1830 struct mdd_object *obj,
1831 const struct lu_attr *attr, int flag)
1837 if (mdd_is_dead_obj(obj))
1840 if (S_ISLNK(attr->la_mode))
1843 mode = accmode(env, attr, flag);
1845 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
1848 if (!(flag & MDS_OPEN_CREATED)) {
1849 rc = mdd_permission_internal(env, obj, attr, mode);
1854 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
1855 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
1856 flag &= ~MDS_OPEN_TRUNC;
1858 /* For writing append-only file must open it with append mode. */
1859 if (attr->la_flags & LUSTRE_APPEND_FL) {
1860 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
1862 if (flag & MDS_OPEN_TRUNC)
1869 static int mdd_open(const struct lu_env *env, struct md_object *obj,
1872 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1873 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1876 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1878 rc = mdd_la_get(env, mdd_obj, attr);
1882 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
1886 mdd_obj->mod_count++;
1889 mdd_write_unlock(env, mdd_obj);
1893 static int mdd_declare_close(const struct lu_env *env,
1894 struct mdd_object *obj,
1896 struct thandle *handle)
1900 rc = orph_declare_index_delete(env, obj, handle);
1904 return mdo_declare_destroy(env, obj, handle);
1908 * No permission check is needed.
1910 static int mdd_close(const struct lu_env *env, struct md_object *obj,
1911 struct md_attr *ma, int mode)
1913 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1914 struct mdd_device *mdd = mdo2mdd(obj);
1915 struct thandle *handle = NULL;
1918 bool blocked = false;
1921 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
1922 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1923 mdd_obj->mod_count--;
1924 mdd_write_unlock(env, mdd_obj);
1926 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
1927 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
1928 "list\n", PFID(mdd_object_fid(mdd_obj)));
1932 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
1933 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
1934 /* check without any lock */
1935 is_orphan = mdd_obj->mod_count == 1 &&
1936 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
1940 /* mdd_trans_create() maybe failed because of barrier_entry(),
1941 * under such case, the orphan MDT-object will be left in the
1942 * orphan list, and when the MDT remount next time, the unused
1943 * orphans will be destroyed automatically.
1945 * One exception: the former mdd_finish_unlink may failed to
1946 * add the orphan MDT-object to the orphan list, then if the
1947 * mdd_trans_create() failed because of barrier_entry(), the
1948 * MDT-object will become real orphan that is neither in the
1949 * namespace nor in the orphan list. Such bad case should be
1950 * very rare and will be handled by e2fsck/lfsck. */
1951 handle = mdd_trans_create(env, mdo2mdd(obj));
1952 if (IS_ERR(handle)) {
1953 rc = PTR_ERR(handle);
1954 if (rc != -EINPROGRESS)
1962 rc = mdd_declare_close(env, mdd_obj, ma, handle);
1966 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1970 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
1976 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1977 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
1979 CERROR("Failed to get lu_attr of "DFID": %d\n",
1980 PFID(mdd_object_fid(mdd_obj)), rc);
1984 /* check again with lock */
1985 is_orphan = (mdd_obj->mod_count == 1) &&
1986 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
1987 ma->ma_attr.la_nlink == 0);
1989 if (is_orphan && !handle && !blocked) {
1990 mdd_write_unlock(env, mdd_obj);
1994 mdd_obj->mod_count--; /*release open count */
1996 if (!is_orphan || blocked)
2000 /* NB: Object maybe not in orphan list originally, it is rare case for
2001 * mdd_finish_unlink() failure, in that case, the object doesn't have
2002 * ORPHAN_OBJ flag */
2003 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2004 /* remove link to object from orphan index */
2005 LASSERT(handle != NULL);
2006 rc = __mdd_orphan_del(env, mdd_obj, handle);
2008 CERROR("%s: unable to delete "DFID" from orphan list: "
2009 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2010 PFID(mdd_object_fid(mdd_obj)), rc);
2011 /* If object was not deleted from orphan list, do not
2012 * destroy OSS objects, which will be done when next
2017 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2018 "list, OSS objects to be destroyed.\n",
2019 PFID(mdd_object_fid(mdd_obj)));
2022 rc = mdo_destroy(env, mdd_obj, handle);
2025 CERROR("%s: unable to delete "DFID" from orphan list: "
2026 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2027 PFID(mdd_object_fid(mdd_obj)), rc);
2032 mdd_write_unlock(env, mdd_obj);
2034 if (!rc && !blocked &&
2035 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2036 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2037 if (handle == NULL) {
2038 handle = mdd_trans_create(env, mdo2mdd(obj));
2040 GOTO(stop, rc = PTR_ERR(handle));
2042 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2047 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2052 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2057 if (handle != NULL && !IS_ERR(handle))
2058 rc = mdd_trans_stop(env, mdd, rc, handle);
2064 * Permission check is done when open,
2065 * no need check again.
2067 static int mdd_readpage_sanity_check(const struct lu_env *env,
2068 struct mdd_object *obj)
2070 struct dt_object *next = mdd_object_child(obj);
2074 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2082 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2083 size_t nob, const struct dt_it_ops *iops,
2084 struct dt_it *it, __u32 attr, void *arg)
2086 struct lu_dirpage *dp = &lp->lp_dir;
2090 struct lu_dirent *ent;
2091 struct lu_dirent *last = NULL;
2095 if (nob < sizeof(*dp))
2098 memset(area, 0, sizeof (*dp));
2099 area += sizeof (*dp);
2100 nob -= sizeof (*dp);
2107 len = iops->key_size(env, it);
2109 /* IAM iterator can return record with zero len. */
2113 hash = iops->store(env, it);
2114 if (unlikely(first)) {
2116 dp->ldp_hash_start = cpu_to_le64(hash);
2119 /* calculate max space required for lu_dirent */
2120 recsize = lu_dirent_calc_size(len, attr);
2122 if (nob >= recsize) {
2123 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2124 if (result == -ESTALE)
2129 /* osd might not able to pack all attributes,
2130 * so recheck rec length */
2131 recsize = le16_to_cpu(ent->lde_reclen);
2133 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2134 fid_le_to_cpu(&fid, &ent->lde_fid);
2135 if (fid_is_dot_lustre(&fid))
2139 result = (last != NULL) ? 0 :-EINVAL;
2143 ent = (void *)ent + recsize;
2147 result = iops->next(env, it);
2148 if (result == -ESTALE)
2150 } while (result == 0);
2153 dp->ldp_hash_end = cpu_to_le64(hash);
2155 if (last->lde_hash == dp->ldp_hash_end)
2156 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2157 last->lde_reclen = 0; /* end mark */
2160 /* end of directory */
2161 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2162 else if (result < 0)
2163 CWARN("build page failed: %d!\n", result);
2167 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2168 const struct lu_rdpg *rdpg)
2170 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2174 if (mdd_object_exists(mdd_obj) == 0) {
2175 CERROR("%s: object "DFID" not found: rc = -2\n",
2176 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2180 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2181 rc = mdd_readpage_sanity_check(env, mdd_obj);
2183 GOTO(out_unlock, rc);
2185 if (mdd_is_dead_obj(mdd_obj)) {
2187 struct lu_dirpage *dp;
2190 * According to POSIX, please do not return any entry to client:
2191 * even dot and dotdot should not be returned.
2193 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2194 PFID(mdd_object_fid(mdd_obj)));
2196 if (rdpg->rp_count <= 0)
2197 GOTO(out_unlock, rc = -EFAULT);
2198 LASSERT(rdpg->rp_pages != NULL);
2200 pg = rdpg->rp_pages[0];
2201 dp = (struct lu_dirpage *)kmap(pg);
2202 memset(dp, 0 , sizeof(struct lu_dirpage));
2203 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2204 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2205 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2207 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2210 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2211 mdd_dir_page_build, NULL);
2213 struct lu_dirpage *dp;
2215 dp = kmap(rdpg->rp_pages[0]);
2216 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2219 * No pages were processed, mark this for first page
2222 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2223 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2224 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2226 kunmap(rdpg->rp_pages[0]);
2229 GOTO(out_unlock, rc);
2231 mdd_read_unlock(env, mdd_obj);
2235 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2237 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2239 if (mdd_object_exists(mdd_obj) == 0) {
2242 CERROR("%s: object "DFID" not found: rc = %d\n",
2243 mdd_obj_dev_name(mdd_obj),
2244 PFID(mdd_object_fid(mdd_obj)), rc);
2247 return dt_object_sync(env, mdd_object_child(mdd_obj),
2251 static int mdd_object_lock(const struct lu_env *env,
2252 struct md_object *obj,
2253 struct lustre_handle *lh,
2254 struct ldlm_enqueue_info *einfo,
2255 union ldlm_policy_data *policy)
2257 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2258 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2262 static int mdd_object_unlock(const struct lu_env *env,
2263 struct md_object *obj,
2264 struct ldlm_enqueue_info *einfo,
2265 union ldlm_policy_data *policy)
2267 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2268 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2271 const struct md_object_operations mdd_obj_ops = {
2272 .moo_permission = mdd_permission,
2273 .moo_attr_get = mdd_attr_get,
2274 .moo_attr_set = mdd_attr_set,
2275 .moo_xattr_get = mdd_xattr_get,
2276 .moo_xattr_set = mdd_xattr_set,
2277 .moo_xattr_list = mdd_xattr_list,
2278 .moo_invalidate = mdd_invalidate,
2279 .moo_xattr_del = mdd_xattr_del,
2280 .moo_swap_layouts = mdd_swap_layouts,
2281 .moo_open = mdd_open,
2282 .moo_close = mdd_close,
2283 .moo_readpage = mdd_readpage,
2284 .moo_readlink = mdd_readlink,
2285 .moo_changelog = mdd_changelog,
2286 .moo_object_sync = mdd_object_sync,
2287 .moo_object_lock = mdd_object_lock,
2288 .moo_object_unlock = mdd_object_unlock,
2289 .moo_layout_change = mdd_layout_change,