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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/mdd/mdd_object.c
38 * Lustre Metadata Server (mdd) routines
40 * Author: Wang Di <wangdi@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_MDS
45 #include <linux/module.h>
47 #include <obd_class.h>
48 #include <obd_support.h>
49 #include <lprocfs_status.h>
50 /* fid_be_cpu(), fid_cpu_to_be(). */
51 #include <lustre_fid.h>
52 #include <lustre_idmap.h>
53 #include <lustre_param.h>
54 #include <lustre_mds.h>
55 #include <lustre/lustre_idl.h>
57 #include "mdd_internal.h"
59 static const struct lu_object_operations mdd_lu_obj_ops;
61 static int mdd_xattr_get(const struct lu_env *env,
62 struct md_object *obj, struct lu_buf *buf,
65 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
70 if (mdd_object_exists(obj) == 0) {
71 CERROR("%s: object "DFID" not found: rc = -2\n",
72 mdd_obj_dev_name(obj), PFID(mdd_object_fid(obj)));
76 rc = mdo_attr_get(env, obj, la);
77 if (unlikely(rc != 0)) {
79 obj->mod_flags |= DEAD_OBJ;
83 if (la->la_valid & LA_FLAGS &&
84 la->la_flags & LUSTRE_ORPHAN_FL)
85 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
90 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
92 struct mdd_thread_info *info;
94 lu_env_refill((struct lu_env *)env);
95 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
96 LASSERT(info != NULL);
100 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
104 buf = &mdd_env_info(env)->mti_buf[0];
110 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
111 const void *area, ssize_t len)
115 buf = &mdd_env_info(env)->mti_buf[0];
116 buf->lb_buf = (void *)area;
121 struct lu_object *mdd_object_alloc(const struct lu_env *env,
122 const struct lu_object_header *hdr,
125 struct mdd_object *mdd_obj;
127 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
128 if (mdd_obj != NULL) {
131 o = mdd2lu_obj(mdd_obj);
132 lu_object_init(o, NULL, d);
133 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
134 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
135 mdd_obj->mod_count = 0;
136 o->lo_ops = &mdd_lu_obj_ops;
143 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
144 const struct lu_object_conf *unused)
146 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
147 struct mdd_object *mdd_obj = lu2mdd_obj(o);
148 struct lu_object *below;
149 struct lu_device *under;
152 mdd_obj->mod_cltime = 0;
153 under = &d->mdd_child->dd_lu_dev;
154 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
156 RETURN(PTR_ERR(below));
158 lu_object_add(o, below);
163 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
167 if (lu_object_exists(o)) {
168 struct mdd_object *mdd_obj = lu2mdd_obj(o);
169 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
171 rc = mdd_la_get(env, mdd_obj, attr);
177 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
179 struct mdd_object *mdd = lu2mdd_obj(o);
182 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
185 static int mdd_object_print(const struct lu_env *env, void *cookie,
186 lu_printer_t p, const struct lu_object *o)
188 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
189 return (*p)(env, cookie, LUSTRE_MDD_NAME"-object@%p(open_count=%d, "
190 "valid=%x, cltime="LPU64", flags=%lx)",
191 mdd, mdd->mod_count, mdd->mod_valid,
192 mdd->mod_cltime, mdd->mod_flags);
195 static const struct lu_object_operations mdd_lu_obj_ops = {
196 .loo_object_init = mdd_object_init,
197 .loo_object_start = mdd_object_start,
198 .loo_object_free = mdd_object_free,
199 .loo_object_print = mdd_object_print,
202 struct mdd_object *mdd_object_find(const struct lu_env *env,
203 struct mdd_device *d,
204 const struct lu_fid *f)
206 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
210 * No permission check is needed.
212 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
215 struct mdd_object *mdd_obj = md2mdd_obj(obj);
220 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
221 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
222 ma->ma_attr.la_nlink = 0;
228 * No permission check is needed.
230 static int mdd_xattr_get(const struct lu_env *env,
231 struct md_object *obj, struct lu_buf *buf,
234 struct mdd_object *mdd_obj = md2mdd_obj(obj);
239 if (mdd_object_exists(mdd_obj) == 0) {
240 CERROR("%s: object "DFID" not found: rc = -2\n",
241 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
245 /* If the object has been destroyed, then do not get LMVEA, because
246 * it needs to load stripes from the iteration of the master object,
247 * and it will cause problem if master object has been destroyed, see
249 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
250 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
251 strcmp(name, XATTR_NAME_LMV) == 0))
254 /* If the object has been delete from the namespace, then
255 * get linkEA should return -ENOENT as well */
256 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
257 strcmp(name, XATTR_NAME_LINK) == 0))
260 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
261 rc = mdo_xattr_get(env, mdd_obj, buf, name);
262 mdd_read_unlock(env, mdd_obj);
268 * Permission check is done when open,
269 * no need check again.
271 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
274 struct mdd_object *mdd_obj = md2mdd_obj(obj);
275 struct dt_object *next;
280 if (mdd_object_exists(mdd_obj) == 0) {
281 CERROR("%s: object "DFID" not found: rc = -2\n",
282 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
286 next = mdd_object_child(mdd_obj);
287 LASSERT(next != NULL);
288 LASSERT(next->do_body_ops != NULL);
289 LASSERT(next->do_body_ops->dbo_read != NULL);
290 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
291 rc = dt_read(env, next, buf, &pos);
292 mdd_read_unlock(env, mdd_obj);
297 * No permission check is needed.
299 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
302 struct mdd_object *mdd_obj = md2mdd_obj(obj);
307 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
308 rc = mdo_xattr_list(env, mdd_obj, buf);
309 mdd_read_unlock(env, mdd_obj);
315 * Filter out XATTR_NAME_LINK if this is an orphan object. See
318 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
319 char *end = (char *)buf->lb_buf + rc;
320 char *p = buf->lb_buf;
323 char *next = p + strlen(p) + 1;
325 if (strcmp(p, XATTR_NAME_LINK) == 0) {
327 memmove(p, next, end - next);
329 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
330 " of orphan "DFID"\n",
331 PFID(mdd_object_fid(mdd_obj)));
342 int mdd_declare_object_create_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;
360 dof->dof_type = dt_mode_to_dft(attr->la_mode);
361 if (dof->dof_type == DFT_REGULAR) {
362 dof->u.dof_reg.striped =
363 md_should_create(spec->sp_cr_flags);
364 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
365 dof->u.dof_reg.striped = 0;
366 /* is this replay? */
368 dof->u.dof_reg.striped = 0;
372 rc = mdo_declare_create_obj(env, c, attr, hint, dof, handle);
377 int mdd_object_create_internal(const struct lu_env *env, struct mdd_object *p,
378 struct mdd_object *c, struct lu_attr *attr,
379 struct thandle *handle,
380 const struct md_op_spec *spec,
381 struct dt_allocation_hint *hint)
383 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
387 LASSERT(!mdd_object_exists(c));
389 rc = mdo_create_obj(env, c, attr, hint, dof, handle);
394 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
395 const struct lu_attr *attr, struct thandle *handle,
401 rc = mdo_attr_set(env, obj, attr, handle);
402 #ifdef CONFIG_FS_POSIX_ACL
403 if (!rc && (attr->la_valid & LA_MODE) && needacl)
404 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
409 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
410 const struct lu_attr *oattr, struct lu_attr *attr,
411 struct thandle *handle)
416 LASSERT(attr->la_valid & LA_CTIME);
417 LASSERT(oattr != NULL);
419 /* Make sure the ctime is increased only, however, it's not strictly
420 * reliable at here because there is not guarantee to hold lock on
421 * object, so we just bypass some unnecessary cmtime setting first
422 * and OSD has to check it again. */
423 if (attr->la_ctime < oattr->la_ctime)
424 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
425 else if (attr->la_valid == LA_CTIME &&
426 attr->la_ctime == oattr->la_ctime)
427 attr->la_valid &= ~LA_CTIME;
429 if (attr->la_valid != 0)
430 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
435 * This gives the same functionality as the code between
436 * sys_chmod and inode_setattr
437 * chown_common and inode_setattr
438 * utimes and inode_setattr
439 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
441 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
442 const struct lu_attr *oattr, struct lu_attr *la,
443 const unsigned long flags)
452 /* Do not permit change file type */
453 if (la->la_valid & LA_TYPE)
456 /* They should not be processed by setattr */
457 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
460 LASSERT(oattr != NULL);
462 uc = lu_ucred_check(env);
466 if (la->la_valid == LA_CTIME) {
467 if (!(flags & MDS_PERM_BYPASS))
468 /* This is only for set ctime when rename's source is
470 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
472 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
473 la->la_valid &= ~LA_CTIME;
477 if (la->la_valid == LA_ATIME) {
478 /* This is an atime-only attribute update for close RPCs. */
479 if (la->la_atime < (oattr->la_atime +
480 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
481 la->la_valid &= ~LA_ATIME;
485 /* Check if flags change. */
486 if (la->la_valid & LA_FLAGS) {
487 unsigned int oldflags = oattr->la_flags &
488 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
489 unsigned int newflags = la->la_flags &
490 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
492 if ((uc->uc_fsuid != oattr->la_uid) &&
493 !md_capable(uc, CFS_CAP_FOWNER))
496 /* The IMMUTABLE and APPEND_ONLY flags can
497 * only be changed by the relevant capability. */
498 if ((oldflags ^ newflags) &&
499 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
502 if (!S_ISDIR(oattr->la_mode))
503 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
506 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
507 (la->la_valid & ~LA_FLAGS) &&
508 !(flags & MDS_PERM_BYPASS))
511 /* Check for setting the obj time. */
512 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
513 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
514 if ((uc->uc_fsuid != oattr->la_uid) &&
515 !md_capable(uc, CFS_CAP_FOWNER)) {
516 rc = mdd_permission_internal(env, obj, oattr,
523 if (la->la_valid & LA_KILL_SUID) {
524 la->la_valid &= ~LA_KILL_SUID;
525 if ((oattr->la_mode & S_ISUID) &&
526 !(la->la_valid & LA_MODE)) {
527 la->la_mode = oattr->la_mode;
528 la->la_valid |= LA_MODE;
530 la->la_mode &= ~S_ISUID;
533 if (la->la_valid & LA_KILL_SGID) {
534 la->la_valid &= ~LA_KILL_SGID;
535 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
536 (S_ISGID | S_IXGRP)) &&
537 !(la->la_valid & LA_MODE)) {
538 la->la_mode = oattr->la_mode;
539 la->la_valid |= LA_MODE;
541 la->la_mode &= ~S_ISGID;
544 /* Make sure a caller can chmod. */
545 if (la->la_valid & LA_MODE) {
546 if (!(flags & MDS_PERM_BYPASS) &&
547 (uc->uc_fsuid != oattr->la_uid) &&
548 !md_capable(uc, CFS_CAP_FOWNER))
551 if (la->la_mode == (umode_t) -1)
552 la->la_mode = oattr->la_mode;
554 la->la_mode = (la->la_mode & S_IALLUGO) |
555 (oattr->la_mode & ~S_IALLUGO);
557 /* Also check the setgid bit! */
558 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
559 la->la_gid : oattr->la_gid) &&
560 !md_capable(uc, CFS_CAP_FSETID))
561 la->la_mode &= ~S_ISGID;
563 la->la_mode = oattr->la_mode;
566 /* Make sure a caller can chown. */
567 if (la->la_valid & LA_UID) {
568 if (la->la_uid == (uid_t) -1)
569 la->la_uid = oattr->la_uid;
570 if (((uc->uc_fsuid != oattr->la_uid) ||
571 (la->la_uid != oattr->la_uid)) &&
572 !md_capable(uc, CFS_CAP_CHOWN))
575 /* If the user or group of a non-directory has been
576 * changed by a non-root user, remove the setuid bit.
577 * 19981026 David C Niemi <niemi@tux.org>
579 * Changed this to apply to all users, including root,
580 * to avoid some races. This is the behavior we had in
581 * 2.0. The check for non-root was definitely wrong
582 * for 2.2 anyway, as it should have been using
583 * CAP_FSETID rather than fsuid -- 19990830 SD. */
584 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
585 !S_ISDIR(oattr->la_mode)) {
586 la->la_mode &= ~S_ISUID;
587 la->la_valid |= LA_MODE;
591 /* Make sure caller can chgrp. */
592 if (la->la_valid & LA_GID) {
593 if (la->la_gid == (gid_t) -1)
594 la->la_gid = oattr->la_gid;
595 if (((uc->uc_fsuid != oattr->la_uid) ||
596 ((la->la_gid != oattr->la_gid) &&
597 !lustre_in_group_p(uc, la->la_gid))) &&
598 !md_capable(uc, CFS_CAP_CHOWN))
601 /* Likewise, if the user or group of a non-directory
602 * has been changed by a non-root user, remove the
603 * setgid bit UNLESS there is no group execute bit
604 * (this would be a file marked for mandatory
605 * locking). 19981026 David C Niemi <niemi@tux.org>
607 * Removed the fsuid check (see the comment above) --
609 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
610 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
611 la->la_mode &= ~S_ISGID;
612 la->la_valid |= LA_MODE;
616 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
617 if (!((flags & MDS_OWNEROVERRIDE) &&
618 (uc->uc_fsuid == oattr->la_uid)) &&
619 !(flags & MDS_PERM_BYPASS)) {
620 rc = mdd_permission_internal(env, obj, oattr,
627 if (la->la_valid & LA_CTIME) {
628 /* The pure setattr, it has the priority over what is
629 * already set, do not drop it if ctime is equal. */
630 if (la->la_ctime < oattr->la_ctime)
631 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
637 /** Store a data change changelog record
638 * If this fails, we must fail the whole transaction; we don't
639 * want the change to commit without the log entry.
640 * \param mdd_obj - mdd_object of change
641 * \param handle - transaction handle
643 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
644 enum changelog_rec_type type, int flags,
645 struct mdd_object *mdd_obj, struct thandle *handle)
647 const struct lu_ucred *uc = lu_ucred(env);
648 const struct lu_fid *tfid;
649 struct llog_changelog_rec *rec;
655 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
657 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
660 LASSERT(mdd_obj != NULL);
661 LASSERT(handle != NULL);
663 tfid = mdo2fid(mdd_obj);
665 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
666 cfs_time_before_64(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
667 /* Don't need multiple updates in this log */
668 /* Don't check under lock - no big deal if we get an extra
673 flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
674 if (uc != NULL && uc->uc_jobid[0] != '\0')
677 reclen = llog_data_len(changelog_rec_offset(flags & CLF_SUPPORTED));
678 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
679 if (buf->lb_buf == NULL)
683 rec->cr.cr_flags = flags;
684 rec->cr.cr_type = (__u32)type;
685 rec->cr.cr_tfid = *tfid;
686 rec->cr.cr_namelen = 0;
687 mdd_obj->mod_cltime = cfs_time_current_64();
689 if (flags & CLF_JOBID)
690 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
692 rc = mdd_changelog_store(env, mdd, rec, handle);
697 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
698 int flags, struct md_object *obj)
700 struct thandle *handle;
701 struct mdd_object *mdd_obj = md2mdd_obj(obj);
702 struct mdd_device *mdd = mdo2mdd(obj);
706 handle = mdd_trans_create(env, mdd);
708 RETURN(PTR_ERR(handle));
710 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
714 rc = mdd_trans_start(env, mdd, handle);
718 rc = mdd_changelog_data_store(env, mdd, type, flags, mdd_obj,
722 mdd_trans_stop(env, mdd, rc, handle);
728 * Save LMA extended attributes with data from \a ma.
730 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
731 * not, LMA EA will be first read from disk, modified and write back.
734 /* Precedence for choosing record type when multiple
735 * attributes change: setattr > mtime > ctime > atime
736 * (ctime changes when mtime does, plus chmod/chown.
737 * atime and ctime are independent.) */
738 static int mdd_attr_set_changelog(const struct lu_env *env,
739 struct md_object *obj, struct thandle *handle,
742 struct mdd_device *mdd = mdo2mdd(obj);
745 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
746 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
747 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
748 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
749 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
750 bits = bits & mdd->mdd_cl.mc_mask;
751 /* This is an implementation limit rather than a protocol limit */
752 CLASSERT(CL_LAST <= sizeof(int) * 8);
756 /* The record type is the lowest non-masked set bit */
759 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
760 return mdd_changelog_data_store(env, mdd, type, (int)valid,
761 md2mdd_obj(obj), handle);
764 static int mdd_declare_attr_set(const struct lu_env *env,
765 struct mdd_device *mdd,
766 struct mdd_object *obj,
767 const struct lu_attr *attr,
768 struct thandle *handle)
772 rc = mdo_declare_attr_set(env, obj, attr, handle);
776 #ifdef CONFIG_FS_POSIX_ACL
777 if (attr->la_valid & LA_MODE) {
778 mdd_read_lock(env, obj, MOR_TGT_CHILD);
779 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
780 XATTR_NAME_ACL_ACCESS);
781 mdd_read_unlock(env, obj);
782 if (rc == -EOPNOTSUPP || rc == -ENODATA)
788 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
789 rc = mdo_declare_xattr_set(env, obj, buf,
790 XATTR_NAME_ACL_ACCESS, 0,
798 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
805 * permission changes may require sync operation, to mitigate performance
806 * impact, only do this for dir and when permission is reduced.
808 * For regular files, version is updated with permission change (see VBR), async
809 * permission won't cause any issue, while missing permission change on
810 * directory may affect accessibility of other objects after recovery.
812 static inline bool permission_needs_sync(const struct lu_attr *old,
813 const struct lu_attr *new)
815 if (!S_ISDIR(old->la_mode))
818 if (new->la_valid & (LA_UID | LA_GID))
821 if (new->la_valid & LA_MODE &&
822 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
825 if ((new->la_valid & LA_MODE) &&
826 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
827 (old->la_mode & S_IRWXUGO))
833 /* set attr and LOV EA at once, return updated attr */
834 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
835 const struct md_attr *ma)
837 struct mdd_object *mdd_obj = md2mdd_obj(obj);
838 struct mdd_device *mdd = mdo2mdd(obj);
839 struct thandle *handle;
840 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
841 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
842 const struct lu_attr *la = &ma->ma_attr;
846 /* we do not use ->attr_set() for LOV/HSM EA any more */
847 LASSERT((ma->ma_valid & MA_LOV) == 0);
848 LASSERT((ma->ma_valid & MA_HSM) == 0);
850 rc = mdd_la_get(env, mdd_obj, attr);
854 *la_copy = ma->ma_attr;
855 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
859 /* no need to setattr anymore */
860 if (la_copy->la_valid == 0) {
861 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
862 "valid is "LPX64"\n", mdd2obd_dev(mdd)->obd_name,
863 PFID(mdo2fid(mdd_obj)), la->la_valid);
868 handle = mdd_trans_create(env, mdd);
870 RETURN(PTR_ERR(handle));
872 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
876 rc = mdd_trans_start(env, mdd, handle);
880 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
883 if (la->la_valid & (LA_MTIME | LA_CTIME))
884 CDEBUG(D_INODE, "setting mtime "LPU64", ctime "LPU64"\n",
885 la->la_mtime, la->la_ctime);
887 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
888 if (la_copy->la_valid & LA_FLAGS)
889 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
890 else if (la_copy->la_valid) /* setattr */
891 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
892 mdd_write_unlock(env, mdd_obj);
895 rc = mdd_attr_set_changelog(env, obj, handle,
901 mdd_trans_stop(env, mdd, rc, handle);
905 static int mdd_xattr_sanity_check(const struct lu_env *env,
906 struct mdd_object *obj,
907 const struct lu_attr *attr)
909 struct lu_ucred *uc = lu_ucred_assert(env);
912 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
915 if ((uc->uc_fsuid != attr->la_uid) && !md_capable(uc, CFS_CAP_FOWNER))
921 static int mdd_declare_xattr_set(const struct lu_env *env,
922 struct mdd_device *mdd,
923 struct mdd_object *obj,
924 const struct lu_buf *buf,
926 int fl, struct thandle *handle)
930 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
934 /* Only record user and layout xattr changes */
935 if (strncmp(XATTR_USER_PREFIX, name,
936 sizeof(XATTR_USER_PREFIX) - 1) == 0 ||
937 strcmp(XATTR_NAME_LOV, name) == 0) {
938 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
943 /* If HSM data is modified, this could add a changelog */
944 if (strcmp(XATTR_NAME_HSM, name) == 0) {
945 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
950 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
955 * Compare current and future data of HSM EA and add a changelog if needed.
957 * Caller should have write-locked \param obj.
959 * \param buf - Future HSM EA content.
960 * \retval 0 if no changelog is needed or changelog was added properly.
961 * \retval -ve errno if there was a problem
963 static int mdd_hsm_update_locked(const struct lu_env *env,
964 struct md_object *obj,
965 const struct lu_buf *buf,
966 struct thandle *handle)
968 struct mdd_thread_info *info = mdd_env_info(env);
969 struct mdd_device *mdd = mdo2mdd(obj);
970 struct mdd_object *mdd_obj = md2mdd_obj(obj);
971 struct lu_buf *current_buf;
972 struct md_hsm *current_mh;
973 struct md_hsm *new_mh;
977 OBD_ALLOC_PTR(current_mh);
978 if (current_mh == NULL)
981 /* Read HSM attrs from disk */
982 CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
983 current_buf = mdd_buf_get(env, info->mti_xattr_buf,
984 sizeof(info->mti_xattr_buf));
985 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
986 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
987 if (rc < 0 && rc != -ENODATA)
989 else if (rc == -ENODATA)
990 current_mh->mh_flags = 0;
992 /* Map future HSM xattr */
993 OBD_ALLOC_PTR(new_mh);
995 GOTO(free, rc = -ENOMEM);
996 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
998 /* If HSM flags are different, add a changelog */
1000 if (current_mh->mh_flags != new_mh->mh_flags) {
1002 hsm_set_cl_event(&flags, HE_STATE);
1003 if (new_mh->mh_flags & HS_DIRTY)
1004 hsm_set_cl_flags(&flags, CLF_HSM_DIRTY);
1006 rc = mdd_changelog_data_store(env, mdd, CL_HSM, flags, mdd_obj,
1010 OBD_FREE_PTR(new_mh);
1012 OBD_FREE_PTR(current_mh);
1016 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1020 * The caller should guarantee to update the object ctime
1021 * after xattr_set if needed.
1023 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1024 const struct lu_buf *buf, const char *name,
1027 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1028 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1029 struct mdd_device *mdd = mdo2mdd(obj);
1030 struct thandle *handle;
1034 rc = mdd_la_get(env, mdd_obj, attr);
1038 rc = mdd_xattr_sanity_check(env, mdd_obj, attr);
1042 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1043 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1044 struct posix_acl *acl;
1046 /* user may set empty ACL, which should be treated as removing
1048 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1051 RETURN(PTR_ERR(acl));
1053 rc = mdd_xattr_del(env, obj, name);
1056 posix_acl_release(acl);
1059 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1060 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1064 handle = mdd_trans_create(env, mdd);
1066 RETURN(PTR_ERR(handle));
1068 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1072 rc = mdd_trans_start(env, mdd, handle);
1076 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1078 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1079 rc = mdd_hsm_update_locked(env, obj, buf, handle);
1081 mdd_write_unlock(env, mdd_obj);
1086 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1087 mdd_write_unlock(env, mdd_obj);
1091 if (strcmp(XATTR_NAME_LOV, name) == 0)
1092 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj,
1094 else if (strncmp(XATTR_USER_PREFIX, name,
1095 sizeof(XATTR_USER_PREFIX) - 1) == 0 ||
1096 strncmp(POSIX_ACL_XATTR_ACCESS, name,
1097 sizeof(POSIX_ACL_XATTR_ACCESS) - 1) == 0 ||
1098 strncmp(POSIX_ACL_XATTR_DEFAULT, name,
1099 sizeof(POSIX_ACL_XATTR_DEFAULT) - 1) == 0)
1100 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj,
1104 mdd_trans_stop(env, mdd, rc, handle);
1109 static int mdd_declare_xattr_del(const struct lu_env *env,
1110 struct mdd_device *mdd,
1111 struct mdd_object *obj,
1113 struct thandle *handle)
1117 rc = mdo_declare_xattr_del(env, obj, name, handle);
1121 /* Only record system & user xattr changes */
1122 if (strncmp(XATTR_USER_PREFIX, name,
1123 sizeof(XATTR_USER_PREFIX) - 1) == 0 ||
1124 strncmp(POSIX_ACL_XATTR_ACCESS, name,
1125 sizeof(POSIX_ACL_XATTR_ACCESS) - 1) == 0 ||
1126 strncmp(POSIX_ACL_XATTR_DEFAULT, name,
1127 sizeof(POSIX_ACL_XATTR_DEFAULT) - 1) == 0)
1128 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1134 * The caller should guarantee to update the object ctime
1135 * after xattr_set if needed.
1137 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1140 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1141 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1142 struct mdd_device *mdd = mdo2mdd(obj);
1143 struct thandle *handle;
1147 rc = mdd_la_get(env, mdd_obj, attr);
1151 rc = mdd_xattr_sanity_check(env, mdd_obj, attr);
1155 handle = mdd_trans_create(env, mdd);
1157 RETURN(PTR_ERR(handle));
1159 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1163 rc = mdd_trans_start(env, mdd, handle);
1167 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1168 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1169 mdd_write_unlock(env, mdd_obj);
1173 /* Only record system & user xattr changes */
1174 if (strncmp(XATTR_USER_PREFIX, name,
1175 sizeof(XATTR_USER_PREFIX) - 1) == 0 ||
1176 strncmp(POSIX_ACL_XATTR_ACCESS, name,
1177 sizeof(POSIX_ACL_XATTR_ACCESS) - 1) == 0 ||
1178 strncmp(POSIX_ACL_XATTR_DEFAULT, name,
1179 sizeof(POSIX_ACL_XATTR_DEFAULT) - 1) == 0)
1180 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj,
1184 mdd_trans_stop(env, mdd, rc, handle);
1190 * read lov EA of an object
1191 * return the lov EA in an allocated lu_buf
1193 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1194 struct lu_buf *lmm_buf)
1196 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1201 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1203 if (rc == -ERANGE) {
1204 /* mti_big_buf is allocated but is too small
1205 * we need to increase it */
1206 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1208 if (buf->lb_buf == NULL)
1209 GOTO(out, rc = -ENOMEM);
1220 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1221 /* mti_big_buf was not allocated, so we have to
1222 * allocate it based on the ea size */
1223 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1225 if (buf->lb_buf == NULL)
1226 GOTO(out, rc = -ENOMEM);
1230 lu_buf_alloc(lmm_buf, bufsize);
1231 if (lmm_buf->lb_buf == NULL)
1232 GOTO(out, rc = -ENOMEM);
1234 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1240 lu_buf_free(lmm_buf);
1244 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1245 struct mdd_object *o, struct lu_buf *buf,
1246 struct thandle *handle)
1248 struct hsm_attrs *attrs;
1254 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1259 attrs = buf->lb_buf;
1260 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1261 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1264 /* Add a changelog record for release. */
1265 hsm_set_cl_event(&flags, HE_RELEASE);
1266 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1272 * check if layout swapping between 2 objects is allowed
1274 * - only normal FIDs or non-system IGIFs
1275 * - same type of objects
1276 * - same owner/group (so quotas are still valid)
1278 static int mdd_layout_swap_allowed(const struct lu_env *env,
1279 struct mdd_object *o1,
1280 const struct lu_attr *attr1,
1281 struct mdd_object *o2,
1282 const struct lu_attr *attr2)
1284 const struct lu_fid *fid1, *fid2;
1290 if (!fid_is_norm(fid1) &&
1291 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1294 if (!fid_is_norm(fid2) &&
1295 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1298 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1299 if (S_ISDIR(mdd_object_type(o1)))
1301 if (S_ISREG(mdd_object_type(o1)))
1306 if ((attr1->la_uid != attr2->la_uid) ||
1307 (attr1->la_gid != attr2->la_gid))
1314 * swap layouts between 2 lustre objects
1316 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1317 struct md_object *obj2, __u64 flags)
1319 struct mdd_thread_info *info = mdd_env_info(env);
1320 struct mdd_object *fst_o = md2mdd_obj(obj1);
1321 struct mdd_object *snd_o = md2mdd_obj(obj2);
1322 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1323 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1324 struct mdd_device *mdd = mdo2mdd(obj1);
1325 struct lov_mds_md *fst_lmm, *snd_lmm;
1326 struct lu_buf *fst_buf = &info->mti_buf[0];
1327 struct lu_buf *snd_buf = &info->mti_buf[1];
1328 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1329 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1330 struct ost_id *saved_oi = NULL;
1331 struct thandle *handle;
1332 __u16 fst_gen, snd_gen;
1338 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1339 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1341 /* we have to sort the 2 obj, so locking will always
1342 * be in the same order, even in case of 2 concurrent swaps */
1343 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1344 if (rc == 0) /* same fid ? */
1350 rc = mdd_la_get(env, fst_o, fst_la);
1354 rc = mdd_la_get(env, snd_o, snd_la);
1358 /* check if layout swapping is allowed */
1359 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la);
1363 handle = mdd_trans_create(env, mdd);
1365 RETURN(PTR_ERR(handle));
1367 /* objects are already sorted */
1368 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1369 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1371 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1372 if (rc < 0 && rc != -ENODATA)
1375 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1376 if (rc < 0 && rc != -ENODATA)
1379 /* swapping 2 non existant layouts is a success */
1380 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1383 /* to help inode migration between MDT, it is better to
1384 * start by the no layout file (if one), so we order the swap */
1385 if (snd_buf->lb_buf == NULL) {
1387 swap(fst_buf, snd_buf);
1390 /* lmm and generation layout initialization */
1391 if (fst_buf->lb_buf != NULL) {
1392 fst_lmm = fst_buf->lb_buf;
1393 fst_gen = le16_to_cpu(fst_lmm->lmm_layout_gen);
1394 fst_fl = LU_XATTR_REPLACE;
1398 fst_fl = LU_XATTR_CREATE;
1401 snd_lmm = snd_buf->lb_buf;
1402 snd_gen = le16_to_cpu(snd_lmm->lmm_layout_gen);
1404 /* increase the generation layout numbers */
1408 /* set the file specific informations in lmm */
1409 if (fst_lmm != NULL) {
1410 saved_oi = &info->mti_oa.o_oi;
1412 *saved_oi = fst_lmm->lmm_oi;
1413 fst_lmm->lmm_layout_gen = cpu_to_le16(snd_gen);
1414 fst_lmm->lmm_oi = snd_lmm->lmm_oi;
1415 snd_lmm->lmm_oi = *saved_oi;
1417 if (snd_lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V1))
1418 snd_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1419 else if (snd_lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3))
1420 snd_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
1422 GOTO(stop, rc = -EPROTO);
1424 snd_lmm->lmm_layout_gen = cpu_to_le16(fst_gen);
1426 /* Prepare HSM attribute if it's required */
1427 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1428 const int buflen = sizeof(struct hsm_attrs);
1430 lu_buf_alloc(fst_hsm_buf, buflen);
1431 lu_buf_alloc(snd_hsm_buf, buflen);
1432 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1433 GOTO(stop, rc = -ENOMEM);
1435 /* Read HSM attribute */
1436 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1440 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1444 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1445 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1450 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1451 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1457 /* prepare transaction */
1458 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1463 if (fst_buf->lb_buf != NULL)
1464 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1465 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1468 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1473 rc = mdd_trans_start(env, mdd, handle);
1477 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1478 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1482 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1484 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1487 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1488 mdd_obj_dev_name(fst_o),
1489 PFID(mdo2fid(fst_o)), rc, rc2);
1494 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1498 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1501 if (fst_buf->lb_buf != NULL)
1502 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1503 LU_XATTR_REPLACE, handle);
1505 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1511 /* failure on second file, but first was done, so we have
1512 * to roll back first. */
1513 if (fst_buf->lb_buf != NULL) {
1514 fst_lmm->lmm_oi = *saved_oi;
1515 fst_lmm->lmm_layout_gen = cpu_to_le16(fst_gen - 1);
1516 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1517 LU_XATTR_REPLACE, handle);
1519 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1525 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1530 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1535 CERROR("%s: unable to roll back layout swap. FIDs: "
1536 DFID" and "DFID "error: %d/%d, steps: %d\n",
1537 mdd_obj_dev_name(fst_o),
1538 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1540 /* a solution to avoid journal commit is to panic,
1541 * but it has strong consequences so we use LBUG to
1542 * allow sysdamin to choose to panic or not
1549 /* Issue one changelog record per file */
1550 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1554 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1560 mdd_trans_stop(env, mdd, rc, handle);
1561 mdd_write_unlock(env, snd_o);
1562 mdd_write_unlock(env, fst_o);
1564 lu_buf_free(fst_buf);
1565 lu_buf_free(snd_buf);
1566 lu_buf_free(fst_hsm_buf);
1567 lu_buf_free(snd_hsm_buf);
1571 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
1572 struct mdd_object *child, const struct lu_attr *attr,
1573 const struct md_op_spec *spec,
1574 struct dt_allocation_hint *hint)
1576 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
1577 struct dt_object *nc = mdd_object_child(child);
1579 memset(hint, 0, sizeof(*hint));
1581 /* For striped directory, give striping EA to lod_ah_init, which will
1582 * decide the stripe_offset and stripe count by it. */
1583 if (S_ISDIR(attr->la_mode) &&
1584 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
1585 hint->dah_eadata = spec->u.sp_ea.eadata;
1586 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
1588 hint->dah_eadata = NULL;
1589 hint->dah_eadata_len = 0;
1592 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
1593 hint->dah_eadata, hint->dah_eadata_len);
1594 /* @hint will be initialized by underlying device. */
1595 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
1599 * do NOT or the MAY_*'s, you'll get the weakest
1601 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
1605 /* Sadly, NFSD reopens a file repeatedly during operation, so the
1606 * "acc_mode = 0" allowance for newly-created files isn't honoured.
1607 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
1608 * owner can write to a file even if it is marked readonly to hide
1609 * its brokenness. (bug 5781) */
1610 if (flags & MDS_OPEN_OWNEROVERRIDE) {
1611 struct lu_ucred *uc = lu_ucred_check(env);
1613 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
1617 if (flags & FMODE_READ)
1619 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
1621 if (flags & MDS_FMODE_EXEC)
1626 static int mdd_open_sanity_check(const struct lu_env *env,
1627 struct mdd_object *obj,
1628 const struct lu_attr *attr, int flag)
1634 if (mdd_is_dead_obj(obj))
1637 if (S_ISLNK(attr->la_mode))
1640 mode = accmode(env, attr, flag);
1642 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
1645 if (!(flag & MDS_OPEN_CREATED)) {
1646 rc = mdd_permission_internal(env, obj, attr, mode);
1651 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
1652 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
1653 flag &= ~MDS_OPEN_TRUNC;
1655 /* For writing append-only file must open it with append mode. */
1656 if (attr->la_flags & LUSTRE_APPEND_FL) {
1657 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
1659 if (flag & MDS_OPEN_TRUNC)
1665 * Now, flag -- O_NOATIME does not be packed by client.
1667 if (flag & O_NOATIME) {
1668 struct lu_ucred *uc = lu_ucred(env);
1670 if (uc && ((uc->uc_valid == UCRED_OLD) ||
1671 (uc->uc_valid == UCRED_NEW)) &&
1672 (uc->uc_fsuid != attr->la_uid) &&
1673 !md_capable(uc, CFS_CAP_FOWNER))
1681 static int mdd_open(const struct lu_env *env, struct md_object *obj,
1684 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1685 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1688 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1690 rc = mdd_la_get(env, mdd_obj, attr);
1694 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
1698 mdd_obj->mod_count++;
1701 mdd_write_unlock(env, mdd_obj);
1705 static int mdd_declare_close(const struct lu_env *env,
1706 struct mdd_object *obj,
1708 struct thandle *handle)
1712 rc = orph_declare_index_delete(env, obj, handle);
1716 return mdo_declare_destroy(env, obj, handle);
1720 * No permission check is needed.
1722 static int mdd_close(const struct lu_env *env, struct md_object *obj,
1723 struct md_attr *ma, int mode)
1725 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1726 struct mdd_device *mdd = mdo2mdd(obj);
1727 struct thandle *handle = NULL;
1728 int rc, is_orphan = 0;
1731 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
1732 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1733 mdd_obj->mod_count--;
1734 mdd_write_unlock(env, mdd_obj);
1736 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
1737 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
1738 "list\n", PFID(mdd_object_fid(mdd_obj)));
1742 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
1743 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
1744 /* check without any lock */
1745 is_orphan = mdd_obj->mod_count == 1 &&
1746 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
1750 handle = mdd_trans_create(env, mdo2mdd(obj));
1752 RETURN(PTR_ERR(handle));
1754 rc = mdd_declare_close(env, mdd_obj, ma, handle);
1758 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1762 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
1767 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1768 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
1770 CERROR("Failed to get lu_attr of "DFID": %d\n",
1771 PFID(mdd_object_fid(mdd_obj)), rc);
1775 /* check again with lock */
1776 is_orphan = (mdd_obj->mod_count == 1) &&
1777 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
1778 ma->ma_attr.la_nlink == 0);
1780 if (is_orphan && handle == NULL) {
1781 mdd_write_unlock(env, mdd_obj);
1785 mdd_obj->mod_count--; /*release open count */
1791 /* NB: Object maybe not in orphan list originally, it is rare case for
1792 * mdd_finish_unlink() failure, in that case, the object doesn't have
1793 * ORPHAN_OBJ flag */
1794 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
1795 /* remove link to object from orphan index */
1796 LASSERT(handle != NULL);
1797 rc = __mdd_orphan_del(env, mdd_obj, handle);
1799 CERROR("%s: unable to delete "DFID" from orphan list: "
1800 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
1801 PFID(mdd_object_fid(mdd_obj)), rc);
1802 /* If object was not deleted from orphan list, do not
1803 * destroy OSS objects, which will be done when next
1808 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
1809 "list, OSS objects to be destroyed.\n",
1810 PFID(mdd_object_fid(mdd_obj)));
1813 rc = mdo_destroy(env, mdd_obj, handle);
1816 CERROR("%s: unable to delete "DFID" from orphan list: "
1817 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
1818 PFID(mdd_object_fid(mdd_obj)), rc);
1823 mdd_write_unlock(env, mdd_obj);
1826 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
1827 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
1828 if (handle == NULL) {
1829 handle = mdd_trans_create(env, mdo2mdd(obj));
1831 GOTO(stop, rc = PTR_ERR(handle));
1833 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
1838 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
1843 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
1848 if (handle != NULL && !IS_ERR(handle))
1849 mdd_trans_stop(env, mdd, rc, handle);
1855 * Permission check is done when open,
1856 * no need check again.
1858 static int mdd_readpage_sanity_check(const struct lu_env *env,
1859 struct mdd_object *obj)
1861 struct dt_object *next = mdd_object_child(obj);
1865 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
1873 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
1874 size_t nob, const struct dt_it_ops *iops,
1875 struct dt_it *it, __u32 attr, void *arg)
1877 struct lu_dirpage *dp = &lp->lp_dir;
1881 struct lu_dirent *ent;
1882 struct lu_dirent *last = NULL;
1886 if (nob < sizeof(*dp))
1889 memset(area, 0, sizeof (*dp));
1890 area += sizeof (*dp);
1891 nob -= sizeof (*dp);
1898 len = iops->key_size(env, it);
1900 /* IAM iterator can return record with zero len. */
1904 hash = iops->store(env, it);
1905 if (unlikely(first)) {
1907 dp->ldp_hash_start = cpu_to_le64(hash);
1910 /* calculate max space required for lu_dirent */
1911 recsize = lu_dirent_calc_size(len, attr);
1913 if (nob >= recsize) {
1914 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
1915 if (result == -ESTALE)
1920 /* osd might not able to pack all attributes,
1921 * so recheck rec length */
1922 recsize = le16_to_cpu(ent->lde_reclen);
1924 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
1925 fid_le_to_cpu(&fid, &ent->lde_fid);
1926 if (fid_is_dot_lustre(&fid))
1930 result = (last != NULL) ? 0 :-EINVAL;
1934 ent = (void *)ent + recsize;
1938 result = iops->next(env, it);
1939 if (result == -ESTALE)
1941 } while (result == 0);
1944 dp->ldp_hash_end = cpu_to_le64(hash);
1946 if (last->lde_hash == dp->ldp_hash_end)
1947 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
1948 last->lde_reclen = 0; /* end mark */
1951 /* end of directory */
1952 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
1953 else if (result < 0)
1954 CWARN("build page failed: %d!\n", result);
1958 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
1959 const struct lu_rdpg *rdpg)
1961 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1965 if (mdd_object_exists(mdd_obj) == 0) {
1966 CERROR("%s: object "DFID" not found: rc = -2\n",
1967 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
1971 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
1972 rc = mdd_readpage_sanity_check(env, mdd_obj);
1974 GOTO(out_unlock, rc);
1976 if (mdd_is_dead_obj(mdd_obj)) {
1978 struct lu_dirpage *dp;
1981 * According to POSIX, please do not return any entry to client:
1982 * even dot and dotdot should not be returned.
1984 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
1985 PFID(mdd_object_fid(mdd_obj)));
1987 if (rdpg->rp_count <= 0)
1988 GOTO(out_unlock, rc = -EFAULT);
1989 LASSERT(rdpg->rp_pages != NULL);
1991 pg = rdpg->rp_pages[0];
1992 dp = (struct lu_dirpage *)kmap(pg);
1993 memset(dp, 0 , sizeof(struct lu_dirpage));
1994 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
1995 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
1996 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
1998 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2001 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2002 mdd_dir_page_build, NULL);
2004 struct lu_dirpage *dp;
2006 dp = kmap(rdpg->rp_pages[0]);
2007 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2010 * No pages were processed, mark this for first page
2013 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2014 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2015 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2017 kunmap(rdpg->rp_pages[0]);
2020 GOTO(out_unlock, rc);
2022 mdd_read_unlock(env, mdd_obj);
2026 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2028 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2030 if (mdd_object_exists(mdd_obj) == 0) {
2033 CERROR("%s: object "DFID" not found: rc = %d\n",
2034 mdd_obj_dev_name(mdd_obj),
2035 PFID(mdd_object_fid(mdd_obj)), rc);
2038 return dt_object_sync(env, mdd_object_child(mdd_obj),
2042 static int mdd_object_lock(const struct lu_env *env,
2043 struct md_object *obj,
2044 struct lustre_handle *lh,
2045 struct ldlm_enqueue_info *einfo,
2046 union ldlm_policy_data *policy)
2048 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2049 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2053 static int mdd_object_unlock(const struct lu_env *env,
2054 struct md_object *obj,
2055 struct ldlm_enqueue_info *einfo,
2056 union ldlm_policy_data *policy)
2058 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2059 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2062 const struct md_object_operations mdd_obj_ops = {
2063 .moo_permission = mdd_permission,
2064 .moo_attr_get = mdd_attr_get,
2065 .moo_attr_set = mdd_attr_set,
2066 .moo_xattr_get = mdd_xattr_get,
2067 .moo_xattr_set = mdd_xattr_set,
2068 .moo_xattr_list = mdd_xattr_list,
2069 .moo_xattr_del = mdd_xattr_del,
2070 .moo_swap_layouts = mdd_swap_layouts,
2071 .moo_open = mdd_open,
2072 .moo_close = mdd_close,
2073 .moo_readpage = mdd_readpage,
2074 .moo_readlink = mdd_readlink,
2075 .moo_changelog = mdd_changelog,
2076 .moo_object_sync = mdd_object_sync,
2077 .moo_object_lock = mdd_object_lock,
2078 .moo_object_unlock = mdd_object_unlock,