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_param.h>
50 #include <lustre_mds.h>
51 #include <lustre/lustre_idl.h>
53 #include "mdd_internal.h"
55 static const struct lu_object_operations mdd_lu_obj_ops;
57 static int mdd_xattr_get(const struct lu_env *env,
58 struct md_object *obj, struct lu_buf *buf,
61 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
66 if (mdd_object_exists(obj) == 0)
69 rc = mdo_attr_get(env, obj, la);
70 if (unlikely(rc != 0)) {
72 obj->mod_flags |= DEAD_OBJ;
76 if (la->la_valid & LA_FLAGS &&
77 la->la_flags & LUSTRE_ORPHAN_FL)
78 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
83 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
85 struct mdd_thread_info *info;
87 lu_env_refill((struct lu_env *)env);
88 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
89 LASSERT(info != NULL);
93 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
97 buf = &mdd_env_info(env)->mti_buf[0];
103 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
104 const void *area, ssize_t len)
108 buf = &mdd_env_info(env)->mti_buf[0];
109 buf->lb_buf = (void *)area;
114 struct lu_object *mdd_object_alloc(const struct lu_env *env,
115 const struct lu_object_header *hdr,
118 struct mdd_object *mdd_obj;
120 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
121 if (mdd_obj != NULL) {
124 o = mdd2lu_obj(mdd_obj);
125 lu_object_init(o, NULL, d);
126 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
127 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
128 mdd_obj->mod_count = 0;
129 o->lo_ops = &mdd_lu_obj_ops;
136 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
137 const struct lu_object_conf *unused)
139 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
140 struct mdd_object *mdd_obj = lu2mdd_obj(o);
141 struct lu_object *below;
142 struct lu_device *under;
145 mdd_obj->mod_cltime = 0;
146 under = &d->mdd_child->dd_lu_dev;
147 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
149 RETURN(PTR_ERR(below));
151 lu_object_add(o, below);
156 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
160 if (lu_object_exists(o)) {
161 struct mdd_object *mdd_obj = lu2mdd_obj(o);
162 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
164 rc = mdd_la_get(env, mdd_obj, attr);
170 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
172 struct mdd_object *mdd = lu2mdd_obj(o);
175 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
178 static int mdd_object_print(const struct lu_env *env, void *cookie,
179 lu_printer_t p, const struct lu_object *o)
181 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
183 return (*p)(env, cookie,
184 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%llu, flags=%lx)",
185 mdd, mdd->mod_count, mdd->mod_valid,
186 mdd->mod_cltime, mdd->mod_flags);
189 static const struct lu_object_operations mdd_lu_obj_ops = {
190 .loo_object_init = mdd_object_init,
191 .loo_object_start = mdd_object_start,
192 .loo_object_free = mdd_object_free,
193 .loo_object_print = mdd_object_print,
196 struct mdd_object *mdd_object_find(const struct lu_env *env,
197 struct mdd_device *d,
198 const struct lu_fid *f)
200 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
204 * No permission check is needed.
206 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
209 struct mdd_object *mdd_obj = md2mdd_obj(obj);
214 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
215 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
216 ma->ma_attr.la_nlink = 0;
222 * No permission check is needed.
224 static int mdd_xattr_get(const struct lu_env *env,
225 struct md_object *obj, struct lu_buf *buf,
228 struct mdd_object *mdd_obj = md2mdd_obj(obj);
233 if (mdd_object_exists(mdd_obj) == 0) {
234 CERROR("%s: object "DFID" not found: rc = -2\n",
235 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
239 /* If the object has been destroyed, then do not get LMVEA, because
240 * it needs to load stripes from the iteration of the master object,
241 * and it will cause problem if master object has been destroyed, see
243 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
244 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
245 strcmp(name, XATTR_NAME_LMV) == 0))
248 /* If the object has been delete from the namespace, then
249 * get linkEA should return -ENOENT as well */
250 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
251 strcmp(name, XATTR_NAME_LINK) == 0))
254 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
255 rc = mdo_xattr_get(env, mdd_obj, buf, name);
256 mdd_read_unlock(env, mdd_obj);
262 * Permission check is done when open,
263 * no need check again.
265 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
268 struct mdd_object *mdd_obj = md2mdd_obj(obj);
269 struct dt_object *next;
274 if (mdd_object_exists(mdd_obj) == 0) {
275 CERROR("%s: object "DFID" not found: rc = -2\n",
276 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
280 next = mdd_object_child(mdd_obj);
281 LASSERT(next != NULL);
282 LASSERT(next->do_body_ops != NULL);
283 LASSERT(next->do_body_ops->dbo_read != NULL);
284 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
285 rc = dt_read(env, next, buf, &pos);
286 mdd_read_unlock(env, mdd_obj);
291 * No permission check is needed.
293 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
296 struct mdd_object *mdd_obj = md2mdd_obj(obj);
301 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
302 rc = mdo_xattr_list(env, mdd_obj, buf);
303 mdd_read_unlock(env, mdd_obj);
309 * Filter out XATTR_NAME_LINK if this is an orphan object. See
312 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
313 char *end = (char *)buf->lb_buf + rc;
314 char *p = buf->lb_buf;
317 char *next = p + strlen(p) + 1;
319 if (strcmp(p, XATTR_NAME_LINK) == 0) {
321 memmove(p, next, end - next);
323 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
324 " of orphan "DFID"\n",
325 PFID(mdd_object_fid(mdd_obj)));
336 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
338 return mdo_invalidate(env, md2mdd_obj(obj));
341 int mdd_declare_create_object_internal(const struct lu_env *env,
342 struct mdd_object *p,
343 struct mdd_object *c,
344 struct lu_attr *attr,
345 struct thandle *handle,
346 const struct md_op_spec *spec,
347 struct dt_allocation_hint *hint)
349 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
350 const struct dt_index_features *feat = spec->sp_feat;
354 if (feat != &dt_directory_features && feat != NULL) {
355 dof->dof_type = DFT_INDEX;
356 dof->u.dof_idx.di_feat = feat;
358 dof->dof_type = dt_mode_to_dft(attr->la_mode);
359 if (dof->dof_type == DFT_REGULAR) {
360 dof->u.dof_reg.striped =
361 md_should_create(spec->sp_cr_flags);
362 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
363 dof->u.dof_reg.striped = 0;
364 /* is this replay? */
366 dof->u.dof_reg.striped = 0;
370 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
375 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
376 struct mdd_object *c, struct lu_attr *attr,
377 struct thandle *handle,
378 const struct md_op_spec *spec,
379 struct dt_allocation_hint *hint)
381 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
385 LASSERT(!mdd_object_exists(c));
387 rc = mdo_create_object(env, c, attr, hint, dof, handle);
392 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
393 const struct lu_attr *attr, struct thandle *handle,
399 rc = mdo_attr_set(env, obj, attr, handle);
400 #ifdef CONFIG_FS_POSIX_ACL
401 if (!rc && (attr->la_valid & LA_MODE) && needacl)
402 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
407 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
408 const struct lu_attr *oattr, struct lu_attr *attr,
409 struct thandle *handle)
414 LASSERT(attr->la_valid & LA_CTIME);
415 LASSERT(oattr != NULL);
417 /* Make sure the ctime is increased only, however, it's not strictly
418 * reliable at here because there is not guarantee to hold lock on
419 * object, so we just bypass some unnecessary cmtime setting first
420 * and OSD has to check it again. */
421 if (attr->la_ctime < oattr->la_ctime)
422 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
423 else if (attr->la_valid == LA_CTIME &&
424 attr->la_ctime == oattr->la_ctime)
425 attr->la_valid &= ~LA_CTIME;
427 if (attr->la_valid != 0)
428 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
433 * This gives the same functionality as the code between
434 * sys_chmod and inode_setattr
435 * chown_common and inode_setattr
436 * utimes and inode_setattr
437 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
439 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
440 const struct lu_attr *oattr, struct lu_attr *la,
441 const unsigned long flags)
450 /* Do not permit change file type */
451 if (la->la_valid & LA_TYPE)
454 /* They should not be processed by setattr */
455 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
458 LASSERT(oattr != NULL);
460 uc = lu_ucred_check(env);
464 if (la->la_valid == LA_CTIME) {
465 if (!(flags & MDS_PERM_BYPASS))
466 /* This is only for set ctime when rename's source is
468 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
470 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
471 la->la_valid &= ~LA_CTIME;
475 if (la->la_valid == LA_ATIME) {
476 /* This is an atime-only attribute update for close RPCs. */
477 if (la->la_atime < (oattr->la_atime +
478 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
479 la->la_valid &= ~LA_ATIME;
483 /* Check if flags change. */
484 if (la->la_valid & LA_FLAGS) {
485 unsigned int oldflags = oattr->la_flags &
486 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
487 unsigned int newflags = la->la_flags &
488 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
490 if ((uc->uc_fsuid != oattr->la_uid) &&
491 !md_capable(uc, CFS_CAP_FOWNER))
494 /* The IMMUTABLE and APPEND_ONLY flags can
495 * only be changed by the relevant capability. */
496 if ((oldflags ^ newflags) &&
497 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
500 if (!S_ISDIR(oattr->la_mode))
501 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
504 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
505 (la->la_valid & ~LA_FLAGS) &&
506 !(flags & MDS_PERM_BYPASS))
509 /* Check for setting the obj time. */
510 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
511 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
512 if ((uc->uc_fsuid != oattr->la_uid) &&
513 !md_capable(uc, CFS_CAP_FOWNER)) {
514 rc = mdd_permission_internal(env, obj, oattr,
521 if (la->la_valid & LA_KILL_SUID) {
522 la->la_valid &= ~LA_KILL_SUID;
523 if ((oattr->la_mode & S_ISUID) &&
524 !(la->la_valid & LA_MODE)) {
525 la->la_mode = oattr->la_mode;
526 la->la_valid |= LA_MODE;
528 la->la_mode &= ~S_ISUID;
531 if (la->la_valid & LA_KILL_SGID) {
532 la->la_valid &= ~LA_KILL_SGID;
533 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
534 (S_ISGID | S_IXGRP)) &&
535 !(la->la_valid & LA_MODE)) {
536 la->la_mode = oattr->la_mode;
537 la->la_valid |= LA_MODE;
539 la->la_mode &= ~S_ISGID;
542 /* Make sure a caller can chmod. */
543 if (la->la_valid & LA_MODE) {
544 if (!(flags & MDS_PERM_BYPASS) &&
545 (uc->uc_fsuid != oattr->la_uid) &&
546 !md_capable(uc, CFS_CAP_FOWNER))
549 if (la->la_mode == (umode_t) -1)
550 la->la_mode = oattr->la_mode;
552 la->la_mode = (la->la_mode & S_IALLUGO) |
553 (oattr->la_mode & ~S_IALLUGO);
555 /* Also check the setgid bit! */
556 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
557 la->la_gid : oattr->la_gid) &&
558 !md_capable(uc, CFS_CAP_FSETID))
559 la->la_mode &= ~S_ISGID;
561 la->la_mode = oattr->la_mode;
564 /* Make sure a caller can chown. */
565 if (la->la_valid & LA_UID) {
566 if (la->la_uid == (uid_t) -1)
567 la->la_uid = oattr->la_uid;
568 if (((uc->uc_fsuid != oattr->la_uid) ||
569 (la->la_uid != oattr->la_uid)) &&
570 !md_capable(uc, CFS_CAP_CHOWN))
573 /* If the user or group of a non-directory has been
574 * changed by a non-root user, remove the setuid bit.
575 * 19981026 David C Niemi <niemi@tux.org>
577 * Changed this to apply to all users, including root,
578 * to avoid some races. This is the behavior we had in
579 * 2.0. The check for non-root was definitely wrong
580 * for 2.2 anyway, as it should have been using
581 * CAP_FSETID rather than fsuid -- 19990830 SD. */
582 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
583 !S_ISDIR(oattr->la_mode)) {
584 la->la_mode &= ~S_ISUID;
585 la->la_valid |= LA_MODE;
589 /* Make sure caller can chgrp. */
590 if (la->la_valid & LA_GID) {
591 if (la->la_gid == (gid_t) -1)
592 la->la_gid = oattr->la_gid;
593 if (((uc->uc_fsuid != oattr->la_uid) ||
594 ((la->la_gid != oattr->la_gid) &&
595 !lustre_in_group_p(uc, la->la_gid))) &&
596 !md_capable(uc, CFS_CAP_CHOWN))
599 /* Likewise, if the user or group of a non-directory
600 * has been changed by a non-root user, remove the
601 * setgid bit UNLESS there is no group execute bit
602 * (this would be a file marked for mandatory
603 * locking). 19981026 David C Niemi <niemi@tux.org>
605 * Removed the fsuid check (see the comment above) --
607 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
608 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
609 la->la_mode &= ~S_ISGID;
610 la->la_valid |= LA_MODE;
614 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
615 if (!((flags & MDS_OWNEROVERRIDE) &&
616 (uc->uc_fsuid == oattr->la_uid)) &&
617 !(flags & MDS_PERM_BYPASS)) {
618 rc = mdd_permission_internal(env, obj, oattr,
625 if (la->la_valid & LA_CTIME) {
626 /* The pure setattr, it has the priority over what is
627 * already set, do not drop it if ctime is equal. */
628 if (la->la_ctime < oattr->la_ctime)
629 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
635 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
636 struct mdd_device *mdd,
637 enum changelog_rec_type type, int flags,
638 const struct lu_fid *fid,
639 struct thandle *handle)
641 const struct lu_ucred *uc = lu_ucred(env);
642 struct llog_changelog_rec *rec;
647 flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
648 if (uc != NULL && uc->uc_jobid[0] != '\0')
651 reclen = llog_data_len(changelog_rec_offset(flags & CLF_SUPPORTED));
652 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
653 if (buf->lb_buf == NULL)
657 rec->cr.cr_flags = flags;
658 rec->cr.cr_type = (__u32)type;
659 rec->cr.cr_tfid = *fid;
660 rec->cr.cr_namelen = 0;
662 if (flags & CLF_JOBID)
663 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
665 rc = mdd_changelog_store(env, mdd, rec, handle);
671 /** Store a data change changelog record
672 * If this fails, we must fail the whole transaction; we don't
673 * want the change to commit without the log entry.
674 * \param mdd_obj - mdd_object of change
675 * \param handle - transaction handle
677 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
678 enum changelog_rec_type type, int flags,
679 struct mdd_object *mdd_obj, struct thandle *handle)
683 LASSERT(mdd_obj != NULL);
684 LASSERT(handle != NULL);
687 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
689 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
692 if (mdd_is_volatile_obj(mdd_obj))
695 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
696 cfs_time_before_64(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
697 /* Don't need multiple updates in this log */
698 /* Don't check under lock - no big deal if we get an extra
703 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
704 mdo2fid(mdd_obj), handle);
706 mdd_obj->mod_cltime = cfs_time_current_64();
711 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
712 int flags, struct md_device *m, const struct lu_fid *fid)
714 struct thandle *handle;
715 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
720 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
722 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
725 LASSERT(fid != NULL);
727 handle = mdd_trans_create(env, mdd);
729 RETURN(PTR_ERR(handle));
731 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
735 rc = mdd_trans_start(env, mdd, handle);
739 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
743 rc = mdd_trans_stop(env, mdd, rc, handle);
749 * Save LMA extended attributes with data from \a ma.
751 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
752 * not, LMA EA will be first read from disk, modified and write back.
755 /* Precedence for choosing record type when multiple
756 * attributes change: setattr > mtime > ctime > atime
757 * (ctime changes when mtime does, plus chmod/chown.
758 * atime and ctime are independent.) */
759 static int mdd_attr_set_changelog(const struct lu_env *env,
760 struct md_object *obj, struct thandle *handle,
763 struct mdd_device *mdd = mdo2mdd(obj);
766 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
767 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
768 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
769 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
770 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
771 bits = bits & mdd->mdd_cl.mc_mask;
772 /* This is an implementation limit rather than a protocol limit */
773 CLASSERT(CL_LAST <= sizeof(int) * 8);
777 /* The record type is the lowest non-masked set bit */
780 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
781 return mdd_changelog_data_store(env, mdd, type, (int)valid,
782 md2mdd_obj(obj), handle);
785 static int mdd_declare_attr_set(const struct lu_env *env,
786 struct mdd_device *mdd,
787 struct mdd_object *obj,
788 const struct lu_attr *attr,
789 struct thandle *handle)
793 rc = mdo_declare_attr_set(env, obj, attr, handle);
797 #ifdef CONFIG_FS_POSIX_ACL
798 if (attr->la_valid & LA_MODE) {
799 mdd_read_lock(env, obj, MOR_TGT_CHILD);
800 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
801 XATTR_NAME_ACL_ACCESS);
802 mdd_read_unlock(env, obj);
803 if (rc == -EOPNOTSUPP || rc == -ENODATA)
809 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
810 rc = mdo_declare_xattr_set(env, obj, buf,
811 XATTR_NAME_ACL_ACCESS, 0,
819 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
826 * permission changes may require sync operation, to mitigate performance
827 * impact, only do this for dir and when permission is reduced.
829 * For regular files, version is updated with permission change (see VBR), async
830 * permission won't cause any issue, while missing permission change on
831 * directory may affect accessibility of other objects after recovery.
833 static inline bool permission_needs_sync(const struct lu_attr *old,
834 const struct lu_attr *new)
836 if (!S_ISDIR(old->la_mode))
839 if (new->la_valid & (LA_UID | LA_GID))
842 if (new->la_valid & LA_MODE &&
843 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
846 if ((new->la_valid & LA_MODE) &&
847 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
848 (old->la_mode & S_IRWXUGO))
854 /* set attr and LOV EA at once, return updated attr */
855 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
856 const struct md_attr *ma)
858 struct mdd_object *mdd_obj = md2mdd_obj(obj);
859 struct mdd_device *mdd = mdo2mdd(obj);
860 struct thandle *handle;
861 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
862 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
863 const struct lu_attr *la = &ma->ma_attr;
867 /* we do not use ->attr_set() for LOV/HSM EA any more */
868 LASSERT((ma->ma_valid & MA_LOV) == 0);
869 LASSERT((ma->ma_valid & MA_HSM) == 0);
871 rc = mdd_la_get(env, mdd_obj, attr);
875 *la_copy = ma->ma_attr;
876 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
880 /* no need to setattr anymore */
881 if (la_copy->la_valid == 0) {
882 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
883 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
884 PFID(mdo2fid(mdd_obj)), la->la_valid);
889 handle = mdd_trans_create(env, mdd);
891 RETURN(PTR_ERR(handle));
893 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
897 rc = mdd_trans_start(env, mdd, handle);
901 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
904 if (la->la_valid & (LA_MTIME | LA_CTIME))
905 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
906 la->la_mtime, la->la_ctime);
908 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
909 if (la_copy->la_valid & LA_FLAGS)
910 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
911 else if (la_copy->la_valid) /* setattr */
912 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
913 mdd_write_unlock(env, mdd_obj);
916 rc = mdd_attr_set_changelog(env, obj, handle,
922 rc = mdd_trans_stop(env, mdd, rc, handle);
927 static int mdd_xattr_sanity_check(const struct lu_env *env,
928 struct mdd_object *obj,
929 const struct lu_attr *attr,
932 struct lu_ucred *uc = lu_ucred_assert(env);
935 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
938 if (strncmp(XATTR_USER_PREFIX, name,
939 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
940 /* For sticky directories, only the owner and privileged user
941 * can write attributes. */
942 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
943 (uc->uc_fsuid != attr->la_uid) &&
944 !md_capable(uc, CFS_CAP_FOWNER))
947 if ((uc->uc_fsuid != attr->la_uid) &&
948 !md_capable(uc, CFS_CAP_FOWNER))
956 * Check if a string begins with a given prefix.
958 * \param str String to check
959 * \param prefix Substring to check at the beginning of \a str
960 * \return true/false whether the condition is verified.
962 static inline bool has_prefix(const char *str, const char *prefix)
964 return strncmp(prefix, str, strlen(prefix)) == 0;
968 * Indicate the kind of changelog to store (if any) for a xattr set/del.
970 * \param[in] xattr_name Full extended attribute name.
972 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
974 static enum changelog_rec_type
975 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
976 const char *xattr_name)
978 /* Layout changes systematically recorded */
979 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
980 strncmp(XATTR_LUSTRE_LOV, xattr_name,
981 strlen(XATTR_LUSTRE_LOV)) == 0)
984 /* HSM information changes systematically recorded */
985 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
988 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
989 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
990 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
996 static int mdd_declare_xattr_set(const struct lu_env *env,
997 struct mdd_device *mdd,
998 struct mdd_object *obj,
999 const struct lu_buf *buf,
1001 int fl, struct thandle *handle)
1005 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1009 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1010 return 0; /* no changelog to store */
1012 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1016 * Compare current and future data of HSM EA and add a changelog if needed.
1018 * Caller should have write-locked \param obj.
1020 * \param buf - Future HSM EA content.
1021 * \retval 0 if no changelog is needed or changelog was added properly.
1022 * \retval -ve errno if there was a problem
1024 static int mdd_hsm_update_locked(const struct lu_env *env,
1025 struct md_object *obj,
1026 const struct lu_buf *buf,
1027 struct thandle *handle, int *cl_flags)
1029 struct mdd_thread_info *info = mdd_env_info(env);
1030 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1031 struct lu_buf *current_buf;
1032 struct md_hsm *current_mh;
1033 struct md_hsm *new_mh;
1037 OBD_ALLOC_PTR(current_mh);
1038 if (current_mh == NULL)
1041 /* Read HSM attrs from disk */
1042 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1043 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1044 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1045 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1046 if (rc < 0 && rc != -ENODATA)
1048 else if (rc == -ENODATA)
1049 current_mh->mh_flags = 0;
1051 /* Map future HSM xattr */
1052 OBD_ALLOC_PTR(new_mh);
1054 GOTO(free, rc = -ENOMEM);
1055 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1059 /* Flags differ, set flags for the changelog that will be added */
1060 if (current_mh->mh_flags != new_mh->mh_flags) {
1061 hsm_set_cl_event(cl_flags, HE_STATE);
1062 if (new_mh->mh_flags & HS_DIRTY)
1063 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1066 OBD_FREE_PTR(new_mh);
1069 OBD_FREE_PTR(current_mh);
1073 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1077 * The caller should guarantee to update the object ctime
1078 * after xattr_set if needed.
1080 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1081 const struct lu_buf *buf, const char *name,
1084 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1085 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1086 struct mdd_device *mdd = mdo2mdd(obj);
1087 struct thandle *handle;
1088 enum changelog_rec_type cl_type;
1093 rc = mdd_la_get(env, mdd_obj, attr);
1097 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1101 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1102 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1103 struct posix_acl *acl;
1105 /* user may set empty ACL, which should be treated as removing
1107 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1110 RETURN(PTR_ERR(acl));
1112 rc = mdd_xattr_del(env, obj, name);
1115 posix_acl_release(acl);
1118 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1119 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1123 handle = mdd_trans_create(env, mdd);
1125 RETURN(PTR_ERR(handle));
1127 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1131 rc = mdd_trans_start(env, mdd, handle);
1135 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1137 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1138 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1140 mdd_write_unlock(env, mdd_obj);
1145 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1146 mdd_write_unlock(env, mdd_obj);
1150 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1154 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1159 return mdd_trans_stop(env, mdd, rc, handle);
1162 static int mdd_declare_xattr_del(const struct lu_env *env,
1163 struct mdd_device *mdd,
1164 struct mdd_object *obj,
1166 struct thandle *handle)
1170 rc = mdo_declare_xattr_del(env, obj, name, handle);
1174 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1175 return 0; /* no changelog to store */
1177 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1181 * The caller should guarantee to update the object ctime
1182 * after xattr_set if needed.
1184 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1187 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1188 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1189 struct mdd_device *mdd = mdo2mdd(obj);
1190 struct thandle *handle;
1194 rc = mdd_la_get(env, mdd_obj, attr);
1198 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1202 handle = mdd_trans_create(env, mdd);
1204 RETURN(PTR_ERR(handle));
1206 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1210 rc = mdd_trans_start(env, mdd, handle);
1214 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1215 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1216 mdd_write_unlock(env, mdd_obj);
1220 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1223 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1227 return mdd_trans_stop(env, mdd, rc, handle);
1231 * read lov EA of an object
1232 * return the lov EA in an allocated lu_buf
1234 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1235 struct lu_buf *lmm_buf)
1237 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1242 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1244 if (rc == -ERANGE) {
1245 /* mti_big_buf is allocated but is too small
1246 * we need to increase it */
1247 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1249 if (buf->lb_buf == NULL)
1250 GOTO(out, rc = -ENOMEM);
1261 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1262 /* mti_big_buf was not allocated, so we have to
1263 * allocate it based on the ea size */
1264 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1266 if (buf->lb_buf == NULL)
1267 GOTO(out, rc = -ENOMEM);
1271 lu_buf_alloc(lmm_buf, bufsize);
1272 if (lmm_buf->lb_buf == NULL)
1273 GOTO(out, rc = -ENOMEM);
1275 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1281 lu_buf_free(lmm_buf);
1285 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1286 struct mdd_object *o, struct lu_buf *buf,
1287 struct thandle *handle)
1289 struct hsm_attrs *attrs;
1295 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1300 attrs = buf->lb_buf;
1301 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1302 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1305 /* Add a changelog record for release. */
1306 hsm_set_cl_event(&flags, HE_RELEASE);
1307 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1313 * check if layout swapping between 2 objects is allowed
1315 * - only normal FIDs or non-system IGIFs
1316 * - same type of objects
1317 * - same owner/group (so quotas are still valid) unless this is from HSM
1320 static int mdd_layout_swap_allowed(const struct lu_env *env,
1321 struct mdd_object *o1,
1322 const struct lu_attr *attr1,
1323 struct mdd_object *o2,
1324 const struct lu_attr *attr2,
1327 const struct lu_fid *fid1, *fid2;
1333 if (!fid_is_norm(fid1) &&
1334 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1337 if (!fid_is_norm(fid2) &&
1338 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1341 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1342 if (S_ISDIR(mdd_object_type(o1)))
1344 if (S_ISREG(mdd_object_type(o1)))
1349 if (flags & SWAP_LAYOUTS_MDS_HSM)
1352 if ((attr1->la_uid != attr2->la_uid) ||
1353 (attr1->la_gid != attr2->la_gid))
1359 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1360 * look into the layout in MDD layer. */
1361 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1363 struct lov_comp_md_v1 *comp_v1;
1364 struct lov_mds_md *v1;
1368 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1369 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1370 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1376 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1377 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1380 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1381 off = le32_to_cpu(comp_v1->lcm_entries[i].
1383 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1388 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1389 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1400 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1402 return mdd_lmm_oi(lmm, oi, true);
1405 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1407 return mdd_lmm_oi(lmm, oi, false);
1410 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1412 struct lov_comp_md_v1 *comp_v1;
1414 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1415 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1417 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1419 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1420 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1421 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1422 __u16 tmp_gen = *gen;
1424 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1426 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1433 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1435 return mdd_lmm_gen(lmm, gen, true);
1438 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1440 return mdd_lmm_gen(lmm, gen, false);
1444 * swap layouts between 2 lustre objects
1446 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1447 struct md_object *obj2, __u64 flags)
1449 struct mdd_thread_info *info = mdd_env_info(env);
1450 struct mdd_object *fst_o = md2mdd_obj(obj1);
1451 struct mdd_object *snd_o = md2mdd_obj(obj2);
1452 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1453 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1454 struct mdd_device *mdd = mdo2mdd(obj1);
1455 struct lov_mds_md *fst_lmm, *snd_lmm;
1456 struct lu_buf *fst_buf = &info->mti_buf[0];
1457 struct lu_buf *snd_buf = &info->mti_buf[1];
1458 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1459 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1460 struct ost_id *saved_oi = NULL;
1461 struct thandle *handle;
1462 __u32 fst_gen, snd_gen, saved_gen;
1468 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1469 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1471 /* we have to sort the 2 obj, so locking will always
1472 * be in the same order, even in case of 2 concurrent swaps */
1473 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1474 if (rc == 0) /* same fid ? */
1480 rc = mdd_la_get(env, fst_o, fst_la);
1484 rc = mdd_la_get(env, snd_o, snd_la);
1488 /* check if layout swapping is allowed */
1489 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1493 handle = mdd_trans_create(env, mdd);
1495 RETURN(PTR_ERR(handle));
1497 /* objects are already sorted */
1498 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1499 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1501 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1502 if (rc < 0 && rc != -ENODATA)
1505 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1506 if (rc < 0 && rc != -ENODATA)
1509 /* swapping 2 non existant layouts is a success */
1510 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1513 /* to help inode migration between MDT, it is better to
1514 * start by the no layout file (if one), so we order the swap */
1515 if (snd_buf->lb_buf == NULL) {
1517 swap(fst_buf, snd_buf);
1520 fst_gen = snd_gen = 0;
1521 /* lmm and generation layout initialization */
1522 if (fst_buf->lb_buf != NULL) {
1523 fst_lmm = fst_buf->lb_buf;
1524 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1525 fst_fl = LU_XATTR_REPLACE;
1529 fst_fl = LU_XATTR_CREATE;
1532 snd_lmm = snd_buf->lb_buf;
1533 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1535 saved_gen = fst_gen;
1536 /* increase the generation layout numbers */
1541 * XXX The layout generation is used to generate component IDs for
1542 * the composite file, we have to do some special tweaks to make
1543 * sure the layout generation is always adequate for that job.
1546 /* Skip invalid generation number for composite layout */
1547 if ((snd_gen & LCME_ID_MASK) == 0)
1549 if ((fst_gen & LCME_ID_MASK) == 0)
1551 /* Make sure the generation is greater than all the component IDs */
1552 if (fst_gen < snd_gen)
1554 else if (fst_gen > snd_gen)
1557 /* set the file specific informations in lmm */
1558 if (fst_lmm != NULL) {
1559 saved_oi = &info->mti_oa.o_oi;
1560 mdd_get_lmm_oi(fst_lmm, saved_oi);
1561 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1562 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1563 mdd_set_lmm_oi(snd_lmm, saved_oi);
1565 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1566 cpu_to_le32(LOV_MAGIC_MAGIC))
1567 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEF);
1569 GOTO(stop, rc = -EPROTO);
1571 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1573 /* Prepare HSM attribute if it's required */
1574 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1575 const int buflen = sizeof(struct hsm_attrs);
1577 lu_buf_alloc(fst_hsm_buf, buflen);
1578 lu_buf_alloc(snd_hsm_buf, buflen);
1579 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1580 GOTO(stop, rc = -ENOMEM);
1582 /* Read HSM attribute */
1583 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1587 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1591 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1592 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1597 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1598 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1604 /* prepare transaction */
1605 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1610 if (fst_buf->lb_buf != NULL)
1611 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1612 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1615 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1620 rc = mdd_trans_start(env, mdd, handle);
1624 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1625 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1629 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1631 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1634 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1635 mdd_obj_dev_name(fst_o),
1636 PFID(mdo2fid(fst_o)), rc, rc2);
1641 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1645 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1648 if (fst_buf->lb_buf != NULL)
1649 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1650 LU_XATTR_REPLACE, handle);
1652 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1658 /* failure on second file, but first was done, so we have
1659 * to roll back first. */
1660 if (fst_buf->lb_buf != NULL) {
1661 mdd_set_lmm_oi(fst_lmm, saved_oi);
1662 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1663 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1664 LU_XATTR_REPLACE, handle);
1666 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1672 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1677 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1682 CERROR("%s: unable to roll back layout swap. FIDs: "
1683 DFID" and "DFID "error: %d/%d, steps: %d\n",
1684 mdd_obj_dev_name(fst_o),
1685 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1687 /* a solution to avoid journal commit is to panic,
1688 * but it has strong consequences so we use LBUG to
1689 * allow sysdamin to choose to panic or not
1696 /* Issue one changelog record per file */
1697 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1701 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1707 rc = mdd_trans_stop(env, mdd, rc, handle);
1709 mdd_write_unlock(env, snd_o);
1710 mdd_write_unlock(env, fst_o);
1712 lu_buf_free(fst_buf);
1713 lu_buf_free(snd_buf);
1714 lu_buf_free(fst_hsm_buf);
1715 lu_buf_free(snd_hsm_buf);
1719 static int mdd_declare_layout_change(const struct lu_env *env,
1720 struct mdd_device *mdd,
1721 struct mdd_object *obj,
1722 struct layout_intent *layout,
1723 const struct lu_buf *buf,
1724 struct thandle *handle)
1728 rc = mdo_declare_layout_change(env, obj, layout, buf, handle);
1732 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1735 /* For PFL, this is used to instantiate necessary component objects. */
1736 int mdd_layout_change(const struct lu_env *env, struct md_object *obj,
1737 struct layout_intent *layout, const struct lu_buf *buf)
1739 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1740 struct mdd_device *mdd = mdo2mdd(obj);
1741 struct thandle *handle;
1745 handle = mdd_trans_create(env, mdd);
1747 RETURN(PTR_ERR(handle));
1749 rc = mdd_declare_layout_change(env, mdd, mdd_obj, layout, buf, handle);
1751 * It's possible that another layout write intent has already
1752 * instantiated our objects, so a -EALREADY returned, and we need to
1756 GOTO(stop, rc = (rc == -EALREADY) ? 0 : rc);
1758 rc = mdd_trans_start(env, mdd, handle);
1762 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1763 rc = mdo_layout_change(env, mdd_obj, layout, buf, handle);
1764 mdd_write_unlock(env, mdd_obj);
1768 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, handle);
1770 RETURN(mdd_trans_stop(env, mdd, rc, handle));
1773 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
1774 struct mdd_object *child, const struct lu_attr *attr,
1775 const struct md_op_spec *spec,
1776 struct dt_allocation_hint *hint)
1778 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
1779 struct dt_object *nc = mdd_object_child(child);
1781 memset(hint, 0, sizeof(*hint));
1783 /* For striped directory, give striping EA to lod_ah_init, which will
1784 * decide the stripe_offset and stripe count by it. */
1785 if (S_ISDIR(attr->la_mode) &&
1786 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
1787 hint->dah_eadata = spec->u.sp_ea.eadata;
1788 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
1790 hint->dah_eadata = NULL;
1791 hint->dah_eadata_len = 0;
1794 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
1795 hint->dah_eadata, hint->dah_eadata_len);
1796 /* @hint will be initialized by underlying device. */
1797 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
1801 * do NOT or the MAY_*'s, you'll get the weakest
1803 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
1807 /* Sadly, NFSD reopens a file repeatedly during operation, so the
1808 * "acc_mode = 0" allowance for newly-created files isn't honoured.
1809 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
1810 * owner can write to a file even if it is marked readonly to hide
1811 * its brokenness. (bug 5781) */
1812 if (flags & MDS_OPEN_OWNEROVERRIDE) {
1813 struct lu_ucred *uc = lu_ucred_check(env);
1815 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
1819 if (flags & FMODE_READ)
1821 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
1823 if (flags & MDS_FMODE_EXEC)
1828 static int mdd_open_sanity_check(const struct lu_env *env,
1829 struct mdd_object *obj,
1830 const struct lu_attr *attr, int flag)
1836 if (mdd_is_dead_obj(obj))
1839 if (S_ISLNK(attr->la_mode))
1842 mode = accmode(env, attr, flag);
1844 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
1847 if (!(flag & MDS_OPEN_CREATED)) {
1848 rc = mdd_permission_internal(env, obj, attr, mode);
1853 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
1854 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
1855 flag &= ~MDS_OPEN_TRUNC;
1857 /* For writing append-only file must open it with append mode. */
1858 if (attr->la_flags & LUSTRE_APPEND_FL) {
1859 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
1861 if (flag & MDS_OPEN_TRUNC)
1868 static int mdd_open(const struct lu_env *env, struct md_object *obj,
1871 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1872 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1875 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1877 rc = mdd_la_get(env, mdd_obj, attr);
1881 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
1885 mdd_obj->mod_count++;
1888 mdd_write_unlock(env, mdd_obj);
1892 static int mdd_declare_close(const struct lu_env *env,
1893 struct mdd_object *obj,
1895 struct thandle *handle)
1899 rc = orph_declare_index_delete(env, obj, handle);
1903 return mdo_declare_destroy(env, obj, handle);
1907 * No permission check is needed.
1909 static int mdd_close(const struct lu_env *env, struct md_object *obj,
1910 struct md_attr *ma, int mode)
1912 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1913 struct mdd_device *mdd = mdo2mdd(obj);
1914 struct thandle *handle = NULL;
1917 bool blocked = false;
1920 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
1921 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1922 mdd_obj->mod_count--;
1923 mdd_write_unlock(env, mdd_obj);
1925 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
1926 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
1927 "list\n", PFID(mdd_object_fid(mdd_obj)));
1931 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
1932 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
1933 /* check without any lock */
1934 is_orphan = mdd_obj->mod_count == 1 &&
1935 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
1939 /* mdd_trans_create() maybe failed because of barrier_entry(),
1940 * under such case, the orphan MDT-object will be left in the
1941 * orphan list, and when the MDT remount next time, the unused
1942 * orphans will be destroyed automatically.
1944 * One exception: the former mdd_finish_unlink may failed to
1945 * add the orphan MDT-object to the orphan list, then if the
1946 * mdd_trans_create() failed because of barrier_entry(), the
1947 * MDT-object will become real orphan that is neither in the
1948 * namespace nor in the orphan list. Such bad case should be
1949 * very rare and will be handled by e2fsck/lfsck. */
1950 handle = mdd_trans_create(env, mdo2mdd(obj));
1951 if (IS_ERR(handle)) {
1952 rc = PTR_ERR(handle);
1953 if (rc != -EINPROGRESS)
1961 rc = mdd_declare_close(env, mdd_obj, ma, handle);
1965 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1969 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
1975 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1976 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
1978 CERROR("Failed to get lu_attr of "DFID": %d\n",
1979 PFID(mdd_object_fid(mdd_obj)), rc);
1983 /* check again with lock */
1984 is_orphan = (mdd_obj->mod_count == 1) &&
1985 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
1986 ma->ma_attr.la_nlink == 0);
1988 if (is_orphan && !handle && !blocked) {
1989 mdd_write_unlock(env, mdd_obj);
1993 mdd_obj->mod_count--; /*release open count */
1995 if (!is_orphan || blocked)
1999 /* NB: Object maybe not in orphan list originally, it is rare case for
2000 * mdd_finish_unlink() failure, in that case, the object doesn't have
2001 * ORPHAN_OBJ flag */
2002 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2003 /* remove link to object from orphan index */
2004 LASSERT(handle != NULL);
2005 rc = __mdd_orphan_del(env, mdd_obj, handle);
2007 CERROR("%s: unable to delete "DFID" from orphan list: "
2008 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2009 PFID(mdd_object_fid(mdd_obj)), rc);
2010 /* If object was not deleted from orphan list, do not
2011 * destroy OSS objects, which will be done when next
2016 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2017 "list, OSS objects to be destroyed.\n",
2018 PFID(mdd_object_fid(mdd_obj)));
2021 rc = mdo_destroy(env, mdd_obj, handle);
2024 CERROR("%s: unable to delete "DFID" from orphan list: "
2025 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2026 PFID(mdd_object_fid(mdd_obj)), rc);
2031 mdd_write_unlock(env, mdd_obj);
2033 if (!rc && !blocked &&
2034 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2035 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2036 if (handle == NULL) {
2037 handle = mdd_trans_create(env, mdo2mdd(obj));
2039 GOTO(stop, rc = PTR_ERR(handle));
2041 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2046 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2051 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2056 if (handle != NULL && !IS_ERR(handle))
2057 rc = mdd_trans_stop(env, mdd, rc, handle);
2063 * Permission check is done when open,
2064 * no need check again.
2066 static int mdd_readpage_sanity_check(const struct lu_env *env,
2067 struct mdd_object *obj)
2069 struct dt_object *next = mdd_object_child(obj);
2073 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2081 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2082 size_t nob, const struct dt_it_ops *iops,
2083 struct dt_it *it, __u32 attr, void *arg)
2085 struct lu_dirpage *dp = &lp->lp_dir;
2089 struct lu_dirent *ent;
2090 struct lu_dirent *last = NULL;
2094 if (nob < sizeof(*dp))
2097 memset(area, 0, sizeof (*dp));
2098 area += sizeof (*dp);
2099 nob -= sizeof (*dp);
2106 len = iops->key_size(env, it);
2108 /* IAM iterator can return record with zero len. */
2112 hash = iops->store(env, it);
2113 if (unlikely(first)) {
2115 dp->ldp_hash_start = cpu_to_le64(hash);
2118 /* calculate max space required for lu_dirent */
2119 recsize = lu_dirent_calc_size(len, attr);
2121 if (nob >= recsize) {
2122 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2123 if (result == -ESTALE)
2128 /* osd might not able to pack all attributes,
2129 * so recheck rec length */
2130 recsize = le16_to_cpu(ent->lde_reclen);
2132 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2133 fid_le_to_cpu(&fid, &ent->lde_fid);
2134 if (fid_is_dot_lustre(&fid))
2138 result = (last != NULL) ? 0 :-EINVAL;
2142 ent = (void *)ent + recsize;
2146 result = iops->next(env, it);
2147 if (result == -ESTALE)
2149 } while (result == 0);
2152 dp->ldp_hash_end = cpu_to_le64(hash);
2154 if (last->lde_hash == dp->ldp_hash_end)
2155 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2156 last->lde_reclen = 0; /* end mark */
2159 /* end of directory */
2160 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2161 else if (result < 0)
2162 CWARN("build page failed: %d!\n", result);
2166 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2167 const struct lu_rdpg *rdpg)
2169 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2173 if (mdd_object_exists(mdd_obj) == 0) {
2174 CERROR("%s: object "DFID" not found: rc = -2\n",
2175 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2179 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2180 rc = mdd_readpage_sanity_check(env, mdd_obj);
2182 GOTO(out_unlock, rc);
2184 if (mdd_is_dead_obj(mdd_obj)) {
2186 struct lu_dirpage *dp;
2189 * According to POSIX, please do not return any entry to client:
2190 * even dot and dotdot should not be returned.
2192 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2193 PFID(mdd_object_fid(mdd_obj)));
2195 if (rdpg->rp_count <= 0)
2196 GOTO(out_unlock, rc = -EFAULT);
2197 LASSERT(rdpg->rp_pages != NULL);
2199 pg = rdpg->rp_pages[0];
2200 dp = (struct lu_dirpage *)kmap(pg);
2201 memset(dp, 0 , sizeof(struct lu_dirpage));
2202 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2203 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2204 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2206 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2209 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2210 mdd_dir_page_build, NULL);
2212 struct lu_dirpage *dp;
2214 dp = kmap(rdpg->rp_pages[0]);
2215 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2218 * No pages were processed, mark this for first page
2221 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2222 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2223 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2225 kunmap(rdpg->rp_pages[0]);
2228 GOTO(out_unlock, rc);
2230 mdd_read_unlock(env, mdd_obj);
2234 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2236 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2238 if (mdd_object_exists(mdd_obj) == 0) {
2241 CERROR("%s: object "DFID" not found: rc = %d\n",
2242 mdd_obj_dev_name(mdd_obj),
2243 PFID(mdd_object_fid(mdd_obj)), rc);
2246 return dt_object_sync(env, mdd_object_child(mdd_obj),
2250 static int mdd_object_lock(const struct lu_env *env,
2251 struct md_object *obj,
2252 struct lustre_handle *lh,
2253 struct ldlm_enqueue_info *einfo,
2254 union ldlm_policy_data *policy)
2256 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2257 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2261 static int mdd_object_unlock(const struct lu_env *env,
2262 struct md_object *obj,
2263 struct ldlm_enqueue_info *einfo,
2264 union ldlm_policy_data *policy)
2266 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2267 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2270 const struct md_object_operations mdd_obj_ops = {
2271 .moo_permission = mdd_permission,
2272 .moo_attr_get = mdd_attr_get,
2273 .moo_attr_set = mdd_attr_set,
2274 .moo_xattr_get = mdd_xattr_get,
2275 .moo_xattr_set = mdd_xattr_set,
2276 .moo_xattr_list = mdd_xattr_list,
2277 .moo_invalidate = mdd_invalidate,
2278 .moo_xattr_del = mdd_xattr_del,
2279 .moo_swap_layouts = mdd_swap_layouts,
2280 .moo_open = mdd_open,
2281 .moo_close = mdd_close,
2282 .moo_readpage = mdd_readpage,
2283 .moo_readlink = mdd_readlink,
2284 .moo_changelog = mdd_changelog,
2285 .moo_object_sync = mdd_object_sync,
2286 .moo_object_lock = mdd_object_lock,
2287 .moo_object_unlock = mdd_object_unlock,
2288 .moo_layout_change = mdd_layout_change,