4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/mdd/mdd_object.c
34 * Lustre Metadata Server (mdd) routines
36 * Author: Wang Di <wangdi@clusterfs.com>
39 #define DEBUG_SUBSYSTEM S_MDS
41 #include <linux/module.h>
43 #include <obd_class.h>
44 #include <obd_support.h>
45 #include <lprocfs_status.h>
46 /* fid_be_cpu(), fid_cpu_to_be(). */
47 #include <lustre_fid.h>
48 #include <lustre_idmap.h>
49 #include <uapi/linux/lustre_param.h>
50 #include <lustre_mds.h>
51 #include <lustre/lustre_idl.h>
53 #include "mdd_internal.h"
55 static const struct lu_object_operations mdd_lu_obj_ops;
57 static int mdd_xattr_get(const struct lu_env *env,
58 struct md_object *obj, struct lu_buf *buf,
61 int mdd_la_get(const struct lu_env *env, struct mdd_object *obj,
66 if (mdd_object_exists(obj) == 0)
69 rc = mdo_attr_get(env, obj, la);
70 if (unlikely(rc != 0)) {
72 obj->mod_flags |= DEAD_OBJ;
76 if (la->la_valid & LA_FLAGS &&
77 la->la_flags & LUSTRE_ORPHAN_FL)
78 obj->mod_flags |= ORPHAN_OBJ | DEAD_OBJ;
83 struct mdd_thread_info *mdd_env_info(const struct lu_env *env)
85 struct mdd_thread_info *info;
87 lu_env_refill((struct lu_env *)env);
88 info = lu_context_key_get(&env->le_ctx, &mdd_thread_key);
89 LASSERT(info != NULL);
93 struct lu_buf *mdd_buf_get(const struct lu_env *env, void *area, ssize_t len)
97 buf = &mdd_env_info(env)->mti_buf[0];
103 const struct lu_buf *mdd_buf_get_const(const struct lu_env *env,
104 const void *area, ssize_t len)
108 buf = &mdd_env_info(env)->mti_buf[0];
109 buf->lb_buf = (void *)area;
114 struct lu_object *mdd_object_alloc(const struct lu_env *env,
115 const struct lu_object_header *hdr,
118 struct mdd_object *mdd_obj;
120 OBD_SLAB_ALLOC_PTR_GFP(mdd_obj, mdd_object_kmem, GFP_NOFS);
121 if (mdd_obj != NULL) {
124 o = mdd2lu_obj(mdd_obj);
125 lu_object_init(o, NULL, d);
126 mdd_obj->mod_obj.mo_ops = &mdd_obj_ops;
127 mdd_obj->mod_obj.mo_dir_ops = &mdd_dir_ops;
128 mdd_obj->mod_count = 0;
129 o->lo_ops = &mdd_lu_obj_ops;
136 static int mdd_object_init(const struct lu_env *env, struct lu_object *o,
137 const struct lu_object_conf *unused)
139 struct mdd_device *d = lu2mdd_dev(o->lo_dev);
140 struct mdd_object *mdd_obj = lu2mdd_obj(o);
141 struct lu_object *below;
142 struct lu_device *under;
145 mdd_obj->mod_cltime = 0;
146 under = &d->mdd_child->dd_lu_dev;
147 below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
149 RETURN(PTR_ERR(below));
151 lu_object_add(o, below);
156 static int mdd_object_start(const struct lu_env *env, struct lu_object *o)
160 if (lu_object_exists(o)) {
161 struct mdd_object *mdd_obj = lu2mdd_obj(o);
162 struct lu_attr *attr = MDD_ENV_VAR(env, la_for_start);
164 rc = mdd_la_get(env, mdd_obj, attr);
170 static void mdd_object_free(const struct lu_env *env, struct lu_object *o)
172 struct mdd_object *mdd = lu2mdd_obj(o);
175 OBD_SLAB_FREE_PTR(mdd, mdd_object_kmem);
178 static int mdd_object_print(const struct lu_env *env, void *cookie,
179 lu_printer_t p, const struct lu_object *o)
181 struct mdd_object *mdd = lu2mdd_obj((struct lu_object *)o);
183 return (*p)(env, cookie,
184 LUSTRE_MDD_NAME"-object@%p(open_count=%d, valid=%x, cltime=%llu, flags=%lx)",
185 mdd, mdd->mod_count, mdd->mod_valid,
186 mdd->mod_cltime, mdd->mod_flags);
189 static const struct lu_object_operations mdd_lu_obj_ops = {
190 .loo_object_init = mdd_object_init,
191 .loo_object_start = mdd_object_start,
192 .loo_object_free = mdd_object_free,
193 .loo_object_print = mdd_object_print,
196 struct mdd_object *mdd_object_find(const struct lu_env *env,
197 struct mdd_device *d,
198 const struct lu_fid *f)
200 return md2mdd_obj(md_object_find_slice(env, &d->mdd_md_dev, f));
204 * No permission check is needed.
206 int mdd_attr_get(const struct lu_env *env, struct md_object *obj,
209 struct mdd_object *mdd_obj = md2mdd_obj(obj);
214 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
215 if ((ma->ma_need & MA_INODE) != 0 && mdd_is_dead_obj(mdd_obj))
216 ma->ma_attr.la_nlink = 0;
222 * No permission check is needed.
224 static int mdd_xattr_get(const struct lu_env *env,
225 struct md_object *obj, struct lu_buf *buf,
228 struct mdd_object *mdd_obj = md2mdd_obj(obj);
233 if (mdd_object_exists(mdd_obj) == 0) {
234 CERROR("%s: object "DFID" not found: rc = -2\n",
235 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
239 /* If the object has been destroyed, then do not get LMVEA, because
240 * it needs to load stripes from the iteration of the master object,
241 * and it will cause problem if master object has been destroyed, see
243 if (unlikely((mdd_obj->mod_flags & DEAD_OBJ) &&
244 !(mdd_obj->mod_flags & ORPHAN_OBJ) &&
245 strcmp(name, XATTR_NAME_LMV) == 0))
248 /* If the object has been delete from the namespace, then
249 * get linkEA should return -ENOENT as well */
250 if (unlikely((mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ)) &&
251 strcmp(name, XATTR_NAME_LINK) == 0))
254 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
255 rc = mdo_xattr_get(env, mdd_obj, buf, name);
256 mdd_read_unlock(env, mdd_obj);
262 * Permission check is done when open,
263 * no need check again.
265 int mdd_readlink(const struct lu_env *env, struct md_object *obj,
268 struct mdd_object *mdd_obj = md2mdd_obj(obj);
269 struct dt_object *next;
274 if (mdd_object_exists(mdd_obj) == 0) {
275 CERROR("%s: object "DFID" not found: rc = -2\n",
276 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
280 next = mdd_object_child(mdd_obj);
281 LASSERT(next != NULL);
282 LASSERT(next->do_body_ops != NULL);
283 LASSERT(next->do_body_ops->dbo_read != NULL);
284 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
285 rc = dt_read(env, next, buf, &pos);
286 mdd_read_unlock(env, mdd_obj);
291 * No permission check is needed.
293 static int mdd_xattr_list(const struct lu_env *env, struct md_object *obj,
296 struct mdd_object *mdd_obj = md2mdd_obj(obj);
301 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
302 rc = mdo_xattr_list(env, mdd_obj, buf);
303 mdd_read_unlock(env, mdd_obj);
305 /* If the buffer is NULL then we are only here to get the
306 * length of the xattr name list. */
307 if (rc < 0 || buf->lb_buf == NULL)
311 * Filter out XATTR_NAME_LINK if this is an orphan object. See
314 if (unlikely(mdd_obj->mod_flags & (DEAD_OBJ | ORPHAN_OBJ))) {
315 char *end = (char *)buf->lb_buf + rc;
316 char *p = buf->lb_buf;
319 char *next = p + strlen(p) + 1;
321 if (strcmp(p, XATTR_NAME_LINK) == 0) {
323 memmove(p, next, end - next);
325 CDEBUG(D_INFO, "Filtered out "XATTR_NAME_LINK
326 " of orphan "DFID"\n",
327 PFID(mdd_object_fid(mdd_obj)));
338 int mdd_invalidate(const struct lu_env *env, struct md_object *obj)
340 return mdo_invalidate(env, md2mdd_obj(obj));
343 int mdd_declare_create_object_internal(const struct lu_env *env,
344 struct mdd_object *p,
345 struct mdd_object *c,
346 struct lu_attr *attr,
347 struct thandle *handle,
348 const struct md_op_spec *spec,
349 struct dt_allocation_hint *hint)
351 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
352 const struct dt_index_features *feat = spec->sp_feat;
356 if (feat != &dt_directory_features && feat != NULL) {
357 dof->dof_type = DFT_INDEX;
358 dof->u.dof_idx.di_feat = feat;
360 dof->dof_type = dt_mode_to_dft(attr->la_mode);
361 if (dof->dof_type == DFT_REGULAR) {
362 dof->u.dof_reg.striped =
363 md_should_create(spec->sp_cr_flags);
364 if (spec->sp_cr_flags & MDS_OPEN_HAS_EA)
365 dof->u.dof_reg.striped = 0;
366 /* is this replay? */
368 dof->u.dof_reg.striped = 0;
372 rc = mdo_declare_create_object(env, c, attr, hint, dof, handle);
377 int mdd_create_object_internal(const struct lu_env *env, struct mdd_object *p,
378 struct mdd_object *c, struct lu_attr *attr,
379 struct thandle *handle,
380 const struct md_op_spec *spec,
381 struct dt_allocation_hint *hint)
383 struct dt_object_format *dof = &mdd_env_info(env)->mti_dof;
387 LASSERT(!mdd_object_exists(c));
389 rc = mdo_create_object(env, c, attr, hint, dof, handle);
394 int mdd_attr_set_internal(const struct lu_env *env, struct mdd_object *obj,
395 const struct lu_attr *attr, struct thandle *handle,
401 rc = mdo_attr_set(env, obj, attr, handle);
402 #ifdef CONFIG_FS_POSIX_ACL
403 if (!rc && (attr->la_valid & LA_MODE) && needacl)
404 rc = mdd_acl_chmod(env, obj, attr->la_mode, handle);
409 int mdd_update_time(const struct lu_env *env, struct mdd_object *obj,
410 const struct lu_attr *oattr, struct lu_attr *attr,
411 struct thandle *handle)
416 LASSERT(attr->la_valid & LA_CTIME);
417 LASSERT(oattr != NULL);
419 /* Make sure the ctime is increased only, however, it's not strictly
420 * reliable at here because there is not guarantee to hold lock on
421 * object, so we just bypass some unnecessary cmtime setting first
422 * and OSD has to check it again. */
423 if (attr->la_ctime < oattr->la_ctime)
424 attr->la_valid &= ~(LA_MTIME | LA_CTIME);
425 else if (attr->la_valid == LA_CTIME &&
426 attr->la_ctime == oattr->la_ctime)
427 attr->la_valid &= ~LA_CTIME;
429 if (attr->la_valid != 0)
430 rc = mdd_attr_set_internal(env, obj, attr, handle, 0);
435 * This gives the same functionality as the code between
436 * sys_chmod and inode_setattr
437 * chown_common and inode_setattr
438 * utimes and inode_setattr
439 * This API is ported from mds_fix_attr but remove some unnecesssary stuff.
441 static int mdd_fix_attr(const struct lu_env *env, struct mdd_object *obj,
442 const struct lu_attr *oattr, struct lu_attr *la,
443 const unsigned long flags)
452 /* Do not permit change file type */
453 if (la->la_valid & LA_TYPE)
456 /* They should not be processed by setattr */
457 if (la->la_valid & (LA_NLINK | LA_RDEV | LA_BLKSIZE))
460 LASSERT(oattr != NULL);
462 uc = lu_ucred_check(env);
466 if (la->la_valid == LA_CTIME) {
467 if (!(flags & MDS_PERM_BYPASS))
468 /* This is only for set ctime when rename's source is
470 rc = mdd_may_delete(env, NULL, NULL, obj, oattr, NULL,
472 if (rc == 0 && la->la_ctime <= oattr->la_ctime)
473 la->la_valid &= ~LA_CTIME;
477 if (la->la_valid == LA_ATIME) {
478 /* This is an atime-only attribute update for close RPCs. */
479 if (la->la_atime < (oattr->la_atime +
480 mdd_obj2mdd_dev(obj)->mdd_atime_diff))
481 la->la_valid &= ~LA_ATIME;
485 /* Check if flags change. */
486 if (la->la_valid & LA_FLAGS) {
487 unsigned int oldflags = oattr->la_flags &
488 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
489 unsigned int newflags = la->la_flags &
490 (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL);
492 if ((uc->uc_fsuid != oattr->la_uid) &&
493 !md_capable(uc, CFS_CAP_FOWNER))
496 /* The IMMUTABLE and APPEND_ONLY flags can
497 * only be changed by the relevant capability. */
498 if ((oldflags ^ newflags) &&
499 !md_capable(uc, CFS_CAP_LINUX_IMMUTABLE))
502 if (!S_ISDIR(oattr->la_mode))
503 la->la_flags &= ~(LUSTRE_DIRSYNC_FL | LUSTRE_TOPDIR_FL);
506 if (oattr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL) &&
507 (la->la_valid & ~LA_FLAGS) &&
508 !(flags & MDS_PERM_BYPASS))
511 /* Check for setting the obj time. */
512 if ((la->la_valid & (LA_MTIME | LA_ATIME | LA_CTIME)) &&
513 !(la->la_valid & ~(LA_MTIME | LA_ATIME | LA_CTIME))) {
514 if ((uc->uc_fsuid != oattr->la_uid) &&
515 !md_capable(uc, CFS_CAP_FOWNER)) {
516 rc = mdd_permission_internal(env, obj, oattr,
523 if (la->la_valid & LA_KILL_SUID) {
524 la->la_valid &= ~LA_KILL_SUID;
525 if ((oattr->la_mode & S_ISUID) &&
526 !(la->la_valid & LA_MODE)) {
527 la->la_mode = oattr->la_mode;
528 la->la_valid |= LA_MODE;
530 la->la_mode &= ~S_ISUID;
533 if (la->la_valid & LA_KILL_SGID) {
534 la->la_valid &= ~LA_KILL_SGID;
535 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
536 (S_ISGID | S_IXGRP)) &&
537 !(la->la_valid & LA_MODE)) {
538 la->la_mode = oattr->la_mode;
539 la->la_valid |= LA_MODE;
541 la->la_mode &= ~S_ISGID;
544 /* Make sure a caller can chmod. */
545 if (la->la_valid & LA_MODE) {
546 if (!(flags & MDS_PERM_BYPASS) &&
547 (uc->uc_fsuid != oattr->la_uid) &&
548 !md_capable(uc, CFS_CAP_FOWNER))
551 if (la->la_mode == (umode_t) -1)
552 la->la_mode = oattr->la_mode;
554 la->la_mode = (la->la_mode & S_IALLUGO) |
555 (oattr->la_mode & ~S_IALLUGO);
557 /* Also check the setgid bit! */
558 if (!lustre_in_group_p(uc, (la->la_valid & LA_GID) ?
559 la->la_gid : oattr->la_gid) &&
560 !md_capable(uc, CFS_CAP_FSETID))
561 la->la_mode &= ~S_ISGID;
563 la->la_mode = oattr->la_mode;
566 /* Make sure a caller can chown. */
567 if (la->la_valid & LA_UID) {
568 if (la->la_uid == (uid_t) -1)
569 la->la_uid = oattr->la_uid;
570 if (((uc->uc_fsuid != oattr->la_uid) ||
571 (la->la_uid != oattr->la_uid)) &&
572 !md_capable(uc, CFS_CAP_CHOWN))
575 /* If the user or group of a non-directory has been
576 * changed by a non-root user, remove the setuid bit.
577 * 19981026 David C Niemi <niemi@tux.org>
579 * Changed this to apply to all users, including root,
580 * to avoid some races. This is the behavior we had in
581 * 2.0. The check for non-root was definitely wrong
582 * for 2.2 anyway, as it should have been using
583 * CAP_FSETID rather than fsuid -- 19990830 SD. */
584 if (((oattr->la_mode & S_ISUID) == S_ISUID) &&
585 !S_ISDIR(oattr->la_mode)) {
586 la->la_mode &= ~S_ISUID;
587 la->la_valid |= LA_MODE;
591 /* Make sure caller can chgrp. */
592 if (la->la_valid & LA_GID) {
593 if (la->la_gid == (gid_t) -1)
594 la->la_gid = oattr->la_gid;
595 if (((uc->uc_fsuid != oattr->la_uid) ||
596 ((la->la_gid != oattr->la_gid) &&
597 !lustre_in_group_p(uc, la->la_gid))) &&
598 !md_capable(uc, CFS_CAP_CHOWN))
601 /* Likewise, if the user or group of a non-directory
602 * has been changed by a non-root user, remove the
603 * setgid bit UNLESS there is no group execute bit
604 * (this would be a file marked for mandatory
605 * locking). 19981026 David C Niemi <niemi@tux.org>
607 * Removed the fsuid check (see the comment above) --
609 if (((oattr->la_mode & (S_ISGID | S_IXGRP)) ==
610 (S_ISGID | S_IXGRP)) && !S_ISDIR(oattr->la_mode)) {
611 la->la_mode &= ~S_ISGID;
612 la->la_valid |= LA_MODE;
616 if (la->la_valid & (LA_SIZE | LA_BLOCKS)) {
617 if (!((flags & MDS_OWNEROVERRIDE) &&
618 (uc->uc_fsuid == oattr->la_uid)) &&
619 !(flags & MDS_PERM_BYPASS)) {
620 rc = mdd_permission_internal(env, obj, oattr,
627 if (la->la_valid & LA_CTIME) {
628 /* The pure setattr, it has the priority over what is
629 * already set, do not drop it if ctime is equal. */
630 if (la->la_ctime < oattr->la_ctime)
631 la->la_valid &= ~(LA_ATIME | LA_MTIME | LA_CTIME);
637 static int mdd_changelog_data_store_by_fid(const struct lu_env *env,
638 struct mdd_device *mdd,
639 enum changelog_rec_type type, int flags,
640 const struct lu_fid *fid,
641 struct thandle *handle)
643 const struct lu_ucred *uc = lu_ucred(env);
644 struct llog_changelog_rec *rec;
649 flags = (flags & CLF_FLAGMASK) | CLF_VERSION;
650 if (uc != NULL && uc->uc_jobid[0] != '\0')
653 reclen = llog_data_len(LLOG_CHANGELOG_HDR_SZ +
654 changelog_rec_offset(flags & CLF_SUPPORTED));
655 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf, reclen);
656 if (buf->lb_buf == NULL)
660 rec->cr_hdr.lrh_len = reclen;
661 rec->cr.cr_flags = flags;
662 rec->cr.cr_type = (__u32)type;
663 rec->cr.cr_tfid = *fid;
664 rec->cr.cr_namelen = 0;
666 if (flags & CLF_JOBID)
667 mdd_changelog_rec_ext_jobid(&rec->cr, uc->uc_jobid);
669 rc = mdd_changelog_store(env, mdd, rec, handle);
675 /** Store a data change changelog record
676 * If this fails, we must fail the whole transaction; we don't
677 * want the change to commit without the log entry.
678 * \param mdd_obj - mdd_object of change
679 * \param handle - transaction handle
681 int mdd_changelog_data_store(const struct lu_env *env, struct mdd_device *mdd,
682 enum changelog_rec_type type, int flags,
683 struct mdd_object *mdd_obj, struct thandle *handle)
687 LASSERT(mdd_obj != NULL);
688 LASSERT(handle != NULL);
691 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
693 if ((mdd->mdd_cl.mc_mask & (1 << type)) == 0)
696 if (mdd_is_volatile_obj(mdd_obj))
699 if ((type >= CL_MTIME) && (type <= CL_ATIME) &&
700 cfs_time_before_64(mdd->mdd_cl.mc_starttime, mdd_obj->mod_cltime)) {
701 /* Don't need multiple updates in this log */
702 /* Don't check under lock - no big deal if we get an extra
707 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
708 mdo2fid(mdd_obj), handle);
710 mdd_obj->mod_cltime = cfs_time_current_64();
715 static int mdd_changelog(const struct lu_env *env, enum changelog_rec_type type,
716 int flags, struct md_device *m, const struct lu_fid *fid)
718 struct thandle *handle;
719 struct mdd_device *mdd = lu2mdd_dev(&m->md_lu_dev);
724 if (!(mdd->mdd_cl.mc_flags & CLM_ON))
726 if (!(mdd->mdd_cl.mc_mask & (1 << type)))
729 LASSERT(fid != NULL);
731 handle = mdd_trans_create(env, mdd);
733 RETURN(PTR_ERR(handle));
735 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
739 rc = mdd_trans_start(env, mdd, handle);
743 rc = mdd_changelog_data_store_by_fid(env, mdd, type, flags,
747 rc = mdd_trans_stop(env, mdd, rc, handle);
753 * Save LMA extended attributes with data from \a ma.
755 * HSM and Size-On-MDS data will be extracted from \ma if they are valid, if
756 * not, LMA EA will be first read from disk, modified and write back.
759 /* Precedence for choosing record type when multiple
760 * attributes change: setattr > mtime > ctime > atime
761 * (ctime changes when mtime does, plus chmod/chown.
762 * atime and ctime are independent.) */
763 static int mdd_attr_set_changelog(const struct lu_env *env,
764 struct md_object *obj, struct thandle *handle,
767 struct mdd_device *mdd = mdo2mdd(obj);
770 bits = (valid & LA_SIZE) ? 1 << CL_TRUNC : 0;
771 bits |= (valid & ~(LA_CTIME|LA_MTIME|LA_ATIME)) ? 1 << CL_SETATTR : 0;
772 bits |= (valid & LA_MTIME) ? 1 << CL_MTIME : 0;
773 bits |= (valid & LA_CTIME) ? 1 << CL_CTIME : 0;
774 bits |= (valid & LA_ATIME) ? 1 << CL_ATIME : 0;
775 bits = bits & mdd->mdd_cl.mc_mask;
776 /* This is an implementation limit rather than a protocol limit */
777 CLASSERT(CL_LAST <= sizeof(int) * 8);
781 /* The record type is the lowest non-masked set bit */
784 /* FYI we only store the first CLF_FLAGMASK bits of la_valid */
785 return mdd_changelog_data_store(env, mdd, type, (int)valid,
786 md2mdd_obj(obj), handle);
789 static int mdd_declare_attr_set(const struct lu_env *env,
790 struct mdd_device *mdd,
791 struct mdd_object *obj,
792 const struct lu_attr *attr,
793 struct thandle *handle)
797 rc = mdo_declare_attr_set(env, obj, attr, handle);
801 #ifdef CONFIG_FS_POSIX_ACL
802 if (attr->la_valid & LA_MODE) {
803 mdd_read_lock(env, obj, MOR_TGT_CHILD);
804 rc = mdo_xattr_get(env, obj, &LU_BUF_NULL,
805 XATTR_NAME_ACL_ACCESS);
806 mdd_read_unlock(env, obj);
807 if (rc == -EOPNOTSUPP || rc == -ENODATA)
813 struct lu_buf *buf = mdd_buf_get(env, NULL, rc);
814 rc = mdo_declare_xattr_set(env, obj, buf,
815 XATTR_NAME_ACL_ACCESS, 0,
823 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
830 * permission changes may require sync operation, to mitigate performance
831 * impact, only do this for dir and when permission is reduced.
833 * For regular files, version is updated with permission change (see VBR), async
834 * permission won't cause any issue, while missing permission change on
835 * directory may affect accessibility of other objects after recovery.
837 static inline bool permission_needs_sync(const struct lu_attr *old,
838 const struct lu_attr *new)
840 if (!S_ISDIR(old->la_mode))
843 if (new->la_valid & (LA_UID | LA_GID))
846 if (new->la_valid & LA_MODE &&
847 new->la_mode & (S_ISUID | S_ISGID | S_ISVTX))
850 if ((new->la_valid & LA_MODE) &&
851 ((new->la_mode & old->la_mode) & S_IRWXUGO) !=
852 (old->la_mode & S_IRWXUGO))
858 /* set attr and LOV EA at once, return updated attr */
859 int mdd_attr_set(const struct lu_env *env, struct md_object *obj,
860 const struct md_attr *ma)
862 struct mdd_object *mdd_obj = md2mdd_obj(obj);
863 struct mdd_device *mdd = mdo2mdd(obj);
864 struct thandle *handle = NULL;
865 struct lu_attr *la_copy = &mdd_env_info(env)->mti_la_for_fix;
866 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
867 const struct lu_attr *la = &ma->ma_attr;
872 /* we do not use ->attr_set() for LOV/HSM EA any more */
873 LASSERT((ma->ma_valid & MA_LOV) == 0);
874 LASSERT((ma->ma_valid & MA_HSM) == 0);
876 rc = mdd_la_get(env, mdd_obj, attr);
880 *la_copy = ma->ma_attr;
881 rc = mdd_fix_attr(env, mdd_obj, attr, la_copy, ma->ma_attr_flags);
885 /* no need to setattr anymore */
886 if (la_copy->la_valid == 0) {
887 CDEBUG(D_INODE, "%s: no valid attribute on "DFID", previous"
888 "valid is %#llx\n", mdd2obd_dev(mdd)->obd_name,
889 PFID(mdo2fid(mdd_obj)), la->la_valid);
894 /* If an unprivileged user changes group of some file,
895 * the setattr operation will be processed synchronously to
896 * honor the quota limit of the corresponding group. see LU-5152 */
897 uc = lu_ucred_check(env);
898 if (S_ISREG(attr->la_mode) && la->la_valid & LA_GID &&
899 la->la_gid != attr->la_gid && uc != NULL && uc->uc_fsuid != 0) {
900 la_copy->la_valid |= LA_FLAGS;
901 la_copy->la_flags |= LUSTRE_SET_SYNC_FL;
903 /* Flush the possible existing sync requests to OSTs to
904 * keep the order of sync for the current setattr operation
905 * will be sent directly to OSTs. see LU-5152 */
906 rc = dt_sync(env, mdd->mdd_child);
911 handle = mdd_trans_create(env, mdd);
912 if (IS_ERR(handle)) {
913 rc = PTR_ERR(handle);
919 rc = mdd_declare_attr_set(env, mdd, mdd_obj, la_copy, handle);
923 rc = mdd_trans_start(env, mdd, handle);
927 if (mdd->mdd_sync_permission && permission_needs_sync(attr, la))
930 if (la->la_valid & (LA_MTIME | LA_CTIME))
931 CDEBUG(D_INODE, "setting mtime %llu, ctime %llu\n",
932 la->la_mtime, la->la_ctime);
934 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
935 if (la_copy->la_valid) {
936 rc = mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
938 if (rc == -EDQUOT && la_copy->la_flags & LUSTRE_SET_SYNC_FL) {
939 /* rollback to the original gid */
940 la_copy->la_flags &= ~LUSTRE_SET_SYNC_FL;
941 la_copy->la_gid = attr->la_gid;
942 mdd_attr_set_internal(env, mdd_obj, la_copy, handle, 1);
945 mdd_write_unlock(env, mdd_obj);
949 rc = mdd_attr_set_changelog(env, obj, handle,
953 rc = mdd_trans_stop(env, mdd, rc, handle);
958 static int mdd_xattr_sanity_check(const struct lu_env *env,
959 struct mdd_object *obj,
960 const struct lu_attr *attr,
963 struct lu_ucred *uc = lu_ucred_assert(env);
966 if (attr->la_flags & (LUSTRE_IMMUTABLE_FL | LUSTRE_APPEND_FL))
969 if (strncmp(XATTR_USER_PREFIX, name,
970 sizeof(XATTR_USER_PREFIX) - 1) == 0) {
971 /* For sticky directories, only the owner and privileged user
972 * can write attributes. */
973 if (S_ISDIR(attr->la_mode) && (attr->la_mode & S_ISVTX) &&
974 (uc->uc_fsuid != attr->la_uid) &&
975 !md_capable(uc, CFS_CAP_FOWNER))
978 if ((uc->uc_fsuid != attr->la_uid) &&
979 !md_capable(uc, CFS_CAP_FOWNER))
987 * Check if a string begins with a given prefix.
989 * \param str String to check
990 * \param prefix Substring to check at the beginning of \a str
991 * \return true/false whether the condition is verified.
993 static inline bool has_prefix(const char *str, const char *prefix)
995 return strncmp(prefix, str, strlen(prefix)) == 0;
999 * Indicate the kind of changelog to store (if any) for a xattr set/del.
1001 * \param[in] xattr_name Full extended attribute name.
1003 * \return The type of changelog to use, or -1 if no changelog is to be emitted.
1005 static enum changelog_rec_type
1006 mdd_xattr_changelog_type(const struct lu_env *env, struct mdd_device *mdd,
1007 const char *xattr_name)
1009 /* Layout changes systematically recorded */
1010 if (strcmp(XATTR_NAME_LOV, xattr_name) == 0 ||
1011 strncmp(XATTR_LUSTRE_LOV, xattr_name,
1012 strlen(XATTR_LUSTRE_LOV)) == 0)
1015 /* HSM information changes systematically recorded */
1016 if (strcmp(XATTR_NAME_HSM, xattr_name) == 0)
1019 if (has_prefix(xattr_name, XATTR_USER_PREFIX) ||
1020 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_ACCESS) ||
1021 has_prefix(xattr_name, XATTR_NAME_POSIX_ACL_DEFAULT) ||
1022 has_prefix(xattr_name, XATTR_TRUSTED_PREFIX) ||
1023 has_prefix(xattr_name, XATTR_SECURITY_PREFIX))
1029 static int mdd_declare_xattr_set(const struct lu_env *env,
1030 struct mdd_device *mdd,
1031 struct mdd_object *obj,
1032 const struct lu_buf *buf,
1034 int fl, struct thandle *handle)
1038 rc = mdo_declare_xattr_set(env, obj, buf, name, fl, handle);
1042 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1043 return 0; /* no changelog to store */
1045 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1049 * Compare current and future data of HSM EA and add a changelog if needed.
1051 * Caller should have write-locked \param obj.
1053 * \param buf - Future HSM EA content.
1054 * \retval 0 if no changelog is needed or changelog was added properly.
1055 * \retval -ve errno if there was a problem
1057 static int mdd_hsm_update_locked(const struct lu_env *env,
1058 struct md_object *obj,
1059 const struct lu_buf *buf,
1060 struct thandle *handle, int *cl_flags)
1062 struct mdd_thread_info *info = mdd_env_info(env);
1063 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1064 struct lu_buf *current_buf;
1065 struct md_hsm *current_mh;
1066 struct md_hsm *new_mh;
1070 OBD_ALLOC_PTR(current_mh);
1071 if (current_mh == NULL)
1074 /* Read HSM attrs from disk */
1075 current_buf = lu_buf_check_and_alloc(&info->mti_xattr_buf,
1076 mdo2mdd(obj)->mdd_dt_conf.ddp_max_ea_size);
1077 rc = mdo_xattr_get(env, mdd_obj, current_buf, XATTR_NAME_HSM);
1078 rc = lustre_buf2hsm(current_buf->lb_buf, rc, current_mh);
1079 if (rc < 0 && rc != -ENODATA)
1081 else if (rc == -ENODATA)
1082 current_mh->mh_flags = 0;
1084 /* Map future HSM xattr */
1085 OBD_ALLOC_PTR(new_mh);
1087 GOTO(free, rc = -ENOMEM);
1088 lustre_buf2hsm(buf->lb_buf, buf->lb_len, new_mh);
1092 /* Flags differ, set flags for the changelog that will be added */
1093 if (current_mh->mh_flags != new_mh->mh_flags) {
1094 hsm_set_cl_event(cl_flags, HE_STATE);
1095 if (new_mh->mh_flags & HS_DIRTY)
1096 hsm_set_cl_flags(cl_flags, CLF_HSM_DIRTY);
1099 OBD_FREE_PTR(new_mh);
1102 OBD_FREE_PTR(current_mh);
1106 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1110 * The caller should guarantee to update the object ctime
1111 * after xattr_set if needed.
1113 static int mdd_xattr_set(const struct lu_env *env, struct md_object *obj,
1114 const struct lu_buf *buf, const char *name,
1117 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1118 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1119 struct mdd_device *mdd = mdo2mdd(obj);
1120 struct thandle *handle;
1121 enum changelog_rec_type cl_type;
1126 rc = mdd_la_get(env, mdd_obj, attr);
1130 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1134 if (strcmp(name, XATTR_NAME_ACL_ACCESS) == 0 ||
1135 strcmp(name, XATTR_NAME_ACL_DEFAULT) == 0) {
1136 struct posix_acl *acl;
1138 /* user may set empty ACL, which should be treated as removing
1140 acl = posix_acl_from_xattr(&init_user_ns, buf->lb_buf,
1143 RETURN(PTR_ERR(acl));
1145 rc = mdd_xattr_del(env, obj, name);
1148 posix_acl_release(acl);
1151 if (!strcmp(name, XATTR_NAME_ACL_ACCESS)) {
1152 rc = mdd_acl_set(env, mdd_obj, attr, buf, fl);
1156 handle = mdd_trans_create(env, mdd);
1158 RETURN(PTR_ERR(handle));
1160 rc = mdd_declare_xattr_set(env, mdd, mdd_obj, buf, name, fl, handle);
1164 rc = mdd_trans_start(env, mdd, handle);
1168 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1170 if (strcmp(XATTR_NAME_HSM, name) == 0) {
1171 rc = mdd_hsm_update_locked(env, obj, buf, handle, &cl_flags);
1173 mdd_write_unlock(env, mdd_obj);
1178 rc = mdo_xattr_set(env, mdd_obj, buf, name, fl, handle);
1179 mdd_write_unlock(env, mdd_obj);
1183 cl_type = mdd_xattr_changelog_type(env, mdd, name);
1187 rc = mdd_changelog_data_store(env, mdd, cl_type, cl_flags, mdd_obj,
1192 return mdd_trans_stop(env, mdd, rc, handle);
1195 static int mdd_declare_xattr_del(const struct lu_env *env,
1196 struct mdd_device *mdd,
1197 struct mdd_object *obj,
1199 struct thandle *handle)
1203 rc = mdo_declare_xattr_del(env, obj, name, handle);
1207 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1208 return 0; /* no changelog to store */
1210 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1214 * The caller should guarantee to update the object ctime
1215 * after xattr_set if needed.
1217 static int mdd_xattr_del(const struct lu_env *env, struct md_object *obj,
1220 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1221 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1222 struct mdd_device *mdd = mdo2mdd(obj);
1223 struct thandle *handle;
1227 rc = mdd_la_get(env, mdd_obj, attr);
1231 rc = mdd_xattr_sanity_check(env, mdd_obj, attr, name);
1235 handle = mdd_trans_create(env, mdd);
1237 RETURN(PTR_ERR(handle));
1239 rc = mdd_declare_xattr_del(env, mdd, mdd_obj, name, handle);
1243 rc = mdd_trans_start(env, mdd, handle);
1247 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1248 rc = mdo_xattr_del(env, mdd_obj, name, handle);
1249 mdd_write_unlock(env, mdd_obj);
1253 if (mdd_xattr_changelog_type(env, mdd, name) < 0)
1256 rc = mdd_changelog_data_store(env, mdd, CL_XATTR, 0, mdd_obj, handle);
1260 return mdd_trans_stop(env, mdd, rc, handle);
1264 * read lov EA of an object
1265 * return the lov EA in an allocated lu_buf
1267 int mdd_get_lov_ea(const struct lu_env *env, struct mdd_object *obj,
1268 struct lu_buf *lmm_buf)
1270 struct lu_buf *buf = &mdd_env_info(env)->mti_big_buf;
1275 rc = mdo_xattr_get(env, obj, buf, XATTR_NAME_LOV);
1277 if (rc == -ERANGE) {
1278 /* mti_big_buf is allocated but is too small
1279 * we need to increase it */
1280 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1282 if (buf->lb_buf == NULL)
1283 GOTO(out, rc = -ENOMEM);
1294 if (memcmp(buf, &LU_BUF_NULL, sizeof(*buf)) == 0) {
1295 /* mti_big_buf was not allocated, so we have to
1296 * allocate it based on the ea size */
1297 buf = lu_buf_check_and_alloc(&mdd_env_info(env)->mti_big_buf,
1299 if (buf->lb_buf == NULL)
1300 GOTO(out, rc = -ENOMEM);
1304 lu_buf_alloc(lmm_buf, bufsize);
1305 if (lmm_buf->lb_buf == NULL)
1306 GOTO(out, rc = -ENOMEM);
1308 memcpy(lmm_buf->lb_buf, buf->lb_buf, bufsize);
1314 lu_buf_free(lmm_buf);
1318 static int mdd_xattr_hsm_replace(const struct lu_env *env,
1319 struct mdd_object *o, struct lu_buf *buf,
1320 struct thandle *handle)
1322 struct hsm_attrs *attrs;
1328 rc = mdo_xattr_set(env, o, buf, XATTR_NAME_HSM, LU_XATTR_REPLACE,
1333 attrs = buf->lb_buf;
1334 hsm_flags = le32_to_cpu(attrs->hsm_flags);
1335 if (!(hsm_flags & HS_RELEASED) || mdd_is_dead_obj(o))
1338 /* Add a changelog record for release. */
1339 hsm_set_cl_event(&flags, HE_RELEASE);
1340 rc = mdd_changelog_data_store(env, mdo2mdd(&o->mod_obj), CL_HSM,
1346 * check if layout swapping between 2 objects is allowed
1348 * - only normal FIDs or non-system IGIFs
1349 * - same type of objects
1350 * - same owner/group (so quotas are still valid) unless this is from HSM
1353 static int mdd_layout_swap_allowed(const struct lu_env *env,
1354 struct mdd_object *o1,
1355 const struct lu_attr *attr1,
1356 struct mdd_object *o2,
1357 const struct lu_attr *attr2,
1360 const struct lu_fid *fid1, *fid2;
1366 if (!fid_is_norm(fid1) &&
1367 (!fid_is_igif(fid1) || IS_ERR(mdd_links_get(env, o1))))
1370 if (!fid_is_norm(fid2) &&
1371 (!fid_is_igif(fid2) || IS_ERR(mdd_links_get(env, o2))))
1374 if (mdd_object_type(o1) != mdd_object_type(o2)) {
1375 if (S_ISDIR(mdd_object_type(o1)))
1377 if (S_ISREG(mdd_object_type(o1)))
1382 if (flags & SWAP_LAYOUTS_MDS_HSM)
1385 if ((attr1->la_uid != attr2->la_uid) ||
1386 (attr1->la_gid != attr2->la_gid))
1392 /* XXX To set the proper lmm_oi & lmm_layout_gen when swap layouts, we have to
1393 * look into the layout in MDD layer. */
1394 static int mdd_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi, bool get)
1396 struct lov_comp_md_v1 *comp_v1;
1397 struct lov_mds_md *v1;
1401 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1402 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1403 ent_count = le16_to_cpu(comp_v1->lcm_entry_count);
1409 off = le32_to_cpu(comp_v1->lcm_entries[0].lcme_offset);
1410 v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1413 for (i = 0; i < le32_to_cpu(ent_count); i++) {
1414 off = le32_to_cpu(comp_v1->lcm_entries[i].
1416 v1 = (struct lov_mds_md *)((char *)comp_v1 +
1421 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1422 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1433 static inline int mdd_get_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1435 return mdd_lmm_oi(lmm, oi, true);
1438 static inline int mdd_set_lmm_oi(struct lov_mds_md *lmm, struct ost_id *oi)
1440 return mdd_lmm_oi(lmm, oi, false);
1443 static int mdd_lmm_gen(struct lov_mds_md *lmm, __u32 *gen, bool get)
1445 struct lov_comp_md_v1 *comp_v1;
1447 if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
1448 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1450 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1452 comp_v1->lcm_layout_gen = cpu_to_le32(*gen);
1453 } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1454 le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1455 __u16 tmp_gen = *gen;
1457 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1459 lmm->lmm_layout_gen = cpu_to_le16(tmp_gen);
1466 static inline int mdd_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1468 return mdd_lmm_gen(lmm, gen, true);
1471 static inline int mdd_set_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1473 return mdd_lmm_gen(lmm, gen, false);
1477 * swap layouts between 2 lustre objects
1479 static int mdd_swap_layouts(const struct lu_env *env, struct md_object *obj1,
1480 struct md_object *obj2, __u64 flags)
1482 struct mdd_thread_info *info = mdd_env_info(env);
1483 struct mdd_object *fst_o = md2mdd_obj(obj1);
1484 struct mdd_object *snd_o = md2mdd_obj(obj2);
1485 struct lu_attr *fst_la = MDD_ENV_VAR(env, cattr);
1486 struct lu_attr *snd_la = MDD_ENV_VAR(env, tattr);
1487 struct mdd_device *mdd = mdo2mdd(obj1);
1488 struct lov_mds_md *fst_lmm, *snd_lmm;
1489 struct lu_buf *fst_buf = &info->mti_buf[0];
1490 struct lu_buf *snd_buf = &info->mti_buf[1];
1491 struct lu_buf *fst_hsm_buf = &info->mti_buf[2];
1492 struct lu_buf *snd_hsm_buf = &info->mti_buf[3];
1493 struct ost_id *saved_oi = NULL;
1494 struct thandle *handle;
1495 __u32 fst_gen, snd_gen, saved_gen;
1501 CLASSERT(ARRAY_SIZE(info->mti_buf) >= 4);
1502 memset(info->mti_buf, 0, sizeof(info->mti_buf));
1504 /* we have to sort the 2 obj, so locking will always
1505 * be in the same order, even in case of 2 concurrent swaps */
1506 rc = lu_fid_cmp(mdo2fid(fst_o), mdo2fid(snd_o));
1507 if (rc == 0) /* same fid ? */
1513 rc = mdd_la_get(env, fst_o, fst_la);
1517 rc = mdd_la_get(env, snd_o, snd_la);
1521 /* check if layout swapping is allowed */
1522 rc = mdd_layout_swap_allowed(env, fst_o, fst_la, snd_o, snd_la, flags);
1526 handle = mdd_trans_create(env, mdd);
1528 RETURN(PTR_ERR(handle));
1530 /* objects are already sorted */
1531 mdd_write_lock(env, fst_o, MOR_TGT_CHILD);
1532 mdd_write_lock(env, snd_o, MOR_TGT_CHILD);
1534 rc = mdd_get_lov_ea(env, fst_o, fst_buf);
1535 if (rc < 0 && rc != -ENODATA)
1538 rc = mdd_get_lov_ea(env, snd_o, snd_buf);
1539 if (rc < 0 && rc != -ENODATA)
1542 /* swapping 2 non existant layouts is a success */
1543 if (fst_buf->lb_buf == NULL && snd_buf->lb_buf == NULL)
1546 /* to help inode migration between MDT, it is better to
1547 * start by the no layout file (if one), so we order the swap */
1548 if (snd_buf->lb_buf == NULL) {
1550 swap(fst_buf, snd_buf);
1553 fst_gen = snd_gen = 0;
1554 /* lmm and generation layout initialization */
1555 if (fst_buf->lb_buf != NULL) {
1556 fst_lmm = fst_buf->lb_buf;
1557 mdd_get_lmm_gen(fst_lmm, &fst_gen);
1558 fst_fl = LU_XATTR_REPLACE;
1562 fst_fl = LU_XATTR_CREATE;
1565 snd_lmm = snd_buf->lb_buf;
1566 mdd_get_lmm_gen(snd_lmm, &snd_gen);
1568 saved_gen = fst_gen;
1569 /* increase the generation layout numbers */
1574 * XXX The layout generation is used to generate component IDs for
1575 * the composite file, we have to do some special tweaks to make
1576 * sure the layout generation is always adequate for that job.
1579 /* Skip invalid generation number for composite layout */
1580 if ((snd_gen & LCME_ID_MASK) == 0)
1582 if ((fst_gen & LCME_ID_MASK) == 0)
1584 /* Make sure the generation is greater than all the component IDs */
1585 if (fst_gen < snd_gen)
1587 else if (fst_gen > snd_gen)
1590 /* set the file specific informations in lmm */
1591 if (fst_lmm != NULL) {
1592 saved_oi = &info->mti_oa.o_oi;
1593 mdd_get_lmm_oi(fst_lmm, saved_oi);
1594 mdd_set_lmm_gen(fst_lmm, &snd_gen);
1595 mdd_set_lmm_oi(fst_lmm, &snd_lmm->lmm_oi);
1596 mdd_set_lmm_oi(snd_lmm, saved_oi);
1598 if ((snd_lmm->lmm_magic & cpu_to_le32(LOV_MAGIC_MASK)) ==
1599 cpu_to_le32(LOV_MAGIC_MAGIC))
1600 snd_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEF);
1602 GOTO(stop, rc = -EPROTO);
1604 mdd_set_lmm_gen(snd_lmm, &fst_gen);
1606 /* Prepare HSM attribute if it's required */
1607 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1608 const int buflen = sizeof(struct hsm_attrs);
1610 lu_buf_alloc(fst_hsm_buf, buflen);
1611 lu_buf_alloc(snd_hsm_buf, buflen);
1612 if (fst_hsm_buf->lb_buf == NULL || snd_hsm_buf->lb_buf == NULL)
1613 GOTO(stop, rc = -ENOMEM);
1615 /* Read HSM attribute */
1616 rc = mdo_xattr_get(env, fst_o, fst_hsm_buf, XATTR_NAME_HSM);
1620 rc = mdo_xattr_get(env, snd_o, snd_hsm_buf, XATTR_NAME_HSM);
1624 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_hsm_buf,
1625 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1630 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_hsm_buf,
1631 XATTR_NAME_HSM, LU_XATTR_REPLACE,
1637 /* prepare transaction */
1638 rc = mdd_declare_xattr_set(env, mdd, fst_o, snd_buf, XATTR_NAME_LOV,
1643 if (fst_buf->lb_buf != NULL)
1644 rc = mdd_declare_xattr_set(env, mdd, snd_o, fst_buf,
1645 XATTR_NAME_LOV, LU_XATTR_REPLACE,
1648 rc = mdd_declare_xattr_del(env, mdd, snd_o, XATTR_NAME_LOV,
1653 rc = mdd_trans_start(env, mdd, handle);
1657 if (flags & SWAP_LAYOUTS_MDS_HSM) {
1658 rc = mdd_xattr_hsm_replace(env, fst_o, snd_hsm_buf, handle);
1662 rc = mdd_xattr_hsm_replace(env, snd_o, fst_hsm_buf, handle);
1664 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf,
1667 CERROR("%s: restore "DFID" HSM error: %d/%d\n",
1668 mdd_obj_dev_name(fst_o),
1669 PFID(mdo2fid(fst_o)), rc, rc2);
1674 rc = mdo_xattr_set(env, fst_o, snd_buf, XATTR_NAME_LOV, fst_fl, handle);
1678 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_MDS_HSM_SWAP_LAYOUTS))) {
1681 if (fst_buf->lb_buf != NULL)
1682 rc = mdo_xattr_set(env, snd_o, fst_buf, XATTR_NAME_LOV,
1683 LU_XATTR_REPLACE, handle);
1685 rc = mdo_xattr_del(env, snd_o, XATTR_NAME_LOV, handle);
1691 /* failure on second file, but first was done, so we have
1692 * to roll back first. */
1693 if (fst_buf->lb_buf != NULL) {
1694 mdd_set_lmm_oi(fst_lmm, saved_oi);
1695 mdd_set_lmm_gen(fst_lmm, &saved_gen);
1696 rc2 = mdo_xattr_set(env, fst_o, fst_buf, XATTR_NAME_LOV,
1697 LU_XATTR_REPLACE, handle);
1699 rc2 = mdo_xattr_del(env, fst_o, XATTR_NAME_LOV, handle);
1705 rc2 = mdd_xattr_hsm_replace(env, fst_o, fst_hsm_buf, handle);
1710 rc2 = mdd_xattr_hsm_replace(env, snd_o, snd_hsm_buf, handle);
1715 CERROR("%s: unable to roll back layout swap. FIDs: "
1716 DFID" and "DFID "error: %d/%d, steps: %d\n",
1717 mdd_obj_dev_name(fst_o),
1718 PFID(mdo2fid(snd_o)), PFID(mdo2fid(fst_o)),
1720 /* a solution to avoid journal commit is to panic,
1721 * but it has strong consequences so we use LBUG to
1722 * allow sysdamin to choose to panic or not
1729 /* Issue one changelog record per file */
1730 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, fst_o, handle);
1734 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, snd_o, handle);
1740 rc = mdd_trans_stop(env, mdd, rc, handle);
1742 mdd_write_unlock(env, snd_o);
1743 mdd_write_unlock(env, fst_o);
1745 lu_buf_free(fst_buf);
1746 lu_buf_free(snd_buf);
1747 lu_buf_free(fst_hsm_buf);
1748 lu_buf_free(snd_hsm_buf);
1752 static int mdd_declare_layout_change(const struct lu_env *env,
1753 struct mdd_device *mdd,
1754 struct mdd_object *obj,
1755 struct layout_intent *layout,
1756 const struct lu_buf *buf,
1757 struct thandle *handle)
1761 rc = mdo_declare_layout_change(env, obj, layout, buf, handle);
1765 return mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
1768 /* For PFL, this is used to instantiate necessary component objects. */
1769 int mdd_layout_change(const struct lu_env *env, struct md_object *obj,
1770 struct layout_intent *layout, const struct lu_buf *buf)
1772 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1773 struct mdd_device *mdd = mdo2mdd(obj);
1774 struct thandle *handle;
1778 handle = mdd_trans_create(env, mdd);
1780 RETURN(PTR_ERR(handle));
1782 rc = mdd_declare_layout_change(env, mdd, mdd_obj, layout, buf, handle);
1784 * It's possible that another layout write intent has already
1785 * instantiated our objects, so a -EALREADY returned, and we need to
1789 GOTO(stop, rc = (rc == -EALREADY) ? 0 : rc);
1791 rc = mdd_trans_start(env, mdd, handle);
1795 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1796 rc = mdo_layout_change(env, mdd_obj, layout, buf, handle);
1797 mdd_write_unlock(env, mdd_obj);
1801 rc = mdd_changelog_data_store(env, mdd, CL_LAYOUT, 0, mdd_obj, handle);
1803 RETURN(mdd_trans_stop(env, mdd, rc, handle));
1806 void mdd_object_make_hint(const struct lu_env *env, struct mdd_object *parent,
1807 struct mdd_object *child, const struct lu_attr *attr,
1808 const struct md_op_spec *spec,
1809 struct dt_allocation_hint *hint)
1811 struct dt_object *np = parent ? mdd_object_child(parent) : NULL;
1812 struct dt_object *nc = mdd_object_child(child);
1814 memset(hint, 0, sizeof(*hint));
1816 /* For striped directory, give striping EA to lod_ah_init, which will
1817 * decide the stripe_offset and stripe count by it. */
1818 if (S_ISDIR(attr->la_mode) &&
1819 unlikely(spec != NULL && spec->sp_cr_flags & MDS_OPEN_HAS_EA)) {
1820 hint->dah_eadata = spec->u.sp_ea.eadata;
1821 hint->dah_eadata_len = spec->u.sp_ea.eadatalen;
1823 hint->dah_eadata = NULL;
1824 hint->dah_eadata_len = 0;
1827 CDEBUG(D_INFO, DFID" eadata %p len %d\n", PFID(mdd_object_fid(child)),
1828 hint->dah_eadata, hint->dah_eadata_len);
1829 /* @hint will be initialized by underlying device. */
1830 nc->do_ops->do_ah_init(env, hint, np, nc, attr->la_mode & S_IFMT);
1834 * do NOT or the MAY_*'s, you'll get the weakest
1836 int accmode(const struct lu_env *env, const struct lu_attr *la, int flags)
1840 /* Sadly, NFSD reopens a file repeatedly during operation, so the
1841 * "acc_mode = 0" allowance for newly-created files isn't honoured.
1842 * NFSD uses the MDS_OPEN_OWNEROVERRIDE flag to say that a file
1843 * owner can write to a file even if it is marked readonly to hide
1844 * its brokenness. (bug 5781) */
1845 if (flags & MDS_OPEN_OWNEROVERRIDE) {
1846 struct lu_ucred *uc = lu_ucred_check(env);
1848 if ((uc == NULL) || (la->la_uid == uc->uc_fsuid))
1852 if (flags & FMODE_READ)
1854 if (flags & (FMODE_WRITE | MDS_OPEN_TRUNC | MDS_OPEN_APPEND))
1856 if (flags & MDS_FMODE_EXEC)
1861 static int mdd_open_sanity_check(const struct lu_env *env,
1862 struct mdd_object *obj,
1863 const struct lu_attr *attr, int flag)
1869 if (mdd_is_dead_obj(obj))
1872 if (S_ISLNK(attr->la_mode))
1875 mode = accmode(env, attr, flag);
1877 if (S_ISDIR(attr->la_mode) && (mode & MAY_WRITE))
1880 if (!(flag & MDS_OPEN_CREATED)) {
1881 rc = mdd_permission_internal(env, obj, attr, mode);
1886 if (S_ISFIFO(attr->la_mode) || S_ISSOCK(attr->la_mode) ||
1887 S_ISBLK(attr->la_mode) || S_ISCHR(attr->la_mode))
1888 flag &= ~MDS_OPEN_TRUNC;
1890 /* For writing append-only file must open it with append mode. */
1891 if (attr->la_flags & LUSTRE_APPEND_FL) {
1892 if ((flag & FMODE_WRITE) && !(flag & MDS_OPEN_APPEND))
1894 if (flag & MDS_OPEN_TRUNC)
1901 static int mdd_open(const struct lu_env *env, struct md_object *obj,
1904 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1905 struct lu_attr *attr = MDD_ENV_VAR(env, cattr);
1908 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1910 rc = mdd_la_get(env, mdd_obj, attr);
1914 rc = mdd_open_sanity_check(env, mdd_obj, attr, flags);
1918 mdd_obj->mod_count++;
1921 mdd_write_unlock(env, mdd_obj);
1925 static int mdd_declare_close(const struct lu_env *env,
1926 struct mdd_object *obj,
1928 struct thandle *handle)
1932 rc = orph_declare_index_delete(env, obj, handle);
1936 return mdo_declare_destroy(env, obj, handle);
1940 * No permission check is needed.
1942 static int mdd_close(const struct lu_env *env, struct md_object *obj,
1943 struct md_attr *ma, int mode)
1945 struct mdd_object *mdd_obj = md2mdd_obj(obj);
1946 struct mdd_device *mdd = mdo2mdd(obj);
1947 struct thandle *handle = NULL;
1950 bool blocked = false;
1953 if (ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_KEEP_ORPHAN) {
1954 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
1955 mdd_obj->mod_count--;
1956 mdd_write_unlock(env, mdd_obj);
1958 if (mdd_obj->mod_flags & ORPHAN_OBJ && !mdd_obj->mod_count)
1959 CDEBUG(D_HA, "Object "DFID" is retained in orphan "
1960 "list\n", PFID(mdd_object_fid(mdd_obj)));
1964 /* mdd_finish_unlink() will always set orphan object as DEAD_OBJ, but
1965 * it might fail to add the object to orphan list (w/o ORPHAN_OBJ). */
1966 /* check without any lock */
1967 is_orphan = mdd_obj->mod_count == 1 &&
1968 (mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0;
1972 /* mdd_trans_create() maybe failed because of barrier_entry(),
1973 * under such case, the orphan MDT-object will be left in the
1974 * orphan list, and when the MDT remount next time, the unused
1975 * orphans will be destroyed automatically.
1977 * One exception: the former mdd_finish_unlink may failed to
1978 * add the orphan MDT-object to the orphan list, then if the
1979 * mdd_trans_create() failed because of barrier_entry(), the
1980 * MDT-object will become real orphan that is neither in the
1981 * namespace nor in the orphan list. Such bad case should be
1982 * very rare and will be handled by e2fsck/lfsck. */
1983 handle = mdd_trans_create(env, mdo2mdd(obj));
1984 if (IS_ERR(handle)) {
1985 rc = PTR_ERR(handle);
1986 if (rc != -EINPROGRESS)
1994 rc = mdd_declare_close(env, mdd_obj, ma, handle);
1998 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL, handle);
2002 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2008 mdd_write_lock(env, mdd_obj, MOR_TGT_CHILD);
2009 rc = mdd_la_get(env, mdd_obj, &ma->ma_attr);
2011 CERROR("Failed to get lu_attr of "DFID": %d\n",
2012 PFID(mdd_object_fid(mdd_obj)), rc);
2016 /* check again with lock */
2017 is_orphan = (mdd_obj->mod_count == 1) &&
2018 ((mdd_obj->mod_flags & (ORPHAN_OBJ | DEAD_OBJ)) != 0 ||
2019 ma->ma_attr.la_nlink == 0);
2021 if (is_orphan && !handle && !blocked) {
2022 mdd_write_unlock(env, mdd_obj);
2026 mdd_obj->mod_count--; /*release open count */
2028 if (!is_orphan || blocked)
2032 /* NB: Object maybe not in orphan list originally, it is rare case for
2033 * mdd_finish_unlink() failure, in that case, the object doesn't have
2034 * ORPHAN_OBJ flag */
2035 if ((mdd_obj->mod_flags & ORPHAN_OBJ) != 0) {
2036 /* remove link to object from orphan index */
2037 LASSERT(handle != NULL);
2038 rc = __mdd_orphan_del(env, mdd_obj, handle);
2040 CERROR("%s: unable to delete "DFID" from orphan list: "
2041 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2042 PFID(mdd_object_fid(mdd_obj)), rc);
2043 /* If object was not deleted from orphan list, do not
2044 * destroy OSS objects, which will be done when next
2049 CDEBUG(D_HA, "Object "DFID" is deleted from orphan "
2050 "list, OSS objects to be destroyed.\n",
2051 PFID(mdd_object_fid(mdd_obj)));
2054 rc = mdo_destroy(env, mdd_obj, handle);
2057 CERROR("%s: unable to delete "DFID" from orphan list: "
2058 "rc = %d\n", lu_dev_name(mdd2lu_dev(mdd)),
2059 PFID(mdd_object_fid(mdd_obj)), rc);
2064 mdd_write_unlock(env, mdd_obj);
2066 if (!rc && !blocked &&
2067 (mode & (FMODE_WRITE | MDS_OPEN_APPEND | MDS_OPEN_TRUNC)) &&
2068 !(ma->ma_valid & MA_FLAGS && ma->ma_attr_flags & MDS_RECOV_OPEN)) {
2069 if (handle == NULL) {
2070 handle = mdd_trans_create(env, mdo2mdd(obj));
2072 GOTO(stop, rc = PTR_ERR(handle));
2074 rc = mdd_declare_changelog_store(env, mdd, NULL, NULL,
2079 rc = mdd_trans_start(env, mdo2mdd(obj), handle);
2084 mdd_changelog_data_store(env, mdd, CL_CLOSE, mode,
2089 if (handle != NULL && !IS_ERR(handle))
2090 rc = mdd_trans_stop(env, mdd, rc, handle);
2096 * Permission check is done when open,
2097 * no need check again.
2099 static int mdd_readpage_sanity_check(const struct lu_env *env,
2100 struct mdd_object *obj)
2102 struct dt_object *next = mdd_object_child(obj);
2106 if (S_ISDIR(mdd_object_type(obj)) && dt_try_as_dir(env, next))
2114 static int mdd_dir_page_build(const struct lu_env *env, union lu_page *lp,
2115 size_t nob, const struct dt_it_ops *iops,
2116 struct dt_it *it, __u32 attr, void *arg)
2118 struct lu_dirpage *dp = &lp->lp_dir;
2122 struct lu_dirent *ent;
2123 struct lu_dirent *last = NULL;
2127 if (nob < sizeof(*dp))
2130 memset(area, 0, sizeof (*dp));
2131 area += sizeof (*dp);
2132 nob -= sizeof (*dp);
2139 len = iops->key_size(env, it);
2141 /* IAM iterator can return record with zero len. */
2145 hash = iops->store(env, it);
2146 if (unlikely(first)) {
2148 dp->ldp_hash_start = cpu_to_le64(hash);
2151 /* calculate max space required for lu_dirent */
2152 recsize = lu_dirent_calc_size(len, attr);
2154 if (nob >= recsize) {
2155 result = iops->rec(env, it, (struct dt_rec *)ent, attr);
2156 if (result == -ESTALE)
2161 /* osd might not able to pack all attributes,
2162 * so recheck rec length */
2163 recsize = le16_to_cpu(ent->lde_reclen);
2165 if (le32_to_cpu(ent->lde_attrs) & LUDA_FID) {
2166 fid_le_to_cpu(&fid, &ent->lde_fid);
2167 if (fid_is_dot_lustre(&fid))
2171 result = (last != NULL) ? 0 :-EINVAL;
2175 ent = (void *)ent + recsize;
2179 result = iops->next(env, it);
2180 if (result == -ESTALE)
2182 } while (result == 0);
2185 dp->ldp_hash_end = cpu_to_le64(hash);
2187 if (last->lde_hash == dp->ldp_hash_end)
2188 dp->ldp_flags |= cpu_to_le32(LDF_COLLIDE);
2189 last->lde_reclen = 0; /* end mark */
2192 /* end of directory */
2193 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2194 else if (result < 0)
2195 CWARN("build page failed: %d!\n", result);
2199 int mdd_readpage(const struct lu_env *env, struct md_object *obj,
2200 const struct lu_rdpg *rdpg)
2202 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2206 if (mdd_object_exists(mdd_obj) == 0) {
2207 CERROR("%s: object "DFID" not found: rc = -2\n",
2208 mdd_obj_dev_name(mdd_obj),PFID(mdd_object_fid(mdd_obj)));
2212 mdd_read_lock(env, mdd_obj, MOR_TGT_CHILD);
2213 rc = mdd_readpage_sanity_check(env, mdd_obj);
2215 GOTO(out_unlock, rc);
2217 if (mdd_is_dead_obj(mdd_obj)) {
2219 struct lu_dirpage *dp;
2222 * According to POSIX, please do not return any entry to client:
2223 * even dot and dotdot should not be returned.
2225 CDEBUG(D_INODE, "readdir from dead object: "DFID"\n",
2226 PFID(mdd_object_fid(mdd_obj)));
2228 if (rdpg->rp_count <= 0)
2229 GOTO(out_unlock, rc = -EFAULT);
2230 LASSERT(rdpg->rp_pages != NULL);
2232 pg = rdpg->rp_pages[0];
2233 dp = (struct lu_dirpage *)kmap(pg);
2234 memset(dp, 0 , sizeof(struct lu_dirpage));
2235 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2236 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2237 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2239 GOTO(out_unlock, rc = LU_PAGE_SIZE);
2242 rc = dt_index_walk(env, mdd_object_child(mdd_obj), rdpg,
2243 mdd_dir_page_build, NULL);
2245 struct lu_dirpage *dp;
2247 dp = kmap(rdpg->rp_pages[0]);
2248 dp->ldp_hash_start = cpu_to_le64(rdpg->rp_hash);
2251 * No pages were processed, mark this for first page
2254 dp->ldp_hash_end = cpu_to_le64(MDS_DIR_END_OFF);
2255 dp->ldp_flags = cpu_to_le32(LDF_EMPTY);
2256 rc = min_t(unsigned int, LU_PAGE_SIZE, rdpg->rp_count);
2258 kunmap(rdpg->rp_pages[0]);
2261 GOTO(out_unlock, rc);
2263 mdd_read_unlock(env, mdd_obj);
2267 static int mdd_object_sync(const struct lu_env *env, struct md_object *obj)
2269 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2271 if (mdd_object_exists(mdd_obj) == 0) {
2274 CERROR("%s: object "DFID" not found: rc = %d\n",
2275 mdd_obj_dev_name(mdd_obj),
2276 PFID(mdd_object_fid(mdd_obj)), rc);
2279 return dt_object_sync(env, mdd_object_child(mdd_obj),
2283 static int mdd_object_lock(const struct lu_env *env,
2284 struct md_object *obj,
2285 struct lustre_handle *lh,
2286 struct ldlm_enqueue_info *einfo,
2287 union ldlm_policy_data *policy)
2289 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2290 return dt_object_lock(env, mdd_object_child(mdd_obj), lh,
2294 static int mdd_object_unlock(const struct lu_env *env,
2295 struct md_object *obj,
2296 struct ldlm_enqueue_info *einfo,
2297 union ldlm_policy_data *policy)
2299 struct mdd_object *mdd_obj = md2mdd_obj(obj);
2300 return dt_object_unlock(env, mdd_object_child(mdd_obj), einfo, policy);
2303 const struct md_object_operations mdd_obj_ops = {
2304 .moo_permission = mdd_permission,
2305 .moo_attr_get = mdd_attr_get,
2306 .moo_attr_set = mdd_attr_set,
2307 .moo_xattr_get = mdd_xattr_get,
2308 .moo_xattr_set = mdd_xattr_set,
2309 .moo_xattr_list = mdd_xattr_list,
2310 .moo_invalidate = mdd_invalidate,
2311 .moo_xattr_del = mdd_xattr_del,
2312 .moo_swap_layouts = mdd_swap_layouts,
2313 .moo_open = mdd_open,
2314 .moo_close = mdd_close,
2315 .moo_readpage = mdd_readpage,
2316 .moo_readlink = mdd_readlink,
2317 .moo_changelog = mdd_changelog,
2318 .moo_object_sync = mdd_object_sync,
2319 .moo_object_lock = mdd_object_lock,
2320 .moo_object_unlock = mdd_object_unlock,
2321 .moo_layout_change = mdd_layout_change,