4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre/lustre_param.h>
50 #include <lustre_mds.h>
52 #include "mdd_internal.h"
54 static const struct lu_object_operations mdd_lu_obj_ops;
56 static int mdd_xattr_get(const struct lu_env *env,
57 struct md_object *obj, struct lu_buf *buf,
60 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
65 if (mdd_object_exists(obj) == 0)
68 rc = mdo_attr_get(env, obj, la);
69 if (unlikely(rc != 0)) {
71 obj->mod_flags |= DEAD_OBJ;
75 if (la->la_valid & LA_FLAGS &&
76 la->la_flags & LUSTRE_ORPHAN_FL)
77 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
82 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
84 struct mdd_thread_info *info;
86 lu_env_refill((struct lu_env *)env);
87 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
88 LASSERT(info != NULL);
92 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
96 buf = &mdd_env_info(env)->mti_buf[0];
102 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
103 const void *area, ssize_t len)
107 buf = &mdd_env_info(env)->mti_buf[0];
108 buf->lb_buf = (void *)area;
113 struct lu_object *mdd_object_alloc(const struct lu_env *env,
114 const struct lu_object_header *hdr,
117 struct mdd_object *mdd_obj;
119 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
120 if (mdd_obj != NULL) {
123 o = mdd2lu_obj(mdd_obj);
124 lu_object_init(o, NULL, d);
125 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
126 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
127 mdd_obj->mod_count = 0;
128 o->lo_ops = &mdd_lu_obj_ops;
135 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
136 const struct lu_object_conf *unused)
138 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
139 struct mdd_object *mdd_obj = lu2mdd_obj(o);
140 struct lu_object *below;
141 struct lu_device *under;
144 mdd_obj->mod_cltime = ktime_set(0, 0);
145 under = &d->mdd_child->dd_lu_dev;
146 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
148 RETURN(PTR_ERR(below));
150 lu_object_add(o, below);
155 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
159 if (lu_object_exists(o)) {
160 struct mdd_object *mdd_obj = lu2mdd_obj(o);
161 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
163 rc = mdd_la_get(env, mdd_obj, attr);
169 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
171 struct mdd_object *mdd = lu2mdd_obj(o);
174 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
177 static int mdd_object_print(const struct lu_env *env, void *cookie,
178 lu_printer_t p, const struct lu_object *o)
180 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
182 return (*p)(env, cookie,
183 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%lldns, flags=%lx)",
184 mdd, mdd->mod_count, mdd->mod_valid,
185 ktime_to_ns(mdd->mod_cltime), mdd->mod_flags);
188 static const struct lu_object_operations mdd_lu_obj_ops = {
189 .loo_object_init = mdd_object_init,
190 .loo_object_start = mdd_object_start,
191 .loo_object_free = mdd_object_free,
192 .loo_object_print = mdd_object_print,
195 struct mdd_object *mdd_object_find(const struct lu_env *env,
196 struct mdd_device *d,
197 const struct lu_fid *f)
199 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
203 * No permission check is needed.
205 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
208 struct mdd_object *mdd_obj = md2mdd_obj(obj);
213 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
214 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
215 ma->ma_attr.la_nlink = 0;
221 * No permission check is needed.
223 static int mdd_xattr_get(const struct lu_env *env,
224 struct md_object *obj, struct lu_buf *buf,
227 struct mdd_object *mdd_obj = md2mdd_obj(obj);
232 if (mdd_object_exists(mdd_obj) == 0) {
233 CERROR("%s: object "DFID" not found: rc = -2\n",
234 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
238 /* If the object has been destroyed, then do not get LMVEA, because
239 * it needs to load stripes from the iteration of the master object,
240 * and it will cause problem if master object has been destroyed, see
242 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
243 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
244 strcmp(name, XATTR_NAME_LMV) == 0))
247 /* If the object has been delete from the namespace, then
248 * get linkEA should return -ENOENT as well */
249 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
250 strcmp(name, XATTR_NAME_LINK) == 0))
253 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
254 rc = mdo_xattr_get(env, mdd_obj, buf, name);
255 mdd_read_unlock(env, mdd_obj);
261 * Permission check is done when open,
262 * no need check again.
264 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
267 struct mdd_object *mdd_obj = md2mdd_obj(obj);
268 struct dt_object *next;
273 if (mdd_object_exists(mdd_obj) == 0) {
274 CERROR("%s: object "DFID" not found: rc = -2\n",
275 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
279 next = mdd_object_child(mdd_obj);
280 LASSERT(next != NULL);
281 LASSERT(next->do_body_ops != NULL);
282 LASSERT(next->do_body_ops->dbo_read != NULL);
283 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
284 rc = dt_read(env, next, buf, &pos);
285 mdd_read_unlock(env, mdd_obj);
290 * No permission check is needed.
292 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
295 struct mdd_object *mdd_obj = md2mdd_obj(obj);
300 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
301 rc = mdo_xattr_list(env, mdd_obj, buf);
302 mdd_read_unlock(env, mdd_obj);
304 /* If the buffer is NULL then we are only here to get the
305 * length of the xattr name list. */
306 if (rc < 0 || buf->lb_buf == NULL)
310 * Filter out XATTR_NAME_LINK if this is an orphan object. See
313 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
314 char *end = (char *)buf->lb_buf + rc;
315 char *p = buf->lb_buf;
318 char *next = p + strlen(p) + 1;
320 if (strcmp(p, XATTR_NAME_LINK) == 0) {
322 memmove(p, next, end - next);
324 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
325 " of orphan "DFID"\n",
326 PFID(mdd_object_fid(mdd_obj)));
337 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
339 return mdo_invalidate(env, md2mdd_obj(obj));
342 int mdd_declare_create_object_internal(const struct lu_env *env,
343 struct mdd_object *p,
344 struct mdd_object *c,
345 struct lu_attr *attr,
346 struct thandle *handle,
347 const struct md_op_spec *spec,
348 struct dt_allocation_hint *hint)
350 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
351 const struct dt_index_features *feat = spec->sp_feat;
355 if (feat != &dt_directory_features && feat != NULL) {
356 dof->dof_type = DFT_INDEX;
357 dof->u.dof_idx.di_feat = feat;
359 dof->dof_type = dt_mode_to_dft(attr->la_mode);
360 if (dof->dof_type == DFT_REGULAR) {
361 dof->u.dof_reg.striped =
362 md_should_create(spec->sp_cr_flags);
363 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
364 dof->u.dof_reg.striped = 0;
365 /* is this replay? */
367 dof->u.dof_reg.striped = 0;
371 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
376 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
377 struct mdd_object *c, struct lu_attr *attr,
378 struct thandle *handle,
379 const struct md_op_spec *spec,
380 struct dt_allocation_hint *hint)
382 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
386 LASSERT(!mdd_object_exists(c));
388 rc = mdo_create_object(env, c, attr, hint, dof, handle);
393 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
394 const struct lu_attr *attr, struct thandle *handle,
400 rc = mdo_attr_set(env, obj, attr, handle);
401 #ifdef CONFIG_FS_POSIX_ACL
402 if (!rc && (attr->la_valid & LA_MODE) && needacl)
403 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
408 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
409 const struct lu_attr *oattr, struct lu_attr *attr,
410 struct thandle *handle)
415 LASSERT(attr->la_valid & LA_CTIME);
416 LASSERT(oattr != NULL);
418 /* Make sure the ctime is increased only, however, it's not strictly
419 * reliable at here because there is not guarantee to hold lock on
420 * object, so we just bypass some unnecessary cmtime setting first
421 * and OSD has to check it again. */
422 if (attr->la_ctime < oattr->la_ctime)
423 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
424 else if (attr->la_valid == LA_CTIME &&
425 attr->la_ctime == oattr->la_ctime)
426 attr->la_valid &= ~LA_CTIME;
428 if (attr->la_valid != 0)
429 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
434 * This gives the same functionality as the code between
435 * sys_chmod and inode_setattr
436 * chown_common and inode_setattr
437 * utimes and inode_setattr
438 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
440 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
441 const struct lu_attr *oattr, struct lu_attr *la,
442 const unsigned long flags)
451 /* Do not permit change file type */
452 if (la->la_valid & LA_TYPE)
455 /* They should not be processed by setattr */
456 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
459 LASSERT(oattr != NULL);
461 uc = lu_ucred_check(env);
465 if (la->la_valid == LA_CTIME) {
466 if (!(flags & MDS_PERM_BYPASS))
467 /* This is only for set ctime when rename's source is
469 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
471 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
472 la->la_valid &= ~LA_CTIME;
476 if (la->la_valid == LA_ATIME) {
477 /* This is an atime-only attribute update for close RPCs. */
478 if (la->la_atime < (oattr->la_atime +
479 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
480 la->la_valid &= ~LA_ATIME;
484 /* Check if flags change. */
485 if (la->la_valid & LA_FLAGS) {
486 unsigned int oldflags = oattr->la_flags &
487 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
488 unsigned int newflags = la->la_flags &
489 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
491 if ((uc->uc_fsuid != oattr->la_uid) &&
492 !md_capable(uc, CFS_CAP_FOWNER))
495 /* The IMMUTABLE and APPEND_ONLY flags can
496 * only be changed by the relevant capability. */
497 if ((oldflags ^ newflags) &&
498 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
501 if (!S_ISDIR(oattr->la_mode))
502 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
505 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
506 (la->la_valid & ~LA_FLAGS) &&
507 !(flags & MDS_PERM_BYPASS))
510 /* Check for setting the obj time. */
511 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
512 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
513 if ((uc->uc_fsuid != oattr->la_uid) &&
514 !md_capable(uc, CFS_CAP_FOWNER)) {
515 rc = mdd_permission_internal(env, obj, oattr,
522 if (la->la_valid & LA_KILL_SUID) {
523 la->la_valid &= ~LA_KILL_SUID;
524 if ((oattr->la_mode & S_ISUID) &&
525 !(la->la_valid & LA_MODE)) {
526 la->la_mode = oattr->la_mode;
527 la->la_valid |= LA_MODE;
529 la->la_mode &= ~S_ISUID;
532 if (la->la_valid & LA_KILL_SGID) {
533 la->la_valid &= ~LA_KILL_SGID;
534 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
535 (S_ISGID | S_IXGRP)) &&
536 !(la->la_valid & LA_MODE)) {
537 la->la_mode = oattr->la_mode;
538 la->la_valid |= LA_MODE;
540 la->la_mode &= ~S_ISGID;
543 /* Make sure a caller can chmod. */
544 if (la->la_valid & LA_MODE) {
545 if (!(flags & MDS_PERM_BYPASS) &&
546 (uc->uc_fsuid != oattr->la_uid) &&
547 !md_capable(uc, CFS_CAP_FOWNER))
550 if (la->la_mode == (umode_t) -1)
551 la->la_mode = oattr->la_mode;
553 la->la_mode = (la->la_mode & S_IALLUGO) |
554 (oattr->la_mode & ~S_IALLUGO);
556 /* Also check the setgid bit! */
557 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
558 la->la_gid : oattr->la_gid) &&
559 !md_capable(uc, CFS_CAP_FSETID))
560 la->la_mode &= ~S_ISGID;
562 la->la_mode = oattr->la_mode;
565 /* Make sure a caller can chown. */
566 if (la->la_valid & LA_UID) {
567 if (la->la_uid == (uid_t) -1)
568 la->la_uid = oattr->la_uid;
569 if (((uc->uc_fsuid != oattr->la_uid) ||
570 (la->la_uid != oattr->la_uid)) &&
571 !md_capable(uc, CFS_CAP_CHOWN))
574 /* If the user or group of a non-directory has been
575 * changed by a non-root user, remove the setuid bit.
576 * 19981026 David C Niemi <niemi@tux.org>
578 * Changed this to apply to all users, including root,
579 * to avoid some races. This is the behavior we had in
580 * 2.0. The check for non-root was definitely wrong
581 * for 2.2 anyway, as it should have been using
582 * CAP_FSETID rather than fsuid -- 19990830 SD. */
583 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
584 !S_ISDIR(oattr->la_mode)) {
585 la->la_mode &= ~S_ISUID;
586 la->la_valid |= LA_MODE;
590 /* Make sure caller can chgrp. */
591 if (la->la_valid & LA_GID) {
592 if (la->la_gid == (gid_t) -1)
593 la->la_gid = oattr->la_gid;
594 if (((uc->uc_fsuid != oattr->la_uid) ||
595 ((la->la_gid != oattr->la_gid) &&
596 !lustre_in_group_p(uc, la->la_gid))) &&
597 !md_capable(uc, CFS_CAP_CHOWN))
600 /* Likewise, if the user or group of a non-directory
601 * has been changed by a non-root user, remove the
602 * setgid bit UNLESS there is no group execute bit
603 * (this would be a file marked for mandatory
604 * locking). 19981026 David C Niemi <niemi@tux.org>
606 * Removed the fsuid check (see the comment above) --
608 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
609 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
610 la->la_mode &= ~S_ISGID;
611 la->la_valid |= LA_MODE;
615 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
616 if (!((flags & MDS_OWNEROVERRIDE) &&
617 (uc->uc_fsuid == oattr->la_uid)) &&
618 !(flags & MDS_PERM_BYPASS)) {
619 rc = mdd_permission_internal(env, obj, oattr,
626 if (la->la_valid & LA_CTIME) {
627 /* The pure setattr, it has the priority over what is
628 * already set, do not drop it if ctime is equal. */
629 if (la->la_ctime < oattr->la_ctime)
630 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
636 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
637 struct mdd_device *mdd,
638 enum changelog_rec_type type, int flags,
639 const struct lu_fid *fid,
640 struct thandle *handle)
642 const struct lu_ucred *uc = lu_ucred(env);
643 struct llog_changelog_rec *rec;
648 flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
649 if (uc != NULL && uc->uc_jobid[0] != '\0')
652 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
653 changelog_rec_offset(flags & CLF_SUPPORTED));
654 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
655 if (buf->lb_buf == NULL)
659 rec->cr_hdr.lrh_len = reclen;
660 rec->cr.cr_flags = flags;
661 rec->cr.cr_type = (__u32)type;
662 rec->cr.cr_tfid = *fid;
663 rec->cr.cr_namelen = 0;
665 if (flags & CLF_JOBID)
666 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
668 rc = mdd_changelog_store(env, mdd, rec, handle);
674 /** Store a data change changelog record
675 * If this fails, we must fail the whole transaction; we don't
676 * want the change to commit without the log entry.
677 * \param mdd_obj - mdd_object of change
678 * \param handle - transaction handle
680 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
681 enum changelog_rec_type type, int flags,
682 struct mdd_object *mdd_obj, struct thandle *handle)
686 LASSERT(mdd_obj != NULL);
687 LASSERT(handle != NULL);
690 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
692 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
695 if (mdd_is_volatile_obj(mdd_obj))
698 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
699 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
700 /* Don't need multiple updates in this log */
701 /* Don't check under lock - no big deal if we get an extra
706 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
707 mdo2fid(mdd_obj), handle);
709 mdd_obj->mod_cltime = ktime_get();
714 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
715 int flags, struct md_device *m, const struct lu_fid *fid)
717 struct thandle *handle;
718 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
723 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
725 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
728 LASSERT(fid != NULL);
730 handle = mdd_trans_create(env, mdd);
732 RETURN(PTR_ERR(handle));
734 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
738 rc = mdd_trans_start(env, mdd, handle);
742 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
746 rc = mdd_trans_stop(env, mdd, rc, handle);
752 * Save LMA extended attributes with data from \a ma.
754 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
755 * not, LMA EA will be first read from disk, modified and write back.
758 /* Precedence for choosing record type when multiple
759 * attributes change: setattr > mtime > ctime > atime
760 * (ctime changes when mtime does, plus chmod/chown.
761 * atime and ctime are independent.) */
762 static int mdd_attr_set_changelog(const struct lu_env *env,
763 struct md_object *obj, struct thandle *handle,
766 struct mdd_device *mdd = mdo2mdd(obj);
769 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
770 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
771 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
772 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
773 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
774 bits = bits & mdd->mdd_cl.mc_mask;
775 /* This is an implementation limit rather than a protocol limit */
776 CLASSERT(CL_LAST <= sizeof(int) * 8);
780 /* The record type is the lowest non-masked set bit */
783 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
784 return mdd_changelog_data_store(env, mdd, type, (int)valid,
785 md2mdd_obj(obj), handle);
788 static int mdd_declare_attr_set(const struct lu_env *env,
789 struct mdd_device *mdd,
790 struct mdd_object *obj,
791 const struct lu_attr *attr,
792 struct thandle *handle)
796 rc = mdo_declare_attr_set(env, obj, attr, handle);
800 #ifdef CONFIG_FS_POSIX_ACL
801 if (attr->la_valid & LA_MODE) {
802 mdd_read_lock(env, obj, MOR_TGT_CHILD);
803 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
804 XATTR_NAME_ACL_ACCESS);
805 mdd_read_unlock(env, obj);
806 if (rc == -EOPNOTSUPP || rc == -ENODATA)
812 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
813 rc = mdo_declare_xattr_set(env, obj, buf,
814 XATTR_NAME_ACL_ACCESS, 0,
822 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
829 * permission changes may require sync operation, to mitigate performance
830 * impact, only do this for dir and when permission is reduced.
832 * For regular files, version is updated with permission change (see VBR), async
833 * permission won't cause any issue, while missing permission change on
834 * directory may affect accessibility of other objects after recovery.
836 static inline bool permission_needs_sync(const struct lu_attr *old,
837 const struct lu_attr *new)
839 if (!S_ISDIR(old->la_mode))
842 if (new->la_valid & (LA_UID | LA_GID))
845 if (new->la_valid & LA_MODE &&
846 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
849 if ((new->la_valid & LA_MODE) &&
850 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
851 (old->la_mode & S_IRWXUGO))
857 /* set attr and LOV EA at once, return updated attr */
858 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
859 const struct md_attr *ma)
861 struct mdd_object *mdd_obj = md2mdd_obj(obj);
862 struct mdd_device *mdd = mdo2mdd(obj);
863 struct thandle *handle;
864 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
865 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
866 const struct lu_attr *la = &ma->ma_attr;
870 /* we do not use ->attr_set() for LOV/HSM EA any more */
871 LASSERT((ma->ma_valid & MA_LOV) == 0);
872 LASSERT((ma->ma_valid & MA_HSM) == 0);
874 rc = mdd_la_get(env, mdd_obj, attr);
878 *la_copy = ma->ma_attr;
879 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
883 /* no need to setattr anymore */
884 if (la_copy->la_valid == 0) {
885 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
886 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
887 PFID(mdo2fid(mdd_obj)), la->la_valid);
892 handle = mdd_trans_create(env, mdd);
894 RETURN(PTR_ERR(handle));
896 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
900 rc = mdd_trans_start(env, mdd, handle);
904 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
907 if (la->la_valid & (LA_MTIME | LA_CTIME))
908 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
909 la->la_mtime, la->la_ctime);
911 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
912 if (la_copy->la_valid & LA_FLAGS)
913 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
914 else if (la_copy->la_valid) /* setattr */
915 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
916 mdd_write_unlock(env, mdd_obj);
919 rc = mdd_attr_set_changelog(env, obj, handle,
925 rc = mdd_trans_stop(env, mdd, rc, handle);
930 static int mdd_xattr_sanity_check(const struct lu_env *env,
931 struct mdd_object *obj,
932 const struct lu_attr *attr,
935 struct lu_ucred *uc = lu_ucred_assert(env);
938 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
941 if (strncmp(XATTR_USER_PREFIX, name,
942 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
943 /* For sticky directories, only the owner and privileged user
944 * can write attributes. */
945 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
946 (uc->uc_fsuid != attr->la_uid) &&
947 !md_capable(uc, CFS_CAP_FOWNER))
950 if ((uc->uc_fsuid != attr->la_uid) &&
951 !md_capable(uc, CFS_CAP_FOWNER))
959 * Check if a string begins with a given prefix.
961 * \param str String to check
962 * \param prefix Substring to check at the beginning of \a str
963 * \return true/false whether the condition is verified.
965 static inline bool has_prefix(const char *str, const char *prefix)
967 return strncmp(prefix, str, strlen(prefix)) == 0;
971 * Indicate the kind of changelog to store (if any) for a xattr set/del.
973 * \param[in] xattr_name Full extended attribute name.
975 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
977 static enum changelog_rec_type
978 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
979 const char *xattr_name)
981 /* Layout changes systematically recorded */
982 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
983 strncmp(XATTR_LUSTRE_LOV, xattr_name,
984 strlen(XATTR_LUSTRE_LOV)) == 0)
987 /* HSM information changes systematically recorded */
988 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
991 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
992 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
993 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
999 static int mdd_declare_xattr_set(const struct lu_env *env,
1000 struct mdd_device *mdd,
1001 struct mdd_object *obj,
1002 const struct lu_buf *buf,
1004 int fl, struct thandle *handle)
1008 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1012 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1013 return 0; /* no changelog to store */
1015 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1019 * Compare current and future data of HSM EA and add a changelog if needed.
1021 * Caller should have write-locked \param obj.
1023 * \param buf - Future HSM EA content.
1024 * \retval 0 if no changelog is needed or changelog was added properly.
1025 * \retval -ve errno if there was a problem
1027 static int mdd_hsm_update_locked(const struct lu_env *env,
1028 struct md_object *obj,
1029 const struct lu_buf *buf,
1030 struct thandle *handle, int *cl_flags)
1032 struct mdd_thread_info *info = mdd_env_info(env);
1033 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1034 struct lu_buf *current_buf;
1035 struct md_hsm *current_mh;
1036 struct md_hsm *new_mh;
1040 OBD_ALLOC_PTR(current_mh);
1041 if (current_mh == NULL)
1044 /* Read HSM attrs from disk */
1045 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1046 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1047 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1048 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1049 if (rc < 0 && rc != -ENODATA)
1051 else if (rc == -ENODATA)
1052 current_mh->mh_flags = 0;
1054 /* Map future HSM xattr */
1055 OBD_ALLOC_PTR(new_mh);
1057 GOTO(free, rc = -ENOMEM);
1058 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1062 /* Flags differ, set flags for the changelog that will be added */
1063 if (current_mh->mh_flags != new_mh->mh_flags) {
1064 hsm_set_cl_event(cl_flags, HE_STATE);
1065 if (new_mh->mh_flags & HS_DIRTY)
1066 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1069 OBD_FREE_PTR(new_mh);
1072 OBD_FREE_PTR(current_mh);
1076 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1080 * The caller should guarantee to update the object ctime
1081 * after xattr_set if needed.
1083 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1084 const struct lu_buf *buf, const char *name,
1087 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1088 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1089 struct mdd_device *mdd = mdo2mdd(obj);
1090 struct thandle *handle;
1091 enum changelog_rec_type cl_type;
1096 rc = mdd_la_get(env, mdd_obj, attr);
1100 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1104 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1105 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1106 struct posix_acl *acl;
1108 /* user may set empty ACL, which should be treated as removing
1110 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1113 RETURN(PTR_ERR(acl));
1115 rc = mdd_xattr_del(env, obj, name);
1118 posix_acl_release(acl);
1121 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1122 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1126 handle = mdd_trans_create(env, mdd);
1128 RETURN(PTR_ERR(handle));
1130 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1134 rc = mdd_trans_start(env, mdd, handle);
1138 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1140 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1141 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1143 mdd_write_unlock(env, mdd_obj);
1148 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1149 mdd_write_unlock(env, mdd_obj);
1153 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1157 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1162 return mdd_trans_stop(env, mdd, rc, handle);
1165 static int mdd_declare_xattr_del(const struct lu_env *env,
1166 struct mdd_device *mdd,
1167 struct mdd_object *obj,
1169 struct thandle *handle)
1173 rc = mdo_declare_xattr_del(env, obj, name, handle);
1177 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1178 return 0; /* no changelog to store */
1180 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1184 * The caller should guarantee to update the object ctime
1185 * after xattr_set if needed.
1187 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1190 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1191 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1192 struct mdd_device *mdd = mdo2mdd(obj);
1193 struct thandle *handle;
1197 rc = mdd_la_get(env, mdd_obj, attr);
1201 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1205 handle = mdd_trans_create(env, mdd);
1207 RETURN(PTR_ERR(handle));
1209 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1213 rc = mdd_trans_start(env, mdd, handle);
1217 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1218 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1219 mdd_write_unlock(env, mdd_obj);
1223 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1226 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1230 return mdd_trans_stop(env, mdd, rc, handle);
1234 * read lov EA of an object
1235 * return the lov EA in an allocated lu_buf
1237 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1238 struct lu_buf *lmm_buf)
1240 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1245 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1247 if (rc == -ERANGE) {
1248 /* mti_big_buf is allocated but is too small
1249 * we need to increase it */
1250 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1252 if (buf->lb_buf == NULL)
1253 GOTO(out, rc = -ENOMEM);
1264 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1265 /* mti_big_buf was not allocated, so we have to
1266 * allocate it based on the ea size */
1267 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1269 if (buf->lb_buf == NULL)
1270 GOTO(out, rc = -ENOMEM);
1274 lu_buf_alloc(lmm_buf, bufsize);
1275 if (lmm_buf->lb_buf == NULL)
1276 GOTO(out, rc = -ENOMEM);
1278 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1284 lu_buf_free(lmm_buf);
1288 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1289 struct mdd_object *o, struct lu_buf *buf,
1290 struct thandle *handle)
1292 struct hsm_attrs *attrs;
1298 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1303 attrs = buf->lb_buf;
1304 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1305 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1308 /* Add a changelog record for release. */
1309 hsm_set_cl_event(&flags, HE_RELEASE);
1310 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1316 * check if layout swapping between 2 objects is allowed
1318 * - only normal FIDs or non-system IGIFs
1319 * - same type of objects
1320 * - same owner/group (so quotas are still valid) unless this is from HSM
1323 static int mdd_layout_swap_allowed(const struct lu_env *env,
1324 struct mdd_object *o1,
1325 const struct lu_attr *attr1,
1326 struct mdd_object *o2,
1327 const struct lu_attr *attr2,
1330 const struct lu_fid *fid1, *fid2;
1336 if (!fid_is_norm(fid1) &&
1337 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1340 if (!fid_is_norm(fid2) &&
1341 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1344 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1345 if (S_ISDIR(mdd_object_type(o1)))
1347 if (S_ISREG(mdd_object_type(o1)))
1352 if (flags & SWAP_LAYOUTS_MDS_HSM)
1355 if ((attr1->la_uid != attr2->la_uid) ||
1356 (attr1->la_gid != attr2->la_gid))
1362 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1363 * look into the layout in MDD layer. */
1364 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1366 struct lov_comp_md_v1 *comp_v1;
1367 struct lov_mds_md *v1;
1371 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1372 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1373 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1379 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1380 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1383 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1384 off = le32_to_cpu(comp_v1->lcm_entries[i].
1386 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1391 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1392 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1403 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1405 return mdd_lmm_oi(lmm, oi, true);
1408 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1410 return mdd_lmm_oi(lmm, oi, false);
1413 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1415 struct lov_comp_md_v1 *comp_v1;
1417 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1418 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1420 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1422 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1423 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1424 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1425 __u16 tmp_gen = *gen;
1427 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1429 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1436 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1438 return mdd_lmm_gen(lmm, gen, true);
1441 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1443 return mdd_lmm_gen(lmm, gen, false);
1447 * swap layouts between 2 lustre objects
1449 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1450 struct md_object *obj2, __u64 flags)
1452 struct mdd_thread_info *info = mdd_env_info(env);
1453 struct mdd_object *fst_o = md2mdd_obj(obj1);
1454 struct mdd_object *snd_o = md2mdd_obj(obj2);
1455 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1456 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1457 struct mdd_device *mdd = mdo2mdd(obj1);
1458 struct lov_mds_md *fst_lmm, *snd_lmm;
1459 struct lu_buf *fst_buf = &info->mti_buf[0];
1460 struct lu_buf *snd_buf = &info->mti_buf[1];
1461 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1462 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1463 struct ost_id *saved_oi = NULL;
1464 struct thandle *handle;
1465 __u32 fst_gen, snd_gen, saved_gen;
1471 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1472 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1474 /* we have to sort the 2 obj, so locking will always
1475 * be in the same order, even in case of 2 concurrent swaps */
1476 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1477 if (rc == 0) /* same fid ? */
1483 rc = mdd_la_get(env, fst_o, fst_la);
1487 rc = mdd_la_get(env, snd_o, snd_la);
1491 /* check if layout swapping is allowed */
1492 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1496 handle = mdd_trans_create(env, mdd);
1498 RETURN(PTR_ERR(handle));
1500 /* objects are already sorted */
1501 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1502 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1504 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1505 if (rc < 0 && rc != -ENODATA)
1508 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1509 if (rc < 0 && rc != -ENODATA)
1512 /* swapping 2 non existant layouts is a success */
1513 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1516 /* to help inode migration between MDT, it is better to
1517 * start by the no layout file (if one), so we order the swap */
1518 if (snd_buf->lb_buf == NULL) {
1520 swap(fst_buf, snd_buf);
1523 fst_gen = snd_gen = 0;
1524 /* lmm and generation layout initialization */
1525 if (fst_buf->lb_buf != NULL) {
1526 fst_lmm = fst_buf->lb_buf;
1527 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1528 fst_fl = LU_XATTR_REPLACE;
1532 fst_fl = LU_XATTR_CREATE;
1535 snd_lmm = snd_buf->lb_buf;
1536 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1538 saved_gen = fst_gen;
1539 /* increase the generation layout numbers */
1544 * XXX The layout generation is used to generate component IDs for
1545 * the composite file, we have to do some special tweaks to make
1546 * sure the layout generation is always adequate for that job.
1549 /* Skip invalid generation number for composite layout */
1550 if ((snd_gen & LCME_ID_MASK) == 0)
1552 if ((fst_gen & LCME_ID_MASK) == 0)
1554 /* Make sure the generation is greater than all the component IDs */
1555 if (fst_gen < snd_gen)
1557 else if (fst_gen > snd_gen)
1560 /* set the file specific informations in lmm */
1561 if (fst_lmm != NULL) {
1562 saved_oi = &info->mti_oa.o_oi;
1563 mdd_get_lmm_oi(fst_lmm, saved_oi);
1564 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1565 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1566 mdd_set_lmm_oi(snd_lmm, saved_oi);
1568 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1569 cpu_to_le32(LOV_MAGIC_MAGIC))
1570 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1572 GOTO(stop, rc = -EPROTO);
1574 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1576 /* Prepare HSM attribute if it's required */
1577 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1578 const int buflen = sizeof(struct hsm_attrs);
1580 lu_buf_alloc(fst_hsm_buf, buflen);
1581 lu_buf_alloc(snd_hsm_buf, buflen);
1582 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1583 GOTO(stop, rc = -ENOMEM);
1585 /* Read HSM attribute */
1586 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1590 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1594 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1595 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1600 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1601 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1607 /* prepare transaction */
1608 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1613 if (fst_buf->lb_buf != NULL)
1614 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1615 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1618 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1623 rc = mdd_trans_start(env, mdd, handle);
1627 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1628 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1632 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1634 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1637 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1638 mdd_obj_dev_name(fst_o),
1639 PFID(mdo2fid(fst_o)), rc, rc2);
1644 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1648 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1651 if (fst_buf->lb_buf != NULL)
1652 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1653 LU_XATTR_REPLACE, handle);
1655 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1661 /* failure on second file, but first was done, so we have
1662 * to roll back first. */
1663 if (fst_buf->lb_buf != NULL) {
1664 mdd_set_lmm_oi(fst_lmm, saved_oi);
1665 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1666 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1667 LU_XATTR_REPLACE, handle);
1669 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1675 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1680 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1685 CERROR("%s: unable to roll back layout swap. FIDs: "
1686 DFID" and "DFID "error: %d/%d, steps: %d\n",
1687 mdd_obj_dev_name(fst_o),
1688 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1690 /* a solution to avoid journal commit is to panic,
1691 * but it has strong consequences so we use LBUG to
1692 * allow sysdamin to choose to panic or not
1699 /* Issue one changelog record per file */
1700 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1704 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1710 rc = mdd_trans_stop(env, mdd, rc, handle);
1712 mdd_write_unlock(env, snd_o);
1713 mdd_write_unlock(env, fst_o);
1715 lu_buf_free(fst_buf);
1716 lu_buf_free(snd_buf);
1717 lu_buf_free(fst_hsm_buf);
1718 lu_buf_free(snd_hsm_buf);
1722 static int mdd_declare_layout_change(const struct lu_env *env,
1723 struct mdd_device *mdd,
1724 struct mdd_object *obj,
1725 struct layout_intent *layout,
1726 const struct lu_buf *buf,
1727 struct thandle *handle)
1731 rc = mdo_declare_layout_change(env, obj, layout, buf, handle);
1735 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1738 /* For PFL, this is used to instantiate necessary component objects. */
1739 int mdd_layout_change(const struct lu_env *env, struct md_object *obj,
1740 struct layout_intent *layout, const struct lu_buf *buf)
1742 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1743 struct mdd_device *mdd = mdo2mdd(obj);
1744 struct thandle *handle;
1748 handle = mdd_trans_create(env, mdd);
1750 RETURN(PTR_ERR(handle));
1752 rc = mdd_declare_layout_change(env, mdd, mdd_obj, layout, buf, handle);
1754 * It's possible that another layout write intent has already
1755 * instantiated our objects, so a -EALREADY returned, and we need to
1759 GOTO(stop, rc = (rc == -EALREADY) ? 0 : rc);
1761 rc = mdd_trans_start(env, mdd, handle);
1765 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1766 rc = mdo_layout_change(env, mdd_obj, layout, buf, handle);
1767 mdd_write_unlock(env, mdd_obj);
1771 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, handle);
1773 RETURN(mdd_trans_stop(env, mdd, rc, handle));
1776 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
1777 struct mdd_object *child, const struct lu_attr *attr,
1778 const struct md_op_spec *spec,
1779 struct dt_allocation_hint *hint)
1781 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
1782 struct dt_object *nc = mdd_object_child(child);
1784 memset(hint, 0, sizeof(*hint));
1786 /* For striped directory, give striping EA to lod_ah_init, which will
1787 * decide the stripe_offset and stripe count by it. */
1788 if (S_ISDIR(attr->la_mode) &&
1789 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
1790 hint->dah_eadata = spec->u.sp_ea.eadata;
1791 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
1793 hint->dah_eadata = NULL;
1794 hint->dah_eadata_len = 0;
1797 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
1798 hint->dah_eadata, hint->dah_eadata_len);
1799 /* @hint will be initialized by underlying device. */
1800 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
1804 * do NOT or the MAY_*'s, you'll get the weakest
1806 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
1810 /* Sadly, NFSD reopens a file repeatedly during operation, so the
1811 * "acc_mode = 0" allowance for newly-created files isn't honoured.
1812 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
1813 * owner can write to a file even if it is marked readonly to hide
1814 * its brokenness. (bug 5781) */
1815 if (flags & MDS_OPEN_OWNEROVERRIDE) {
1816 struct lu_ucred *uc = lu_ucred_check(env);
1818 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
1822 if (flags & FMODE_READ)
1824 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
1826 if (flags & MDS_FMODE_EXEC)
1831 static int mdd_open_sanity_check(const struct lu_env *env,
1832 struct mdd_object *obj,
1833 const struct lu_attr *attr, int flag)
1839 if (mdd_is_dead_obj(obj))
1842 if (S_ISLNK(attr->la_mode))
1845 mode = accmode(env, attr, flag);
1847 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
1850 if (!(flag & MDS_OPEN_CREATED)) {
1851 rc = mdd_permission_internal(env, obj, attr, mode);
1856 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
1857 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
1858 flag &= ~MDS_OPEN_TRUNC;
1860 /* For writing append-only file must open it with append mode. */
1861 if (attr->la_flags & LUSTRE_APPEND_FL) {
1862 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
1864 if (flag & MDS_OPEN_TRUNC)
1871 static int mdd_open(const struct lu_env *env, struct md_object *obj,
1874 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1875 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1878 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1880 rc = mdd_la_get(env, mdd_obj, attr);
1884 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
1888 mdd_obj->mod_count++;
1891 mdd_write_unlock(env, mdd_obj);
1895 static int mdd_declare_close(const struct lu_env *env,
1896 struct mdd_object *obj,
1898 struct thandle *handle)
1902 rc = orph_declare_index_delete(env, obj, handle);
1906 return mdo_declare_destroy(env, obj, handle);
1910 * No permission check is needed.
1912 static int mdd_close(const struct lu_env *env, struct md_object *obj,
1913 struct md_attr *ma, int mode)
1915 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1916 struct mdd_device *mdd = mdo2mdd(obj);
1917 struct thandle *handle = NULL;
1920 bool blocked = false;
1923 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
1924 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1925 mdd_obj->mod_count--;
1926 mdd_write_unlock(env, mdd_obj);
1928 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
1929 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
1930 "list\n", PFID(mdd_object_fid(mdd_obj)));
1934 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
1935 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
1936 /* check without any lock */
1937 is_orphan = mdd_obj->mod_count == 1 &&
1938 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
1942 /* mdd_trans_create() maybe failed because of barrier_entry(),
1943 * under such case, the orphan MDT-object will be left in the
1944 * orphan list, and when the MDT remount next time, the unused
1945 * orphans will be destroyed automatically.
1947 * One exception: the former mdd_finish_unlink may failed to
1948 * add the orphan MDT-object to the orphan list, then if the
1949 * mdd_trans_create() failed because of barrier_entry(), the
1950 * MDT-object will become real orphan that is neither in the
1951 * namespace nor in the orphan list. Such bad case should be
1952 * very rare and will be handled by e2fsck/lfsck. */
1953 handle = mdd_trans_create(env, mdo2mdd(obj));
1954 if (IS_ERR(handle)) {
1955 rc = PTR_ERR(handle);
1956 if (rc != -EINPROGRESS)
1964 rc = mdd_declare_close(env, mdd_obj, ma, handle);
1968 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1972 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
1978 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1979 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
1981 CERROR("Failed to get lu_attr of "DFID": %d\n",
1982 PFID(mdd_object_fid(mdd_obj)), rc);
1986 /* check again with lock */
1987 is_orphan = (mdd_obj->mod_count == 1) &&
1988 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
1989 ma->ma_attr.la_nlink == 0);
1991 if (is_orphan && !handle && !blocked) {
1992 mdd_write_unlock(env, mdd_obj);
1996 mdd_obj->mod_count--; /*release open count */
1998 if (!is_orphan || blocked)
2002 /* NB: Object maybe not in orphan list originally, it is rare case for
2003 * mdd_finish_unlink() failure, in that case, the object doesn't have
2004 * ORPHAN_OBJ flag */
2005 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2006 /* remove link to object from orphan index */
2007 LASSERT(handle != NULL);
2008 rc = __mdd_orphan_del(env, mdd_obj, handle);
2010 CERROR("%s: unable to delete "DFID" from orphan list: "
2011 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2012 PFID(mdd_object_fid(mdd_obj)), rc);
2013 /* If object was not deleted from orphan list, do not
2014 * destroy OSS objects, which will be done when next
2019 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2020 "list, OSS objects to be destroyed.\n",
2021 PFID(mdd_object_fid(mdd_obj)));
2024 rc = mdo_destroy(env, mdd_obj, handle);
2027 CERROR("%s: unable to delete "DFID" from orphan list: "
2028 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2029 PFID(mdd_object_fid(mdd_obj)), rc);
2034 mdd_write_unlock(env, mdd_obj);
2036 if (!rc && !blocked &&
2037 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2038 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2039 if (handle == NULL) {
2040 handle = mdd_trans_create(env, mdo2mdd(obj));
2042 GOTO(stop, rc = PTR_ERR(handle));
2044 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2049 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2054 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2059 if (handle != NULL && !IS_ERR(handle))
2060 rc = mdd_trans_stop(env, mdd, rc, handle);
2066 * Permission check is done when open,
2067 * no need check again.
2069 static int mdd_readpage_sanity_check(const struct lu_env *env,
2070 struct mdd_object *obj)
2072 struct dt_object *next = mdd_object_child(obj);
2076 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2084 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2085 size_t nob, const struct dt_it_ops *iops,
2086 struct dt_it *it, __u32 attr, void *arg)
2088 struct lu_dirpage *dp = &lp->lp_dir;
2092 struct lu_dirent *ent;
2093 struct lu_dirent *last = NULL;
2097 if (nob < sizeof(*dp))
2100 memset(area, 0, sizeof (*dp));
2101 area += sizeof (*dp);
2102 nob -= sizeof (*dp);
2109 len = iops->key_size(env, it);
2111 /* IAM iterator can return record with zero len. */
2115 hash = iops->store(env, it);
2116 if (unlikely(first)) {
2118 dp->ldp_hash_start = cpu_to_le64(hash);
2121 /* calculate max space required for lu_dirent */
2122 recsize = lu_dirent_calc_size(len, attr);
2124 if (nob >= recsize) {
2125 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2126 if (result == -ESTALE)
2131 /* osd might not able to pack all attributes,
2132 * so recheck rec length */
2133 recsize = le16_to_cpu(ent->lde_reclen);
2135 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2136 fid_le_to_cpu(&fid, &ent->lde_fid);
2137 if (fid_is_dot_lustre(&fid))
2141 result = (last != NULL) ? 0 :-EINVAL;
2145 ent = (void *)ent + recsize;
2149 result = iops->next(env, it);
2150 if (result == -ESTALE)
2152 } while (result == 0);
2155 dp->ldp_hash_end = cpu_to_le64(hash);
2157 if (last->lde_hash == dp->ldp_hash_end)
2158 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2159 last->lde_reclen = 0; /* end mark */
2162 /* end of directory */
2163 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2164 else if (result < 0)
2165 CWARN("build page failed: %d!\n", result);
2169 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2170 const struct lu_rdpg *rdpg)
2172 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2176 if (mdd_object_exists(mdd_obj) == 0) {
2177 CERROR("%s: object "DFID" not found: rc = -2\n",
2178 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2182 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2183 rc = mdd_readpage_sanity_check(env, mdd_obj);
2185 GOTO(out_unlock, rc);
2187 if (mdd_is_dead_obj(mdd_obj)) {
2189 struct lu_dirpage *dp;
2192 * According to POSIX, please do not return any entry to client:
2193 * even dot and dotdot should not be returned.
2195 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2196 PFID(mdd_object_fid(mdd_obj)));
2198 if (rdpg->rp_count <= 0)
2199 GOTO(out_unlock, rc = -EFAULT);
2200 LASSERT(rdpg->rp_pages != NULL);
2202 pg = rdpg->rp_pages[0];
2203 dp = (struct lu_dirpage *)kmap(pg);
2204 memset(dp, 0 , sizeof(struct lu_dirpage));
2205 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2206 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2207 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2209 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2212 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2213 mdd_dir_page_build, NULL);
2215 struct lu_dirpage *dp;
2217 dp = kmap(rdpg->rp_pages[0]);
2218 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2221 * No pages were processed, mark this for first page
2224 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2225 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2226 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2228 kunmap(rdpg->rp_pages[0]);
2231 GOTO(out_unlock, rc);
2233 mdd_read_unlock(env, mdd_obj);
2237 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2239 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2241 if (mdd_object_exists(mdd_obj) == 0) {
2244 CERROR("%s: object "DFID" not found: rc = %d\n",
2245 mdd_obj_dev_name(mdd_obj),
2246 PFID(mdd_object_fid(mdd_obj)), rc);
2249 return dt_object_sync(env, mdd_object_child(mdd_obj),
2253 static int mdd_object_lock(const struct lu_env *env,
2254 struct md_object *obj,
2255 struct lustre_handle *lh,
2256 struct ldlm_enqueue_info *einfo,
2257 union ldlm_policy_data *policy)
2259 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2260 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2264 static int mdd_object_unlock(const struct lu_env *env,
2265 struct md_object *obj,
2266 struct ldlm_enqueue_info *einfo,
2267 union ldlm_policy_data *policy)
2269 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2270 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2273 const struct md_object_operations mdd_obj_ops = {
2274 .moo_permission = mdd_permission,
2275 .moo_attr_get = mdd_attr_get,
2276 .moo_attr_set = mdd_attr_set,
2277 .moo_xattr_get = mdd_xattr_get,
2278 .moo_xattr_set = mdd_xattr_set,
2279 .moo_xattr_list = mdd_xattr_list,
2280 .moo_invalidate = mdd_invalidate,
2281 .moo_xattr_del = mdd_xattr_del,
2282 .moo_swap_layouts = mdd_swap_layouts,
2283 .moo_open = mdd_open,
2284 .moo_close = mdd_close,
2285 .moo_readpage = mdd_readpage,
2286 .moo_readlink = mdd_readlink,
2287 .moo_changelog = mdd_changelog,
2288 .moo_object_sync = mdd_object_sync,
2289 .moo_object_lock = mdd_object_lock,
2290 .moo_object_unlock = mdd_object_unlock,
2291 .moo_layout_change = mdd_layout_change,