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_def_qsd(osd);
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_def_qsd(osd_obj2dev(obj));
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.");
2317 char type_number = name[type_off - 2];
2319 if (interval != 512 && interval != 4096) {
2320 CERROR("%s: unsupported T10PI sector size %u\n",
2321 d->od_svname, interval);
2322 } else if (type_number != '1' && type_number != '3') {
2323 CERROR("%s: unsupported T10PI type %s\n",
2324 d->od_svname, name);
2325 } else if (strcmp(name + type_off, "CRC") == 0) {
2326 d->od_t10_type = type_number == '1' ?
2327 OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2328 param->ddp_t10_cksum_type = interval == 512 ?
2329 OBD_CKSUM_T10CRC512 :
2331 } else if (strcmp(name + type_off, "IP") == 0) {
2332 d->od_t10_type = type_number == '1' ?
2333 OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2334 param->ddp_t10_cksum_type = interval == 512 ?
2335 OBD_CKSUM_T10IP512 :
2338 CERROR("%s: unsupported checksum type of "
2340 d->od_svname, name);
2344 CERROR("%s: unsupported T10PI type '%s'",
2345 d->od_svname, name);
2351 * Concurrency: shouldn't matter.
2353 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2356 struct super_block *s = osd_sb(osd_dt_dev(d));
2359 down_read(&s->s_umount);
2360 rc = s->s_op->sync_fs(s, 1);
2361 up_read(&s->s_umount);
2363 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2370 * Start commit for OSD device.
2372 * An implementation of dt_commit_async method for OSD device.
2373 * Asychronously starts underlayng fs sync and thereby a transaction
2376 * \param env environment
2377 * \param d dt device
2379 * \see dt_device_operations
2381 static int osd_commit_async(const struct lu_env *env,
2382 struct dt_device *d)
2384 struct super_block *s = osd_sb(osd_dt_dev(d));
2389 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2390 down_read(&s->s_umount);
2391 rc = s->s_op->sync_fs(s, 0);
2392 up_read(&s->s_umount);
2397 /* Our own copy of the set readonly functions if present, or NU if not. */
2398 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2399 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2400 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2403 * Concurrency: shouldn't matter.
2405 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2407 struct super_block *sb = osd_sb(osd_dt_dev(d));
2408 struct block_device *dev = sb->s_bdev;
2409 int rc = -EOPNOTSUPP;
2413 if (priv_dev_set_rdonly) {
2414 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2417 CERROR("*** setting %s read-only ***\n",
2418 osd_dt_dev(d)->od_svname);
2420 if (sb->s_op->freeze_fs) {
2421 rc = sb->s_op->freeze_fs(sb);
2426 if (jdev && (jdev != dev)) {
2427 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2429 priv_dev_set_rdonly(jdev);
2431 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2432 priv_dev_set_rdonly(dev);
2434 if (sb->s_op->unfreeze_fs)
2435 sb->s_op->unfreeze_fs(sb);
2440 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2441 osd_dt_dev(d)->od_svname, (long)dev, rc);
2447 * Note: we do not count into QUOTA here.
2448 * If we mount with --data_journal we may need more.
2450 const int osd_dto_credits_noquota[DTO_NR] = {
2453 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2454 * SINGLEDATA_TRANS_BLOCKS(8)
2455 * XXX Note: maybe iam need more, since iam have more level than
2458 [DTO_INDEX_INSERT] = 16,
2461 * just modify a single entry, probably merge few within a block
2463 [DTO_INDEX_DELETE] = 1,
2467 [DTO_INDEX_UPDATE] = 16,
2469 * 4(inode, inode bits, groups, GDT)
2470 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2472 [DTO_OBJECT_CREATE] = 4,
2474 * 4(inode, inode bits, groups, GDT)
2475 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2477 [DTO_OBJECT_DELETE] = 4,
2479 * Attr set credits (inode)
2481 [DTO_ATTR_SET_BASE] = 1,
2483 * Xattr set. The same as xattr of EXT3.
2484 * DATA_TRANS_BLOCKS(14)
2485 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2486 * are also counted in. Do not know why?
2488 [DTO_XATTR_SET] = 14,
2490 * credits for inode change during write.
2492 [DTO_WRITE_BASE] = 3,
2494 * credits for single block write.
2496 [DTO_WRITE_BLOCK] = 14,
2498 * Attr set credits for chown.
2499 * This is extra credits for setattr, and it is null without quota
2501 [DTO_ATTR_SET_CHOWN] = 0
2504 static const struct dt_device_operations osd_dt_ops = {
2505 .dt_root_get = osd_root_get,
2506 .dt_statfs = osd_statfs,
2507 .dt_trans_create = osd_trans_create,
2508 .dt_trans_start = osd_trans_start,
2509 .dt_trans_stop = osd_trans_stop,
2510 .dt_trans_cb_add = osd_trans_cb_add,
2511 .dt_conf_get = osd_conf_get,
2512 .dt_sync = osd_sync,
2514 .dt_commit_async = osd_commit_async,
2517 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2520 struct osd_object *obj = osd_dt_obj(dt);
2521 struct osd_thread_info *oti = osd_oti_get(env);
2523 LINVRNT(osd_invariant(obj));
2525 LASSERT(obj->oo_owner != env);
2526 down_read_nested(&obj->oo_sem, role);
2528 LASSERT(obj->oo_owner == NULL);
2532 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2535 struct osd_object *obj = osd_dt_obj(dt);
2536 struct osd_thread_info *oti = osd_oti_get(env);
2538 LINVRNT(osd_invariant(obj));
2540 LASSERT(obj->oo_owner != env);
2541 down_write_nested(&obj->oo_sem, role);
2543 LASSERT(obj->oo_owner == NULL);
2544 obj->oo_owner = env;
2548 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2550 struct osd_object *obj = osd_dt_obj(dt);
2551 struct osd_thread_info *oti = osd_oti_get(env);
2553 LINVRNT(osd_invariant(obj));
2555 LASSERT(oti->oti_r_locks > 0);
2557 up_read(&obj->oo_sem);
2560 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2562 struct osd_object *obj = osd_dt_obj(dt);
2563 struct osd_thread_info *oti = osd_oti_get(env);
2565 LINVRNT(osd_invariant(obj));
2567 LASSERT(obj->oo_owner == env);
2568 LASSERT(oti->oti_w_locks > 0);
2570 obj->oo_owner = NULL;
2571 up_write(&obj->oo_sem);
2574 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2576 struct osd_object *obj = osd_dt_obj(dt);
2578 LINVRNT(osd_invariant(obj));
2580 return obj->oo_owner == env;
2583 static struct timespec *osd_inode_time(const struct lu_env *env,
2584 struct inode *inode, __u64 seconds)
2586 struct osd_thread_info *oti = osd_oti_get(env);
2587 struct timespec *t = &oti->oti_time;
2589 t->tv_sec = seconds;
2591 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2595 static void osd_inode_getattr(const struct lu_env *env,
2596 struct inode *inode, struct lu_attr *attr)
2598 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2599 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2600 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2601 LA_BLKSIZE | LA_TYPE;
2603 attr->la_atime = LTIME_S(inode->i_atime);
2604 attr->la_mtime = LTIME_S(inode->i_mtime);
2605 attr->la_ctime = LTIME_S(inode->i_ctime);
2606 attr->la_mode = inode->i_mode;
2607 attr->la_size = i_size_read(inode);
2608 attr->la_blocks = inode->i_blocks;
2609 attr->la_uid = i_uid_read(inode);
2610 attr->la_gid = i_gid_read(inode);
2611 attr->la_projid = i_projid_read(inode);
2612 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2613 attr->la_nlink = inode->i_nlink;
2614 attr->la_rdev = inode->i_rdev;
2615 attr->la_blksize = 1 << inode->i_blkbits;
2616 attr->la_blkbits = inode->i_blkbits;
2618 * Ext4 did not transfer inherit flags from raw inode
2619 * to inode flags, and ext4 internally test raw inode
2620 * @i_flags directly. Instead of patching ext4, we do it here.
2622 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2623 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2626 static int osd_attr_get(const struct lu_env *env,
2627 struct dt_object *dt,
2628 struct lu_attr *attr)
2630 struct osd_object *obj = osd_dt_obj(dt);
2632 if (unlikely(!dt_object_exists(dt)))
2634 if (unlikely(obj->oo_destroyed))
2637 LASSERT(!dt_object_remote(dt));
2638 LINVRNT(osd_invariant(obj));
2640 spin_lock(&obj->oo_guard);
2641 osd_inode_getattr(env, obj->oo_inode, attr);
2642 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2643 attr->la_flags |= LUSTRE_ORPHAN_FL;
2644 spin_unlock(&obj->oo_guard);
2649 static int osd_declare_attr_qid(const struct lu_env *env,
2650 struct osd_object *obj,
2651 struct osd_thandle *oh, long long bspace,
2652 qid_t old_id, qid_t new_id, bool enforce,
2653 unsigned int type, bool ignore_edquot)
2656 struct osd_thread_info *info = osd_oti_get(env);
2657 struct lquota_id_info *qi = &info->oti_qi;
2659 qi->lqi_type = type;
2660 /* inode accounting */
2661 qi->lqi_is_blk = false;
2663 /* one more inode for the new id ... */
2664 qi->lqi_id.qid_uid = new_id;
2666 /* Reserve credits for the new id */
2667 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2668 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2673 /* and one less inode for the current id */
2674 qi->lqi_id.qid_uid = old_id;
2676 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2677 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2682 /* block accounting */
2683 qi->lqi_is_blk = true;
2685 /* more blocks for the new id ... */
2686 qi->lqi_id.qid_uid = new_id;
2687 qi->lqi_space = bspace;
2689 * Credits for the new uid has been reserved, re-use "obj"
2690 * to save credit reservation.
2692 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2693 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2698 /* and finally less blocks for the current uid */
2699 qi->lqi_id.qid_uid = old_id;
2700 qi->lqi_space = -bspace;
2701 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2702 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2708 static int osd_declare_attr_set(const struct lu_env *env,
2709 struct dt_object *dt,
2710 const struct lu_attr *attr,
2711 struct thandle *handle)
2713 struct osd_thandle *oh;
2714 struct osd_object *obj;
2723 LASSERT(dt != NULL);
2724 LASSERT(handle != NULL);
2726 obj = osd_dt_obj(dt);
2727 LASSERT(osd_invariant(obj));
2729 oh = container_of0(handle, struct osd_thandle, ot_super);
2730 LASSERT(oh->ot_handle == NULL);
2732 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2733 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2735 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2736 osd_dto_credits_noquota[DTO_XATTR_SET]);
2738 if (attr == NULL || obj->oo_inode == NULL)
2741 bspace = obj->oo_inode->i_blocks << 9;
2742 bspace = toqb(bspace);
2745 * Changing ownership is always preformed by super user, it should not
2746 * fail with EDQUOT unless required explicitly.
2748 * We still need to call the osd_declare_qid() to calculate the journal
2749 * credits for updating quota accounting files and to trigger quota
2750 * space adjustment once the operation is completed.
2752 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2753 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2756 CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2757 "(the quota space is %lld)\n",
2758 obj->oo_inode->i_sb->s_id, attr->la_uid,
2759 attr->la_gid, bspace);
2762 uid = i_uid_read(obj->oo_inode);
2763 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2764 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2765 attr->la_uid, enforce, USRQUOTA,
2770 gid = i_gid_read(obj->oo_inode);
2771 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2772 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2773 i_gid_read(obj->oo_inode),
2774 attr->la_gid, enforce, GRPQUOTA,
2780 #ifdef HAVE_PROJECT_QUOTA
2781 if (attr->la_valid & LA_PROJID) {
2782 __u32 projid = i_projid_read(obj->oo_inode);
2784 enforce = (attr->la_valid & LA_PROJID) &&
2785 (attr->la_projid != projid);
2786 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2787 (qid_t)projid, (qid_t)attr->la_projid,
2788 enforce, PRJQUOTA, true);
2796 static int osd_inode_setattr(const struct lu_env *env,
2797 struct inode *inode, const struct lu_attr *attr)
2799 __u64 bits = attr->la_valid;
2801 /* Only allow set size for regular file */
2802 if (!S_ISREG(inode->i_mode))
2803 bits &= ~(LA_SIZE | LA_BLOCKS);
2808 if (bits & LA_ATIME)
2809 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2810 if (bits & LA_CTIME)
2811 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2812 if (bits & LA_MTIME)
2813 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2814 if (bits & LA_SIZE) {
2815 spin_lock(&inode->i_lock);
2816 LDISKFS_I(inode)->i_disksize = attr->la_size;
2817 i_size_write(inode, attr->la_size);
2818 spin_unlock(&inode->i_lock);
2822 * OSD should not change "i_blocks" which is used by quota.
2823 * "i_blocks" should be changed by ldiskfs only.
2826 inode->i_mode = (inode->i_mode & S_IFMT) |
2827 (attr->la_mode & ~S_IFMT);
2829 i_uid_write(inode, attr->la_uid);
2831 i_gid_write(inode, attr->la_gid);
2832 if (bits & LA_PROJID)
2833 i_projid_write(inode, attr->la_projid);
2834 if (bits & LA_NLINK)
2835 set_nlink(inode, attr->la_nlink);
2837 inode->i_rdev = attr->la_rdev;
2839 if (bits & LA_FLAGS) {
2840 /* always keep S_NOCMTIME */
2841 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2844 * Ext4 did not transfer inherit flags from
2845 * @inode->i_flags to raw inode i_flags when writing
2846 * flags, we do it explictly here.
2848 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2849 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2851 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2856 #ifdef HAVE_PROJECT_QUOTA
2857 static int osd_transfer_project(struct inode *inode, __u32 projid)
2859 struct super_block *sb = inode->i_sb;
2860 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2863 struct ldiskfs_iloc iloc;
2864 struct ldiskfs_inode *raw_inode;
2865 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2867 if (!ldiskfs_has_feature_project(sb)) {
2868 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2869 == LDISKFS_DEF_PROJID);
2870 if (projid != LDISKFS_DEF_PROJID)
2876 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2879 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2880 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2883 err = ldiskfs_get_inode_loc(inode, &iloc);
2887 raw_inode = ldiskfs_raw_inode(&iloc);
2888 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2895 dquot_initialize(inode);
2896 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2897 if (transfer_to[PRJQUOTA]) {
2898 err = __dquot_transfer(inode, transfer_to);
2899 dqput(transfer_to[PRJQUOTA]);
2908 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2912 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2913 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2916 ll_vfs_dq_init(inode);
2918 if (attr->la_valid & LA_UID)
2919 iattr.ia_valid |= ATTR_UID;
2920 if (attr->la_valid & LA_GID)
2921 iattr.ia_valid |= ATTR_GID;
2922 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2923 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2925 rc = ll_vfs_dq_transfer(inode, &iattr);
2927 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2928 "enforcement enabled on the ldiskfs "
2929 "filesystem?\n", inode->i_sb->s_id, rc);
2934 /* Handle project id transfer here properly */
2935 if (attr->la_valid & LA_PROJID &&
2936 attr->la_projid != i_projid_read(inode)) {
2937 #ifdef HAVE_PROJECT_QUOTA
2938 rc = osd_transfer_project(inode, attr->la_projid);
2943 CERROR("%s: quota transfer failed: rc = %d. Is project "
2944 "enforcement enabled on the ldiskfs "
2945 "filesystem?\n", inode->i_sb->s_id, rc);
2952 static int osd_attr_set(const struct lu_env *env,
2953 struct dt_object *dt,
2954 const struct lu_attr *attr,
2955 struct thandle *handle)
2957 struct osd_object *obj = osd_dt_obj(dt);
2958 struct inode *inode;
2961 if (!dt_object_exists(dt))
2964 LASSERT(handle != NULL);
2965 LASSERT(!dt_object_remote(dt));
2966 LASSERT(osd_invariant(obj));
2968 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2970 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
2971 !osd_obj2dev(obj)->od_is_ost) {
2972 struct osd_thread_info *oti = osd_oti_get(env);
2973 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2974 struct lu_fid *fid1 = &oti->oti_fid;
2975 struct osd_inode_id *id = &oti->oti_id;
2976 struct iam_path_descr *ipd;
2977 struct iam_container *bag;
2978 struct osd_thandle *oh;
2981 fid_cpu_to_be(fid1, fid0);
2982 memset(id, 1, sizeof(*id));
2983 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2984 fid0)->oi_dir.od_container;
2985 ipd = osd_idx_ipd_get(env, bag);
2986 if (unlikely(ipd == NULL))
2989 oh = container_of0(handle, struct osd_thandle, ot_super);
2990 rc = iam_update(oh->ot_handle, bag,
2991 (const struct iam_key *)fid1,
2992 (const struct iam_rec *)id, ipd);
2993 osd_ipd_put(env, bag, ipd);
2994 return(rc > 0 ? 0 : rc);
2997 inode = obj->oo_inode;
2999 rc = osd_quota_transfer(inode, attr);
3003 spin_lock(&obj->oo_guard);
3004 rc = osd_inode_setattr(env, inode, attr);
3005 spin_unlock(&obj->oo_guard);
3009 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3011 if (!(attr->la_valid & LA_FLAGS))
3014 /* Let's check if there are extra flags need to be set into LMA */
3015 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3016 struct osd_thread_info *info = osd_oti_get(env);
3017 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3019 LASSERT(!obj->oo_pfid_in_lma);
3021 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3022 &info->oti_ost_attrs);
3026 lma->lma_incompat |=
3027 lustre_to_lma_flags(attr->la_flags);
3028 lustre_lma_swab(lma);
3029 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3030 lma, sizeof(*lma), XATTR_REPLACE);
3032 struct osd_device *osd = osd_obj2dev(obj);
3034 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3035 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3036 lma->lma_incompat, rc);
3039 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3041 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3044 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3049 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3050 struct osd_object *obj,
3051 const char *name, const int namelen)
3053 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3056 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3057 umode_t mode, struct dt_allocation_hint *hint,
3061 struct osd_device *osd = osd_obj2dev(obj);
3062 struct osd_thandle *oth;
3063 struct dt_object *parent = NULL;
3064 struct inode *inode;
3066 LINVRNT(osd_invariant(obj));
3067 LASSERT(obj->oo_inode == NULL);
3068 LASSERT(obj->oo_hl_head == NULL);
3070 if (S_ISDIR(mode) && ldiskfs_pdo) {
3072 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3073 if (obj->oo_hl_head == NULL)
3077 oth = container_of(th, struct osd_thandle, ot_super);
3078 LASSERT(oth->ot_handle->h_transaction != NULL);
3080 if (hint != NULL && hint->dah_parent != NULL &&
3081 !dt_object_remote(hint->dah_parent))
3082 parent = hint->dah_parent;
3084 inode = ldiskfs_create_inode(oth->ot_handle,
3085 parent ? osd_dt_obj(parent)->oo_inode :
3086 osd_sb(osd)->s_root->d_inode,
3088 if (!IS_ERR(inode)) {
3089 /* Do not update file c/mtime in ldiskfs. */
3090 inode->i_flags |= S_NOCMTIME;
3093 * For new created object, it must be consistent,
3094 * and it is unnecessary to scrub against it.
3096 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3098 obj->oo_inode = inode;
3101 if (obj->oo_hl_head != NULL) {
3102 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3103 obj->oo_hl_head = NULL;
3105 result = PTR_ERR(inode);
3107 LINVRNT(osd_invariant(obj));
3115 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3116 struct lu_attr *attr,
3117 struct dt_allocation_hint *hint,
3118 struct dt_object_format *dof,
3122 struct osd_thandle *oth;
3123 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3125 LASSERT(S_ISDIR(attr->la_mode));
3127 oth = container_of(th, struct osd_thandle, ot_super);
3128 LASSERT(oth->ot_handle->h_transaction != NULL);
3129 result = osd_mkfile(info, obj, mode, hint, th);
3134 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3135 struct lu_attr *attr,
3136 struct dt_allocation_hint *hint,
3137 struct dt_object_format *dof,
3141 struct osd_thandle *oth;
3142 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3144 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3146 LASSERT(S_ISREG(attr->la_mode));
3148 oth = container_of(th, struct osd_thandle, ot_super);
3149 LASSERT(oth->ot_handle->h_transaction != NULL);
3151 result = osd_mkfile(info, obj, mode, hint, th);
3153 LASSERT(obj->oo_inode != NULL);
3154 if (feat->dif_flags & DT_IND_VARKEY)
3155 result = iam_lvar_create(obj->oo_inode,
3156 feat->dif_keysize_max,
3158 feat->dif_recsize_max,
3161 result = iam_lfix_create(obj->oo_inode,
3162 feat->dif_keysize_max,
3164 feat->dif_recsize_max,
3170 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3171 struct lu_attr *attr,
3172 struct dt_allocation_hint *hint,
3173 struct dt_object_format *dof,
3176 LASSERT(S_ISREG(attr->la_mode));
3177 return osd_mkfile(info, obj, (attr->la_mode &
3178 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
3181 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3182 struct lu_attr *attr,
3183 struct dt_allocation_hint *hint,
3184 struct dt_object_format *dof,
3187 LASSERT(S_ISLNK(attr->la_mode));
3188 return osd_mkfile(info, obj, (attr->la_mode &
3189 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
3192 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3193 struct lu_attr *attr,
3194 struct dt_allocation_hint *hint,
3195 struct dt_object_format *dof,
3198 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3201 LINVRNT(osd_invariant(obj));
3202 LASSERT(obj->oo_inode == NULL);
3203 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3204 S_ISFIFO(mode) || S_ISSOCK(mode));
3206 result = osd_mkfile(info, obj, mode, hint, th);
3208 LASSERT(obj->oo_inode != NULL);
3210 * This inode should be marked dirty for i_rdev. Currently
3211 * that is done in the osd_attr_init().
3213 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3216 LINVRNT(osd_invariant(obj));
3220 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3222 struct dt_allocation_hint *hint,
3223 struct dt_object_format *dof,
3226 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3228 osd_obj_type_f result;
3244 result = osd_mk_index;
3255 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3256 struct dt_object *parent, struct dt_object *child,
3261 ah->dah_parent = parent;
3262 ah->dah_mode = child_mode;
3264 if (parent != NULL && !dt_object_remote(parent)) {
3265 /* will help to find FID->ino at dt_insert("..") */
3266 struct osd_object *pobj = osd_dt_obj(parent);
3268 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3272 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3273 struct lu_attr *attr, struct dt_object_format *dof)
3275 struct inode *inode = obj->oo_inode;
3276 __u64 valid = attr->la_valid;
3279 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3281 if (dof->dof_type != DFT_NODE)
3282 attr->la_valid &= ~LA_RDEV;
3283 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
3284 attr->la_valid &= ~LA_ATIME;
3285 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
3286 attr->la_valid &= ~LA_CTIME;
3287 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
3288 attr->la_valid &= ~LA_MTIME;
3290 result = osd_quota_transfer(inode, attr);
3294 if (attr->la_valid != 0) {
3295 result = osd_inode_setattr(info->oti_env, inode, attr);
3297 * The osd_inode_setattr() should always succeed here. The
3298 * only error that could be returned is EDQUOT when we are
3299 * trying to change the UID or GID of the inode. However, this
3300 * should not happen since quota enforcement is no longer
3301 * enabled on ldiskfs (lquota takes care of it).
3303 LASSERTF(result == 0, "%d\n", result);
3304 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3307 attr->la_valid = valid;
3311 * Helper function for osd_create()
3313 * \retval 0, on success
3315 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3316 struct lu_attr *attr, struct dt_allocation_hint *hint,
3317 struct dt_object_format *dof, struct thandle *th)
3322 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3324 /* we drop umask so that permissions we pass are not affected */
3325 umask = current->fs->umask;
3326 current->fs->umask = 0;
3328 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3330 if (likely(obj->oo_inode != NULL)) {
3331 LASSERT(obj->oo_inode->i_state & I_NEW);
3334 * Unlock the inode before attr initialization to avoid
3335 * unnecessary dqget operations. LU-6378
3337 unlock_new_inode(obj->oo_inode);
3340 if (likely(result == 0)) {
3341 osd_attr_init(info, obj, attr, dof);
3342 osd_object_init0(obj);
3345 /* restore previous umask value */
3346 current->fs->umask = umask;
3348 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3354 * Helper function for osd_create()
3356 * \retval 0, on success
3358 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3359 const struct lu_fid *fid, struct thandle *th)
3361 struct osd_thread_info *info = osd_oti_get(env);
3362 struct osd_inode_id *id = &info->oti_id;
3363 struct osd_device *osd = osd_obj2dev(obj);
3364 struct osd_thandle *oh;
3367 LASSERT(obj->oo_inode != NULL);
3369 oh = container_of0(th, struct osd_thandle, ot_super);
3370 LASSERT(oh->ot_handle);
3371 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3373 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3374 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3375 OI_CHECK_FLD, NULL);
3376 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3381 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3382 u64 seq, struct lu_seq_range *range)
3384 struct seq_server_site *ss = osd_seq_site(osd);
3386 if (fid_seq_is_idif(seq)) {
3387 fld_range_set_ost(range);
3388 range->lsr_index = idif_ost_idx(seq);
3392 if (!fid_seq_in_fldb(seq)) {
3393 fld_range_set_mdt(range);
3396 * FIXME: If ss is NULL, it suppose not get lsr_index
3399 range->lsr_index = ss->ss_node_id;
3403 LASSERT(ss != NULL);
3404 fld_range_set_any(range);
3405 /* OSD will only do local fld lookup */
3406 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3409 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3410 struct lu_attr *attr,
3411 struct dt_allocation_hint *hint,
3412 struct dt_object_format *dof,
3413 struct thandle *handle)
3415 struct osd_thandle *oh;
3420 LASSERT(handle != NULL);
3422 oh = container_of0(handle, struct osd_thandle, ot_super);
3423 LASSERT(oh->ot_handle == NULL);
3426 * EA object consumes more credits than regular object: osd_mk_index
3427 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3428 * leaf_node, could involves the block, block bitmap, groups, GDT
3429 * change for each block, so add 4 * 2 credits in that case.
3431 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3432 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3433 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3435 * Reuse idle OI block may cause additional one OI block
3438 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3439 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3441 /* will help to find FID->ino mapping at dt_insert() */
3442 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3450 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3451 attr->la_projid, 1, oh, osd_dt_obj(dt),
3452 NULL, OSD_QID_INODE);
3460 * Called to destroy on-disk representation of the object
3462 * Concurrency: must be locked
3464 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3467 struct osd_object *obj = osd_dt_obj(dt);
3468 struct inode *inode = obj->oo_inode;
3469 struct osd_thandle *oh;
3477 oh = container_of0(th, struct osd_thandle, ot_super);
3478 LASSERT(oh->ot_handle == NULL);
3480 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3481 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3483 /* For removing agent entry */
3484 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3485 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3488 * Recycle idle OI leaf may cause additional three OI blocks
3491 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3492 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3493 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3494 /* one less inode */
3495 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3496 i_projid_read(inode), -1, oh, obj, NULL,
3500 /* data to be truncated */
3501 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3502 i_projid_read(inode), 0, oh, obj, NULL,
3508 * will help to find FID->ino when this object is being
3511 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3516 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3519 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3520 struct osd_object *obj = osd_dt_obj(dt);
3521 struct inode *inode = obj->oo_inode;
3522 struct osd_device *osd = osd_obj2dev(obj);
3523 struct osd_thandle *oh;
3528 oh = container_of0(th, struct osd_thandle, ot_super);
3529 LASSERT(oh->ot_handle);
3531 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3533 if (unlikely(fid_is_acct(fid)))
3536 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3537 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3539 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3540 osd_name(osd), PFID(fid), result);
3543 if (S_ISDIR(inode->i_mode)) {
3544 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3545 inode->i_nlink == 2);
3547 spin_lock(&obj->oo_guard);
3549 spin_unlock(&obj->oo_guard);
3550 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3553 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3555 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3557 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3558 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3559 oh->ot_handle, OI_CHECK_FLD);
3561 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3562 /* XXX: add to ext3 orphan list */
3563 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3565 /* not needed in the cache anymore */
3566 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3567 obj->oo_destroyed = 1;
3573 * Put the fid into lustre_mdt_attrs, and then place the structure
3574 * inode's ea. This fid should not be altered during the life time
3577 * \retval +ve, on success
3578 * \retval -ve, on error
3580 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3582 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3583 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3585 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3586 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3591 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3594 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3597 lustre_loa_init(loa, fid, compat, incompat);
3598 lustre_loa_swab(loa, false);
3601 * For the OST device with 256 bytes inode size by default,
3602 * the PFID EA will be stored together with LMA EA to avoid
3603 * performance trouble. Otherwise the PFID EA can be stored
3604 * independently. LU-8998
3606 if ((compat & LMAC_FID_ON_OST) &&
3607 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3608 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3609 sizeof(*loa), XATTR_CREATE);
3611 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3612 sizeof(*lma), XATTR_CREATE);
3614 * LMA may already exist, but we need to check that all the
3615 * desired compat/incompat flags have been added.
3617 if (unlikely(rc == -EEXIST)) {
3618 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3619 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3623 if (rc < sizeof(*lma))
3626 lustre_loa_swab(loa, true);
3627 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3628 ((compat == 0 && incompat == 0) ||
3629 (!(~lma->lma_compat & compat) &&
3630 !(~lma->lma_incompat & incompat))))
3633 lma->lma_self_fid = *fid;
3634 lma->lma_compat |= compat;
3635 lma->lma_incompat |= incompat;
3636 if (rc == sizeof(*lma)) {
3637 lustre_lma_swab(lma);
3638 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3639 sizeof(*lma), XATTR_REPLACE);
3641 lustre_loa_swab(loa, false);
3642 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3643 sizeof(*loa), XATTR_REPLACE);
3651 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3652 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3653 * To have compatilibility with 1.8 ldiskfs driver we need to have
3654 * magic number at start of fid data.
3655 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3658 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3659 const struct lu_fid *fid)
3661 if (!fid_is_namespace_visible(fid) ||
3662 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3663 param->edp_magic = 0;
3667 param->edp_magic = LDISKFS_LUFID_MAGIC;
3668 param->edp_len = sizeof(struct lu_fid) + 1;
3669 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3673 * Try to read the fid from inode ea into dt_rec.
3675 * \param fid object fid.
3677 * \retval 0 on success
3679 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3680 __u32 ino, struct lu_fid *fid,
3681 struct osd_inode_id *id)
3683 struct osd_thread_info *info = osd_oti_get(env);
3684 struct inode *inode;
3688 osd_id_gen(id, ino, OSD_OII_NOGEN);
3689 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3691 RETURN(PTR_ERR(inode));
3697 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3699 struct inode *parent_dir,
3700 const struct lu_fid *dot_fid,
3701 const struct lu_fid *dot_dot_fid,
3702 struct osd_thandle *oth)
3704 struct ldiskfs_dentry_param *dot_ldp;
3705 struct ldiskfs_dentry_param *dot_dot_ldp;
3706 __u32 saved_nlink = dir->i_nlink;
3709 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3710 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3712 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3713 dot_ldp->edp_magic = 0;
3715 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3716 dir, dot_ldp, dot_dot_ldp);
3718 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3719 * the subseqent ref_add() will increase the dir->i_nlink
3720 * as 3. That is incorrect for new created directory.
3722 * It looks like hack, because we want to make the OSD API
3723 * to be order-independent for new created directory object
3724 * between dt_insert(..) and ref_add() operations.
3726 * Here, we only restore the in-RAM dir-inode's nlink attr,
3727 * becuase if the nlink attr is not 2, then there will be
3728 * ref_add() called following the dt_insert(..), such call
3729 * will make both the in-RAM and on-disk dir-inode's nlink
3730 * attr to be set as 2. LU-7447
3732 set_nlink(dir, saved_nlink);
3737 * Create an local agent inode for remote entry
3739 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3740 struct osd_device *osd,
3741 struct osd_object *pobj,
3742 const struct lu_fid *fid,
3746 struct osd_thread_info *info = osd_oti_get(env);
3747 struct inode *local;
3748 struct osd_thandle *oh;
3754 oh = container_of(th, struct osd_thandle, ot_super);
3755 LASSERT(oh->ot_handle->h_transaction != NULL);
3757 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3758 if (IS_ERR(local)) {
3759 CERROR("%s: create local error %d\n", osd_name(osd),
3760 (int)PTR_ERR(local));
3765 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3766 * correct gid on remote file, not agent here
3768 local->i_gid = current_fsgid();
3769 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3771 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
3772 * That gives e2fsck something to look at and be happy, and allows
3773 * debugging if we need to determine where this symlink came from.
3775 if (S_ISLNK(type)) {
3776 CLASSERT(LDISKFS_N_BLOCKS * 4 >= FID_LEN + 1);
3777 rc = snprintf((char *)LDISKFS_I(local)->i_data,
3778 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
3780 i_size_write(local, rc);
3781 LDISKFS_I(local)->i_disksize = rc;
3783 unlock_new_inode(local);
3785 /* Agent inode should not have project ID */
3786 #ifdef HAVE_PROJECT_QUOTA
3787 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3788 i_projid_read(pobj->oo_inode) != 0) {
3789 rc = osd_transfer_project(local, 0);
3791 CERROR("%s: quota transfer failed: rc = %d. Is project "
3792 "quota enforcement enabled on the ldiskfs "
3793 "filesystem?\n", local->i_sb->s_id, rc);
3794 RETURN(ERR_PTR(rc));
3798 /* Set special LMA flag for local agent inode */
3799 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3801 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3802 osd_name(osd), PFID(fid), rc);
3803 RETURN(ERR_PTR(rc));
3809 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3810 lu_object_fid(&pobj->oo_dt.do_lu),
3813 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3814 osd_name(osd), PFID(fid), rc);
3815 RETURN(ERR_PTR(rc));
3822 * when direntry is deleted, we have to take care of possible agent inode
3823 * referenced by that. unfortunately we can't do this at that point:
3824 * iget() within a running transaction leads to deadlock and we better do
3825 * not call that every delete declaration to save performance. so we put
3826 * a potention agent inode on a list and process that once the transaction
3827 * is over. Notice it's not any worse in terms of real orphans as regular
3828 * object destroy doesn't put inodes on the on-disk orphan list. this should
3829 * be addressed separately
3831 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3832 struct osd_thandle *oh,
3835 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3836 struct osd_obj_orphan *oor;
3843 oor->oor_env = (struct lu_env *)env;
3844 spin_lock(&osd->od_osfs_lock);
3845 list_add(&oor->oor_list, &osd->od_orphan_list);
3846 spin_unlock(&osd->od_osfs_lock);
3848 oh->ot_remove_agents = 1;
3854 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3855 struct osd_device *osd)
3857 struct osd_thread_info *info = osd_oti_get(env);
3858 struct osd_obj_orphan *oor, *tmp;
3859 struct osd_inode_id id;
3860 struct list_head list;
3861 struct inode *inode;
3866 INIT_LIST_HEAD(&list);
3868 spin_lock(&osd->od_osfs_lock);
3869 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3870 if (oor->oor_env == env) {
3871 list_del(&oor->oor_list);
3872 list_add(&oor->oor_list, &list);
3875 spin_unlock(&osd->od_osfs_lock);
3877 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3881 list_del(&oor->oor_list);
3884 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3885 inode = osd_iget_fid(info, osd, &id, &fid);
3889 if (!osd_remote_fid(env, osd, &fid)) {
3894 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3896 mark_inode_dirty(inode);
3897 ldiskfs_journal_stop(jh);
3905 * OSD layer object create function for OST objects (b=11826).
3907 * The FID is inserted into inode xattr here.
3909 * \retval 0, on success
3910 * \retval -ve, on error
3912 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3913 struct lu_attr *attr, struct dt_allocation_hint *hint,
3914 struct dt_object_format *dof, struct thandle *th)
3916 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3917 struct osd_object *obj = osd_dt_obj(dt);
3918 struct osd_thread_info *info = osd_oti_get(env);
3919 int result, on_ost = 0;
3923 if (dt_object_exists(dt))
3926 LINVRNT(osd_invariant(obj));
3927 LASSERT(!dt_object_remote(dt));
3928 LASSERT(osd_is_write_locked(env, obj));
3929 LASSERT(th != NULL);
3931 if (unlikely(fid_is_acct(fid)))
3933 * Quota files can't be created from the kernel any more,
3934 * 'tune2fs -O quota' will take care of creating them
3938 result = __osd_create(info, obj, attr, hint, dof, th);
3940 if (fid_is_idif(fid) &&
3941 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3942 struct lu_fid *tfid = &info->oti_fid;
3943 struct ost_id *oi = &info->oti_ostid;
3945 fid_to_ostid(fid, oi);
3946 ostid_to_fid(tfid, oi, 0);
3948 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3949 LMAC_FID_ON_OST, 0);
3951 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3953 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3954 on_ost ? LMAC_FID_ON_OST : 0,
3957 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3958 obj->oo_dt.do_body_ops = &osd_body_ops;
3961 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY))
3962 result = __osd_oi_insert(env, obj, fid, th);
3965 * a small optimization - dt_insert() isn't usually applied
3966 * to OST objects, so we don't need to cache OI mapping for
3969 if (result == 0 && on_ost == 0) {
3970 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3972 result = osd_idc_find_and_init(env, osd, obj);
3973 LASSERT(result == 0);
3976 LASSERT(ergo(result == 0,
3977 dt_object_exists(dt) && !dt_object_remote(dt)));
3978 LINVRNT(osd_invariant(obj));
3982 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3983 struct thandle *handle)
3985 struct osd_thandle *oh;
3987 /* it's possible that object doesn't exist yet */
3988 LASSERT(handle != NULL);
3990 oh = container_of0(handle, struct osd_thandle, ot_super);
3991 LASSERT(oh->ot_handle == NULL);
3993 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3994 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3996 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
4002 * Concurrency: @dt is write locked.
4004 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4007 struct osd_object *obj = osd_dt_obj(dt);
4008 struct inode *inode = obj->oo_inode;
4009 struct osd_thandle *oh;
4012 if (!dt_object_exists(dt) || obj->oo_destroyed)
4015 LINVRNT(osd_invariant(obj));
4016 LASSERT(!dt_object_remote(dt));
4017 LASSERT(osd_is_write_locked(env, obj));
4018 LASSERT(th != NULL);
4020 oh = container_of0(th, struct osd_thandle, ot_super);
4021 LASSERT(oh->ot_handle != NULL);
4023 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4025 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4026 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4028 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4029 * (65000) subdirectories by storing "1" in i_nlink if the link count
4030 * would otherwise overflow. Directory tranversal tools understand
4031 * that (st_nlink == 1) indicates that the filesystem dose not track
4032 * hard links count on the directory, and will not abort subdirectory
4033 * scanning early once (st_nlink - 2) subdirs have been found.
4035 * This also has to properly handle the case of inodes with nlink == 0
4036 * in case they are being linked into the PENDING directory
4038 spin_lock(&obj->oo_guard);
4039 if (unlikely(inode->i_nlink == 0))
4040 /* inc_nlink from 0 may cause WARN_ON */
4041 set_nlink(inode, 1);
4043 ldiskfs_inc_count(oh->ot_handle, inode);
4044 if (!S_ISDIR(inode->i_mode))
4045 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4047 spin_unlock(&obj->oo_guard);
4049 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4050 LINVRNT(osd_invariant(obj));
4052 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4057 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4058 struct thandle *handle)
4060 struct osd_thandle *oh;
4062 if (!dt_object_exists(dt))
4065 LASSERT(!dt_object_remote(dt));
4066 LASSERT(handle != NULL);
4068 oh = container_of0(handle, struct osd_thandle, ot_super);
4069 LASSERT(oh->ot_handle == NULL);
4071 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4072 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4078 * Concurrency: @dt is write locked.
4080 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4083 struct osd_object *obj = osd_dt_obj(dt);
4084 struct inode *inode = obj->oo_inode;
4085 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4086 struct osd_thandle *oh;
4088 if (!dt_object_exists(dt))
4091 LINVRNT(osd_invariant(obj));
4092 LASSERT(!dt_object_remote(dt));
4093 LASSERT(osd_is_write_locked(env, obj));
4094 LASSERT(th != NULL);
4096 oh = container_of0(th, struct osd_thandle, ot_super);
4097 LASSERT(oh->ot_handle != NULL);
4099 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4101 spin_lock(&obj->oo_guard);
4103 * That can be result of upgrade from old Lustre version and
4104 * applied only to local files. Just skip this ref_del call.
4105 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4107 if (inode->i_nlink == 0) {
4108 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4109 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4110 ", maybe an upgraded file? (LU-3915)\n",
4111 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4112 spin_unlock(&obj->oo_guard);
4116 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4117 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4119 ldiskfs_dec_count(oh->ot_handle, inode);
4120 spin_unlock(&obj->oo_guard);
4122 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4123 LINVRNT(osd_invariant(obj));
4125 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4131 * Concurrency: @dt is read locked.
4133 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4134 struct lu_buf *buf, const char *name)
4136 struct osd_object *obj = osd_dt_obj(dt);
4137 struct inode *inode = obj->oo_inode;
4138 struct osd_thread_info *info = osd_oti_get(env);
4139 struct dentry *dentry = &info->oti_obj_dentry;
4140 bool cache_xattr = false;
4145 /* version get is not real XATTR but uses xattr API */
4146 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4147 dt_obj_version_t *ver = buf->lb_buf;
4150 * for version we are just using xattr API but change inode
4153 if (buf->lb_len == 0)
4154 return sizeof(dt_obj_version_t);
4156 if (buf->lb_len < sizeof(dt_obj_version_t))
4159 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4160 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4162 *ver = LDISKFS_I(inode)->i_fs_version;
4164 return sizeof(dt_obj_version_t);
4167 if (!dt_object_exists(dt))
4170 LASSERT(!dt_object_remote(dt));
4171 LASSERT(inode->i_op != NULL);
4172 #ifdef HAVE_IOP_XATTR
4173 LASSERT(inode->i_op->getxattr != NULL);
4176 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4177 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4181 rc = osd_oxc_get(obj, name, buf);
4186 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4187 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4188 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4189 struct filter_fid *ff;
4190 struct ost_layout *ol;
4192 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4196 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4199 if (buf->lb_len == 0 || !buf->lb_buf)
4202 if (buf->lb_len < rc)
4206 ol = &ff->ff_layout;
4207 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4208 PFID_STRIPE_IDX_BITS);
4209 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4210 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4211 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4212 if (lma->lma_compat & LMAC_COMP_INFO) {
4213 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4214 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4215 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4217 ol->ol_comp_start = 0;
4218 ol->ol_comp_end = 0;
4222 rc = __osd_xattr_get(inode, dentry, name,
4223 buf->lb_buf, buf->lb_len);
4227 if (rc == -ENOENT || rc == -ENODATA)
4228 osd_oxc_add(obj, name, NULL, 0);
4229 else if (rc > 0 && buf->lb_buf != NULL)
4230 osd_oxc_add(obj, name, buf->lb_buf, rc);
4236 static int osd_declare_xattr_set(const struct lu_env *env,
4237 struct dt_object *dt,
4238 const struct lu_buf *buf, const char *name,
4239 int fl, struct thandle *handle)
4241 struct osd_thandle *oh;
4243 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4245 LASSERT(handle != NULL);
4247 oh = container_of0(handle, struct osd_thandle, ot_super);
4248 LASSERT(oh->ot_handle == NULL);
4250 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4252 * For non-upgrading case, the LMA is set first and
4253 * usually fit inode. But for upgrade case, the LMA
4254 * may be in another separated EA block.
4256 if (dt_object_exists(dt)) {
4257 if (fl == LU_XATTR_REPLACE)
4262 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4264 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4265 /* We may need to delete the old PFID EA. */
4266 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4267 if (fl == LU_XATTR_REPLACE)
4273 * If some name entry resides on remote MDT, then will create
4274 * agent entry under remote parent. On the other hand, if the
4275 * remote entry will be removed, then related agent entry may
4276 * need to be removed from the remote parent. So there may be
4277 * kinds of cases, let's declare enough credits. The credits
4278 * for create agent entry is enough for remove case.
4280 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4281 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4282 if (dt_object_exists(dt))
4283 credits += 1; /* For updating LMA */
4287 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4292 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4294 * learn xattr size from osd_xattr_get if
4295 * attribute has not been read yet
4297 buflen = __osd_xattr_get(
4298 osd_dt_obj(dt)->oo_inode,
4299 &osd_oti_get(env)->oti_obj_dentry,
4304 buflen = buf->lb_len;
4307 if (buflen > sb->s_blocksize) {
4308 credits += osd_calc_bkmap_credits(
4310 (buflen + sb->s_blocksize - 1) >>
4311 sb->s_blocksize_bits);
4315 * xattr set may involve inode quota change, reserve credits for
4316 * dquot_initialize()
4318 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4321 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4326 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4327 const struct lu_buf *buf, int fl,
4328 struct thandle *handle)
4330 struct osd_thread_info *info = osd_oti_get(env);
4331 struct dentry *dentry = &info->oti_obj_dentry;
4332 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4333 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4334 struct inode *inode = obj->oo_inode;
4335 struct filter_fid *ff = buf->lb_buf;
4336 struct ost_layout *ol = &ff->ff_layout;
4337 int flags = XATTR_REPLACE;
4342 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4345 rc = osd_get_lma(info, inode, dentry, loa);
4346 if (rc == -ENODATA) {
4347 /* Usually for upgarding from old device */
4348 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4349 LMAC_FID_ON_OST, 0);
4350 flags = XATTR_CREATE;
4355 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4356 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4359 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4360 /* Separate PFID EA from LMA */
4361 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4362 lustre_lma_swab(lma);
4363 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4364 sizeof(*lma), XATTR_REPLACE);
4366 obj->oo_pfid_in_lma = 0;
4367 rc = LU_XATTR_CREATE;
4373 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4377 * Old client does not send stripe information,
4378 * then store the PFID EA on disk separatedly.
4380 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4381 ol->ol_stripe_size == 0))
4384 /* Remove old PFID EA entry firstly. */
4385 ll_vfs_dq_init(inode);
4386 rc = osd_removexattr(dentry, inode, XATTR_NAME_FID);
4387 if (rc == -ENODATA) {
4388 if ((fl & LU_XATTR_REPLACE) && !(fl & LU_XATTR_CREATE))
4395 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4396 if (likely(ol->ol_stripe_size != 0)) {
4397 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4398 PFID_STRIPE_IDX_BITS;
4399 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4400 lma->lma_compat |= LMAC_STRIPE_INFO;
4401 if (ol->ol_comp_id != 0) {
4402 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4403 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4404 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4405 lma->lma_compat |= LMAC_COMP_INFO;
4409 lustre_loa_swab(loa, false);
4411 /* Store the PFID EA inside LMA. */
4412 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4415 obj->oo_pfid_in_lma = 1;
4421 * In DNE environment, the object (in spite of regular file or directory)
4422 * and its name entry may reside on different MDTs. Under such case, we will
4423 * create an agent entry on the MDT where the object resides. The agent entry
4424 * references the object locally, that makes the object to be visible to the
4425 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4426 * such as 'tar' can handle the object properly.
4428 * We handle the agent entry during set linkEA that is the common interface
4429 * for both regular file and directroy, can handle kinds of cases, such as
4430 * create/link/unlink/rename, and so on.
4432 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4434 * XXX: There are two known issues:
4435 * 1. For one object, we will create at most one agent entry even if there
4436 * may be more than one cross-MDTs hard links on the object. So the local
4437 * e2fsck may claim that the object's nlink is larger than the name entries
4438 * that reference such inode. And in further, the e2fsck will fix the nlink
4439 * attribute to match the local references. Then it will cause the object's
4440 * nlink attribute to be inconsistent with the global references. it is bad
4441 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4442 * other hand, the global namespace LFSCK can repair the object's attribute
4443 * according to the linkEA.
4444 * 2. There may be too many hard links on the object as to its linkEA overflow,
4445 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4446 * case happened, then at this point, we do not know whether there are some
4447 * cross-MDTs reference. But there are local references, it guarantees that
4448 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4450 static int osd_xattr_handle_linkea(const struct lu_env *env,
4451 struct osd_device *osd,
4452 struct osd_object *obj,
4453 const struct lu_buf *buf,
4454 struct thandle *handle)
4456 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4457 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4458 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4459 struct lu_name tmpname;
4460 struct osd_thandle *oh;
4462 bool remote = false;
4466 oh = container_of0(handle, struct osd_thandle, ot_super);
4467 LASSERT(oh->ot_handle != NULL);
4469 rc = linkea_init_with_rec(&ldata);
4471 linkea_first_entry(&ldata);
4472 while (ldata.ld_lee != NULL && !remote) {
4473 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4475 if (osd_remote_fid(env, osd, tfid) > 0)
4478 linkea_next_entry(&ldata);
4480 } else if (rc == -ENODATA) {
4486 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4487 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4489 CERROR("%s: failed to remove agent entry for "DFID
4490 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4491 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4492 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4494 CERROR("%s: failed to create agent entry for "DFID
4495 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4502 * Concurrency: @dt is write locked.
4504 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4505 const struct lu_buf *buf, const char *name, int fl,
4506 struct thandle *handle)
4508 struct osd_object *obj = osd_dt_obj(dt);
4509 struct osd_device *osd = osd_obj2dev(obj);
4510 struct inode *inode = obj->oo_inode;
4511 struct osd_thread_info *info = osd_oti_get(env);
4521 /* version set is not real XATTR */
4522 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4523 dt_obj_version_t *version = buf->lb_buf;
4526 * for version we are just using xattr API but change inode
4529 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4531 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4532 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4534 LDISKFS_I(inode)->i_fs_version = *version;
4536 * Version is set after all inode operations are finished,
4537 * so we should mark it dirty here
4539 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4544 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4545 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4548 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4551 * For the OST device with 256 bytes inode size by default,
4552 * the PFID EA will be stored together with LMA EA to avoid
4553 * performance trouble. Otherwise the PFID EA can be stored
4554 * independently. LU-8998
4556 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4557 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4558 LASSERT(buf->lb_buf);
4560 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4563 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4564 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4565 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4567 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4571 lma->lma_incompat |= LMAI_STRIPED;
4572 lustre_lma_swab(lma);
4573 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4574 sizeof(*lma), XATTR_REPLACE);
4577 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4578 LASSERT(!osd->od_is_ost);
4580 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4585 if (fl & LU_XATTR_REPLACE)
4586 fs_flags |= XATTR_REPLACE;
4588 if (fl & LU_XATTR_CREATE)
4589 fs_flags |= XATTR_CREATE;
4591 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4592 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4595 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4596 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4597 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4603 * Concurrency: @dt is read locked.
4605 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4606 const struct lu_buf *buf)
4608 struct osd_object *obj = osd_dt_obj(dt);
4609 struct inode *inode = obj->oo_inode;
4610 struct osd_thread_info *info = osd_oti_get(env);
4611 struct dentry *dentry = &info->oti_obj_dentry;
4613 if (!dt_object_exists(dt))
4616 LASSERT(!dt_object_remote(dt));
4617 LASSERT(inode->i_op != NULL);
4618 LASSERT(inode->i_op->listxattr != NULL);
4620 dentry->d_inode = inode;
4621 dentry->d_sb = inode->i_sb;
4622 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4625 static int osd_declare_xattr_del(const struct lu_env *env,
4626 struct dt_object *dt, const char *name,
4627 struct thandle *handle)
4629 struct osd_thandle *oh;
4630 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4632 LASSERT(!dt_object_remote(dt));
4633 LASSERT(handle != NULL);
4635 oh = container_of0(handle, struct osd_thandle, ot_super);
4636 LASSERT(oh->ot_handle == NULL);
4638 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4639 osd_dto_credits_noquota[DTO_XATTR_SET]);
4641 * xattr del may involve inode quota change, reserve credits for
4642 * dquot_initialize()
4644 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4650 * Concurrency: @dt is write locked.
4652 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4653 const char *name, struct thandle *handle)
4655 struct osd_object *obj = osd_dt_obj(dt);
4656 struct inode *inode = obj->oo_inode;
4657 struct osd_thread_info *info = osd_oti_get(env);
4658 struct dentry *dentry = &info->oti_obj_dentry;
4661 if (!dt_object_exists(dt))
4664 LASSERT(!dt_object_remote(dt));
4665 LASSERT(inode->i_op != NULL);
4666 LASSERT(handle != NULL);
4667 #ifdef HAVE_IOP_XATTR
4668 LASSERT(inode->i_op->removexattr != NULL);
4671 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4673 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4674 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4676 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4677 &info->oti_ost_attrs);
4679 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4681 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4682 lustre_lma_swab(lma);
4683 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4684 sizeof(*lma), XATTR_REPLACE);
4686 obj->oo_pfid_in_lma = 0;
4689 ll_vfs_dq_init(inode);
4690 dentry->d_inode = inode;
4691 dentry->d_sb = inode->i_sb;
4692 rc = osd_removexattr(dentry, inode, name);
4695 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4698 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4699 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4700 osd_oxc_del(obj, name);
4705 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4706 __u64 start, __u64 end)
4708 struct osd_object *obj = osd_dt_obj(dt);
4709 struct inode *inode = obj->oo_inode;
4710 struct osd_thread_info *info = osd_oti_get(env);
4711 struct dentry *dentry = &info->oti_obj_dentry;
4712 struct file *file = &info->oti_file;
4717 dentry->d_inode = inode;
4718 dentry->d_sb = inode->i_sb;
4719 file->f_path.dentry = dentry;
4720 file->f_mapping = inode->i_mapping;
4721 file->f_op = inode->i_fop;
4722 set_file_inode(file, inode);
4724 rc = ll_vfs_fsync_range(file, start, end, 0);
4729 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4738 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4739 const struct dt_index_features *feat)
4741 struct iam_descr *descr;
4743 if (osd_object_is_root(o))
4744 return feat == &dt_directory_features;
4746 LASSERT(o->oo_dir != NULL);
4748 descr = o->oo_dir->od_container.ic_descr;
4749 if (feat == &dt_directory_features) {
4750 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4755 return feat->dif_keysize_min <= descr->id_key_size &&
4756 descr->id_key_size <= feat->dif_keysize_max &&
4757 feat->dif_recsize_min <= descr->id_rec_size &&
4758 descr->id_rec_size <= feat->dif_recsize_max &&
4759 !(feat->dif_flags & (DT_IND_VARKEY |
4760 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4761 ergo(feat->dif_flags & DT_IND_UPDATE,
4762 1 /* XXX check that object (and fs) is writable */);
4766 static int osd_iam_container_init(const struct lu_env *env,
4767 struct osd_object *obj,
4768 struct osd_directory *dir)
4770 struct iam_container *bag = &dir->od_container;
4773 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4777 result = iam_container_setup(bag);
4779 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4781 iam_container_fini(bag);
4788 * Concurrency: no external locking is necessary.
4790 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4791 const struct dt_index_features *feat)
4795 struct osd_object *obj = osd_dt_obj(dt);
4797 LINVRNT(osd_invariant(obj));
4799 if (osd_object_is_root(obj)) {
4800 dt->do_index_ops = &osd_index_ea_ops;
4802 } else if (feat == &dt_directory_features) {
4803 dt->do_index_ops = &osd_index_ea_ops;
4804 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4809 } else if (unlikely(feat == &dt_otable_features)) {
4810 dt->do_index_ops = &osd_otable_ops;
4812 } else if (unlikely(feat == &dt_acct_features)) {
4813 dt->do_index_ops = &osd_acct_index_ops;
4816 } else if (!osd_has_index(obj)) {
4817 struct osd_directory *dir;
4818 struct osd_device *osd = osd_obj2dev(obj);
4819 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4824 spin_lock(&obj->oo_guard);
4825 if (obj->oo_dir == NULL)
4829 * Concurrent thread allocated container data.
4832 spin_unlock(&obj->oo_guard);
4834 * Now, that we have container data, serialize its
4837 down_write(&obj->oo_ext_idx_sem);
4839 * recheck under lock.
4842 if (osd_has_index(obj)) {
4847 result = osd_iam_container_init(env, obj, obj->oo_dir);
4848 if (result || feat == &dt_lfsck_namespace_features ||
4849 feat == &dt_lfsck_layout_orphan_features ||
4850 feat == &dt_lfsck_layout_dangling_features)
4853 result = osd_index_register(osd, fid,
4854 feat->dif_keysize_max,
4855 feat->dif_recsize_max);
4857 CWARN("%s: failed to register index "
4859 osd_name(osd), PFID(fid), result);
4860 else if (result > 0)
4863 CDEBUG(D_LFSCK, "%s: index object "DFID
4864 " (%d/%d) registered\n",
4865 osd_name(osd), PFID(fid),
4866 (int)feat->dif_keysize_max,
4867 (int)feat->dif_recsize_max);
4870 up_write(&obj->oo_ext_idx_sem);
4878 if (result == 0 && skip_iam == 0) {
4879 if (!osd_iam_index_probe(env, obj, feat))
4882 LINVRNT(osd_invariant(obj));
4887 static int osd_otable_it_attr_get(const struct lu_env *env,
4888 struct dt_object *dt,
4889 struct lu_attr *attr)
4895 static const struct dt_object_operations osd_obj_ops = {
4896 .do_read_lock = osd_read_lock,
4897 .do_write_lock = osd_write_lock,
4898 .do_read_unlock = osd_read_unlock,
4899 .do_write_unlock = osd_write_unlock,
4900 .do_write_locked = osd_write_locked,
4901 .do_attr_get = osd_attr_get,
4902 .do_declare_attr_set = osd_declare_attr_set,
4903 .do_attr_set = osd_attr_set,
4904 .do_ah_init = osd_ah_init,
4905 .do_declare_create = osd_declare_create,
4906 .do_create = osd_create,
4907 .do_declare_destroy = osd_declare_destroy,
4908 .do_destroy = osd_destroy,
4909 .do_index_try = osd_index_try,
4910 .do_declare_ref_add = osd_declare_ref_add,
4911 .do_ref_add = osd_ref_add,
4912 .do_declare_ref_del = osd_declare_ref_del,
4913 .do_ref_del = osd_ref_del,
4914 .do_xattr_get = osd_xattr_get,
4915 .do_declare_xattr_set = osd_declare_xattr_set,
4916 .do_xattr_set = osd_xattr_set,
4917 .do_declare_xattr_del = osd_declare_xattr_del,
4918 .do_xattr_del = osd_xattr_del,
4919 .do_xattr_list = osd_xattr_list,
4920 .do_object_sync = osd_object_sync,
4921 .do_invalidate = osd_invalidate,
4924 static const struct dt_object_operations osd_obj_otable_it_ops = {
4925 .do_attr_get = osd_otable_it_attr_get,
4926 .do_index_try = osd_index_try,
4929 static int osd_index_declare_iam_delete(const struct lu_env *env,
4930 struct dt_object *dt,
4931 const struct dt_key *key,
4932 struct thandle *handle)
4934 struct osd_thandle *oh;
4936 oh = container_of0(handle, struct osd_thandle, ot_super);
4937 LASSERT(oh->ot_handle == NULL);
4939 /* Recycle may cause additional three blocks to be changed. */
4940 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4941 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4947 * delete a (key, value) pair from index \a dt specified by \a key
4949 * \param dt osd index object
4950 * \param key key for index
4951 * \param rec record reference
4952 * \param handle transaction handler
4955 * \retval -ve failure
4957 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4958 const struct dt_key *key,
4959 struct thandle *handle)
4961 struct osd_thread_info *oti = osd_oti_get(env);
4962 struct osd_object *obj = osd_dt_obj(dt);
4963 struct osd_thandle *oh;
4964 struct iam_path_descr *ipd;
4965 struct iam_container *bag = &obj->oo_dir->od_container;
4970 if (!dt_object_exists(dt))
4973 LINVRNT(osd_invariant(obj));
4974 LASSERT(!dt_object_remote(dt));
4975 LASSERT(bag->ic_object == obj->oo_inode);
4976 LASSERT(handle != NULL);
4978 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4980 ipd = osd_idx_ipd_get(env, bag);
4981 if (unlikely(ipd == NULL))
4984 oh = container_of0(handle, struct osd_thandle, ot_super);
4985 LASSERT(oh->ot_handle != NULL);
4986 LASSERT(oh->ot_handle->h_transaction != NULL);
4988 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4989 /* swab quota uid/gid provided by caller */
4990 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4991 key = (const struct dt_key *)&oti->oti_quota_id;
4994 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4995 osd_ipd_put(env, bag, ipd);
4996 LINVRNT(osd_invariant(obj));
4997 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5001 static int osd_index_declare_ea_delete(const struct lu_env *env,
5002 struct dt_object *dt,
5003 const struct dt_key *key,
5004 struct thandle *handle)
5006 struct osd_thandle *oh;
5007 struct inode *inode;
5012 LASSERT(!dt_object_remote(dt));
5013 LASSERT(handle != NULL);
5015 oh = container_of0(handle, struct osd_thandle, ot_super);
5016 LASSERT(oh->ot_handle == NULL);
5018 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5019 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5021 inode = osd_dt_obj(dt)->oo_inode;
5025 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5026 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5031 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5034 struct osd_fid_pack *rec;
5037 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5038 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5039 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5040 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5046 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5047 const struct lu_fid *fid)
5049 struct seq_server_site *ss = osd_seq_site(osd);
5053 /* FID seqs not in FLDB, must be local seq */
5054 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5058 * If FLD is not being initialized yet, it only happens during the
5059 * initialization, likely during mgs initialization, and we assume
5060 * this is local FID.
5062 if (ss == NULL || ss->ss_server_fld == NULL)
5065 /* Only check the local FLDB here */
5066 if (osd_seq_exists(env, osd, fid_seq(fid)))
5072 static void osd_take_care_of_agent(const struct lu_env *env,
5073 struct osd_device *osd,
5074 struct osd_thandle *oh,
5075 struct ldiskfs_dir_entry_2 *de)
5077 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5078 struct osd_idmap_cache *idc;
5079 int rc, schedule = 0;
5081 LASSERT(de != NULL);
5083 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5084 if (likely(rc == 0)) {
5085 idc = osd_idc_find_or_init(env, osd, fid);
5086 if (IS_ERR(idc) || idc->oic_remote)
5088 } else if (rc == -ENODATA) {
5090 * can't get FID, postpone to the end of the
5091 * transaction when iget() is safe
5095 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5098 osd_schedule_agent_inode_removal(env, oh,
5099 le32_to_cpu(de->inode));
5103 * Index delete function for interoperability mode (b11826).
5104 * It will remove the directory entry added by osd_index_ea_insert().
5105 * This entry is needed to maintain name->fid mapping.
5107 * \param key, key i.e. file entry to be deleted
5109 * \retval 0, on success
5110 * \retval -ve, on error
5112 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5113 const struct dt_key *key, struct thandle *handle)
5115 struct osd_object *obj = osd_dt_obj(dt);
5116 struct inode *dir = obj->oo_inode;
5117 struct dentry *dentry;
5118 struct osd_thandle *oh;
5119 struct ldiskfs_dir_entry_2 *de = NULL;
5120 struct buffer_head *bh;
5121 struct htree_lock *hlock = NULL;
5122 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5127 if (!dt_object_exists(dt))
5130 LINVRNT(osd_invariant(obj));
5131 LASSERT(!dt_object_remote(dt));
5132 LASSERT(handle != NULL);
5134 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5136 oh = container_of(handle, struct osd_thandle, ot_super);
5137 LASSERT(oh->ot_handle != NULL);
5138 LASSERT(oh->ot_handle->h_transaction != NULL);
5140 ll_vfs_dq_init(dir);
5141 dentry = osd_child_dentry_get(env, obj,
5142 (char *)key, strlen((char *)key));
5144 if (obj->oo_hl_head != NULL) {
5145 hlock = osd_oti_get(env)->oti_hlock;
5146 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5147 dir, LDISKFS_HLOCK_DEL);
5149 down_write(&obj->oo_ext_idx_sem);
5152 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5155 * If this is not the ".." entry, it might be a remote DNE
5156 * entry and we need to check if the FID is for a remote
5157 * MDT. If the FID is not in the directory entry (e.g.
5158 * upgraded 1.8 filesystem without dirdata enabled) then
5159 * we need to get the FID from the LMA. For a remote directory
5160 * there HAS to be an LMA, it cannot be an IGIF inode in this
5163 * Delete the entry before the agent inode in order to
5164 * simplify error handling. At worst an error after deleting
5165 * the entry first might leak the agent inode afterward. The
5166 * reverse would need filesystem abort in case of error deleting
5167 * the entry after the agent had been removed, or leave a
5168 * dangling entry pointing at a random inode.
5170 if (strcmp((char *)key, dotdot) != 0)
5171 osd_take_care_of_agent(env, osd, oh, de);
5172 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5178 ldiskfs_htree_unlock(hlock);
5180 up_write(&obj->oo_ext_idx_sem);
5184 LASSERT(osd_invariant(obj));
5185 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5190 * Lookup index for \a key and copy record to \a rec.
5192 * \param dt osd index object
5193 * \param key key for index
5194 * \param rec record reference
5196 * \retval +ve success : exact mach
5197 * \retval 0 return record with key not greater than \a key
5198 * \retval -ve failure
5200 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5201 struct dt_rec *rec, const struct dt_key *key)
5203 struct osd_object *obj = osd_dt_obj(dt);
5204 struct iam_path_descr *ipd;
5205 struct iam_container *bag = &obj->oo_dir->od_container;
5206 struct osd_thread_info *oti = osd_oti_get(env);
5207 struct iam_iterator *it = &oti->oti_idx_it;
5208 struct iam_rec *iam_rec;
5213 if (!dt_object_exists(dt))
5216 LASSERT(osd_invariant(obj));
5217 LASSERT(!dt_object_remote(dt));
5218 LASSERT(bag->ic_object == obj->oo_inode);
5220 ipd = osd_idx_ipd_get(env, bag);
5224 /* got ipd now we can start iterator. */
5225 iam_it_init(it, bag, 0, ipd);
5227 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5228 /* swab quota uid/gid provided by caller */
5229 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5230 key = (const struct dt_key *)&oti->oti_quota_id;
5233 rc = iam_it_get(it, (struct iam_key *)key);
5235 if (S_ISDIR(obj->oo_inode->i_mode))
5236 iam_rec = (struct iam_rec *)oti->oti_ldp;
5238 iam_rec = (struct iam_rec *)rec;
5240 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5242 if (S_ISDIR(obj->oo_inode->i_mode))
5243 osd_fid_unpack((struct lu_fid *)rec,
5244 (struct osd_fid_pack *)iam_rec);
5245 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5246 osd_quota_unpack(obj, rec);
5251 osd_ipd_put(env, bag, ipd);
5253 LINVRNT(osd_invariant(obj));
5258 static int osd_index_declare_iam_insert(const struct lu_env *env,
5259 struct dt_object *dt,
5260 const struct dt_rec *rec,
5261 const struct dt_key *key,
5262 struct thandle *handle)
5264 struct osd_thandle *oh;
5266 LASSERT(handle != NULL);
5268 oh = container_of0(handle, struct osd_thandle, ot_super);
5269 LASSERT(oh->ot_handle == NULL);
5271 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5272 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5278 * Inserts (key, value) pair in \a dt index object.
5280 * \param dt osd index object
5281 * \param key key for index
5282 * \param rec record reference
5283 * \param th transaction handler
5286 * \retval -ve failure
5288 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5289 const struct dt_rec *rec,
5290 const struct dt_key *key, struct thandle *th)
5292 struct osd_object *obj = osd_dt_obj(dt);
5293 struct iam_path_descr *ipd;
5294 struct osd_thandle *oh;
5295 struct iam_container *bag;
5296 struct osd_thread_info *oti = osd_oti_get(env);
5297 struct iam_rec *iam_rec;
5302 if (!dt_object_exists(dt))
5305 LINVRNT(osd_invariant(obj));
5306 LASSERT(!dt_object_remote(dt));
5308 bag = &obj->oo_dir->od_container;
5309 LASSERT(bag->ic_object == obj->oo_inode);
5310 LASSERT(th != NULL);
5312 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5314 ipd = osd_idx_ipd_get(env, bag);
5315 if (unlikely(ipd == NULL))
5318 oh = container_of0(th, struct osd_thandle, ot_super);
5319 LASSERT(oh->ot_handle != NULL);
5320 LASSERT(oh->ot_handle->h_transaction != NULL);
5321 if (S_ISDIR(obj->oo_inode->i_mode)) {
5322 iam_rec = (struct iam_rec *)oti->oti_ldp;
5323 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5325 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5326 /* pack quota uid/gid */
5327 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5328 key = (const struct dt_key *)&oti->oti_quota_id;
5329 /* pack quota record */
5330 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5331 iam_rec = (struct iam_rec *)rec;
5333 iam_rec = (struct iam_rec *)rec;
5336 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5338 osd_ipd_put(env, bag, ipd);
5339 LINVRNT(osd_invariant(obj));
5340 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5345 * Calls ldiskfs_add_entry() to add directory entry
5346 * into the directory. This is required for
5347 * interoperability mode (b11826)
5349 * \retval 0, on success
5350 * \retval -ve, on error
5352 static int __osd_ea_add_rec(struct osd_thread_info *info,
5353 struct osd_object *pobj, struct inode *cinode,
5354 const char *name, const struct lu_fid *fid,
5355 struct htree_lock *hlock, struct thandle *th)
5357 struct ldiskfs_dentry_param *ldp;
5358 struct dentry *child;
5359 struct osd_thandle *oth;
5362 oth = container_of(th, struct osd_thandle, ot_super);
5363 LASSERT(oth->ot_handle != NULL);
5364 LASSERT(oth->ot_handle->h_transaction != NULL);
5365 LASSERT(pobj->oo_inode);
5367 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5368 if (unlikely(pobj->oo_inode ==
5369 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
5372 osd_get_ldiskfs_dirent_param(ldp, fid);
5373 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5374 child->d_fsdata = (void *)ldp;
5375 ll_vfs_dq_init(pobj->oo_inode);
5376 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5377 child, cinode, hlock);
5378 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5379 struct ldiskfs_dir_entry_2 *de;
5380 struct buffer_head *bh;
5383 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5386 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5389 if (S_ISDIR(cinode->i_mode))
5390 de->file_type = LDISKFS_DIRENT_LUFID |
5391 LDISKFS_FT_REG_FILE;
5393 de->file_type = LDISKFS_DIRENT_LUFID |
5395 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5406 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5407 * into the directory.Also sets flags into osd object to
5408 * indicate dot and dotdot are created. This is required for
5409 * interoperability mode (b11826)
5411 * \param dir directory for dot and dotdot fixup.
5412 * \param obj child object for linking
5414 * \retval 0, on success
5415 * \retval -ve, on error
5417 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5418 struct osd_object *dir,
5419 struct inode *parent_dir, const char *name,
5420 const struct lu_fid *dot_fid,
5421 const struct lu_fid *dot_dot_fid,
5424 struct inode *inode = dir->oo_inode;
5425 struct osd_thandle *oth;
5428 oth = container_of(th, struct osd_thandle, ot_super);
5429 LASSERT(oth->ot_handle->h_transaction != NULL);
5430 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5432 if (strcmp(name, dot) == 0) {
5433 if (dir->oo_compat_dot_created) {
5436 LASSERT(inode->i_ino == parent_dir->i_ino);
5437 dir->oo_compat_dot_created = 1;
5440 } else if (strcmp(name, dotdot) == 0) {
5441 if (!dir->oo_compat_dot_created)
5443 /* in case of rename, dotdot is already created */
5444 if (dir->oo_compat_dotdot_created) {
5445 return __osd_ea_add_rec(info, dir, parent_dir, name,
5446 dot_dot_fid, NULL, th);
5449 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5450 struct lu_fid tfid = *dot_dot_fid;
5453 result = osd_add_dot_dotdot_internal(info,
5454 dir->oo_inode, parent_dir, dot_fid,
5457 result = osd_add_dot_dotdot_internal(info,
5458 dir->oo_inode, parent_dir, dot_fid,
5463 dir->oo_compat_dotdot_created = 1;
5471 * It will call the appropriate osd_add* function and return the
5472 * value, return by respective functions.
5474 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5475 struct inode *cinode, const char *name,
5476 const struct lu_fid *fid, struct thandle *th)
5478 struct osd_thread_info *info = osd_oti_get(env);
5479 struct htree_lock *hlock;
5482 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5484 if (name[0] == '.' && (name[1] == '\0' ||
5485 (name[1] == '.' && name[2] == '\0'))) {
5486 if (hlock != NULL) {
5487 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5490 down_write(&pobj->oo_ext_idx_sem);
5493 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5494 lu_object_fid(&pobj->oo_dt.do_lu),
5497 if (hlock != NULL) {
5498 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5499 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5501 down_write(&pobj->oo_ext_idx_sem);
5504 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5505 struct lu_fid *tfid = &info->oti_fid;
5509 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5512 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5517 ldiskfs_htree_unlock(hlock);
5519 up_write(&pobj->oo_ext_idx_sem);
5525 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5526 struct osd_idmap_cache *oic)
5528 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5529 struct lu_fid *fid = &oic->oic_fid;
5530 struct osd_inode_id *id = &oic->oic_lid;
5531 struct inode *inode = NULL;
5538 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5541 if (thread_is_running(&scrub->os_thread) &&
5542 scrub->os_pos_current > id->oii_ino)
5545 if (dev->od_auto_scrub_interval == AS_NEVER ||
5546 ktime_get_real_seconds() <
5547 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5551 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5552 if (rc == -ENOENT) {
5553 __u32 gen = id->oii_gen;
5559 inode = osd_iget(oti, dev, id);
5560 /* The inode has been removed (by race maybe). */
5561 if (IS_ERR(inode)) {
5562 rc = PTR_ERR(inode);
5564 RETURN(rc == -ESTALE ? -ENOENT : rc);
5567 /* The OI mapping is lost. */
5568 if (gen != OSD_OII_NOGEN)
5572 * The inode may has been reused by others, we do not know,
5573 * leave it to be handled by subsequent osd_fid_lookup().
5576 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5583 if (thread_is_running(&scrub->os_thread)) {
5584 if (inode == NULL) {
5585 inode = osd_iget(oti, dev, id);
5586 /* The inode has been removed (by race maybe). */
5587 if (IS_ERR(inode)) {
5588 rc = PTR_ERR(inode);
5590 RETURN(rc == -ESTALE ? -ENOENT : rc);
5594 rc = osd_oii_insert(dev, oic, insert);
5596 * There is race condition between osd_oi_lookup and OI scrub.
5597 * The OI scrub finished just after osd_oi_lookup() failure.
5598 * Under such case, it is unnecessary to trigger OI scrub again,
5599 * but try to call osd_oi_lookup() again.
5601 if (unlikely(rc == -EAGAIN))
5604 if (!S_ISDIR(inode->i_mode))
5607 rc = osd_check_lmv(oti, dev, inode, oic);
5612 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5613 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5614 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5615 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5616 "%s: trigger partial OI scrub for RPC inconsistency "
5617 "checking FID "DFID": rc = %d\n",
5618 osd_dev2name(dev), PFID(fid), rc);
5619 if (rc == 0 || rc == -EALREADY)
5632 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5633 struct osd_device *dev,
5634 struct osd_idmap_cache *oic,
5635 struct lu_fid *fid, __u32 ino)
5637 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5638 struct inode *inode;
5641 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5642 inode = osd_iget(oti, dev, &oic->oic_lid);
5643 if (IS_ERR(inode)) {
5644 fid_zero(&oic->oic_fid);
5645 return PTR_ERR(inode);
5648 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5651 fid_zero(&oic->oic_fid);
5653 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5657 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5658 struct osd_inode_id *id, const struct lu_fid *fid)
5660 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5661 id->oii_ino, id->oii_gen, info);
5662 info->oti_cache.oic_lid = *id;
5663 info->oti_cache.oic_fid = *fid;
5664 info->oti_cache.oic_dev = osd;
5668 * Get parent FID from the linkEA.
5670 * For a directory which parent resides on remote MDT, to satisfy the
5671 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5672 * the other hand, to make the lookup(..) on the directory can return
5673 * the real parent FID, we append the real parent FID after its ".."
5674 * name entry in the /REMOTE_PARENT_DIR.
5676 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5677 * backup. So after the restore, we cannot get the right parent FID from
5678 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5679 * we have stored the real parent FID in the directory object's linkEA,
5680 * we can parse the linkEA for the real parent FID.
5682 * \param[in] env pointer to the thread context
5683 * \param[in] obj pointer to the object to be handled
5684 * \param[out]fid pointer to the buffer to hold the parent FID
5686 * \retval 0 for getting the real parent FID successfully
5687 * \retval negative error number on failure
5689 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5690 struct osd_object *obj,
5693 struct osd_thread_info *oti = osd_oti_get(env);
5694 struct lu_buf *buf = &oti->oti_big_buf;
5695 struct dentry *dentry = &oti->oti_obj_dentry;
5696 struct inode *inode = obj->oo_inode;
5697 struct linkea_data ldata = { NULL };
5703 if (!S_ISDIR(inode->i_mode))
5707 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5708 buf->lb_buf, buf->lb_len);
5709 if (rc == -ERANGE) {
5710 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5713 lu_buf_realloc(buf, rc);
5714 if (buf->lb_buf == NULL)
5721 if (unlikely(rc == 0))
5727 if (unlikely(buf->lb_buf == NULL)) {
5728 lu_buf_realloc(buf, rc);
5729 if (buf->lb_buf == NULL)
5736 rc = linkea_init_with_rec(&ldata);
5738 linkea_first_entry(&ldata);
5739 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5745 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5746 struct inode *inode,
5747 const struct lu_fid *pfid,
5748 const char *name, const int namelen)
5750 struct osd_thread_info *oti = osd_oti_get(env);
5751 struct lu_buf *buf = &oti->oti_big_buf;
5752 struct dentry *dentry = &oti->oti_obj_dentry;
5753 struct linkea_data ldata = { NULL };
5754 struct lu_name cname = { .ln_name = name,
5755 .ln_namelen = namelen };
5761 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5762 buf->lb_buf, buf->lb_len);
5764 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5769 if (unlikely(rc == 0))
5772 if (buf->lb_len < rc) {
5773 lu_buf_realloc(buf, rc);
5774 if (buf->lb_buf == NULL)
5781 rc = linkea_init_with_rec(&ldata);
5783 rc = linkea_links_find(&ldata, &cname, pfid);
5789 * Calls ->lookup() to find dentry. From dentry get inode and
5790 * read inode's ea to get fid. This is required for interoperability
5793 * \retval 0, on success
5794 * \retval -ve, on error
5796 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5797 struct dt_rec *rec, const struct dt_key *key)
5799 struct inode *dir = obj->oo_inode;
5800 struct dentry *dentry;
5801 struct ldiskfs_dir_entry_2 *de;
5802 struct buffer_head *bh;
5803 struct lu_fid *fid = (struct lu_fid *)rec;
5804 struct htree_lock *hlock = NULL;
5810 LASSERT(dir->i_op != NULL);
5811 LASSERT(dir->i_op->lookup != NULL);
5813 dentry = osd_child_dentry_get(env, obj,
5814 (char *)key, strlen((char *)key));
5816 if (obj->oo_hl_head != NULL) {
5817 hlock = osd_oti_get(env)->oti_hlock;
5818 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5819 dir, LDISKFS_HLOCK_LOOKUP);
5821 down_read(&obj->oo_ext_idx_sem);
5824 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5826 struct osd_thread_info *oti = osd_oti_get(env);
5827 struct osd_inode_id *id = &oti->oti_id;
5828 struct osd_idmap_cache *oic = &oti->oti_cache;
5829 struct osd_device *dev = osd_obj2dev(obj);
5831 ino = le32_to_cpu(de->inode);
5832 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5834 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5838 rc = osd_get_fid_from_dentry(de, rec);
5840 /* done with de, release bh */
5843 if (unlikely(is_remote_parent_ino(dev, ino))) {
5844 const char *name = (const char *)key;
5847 * If the parent is on remote MDT, and there
5848 * is no FID-in-dirent, then we have to get
5849 * the parent FID from the linkEA.
5851 if (likely(strlen(name) == 2 &&
5852 name[0] == '.' && name[1] == '.'))
5853 rc = osd_get_pfid_from_linkea(env, obj,
5856 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5859 osd_id_gen(id, ino, OSD_OII_NOGEN);
5862 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5863 fid_zero(&oic->oic_fid);
5868 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5869 rc = osd_consistency_check(oti, dev, oic);
5871 fid_zero(&oic->oic_fid);
5873 /* Other error should not affect lookup result. */
5883 ldiskfs_htree_unlock(hlock);
5885 up_read(&obj->oo_ext_idx_sem);
5889 static int osd_index_declare_ea_insert(const struct lu_env *env,
5890 struct dt_object *dt,
5891 const struct dt_rec *rec,
5892 const struct dt_key *key,
5893 struct thandle *handle)
5895 struct osd_thandle *oh;
5896 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5897 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5898 const struct lu_fid *fid = rec1->rec_fid;
5899 int credits, rc = 0;
5900 struct osd_idmap_cache *idc;
5904 LASSERT(!dt_object_remote(dt));
5905 LASSERT(handle != NULL);
5906 LASSERT(fid != NULL);
5907 LASSERT(rec1->rec_type != 0);
5909 oh = container_of0(handle, struct osd_thandle, ot_super);
5910 LASSERT(oh->ot_handle == NULL);
5912 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5915 * we can't call iget() while a transactions is running
5916 * (this can lead to a deadlock), but we need to know
5917 * inum and object type. so we find this information at
5918 * declaration and cache in per-thread info
5920 idc = osd_idc_find_or_init(env, osd, fid);
5922 RETURN(PTR_ERR(idc));
5923 if (idc->oic_remote) {
5925 * a reference to remote inode is represented by an
5926 * agent inode which we have to create
5928 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5929 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5932 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5934 if (osd_dt_obj(dt)->oo_inode != NULL) {
5935 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5938 * We ignore block quota on meta pool (MDTs), so needn't
5939 * calculate how many blocks will be consumed by this index
5942 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5944 i_projid_read(inode), 0,
5945 oh, osd_dt_obj(dt), NULL,
5950 #ifdef HAVE_PROJECT_QUOTA
5952 * Reserve credits for local agent inode to transfer
5953 * to 0, quota enforcement is ignored in this case.
5955 if (idc->oic_remote &&
5956 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
5957 i_projid_read(inode) != 0)
5958 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
5959 0, i_projid_read(inode),
5960 0, false, PRJQUOTA, true);
5968 * Index add function for interoperability mode (b11826).
5969 * It will add the directory entry.This entry is needed to
5970 * maintain name->fid mapping.
5972 * \param key it is key i.e. file entry to be inserted
5973 * \param rec it is value of given key i.e. fid
5975 * \retval 0, on success
5976 * \retval -ve, on error
5978 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5979 const struct dt_rec *rec,
5980 const struct dt_key *key, struct thandle *th)
5982 struct osd_object *obj = osd_dt_obj(dt);
5983 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5984 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5985 const struct lu_fid *fid = rec1->rec_fid;
5986 const char *name = (const char *)key;
5987 struct osd_thread_info *oti = osd_oti_get(env);
5988 struct inode *child_inode = NULL;
5989 struct osd_idmap_cache *idc;
5994 if (!dt_object_exists(dt))
5997 LASSERT(osd_invariant(obj));
5998 LASSERT(!dt_object_remote(dt));
5999 LASSERT(th != NULL);
6001 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6003 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6005 idc = osd_idc_find(env, osd, fid);
6006 if (unlikely(idc == NULL)) {
6007 idc = osd_idc_find_or_init(env, osd, fid);
6010 * this dt_insert() wasn't declared properly, so
6011 * FID is missing in OI cache. we better do not
6012 * lookup FID in FLDB/OI and don't risk to deadlock,
6013 * but in some special cases (lfsck testing, etc)
6014 * it's much simpler than fixing a caller.
6016 * normally this error should be placed after the first
6017 * find, but migrate may attach source stripes to
6018 * target, which doesn't create stripes.
6020 CERROR("%s: "DFID" wasn't declared for insert\n",
6021 osd_name(osd), PFID(fid));
6023 RETURN(PTR_ERR(idc));
6027 if (idc->oic_remote) {
6028 /* Insert remote entry */
6029 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6031 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6033 child_inode = osd_create_local_agent_inode(env, osd,
6034 obj, fid, rec1->rec_type & S_IFMT, th);
6035 if (IS_ERR(child_inode))
6036 RETURN(PTR_ERR(child_inode));
6039 /* Insert local entry */
6040 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6041 /* for a reason OI cache wasn't filled properly */
6042 CERROR("%s: OIC for "DFID" isn't filled\n",
6043 osd_name(osd), PFID(fid));
6046 child_inode = oti->oti_inode;
6047 if (unlikely(child_inode == NULL)) {
6048 struct ldiskfs_inode_info *lii;
6053 child_inode = oti->oti_inode = &lii->vfs_inode;
6055 child_inode->i_sb = osd_sb(osd);
6056 child_inode->i_ino = idc->oic_lid.oii_ino;
6057 child_inode->i_mode = rec1->rec_type & S_IFMT;
6060 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6062 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6063 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6065 if (child_inode && child_inode != oti->oti_inode)
6067 LASSERT(osd_invariant(obj));
6068 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6073 * Initialize osd Iterator for given osd index object.
6075 * \param dt osd index object
6078 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6079 struct dt_object *dt,
6082 struct osd_it_iam *it;
6083 struct osd_object *obj = osd_dt_obj(dt);
6084 struct lu_object *lo = &dt->do_lu;
6085 struct iam_path_descr *ipd;
6086 struct iam_container *bag = &obj->oo_dir->od_container;
6088 if (!dt_object_exists(dt))
6089 return ERR_PTR(-ENOENT);
6093 return ERR_PTR(-ENOMEM);
6095 ipd = osd_it_ipd_get(env, bag);
6096 if (likely(ipd != NULL)) {
6100 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6101 return (struct dt_it *)it;
6104 return ERR_PTR(-ENOMEM);
6109 * free given Iterator.
6111 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6113 struct osd_it_iam *it = (struct osd_it_iam *)di;
6114 struct osd_object *obj = it->oi_obj;
6116 iam_it_fini(&it->oi_it);
6117 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6118 osd_object_put(env, obj);
6123 * Move Iterator to record specified by \a key
6125 * \param di osd iterator
6126 * \param key key for index
6128 * \retval +ve di points to record with least key not larger than key
6129 * \retval 0 di points to exact matched key
6130 * \retval -ve failure
6133 static int osd_it_iam_get(const struct lu_env *env,
6134 struct dt_it *di, const struct dt_key *key)
6136 struct osd_thread_info *oti = osd_oti_get(env);
6137 struct osd_it_iam *it = (struct osd_it_iam *)di;
6139 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6140 /* swab quota uid/gid */
6141 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6142 key = (struct dt_key *)&oti->oti_quota_id;
6145 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6151 * \param di osd iterator
6153 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6155 struct osd_it_iam *it = (struct osd_it_iam *)di;
6157 iam_it_put(&it->oi_it);
6161 * Move iterator by one record
6163 * \param di osd iterator
6165 * \retval +1 end of container reached
6167 * \retval -ve failure
6170 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6172 struct osd_it_iam *it = (struct osd_it_iam *)di;
6174 return iam_it_next(&it->oi_it);
6178 * Return pointer to the key under iterator.
6181 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6182 const struct dt_it *di)
6184 struct osd_thread_info *oti = osd_oti_get(env);
6185 struct osd_it_iam *it = (struct osd_it_iam *)di;
6186 struct osd_object *obj = it->oi_obj;
6189 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6191 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6192 /* swab quota uid/gid */
6193 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6194 key = (struct dt_key *)&oti->oti_quota_id;
6201 * Return size of key under iterator (in bytes)
6204 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6206 struct osd_it_iam *it = (struct osd_it_iam *)di;
6208 return iam_it_key_size(&it->oi_it);
6212 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6214 /* check if file type is required */
6215 if (ent->lde_attrs & LUDA_TYPE) {
6216 struct luda_type *lt;
6217 int align = sizeof(*lt) - 1;
6219 len = (len + align) & ~align;
6220 lt = (struct luda_type *)(ent->lde_name + len);
6221 lt->lt_type = cpu_to_le16(DTTOIF(type));
6224 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6228 * build lu direct from backend fs dirent.
6232 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6233 char *name, __u16 namelen, __u16 type, __u32 attr)
6235 ent->lde_attrs = attr | LUDA_FID;
6236 fid_cpu_to_le(&ent->lde_fid, fid);
6238 ent->lde_hash = cpu_to_le64(offset);
6239 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6241 strncpy(ent->lde_name, name, namelen);
6242 ent->lde_name[namelen] = '\0';
6243 ent->lde_namelen = cpu_to_le16(namelen);
6245 /* append lustre attributes */
6246 osd_it_append_attrs(ent, namelen, type);
6250 * Return pointer to the record under iterator.
6252 static int osd_it_iam_rec(const struct lu_env *env,
6253 const struct dt_it *di,
6254 struct dt_rec *dtrec, __u32 attr)
6256 struct osd_it_iam *it = (struct osd_it_iam *)di;
6257 struct osd_thread_info *info = osd_oti_get(env);
6261 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6262 const struct osd_fid_pack *rec;
6263 struct lu_fid *fid = &info->oti_fid;
6264 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6270 name = (char *)iam_it_key_get(&it->oi_it);
6272 RETURN(PTR_ERR(name));
6274 namelen = iam_it_key_size(&it->oi_it);
6276 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6278 RETURN(PTR_ERR(rec));
6280 rc = osd_fid_unpack(fid, rec);
6284 hash = iam_it_store(&it->oi_it);
6286 /* IAM does not store object type in IAM index (dir) */
6287 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6289 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6290 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6291 (struct iam_rec *)dtrec);
6292 osd_quota_unpack(it->oi_obj, dtrec);
6294 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6295 (struct iam_rec *)dtrec);
6302 * Returns cookie for current Iterator position.
6304 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6306 struct osd_it_iam *it = (struct osd_it_iam *)di;
6308 return iam_it_store(&it->oi_it);
6312 * Restore iterator from cookie.
6314 * \param di osd iterator
6315 * \param hash Iterator location cookie
6317 * \retval +ve di points to record with least key not larger than key.
6318 * \retval 0 di points to exact matched key
6319 * \retval -ve failure
6322 static int osd_it_iam_load(const struct lu_env *env,
6323 const struct dt_it *di, __u64 hash)
6325 struct osd_it_iam *it = (struct osd_it_iam *)di;
6327 return iam_it_load(&it->oi_it, hash);
6330 static const struct dt_index_operations osd_index_iam_ops = {
6331 .dio_lookup = osd_index_iam_lookup,
6332 .dio_declare_insert = osd_index_declare_iam_insert,
6333 .dio_insert = osd_index_iam_insert,
6334 .dio_declare_delete = osd_index_declare_iam_delete,
6335 .dio_delete = osd_index_iam_delete,
6337 .init = osd_it_iam_init,
6338 .fini = osd_it_iam_fini,
6339 .get = osd_it_iam_get,
6340 .put = osd_it_iam_put,
6341 .next = osd_it_iam_next,
6342 .key = osd_it_iam_key,
6343 .key_size = osd_it_iam_key_size,
6344 .rec = osd_it_iam_rec,
6345 .store = osd_it_iam_store,
6346 .load = osd_it_iam_load
6352 * Creates or initializes iterator context.
6354 * \retval struct osd_it_ea, iterator structure on success
6357 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6358 struct dt_object *dt,
6361 struct osd_object *obj = osd_dt_obj(dt);
6362 struct osd_thread_info *info = osd_oti_get(env);
6363 struct osd_it_ea *oie;
6365 struct lu_object *lo = &dt->do_lu;
6366 struct dentry *obj_dentry;
6370 if (!dt_object_exists(dt) || obj->oo_destroyed)
6371 RETURN(ERR_PTR(-ENOENT));
6373 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6375 RETURN(ERR_PTR(-ENOMEM));
6376 obj_dentry = &oie->oie_dentry;
6378 obj_dentry->d_inode = obj->oo_inode;
6379 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
6380 obj_dentry->d_name.hash = 0;
6382 oie->oie_rd_dirent = 0;
6383 oie->oie_it_dirent = 0;
6384 oie->oie_dirent = NULL;
6385 if (unlikely(!info->oti_it_ea_buf_used)) {
6386 oie->oie_buf = info->oti_it_ea_buf;
6387 info->oti_it_ea_buf_used = 1;
6389 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6390 if (oie->oie_buf == NULL)
6391 RETURN(ERR_PTR(-ENOMEM));
6395 file = &oie->oie_file;
6397 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6398 if (attr & LUDA_64BITHASH)
6399 file->f_mode = FMODE_64BITHASH;
6401 file->f_mode = FMODE_32BITHASH;
6402 file->f_path.dentry = obj_dentry;
6403 file->f_mapping = obj->oo_inode->i_mapping;
6404 file->f_op = obj->oo_inode->i_fop;
6405 set_file_inode(file, obj->oo_inode);
6408 RETURN((struct dt_it *)oie);
6412 * Destroy or finishes iterator context.
6414 * \param di iterator structure to be destroyed
6416 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6418 struct osd_thread_info *info = osd_oti_get(env);
6419 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6420 struct osd_object *obj = oie->oie_obj;
6421 struct inode *inode = obj->oo_inode;
6424 oie->oie_file.f_op->release(inode, &oie->oie_file);
6425 osd_object_put(env, obj);
6426 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6427 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6429 info->oti_it_ea_buf_used = 0;
6430 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6435 * It position the iterator at given key, so that next lookup continues from
6436 * that key Or it is similar to dio_it->load() but based on a key,
6437 * rather than file position.
6439 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6442 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6444 static int osd_it_ea_get(const struct lu_env *env,
6445 struct dt_it *di, const struct dt_key *key)
6447 struct osd_it_ea *it = (struct osd_it_ea *)di;
6450 LASSERT(((const char *)key)[0] == '\0');
6451 it->oie_file.f_pos = 0;
6452 it->oie_rd_dirent = 0;
6453 it->oie_it_dirent = 0;
6454 it->oie_dirent = NULL;
6462 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6466 struct osd_filldir_cbs {
6467 #ifdef HAVE_DIR_CONTEXT
6468 struct dir_context ctx;
6470 struct osd_it_ea *it;
6473 * It is called internally by ->readdir(). It fills the
6474 * iterator's in-memory data structure with required
6475 * information i.e. name, namelen, rec_size etc.
6477 * \param buf in which information to be filled in.
6478 * \param name name of the file in given dir
6480 * \retval 0 on success
6481 * \retval 1 on buffer full
6483 #ifdef HAVE_FILLDIR_USE_CTX
6484 static int osd_ldiskfs_filldir(struct dir_context *buf,
6486 static int osd_ldiskfs_filldir(void *buf,
6488 const char *name, int namelen,
6489 loff_t offset, __u64 ino, unsigned int d_type)
6491 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6492 struct osd_object *obj = it->oie_obj;
6493 struct osd_it_ea_dirent *ent = it->oie_dirent;
6494 struct lu_fid *fid = &ent->oied_fid;
6495 struct osd_fid_pack *rec;
6499 /* this should never happen */
6500 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6501 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6505 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6509 /* "." is just the object itself. */
6510 if (namelen == 1 && name[0] == '.') {
6511 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6512 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6513 rec = (struct osd_fid_pack *)(name + namelen + 1);
6514 if (osd_fid_unpack(fid, rec) != 0)
6519 d_type &= ~LDISKFS_DIRENT_LUFID;
6521 /* NOT export local root. */
6522 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6523 ino = obj->oo_inode->i_ino;
6524 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6527 ent->oied_ino = ino;
6528 ent->oied_off = offset;
6529 ent->oied_namelen = namelen;
6530 ent->oied_type = d_type;
6532 memcpy(ent->oied_name, name, namelen);
6534 it->oie_rd_dirent++;
6535 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6540 * Calls ->readdir() to load a directory entry at a time
6541 * and stored it in iterator's in-memory data structure.
6543 * \param di iterator's in memory structure
6545 * \retval 0 on success
6546 * \retval -ve on error
6547 * \retval +1 reach the end of entry
6549 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6550 const struct dt_it *di)
6552 struct osd_it_ea *it = (struct osd_it_ea *)di;
6553 struct osd_object *obj = it->oie_obj;
6554 struct inode *inode = obj->oo_inode;
6555 struct htree_lock *hlock = NULL;
6556 struct file *filp = &it->oie_file;
6558 struct osd_filldir_cbs buf = {
6559 #ifdef HAVE_DIR_CONTEXT
6560 .ctx.actor = osd_ldiskfs_filldir,
6566 it->oie_dirent = it->oie_buf;
6567 it->oie_rd_dirent = 0;
6569 if (obj->oo_hl_head != NULL) {
6570 hlock = osd_oti_get(env)->oti_hlock;
6571 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6572 inode, LDISKFS_HLOCK_READDIR);
6574 down_read(&obj->oo_ext_idx_sem);
6577 #ifdef HAVE_DIR_CONTEXT
6578 buf.ctx.pos = filp->f_pos;
6579 #ifdef HAVE_ITERATE_SHARED
6580 rc = inode->i_fop->iterate_shared(filp, &buf.ctx);
6582 rc = inode->i_fop->iterate(filp, &buf.ctx);
6584 filp->f_pos = buf.ctx.pos;
6586 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6590 ldiskfs_htree_unlock(hlock);
6592 up_read(&obj->oo_ext_idx_sem);
6594 if (it->oie_rd_dirent == 0) {
6596 * If it does not get any dirent, it means it has been reached
6597 * to the end of the dir
6599 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6603 it->oie_dirent = it->oie_buf;
6604 it->oie_it_dirent = 1;
6611 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6612 * to load a directory entry at a time and stored it in
6613 * iterator's in-memory data structure.
6615 * \param di iterator's in memory structure
6617 * \retval +ve iterator reached to end
6618 * \retval 0 iterator not reached to end
6619 * \retval -ve on error
6621 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6623 struct osd_it_ea *it = (struct osd_it_ea *)di;
6628 if (it->oie_it_dirent < it->oie_rd_dirent) {
6630 (void *)it->oie_dirent +
6631 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6632 it->oie_dirent->oied_namelen);
6633 it->oie_it_dirent++;
6636 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6639 rc = osd_ldiskfs_it_fill(env, di);
6646 * Returns the key at current position from iterator's in memory structure.
6648 * \param di iterator's in memory structure
6650 * \retval key i.e. struct dt_key on success
6652 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6653 const struct dt_it *di)
6655 struct osd_it_ea *it = (struct osd_it_ea *)di;
6657 return (struct dt_key *)it->oie_dirent->oied_name;
6661 * Returns key's size at current position from iterator's in memory structure.
6663 * \param di iterator's in memory structure
6665 * \retval key_size i.e. struct dt_key on success
6667 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6669 struct osd_it_ea *it = (struct osd_it_ea *)di;
6671 return it->oie_dirent->oied_namelen;
6674 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6676 if (LDISKFS_DIR_REC_LEN(de) >=
6677 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6684 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6685 unsigned int blocksize, bool dotdot)
6688 return osd_dotdot_has_space(de);
6690 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6691 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6698 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6699 handle_t *jh, struct dentry *dentry,
6700 const struct lu_fid *fid, struct buffer_head *bh,
6701 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6704 struct inode *dir = dentry->d_parent->d_inode;
6705 struct inode *inode = dentry->d_inode;
6706 struct osd_fid_pack *rec;
6707 struct ldiskfs_dentry_param *ldp;
6708 int namelen = dentry->d_name.len;
6710 struct osd_thread_info *info = osd_oti_get(env);
6714 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6717 /* There is enough space to hold the FID-in-dirent. */
6718 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6719 rc = ldiskfs_journal_get_write_access(jh, bh);
6723 de->name[namelen] = 0;
6724 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6725 rec->fp_len = sizeof(struct lu_fid) + 1;
6726 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6727 de->file_type |= LDISKFS_DIRENT_LUFID;
6728 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6735 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6739 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6740 osd_get_ldiskfs_dirent_param(ldp, fid);
6741 dentry->d_fsdata = (void *)ldp;
6742 ll_vfs_dq_init(dir);
6743 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6745 * It is too bad, we cannot reinsert the name entry back.
6746 * That means we lose it!
6749 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6750 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6751 osd_ino2name(inode),
6752 dir->i_ino, dir->i_generation, namelen,
6753 dentry->d_name.name, PFID(fid), rc);
6759 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6760 struct osd_it_ea *it, struct lu_fid *fid,
6761 struct osd_inode_id *id, __u32 *attr)
6763 struct osd_thread_info *info = osd_oti_get(env);
6764 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6765 struct osd_device *dev = osd_obj2dev(obj);
6766 struct super_block *sb = osd_sb(dev);
6767 const char *devname = osd_name(dev);
6768 struct osd_it_ea_dirent *ent = it->oie_dirent;
6769 struct inode *dir = obj->oo_inode;
6770 struct htree_lock *hlock = NULL;
6771 struct buffer_head *bh = NULL;
6772 handle_t *jh = NULL;
6773 struct ldiskfs_dir_entry_2 *de;
6774 struct dentry *dentry;
6775 struct inode *inode;
6776 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6779 bool dotdot = false;
6784 if (ent->oied_name[0] == '.') {
6785 if (ent->oied_namelen == 1)
6788 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6792 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6793 inode = osd_iget(info, dev, id);
6794 if (IS_ERR(inode)) {
6795 rc = PTR_ERR(inode);
6796 if (rc == -ENOENT || rc == -ESTALE) {
6798 * Maybe dangling name entry, or
6799 * corrupted directory entry.
6801 *attr |= LUDA_UNKNOWN;
6804 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6805 "check_repair, dir = %lu/%u, name = %.*s, "
6806 "ino = %llu, rc = %d\n",
6807 devname, dir->i_ino, dir->i_generation,
6808 ent->oied_namelen, ent->oied_name,
6815 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6817 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6818 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6824 * We need to ensure that the name entry is still valid.
6825 * Because it may be removed or renamed by other already.
6827 * The unlink or rename operation will start journal before PDO lock,
6828 * so to avoid deadlock, here we need to start journal handle before
6829 * related PDO lock also. But because we do not know whether there
6830 * will be something to be repaired before PDO lock, we just start
6831 * journal without conditions.
6833 * We may need to remove the name entry firstly, then insert back.
6834 * One credit is for user quota file update.
6835 * One credit is for group quota file update.
6836 * Two credits are for dirty inode.
6838 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6839 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6841 if (dev->od_dirent_journal != 0) {
6844 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6847 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6848 "check_repair, dir = %lu/%u, credits = %d, "
6849 "name = %.*s, ino = %llu: rc = %d\n",
6850 devname, dir->i_ino, dir->i_generation, credits,
6851 ent->oied_namelen, ent->oied_name,
6854 GOTO(out_inode, rc);
6857 if (obj->oo_hl_head != NULL) {
6858 hlock = osd_oti_get(env)->oti_hlock;
6860 * "0" means exclusive lock for the whole directory.
6861 * We need to prevent others access such name entry
6862 * during the delete + insert. Neither HLOCK_ADD nor
6863 * HLOCK_DEL cannot guarantee the atomicity.
6865 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6867 down_write(&obj->oo_ext_idx_sem);
6870 if (obj->oo_hl_head != NULL) {
6871 hlock = osd_oti_get(env)->oti_hlock;
6872 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6873 LDISKFS_HLOCK_LOOKUP);
6875 down_read(&obj->oo_ext_idx_sem);
6879 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6880 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6881 *attr |= LUDA_IGNORE;
6887 * For dotdot entry, if there is not enough space to hold the
6888 * FID-in-dirent, just keep them there. It only happens when the
6889 * device upgraded from 1.8 or restored from MDT file-level backup.
6890 * For the whole directory, only dotdot entry have no FID-in-dirent
6891 * and needs to get FID from LMA when readdir, it will not affect the
6894 if (dotdot && !osd_dotdot_has_space(de)) {
6895 *attr |= LUDA_UNKNOWN;
6901 if (lu_fid_eq(fid, &lma->lma_self_fid))
6904 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6905 struct lu_fid *tfid = &lma->lma_self_fid;
6907 if (likely(dotdot &&
6908 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6909 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6911 * It must be REMOTE_PARENT_DIR and as the
6912 * 'dotdot' entry of remote directory
6914 *attr |= LUDA_IGNORE;
6916 CDEBUG(D_LFSCK, "%s: expect remote agent "
6917 "parent directory, but got %.*s under "
6918 "dir = %lu/%u with the FID "DFID"\n",
6919 devname, ent->oied_namelen,
6920 ent->oied_name, dir->i_ino,
6921 dir->i_generation, PFID(tfid));
6923 *attr |= LUDA_UNKNOWN;
6930 if (!fid_is_zero(fid)) {
6931 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6933 if (rc == -ENOENT ||
6935 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
6937 * linkEA does not recognize the dirent entry,
6938 * it may because the dirent entry corruption
6939 * and points to other's inode.
6941 CDEBUG(D_LFSCK, "%s: the target inode does not "
6942 "recognize the dirent, dir = %lu/%u, "
6943 " name = %.*s, ino = %llu, "
6944 DFID": rc = %d\n", devname, dir->i_ino,
6945 dir->i_generation, ent->oied_namelen,
6946 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6947 *attr |= LUDA_UNKNOWN;
6952 if (rc && rc != -ENODATA) {
6953 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6954 "dir = %lu/%u, name = %.*s, ino = %llu, "
6955 DFID": rc = %d\n", devname, dir->i_ino,
6956 dir->i_generation, ent->oied_namelen,
6957 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6958 *attr |= LUDA_UNKNOWN;
6966 * linkEA recognizes the dirent entry, the FID-in-LMA is
6967 * valid, trusted, in spite of fid_is_sane(fid) or not.
6969 if (*attr & LUDA_VERIFY_DRYRUN) {
6970 *fid = lma->lma_self_fid;
6971 *attr |= LUDA_REPAIR;
6978 dev->od_dirent_journal = 1;
6979 if (hlock != NULL) {
6980 ldiskfs_htree_unlock(hlock);
6983 up_read(&obj->oo_ext_idx_sem);
6989 *fid = lma->lma_self_fid;
6991 /* Update or append the FID-in-dirent. */
6992 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6993 bh, de, hlock, dotdot);
6995 *attr |= LUDA_REPAIR;
6997 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6998 "the dirent, dir = %lu/%u, name = %.*s, "
6999 "ino = %llu, "DFID": rc = %d\n",
7000 devname, dir->i_ino, dir->i_generation,
7001 ent->oied_namelen, ent->oied_name,
7002 ent->oied_ino, PFID(fid), rc);
7004 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7005 if (*attr & LUDA_VERIFY_DRYRUN) {
7006 if (fid_is_sane(fid)) {
7007 *attr |= LUDA_REPAIR;
7008 } else if (dev->od_index == 0) {
7009 lu_igif_build(fid, inode->i_ino,
7010 inode->i_generation);
7011 *attr |= LUDA_UPGRADE;
7019 dev->od_dirent_journal = 1;
7020 if (hlock != NULL) {
7021 ldiskfs_htree_unlock(hlock);
7024 up_read(&obj->oo_ext_idx_sem);
7031 if (unlikely(fid_is_sane(fid))) {
7033 * FID-in-dirent exists, but FID-in-LMA is lost.
7034 * Trust the FID-in-dirent, and add FID-in-LMA.
7036 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7038 *attr |= LUDA_REPAIR;
7040 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7041 "update dirent, dir = %lu/%u, "
7042 "name = %.*s, ino = %llu, "
7044 devname, dir->i_ino, dir->i_generation,
7045 ent->oied_namelen, ent->oied_name,
7046 ent->oied_ino, PFID(fid), rc);
7047 } else if (dev->od_index == 0) {
7048 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7050 * It is probably IGIF object. Only aappend the
7051 * FID-in-dirent. OI scrub will process FID-in-LMA.
7053 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7054 bh, de, hlock, dotdot);
7056 *attr |= LUDA_UPGRADE;
7058 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7059 "after the dirent, dir = %lu/%u, "
7060 "name = %.*s, ino = %llu, "
7062 devname, dir->i_ino, dir->i_generation,
7063 ent->oied_namelen, ent->oied_name,
7064 ent->oied_ino, PFID(fid), rc);
7073 if (hlock != NULL) {
7074 ldiskfs_htree_unlock(hlock);
7076 if (dev->od_dirent_journal != 0)
7077 up_write(&obj->oo_ext_idx_sem);
7079 up_read(&obj->oo_ext_idx_sem);
7083 ldiskfs_journal_stop(jh);
7087 if (rc >= 0 && !dirty)
7088 dev->od_dirent_journal = 0;
7094 * Returns the value at current position from iterator's in memory structure.
7096 * \param di struct osd_it_ea, iterator's in memory structure
7097 * \param attr attr requested for dirent.
7098 * \param lde lustre dirent
7100 * \retval 0 no error and \param lde has correct lustre dirent.
7101 * \retval -ve on error
7103 static inline int osd_it_ea_rec(const struct lu_env *env,
7104 const struct dt_it *di,
7105 struct dt_rec *dtrec, __u32 attr)
7107 struct osd_it_ea *it = (struct osd_it_ea *)di;
7108 struct osd_object *obj = it->oie_obj;
7109 struct osd_device *dev = osd_obj2dev(obj);
7110 struct osd_thread_info *oti = osd_oti_get(env);
7111 struct osd_inode_id *id = &oti->oti_id;
7112 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7113 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7114 __u32 ino = it->oie_dirent->oied_ino;
7119 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7121 if (attr & LUDA_VERIFY) {
7122 if (unlikely(is_remote_parent_ino(dev, ino))) {
7123 attr |= LUDA_IGNORE;
7125 * If the parent is on remote MDT, and there
7126 * is no FID-in-dirent, then we have to get
7127 * the parent FID from the linkEA.
7129 if (!fid_is_sane(fid) &&
7130 it->oie_dirent->oied_namelen == 2 &&
7131 it->oie_dirent->oied_name[0] == '.' &&
7132 it->oie_dirent->oied_name[1] == '.')
7133 osd_get_pfid_from_linkea(env, obj, fid);
7135 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7139 if (!fid_is_sane(fid))
7140 attr |= LUDA_UNKNOWN;
7142 attr &= ~LU_DIRENT_ATTRS_MASK;
7143 if (!fid_is_sane(fid)) {
7144 bool is_dotdot = false;
7146 if (it->oie_dirent->oied_namelen == 2 &&
7147 it->oie_dirent->oied_name[0] == '.' &&
7148 it->oie_dirent->oied_name[1] == '.')
7151 * If the parent is on remote MDT, and there
7152 * is no FID-in-dirent, then we have to get
7153 * the parent FID from the linkEA.
7155 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7156 rc = osd_get_pfid_from_linkea(env, obj, fid);
7158 if (is_dotdot == false &&
7159 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7162 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7165 osd_id_gen(id, ino, OSD_OII_NOGEN);
7169 /* Pack the entry anyway, at least the offset is right. */
7170 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7171 it->oie_dirent->oied_name,
7172 it->oie_dirent->oied_namelen,
7173 it->oie_dirent->oied_type, attr);
7178 if (osd_remote_fid(env, dev, fid))
7181 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
7182 osd_add_oi_cache(oti, dev, id, fid);
7184 RETURN(rc > 0 ? 0 : rc);
7188 * Returns the record size size at current position.
7190 * This function will return record(lu_dirent) size in bytes.
7192 * \param[in] env execution environment
7193 * \param[in] di iterator's in memory structure
7194 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7195 * calculate the lu_dirent size.
7197 * \retval record size(in bytes & in memory) of the current lu_dirent
7200 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7203 struct osd_it_ea *it = (struct osd_it_ea *)di;
7205 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7209 * Returns a cookie for current position of the iterator head, so that
7210 * user can use this cookie to load/start the iterator next time.
7212 * \param di iterator's in memory structure
7214 * \retval cookie for current position, on success
7216 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7218 struct osd_it_ea *it = (struct osd_it_ea *)di;
7220 return it->oie_dirent->oied_off;
7224 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
7225 * to load a directory entry at a time and stored it i inn,
7226 * in iterator's in-memory data structure.
7228 * \param di struct osd_it_ea, iterator's in memory structure
7230 * \retval +ve on success
7231 * \retval -ve on error
7233 static int osd_it_ea_load(const struct lu_env *env,
7234 const struct dt_it *di, __u64 hash)
7236 struct osd_it_ea *it = (struct osd_it_ea *)di;
7240 it->oie_file.f_pos = hash;
7242 rc = osd_ldiskfs_it_fill(env, di);
7253 * Index lookup function for interoperability mode (b11826).
7255 * \param key, key i.e. file name to be searched
7257 * \retval +ve, on success
7258 * \retval -ve, on error
7260 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7261 struct dt_rec *rec, const struct dt_key *key)
7263 struct osd_object *obj = osd_dt_obj(dt);
7268 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7269 LINVRNT(osd_invariant(obj));
7271 rc = osd_ea_lookup_rec(env, obj, rec, key);
7278 * Index and Iterator operations for interoperability
7279 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7281 static const struct dt_index_operations osd_index_ea_ops = {
7282 .dio_lookup = osd_index_ea_lookup,
7283 .dio_declare_insert = osd_index_declare_ea_insert,
7284 .dio_insert = osd_index_ea_insert,
7285 .dio_declare_delete = osd_index_declare_ea_delete,
7286 .dio_delete = osd_index_ea_delete,
7288 .init = osd_it_ea_init,
7289 .fini = osd_it_ea_fini,
7290 .get = osd_it_ea_get,
7291 .put = osd_it_ea_put,
7292 .next = osd_it_ea_next,
7293 .key = osd_it_ea_key,
7294 .key_size = osd_it_ea_key_size,
7295 .rec = osd_it_ea_rec,
7296 .rec_size = osd_it_ea_rec_size,
7297 .store = osd_it_ea_store,
7298 .load = osd_it_ea_load
7302 static void *osd_key_init(const struct lu_context *ctx,
7303 struct lu_context_key *key)
7305 struct osd_thread_info *info;
7307 OBD_ALLOC_PTR(info);
7309 return ERR_PTR(-ENOMEM);
7311 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7312 if (info->oti_it_ea_buf == NULL)
7315 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7317 info->oti_hlock = ldiskfs_htree_lock_alloc();
7318 if (info->oti_hlock == NULL)
7324 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7327 return ERR_PTR(-ENOMEM);
7330 static void osd_key_fini(const struct lu_context *ctx,
7331 struct lu_context_key *key, void *data)
7333 struct osd_thread_info *info = data;
7334 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7335 struct osd_idmap_cache *idc = info->oti_ins_cache;
7337 if (info->oti_dio_pages) {
7339 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7340 if (info->oti_dio_pages[i])
7341 __free_page(info->oti_dio_pages[i]);
7343 OBD_FREE(info->oti_dio_pages,
7344 sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
7347 if (info->oti_inode != NULL)
7349 if (info->oti_hlock != NULL)
7350 ldiskfs_htree_lock_free(info->oti_hlock);
7351 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7352 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7353 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7354 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7355 lu_buf_free(&info->oti_big_buf);
7357 LASSERT(info->oti_ins_cache_size > 0);
7358 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
7359 info->oti_ins_cache = NULL;
7360 info->oti_ins_cache_size = 0;
7365 static void osd_key_exit(const struct lu_context *ctx,
7366 struct lu_context_key *key, void *data)
7368 struct osd_thread_info *info = data;
7370 LASSERT(info->oti_r_locks == 0);
7371 LASSERT(info->oti_w_locks == 0);
7372 LASSERT(info->oti_txns == 0);
7375 /* type constructor/destructor: osd_type_init, osd_type_fini */
7376 LU_TYPE_INIT_FINI(osd, &osd_key);
7378 struct lu_context_key osd_key = {
7379 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7380 .lct_init = osd_key_init,
7381 .lct_fini = osd_key_fini,
7382 .lct_exit = osd_key_exit
7386 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7387 const char *name, struct lu_device *next)
7389 struct osd_device *osd = osd_dev(d);
7391 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7392 sizeof(osd->od_svname))
7394 return osd_procfs_init(osd, name);
7397 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7399 struct seq_server_site *ss = osd_seq_site(osd);
7404 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7407 if (unlikely(ss == NULL))
7410 OBD_ALLOC_PTR(osd->od_cl_seq);
7411 if (osd->od_cl_seq == NULL)
7414 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7415 osd->od_svname, ss->ss_server_seq);
7417 OBD_FREE_PTR(osd->od_cl_seq);
7418 osd->od_cl_seq = NULL;
7422 if (ss->ss_node_id == 0) {
7424 * If the OSD on the sequence controller(MDT0), then allocate
7425 * sequence here, otherwise allocate sequence after connected
7426 * to MDT0 (see mdt_register_lwp_callback()).
7428 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7429 &osd->od_cl_seq->lcs_space, env);
7435 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7437 if (osd->od_cl_seq == NULL)
7440 seq_client_fini(osd->od_cl_seq);
7441 OBD_FREE_PTR(osd->od_cl_seq);
7442 osd->od_cl_seq = NULL;
7445 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7449 /* shutdown quota slave instance associated with the device */
7450 if (o->od_quota_slave_md != NULL) {
7451 struct qsd_instance *qsd = o->od_quota_slave_md;
7453 o->od_quota_slave_md = NULL;
7457 if (o->od_quota_slave_dt != NULL) {
7458 struct qsd_instance *qsd = o->od_quota_slave_dt;
7460 o->od_quota_slave_dt = NULL;
7464 osd_fid_fini(env, o);
7465 osd_scrub_cleanup(env, o);
7470 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7474 if (o->od_mnt != NULL) {
7475 shrink_dcache_sb(osd_sb(o));
7476 osd_sync(env, &o->od_dt_dev);
7485 static int osd_mount(const struct lu_env *env,
7486 struct osd_device *o, struct lustre_cfg *cfg)
7488 const char *name = lustre_cfg_string(cfg, 0);
7489 const char *dev = lustre_cfg_string(cfg, 1);
7491 unsigned long page, s_flags, lmd_flags = 0;
7492 struct page *__page;
7493 struct file_system_type *type;
7494 char *options = NULL;
7496 struct osd_thread_info *info = osd_oti_get(env);
7497 struct lu_fid *fid = &info->oti_fid;
7498 struct inode *inode;
7499 int rc = 0, force_over_512tb = 0;
7503 if (o->od_mnt != NULL)
7506 if (strlen(dev) >= sizeof(o->od_mntdev))
7508 strcpy(o->od_mntdev, dev);
7510 str = lustre_cfg_string(cfg, 2);
7511 s_flags = simple_strtoul(str, NULL, 0);
7512 str = strstr(str, ":");
7514 lmd_flags = simple_strtoul(str + 1, NULL, 0);
7515 opts = lustre_cfg_string(cfg, 3);
7517 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7518 CERROR("%s: device %s extents feature is not guaranteed to "
7519 "work on big-endian systems. Use \"bigendian_extents\" "
7520 "mount option to override.\n", name, dev);
7524 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7525 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7526 CWARN("force_over_128tb option is deprecated. "
7527 "Filesystems less than 512TB can be created without any "
7528 "force options. Use force_over_512tb option for "
7529 "filesystems greater than 512TB.\n");
7532 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7533 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7534 CWARN("force_over_256tb option is deprecated. "
7535 "Filesystems less than 512TB can be created without any "
7536 "force options. Use force_over_512tb option for "
7537 "filesystems greater than 512TB.\n");
7541 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
7542 force_over_512tb = 1;
7544 __page = alloc_page(GFP_KERNEL);
7546 GOTO(out, rc = -ENOMEM);
7547 page = (unsigned long)page_address(__page);
7548 options = (char *)page;
7551 /* strip out the options for back compatiblity */
7552 static char *sout[] = {
7559 /* strip out option we processed in osd */
7560 "bigendian_extents",
7566 strcat(options, opts);
7567 for (rc = 0, str = options; sout[rc]; ) {
7568 char *op = strstr(str, sout[rc]);
7575 if (op == options || *(op - 1) == ',') {
7576 str = op + strlen(sout[rc]);
7577 if (*str == ',' || *str == '\0') {
7578 *str == ',' ? str++ : str;
7579 memmove(op, str, strlen(str) + 1);
7582 for (str = op; *str != ',' && *str != '\0'; str++)
7586 strncat(options, "user_xattr,acl", 14);
7589 /* Glom up mount options */
7590 if (*options != '\0')
7591 strcat(options, ",");
7592 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7594 type = get_fs_type("ldiskfs");
7596 CERROR("%s: cannot find ldiskfs module\n", name);
7597 GOTO(out, rc = -ENODEV);
7600 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7601 module_put(type->owner);
7603 if (IS_ERR(o->od_mnt)) {
7604 rc = PTR_ERR(o->od_mnt);
7606 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7610 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7611 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7612 force_over_512tb == 0) {
7613 CERROR("%s: device %s LDISKFS does not support filesystems "
7614 "greater than 512TB and can cause data corruption. "
7615 "Use \"force_over_512tb\" mount option to override.\n",
7617 GOTO(out_mnt, rc = -EINVAL);
7620 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7621 if (priv_dev_set_rdonly) {
7622 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
7623 o->od_dt_dev.dd_rdonly = 1;
7624 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
7627 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7630 GOTO(out_mnt, rc = -EOPNOTSUPP);
7632 } else if (priv_dev_check_rdonly &&
7633 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
7634 CERROR("%s: underlying device %s is marked as "
7635 "read-only. Setup failed\n", name, dev);
7637 GOTO(out_mnt, rc = -EROFS);
7640 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7641 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7642 GOTO(out_mnt, rc = -EINVAL);
7645 #ifdef LDISKFS_MOUNT_DIRDATA
7646 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7647 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7648 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7649 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7650 "enabling the dirdata feature. If you do not want to "
7651 "downgrade to Lustre-1.x again, you can enable it via "
7652 "'tune2fs -O dirdata device'\n", name, dev);
7654 inode = osd_sb(o)->s_root->d_inode;
7655 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7656 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7658 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7662 if (lmd_flags & LMD_FLG_NOSCRUB)
7663 o->od_auto_scrub_interval = AS_NEVER;
7665 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7666 /* do not use pagecache with flash-backed storage */
7667 o->od_writethrough_cache = 0;
7668 o->od_read_cache = 0;
7679 __free_page(__page);
7684 static struct lu_device *osd_device_fini(const struct lu_env *env,
7685 struct lu_device *d)
7687 struct osd_device *o = osd_dev(d);
7691 osd_index_backup(env, o, false);
7692 osd_shutdown(env, o);
7694 osd_obj_map_fini(o);
7700 static int osd_device_init0(const struct lu_env *env,
7701 struct osd_device *o,
7702 struct lustre_cfg *cfg)
7704 struct lu_device *l = osd2lu_dev(o);
7705 struct osd_thread_info *info;
7709 /* if the module was re-loaded, env can loose its keys */
7710 rc = lu_env_refill((struct lu_env *)env);
7713 info = osd_oti_get(env);
7716 l->ld_ops = &osd_lu_ops;
7717 o->od_dt_dev.dd_ops = &osd_dt_ops;
7719 spin_lock_init(&o->od_osfs_lock);
7720 mutex_init(&o->od_otable_mutex);
7721 INIT_LIST_HEAD(&o->od_orphan_list);
7722 INIT_LIST_HEAD(&o->od_index_backup_list);
7723 INIT_LIST_HEAD(&o->od_index_restore_list);
7724 spin_lock_init(&o->od_lock);
7725 o->od_index_backup_policy = LIBP_NONE;
7728 o->od_read_cache = 1;
7729 o->od_writethrough_cache = 1;
7730 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7731 o->od_auto_scrub_interval = AS_DEFAULT;
7733 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7734 sizeof(o->od_svname));
7735 if (cplen >= sizeof(o->od_svname)) {
7740 o->od_index_backup_stop = 0;
7741 o->od_index = -1; /* -1 means index is invalid */
7742 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7743 if (rc == LDD_F_SV_TYPE_OST)
7746 o->od_full_scrub_ratio = OFSR_DEFAULT;
7747 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7748 rc = osd_mount(env, o, cfg);
7752 rc = osd_obj_map_init(env, o);
7756 rc = lu_site_init(&o->od_site, l);
7758 GOTO(out_compat, rc);
7759 o->od_site.ls_bottom_dev = l;
7761 rc = lu_site_init_finish(&o->od_site);
7765 INIT_LIST_HEAD(&o->od_ios_list);
7766 /* setup scrub, including OI files initialization */
7768 rc = osd_scrub_setup(env, o);
7773 rc = osd_procfs_init(o, o->od_svname);
7775 CERROR("%s: can't initialize procfs: rc = %d\n",
7777 GOTO(out_scrub, rc);
7780 LASSERT(l->ld_site->ls_linkage.next != NULL);
7781 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7783 /* initialize quota slave instance */
7784 /* currently it's no need to prepare qsd_instance_md for OST */
7785 if (!o->od_is_ost) {
7786 o->od_quota_slave_md = qsd_init(env, o->od_svname,
7788 o->od_proc_entry, true);
7789 if (IS_ERR(o->od_quota_slave_md)) {
7790 rc = PTR_ERR(o->od_quota_slave_md);
7791 o->od_quota_slave_md = NULL;
7792 GOTO(out_procfs, rc);
7796 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
7797 o->od_proc_entry, false);
7799 if (IS_ERR(o->od_quota_slave_dt)) {
7800 if (o->od_quota_slave_md != NULL) {
7801 qsd_fini(env, o->od_quota_slave_md);
7802 o->od_quota_slave_md = NULL;
7805 rc = PTR_ERR(o->od_quota_slave_dt);
7806 o->od_quota_slave_dt = NULL;
7807 GOTO(out_procfs, rc);
7815 osd_scrub_cleanup(env, o);
7817 lu_site_fini(&o->od_site);
7819 osd_obj_map_fini(o);
7826 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7827 struct lu_device_type *t,
7828 struct lustre_cfg *cfg)
7830 struct osd_device *o;
7835 return ERR_PTR(-ENOMEM);
7837 rc = dt_device_init(&o->od_dt_dev, t);
7840 * Because the ctx might be revived in dt_device_init,
7841 * refill the env here
7843 lu_env_refill((struct lu_env *)env);
7844 rc = osd_device_init0(env, o, cfg);
7846 dt_device_fini(&o->od_dt_dev);
7849 if (unlikely(rc != 0))
7852 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7855 static struct lu_device *osd_device_free(const struct lu_env *env,
7856 struct lu_device *d)
7858 struct osd_device *o = osd_dev(d);
7862 /* XXX: make osd top device in order to release reference */
7863 d->ld_site->ls_top_dev = d;
7864 lu_site_purge(env, d->ld_site, -1);
7865 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7866 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7867 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7869 lu_site_fini(&o->od_site);
7870 dt_device_fini(&o->od_dt_dev);
7875 static int osd_process_config(const struct lu_env *env,
7876 struct lu_device *d, struct lustre_cfg *cfg)
7878 struct osd_device *o = osd_dev(d);
7883 switch (cfg->lcfg_command) {
7885 rc = osd_mount(env, o, cfg);
7889 * For the case LCFG_PRE_CLEANUP is not called in advance,
7890 * that may happend if hit failure during mount process.
7892 osd_index_backup(env, o, false);
7893 lu_dev_del_linkage(d->ld_site, d);
7894 rc = osd_shutdown(env, o);
7897 LASSERT(&o->od_dt_dev);
7898 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7899 cfg, &o->od_dt_dev);
7900 if (rc > 0 || rc == -ENOSYS) {
7901 rc = class_process_proc_param(PARAM_OST,
7902 lprocfs_osd_obd_vars,
7903 cfg, &o->od_dt_dev);
7908 case LCFG_PRE_CLEANUP:
7910 osd_index_backup(env, o,
7911 o->od_index_backup_policy != LIBP_NONE);
7921 static int osd_recovery_complete(const struct lu_env *env,
7922 struct lu_device *d)
7924 struct osd_device *osd = osd_dev(d);
7929 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
7933 * start qsd instance on recovery completion, this notifies the quota
7934 * slave code that we are about to process new requests now
7936 rc = qsd_start(env, osd->od_quota_slave_dt);
7937 if (rc == 0 && osd->od_quota_slave_md != NULL)
7938 rc = qsd_start(env, osd->od_quota_slave_md);
7944 * we use exports to track all osd users
7946 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7947 struct obd_device *obd, struct obd_uuid *cluuid,
7948 struct obd_connect_data *data, void *localdata)
7950 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7951 struct lustre_handle conn;
7956 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7958 rc = class_connect(&conn, obd, cluuid);
7962 *exp = class_conn2export(&conn);
7964 spin_lock(&osd->od_osfs_lock);
7966 spin_unlock(&osd->od_osfs_lock);
7972 * once last export (we don't count self-export) disappeared
7973 * osd can be released
7975 static int osd_obd_disconnect(struct obd_export *exp)
7977 struct obd_device *obd = exp->exp_obd;
7978 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7979 int rc, release = 0;
7983 /* Only disconnect the underlying layers on the final disconnect. */
7984 spin_lock(&osd->od_osfs_lock);
7986 if (osd->od_connects == 0)
7988 spin_unlock(&osd->od_osfs_lock);
7990 rc = class_disconnect(exp); /* bz 9811 */
7992 if (rc == 0 && release)
7993 class_manual_cleanup(obd);
7997 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7998 struct lu_device *dev)
8000 struct osd_device *osd = osd_dev(dev);
8001 struct lr_server_data *lsd =
8002 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8007 if (osd->od_quota_slave_md != NULL) {
8008 /* set up quota slave objects for inode */
8009 result = qsd_prepare(env, osd->od_quota_slave_md);
8014 if (osd->od_quota_slave_dt != NULL) {
8015 /* set up quota slave objects for block */
8016 result = qsd_prepare(env, osd->od_quota_slave_dt);
8022 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8023 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8024 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8025 osd->od_index_in_idif = 1;
8027 osd->od_index_in_idif = 0;
8028 result = osd_register_proc_index_in_idif(osd);
8033 osd->od_index_in_idif = 1;
8037 result = osd_fid_init(env, osd);
8042 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
8043 struct lu_fid *fid, struct md_op_data *op_data)
8045 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
8047 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8050 static const struct lu_object_operations osd_lu_obj_ops = {
8051 .loo_object_init = osd_object_init,
8052 .loo_object_delete = osd_object_delete,
8053 .loo_object_release = osd_object_release,
8054 .loo_object_free = osd_object_free,
8055 .loo_object_print = osd_object_print,
8056 .loo_object_invariant = osd_object_invariant
8059 const struct lu_device_operations osd_lu_ops = {
8060 .ldo_object_alloc = osd_object_alloc,
8061 .ldo_process_config = osd_process_config,
8062 .ldo_recovery_complete = osd_recovery_complete,
8063 .ldo_prepare = osd_prepare,
8066 static const struct lu_device_type_operations osd_device_type_ops = {
8067 .ldto_init = osd_type_init,
8068 .ldto_fini = osd_type_fini,
8070 .ldto_start = osd_type_start,
8071 .ldto_stop = osd_type_stop,
8073 .ldto_device_alloc = osd_device_alloc,
8074 .ldto_device_free = osd_device_free,
8076 .ldto_device_init = osd_device_init,
8077 .ldto_device_fini = osd_device_fini
8080 static struct lu_device_type osd_device_type = {
8081 .ldt_tags = LU_DEVICE_DT,
8082 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8083 .ldt_ops = &osd_device_type_ops,
8084 .ldt_ctx_tags = LCT_LOCAL,
8087 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8089 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8090 struct super_block *sb = osd_sb(osd);
8092 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
8096 * lprocfs legacy support.
8098 static struct obd_ops osd_obd_device_ops = {
8099 .o_owner = THIS_MODULE,
8100 .o_connect = osd_obd_connect,
8101 .o_disconnect = osd_obd_disconnect,
8102 .o_fid_alloc = osd_fid_alloc,
8103 .o_health_check = osd_health_check,
8106 static int __init osd_init(void)
8110 CLASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
8111 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8112 /* please, try to keep osd_thread_info smaller than a page */
8113 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
8118 rc = lu_kmem_init(ldiskfs_caches);
8122 #ifdef CONFIG_KALLSYMS
8123 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
8124 priv_dev_check_rdonly =
8125 (void *)kallsyms_lookup_name("dev_check_rdonly");
8128 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8129 lprocfs_osd_module_vars,
8130 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8132 lu_kmem_fini(ldiskfs_caches);
8136 static void __exit osd_exit(void)
8138 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8139 lu_kmem_fini(ldiskfs_caches);
8142 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8143 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8144 MODULE_VERSION(LUSTRE_VERSION_STRING);
8145 MODULE_LICENSE("GPL");
8147 module_init(osd_init);
8148 module_exit(osd_exit);