4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <lustre_mds.h>
52 #include "mdd_internal.h"
54 static const struct lu_object_operations mdd_lu_obj_ops;
56 static int mdd_xattr_get(const struct lu_env *env,
57 struct md_object *obj, struct lu_buf *buf,
60 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
65 if (mdd_object_exists(obj) == 0)
68 rc = mdo_attr_get(env, obj, la);
69 if (unlikely(rc != 0)) {
71 obj->mod_flags |= DEAD_OBJ;
75 if (la->la_valid & LA_FLAGS &&
76 la->la_flags & LUSTRE_ORPHAN_FL)
77 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
82 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
84 struct mdd_thread_info *info;
86 lu_env_refill((struct lu_env *)env);
87 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
88 LASSERT(info != NULL);
92 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
96 buf = &mdd_env_info(env)->mti_buf[0];
102 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
103 const void *area, ssize_t len)
107 buf = &mdd_env_info(env)->mti_buf[0];
108 buf->lb_buf = (void *)area;
113 struct lu_object *mdd_object_alloc(const struct lu_env *env,
114 const struct lu_object_header *hdr,
117 struct mdd_object *mdd_obj;
119 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
120 if (mdd_obj != NULL) {
123 o = mdd2lu_obj(mdd_obj);
124 lu_object_init(o, NULL, d);
125 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
126 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
127 mdd_obj->mod_count = 0;
128 o->lo_ops = &mdd_lu_obj_ops;
135 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
136 const struct lu_object_conf *unused)
138 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
139 struct mdd_object *mdd_obj = lu2mdd_obj(o);
140 struct lu_object *below;
141 struct lu_device *under;
144 mdd_obj->mod_cltime = ktime_set(0, 0);
145 under = &d->mdd_child->dd_lu_dev;
146 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
148 RETURN(PTR_ERR(below));
150 lu_object_add(o, below);
155 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
159 if (lu_object_exists(o)) {
160 struct mdd_object *mdd_obj = lu2mdd_obj(o);
161 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
163 rc = mdd_la_get(env, mdd_obj, attr);
169 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
171 struct mdd_object *mdd = lu2mdd_obj(o);
174 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
177 static int mdd_object_print(const struct lu_env *env, void *cookie,
178 lu_printer_t p, const struct lu_object *o)
180 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
182 return (*p)(env, cookie,
183 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
184 mdd, mdd->mod_count, mdd->mod_valid,
185 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
188 static const struct lu_object_operations mdd_lu_obj_ops = {
189 .loo_object_init = mdd_object_init,
190 .loo_object_start = mdd_object_start,
191 .loo_object_free = mdd_object_free,
192 .loo_object_print = mdd_object_print,
195 struct mdd_object *mdd_object_find(const struct lu_env *env,
196 struct mdd_device *d,
197 const struct lu_fid *f)
199 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
203 * No permission check is needed.
205 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
208 struct mdd_object *mdd_obj = md2mdd_obj(obj);
213 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
214 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
215 ma->ma_attr.la_nlink = 0;
221 * No permission check is needed.
223 static int mdd_xattr_get(const struct lu_env *env,
224 struct md_object *obj, struct lu_buf *buf,
227 struct mdd_object *mdd_obj = md2mdd_obj(obj);
232 if (mdd_object_exists(mdd_obj) == 0) {
233 CERROR("%s: object "DFID" not found: rc = -2\n",
234 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
238 /* If the object has been destroyed, then do not get LMVEA, because
239 * it needs to load stripes from the iteration of the master object,
240 * and it will cause problem if master object has been destroyed, see
242 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
243 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
244 strcmp(name, XATTR_NAME_LMV) == 0))
247 /* If the object has been delete from the namespace, then
248 * get linkEA should return -ENOENT as well */
249 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
250 strcmp(name, XATTR_NAME_LINK) == 0))
253 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
254 rc = mdo_xattr_get(env, mdd_obj, buf, name);
255 mdd_read_unlock(env, mdd_obj);
261 * Permission check is done when open,
262 * no need check again.
264 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
267 struct mdd_object *mdd_obj = md2mdd_obj(obj);
268 struct dt_object *next;
273 if (mdd_object_exists(mdd_obj) == 0) {
274 CERROR("%s: object "DFID" not found: rc = -2\n",
275 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
279 next = mdd_object_child(mdd_obj);
280 LASSERT(next != NULL);
281 LASSERT(next->do_body_ops != NULL);
282 LASSERT(next->do_body_ops->dbo_read != NULL);
283 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
284 rc = dt_read(env, next, buf, &pos);
285 mdd_read_unlock(env, mdd_obj);
290 * No permission check is needed.
292 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
295 struct mdd_object *mdd_obj = md2mdd_obj(obj);
300 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
301 rc = mdo_xattr_list(env, mdd_obj, buf);
302 mdd_read_unlock(env, mdd_obj);
308 * Filter out XATTR_NAME_LINK if this is an orphan object. See
311 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
312 char *end = (char *)buf->lb_buf + rc;
313 char *p = buf->lb_buf;
316 char *next = p + strlen(p) + 1;
318 if (strcmp(p, XATTR_NAME_LINK) == 0) {
320 memmove(p, next, end - next);
322 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
323 " of orphan "DFID"\n",
324 PFID(mdd_object_fid(mdd_obj)));
335 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
337 return mdo_invalidate(env, md2mdd_obj(obj));
340 int mdd_declare_create_object_internal(const struct lu_env *env,
341 struct mdd_object *p,
342 struct mdd_object *c,
343 struct lu_attr *attr,
344 struct thandle *handle,
345 const struct md_op_spec *spec,
346 struct dt_allocation_hint *hint)
348 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
349 const struct dt_index_features *feat = spec->sp_feat;
353 if (feat != &dt_directory_features && feat != NULL) {
354 dof->dof_type = DFT_INDEX;
355 dof->u.dof_idx.di_feat = feat;
357 dof->dof_type = dt_mode_to_dft(attr->la_mode);
358 if (dof->dof_type == DFT_REGULAR) {
359 dof->u.dof_reg.striped =
360 md_should_create(spec->sp_cr_flags);
361 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
362 dof->u.dof_reg.striped = 0;
363 /* is this replay? */
365 dof->u.dof_reg.striped = 0;
369 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
374 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
375 struct mdd_object *c, struct lu_attr *attr,
376 struct thandle *handle,
377 const struct md_op_spec *spec,
378 struct dt_allocation_hint *hint)
380 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
384 LASSERT(!mdd_object_exists(c));
386 rc = mdo_create_object(env, c, attr, hint, dof, handle);
391 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
392 const struct lu_attr *attr, struct thandle *handle,
398 rc = mdo_attr_set(env, obj, attr, handle);
399 #ifdef CONFIG_FS_POSIX_ACL
400 if (!rc && (attr->la_valid & LA_MODE) && needacl)
401 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
406 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
407 const struct lu_attr *oattr, struct lu_attr *attr,
408 struct thandle *handle)
413 LASSERT(attr->la_valid & LA_CTIME);
414 LASSERT(oattr != NULL);
416 /* Make sure the ctime is increased only, however, it's not strictly
417 * reliable at here because there is not guarantee to hold lock on
418 * object, so we just bypass some unnecessary cmtime setting first
419 * and OSD has to check it again. */
420 if (attr->la_ctime < oattr->la_ctime)
421 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
422 else if (attr->la_valid == LA_CTIME &&
423 attr->la_ctime == oattr->la_ctime)
424 attr->la_valid &= ~LA_CTIME;
426 if (attr->la_valid != 0)
427 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
432 * This gives the same functionality as the code between
433 * sys_chmod and inode_setattr
434 * chown_common and inode_setattr
435 * utimes and inode_setattr
436 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
438 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
439 const struct lu_attr *oattr, struct lu_attr *la,
440 const unsigned long flags)
449 /* Do not permit change file type */
450 if (la->la_valid & LA_TYPE)
453 /* They should not be processed by setattr */
454 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
457 LASSERT(oattr != NULL);
459 uc = lu_ucred_check(env);
463 if (la->la_valid == LA_CTIME) {
464 if (!(flags & MDS_PERM_BYPASS))
465 /* This is only for set ctime when rename's source is
467 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
469 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
470 la->la_valid &= ~LA_CTIME;
474 if (la->la_valid == LA_ATIME) {
475 /* This is an atime-only attribute update for close RPCs. */
476 if (la->la_atime < (oattr->la_atime +
477 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
478 la->la_valid &= ~LA_ATIME;
482 /* Check if flags change. */
483 if (la->la_valid & LA_FLAGS) {
484 unsigned int oldflags = oattr->la_flags &
485 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
486 unsigned int newflags = la->la_flags &
487 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
489 if ((uc->uc_fsuid != oattr->la_uid) &&
490 !md_capable(uc, CFS_CAP_FOWNER))
493 /* The IMMUTABLE and APPEND_ONLY flags can
494 * only be changed by the relevant capability. */
495 if ((oldflags ^ newflags) &&
496 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
499 if (!S_ISDIR(oattr->la_mode))
500 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
503 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
504 (la->la_valid & ~LA_FLAGS) &&
505 !(flags & MDS_PERM_BYPASS))
508 /* Check for setting the obj time. */
509 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
510 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
511 if ((uc->uc_fsuid != oattr->la_uid) &&
512 !md_capable(uc, CFS_CAP_FOWNER)) {
513 rc = mdd_permission_internal(env, obj, oattr,
520 if (la->la_valid & LA_KILL_SUID) {
521 la->la_valid &= ~LA_KILL_SUID;
522 if ((oattr->la_mode & S_ISUID) &&
523 !(la->la_valid & LA_MODE)) {
524 la->la_mode = oattr->la_mode;
525 la->la_valid |= LA_MODE;
527 la->la_mode &= ~S_ISUID;
530 if (la->la_valid & LA_KILL_SGID) {
531 la->la_valid &= ~LA_KILL_SGID;
532 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
533 (S_ISGID | S_IXGRP)) &&
534 !(la->la_valid & LA_MODE)) {
535 la->la_mode = oattr->la_mode;
536 la->la_valid |= LA_MODE;
538 la->la_mode &= ~S_ISGID;
541 /* Make sure a caller can chmod. */
542 if (la->la_valid & LA_MODE) {
543 if (!(flags & MDS_PERM_BYPASS) &&
544 (uc->uc_fsuid != oattr->la_uid) &&
545 !md_capable(uc, CFS_CAP_FOWNER))
548 if (la->la_mode == (umode_t) -1)
549 la->la_mode = oattr->la_mode;
551 la->la_mode = (la->la_mode & S_IALLUGO) |
552 (oattr->la_mode & ~S_IALLUGO);
554 /* Also check the setgid bit! */
555 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
556 la->la_gid : oattr->la_gid) &&
557 !md_capable(uc, CFS_CAP_FSETID))
558 la->la_mode &= ~S_ISGID;
560 la->la_mode = oattr->la_mode;
563 /* Make sure a caller can chown. */
564 if (la->la_valid & LA_UID) {
565 if (la->la_uid == (uid_t) -1)
566 la->la_uid = oattr->la_uid;
567 if (((uc->uc_fsuid != oattr->la_uid) ||
568 (la->la_uid != oattr->la_uid)) &&
569 !md_capable(uc, CFS_CAP_CHOWN))
572 /* If the user or group of a non-directory has been
573 * changed by a non-root user, remove the setuid bit.
574 * 19981026 David C Niemi <niemi@tux.org>
576 * Changed this to apply to all users, including root,
577 * to avoid some races. This is the behavior we had in
578 * 2.0. The check for non-root was definitely wrong
579 * for 2.2 anyway, as it should have been using
580 * CAP_FSETID rather than fsuid -- 19990830 SD. */
581 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
582 !S_ISDIR(oattr->la_mode)) {
583 la->la_mode &= ~S_ISUID;
584 la->la_valid |= LA_MODE;
588 /* Make sure caller can chgrp. */
589 if (la->la_valid & LA_GID) {
590 if (la->la_gid == (gid_t) -1)
591 la->la_gid = oattr->la_gid;
592 if (((uc->uc_fsuid != oattr->la_uid) ||
593 ((la->la_gid != oattr->la_gid) &&
594 !lustre_in_group_p(uc, la->la_gid))) &&
595 !md_capable(uc, CFS_CAP_CHOWN))
598 /* Likewise, if the user or group of a non-directory
599 * has been changed by a non-root user, remove the
600 * setgid bit UNLESS there is no group execute bit
601 * (this would be a file marked for mandatory
602 * locking). 19981026 David C Niemi <niemi@tux.org>
604 * Removed the fsuid check (see the comment above) --
606 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
607 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
608 la->la_mode &= ~S_ISGID;
609 la->la_valid |= LA_MODE;
613 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
614 if (!((flags & MDS_OWNEROVERRIDE) &&
615 (uc->uc_fsuid == oattr->la_uid)) &&
616 !(flags & MDS_PERM_BYPASS)) {
617 rc = mdd_permission_internal(env, obj, oattr,
624 if (la->la_valid & LA_CTIME) {
625 /* The pure setattr, it has the priority over what is
626 * already set, do not drop it if ctime is equal. */
627 if (la->la_ctime < oattr->la_ctime)
628 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
634 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
635 struct mdd_device *mdd,
636 enum changelog_rec_type type, int flags,
637 const struct lu_fid *fid,
638 struct thandle *handle)
640 const struct lu_ucred *uc = lu_ucred(env);
641 struct llog_changelog_rec *rec;
646 flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
647 if (uc != NULL && uc->uc_jobid[0] != '\0')
650 reclen = llog_data_len(changelog_rec_offset(flags & CLF_SUPPORTED));
651 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
652 if (buf->lb_buf == NULL)
656 rec->cr.cr_flags = flags;
657 rec->cr.cr_type = (__u32)type;
658 rec->cr.cr_tfid = *fid;
659 rec->cr.cr_namelen = 0;
661 if (flags & CLF_JOBID)
662 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
664 rc = mdd_changelog_store(env, mdd, rec, handle);
670 /** Store a data change changelog record
671 * If this fails, we must fail the whole transaction; we don't
672 * want the change to commit without the log entry.
673 * \param mdd_obj - mdd_object of change
674 * \param handle - transaction handle
676 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
677 enum changelog_rec_type type, int flags,
678 struct mdd_object *mdd_obj, struct thandle *handle)
682 LASSERT(mdd_obj != NULL);
683 LASSERT(handle != NULL);
686 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
688 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
691 if (mdd_is_volatile_obj(mdd_obj))
694 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
695 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
696 /* Don't need multiple updates in this log */
697 /* Don't check under lock - no big deal if we get an extra
702 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
703 mdo2fid(mdd_obj), handle);
705 mdd_obj->mod_cltime = ktime_get();
710 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
711 int flags, struct md_device *m, const struct lu_fid *fid)
713 struct thandle *handle;
714 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
719 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
721 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
724 LASSERT(fid != NULL);
726 handle = mdd_trans_create(env, mdd);
728 RETURN(PTR_ERR(handle));
730 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
734 rc = mdd_trans_start(env, mdd, handle);
738 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
742 rc = mdd_trans_stop(env, mdd, rc, handle);
748 * Save LMA extended attributes with data from \a ma.
750 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
751 * not, LMA EA will be first read from disk, modified and write back.
754 /* Precedence for choosing record type when multiple
755 * attributes change: setattr > mtime > ctime > atime
756 * (ctime changes when mtime does, plus chmod/chown.
757 * atime and ctime are independent.) */
758 static int mdd_attr_set_changelog(const struct lu_env *env,
759 struct md_object *obj, struct thandle *handle,
762 struct mdd_device *mdd = mdo2mdd(obj);
765 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
766 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
767 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
768 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
769 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
770 bits = bits & mdd->mdd_cl.mc_mask;
771 /* This is an implementation limit rather than a protocol limit */
772 CLASSERT(CL_LAST <= sizeof(int) * 8);
776 /* The record type is the lowest non-masked set bit */
779 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
780 return mdd_changelog_data_store(env, mdd, type, (int)valid,
781 md2mdd_obj(obj), handle);
784 static int mdd_declare_attr_set(const struct lu_env *env,
785 struct mdd_device *mdd,
786 struct mdd_object *obj,
787 const struct lu_attr *attr,
788 struct thandle *handle)
792 rc = mdo_declare_attr_set(env, obj, attr, handle);
796 #ifdef CONFIG_FS_POSIX_ACL
797 if (attr->la_valid & LA_MODE) {
798 mdd_read_lock(env, obj, MOR_TGT_CHILD);
799 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
800 XATTR_NAME_ACL_ACCESS);
801 mdd_read_unlock(env, obj);
802 if (rc == -EOPNOTSUPP || rc == -ENODATA)
808 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
809 rc = mdo_declare_xattr_set(env, obj, buf,
810 XATTR_NAME_ACL_ACCESS, 0,
818 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
825 * permission changes may require sync operation, to mitigate performance
826 * impact, only do this for dir and when permission is reduced.
828 * For regular files, version is updated with permission change (see VBR), async
829 * permission won't cause any issue, while missing permission change on
830 * directory may affect accessibility of other objects after recovery.
832 static inline bool permission_needs_sync(const struct lu_attr *old,
833 const struct lu_attr *new)
835 if (!S_ISDIR(old->la_mode))
838 if (new->la_valid & (LA_UID | LA_GID))
841 if (new->la_valid & LA_MODE &&
842 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
845 if ((new->la_valid & LA_MODE) &&
846 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
847 (old->la_mode & S_IRWXUGO))
853 /* set attr and LOV EA at once, return updated attr */
854 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
855 const struct md_attr *ma)
857 struct mdd_object *mdd_obj = md2mdd_obj(obj);
858 struct mdd_device *mdd = mdo2mdd(obj);
859 struct thandle *handle;
860 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
861 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
862 const struct lu_attr *la = &ma->ma_attr;
866 /* we do not use ->attr_set() for LOV/HSM EA any more */
867 LASSERT((ma->ma_valid & MA_LOV) == 0);
868 LASSERT((ma->ma_valid & MA_HSM) == 0);
870 rc = mdd_la_get(env, mdd_obj, attr);
874 *la_copy = ma->ma_attr;
875 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
879 /* no need to setattr anymore */
880 if (la_copy->la_valid == 0) {
881 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
882 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
883 PFID(mdo2fid(mdd_obj)), la->la_valid);
888 handle = mdd_trans_create(env, mdd);
890 RETURN(PTR_ERR(handle));
892 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
896 rc = mdd_trans_start(env, mdd, handle);
900 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
903 if (la->la_valid & (LA_MTIME | LA_CTIME))
904 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
905 la->la_mtime, la->la_ctime);
907 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
908 if (la_copy->la_valid & LA_FLAGS)
909 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
910 else if (la_copy->la_valid) /* setattr */
911 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
912 mdd_write_unlock(env, mdd_obj);
915 rc = mdd_attr_set_changelog(env, obj, handle,
921 rc = mdd_trans_stop(env, mdd, rc, handle);
926 static int mdd_xattr_sanity_check(const struct lu_env *env,
927 struct mdd_object *obj,
928 const struct lu_attr *attr,
931 struct lu_ucred *uc = lu_ucred_assert(env);
934 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
937 if (strncmp(XATTR_USER_PREFIX, name,
938 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
939 /* For sticky directories, only the owner and privileged user
940 * can write attributes. */
941 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
942 (uc->uc_fsuid != attr->la_uid) &&
943 !md_capable(uc, CFS_CAP_FOWNER))
946 if ((uc->uc_fsuid != attr->la_uid) &&
947 !md_capable(uc, CFS_CAP_FOWNER))
955 * Check if a string begins with a given prefix.
957 * \param str String to check
958 * \param prefix Substring to check at the beginning of \a str
959 * \return true/false whether the condition is verified.
961 static inline bool has_prefix(const char *str, const char *prefix)
963 return strncmp(prefix, str, strlen(prefix)) == 0;
967 * Indicate the kind of changelog to store (if any) for a xattr set/del.
969 * \param[in] xattr_name Full extended attribute name.
971 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
973 static enum changelog_rec_type
974 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
975 const char *xattr_name)
977 /* Layout changes systematically recorded */
978 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
979 strncmp(XATTR_LUSTRE_LOV, xattr_name,
980 strlen(XATTR_LUSTRE_LOV)) == 0)
983 /* HSM information changes systematically recorded */
984 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
987 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
988 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
989 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
995 static int mdd_declare_xattr_set(const struct lu_env *env,
996 struct mdd_device *mdd,
997 struct mdd_object *obj,
998 const struct lu_buf *buf,
1000 int fl, struct thandle *handle)
1004 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1008 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1009 return 0; /* no changelog to store */
1011 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1015 * Compare current and future data of HSM EA and add a changelog if needed.
1017 * Caller should have write-locked \param obj.
1019 * \param buf - Future HSM EA content.
1020 * \retval 0 if no changelog is needed or changelog was added properly.
1021 * \retval -ve errno if there was a problem
1023 static int mdd_hsm_update_locked(const struct lu_env *env,
1024 struct md_object *obj,
1025 const struct lu_buf *buf,
1026 struct thandle *handle, int *cl_flags)
1028 struct mdd_thread_info *info = mdd_env_info(env);
1029 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1030 struct lu_buf *current_buf;
1031 struct md_hsm *current_mh;
1032 struct md_hsm *new_mh;
1036 OBD_ALLOC_PTR(current_mh);
1037 if (current_mh == NULL)
1040 /* Read HSM attrs from disk */
1041 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1042 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1043 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1044 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1045 if (rc < 0 && rc != -ENODATA)
1047 else if (rc == -ENODATA)
1048 current_mh->mh_flags = 0;
1050 /* Map future HSM xattr */
1051 OBD_ALLOC_PTR(new_mh);
1053 GOTO(free, rc = -ENOMEM);
1054 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1058 /* Flags differ, set flags for the changelog that will be added */
1059 if (current_mh->mh_flags != new_mh->mh_flags) {
1060 hsm_set_cl_event(cl_flags, HE_STATE);
1061 if (new_mh->mh_flags & HS_DIRTY)
1062 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1065 OBD_FREE_PTR(new_mh);
1068 OBD_FREE_PTR(current_mh);
1072 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1076 * The caller should guarantee to update the object ctime
1077 * after xattr_set if needed.
1079 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1080 const struct lu_buf *buf, const char *name,
1083 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1084 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1085 struct mdd_device *mdd = mdo2mdd(obj);
1086 struct thandle *handle;
1087 enum changelog_rec_type cl_type;
1092 rc = mdd_la_get(env, mdd_obj, attr);
1096 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1100 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1101 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1102 struct posix_acl *acl;
1104 /* user may set empty ACL, which should be treated as removing
1106 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1109 RETURN(PTR_ERR(acl));
1111 rc = mdd_xattr_del(env, obj, name);
1114 posix_acl_release(acl);
1117 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1118 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1122 handle = mdd_trans_create(env, mdd);
1124 RETURN(PTR_ERR(handle));
1126 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1130 rc = mdd_trans_start(env, mdd, handle);
1134 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1136 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1137 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1139 mdd_write_unlock(env, mdd_obj);
1144 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1145 mdd_write_unlock(env, mdd_obj);
1149 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1153 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1158 return mdd_trans_stop(env, mdd, rc, handle);
1161 static int mdd_declare_xattr_del(const struct lu_env *env,
1162 struct mdd_device *mdd,
1163 struct mdd_object *obj,
1165 struct thandle *handle)
1169 rc = mdo_declare_xattr_del(env, obj, name, handle);
1173 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1174 return 0; /* no changelog to store */
1176 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1180 * The caller should guarantee to update the object ctime
1181 * after xattr_set if needed.
1183 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1186 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1187 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1188 struct mdd_device *mdd = mdo2mdd(obj);
1189 struct thandle *handle;
1193 rc = mdd_la_get(env, mdd_obj, attr);
1197 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1201 handle = mdd_trans_create(env, mdd);
1203 RETURN(PTR_ERR(handle));
1205 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1209 rc = mdd_trans_start(env, mdd, handle);
1213 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1214 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1215 mdd_write_unlock(env, mdd_obj);
1219 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1222 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1226 return mdd_trans_stop(env, mdd, rc, handle);
1230 * read lov EA of an object
1231 * return the lov EA in an allocated lu_buf
1233 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1234 struct lu_buf *lmm_buf)
1236 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1241 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1243 if (rc == -ERANGE) {
1244 /* mti_big_buf is allocated but is too small
1245 * we need to increase it */
1246 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1248 if (buf->lb_buf == NULL)
1249 GOTO(out, rc = -ENOMEM);
1260 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1261 /* mti_big_buf was not allocated, so we have to
1262 * allocate it based on the ea size */
1263 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1265 if (buf->lb_buf == NULL)
1266 GOTO(out, rc = -ENOMEM);
1270 lu_buf_alloc(lmm_buf, bufsize);
1271 if (lmm_buf->lb_buf == NULL)
1272 GOTO(out, rc = -ENOMEM);
1274 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1280 lu_buf_free(lmm_buf);
1284 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1285 struct mdd_object *o, struct lu_buf *buf,
1286 struct thandle *handle)
1288 struct hsm_attrs *attrs;
1294 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1299 attrs = buf->lb_buf;
1300 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1301 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1304 /* Add a changelog record for release. */
1305 hsm_set_cl_event(&flags, HE_RELEASE);
1306 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1312 * check if layout swapping between 2 objects is allowed
1314 * - only normal FIDs or non-system IGIFs
1315 * - same type of objects
1316 * - same owner/group (so quotas are still valid) unless this is from HSM
1319 static int mdd_layout_swap_allowed(const struct lu_env *env,
1320 struct mdd_object *o1,
1321 const struct lu_attr *attr1,
1322 struct mdd_object *o2,
1323 const struct lu_attr *attr2,
1326 const struct lu_fid *fid1, *fid2;
1332 if (!fid_is_norm(fid1) &&
1333 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1336 if (!fid_is_norm(fid2) &&
1337 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1340 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1341 if (S_ISDIR(mdd_object_type(o1)))
1343 if (S_ISREG(mdd_object_type(o1)))
1348 if (flags & SWAP_LAYOUTS_MDS_HSM)
1351 if ((attr1->la_uid != attr2->la_uid) ||
1352 (attr1->la_gid != attr2->la_gid))
1358 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1359 * look into the layout in MDD layer. */
1360 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1362 struct lov_comp_md_v1 *comp_v1;
1363 struct lov_mds_md *v1;
1367 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1368 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1369 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1375 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1376 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1379 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1380 off = le32_to_cpu(comp_v1->lcm_entries[i].
1382 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1387 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1388 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1399 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1401 return mdd_lmm_oi(lmm, oi, true);
1404 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1406 return mdd_lmm_oi(lmm, oi, false);
1409 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1411 struct lov_comp_md_v1 *comp_v1;
1413 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1414 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1416 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1418 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1419 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1420 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1421 __u16 tmp_gen = *gen;
1423 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1425 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1432 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1434 return mdd_lmm_gen(lmm, gen, true);
1437 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1439 return mdd_lmm_gen(lmm, gen, false);
1443 * swap layouts between 2 lustre objects
1445 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1446 struct md_object *obj2, __u64 flags)
1448 struct mdd_thread_info *info = mdd_env_info(env);
1449 struct mdd_object *fst_o = md2mdd_obj(obj1);
1450 struct mdd_object *snd_o = md2mdd_obj(obj2);
1451 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1452 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1453 struct mdd_device *mdd = mdo2mdd(obj1);
1454 struct lov_mds_md *fst_lmm, *snd_lmm;
1455 struct lu_buf *fst_buf = &info->mti_buf[0];
1456 struct lu_buf *snd_buf = &info->mti_buf[1];
1457 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1458 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1459 struct ost_id *saved_oi = NULL;
1460 struct thandle *handle;
1461 __u32 fst_gen, snd_gen, saved_gen;
1467 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1468 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1470 /* we have to sort the 2 obj, so locking will always
1471 * be in the same order, even in case of 2 concurrent swaps */
1472 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1473 if (rc == 0) /* same fid ? */
1479 rc = mdd_la_get(env, fst_o, fst_la);
1483 rc = mdd_la_get(env, snd_o, snd_la);
1487 /* check if layout swapping is allowed */
1488 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1492 handle = mdd_trans_create(env, mdd);
1494 RETURN(PTR_ERR(handle));
1496 /* objects are already sorted */
1497 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1498 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1500 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1501 if (rc < 0 && rc != -ENODATA)
1504 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1505 if (rc < 0 && rc != -ENODATA)
1508 /* swapping 2 non existant layouts is a success */
1509 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1512 /* to help inode migration between MDT, it is better to
1513 * start by the no layout file (if one), so we order the swap */
1514 if (snd_buf->lb_buf == NULL) {
1516 swap(fst_buf, snd_buf);
1519 fst_gen = snd_gen = 0;
1520 /* lmm and generation layout initialization */
1521 if (fst_buf->lb_buf != NULL) {
1522 fst_lmm = fst_buf->lb_buf;
1523 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1524 fst_fl = LU_XATTR_REPLACE;
1528 fst_fl = LU_XATTR_CREATE;
1531 snd_lmm = snd_buf->lb_buf;
1532 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1534 saved_gen = fst_gen;
1535 /* increase the generation layout numbers */
1540 * XXX The layout generation is used to generate component IDs for
1541 * the composite file, we have to do some special tweaks to make
1542 * sure the layout generation is always adequate for that job.
1545 /* Skip invalid generation number for composite layout */
1546 if ((snd_gen & LCME_ID_MASK) == 0)
1548 if ((fst_gen & LCME_ID_MASK) == 0)
1550 /* Make sure the generation is greater than all the component IDs */
1551 if (fst_gen < snd_gen)
1553 else if (fst_gen > snd_gen)
1556 /* set the file specific informations in lmm */
1557 if (fst_lmm != NULL) {
1558 saved_oi = &info->mti_oa.o_oi;
1559 mdd_get_lmm_oi(fst_lmm, saved_oi);
1560 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1561 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1562 mdd_set_lmm_oi(snd_lmm, saved_oi);
1564 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1565 cpu_to_le32(LOV_MAGIC_MAGIC))
1566 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEF);
1568 GOTO(stop, rc = -EPROTO);
1570 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1572 /* Prepare HSM attribute if it's required */
1573 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1574 const int buflen = sizeof(struct hsm_attrs);
1576 lu_buf_alloc(fst_hsm_buf, buflen);
1577 lu_buf_alloc(snd_hsm_buf, buflen);
1578 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1579 GOTO(stop, rc = -ENOMEM);
1581 /* Read HSM attribute */
1582 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1586 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1590 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1591 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1596 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1597 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1603 /* prepare transaction */
1604 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1609 if (fst_buf->lb_buf != NULL)
1610 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1611 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1614 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1619 rc = mdd_trans_start(env, mdd, handle);
1623 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1624 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1628 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1630 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1633 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1634 mdd_obj_dev_name(fst_o),
1635 PFID(mdo2fid(fst_o)), rc, rc2);
1640 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1644 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1647 if (fst_buf->lb_buf != NULL)
1648 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1649 LU_XATTR_REPLACE, handle);
1651 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1657 /* failure on second file, but first was done, so we have
1658 * to roll back first. */
1659 if (fst_buf->lb_buf != NULL) {
1660 mdd_set_lmm_oi(fst_lmm, saved_oi);
1661 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1662 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1663 LU_XATTR_REPLACE, handle);
1665 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1671 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1676 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1681 CERROR("%s: unable to roll back layout swap. FIDs: "
1682 DFID" and "DFID "error: %d/%d, steps: %d\n",
1683 mdd_obj_dev_name(fst_o),
1684 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1686 /* a solution to avoid journal commit is to panic,
1687 * but it has strong consequences so we use LBUG to
1688 * allow sysdamin to choose to panic or not
1695 /* Issue one changelog record per file */
1696 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1700 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1706 rc = mdd_trans_stop(env, mdd, rc, handle);
1708 mdd_write_unlock(env, snd_o);
1709 mdd_write_unlock(env, fst_o);
1711 lu_buf_free(fst_buf);
1712 lu_buf_free(snd_buf);
1713 lu_buf_free(fst_hsm_buf);
1714 lu_buf_free(snd_hsm_buf);
1718 static int mdd_declare_layout_change(const struct lu_env *env,
1719 struct mdd_device *mdd,
1720 struct mdd_object *obj,
1721 struct layout_intent *layout,
1722 const struct lu_buf *buf,
1723 struct thandle *handle)
1727 rc = mdo_declare_layout_change(env, obj, layout, buf, handle);
1731 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1734 /* For PFL, this is used to instantiate necessary component objects. */
1735 int mdd_layout_change(const struct lu_env *env, struct md_object *obj,
1736 struct layout_intent *layout, const struct lu_buf *buf)
1738 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1739 struct mdd_device *mdd = mdo2mdd(obj);
1740 struct thandle *handle;
1744 handle = mdd_trans_create(env, mdd);
1746 RETURN(PTR_ERR(handle));
1748 rc = mdd_declare_layout_change(env, mdd, mdd_obj, layout, buf, handle);
1750 * It's possible that another layout write intent has already
1751 * instantiated our objects, so a -EALREADY returned, and we need to
1755 GOTO(stop, rc = (rc == -EALREADY) ? 0 : rc);
1757 rc = mdd_trans_start(env, mdd, handle);
1761 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1762 rc = mdo_layout_change(env, mdd_obj, layout, buf, handle);
1763 mdd_write_unlock(env, mdd_obj);
1767 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, handle);
1769 RETURN(mdd_trans_stop(env, mdd, rc, handle));
1772 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
1773 struct mdd_object *child, const struct lu_attr *attr,
1774 const struct md_op_spec *spec,
1775 struct dt_allocation_hint *hint)
1777 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
1778 struct dt_object *nc = mdd_object_child(child);
1780 memset(hint, 0, sizeof(*hint));
1782 /* For striped directory, give striping EA to lod_ah_init, which will
1783 * decide the stripe_offset and stripe count by it. */
1784 if (S_ISDIR(attr->la_mode) &&
1785 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
1786 hint->dah_eadata = spec->u.sp_ea.eadata;
1787 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
1789 hint->dah_eadata = NULL;
1790 hint->dah_eadata_len = 0;
1793 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
1794 hint->dah_eadata, hint->dah_eadata_len);
1795 /* @hint will be initialized by underlying device. */
1796 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
1800 * do NOT or the MAY_*'s, you'll get the weakest
1802 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
1806 /* Sadly, NFSD reopens a file repeatedly during operation, so the
1807 * "acc_mode = 0" allowance for newly-created files isn't honoured.
1808 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
1809 * owner can write to a file even if it is marked readonly to hide
1810 * its brokenness. (bug 5781) */
1811 if (flags & MDS_OPEN_OWNEROVERRIDE) {
1812 struct lu_ucred *uc = lu_ucred_check(env);
1814 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
1818 if (flags & FMODE_READ)
1820 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
1822 if (flags & MDS_FMODE_EXEC)
1827 static int mdd_open_sanity_check(const struct lu_env *env,
1828 struct mdd_object *obj,
1829 const struct lu_attr *attr, int flag)
1835 if (mdd_is_dead_obj(obj))
1838 if (S_ISLNK(attr->la_mode))
1841 mode = accmode(env, attr, flag);
1843 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
1846 if (!(flag & MDS_OPEN_CREATED)) {
1847 rc = mdd_permission_internal(env, obj, attr, mode);
1852 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
1853 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
1854 flag &= ~MDS_OPEN_TRUNC;
1856 /* For writing append-only file must open it with append mode. */
1857 if (attr->la_flags & LUSTRE_APPEND_FL) {
1858 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
1860 if (flag & MDS_OPEN_TRUNC)
1867 static int mdd_open(const struct lu_env *env, struct md_object *obj,
1870 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1871 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1874 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1876 rc = mdd_la_get(env, mdd_obj, attr);
1880 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
1884 mdd_obj->mod_count++;
1887 mdd_write_unlock(env, mdd_obj);
1891 static int mdd_declare_close(const struct lu_env *env,
1892 struct mdd_object *obj,
1894 struct thandle *handle)
1898 rc = orph_declare_index_delete(env, obj, handle);
1902 return mdo_declare_destroy(env, obj, handle);
1906 * No permission check is needed.
1908 static int mdd_close(const struct lu_env *env, struct md_object *obj,
1909 struct md_attr *ma, int mode)
1911 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1912 struct mdd_device *mdd = mdo2mdd(obj);
1913 struct thandle *handle = NULL;
1916 bool blocked = false;
1919 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
1920 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1921 mdd_obj->mod_count--;
1922 mdd_write_unlock(env, mdd_obj);
1924 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
1925 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
1926 "list\n", PFID(mdd_object_fid(mdd_obj)));
1930 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
1931 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
1932 /* check without any lock */
1933 is_orphan = mdd_obj->mod_count == 1 &&
1934 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
1938 /* mdd_trans_create() maybe failed because of barrier_entry(),
1939 * under such case, the orphan MDT-object will be left in the
1940 * orphan list, and when the MDT remount next time, the unused
1941 * orphans will be destroyed automatically.
1943 * One exception: the former mdd_finish_unlink may failed to
1944 * add the orphan MDT-object to the orphan list, then if the
1945 * mdd_trans_create() failed because of barrier_entry(), the
1946 * MDT-object will become real orphan that is neither in the
1947 * namespace nor in the orphan list. Such bad case should be
1948 * very rare and will be handled by e2fsck/lfsck. */
1949 handle = mdd_trans_create(env, mdo2mdd(obj));
1950 if (IS_ERR(handle)) {
1951 rc = PTR_ERR(handle);
1952 if (rc != -EINPROGRESS)
1960 rc = mdd_declare_close(env, mdd_obj, ma, handle);
1964 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1968 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
1974 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1975 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
1977 CERROR("Failed to get lu_attr of "DFID": %d\n",
1978 PFID(mdd_object_fid(mdd_obj)), rc);
1982 /* check again with lock */
1983 is_orphan = (mdd_obj->mod_count == 1) &&
1984 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
1985 ma->ma_attr.la_nlink == 0);
1987 if (is_orphan && !handle && !blocked) {
1988 mdd_write_unlock(env, mdd_obj);
1992 mdd_obj->mod_count--; /*release open count */
1994 if (!is_orphan || blocked)
1998 /* NB: Object maybe not in orphan list originally, it is rare case for
1999 * mdd_finish_unlink() failure, in that case, the object doesn't have
2000 * ORPHAN_OBJ flag */
2001 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2002 /* remove link to object from orphan index */
2003 LASSERT(handle != NULL);
2004 rc = __mdd_orphan_del(env, mdd_obj, handle);
2006 CERROR("%s: unable to delete "DFID" from orphan list: "
2007 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2008 PFID(mdd_object_fid(mdd_obj)), rc);
2009 /* If object was not deleted from orphan list, do not
2010 * destroy OSS objects, which will be done when next
2015 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2016 "list, OSS objects to be destroyed.\n",
2017 PFID(mdd_object_fid(mdd_obj)));
2020 rc = mdo_destroy(env, mdd_obj, handle);
2023 CERROR("%s: unable to delete "DFID" from orphan list: "
2024 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2025 PFID(mdd_object_fid(mdd_obj)), rc);
2030 mdd_write_unlock(env, mdd_obj);
2032 if (!rc && !blocked &&
2033 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2034 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2035 if (handle == NULL) {
2036 handle = mdd_trans_create(env, mdo2mdd(obj));
2038 GOTO(stop, rc = PTR_ERR(handle));
2040 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2045 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2050 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2055 if (handle != NULL && !IS_ERR(handle))
2056 rc = mdd_trans_stop(env, mdd, rc, handle);
2062 * Permission check is done when open,
2063 * no need check again.
2065 static int mdd_readpage_sanity_check(const struct lu_env *env,
2066 struct mdd_object *obj)
2068 struct dt_object *next = mdd_object_child(obj);
2072 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2080 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2081 size_t nob, const struct dt_it_ops *iops,
2082 struct dt_it *it, __u32 attr, void *arg)
2084 struct lu_dirpage *dp = &lp->lp_dir;
2088 struct lu_dirent *ent;
2089 struct lu_dirent *last = NULL;
2093 if (nob < sizeof(*dp))
2096 memset(area, 0, sizeof (*dp));
2097 area += sizeof (*dp);
2098 nob -= sizeof (*dp);
2105 len = iops->key_size(env, it);
2107 /* IAM iterator can return record with zero len. */
2111 hash = iops->store(env, it);
2112 if (unlikely(first)) {
2114 dp->ldp_hash_start = cpu_to_le64(hash);
2117 /* calculate max space required for lu_dirent */
2118 recsize = lu_dirent_calc_size(len, attr);
2120 if (nob >= recsize) {
2121 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2122 if (result == -ESTALE)
2127 /* osd might not able to pack all attributes,
2128 * so recheck rec length */
2129 recsize = le16_to_cpu(ent->lde_reclen);
2131 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2132 fid_le_to_cpu(&fid, &ent->lde_fid);
2133 if (fid_is_dot_lustre(&fid))
2137 result = (last != NULL) ? 0 :-EINVAL;
2141 ent = (void *)ent + recsize;
2145 result = iops->next(env, it);
2146 if (result == -ESTALE)
2148 } while (result == 0);
2151 dp->ldp_hash_end = cpu_to_le64(hash);
2153 if (last->lde_hash == dp->ldp_hash_end)
2154 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2155 last->lde_reclen = 0; /* end mark */
2158 /* end of directory */
2159 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2160 else if (result < 0)
2161 CWARN("build page failed: %d!\n", result);
2165 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2166 const struct lu_rdpg *rdpg)
2168 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2172 if (mdd_object_exists(mdd_obj) == 0) {
2173 CERROR("%s: object "DFID" not found: rc = -2\n",
2174 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2178 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2179 rc = mdd_readpage_sanity_check(env, mdd_obj);
2181 GOTO(out_unlock, rc);
2183 if (mdd_is_dead_obj(mdd_obj)) {
2185 struct lu_dirpage *dp;
2188 * According to POSIX, please do not return any entry to client:
2189 * even dot and dotdot should not be returned.
2191 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2192 PFID(mdd_object_fid(mdd_obj)));
2194 if (rdpg->rp_count <= 0)
2195 GOTO(out_unlock, rc = -EFAULT);
2196 LASSERT(rdpg->rp_pages != NULL);
2198 pg = rdpg->rp_pages[0];
2199 dp = (struct lu_dirpage *)kmap(pg);
2200 memset(dp, 0 , sizeof(struct lu_dirpage));
2201 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2202 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2203 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2205 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2208 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2209 mdd_dir_page_build, NULL);
2211 struct lu_dirpage *dp;
2213 dp = kmap(rdpg->rp_pages[0]);
2214 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2217 * No pages were processed, mark this for first page
2220 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2221 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2222 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2224 kunmap(rdpg->rp_pages[0]);
2227 GOTO(out_unlock, rc);
2229 mdd_read_unlock(env, mdd_obj);
2233 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2235 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2237 if (mdd_object_exists(mdd_obj) == 0) {
2240 CERROR("%s: object "DFID" not found: rc = %d\n",
2241 mdd_obj_dev_name(mdd_obj),
2242 PFID(mdd_object_fid(mdd_obj)), rc);
2245 return dt_object_sync(env, mdd_object_child(mdd_obj),
2249 static int mdd_object_lock(const struct lu_env *env,
2250 struct md_object *obj,
2251 struct lustre_handle *lh,
2252 struct ldlm_enqueue_info *einfo,
2253 union ldlm_policy_data *policy)
2255 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2256 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2260 static int mdd_object_unlock(const struct lu_env *env,
2261 struct md_object *obj,
2262 struct ldlm_enqueue_info *einfo,
2263 union ldlm_policy_data *policy)
2265 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2266 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2269 const struct md_object_operations mdd_obj_ops = {
2270 .moo_permission = mdd_permission,
2271 .moo_attr_get = mdd_attr_get,
2272 .moo_attr_set = mdd_attr_set,
2273 .moo_xattr_get = mdd_xattr_get,
2274 .moo_xattr_set = mdd_xattr_set,
2275 .moo_xattr_list = mdd_xattr_list,
2276 .moo_invalidate = mdd_invalidate,
2277 .moo_xattr_del = mdd_xattr_del,
2278 .moo_swap_layouts = mdd_swap_layouts,
2279 .moo_open = mdd_open,
2280 .moo_close = mdd_close,
2281 .moo_readpage = mdd_readpage,
2282 .moo_readlink = mdd_readlink,
2283 .moo_changelog = mdd_changelog,
2284 .moo_object_sync = mdd_object_sync,
2285 .moo_object_lock = mdd_object_lock,
2286 .moo_object_unlock = mdd_object_unlock,
2287 .moo_layout_change = mdd_layout_change,