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/osd/osd_handler.c
34 * Top-level entry points into osd module
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <linux/kallsyms.h>
43 #include <linux/module.h>
44 #include <linux/user_namespace.h>
45 #ifdef HAVE_UIDGID_HEADER
46 # include <linux/uidgid.h>
49 /* prerequisite for linux/xattr.h */
50 #include <linux/types.h>
51 /* prerequisite for linux/xattr.h */
53 /* XATTR_{REPLACE,CREATE} */
54 #include <linux/xattr.h>
56 #include <ldiskfs/ldiskfs.h>
57 #include <ldiskfs/xattr.h>
58 #include <ldiskfs/ldiskfs_extents.h>
61 * struct OBD_{ALLOC,FREE}*()
64 #include <obd_support.h>
65 /* struct ptlrpc_thread */
66 #include <lustre_net.h>
67 #include <lustre_fid.h>
69 #include <uapi/linux/lustre/lustre_param.h>
71 #include "osd_internal.h"
72 #include "osd_dynlocks.h"
74 /* llo_* api support */
75 #include <md_object.h>
76 #include <lustre_quota.h>
78 #include <lustre_linkea.h>
80 /* Maximum EA size is limited by LNET_MTU for remote objects */
81 #define OSD_MAX_EA_SIZE 1048364
84 module_param(ldiskfs_pdo, int, 0644);
85 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
87 int ldiskfs_track_declares_assert;
88 module_param(ldiskfs_track_declares_assert, int, 0644);
89 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
91 /* Slab to allocate dynlocks */
92 struct kmem_cache *dynlock_cachep;
94 /* Slab to allocate osd_it_ea */
95 struct kmem_cache *osd_itea_cachep;
97 static struct lu_kmem_descr ldiskfs_caches[] = {
99 .ckd_cache = &dynlock_cachep,
100 .ckd_name = "dynlock_cache",
101 .ckd_size = sizeof(struct dynlock_handle)
104 .ckd_cache = &osd_itea_cachep,
105 .ckd_name = "osd_itea_cache",
106 .ckd_size = sizeof(struct osd_it_ea)
113 static const char dot[] = ".";
114 static const char dotdot[] = "..";
116 static const struct lu_object_operations osd_lu_obj_ops;
117 static const struct dt_object_operations osd_obj_ops;
118 static const struct dt_object_operations osd_obj_otable_it_ops;
119 static const struct dt_index_operations osd_index_iam_ops;
120 static const struct dt_index_operations osd_index_ea_ops;
122 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
123 const struct lu_fid *fid);
124 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
125 struct osd_device *osd);
127 int osd_trans_declare_op2rb[] = {
128 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
129 [OSD_OT_PUNCH] = OSD_OT_MAX,
130 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
131 [OSD_OT_CREATE] = OSD_OT_DESTROY,
132 [OSD_OT_DESTROY] = OSD_OT_CREATE,
133 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
134 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
135 [OSD_OT_WRITE] = OSD_OT_WRITE,
136 [OSD_OT_INSERT] = OSD_OT_DELETE,
137 [OSD_OT_DELETE] = OSD_OT_INSERT,
138 [OSD_OT_QUOTA] = OSD_OT_MAX,
141 static int osd_has_index(const struct osd_object *obj)
143 return obj->oo_dt.do_index_ops != NULL;
146 static int osd_object_invariant(const struct lu_object *l)
148 return osd_invariant(osd_obj(l));
152 * Concurrency: doesn't matter
154 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
156 struct osd_thread_info *oti = osd_oti_get(env);
158 return oti->oti_w_locks > 0 && o->oo_owner == env;
162 * Concurrency: doesn't access mutable data
164 static int osd_root_get(const struct lu_env *env,
165 struct dt_device *dev, struct lu_fid *f)
167 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
172 * the following set of functions are used to maintain per-thread
173 * cache of FID->ino mapping. this mechanism is needed to resolve
174 * FID to inode at dt_insert() which in turn stores ino in the
175 * directory entries to keep ldiskfs compatible with ext[34].
176 * due to locking-originated restrictions we can't lookup ino
177 * using LU cache (deadlock is possible). lookup using OI is quite
178 * expensive. so instead we maintain this cache and methods like
179 * dt_create() fill it. so in the majority of cases dt_insert() is
180 * able to find needed mapping in lockless manner.
182 static struct osd_idmap_cache *
183 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
184 const struct lu_fid *fid)
186 struct osd_thread_info *oti = osd_oti_get(env);
187 struct osd_idmap_cache *idc = oti->oti_ins_cache;
190 for (i = 0; i < oti->oti_ins_cache_used; i++) {
191 if (!lu_fid_eq(&idc[i].oic_fid, fid))
193 if (idc[i].oic_dev != osd)
202 static struct osd_idmap_cache *
203 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
204 const struct lu_fid *fid)
206 struct osd_thread_info *oti = osd_oti_get(env);
207 struct osd_idmap_cache *idc;
210 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
211 i = oti->oti_ins_cache_size * 2;
213 i = OSD_INS_CACHE_SIZE;
214 OBD_ALLOC(idc, sizeof(*idc) * i);
216 return ERR_PTR(-ENOMEM);
217 if (oti->oti_ins_cache != NULL) {
218 memcpy(idc, oti->oti_ins_cache,
219 oti->oti_ins_cache_used * sizeof(*idc));
220 OBD_FREE(oti->oti_ins_cache,
221 oti->oti_ins_cache_used * sizeof(*idc));
223 oti->oti_ins_cache = idc;
224 oti->oti_ins_cache_size = i;
227 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
230 idc->oic_lid.oii_ino = 0;
231 idc->oic_lid.oii_gen = 0;
238 * lookup mapping for the given fid in the cache, initialize a
239 * new one if not found. the initialization checks whether the
240 * object is local or remote. for local objects, OI is used to
241 * learn ino/generation. the function is used when the caller
242 * has no information about the object, e.g. at dt_insert().
244 static struct osd_idmap_cache *
245 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
246 const struct lu_fid *fid)
248 struct osd_idmap_cache *idc;
251 idc = osd_idc_find(env, osd, fid);
252 LASSERT(!IS_ERR(idc));
256 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
257 osd->od_svname, PFID(fid));
259 /* new mapping is needed */
260 idc = osd_idc_add(env, osd, fid);
262 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
263 osd->od_svname, PFID(fid), PTR_ERR(idc));
268 rc = osd_remote_fid(env, osd, fid);
269 if (unlikely(rc < 0))
273 /* the object is local, lookup in OI */
274 /* XXX: probably cheaper to lookup in LU first? */
275 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
277 if (unlikely(rc < 0)) {
278 CERROR("can't lookup: rc = %d\n", rc);
282 /* the object is remote */
290 * lookup mapping for given FID and fill it from the given object.
291 * the object is lolcal by definition.
293 static int osd_idc_find_and_init(const struct lu_env *env,
294 struct osd_device *osd,
295 struct osd_object *obj)
297 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
298 struct osd_idmap_cache *idc;
300 idc = osd_idc_find(env, osd, fid);
301 LASSERT(!IS_ERR(idc));
303 if (obj->oo_inode == NULL)
305 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
306 LASSERT(idc->oic_lid.oii_ino == 0);
307 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
308 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
313 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
314 osd->od_svname, PFID(fid));
316 /* new mapping is needed */
317 idc = osd_idc_add(env, osd, fid);
319 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
320 osd->od_svname, PFID(fid), PTR_ERR(idc));
324 if (obj->oo_inode != NULL) {
325 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
326 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
332 * OSD object methods.
336 * Concurrency: no concurrent access is possible that early in object
339 static struct lu_object *osd_object_alloc(const struct lu_env *env,
340 const struct lu_object_header *hdr,
343 struct osd_object *mo;
348 struct lu_object_header *h;
349 struct osd_device *o = osd_dev(d);
351 l = &mo->oo_dt.do_lu;
352 if (unlikely(o->od_in_init)) {
359 lu_object_header_init(h);
360 lu_object_init(l, h, d);
361 lu_object_add_top(h, l);
364 dt_object_init(&mo->oo_dt, NULL, d);
365 mo->oo_header = NULL;
368 mo->oo_dt.do_ops = &osd_obj_ops;
369 l->lo_ops = &osd_lu_obj_ops;
370 init_rwsem(&mo->oo_sem);
371 init_rwsem(&mo->oo_ext_idx_sem);
372 spin_lock_init(&mo->oo_guard);
373 INIT_LIST_HEAD(&mo->oo_xattr_list);
379 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
380 struct dentry *dentry, struct lustre_ost_attrs *loa)
384 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
385 (void *)loa, sizeof(*loa));
387 struct lustre_mdt_attrs *lma = &loa->loa_lma;
389 if (rc < sizeof(*lma))
393 lustre_loa_swab(loa, true);
394 /* Check LMA compatibility */
395 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
396 CWARN("%s: unsupported incompat LMA feature(s) %#x "
397 "for fid = "DFID", ino = %lu\n",
399 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
400 PFID(&lma->lma_self_fid), inode->i_ino);
403 } else if (rc == 0) {
411 * retrieve object from backend ext fs.
413 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
414 struct osd_inode_id *id)
417 struct inode *inode = NULL;
420 * if we look for an inode withing a running
421 * transaction, then we risk to deadlock
422 * osd_dirent_check_repair() breaks this
424 /* LASSERT(current->journal_info == NULL); */
426 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
428 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
429 id->oii_ino, PTR_ERR(inode));
430 } else if (id->oii_gen != OSD_OII_NOGEN &&
431 inode->i_generation != id->oii_gen) {
432 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
433 "i_generation = %u\n",
434 id->oii_ino, id->oii_gen, inode->i_generation);
436 inode = ERR_PTR(-ESTALE);
437 } else if (inode->i_nlink == 0) {
439 * due to parallel readdir and unlink,
440 * we can have dead inode here.
442 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
444 inode = ERR_PTR(-ESTALE);
445 } else if (is_bad_inode(inode)) {
446 CWARN("%s: bad inode: ino = %u\n",
447 osd_dev2name(dev), id->oii_ino);
449 inode = ERR_PTR(-ENOENT);
450 } else if ((rc = osd_attach_jinode(inode))) {
454 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
455 if (id->oii_gen == OSD_OII_NOGEN)
456 osd_id_gen(id, inode->i_ino, inode->i_generation);
459 * Do not update file c/mtime in ldiskfs.
460 * NB: we don't have any lock to protect this because we don't
461 * have reference on osd_object now, but contention with
462 * another lookup + attr_set can't happen in the tiny window
463 * between if (...) and set S_NOCMTIME.
465 if (!(inode->i_flags & S_NOCMTIME))
466 inode->i_flags |= S_NOCMTIME;
471 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
472 handle_t *handle, struct dentry *child,
473 struct inode *inode, struct htree_lock *hlock)
477 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
478 if (rc == -ENOBUFS || rc == -ENOSPC) {
479 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
480 struct inode *parent = child->d_parent->d_inode;
481 struct lu_fid *fid = NULL;
483 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
485 fid = &loa->loa_lma.lma_self_fid;
486 } else if (rc2 == -ENODATA) {
487 if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
488 fid = &info->oti_fid3;
489 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
490 } else if (!osd->od_is_ost && osd->od_index == 0) {
491 fid = &info->oti_fid3;
492 lu_igif_build(fid, parent->i_ino,
493 parent->i_generation);
498 CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
499 "maximum entry limit\n",
500 osd_name(osd), parent->i_ino, PFID(fid),
501 rc == -ENOSPC ? "has reached" : "is approaching");
503 CWARN("%s: directory (inode: %lu, FID: unknown) %s "
504 "maximum entry limit\n",
505 osd_name(osd), parent->i_ino,
506 rc == -ENOSPC ? "has reached" : "is approaching");
508 /* ignore such error now */
518 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
519 struct osd_inode_id *id, struct lu_fid *fid)
521 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
525 inode = osd_iget(info, dev, id);
529 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
531 *fid = loa->loa_lma.lma_self_fid;
532 } else if (rc == -ENODATA) {
533 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
534 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
536 lu_igif_build(fid, inode->i_ino, inode->i_generation);
544 static struct inode *osd_iget_check(struct osd_thread_info *info,
545 struct osd_device *dev,
546 const struct lu_fid *fid,
547 struct osd_inode_id *id,
556 * The cached OI mapping is trustable. If we cannot locate the inode
557 * via the cached OI mapping, then return the failure to the caller
558 * directly without further OI checking.
562 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
565 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
568 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
569 PFID(fid), id->oii_ino, rc);
573 if (is_bad_inode(inode)) {
578 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
579 PFID(fid), id->oii_ino);
583 if (id->oii_gen != OSD_OII_NOGEN &&
584 inode->i_generation != id->oii_gen) {
589 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
590 "oii_gen = %u, i_generation = %u\n", PFID(fid),
591 id->oii_ino, id->oii_gen, inode->i_generation);
595 if (inode->i_nlink == 0) {
600 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
601 PFID(fid), id->oii_ino);
605 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
609 __u32 saved_ino = id->oii_ino;
610 __u32 saved_gen = id->oii_gen;
613 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
615 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
617 * XXX: There are four possible cases:
619 * Backup/restore caused the OI invalid.
621 * Someone unlinked the object but NOT removed
622 * the OI mapping, such as mount target device
623 * as ldiskfs, and modify something directly.
625 * Someone just removed the object between the
626 * former oi_lookup and the iget. It is normal.
627 * 4. Other failure cases.
629 * Generally, when the device is mounted, it will
630 * auto check whether the system is restored from
631 * file-level backup or not. We trust such detect
632 * to distinguish the 1st case from the 2nd case:
633 * if the OI files are consistent but may contain
634 * stale OI mappings because of case 2, if iget()
635 * returns -ENOENT or -ESTALE, then it should be
640 * If the OI mapping was in OI file before the
641 * osd_iget_check(), but now, it is disappear,
642 * then it must be removed by race. That is a
648 * It is the OI scrub updated the OI mapping by race.
649 * The new OI mapping must be valid.
651 if (saved_ino != id->oii_ino ||
652 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
661 if (dev->od_scrub.os_scrub.os_file.sf_flags &
664 * It still can be the case 2, but we cannot
665 * distinguish it from the case 1. So return
666 * -EREMCHG to block current operation until
667 * OI scrub rebuilt the OI mappings.
676 if (inode->i_generation == id->oii_gen)
681 if (id->oii_gen == OSD_OII_NOGEN)
682 osd_id_gen(id, inode->i_ino, inode->i_generation);
685 * Do not update file c/mtime in ldiskfs.
686 * NB: we don't have any lock to protect this because we don't
687 * have reference on osd_object now, but contention with
688 * another lookup + attr_set can't happen in the tiny window
689 * between if (...) and set S_NOCMTIME.
691 if (!(inode->i_flags & S_NOCMTIME))
692 inode->i_flags |= S_NOCMTIME;
709 * \retval +v: new filter_fid, does not contain self-fid
710 * \retval 0: filter_fid_old, contains self-fid
711 * \retval -v: other failure cases
713 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
714 struct dentry *dentry, struct lu_fid *fid)
716 struct filter_fid_old *ff = &info->oti_ff;
717 struct ost_id *ostid = &info->oti_ostid;
720 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
721 if (rc == sizeof(*ff)) {
723 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
724 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
726 * XXX: use 0 as the index for compatibility, the caller will
727 * handle index related issues when necessary.
730 ostid_to_fid(fid, ostid, 0);
731 } else if (rc == sizeof(struct filter_fid)) {
733 } else if (rc >= 0) {
740 static int osd_lma_self_repair(struct osd_thread_info *info,
741 struct osd_device *osd, struct inode *inode,
742 const struct lu_fid *fid, __u32 compat)
747 LASSERT(current->journal_info == NULL);
749 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
750 osd_dto_credits_noquota[DTO_XATTR_SET]);
753 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
758 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
760 CWARN("%s: cannot self repair the LMA: rc = %d\n",
762 ldiskfs_journal_stop(jh);
766 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
768 struct osd_thread_info *info = osd_oti_get(env);
769 struct osd_device *osd = osd_obj2dev(obj);
770 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
771 struct lustre_mdt_attrs *lma = &loa->loa_lma;
772 struct inode *inode = obj->oo_inode;
773 struct dentry *dentry = &info->oti_obj_dentry;
774 struct lu_fid *fid = NULL;
775 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
780 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
781 (void *)loa, sizeof(*loa));
782 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
783 fid = &lma->lma_self_fid;
784 rc = osd_get_idif(info, inode, dentry, fid);
785 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
787 * For the given OST-object, if it has neither LMA nor
788 * FID in XATTR_NAME_FID, then the given FID (which is
789 * contained in the @obj, from client RPC for locating
790 * the OST-object) is trusted. We use it to generate
793 osd_lma_self_repair(info, osd, inode, rfid,
804 lustre_lma_swab(lma);
805 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
806 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
807 S_ISREG(inode->i_mode)))) {
808 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
809 "fid = "DFID", ino = %lu\n", osd_name(osd),
810 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
811 PFID(rfid), inode->i_ino);
814 fid = &lma->lma_self_fid;
815 if (lma->lma_compat & LMAC_STRIPE_INFO &&
817 obj->oo_pfid_in_lma = 1;
818 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
820 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
824 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
825 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
826 struct ost_id *oi = &info->oti_ostid;
827 struct lu_fid *fid1 = &info->oti_fid3;
828 __u32 idx = fid_idif_ost_idx(rfid);
831 * For old IDIF, the OST index is not part of the IDIF,
832 * Means that different OSTs may have the same IDIFs.
833 * Under such case, we need to make some compatible
834 * check to make sure to trigger OI scrub properly.
836 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
837 /* Given @rfid is new, LMA is old. */
838 fid_to_ostid(fid, oi);
839 ostid_to_fid(fid1, oi, idx);
840 if (lu_fid_eq(fid1, rfid)) {
841 if (osd->od_index_in_idif)
842 osd_lma_self_repair(info, osd,
856 struct osd_check_lmv_buf {
857 #ifdef HAVE_DIR_CONTEXT
858 /* please keep it as first member */
859 struct dir_context ctx;
861 struct osd_thread_info *oclb_info;
862 struct osd_device *oclb_dev;
863 struct osd_idmap_cache *oclb_oic;
869 * It is called internally by ->readdir() to filter out the
870 * local slave object's FID of the striped directory.
872 * \retval 1 found the local slave's FID
873 * \retval 0 continue to check next item
874 * \retval -ve for failure
876 #ifdef HAVE_FILLDIR_USE_CTX
877 static int osd_stripe_dir_filldir(struct dir_context *buf,
879 static int osd_stripe_dir_filldir(void *buf,
881 const char *name, int namelen,
882 loff_t offset, __u64 ino, unsigned int d_type)
884 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
885 struct osd_thread_info *oti = oclb->oclb_info;
886 struct lu_fid *fid = &oti->oti_fid3;
887 struct osd_inode_id *id = &oti->oti_id3;
888 struct osd_device *dev = oclb->oclb_dev;
889 struct osd_idmap_cache *oic = oclb->oclb_oic;
898 sscanf(name + 1, SFID, RFID(fid));
899 if (!fid_is_sane(fid))
902 if (osd_remote_fid(oti->oti_env, dev, fid))
905 osd_id_gen(id, ino, OSD_OII_NOGEN);
906 inode = osd_iget(oti, dev, id);
908 return PTR_ERR(inode);
911 osd_add_oi_cache(oti, dev, id, fid);
915 osd_oii_insert(dev, oic, true);
916 oclb->oclb_found = true;
922 * When lookup item under striped directory, we need to locate the master
923 * MDT-object of the striped directory firstly, then the client will send
924 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
925 * and the item's name. If the system is restored from MDT file level backup,
926 * then before the OI scrub completely built the OI files, the OI mappings of
927 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
928 * not a problem for the master MDT-object. Because when locate the master
929 * MDT-object, we will do name based lookup (for the striped directory itself)
930 * firstly, during such process we can setup the correct OI mapping for the
931 * master MDT-object. But it will be trouble for the slave MDT-object. Because
932 * the client will not trigger name based lookup on the MDT to locate the slave
933 * MDT-object before locating item under the striped directory, then when
934 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
935 * is invalid and does not know what the right OI mapping is, then the MDT has
936 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
937 * the OI file, related OI mapping is unknown yet, please try again later. And
938 * then client will re-try the RPC again and again until related OI mapping has
939 * been updated. That is quite inefficient.
941 * To resolve above trouble, we will handle it as the following two cases:
943 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
944 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
945 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
946 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
947 * directly. Please check osd_fid_lookup().
949 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
950 * Under such case, during lookup the master MDT-object, we will lookup the
951 * slave MDT-object via readdir against the master MDT-object, because the
952 * slave MDT-objects information are stored as sub-directories with the name
953 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
954 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
955 * the correct OI mapping for the slave MDT-object.
957 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
958 struct inode *inode, struct osd_idmap_cache *oic)
960 struct lu_buf *buf = &oti->oti_big_buf;
961 struct dentry *dentry = &oti->oti_obj_dentry;
962 struct file *filp = &oti->oti_file;
963 const struct file_operations *fops;
964 struct lmv_mds_md_v1 *lmv1;
965 struct osd_check_lmv_buf oclb = {
966 #ifdef HAVE_DIR_CONTEXT
967 .ctx.actor = osd_stripe_dir_filldir,
979 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
982 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
984 lu_buf_realloc(buf, rc);
985 if (buf->lb_buf == NULL)
986 GOTO(out, rc = -ENOMEM);
992 if (unlikely(rc == 0 || rc == -ENODATA))
998 if (unlikely(buf->lb_buf == NULL)) {
999 lu_buf_realloc(buf, rc);
1000 if (buf->lb_buf == NULL)
1001 GOTO(out, rc = -ENOMEM);
1007 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1010 fops = inode->i_fop;
1011 dentry->d_inode = inode;
1012 dentry->d_sb = inode->i_sb;
1014 filp->f_path.dentry = dentry;
1015 filp->f_mode = FMODE_64BITHASH;
1016 filp->f_mapping = inode->i_mapping;
1018 filp->private_data = NULL;
1019 set_file_inode(filp, inode);
1022 oclb.oclb_items = 0;
1023 #ifdef HAVE_DIR_CONTEXT
1024 oclb.ctx.pos = filp->f_pos;
1025 #ifdef HAVE_ITERATE_SHARED
1026 rc = fops->iterate_shared(filp, &oclb.ctx);
1028 rc = fops->iterate(filp, &oclb.ctx);
1030 filp->f_pos = oclb.ctx.pos;
1032 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
1034 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1035 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1036 fops->release(inode, filp);
1040 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
1041 DFID": rc = %d\n", osd_ino2name(inode),
1042 inode->i_ino, inode->i_generation,
1043 PFID(&oic->oic_fid), rc);
1050 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1051 const struct lu_fid *fid,
1052 const struct lu_object_conf *conf)
1054 struct osd_thread_info *info;
1055 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1056 struct osd_device *dev;
1057 struct osd_idmap_cache *oic;
1058 struct osd_inode_id *id;
1059 struct inode *inode = NULL;
1060 struct lustre_scrub *scrub;
1061 struct scrub_file *sf;
1062 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1067 bool remote = false;
1068 bool trusted = true;
1069 bool updated = false;
1070 bool checked = false;
1074 LINVRNT(osd_invariant(obj));
1075 LASSERT(obj->oo_inode == NULL);
1076 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1078 dev = osd_dev(ldev);
1079 scrub = &dev->od_scrub.os_scrub;
1080 sf = &scrub->os_file;
1081 info = osd_oti_get(env);
1083 oic = &info->oti_cache;
1085 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1089 * For the object is created as locking anchor, or for the object to
1090 * be created on disk. No need to osd_oi_lookup() at here because FID
1091 * shouldn't never be re-used, if it's really a duplicate FID from
1092 * unexpected reason, we should be able to detect it later by calling
1093 * do_create->osd_oi_insert().
1095 if (conf && conf->loc_flags & LOC_F_NEW)
1096 GOTO(out, result = 0);
1098 /* Search order: 1. per-thread cache. */
1099 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1105 if (!list_empty(&scrub->os_inconsistent_items)) {
1106 /* Search order: 2. OI scrub pending list. */
1107 result = osd_oii_lookup(dev, fid, id);
1113 * The OI mapping in the OI file can be updated by the OI scrub
1114 * when we locate the inode via FID. So it may be not trustable.
1118 /* Search order: 3. OI files. */
1119 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1120 if (result == -ENOENT) {
1121 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1122 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1123 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1125 GOTO(out, result = 0);
1134 obj->oo_inode = NULL;
1135 /* for later passes through checks, not true on first pass */
1136 if (!IS_ERR_OR_NULL(inode))
1139 inode = osd_iget_check(info, dev, fid, id, trusted);
1140 if (!IS_ERR(inode)) {
1141 obj->oo_inode = inode;
1149 result = PTR_ERR(inode);
1150 if (result == -ENOENT || result == -ESTALE)
1151 GOTO(out, result = 0);
1153 if (result != -EREMCHG)
1158 * We still have chance to get the valid inode: for the
1159 * object which is referenced by remote name entry, the
1160 * object on the local MDT will be linked under the dir
1161 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1163 * We do not know whether the object for the given FID
1164 * is referenced by some remote name entry or not, and
1165 * especially for DNE II, a multiple-linked object may
1166 * have many name entries reside on many MDTs.
1168 * To simplify the operation, OSD will not distinguish
1169 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1170 * only happened for the RPC from other MDT during the
1171 * OI scrub, or for the client side RPC with FID only,
1172 * such as FID to path, or from old connected client.
1175 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1179 flags |= SS_AUTO_PARTIAL;
1180 flags &= ~SS_AUTO_FULL;
1185 if (thread_is_running(&scrub->os_thread)) {
1186 if (scrub->os_partial_scan && !scrub->os_in_join)
1189 if (IS_ERR_OR_NULL(inode) || result)
1190 GOTO(out, result = -EINPROGRESS);
1193 LASSERT(obj->oo_inode == inode);
1195 osd_add_oi_cache(info, dev, id, fid);
1196 osd_oii_insert(dev, oic, true);
1200 if (dev->od_auto_scrub_interval == AS_NEVER) {
1202 GOTO(out, result = -EREMCHG);
1205 LASSERT(obj->oo_inode == inode);
1207 osd_add_oi_cache(info, dev, id, fid);
1212 rc1 = osd_scrub_start(env, dev, flags);
1213 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1214 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1215 if (rc1 && rc1 != -EALREADY)
1216 GOTO(out, result = -EREMCHG);
1218 if (IS_ERR_OR_NULL(inode) || result)
1219 GOTO(out, result = -EINPROGRESS);
1222 LASSERT(obj->oo_inode == inode);
1224 osd_add_oi_cache(info, dev, id, fid);
1225 osd_oii_insert(dev, oic, true);
1230 if (unlikely(obj->oo_header))
1233 result = osd_check_lma(env, obj);
1237 LASSERTF(id->oii_ino == inode->i_ino &&
1238 id->oii_gen == inode->i_generation,
1239 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1240 PFID(fid), id->oii_ino, id->oii_gen,
1241 inode->i_ino, inode->i_generation);
1243 saved_ino = inode->i_ino;
1244 saved_gen = inode->i_generation;
1246 if (unlikely(result == -ENODATA)) {
1248 * If the OI scrub updated the OI mapping by race, it
1249 * must be valid. Trust the inode that has no LMA EA.
1254 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1257 * The OI mapping is still there, the inode is still
1258 * valid. It is just becaues the inode has no LMA EA.
1260 if (saved_ino == id->oii_ino &&
1261 saved_gen == id->oii_gen)
1265 * It is the OI scrub updated the OI mapping by race.
1266 * The new OI mapping must be valid.
1274 * "result == -ENOENT" means that the OI mappinghas been
1275 * removed by race, so the inode belongs to other object.
1277 * Others error can be returned directly.
1279 if (result == -ENOENT) {
1282 obj->oo_inode = NULL;
1287 if (result != -EREMCHG)
1292 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1294 * "result == -ENOENT" means the cached OI mapping has been removed
1295 * from the OI file by race, above inode belongs to other object.
1297 if (result == -ENOENT) {
1300 obj->oo_inode = NULL;
1301 GOTO(out, result = 0);
1307 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1313 * It is the OI scrub updated the OI mapping by race.
1314 * The new OI mapping must be valid.
1322 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1323 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1325 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1327 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1329 obj->oo_pfid_in_lma = 1;
1330 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1332 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1333 } else if (result != -ENODATA) {
1338 obj->oo_compat_dot_created = 1;
1339 obj->oo_compat_dotdot_created = 1;
1341 if (S_ISDIR(inode->i_mode) &&
1342 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1343 osd_check_lmv(info, dev, inode, oic);
1345 result = osd_attach_jinode(inode);
1350 GOTO(out, result = 0);
1352 LASSERT(!obj->oo_hl_head);
1353 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1355 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1358 if (result || !obj->oo_inode) {
1359 if (!IS_ERR_OR_NULL(inode))
1362 obj->oo_inode = NULL;
1364 fid_zero(&oic->oic_fid);
1367 LINVRNT(osd_invariant(obj));
1372 * Concurrency: shouldn't matter.
1374 static void osd_object_init0(struct osd_object *obj)
1376 LASSERT(obj->oo_inode != NULL);
1377 obj->oo_dt.do_body_ops = &osd_body_ops;
1378 obj->oo_dt.do_lu.lo_header->loh_attr |=
1379 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1383 * Concurrency: no concurrent access is possible that early in object
1386 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1387 const struct lu_object_conf *conf)
1389 struct osd_object *obj = osd_obj(l);
1392 LINVRNT(osd_invariant(obj));
1394 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1395 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1396 l->lo_header->loh_attr |= LOHA_EXISTS;
1400 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1401 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1402 if (result == 0 && obj->oo_inode != NULL) {
1403 struct osd_thread_info *oti = osd_oti_get(env);
1404 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1406 osd_object_init0(obj);
1407 if (unlikely(obj->oo_header))
1410 result = osd_get_lma(oti, obj->oo_inode,
1411 &oti->oti_obj_dentry, loa);
1414 * Convert LMAI flags to lustre LMA flags
1415 * and cache it to oo_lma_flags
1418 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1419 } else if (result == -ENODATA) {
1424 LINVRNT(osd_invariant(obj));
1429 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1430 * The left part is for value, binary mode.
1432 struct osd_xattr_entry {
1433 struct list_head oxe_list;
1437 struct rcu_head oxe_rcu;
1441 static int osd_oxc_get(struct osd_object *obj, const char *name,
1444 struct osd_xattr_entry *tmp;
1445 struct osd_xattr_entry *oxe = NULL;
1446 size_t namelen = strlen(name);
1452 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1453 if (namelen == tmp->oxe_namelen &&
1454 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1461 GOTO(out, rc = -ENOENT);
1463 if (!oxe->oxe_exist)
1464 GOTO(out, rc = -ENODATA);
1467 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1470 if (buf->lb_buf == NULL)
1473 if (buf->lb_len < rc)
1474 GOTO(out, rc = -ERANGE);
1476 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1484 static void osd_oxc_free(struct rcu_head *head)
1486 struct osd_xattr_entry *oxe;
1488 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1489 OBD_FREE(oxe, oxe->oxe_len);
1492 static void osd_oxc_add(struct osd_object *obj, const char *name,
1493 const char *buf, int buflen)
1495 struct osd_xattr_entry *oxe;
1496 struct osd_xattr_entry *old = NULL;
1497 struct osd_xattr_entry *tmp;
1498 size_t namelen = strlen(name);
1499 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1501 OBD_ALLOC(oxe, len);
1505 INIT_LIST_HEAD(&oxe->oxe_list);
1507 oxe->oxe_namelen = namelen;
1508 memcpy(oxe->oxe_buf, name, namelen);
1510 LASSERT(buf != NULL);
1511 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1512 oxe->oxe_exist = true;
1514 oxe->oxe_exist = false;
1517 /* this should be rarely called, just remove old and add new */
1518 spin_lock(&obj->oo_guard);
1519 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1520 if (namelen == tmp->oxe_namelen &&
1521 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1527 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1528 call_rcu(&old->oxe_rcu, osd_oxc_free);
1530 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1532 spin_unlock(&obj->oo_guard);
1535 static void osd_oxc_del(struct osd_object *obj, const char *name)
1537 struct osd_xattr_entry *oxe;
1538 size_t namelen = strlen(name);
1540 spin_lock(&obj->oo_guard);
1541 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1542 if (namelen == oxe->oxe_namelen &&
1543 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1544 list_del_rcu(&oxe->oxe_list);
1545 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1549 spin_unlock(&obj->oo_guard);
1552 static void osd_oxc_fini(struct osd_object *obj)
1554 struct osd_xattr_entry *oxe, *next;
1556 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1557 list_del(&oxe->oxe_list);
1558 OBD_FREE(oxe, oxe->oxe_len);
1563 * Concurrency: no concurrent access is possible that late in object
1566 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1568 struct osd_object *obj = osd_obj(l);
1569 struct lu_object_header *h = obj->oo_header;
1571 LINVRNT(osd_invariant(obj));
1574 dt_object_fini(&obj->oo_dt);
1575 if (obj->oo_hl_head != NULL)
1576 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1579 lu_object_header_fini(h);
1585 * Concurrency: no concurrent access is possible that late in object
1588 static void osd_index_fini(struct osd_object *o)
1590 struct iam_container *bag;
1592 if (o->oo_dir != NULL) {
1593 bag = &o->oo_dir->od_container;
1594 if (o->oo_inode != NULL) {
1595 if (bag->ic_object == o->oo_inode)
1596 iam_container_fini(bag);
1598 OBD_FREE_PTR(o->oo_dir);
1604 * Concurrency: no concurrent access is possible that late in object
1605 * life-cycle (for all existing callers, that is. New callers have to provide
1606 * their own locking.)
1608 static int osd_inode_unlinked(const struct inode *inode)
1610 return inode->i_nlink == 0;
1614 OSD_TXN_OI_DELETE_CREDITS = 20,
1615 OSD_TXN_INODE_DELETE_CREDITS = 20
1622 #if OSD_THANDLE_STATS
1624 * Set time when the handle is allocated
1626 static void osd_th_alloced(struct osd_thandle *oth)
1628 oth->oth_alloced = ktime_get();
1632 * Set time when the handle started
1634 static void osd_th_started(struct osd_thandle *oth)
1636 oth->oth_started = ktime_get();
1640 * Check whether the we deal with this handle for too long.
1642 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1643 ktime_t alloced, ktime_t started,
1646 ktime_t now = ktime_get();
1648 LASSERT(dev != NULL);
1650 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1651 ktime_us_delta(started, alloced));
1652 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1653 ktime_us_delta(closed, started));
1654 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1655 ktime_us_delta(now, closed));
1657 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1658 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1659 oth, now, alloced, started, closed);
1660 libcfs_debug_dumpstack(NULL);
1664 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1666 ktime_t __closed = ktime_get(); \
1667 ktime_t __alloced = oth->oth_alloced; \
1668 ktime_t __started = oth->oth_started; \
1671 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1674 #else /* OSD_THANDLE_STATS */
1676 #define osd_th_alloced(h) do {} while(0)
1677 #define osd_th_started(h) do {} while(0)
1678 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1680 #endif /* OSD_THANDLE_STATS */
1683 * Concurrency: doesn't access mutable data.
1685 static int osd_param_is_not_sane(const struct osd_device *dev,
1686 const struct thandle *th)
1688 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1690 return oh->ot_credits > osd_transaction_size(dev);
1694 * Concurrency: shouldn't matter.
1696 static void osd_trans_commit_cb(struct super_block *sb,
1697 struct ldiskfs_journal_cb_entry *jcb, int error)
1699 struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1700 struct thandle *th = &oh->ot_super;
1701 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1702 struct dt_txn_commit_cb *dcb, *tmp;
1704 LASSERT(oh->ot_handle == NULL);
1707 CERROR("transaction @0x%p commit error: %d\n", th, error);
1709 dt_txn_hook_commit(th);
1711 /* call per-transaction callbacks if any */
1712 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1714 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1715 "commit callback entry: magic=%x name='%s'\n",
1716 dcb->dcb_magic, dcb->dcb_name);
1717 list_del_init(&dcb->dcb_linkage);
1718 dcb->dcb_func(NULL, th, dcb, error);
1721 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1728 #ifndef HAVE_SB_START_WRITE
1729 # define sb_start_write(sb) do {} while (0)
1730 # define sb_end_write(sb) do {} while (0)
1733 static struct thandle *osd_trans_create(const struct lu_env *env,
1734 struct dt_device *d)
1736 struct osd_thread_info *oti = osd_oti_get(env);
1737 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1738 struct osd_thandle *oh;
1744 CERROR("%s: someone try to start transaction under "
1745 "readonly mode, should be disabled.\n",
1746 osd_name(osd_dt_dev(d)));
1748 RETURN(ERR_PTR(-EROFS));
1751 /* on pending IO in this thread should left from prev. request */
1752 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1754 sb_start_write(osd_sb(osd_dt_dev(d)));
1756 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1758 sb_end_write(osd_sb(osd_dt_dev(d)));
1759 RETURN(ERR_PTR(-ENOMEM));
1762 oh->ot_quota_trans = &oti->oti_quota_trans;
1763 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1768 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1769 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1770 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1773 memset(oti->oti_declare_ops, 0,
1774 sizeof(oti->oti_declare_ops));
1775 memset(oti->oti_declare_ops_cred, 0,
1776 sizeof(oti->oti_declare_ops_cred));
1777 memset(oti->oti_declare_ops_used, 0,
1778 sizeof(oti->oti_declare_ops_used));
1780 oti->oti_ins_cache_depth++;
1785 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1787 struct osd_thread_info *oti = osd_oti_get(env);
1788 struct osd_thandle *oh;
1790 oh = container_of0(th, struct osd_thandle, ot_super);
1791 LASSERT(oh != NULL);
1793 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1794 oti->oti_declare_ops[OSD_OT_CREATE],
1795 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1796 oti->oti_declare_ops_used[OSD_OT_CREATE],
1797 oti->oti_declare_ops[OSD_OT_DESTROY],
1798 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1799 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1800 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1801 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1802 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1803 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1804 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1805 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1806 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1807 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1808 oti->oti_declare_ops[OSD_OT_WRITE],
1809 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1810 oti->oti_declare_ops_used[OSD_OT_WRITE],
1811 oti->oti_declare_ops[OSD_OT_PUNCH],
1812 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1813 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1814 oti->oti_declare_ops[OSD_OT_QUOTA],
1815 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1816 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1817 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1818 oti->oti_declare_ops[OSD_OT_INSERT],
1819 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1820 oti->oti_declare_ops_used[OSD_OT_INSERT],
1821 oti->oti_declare_ops[OSD_OT_DELETE],
1822 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1823 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1824 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1825 oti->oti_declare_ops[OSD_OT_REF_ADD],
1826 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1827 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1828 oti->oti_declare_ops[OSD_OT_REF_DEL],
1829 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1830 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1834 * Concurrency: shouldn't matter.
1836 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1839 struct osd_thread_info *oti = osd_oti_get(env);
1840 struct osd_device *dev = osd_dt_dev(d);
1842 struct osd_thandle *oh;
1847 LASSERT(current->journal_info == NULL);
1849 oh = container_of0(th, struct osd_thandle, ot_super);
1850 LASSERT(oh != NULL);
1851 LASSERT(oh->ot_handle == NULL);
1853 rc = dt_txn_hook_start(env, d, th);
1857 if (unlikely(osd_param_is_not_sane(dev, th))) {
1858 static unsigned long last_printed;
1859 static int last_credits;
1862 * don't make noise on a tiny testing systems
1863 * actual credits misuse will be caught anyway
1865 if (last_credits != oh->ot_credits &&
1866 time_after(jiffies, last_printed +
1867 msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1868 osd_transaction_size(dev) > 512) {
1869 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1870 oh->ot_credits, osd_transaction_size(dev));
1871 osd_trans_dump_creds(env, th);
1872 libcfs_debug_dumpstack(NULL);
1873 last_credits = oh->ot_credits;
1874 last_printed = jiffies;
1877 * XXX Limit the credits to 'max_transaction_buffers', and
1878 * let the underlying filesystem to catch the error if
1879 * we really need so many credits.
1881 * This should be removed when we can calculate the
1882 * credits precisely.
1884 oh->ot_credits = osd_transaction_size(dev);
1885 } else if (ldiskfs_track_declares_assert != 0) {
1887 * reserve few credits to prevent an assertion in JBD
1888 * our debugging mechanism will be able to detected
1889 * overuse. this can help to debug single-update
1892 oh->ot_credits += 10;
1893 if (unlikely(osd_param_is_not_sane(dev, th)))
1894 oh->ot_credits = osd_transaction_size(dev);
1897 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1898 GOTO(out, rc = -EIO);
1901 * XXX temporary stuff. Some abstraction layer should
1904 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1908 LASSERT(oti->oti_txns == 0);
1910 lu_device_get(&d->dd_lu_dev);
1911 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1922 static int osd_seq_exists(const struct lu_env *env,
1923 struct osd_device *osd, u64 seq)
1925 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1926 struct seq_server_site *ss = osd_seq_site(osd);
1931 LASSERT(ss != NULL);
1932 LASSERT(ss->ss_server_fld != NULL);
1934 rc = osd_fld_lookup(env, osd, seq, range);
1937 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1938 osd_name(osd), seq, rc);
1942 RETURN(ss->ss_node_id == range->lsr_index);
1945 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1947 struct dt_txn_commit_cb *dcb;
1948 struct dt_txn_commit_cb *tmp;
1950 /* call per-transaction stop callbacks if any */
1951 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1953 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1954 "commit callback entry: magic=%x name='%s'\n",
1955 dcb->dcb_magic, dcb->dcb_name);
1956 list_del_init(&dcb->dcb_linkage);
1957 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1962 * Concurrency: shouldn't matter.
1964 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1967 struct osd_thread_info *oti = osd_oti_get(env);
1968 struct osd_thandle *oh;
1969 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1970 struct osd_device *osd = osd_dt_dev(th->th_dev);
1971 struct qsd_instance *qsd = osd->od_quota_slave;
1972 struct lquota_trans *qtrans;
1973 struct list_head truncates = LIST_HEAD_INIT(truncates);
1974 int rc = 0, remove_agents = 0;
1978 oh = container_of0(th, struct osd_thandle, ot_super);
1980 remove_agents = oh->ot_remove_agents;
1982 qtrans = oh->ot_quota_trans;
1983 oh->ot_quota_trans = NULL;
1985 /* move locks to local list, stop tx, execute truncates */
1986 list_splice(&oh->ot_trunc_locks, &truncates);
1988 if (oh->ot_handle != NULL) {
1991 handle_t *hdl = oh->ot_handle;
1994 * add commit callback
1995 * notice we don't do this in osd_trans_start()
1996 * as underlying transaction can change during truncate
1998 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2001 LASSERT(oti->oti_txns == 1);
2004 rc = dt_txn_hook_stop(env, th);
2006 CERROR("%s: failed in transaction hook: rc = %d\n",
2009 osd_trans_stop_cb(oh, rc);
2010 /* hook functions might modify th_sync */
2011 hdl->h_sync = th->th_sync;
2013 oh->ot_handle = NULL;
2014 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2016 CERROR("%s: failed to stop transaction: rc = %d\n",
2017 osd_name(osd), rc2);
2021 osd_process_truncates(&truncates);
2023 osd_trans_stop_cb(oh, th->th_result);
2027 osd_trunc_unlock_all(&truncates);
2029 /* inform the quota slave device that the transaction is stopping */
2030 qsd_op_end(env, qsd, qtrans);
2033 * as we want IO to journal and data IO be concurrent, we don't block
2034 * awaiting data IO completion in osd_do_bio(), instead we wait here
2035 * once transaction is submitted to the journal. all reqular requests
2036 * don't do direct IO (except read/write), thus this wait_event becomes
2039 * IMPORTANT: we have to wait till any IO submited by the thread is
2040 * completed otherwise iobuf may be corrupted by different request
2042 wait_event(iobuf->dr_wait,
2043 atomic_read(&iobuf->dr_numreqs) == 0);
2044 osd_fini_iobuf(osd, iobuf);
2046 rc = iobuf->dr_error;
2048 if (unlikely(remove_agents != 0))
2049 osd_process_scheduled_agent_removals(env, osd);
2051 oti->oti_ins_cache_depth--;
2052 /* reset OI cache for safety */
2053 if (oti->oti_ins_cache_depth == 0)
2054 oti->oti_ins_cache_used = 0;
2056 sb_end_write(osd_sb(osd));
2061 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2063 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
2066 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2067 LASSERT(&dcb->dcb_func != NULL);
2068 if (dcb->dcb_flags & DCB_TRANS_STOP)
2069 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2071 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2077 * Called just before object is freed. Releases all resources except for
2078 * object itself (that is released by osd_object_free()).
2080 * Concurrency: no concurrent access is possible that late in object
2083 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2085 struct osd_object *obj = osd_obj(l);
2086 struct inode *inode = obj->oo_inode;
2088 LINVRNT(osd_invariant(obj));
2091 * If object is unlinked remove fid->ino mapping from object index.
2094 osd_index_fini(obj);
2095 if (inode != NULL) {
2096 struct qsd_instance *qsd = osd_obj2dev(obj)->od_quota_slave;
2097 qid_t uid = i_uid_read(inode);
2098 qid_t gid = i_gid_read(inode);
2100 obj->oo_inode = NULL;
2102 if (!obj->oo_header && qsd) {
2103 struct osd_thread_info *info = osd_oti_get(env);
2104 struct lquota_id_info *qi = &info->oti_qi;
2106 /* Release granted quota to master if necessary */
2107 qi->lqi_id.qid_uid = uid;
2108 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2110 qi->lqi_id.qid_uid = gid;
2111 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2113 qi->lqi_id.qid_uid = i_projid_read(inode);
2114 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2120 * Concurrency: ->loo_object_release() is called under site spin-lock.
2122 static void osd_object_release(const struct lu_env *env,
2123 struct lu_object *l)
2125 struct osd_object *o = osd_obj(l);
2128 * nobody should be releasing a non-destroyed object with nlink=0
2129 * the API allows this, but ldiskfs doesn't like and then report
2130 * this inode as deleted
2132 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2133 o->oo_inode->i_nlink == 0));
2137 * Concurrency: shouldn't matter.
2139 static int osd_object_print(const struct lu_env *env, void *cookie,
2140 lu_printer_t p, const struct lu_object *l)
2142 struct osd_object *o = osd_obj(l);
2143 struct iam_descr *d;
2145 if (o->oo_dir != NULL)
2146 d = o->oo_dir->od_container.ic_descr;
2149 return (*p)(env, cookie,
2150 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2152 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2153 o->oo_inode ? o->oo_inode->i_generation : 0,
2154 d ? d->id_ops->id_name : "plain");
2158 * Concurrency: shouldn't matter.
2160 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2161 struct obd_statfs *sfs)
2163 struct osd_device *osd = osd_dt_dev(d);
2164 struct super_block *sb = osd_sb(osd);
2165 struct kstatfs *ksfs;
2169 if (unlikely(osd->od_mnt == NULL))
2170 return -EINPROGRESS;
2172 /* osd_lproc.c call this without env, allocate ksfs for that case */
2173 if (unlikely(env == NULL)) {
2174 OBD_ALLOC_PTR(ksfs);
2178 ksfs = &osd_oti_get(env)->oti_ksfs;
2181 result = sb->s_op->statfs(sb->s_root, ksfs);
2185 statfs_pack(sfs, ksfs);
2186 if (unlikely(sb->s_flags & MS_RDONLY))
2187 sfs->os_state |= OS_STATE_READONLY;
2188 if (ldiskfs_has_feature_extents(sb))
2189 sfs->os_maxbytes = sb->s_maxbytes;
2191 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2194 * Reserve some space so to avoid fragmenting the filesystem too much.
2195 * Fragmentation not only impacts performance, but can also increase
2196 * metadata overhead significantly, causing grant calculation to be
2199 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2201 CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2202 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2203 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2204 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2206 sfs->os_blocks -= reserved;
2207 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2208 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2211 if (unlikely(env == NULL))
2217 * Estimate space needed for file creations. We assume the largest filename
2218 * which is 2^64 - 1, hence a filename of 20 chars.
2219 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2221 #ifdef __LDISKFS_DIR_REC_LEN
2222 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2224 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2228 * Concurrency: doesn't access mutable data.
2230 static void osd_conf_get(const struct lu_env *env,
2231 const struct dt_device *dev,
2232 struct dt_device_param *param)
2234 struct osd_device *d = osd_dt_dev(dev);
2235 struct super_block *sb = osd_sb(d);
2236 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2241 * XXX should be taken from not-yet-existing fs abstraction layer.
2243 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2244 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2245 param->ddp_symlink_max = sb->s_blocksize;
2246 param->ddp_mount_type = LDD_MT_LDISKFS;
2247 if (ldiskfs_has_feature_extents(sb))
2248 param->ddp_maxbytes = sb->s_maxbytes;
2250 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2252 * inode are statically allocated, so per-inode space consumption
2253 * is the space consumed by the directory entry
2255 param->ddp_inodespace = PER_OBJ_USAGE;
2257 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2258 * = 128MB) which is unlikely to be hit in real life. Report a smaller
2259 * maximum length to not under count the actual number of extents
2260 * needed for writing a file.
2262 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2263 /* worst-case extent insertion metadata overhead */
2264 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2265 param->ddp_mntopts = 0;
2266 if (test_opt(sb, XATTR_USER))
2267 param->ddp_mntopts |= MNTOPT_USERXATTR;
2268 if (test_opt(sb, POSIX_ACL))
2269 param->ddp_mntopts |= MNTOPT_ACL;
2272 * LOD might calculate the max stripe count based on max_ea_size,
2273 * so we need take account in the overhead as well,
2274 * xattr_header + magic + xattr_entry_head
2276 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2277 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2279 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2280 if (ldiskfs_has_feature_ea_inode(sb))
2281 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2285 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2287 if (param->ddp_max_ea_size > OSD_MAX_EA_SIZE)
2288 param->ddp_max_ea_size = OSD_MAX_EA_SIZE;
2291 * Preferred RPC size for efficient disk IO. 4MB shows good
2292 * all-around performance for ldiskfs, but use bigalloc chunk size
2293 * by default if larger.
2295 #if defined(LDISKFS_CLUSTER_SIZE)
2296 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2297 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2300 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2302 param->ddp_t10_cksum_type = 0;
2304 unsigned short interval = blk_integrity_interval(bi);
2305 name = blk_integrity_name(bi);
2313 if (strncmp(name, "T10-DIF-TYPE",
2314 sizeof("T10-DIF-TYPE") - 1) == 0) {
2315 /* also skip "1/3-" at end */
2316 const int type_off = sizeof("T10-DIF-TYPE.");
2318 if (interval != 512 && interval != 4096)
2319 CERROR("%s: unsupported T10PI sector size %u\n",
2320 d->od_svname, interval);
2321 else if (strcmp(name + type_off, "CRC") == 0)
2322 param->ddp_t10_cksum_type = interval == 512 ?
2323 OBD_CKSUM_T10CRC512 :
2325 else if (strcmp(name + type_off, "IP") == 0)
2326 param->ddp_t10_cksum_type = interval == 512 ?
2327 OBD_CKSUM_T10IP512 :
2330 CERROR("%s: unsupported checksum type of "
2332 d->od_svname, name);
2334 CERROR("%s: unsupported T10PI type '%s'",
2335 d->od_svname, name);
2341 * Concurrency: shouldn't matter.
2343 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2346 struct super_block *s = osd_sb(osd_dt_dev(d));
2349 down_read(&s->s_umount);
2350 rc = s->s_op->sync_fs(s, 1);
2351 up_read(&s->s_umount);
2353 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2360 * Start commit for OSD device.
2362 * An implementation of dt_commit_async method for OSD device.
2363 * Asychronously starts underlayng fs sync and thereby a transaction
2366 * \param env environment
2367 * \param d dt device
2369 * \see dt_device_operations
2371 static int osd_commit_async(const struct lu_env *env,
2372 struct dt_device *d)
2374 struct super_block *s = osd_sb(osd_dt_dev(d));
2379 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2380 down_read(&s->s_umount);
2381 rc = s->s_op->sync_fs(s, 0);
2382 up_read(&s->s_umount);
2387 /* Our own copy of the set readonly functions if present, or NU if not. */
2388 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2389 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2390 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2393 * Concurrency: shouldn't matter.
2395 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2397 struct super_block *sb = osd_sb(osd_dt_dev(d));
2398 struct block_device *dev = sb->s_bdev;
2399 int rc = -EOPNOTSUPP;
2403 if (priv_dev_set_rdonly) {
2404 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2407 CERROR("*** setting %s read-only ***\n",
2408 osd_dt_dev(d)->od_svname);
2410 if (sb->s_op->freeze_fs) {
2411 rc = sb->s_op->freeze_fs(sb);
2416 if (jdev && (jdev != dev)) {
2417 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2419 priv_dev_set_rdonly(jdev);
2421 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2422 priv_dev_set_rdonly(dev);
2424 if (sb->s_op->unfreeze_fs)
2425 sb->s_op->unfreeze_fs(sb);
2430 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2431 osd_dt_dev(d)->od_svname, (long)dev, rc);
2437 * Note: we do not count into QUOTA here.
2438 * If we mount with --data_journal we may need more.
2440 const int osd_dto_credits_noquota[DTO_NR] = {
2443 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2444 * SINGLEDATA_TRANS_BLOCKS(8)
2445 * XXX Note: maybe iam need more, since iam have more level than
2448 [DTO_INDEX_INSERT] = 16,
2451 * just modify a single entry, probably merge few within a block
2453 [DTO_INDEX_DELETE] = 1,
2457 [DTO_INDEX_UPDATE] = 16,
2459 * 4(inode, inode bits, groups, GDT)
2460 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2462 [DTO_OBJECT_CREATE] = 4,
2464 * 4(inode, inode bits, groups, GDT)
2465 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2467 [DTO_OBJECT_DELETE] = 4,
2469 * Attr set credits (inode)
2471 [DTO_ATTR_SET_BASE] = 1,
2473 * Xattr set. The same as xattr of EXT3.
2474 * DATA_TRANS_BLOCKS(14)
2475 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2476 * are also counted in. Do not know why?
2478 [DTO_XATTR_SET] = 14,
2480 * credits for inode change during write.
2482 [DTO_WRITE_BASE] = 3,
2484 * credits for single block write.
2486 [DTO_WRITE_BLOCK] = 14,
2488 * Attr set credits for chown.
2489 * This is extra credits for setattr, and it is null without quota
2491 [DTO_ATTR_SET_CHOWN] = 0
2494 static const struct dt_device_operations osd_dt_ops = {
2495 .dt_root_get = osd_root_get,
2496 .dt_statfs = osd_statfs,
2497 .dt_trans_create = osd_trans_create,
2498 .dt_trans_start = osd_trans_start,
2499 .dt_trans_stop = osd_trans_stop,
2500 .dt_trans_cb_add = osd_trans_cb_add,
2501 .dt_conf_get = osd_conf_get,
2502 .dt_sync = osd_sync,
2504 .dt_commit_async = osd_commit_async,
2507 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2510 struct osd_object *obj = osd_dt_obj(dt);
2511 struct osd_thread_info *oti = osd_oti_get(env);
2513 LINVRNT(osd_invariant(obj));
2515 LASSERT(obj->oo_owner != env);
2516 down_read_nested(&obj->oo_sem, role);
2518 LASSERT(obj->oo_owner == NULL);
2522 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2525 struct osd_object *obj = osd_dt_obj(dt);
2526 struct osd_thread_info *oti = osd_oti_get(env);
2528 LINVRNT(osd_invariant(obj));
2530 LASSERT(obj->oo_owner != env);
2531 down_write_nested(&obj->oo_sem, role);
2533 LASSERT(obj->oo_owner == NULL);
2534 obj->oo_owner = env;
2538 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2540 struct osd_object *obj = osd_dt_obj(dt);
2541 struct osd_thread_info *oti = osd_oti_get(env);
2543 LINVRNT(osd_invariant(obj));
2545 LASSERT(oti->oti_r_locks > 0);
2547 up_read(&obj->oo_sem);
2550 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2552 struct osd_object *obj = osd_dt_obj(dt);
2553 struct osd_thread_info *oti = osd_oti_get(env);
2555 LINVRNT(osd_invariant(obj));
2557 LASSERT(obj->oo_owner == env);
2558 LASSERT(oti->oti_w_locks > 0);
2560 obj->oo_owner = NULL;
2561 up_write(&obj->oo_sem);
2564 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2566 struct osd_object *obj = osd_dt_obj(dt);
2568 LINVRNT(osd_invariant(obj));
2570 return obj->oo_owner == env;
2573 static struct timespec *osd_inode_time(const struct lu_env *env,
2574 struct inode *inode, __u64 seconds)
2576 struct osd_thread_info *oti = osd_oti_get(env);
2577 struct timespec *t = &oti->oti_time;
2579 t->tv_sec = seconds;
2581 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2585 static void osd_inode_getattr(const struct lu_env *env,
2586 struct inode *inode, struct lu_attr *attr)
2588 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2589 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2590 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2591 LA_BLKSIZE | LA_TYPE;
2593 attr->la_atime = LTIME_S(inode->i_atime);
2594 attr->la_mtime = LTIME_S(inode->i_mtime);
2595 attr->la_ctime = LTIME_S(inode->i_ctime);
2596 attr->la_mode = inode->i_mode;
2597 attr->la_size = i_size_read(inode);
2598 attr->la_blocks = inode->i_blocks;
2599 attr->la_uid = i_uid_read(inode);
2600 attr->la_gid = i_gid_read(inode);
2601 attr->la_projid = i_projid_read(inode);
2602 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2603 attr->la_nlink = inode->i_nlink;
2604 attr->la_rdev = inode->i_rdev;
2605 attr->la_blksize = 1 << inode->i_blkbits;
2606 attr->la_blkbits = inode->i_blkbits;
2608 * Ext4 did not transfer inherit flags from raw inode
2609 * to inode flags, and ext4 internally test raw inode
2610 * @i_flags directly. Instead of patching ext4, we do it here.
2612 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2613 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2616 static int osd_attr_get(const struct lu_env *env,
2617 struct dt_object *dt,
2618 struct lu_attr *attr)
2620 struct osd_object *obj = osd_dt_obj(dt);
2622 if (unlikely(!dt_object_exists(dt)))
2624 if (unlikely(obj->oo_destroyed))
2627 LASSERT(!dt_object_remote(dt));
2628 LINVRNT(osd_invariant(obj));
2630 spin_lock(&obj->oo_guard);
2631 osd_inode_getattr(env, obj->oo_inode, attr);
2632 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2633 attr->la_flags |= LUSTRE_ORPHAN_FL;
2634 spin_unlock(&obj->oo_guard);
2639 static int osd_declare_attr_qid(const struct lu_env *env,
2640 struct osd_object *obj,
2641 struct osd_thandle *oh, long long bspace,
2642 qid_t old_id, qid_t new_id, bool enforce,
2643 unsigned int type, bool ignore_edquot)
2646 struct osd_thread_info *info = osd_oti_get(env);
2647 struct lquota_id_info *qi = &info->oti_qi;
2649 qi->lqi_type = type;
2650 /* inode accounting */
2651 qi->lqi_is_blk = false;
2653 /* one more inode for the new id ... */
2654 qi->lqi_id.qid_uid = new_id;
2656 /* Reserve credits for the new id */
2657 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2658 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2663 /* and one less inode for the current id */
2664 qi->lqi_id.qid_uid = old_id;
2666 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2667 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2672 /* block accounting */
2673 qi->lqi_is_blk = true;
2675 /* more blocks for the new id ... */
2676 qi->lqi_id.qid_uid = new_id;
2677 qi->lqi_space = bspace;
2679 * Credits for the new uid has been reserved, re-use "obj"
2680 * to save credit reservation.
2682 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2683 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2688 /* and finally less blocks for the current uid */
2689 qi->lqi_id.qid_uid = old_id;
2690 qi->lqi_space = -bspace;
2691 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2692 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2698 static int osd_declare_attr_set(const struct lu_env *env,
2699 struct dt_object *dt,
2700 const struct lu_attr *attr,
2701 struct thandle *handle)
2703 struct osd_thandle *oh;
2704 struct osd_object *obj;
2713 LASSERT(dt != NULL);
2714 LASSERT(handle != NULL);
2716 obj = osd_dt_obj(dt);
2717 LASSERT(osd_invariant(obj));
2719 oh = container_of0(handle, struct osd_thandle, ot_super);
2720 LASSERT(oh->ot_handle == NULL);
2722 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2723 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2725 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2726 osd_dto_credits_noquota[DTO_XATTR_SET]);
2728 if (attr == NULL || obj->oo_inode == NULL)
2731 bspace = obj->oo_inode->i_blocks << 9;
2732 bspace = toqb(bspace);
2735 * Changing ownership is always preformed by super user, it should not
2736 * fail with EDQUOT unless required explicitly.
2738 * We still need to call the osd_declare_qid() to calculate the journal
2739 * credits for updating quota accounting files and to trigger quota
2740 * space adjustment once the operation is completed.
2742 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2743 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2746 CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2747 "(the quota space is %lld)\n",
2748 obj->oo_inode->i_sb->s_id, attr->la_uid,
2749 attr->la_gid, bspace);
2752 uid = i_uid_read(obj->oo_inode);
2753 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2754 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2755 attr->la_uid, enforce, USRQUOTA,
2760 gid = i_gid_read(obj->oo_inode);
2761 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2762 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2763 i_gid_read(obj->oo_inode),
2764 attr->la_gid, enforce, GRPQUOTA,
2770 #ifdef HAVE_PROJECT_QUOTA
2771 if (attr->la_valid & LA_PROJID) {
2772 __u32 projid = i_projid_read(obj->oo_inode);
2774 enforce = (attr->la_valid & LA_PROJID) &&
2775 (attr->la_projid != projid);
2776 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2777 (qid_t)projid, (qid_t)attr->la_projid,
2778 enforce, PRJQUOTA, true);
2786 static int osd_inode_setattr(const struct lu_env *env,
2787 struct inode *inode, const struct lu_attr *attr)
2789 __u64 bits = attr->la_valid;
2791 /* Only allow set size for regular file */
2792 if (!S_ISREG(inode->i_mode))
2793 bits &= ~(LA_SIZE | LA_BLOCKS);
2798 if (bits & LA_ATIME)
2799 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2800 if (bits & LA_CTIME)
2801 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2802 if (bits & LA_MTIME)
2803 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2804 if (bits & LA_SIZE) {
2805 spin_lock(&inode->i_lock);
2806 LDISKFS_I(inode)->i_disksize = attr->la_size;
2807 i_size_write(inode, attr->la_size);
2808 spin_unlock(&inode->i_lock);
2812 * OSD should not change "i_blocks" which is used by quota.
2813 * "i_blocks" should be changed by ldiskfs only.
2816 inode->i_mode = (inode->i_mode & S_IFMT) |
2817 (attr->la_mode & ~S_IFMT);
2819 i_uid_write(inode, attr->la_uid);
2821 i_gid_write(inode, attr->la_gid);
2822 if (bits & LA_PROJID)
2823 i_projid_write(inode, attr->la_projid);
2824 if (bits & LA_NLINK)
2825 set_nlink(inode, attr->la_nlink);
2827 inode->i_rdev = attr->la_rdev;
2829 if (bits & LA_FLAGS) {
2830 /* always keep S_NOCMTIME */
2831 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2834 * Ext4 did not transfer inherit flags from
2835 * @inode->i_flags to raw inode i_flags when writing
2836 * flags, we do it explictly here.
2838 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2839 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2841 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2846 #ifdef HAVE_PROJECT_QUOTA
2847 static int osd_transfer_project(struct inode *inode, __u32 projid)
2849 struct super_block *sb = inode->i_sb;
2850 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2853 struct ldiskfs_iloc iloc;
2854 struct ldiskfs_inode *raw_inode;
2855 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2857 if (!ldiskfs_has_feature_project(sb)) {
2858 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2859 == LDISKFS_DEF_PROJID);
2860 if (projid != LDISKFS_DEF_PROJID)
2866 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2869 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2870 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2873 err = ldiskfs_get_inode_loc(inode, &iloc);
2877 raw_inode = ldiskfs_raw_inode(&iloc);
2878 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2885 dquot_initialize(inode);
2886 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2887 if (transfer_to[PRJQUOTA]) {
2888 err = __dquot_transfer(inode, transfer_to);
2889 dqput(transfer_to[PRJQUOTA]);
2898 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2902 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2903 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2906 ll_vfs_dq_init(inode);
2908 if (attr->la_valid & LA_UID)
2909 iattr.ia_valid |= ATTR_UID;
2910 if (attr->la_valid & LA_GID)
2911 iattr.ia_valid |= ATTR_GID;
2912 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2913 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2915 rc = ll_vfs_dq_transfer(inode, &iattr);
2917 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2918 "enforcement enabled on the ldiskfs "
2919 "filesystem?\n", inode->i_sb->s_id, rc);
2924 /* Handle project id transfer here properly */
2925 if (attr->la_valid & LA_PROJID &&
2926 attr->la_projid != i_projid_read(inode)) {
2927 #ifdef HAVE_PROJECT_QUOTA
2928 rc = osd_transfer_project(inode, attr->la_projid);
2933 CERROR("%s: quota transfer failed: rc = %d. Is project "
2934 "enforcement enabled on the ldiskfs "
2935 "filesystem?\n", inode->i_sb->s_id, rc);
2942 static int osd_attr_set(const struct lu_env *env,
2943 struct dt_object *dt,
2944 const struct lu_attr *attr,
2945 struct thandle *handle)
2947 struct osd_object *obj = osd_dt_obj(dt);
2948 struct inode *inode;
2951 if (!dt_object_exists(dt))
2954 LASSERT(handle != NULL);
2955 LASSERT(!dt_object_remote(dt));
2956 LASSERT(osd_invariant(obj));
2958 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2960 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
2961 !osd_obj2dev(obj)->od_is_ost) {
2962 struct osd_thread_info *oti = osd_oti_get(env);
2963 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2964 struct lu_fid *fid1 = &oti->oti_fid;
2965 struct osd_inode_id *id = &oti->oti_id;
2966 struct iam_path_descr *ipd;
2967 struct iam_container *bag;
2968 struct osd_thandle *oh;
2971 fid_cpu_to_be(fid1, fid0);
2972 memset(id, 1, sizeof(*id));
2973 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2974 fid0)->oi_dir.od_container;
2975 ipd = osd_idx_ipd_get(env, bag);
2976 if (unlikely(ipd == NULL))
2979 oh = container_of0(handle, struct osd_thandle, ot_super);
2980 rc = iam_update(oh->ot_handle, bag,
2981 (const struct iam_key *)fid1,
2982 (const struct iam_rec *)id, ipd);
2983 osd_ipd_put(env, bag, ipd);
2984 return(rc > 0 ? 0 : rc);
2987 inode = obj->oo_inode;
2989 rc = osd_quota_transfer(inode, attr);
2993 spin_lock(&obj->oo_guard);
2994 rc = osd_inode_setattr(env, inode, attr);
2995 spin_unlock(&obj->oo_guard);
2999 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3001 if (!(attr->la_valid & LA_FLAGS))
3004 /* Let's check if there are extra flags need to be set into LMA */
3005 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3006 struct osd_thread_info *info = osd_oti_get(env);
3007 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3009 LASSERT(!obj->oo_pfid_in_lma);
3011 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3012 &info->oti_ost_attrs);
3016 lma->lma_incompat |=
3017 lustre_to_lma_flags(attr->la_flags);
3018 lustre_lma_swab(lma);
3019 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3020 lma, sizeof(*lma), XATTR_REPLACE);
3022 struct osd_device *osd = osd_obj2dev(obj);
3024 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3025 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3026 lma->lma_incompat, rc);
3029 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3031 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3034 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3039 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3040 struct osd_object *obj,
3041 const char *name, const int namelen)
3043 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3046 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3047 umode_t mode, struct dt_allocation_hint *hint,
3051 struct osd_device *osd = osd_obj2dev(obj);
3052 struct osd_thandle *oth;
3053 struct dt_object *parent = NULL;
3054 struct inode *inode;
3056 LINVRNT(osd_invariant(obj));
3057 LASSERT(obj->oo_inode == NULL);
3058 LASSERT(obj->oo_hl_head == NULL);
3060 if (S_ISDIR(mode) && ldiskfs_pdo) {
3062 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3063 if (obj->oo_hl_head == NULL)
3067 oth = container_of(th, struct osd_thandle, ot_super);
3068 LASSERT(oth->ot_handle->h_transaction != NULL);
3070 if (hint != NULL && hint->dah_parent != NULL &&
3071 !dt_object_remote(hint->dah_parent))
3072 parent = hint->dah_parent;
3074 inode = ldiskfs_create_inode(oth->ot_handle,
3075 parent ? osd_dt_obj(parent)->oo_inode :
3076 osd_sb(osd)->s_root->d_inode,
3078 if (!IS_ERR(inode)) {
3079 /* Do not update file c/mtime in ldiskfs. */
3080 inode->i_flags |= S_NOCMTIME;
3083 * For new created object, it must be consistent,
3084 * and it is unnecessary to scrub against it.
3086 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3088 obj->oo_inode = inode;
3091 if (obj->oo_hl_head != NULL) {
3092 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3093 obj->oo_hl_head = NULL;
3095 result = PTR_ERR(inode);
3097 LINVRNT(osd_invariant(obj));
3105 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3106 struct lu_attr *attr,
3107 struct dt_allocation_hint *hint,
3108 struct dt_object_format *dof,
3112 struct osd_thandle *oth;
3113 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3115 LASSERT(S_ISDIR(attr->la_mode));
3117 oth = container_of(th, struct osd_thandle, ot_super);
3118 LASSERT(oth->ot_handle->h_transaction != NULL);
3119 result = osd_mkfile(info, obj, mode, hint, th);
3124 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3125 struct lu_attr *attr,
3126 struct dt_allocation_hint *hint,
3127 struct dt_object_format *dof,
3131 struct osd_thandle *oth;
3132 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3134 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3136 LASSERT(S_ISREG(attr->la_mode));
3138 oth = container_of(th, struct osd_thandle, ot_super);
3139 LASSERT(oth->ot_handle->h_transaction != NULL);
3141 result = osd_mkfile(info, obj, mode, hint, th);
3143 LASSERT(obj->oo_inode != NULL);
3144 if (feat->dif_flags & DT_IND_VARKEY)
3145 result = iam_lvar_create(obj->oo_inode,
3146 feat->dif_keysize_max,
3148 feat->dif_recsize_max,
3151 result = iam_lfix_create(obj->oo_inode,
3152 feat->dif_keysize_max,
3154 feat->dif_recsize_max,
3160 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3161 struct lu_attr *attr,
3162 struct dt_allocation_hint *hint,
3163 struct dt_object_format *dof,
3166 LASSERT(S_ISREG(attr->la_mode));
3167 return osd_mkfile(info, obj, (attr->la_mode &
3168 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
3171 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3172 struct lu_attr *attr,
3173 struct dt_allocation_hint *hint,
3174 struct dt_object_format *dof,
3177 LASSERT(S_ISLNK(attr->la_mode));
3178 return osd_mkfile(info, obj, (attr->la_mode &
3179 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
3182 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3183 struct lu_attr *attr,
3184 struct dt_allocation_hint *hint,
3185 struct dt_object_format *dof,
3188 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3191 LINVRNT(osd_invariant(obj));
3192 LASSERT(obj->oo_inode == NULL);
3193 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3194 S_ISFIFO(mode) || S_ISSOCK(mode));
3196 result = osd_mkfile(info, obj, mode, hint, th);
3198 LASSERT(obj->oo_inode != NULL);
3200 * This inode should be marked dirty for i_rdev. Currently
3201 * that is done in the osd_attr_init().
3203 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3206 LINVRNT(osd_invariant(obj));
3210 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3212 struct dt_allocation_hint *hint,
3213 struct dt_object_format *dof,
3216 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3218 osd_obj_type_f result;
3234 result = osd_mk_index;
3245 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3246 struct dt_object *parent, struct dt_object *child,
3251 ah->dah_parent = parent;
3252 ah->dah_mode = child_mode;
3254 if (parent != NULL && !dt_object_remote(parent)) {
3255 /* will help to find FID->ino at dt_insert("..") */
3256 struct osd_object *pobj = osd_dt_obj(parent);
3258 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3262 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3263 struct lu_attr *attr, struct dt_object_format *dof)
3265 struct inode *inode = obj->oo_inode;
3266 __u64 valid = attr->la_valid;
3269 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3271 if (dof->dof_type != DFT_NODE)
3272 attr->la_valid &= ~LA_RDEV;
3273 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
3274 attr->la_valid &= ~LA_ATIME;
3275 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
3276 attr->la_valid &= ~LA_CTIME;
3277 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
3278 attr->la_valid &= ~LA_MTIME;
3280 result = osd_quota_transfer(inode, attr);
3284 if (attr->la_valid != 0) {
3285 result = osd_inode_setattr(info->oti_env, inode, attr);
3287 * The osd_inode_setattr() should always succeed here. The
3288 * only error that could be returned is EDQUOT when we are
3289 * trying to change the UID or GID of the inode. However, this
3290 * should not happen since quota enforcement is no longer
3291 * enabled on ldiskfs (lquota takes care of it).
3293 LASSERTF(result == 0, "%d\n", result);
3294 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3297 attr->la_valid = valid;
3301 * Helper function for osd_create()
3303 * \retval 0, on success
3305 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3306 struct lu_attr *attr, struct dt_allocation_hint *hint,
3307 struct dt_object_format *dof, struct thandle *th)
3312 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3314 /* we drop umask so that permissions we pass are not affected */
3315 umask = current->fs->umask;
3316 current->fs->umask = 0;
3318 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3320 if (likely(obj->oo_inode != NULL)) {
3321 LASSERT(obj->oo_inode->i_state & I_NEW);
3324 * Unlock the inode before attr initialization to avoid
3325 * unnecessary dqget operations. LU-6378
3327 unlock_new_inode(obj->oo_inode);
3330 if (likely(result == 0)) {
3331 osd_attr_init(info, obj, attr, dof);
3332 osd_object_init0(obj);
3335 /* restore previous umask value */
3336 current->fs->umask = umask;
3338 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3344 * Helper function for osd_create()
3346 * \retval 0, on success
3348 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3349 const struct lu_fid *fid, struct thandle *th)
3351 struct osd_thread_info *info = osd_oti_get(env);
3352 struct osd_inode_id *id = &info->oti_id;
3353 struct osd_device *osd = osd_obj2dev(obj);
3354 struct osd_thandle *oh;
3357 LASSERT(obj->oo_inode != NULL);
3359 oh = container_of0(th, struct osd_thandle, ot_super);
3360 LASSERT(oh->ot_handle);
3361 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3363 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3364 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3365 OI_CHECK_FLD, NULL);
3366 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3371 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3372 u64 seq, struct lu_seq_range *range)
3374 struct seq_server_site *ss = osd_seq_site(osd);
3376 if (fid_seq_is_idif(seq)) {
3377 fld_range_set_ost(range);
3378 range->lsr_index = idif_ost_idx(seq);
3382 if (!fid_seq_in_fldb(seq)) {
3383 fld_range_set_mdt(range);
3386 * FIXME: If ss is NULL, it suppose not get lsr_index
3389 range->lsr_index = ss->ss_node_id;
3393 LASSERT(ss != NULL);
3394 fld_range_set_any(range);
3395 /* OSD will only do local fld lookup */
3396 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3399 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3400 struct lu_attr *attr,
3401 struct dt_allocation_hint *hint,
3402 struct dt_object_format *dof,
3403 struct thandle *handle)
3405 struct osd_thandle *oh;
3410 LASSERT(handle != NULL);
3412 oh = container_of0(handle, struct osd_thandle, ot_super);
3413 LASSERT(oh->ot_handle == NULL);
3416 * EA object consumes more credits than regular object: osd_mk_index
3417 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3418 * leaf_node, could involves the block, block bitmap, groups, GDT
3419 * change for each block, so add 4 * 2 credits in that case.
3421 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3422 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3423 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3425 * Reuse idle OI block may cause additional one OI block
3428 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3429 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3431 /* will help to find FID->ino mapping at dt_insert() */
3432 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3440 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3441 attr->la_projid, 1, oh, osd_dt_obj(dt),
3442 NULL, OSD_QID_INODE);
3450 * Called to destroy on-disk representation of the object
3452 * Concurrency: must be locked
3454 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3457 struct osd_object *obj = osd_dt_obj(dt);
3458 struct inode *inode = obj->oo_inode;
3459 struct osd_thandle *oh;
3467 oh = container_of0(th, struct osd_thandle, ot_super);
3468 LASSERT(oh->ot_handle == NULL);
3470 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3471 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3473 /* For removing agent entry */
3474 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3475 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3478 * Recycle idle OI leaf may cause additional three OI blocks
3481 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3482 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3483 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3484 /* one less inode */
3485 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3486 i_projid_read(inode), -1, oh, obj, NULL,
3490 /* data to be truncated */
3491 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3492 i_projid_read(inode), 0, oh, obj, NULL,
3498 * will help to find FID->ino when this object is being
3501 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3506 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3509 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3510 struct osd_object *obj = osd_dt_obj(dt);
3511 struct inode *inode = obj->oo_inode;
3512 struct osd_device *osd = osd_obj2dev(obj);
3513 struct osd_thandle *oh;
3518 oh = container_of0(th, struct osd_thandle, ot_super);
3519 LASSERT(oh->ot_handle);
3521 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3523 if (unlikely(fid_is_acct(fid)))
3526 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3527 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3529 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3530 osd_name(osd), PFID(fid), result);
3533 if (S_ISDIR(inode->i_mode)) {
3534 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3535 inode->i_nlink == 2);
3537 spin_lock(&obj->oo_guard);
3539 spin_unlock(&obj->oo_guard);
3540 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3543 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3545 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3547 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3548 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3549 oh->ot_handle, OI_CHECK_FLD);
3551 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3552 /* XXX: add to ext3 orphan list */
3553 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3555 /* not needed in the cache anymore */
3556 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3557 obj->oo_destroyed = 1;
3563 * Put the fid into lustre_mdt_attrs, and then place the structure
3564 * inode's ea. This fid should not be altered during the life time
3567 * \retval +ve, on success
3568 * \retval -ve, on error
3570 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3572 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3573 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3575 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3576 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3581 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3584 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3587 lustre_loa_init(loa, fid, compat, incompat);
3588 lustre_loa_swab(loa, false);
3591 * For the OST device with 256 bytes inode size by default,
3592 * the PFID EA will be stored together with LMA EA to avoid
3593 * performance trouble. Otherwise the PFID EA can be stored
3594 * independently. LU-8998
3596 if ((compat & LMAC_FID_ON_OST) &&
3597 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3598 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3599 sizeof(*loa), XATTR_CREATE);
3601 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3602 sizeof(*lma), XATTR_CREATE);
3604 * LMA may already exist, but we need to check that all the
3605 * desired compat/incompat flags have been added.
3607 if (unlikely(rc == -EEXIST)) {
3608 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3609 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3613 if (rc < sizeof(*lma))
3616 lustre_loa_swab(loa, true);
3617 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3618 ((compat == 0 && incompat == 0) ||
3619 (!(~lma->lma_compat & compat) &&
3620 !(~lma->lma_incompat & incompat))))
3623 lma->lma_self_fid = *fid;
3624 lma->lma_compat |= compat;
3625 lma->lma_incompat |= incompat;
3626 if (rc == sizeof(*lma)) {
3627 lustre_lma_swab(lma);
3628 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3629 sizeof(*lma), XATTR_REPLACE);
3631 lustre_loa_swab(loa, false);
3632 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3633 sizeof(*loa), XATTR_REPLACE);
3641 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3642 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3643 * To have compatilibility with 1.8 ldiskfs driver we need to have
3644 * magic number at start of fid data.
3645 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3648 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3649 const struct lu_fid *fid)
3651 if (!fid_is_namespace_visible(fid) ||
3652 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3653 param->edp_magic = 0;
3657 param->edp_magic = LDISKFS_LUFID_MAGIC;
3658 param->edp_len = sizeof(struct lu_fid) + 1;
3659 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3663 * Try to read the fid from inode ea into dt_rec.
3665 * \param fid object fid.
3667 * \retval 0 on success
3669 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3670 __u32 ino, struct lu_fid *fid,
3671 struct osd_inode_id *id)
3673 struct osd_thread_info *info = osd_oti_get(env);
3674 struct inode *inode;
3678 osd_id_gen(id, ino, OSD_OII_NOGEN);
3679 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3681 RETURN(PTR_ERR(inode));
3687 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3689 struct inode *parent_dir,
3690 const struct lu_fid *dot_fid,
3691 const struct lu_fid *dot_dot_fid,
3692 struct osd_thandle *oth)
3694 struct ldiskfs_dentry_param *dot_ldp;
3695 struct ldiskfs_dentry_param *dot_dot_ldp;
3696 __u32 saved_nlink = dir->i_nlink;
3699 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3700 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3702 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3703 dot_ldp->edp_magic = 0;
3705 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3706 dir, dot_ldp, dot_dot_ldp);
3708 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3709 * the subseqent ref_add() will increase the dir->i_nlink
3710 * as 3. That is incorrect for new created directory.
3712 * It looks like hack, because we want to make the OSD API
3713 * to be order-independent for new created directory object
3714 * between dt_insert(..) and ref_add() operations.
3716 * Here, we only restore the in-RAM dir-inode's nlink attr,
3717 * becuase if the nlink attr is not 2, then there will be
3718 * ref_add() called following the dt_insert(..), such call
3719 * will make both the in-RAM and on-disk dir-inode's nlink
3720 * attr to be set as 2. LU-7447
3722 set_nlink(dir, saved_nlink);
3727 * Create an local agent inode for remote entry
3729 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3730 struct osd_device *osd,
3731 struct osd_object *pobj,
3732 const struct lu_fid *fid,
3736 struct osd_thread_info *info = osd_oti_get(env);
3737 struct inode *local;
3738 struct osd_thandle *oh;
3744 oh = container_of(th, struct osd_thandle, ot_super);
3745 LASSERT(oh->ot_handle->h_transaction != NULL);
3747 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3748 if (IS_ERR(local)) {
3749 CERROR("%s: create local error %d\n", osd_name(osd),
3750 (int)PTR_ERR(local));
3755 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3756 * correct gid on remote file, not agent here
3758 local->i_gid = current_fsgid();
3759 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3760 unlock_new_inode(local);
3762 /* Agent inode should not have project ID */
3763 #ifdef HAVE_PROJECT_QUOTA
3764 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3765 i_projid_read(pobj->oo_inode) != 0) {
3766 rc = osd_transfer_project(local, 0);
3768 CERROR("%s: quota transfer failed: rc = %d. Is project "
3769 "quota enforcement enabled on the ldiskfs "
3770 "filesystem?\n", local->i_sb->s_id, rc);
3771 RETURN(ERR_PTR(rc));
3775 /* Set special LMA flag for local agent inode */
3776 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3778 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3779 osd_name(osd), PFID(fid), rc);
3780 RETURN(ERR_PTR(rc));
3786 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3787 lu_object_fid(&pobj->oo_dt.do_lu),
3790 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3791 osd_name(osd), PFID(fid), rc);
3792 RETURN(ERR_PTR(rc));
3799 * when direntry is deleted, we have to take care of possible agent inode
3800 * referenced by that. unfortunately we can't do this at that point:
3801 * iget() within a running transaction leads to deadlock and we better do
3802 * not call that every delete declaration to save performance. so we put
3803 * a potention agent inode on a list and process that once the transaction
3804 * is over. Notice it's not any worse in terms of real orphans as regular
3805 * object destroy doesn't put inodes on the on-disk orphan list. this should
3806 * be addressed separately
3808 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3809 struct osd_thandle *oh,
3812 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3813 struct osd_obj_orphan *oor;
3820 oor->oor_env = (struct lu_env *)env;
3821 spin_lock(&osd->od_osfs_lock);
3822 list_add(&oor->oor_list, &osd->od_orphan_list);
3823 spin_unlock(&osd->od_osfs_lock);
3825 oh->ot_remove_agents = 1;
3831 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3832 struct osd_device *osd)
3834 struct osd_thread_info *info = osd_oti_get(env);
3835 struct osd_obj_orphan *oor, *tmp;
3836 struct osd_inode_id id;
3837 struct list_head list;
3838 struct inode *inode;
3843 INIT_LIST_HEAD(&list);
3845 spin_lock(&osd->od_osfs_lock);
3846 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3847 if (oor->oor_env == env) {
3848 list_del(&oor->oor_list);
3849 list_add(&oor->oor_list, &list);
3852 spin_unlock(&osd->od_osfs_lock);
3854 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3858 list_del(&oor->oor_list);
3861 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3862 inode = osd_iget_fid(info, osd, &id, &fid);
3866 if (!osd_remote_fid(env, osd, &fid)) {
3871 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3873 mark_inode_dirty(inode);
3874 ldiskfs_journal_stop(jh);
3882 * OSD layer object create function for OST objects (b=11826).
3884 * The FID is inserted into inode xattr here.
3886 * \retval 0, on success
3887 * \retval -ve, on error
3889 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3890 struct lu_attr *attr, struct dt_allocation_hint *hint,
3891 struct dt_object_format *dof, struct thandle *th)
3893 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3894 struct osd_object *obj = osd_dt_obj(dt);
3895 struct osd_thread_info *info = osd_oti_get(env);
3896 int result, on_ost = 0;
3900 if (dt_object_exists(dt))
3903 LINVRNT(osd_invariant(obj));
3904 LASSERT(!dt_object_remote(dt));
3905 LASSERT(osd_is_write_locked(env, obj));
3906 LASSERT(th != NULL);
3908 if (unlikely(fid_is_acct(fid)))
3910 * Quota files can't be created from the kernel any more,
3911 * 'tune2fs -O quota' will take care of creating them
3915 result = __osd_create(info, obj, attr, hint, dof, th);
3917 if (fid_is_idif(fid) &&
3918 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3919 struct lu_fid *tfid = &info->oti_fid;
3920 struct ost_id *oi = &info->oti_ostid;
3922 fid_to_ostid(fid, oi);
3923 ostid_to_fid(tfid, oi, 0);
3925 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3926 LMAC_FID_ON_OST, 0);
3928 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3930 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3931 on_ost ? LMAC_FID_ON_OST : 0,
3934 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3935 obj->oo_dt.do_body_ops = &osd_body_ops;
3938 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY))
3939 result = __osd_oi_insert(env, obj, fid, th);
3942 * a small optimization - dt_insert() isn't usually applied
3943 * to OST objects, so we don't need to cache OI mapping for
3946 if (result == 0 && on_ost == 0) {
3947 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3949 result = osd_idc_find_and_init(env, osd, obj);
3950 LASSERT(result == 0);
3953 LASSERT(ergo(result == 0,
3954 dt_object_exists(dt) && !dt_object_remote(dt)));
3955 LINVRNT(osd_invariant(obj));
3959 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3960 struct thandle *handle)
3962 struct osd_thandle *oh;
3964 /* it's possible that object doesn't exist yet */
3965 LASSERT(handle != NULL);
3967 oh = container_of0(handle, struct osd_thandle, ot_super);
3968 LASSERT(oh->ot_handle == NULL);
3970 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3971 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3973 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
3979 * Concurrency: @dt is write locked.
3981 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3984 struct osd_object *obj = osd_dt_obj(dt);
3985 struct inode *inode = obj->oo_inode;
3986 struct osd_thandle *oh;
3989 if (!dt_object_exists(dt) || obj->oo_destroyed)
3992 LINVRNT(osd_invariant(obj));
3993 LASSERT(!dt_object_remote(dt));
3994 LASSERT(osd_is_write_locked(env, obj));
3995 LASSERT(th != NULL);
3997 oh = container_of0(th, struct osd_thandle, ot_super);
3998 LASSERT(oh->ot_handle != NULL);
4000 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4002 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4003 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4005 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4006 * (65000) subdirectories by storing "1" in i_nlink if the link count
4007 * would otherwise overflow. Directory tranversal tools understand
4008 * that (st_nlink == 1) indicates that the filesystem dose not track
4009 * hard links count on the directory, and will not abort subdirectory
4010 * scanning early once (st_nlink - 2) subdirs have been found.
4012 * This also has to properly handle the case of inodes with nlink == 0
4013 * in case they are being linked into the PENDING directory
4015 spin_lock(&obj->oo_guard);
4016 if (unlikely(inode->i_nlink == 0))
4017 /* inc_nlink from 0 may cause WARN_ON */
4018 set_nlink(inode, 1);
4020 ldiskfs_inc_count(oh->ot_handle, inode);
4021 if (!S_ISDIR(inode->i_mode))
4022 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4024 spin_unlock(&obj->oo_guard);
4026 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4027 LINVRNT(osd_invariant(obj));
4029 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4034 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4035 struct thandle *handle)
4037 struct osd_thandle *oh;
4039 if (!dt_object_exists(dt))
4042 LASSERT(!dt_object_remote(dt));
4043 LASSERT(handle != NULL);
4045 oh = container_of0(handle, struct osd_thandle, ot_super);
4046 LASSERT(oh->ot_handle == NULL);
4048 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4049 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4055 * Concurrency: @dt is write locked.
4057 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4060 struct osd_object *obj = osd_dt_obj(dt);
4061 struct inode *inode = obj->oo_inode;
4062 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4063 struct osd_thandle *oh;
4065 if (!dt_object_exists(dt))
4068 LINVRNT(osd_invariant(obj));
4069 LASSERT(!dt_object_remote(dt));
4070 LASSERT(osd_is_write_locked(env, obj));
4071 LASSERT(th != NULL);
4073 oh = container_of0(th, struct osd_thandle, ot_super);
4074 LASSERT(oh->ot_handle != NULL);
4076 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4078 spin_lock(&obj->oo_guard);
4080 * That can be result of upgrade from old Lustre version and
4081 * applied only to local files. Just skip this ref_del call.
4082 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4084 if (inode->i_nlink == 0) {
4085 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4086 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4087 ", maybe an upgraded file? (LU-3915)\n",
4088 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4089 spin_unlock(&obj->oo_guard);
4093 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4094 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4096 ldiskfs_dec_count(oh->ot_handle, inode);
4097 spin_unlock(&obj->oo_guard);
4099 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4100 LINVRNT(osd_invariant(obj));
4102 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4108 * Concurrency: @dt is read locked.
4110 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4111 struct lu_buf *buf, const char *name)
4113 struct osd_object *obj = osd_dt_obj(dt);
4114 struct inode *inode = obj->oo_inode;
4115 struct osd_thread_info *info = osd_oti_get(env);
4116 struct dentry *dentry = &info->oti_obj_dentry;
4117 bool cache_xattr = false;
4122 /* version get is not real XATTR but uses xattr API */
4123 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4124 dt_obj_version_t *ver = buf->lb_buf;
4127 * for version we are just using xattr API but change inode
4130 if (buf->lb_len == 0)
4131 return sizeof(dt_obj_version_t);
4133 if (buf->lb_len < sizeof(dt_obj_version_t))
4136 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4137 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4139 *ver = LDISKFS_I(inode)->i_fs_version;
4141 return sizeof(dt_obj_version_t);
4144 if (!dt_object_exists(dt))
4147 LASSERT(!dt_object_remote(dt));
4148 LASSERT(inode->i_op != NULL);
4149 #ifdef HAVE_IOP_XATTR
4150 LASSERT(inode->i_op->getxattr != NULL);
4153 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4154 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4158 rc = osd_oxc_get(obj, name, buf);
4163 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4164 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4165 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4166 struct filter_fid *ff;
4167 struct ost_layout *ol;
4169 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4173 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4176 if (buf->lb_len == 0 || !buf->lb_buf)
4179 if (buf->lb_len < rc)
4183 ol = &ff->ff_layout;
4184 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4185 PFID_STRIPE_IDX_BITS);
4186 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4187 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4188 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4189 if (lma->lma_compat & LMAC_COMP_INFO) {
4190 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4191 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4192 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4194 ol->ol_comp_start = 0;
4195 ol->ol_comp_end = 0;
4199 rc = __osd_xattr_get(inode, dentry, name,
4200 buf->lb_buf, buf->lb_len);
4204 if (rc == -ENOENT || rc == -ENODATA)
4205 osd_oxc_add(obj, name, NULL, 0);
4206 else if (rc > 0 && buf->lb_buf != NULL)
4207 osd_oxc_add(obj, name, buf->lb_buf, rc);
4213 static int osd_declare_xattr_set(const struct lu_env *env,
4214 struct dt_object *dt,
4215 const struct lu_buf *buf, const char *name,
4216 int fl, struct thandle *handle)
4218 struct osd_thandle *oh;
4220 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4222 LASSERT(handle != NULL);
4224 oh = container_of0(handle, struct osd_thandle, ot_super);
4225 LASSERT(oh->ot_handle == NULL);
4227 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4229 * For non-upgrading case, the LMA is set first and
4230 * usually fit inode. But for upgrade case, the LMA
4231 * may be in another separated EA block.
4233 if (dt_object_exists(dt)) {
4234 if (fl == LU_XATTR_REPLACE)
4239 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4241 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4242 /* We may need to delete the old PFID EA. */
4243 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4244 if (fl == LU_XATTR_REPLACE)
4250 * If some name entry resides on remote MDT, then will create
4251 * agent entry under remote parent. On the other hand, if the
4252 * remote entry will be removed, then related agent entry may
4253 * need to be removed from the remote parent. So there may be
4254 * kinds of cases, let's declare enough credits. The credits
4255 * for create agent entry is enough for remove case.
4257 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4258 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4259 if (dt_object_exists(dt))
4260 credits += 1; /* For updating LMA */
4264 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4269 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4271 * learn xattr size from osd_xattr_get if
4272 * attribute has not been read yet
4274 buflen = __osd_xattr_get(
4275 osd_dt_obj(dt)->oo_inode,
4276 &osd_oti_get(env)->oti_obj_dentry,
4281 buflen = buf->lb_len;
4284 if (buflen > sb->s_blocksize) {
4285 credits += osd_calc_bkmap_credits(
4287 (buflen + sb->s_blocksize - 1) >>
4288 sb->s_blocksize_bits);
4292 * xattr set may involve inode quota change, reserve credits for
4293 * dquot_initialize()
4295 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4298 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4303 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4304 const struct lu_buf *buf, int fl,
4305 struct thandle *handle)
4307 struct osd_thread_info *info = osd_oti_get(env);
4308 struct dentry *dentry = &info->oti_obj_dentry;
4309 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4310 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4311 struct inode *inode = obj->oo_inode;
4312 struct filter_fid *ff = buf->lb_buf;
4313 struct ost_layout *ol = &ff->ff_layout;
4314 int flags = XATTR_REPLACE;
4319 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4322 rc = osd_get_lma(info, inode, dentry, loa);
4323 if (rc == -ENODATA) {
4324 /* Usually for upgarding from old device */
4325 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4326 LMAC_FID_ON_OST, 0);
4327 flags = XATTR_CREATE;
4332 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4333 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4336 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4337 /* Separate PFID EA from LMA */
4338 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4339 lustre_lma_swab(lma);
4340 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4341 sizeof(*lma), XATTR_REPLACE);
4343 obj->oo_pfid_in_lma = 0;
4344 rc = LU_XATTR_CREATE;
4350 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4354 * Old client does not send stripe information,
4355 * then store the PFID EA on disk separatedly.
4357 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4358 ol->ol_stripe_size == 0))
4361 /* Remove old PFID EA entry firstly. */
4362 ll_vfs_dq_init(inode);
4363 rc = osd_removexattr(dentry, inode, XATTR_NAME_FID);
4364 if (rc == -ENODATA) {
4365 if ((fl & LU_XATTR_REPLACE) && !(fl & LU_XATTR_CREATE))
4372 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4373 if (likely(ol->ol_stripe_size != 0)) {
4374 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4375 PFID_STRIPE_IDX_BITS;
4376 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4377 lma->lma_compat |= LMAC_STRIPE_INFO;
4378 if (ol->ol_comp_id != 0) {
4379 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4380 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4381 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4382 lma->lma_compat |= LMAC_COMP_INFO;
4386 lustre_loa_swab(loa, false);
4388 /* Store the PFID EA inside LMA. */
4389 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4392 obj->oo_pfid_in_lma = 1;
4398 * In DNE environment, the object (in spite of regular file or directory)
4399 * and its name entry may reside on different MDTs. Under such case, we will
4400 * create an agent entry on the MDT where the object resides. The agent entry
4401 * references the object locally, that makes the object to be visible to the
4402 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4403 * such as 'tar' can handle the object properly.
4405 * We handle the agent entry during set linkEA that is the common interface
4406 * for both regular file and directroy, can handle kinds of cases, such as
4407 * create/link/unlink/rename, and so on.
4409 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4411 * XXX: There are two known issues:
4412 * 1. For one object, we will create at most one agent entry even if there
4413 * may be more than one cross-MDTs hard links on the object. So the local
4414 * e2fsck may claim that the object's nlink is larger than the name entries
4415 * that reference such inode. And in further, the e2fsck will fix the nlink
4416 * attribute to match the local references. Then it will cause the object's
4417 * nlink attribute to be inconsistent with the global references. it is bad
4418 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4419 * other hand, the global namespace LFSCK can repair the object's attribute
4420 * according to the linkEA.
4421 * 2. There may be too many hard links on the object as to its linkEA overflow,
4422 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4423 * case happened, then at this point, we do not know whether there are some
4424 * cross-MDTs reference. But there are local references, it guarantees that
4425 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4427 static int osd_xattr_handle_linkea(const struct lu_env *env,
4428 struct osd_device *osd,
4429 struct osd_object *obj,
4430 const struct lu_buf *buf,
4431 struct thandle *handle)
4433 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4434 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4435 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4436 struct lu_name tmpname;
4437 struct osd_thandle *oh;
4439 bool remote = false;
4443 oh = container_of0(handle, struct osd_thandle, ot_super);
4444 LASSERT(oh->ot_handle != NULL);
4446 rc = linkea_init_with_rec(&ldata);
4448 linkea_first_entry(&ldata);
4449 while (ldata.ld_lee != NULL && !remote) {
4450 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4452 if (osd_remote_fid(env, osd, tfid) > 0)
4455 linkea_next_entry(&ldata);
4457 } else if (rc == -ENODATA) {
4463 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4464 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4466 CERROR("%s: failed to remove agent entry for "DFID
4467 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4468 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4469 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4471 CERROR("%s: failed to create agent entry for "DFID
4472 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4479 * Concurrency: @dt is write locked.
4481 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4482 const struct lu_buf *buf, const char *name, int fl,
4483 struct thandle *handle)
4485 struct osd_object *obj = osd_dt_obj(dt);
4486 struct osd_device *osd = osd_obj2dev(obj);
4487 struct inode *inode = obj->oo_inode;
4488 struct osd_thread_info *info = osd_oti_get(env);
4498 /* version set is not real XATTR */
4499 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4500 dt_obj_version_t *version = buf->lb_buf;
4503 * for version we are just using xattr API but change inode
4506 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4508 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4509 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4511 LDISKFS_I(inode)->i_fs_version = *version;
4513 * Version is set after all inode operations are finished,
4514 * so we should mark it dirty here
4516 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4521 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4522 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4525 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4528 * For the OST device with 256 bytes inode size by default,
4529 * the PFID EA will be stored together with LMA EA to avoid
4530 * performance trouble. Otherwise the PFID EA can be stored
4531 * independently. LU-8998
4533 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4534 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4535 LASSERT(buf->lb_buf);
4537 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4540 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4541 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4542 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4544 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4548 lma->lma_incompat |= LMAI_STRIPED;
4549 lustre_lma_swab(lma);
4550 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4551 sizeof(*lma), XATTR_REPLACE);
4554 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4555 LASSERT(!osd->od_is_ost);
4557 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4562 if (fl & LU_XATTR_REPLACE)
4563 fs_flags |= XATTR_REPLACE;
4565 if (fl & LU_XATTR_CREATE)
4566 fs_flags |= XATTR_CREATE;
4568 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4569 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4572 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4573 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4574 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4580 * Concurrency: @dt is read locked.
4582 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4583 const struct lu_buf *buf)
4585 struct osd_object *obj = osd_dt_obj(dt);
4586 struct inode *inode = obj->oo_inode;
4587 struct osd_thread_info *info = osd_oti_get(env);
4588 struct dentry *dentry = &info->oti_obj_dentry;
4590 if (!dt_object_exists(dt))
4593 LASSERT(!dt_object_remote(dt));
4594 LASSERT(inode->i_op != NULL);
4595 LASSERT(inode->i_op->listxattr != NULL);
4597 dentry->d_inode = inode;
4598 dentry->d_sb = inode->i_sb;
4599 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4602 static int osd_declare_xattr_del(const struct lu_env *env,
4603 struct dt_object *dt, const char *name,
4604 struct thandle *handle)
4606 struct osd_thandle *oh;
4607 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4609 LASSERT(!dt_object_remote(dt));
4610 LASSERT(handle != NULL);
4612 oh = container_of0(handle, struct osd_thandle, ot_super);
4613 LASSERT(oh->ot_handle == NULL);
4615 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4616 osd_dto_credits_noquota[DTO_XATTR_SET]);
4618 * xattr del may involve inode quota change, reserve credits for
4619 * dquot_initialize()
4621 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4627 * Concurrency: @dt is write locked.
4629 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4630 const char *name, struct thandle *handle)
4632 struct osd_object *obj = osd_dt_obj(dt);
4633 struct inode *inode = obj->oo_inode;
4634 struct osd_thread_info *info = osd_oti_get(env);
4635 struct dentry *dentry = &info->oti_obj_dentry;
4638 if (!dt_object_exists(dt))
4641 LASSERT(!dt_object_remote(dt));
4642 LASSERT(inode->i_op != NULL);
4643 LASSERT(handle != NULL);
4644 #ifdef HAVE_IOP_XATTR
4645 LASSERT(inode->i_op->removexattr != NULL);
4648 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4650 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4651 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4653 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4654 &info->oti_ost_attrs);
4656 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4658 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4659 lustre_lma_swab(lma);
4660 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4661 sizeof(*lma), XATTR_REPLACE);
4663 obj->oo_pfid_in_lma = 0;
4666 ll_vfs_dq_init(inode);
4667 dentry->d_inode = inode;
4668 dentry->d_sb = inode->i_sb;
4669 rc = osd_removexattr(dentry, inode, name);
4672 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4675 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4676 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4677 osd_oxc_del(obj, name);
4682 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4683 __u64 start, __u64 end)
4685 struct osd_object *obj = osd_dt_obj(dt);
4686 struct inode *inode = obj->oo_inode;
4687 struct osd_thread_info *info = osd_oti_get(env);
4688 struct dentry *dentry = &info->oti_obj_dentry;
4689 struct file *file = &info->oti_file;
4694 dentry->d_inode = inode;
4695 dentry->d_sb = inode->i_sb;
4696 file->f_path.dentry = dentry;
4697 file->f_mapping = inode->i_mapping;
4698 file->f_op = inode->i_fop;
4699 set_file_inode(file, inode);
4701 rc = ll_vfs_fsync_range(file, start, end, 0);
4706 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4715 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4716 const struct dt_index_features *feat)
4718 struct iam_descr *descr;
4720 if (osd_object_is_root(o))
4721 return feat == &dt_directory_features;
4723 LASSERT(o->oo_dir != NULL);
4725 descr = o->oo_dir->od_container.ic_descr;
4726 if (feat == &dt_directory_features) {
4727 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4732 return feat->dif_keysize_min <= descr->id_key_size &&
4733 descr->id_key_size <= feat->dif_keysize_max &&
4734 feat->dif_recsize_min <= descr->id_rec_size &&
4735 descr->id_rec_size <= feat->dif_recsize_max &&
4736 !(feat->dif_flags & (DT_IND_VARKEY |
4737 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4738 ergo(feat->dif_flags & DT_IND_UPDATE,
4739 1 /* XXX check that object (and fs) is writable */);
4743 static int osd_iam_container_init(const struct lu_env *env,
4744 struct osd_object *obj,
4745 struct osd_directory *dir)
4747 struct iam_container *bag = &dir->od_container;
4750 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4754 result = iam_container_setup(bag);
4756 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4758 iam_container_fini(bag);
4765 * Concurrency: no external locking is necessary.
4767 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4768 const struct dt_index_features *feat)
4772 struct osd_object *obj = osd_dt_obj(dt);
4774 LINVRNT(osd_invariant(obj));
4776 if (osd_object_is_root(obj)) {
4777 dt->do_index_ops = &osd_index_ea_ops;
4779 } else if (feat == &dt_directory_features) {
4780 dt->do_index_ops = &osd_index_ea_ops;
4781 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4786 } else if (unlikely(feat == &dt_otable_features)) {
4787 dt->do_index_ops = &osd_otable_ops;
4789 } else if (unlikely(feat == &dt_acct_features)) {
4790 dt->do_index_ops = &osd_acct_index_ops;
4793 } else if (!osd_has_index(obj)) {
4794 struct osd_directory *dir;
4795 struct osd_device *osd = osd_obj2dev(obj);
4796 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4801 spin_lock(&obj->oo_guard);
4802 if (obj->oo_dir == NULL)
4806 * Concurrent thread allocated container data.
4809 spin_unlock(&obj->oo_guard);
4811 * Now, that we have container data, serialize its
4814 down_write(&obj->oo_ext_idx_sem);
4816 * recheck under lock.
4819 if (osd_has_index(obj)) {
4824 result = osd_iam_container_init(env, obj, obj->oo_dir);
4825 if (result || feat == &dt_lfsck_namespace_features ||
4826 feat == &dt_lfsck_layout_orphan_features ||
4827 feat == &dt_lfsck_layout_dangling_features)
4830 result = osd_index_register(osd, fid,
4831 feat->dif_keysize_max,
4832 feat->dif_recsize_max);
4834 CWARN("%s: failed to register index "
4836 osd_name(osd), PFID(fid), result);
4837 else if (result > 0)
4840 CDEBUG(D_LFSCK, "%s: index object "DFID
4841 " (%d/%d) registered\n",
4842 osd_name(osd), PFID(fid),
4843 (int)feat->dif_keysize_max,
4844 (int)feat->dif_recsize_max);
4847 up_write(&obj->oo_ext_idx_sem);
4855 if (result == 0 && skip_iam == 0) {
4856 if (!osd_iam_index_probe(env, obj, feat))
4859 LINVRNT(osd_invariant(obj));
4864 static int osd_otable_it_attr_get(const struct lu_env *env,
4865 struct dt_object *dt,
4866 struct lu_attr *attr)
4872 static const struct dt_object_operations osd_obj_ops = {
4873 .do_read_lock = osd_read_lock,
4874 .do_write_lock = osd_write_lock,
4875 .do_read_unlock = osd_read_unlock,
4876 .do_write_unlock = osd_write_unlock,
4877 .do_write_locked = osd_write_locked,
4878 .do_attr_get = osd_attr_get,
4879 .do_declare_attr_set = osd_declare_attr_set,
4880 .do_attr_set = osd_attr_set,
4881 .do_ah_init = osd_ah_init,
4882 .do_declare_create = osd_declare_create,
4883 .do_create = osd_create,
4884 .do_declare_destroy = osd_declare_destroy,
4885 .do_destroy = osd_destroy,
4886 .do_index_try = osd_index_try,
4887 .do_declare_ref_add = osd_declare_ref_add,
4888 .do_ref_add = osd_ref_add,
4889 .do_declare_ref_del = osd_declare_ref_del,
4890 .do_ref_del = osd_ref_del,
4891 .do_xattr_get = osd_xattr_get,
4892 .do_declare_xattr_set = osd_declare_xattr_set,
4893 .do_xattr_set = osd_xattr_set,
4894 .do_declare_xattr_del = osd_declare_xattr_del,
4895 .do_xattr_del = osd_xattr_del,
4896 .do_xattr_list = osd_xattr_list,
4897 .do_object_sync = osd_object_sync,
4898 .do_invalidate = osd_invalidate,
4901 static const struct dt_object_operations osd_obj_otable_it_ops = {
4902 .do_attr_get = osd_otable_it_attr_get,
4903 .do_index_try = osd_index_try,
4906 static int osd_index_declare_iam_delete(const struct lu_env *env,
4907 struct dt_object *dt,
4908 const struct dt_key *key,
4909 struct thandle *handle)
4911 struct osd_thandle *oh;
4913 oh = container_of0(handle, struct osd_thandle, ot_super);
4914 LASSERT(oh->ot_handle == NULL);
4916 /* Recycle may cause additional three blocks to be changed. */
4917 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4918 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4924 * delete a (key, value) pair from index \a dt specified by \a key
4926 * \param dt osd index object
4927 * \param key key for index
4928 * \param rec record reference
4929 * \param handle transaction handler
4932 * \retval -ve failure
4934 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4935 const struct dt_key *key,
4936 struct thandle *handle)
4938 struct osd_thread_info *oti = osd_oti_get(env);
4939 struct osd_object *obj = osd_dt_obj(dt);
4940 struct osd_thandle *oh;
4941 struct iam_path_descr *ipd;
4942 struct iam_container *bag = &obj->oo_dir->od_container;
4947 if (!dt_object_exists(dt))
4950 LINVRNT(osd_invariant(obj));
4951 LASSERT(!dt_object_remote(dt));
4952 LASSERT(bag->ic_object == obj->oo_inode);
4953 LASSERT(handle != NULL);
4955 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4957 ipd = osd_idx_ipd_get(env, bag);
4958 if (unlikely(ipd == NULL))
4961 oh = container_of0(handle, struct osd_thandle, ot_super);
4962 LASSERT(oh->ot_handle != NULL);
4963 LASSERT(oh->ot_handle->h_transaction != NULL);
4965 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4966 /* swab quota uid/gid provided by caller */
4967 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4968 key = (const struct dt_key *)&oti->oti_quota_id;
4971 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4972 osd_ipd_put(env, bag, ipd);
4973 LINVRNT(osd_invariant(obj));
4974 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4978 static int osd_index_declare_ea_delete(const struct lu_env *env,
4979 struct dt_object *dt,
4980 const struct dt_key *key,
4981 struct thandle *handle)
4983 struct osd_thandle *oh;
4984 struct inode *inode;
4989 LASSERT(!dt_object_remote(dt));
4990 LASSERT(handle != NULL);
4992 oh = container_of0(handle, struct osd_thandle, ot_super);
4993 LASSERT(oh->ot_handle == NULL);
4995 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4996 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4998 inode = osd_dt_obj(dt)->oo_inode;
5002 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5003 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5008 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5011 struct osd_fid_pack *rec;
5014 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5015 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5016 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5017 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5023 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5024 const struct lu_fid *fid)
5026 struct seq_server_site *ss = osd_seq_site(osd);
5030 /* FID seqs not in FLDB, must be local seq */
5031 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5035 * If FLD is not being initialized yet, it only happens during the
5036 * initialization, likely during mgs initialization, and we assume
5037 * this is local FID.
5039 if (ss == NULL || ss->ss_server_fld == NULL)
5042 /* Only check the local FLDB here */
5043 if (osd_seq_exists(env, osd, fid_seq(fid)))
5049 static void osd_take_care_of_agent(const struct lu_env *env,
5050 struct osd_device *osd,
5051 struct osd_thandle *oh,
5052 struct ldiskfs_dir_entry_2 *de)
5054 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5055 struct osd_idmap_cache *idc;
5056 int rc, schedule = 0;
5058 LASSERT(de != NULL);
5060 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5061 if (likely(rc == 0)) {
5062 idc = osd_idc_find_or_init(env, osd, fid);
5063 if (IS_ERR(idc) || idc->oic_remote)
5065 } else if (rc == -ENODATA) {
5067 * can't get FID, postpone to the end of the
5068 * transaction when iget() is safe
5072 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5075 osd_schedule_agent_inode_removal(env, oh,
5076 le32_to_cpu(de->inode));
5080 * Index delete function for interoperability mode (b11826).
5081 * It will remove the directory entry added by osd_index_ea_insert().
5082 * This entry is needed to maintain name->fid mapping.
5084 * \param key, key i.e. file entry to be deleted
5086 * \retval 0, on success
5087 * \retval -ve, on error
5089 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5090 const struct dt_key *key, struct thandle *handle)
5092 struct osd_object *obj = osd_dt_obj(dt);
5093 struct inode *dir = obj->oo_inode;
5094 struct dentry *dentry;
5095 struct osd_thandle *oh;
5096 struct ldiskfs_dir_entry_2 *de = NULL;
5097 struct buffer_head *bh;
5098 struct htree_lock *hlock = NULL;
5099 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5104 if (!dt_object_exists(dt))
5107 LINVRNT(osd_invariant(obj));
5108 LASSERT(!dt_object_remote(dt));
5109 LASSERT(handle != NULL);
5111 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5113 oh = container_of(handle, struct osd_thandle, ot_super);
5114 LASSERT(oh->ot_handle != NULL);
5115 LASSERT(oh->ot_handle->h_transaction != NULL);
5117 ll_vfs_dq_init(dir);
5118 dentry = osd_child_dentry_get(env, obj,
5119 (char *)key, strlen((char *)key));
5121 if (obj->oo_hl_head != NULL) {
5122 hlock = osd_oti_get(env)->oti_hlock;
5123 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5124 dir, LDISKFS_HLOCK_DEL);
5126 down_write(&obj->oo_ext_idx_sem);
5129 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5132 * If this is not the ".." entry, it might be a remote DNE
5133 * entry and we need to check if the FID is for a remote
5134 * MDT. If the FID is not in the directory entry (e.g.
5135 * upgraded 1.8 filesystem without dirdata enabled) then
5136 * we need to get the FID from the LMA. For a remote directory
5137 * there HAS to be an LMA, it cannot be an IGIF inode in this
5140 * Delete the entry before the agent inode in order to
5141 * simplify error handling. At worst an error after deleting
5142 * the entry first might leak the agent inode afterward. The
5143 * reverse would need filesystem abort in case of error deleting
5144 * the entry after the agent had been removed, or leave a
5145 * dangling entry pointing at a random inode.
5147 if (strcmp((char *)key, dotdot) != 0)
5148 osd_take_care_of_agent(env, osd, oh, de);
5149 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5155 ldiskfs_htree_unlock(hlock);
5157 up_write(&obj->oo_ext_idx_sem);
5161 LASSERT(osd_invariant(obj));
5162 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5167 * Lookup index for \a key and copy record to \a rec.
5169 * \param dt osd index object
5170 * \param key key for index
5171 * \param rec record reference
5173 * \retval +ve success : exact mach
5174 * \retval 0 return record with key not greater than \a key
5175 * \retval -ve failure
5177 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5178 struct dt_rec *rec, const struct dt_key *key)
5180 struct osd_object *obj = osd_dt_obj(dt);
5181 struct iam_path_descr *ipd;
5182 struct iam_container *bag = &obj->oo_dir->od_container;
5183 struct osd_thread_info *oti = osd_oti_get(env);
5184 struct iam_iterator *it = &oti->oti_idx_it;
5185 struct iam_rec *iam_rec;
5190 if (!dt_object_exists(dt))
5193 LASSERT(osd_invariant(obj));
5194 LASSERT(!dt_object_remote(dt));
5195 LASSERT(bag->ic_object == obj->oo_inode);
5197 ipd = osd_idx_ipd_get(env, bag);
5201 /* got ipd now we can start iterator. */
5202 iam_it_init(it, bag, 0, ipd);
5204 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5205 /* swab quota uid/gid provided by caller */
5206 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5207 key = (const struct dt_key *)&oti->oti_quota_id;
5210 rc = iam_it_get(it, (struct iam_key *)key);
5212 if (S_ISDIR(obj->oo_inode->i_mode))
5213 iam_rec = (struct iam_rec *)oti->oti_ldp;
5215 iam_rec = (struct iam_rec *)rec;
5217 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5219 if (S_ISDIR(obj->oo_inode->i_mode))
5220 osd_fid_unpack((struct lu_fid *)rec,
5221 (struct osd_fid_pack *)iam_rec);
5222 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5223 osd_quota_unpack(obj, rec);
5228 osd_ipd_put(env, bag, ipd);
5230 LINVRNT(osd_invariant(obj));
5235 static int osd_index_declare_iam_insert(const struct lu_env *env,
5236 struct dt_object *dt,
5237 const struct dt_rec *rec,
5238 const struct dt_key *key,
5239 struct thandle *handle)
5241 struct osd_thandle *oh;
5243 LASSERT(handle != NULL);
5245 oh = container_of0(handle, struct osd_thandle, ot_super);
5246 LASSERT(oh->ot_handle == NULL);
5248 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5249 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5255 * Inserts (key, value) pair in \a dt index object.
5257 * \param dt osd index object
5258 * \param key key for index
5259 * \param rec record reference
5260 * \param th transaction handler
5263 * \retval -ve failure
5265 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5266 const struct dt_rec *rec,
5267 const struct dt_key *key, struct thandle *th)
5269 struct osd_object *obj = osd_dt_obj(dt);
5270 struct iam_path_descr *ipd;
5271 struct osd_thandle *oh;
5272 struct iam_container *bag;
5273 struct osd_thread_info *oti = osd_oti_get(env);
5274 struct iam_rec *iam_rec;
5279 if (!dt_object_exists(dt))
5282 LINVRNT(osd_invariant(obj));
5283 LASSERT(!dt_object_remote(dt));
5285 bag = &obj->oo_dir->od_container;
5286 LASSERT(bag->ic_object == obj->oo_inode);
5287 LASSERT(th != NULL);
5289 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5291 ipd = osd_idx_ipd_get(env, bag);
5292 if (unlikely(ipd == NULL))
5295 oh = container_of0(th, struct osd_thandle, ot_super);
5296 LASSERT(oh->ot_handle != NULL);
5297 LASSERT(oh->ot_handle->h_transaction != NULL);
5298 if (S_ISDIR(obj->oo_inode->i_mode)) {
5299 iam_rec = (struct iam_rec *)oti->oti_ldp;
5300 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5302 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5303 /* pack quota uid/gid */
5304 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5305 key = (const struct dt_key *)&oti->oti_quota_id;
5306 /* pack quota record */
5307 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5308 iam_rec = (struct iam_rec *)rec;
5310 iam_rec = (struct iam_rec *)rec;
5313 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5315 osd_ipd_put(env, bag, ipd);
5316 LINVRNT(osd_invariant(obj));
5317 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5322 * Calls ldiskfs_add_entry() to add directory entry
5323 * into the directory. This is required for
5324 * interoperability mode (b11826)
5326 * \retval 0, on success
5327 * \retval -ve, on error
5329 static int __osd_ea_add_rec(struct osd_thread_info *info,
5330 struct osd_object *pobj, struct inode *cinode,
5331 const char *name, const struct lu_fid *fid,
5332 struct htree_lock *hlock, struct thandle *th)
5334 struct ldiskfs_dentry_param *ldp;
5335 struct dentry *child;
5336 struct osd_thandle *oth;
5339 oth = container_of(th, struct osd_thandle, ot_super);
5340 LASSERT(oth->ot_handle != NULL);
5341 LASSERT(oth->ot_handle->h_transaction != NULL);
5342 LASSERT(pobj->oo_inode);
5344 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5345 if (unlikely(pobj->oo_inode ==
5346 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
5349 osd_get_ldiskfs_dirent_param(ldp, fid);
5350 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5351 child->d_fsdata = (void *)ldp;
5352 ll_vfs_dq_init(pobj->oo_inode);
5353 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5354 child, cinode, hlock);
5355 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5356 struct ldiskfs_dir_entry_2 *de;
5357 struct buffer_head *bh;
5360 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5363 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5366 if (S_ISDIR(cinode->i_mode))
5367 de->file_type = LDISKFS_DIRENT_LUFID |
5368 LDISKFS_FT_REG_FILE;
5370 de->file_type = LDISKFS_DIRENT_LUFID |
5372 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5383 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5384 * into the directory.Also sets flags into osd object to
5385 * indicate dot and dotdot are created. This is required for
5386 * interoperability mode (b11826)
5388 * \param dir directory for dot and dotdot fixup.
5389 * \param obj child object for linking
5391 * \retval 0, on success
5392 * \retval -ve, on error
5394 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5395 struct osd_object *dir,
5396 struct inode *parent_dir, const char *name,
5397 const struct lu_fid *dot_fid,
5398 const struct lu_fid *dot_dot_fid,
5401 struct inode *inode = dir->oo_inode;
5402 struct osd_thandle *oth;
5405 oth = container_of(th, struct osd_thandle, ot_super);
5406 LASSERT(oth->ot_handle->h_transaction != NULL);
5407 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5409 if (strcmp(name, dot) == 0) {
5410 if (dir->oo_compat_dot_created) {
5413 LASSERT(inode->i_ino == parent_dir->i_ino);
5414 dir->oo_compat_dot_created = 1;
5417 } else if (strcmp(name, dotdot) == 0) {
5418 if (!dir->oo_compat_dot_created)
5420 /* in case of rename, dotdot is already created */
5421 if (dir->oo_compat_dotdot_created) {
5422 return __osd_ea_add_rec(info, dir, parent_dir, name,
5423 dot_dot_fid, NULL, th);
5426 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5427 struct lu_fid tfid = *dot_dot_fid;
5430 result = osd_add_dot_dotdot_internal(info,
5431 dir->oo_inode, parent_dir, dot_fid,
5434 result = osd_add_dot_dotdot_internal(info,
5435 dir->oo_inode, parent_dir, dot_fid,
5440 dir->oo_compat_dotdot_created = 1;
5448 * It will call the appropriate osd_add* function and return the
5449 * value, return by respective functions.
5451 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5452 struct inode *cinode, const char *name,
5453 const struct lu_fid *fid, struct thandle *th)
5455 struct osd_thread_info *info = osd_oti_get(env);
5456 struct htree_lock *hlock;
5459 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5461 if (name[0] == '.' && (name[1] == '\0' ||
5462 (name[1] == '.' && name[2] == '\0'))) {
5463 if (hlock != NULL) {
5464 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5467 down_write(&pobj->oo_ext_idx_sem);
5470 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5471 lu_object_fid(&pobj->oo_dt.do_lu),
5474 if (hlock != NULL) {
5475 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5476 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5478 down_write(&pobj->oo_ext_idx_sem);
5481 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5482 struct lu_fid *tfid = &info->oti_fid;
5486 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5489 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5494 ldiskfs_htree_unlock(hlock);
5496 up_write(&pobj->oo_ext_idx_sem);
5502 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5503 struct osd_idmap_cache *oic)
5505 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5506 struct lu_fid *fid = &oic->oic_fid;
5507 struct osd_inode_id *id = &oic->oic_lid;
5508 struct inode *inode = NULL;
5515 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5518 if (thread_is_running(&scrub->os_thread) &&
5519 scrub->os_pos_current > id->oii_ino)
5522 if (dev->od_auto_scrub_interval == AS_NEVER ||
5523 ktime_get_real_seconds() <
5524 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5528 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5529 if (rc == -ENOENT) {
5530 __u32 gen = id->oii_gen;
5536 inode = osd_iget(oti, dev, id);
5537 /* The inode has been removed (by race maybe). */
5538 if (IS_ERR(inode)) {
5539 rc = PTR_ERR(inode);
5541 RETURN(rc == -ESTALE ? -ENOENT : rc);
5544 /* The OI mapping is lost. */
5545 if (gen != OSD_OII_NOGEN)
5549 * The inode may has been reused by others, we do not know,
5550 * leave it to be handled by subsequent osd_fid_lookup().
5553 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5560 if (thread_is_running(&scrub->os_thread)) {
5561 if (inode == NULL) {
5562 inode = osd_iget(oti, dev, id);
5563 /* The inode has been removed (by race maybe). */
5564 if (IS_ERR(inode)) {
5565 rc = PTR_ERR(inode);
5567 RETURN(rc == -ESTALE ? -ENOENT : rc);
5571 rc = osd_oii_insert(dev, oic, insert);
5573 * There is race condition between osd_oi_lookup and OI scrub.
5574 * The OI scrub finished just after osd_oi_lookup() failure.
5575 * Under such case, it is unnecessary to trigger OI scrub again,
5576 * but try to call osd_oi_lookup() again.
5578 if (unlikely(rc == -EAGAIN))
5581 if (!S_ISDIR(inode->i_mode))
5584 rc = osd_check_lmv(oti, dev, inode, oic);
5589 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5590 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5591 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5592 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5593 "%s: trigger partial OI scrub for RPC inconsistency "
5594 "checking FID "DFID": rc = %d\n",
5595 osd_dev2name(dev), PFID(fid), rc);
5596 if (rc == 0 || rc == -EALREADY)
5609 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5610 struct osd_device *dev,
5611 struct osd_idmap_cache *oic,
5612 struct lu_fid *fid, __u32 ino)
5614 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5615 struct inode *inode;
5618 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5619 inode = osd_iget(oti, dev, &oic->oic_lid);
5620 if (IS_ERR(inode)) {
5621 fid_zero(&oic->oic_fid);
5622 return PTR_ERR(inode);
5625 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5628 fid_zero(&oic->oic_fid);
5630 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5634 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5635 struct osd_inode_id *id, const struct lu_fid *fid)
5637 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5638 id->oii_ino, id->oii_gen, info);
5639 info->oti_cache.oic_lid = *id;
5640 info->oti_cache.oic_fid = *fid;
5641 info->oti_cache.oic_dev = osd;
5645 * Get parent FID from the linkEA.
5647 * For a directory which parent resides on remote MDT, to satisfy the
5648 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5649 * the other hand, to make the lookup(..) on the directory can return
5650 * the real parent FID, we append the real parent FID after its ".."
5651 * name entry in the /REMOTE_PARENT_DIR.
5653 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5654 * backup. So after the restore, we cannot get the right parent FID from
5655 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5656 * we have stored the real parent FID in the directory object's linkEA,
5657 * we can parse the linkEA for the real parent FID.
5659 * \param[in] env pointer to the thread context
5660 * \param[in] obj pointer to the object to be handled
5661 * \param[out]fid pointer to the buffer to hold the parent FID
5663 * \retval 0 for getting the real parent FID successfully
5664 * \retval negative error number on failure
5666 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5667 struct osd_object *obj,
5670 struct osd_thread_info *oti = osd_oti_get(env);
5671 struct lu_buf *buf = &oti->oti_big_buf;
5672 struct dentry *dentry = &oti->oti_obj_dentry;
5673 struct inode *inode = obj->oo_inode;
5674 struct linkea_data ldata = { NULL };
5680 if (!S_ISDIR(inode->i_mode))
5684 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5685 buf->lb_buf, buf->lb_len);
5686 if (rc == -ERANGE) {
5687 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5690 lu_buf_realloc(buf, rc);
5691 if (buf->lb_buf == NULL)
5698 if (unlikely(rc == 0))
5704 if (unlikely(buf->lb_buf == NULL)) {
5705 lu_buf_realloc(buf, rc);
5706 if (buf->lb_buf == NULL)
5713 rc = linkea_init_with_rec(&ldata);
5715 linkea_first_entry(&ldata);
5716 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5722 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5723 struct inode *inode,
5724 const struct lu_fid *pfid,
5725 const char *name, const int namelen)
5727 struct osd_thread_info *oti = osd_oti_get(env);
5728 struct lu_buf *buf = &oti->oti_big_buf;
5729 struct dentry *dentry = &oti->oti_obj_dentry;
5730 struct linkea_data ldata = { NULL };
5731 struct lu_name cname = { .ln_name = name,
5732 .ln_namelen = namelen };
5738 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5739 buf->lb_buf, buf->lb_len);
5741 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5746 if (unlikely(rc == 0))
5749 if (buf->lb_len < rc) {
5750 lu_buf_realloc(buf, rc);
5751 if (buf->lb_buf == NULL)
5758 rc = linkea_init_with_rec(&ldata);
5760 rc = linkea_links_find(&ldata, &cname, pfid);
5766 * Calls ->lookup() to find dentry. From dentry get inode and
5767 * read inode's ea to get fid. This is required for interoperability
5770 * \retval 0, on success
5771 * \retval -ve, on error
5773 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5774 struct dt_rec *rec, const struct dt_key *key)
5776 struct inode *dir = obj->oo_inode;
5777 struct dentry *dentry;
5778 struct ldiskfs_dir_entry_2 *de;
5779 struct buffer_head *bh;
5780 struct lu_fid *fid = (struct lu_fid *)rec;
5781 struct htree_lock *hlock = NULL;
5787 LASSERT(dir->i_op != NULL);
5788 LASSERT(dir->i_op->lookup != NULL);
5790 dentry = osd_child_dentry_get(env, obj,
5791 (char *)key, strlen((char *)key));
5793 if (obj->oo_hl_head != NULL) {
5794 hlock = osd_oti_get(env)->oti_hlock;
5795 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5796 dir, LDISKFS_HLOCK_LOOKUP);
5798 down_read(&obj->oo_ext_idx_sem);
5801 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5803 struct osd_thread_info *oti = osd_oti_get(env);
5804 struct osd_inode_id *id = &oti->oti_id;
5805 struct osd_idmap_cache *oic = &oti->oti_cache;
5806 struct osd_device *dev = osd_obj2dev(obj);
5808 ino = le32_to_cpu(de->inode);
5809 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5811 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5815 rc = osd_get_fid_from_dentry(de, rec);
5817 /* done with de, release bh */
5820 if (unlikely(is_remote_parent_ino(dev, ino))) {
5821 const char *name = (const char *)key;
5824 * If the parent is on remote MDT, and there
5825 * is no FID-in-dirent, then we have to get
5826 * the parent FID from the linkEA.
5828 if (likely(strlen(name) == 2 &&
5829 name[0] == '.' && name[1] == '.'))
5830 rc = osd_get_pfid_from_linkea(env, obj,
5833 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5836 osd_id_gen(id, ino, OSD_OII_NOGEN);
5839 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5840 fid_zero(&oic->oic_fid);
5845 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5846 rc = osd_consistency_check(oti, dev, oic);
5848 fid_zero(&oic->oic_fid);
5850 /* Other error should not affect lookup result. */
5860 ldiskfs_htree_unlock(hlock);
5862 up_read(&obj->oo_ext_idx_sem);
5866 static int osd_index_declare_ea_insert(const struct lu_env *env,
5867 struct dt_object *dt,
5868 const struct dt_rec *rec,
5869 const struct dt_key *key,
5870 struct thandle *handle)
5872 struct osd_thandle *oh;
5873 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5874 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5875 const struct lu_fid *fid = rec1->rec_fid;
5876 int credits, rc = 0;
5877 struct osd_idmap_cache *idc;
5881 LASSERT(!dt_object_remote(dt));
5882 LASSERT(handle != NULL);
5883 LASSERT(fid != NULL);
5884 LASSERT(rec1->rec_type != 0);
5886 oh = container_of0(handle, struct osd_thandle, ot_super);
5887 LASSERT(oh->ot_handle == NULL);
5889 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5892 * we can't call iget() while a transactions is running
5893 * (this can lead to a deadlock), but we need to know
5894 * inum and object type. so we find this information at
5895 * declaration and cache in per-thread info
5897 idc = osd_idc_find_or_init(env, osd, fid);
5899 RETURN(PTR_ERR(idc));
5900 if (idc->oic_remote) {
5902 * a reference to remote inode is represented by an
5903 * agent inode which we have to create
5905 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5906 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5909 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5911 if (osd_dt_obj(dt)->oo_inode != NULL) {
5912 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5915 * We ignore block quota on meta pool (MDTs), so needn't
5916 * calculate how many blocks will be consumed by this index
5919 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5921 i_projid_read(inode), 0,
5922 oh, osd_dt_obj(dt), NULL,
5927 #ifdef HAVE_PROJECT_QUOTA
5929 * Reserve credits for local agent inode to transfer
5930 * to 0, quota enforcement is ignored in this case.
5932 if (idc->oic_remote &&
5933 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
5934 i_projid_read(inode) != 0)
5935 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
5936 0, i_projid_read(inode),
5937 0, false, PRJQUOTA, true);
5945 * Index add function for interoperability mode (b11826).
5946 * It will add the directory entry.This entry is needed to
5947 * maintain name->fid mapping.
5949 * \param key it is key i.e. file entry to be inserted
5950 * \param rec it is value of given key i.e. fid
5952 * \retval 0, on success
5953 * \retval -ve, on error
5955 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5956 const struct dt_rec *rec,
5957 const struct dt_key *key, struct thandle *th)
5959 struct osd_object *obj = osd_dt_obj(dt);
5960 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5961 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5962 const struct lu_fid *fid = rec1->rec_fid;
5963 const char *name = (const char *)key;
5964 struct osd_thread_info *oti = osd_oti_get(env);
5965 struct inode *child_inode = NULL;
5966 struct osd_idmap_cache *idc;
5971 if (!dt_object_exists(dt))
5974 LASSERT(osd_invariant(obj));
5975 LASSERT(!dt_object_remote(dt));
5976 LASSERT(th != NULL);
5978 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5980 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5982 idc = osd_idc_find(env, osd, fid);
5983 if (unlikely(idc == NULL)) {
5984 idc = osd_idc_find_or_init(env, osd, fid);
5987 * this dt_insert() wasn't declared properly, so
5988 * FID is missing in OI cache. we better do not
5989 * lookup FID in FLDB/OI and don't risk to deadlock,
5990 * but in some special cases (lfsck testing, etc)
5991 * it's much simpler than fixing a caller.
5993 * normally this error should be placed after the first
5994 * find, but migrate may attach source stripes to
5995 * target, which doesn't create stripes.
5997 CERROR("%s: "DFID" wasn't declared for insert\n",
5998 osd_name(osd), PFID(fid));
6000 RETURN(PTR_ERR(idc));
6004 if (idc->oic_remote) {
6005 /* Insert remote entry */
6006 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6008 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6010 child_inode = osd_create_local_agent_inode(env, osd,
6011 obj, fid, rec1->rec_type & S_IFMT, th);
6012 if (IS_ERR(child_inode))
6013 RETURN(PTR_ERR(child_inode));
6016 /* Insert local entry */
6017 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6018 /* for a reason OI cache wasn't filled properly */
6019 CERROR("%s: OIC for "DFID" isn't filled\n",
6020 osd_name(osd), PFID(fid));
6023 child_inode = oti->oti_inode;
6024 if (unlikely(child_inode == NULL)) {
6025 struct ldiskfs_inode_info *lii;
6030 child_inode = oti->oti_inode = &lii->vfs_inode;
6032 child_inode->i_sb = osd_sb(osd);
6033 child_inode->i_ino = idc->oic_lid.oii_ino;
6034 child_inode->i_mode = rec1->rec_type & S_IFMT;
6037 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6039 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6040 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6042 if (child_inode && child_inode != oti->oti_inode)
6044 LASSERT(osd_invariant(obj));
6045 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6050 * Initialize osd Iterator for given osd index object.
6052 * \param dt osd index object
6055 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6056 struct dt_object *dt,
6059 struct osd_it_iam *it;
6060 struct osd_object *obj = osd_dt_obj(dt);
6061 struct lu_object *lo = &dt->do_lu;
6062 struct iam_path_descr *ipd;
6063 struct iam_container *bag = &obj->oo_dir->od_container;
6065 if (!dt_object_exists(dt))
6066 return ERR_PTR(-ENOENT);
6070 return ERR_PTR(-ENOMEM);
6072 ipd = osd_it_ipd_get(env, bag);
6073 if (likely(ipd != NULL)) {
6077 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6078 return (struct dt_it *)it;
6081 return ERR_PTR(-ENOMEM);
6086 * free given Iterator.
6088 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6090 struct osd_it_iam *it = (struct osd_it_iam *)di;
6091 struct osd_object *obj = it->oi_obj;
6093 iam_it_fini(&it->oi_it);
6094 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6095 osd_object_put(env, obj);
6100 * Move Iterator to record specified by \a key
6102 * \param di osd iterator
6103 * \param key key for index
6105 * \retval +ve di points to record with least key not larger than key
6106 * \retval 0 di points to exact matched key
6107 * \retval -ve failure
6110 static int osd_it_iam_get(const struct lu_env *env,
6111 struct dt_it *di, const struct dt_key *key)
6113 struct osd_thread_info *oti = osd_oti_get(env);
6114 struct osd_it_iam *it = (struct osd_it_iam *)di;
6116 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6117 /* swab quota uid/gid */
6118 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6119 key = (struct dt_key *)&oti->oti_quota_id;
6122 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6128 * \param di osd iterator
6130 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6132 struct osd_it_iam *it = (struct osd_it_iam *)di;
6134 iam_it_put(&it->oi_it);
6138 * Move iterator by one record
6140 * \param di osd iterator
6142 * \retval +1 end of container reached
6144 * \retval -ve failure
6147 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6149 struct osd_it_iam *it = (struct osd_it_iam *)di;
6151 return iam_it_next(&it->oi_it);
6155 * Return pointer to the key under iterator.
6158 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6159 const struct dt_it *di)
6161 struct osd_thread_info *oti = osd_oti_get(env);
6162 struct osd_it_iam *it = (struct osd_it_iam *)di;
6163 struct osd_object *obj = it->oi_obj;
6166 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6168 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6169 /* swab quota uid/gid */
6170 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6171 key = (struct dt_key *)&oti->oti_quota_id;
6178 * Return size of key under iterator (in bytes)
6181 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6183 struct osd_it_iam *it = (struct osd_it_iam *)di;
6185 return iam_it_key_size(&it->oi_it);
6189 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6191 /* check if file type is required */
6192 if (ent->lde_attrs & LUDA_TYPE) {
6193 struct luda_type *lt;
6194 int align = sizeof(*lt) - 1;
6196 len = (len + align) & ~align;
6197 lt = (struct luda_type *)(ent->lde_name + len);
6198 lt->lt_type = cpu_to_le16(DTTOIF(type));
6201 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6205 * build lu direct from backend fs dirent.
6209 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6210 char *name, __u16 namelen, __u16 type, __u32 attr)
6212 ent->lde_attrs = attr | LUDA_FID;
6213 fid_cpu_to_le(&ent->lde_fid, fid);
6215 ent->lde_hash = cpu_to_le64(offset);
6216 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6218 strncpy(ent->lde_name, name, namelen);
6219 ent->lde_name[namelen] = '\0';
6220 ent->lde_namelen = cpu_to_le16(namelen);
6222 /* append lustre attributes */
6223 osd_it_append_attrs(ent, namelen, type);
6227 * Return pointer to the record under iterator.
6229 static int osd_it_iam_rec(const struct lu_env *env,
6230 const struct dt_it *di,
6231 struct dt_rec *dtrec, __u32 attr)
6233 struct osd_it_iam *it = (struct osd_it_iam *)di;
6234 struct osd_thread_info *info = osd_oti_get(env);
6238 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6239 const struct osd_fid_pack *rec;
6240 struct lu_fid *fid = &info->oti_fid;
6241 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6247 name = (char *)iam_it_key_get(&it->oi_it);
6249 RETURN(PTR_ERR(name));
6251 namelen = iam_it_key_size(&it->oi_it);
6253 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6255 RETURN(PTR_ERR(rec));
6257 rc = osd_fid_unpack(fid, rec);
6261 hash = iam_it_store(&it->oi_it);
6263 /* IAM does not store object type in IAM index (dir) */
6264 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6266 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6267 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6268 (struct iam_rec *)dtrec);
6269 osd_quota_unpack(it->oi_obj, dtrec);
6271 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6272 (struct iam_rec *)dtrec);
6279 * Returns cookie for current Iterator position.
6281 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6283 struct osd_it_iam *it = (struct osd_it_iam *)di;
6285 return iam_it_store(&it->oi_it);
6289 * Restore iterator from cookie.
6291 * \param di osd iterator
6292 * \param hash Iterator location cookie
6294 * \retval +ve di points to record with least key not larger than key.
6295 * \retval 0 di points to exact matched key
6296 * \retval -ve failure
6299 static int osd_it_iam_load(const struct lu_env *env,
6300 const struct dt_it *di, __u64 hash)
6302 struct osd_it_iam *it = (struct osd_it_iam *)di;
6304 return iam_it_load(&it->oi_it, hash);
6307 static const struct dt_index_operations osd_index_iam_ops = {
6308 .dio_lookup = osd_index_iam_lookup,
6309 .dio_declare_insert = osd_index_declare_iam_insert,
6310 .dio_insert = osd_index_iam_insert,
6311 .dio_declare_delete = osd_index_declare_iam_delete,
6312 .dio_delete = osd_index_iam_delete,
6314 .init = osd_it_iam_init,
6315 .fini = osd_it_iam_fini,
6316 .get = osd_it_iam_get,
6317 .put = osd_it_iam_put,
6318 .next = osd_it_iam_next,
6319 .key = osd_it_iam_key,
6320 .key_size = osd_it_iam_key_size,
6321 .rec = osd_it_iam_rec,
6322 .store = osd_it_iam_store,
6323 .load = osd_it_iam_load
6329 * Creates or initializes iterator context.
6331 * \retval struct osd_it_ea, iterator structure on success
6334 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6335 struct dt_object *dt,
6338 struct osd_object *obj = osd_dt_obj(dt);
6339 struct osd_thread_info *info = osd_oti_get(env);
6340 struct osd_it_ea *oie;
6342 struct lu_object *lo = &dt->do_lu;
6343 struct dentry *obj_dentry;
6347 if (!dt_object_exists(dt) || obj->oo_destroyed)
6348 RETURN(ERR_PTR(-ENOENT));
6350 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6352 RETURN(ERR_PTR(-ENOMEM));
6353 obj_dentry = &oie->oie_dentry;
6355 obj_dentry->d_inode = obj->oo_inode;
6356 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
6357 obj_dentry->d_name.hash = 0;
6359 oie->oie_rd_dirent = 0;
6360 oie->oie_it_dirent = 0;
6361 oie->oie_dirent = NULL;
6362 if (unlikely(!info->oti_it_ea_buf_used)) {
6363 oie->oie_buf = info->oti_it_ea_buf;
6364 info->oti_it_ea_buf_used = 1;
6366 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6367 if (oie->oie_buf == NULL)
6368 RETURN(ERR_PTR(-ENOMEM));
6372 file = &oie->oie_file;
6374 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6375 if (attr & LUDA_64BITHASH)
6376 file->f_mode = FMODE_64BITHASH;
6378 file->f_mode = FMODE_32BITHASH;
6379 file->f_path.dentry = obj_dentry;
6380 file->f_mapping = obj->oo_inode->i_mapping;
6381 file->f_op = obj->oo_inode->i_fop;
6382 set_file_inode(file, obj->oo_inode);
6385 RETURN((struct dt_it *)oie);
6389 * Destroy or finishes iterator context.
6391 * \param di iterator structure to be destroyed
6393 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6395 struct osd_thread_info *info = osd_oti_get(env);
6396 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6397 struct osd_object *obj = oie->oie_obj;
6398 struct inode *inode = obj->oo_inode;
6401 oie->oie_file.f_op->release(inode, &oie->oie_file);
6402 osd_object_put(env, obj);
6403 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6404 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6406 info->oti_it_ea_buf_used = 0;
6407 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6412 * It position the iterator at given key, so that next lookup continues from
6413 * that key Or it is similar to dio_it->load() but based on a key,
6414 * rather than file position.
6416 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6419 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6421 static int osd_it_ea_get(const struct lu_env *env,
6422 struct dt_it *di, const struct dt_key *key)
6424 struct osd_it_ea *it = (struct osd_it_ea *)di;
6427 LASSERT(((const char *)key)[0] == '\0');
6428 it->oie_file.f_pos = 0;
6429 it->oie_rd_dirent = 0;
6430 it->oie_it_dirent = 0;
6431 it->oie_dirent = NULL;
6439 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6443 struct osd_filldir_cbs {
6444 #ifdef HAVE_DIR_CONTEXT
6445 struct dir_context ctx;
6447 struct osd_it_ea *it;
6450 * It is called internally by ->readdir(). It fills the
6451 * iterator's in-memory data structure with required
6452 * information i.e. name, namelen, rec_size etc.
6454 * \param buf in which information to be filled in.
6455 * \param name name of the file in given dir
6457 * \retval 0 on success
6458 * \retval 1 on buffer full
6460 #ifdef HAVE_FILLDIR_USE_CTX
6461 static int osd_ldiskfs_filldir(struct dir_context *buf,
6463 static int osd_ldiskfs_filldir(void *buf,
6465 const char *name, int namelen,
6466 loff_t offset, __u64 ino, unsigned int d_type)
6468 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6469 struct osd_object *obj = it->oie_obj;
6470 struct osd_it_ea_dirent *ent = it->oie_dirent;
6471 struct lu_fid *fid = &ent->oied_fid;
6472 struct osd_fid_pack *rec;
6476 /* this should never happen */
6477 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6478 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6482 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6486 /* "." is just the object itself. */
6487 if (namelen == 1 && name[0] == '.') {
6488 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6489 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6490 rec = (struct osd_fid_pack *)(name + namelen + 1);
6491 if (osd_fid_unpack(fid, rec) != 0)
6496 d_type &= ~LDISKFS_DIRENT_LUFID;
6498 /* NOT export local root. */
6499 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6500 ino = obj->oo_inode->i_ino;
6501 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6504 ent->oied_ino = ino;
6505 ent->oied_off = offset;
6506 ent->oied_namelen = namelen;
6507 ent->oied_type = d_type;
6509 memcpy(ent->oied_name, name, namelen);
6511 it->oie_rd_dirent++;
6512 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6517 * Calls ->readdir() to load a directory entry at a time
6518 * and stored it in iterator's in-memory data structure.
6520 * \param di iterator's in memory structure
6522 * \retval 0 on success
6523 * \retval -ve on error
6524 * \retval +1 reach the end of entry
6526 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6527 const struct dt_it *di)
6529 struct osd_it_ea *it = (struct osd_it_ea *)di;
6530 struct osd_object *obj = it->oie_obj;
6531 struct inode *inode = obj->oo_inode;
6532 struct htree_lock *hlock = NULL;
6533 struct file *filp = &it->oie_file;
6535 struct osd_filldir_cbs buf = {
6536 #ifdef HAVE_DIR_CONTEXT
6537 .ctx.actor = osd_ldiskfs_filldir,
6543 it->oie_dirent = it->oie_buf;
6544 it->oie_rd_dirent = 0;
6546 if (obj->oo_hl_head != NULL) {
6547 hlock = osd_oti_get(env)->oti_hlock;
6548 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6549 inode, LDISKFS_HLOCK_READDIR);
6551 down_read(&obj->oo_ext_idx_sem);
6554 #ifdef HAVE_DIR_CONTEXT
6555 buf.ctx.pos = filp->f_pos;
6556 #ifdef HAVE_ITERATE_SHARED
6557 rc = inode->i_fop->iterate_shared(filp, &buf.ctx);
6559 rc = inode->i_fop->iterate(filp, &buf.ctx);
6561 filp->f_pos = buf.ctx.pos;
6563 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6567 ldiskfs_htree_unlock(hlock);
6569 up_read(&obj->oo_ext_idx_sem);
6571 if (it->oie_rd_dirent == 0) {
6573 * If it does not get any dirent, it means it has been reached
6574 * to the end of the dir
6576 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6580 it->oie_dirent = it->oie_buf;
6581 it->oie_it_dirent = 1;
6588 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6589 * to load a directory entry at a time and stored it in
6590 * iterator's in-memory data structure.
6592 * \param di iterator's in memory structure
6594 * \retval +ve iterator reached to end
6595 * \retval 0 iterator not reached to end
6596 * \retval -ve on error
6598 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6600 struct osd_it_ea *it = (struct osd_it_ea *)di;
6605 if (it->oie_it_dirent < it->oie_rd_dirent) {
6607 (void *)it->oie_dirent +
6608 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6609 it->oie_dirent->oied_namelen);
6610 it->oie_it_dirent++;
6613 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6616 rc = osd_ldiskfs_it_fill(env, di);
6623 * Returns the key at current position from iterator's in memory structure.
6625 * \param di iterator's in memory structure
6627 * \retval key i.e. struct dt_key on success
6629 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6630 const struct dt_it *di)
6632 struct osd_it_ea *it = (struct osd_it_ea *)di;
6634 return (struct dt_key *)it->oie_dirent->oied_name;
6638 * Returns key's size at current position from iterator's in memory structure.
6640 * \param di iterator's in memory structure
6642 * \retval key_size i.e. struct dt_key on success
6644 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6646 struct osd_it_ea *it = (struct osd_it_ea *)di;
6648 return it->oie_dirent->oied_namelen;
6651 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6653 if (LDISKFS_DIR_REC_LEN(de) >=
6654 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6661 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6662 unsigned int blocksize, bool dotdot)
6665 return osd_dotdot_has_space(de);
6667 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6668 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6675 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6676 handle_t *jh, struct dentry *dentry,
6677 const struct lu_fid *fid, struct buffer_head *bh,
6678 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6681 struct inode *dir = dentry->d_parent->d_inode;
6682 struct inode *inode = dentry->d_inode;
6683 struct osd_fid_pack *rec;
6684 struct ldiskfs_dentry_param *ldp;
6685 int namelen = dentry->d_name.len;
6687 struct osd_thread_info *info = osd_oti_get(env);
6691 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6694 /* There is enough space to hold the FID-in-dirent. */
6695 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6696 rc = ldiskfs_journal_get_write_access(jh, bh);
6700 de->name[namelen] = 0;
6701 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6702 rec->fp_len = sizeof(struct lu_fid) + 1;
6703 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6704 de->file_type |= LDISKFS_DIRENT_LUFID;
6705 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6712 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6716 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6717 osd_get_ldiskfs_dirent_param(ldp, fid);
6718 dentry->d_fsdata = (void *)ldp;
6719 ll_vfs_dq_init(dir);
6720 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6722 * It is too bad, we cannot reinsert the name entry back.
6723 * That means we lose it!
6726 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6727 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6728 osd_ino2name(inode),
6729 dir->i_ino, dir->i_generation, namelen,
6730 dentry->d_name.name, PFID(fid), rc);
6736 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6737 struct osd_it_ea *it, struct lu_fid *fid,
6738 struct osd_inode_id *id, __u32 *attr)
6740 struct osd_thread_info *info = osd_oti_get(env);
6741 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6742 struct osd_device *dev = osd_obj2dev(obj);
6743 struct super_block *sb = osd_sb(dev);
6744 const char *devname = osd_name(dev);
6745 struct osd_it_ea_dirent *ent = it->oie_dirent;
6746 struct inode *dir = obj->oo_inode;
6747 struct htree_lock *hlock = NULL;
6748 struct buffer_head *bh = NULL;
6749 handle_t *jh = NULL;
6750 struct ldiskfs_dir_entry_2 *de;
6751 struct dentry *dentry;
6752 struct inode *inode;
6753 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6756 bool dotdot = false;
6761 if (ent->oied_name[0] == '.') {
6762 if (ent->oied_namelen == 1)
6765 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6769 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6770 inode = osd_iget(info, dev, id);
6771 if (IS_ERR(inode)) {
6772 rc = PTR_ERR(inode);
6773 if (rc == -ENOENT || rc == -ESTALE) {
6775 * Maybe dangling name entry, or
6776 * corrupted directory entry.
6778 *attr |= LUDA_UNKNOWN;
6781 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6782 "check_repair, dir = %lu/%u, name = %.*s, "
6783 "ino = %llu, rc = %d\n",
6784 devname, dir->i_ino, dir->i_generation,
6785 ent->oied_namelen, ent->oied_name,
6792 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6794 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6795 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6801 * We need to ensure that the name entry is still valid.
6802 * Because it may be removed or renamed by other already.
6804 * The unlink or rename operation will start journal before PDO lock,
6805 * so to avoid deadlock, here we need to start journal handle before
6806 * related PDO lock also. But because we do not know whether there
6807 * will be something to be repaired before PDO lock, we just start
6808 * journal without conditions.
6810 * We may need to remove the name entry firstly, then insert back.
6811 * One credit is for user quota file update.
6812 * One credit is for group quota file update.
6813 * Two credits are for dirty inode.
6815 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6816 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6818 if (dev->od_dirent_journal != 0) {
6821 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6824 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6825 "check_repair, dir = %lu/%u, credits = %d, "
6826 "name = %.*s, ino = %llu: rc = %d\n",
6827 devname, dir->i_ino, dir->i_generation, credits,
6828 ent->oied_namelen, ent->oied_name,
6831 GOTO(out_inode, rc);
6834 if (obj->oo_hl_head != NULL) {
6835 hlock = osd_oti_get(env)->oti_hlock;
6837 * "0" means exclusive lock for the whole directory.
6838 * We need to prevent others access such name entry
6839 * during the delete + insert. Neither HLOCK_ADD nor
6840 * HLOCK_DEL cannot guarantee the atomicity.
6842 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6844 down_write(&obj->oo_ext_idx_sem);
6847 if (obj->oo_hl_head != NULL) {
6848 hlock = osd_oti_get(env)->oti_hlock;
6849 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6850 LDISKFS_HLOCK_LOOKUP);
6852 down_read(&obj->oo_ext_idx_sem);
6856 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6857 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6858 *attr |= LUDA_IGNORE;
6864 * For dotdot entry, if there is not enough space to hold the
6865 * FID-in-dirent, just keep them there. It only happens when the
6866 * device upgraded from 1.8 or restored from MDT file-level backup.
6867 * For the whole directory, only dotdot entry have no FID-in-dirent
6868 * and needs to get FID from LMA when readdir, it will not affect the
6871 if (dotdot && !osd_dotdot_has_space(de)) {
6872 *attr |= LUDA_UNKNOWN;
6878 if (lu_fid_eq(fid, &lma->lma_self_fid))
6881 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6882 struct lu_fid *tfid = &lma->lma_self_fid;
6884 if (likely(dotdot &&
6885 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6886 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6888 * It must be REMOTE_PARENT_DIR and as the
6889 * 'dotdot' entry of remote directory
6891 *attr |= LUDA_IGNORE;
6893 CDEBUG(D_LFSCK, "%s: expect remote agent "
6894 "parent directory, but got %.*s under "
6895 "dir = %lu/%u with the FID "DFID"\n",
6896 devname, ent->oied_namelen,
6897 ent->oied_name, dir->i_ino,
6898 dir->i_generation, PFID(tfid));
6900 *attr |= LUDA_UNKNOWN;
6907 if (!fid_is_zero(fid)) {
6908 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6910 if (rc == -ENOENT ||
6912 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
6914 * linkEA does not recognize the dirent entry,
6915 * it may because the dirent entry corruption
6916 * and points to other's inode.
6918 CDEBUG(D_LFSCK, "%s: the target inode does not "
6919 "recognize the dirent, dir = %lu/%u, "
6920 " name = %.*s, ino = %llu, "
6921 DFID": rc = %d\n", devname, dir->i_ino,
6922 dir->i_generation, ent->oied_namelen,
6923 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6924 *attr |= LUDA_UNKNOWN;
6929 if (rc && rc != -ENODATA) {
6930 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6931 "dir = %lu/%u, name = %.*s, ino = %llu, "
6932 DFID": rc = %d\n", devname, dir->i_ino,
6933 dir->i_generation, ent->oied_namelen,
6934 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6935 *attr |= LUDA_UNKNOWN;
6943 * linkEA recognizes the dirent entry, the FID-in-LMA is
6944 * valid, trusted, in spite of fid_is_sane(fid) or not.
6946 if (*attr & LUDA_VERIFY_DRYRUN) {
6947 *fid = lma->lma_self_fid;
6948 *attr |= LUDA_REPAIR;
6955 dev->od_dirent_journal = 1;
6956 if (hlock != NULL) {
6957 ldiskfs_htree_unlock(hlock);
6960 up_read(&obj->oo_ext_idx_sem);
6966 *fid = lma->lma_self_fid;
6968 /* Update or append the FID-in-dirent. */
6969 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6970 bh, de, hlock, dotdot);
6972 *attr |= LUDA_REPAIR;
6974 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6975 "the dirent, dir = %lu/%u, name = %.*s, "
6976 "ino = %llu, "DFID": rc = %d\n",
6977 devname, dir->i_ino, dir->i_generation,
6978 ent->oied_namelen, ent->oied_name,
6979 ent->oied_ino, PFID(fid), rc);
6981 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6982 if (*attr & LUDA_VERIFY_DRYRUN) {
6983 if (fid_is_sane(fid)) {
6984 *attr |= LUDA_REPAIR;
6985 } else if (dev->od_index == 0) {
6986 lu_igif_build(fid, inode->i_ino,
6987 inode->i_generation);
6988 *attr |= LUDA_UPGRADE;
6996 dev->od_dirent_journal = 1;
6997 if (hlock != NULL) {
6998 ldiskfs_htree_unlock(hlock);
7001 up_read(&obj->oo_ext_idx_sem);
7008 if (unlikely(fid_is_sane(fid))) {
7010 * FID-in-dirent exists, but FID-in-LMA is lost.
7011 * Trust the FID-in-dirent, and add FID-in-LMA.
7013 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7015 *attr |= LUDA_REPAIR;
7017 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7018 "update dirent, dir = %lu/%u, "
7019 "name = %.*s, ino = %llu, "
7021 devname, dir->i_ino, dir->i_generation,
7022 ent->oied_namelen, ent->oied_name,
7023 ent->oied_ino, PFID(fid), rc);
7024 } else if (dev->od_index == 0) {
7025 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7027 * It is probably IGIF object. Only aappend the
7028 * FID-in-dirent. OI scrub will process FID-in-LMA.
7030 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7031 bh, de, hlock, dotdot);
7033 *attr |= LUDA_UPGRADE;
7035 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7036 "after the dirent, dir = %lu/%u, "
7037 "name = %.*s, ino = %llu, "
7039 devname, dir->i_ino, dir->i_generation,
7040 ent->oied_namelen, ent->oied_name,
7041 ent->oied_ino, PFID(fid), rc);
7050 if (hlock != NULL) {
7051 ldiskfs_htree_unlock(hlock);
7053 if (dev->od_dirent_journal != 0)
7054 up_write(&obj->oo_ext_idx_sem);
7056 up_read(&obj->oo_ext_idx_sem);
7060 ldiskfs_journal_stop(jh);
7064 if (rc >= 0 && !dirty)
7065 dev->od_dirent_journal = 0;
7071 * Returns the value at current position from iterator's in memory structure.
7073 * \param di struct osd_it_ea, iterator's in memory structure
7074 * \param attr attr requested for dirent.
7075 * \param lde lustre dirent
7077 * \retval 0 no error and \param lde has correct lustre dirent.
7078 * \retval -ve on error
7080 static inline int osd_it_ea_rec(const struct lu_env *env,
7081 const struct dt_it *di,
7082 struct dt_rec *dtrec, __u32 attr)
7084 struct osd_it_ea *it = (struct osd_it_ea *)di;
7085 struct osd_object *obj = it->oie_obj;
7086 struct osd_device *dev = osd_obj2dev(obj);
7087 struct osd_thread_info *oti = osd_oti_get(env);
7088 struct osd_inode_id *id = &oti->oti_id;
7089 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7090 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7091 __u32 ino = it->oie_dirent->oied_ino;
7096 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7098 if (attr & LUDA_VERIFY) {
7099 if (unlikely(is_remote_parent_ino(dev, ino))) {
7100 attr |= LUDA_IGNORE;
7102 * If the parent is on remote MDT, and there
7103 * is no FID-in-dirent, then we have to get
7104 * the parent FID from the linkEA.
7106 if (!fid_is_sane(fid) &&
7107 it->oie_dirent->oied_namelen == 2 &&
7108 it->oie_dirent->oied_name[0] == '.' &&
7109 it->oie_dirent->oied_name[1] == '.')
7110 osd_get_pfid_from_linkea(env, obj, fid);
7112 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7116 if (!fid_is_sane(fid))
7117 attr |= LUDA_UNKNOWN;
7119 attr &= ~LU_DIRENT_ATTRS_MASK;
7120 if (!fid_is_sane(fid)) {
7121 bool is_dotdot = false;
7123 if (it->oie_dirent->oied_namelen == 2 &&
7124 it->oie_dirent->oied_name[0] == '.' &&
7125 it->oie_dirent->oied_name[1] == '.')
7128 * If the parent is on remote MDT, and there
7129 * is no FID-in-dirent, then we have to get
7130 * the parent FID from the linkEA.
7132 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7133 rc = osd_get_pfid_from_linkea(env, obj, fid);
7135 if (is_dotdot == false &&
7136 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7139 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7142 osd_id_gen(id, ino, OSD_OII_NOGEN);
7146 /* Pack the entry anyway, at least the offset is right. */
7147 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7148 it->oie_dirent->oied_name,
7149 it->oie_dirent->oied_namelen,
7150 it->oie_dirent->oied_type, attr);
7155 if (osd_remote_fid(env, dev, fid))
7158 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
7159 osd_add_oi_cache(oti, dev, id, fid);
7161 RETURN(rc > 0 ? 0 : rc);
7165 * Returns the record size size at current position.
7167 * This function will return record(lu_dirent) size in bytes.
7169 * \param[in] env execution environment
7170 * \param[in] di iterator's in memory structure
7171 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7172 * calculate the lu_dirent size.
7174 * \retval record size(in bytes & in memory) of the current lu_dirent
7177 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7180 struct osd_it_ea *it = (struct osd_it_ea *)di;
7182 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7186 * Returns a cookie for current position of the iterator head, so that
7187 * user can use this cookie to load/start the iterator next time.
7189 * \param di iterator's in memory structure
7191 * \retval cookie for current position, on success
7193 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7195 struct osd_it_ea *it = (struct osd_it_ea *)di;
7197 return it->oie_dirent->oied_off;
7201 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
7202 * to load a directory entry at a time and stored it i inn,
7203 * in iterator's in-memory data structure.
7205 * \param di struct osd_it_ea, iterator's in memory structure
7207 * \retval +ve on success
7208 * \retval -ve on error
7210 static int osd_it_ea_load(const struct lu_env *env,
7211 const struct dt_it *di, __u64 hash)
7213 struct osd_it_ea *it = (struct osd_it_ea *)di;
7217 it->oie_file.f_pos = hash;
7219 rc = osd_ldiskfs_it_fill(env, di);
7230 * Index lookup function for interoperability mode (b11826).
7232 * \param key, key i.e. file name to be searched
7234 * \retval +ve, on success
7235 * \retval -ve, on error
7237 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7238 struct dt_rec *rec, const struct dt_key *key)
7240 struct osd_object *obj = osd_dt_obj(dt);
7245 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7246 LINVRNT(osd_invariant(obj));
7248 rc = osd_ea_lookup_rec(env, obj, rec, key);
7255 * Index and Iterator operations for interoperability
7256 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7258 static const struct dt_index_operations osd_index_ea_ops = {
7259 .dio_lookup = osd_index_ea_lookup,
7260 .dio_declare_insert = osd_index_declare_ea_insert,
7261 .dio_insert = osd_index_ea_insert,
7262 .dio_declare_delete = osd_index_declare_ea_delete,
7263 .dio_delete = osd_index_ea_delete,
7265 .init = osd_it_ea_init,
7266 .fini = osd_it_ea_fini,
7267 .get = osd_it_ea_get,
7268 .put = osd_it_ea_put,
7269 .next = osd_it_ea_next,
7270 .key = osd_it_ea_key,
7271 .key_size = osd_it_ea_key_size,
7272 .rec = osd_it_ea_rec,
7273 .rec_size = osd_it_ea_rec_size,
7274 .store = osd_it_ea_store,
7275 .load = osd_it_ea_load
7279 static void *osd_key_init(const struct lu_context *ctx,
7280 struct lu_context_key *key)
7282 struct osd_thread_info *info;
7284 OBD_ALLOC_PTR(info);
7286 return ERR_PTR(-ENOMEM);
7288 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7289 if (info->oti_it_ea_buf == NULL)
7292 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7294 info->oti_hlock = ldiskfs_htree_lock_alloc();
7295 if (info->oti_hlock == NULL)
7301 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7304 return ERR_PTR(-ENOMEM);
7307 static void osd_key_fini(const struct lu_context *ctx,
7308 struct lu_context_key *key, void *data)
7310 struct osd_thread_info *info = data;
7311 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7312 struct osd_idmap_cache *idc = info->oti_ins_cache;
7314 if (info->oti_dio_pages) {
7316 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7317 if (info->oti_dio_pages[i])
7318 __free_page(info->oti_dio_pages[i]);
7320 OBD_FREE(info->oti_dio_pages,
7321 sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
7324 if (info->oti_inode != NULL)
7326 if (info->oti_hlock != NULL)
7327 ldiskfs_htree_lock_free(info->oti_hlock);
7328 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7329 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7330 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7331 lu_buf_free(&info->oti_big_buf);
7333 LASSERT(info->oti_ins_cache_size > 0);
7334 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
7335 info->oti_ins_cache = NULL;
7336 info->oti_ins_cache_size = 0;
7341 static void osd_key_exit(const struct lu_context *ctx,
7342 struct lu_context_key *key, void *data)
7344 struct osd_thread_info *info = data;
7346 LASSERT(info->oti_r_locks == 0);
7347 LASSERT(info->oti_w_locks == 0);
7348 LASSERT(info->oti_txns == 0);
7351 /* type constructor/destructor: osd_type_init, osd_type_fini */
7352 LU_TYPE_INIT_FINI(osd, &osd_key);
7354 struct lu_context_key osd_key = {
7355 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7356 .lct_init = osd_key_init,
7357 .lct_fini = osd_key_fini,
7358 .lct_exit = osd_key_exit
7362 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7363 const char *name, struct lu_device *next)
7365 struct osd_device *osd = osd_dev(d);
7367 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7368 sizeof(osd->od_svname))
7370 return osd_procfs_init(osd, name);
7373 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7375 struct seq_server_site *ss = osd_seq_site(osd);
7380 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7383 if (unlikely(ss == NULL))
7386 OBD_ALLOC_PTR(osd->od_cl_seq);
7387 if (osd->od_cl_seq == NULL)
7390 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7391 osd->od_svname, ss->ss_server_seq);
7393 OBD_FREE_PTR(osd->od_cl_seq);
7394 osd->od_cl_seq = NULL;
7398 if (ss->ss_node_id == 0) {
7400 * If the OSD on the sequence controller(MDT0), then allocate
7401 * sequence here, otherwise allocate sequence after connected
7402 * to MDT0 (see mdt_register_lwp_callback()).
7404 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7405 &osd->od_cl_seq->lcs_space, env);
7411 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7413 if (osd->od_cl_seq == NULL)
7416 seq_client_fini(osd->od_cl_seq);
7417 OBD_FREE_PTR(osd->od_cl_seq);
7418 osd->od_cl_seq = NULL;
7421 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7425 /* shutdown quota slave instance associated with the device */
7426 if (o->od_quota_slave != NULL) {
7427 struct qsd_instance *qsd = o->od_quota_slave;
7429 o->od_quota_slave = NULL;
7433 osd_fid_fini(env, o);
7434 osd_scrub_cleanup(env, o);
7439 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7443 if (o->od_mnt != NULL) {
7444 shrink_dcache_sb(osd_sb(o));
7445 osd_sync(env, &o->od_dt_dev);
7454 static int osd_mount(const struct lu_env *env,
7455 struct osd_device *o, struct lustre_cfg *cfg)
7457 const char *name = lustre_cfg_string(cfg, 0);
7458 const char *dev = lustre_cfg_string(cfg, 1);
7460 unsigned long page, s_flags, lmd_flags = 0;
7461 struct page *__page;
7462 struct file_system_type *type;
7463 char *options = NULL;
7465 struct osd_thread_info *info = osd_oti_get(env);
7466 struct lu_fid *fid = &info->oti_fid;
7467 struct inode *inode;
7468 int rc = 0, force_over_512tb = 0;
7472 if (o->od_mnt != NULL)
7475 if (strlen(dev) >= sizeof(o->od_mntdev))
7477 strcpy(o->od_mntdev, dev);
7479 str = lustre_cfg_string(cfg, 2);
7480 s_flags = simple_strtoul(str, NULL, 0);
7481 str = strstr(str, ":");
7483 lmd_flags = simple_strtoul(str + 1, NULL, 0);
7484 opts = lustre_cfg_string(cfg, 3);
7486 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7487 CERROR("%s: device %s extents feature is not guaranteed to "
7488 "work on big-endian systems. Use \"bigendian_extents\" "
7489 "mount option to override.\n", name, dev);
7493 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7494 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7495 CWARN("force_over_128tb option is deprecated. "
7496 "Filesystems less than 512TB can be created without any "
7497 "force options. Use force_over_512tb option for "
7498 "filesystems greater than 512TB.\n");
7501 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7502 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7503 CWARN("force_over_256tb option is deprecated. "
7504 "Filesystems less than 512TB can be created without any "
7505 "force options. Use force_over_512tb option for "
7506 "filesystems greater than 512TB.\n");
7510 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
7511 force_over_512tb = 1;
7513 __page = alloc_page(GFP_KERNEL);
7515 GOTO(out, rc = -ENOMEM);
7516 page = (unsigned long)page_address(__page);
7517 options = (char *)page;
7520 /* strip out the options for back compatiblity */
7521 static char *sout[] = {
7528 /* strip out option we processed in osd */
7529 "bigendian_extents",
7535 strcat(options, opts);
7536 for (rc = 0, str = options; sout[rc]; ) {
7537 char *op = strstr(str, sout[rc]);
7544 if (op == options || *(op - 1) == ',') {
7545 str = op + strlen(sout[rc]);
7546 if (*str == ',' || *str == '\0') {
7547 *str == ',' ? str++ : str;
7548 memmove(op, str, strlen(str) + 1);
7551 for (str = op; *str != ',' && *str != '\0'; str++)
7555 strncat(options, "user_xattr,acl", 14);
7558 /* Glom up mount options */
7559 if (*options != '\0')
7560 strcat(options, ",");
7561 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7563 type = get_fs_type("ldiskfs");
7565 CERROR("%s: cannot find ldiskfs module\n", name);
7566 GOTO(out, rc = -ENODEV);
7569 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7570 module_put(type->owner);
7572 if (IS_ERR(o->od_mnt)) {
7573 rc = PTR_ERR(o->od_mnt);
7575 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7579 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7580 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7581 force_over_512tb == 0) {
7582 CERROR("%s: device %s LDISKFS does not support filesystems "
7583 "greater than 512TB and can cause data corruption. "
7584 "Use \"force_over_512tb\" mount option to override.\n",
7586 GOTO(out_mnt, rc = -EINVAL);
7589 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7590 if (priv_dev_set_rdonly) {
7591 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
7592 o->od_dt_dev.dd_rdonly = 1;
7593 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
7596 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7599 GOTO(out_mnt, rc = -EOPNOTSUPP);
7601 } else if (priv_dev_check_rdonly &&
7602 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
7603 CERROR("%s: underlying device %s is marked as "
7604 "read-only. Setup failed\n", name, dev);
7606 GOTO(out_mnt, rc = -EROFS);
7609 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7610 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7611 GOTO(out_mnt, rc = -EINVAL);
7614 #ifdef LDISKFS_MOUNT_DIRDATA
7615 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7616 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7617 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7618 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7619 "enabling the dirdata feature. If you do not want to "
7620 "downgrade to Lustre-1.x again, you can enable it via "
7621 "'tune2fs -O dirdata device'\n", name, dev);
7623 inode = osd_sb(o)->s_root->d_inode;
7624 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7625 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7627 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7631 if (lmd_flags & LMD_FLG_NOSCRUB)
7632 o->od_auto_scrub_interval = AS_NEVER;
7634 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7635 /* do not use pagecache with flash-backed storage */
7636 o->od_writethrough_cache = 0;
7637 o->od_read_cache = 0;
7648 __free_page(__page);
7653 static struct lu_device *osd_device_fini(const struct lu_env *env,
7654 struct lu_device *d)
7656 struct osd_device *o = osd_dev(d);
7660 osd_index_backup(env, o, false);
7661 osd_shutdown(env, o);
7663 osd_obj_map_fini(o);
7669 static int osd_device_init0(const struct lu_env *env,
7670 struct osd_device *o,
7671 struct lustre_cfg *cfg)
7673 struct lu_device *l = osd2lu_dev(o);
7674 struct osd_thread_info *info;
7678 /* if the module was re-loaded, env can loose its keys */
7679 rc = lu_env_refill((struct lu_env *)env);
7682 info = osd_oti_get(env);
7685 l->ld_ops = &osd_lu_ops;
7686 o->od_dt_dev.dd_ops = &osd_dt_ops;
7688 spin_lock_init(&o->od_osfs_lock);
7689 mutex_init(&o->od_otable_mutex);
7690 INIT_LIST_HEAD(&o->od_orphan_list);
7691 INIT_LIST_HEAD(&o->od_index_backup_list);
7692 INIT_LIST_HEAD(&o->od_index_restore_list);
7693 spin_lock_init(&o->od_lock);
7694 o->od_index_backup_policy = LIBP_NONE;
7696 o->od_read_cache = 1;
7697 o->od_writethrough_cache = 1;
7698 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7699 o->od_auto_scrub_interval = AS_DEFAULT;
7701 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7702 sizeof(o->od_svname));
7703 if (cplen >= sizeof(o->od_svname)) {
7708 o->od_index_backup_stop = 0;
7709 o->od_index = -1; /* -1 means index is invalid */
7710 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7711 if (rc == LDD_F_SV_TYPE_OST)
7714 o->od_full_scrub_ratio = OFSR_DEFAULT;
7715 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7716 rc = osd_mount(env, o, cfg);
7720 rc = osd_obj_map_init(env, o);
7724 rc = lu_site_init(&o->od_site, l);
7726 GOTO(out_compat, rc);
7727 o->od_site.ls_bottom_dev = l;
7729 rc = lu_site_init_finish(&o->od_site);
7733 INIT_LIST_HEAD(&o->od_ios_list);
7734 /* setup scrub, including OI files initialization */
7736 rc = osd_scrub_setup(env, o);
7741 rc = osd_procfs_init(o, o->od_svname);
7743 CERROR("%s: can't initialize procfs: rc = %d\n",
7745 GOTO(out_scrub, rc);
7748 LASSERT(l->ld_site->ls_linkage.next != NULL);
7749 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7751 /* initialize quota slave instance */
7752 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7754 if (IS_ERR(o->od_quota_slave)) {
7755 rc = PTR_ERR(o->od_quota_slave);
7756 o->od_quota_slave = NULL;
7757 GOTO(out_procfs, rc);
7765 osd_scrub_cleanup(env, o);
7767 lu_site_fini(&o->od_site);
7769 osd_obj_map_fini(o);
7776 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7777 struct lu_device_type *t,
7778 struct lustre_cfg *cfg)
7780 struct osd_device *o;
7785 return ERR_PTR(-ENOMEM);
7787 rc = dt_device_init(&o->od_dt_dev, t);
7790 * Because the ctx might be revived in dt_device_init,
7791 * refill the env here
7793 lu_env_refill((struct lu_env *)env);
7794 rc = osd_device_init0(env, o, cfg);
7796 dt_device_fini(&o->od_dt_dev);
7799 if (unlikely(rc != 0))
7802 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7805 static struct lu_device *osd_device_free(const struct lu_env *env,
7806 struct lu_device *d)
7808 struct osd_device *o = osd_dev(d);
7812 /* XXX: make osd top device in order to release reference */
7813 d->ld_site->ls_top_dev = d;
7814 lu_site_purge(env, d->ld_site, -1);
7815 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7816 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7817 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7819 lu_site_fini(&o->od_site);
7820 dt_device_fini(&o->od_dt_dev);
7825 static int osd_process_config(const struct lu_env *env,
7826 struct lu_device *d, struct lustre_cfg *cfg)
7828 struct osd_device *o = osd_dev(d);
7833 switch (cfg->lcfg_command) {
7835 rc = osd_mount(env, o, cfg);
7839 * For the case LCFG_PRE_CLEANUP is not called in advance,
7840 * that may happend if hit failure during mount process.
7842 osd_index_backup(env, o, false);
7843 lu_dev_del_linkage(d->ld_site, d);
7844 rc = osd_shutdown(env, o);
7847 LASSERT(&o->od_dt_dev);
7848 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7849 cfg, &o->od_dt_dev);
7850 if (rc > 0 || rc == -ENOSYS) {
7851 rc = class_process_proc_param(PARAM_OST,
7852 lprocfs_osd_obd_vars,
7853 cfg, &o->od_dt_dev);
7858 case LCFG_PRE_CLEANUP:
7860 osd_index_backup(env, o,
7861 o->od_index_backup_policy != LIBP_NONE);
7871 static int osd_recovery_complete(const struct lu_env *env,
7872 struct lu_device *d)
7874 struct osd_device *osd = osd_dev(d);
7879 if (osd->od_quota_slave == NULL)
7883 * start qsd instance on recovery completion, this notifies the quota
7884 * slave code that we are about to process new requests now
7886 rc = qsd_start(env, osd->od_quota_slave);
7891 * we use exports to track all osd users
7893 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7894 struct obd_device *obd, struct obd_uuid *cluuid,
7895 struct obd_connect_data *data, void *localdata)
7897 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7898 struct lustre_handle conn;
7903 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7905 rc = class_connect(&conn, obd, cluuid);
7909 *exp = class_conn2export(&conn);
7911 spin_lock(&osd->od_osfs_lock);
7913 spin_unlock(&osd->od_osfs_lock);
7919 * once last export (we don't count self-export) disappeared
7920 * osd can be released
7922 static int osd_obd_disconnect(struct obd_export *exp)
7924 struct obd_device *obd = exp->exp_obd;
7925 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7926 int rc, release = 0;
7930 /* Only disconnect the underlying layers on the final disconnect. */
7931 spin_lock(&osd->od_osfs_lock);
7933 if (osd->od_connects == 0)
7935 spin_unlock(&osd->od_osfs_lock);
7937 rc = class_disconnect(exp); /* bz 9811 */
7939 if (rc == 0 && release)
7940 class_manual_cleanup(obd);
7944 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7945 struct lu_device *dev)
7947 struct osd_device *osd = osd_dev(dev);
7948 struct lr_server_data *lsd =
7949 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7954 if (osd->od_quota_slave != NULL) {
7955 /* set up quota slave objects */
7956 result = qsd_prepare(env, osd->od_quota_slave);
7961 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7962 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7963 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7964 osd->od_index_in_idif = 1;
7966 osd->od_index_in_idif = 0;
7967 result = osd_register_proc_index_in_idif(osd);
7972 osd->od_index_in_idif = 1;
7976 result = osd_fid_init(env, osd);
7981 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7982 struct lu_fid *fid, struct md_op_data *op_data)
7984 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7986 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7989 static const struct lu_object_operations osd_lu_obj_ops = {
7990 .loo_object_init = osd_object_init,
7991 .loo_object_delete = osd_object_delete,
7992 .loo_object_release = osd_object_release,
7993 .loo_object_free = osd_object_free,
7994 .loo_object_print = osd_object_print,
7995 .loo_object_invariant = osd_object_invariant
7998 const struct lu_device_operations osd_lu_ops = {
7999 .ldo_object_alloc = osd_object_alloc,
8000 .ldo_process_config = osd_process_config,
8001 .ldo_recovery_complete = osd_recovery_complete,
8002 .ldo_prepare = osd_prepare,
8005 static const struct lu_device_type_operations osd_device_type_ops = {
8006 .ldto_init = osd_type_init,
8007 .ldto_fini = osd_type_fini,
8009 .ldto_start = osd_type_start,
8010 .ldto_stop = osd_type_stop,
8012 .ldto_device_alloc = osd_device_alloc,
8013 .ldto_device_free = osd_device_free,
8015 .ldto_device_init = osd_device_init,
8016 .ldto_device_fini = osd_device_fini
8019 static struct lu_device_type osd_device_type = {
8020 .ldt_tags = LU_DEVICE_DT,
8021 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8022 .ldt_ops = &osd_device_type_ops,
8023 .ldt_ctx_tags = LCT_LOCAL,
8026 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8028 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8029 struct super_block *sb = osd_sb(osd);
8031 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
8035 * lprocfs legacy support.
8037 static struct obd_ops osd_obd_device_ops = {
8038 .o_owner = THIS_MODULE,
8039 .o_connect = osd_obd_connect,
8040 .o_disconnect = osd_obd_disconnect,
8041 .o_fid_alloc = osd_fid_alloc,
8042 .o_health_check = osd_health_check,
8045 static int __init osd_init(void)
8049 CLASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
8050 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8051 /* please, try to keep osd_thread_info smaller than a page */
8052 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
8057 rc = lu_kmem_init(ldiskfs_caches);
8061 #ifdef CONFIG_KALLSYMS
8062 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
8063 priv_dev_check_rdonly =
8064 (void *)kallsyms_lookup_name("dev_check_rdonly");
8067 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8068 lprocfs_osd_module_vars,
8069 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8071 lu_kmem_fini(ldiskfs_caches);
8075 static void __exit osd_exit(void)
8077 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8078 lu_kmem_fini(ldiskfs_caches);
8081 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8082 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8083 MODULE_VERSION(LUSTRE_VERSION_STRING);
8084 MODULE_LICENSE("GPL");
8086 module_init(osd_init);
8087 module_exit(osd_exit);