4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/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/module.h>
43 #include <linux/user_namespace.h>
44 #ifdef HAVE_UIDGID_HEADER
45 # include <linux/uidgid.h>
48 /* LUSTRE_VERSION_CODE */
49 #include <lustre_ver.h>
50 /* prerequisite for linux/xattr.h */
51 #include <linux/types.h>
52 /* prerequisite for linux/xattr.h */
54 /* XATTR_{REPLACE,CREATE} */
55 #include <linux/xattr.h>
57 #include <ldiskfs/ldiskfs.h>
58 #include <ldiskfs/xattr.h>
59 #include <ldiskfs/ldiskfs_extents.h>
62 * struct OBD_{ALLOC,FREE}*()
65 #include <obd_support.h>
66 /* struct ptlrpc_thread */
67 #include <lustre_net.h>
68 #include <lustre_fid.h>
70 #include <lustre_param.h>
72 #include "osd_internal.h"
73 #include "osd_dynlocks.h"
75 /* llo_* api support */
76 #include <md_object.h>
77 #include <lustre_quota.h>
79 #include <lustre_linkea.h>
81 #define PFID_STRIPE_IDX_BITS 16
82 #define PFID_STRIPE_COUNT_MASK ((1 << PFID_STRIPE_IDX_BITS) - 1)
85 module_param(ldiskfs_pdo, int, 0644);
86 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
88 int ldiskfs_track_declares_assert;
89 module_param(ldiskfs_track_declares_assert, int, 0644);
90 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
92 /* Slab to allocate dynlocks */
93 struct kmem_cache *dynlock_cachep;
95 /* Slab to allocate osd_it_ea */
96 struct kmem_cache *osd_itea_cachep;
98 static struct lu_kmem_descr ldiskfs_caches[] = {
100 .ckd_cache = &dynlock_cachep,
101 .ckd_name = "dynlock_cache",
102 .ckd_size = sizeof(struct dynlock_handle)
105 .ckd_cache = &osd_itea_cachep,
106 .ckd_name = "osd_itea_cache",
107 .ckd_size = sizeof(struct osd_it_ea)
114 static const char dot[] = ".";
115 static const char dotdot[] = "..";
116 static const char remote_obj_dir[] = "REM_OBJ_DIR";
118 static const struct lu_object_operations osd_lu_obj_ops;
119 static const struct dt_object_operations osd_obj_ops;
120 static const struct dt_object_operations osd_obj_ea_ops;
121 static const struct dt_object_operations osd_obj_otable_it_ops;
122 static const struct dt_index_operations osd_index_iam_ops;
123 static const struct dt_index_operations osd_index_ea_ops;
125 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
126 const struct lu_fid *fid);
127 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
128 struct osd_device *osd);
130 int osd_trans_declare_op2rb[] = {
131 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
132 [OSD_OT_PUNCH] = OSD_OT_MAX,
133 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
134 [OSD_OT_CREATE] = OSD_OT_DESTROY,
135 [OSD_OT_DESTROY] = OSD_OT_CREATE,
136 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
137 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
138 [OSD_OT_WRITE] = OSD_OT_WRITE,
139 [OSD_OT_INSERT] = OSD_OT_DELETE,
140 [OSD_OT_DELETE] = OSD_OT_INSERT,
141 [OSD_OT_QUOTA] = OSD_OT_MAX,
144 static int osd_has_index(const struct osd_object *obj)
146 return obj->oo_dt.do_index_ops != NULL;
149 static int osd_object_invariant(const struct lu_object *l)
151 return osd_invariant(osd_obj(l));
155 * Concurrency: doesn't matter
159 * Concurrency: doesn't matter
161 static int osd_write_locked(const struct lu_env *env, struct osd_object *o)
163 struct osd_thread_info *oti = osd_oti_get(env);
164 return oti->oti_w_locks > 0 && o->oo_owner == env;
168 * Concurrency: doesn't access mutable data
170 static int osd_root_get(const struct lu_env *env,
171 struct dt_device *dev, struct lu_fid *f)
173 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
178 * the following set of functions are used to maintain per-thread
179 * cache of FID->ino mapping. this mechanism is needed to resolve
180 * FID to inode at dt_insert() which in turn stores ino in the
181 * directory entries to keep ldiskfs compatible with ext[34].
182 * due to locking-originated restrictions we can't lookup ino
183 * using LU cache (deadlock is possible). lookup using OI is quite
184 * expensive. so instead we maintain this cache and methods like
185 * dt_create() fill it. so in the majority of cases dt_insert() is
186 * able to find needed mapping in lockless manner.
188 static struct osd_idmap_cache *
189 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
190 const struct lu_fid *fid)
192 struct osd_thread_info *oti = osd_oti_get(env);
193 struct osd_idmap_cache *idc = oti->oti_ins_cache;
195 for (i = 0; i < oti->oti_ins_cache_used; i++) {
196 if (!lu_fid_eq(&idc[i].oic_fid, fid))
198 if (idc[i].oic_dev != osd)
207 static struct osd_idmap_cache *
208 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
209 const struct lu_fid *fid)
211 struct osd_thread_info *oti = osd_oti_get(env);
212 struct osd_idmap_cache *idc;
215 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
216 i = oti->oti_ins_cache_size * 2;
218 i = OSD_INS_CACHE_SIZE;
219 OBD_ALLOC(idc, sizeof(*idc) * i);
221 return ERR_PTR(-ENOMEM);
222 if (oti->oti_ins_cache != NULL) {
223 memcpy(idc, oti->oti_ins_cache,
224 oti->oti_ins_cache_used * sizeof(*idc));
225 OBD_FREE(oti->oti_ins_cache,
226 oti->oti_ins_cache_used * sizeof(*idc));
228 oti->oti_ins_cache = idc;
229 oti->oti_ins_cache_size = i;
232 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
235 idc->oic_lid.oii_ino = 0;
236 idc->oic_lid.oii_gen = 0;
243 * lookup mapping for the given fid in the cache, initialize a
244 * new one if not found. the initialization checks whether the
245 * object is local or remote. for local objects, OI is used to
246 * learn ino/generation. the function is used when the caller
247 * has no information about the object, e.g. at dt_insert().
249 static struct osd_idmap_cache *
250 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
251 const struct lu_fid *fid)
253 struct osd_idmap_cache *idc;
256 idc = osd_idc_find(env, osd, fid);
257 LASSERT(!IS_ERR(idc));
261 /* new mapping is needed */
262 idc = osd_idc_add(env, osd, fid);
267 rc = osd_remote_fid(env, osd, fid);
268 if (unlikely(rc < 0))
272 /* the object is local, lookup in OI */
273 /* XXX: probably cheaper to lookup in LU first? */
274 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
276 if (unlikely(rc < 0)) {
277 CERROR("can't lookup: rc = %d\n", rc);
281 /* the object is remote */
289 * lookup mapping for given FID and fill it from the given object.
290 * the object is lolcal by definition.
292 static int osd_idc_find_and_init(const struct lu_env *env,
293 struct osd_device *osd,
294 struct osd_object *obj)
296 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
297 struct osd_idmap_cache *idc;
299 idc = osd_idc_find(env, osd, fid);
300 LASSERT(!IS_ERR(idc));
302 if (obj->oo_inode == NULL)
304 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
305 LASSERT(idc->oic_lid.oii_ino == 0);
306 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
307 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
312 /* new mapping is needed */
313 idc = osd_idc_add(env, osd, fid);
317 if (obj->oo_inode != NULL) {
318 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
319 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
325 * OSD object methods.
329 * Concurrency: no concurrent access is possible that early in object
332 static struct lu_object *osd_object_alloc(const struct lu_env *env,
333 const struct lu_object_header *hdr,
336 struct osd_object *mo;
342 l = &mo->oo_dt.do_lu;
343 dt_object_init(&mo->oo_dt, NULL, d);
344 mo->oo_dt.do_ops = &osd_obj_ea_ops;
345 l->lo_ops = &osd_lu_obj_ops;
346 init_rwsem(&mo->oo_sem);
347 init_rwsem(&mo->oo_ext_idx_sem);
348 spin_lock_init(&mo->oo_guard);
349 INIT_LIST_HEAD(&mo->oo_xattr_list);
356 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
357 struct dentry *dentry, struct lustre_ost_attrs *loa)
361 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
362 (void *)loa, sizeof(*loa));
364 struct lustre_mdt_attrs *lma = &loa->loa_lma;
366 if (rc < sizeof(*lma))
370 lustre_loa_swab(loa, true);
371 /* Check LMA compatibility */
372 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
373 CWARN("%.16s: unsupported incompat LMA feature(s) %#x "
374 "for fid = "DFID", ino = %lu\n",
375 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
376 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
377 PFID(&lma->lma_self_fid), inode->i_ino);
380 } else if (rc == 0) {
388 * retrieve object from backend ext fs.
390 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
391 struct osd_inode_id *id)
394 struct inode *inode = NULL;
396 /* if we look for an inode withing a running
397 * transaction, then we risk to deadlock */
398 /* osd_dirent_check_repair() breaks this */
399 /*LASSERT(current->journal_info == NULL);*/
401 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
403 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
404 id->oii_ino, PTR_ERR(inode));
405 } else if (id->oii_gen != OSD_OII_NOGEN &&
406 inode->i_generation != id->oii_gen) {
407 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
408 "i_generation = %u\n",
409 id->oii_ino, id->oii_gen, inode->i_generation);
411 inode = ERR_PTR(-ESTALE);
412 } else if (inode->i_nlink == 0) {
413 /* due to parallel readdir and unlink,
414 * we can have dead inode here. */
415 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
417 inode = ERR_PTR(-ESTALE);
418 } else if (is_bad_inode(inode)) {
419 CWARN("%.16s: bad inode: ino = %u\n",
420 LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name, id->oii_ino);
422 inode = ERR_PTR(-ENOENT);
423 } else if ((rc = osd_attach_jinode(inode))) {
427 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
428 if (id->oii_gen == OSD_OII_NOGEN)
429 osd_id_gen(id, inode->i_ino, inode->i_generation);
431 /* Do not update file c/mtime in ldiskfs.
432 * NB: we don't have any lock to protect this because we don't
433 * have reference on osd_object now, but contention with
434 * another lookup + attr_set can't happen in the tiny window
435 * between if (...) and set S_NOCMTIME. */
436 if (!(inode->i_flags & S_NOCMTIME))
437 inode->i_flags |= S_NOCMTIME;
442 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
443 handle_t *handle, struct dentry *child,
444 struct inode *inode, struct htree_lock *hlock)
448 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
449 if (rc == -ENOBUFS || rc == -ENOSPC) {
450 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
451 struct inode *parent = child->d_parent->d_inode;
452 struct lu_fid *fid = NULL;
454 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
456 fid = &loa->loa_lma.lma_self_fid;
457 } else if (rc2 == -ENODATA) {
458 if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
459 fid = &info->oti_fid3;
460 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
461 } else if (!osd->od_is_ost && osd->od_index == 0) {
462 fid = &info->oti_fid3;
463 lu_igif_build(fid, parent->i_ino,
464 parent->i_generation);
469 CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
470 "maximum entry limit\n",
471 osd_name(osd), parent->i_ino, PFID(fid),
472 rc == -ENOSPC ? "has reached" : "is approaching");
474 CWARN("%s: directory (inode: %lu, FID: unknown) %s "
475 "maximum entry limit\n",
476 osd_name(osd), parent->i_ino,
477 rc == -ENOSPC ? "has reached" : "is approaching");
479 /* ignore such error now */
488 static struct inode *
489 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
490 struct osd_inode_id *id, struct lu_fid *fid)
492 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
496 inode = osd_iget(info, dev, id);
500 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
502 *fid = loa->loa_lma.lma_self_fid;
503 } else if (rc == -ENODATA) {
504 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
505 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
507 lu_igif_build(fid, inode->i_ino, inode->i_generation);
515 static struct inode *osd_iget_check(struct osd_thread_info *info,
516 struct osd_device *dev,
517 const struct lu_fid *fid,
518 struct osd_inode_id *id,
525 /* The cached OI mapping is trustable. If we cannot locate the inode
526 * via the cached OI mapping, then return the failure to the caller
527 * directly without further OI checking. */
530 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
533 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
536 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
537 PFID(fid), id->oii_ino, rc);
541 if (is_bad_inode(inode)) {
546 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
547 PFID(fid), id->oii_ino);
551 if (id->oii_gen != OSD_OII_NOGEN &&
552 inode->i_generation != id->oii_gen) {
557 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
558 "oii_gen = %u, i_generation = %u\n", PFID(fid),
559 id->oii_ino, id->oii_gen, inode->i_generation);
563 if (inode->i_nlink == 0) {
568 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
569 PFID(fid), id->oii_ino);
573 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
577 __u32 saved_ino = id->oii_ino;
578 __u32 saved_gen = id->oii_gen;
581 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
583 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
584 /* XXX: There are four possible cases:
586 * Backup/restore caused the OI invalid.
588 * Someone unlinked the object but NOT removed
589 * the OI mapping, such as mount target device
590 * as ldiskfs, and modify something directly.
592 * Someone just removed the object between the
593 * former oi_lookup and the iget. It is normal.
594 * 4. Other failure cases.
596 * Generally, when the device is mounted, it will
597 * auto check whether the system is restored from
598 * file-level backup or not. We trust such detect
599 * to distinguish the 1st case from the 2nd case:
600 * if the OI files are consistent but may contain
601 * stale OI mappings because of case 2, if iget()
602 * returns -ENOENT or -ESTALE, then it should be
605 /* If the OI mapping was in OI file before the
606 * osd_iget_check(), but now, it is disappear,
607 * then it must be removed by race. That is a
608 * normal race case. */
611 /* It is the OI scrub updated the OI mapping by race.
612 * The new OI mapping must be valid. */
613 if (saved_ino != id->oii_ino || saved_gen != id->oii_gen) {
619 if (dev->od_scrub.os_file.sf_flags & SF_INCONSISTENT)
620 /* It still can be the case 2, but we cannot
621 * distinguish it from the case 1. So return
622 * -EREMCHG to block current operation until
623 * OI scrub rebuilt the OI mappings. */
631 if (inode->i_generation == id->oii_gen)
636 if (id->oii_gen == OSD_OII_NOGEN)
637 osd_id_gen(id, inode->i_ino, inode->i_generation);
639 /* Do not update file c/mtime in ldiskfs.
640 * NB: we don't have any lock to protect this because we don't
641 * have reference on osd_object now, but contention with
642 * another lookup + attr_set can't happen in the tiny window
643 * between if (...) and set S_NOCMTIME. */
644 if (!(inode->i_flags & S_NOCMTIME))
645 inode->i_flags |= S_NOCMTIME;
662 * \retval +v: new filter_fid, does not contain self-fid
663 * \retval 0: filter_fid_old, contains self-fid
664 * \retval -v: other failure cases
666 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
667 struct dentry *dentry, struct lu_fid *fid)
669 struct filter_fid_old *ff = &info->oti_ff;
670 struct ost_id *ostid = &info->oti_ostid;
673 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
674 if (rc == sizeof(*ff)) {
676 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
677 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
679 * XXX: use 0 as the index for compatibility, the caller will
680 * handle index related issues when necessary.
683 ostid_to_fid(fid, ostid, 0);
684 } else if (rc == sizeof(struct filter_fid)) {
686 } else if (rc >= 0) {
693 static int osd_lma_self_repair(struct osd_thread_info *info,
694 struct osd_device *osd, struct inode *inode,
695 const struct lu_fid *fid, __u32 compat)
700 LASSERT(current->journal_info == NULL);
702 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
703 osd_dto_credits_noquota[DTO_XATTR_SET]);
706 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
711 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
713 CWARN("%s: cannot self repair the LMA: rc = %d\n",
715 ldiskfs_journal_stop(jh);
719 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
721 struct osd_thread_info *info = osd_oti_get(env);
722 struct osd_device *osd = osd_obj2dev(obj);
723 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
724 struct lustre_mdt_attrs *lma = &loa->loa_lma;
725 struct inode *inode = obj->oo_inode;
726 struct dentry *dentry = &info->oti_obj_dentry;
727 struct lu_fid *fid = NULL;
728 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
732 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
733 (void *)loa, sizeof(*loa));
734 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
735 fid = &lma->lma_self_fid;
736 rc = osd_get_idif(info, inode, dentry, fid);
737 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
738 /* For the given OST-object, if it has neither LMA nor
739 * FID in XATTR_NAME_FID, then the given FID (which is
740 * contained in the @obj, from client RPC for locating
741 * the OST-object) is trusted. We use it to generate
743 osd_lma_self_repair(info, osd, inode, rfid,
754 lustre_lma_swab(lma);
755 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
756 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
757 S_ISREG(inode->i_mode)))) {
758 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
759 "fid = "DFID", ino = %lu\n", osd_name(osd),
760 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
761 PFID(rfid), inode->i_ino);
764 fid = &lma->lma_self_fid;
768 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
769 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
770 struct ost_id *oi = &info->oti_ostid;
771 struct lu_fid *fid1 = &info->oti_fid3;
772 __u32 idx = fid_idif_ost_idx(rfid);
774 /* For old IDIF, the OST index is not part of the IDIF,
775 * Means that different OSTs may have the same IDIFs.
776 * Under such case, we need to make some compatible
777 * check to make sure to trigger OI scrub properly. */
778 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
779 /* Given @rfid is new, LMA is old. */
780 fid_to_ostid(fid, oi);
781 ostid_to_fid(fid1, oi, idx);
782 if (lu_fid_eq(fid1, rfid)) {
783 if (osd->od_index_in_idif)
784 osd_lma_self_repair(info, osd,
798 struct osd_check_lmv_buf {
799 #ifdef HAVE_DIR_CONTEXT
800 /* please keep it as first member */
801 struct dir_context ctx;
803 struct osd_thread_info *oclb_info;
804 struct osd_device *oclb_dev;
805 struct osd_idmap_cache *oclb_oic;
809 * It is called internally by ->readdir() to filter out the
810 * local slave object's FID of the striped directory.
812 * \retval 1 found the local slave's FID
813 * \retval 0 continue to check next item
814 * \retval -ve for failure
816 #ifdef HAVE_FILLDIR_USE_CTX
817 static int osd_stripe_dir_filldir(struct dir_context *buf,
819 static int osd_stripe_dir_filldir(void *buf,
821 const char *name, int namelen,
822 loff_t offset, __u64 ino, unsigned d_type)
824 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
825 struct osd_thread_info *oti = oclb->oclb_info;
826 struct lu_fid *fid = &oti->oti_fid3;
827 struct osd_inode_id *id = &oti->oti_id3;
828 struct osd_device *dev = oclb->oclb_dev;
829 struct osd_idmap_cache *oic = oclb->oclb_oic;
837 sscanf(name + 1, SFID, RFID(fid));
838 if (!fid_is_sane(fid))
841 if (osd_remote_fid(oti->oti_env, dev, fid))
844 osd_id_gen(id, ino, OSD_OII_NOGEN);
845 inode = osd_iget(oti, dev, id);
847 return PTR_ERR(inode);
850 osd_add_oi_cache(oti, dev, id, fid);
854 rc = osd_oii_insert(dev, oic, true);
856 return rc == 0 ? 1 : rc;
859 /* When lookup item under striped directory, we need to locate the master
860 * MDT-object of the striped directory firstly, then the client will send
861 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
862 * and the item's name. If the system is restored from MDT file level backup,
863 * then before the OI scrub completely built the OI files, the OI mappings of
864 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
865 * not a problem for the master MDT-object. Because when locate the master
866 * MDT-object, we will do name based lookup (for the striped directory itself)
867 * firstly, during such process we can setup the correct OI mapping for the
868 * master MDT-object. But it will be trouble for the slave MDT-object. Because
869 * the client will not trigger name based lookup on the MDT to locate the slave
870 * MDT-object before locating item under the striped directory, then when
871 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
872 * is invalid and does not know what the right OI mapping is, then the MDT has
873 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
874 * the OI file, related OI mapping is unknown yet, please try again later. And
875 * then client will re-try the RPC again and again until related OI mapping has
876 * been updated. That is quite inefficient.
878 * To resolve above trouble, we will handle it as the following two cases:
880 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
881 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
882 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
883 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
884 * directly. Please check osd_fid_lookup().
886 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
887 * Under such case, during lookup the master MDT-object, we will lookup the
888 * slave MDT-object via readdir against the master MDT-object, because the
889 * slave MDT-objects information are stored as sub-directories with the name
890 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
891 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
892 * the correct OI mapping for the slave MDT-object. */
893 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
894 struct inode *inode, struct osd_idmap_cache *oic)
896 struct lu_buf *buf = &oti->oti_big_buf;
897 struct dentry *dentry = &oti->oti_obj_dentry;
898 struct file *filp = &oti->oti_file;
899 const struct file_operations *fops;
900 struct lmv_mds_md_v1 *lmv1;
901 struct osd_check_lmv_buf oclb = {
902 #ifdef HAVE_DIR_CONTEXT
903 .ctx.actor = osd_stripe_dir_filldir,
913 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
916 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
918 lu_buf_realloc(buf, rc);
919 if (buf->lb_buf == NULL)
920 GOTO(out, rc = -ENOMEM);
926 if (unlikely(rc == 0 || rc == -ENODATA))
932 if (unlikely(buf->lb_buf == NULL)) {
933 lu_buf_realloc(buf, rc);
934 if (buf->lb_buf == NULL)
935 GOTO(out, rc = -ENOMEM);
941 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
945 dentry->d_inode = inode;
946 dentry->d_sb = inode->i_sb;
948 filp->f_path.dentry = dentry;
949 filp->f_mode = FMODE_64BITHASH;
950 filp->f_mapping = inode->i_mapping;
952 filp->private_data = NULL;
953 set_file_inode(filp, inode);
955 #ifdef HAVE_DIR_CONTEXT
956 oclb.ctx.pos = filp->f_pos;
957 rc = fops->iterate(filp, &oclb.ctx);
958 filp->f_pos = oclb.ctx.pos;
960 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
962 fops->release(inode, filp);
966 CDEBUG(D_LFSCK, "%.16s: fail to check LMV EA, inode = %lu/%u,"
968 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
969 inode->i_ino, inode->i_generation,
970 PFID(&oic->oic_fid), rc);
977 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
978 const struct lu_fid *fid,
979 const struct lu_object_conf *conf)
981 struct osd_thread_info *info;
982 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
983 struct osd_device *dev;
984 struct osd_idmap_cache *oic;
985 struct osd_inode_id *id;
986 struct inode *inode = NULL;
987 struct osd_scrub *scrub;
988 struct scrub_file *sf;
989 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
996 bool updated = false;
999 LINVRNT(osd_invariant(obj));
1000 LASSERT(obj->oo_inode == NULL);
1001 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1003 dev = osd_dev(ldev);
1004 scrub = &dev->od_scrub;
1005 sf = &scrub->os_file;
1006 info = osd_oti_get(env);
1008 oic = &info->oti_cache;
1010 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1013 /* For the object is created as locking anchor, or for the object to
1014 * be created on disk. No need to osd_oi_lookup() at here because FID
1015 * shouldn't never be re-used, if it's really a duplicate FID from
1016 * unexpected reason, we should be able to detect it later by calling
1017 * do_create->osd_oi_insert(). */
1018 if (conf != NULL && conf->loc_flags & LOC_F_NEW)
1019 GOTO(out, result = 0);
1021 /* Search order: 1. per-thread cache. */
1022 if (lu_fid_eq(fid, &oic->oic_fid) &&
1023 likely(oic->oic_dev == dev)) {
1029 if (!list_empty(&scrub->os_inconsistent_items)) {
1030 /* Search order: 2. OI scrub pending list. */
1031 result = osd_oii_lookup(dev, fid, id);
1036 /* The OI mapping in the OI file can be updated by the OI scrub
1037 * when we locate the inode via FID. So it may be not trustable. */
1040 /* Search order: 3. OI files. */
1041 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1042 if (result == -ENOENT) {
1043 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1044 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1045 !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1047 GOTO(out, result = 0);
1056 inode = osd_iget_check(info, dev, fid, id, trusted);
1057 if (IS_ERR(inode)) {
1058 result = PTR_ERR(inode);
1059 if (result == -ENOENT || result == -ESTALE)
1060 GOTO(out, result = 0);
1062 if (result == -EREMCHG) {
1065 /* We still have chance to get the valid inode: for the
1066 * object which is referenced by remote name entry, the
1067 * object on the local MDT will be linked under the dir
1068 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1070 * We do not know whether the object for the given FID
1071 * is referenced by some remote name entry or not, and
1072 * especially for DNE II, a multiple-linked object may
1073 * have many name entries reside on many MDTs.
1075 * To simplify the operation, OSD will not distinguish
1076 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1077 * only happened for the RPC from other MDT during the
1078 * OI scrub, or for the client side RPC with FID only,
1079 * such as FID to path, or from old connected client. */
1081 !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1082 rc1 = osd_lookup_in_remote_parent(info, dev,
1087 flags |= SS_AUTO_PARTIAL;
1088 flags &= ~SS_AUTO_FULL;
1093 if (thread_is_running(&scrub->os_thread)) {
1094 if (scrub->os_partial_scan &&
1095 !scrub->os_in_join) {
1098 if (inode != NULL && !IS_ERR(inode)) {
1101 osd_add_oi_cache(info, dev, id,
1103 osd_oii_insert(dev, oic, true);
1105 result = -EINPROGRESS;
1108 } else if (!dev->od_noscrub) {
1111 rc1 = osd_scrub_start(dev, flags);
1112 LCONSOLE_WARN("%.16s: trigger OI scrub by RPC "
1113 "for the "DFID" with flags 0x%x,"
1114 " rc = %d\n", osd_name(dev),
1115 PFID(fid), flags, rc1);
1116 if (rc1 == 0 || rc1 == -EALREADY) {
1117 if (inode != NULL && !IS_ERR(inode)) {
1120 osd_add_oi_cache(info, dev, id,
1122 osd_oii_insert(dev, oic, true);
1124 result = -EINPROGRESS;
1134 if (inode == NULL || IS_ERR(inode))
1136 } else if (remote) {
1140 obj->oo_inode = inode;
1141 LASSERT(obj->oo_inode->i_sb == osd_sb(dev));
1143 result = osd_check_lma(env, obj);
1147 LASSERTF(id->oii_ino == inode->i_ino &&
1148 id->oii_gen == inode->i_generation,
1149 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1150 PFID(fid), id->oii_ino, id->oii_gen,
1151 inode->i_ino, inode->i_generation);
1153 saved_ino = inode->i_ino;
1154 saved_gen = inode->i_generation;
1156 if (unlikely(result == -ENODATA)) {
1157 /* If the OI scrub updated the OI mapping by race, it
1158 * must be valid. Trust the inode that has no LMA EA. */
1162 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1164 /* The OI mapping is still there, the inode is still
1165 * valid. It is just becaues the inode has no LMA EA. */
1166 if (saved_ino == id->oii_ino &&
1167 saved_gen == id->oii_gen)
1170 /* It is the OI scrub updated the OI mapping by race.
1171 * The new OI mapping must be valid. */
1174 obj->oo_inode = NULL;
1180 /* "result == -ENOENT" means that the OI mappinghas been
1181 * removed by race, so the inode belongs to other object.
1183 * Others error can be returned directly. */
1184 if (result == -ENOENT) {
1193 obj->oo_inode = NULL;
1195 if (result != -EREMCHG)
1200 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1201 /* "result == -ENOENT" means the cached OI mapping has been removed
1202 * from the OI file by race, above inode belongs to other object. */
1203 if (result == -ENOENT) {
1206 GOTO(out, result = 0);
1212 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen)
1215 /* It is the OI scrub updated the OI mapping by race.
1216 * The new OI mapping must be valid. */
1222 obj->oo_compat_dot_created = 1;
1223 obj->oo_compat_dotdot_created = 1;
1225 if (S_ISDIR(inode->i_mode) &&
1226 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1227 osd_check_lmv(info, dev, inode, oic);
1229 result = osd_attach_jinode(inode);
1231 obj->oo_inode = NULL;
1237 GOTO(out, result = 0);
1239 LASSERT(obj->oo_hl_head == NULL);
1240 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1241 if (obj->oo_hl_head == NULL) {
1242 obj->oo_inode = NULL;
1244 GOTO(out, result = -ENOMEM);
1246 GOTO(out, result = 0);
1249 if (result != 0 && trusted)
1250 fid_zero(&oic->oic_fid);
1252 LINVRNT(osd_invariant(obj));
1257 * Concurrency: shouldn't matter.
1259 static void osd_object_init0(struct osd_object *obj)
1261 LASSERT(obj->oo_inode != NULL);
1262 obj->oo_dt.do_body_ops = &osd_body_ops;
1263 obj->oo_dt.do_lu.lo_header->loh_attr |=
1264 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1268 * Concurrency: no concurrent access is possible that early in object
1271 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1272 const struct lu_object_conf *conf)
1274 struct osd_object *obj = osd_obj(l);
1277 LINVRNT(osd_invariant(obj));
1279 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1280 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1281 l->lo_header->loh_attr |= LOHA_EXISTS;
1285 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1286 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1287 if (result == 0 && obj->oo_inode != NULL) {
1288 struct osd_thread_info *oti = osd_oti_get(env);
1289 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1291 osd_object_init0(obj);
1292 result = osd_get_lma(oti, obj->oo_inode,
1293 &oti->oti_obj_dentry, loa);
1295 /* Convert LMAI flags to lustre LMA flags
1296 * and cache it to oo_lma_flags */
1298 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1299 } else if (result == -ENODATA) {
1304 LINVRNT(osd_invariant(obj));
1308 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1309 * The left part is for value, binary mode. */
1310 struct osd_xattr_entry {
1311 struct list_head oxe_list;
1315 struct rcu_head oxe_rcu;
1319 static int osd_oxc_get(struct osd_object *obj, const char *name,
1322 struct osd_xattr_entry *tmp;
1323 struct osd_xattr_entry *oxe = NULL;
1324 size_t namelen = strlen(name);
1329 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1330 if (namelen == tmp->oxe_namelen &&
1331 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1338 GOTO(out, rc = -ENOENT);
1340 if (!oxe->oxe_exist)
1341 GOTO(out, rc = -ENODATA);
1344 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1347 if (buf->lb_buf == NULL)
1350 if (buf->lb_len < rc)
1351 GOTO(out, rc = -ERANGE);
1353 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1361 static void osd_oxc_free(struct rcu_head *head)
1363 struct osd_xattr_entry *oxe;
1365 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1366 OBD_FREE(oxe, oxe->oxe_len);
1369 static void osd_oxc_add(struct osd_object *obj, const char *name,
1370 const char *buf, int buflen)
1372 struct osd_xattr_entry *oxe;
1373 struct osd_xattr_entry *old = NULL;
1374 struct osd_xattr_entry *tmp;
1375 size_t namelen = strlen(name);
1376 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1378 OBD_ALLOC(oxe, len);
1382 INIT_LIST_HEAD(&oxe->oxe_list);
1384 oxe->oxe_namelen = namelen;
1385 memcpy(oxe->oxe_buf, name, namelen);
1387 LASSERT(buf != NULL);
1388 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1389 oxe->oxe_exist = true;
1391 oxe->oxe_exist = false;
1394 /* this should be rarely called, just remove old and add new */
1395 spin_lock(&obj->oo_guard);
1396 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1397 if (namelen == tmp->oxe_namelen &&
1398 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1404 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1405 call_rcu(&old->oxe_rcu, osd_oxc_free);
1407 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1409 spin_unlock(&obj->oo_guard);
1412 static void osd_oxc_del(struct osd_object *obj, const char *name)
1414 struct osd_xattr_entry *oxe;
1415 size_t namelen = strlen(name);
1417 spin_lock(&obj->oo_guard);
1418 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1419 if (namelen == oxe->oxe_namelen &&
1420 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1421 list_del_rcu(&oxe->oxe_list);
1422 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1426 spin_unlock(&obj->oo_guard);
1429 static void osd_oxc_fini(struct osd_object *obj)
1431 struct osd_xattr_entry *oxe, *next;
1433 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1434 list_del(&oxe->oxe_list);
1435 OBD_FREE(oxe, oxe->oxe_len);
1440 * Concurrency: no concurrent access is possible that late in object
1443 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1445 struct osd_object *obj = osd_obj(l);
1447 LINVRNT(osd_invariant(obj));
1450 dt_object_fini(&obj->oo_dt);
1451 if (obj->oo_hl_head != NULL)
1452 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1457 * Concurrency: no concurrent access is possible that late in object
1460 static void osd_index_fini(struct osd_object *o)
1462 struct iam_container *bag;
1464 if (o->oo_dir != NULL) {
1465 bag = &o->oo_dir->od_container;
1466 if (o->oo_inode != NULL) {
1467 if (bag->ic_object == o->oo_inode)
1468 iam_container_fini(bag);
1470 OBD_FREE_PTR(o->oo_dir);
1476 * Concurrency: no concurrent access is possible that late in object
1477 * life-cycle (for all existing callers, that is. New callers have to provide
1478 * their own locking.)
1480 static int osd_inode_unlinked(const struct inode *inode)
1482 return inode->i_nlink == 0;
1486 OSD_TXN_OI_DELETE_CREDITS = 20,
1487 OSD_TXN_INODE_DELETE_CREDITS = 20
1494 #if OSD_THANDLE_STATS
1496 * Set time when the handle is allocated
1498 static void osd_th_alloced(struct osd_thandle *oth)
1500 oth->oth_alloced = ktime_get();
1504 * Set time when the handle started
1506 static void osd_th_started(struct osd_thandle *oth)
1508 oth->oth_started = ktime_get();
1512 * Check whether the we deal with this handle for too long.
1514 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1515 ktime_t alloced, ktime_t started,
1518 ktime_t now = ktime_get();
1520 LASSERT(dev != NULL);
1522 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1523 ktime_us_delta(started, alloced));
1524 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1525 ktime_us_delta(closed, started));
1526 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1527 ktime_us_delta(now, closed));
1529 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1530 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1531 oth, now, alloced, started, closed);
1532 libcfs_debug_dumpstack(NULL);
1536 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1538 ktime_t __closed = ktime_get(); \
1539 ktime_t __alloced = oth->oth_alloced; \
1540 ktime_t __started = oth->oth_started; \
1543 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1546 #else /* OSD_THANDLE_STATS */
1548 #define osd_th_alloced(h) do {} while(0)
1549 #define osd_th_started(h) do {} while(0)
1550 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1552 #endif /* OSD_THANDLE_STATS */
1555 * Concurrency: doesn't access mutable data.
1557 static int osd_param_is_not_sane(const struct osd_device *dev,
1558 const struct thandle *th)
1560 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1562 return oh->ot_credits > osd_transaction_size(dev);
1566 * Concurrency: shouldn't matter.
1568 static void osd_trans_commit_cb(struct super_block *sb,
1569 struct ldiskfs_journal_cb_entry *jcb, int error)
1571 struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1572 struct thandle *th = &oh->ot_super;
1573 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1574 struct dt_txn_commit_cb *dcb, *tmp;
1576 LASSERT(oh->ot_handle == NULL);
1579 CERROR("transaction @0x%p commit error: %d\n", th, error);
1581 dt_txn_hook_commit(th);
1583 /* call per-transaction callbacks if any */
1584 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1586 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1587 "commit callback entry: magic=%x name='%s'\n",
1588 dcb->dcb_magic, dcb->dcb_name);
1589 list_del_init(&dcb->dcb_linkage);
1590 dcb->dcb_func(NULL, th, dcb, error);
1593 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1597 lu_context_exit(&th->th_ctx);
1598 lu_context_fini(&th->th_ctx);
1602 #ifndef HAVE_SB_START_WRITE
1603 # define sb_start_write(sb) do {} while (0)
1604 # define sb_end_write(sb) do {} while (0)
1607 static struct thandle *osd_trans_create(const struct lu_env *env,
1608 struct dt_device *d)
1610 struct osd_thread_info *oti = osd_oti_get(env);
1611 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1612 struct osd_thandle *oh;
1617 CERROR("%s: someone try to start transaction under "
1618 "readonly mode, should be disabled.\n",
1619 osd_name(osd_dt_dev(d)));
1621 RETURN(ERR_PTR(-EROFS));
1624 /* on pending IO in this thread should left from prev. request */
1625 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1627 sb_start_write(osd_sb(osd_dt_dev(d)));
1629 OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1631 oh->ot_quota_trans = &oti->oti_quota_trans;
1632 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1636 th->th_tags = LCT_TX_HANDLE;
1638 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1639 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1642 memset(oti->oti_declare_ops, 0,
1643 sizeof(oti->oti_declare_ops));
1644 memset(oti->oti_declare_ops_cred, 0,
1645 sizeof(oti->oti_declare_ops_cred));
1646 memset(oti->oti_declare_ops_used, 0,
1647 sizeof(oti->oti_declare_ops_used));
1649 sb_end_write(osd_sb(osd_dt_dev(d)));
1650 th = ERR_PTR(-ENOMEM);
1655 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1657 struct osd_thread_info *oti = osd_oti_get(env);
1658 struct osd_thandle *oh;
1660 oh = container_of0(th, struct osd_thandle, ot_super);
1661 LASSERT(oh != NULL);
1663 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1664 oti->oti_declare_ops[OSD_OT_CREATE],
1665 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1666 oti->oti_declare_ops_used[OSD_OT_CREATE],
1667 oti->oti_declare_ops[OSD_OT_DESTROY],
1668 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1669 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1670 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1671 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1672 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1673 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1674 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1675 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1676 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1677 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1678 oti->oti_declare_ops[OSD_OT_WRITE],
1679 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1680 oti->oti_declare_ops_used[OSD_OT_WRITE],
1681 oti->oti_declare_ops[OSD_OT_PUNCH],
1682 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1683 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1684 oti->oti_declare_ops[OSD_OT_QUOTA],
1685 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1686 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1687 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1688 oti->oti_declare_ops[OSD_OT_INSERT],
1689 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1690 oti->oti_declare_ops_used[OSD_OT_INSERT],
1691 oti->oti_declare_ops[OSD_OT_DELETE],
1692 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1693 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1694 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1695 oti->oti_declare_ops[OSD_OT_REF_ADD],
1696 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1697 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1698 oti->oti_declare_ops[OSD_OT_REF_DEL],
1699 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1700 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1704 * Concurrency: shouldn't matter.
1706 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1709 struct osd_thread_info *oti = osd_oti_get(env);
1710 struct osd_device *dev = osd_dt_dev(d);
1712 struct osd_thandle *oh;
1717 LASSERT(current->journal_info == NULL);
1719 oh = container_of0(th, struct osd_thandle, ot_super);
1720 LASSERT(oh != NULL);
1721 LASSERT(oh->ot_handle == NULL);
1723 rc = dt_txn_hook_start(env, d, th);
1727 if (unlikely(osd_param_is_not_sane(dev, th))) {
1728 static unsigned long last_printed;
1729 static int last_credits;
1731 /* don't make noise on a tiny testing systems
1732 * actual credits misuse will be caught anyway */
1733 if (last_credits != oh->ot_credits &&
1734 time_after(jiffies, last_printed +
1735 msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1736 osd_transaction_size(dev) > 512) {
1737 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1738 oh->ot_credits, osd_transaction_size(dev));
1739 osd_trans_dump_creds(env, th);
1740 libcfs_debug_dumpstack(NULL);
1741 last_credits = oh->ot_credits;
1742 last_printed = jiffies;
1744 /* XXX Limit the credits to 'max_transaction_buffers', and
1745 * let the underlying filesystem to catch the error if
1746 * we really need so many credits.
1748 * This should be removed when we can calculate the
1749 * credits precisely. */
1750 oh->ot_credits = osd_transaction_size(dev);
1751 } else if (ldiskfs_track_declares_assert != 0) {
1752 /* reserve few credits to prevent an assertion in JBD
1753 * our debugging mechanism will be able to detected
1754 * overuse. this can help to debug single-update
1756 oh->ot_credits += 10;
1757 if (unlikely(osd_param_is_not_sane(dev, th)))
1758 oh->ot_credits = osd_transaction_size(dev);
1762 * XXX temporary stuff. Some abstraction layer should
1765 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1769 LASSERT(oti->oti_txns == 0);
1770 lu_context_init(&th->th_ctx, th->th_tags);
1771 lu_context_enter(&th->th_ctx);
1773 lu_device_get(&d->dd_lu_dev);
1774 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1785 static int osd_seq_exists(const struct lu_env *env,
1786 struct osd_device *osd, u64 seq)
1788 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1789 struct seq_server_site *ss = osd_seq_site(osd);
1793 LASSERT(ss != NULL);
1794 LASSERT(ss->ss_server_fld != NULL);
1796 rc = osd_fld_lookup(env, osd, seq, range);
1799 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1800 osd_name(osd), seq, rc);
1804 RETURN(ss->ss_node_id == range->lsr_index);
1807 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1809 struct dt_txn_commit_cb *dcb;
1810 struct dt_txn_commit_cb *tmp;
1812 /* call per-transaction stop callbacks if any */
1813 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1815 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1816 "commit callback entry: magic=%x name='%s'\n",
1817 dcb->dcb_magic, dcb->dcb_name);
1818 list_del_init(&dcb->dcb_linkage);
1819 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1824 * Concurrency: shouldn't matter.
1826 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1829 int rc = 0, remove_agents = 0;
1830 struct osd_thandle *oh;
1831 struct osd_thread_info *oti = osd_oti_get(env);
1832 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1833 struct osd_device *osd = osd_dt_dev(th->th_dev);
1834 struct qsd_instance *qsd = osd->od_quota_slave;
1835 struct lquota_trans *qtrans;
1838 oh = container_of0(th, struct osd_thandle, ot_super);
1840 /* reset OI cache for safety */
1841 oti->oti_ins_cache_used = 0;
1843 remove_agents = oh->ot_remove_agents;
1845 qtrans = oh->ot_quota_trans;
1846 oh->ot_quota_trans = NULL;
1848 if (oh->ot_handle != NULL) {
1850 handle_t *hdl = oh->ot_handle;
1853 * add commit callback
1854 * notice we don't do this in osd_trans_start()
1855 * as underlying transaction can change during truncate
1857 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1860 LASSERT(oti->oti_txns == 1);
1863 rc = dt_txn_hook_stop(env, th);
1865 CERROR("%s: failed in transaction hook: rc = %d\n",
1868 osd_trans_stop_cb(oh, rc);
1869 /* hook functions might modify th_sync */
1870 hdl->h_sync = th->th_sync;
1872 oh->ot_handle = NULL;
1873 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1875 CERROR("%s: failed to stop transaction: rc = %d\n",
1876 osd_name(osd), rc2);
1880 osd_trans_stop_cb(oh, th->th_result);
1884 /* inform the quota slave device that the transaction is stopping */
1885 qsd_op_end(env, qsd, qtrans);
1887 /* as we want IO to journal and data IO be concurrent, we don't block
1888 * awaiting data IO completion in osd_do_bio(), instead we wait here
1889 * once transaction is submitted to the journal. all reqular requests
1890 * don't do direct IO (except read/write), thus this wait_event becomes
1893 * IMPORTANT: we have to wait till any IO submited by the thread is
1894 * completed otherwise iobuf may be corrupted by different request
1896 wait_event(iobuf->dr_wait,
1897 atomic_read(&iobuf->dr_numreqs) == 0);
1898 osd_fini_iobuf(osd, iobuf);
1900 rc = iobuf->dr_error;
1902 if (unlikely(remove_agents != 0))
1903 osd_process_scheduled_agent_removals(env, osd);
1905 sb_end_write(osd_sb(osd));
1910 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1912 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1915 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1916 LASSERT(&dcb->dcb_func != NULL);
1917 if (dcb->dcb_flags & DCB_TRANS_STOP)
1918 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1920 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1926 * Called just before object is freed. Releases all resources except for
1927 * object itself (that is released by osd_object_free()).
1929 * Concurrency: no concurrent access is possible that late in object
1932 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1934 struct osd_object *obj = osd_obj(l);
1935 struct inode *inode = obj->oo_inode;
1937 LINVRNT(osd_invariant(obj));
1940 * If object is unlinked remove fid->ino mapping from object index.
1943 osd_index_fini(obj);
1944 if (inode != NULL) {
1945 struct qsd_instance *qsd = osd_obj2dev(obj)->od_quota_slave;
1946 qid_t uid = i_uid_read(inode);
1947 qid_t gid = i_gid_read(inode);
1950 obj->oo_inode = NULL;
1953 struct osd_thread_info *info = osd_oti_get(env);
1954 struct lquota_id_info *qi = &info->oti_qi;
1956 /* Release granted quota to master if necessary */
1957 qi->lqi_id.qid_uid = uid;
1958 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1960 qi->lqi_id.qid_uid = gid;
1961 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1963 qi->lqi_id.qid_uid = i_projid_read(inode);
1964 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1970 * Concurrency: ->loo_object_release() is called under site spin-lock.
1972 static void osd_object_release(const struct lu_env *env,
1973 struct lu_object *l)
1975 struct osd_object *o = osd_obj(l);
1976 /* nobody should be releasing a non-destroyed object with nlink=0
1977 * the API allows this, but ldiskfs doesn't like and then report
1978 * this inode as deleted */
1979 if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1984 * Concurrency: shouldn't matter.
1986 static int osd_object_print(const struct lu_env *env, void *cookie,
1987 lu_printer_t p, const struct lu_object *l)
1989 struct osd_object *o = osd_obj(l);
1990 struct iam_descr *d;
1992 if (o->oo_dir != NULL)
1993 d = o->oo_dir->od_container.ic_descr;
1996 return (*p)(env, cookie,
1997 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1999 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2000 o->oo_inode ? o->oo_inode->i_generation : 0,
2001 d ? d->id_ops->id_name : "plain");
2005 * Concurrency: shouldn't matter.
2007 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2008 struct obd_statfs *sfs)
2010 struct osd_device *osd = osd_dt_dev(d);
2011 struct super_block *sb = osd_sb(osd);
2012 struct kstatfs *ksfs;
2016 if (unlikely(osd->od_mnt == NULL))
2017 return -EINPROGRESS;
2019 /* osd_lproc.c call this without env, allocate ksfs for that case */
2020 if (unlikely(env == NULL)) {
2021 OBD_ALLOC_PTR(ksfs);
2025 ksfs = &osd_oti_get(env)->oti_ksfs;
2028 result = sb->s_op->statfs(sb->s_root, ksfs);
2032 statfs_pack(sfs, ksfs);
2033 if (unlikely(sb->s_flags & MS_RDONLY))
2034 sfs->os_state |= OS_STATE_READONLY;
2035 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2036 LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2037 sfs->os_maxbytes = sb->s_maxbytes;
2039 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2042 * Reserve some space so to avoid fragmenting the filesystem too much.
2043 * Fragmentation not only impacts performance, but can also increase
2044 * metadata overhead significantly, causing grant calculation to be
2047 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2049 CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2050 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2051 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2052 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2054 sfs->os_blocks -= reserved;
2055 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2056 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2059 if (unlikely(env == NULL))
2065 * Estimate space needed for file creations. We assume the largest filename
2066 * which is 2^64 - 1, hence a filename of 20 chars.
2067 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2069 #ifdef __LDISKFS_DIR_REC_LEN
2070 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2072 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2076 * Concurrency: doesn't access mutable data.
2078 static void osd_conf_get(const struct lu_env *env,
2079 const struct dt_device *dev,
2080 struct dt_device_param *param)
2082 struct super_block *sb = osd_sb(osd_dt_dev(dev));
2086 * XXX should be taken from not-yet-existing fs abstraction layer.
2088 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2089 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2090 param->ddp_symlink_max = sb->s_blocksize;
2091 param->ddp_mount_type = LDD_MT_LDISKFS;
2092 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2093 param->ddp_maxbytes = sb->s_maxbytes;
2095 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2096 /* inode are statically allocated, so per-inode space consumption
2097 * is the space consumed by the directory entry */
2098 param->ddp_inodespace = PER_OBJ_USAGE;
2099 /* EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2100 * = 128MB) which is unlikely to be hit in real life. Report a smaller
2101 * maximum length to not under count the actual number of extents
2102 * needed for writing a file. */
2103 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2104 /* worst-case extent insertion metadata overhead */
2105 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2106 param->ddp_mntopts = 0;
2107 if (test_opt(sb, XATTR_USER))
2108 param->ddp_mntopts |= MNTOPT_USERXATTR;
2109 if (test_opt(sb, POSIX_ACL))
2110 param->ddp_mntopts |= MNTOPT_ACL;
2112 /* LOD might calculate the max stripe count based on max_ea_size,
2113 * so we need take account in the overhead as well,
2114 * xattr_header + magic + xattr_entry_head */
2115 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2116 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2118 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2119 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2120 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2124 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2128 * Concurrency: shouldn't matter.
2130 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2134 CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
2136 rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2138 CDEBUG(D_CACHE, "synced OSD %s: rc = %d\n",
2139 LUSTRE_OSD_LDISKFS_NAME, rc);
2145 * Start commit for OSD device.
2147 * An implementation of dt_commit_async method for OSD device.
2148 * Asychronously starts underlayng fs sync and thereby a transaction
2151 * \param env environment
2152 * \param d dt device
2154 * \see dt_device_operations
2156 static int osd_commit_async(const struct lu_env *env,
2157 struct dt_device *d)
2159 struct super_block *s = osd_sb(osd_dt_dev(d));
2162 CDEBUG(D_HA, "async commit OSD %s\n", LUSTRE_OSD_LDISKFS_NAME);
2163 RETURN(s->s_op->sync_fs(s, 0));
2167 * Concurrency: shouldn't matter.
2170 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2172 struct super_block *sb = osd_sb(osd_dt_dev(d));
2173 struct block_device *dev = sb->s_bdev;
2174 #ifdef HAVE_DEV_SET_RDONLY
2175 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2178 int rc = -EOPNOTSUPP;
2182 #ifdef HAVE_DEV_SET_RDONLY
2183 CERROR("*** setting %s read-only ***\n", osd_dt_dev(d)->od_svname);
2185 if (sb->s_op->freeze_fs) {
2186 rc = sb->s_op->freeze_fs(sb);
2191 if (jdev && (jdev != dev)) {
2192 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2194 dev_set_rdonly(jdev);
2196 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2197 dev_set_rdonly(dev);
2199 if (sb->s_op->unfreeze_fs)
2200 sb->s_op->unfreeze_fs(sb);
2205 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2206 osd_dt_dev(d)->od_svname, (long)dev, rc);
2212 * Note: we do not count into QUOTA here.
2213 * If we mount with --data_journal we may need more.
2215 const int osd_dto_credits_noquota[DTO_NR] = {
2218 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2219 * SINGLEDATA_TRANS_BLOCKS(8)
2220 * XXX Note: maybe iam need more, since iam have more level than
2223 [DTO_INDEX_INSERT] = 16,
2226 * just modify a single entry, probably merge few within a block
2228 [DTO_INDEX_DELETE] = 1,
2232 [DTO_INDEX_UPDATE] = 16,
2234 * 4(inode, inode bits, groups, GDT)
2235 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2237 [DTO_OBJECT_CREATE] = 4,
2239 * 4(inode, inode bits, groups, GDT)
2240 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2242 [DTO_OBJECT_DELETE] = 4,
2244 * Attr set credits (inode)
2246 [DTO_ATTR_SET_BASE] = 1,
2248 * Xattr set. The same as xattr of EXT3.
2249 * DATA_TRANS_BLOCKS(14)
2250 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2251 * are also counted in. Do not know why?
2253 [DTO_XATTR_SET] = 14,
2255 * credits for inode change during write.
2257 [DTO_WRITE_BASE] = 3,
2259 * credits for single block write.
2261 [DTO_WRITE_BLOCK] = 14,
2263 * Attr set credits for chown.
2264 * This is extra credits for setattr, and it is null without quota
2266 [DTO_ATTR_SET_CHOWN] = 0
2269 static const struct dt_device_operations osd_dt_ops = {
2270 .dt_root_get = osd_root_get,
2271 .dt_statfs = osd_statfs,
2272 .dt_trans_create = osd_trans_create,
2273 .dt_trans_start = osd_trans_start,
2274 .dt_trans_stop = osd_trans_stop,
2275 .dt_trans_cb_add = osd_trans_cb_add,
2276 .dt_conf_get = osd_conf_get,
2277 .dt_sync = osd_sync,
2279 .dt_commit_async = osd_commit_async,
2282 static void osd_object_read_lock(const struct lu_env *env,
2283 struct dt_object *dt, unsigned role)
2285 struct osd_object *obj = osd_dt_obj(dt);
2286 struct osd_thread_info *oti = osd_oti_get(env);
2288 LINVRNT(osd_invariant(obj));
2290 LASSERT(obj->oo_owner != env);
2291 down_read_nested(&obj->oo_sem, role);
2293 LASSERT(obj->oo_owner == NULL);
2297 static void osd_object_write_lock(const struct lu_env *env,
2298 struct dt_object *dt, unsigned role)
2300 struct osd_object *obj = osd_dt_obj(dt);
2301 struct osd_thread_info *oti = osd_oti_get(env);
2303 LINVRNT(osd_invariant(obj));
2305 LASSERT(obj->oo_owner != env);
2306 down_write_nested(&obj->oo_sem, role);
2308 LASSERT(obj->oo_owner == NULL);
2309 obj->oo_owner = env;
2313 static void osd_object_read_unlock(const struct lu_env *env,
2314 struct dt_object *dt)
2316 struct osd_object *obj = osd_dt_obj(dt);
2317 struct osd_thread_info *oti = osd_oti_get(env);
2319 LINVRNT(osd_invariant(obj));
2321 LASSERT(oti->oti_r_locks > 0);
2323 up_read(&obj->oo_sem);
2326 static void osd_object_write_unlock(const struct lu_env *env,
2327 struct dt_object *dt)
2329 struct osd_object *obj = osd_dt_obj(dt);
2330 struct osd_thread_info *oti = osd_oti_get(env);
2332 LINVRNT(osd_invariant(obj));
2334 LASSERT(obj->oo_owner == env);
2335 LASSERT(oti->oti_w_locks > 0);
2337 obj->oo_owner = NULL;
2338 up_write(&obj->oo_sem);
2341 static int osd_object_write_locked(const struct lu_env *env,
2342 struct dt_object *dt)
2344 struct osd_object *obj = osd_dt_obj(dt);
2346 LINVRNT(osd_invariant(obj));
2348 return obj->oo_owner == env;
2351 static struct timespec *osd_inode_time(const struct lu_env *env,
2352 struct inode *inode, __u64 seconds)
2354 struct osd_thread_info *oti = osd_oti_get(env);
2355 struct timespec *t = &oti->oti_time;
2357 t->tv_sec = seconds;
2359 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2363 static void osd_inode_getattr(const struct lu_env *env,
2364 struct inode *inode, struct lu_attr *attr)
2366 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2367 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2368 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2369 LA_BLKSIZE | LA_TYPE;
2371 attr->la_atime = LTIME_S(inode->i_atime);
2372 attr->la_mtime = LTIME_S(inode->i_mtime);
2373 attr->la_ctime = LTIME_S(inode->i_ctime);
2374 attr->la_mode = inode->i_mode;
2375 attr->la_size = i_size_read(inode);
2376 attr->la_blocks = inode->i_blocks;
2377 attr->la_uid = i_uid_read(inode);
2378 attr->la_gid = i_gid_read(inode);
2379 attr->la_projid = i_projid_read(inode);
2380 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2381 attr->la_nlink = inode->i_nlink;
2382 attr->la_rdev = inode->i_rdev;
2383 attr->la_blksize = 1 << inode->i_blkbits;
2384 attr->la_blkbits = inode->i_blkbits;
2386 * Ext4 did not transfer inherit flags from raw inode
2387 * to inode flags, and ext4 internally test raw inode
2388 * @i_flags directly. Instead of patching ext4, we do it here.
2390 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2391 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2394 static int osd_attr_get(const struct lu_env *env,
2395 struct dt_object *dt,
2396 struct lu_attr *attr)
2398 struct osd_object *obj = osd_dt_obj(dt);
2400 if (unlikely(!dt_object_exists(dt)))
2402 if (unlikely(obj->oo_destroyed))
2405 LASSERT(!dt_object_remote(dt));
2406 LINVRNT(osd_invariant(obj));
2408 spin_lock(&obj->oo_guard);
2409 osd_inode_getattr(env, obj->oo_inode, attr);
2410 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2411 attr->la_flags |= LUSTRE_ORPHAN_FL;
2412 spin_unlock(&obj->oo_guard);
2417 static int osd_declare_attr_qid(const struct lu_env *env,
2418 struct osd_object *obj,
2419 struct osd_thandle *oh, long long bspace,
2420 qid_t old_id, qid_t new_id, bool enforce,
2424 struct osd_thread_info *info = osd_oti_get(env);
2425 struct lquota_id_info *qi = &info->oti_qi;
2427 qi->lqi_type = type;
2428 /* inode accounting */
2429 qi->lqi_is_blk = false;
2431 /* one more inode for the new id ... */
2432 qi->lqi_id.qid_uid = new_id;
2434 /* Reserve credits for the new id */
2435 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2436 if (rc == -EDQUOT || rc == -EINPROGRESS)
2441 /* and one less inode for the current id */
2442 qi->lqi_id.qid_uid = old_id;
2444 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2445 if (rc == -EDQUOT || rc == -EINPROGRESS)
2450 /* block accounting */
2451 qi->lqi_is_blk = true;
2453 /* more blocks for the new uid ... */
2454 qi->lqi_id.qid_uid = new_id;
2455 qi->lqi_space = bspace;
2457 * Credits for the new uid has been reserved, re-use "obj"
2458 * to save credit reservation.
2460 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2461 if (rc == -EDQUOT || rc == -EINPROGRESS)
2466 /* and finally less blocks for the current uid */
2467 qi->lqi_id.qid_uid = old_id;
2468 qi->lqi_space = -bspace;
2469 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2470 if (rc == -EDQUOT || rc == -EINPROGRESS)
2476 static int osd_declare_attr_set(const struct lu_env *env,
2477 struct dt_object *dt,
2478 const struct lu_attr *attr,
2479 struct thandle *handle)
2481 struct osd_thandle *oh;
2482 struct osd_object *obj;
2490 LASSERT(dt != NULL);
2491 LASSERT(handle != NULL);
2493 obj = osd_dt_obj(dt);
2494 LASSERT(osd_invariant(obj));
2496 oh = container_of0(handle, struct osd_thandle, ot_super);
2497 LASSERT(oh->ot_handle == NULL);
2499 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2500 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2502 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2503 osd_dto_credits_noquota[DTO_XATTR_SET]);
2505 if (attr == NULL || obj->oo_inode == NULL)
2508 bspace = obj->oo_inode->i_blocks;
2509 bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2510 bspace = toqb(bspace);
2512 /* Changing ownership is always preformed by super user, it should not
2515 * We still need to call the osd_declare_qid() to calculate the journal
2516 * credits for updating quota accounting files and to trigger quota
2517 * space adjustment once the operation is completed.*/
2518 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2520 uid = i_uid_read(obj->oo_inode);
2521 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2522 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2523 attr->la_uid, enforce, USRQUOTA);
2527 gid = i_gid_read(obj->oo_inode);
2528 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2529 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2530 i_gid_read(obj->oo_inode), attr->la_gid,
2536 #ifdef HAVE_PROJECT_QUOTA
2537 if (attr->la_valid & LA_PROJID) {
2538 __u32 projid = i_projid_read(obj->oo_inode);
2539 enforce = (attr->la_valid & LA_PROJID) &&
2540 (attr->la_projid != projid);
2541 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2542 (qid_t)projid, (qid_t)attr->la_projid,
2551 static int osd_inode_setattr(const struct lu_env *env,
2552 struct inode *inode, const struct lu_attr *attr)
2554 __u64 bits = attr->la_valid;
2556 /* Only allow set size for regular file */
2557 if (!S_ISREG(inode->i_mode))
2558 bits &= ~(LA_SIZE | LA_BLOCKS);
2563 if (bits & LA_ATIME)
2564 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2565 if (bits & LA_CTIME)
2566 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2567 if (bits & LA_MTIME)
2568 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2569 if (bits & LA_SIZE) {
2570 spin_lock(&inode->i_lock);
2571 LDISKFS_I(inode)->i_disksize = attr->la_size;
2572 i_size_write(inode, attr->la_size);
2573 spin_unlock(&inode->i_lock);
2576 /* OSD should not change "i_blocks" which is used by quota.
2577 * "i_blocks" should be changed by ldiskfs only. */
2579 inode->i_mode = (inode->i_mode & S_IFMT) |
2580 (attr->la_mode & ~S_IFMT);
2582 i_uid_write(inode, attr->la_uid);
2584 i_gid_write(inode, attr->la_gid);
2585 if (bits & LA_PROJID)
2586 i_projid_write(inode, attr->la_projid);
2587 if (bits & LA_NLINK)
2588 set_nlink(inode, attr->la_nlink);
2590 inode->i_rdev = attr->la_rdev;
2592 if (bits & LA_FLAGS) {
2593 /* always keep S_NOCMTIME */
2594 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2597 * Ext4 did not transfer inherit flags from
2598 * @inode->i_flags to raw inode i_flags when writing
2599 * flags, we do it explictly here.
2601 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2602 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2607 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2611 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2612 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2615 ll_vfs_dq_init(inode);
2617 if (attr->la_valid & LA_UID)
2618 iattr.ia_valid |= ATTR_UID;
2619 if (attr->la_valid & LA_GID)
2620 iattr.ia_valid |= ATTR_GID;
2621 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2622 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2624 rc = ll_vfs_dq_transfer(inode, &iattr);
2626 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2627 "enforcement enabled on the ldiskfs "
2628 "filesystem?\n", inode->i_sb->s_id, rc);
2633 #ifdef HAVE_PROJECT_QUOTA
2634 /* Handle project id transfer here properly */
2635 if (attr->la_valid & LA_PROJID && attr->la_projid !=
2636 i_projid_read(inode)) {
2637 rc = __ldiskfs_ioctl_setproject(inode, attr->la_projid);
2639 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2640 "enforcement enabled on the ldiskfs "
2641 "filesystem?\n", inode->i_sb->s_id, rc);
2649 static int osd_attr_set(const struct lu_env *env,
2650 struct dt_object *dt,
2651 const struct lu_attr *attr,
2652 struct thandle *handle)
2654 struct osd_object *obj = osd_dt_obj(dt);
2655 struct inode *inode;
2658 if (!dt_object_exists(dt))
2661 LASSERT(handle != NULL);
2662 LASSERT(!dt_object_remote(dt));
2663 LASSERT(osd_invariant(obj));
2665 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2667 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2668 struct osd_thread_info *oti = osd_oti_get(env);
2669 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2670 struct lu_fid *fid1 = &oti->oti_fid;
2671 struct osd_inode_id *id = &oti->oti_id;
2672 struct iam_path_descr *ipd;
2673 struct iam_container *bag;
2674 struct osd_thandle *oh;
2677 fid_cpu_to_be(fid1, fid0);
2678 memset(id, 1, sizeof(*id));
2679 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2680 fid0)->oi_dir.od_container;
2681 ipd = osd_idx_ipd_get(env, bag);
2682 if (unlikely(ipd == NULL))
2685 oh = container_of0(handle, struct osd_thandle, ot_super);
2686 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2687 (const struct iam_rec *)id, ipd);
2688 osd_ipd_put(env, bag, ipd);
2689 return(rc > 0 ? 0 : rc);
2692 inode = obj->oo_inode;
2694 rc = osd_quota_transfer(inode, attr);
2698 spin_lock(&obj->oo_guard);
2699 rc = osd_inode_setattr(env, inode, attr);
2700 spin_unlock(&obj->oo_guard);
2704 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2706 if (!(attr->la_valid & LA_FLAGS))
2709 /* Let's check if there are extra flags need to be set into LMA */
2710 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2711 struct osd_thread_info *info = osd_oti_get(env);
2712 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2714 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2715 &info->oti_ost_attrs);
2719 lma->lma_incompat |=
2720 lustre_to_lma_flags(attr->la_flags);
2721 lustre_lma_swab(lma);
2722 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2723 lma, sizeof(*lma), XATTR_REPLACE);
2725 struct osd_device *osd = osd_obj2dev(obj);
2727 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2728 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2729 lma->lma_incompat, rc);
2732 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2734 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2737 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2742 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2743 struct osd_object *obj,
2744 const char *name, const int namelen)
2746 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2749 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2750 umode_t mode, struct dt_allocation_hint *hint,
2754 struct osd_device *osd = osd_obj2dev(obj);
2755 struct osd_thandle *oth;
2756 struct dt_object *parent = NULL;
2757 struct inode *inode;
2759 LINVRNT(osd_invariant(obj));
2760 LASSERT(obj->oo_inode == NULL);
2761 LASSERT(obj->oo_hl_head == NULL);
2763 if (S_ISDIR(mode) && ldiskfs_pdo) {
2764 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2765 if (obj->oo_hl_head == NULL)
2769 oth = container_of(th, struct osd_thandle, ot_super);
2770 LASSERT(oth->ot_handle->h_transaction != NULL);
2772 if (hint != NULL && hint->dah_parent != NULL &&
2773 !dt_object_remote(hint->dah_parent))
2774 parent = hint->dah_parent;
2776 inode = ldiskfs_create_inode(oth->ot_handle,
2777 parent ? osd_dt_obj(parent)->oo_inode :
2778 osd_sb(osd)->s_root->d_inode,
2780 if (!IS_ERR(inode)) {
2781 /* Do not update file c/mtime in ldiskfs. */
2782 inode->i_flags |= S_NOCMTIME;
2784 /* For new created object, it must be consistent,
2785 * and it is unnecessary to scrub against it. */
2786 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2788 obj->oo_inode = inode;
2791 if (obj->oo_hl_head != NULL) {
2792 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2793 obj->oo_hl_head = NULL;
2795 result = PTR_ERR(inode);
2797 LINVRNT(osd_invariant(obj));
2805 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2806 struct lu_attr *attr,
2807 struct dt_allocation_hint *hint,
2808 struct dt_object_format *dof,
2812 struct osd_thandle *oth;
2813 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2815 LASSERT(S_ISDIR(attr->la_mode));
2817 oth = container_of(th, struct osd_thandle, ot_super);
2818 LASSERT(oth->ot_handle->h_transaction != NULL);
2819 result = osd_mkfile(info, obj, mode, hint, th);
2824 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2825 struct lu_attr *attr,
2826 struct dt_allocation_hint *hint,
2827 struct dt_object_format *dof,
2831 struct osd_thandle *oth;
2832 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2834 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2836 LASSERT(S_ISREG(attr->la_mode));
2838 oth = container_of(th, struct osd_thandle, ot_super);
2839 LASSERT(oth->ot_handle->h_transaction != NULL);
2841 result = osd_mkfile(info, obj, mode, hint, th);
2843 LASSERT(obj->oo_inode != NULL);
2844 if (feat->dif_flags & DT_IND_VARKEY)
2845 result = iam_lvar_create(obj->oo_inode,
2846 feat->dif_keysize_max,
2848 feat->dif_recsize_max,
2851 result = iam_lfix_create(obj->oo_inode,
2852 feat->dif_keysize_max,
2854 feat->dif_recsize_max,
2861 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2862 struct lu_attr *attr,
2863 struct dt_allocation_hint *hint,
2864 struct dt_object_format *dof,
2867 LASSERT(S_ISREG(attr->la_mode));
2868 return osd_mkfile(info, obj, (attr->la_mode &
2869 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2872 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2873 struct lu_attr *attr,
2874 struct dt_allocation_hint *hint,
2875 struct dt_object_format *dof,
2878 LASSERT(S_ISLNK(attr->la_mode));
2879 return osd_mkfile(info, obj, (attr->la_mode &
2880 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2883 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2884 struct lu_attr *attr,
2885 struct dt_allocation_hint *hint,
2886 struct dt_object_format *dof,
2889 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2892 LINVRNT(osd_invariant(obj));
2893 LASSERT(obj->oo_inode == NULL);
2894 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2895 S_ISFIFO(mode) || S_ISSOCK(mode));
2897 result = osd_mkfile(info, obj, mode, hint, th);
2899 LASSERT(obj->oo_inode != NULL);
2901 * This inode should be marked dirty for i_rdev. Currently
2902 * that is done in the osd_attr_init().
2904 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2907 LINVRNT(osd_invariant(obj));
2911 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2913 struct dt_allocation_hint *hint,
2914 struct dt_object_format *dof,
2917 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2919 osd_obj_type_f result;
2935 result = osd_mk_index;
2946 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2947 struct dt_object *parent, struct dt_object *child,
2952 ah->dah_parent = parent;
2953 ah->dah_mode = child_mode;
2955 if (parent != NULL && !dt_object_remote(parent)) {
2956 /* will help to find FID->ino at dt_insert("..") */
2957 struct osd_object *pobj = osd_dt_obj(parent);
2958 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2962 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2963 struct lu_attr *attr, struct dt_object_format *dof)
2965 struct inode *inode = obj->oo_inode;
2966 __u64 valid = attr->la_valid;
2969 attr->la_valid &= ~(LA_TYPE | LA_MODE);
2971 if (dof->dof_type != DFT_NODE)
2972 attr->la_valid &= ~LA_RDEV;
2973 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2974 attr->la_valid &= ~LA_ATIME;
2975 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2976 attr->la_valid &= ~LA_CTIME;
2977 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2978 attr->la_valid &= ~LA_MTIME;
2980 result = osd_quota_transfer(inode, attr);
2984 if (attr->la_valid != 0) {
2985 result = osd_inode_setattr(info->oti_env, inode, attr);
2987 * The osd_inode_setattr() should always succeed here. The
2988 * only error that could be returned is EDQUOT when we are
2989 * trying to change the UID or GID of the inode. However, this
2990 * should not happen since quota enforcement is no longer
2991 * enabled on ldiskfs (lquota takes care of it).
2993 LASSERTF(result == 0, "%d\n", result);
2994 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2997 attr->la_valid = valid;
3001 * Helper function for osd_object_create()
3003 * \retval 0, on success
3005 static int __osd_object_create(struct osd_thread_info *info,
3006 struct osd_object *obj, struct lu_attr *attr,
3007 struct dt_allocation_hint *hint,
3008 struct dt_object_format *dof,
3014 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3016 /* we drop umask so that permissions we pass are not affected */
3017 umask = current->fs->umask;
3018 current->fs->umask = 0;
3020 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3022 if (likely(obj->oo_inode != NULL)) {
3023 LASSERT(obj->oo_inode->i_state & I_NEW);
3025 /* Unlock the inode before attr initialization to avoid
3026 * unnecessary dqget operations. LU-6378 */
3027 unlock_new_inode(obj->oo_inode);
3030 if (likely(result == 0)) {
3031 osd_attr_init(info, obj, attr, dof);
3032 osd_object_init0(obj);
3035 /* restore previous umask value */
3036 current->fs->umask = umask;
3038 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3044 * Helper function for osd_object_create()
3046 * \retval 0, on success
3048 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3049 const struct lu_fid *fid, struct thandle *th)
3051 struct osd_thread_info *info = osd_oti_get(env);
3052 struct osd_inode_id *id = &info->oti_id;
3053 struct osd_device *osd = osd_obj2dev(obj);
3054 struct osd_thandle *oh;
3057 LASSERT(obj->oo_inode != NULL);
3059 oh = container_of0(th, struct osd_thandle, ot_super);
3060 LASSERT(oh->ot_handle);
3061 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3063 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3064 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3065 OI_CHECK_FLD, NULL);
3066 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3071 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3072 u64 seq, struct lu_seq_range *range)
3074 struct seq_server_site *ss = osd_seq_site(osd);
3076 if (fid_seq_is_idif(seq)) {
3077 fld_range_set_ost(range);
3078 range->lsr_index = idif_ost_idx(seq);
3082 if (!fid_seq_in_fldb(seq)) {
3083 fld_range_set_mdt(range);
3085 /* FIXME: If ss is NULL, it suppose not get lsr_index
3087 range->lsr_index = ss->ss_node_id;
3091 LASSERT(ss != NULL);
3092 fld_range_set_any(range);
3093 /* OSD will only do local fld lookup */
3094 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3098 * Concurrency: no external locking is necessary.
3100 static int osd_declare_object_create(const struct lu_env *env,
3101 struct dt_object *dt,
3102 struct lu_attr *attr,
3103 struct dt_allocation_hint *hint,
3104 struct dt_object_format *dof,
3105 struct thandle *handle)
3107 struct osd_thandle *oh;
3111 LASSERT(handle != NULL);
3113 oh = container_of0(handle, struct osd_thandle, ot_super);
3114 LASSERT(oh->ot_handle == NULL);
3116 /* EA object consumes more credits than regular object: osd_mk_index
3117 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3118 * leaf_node, could involves the block, block bitmap, groups, GDT
3119 * change for each block, so add 4 * 2 credits in that case. */
3120 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3121 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3122 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3123 /* Reuse idle OI block may cause additional one OI block
3125 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3126 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3131 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3132 attr->la_projid, 1, oh, osd_dt_obj(dt),
3133 false, NULL, false);
3137 /* will help to find FID->ino mapping at dt_insert() */
3138 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3144 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
3145 struct lu_attr *attr,
3146 struct dt_allocation_hint *hint,
3147 struct dt_object_format *dof, struct thandle *th)
3149 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3150 struct osd_object *obj = osd_dt_obj(dt);
3151 struct osd_thread_info *info = osd_oti_get(env);
3155 if (dt_object_exists(dt))
3158 LINVRNT(osd_invariant(obj));
3159 LASSERT(!dt_object_remote(dt));
3160 LASSERT(osd_write_locked(env, obj));
3161 LASSERT(th != NULL);
3163 if (unlikely(fid_is_acct(fid)))
3164 /* Quota files can't be created from the kernel any more,
3165 * 'tune2fs -O quota' will take care of creating them */
3168 result = __osd_object_create(info, obj, attr, hint, dof, th);
3170 result = __osd_oi_insert(env, obj, fid, th);
3171 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3172 obj->oo_dt.do_body_ops = &osd_body_ops;
3174 LASSERT(ergo(result == 0,
3175 dt_object_exists(dt) && !dt_object_remote(dt)));
3177 LASSERT(osd_invariant(obj));
3182 * Called to destroy on-disk representation of the object
3184 * Concurrency: must be locked
3186 static int osd_declare_object_destroy(const struct lu_env *env,
3187 struct dt_object *dt,
3190 struct osd_object *obj = osd_dt_obj(dt);
3191 struct inode *inode = obj->oo_inode;
3192 struct osd_thandle *oh;
3199 oh = container_of0(th, struct osd_thandle, ot_super);
3200 LASSERT(oh->ot_handle == NULL);
3202 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3203 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3204 /* Recycle idle OI leaf may cause additional three OI blocks
3206 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3207 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3208 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3209 /* one less inode */
3210 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3211 i_projid_read(inode), -1, oh, obj, false,
3215 /* data to be truncated */
3216 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3217 i_projid_read(inode), 0, oh, obj, true,
3222 /* will help to find FID->ino when this object is being
3223 * added to PENDING/ */
3224 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3229 static int osd_object_destroy(const struct lu_env *env,
3230 struct dt_object *dt,
3233 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3234 struct osd_object *obj = osd_dt_obj(dt);
3235 struct inode *inode = obj->oo_inode;
3236 struct osd_device *osd = osd_obj2dev(obj);
3237 struct osd_thandle *oh;
3241 oh = container_of0(th, struct osd_thandle, ot_super);
3242 LASSERT(oh->ot_handle);
3244 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3246 if (unlikely(fid_is_acct(fid)))
3249 if (S_ISDIR(inode->i_mode)) {
3250 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3251 inode->i_nlink == 2);
3252 /* it will check/delete the inode from remote parent,
3253 * how to optimize it? unlink performance impaction XXX */
3254 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3256 CERROR("%s: delete inode "DFID": rc = %d\n",
3257 osd_name(osd), PFID(fid), result);
3259 spin_lock(&obj->oo_guard);
3261 spin_unlock(&obj->oo_guard);
3262 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3265 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3267 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3269 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3270 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3271 oh->ot_handle, OI_CHECK_FLD);
3273 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3274 /* XXX: add to ext3 orphan list */
3275 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3277 /* not needed in the cache anymore */
3278 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3279 obj->oo_destroyed = 1;
3285 * Put the fid into lustre_mdt_attrs, and then place the structure
3286 * inode's ea. This fid should not be altered during the life time
3289 * \retval +ve, on success
3290 * \retval -ve, on error
3292 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3294 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3295 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3297 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3298 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3302 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3305 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3308 lustre_loa_init(loa, fid, compat, incompat);
3309 lustre_loa_swab(loa, false);
3311 /* For the OST device with 256 bytes inode size by default,
3312 * the PFID EA will be stored together with LMA EA to avoid
3313 * performance trouble. Otherwise the PFID EA can be stored
3314 * independently. LU-8998 */
3315 if ((compat & LMAC_FID_ON_OST) &&
3316 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3317 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3318 sizeof(*loa), XATTR_CREATE);
3320 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3321 sizeof(*lma), XATTR_CREATE);
3322 /* LMA may already exist, but we need to check that all the
3323 * desired compat/incompat flags have been added. */
3324 if (unlikely(rc == -EEXIST)) {
3325 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3326 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3330 if (rc < sizeof(*lma))
3333 lustre_loa_swab(loa, true);
3334 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3335 ((compat == 0 && incompat == 0) ||
3336 (!(~lma->lma_compat & compat) &&
3337 !(~lma->lma_incompat & incompat))))
3340 lma->lma_self_fid = *fid;
3341 lma->lma_compat |= compat;
3342 lma->lma_incompat |= incompat;
3343 if (rc == sizeof(*lma)) {
3344 lustre_lma_swab(lma);
3345 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3346 sizeof(*lma), XATTR_REPLACE);
3348 lustre_loa_swab(loa, false);
3349 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3350 sizeof(*loa), XATTR_REPLACE);
3358 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3359 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3360 * To have compatilibility with 1.8 ldiskfs driver we need to have
3361 * magic number at start of fid data.
3362 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3365 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3366 const struct lu_fid *fid)
3368 if (!fid_is_namespace_visible(fid) ||
3369 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3370 param->edp_magic = 0;
3374 param->edp_magic = LDISKFS_LUFID_MAGIC;
3375 param->edp_len = sizeof(struct lu_fid) + 1;
3376 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3380 * Try to read the fid from inode ea into dt_rec.
3382 * \param fid object fid.
3384 * \retval 0 on success
3386 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3387 __u32 ino, struct lu_fid *fid,
3388 struct osd_inode_id *id)
3390 struct osd_thread_info *info = osd_oti_get(env);
3391 struct inode *inode;
3394 osd_id_gen(id, ino, OSD_OII_NOGEN);
3395 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3397 RETURN(PTR_ERR(inode));
3403 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3405 struct inode *parent_dir,
3406 const struct lu_fid *dot_fid,
3407 const struct lu_fid *dot_dot_fid,
3408 struct osd_thandle *oth)
3410 struct ldiskfs_dentry_param *dot_ldp;
3411 struct ldiskfs_dentry_param *dot_dot_ldp;
3412 __u32 saved_nlink = dir->i_nlink;
3415 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3416 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3418 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3419 dot_ldp->edp_magic = 0;
3421 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3422 dir, dot_ldp, dot_dot_ldp);
3423 /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3424 * the subseqent ref_add() will increase the dir->i_nlink
3425 * as 3. That is incorrect for new created directory.
3427 * It looks like hack, because we want to make the OSD API
3428 * to be order-independent for new created directory object
3429 * between dt_insert(..) and ref_add() operations.
3431 * Here, we only restore the in-RAM dir-inode's nlink attr,
3432 * becuase if the nlink attr is not 2, then there will be
3433 * ref_add() called following the dt_insert(..), such call
3434 * will make both the in-RAM and on-disk dir-inode's nlink
3435 * attr to be set as 2. LU-7447 */
3436 set_nlink(dir, saved_nlink);
3441 * Create an local agent inode for remote entry
3443 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3444 struct osd_device *osd,
3445 struct osd_object *pobj,
3446 const struct lu_fid *fid,
3450 struct osd_thread_info *info = osd_oti_get(env);
3451 struct inode *local;
3452 struct osd_thandle *oh;
3457 oh = container_of(th, struct osd_thandle, ot_super);
3458 LASSERT(oh->ot_handle->h_transaction != NULL);
3460 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3461 if (IS_ERR(local)) {
3462 CERROR("%s: create local error %d\n", osd_name(osd),
3463 (int)PTR_ERR(local));
3467 /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3468 * correct gid on remote file, not agent here */
3469 local->i_gid = current_fsgid();
3470 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3471 unlock_new_inode(local);
3473 /* Agent inode should not have project ID*/
3474 #ifdef HAVE_PROJECT_QUOTA
3475 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL) {
3476 rc = __ldiskfs_ioctl_setproject(local, 0);
3478 CERROR("%s: quota transfer failed: rc = %d. Is project "
3479 "quota enforcement enabled on the ldiskfs "
3480 "filesystem?\n", local->i_sb->s_id, rc);
3481 RETURN(ERR_PTR(rc));
3486 /* Set special LMA flag for local agent inode */
3487 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3489 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3490 osd_name(osd), PFID(fid), rc);
3491 RETURN(ERR_PTR(rc));
3497 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3498 lu_object_fid(&pobj->oo_dt.do_lu),
3501 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3502 osd_name(osd), PFID(fid), rc);
3503 RETURN(ERR_PTR(rc));
3510 * when direntry is deleted, we have to take care of possible agent inode
3511 * referenced by that. unfortunately we can't do this at that point:
3512 * iget() within a running transaction leads to deadlock and we better do
3513 * not call that every delete declaration to save performance. so we put
3514 * a potention agent inode on a list and process that once the transaction
3515 * is over. Notice it's not any worse in terms of real orphans as regular
3516 * object destroy doesn't put inodes on the on-disk orphan list. this should
3517 * be addressed separately
3519 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3520 struct osd_thandle *oh,
3523 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3524 struct osd_obj_orphan *oor;
3531 oor->oor_env = (struct lu_env *)env;
3532 spin_lock(&osd->od_osfs_lock);
3533 list_add(&oor->oor_list, &osd->od_orphan_list);
3534 spin_unlock(&osd->od_osfs_lock);
3536 oh->ot_remove_agents = 1;
3542 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3543 struct osd_device *osd)
3545 struct osd_thread_info *info = osd_oti_get(env);
3546 struct osd_obj_orphan *oor, *tmp;
3547 struct osd_inode_id id;
3548 struct list_head list;
3549 struct inode *inode;
3554 INIT_LIST_HEAD(&list);
3556 spin_lock(&osd->od_osfs_lock);
3557 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3558 if (oor->oor_env == env) {
3559 list_del(&oor->oor_list);
3560 list_add(&oor->oor_list, &list);
3563 spin_unlock(&osd->od_osfs_lock);
3565 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3569 list_del(&oor->oor_list);
3572 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3573 inode = osd_iget_fid(info, osd, &id, &fid);
3577 if (!osd_remote_fid(env, osd, &fid)) {
3582 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3584 mark_inode_dirty(inode);
3585 ldiskfs_journal_stop(jh);
3593 * OSD layer object create function for interoperability mode (b11826).
3594 * This is mostly similar to osd_object_create(). Only difference being, fid is
3595 * inserted into inode ea here.
3597 * \retval 0, on success
3598 * \retval -ve, on error
3600 static int osd_object_ea_create(const struct lu_env *env, struct dt_object *dt,
3601 struct lu_attr *attr,
3602 struct dt_allocation_hint *hint,
3603 struct dt_object_format *dof,
3606 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3607 struct osd_object *obj = osd_dt_obj(dt);
3608 struct osd_thread_info *info = osd_oti_get(env);
3609 int result, on_ost = 0;
3613 if (dt_object_exists(dt))
3616 LASSERT(osd_invariant(obj));
3617 LASSERT(!dt_object_remote(dt));
3618 LASSERT(osd_write_locked(env, obj));
3619 LASSERT(th != NULL);
3621 if (unlikely(fid_is_acct(fid)))
3622 /* Quota files can't be created from the kernel any more,
3623 * 'tune2fs -O quota' will take care of creating them */
3626 result = __osd_object_create(info, obj, attr, hint, dof, th);
3628 if (fid_is_idif(fid) &&
3629 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3630 struct lu_fid *tfid = &info->oti_fid;
3631 struct ost_id *oi = &info->oti_ostid;
3633 fid_to_ostid(fid, oi);
3634 ostid_to_fid(tfid, oi, 0);
3636 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3637 LMAC_FID_ON_OST, 0);
3639 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3641 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3642 on_ost ? LMAC_FID_ON_OST : 0,
3645 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3646 obj->oo_dt.do_body_ops = &osd_body_ops;
3650 result = __osd_oi_insert(env, obj, fid, th);
3652 /* a small optimization - dt_insert() isn't usually applied
3653 * to OST objects, so we don't need to cache OI mapping for
3655 if (result == 0 && on_ost == 0) {
3656 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3657 result = osd_idc_find_and_init(env, osd, obj);
3658 LASSERT(result == 0);
3661 LASSERT(ergo(result == 0,
3662 dt_object_exists(dt) && !dt_object_remote(dt)));
3663 LINVRNT(osd_invariant(obj));
3667 static int osd_declare_object_ref_add(const struct lu_env *env,
3668 struct dt_object *dt,
3669 struct thandle *handle)
3671 struct osd_thandle *oh;
3673 /* it's possible that object doesn't exist yet */
3674 LASSERT(handle != NULL);
3676 oh = container_of0(handle, struct osd_thandle, ot_super);
3677 LASSERT(oh->ot_handle == NULL);
3679 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3680 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3686 * Concurrency: @dt is write locked.
3688 static int osd_object_ref_add(const struct lu_env *env,
3689 struct dt_object *dt, struct thandle *th)
3691 struct osd_object *obj = osd_dt_obj(dt);
3692 struct inode *inode = obj->oo_inode;
3693 struct osd_thandle *oh;
3696 if (!dt_object_exists(dt) || obj->oo_destroyed)
3699 LINVRNT(osd_invariant(obj));
3700 LASSERT(!dt_object_remote(dt));
3701 LASSERT(osd_write_locked(env, obj));
3702 LASSERT(th != NULL);
3704 oh = container_of0(th, struct osd_thandle, ot_super);
3705 LASSERT(oh->ot_handle != NULL);
3707 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3709 CDEBUG(D_INODE, DFID" increase nlink %d\n",
3710 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3712 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3713 * (65000) subdirectories by storing "1" in i_nlink if the link count
3714 * would otherwise overflow. Directory tranversal tools understand
3715 * that (st_nlink == 1) indicates that the filesystem dose not track
3716 * hard links count on the directory, and will not abort subdirectory
3717 * scanning early once (st_nlink - 2) subdirs have been found.
3719 * This also has to properly handle the case of inodes with nlink == 0
3720 * in case they are being linked into the PENDING directory
3722 spin_lock(&obj->oo_guard);
3723 if (unlikely(inode->i_nlink == 0))
3724 /* inc_nlink from 0 may cause WARN_ON */
3725 set_nlink(inode, 1);
3727 ldiskfs_inc_count(oh->ot_handle, inode);
3728 if (!S_ISDIR(inode->i_mode))
3729 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3731 spin_unlock(&obj->oo_guard);
3733 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3734 LINVRNT(osd_invariant(obj));
3736 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3741 static int osd_declare_object_ref_del(const struct lu_env *env,
3742 struct dt_object *dt,
3743 struct thandle *handle)
3745 struct osd_thandle *oh;
3747 if (!dt_object_exists(dt))
3750 LASSERT(!dt_object_remote(dt));
3751 LASSERT(handle != NULL);
3753 oh = container_of0(handle, struct osd_thandle, ot_super);
3754 LASSERT(oh->ot_handle == NULL);
3756 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3757 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3763 * Concurrency: @dt is write locked.
3765 static int osd_object_ref_del(const struct lu_env *env, struct dt_object *dt,
3768 struct osd_object *obj = osd_dt_obj(dt);
3769 struct inode *inode = obj->oo_inode;
3770 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3771 struct osd_thandle *oh;
3773 if (!dt_object_exists(dt))
3776 LINVRNT(osd_invariant(obj));
3777 LASSERT(!dt_object_remote(dt));
3778 LASSERT(osd_write_locked(env, obj));
3779 LASSERT(th != NULL);
3781 oh = container_of0(th, struct osd_thandle, ot_super);
3782 LASSERT(oh->ot_handle != NULL);
3784 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3786 spin_lock(&obj->oo_guard);
3787 /* That can be result of upgrade from old Lustre version and
3788 * applied only to local files. Just skip this ref_del call.
3789 * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3790 if (inode->i_nlink == 0) {
3791 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3792 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3793 ", maybe an upgraded file? (LU-3915)\n",
3794 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3795 spin_unlock(&obj->oo_guard);
3799 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3800 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3802 ldiskfs_dec_count(oh->ot_handle, inode);
3803 spin_unlock(&obj->oo_guard);
3805 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3806 LINVRNT(osd_invariant(obj));
3808 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3814 * Concurrency: @dt is read locked.
3816 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3817 struct lu_buf *buf, const char *name)
3819 struct osd_object *obj = osd_dt_obj(dt);
3820 struct inode *inode = obj->oo_inode;
3821 struct osd_thread_info *info = osd_oti_get(env);
3822 struct dentry *dentry = &info->oti_obj_dentry;
3823 bool cache_xattr = false;
3828 /* version get is not real XATTR but uses xattr API */
3829 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3830 dt_obj_version_t *ver = buf->lb_buf;
3832 /* for version we are just using xattr API but change inode
3834 if (buf->lb_len == 0)
3835 return sizeof(dt_obj_version_t);
3837 if (buf->lb_len < sizeof(dt_obj_version_t))
3840 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
3841 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3843 *ver = LDISKFS_I(inode)->i_fs_version;
3845 return sizeof(dt_obj_version_t);
3848 if (!dt_object_exists(dt))
3851 LASSERT(!dt_object_remote(dt));
3852 LASSERT(inode->i_op != NULL);
3853 LASSERT(inode->i_op->getxattr != NULL);
3855 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
3856 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
3860 rc = osd_oxc_get(obj, name, buf);
3865 rc = __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3866 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
3867 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3868 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3869 struct filter_fid *ff;
3870 struct ost_layout *ol;
3872 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
3874 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
3878 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
3884 if (buf->lb_len == 0 || !buf->lb_buf)
3887 if (buf->lb_len < rc)
3891 ol = &ff->ff_layout;
3892 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
3893 PFID_STRIPE_IDX_BITS);
3894 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
3895 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
3896 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
3897 if (lma->lma_compat & LMAC_COMP_INFO) {
3898 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
3899 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
3900 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
3902 ol->ol_comp_start = 0;
3903 ol->ol_comp_end = 0;
3910 if (rc == -ENOENT || rc == -ENODATA)
3911 osd_oxc_add(obj, name, NULL, 0);
3912 else if (rc > 0 && buf->lb_buf != NULL)
3913 osd_oxc_add(obj, name, buf->lb_buf, rc);
3919 static int osd_declare_xattr_set(const struct lu_env *env,
3920 struct dt_object *dt,
3921 const struct lu_buf *buf, const char *name,
3922 int fl, struct thandle *handle)
3924 struct osd_thandle *oh;
3926 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3928 LASSERT(handle != NULL);
3930 oh = container_of0(handle, struct osd_thandle, ot_super);
3931 LASSERT(oh->ot_handle == NULL);
3933 if (strcmp(name, XATTR_NAME_LMA) == 0) {
3934 /* For non-upgrading case, the LMA is set first and
3935 * usually fit inode. But for upgrade case, the LMA
3936 * may be in another separated EA block. */
3937 if (dt_object_exists(dt)) {
3938 if (fl == LU_XATTR_REPLACE)
3943 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3945 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
3946 /* We may need to delete the old PFID EA. */
3947 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
3948 if (fl == LU_XATTR_REPLACE)
3955 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
3960 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3961 /* learn xattr size from osd_xattr_get if
3962 attribute has not been read yet */
3963 buflen = __osd_xattr_get(
3964 osd_dt_obj(dt)->oo_inode,
3965 &osd_oti_get(env)->oti_obj_dentry,
3970 buflen = buf->lb_len;
3973 if (buflen > sb->s_blocksize) {
3974 credits += osd_calc_bkmap_credits(
3976 (buflen + sb->s_blocksize - 1) >>
3977 sb->s_blocksize_bits);
3981 * xattr set may involve inode quota change, reserve credits for
3982 * dquot_initialize()
3984 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3987 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3993 * Concurrency: @dt is write locked.
3995 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3996 const struct lu_buf *buf, const char *name, int fl,
3997 struct thandle *handle)
3999 struct osd_object *obj = osd_dt_obj(dt);
4000 struct inode *inode = obj->oo_inode;
4001 struct osd_thread_info *info = osd_oti_get(env);
4002 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4003 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4012 /* version set is not real XATTR */
4013 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4014 dt_obj_version_t *version = buf->lb_buf;
4016 /* for version we are just using xattr API but change inode
4018 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4020 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4021 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4023 LDISKFS_I(inode)->i_fs_version = *version;
4024 /* Version is set after all inode operations are finished,
4025 * so we should mark it dirty here */
4026 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4031 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4032 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4035 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4036 if (fl & LU_XATTR_REPLACE)
4037 fs_flags |= XATTR_REPLACE;
4039 if (fl & LU_XATTR_CREATE)
4040 fs_flags |= XATTR_CREATE;
4042 /* For the OST device with 256 bytes inode size by default,
4043 * the PFID EA will be stored together with LMA EA to avoid
4044 * performance trouble. Otherwise the PFID EA can be stored
4045 * independently. LU-8998 */
4046 if (strcmp(name, XATTR_NAME_FID) == 0 &&
4047 LDISKFS_INODE_SIZE(inode->i_sb) <= 256) {
4048 struct dentry *dentry = &info->oti_obj_dentry;
4049 struct filter_fid *ff;
4050 struct ost_layout *ol;
4053 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4056 ol = &ff->ff_layout;
4057 /* Old client does not send stripe information, store
4058 * the PFID EA on disk directly. */
4059 if (buf->lb_len == sizeof(struct lu_fid) ||
4060 ol->ol_stripe_size == 0) {
4061 len = sizeof(struct lu_fid);
4065 if (buf->lb_len != sizeof(*ff))
4068 rc = osd_get_lma(info, inode, dentry, loa);
4069 if (unlikely(rc == -ENODATA)) {
4070 /* Usually for upgarding from old device */
4071 lustre_loa_init(loa, lu_object_fid(&dt->do_lu),
4072 LMAC_FID_ON_OST, 0);
4080 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4081 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4082 PFID_STRIPE_IDX_BITS;
4083 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4084 lma->lma_compat |= LMAC_STRIPE_INFO;
4085 if (ol->ol_comp_id != 0) {
4086 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4087 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4088 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4089 lma->lma_compat |= LMAC_COMP_INFO;
4092 lustre_loa_swab(loa, false);
4094 /* Remove old PFID EA entry firstly. */
4095 ll_vfs_dq_init(inode);
4096 rc = inode->i_op->removexattr(dentry, name);
4097 if (rc && rc != -ENODATA)
4100 /* Store the PFID EA inside the LMA EA. */
4101 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4105 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4106 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4110 lma->lma_incompat |= LMAI_STRIPED;
4111 lustre_lma_swab(lma);
4112 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4113 sizeof(*lma), XATTR_REPLACE);
4119 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4120 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4123 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4124 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4125 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4131 * Concurrency: @dt is read locked.
4133 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4134 const struct lu_buf *buf)
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;
4141 if (!dt_object_exists(dt))
4144 LASSERT(!dt_object_remote(dt));
4145 LASSERT(inode->i_op != NULL);
4146 LASSERT(inode->i_op->listxattr != NULL);
4148 dentry->d_inode = inode;
4149 dentry->d_sb = inode->i_sb;
4150 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4153 static int osd_declare_xattr_del(const struct lu_env *env,
4154 struct dt_object *dt, const char *name,
4155 struct thandle *handle)
4157 struct osd_thandle *oh;
4158 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4160 LASSERT(!dt_object_remote(dt));
4161 LASSERT(handle != NULL);
4163 oh = container_of0(handle, struct osd_thandle, ot_super);
4164 LASSERT(oh->ot_handle == NULL);
4166 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4167 osd_dto_credits_noquota[DTO_XATTR_SET]);
4169 * xattr del may involve inode quota change, reserve credits for
4170 * dquot_initialize()
4172 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4178 * Concurrency: @dt is write locked.
4180 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4181 const char *name, struct thandle *handle)
4183 struct osd_object *obj = osd_dt_obj(dt);
4184 struct inode *inode = obj->oo_inode;
4185 struct osd_thread_info *info = osd_oti_get(env);
4186 struct dentry *dentry = &info->oti_obj_dentry;
4189 if (!dt_object_exists(dt))
4192 LASSERT(!dt_object_remote(dt));
4193 LASSERT(inode->i_op != NULL);
4194 LASSERT(inode->i_op->removexattr != NULL);
4195 LASSERT(handle != NULL);
4197 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4199 ll_vfs_dq_init(inode);
4200 dentry->d_inode = inode;
4201 dentry->d_sb = inode->i_sb;
4202 rc = inode->i_op->removexattr(dentry, name);
4203 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
4204 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4206 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4208 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4209 &info->oti_ost_attrs);
4211 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
4216 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4217 lustre_lma_swab(lma);
4218 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4219 sizeof(*lma), XATTR_REPLACE);
4224 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4227 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4228 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4229 osd_oxc_del(obj, name);
4234 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4235 __u64 start, __u64 end)
4237 struct osd_object *obj = osd_dt_obj(dt);
4238 struct inode *inode = obj->oo_inode;
4239 struct osd_thread_info *info = osd_oti_get(env);
4240 struct dentry *dentry = &info->oti_obj_dentry;
4241 struct file *file = &info->oti_file;
4246 dentry->d_inode = inode;
4247 dentry->d_sb = inode->i_sb;
4248 file->f_path.dentry = dentry;
4249 file->f_mapping = inode->i_mapping;
4250 file->f_op = inode->i_fop;
4251 set_file_inode(file, inode);
4253 rc = ll_vfs_fsync_range(file, start, end, 0);
4258 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4267 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4268 const struct dt_index_features *feat)
4270 struct iam_descr *descr;
4272 if (osd_object_is_root(o))
4273 return feat == &dt_directory_features;
4275 LASSERT(o->oo_dir != NULL);
4277 descr = o->oo_dir->od_container.ic_descr;
4278 if (feat == &dt_directory_features) {
4279 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4285 feat->dif_keysize_min <= descr->id_key_size &&
4286 descr->id_key_size <= feat->dif_keysize_max &&
4287 feat->dif_recsize_min <= descr->id_rec_size &&
4288 descr->id_rec_size <= feat->dif_recsize_max &&
4289 !(feat->dif_flags & (DT_IND_VARKEY |
4290 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4291 ergo(feat->dif_flags & DT_IND_UPDATE,
4292 1 /* XXX check that object (and file system) is
4297 static int osd_iam_container_init(const struct lu_env *env,
4298 struct osd_object *obj,
4299 struct osd_directory *dir)
4301 struct iam_container *bag = &dir->od_container;
4304 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4308 result = iam_container_setup(bag);
4310 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4312 iam_container_fini(bag);
4319 * Concurrency: no external locking is necessary.
4321 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4322 const struct dt_index_features *feat)
4326 struct osd_object *obj = osd_dt_obj(dt);
4328 LINVRNT(osd_invariant(obj));
4330 if (osd_object_is_root(obj)) {
4331 dt->do_index_ops = &osd_index_ea_ops;
4333 } else if (feat == &dt_directory_features) {
4334 dt->do_index_ops = &osd_index_ea_ops;
4335 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4340 } else if (unlikely(feat == &dt_otable_features)) {
4341 dt->do_index_ops = &osd_otable_ops;
4343 } else if (unlikely(feat == &dt_acct_features)) {
4344 dt->do_index_ops = &osd_acct_index_ops;
4347 } else if (!osd_has_index(obj)) {
4348 struct osd_directory *dir;
4353 spin_lock(&obj->oo_guard);
4354 if (obj->oo_dir == NULL)
4358 * Concurrent thread allocated container data.
4361 spin_unlock(&obj->oo_guard);
4363 * Now, that we have container data, serialize its
4366 down_write(&obj->oo_ext_idx_sem);
4368 * recheck under lock.
4370 if (!osd_has_index(obj))
4371 result = osd_iam_container_init(env, obj,
4375 up_write(&obj->oo_ext_idx_sem);
4383 if (result == 0 && skip_iam == 0) {
4384 if (!osd_iam_index_probe(env, obj, feat))
4387 LINVRNT(osd_invariant(obj));
4392 static int osd_otable_it_attr_get(const struct lu_env *env,
4393 struct dt_object *dt,
4394 struct lu_attr *attr)
4400 static const struct dt_object_operations osd_obj_ops = {
4401 .do_read_lock = osd_object_read_lock,
4402 .do_write_lock = osd_object_write_lock,
4403 .do_read_unlock = osd_object_read_unlock,
4404 .do_write_unlock = osd_object_write_unlock,
4405 .do_write_locked = osd_object_write_locked,
4406 .do_attr_get = osd_attr_get,
4407 .do_declare_attr_set = osd_declare_attr_set,
4408 .do_attr_set = osd_attr_set,
4409 .do_ah_init = osd_ah_init,
4410 .do_declare_create = osd_declare_object_create,
4411 .do_create = osd_object_create,
4412 .do_declare_destroy = osd_declare_object_destroy,
4413 .do_destroy = osd_object_destroy,
4414 .do_index_try = osd_index_try,
4415 .do_declare_ref_add = osd_declare_object_ref_add,
4416 .do_ref_add = osd_object_ref_add,
4417 .do_declare_ref_del = osd_declare_object_ref_del,
4418 .do_ref_del = osd_object_ref_del,
4419 .do_xattr_get = osd_xattr_get,
4420 .do_declare_xattr_set = osd_declare_xattr_set,
4421 .do_xattr_set = osd_xattr_set,
4422 .do_declare_xattr_del = osd_declare_xattr_del,
4423 .do_xattr_del = osd_xattr_del,
4424 .do_xattr_list = osd_xattr_list,
4425 .do_object_sync = osd_object_sync,
4426 .do_invalidate = osd_invalidate,
4430 * dt_object_operations for interoperability mode
4431 * (i.e. to run 2.0 mds on 1.8 disk) (b11826)
4433 static const struct dt_object_operations osd_obj_ea_ops = {
4434 .do_read_lock = osd_object_read_lock,
4435 .do_write_lock = osd_object_write_lock,
4436 .do_read_unlock = osd_object_read_unlock,
4437 .do_write_unlock = osd_object_write_unlock,
4438 .do_write_locked = osd_object_write_locked,
4439 .do_attr_get = osd_attr_get,
4440 .do_declare_attr_set = osd_declare_attr_set,
4441 .do_attr_set = osd_attr_set,
4442 .do_ah_init = osd_ah_init,
4443 .do_declare_create = osd_declare_object_create,
4444 .do_create = osd_object_ea_create,
4445 .do_declare_destroy = osd_declare_object_destroy,
4446 .do_destroy = osd_object_destroy,
4447 .do_index_try = osd_index_try,
4448 .do_declare_ref_add = osd_declare_object_ref_add,
4449 .do_ref_add = osd_object_ref_add,
4450 .do_declare_ref_del = osd_declare_object_ref_del,
4451 .do_ref_del = osd_object_ref_del,
4452 .do_xattr_get = osd_xattr_get,
4453 .do_declare_xattr_set = osd_declare_xattr_set,
4454 .do_xattr_set = osd_xattr_set,
4455 .do_declare_xattr_del = osd_declare_xattr_del,
4456 .do_xattr_del = osd_xattr_del,
4457 .do_xattr_list = osd_xattr_list,
4458 .do_object_sync = osd_object_sync,
4459 .do_invalidate = osd_invalidate,
4462 static const struct dt_object_operations osd_obj_otable_it_ops = {
4463 .do_attr_get = osd_otable_it_attr_get,
4464 .do_index_try = osd_index_try,
4467 static int osd_index_declare_iam_delete(const struct lu_env *env,
4468 struct dt_object *dt,
4469 const struct dt_key *key,
4470 struct thandle *handle)
4472 struct osd_thandle *oh;
4474 oh = container_of0(handle, struct osd_thandle, ot_super);
4475 LASSERT(oh->ot_handle == NULL);
4477 /* Recycle may cause additional three blocks to be changed. */
4478 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4479 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4485 * delete a (key, value) pair from index \a dt specified by \a key
4487 * \param dt osd index object
4488 * \param key key for index
4489 * \param rec record reference
4490 * \param handle transaction handler
4493 * \retval -ve failure
4495 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4496 const struct dt_key *key,
4497 struct thandle *handle)
4499 struct osd_thread_info *oti = osd_oti_get(env);
4500 struct osd_object *obj = osd_dt_obj(dt);
4501 struct osd_thandle *oh;
4502 struct iam_path_descr *ipd;
4503 struct iam_container *bag = &obj->oo_dir->od_container;
4507 if (!dt_object_exists(dt))
4510 LINVRNT(osd_invariant(obj));
4511 LASSERT(!dt_object_remote(dt));
4512 LASSERT(bag->ic_object == obj->oo_inode);
4513 LASSERT(handle != NULL);
4515 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4517 ipd = osd_idx_ipd_get(env, bag);
4518 if (unlikely(ipd == NULL))
4521 oh = container_of0(handle, struct osd_thandle, ot_super);
4522 LASSERT(oh->ot_handle != NULL);
4523 LASSERT(oh->ot_handle->h_transaction != NULL);
4525 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4526 /* swab quota uid/gid provided by caller */
4527 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4528 key = (const struct dt_key *)&oti->oti_quota_id;
4531 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4532 osd_ipd_put(env, bag, ipd);
4533 LINVRNT(osd_invariant(obj));
4534 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4538 static int osd_index_declare_ea_delete(const struct lu_env *env,
4539 struct dt_object *dt,
4540 const struct dt_key *key,
4541 struct thandle *handle)
4543 struct osd_thandle *oh;
4544 struct inode *inode;
4548 LASSERT(!dt_object_remote(dt));
4549 LASSERT(handle != NULL);
4551 oh = container_of0(handle, struct osd_thandle, ot_super);
4552 LASSERT(oh->ot_handle == NULL);
4554 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4555 if (key != NULL && unlikely(strcmp((char *)key, dotdot) == 0)) {
4556 /* '..' to a remote object has a local representative */
4557 credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
4558 /* to reset LMAI_REMOTE_PARENT */
4561 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4563 inode = osd_dt_obj(dt)->oo_inode;
4567 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4568 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
4573 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
4576 struct osd_fid_pack *rec;
4579 if (de->file_type & LDISKFS_DIRENT_LUFID) {
4580 rec = (struct osd_fid_pack *) (de->name + de->name_len + 1);
4581 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
4582 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
4588 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
4589 const struct lu_fid *fid)
4591 struct seq_server_site *ss = osd_seq_site(osd);
4594 /* FID seqs not in FLDB, must be local seq */
4595 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
4598 /* If FLD is not being initialized yet, it only happens during the
4599 * initialization, likely during mgs initialization, and we assume
4600 * this is local FID. */
4601 if (ss == NULL || ss->ss_server_fld == NULL)
4604 /* Only check the local FLDB here */
4605 if (osd_seq_exists(env, osd, fid_seq(fid)))
4612 * Index delete function for interoperability mode (b11826).
4613 * It will remove the directory entry added by osd_index_ea_insert().
4614 * This entry is needed to maintain name->fid mapping.
4616 * \param key, key i.e. file entry to be deleted
4618 * \retval 0, on success
4619 * \retval -ve, on error
4621 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
4622 const struct dt_key *key, struct thandle *handle)
4624 struct osd_object *obj = osd_dt_obj(dt);
4625 struct inode *dir = obj->oo_inode;
4626 struct dentry *dentry;
4627 struct osd_thandle *oh;
4628 struct ldiskfs_dir_entry_2 *de = NULL;
4629 struct buffer_head *bh;
4630 struct htree_lock *hlock = NULL;
4631 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4632 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4636 if (!dt_object_exists(dt))
4639 LINVRNT(osd_invariant(obj));
4640 LASSERT(!dt_object_remote(dt));
4641 LASSERT(handle != NULL);
4643 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4645 oh = container_of(handle, struct osd_thandle, ot_super);
4646 LASSERT(oh->ot_handle != NULL);
4647 LASSERT(oh->ot_handle->h_transaction != NULL);
4649 ll_vfs_dq_init(dir);
4650 dentry = osd_child_dentry_get(env, obj,
4651 (char *)key, strlen((char *)key));
4653 if (obj->oo_hl_head != NULL) {
4654 hlock = osd_oti_get(env)->oti_hlock;
4655 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
4656 dir, LDISKFS_HLOCK_DEL);
4658 down_write(&obj->oo_ext_idx_sem);
4661 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
4663 /* If this is not the ".." entry, it might be a remote DNE
4664 * entry and we need to check if the FID is for a remote
4665 * MDT. If the FID is not in the directory entry (e.g.
4666 * upgraded 1.8 filesystem without dirdata enabled) then
4667 * we need to get the FID from the LMA. For a remote directory
4668 * there HAS to be an LMA, it cannot be an IGIF inode in this
4671 * Delete the entry before the agent inode in order to
4672 * simplify error handling. At worst an error after deleting
4673 * the entry first might leak the agent inode afterward. The
4674 * reverse would need filesystem abort in case of error deleting
4675 * the entry after the agent had been removed, or leave a
4676 * dangling entry pointing at a random inode. */
4677 if (strcmp((char *)key, dotdot) != 0) {
4678 LASSERT(de != NULL);
4679 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
4680 if (rc == -ENODATA) {
4681 /* can't get FID, postpone to the end of the
4682 * transaction when iget() is safe */
4683 osd_schedule_agent_inode_removal(env, oh,
4684 le32_to_cpu(de->inode));
4685 } else if (rc == 0 &&
4686 unlikely(osd_remote_fid(env, osd, fid))) {
4687 osd_schedule_agent_inode_removal(env, oh,
4688 le32_to_cpu(de->inode));
4691 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
4697 ldiskfs_htree_unlock(hlock);
4699 up_write(&obj->oo_ext_idx_sem);
4704 /* For inode on the remote MDT, .. will point to
4705 * /Agent directory, Check whether it needs to delete
4706 * from agent directory */
4707 if (unlikely(strcmp((char *)key, dotdot) == 0)) {
4710 ret = osd_delete_from_remote_parent(env, osd_obj2dev(obj),
4713 /* Sigh, the entry has been deleted, and
4714 * it is not easy to revert it back, so
4715 * let's keep this error private, and let
4716 * LFSCK fix it. XXX */
4717 CERROR("%s: delete remote parent "DFID": rc = %d\n",
4718 osd_name(osd), PFID(fid), ret);
4721 LASSERT(osd_invariant(obj));
4722 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4727 * Lookup index for \a key and copy record to \a rec.
4729 * \param dt osd index object
4730 * \param key key for index
4731 * \param rec record reference
4733 * \retval +ve success : exact mach
4734 * \retval 0 return record with key not greater than \a key
4735 * \retval -ve failure
4737 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
4738 struct dt_rec *rec, const struct dt_key *key)
4740 struct osd_object *obj = osd_dt_obj(dt);
4741 struct iam_path_descr *ipd;
4742 struct iam_container *bag = &obj->oo_dir->od_container;
4743 struct osd_thread_info *oti = osd_oti_get(env);
4744 struct iam_iterator *it = &oti->oti_idx_it;
4745 struct iam_rec *iam_rec;
4749 if (!dt_object_exists(dt))
4752 LASSERT(osd_invariant(obj));
4753 LASSERT(!dt_object_remote(dt));
4754 LASSERT(bag->ic_object == obj->oo_inode);
4756 ipd = osd_idx_ipd_get(env, bag);
4760 /* got ipd now we can start iterator. */
4761 iam_it_init(it, bag, 0, ipd);
4763 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4764 /* swab quota uid/gid provided by caller */
4765 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4766 key = (const struct dt_key *)&oti->oti_quota_id;
4769 rc = iam_it_get(it, (struct iam_key *)key);
4771 if (S_ISDIR(obj->oo_inode->i_mode))
4772 iam_rec = (struct iam_rec *)oti->oti_ldp;
4774 iam_rec = (struct iam_rec *) rec;
4776 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
4778 if (S_ISDIR(obj->oo_inode->i_mode))
4779 osd_fid_unpack((struct lu_fid *) rec,
4780 (struct osd_fid_pack *)iam_rec);
4781 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
4782 osd_quota_unpack(obj, rec);
4787 osd_ipd_put(env, bag, ipd);
4789 LINVRNT(osd_invariant(obj));
4794 static int osd_index_declare_iam_insert(const struct lu_env *env,
4795 struct dt_object *dt,
4796 const struct dt_rec *rec,
4797 const struct dt_key *key,
4798 struct thandle *handle)
4800 struct osd_thandle *oh;
4802 LASSERT(handle != NULL);
4804 oh = container_of0(handle, struct osd_thandle, ot_super);
4805 LASSERT(oh->ot_handle == NULL);
4807 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
4808 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
4814 * Inserts (key, value) pair in \a dt index object.
4816 * \param dt osd index object
4817 * \param key key for index
4818 * \param rec record reference
4819 * \param th transaction handler
4822 * \retval -ve failure
4824 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
4825 const struct dt_rec *rec,
4826 const struct dt_key *key, struct thandle *th,
4829 struct osd_object *obj = osd_dt_obj(dt);
4830 struct iam_path_descr *ipd;
4831 struct osd_thandle *oh;
4832 struct iam_container *bag;
4833 struct osd_thread_info *oti = osd_oti_get(env);
4834 struct iam_rec *iam_rec;
4838 if (!dt_object_exists(dt))
4841 LINVRNT(osd_invariant(obj));
4842 LASSERT(!dt_object_remote(dt));
4844 bag = &obj->oo_dir->od_container;
4845 LASSERT(bag->ic_object == obj->oo_inode);
4846 LASSERT(th != NULL);
4848 osd_trans_exec_op(env, th, OSD_OT_INSERT);
4850 ipd = osd_idx_ipd_get(env, bag);
4851 if (unlikely(ipd == NULL))
4854 oh = container_of0(th, struct osd_thandle, ot_super);
4855 LASSERT(oh->ot_handle != NULL);
4856 LASSERT(oh->ot_handle->h_transaction != NULL);
4857 if (S_ISDIR(obj->oo_inode->i_mode)) {
4858 iam_rec = (struct iam_rec *)oti->oti_ldp;
4859 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec, &oti->oti_fid);
4860 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4861 /* pack quota uid/gid */
4862 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4863 key = (const struct dt_key *)&oti->oti_quota_id;
4864 /* pack quota record */
4865 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
4866 iam_rec = (struct iam_rec *)rec;
4868 iam_rec = (struct iam_rec *)rec;
4871 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
4873 osd_ipd_put(env, bag, ipd);
4874 LINVRNT(osd_invariant(obj));
4875 osd_trans_exec_check(env, th, OSD_OT_INSERT);
4880 * Calls ldiskfs_add_entry() to add directory entry
4881 * into the directory. This is required for
4882 * interoperability mode (b11826)
4884 * \retval 0, on success
4885 * \retval -ve, on error
4887 static int __osd_ea_add_rec(struct osd_thread_info *info,
4888 struct osd_object *pobj, struct inode *cinode,
4889 const char *name, const struct lu_fid *fid,
4890 struct htree_lock *hlock, struct thandle *th)
4892 struct ldiskfs_dentry_param *ldp;
4893 struct dentry *child;
4894 struct osd_thandle *oth;
4897 oth = container_of(th, struct osd_thandle, ot_super);
4898 LASSERT(oth->ot_handle != NULL);
4899 LASSERT(oth->ot_handle->h_transaction != NULL);
4900 LASSERT(pobj->oo_inode);
4902 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4903 if (unlikely(pobj->oo_inode ==
4904 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
4907 osd_get_ldiskfs_dirent_param(ldp, fid);
4908 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
4909 child->d_fsdata = (void *)ldp;
4910 ll_vfs_dq_init(pobj->oo_inode);
4911 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
4912 child, cinode, hlock);
4913 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
4914 struct ldiskfs_dir_entry_2 *de;
4915 struct buffer_head *bh;
4918 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
4921 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
4924 if (S_ISDIR(cinode->i_mode))
4925 de->file_type = LDISKFS_DIRENT_LUFID |
4926 LDISKFS_FT_REG_FILE;
4928 de->file_type = LDISKFS_DIRENT_LUFID |
4930 ldiskfs_handle_dirty_metadata(oth->ot_handle,
4941 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
4942 * into the directory.Also sets flags into osd object to
4943 * indicate dot and dotdot are created. This is required for
4944 * interoperability mode (b11826)
4946 * \param dir directory for dot and dotdot fixup.
4947 * \param obj child object for linking
4949 * \retval 0, on success
4950 * \retval -ve, on error
4952 static int osd_add_dot_dotdot(struct osd_thread_info *info,
4953 struct osd_object *dir,
4954 struct inode *parent_dir, const char *name,
4955 const struct lu_fid *dot_fid,
4956 const struct lu_fid *dot_dot_fid,
4959 struct inode *inode = dir->oo_inode;
4960 struct osd_thandle *oth;
4963 oth = container_of(th, struct osd_thandle, ot_super);
4964 LASSERT(oth->ot_handle->h_transaction != NULL);
4965 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
4967 if (strcmp(name, dot) == 0) {
4968 if (dir->oo_compat_dot_created) {
4971 LASSERT(inode->i_ino == parent_dir->i_ino);
4972 dir->oo_compat_dot_created = 1;
4975 } else if (strcmp(name, dotdot) == 0) {
4976 if (!dir->oo_compat_dot_created)
4978 /* in case of rename, dotdot is already created */
4979 if (dir->oo_compat_dotdot_created) {
4980 return __osd_ea_add_rec(info, dir, parent_dir, name,
4981 dot_dot_fid, NULL, th);
4984 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
4985 struct lu_fid tfid = *dot_dot_fid;
4988 result = osd_add_dot_dotdot_internal(info,
4989 dir->oo_inode, parent_dir, dot_fid,
4992 result = osd_add_dot_dotdot_internal(info,
4993 dir->oo_inode, parent_dir, dot_fid,
4998 dir->oo_compat_dotdot_created = 1;
5006 * It will call the appropriate osd_add* function and return the
5007 * value, return by respective functions.
5009 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5010 struct inode *cinode, const char *name,
5011 const struct lu_fid *fid, struct thandle *th)
5013 struct osd_thread_info *info = osd_oti_get(env);
5014 struct htree_lock *hlock;
5017 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5019 if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' &&
5021 if (hlock != NULL) {
5022 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5025 down_write(&pobj->oo_ext_idx_sem);
5028 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5029 lu_object_fid(&pobj->oo_dt.do_lu),
5032 if (hlock != NULL) {
5033 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5034 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5036 down_write(&pobj->oo_ext_idx_sem);
5039 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5040 struct lu_fid *tfid = &info->oti_fid;
5044 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5047 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5052 ldiskfs_htree_unlock(hlock);
5054 up_write(&pobj->oo_ext_idx_sem);
5060 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5061 struct osd_idmap_cache *oic)
5063 struct osd_scrub *scrub = &dev->od_scrub;
5064 struct lu_fid *fid = &oic->oic_fid;
5065 struct osd_inode_id *id = &oic->oic_lid;
5066 struct inode *inode = NULL;
5072 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5075 if (scrub->os_pos_current > id->oii_ino)
5079 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5080 if (rc == -ENOENT) {
5081 __u32 gen = id->oii_gen;
5087 inode = osd_iget(oti, dev, id);
5088 /* The inode has been removed (by race maybe). */
5089 if (IS_ERR(inode)) {
5090 rc = PTR_ERR(inode);
5092 RETURN(rc == -ESTALE ? -ENOENT : rc);
5095 /* The OI mapping is lost. */
5096 if (gen != OSD_OII_NOGEN)
5100 /* The inode may has been reused by others, we do not know,
5101 * leave it to be handled by subsequent osd_fid_lookup(). */
5103 } else if (rc != 0 || osd_id_eq(id, &oti->oti_id)) {
5110 if (thread_is_running(&scrub->os_thread)) {
5111 if (inode == NULL) {
5112 inode = osd_iget(oti, dev, id);
5113 /* The inode has been removed (by race maybe). */
5114 if (IS_ERR(inode)) {
5115 rc = PTR_ERR(inode);
5117 RETURN(rc == -ESTALE ? -ENOENT : rc);
5121 rc = osd_oii_insert(dev, oic, insert);
5122 /* There is race condition between osd_oi_lookup and OI scrub.
5123 * The OI scrub finished just after osd_oi_lookup() failure.
5124 * Under such case, it is unnecessary to trigger OI scrub again,
5125 * but try to call osd_oi_lookup() again. */
5126 if (unlikely(rc == -EAGAIN))
5129 if (!S_ISDIR(inode->i_mode))
5132 rc = osd_check_lmv(oti, dev, inode, oic);
5138 if (!dev->od_noscrub && ++once == 1) {
5139 rc = osd_scrub_start(dev, SS_AUTO_PARTIAL | SS_CLEAR_DRYRUN |
5141 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5142 "%.16s: trigger partial OI scrub for RPC inconsistency "
5143 "checking FID "DFID": rc = %d\n",
5144 LDISKFS_SB(osd_sb(dev))->s_es->s_volume_name,
5146 if (rc == 0 || rc == -EALREADY)
5156 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5157 struct osd_device *dev,
5158 struct osd_idmap_cache *oic,
5159 struct lu_fid *fid, __u32 ino)
5161 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5162 struct inode *inode;
5165 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5166 inode = osd_iget(oti, dev, &oic->oic_lid);
5167 if (IS_ERR(inode)) {
5168 fid_zero(&oic->oic_fid);
5169 return PTR_ERR(inode);
5172 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5175 fid_zero(&oic->oic_fid);
5177 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5181 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5182 struct osd_inode_id *id, const struct lu_fid *fid)
5184 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5185 id->oii_ino, id->oii_gen, info);
5186 info->oti_cache.oic_lid = *id;
5187 info->oti_cache.oic_fid = *fid;
5188 info->oti_cache.oic_dev = osd;
5192 * Get parent FID from the linkEA.
5194 * For a directory which parent resides on remote MDT, to satisfy the
5195 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5196 * the other hand, to make the lookup(..) on the directory can return
5197 * the real parent FID, we append the real parent FID after its ".."
5198 * name entry in the /REMOTE_PARENT_DIR.
5200 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5201 * backup. So after the restore, we cannot get the right parent FID from
5202 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5203 * we have stored the real parent FID in the directory object's linkEA,
5204 * we can parse the linkEA for the real parent FID.
5206 * \param[in] env pointer to the thread context
5207 * \param[in] obj pointer to the object to be handled
5208 * \param[out]fid pointer to the buffer to hold the parent FID
5210 * \retval 0 for getting the real parent FID successfully
5211 * \retval negative error number on failure
5213 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5214 struct osd_object *obj,
5217 struct osd_thread_info *oti = osd_oti_get(env);
5218 struct lu_buf *buf = &oti->oti_big_buf;
5219 struct dentry *dentry = &oti->oti_obj_dentry;
5220 struct inode *inode = obj->oo_inode;
5221 struct linkea_data ldata = { NULL };
5226 if (!S_ISDIR(inode->i_mode))
5230 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5231 buf->lb_buf, buf->lb_len);
5232 if (rc == -ERANGE) {
5233 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5236 lu_buf_realloc(buf, rc);
5237 if (buf->lb_buf == NULL)
5244 if (unlikely(rc == 0))
5250 if (unlikely(buf->lb_buf == NULL)) {
5251 lu_buf_realloc(buf, rc);
5252 if (buf->lb_buf == NULL)
5259 rc = linkea_init_with_rec(&ldata);
5261 linkea_first_entry(&ldata);
5262 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5268 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5269 struct inode *inode,
5270 const struct lu_fid *pfid,
5271 const char *name, const int namelen)
5273 struct osd_thread_info *oti = osd_oti_get(env);
5274 struct lu_buf *buf = &oti->oti_big_buf;
5275 struct dentry *dentry = &oti->oti_obj_dentry;
5276 struct linkea_data ldata = { NULL };
5277 struct lu_name cname = { .ln_name = name,
5278 .ln_namelen = namelen };
5283 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5284 buf->lb_buf, buf->lb_len);
5286 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5291 if (unlikely(rc == 0))
5294 if (buf->lb_len < rc) {
5295 lu_buf_realloc(buf, rc);
5296 if (buf->lb_buf == NULL)
5303 rc = linkea_init_with_rec(&ldata);
5305 rc = linkea_links_find(&ldata, &cname, pfid);
5311 * Calls ->lookup() to find dentry. From dentry get inode and
5312 * read inode's ea to get fid. This is required for interoperability
5315 * \retval 0, on success
5316 * \retval -ve, on error
5318 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5319 struct dt_rec *rec, const struct dt_key *key)
5321 struct inode *dir = obj->oo_inode;
5322 struct dentry *dentry;
5323 struct ldiskfs_dir_entry_2 *de;
5324 struct buffer_head *bh;
5325 struct lu_fid *fid = (struct lu_fid *) rec;
5326 struct htree_lock *hlock = NULL;
5331 LASSERT(dir->i_op != NULL);
5332 LASSERT(dir->i_op->lookup != NULL);
5334 dentry = osd_child_dentry_get(env, obj,
5335 (char *)key, strlen((char *)key));
5337 if (obj->oo_hl_head != NULL) {
5338 hlock = osd_oti_get(env)->oti_hlock;
5339 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5340 dir, LDISKFS_HLOCK_LOOKUP);
5342 down_read(&obj->oo_ext_idx_sem);
5345 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5347 struct osd_thread_info *oti = osd_oti_get(env);
5348 struct osd_inode_id *id = &oti->oti_id;
5349 struct osd_idmap_cache *oic = &oti->oti_cache;
5350 struct osd_device *dev = osd_obj2dev(obj);
5352 ino = le32_to_cpu(de->inode);
5353 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5355 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5359 rc = osd_get_fid_from_dentry(de, rec);
5361 /* done with de, release bh */
5364 if (unlikely(ino == osd_remote_parent_ino(dev))) {
5365 const char *name = (const char *)key;
5367 /* If the parent is on remote MDT, and there
5368 * is no FID-in-dirent, then we have to get
5369 * the parent FID from the linkEA. */
5370 if (likely(strlen(name) == 2 &&
5371 name[0] == '.' && name[1] == '.'))
5372 rc = osd_get_pfid_from_linkea(env, obj,
5375 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5378 osd_id_gen(id, ino, OSD_OII_NOGEN);
5381 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5382 fid_zero(&oic->oic_fid);
5387 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5388 rc = osd_consistency_check(oti, dev, oic);
5390 fid_zero(&oic->oic_fid);
5399 ldiskfs_htree_unlock(hlock);
5401 up_read(&obj->oo_ext_idx_sem);
5405 static int osd_index_declare_ea_insert(const struct lu_env *env,
5406 struct dt_object *dt,
5407 const struct dt_rec *rec,
5408 const struct dt_key *key,
5409 struct thandle *handle)
5411 struct osd_thandle *oh;
5412 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5413 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5414 const struct lu_fid *fid = rec1->rec_fid;
5415 int credits, rc = 0;
5416 struct osd_idmap_cache *idc;
5419 LASSERT(!dt_object_remote(dt));
5420 LASSERT(handle != NULL);
5421 LASSERT(fid != NULL);
5422 LASSERT(rec1->rec_type != 0);
5424 oh = container_of0(handle, struct osd_thandle, ot_super);
5425 LASSERT(oh->ot_handle == NULL);
5427 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5429 /* we can't call iget() while a transactions is running
5430 * (this can lead to a deadlock), but we need to know
5431 * inum and object type. so we find this information at
5432 * declaration and cache in per-thread info */
5433 idc = osd_idc_find_or_init(env, osd, fid);
5435 RETURN(PTR_ERR(idc));
5436 if (idc->oic_remote) {
5437 /* a reference to remote inode is represented by an
5438 * agent inode which we have to create */
5439 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5440 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5443 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5445 if (osd_dt_obj(dt)->oo_inode != NULL) {
5446 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5448 /* We ignore block quota on meta pool (MDTs), so needn't
5449 * calculate how many blocks will be consumed by this index
5451 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5453 i_projid_read(inode), 0,
5454 oh, osd_dt_obj(dt), true,
5462 * Index add function for interoperability mode (b11826).
5463 * It will add the directory entry.This entry is needed to
5464 * maintain name->fid mapping.
5466 * \param key it is key i.e. file entry to be inserted
5467 * \param rec it is value of given key i.e. fid
5469 * \retval 0, on success
5470 * \retval -ve, on error
5472 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5473 const struct dt_rec *rec,
5474 const struct dt_key *key, struct thandle *th,
5477 struct osd_object *obj = osd_dt_obj(dt);
5478 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5479 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5480 const struct lu_fid *fid = rec1->rec_fid;
5481 const char *name = (const char *)key;
5482 struct osd_thread_info *oti = osd_oti_get(env);
5483 struct inode *child_inode = NULL;
5484 struct osd_idmap_cache *idc;
5488 if (!dt_object_exists(dt))
5491 LASSERT(osd_invariant(obj));
5492 LASSERT(!dt_object_remote(dt));
5493 LASSERT(th != NULL);
5495 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5497 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5499 idc = osd_idc_find(env, osd, fid);
5500 if (unlikely(idc == NULL)) {
5501 /* this dt_insert() wasn't declared properly, so
5502 * FID is missing in OI cache. we better do not
5503 * lookup FID in FLDB/OI and don't risk to deadlock,
5504 * but in some special cases (lfsck testing, etc)
5505 * it's much simpler than fixing a caller */
5506 CERROR("%s: "DFID" wasn't declared for insert\n",
5507 osd_name(osd), PFID(fid));
5509 idc = osd_idc_find_or_init(env, osd, fid);
5511 RETURN(PTR_ERR(idc));
5514 if (idc->oic_remote) {
5515 /* Insert remote entry */
5516 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
5517 struct osd_mdobj_map *omm = osd->od_mdt_map;
5518 struct osd_thandle *oh;
5520 /* If parent on remote MDT, we need put this object
5522 oh = container_of(th, typeof(*oh), ot_super);
5523 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5525 CERROR("%s: add "DFID" error: rc = %d\n",
5527 PFID(lu_object_fid(&dt->do_lu)), rc);
5531 child_inode = igrab(omm->omm_remote_parent->d_inode);
5533 child_inode = osd_create_local_agent_inode(env, osd,
5534 obj, fid, rec1->rec_type & S_IFMT, th);
5535 if (IS_ERR(child_inode))
5536 RETURN(PTR_ERR(child_inode));
5539 /* Insert local entry */
5540 if (unlikely(idc->oic_lid.oii_ino == 0)) {
5541 /* for a reason OI cache wasn't filled properly */
5542 CERROR("%s: OIC for "DFID" isn't filled\n",
5543 osd_name(osd), PFID(fid));
5546 child_inode = oti->oti_inode;
5547 if (unlikely(child_inode == NULL)) {
5548 struct ldiskfs_inode_info *lii;
5552 child_inode = oti->oti_inode = &lii->vfs_inode;
5554 child_inode->i_sb = osd_sb(osd);
5555 child_inode->i_ino = idc->oic_lid.oii_ino;
5556 child_inode->i_mode = rec1->rec_type & S_IFMT;
5559 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
5561 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
5562 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
5564 if (child_inode && child_inode != oti->oti_inode)
5566 LASSERT(osd_invariant(obj));
5567 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5572 * Initialize osd Iterator for given osd index object.
5574 * \param dt osd index object
5577 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
5578 struct dt_object *dt,
5581 struct osd_it_iam *it;
5582 struct osd_object *obj = osd_dt_obj(dt);
5583 struct lu_object *lo = &dt->do_lu;
5584 struct iam_path_descr *ipd;
5585 struct iam_container *bag = &obj->oo_dir->od_container;
5587 if (!dt_object_exists(dt))
5588 return ERR_PTR(-ENOENT);
5592 return ERR_PTR(-ENOMEM);
5594 ipd = osd_it_ipd_get(env, bag);
5595 if (likely(ipd != NULL)) {
5599 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
5600 return (struct dt_it *)it;
5603 return ERR_PTR(-ENOMEM);
5608 * free given Iterator.
5610 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
5612 struct osd_it_iam *it = (struct osd_it_iam *)di;
5613 struct osd_object *obj = it->oi_obj;
5615 iam_it_fini(&it->oi_it);
5616 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
5617 osd_object_put(env, obj);
5622 * Move Iterator to record specified by \a key
5624 * \param di osd iterator
5625 * \param key key for index
5627 * \retval +ve di points to record with least key not larger than key
5628 * \retval 0 di points to exact matched key
5629 * \retval -ve failure
5632 static int osd_it_iam_get(const struct lu_env *env,
5633 struct dt_it *di, const struct dt_key *key)
5635 struct osd_thread_info *oti = osd_oti_get(env);
5636 struct osd_it_iam *it = (struct osd_it_iam *)di;
5638 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5639 /* swab quota uid/gid */
5640 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5641 key = (struct dt_key *)&oti->oti_quota_id;
5644 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
5650 * \param di osd iterator
5652 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
5654 struct osd_it_iam *it = (struct osd_it_iam *)di;
5656 iam_it_put(&it->oi_it);
5660 * Move iterator by one record
5662 * \param di osd iterator
5664 * \retval +1 end of container reached
5666 * \retval -ve failure
5669 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
5671 struct osd_it_iam *it = (struct osd_it_iam *)di;
5673 return iam_it_next(&it->oi_it);
5677 * Return pointer to the key under iterator.
5680 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
5681 const struct dt_it *di)
5683 struct osd_thread_info *oti = osd_oti_get(env);
5684 struct osd_it_iam *it = (struct osd_it_iam *)di;
5685 struct osd_object *obj = it->oi_obj;
5688 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
5690 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
5691 /* swab quota uid/gid */
5692 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
5693 key = (struct dt_key *)&oti->oti_quota_id;
5700 * Return size of key under iterator (in bytes)
5703 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
5705 struct osd_it_iam *it = (struct osd_it_iam *)di;
5707 return iam_it_key_size(&it->oi_it);
5711 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
5713 /* check if file type is required */
5714 if (ent->lde_attrs & LUDA_TYPE) {
5715 struct luda_type *lt;
5716 int align = sizeof(*lt) - 1;
5718 len = (len + align) & ~align;
5719 lt = (struct luda_type *)(ent->lde_name + len);
5720 lt->lt_type = cpu_to_le16(DTTOIF(type));
5723 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
5727 * build lu direct from backend fs dirent.
5731 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
5732 char *name, __u16 namelen, __u16 type, __u32 attr)
5734 ent->lde_attrs = attr | LUDA_FID;
5735 fid_cpu_to_le(&ent->lde_fid, fid);
5737 ent->lde_hash = cpu_to_le64(offset);
5738 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
5740 strncpy(ent->lde_name, name, namelen);
5741 ent->lde_name[namelen] = '\0';
5742 ent->lde_namelen = cpu_to_le16(namelen);
5744 /* append lustre attributes */
5745 osd_it_append_attrs(ent, namelen, type);
5749 * Return pointer to the record under iterator.
5751 static int osd_it_iam_rec(const struct lu_env *env,
5752 const struct dt_it *di,
5753 struct dt_rec *dtrec, __u32 attr)
5755 struct osd_it_iam *it = (struct osd_it_iam *)di;
5756 struct osd_thread_info *info = osd_oti_get(env);
5759 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
5760 const struct osd_fid_pack *rec;
5761 struct lu_fid *fid = &info->oti_fid;
5762 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
5768 name = (char *)iam_it_key_get(&it->oi_it);
5770 RETURN(PTR_ERR(name));
5772 namelen = iam_it_key_size(&it->oi_it);
5774 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
5776 RETURN(PTR_ERR(rec));
5778 rc = osd_fid_unpack(fid, rec);
5782 hash = iam_it_store(&it->oi_it);
5784 /* IAM does not store object type in IAM index (dir) */
5785 osd_it_pack_dirent(lde, fid, hash, name, namelen,
5787 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5788 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5789 (struct iam_rec *)dtrec);
5790 osd_quota_unpack(it->oi_obj, dtrec);
5792 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5793 (struct iam_rec *)dtrec);
5800 * Returns cookie for current Iterator position.
5802 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
5804 struct osd_it_iam *it = (struct osd_it_iam *)di;
5806 return iam_it_store(&it->oi_it);
5810 * Restore iterator from cookie.
5812 * \param di osd iterator
5813 * \param hash Iterator location cookie
5815 * \retval +ve di points to record with least key not larger than key.
5816 * \retval 0 di points to exact matched key
5817 * \retval -ve failure
5820 static int osd_it_iam_load(const struct lu_env *env,
5821 const struct dt_it *di, __u64 hash)
5823 struct osd_it_iam *it = (struct osd_it_iam *)di;
5825 return iam_it_load(&it->oi_it, hash);
5828 static const struct dt_index_operations osd_index_iam_ops = {
5829 .dio_lookup = osd_index_iam_lookup,
5830 .dio_declare_insert = osd_index_declare_iam_insert,
5831 .dio_insert = osd_index_iam_insert,
5832 .dio_declare_delete = osd_index_declare_iam_delete,
5833 .dio_delete = osd_index_iam_delete,
5835 .init = osd_it_iam_init,
5836 .fini = osd_it_iam_fini,
5837 .get = osd_it_iam_get,
5838 .put = osd_it_iam_put,
5839 .next = osd_it_iam_next,
5840 .key = osd_it_iam_key,
5841 .key_size = osd_it_iam_key_size,
5842 .rec = osd_it_iam_rec,
5843 .store = osd_it_iam_store,
5844 .load = osd_it_iam_load
5850 * Creates or initializes iterator context.
5852 * \retval struct osd_it_ea, iterator structure on success
5855 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
5856 struct dt_object *dt,
5859 struct osd_object *obj = osd_dt_obj(dt);
5860 struct osd_thread_info *info = osd_oti_get(env);
5861 struct osd_it_ea *oie;
5863 struct lu_object *lo = &dt->do_lu;
5864 struct dentry *obj_dentry;
5867 if (!dt_object_exists(dt) || obj->oo_destroyed)
5868 RETURN(ERR_PTR(-ENOENT));
5870 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
5872 RETURN(ERR_PTR(-ENOMEM));
5873 obj_dentry = &oie->oie_dentry;
5875 obj_dentry->d_inode = obj->oo_inode;
5876 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
5877 obj_dentry->d_name.hash = 0;
5879 oie->oie_rd_dirent = 0;
5880 oie->oie_it_dirent = 0;
5881 oie->oie_dirent = NULL;
5882 if (unlikely(!info->oti_it_ea_buf_used)) {
5883 oie->oie_buf = info->oti_it_ea_buf;
5884 info->oti_it_ea_buf_used = 1;
5886 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5887 if (oie->oie_buf == NULL)
5888 RETURN(ERR_PTR(-ENOMEM));
5892 file = &oie->oie_file;
5894 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
5895 if (attr & LUDA_64BITHASH)
5896 file->f_mode = FMODE_64BITHASH;
5898 file->f_mode = FMODE_32BITHASH;
5899 file->f_path.dentry = obj_dentry;
5900 file->f_mapping = obj->oo_inode->i_mapping;
5901 file->f_op = obj->oo_inode->i_fop;
5902 set_file_inode(file, obj->oo_inode);
5905 RETURN((struct dt_it *) oie);
5909 * Destroy or finishes iterator context.
5911 * \param di iterator structure to be destroyed
5913 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
5915 struct osd_thread_info *info = osd_oti_get(env);
5916 struct osd_it_ea *oie = (struct osd_it_ea *)di;
5917 struct osd_object *obj = oie->oie_obj;
5918 struct inode *inode = obj->oo_inode;
5921 oie->oie_file.f_op->release(inode, &oie->oie_file);
5922 osd_object_put(env, obj);
5923 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
5924 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5926 info->oti_it_ea_buf_used = 0;
5927 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
5932 * It position the iterator at given key, so that next lookup continues from
5933 * that key Or it is similar to dio_it->load() but based on a key,
5934 * rather than file position.
5936 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
5939 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
5941 static int osd_it_ea_get(const struct lu_env *env,
5942 struct dt_it *di, const struct dt_key *key)
5944 struct osd_it_ea *it = (struct osd_it_ea *)di;
5947 LASSERT(((const char *)key)[0] == '\0');
5948 it->oie_file.f_pos = 0;
5949 it->oie_rd_dirent = 0;
5950 it->oie_it_dirent = 0;
5951 it->oie_dirent = NULL;
5959 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
5963 struct osd_filldir_cbs {
5964 #ifdef HAVE_DIR_CONTEXT
5965 struct dir_context ctx;
5967 struct osd_it_ea *it;
5970 * It is called internally by ->readdir(). It fills the
5971 * iterator's in-memory data structure with required
5972 * information i.e. name, namelen, rec_size etc.
5974 * \param buf in which information to be filled in.
5975 * \param name name of the file in given dir
5977 * \retval 0 on success
5978 * \retval 1 on buffer full
5980 #ifdef HAVE_FILLDIR_USE_CTX
5981 static int osd_ldiskfs_filldir(struct dir_context *buf,
5983 static int osd_ldiskfs_filldir(void *buf,
5985 const char *name, int namelen,
5986 loff_t offset, __u64 ino, unsigned d_type)
5988 struct osd_it_ea *it =
5989 ((struct osd_filldir_cbs *)buf)->it;
5990 struct osd_object *obj = it->oie_obj;
5991 struct osd_it_ea_dirent *ent = it->oie_dirent;
5992 struct lu_fid *fid = &ent->oied_fid;
5993 struct osd_fid_pack *rec;
5996 /* this should never happen */
5997 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
5998 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6002 if ((void *) ent - it->oie_buf + sizeof(*ent) + namelen >
6006 /* "." is just the object itself. */
6007 if (namelen == 1 && name[0] == '.') {
6008 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6009 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6010 rec = (struct osd_fid_pack*) (name + namelen + 1);
6011 if (osd_fid_unpack(fid, rec) != 0)
6016 d_type &= ~LDISKFS_DIRENT_LUFID;
6018 /* NOT export local root. */
6019 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6020 ino = obj->oo_inode->i_ino;
6021 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6024 ent->oied_ino = ino;
6025 ent->oied_off = offset;
6026 ent->oied_namelen = namelen;
6027 ent->oied_type = d_type;
6029 memcpy(ent->oied_name, name, namelen);
6031 it->oie_rd_dirent++;
6032 it->oie_dirent = (void *) ent + cfs_size_round(sizeof(*ent) + namelen);
6037 * Calls ->readdir() to load a directory entry at a time
6038 * and stored it in iterator's in-memory data structure.
6040 * \param di iterator's in memory structure
6042 * \retval 0 on success
6043 * \retval -ve on error
6044 * \retval +1 reach the end of entry
6046 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6047 const struct dt_it *di)
6049 struct osd_it_ea *it = (struct osd_it_ea *)di;
6050 struct osd_object *obj = it->oie_obj;
6051 struct inode *inode = obj->oo_inode;
6052 struct htree_lock *hlock = NULL;
6053 struct file *filp = &it->oie_file;
6055 struct osd_filldir_cbs buf = {
6056 #ifdef HAVE_DIR_CONTEXT
6057 .ctx.actor = osd_ldiskfs_filldir,
6063 it->oie_dirent = it->oie_buf;
6064 it->oie_rd_dirent = 0;
6066 if (obj->oo_hl_head != NULL) {
6067 hlock = osd_oti_get(env)->oti_hlock;
6068 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6069 inode, LDISKFS_HLOCK_READDIR);
6071 down_read(&obj->oo_ext_idx_sem);
6074 #ifdef HAVE_DIR_CONTEXT
6075 buf.ctx.pos = filp->f_pos;
6076 rc = inode->i_fop->iterate(filp, &buf.ctx);
6077 filp->f_pos = buf.ctx.pos;
6079 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6083 ldiskfs_htree_unlock(hlock);
6085 up_read(&obj->oo_ext_idx_sem);
6087 if (it->oie_rd_dirent == 0) {
6088 /*If it does not get any dirent, it means it has been reached
6089 *to the end of the dir */
6090 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6094 it->oie_dirent = it->oie_buf;
6095 it->oie_it_dirent = 1;
6102 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6103 * to load a directory entry at a time and stored it in
6104 * iterator's in-memory data structure.
6106 * \param di iterator's in memory structure
6108 * \retval +ve iterator reached to end
6109 * \retval 0 iterator not reached to end
6110 * \retval -ve on error
6112 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6114 struct osd_it_ea *it = (struct osd_it_ea *)di;
6119 if (it->oie_it_dirent < it->oie_rd_dirent) {
6121 (void *) it->oie_dirent +
6122 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6123 it->oie_dirent->oied_namelen);
6124 it->oie_it_dirent++;
6127 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6130 rc = osd_ldiskfs_it_fill(env, di);
6137 * Returns the key at current position from iterator's in memory structure.
6139 * \param di iterator's in memory structure
6141 * \retval key i.e. struct dt_key on success
6143 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6144 const struct dt_it *di)
6146 struct osd_it_ea *it = (struct osd_it_ea *)di;
6148 return (struct dt_key *)it->oie_dirent->oied_name;
6152 * Returns the key's size at current position from iterator's in memory structure.
6154 * \param di iterator's in memory structure
6156 * \retval key_size i.e. struct dt_key on success
6158 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6160 struct osd_it_ea *it = (struct osd_it_ea *)di;
6162 return it->oie_dirent->oied_namelen;
6165 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6167 if (LDISKFS_DIR_REC_LEN(de) >=
6168 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6175 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6176 unsigned blocksize, bool dotdot)
6179 return osd_dotdot_has_space(de);
6181 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6182 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6189 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6190 handle_t *jh, struct dentry *dentry,
6191 const struct lu_fid *fid, struct buffer_head *bh,
6192 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6195 struct inode *dir = dentry->d_parent->d_inode;
6196 struct inode *inode = dentry->d_inode;
6197 struct osd_fid_pack *rec;
6198 struct ldiskfs_dentry_param *ldp;
6199 int namelen = dentry->d_name.len;
6201 struct osd_thread_info *info = osd_oti_get(env);
6204 if (!LDISKFS_HAS_INCOMPAT_FEATURE(inode->i_sb,
6205 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
6208 /* There is enough space to hold the FID-in-dirent. */
6209 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6210 rc = ldiskfs_journal_get_write_access(jh, bh);
6214 de->name[namelen] = 0;
6215 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6216 rec->fp_len = sizeof(struct lu_fid) + 1;
6217 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6218 de->file_type |= LDISKFS_DIRENT_LUFID;
6219 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6226 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6230 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6231 osd_get_ldiskfs_dirent_param(ldp, fid);
6232 dentry->d_fsdata = (void *)ldp;
6233 ll_vfs_dq_init(dir);
6234 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6235 /* It is too bad, we cannot reinsert the name entry back.
6236 * That means we lose it! */
6238 CDEBUG(D_LFSCK, "%.16s: fail to reinsert the dirent, "
6239 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6240 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
6241 dir->i_ino, dir->i_generation, namelen,
6242 dentry->d_name.name, PFID(fid), rc);
6248 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6249 struct osd_it_ea *it, struct lu_fid *fid,
6250 struct osd_inode_id *id, __u32 *attr)
6252 struct osd_thread_info *info = osd_oti_get(env);
6253 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6254 struct osd_device *dev = osd_obj2dev(obj);
6255 struct super_block *sb = osd_sb(dev);
6256 const char *devname = osd_name(dev);
6257 struct osd_it_ea_dirent *ent = it->oie_dirent;
6258 struct inode *dir = obj->oo_inode;
6259 struct htree_lock *hlock = NULL;
6260 struct buffer_head *bh = NULL;
6261 handle_t *jh = NULL;
6262 struct ldiskfs_dir_entry_2 *de;
6263 struct dentry *dentry;
6264 struct inode *inode;
6265 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6268 bool dotdot = false;
6272 if (ent->oied_name[0] == '.') {
6273 if (ent->oied_namelen == 1)
6276 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6280 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6281 inode = osd_iget(info, dev, id);
6282 if (IS_ERR(inode)) {
6283 rc = PTR_ERR(inode);
6284 if (rc == -ENOENT || rc == -ESTALE) {
6285 /* Maybe dangling name entry, or
6286 * corrupted directory entry. */
6287 *attr |= LUDA_UNKNOWN;
6290 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6291 "check_repair, dir = %lu/%u, name = %.*s, "
6292 "ino = %llu, rc = %d\n",
6293 devname, dir->i_ino, dir->i_generation,
6294 ent->oied_namelen, ent->oied_name,
6301 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6303 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6304 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6309 /* We need to ensure that the name entry is still valid.
6310 * Because it may be removed or renamed by other already.
6312 * The unlink or rename operation will start journal before PDO lock,
6313 * so to avoid deadlock, here we need to start journal handle before
6314 * related PDO lock also. But because we do not know whether there
6315 * will be something to be repaired before PDO lock, we just start
6316 * journal without conditions.
6318 * We may need to remove the name entry firstly, then insert back.
6319 * One credit is for user quota file update.
6320 * One credit is for group quota file update.
6321 * Two credits are for dirty inode. */
6322 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6323 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6325 if (dev->od_dirent_journal != 0) {
6328 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6331 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6332 "check_repair, dir = %lu/%u, credits = %d, "
6333 "name = %.*s, ino = %llu: rc = %d\n",
6334 devname, dir->i_ino, dir->i_generation, credits,
6335 ent->oied_namelen, ent->oied_name,
6338 GOTO(out_inode, rc);
6341 if (obj->oo_hl_head != NULL) {
6342 hlock = osd_oti_get(env)->oti_hlock;
6343 /* "0" means exclusive lock for the whole directory.
6344 * We need to prevent others access such name entry
6345 * during the delete + insert. Neither HLOCK_ADD nor
6346 * HLOCK_DEL cannot guarantee the atomicity. */
6347 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6349 down_write(&obj->oo_ext_idx_sem);
6352 if (obj->oo_hl_head != NULL) {
6353 hlock = osd_oti_get(env)->oti_hlock;
6354 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6355 LDISKFS_HLOCK_LOOKUP);
6357 down_read(&obj->oo_ext_idx_sem);
6361 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6362 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6363 *attr |= LUDA_IGNORE;
6368 /* For dotdot entry, if there is not enough space to hold the
6369 * FID-in-dirent, just keep them there. It only happens when the
6370 * device upgraded from 1.8 or restored from MDT file-level backup.
6371 * For the whole directory, only dotdot entry have no FID-in-dirent
6372 * and needs to get FID from LMA when readdir, it will not affect the
6373 * performance much. */
6374 if (dotdot && !osd_dotdot_has_space(de)) {
6375 *attr |= LUDA_UNKNOWN;
6381 if (lu_fid_eq(fid, &lma->lma_self_fid))
6384 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6385 struct lu_fid *tfid = &lma->lma_self_fid;
6387 if (likely(dotdot &&
6388 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6389 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6390 /* It must be REMOTE_PARENT_DIR and as the
6391 * 'dotdot' entry of remote directory */
6392 *attr |= LUDA_IGNORE;
6394 CDEBUG(D_LFSCK, "%s: expect remote agent "
6395 "parent directory, but got %.*s under "
6396 "dir = %lu/%u with the FID "DFID"\n",
6397 devname, ent->oied_namelen,
6398 ent->oied_name, dir->i_ino,
6399 dir->i_generation, PFID(tfid));
6401 *attr |= LUDA_UNKNOWN;
6408 if (!fid_is_zero(fid)) {
6409 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6411 if (rc == -ENOENT ||
6413 !(dev->od_scrub.os_file.sf_flags & SF_UPGRADE))) {
6414 /* linkEA does not recognize the dirent entry,
6415 * it may because the dirent entry corruption
6416 * and points to other's inode. */
6417 CDEBUG(D_LFSCK, "%s: the target inode does not "
6418 "recognize the dirent, dir = %lu/%u, "
6419 " name = %.*s, ino = %llu, "
6420 DFID": rc = %d\n", devname, dir->i_ino,
6421 dir->i_generation, ent->oied_namelen,
6422 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6423 *attr |= LUDA_UNKNOWN;
6428 if (rc && rc != -ENODATA) {
6429 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6430 "dir = %lu/%u, name = %.*s, ino = %llu, "
6431 DFID": rc = %d\n", devname, dir->i_ino,
6432 dir->i_generation, ent->oied_namelen,
6433 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6434 *attr |= LUDA_UNKNOWN;
6441 /* linkEA recognizes the dirent entry, the FID-in-LMA is
6442 * valid, trusted, in spite of fid_is_sane(fid) or not. */
6443 if (*attr & LUDA_VERIFY_DRYRUN) {
6444 *fid = lma->lma_self_fid;
6445 *attr |= LUDA_REPAIR;
6452 dev->od_dirent_journal = 1;
6453 if (hlock != NULL) {
6454 ldiskfs_htree_unlock(hlock);
6457 up_read(&obj->oo_ext_idx_sem);
6463 *fid = lma->lma_self_fid;
6465 /* Update or append the FID-in-dirent. */
6466 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6467 bh, de, hlock, dotdot);
6469 *attr |= LUDA_REPAIR;
6471 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6472 "the dirent, dir = %lu/%u, name = %.*s, "
6473 "ino = %llu, "DFID": rc = %d\n",
6474 devname, dir->i_ino, dir->i_generation,
6475 ent->oied_namelen, ent->oied_name,
6476 ent->oied_ino, PFID(fid), rc);
6478 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6479 if (*attr & LUDA_VERIFY_DRYRUN) {
6480 if (fid_is_sane(fid)) {
6481 *attr |= LUDA_REPAIR;
6482 } else if (dev->od_index == 0) {
6483 lu_igif_build(fid, inode->i_ino,
6484 inode->i_generation);
6485 *attr |= LUDA_UPGRADE;
6493 dev->od_dirent_journal = 1;
6494 if (hlock != NULL) {
6495 ldiskfs_htree_unlock(hlock);
6498 up_read(&obj->oo_ext_idx_sem);
6505 if (unlikely(fid_is_sane(fid))) {
6506 /* FID-in-dirent exists, but FID-in-LMA is lost.
6507 * Trust the FID-in-dirent, and add FID-in-LMA. */
6508 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
6510 *attr |= LUDA_REPAIR;
6512 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
6513 "update dirent, dir = %lu/%u, "
6514 "name = %.*s, ino = %llu, "
6516 devname, dir->i_ino, dir->i_generation,
6517 ent->oied_namelen, ent->oied_name,
6518 ent->oied_ino, PFID(fid), rc);
6519 } else if (dev->od_index == 0) {
6520 lu_igif_build(fid, inode->i_ino, inode->i_generation);
6521 /* It is probably IGIF object. Only aappend the
6522 * FID-in-dirent. OI scrub will process FID-in-LMA. */
6523 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6524 bh, de, hlock, dotdot);
6526 *attr |= LUDA_UPGRADE;
6528 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
6529 "after the dirent, dir = %lu/%u, "
6530 "name = %.*s, ino = %llu, "
6532 devname, dir->i_ino, dir->i_generation,
6533 ent->oied_namelen, ent->oied_name,
6534 ent->oied_ino, PFID(fid), rc);
6543 if (hlock != NULL) {
6544 ldiskfs_htree_unlock(hlock);
6546 if (dev->od_dirent_journal != 0)
6547 up_write(&obj->oo_ext_idx_sem);
6549 up_read(&obj->oo_ext_idx_sem);
6553 ldiskfs_journal_stop(jh);
6557 if (rc >= 0 && !dirty)
6558 dev->od_dirent_journal = 0;
6564 * Returns the value at current position from iterator's in memory structure.
6566 * \param di struct osd_it_ea, iterator's in memory structure
6567 * \param attr attr requested for dirent.
6568 * \param lde lustre dirent
6570 * \retval 0 no error and \param lde has correct lustre dirent.
6571 * \retval -ve on error
6573 static inline int osd_it_ea_rec(const struct lu_env *env,
6574 const struct dt_it *di,
6575 struct dt_rec *dtrec, __u32 attr)
6577 struct osd_it_ea *it = (struct osd_it_ea *)di;
6578 struct osd_object *obj = it->oie_obj;
6579 struct osd_device *dev = osd_obj2dev(obj);
6580 struct osd_thread_info *oti = osd_oti_get(env);
6581 struct osd_inode_id *id = &oti->oti_id;
6582 struct lu_fid *fid = &it->oie_dirent->oied_fid;
6583 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6584 __u32 ino = it->oie_dirent->oied_ino;
6588 LASSERT(obj->oo_inode != dev->od_mdt_map->omm_remote_parent->d_inode);
6590 if (attr & LUDA_VERIFY) {
6591 if (unlikely(ino == osd_remote_parent_ino(dev))) {
6592 attr |= LUDA_IGNORE;
6593 /* If the parent is on remote MDT, and there
6594 * is no FID-in-dirent, then we have to get
6595 * the parent FID from the linkEA. */
6596 if (!fid_is_sane(fid) &&
6597 it->oie_dirent->oied_namelen == 2 &&
6598 it->oie_dirent->oied_name[0] == '.' &&
6599 it->oie_dirent->oied_name[1] == '.')
6600 osd_get_pfid_from_linkea(env, obj, fid);
6602 rc = osd_dirent_check_repair(env, obj, it, fid, id,
6606 if (!fid_is_sane(fid))
6607 attr |= LUDA_UNKNOWN;
6609 attr &= ~LU_DIRENT_ATTRS_MASK;
6610 if (!fid_is_sane(fid)) {
6611 bool is_dotdot = false;
6612 if (it->oie_dirent->oied_namelen == 2 &&
6613 it->oie_dirent->oied_name[0] == '.' &&
6614 it->oie_dirent->oied_name[1] == '.')
6616 /* If the parent is on remote MDT, and there
6617 * is no FID-in-dirent, then we have to get
6618 * the parent FID from the linkEA. */
6619 if (ino == osd_remote_parent_ino(dev) && is_dotdot) {
6620 rc = osd_get_pfid_from_linkea(env, obj, fid);
6622 if (is_dotdot == false &&
6623 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
6626 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6629 osd_id_gen(id, ino, OSD_OII_NOGEN);
6633 /* Pack the entry anyway, at least the offset is right. */
6634 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
6635 it->oie_dirent->oied_name,
6636 it->oie_dirent->oied_namelen,
6637 it->oie_dirent->oied_type, attr);
6642 if (osd_remote_fid(env, dev, fid))
6645 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
6646 osd_add_oi_cache(oti, dev, id, fid);
6648 RETURN(rc > 0 ? 0 : rc);
6652 * Returns the record size size at current position.
6654 * This function will return record(lu_dirent) size in bytes.
6656 * \param[in] env execution environment
6657 * \param[in] di iterator's in memory structure
6658 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
6659 * calculate the lu_dirent size.
6661 * \retval record size(in bytes & in memory) of the current lu_dirent
6664 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
6667 struct osd_it_ea *it = (struct osd_it_ea *)di;
6669 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
6673 * Returns a cookie for current position of the iterator head, so that
6674 * user can use this cookie to load/start the iterator next time.
6676 * \param di iterator's in memory structure
6678 * \retval cookie for current position, on success
6680 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
6682 struct osd_it_ea *it = (struct osd_it_ea *)di;
6684 return it->oie_dirent->oied_off;
6688 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6689 * to load a directory entry at a time and stored it i inn,
6690 * in iterator's in-memory data structure.
6692 * \param di struct osd_it_ea, iterator's in memory structure
6694 * \retval +ve on success
6695 * \retval -ve on error
6697 static int osd_it_ea_load(const struct lu_env *env,
6698 const struct dt_it *di, __u64 hash)
6700 struct osd_it_ea *it = (struct osd_it_ea *)di;
6704 it->oie_file.f_pos = hash;
6706 rc = osd_ldiskfs_it_fill(env, di);
6717 * Index lookup function for interoperability mode (b11826).
6719 * \param key, key i.e. file name to be searched
6721 * \retval +ve, on success
6722 * \retval -ve, on error
6724 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
6725 struct dt_rec *rec, const struct dt_key *key)
6727 struct osd_object *obj = osd_dt_obj(dt);
6732 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
6733 LINVRNT(osd_invariant(obj));
6735 rc = osd_ea_lookup_rec(env, obj, rec, key);
6742 * Index and Iterator operations for interoperability
6743 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
6745 static const struct dt_index_operations osd_index_ea_ops = {
6746 .dio_lookup = osd_index_ea_lookup,
6747 .dio_declare_insert = osd_index_declare_ea_insert,
6748 .dio_insert = osd_index_ea_insert,
6749 .dio_declare_delete = osd_index_declare_ea_delete,
6750 .dio_delete = osd_index_ea_delete,
6752 .init = osd_it_ea_init,
6753 .fini = osd_it_ea_fini,
6754 .get = osd_it_ea_get,
6755 .put = osd_it_ea_put,
6756 .next = osd_it_ea_next,
6757 .key = osd_it_ea_key,
6758 .key_size = osd_it_ea_key_size,
6759 .rec = osd_it_ea_rec,
6760 .rec_size = osd_it_ea_rec_size,
6761 .store = osd_it_ea_store,
6762 .load = osd_it_ea_load
6766 static void *osd_key_init(const struct lu_context *ctx,
6767 struct lu_context_key *key)
6769 struct osd_thread_info *info;
6771 OBD_ALLOC_PTR(info);
6773 return ERR_PTR(-ENOMEM);
6775 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6776 if (info->oti_it_ea_buf == NULL)
6779 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
6781 info->oti_hlock = ldiskfs_htree_lock_alloc();
6782 if (info->oti_hlock == NULL)
6788 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6791 return ERR_PTR(-ENOMEM);
6794 static void osd_key_fini(const struct lu_context *ctx,
6795 struct lu_context_key *key, void* data)
6797 struct osd_thread_info *info = data;
6798 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
6799 struct osd_idmap_cache *idc = info->oti_ins_cache;
6801 if (info->oti_inode != NULL)
6803 if (info->oti_hlock != NULL)
6804 ldiskfs_htree_lock_free(info->oti_hlock);
6805 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6806 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
6807 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
6808 lu_buf_free(&info->oti_big_buf);
6810 LASSERT(info->oti_ins_cache_size > 0);
6811 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
6812 info->oti_ins_cache = NULL;
6813 info->oti_ins_cache_size = 0;
6818 static void osd_key_exit(const struct lu_context *ctx,
6819 struct lu_context_key *key, void *data)
6821 struct osd_thread_info *info = data;
6823 LASSERT(info->oti_r_locks == 0);
6824 LASSERT(info->oti_w_locks == 0);
6825 LASSERT(info->oti_txns == 0);
6828 /* type constructor/destructor: osd_type_init, osd_type_fini */
6829 LU_TYPE_INIT_FINI(osd, &osd_key);
6831 struct lu_context_key osd_key = {
6832 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
6833 .lct_init = osd_key_init,
6834 .lct_fini = osd_key_fini,
6835 .lct_exit = osd_key_exit
6839 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
6840 const char *name, struct lu_device *next)
6842 struct osd_device *osd = osd_dev(d);
6844 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname))
6845 >= sizeof(osd->od_svname))
6847 return osd_procfs_init(osd, name);
6850 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
6852 struct seq_server_site *ss = osd_seq_site(osd);
6856 if (osd->od_is_ost || osd->od_cl_seq != NULL)
6859 if (unlikely(ss == NULL))
6862 OBD_ALLOC_PTR(osd->od_cl_seq);
6863 if (osd->od_cl_seq == NULL)
6866 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
6867 osd->od_svname, ss->ss_server_seq);
6869 OBD_FREE_PTR(osd->od_cl_seq);
6870 osd->od_cl_seq = NULL;
6874 if (ss->ss_node_id == 0) {
6875 /* If the OSD on the sequence controller(MDT0), then allocate
6876 * sequence here, otherwise allocate sequence after connected
6877 * to MDT0 (see mdt_register_lwp_callback()). */
6878 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
6879 &osd->od_cl_seq->lcs_space, env);
6885 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
6887 if (osd->od_cl_seq == NULL)
6890 seq_client_fini(osd->od_cl_seq);
6891 OBD_FREE_PTR(osd->od_cl_seq);
6892 osd->od_cl_seq = NULL;
6895 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
6899 /* shutdown quota slave instance associated with the device */
6900 if (o->od_quota_slave != NULL) {
6901 qsd_fini(env, o->od_quota_slave);
6902 o->od_quota_slave = NULL;
6905 osd_fid_fini(env, o);
6910 static void osd_umount(const struct lu_env *env, struct osd_device *o)
6914 if (o->od_mnt != NULL) {
6915 shrink_dcache_sb(osd_sb(o));
6916 osd_sync(env, &o->od_dt_dev);
6925 static int osd_mount(const struct lu_env *env,
6926 struct osd_device *o, struct lustre_cfg *cfg)
6928 const char *name = lustre_cfg_string(cfg, 0);
6929 const char *dev = lustre_cfg_string(cfg, 1);
6931 unsigned long page, s_flags, lmd_flags = 0;
6932 struct page *__page;
6933 struct file_system_type *type;
6934 char *options = NULL;
6936 struct osd_thread_info *info = osd_oti_get(env);
6937 struct lu_fid *fid = &info->oti_fid;
6938 struct inode *inode;
6939 int rc = 0, force_over_512tb = 0;
6942 if (o->od_mnt != NULL)
6945 if (strlen(dev) >= sizeof(o->od_mntdev))
6947 strcpy(o->od_mntdev, dev);
6949 str = lustre_cfg_string(cfg, 2);
6950 s_flags = simple_strtoul(str, NULL, 0);
6951 str = strstr(str, ":");
6953 lmd_flags = simple_strtoul(str + 1, NULL, 0);
6954 opts = lustre_cfg_string(cfg, 3);
6956 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
6957 CERROR("%s: device %s extents feature is not guaranteed to "
6958 "work on big-endian systems. Use \"bigendian_extents\" "
6959 "mount option to override.\n", name, dev);
6963 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
6964 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
6965 CWARN("force_over_128tb option is deprecated. "
6966 "Filesystems less than 512TB can be created without any "
6967 "force options. Use force_over_512tb option for "
6968 "filesystems greater than 512TB.\n");
6971 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
6972 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
6973 CWARN("force_over_256tb option is deprecated. "
6974 "Filesystems less than 512TB can be created without any "
6975 "force options. Use force_over_512tb option for "
6976 "filesystems greater than 512TB.\n");
6980 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
6981 force_over_512tb = 1;
6983 __page = alloc_page(GFP_KERNEL);
6985 GOTO(out, rc = -ENOMEM);
6986 page = (unsigned long)page_address(__page);
6987 options = (char *)page;
6990 /* strip out the options for back compatiblity */
6991 static char *sout[] = {
6998 /* strip out option we processed in osd */
6999 "bigendian_extents",
7005 strcat(options, opts);
7006 for (rc = 0, str = options; sout[rc]; ) {
7007 char *op = strstr(str, sout[rc]);
7013 if (op == options || *(op - 1) == ',') {
7014 str = op + strlen(sout[rc]);
7015 if (*str == ',' || *str == '\0') {
7016 *str == ',' ? str++ : str;
7017 memmove(op, str, strlen(str) + 1);
7020 for (str = op; *str != ',' && *str != '\0'; str++)
7024 strncat(options, "user_xattr,acl", 14);
7027 /* Glom up mount options */
7028 if (*options != '\0')
7029 strcat(options, ",");
7030 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7032 type = get_fs_type("ldiskfs");
7034 CERROR("%s: cannot find ldiskfs module\n", name);
7035 GOTO(out, rc = -ENODEV);
7038 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7039 module_put(type->owner);
7041 if (IS_ERR(o->od_mnt)) {
7042 rc = PTR_ERR(o->od_mnt);
7044 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7048 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7049 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7050 force_over_512tb == 0) {
7051 CERROR("%s: device %s LDISKFS does not support filesystems "
7052 "greater than 512TB and can cause data corruption. "
7053 "Use \"force_over_512tb\" mount option to override.\n",
7055 GOTO(out_mnt, rc = -EINVAL);
7058 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7059 #ifdef HAVE_DEV_SET_RDONLY
7060 dev_set_rdonly(osd_sb(o)->s_bdev);
7061 o->od_dt_dev.dd_rdonly = 1;
7062 LCONSOLE_WARN("%s: set dev_rdonly on this device\n", name);
7064 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7067 GOTO(out_mnt, rc = -EOPNOTSUPP);
7070 #ifdef HAVE_DEV_SET_RDONLY
7071 if (dev_check_rdonly(osd_sb(o)->s_bdev)) {
7072 CERROR("%s: underlying device %s is marked as "
7073 "read-only. Setup failed\n", name, dev);
7075 GOTO(out_mnt, rc = -EROFS);
7080 if (!LDISKFS_HAS_COMPAT_FEATURE(o->od_mnt->mnt_sb,
7081 LDISKFS_FEATURE_COMPAT_HAS_JOURNAL)) {
7082 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7083 GOTO(out_mnt, rc = -EINVAL);
7086 #ifdef LDISKFS_MOUNT_DIRDATA
7087 if (LDISKFS_HAS_INCOMPAT_FEATURE(o->od_mnt->mnt_sb,
7088 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
7089 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7090 else if (!o->od_is_ost)
7091 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7092 "enabling the dirdata feature. If you do not want to "
7093 "downgrade to Lustre-1.x again, you can enable it via "
7094 "'tune2fs -O dirdata device'\n", name, dev);
7096 inode = osd_sb(o)->s_root->d_inode;
7097 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7098 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7100 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7104 if (lmd_flags & LMD_FLG_NOSCRUB)
7115 __free_page(__page);
7120 static struct lu_device *osd_device_fini(const struct lu_env *env,
7121 struct lu_device *d)
7123 struct osd_device *o = osd_dev(d);
7126 osd_shutdown(env, o);
7128 osd_scrub_cleanup(env, o);
7129 osd_obj_map_fini(o);
7135 static int osd_device_init0(const struct lu_env *env,
7136 struct osd_device *o,
7137 struct lustre_cfg *cfg)
7139 struct lu_device *l = osd2lu_dev(o);
7140 struct osd_thread_info *info;
7144 /* if the module was re-loaded, env can loose its keys */
7145 rc = lu_env_refill((struct lu_env *) env);
7148 info = osd_oti_get(env);
7151 l->ld_ops = &osd_lu_ops;
7152 o->od_dt_dev.dd_ops = &osd_dt_ops;
7154 spin_lock_init(&o->od_osfs_lock);
7155 mutex_init(&o->od_otable_mutex);
7156 INIT_LIST_HEAD(&o->od_orphan_list);
7158 o->od_read_cache = 1;
7159 o->od_writethrough_cache = 1;
7160 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7162 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7163 sizeof(o->od_svname));
7164 if (cplen >= sizeof(o->od_svname)) {
7169 o->od_index = -1; /* -1 means index is invalid */
7170 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7171 if (rc == LDD_F_SV_TYPE_OST)
7174 o->od_full_scrub_ratio = OFSR_DEFAULT;
7175 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7176 rc = osd_mount(env, o, cfg);
7180 rc = osd_obj_map_init(env, o);
7184 rc = lu_site_init(&o->od_site, l);
7186 GOTO(out_compat, rc);
7187 o->od_site.ls_bottom_dev = l;
7189 rc = lu_site_init_finish(&o->od_site);
7193 INIT_LIST_HEAD(&o->od_ios_list);
7194 /* setup scrub, including OI files initialization */
7195 rc = osd_scrub_setup(env, o);
7199 rc = osd_procfs_init(o, o->od_svname);
7201 CERROR("%s: can't initialize procfs: rc = %d\n",
7203 GOTO(out_scrub, rc);
7206 LASSERT(l->ld_site->ls_linkage.next != NULL);
7207 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7209 /* initialize quota slave instance */
7210 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7212 if (IS_ERR(o->od_quota_slave)) {
7213 rc = PTR_ERR(o->od_quota_slave);
7214 o->od_quota_slave = NULL;
7215 GOTO(out_procfs, rc);
7223 osd_scrub_cleanup(env, o);
7225 lu_site_fini(&o->od_site);
7227 osd_obj_map_fini(o);
7234 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7235 struct lu_device_type *t,
7236 struct lustre_cfg *cfg)
7238 struct osd_device *o;
7243 return ERR_PTR(-ENOMEM);
7245 rc = dt_device_init(&o->od_dt_dev, t);
7247 /* Because the ctx might be revived in dt_device_init,
7248 * refill the env here */
7249 lu_env_refill((struct lu_env *)env);
7250 rc = osd_device_init0(env, o, cfg);
7252 dt_device_fini(&o->od_dt_dev);
7255 if (unlikely(rc != 0))
7258 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7261 static struct lu_device *osd_device_free(const struct lu_env *env,
7262 struct lu_device *d)
7264 struct osd_device *o = osd_dev(d);
7267 /* XXX: make osd top device in order to release reference */
7268 d->ld_site->ls_top_dev = d;
7269 lu_site_purge(env, d->ld_site, -1);
7270 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7271 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7272 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7274 lu_site_fini(&o->od_site);
7275 dt_device_fini(&o->od_dt_dev);
7280 static int osd_process_config(const struct lu_env *env,
7281 struct lu_device *d, struct lustre_cfg *cfg)
7283 struct osd_device *o = osd_dev(d);
7287 switch (cfg->lcfg_command) {
7289 rc = osd_mount(env, o, cfg);
7292 lu_dev_del_linkage(d->ld_site, d);
7293 rc = osd_shutdown(env, o);
7296 LASSERT(&o->od_dt_dev);
7297 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7298 cfg, &o->od_dt_dev);
7299 if (rc > 0 || rc == -ENOSYS)
7300 rc = class_process_proc_param(PARAM_OST,
7301 lprocfs_osd_obd_vars,
7302 cfg, &o->od_dt_dev);
7311 static int osd_recovery_complete(const struct lu_env *env,
7312 struct lu_device *d)
7314 struct osd_device *osd = osd_dev(d);
7318 if (osd->od_quota_slave == NULL)
7321 /* start qsd instance on recovery completion, this notifies the quota
7322 * slave code that we are about to process new requests now */
7323 rc = qsd_start(env, osd->od_quota_slave);
7328 * we use exports to track all osd users
7330 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7331 struct obd_device *obd, struct obd_uuid *cluuid,
7332 struct obd_connect_data *data, void *localdata)
7334 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7335 struct lustre_handle conn;
7339 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7341 rc = class_connect(&conn, obd, cluuid);
7345 *exp = class_conn2export(&conn);
7347 spin_lock(&osd->od_osfs_lock);
7349 spin_unlock(&osd->od_osfs_lock);
7355 * once last export (we don't count self-export) disappeared
7356 * osd can be released
7358 static int osd_obd_disconnect(struct obd_export *exp)
7360 struct obd_device *obd = exp->exp_obd;
7361 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7362 int rc, release = 0;
7365 /* Only disconnect the underlying layers on the final disconnect. */
7366 spin_lock(&osd->od_osfs_lock);
7368 if (osd->od_connects == 0)
7370 spin_unlock(&osd->od_osfs_lock);
7372 rc = class_disconnect(exp); /* bz 9811 */
7374 if (rc == 0 && release)
7375 class_manual_cleanup(obd);
7379 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7380 struct lu_device *dev)
7382 struct osd_device *osd = osd_dev(dev);
7383 struct lr_server_data *lsd =
7384 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7388 if (osd->od_quota_slave != NULL) {
7389 /* set up quota slave objects */
7390 result = qsd_prepare(env, osd->od_quota_slave);
7395 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7396 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7397 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7398 osd->od_index_in_idif = 1;
7400 osd->od_index_in_idif = 0;
7401 result = osd_register_proc_index_in_idif(osd);
7406 osd->od_index_in_idif = 1;
7410 result = osd_fid_init(env, osd);
7415 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7416 struct lu_fid *fid, struct md_op_data *op_data)
7418 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7420 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7423 static const struct lu_object_operations osd_lu_obj_ops = {
7424 .loo_object_init = osd_object_init,
7425 .loo_object_delete = osd_object_delete,
7426 .loo_object_release = osd_object_release,
7427 .loo_object_free = osd_object_free,
7428 .loo_object_print = osd_object_print,
7429 .loo_object_invariant = osd_object_invariant
7432 const struct lu_device_operations osd_lu_ops = {
7433 .ldo_object_alloc = osd_object_alloc,
7434 .ldo_process_config = osd_process_config,
7435 .ldo_recovery_complete = osd_recovery_complete,
7436 .ldo_prepare = osd_prepare,
7439 static const struct lu_device_type_operations osd_device_type_ops = {
7440 .ldto_init = osd_type_init,
7441 .ldto_fini = osd_type_fini,
7443 .ldto_start = osd_type_start,
7444 .ldto_stop = osd_type_stop,
7446 .ldto_device_alloc = osd_device_alloc,
7447 .ldto_device_free = osd_device_free,
7449 .ldto_device_init = osd_device_init,
7450 .ldto_device_fini = osd_device_fini
7453 static struct lu_device_type osd_device_type = {
7454 .ldt_tags = LU_DEVICE_DT,
7455 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
7456 .ldt_ops = &osd_device_type_ops,
7457 .ldt_ctx_tags = LCT_LOCAL,
7460 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
7462 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7463 struct super_block *sb = osd_sb(osd);
7465 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
7469 * lprocfs legacy support.
7471 static struct obd_ops osd_obd_device_ops = {
7472 .o_owner = THIS_MODULE,
7473 .o_connect = osd_obd_connect,
7474 .o_disconnect = osd_obd_disconnect,
7475 .o_fid_alloc = osd_fid_alloc,
7476 .o_health_check = osd_health_check,
7479 static int __init osd_init(void)
7483 LASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
7484 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
7485 /* please, try to keep osd_thread_info smaller than a page */
7486 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
7491 rc = lu_kmem_init(ldiskfs_caches);
7495 rc = class_register_type(&osd_obd_device_ops, NULL, true,
7496 lprocfs_osd_module_vars,
7497 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
7499 lu_kmem_fini(ldiskfs_caches);
7503 static void __exit osd_exit(void)
7505 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
7506 lu_kmem_fini(ldiskfs_caches);
7509 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7510 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
7511 MODULE_VERSION(LUSTRE_VERSION_STRING);
7512 MODULE_LICENSE("GPL");
7514 module_init(osd_init);
7515 module_exit(osd_exit);