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, 2017, 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;
646 int xflags = CLFE_INVALID;
649 flags = (flags & CLF_FLAGMASK) | CLF_VERSION | CLF_EXTRA_FLAGS;
652 if (uc->uc_jobid[0] != '\0')
654 xflags |= CLFE_UIDGID;
660 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
661 changelog_rec_offset(flags & CLF_SUPPORTED,
662 xflags & CLFE_SUPPORTED));
663 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
664 if (buf->lb_buf == NULL)
668 rec->cr_hdr.lrh_len = reclen;
669 rec->cr.cr_flags = flags;
670 rec->cr.cr_type = (__u32)type;
671 rec->cr.cr_tfid = *fid;
672 rec->cr.cr_namelen = 0;
674 if (flags & CLF_JOBID)
675 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
677 if (flags & CLF_EXTRA_FLAGS) {
678 mdd_changelog_rec_ext_extra_flags(&rec->cr, xflags);
679 if (xflags & CLFE_UIDGID)
680 mdd_changelog_rec_extra_uidgid(&rec->cr,
681 uc->uc_uid, uc->uc_gid);
682 if (xflags & CLFE_NID)
683 mdd_changelog_rec_extra_nid(&rec->cr, uc->uc_nid);
684 if (xflags & CLFE_OPEN)
685 mdd_changelog_rec_extra_omode(&rec->cr, flags);
688 rc = mdd_changelog_store(env, mdd, rec, handle);
693 /** Store a data change changelog record
694 * If this fails, we must fail the whole transaction; we don't
695 * want the change to commit without the log entry.
696 * \param mdd_obj - mdd_object of change
697 * \param handle - transaction handle
699 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
700 enum changelog_rec_type type, int flags,
701 struct mdd_object *mdd_obj, struct thandle *handle)
705 LASSERT(mdd_obj != NULL);
706 LASSERT(handle != NULL);
709 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
711 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
714 if (mdd_is_volatile_obj(mdd_obj))
717 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
718 ktime_before(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
719 /* Don't need multiple updates in this log */
720 /* Don't check under lock - no big deal if we get an extra
725 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
726 mdo2fid(mdd_obj), handle);
728 mdd_obj->mod_cltime = ktime_get();
733 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
734 int flags, struct md_device *m, const struct lu_fid *fid)
736 struct thandle *handle;
737 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
742 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
744 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
747 LASSERT(fid != NULL);
749 handle = mdd_trans_create(env, mdd);
751 RETURN(PTR_ERR(handle));
753 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
757 rc = mdd_trans_start(env, mdd, handle);
761 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
765 rc = mdd_trans_stop(env, mdd, rc, handle);
771 * Save LMA extended attributes with data from \a ma.
773 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
774 * not, LMA EA will be first read from disk, modified and write back.
777 /* Precedence for choosing record type when multiple
778 * attributes change: setattr > mtime > ctime > atime
779 * (ctime changes when mtime does, plus chmod/chown.
780 * atime and ctime are independent.) */
781 static int mdd_attr_set_changelog(const struct lu_env *env,
782 struct md_object *obj, struct thandle *handle,
785 struct mdd_device *mdd = mdo2mdd(obj);
788 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
789 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
790 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
791 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
792 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
793 bits = bits & mdd->mdd_cl.mc_mask;
794 /* This is an implementation limit rather than a protocol limit */
795 CLASSERT(CL_LAST <= sizeof(int) * 8);
799 /* The record type is the lowest non-masked set bit */
802 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
803 return mdd_changelog_data_store(env, mdd, type, (int)valid,
804 md2mdd_obj(obj), handle);
807 static int mdd_declare_attr_set(const struct lu_env *env,
808 struct mdd_device *mdd,
809 struct mdd_object *obj,
810 const struct lu_attr *attr,
811 struct thandle *handle)
815 rc = mdo_declare_attr_set(env, obj, attr, handle);
819 #ifdef CONFIG_FS_POSIX_ACL
820 if (attr->la_valid & LA_MODE) {
821 mdd_read_lock(env, obj, MOR_TGT_CHILD);
822 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
823 XATTR_NAME_ACL_ACCESS);
824 mdd_read_unlock(env, obj);
825 if (rc == -EOPNOTSUPP || rc == -ENODATA)
831 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
832 rc = mdo_declare_xattr_set(env, obj, buf,
833 XATTR_NAME_ACL_ACCESS, 0,
841 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
848 * permission changes may require sync operation, to mitigate performance
849 * impact, only do this for dir and when permission is reduced.
851 * For regular files, version is updated with permission change (see VBR), async
852 * permission won't cause any issue, while missing permission change on
853 * directory may affect accessibility of other objects after recovery.
855 static inline bool permission_needs_sync(const struct lu_attr *old,
856 const struct lu_attr *new)
858 if (!S_ISDIR(old->la_mode))
861 if (new->la_valid & (LA_UID | LA_GID))
864 if (new->la_valid & LA_MODE &&
865 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
868 if ((new->la_valid & LA_MODE) &&
869 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
870 (old->la_mode & S_IRWXUGO))
876 /* set attr and LOV EA at once, return updated attr */
877 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
878 const struct md_attr *ma)
880 struct mdd_object *mdd_obj = md2mdd_obj(obj);
881 struct mdd_device *mdd = mdo2mdd(obj);
882 struct thandle *handle = NULL;
883 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
884 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
885 const struct lu_attr *la = &ma->ma_attr;
890 /* we do not use ->attr_set() for LOV/HSM EA any more */
891 LASSERT((ma->ma_valid & MA_LOV) == 0);
892 LASSERT((ma->ma_valid & MA_HSM) == 0);
894 rc = mdd_la_get(env, mdd_obj, attr);
898 *la_copy = ma->ma_attr;
899 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
903 /* no need to setattr anymore */
904 if (la_copy->la_valid == 0) {
905 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
906 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
907 PFID(mdo2fid(mdd_obj)), la->la_valid);
912 /* If an unprivileged user changes group of some file,
913 * the setattr operation will be processed synchronously to
914 * honor the quota limit of the corresponding group. see LU-5152 */
915 uc = lu_ucred_check(env);
916 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
917 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
918 la_copy->la_valid |= LA_FLAGS;
919 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
921 /* Flush the possible existing sync requests to OSTs to
922 * keep the order of sync for the current setattr operation
923 * will be sent directly to OSTs. see LU-5152 */
924 rc = dt_sync(env, mdd->mdd_child);
929 handle = mdd_trans_create(env, mdd);
930 if (IS_ERR(handle)) {
931 rc = PTR_ERR(handle);
937 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
941 rc = mdd_trans_start(env, mdd, handle);
945 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
948 if (la->la_valid & (LA_MTIME | LA_CTIME))
949 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
950 la->la_mtime, la->la_ctime);
952 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
953 if (la_copy->la_valid) {
954 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
956 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
957 /* rollback to the original gid */
958 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
959 la_copy->la_gid = attr->la_gid;
960 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
963 mdd_write_unlock(env, mdd_obj);
967 rc = mdd_attr_set_changelog(env, obj, handle,
971 rc = mdd_trans_stop(env, mdd, rc, handle);
976 static int mdd_xattr_sanity_check(const struct lu_env *env,
977 struct mdd_object *obj,
978 const struct lu_attr *attr,
981 struct lu_ucred *uc = lu_ucred_assert(env);
984 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
987 if (strncmp(XATTR_USER_PREFIX, name,
988 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
989 /* For sticky directories, only the owner and privileged user
990 * can write attributes. */
991 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
992 (uc->uc_fsuid != attr->la_uid) &&
993 !md_capable(uc, CFS_CAP_FOWNER))
996 if ((uc->uc_fsuid != attr->la_uid) &&
997 !md_capable(uc, CFS_CAP_FOWNER))
1005 * Check if a string begins with a given prefix.
1007 * \param str String to check
1008 * \param prefix Substring to check at the beginning of \a str
1009 * \return true/false whether the condition is verified.
1011 static inline bool has_prefix(const char *str, const char *prefix)
1013 return strncmp(prefix, str, strlen(prefix)) == 0;
1017 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1019 * \param[in] xattr_name Full extended attribute name.
1021 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1023 static enum changelog_rec_type
1024 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1025 const char *xattr_name)
1027 /* Layout changes systematically recorded */
1028 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1029 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1030 strlen(XATTR_LUSTRE_LOV)) == 0)
1033 /* HSM information changes systematically recorded */
1034 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1037 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1038 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1039 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT))
1045 static int mdd_declare_xattr_set(const struct lu_env *env,
1046 struct mdd_device *mdd,
1047 struct mdd_object *obj,
1048 const struct lu_buf *buf,
1050 int fl, struct thandle *handle)
1054 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1058 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1059 return 0; /* no changelog to store */
1061 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1065 * Compare current and future data of HSM EA and add a changelog if needed.
1067 * Caller should have write-locked \param obj.
1069 * \param buf - Future HSM EA content.
1070 * \retval 0 if no changelog is needed or changelog was added properly.
1071 * \retval -ve errno if there was a problem
1073 static int mdd_hsm_update_locked(const struct lu_env *env,
1074 struct md_object *obj,
1075 const struct lu_buf *buf,
1076 struct thandle *handle, int *cl_flags)
1078 struct mdd_thread_info *info = mdd_env_info(env);
1079 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1080 struct lu_buf *current_buf;
1081 struct md_hsm *current_mh;
1082 struct md_hsm *new_mh;
1086 OBD_ALLOC_PTR(current_mh);
1087 if (current_mh == NULL)
1090 /* Read HSM attrs from disk */
1091 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1092 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1093 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1094 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1095 if (rc < 0 && rc != -ENODATA)
1097 else if (rc == -ENODATA)
1098 current_mh->mh_flags = 0;
1100 /* Map future HSM xattr */
1101 OBD_ALLOC_PTR(new_mh);
1103 GOTO(free, rc = -ENOMEM);
1104 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1108 /* Flags differ, set flags for the changelog that will be added */
1109 if (current_mh->mh_flags != new_mh->mh_flags) {
1110 hsm_set_cl_event(cl_flags, HE_STATE);
1111 if (new_mh->mh_flags & HS_DIRTY)
1112 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1115 OBD_FREE_PTR(new_mh);
1118 OBD_FREE_PTR(current_mh);
1122 static int mdd_object_pfid_replace(const struct lu_env *env,
1123 struct mdd_object *o)
1125 struct mdd_device *mdd = mdo2mdd(&o->mod_obj);
1126 struct thandle *handle;
1129 handle = mdd_trans_create(env, mdd);
1131 RETURN(PTR_ERR(handle));
1133 handle->th_complex = 1;
1135 /* it doesn't need to track the PFID update via llog, because LFSCK
1136 * will repair it even it goes wrong */
1137 rc = mdd_declare_xattr_set(env, mdd, o, NULL, XATTR_NAME_FID,
1142 rc = mdd_trans_start(env, mdd, handle);
1146 rc = mdo_xattr_set(env, o, NULL, XATTR_NAME_FID, 0, handle);
1151 mdd_trans_stop(env, mdd, rc, handle);
1156 static int mdd_declare_xattr_del(const struct lu_env *env,
1157 struct mdd_device *mdd,
1158 struct mdd_object *obj,
1160 struct thandle *handle);
1162 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1165 static int mdd_xattr_merge(const struct lu_env *env, struct md_object *md_obj,
1166 struct md_object *md_vic)
1168 struct mdd_device *mdd = mdo2mdd(md_obj);
1169 struct mdd_object *obj = md2mdd_obj(md_obj);
1170 struct mdd_object *vic = md2mdd_obj(md_vic);
1171 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1172 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[1];
1173 struct lov_mds_md *lmm;
1174 struct thandle *handle;
1178 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1179 if (rc == 0) /* same fid */
1182 handle = mdd_trans_create(env, mdd);
1184 RETURN(PTR_ERR(handle));
1187 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1188 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1190 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1191 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1194 /* get EA of victim file */
1195 memset(buf_vic, 0, sizeof(*buf_vic));
1196 rc = mdd_get_lov_ea(env, vic, buf_vic);
1203 /* parse the layout of victim file */
1204 lmm = buf_vic->lb_buf;
1205 if (le32_to_cpu(lmm->lmm_magic) != LOV_MAGIC_COMP_V1)
1206 GOTO(out, rc = -EINVAL);
1208 /* save EA of target file for restore */
1209 memset(buf, 0, sizeof(*buf));
1210 rc = mdd_get_lov_ea(env, obj, buf);
1214 /* Get rid of the layout from victim object */
1215 rc = mdd_declare_xattr_del(env, mdd, vic, XATTR_NAME_LOV, handle);
1219 rc = mdd_declare_xattr_set(env, mdd, obj, buf_vic, XATTR_LUSTRE_LOV,
1220 LU_XATTR_MERGE, handle);
1224 rc = mdd_trans_start(env, mdd, handle);
1228 rc = mdo_xattr_set(env, obj, buf_vic, XATTR_LUSTRE_LOV, LU_XATTR_MERGE,
1233 rc = mdo_xattr_del(env, vic, XATTR_NAME_LOV, handle);
1235 GOTO(out_restore, rc);
1237 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1238 (void)mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1243 int rc2 = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV,
1244 LU_XATTR_REPLACE, handle);
1246 CERROR("%s: failed to rollback of layout of: "DFID
1247 ": %d, file state unknown\n",
1248 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1252 mdd_trans_stop(env, mdd, rc, handle);
1253 mdd_write_unlock(env, obj);
1254 mdd_write_unlock(env, vic);
1256 lu_buf_free(buf_vic);
1259 (void) mdd_object_pfid_replace(env, obj);
1265 * Extract the mirror with specified mirror id, and store the splitted
1266 * mirror layout to @buf.
1268 * \param[in] comp_v1 mirrored layout
1269 * \param[in] mirror_id the mirror with mirror_id to be extracted
1270 * \param[out] buf store the layout excluding the extracted mirror,
1271 * caller free the buffer we allocated in this function
1272 * \param[out] buf_vic store the extracted layout, caller free the buffer
1273 * we allocated in this function
1275 * \retval 0 on success; < 0 if error happens
1277 static int mdd_split_ea(struct lov_comp_md_v1 *comp_v1, __u16 mirror_id,
1278 struct lu_buf *buf, struct lu_buf *buf_vic)
1280 struct lov_comp_md_v1 *comp_rem;
1281 struct lov_comp_md_v1 *comp_vic;
1282 struct lov_comp_md_entry_v1 *entry;
1283 struct lov_comp_md_entry_v1 *entry_rem;
1284 struct lov_comp_md_entry_v1 *entry_vic;
1286 __u16 comp_cnt, count = 0;
1287 int lmm_size, lmm_size_vic = 0;
1289 int offset, offset_rem, offset_vic;
1291 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1292 /* comp_v1 should contains more than 1 mirror */
1293 if (mirror_cnt <= 1)
1295 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1296 lmm_size = le32_to_cpu(comp_v1->lcm_size);
1298 for (i = 0; i < comp_cnt; i++) {
1299 entry = &comp_v1->lcm_entries[i];
1300 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id) {
1302 lmm_size_vic += sizeof(*entry);
1303 lmm_size_vic += le32_to_cpu(entry->lcme_size);
1304 } else if (count > 0) {
1305 /* find the specified mirror */
1313 lu_buf_alloc(buf, lmm_size - lmm_size_vic);
1317 lu_buf_alloc(buf_vic, sizeof(*comp_vic) + lmm_size_vic);
1318 if (!buf_vic->lb_buf) {
1323 comp_rem = (struct lov_comp_md_v1 *)buf->lb_buf;
1324 comp_vic = (struct lov_comp_md_v1 *)buf_vic->lb_buf;
1326 memcpy(comp_rem, comp_v1, sizeof(*comp_v1));
1327 comp_rem->lcm_mirror_count = cpu_to_le16(mirror_cnt - 2);
1328 comp_rem->lcm_entry_count = cpu_to_le32(comp_cnt - count);
1329 comp_rem->lcm_size = cpu_to_le32(lmm_size - lmm_size_vic);
1330 if (!comp_rem->lcm_mirror_count)
1331 comp_rem->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1333 memset(comp_vic, 0, sizeof(*comp_v1));
1334 comp_vic->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1335 comp_vic->lcm_mirror_count = 0;
1336 comp_vic->lcm_entry_count = cpu_to_le32(count);
1337 comp_vic->lcm_size = cpu_to_le32(lmm_size_vic + sizeof(*comp_vic));
1338 comp_vic->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1339 comp_vic->lcm_layout_gen = 0;
1341 offset = sizeof(*comp_v1) + sizeof(*entry) * comp_cnt;
1342 offset_rem = sizeof(*comp_rem) +
1343 sizeof(*entry_rem) * (comp_cnt - count);
1344 offset_vic = sizeof(*comp_vic) + sizeof(*entry_vic) * count;
1345 for (i = j = k = 0; i < comp_cnt; i++) {
1346 struct lov_mds_md *lmm, *lmm_dst;
1349 entry = &comp_v1->lcm_entries[i];
1350 entry_vic = &comp_vic->lcm_entries[j];
1351 entry_rem = &comp_rem->lcm_entries[k];
1353 if (mirror_id_of(le32_to_cpu(entry->lcme_id)) == mirror_id)
1356 /* copy component entry */
1358 memcpy(entry_vic, entry, sizeof(*entry));
1359 entry_vic->lcme_flags &= cpu_to_le32(LCME_FL_INIT);
1360 entry_vic->lcme_offset = cpu_to_le32(offset_vic);
1363 memcpy(entry_rem, entry, sizeof(*entry));
1364 entry_rem->lcme_offset = cpu_to_le32(offset_rem);
1368 lmm = (struct lov_mds_md *)((char *)comp_v1 + offset);
1370 lmm_dst = (struct lov_mds_md *)
1371 ((char *)comp_vic + offset_vic);
1373 lmm_dst = (struct lov_mds_md *)
1374 ((char *)comp_rem + offset_rem);
1376 /* copy component entry blob */
1377 memcpy(lmm_dst, lmm, le32_to_cpu(entry->lcme_size));
1379 /* blob offset advance */
1380 offset += le32_to_cpu(entry->lcme_size);
1382 offset_vic += le32_to_cpu(entry->lcme_size);
1384 offset_rem += le32_to_cpu(entry->lcme_size);
1390 static int mdd_xattr_split(const struct lu_env *env, struct md_object *md_obj,
1391 struct md_rejig_data *mrd)
1393 struct mdd_device *mdd = mdo2mdd(md_obj);
1394 struct mdd_object *obj = md2mdd_obj(md_obj);
1395 struct mdd_object *vic = md2mdd_obj(mrd->mrd_obj);
1396 struct lu_buf *buf = &mdd_env_info(env)->mti_buf[0];
1397 struct lu_buf *buf_save = &mdd_env_info(env)->mti_buf[1];
1398 struct lu_buf *buf_vic = &mdd_env_info(env)->mti_buf[2];
1399 struct lov_comp_md_v1 *lcm;
1400 struct thandle *handle;
1404 rc = lu_fid_cmp(mdo2fid(obj), mdo2fid(vic));
1405 if (rc == 0) /* same fid */
1408 handle = mdd_trans_create(env, mdd);
1410 RETURN(PTR_ERR(handle));
1413 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1414 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1416 mdd_write_lock(env, vic, MOR_TGT_CHILD);
1417 mdd_write_lock(env, obj, MOR_TGT_CHILD);
1420 /* get EA of mirrored file */
1421 memset(buf_save, 0, sizeof(*buf));
1422 rc = mdd_get_lov_ea(env, obj, buf_save);
1426 lcm = buf_save->lb_buf;
1427 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
1428 GOTO(out, rc = -EINVAL);
1431 * Extract the mirror with specified mirror id, and store the splitted
1432 * mirror layout to the victim file.
1434 memset(buf, 0, sizeof(*buf));
1435 memset(buf_vic, 0, sizeof(*buf_vic));
1436 rc = mdd_split_ea(lcm, mrd->mrd_mirror_id, buf, buf_vic);
1440 rc = mdd_declare_xattr_set(env, mdd, obj, buf, XATTR_NAME_LOV,
1441 LU_XATTR_SPLIT, handle);
1444 rc = mdd_declare_xattr_set(env, mdd, vic, buf_vic, XATTR_NAME_LOV,
1445 LU_XATTR_SPLIT, handle);
1449 rc = mdd_trans_start(env, mdd, handle);
1453 rc = mdo_xattr_set(env, obj, buf, XATTR_NAME_LOV, LU_XATTR_REPLACE,
1458 rc = mdo_xattr_set(env, vic, buf_vic, XATTR_NAME_LOV, LU_XATTR_CREATE,
1461 GOTO(out_restore, rc);
1463 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
1467 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, vic, handle);
1474 /* restore obj's layout */
1475 int rc2 = mdo_xattr_set(env, obj, buf_save, XATTR_NAME_LOV,
1476 LU_XATTR_REPLACE, handle);
1478 CERROR("%s: failed to rollback of layout of: "DFID
1479 ": %d, file state unkonwn.\n",
1480 mdd_obj_dev_name(obj), PFID(mdo2fid(obj)), rc2);
1483 mdd_trans_stop(env, mdd, rc, handle);
1484 mdd_write_unlock(env, obj);
1485 mdd_write_unlock(env, vic);
1486 lu_buf_free(buf_save);
1488 lu_buf_free(buf_vic);
1491 (void) mdd_object_pfid_replace(env, obj);
1496 static int mdd_layout_merge_allowed(const struct lu_env *env,
1497 struct md_object *target,
1498 struct md_object *victim)
1500 struct mdd_object *o1 = md2mdd_obj(target);
1502 /* cannot extend directory's LOVEA */
1503 if (S_ISDIR(mdd_object_type(o1))) {
1504 CERROR("%s: Don't extend directory's LOVEA, just set it.\n",
1505 mdd_obj_dev_name(o1));
1513 * The caller should guarantee to update the object ctime
1514 * after xattr_set if needed.
1516 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1517 const struct lu_buf *buf, const char *name,
1520 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1521 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1522 struct mdd_device *mdd = mdo2mdd(obj);
1523 struct thandle *handle;
1524 enum changelog_rec_type cl_type;
1529 rc = mdd_la_get(env, mdd_obj, attr);
1533 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1537 if (strcmp(name, XATTR_LUSTRE_LOV) == 0 &&
1538 (fl == LU_XATTR_MERGE || fl == LU_XATTR_SPLIT)) {
1539 struct md_rejig_data *mrd = buf->lb_buf;
1540 struct md_object *victim = mrd->mrd_obj;
1542 if (buf->lb_len != sizeof(*mrd))
1545 rc = mdd_layout_merge_allowed(env, obj, victim);
1549 if (fl == LU_XATTR_MERGE)
1550 /* merge layout of victim as a mirror of obj's. */
1551 rc = mdd_xattr_merge(env, obj, victim);
1553 rc = mdd_xattr_split(env, obj, mrd);
1557 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1558 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1559 struct posix_acl *acl;
1561 /* user may set empty ACL, which should be treated as removing
1563 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1566 RETURN(PTR_ERR(acl));
1568 rc = mdd_xattr_del(env, obj, name);
1571 posix_acl_release(acl);
1574 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1575 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1579 handle = mdd_trans_create(env, mdd);
1581 RETURN(PTR_ERR(handle));
1583 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1587 rc = mdd_trans_start(env, mdd, handle);
1591 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1593 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1594 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1596 mdd_write_unlock(env, mdd_obj);
1601 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1602 mdd_write_unlock(env, mdd_obj);
1606 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1610 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1615 return mdd_trans_stop(env, mdd, rc, handle);
1618 static int mdd_declare_xattr_del(const struct lu_env *env,
1619 struct mdd_device *mdd,
1620 struct mdd_object *obj,
1622 struct thandle *handle)
1626 rc = mdo_declare_xattr_del(env, obj, name, handle);
1630 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1631 return 0; /* no changelog to store */
1633 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1637 * The caller should guarantee to update the object ctime
1638 * after xattr_set if needed.
1640 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1643 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1644 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1645 struct mdd_device *mdd = mdo2mdd(obj);
1646 struct thandle *handle;
1650 rc = mdd_la_get(env, mdd_obj, attr);
1654 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1658 handle = mdd_trans_create(env, mdd);
1660 RETURN(PTR_ERR(handle));
1662 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1666 rc = mdd_trans_start(env, mdd, handle);
1670 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1671 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1672 mdd_write_unlock(env, mdd_obj);
1676 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1679 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1683 return mdd_trans_stop(env, mdd, rc, handle);
1687 * read lov EA of an object
1688 * return the lov EA in an allocated lu_buf
1690 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1691 struct lu_buf *lmm_buf)
1693 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1698 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1700 if (rc == -ERANGE) {
1701 /* mti_big_buf is allocated but is too small
1702 * we need to increase it */
1703 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1705 if (buf->lb_buf == NULL)
1706 GOTO(out, rc = -ENOMEM);
1717 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1718 /* mti_big_buf was not allocated, so we have to
1719 * allocate it based on the ea size */
1720 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1722 if (buf->lb_buf == NULL)
1723 GOTO(out, rc = -ENOMEM);
1727 lu_buf_alloc(lmm_buf, bufsize);
1728 if (lmm_buf->lb_buf == NULL)
1729 GOTO(out, rc = -ENOMEM);
1731 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1737 lu_buf_free(lmm_buf);
1741 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1742 struct mdd_object *o, struct lu_buf *buf,
1743 struct thandle *handle)
1745 struct hsm_attrs *attrs;
1751 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1756 attrs = buf->lb_buf;
1757 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1758 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1761 /* Add a changelog record for release. */
1762 hsm_set_cl_event(&flags, HE_RELEASE);
1763 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1769 * check if layout swapping between 2 objects is allowed
1771 * - only normal FIDs or non-system IGIFs
1772 * - same type of objects
1773 * - same owner/group (so quotas are still valid) unless this is from HSM
1776 static int mdd_layout_swap_allowed(const struct lu_env *env,
1777 struct mdd_object *o1,
1778 const struct lu_attr *attr1,
1779 struct mdd_object *o2,
1780 const struct lu_attr *attr2,
1783 const struct lu_fid *fid1, *fid2;
1789 if (!fid_is_norm(fid1) &&
1790 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1793 if (!fid_is_norm(fid2) &&
1794 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1797 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1798 if (S_ISDIR(mdd_object_type(o1)))
1800 if (S_ISREG(mdd_object_type(o1)))
1805 if (flags & SWAP_LAYOUTS_MDS_HSM)
1808 if ((attr1->la_uid != attr2->la_uid) ||
1809 (attr1->la_gid != attr2->la_gid))
1815 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1816 * look into the layout in MDD layer. */
1817 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1819 struct lov_comp_md_v1 *comp_v1;
1820 struct lov_mds_md *v1;
1824 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1825 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1826 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1832 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1833 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1836 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1837 off = le32_to_cpu(comp_v1->lcm_entries[i].
1839 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1844 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1845 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1856 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1858 return mdd_lmm_oi(lmm, oi, true);
1861 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1863 return mdd_lmm_oi(lmm, oi, false);
1866 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1868 struct lov_comp_md_v1 *comp_v1;
1870 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1871 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1873 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1875 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1876 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1877 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1878 __u16 tmp_gen = *gen;
1880 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1882 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1889 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1891 return mdd_lmm_gen(lmm, gen, true);
1894 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1896 return mdd_lmm_gen(lmm, gen, false);
1900 * swap layouts between 2 lustre objects
1902 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1903 struct md_object *obj2, __u64 flags)
1905 struct mdd_thread_info *info = mdd_env_info(env);
1906 struct mdd_object *fst_o = md2mdd_obj(obj1);
1907 struct mdd_object *snd_o = md2mdd_obj(obj2);
1908 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1909 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1910 struct mdd_device *mdd = mdo2mdd(obj1);
1911 struct lov_mds_md *fst_lmm, *snd_lmm;
1912 struct lu_buf *fst_buf = &info->mti_buf[0];
1913 struct lu_buf *snd_buf = &info->mti_buf[1];
1914 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1915 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1916 struct ost_id *saved_oi = NULL;
1917 struct thandle *handle;
1918 __u32 fst_gen, snd_gen, saved_gen;
1924 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1925 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1927 /* we have to sort the 2 obj, so locking will always
1928 * be in the same order, even in case of 2 concurrent swaps */
1929 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1930 if (rc == 0) /* same fid ? */
1936 rc = mdd_la_get(env, fst_o, fst_la);
1940 rc = mdd_la_get(env, snd_o, snd_la);
1944 /* check if layout swapping is allowed */
1945 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1949 handle = mdd_trans_create(env, mdd);
1951 RETURN(PTR_ERR(handle));
1953 /* objects are already sorted */
1954 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1955 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1957 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1958 if (rc < 0 && rc != -ENODATA)
1961 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1962 if (rc < 0 && rc != -ENODATA)
1965 /* swapping 2 non existant layouts is a success */
1966 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1969 /* to help inode migration between MDT, it is better to
1970 * start by the no layout file (if one), so we order the swap */
1971 if (snd_buf->lb_buf == NULL) {
1973 swap(fst_buf, snd_buf);
1976 fst_gen = snd_gen = 0;
1977 /* lmm and generation layout initialization */
1978 if (fst_buf->lb_buf != NULL) {
1979 fst_lmm = fst_buf->lb_buf;
1980 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1981 fst_fl = LU_XATTR_REPLACE;
1985 fst_fl = LU_XATTR_CREATE;
1988 snd_lmm = snd_buf->lb_buf;
1989 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1991 saved_gen = fst_gen;
1992 /* increase the generation layout numbers */
1997 * XXX The layout generation is used to generate component IDs for
1998 * the composite file, we have to do some special tweaks to make
1999 * sure the layout generation is always adequate for that job.
2002 /* Skip invalid generation number for composite layout */
2003 if ((snd_gen & LCME_ID_MASK) == 0)
2005 if ((fst_gen & LCME_ID_MASK) == 0)
2007 /* Make sure the generation is greater than all the component IDs */
2008 if (fst_gen < snd_gen)
2010 else if (fst_gen > snd_gen)
2013 /* set the file specific informations in lmm */
2014 if (fst_lmm != NULL) {
2015 saved_oi = &info->mti_oa.o_oi;
2016 mdd_get_lmm_oi(fst_lmm, saved_oi);
2017 mdd_set_lmm_gen(fst_lmm, &snd_gen);
2018 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
2019 mdd_set_lmm_oi(snd_lmm, saved_oi);
2021 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
2022 cpu_to_le32(LOV_MAGIC_MAGIC))
2023 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2025 GOTO(stop, rc = -EPROTO);
2027 mdd_set_lmm_gen(snd_lmm, &fst_gen);
2029 /* Prepare HSM attribute if it's required */
2030 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2031 const int buflen = sizeof(struct hsm_attrs);
2033 lu_buf_alloc(fst_hsm_buf, buflen);
2034 lu_buf_alloc(snd_hsm_buf, buflen);
2035 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
2036 GOTO(stop, rc = -ENOMEM);
2038 /* Read HSM attribute */
2039 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
2043 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
2047 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
2048 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2053 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
2054 XATTR_NAME_HSM, LU_XATTR_REPLACE,
2060 /* prepare transaction */
2061 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
2066 if (fst_buf->lb_buf != NULL)
2067 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
2068 XATTR_NAME_LOV, LU_XATTR_REPLACE,
2071 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
2076 rc = mdd_trans_start(env, mdd, handle);
2080 if (flags & SWAP_LAYOUTS_MDS_HSM) {
2081 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
2085 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
2087 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
2090 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
2091 mdd_obj_dev_name(fst_o),
2092 PFID(mdo2fid(fst_o)), rc, rc2);
2097 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
2101 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
2104 if (fst_buf->lb_buf != NULL)
2105 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
2106 LU_XATTR_REPLACE, handle);
2108 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
2111 GOTO(out_restore, rc);
2113 /* Issue one changelog record per file */
2114 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
2118 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
2127 /* failure on second file, but first was done, so we have
2128 * to roll back first. */
2129 if (fst_buf->lb_buf != NULL) {
2130 mdd_set_lmm_oi(fst_lmm, saved_oi);
2131 mdd_set_lmm_gen(fst_lmm, &saved_gen);
2132 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
2133 LU_XATTR_REPLACE, handle);
2135 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
2141 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
2146 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
2151 CERROR("%s: unable to roll back layout swap. FIDs: "
2152 DFID" and "DFID "error: %d/%d, steps: %d\n",
2153 mdd_obj_dev_name(fst_o),
2154 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
2156 /* a solution to avoid journal commit is to panic,
2157 * but it has strong consequences so we use LBUG to
2158 * allow sysdamin to choose to panic or not
2165 rc = mdd_trans_stop(env, mdd, rc, handle);
2167 mdd_write_unlock(env, snd_o);
2168 mdd_write_unlock(env, fst_o);
2170 lu_buf_free(fst_buf);
2171 lu_buf_free(snd_buf);
2172 lu_buf_free(fst_hsm_buf);
2173 lu_buf_free(snd_hsm_buf);
2176 (void) mdd_object_pfid_replace(env, fst_o);
2177 (void) mdd_object_pfid_replace(env, snd_o);
2182 static int mdd_declare_layout_change(const struct lu_env *env,
2183 struct mdd_device *mdd,
2184 struct mdd_object *obj,
2185 struct md_layout_change *mlc,
2186 struct thandle *handle)
2190 rc = mdo_declare_layout_change(env, obj, mlc, handle);
2194 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2197 /* For PFL, this is used to instantiate necessary component objects. */
2199 mdd_layout_instantiate_component(const struct lu_env *env,
2200 struct mdd_object *obj, struct md_layout_change *mlc,
2201 struct thandle *handle)
2203 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2207 if (mlc->mlc_opc != MD_LAYOUT_WRITE)
2210 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2212 * It's possible that another layout write intent has already
2213 * instantiated our objects, so a -EALREADY returned, and we need to
2217 RETURN(rc == -EALREADY ? 0 : rc);
2219 rc = mdd_trans_start(env, mdd, handle);
2223 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2224 rc = mdo_layout_change(env, obj, mlc, handle);
2225 mdd_write_unlock(env, obj);
2229 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, obj, handle);
2234 * Change the FLR layout from RDONLY to WRITE_PENDING.
2236 * It picks the primary mirror, and bumps the layout version, and set
2237 * layout version xattr to OST objects in a sync tx. In order to facilitate
2238 * the handling of phantom writers from evicted clients, the clients carry
2239 * layout version of the file with write RPC, so that the OSTs can verify
2240 * if the write RPCs are legitimate, meaning not from evicted clients.
2243 mdd_layout_update_rdonly(const struct lu_env *env, struct mdd_object *obj,
2244 struct md_layout_change *mlc, struct thandle *handle)
2246 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2250 /* Verify acceptable operations */
2251 switch (mlc->mlc_opc) {
2252 case MD_LAYOUT_WRITE:
2254 case MD_LAYOUT_RESYNC:
2255 /* these are legal operations - this represents the case that
2256 * a few mirrors were missed in the last resync.
2257 * XXX: it will be supported later */
2258 case MD_LAYOUT_RESYNC_DONE:
2263 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2267 rc = mdd_declare_xattr_del(env, mdd, obj, XATTR_NAME_SOM, handle);
2271 /* record a changelog for data mover to consume */
2272 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2276 rc = mdd_trans_start(env, mdd, handle);
2280 /* it needs a sync tx to make FLR to work properly */
2281 handle->th_sync = 1;
2283 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2284 rc = mdo_layout_change(env, obj, mlc, handle);
2286 rc = mdo_xattr_del(env, obj, XATTR_NAME_SOM, handle);
2290 mdd_write_unlock(env, obj);
2294 rc = mdd_changelog_data_store(env, mdd, CL_FLRW, 0, obj, handle);
2305 * Handle mirrored file state transition when it's in WRITE_PENDING.
2307 * Only MD_LAYOUT_RESYNC, which represents start of resync, is allowed when
2308 * the file is in WRITE_PENDING state. If everything goes fine, the file's
2309 * layout version will be increased, and the file's state will be changed to
2313 mdd_layout_update_write_pending(const struct lu_env *env,
2314 struct mdd_object *obj, struct md_layout_change *mlc,
2315 struct thandle *handle)
2317 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2321 switch (mlc->mlc_opc) {
2322 case MD_LAYOUT_RESYNC:
2323 /* Upon receiving the resync request, it should
2324 * instantiate all stale components right away to get ready
2325 * for mirror copy. In order to avoid layout version change,
2326 * client should avoid sending LAYOUT_WRITE request at the
2329 case MD_LAYOUT_WRITE:
2330 /* legal race for concurrent write, the file state has been
2331 * changed by another client. */
2337 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2341 rc = mdd_trans_start(env, mdd, handle);
2345 /* it needs a sync tx to make FLR to work properly */
2346 handle->th_sync = 1;
2348 mdd_write_lock(env, obj, MOR_TGT_CHILD);
2349 rc = mdo_layout_change(env, obj, mlc, handle);
2350 mdd_write_unlock(env, obj);
2361 * Handle the requests when a FLR file's state is in SYNC_PENDING.
2363 * Only concurrent write and sync complete requests are possible when the
2364 * file is in SYNC_PENDING. For the latter request, it will pass in the
2365 * mirrors that have been synchronized, then the stale bit will be cleared
2366 * to make the file's state turn into RDONLY.
2367 * For concurrent write reqeust, it just needs to change the file's state
2368 * to WRITE_PENDING in a sync tx. It doesn't have to change the layout
2369 * version because the version will be increased in the transition to
2370 * SYNC_PENDING later so that it can deny the write request from potential
2371 * evicted SYNC clients. */
2373 mdd_object_update_sync_pending(const struct lu_env *env, struct mdd_object *obj,
2374 struct md_layout_change *mlc, struct thandle *handle)
2376 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2377 struct lu_buf *som_buf = &mdd_env_info(env)->mti_buf[1];
2382 /* operation validation */
2383 switch (mlc->mlc_opc) {
2384 case MD_LAYOUT_RESYNC_DONE:
2385 /* resync complete. */
2386 case MD_LAYOUT_WRITE:
2387 /* concurrent write. */
2389 case MD_LAYOUT_RESYNC:
2390 /* resync again, most likely the previous run failed.
2391 * no-op if it's already in SYNC_PENDING state */
2397 if (mlc->mlc_som.lsa_valid & LSOM_FL_VALID) {
2398 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_SOM);
2399 if (rc && rc != -ENODATA)
2402 fl = rc == -ENODATA ? LU_XATTR_CREATE : LU_XATTR_REPLACE;
2403 som_buf->lb_buf = &mlc->mlc_som;
2404 som_buf->lb_len = sizeof(mlc->mlc_som);
2407 rc = mdd_declare_layout_change(env, mdd, obj, mlc, handle);
2411 /* record a changelog for the completion of resync */
2412 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2416 /* RESYNC_DONE has piggybacked size and blocks */
2418 rc = mdd_declare_xattr_set(env, mdd, obj, som_buf,
2419 XATTR_NAME_SOM, fl, handle);
2424 rc = mdd_trans_start(env, mdd, handle);
2428 /* it needs a sync tx to make FLR to work properly */
2429 handle->th_sync = 1;
2431 rc = mdo_layout_change(env, obj, mlc, handle);
2436 rc = mdo_xattr_set(env, obj, som_buf, XATTR_NAME_SOM,
2442 rc = mdd_changelog_data_store(env, mdd, CL_RESYNC, 0, obj, handle);
2451 * Layout change callback for object.
2453 * This is only used by FLR for now. In the future, it can be exteneded to
2454 * handle all layout change.
2457 mdd_layout_change(const struct lu_env *env, struct md_object *o,
2458 struct md_layout_change *mlc)
2460 struct mdd_object *obj = md2mdd_obj(o);
2461 struct mdd_device *mdd = mdd_obj2mdd_dev(obj);
2462 struct lu_buf *buf = mdd_buf_get(env, NULL, 0);
2463 struct lov_comp_md_v1 *lcm;
2464 struct thandle *handle;
2469 /* Verify acceptable operations */
2470 switch (mlc->mlc_opc) {
2471 case MD_LAYOUT_WRITE:
2472 case MD_LAYOUT_RESYNC:
2473 case MD_LAYOUT_RESYNC_DONE:
2479 handle = mdd_trans_create(env, mdd);
2481 RETURN(PTR_ERR(handle));
2483 rc = mdd_get_lov_ea(env, obj, buf);
2490 /* analyze the layout to make sure it's a FLR file */
2492 if (le32_to_cpu(lcm->lcm_magic) != LOV_MAGIC_COMP_V1)
2493 GOTO(out, rc = -EINVAL);
2495 flr_state = le16_to_cpu(lcm->lcm_flags) & LCM_FL_FLR_MASK;
2497 /* please refer to HLD of FLR for state transition */
2498 switch (flr_state) {
2500 rc = mdd_layout_instantiate_component(env, obj, mlc, handle);
2502 case LCM_FL_WRITE_PENDING:
2503 rc = mdd_layout_update_write_pending(env, obj, mlc, handle);
2506 rc = mdd_layout_update_rdonly(env, obj, mlc, handle);
2508 case LCM_FL_SYNC_PENDING:
2509 rc = mdd_object_update_sync_pending(env, obj, mlc, handle);
2518 mdd_trans_stop(env, mdd, rc, handle);
2523 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
2524 struct mdd_object *child, const struct lu_attr *attr,
2525 const struct md_op_spec *spec,
2526 struct dt_allocation_hint *hint)
2528 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
2529 struct dt_object *nc = mdd_object_child(child);
2531 memset(hint, 0, sizeof(*hint));
2533 /* For striped directory, give striping EA to lod_ah_init, which will
2534 * decide the stripe_offset and stripe count by it. */
2535 if (S_ISDIR(attr->la_mode) &&
2536 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
2537 hint->dah_eadata = spec->u.sp_ea.eadata;
2538 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
2540 hint->dah_eadata = NULL;
2541 hint->dah_eadata_len = 0;
2544 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
2545 hint->dah_eadata, hint->dah_eadata_len);
2546 /* @hint will be initialized by underlying device. */
2547 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
2551 * do NOT or the MAY_*'s, you'll get the weakest
2553 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
2557 /* Sadly, NFSD reopens a file repeatedly during operation, so the
2558 * "acc_mode = 0" allowance for newly-created files isn't honoured.
2559 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
2560 * owner can write to a file even if it is marked readonly to hide
2561 * its brokenness. (bug 5781) */
2562 if (flags & MDS_OPEN_OWNEROVERRIDE) {
2563 struct lu_ucred *uc = lu_ucred_check(env);
2565 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
2569 if (flags & FMODE_READ)
2571 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
2573 if (flags & MDS_FMODE_EXEC)
2578 static int mdd_open_sanity_check(const struct lu_env *env,
2579 struct mdd_object *obj,
2580 const struct lu_attr *attr, int flag)
2586 if (mdd_is_dead_obj(obj))
2589 if (S_ISLNK(attr->la_mode))
2592 mode = accmode(env, attr, flag);
2594 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
2597 if (!(flag & MDS_OPEN_CREATED)) {
2598 rc = mdd_permission_internal(env, obj, attr, mode);
2603 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
2604 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
2605 flag &= ~MDS_OPEN_TRUNC;
2607 /* For writing append-only file must open it with append mode. */
2608 if (attr->la_flags & LUSTRE_APPEND_FL) {
2609 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
2611 if (flag & MDS_OPEN_TRUNC)
2618 static int mdd_open(const struct lu_env *env, struct md_object *obj,
2621 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2622 struct md_device *md_dev = lu2md_dev(mdd2lu_dev(mdo2mdd(obj)));
2623 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
2626 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2628 rc = mdd_la_get(env, mdd_obj, attr);
2632 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
2636 mdd_obj->mod_count++;
2638 mdd_changelog(env, CL_OPEN, flags, md_dev, mdo2fid(mdd_obj));
2642 mdd_write_unlock(env, mdd_obj);
2646 static int mdd_declare_close(const struct lu_env *env,
2647 struct mdd_object *obj,
2649 struct thandle *handle)
2653 rc = orph_declare_index_delete(env, obj, handle);
2657 return mdo_declare_destroy(env, obj, handle);
2661 * No permission check is needed.
2663 static int mdd_close(const struct lu_env *env, struct md_object *obj,
2664 struct md_attr *ma, int mode)
2666 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2667 struct mdd_device *mdd = mdo2mdd(obj);
2668 struct thandle *handle = NULL;
2671 bool blocked = false;
2674 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
2675 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2676 mdd_obj->mod_count--;
2677 mdd_write_unlock(env, mdd_obj);
2679 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
2680 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
2681 "list\n", PFID(mdd_object_fid(mdd_obj)));
2685 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
2686 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
2687 /* check without any lock */
2688 is_orphan = mdd_obj->mod_count == 1 &&
2689 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
2693 /* mdd_trans_create() maybe failed because of barrier_entry(),
2694 * under such case, the orphan MDT-object will be left in the
2695 * orphan list, and when the MDT remount next time, the unused
2696 * orphans will be destroyed automatically.
2698 * One exception: the former mdd_finish_unlink may failed to
2699 * add the orphan MDT-object to the orphan list, then if the
2700 * mdd_trans_create() failed because of barrier_entry(), the
2701 * MDT-object will become real orphan that is neither in the
2702 * namespace nor in the orphan list. Such bad case should be
2703 * very rare and will be handled by e2fsck/lfsck. */
2704 handle = mdd_trans_create(env, mdo2mdd(obj));
2705 if (IS_ERR(handle)) {
2706 rc = PTR_ERR(handle);
2707 if (rc != -EINPROGRESS)
2715 rc = mdd_declare_close(env, mdd_obj, ma, handle);
2719 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2723 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2729 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2730 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
2732 CERROR("Failed to get lu_attr of "DFID": %d\n",
2733 PFID(mdd_object_fid(mdd_obj)), rc);
2737 /* check again with lock */
2738 is_orphan = (mdd_obj->mod_count == 1) &&
2739 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
2740 ma->ma_attr.la_nlink == 0);
2742 if (is_orphan && !handle && !blocked) {
2743 mdd_write_unlock(env, mdd_obj);
2747 mdd_obj->mod_count--; /*release open count */
2749 if (!is_orphan || blocked)
2753 /* NB: Object maybe not in orphan list originally, it is rare case for
2754 * mdd_finish_unlink() failure, in that case, the object doesn't have
2755 * ORPHAN_OBJ flag */
2756 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2757 /* remove link to object from orphan index */
2758 LASSERT(handle != NULL);
2759 rc = __mdd_orphan_del(env, mdd_obj, handle);
2761 CERROR("%s: unable to delete "DFID" from orphan list: "
2762 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2763 PFID(mdd_object_fid(mdd_obj)), rc);
2764 /* If object was not deleted from orphan list, do not
2765 * destroy OSS objects, which will be done when next
2770 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2771 "list, OSS objects to be destroyed.\n",
2772 PFID(mdd_object_fid(mdd_obj)));
2775 rc = mdo_destroy(env, mdd_obj, handle);
2778 CERROR("%s: unable to delete "DFID" from orphan list: "
2779 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2780 PFID(mdd_object_fid(mdd_obj)), rc);
2785 mdd_write_unlock(env, mdd_obj);
2787 /* Record CL_CLOSE in changelog only if file was opened in write mode,
2788 * or if CL_OPEN was recorded.
2789 * Changelogs mask may change between open and close operations, but
2790 * this is not a big deal if we have a CL_CLOSE entry with no matching
2791 * CL_OPEN. Plus Changelogs mask may not change often.
2793 if (!rc && !blocked &&
2794 ((mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) ||
2795 (mdd->mdd_cl.mc_mask & (1 << CL_OPEN))) &&
2796 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2797 if (handle == NULL) {
2798 handle = mdd_trans_create(env, mdo2mdd(obj));
2800 GOTO(stop, rc = PTR_ERR(handle));
2802 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2807 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2812 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2817 if (handle != NULL && !IS_ERR(handle))
2818 rc = mdd_trans_stop(env, mdd, rc, handle);
2824 * Permission check is done when open,
2825 * no need check again.
2827 static int mdd_readpage_sanity_check(const struct lu_env *env,
2828 struct mdd_object *obj)
2830 struct dt_object *next = mdd_object_child(obj);
2834 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2842 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2843 size_t nob, const struct dt_it_ops *iops,
2844 struct dt_it *it, __u32 attr, void *arg)
2846 struct lu_dirpage *dp = &lp->lp_dir;
2850 struct lu_dirent *ent;
2851 struct lu_dirent *last = NULL;
2855 if (nob < sizeof(*dp))
2858 memset(area, 0, sizeof (*dp));
2859 area += sizeof (*dp);
2860 nob -= sizeof (*dp);
2867 len = iops->key_size(env, it);
2869 /* IAM iterator can return record with zero len. */
2873 hash = iops->store(env, it);
2874 if (unlikely(first)) {
2876 dp->ldp_hash_start = cpu_to_le64(hash);
2879 /* calculate max space required for lu_dirent */
2880 recsize = lu_dirent_calc_size(len, attr);
2882 if (nob >= recsize) {
2883 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2884 if (result == -ESTALE)
2889 /* osd might not able to pack all attributes,
2890 * so recheck rec length */
2891 recsize = le16_to_cpu(ent->lde_reclen);
2893 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2894 fid_le_to_cpu(&fid, &ent->lde_fid);
2895 if (fid_is_dot_lustre(&fid))
2899 result = (last != NULL) ? 0 :-EINVAL;
2903 ent = (void *)ent + recsize;
2907 result = iops->next(env, it);
2908 if (result == -ESTALE)
2910 } while (result == 0);
2913 dp->ldp_hash_end = cpu_to_le64(hash);
2915 if (last->lde_hash == dp->ldp_hash_end)
2916 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2917 last->lde_reclen = 0; /* end mark */
2920 /* end of directory */
2921 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2922 else if (result < 0)
2923 CWARN("build page failed: %d!\n", result);
2927 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2928 const struct lu_rdpg *rdpg)
2930 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2934 if (mdd_object_exists(mdd_obj) == 0) {
2935 CERROR("%s: object "DFID" not found: rc = -2\n",
2936 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2940 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2941 rc = mdd_readpage_sanity_check(env, mdd_obj);
2943 GOTO(out_unlock, rc);
2945 if (mdd_is_dead_obj(mdd_obj)) {
2947 struct lu_dirpage *dp;
2950 * According to POSIX, please do not return any entry to client:
2951 * even dot and dotdot should not be returned.
2953 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2954 PFID(mdd_object_fid(mdd_obj)));
2956 if (rdpg->rp_count <= 0)
2957 GOTO(out_unlock, rc = -EFAULT);
2958 LASSERT(rdpg->rp_pages != NULL);
2960 pg = rdpg->rp_pages[0];
2961 dp = (struct lu_dirpage *)kmap(pg);
2962 memset(dp, 0 , sizeof(struct lu_dirpage));
2963 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2964 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2965 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2967 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2970 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2971 mdd_dir_page_build, NULL);
2973 struct lu_dirpage *dp;
2975 dp = kmap(rdpg->rp_pages[0]);
2976 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2979 * No pages were processed, mark this for first page
2982 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2983 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2984 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2986 kunmap(rdpg->rp_pages[0]);
2989 GOTO(out_unlock, rc);
2991 mdd_read_unlock(env, mdd_obj);
2995 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2997 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2999 if (mdd_object_exists(mdd_obj) == 0) {
3002 CERROR("%s: object "DFID" not found: rc = %d\n",
3003 mdd_obj_dev_name(mdd_obj),
3004 PFID(mdd_object_fid(mdd_obj)), rc);
3007 return dt_object_sync(env, mdd_object_child(mdd_obj),
3011 static int mdd_object_lock(const struct lu_env *env,
3012 struct md_object *obj,
3013 struct lustre_handle *lh,
3014 struct ldlm_enqueue_info *einfo,
3015 union ldlm_policy_data *policy)
3017 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3018 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
3022 static int mdd_object_unlock(const struct lu_env *env,
3023 struct md_object *obj,
3024 struct ldlm_enqueue_info *einfo,
3025 union ldlm_policy_data *policy)
3027 struct mdd_object *mdd_obj = md2mdd_obj(obj);
3028 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
3031 const struct md_object_operations mdd_obj_ops = {
3032 .moo_permission = mdd_permission,
3033 .moo_attr_get = mdd_attr_get,
3034 .moo_attr_set = mdd_attr_set,
3035 .moo_xattr_get = mdd_xattr_get,
3036 .moo_xattr_set = mdd_xattr_set,
3037 .moo_xattr_list = mdd_xattr_list,
3038 .moo_invalidate = mdd_invalidate,
3039 .moo_xattr_del = mdd_xattr_del,
3040 .moo_swap_layouts = mdd_swap_layouts,
3041 .moo_open = mdd_open,
3042 .moo_close = mdd_close,
3043 .moo_readpage = mdd_readpage,
3044 .moo_readlink = mdd_readlink,
3045 .moo_changelog = mdd_changelog,
3046 .moo_object_sync = mdd_object_sync,
3047 .moo_object_lock = mdd_object_lock,
3048 .moo_object_unlock = mdd_object_unlock,
3049 .moo_layout_change = mdd_layout_change,