4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd/osd_handler.c
34 * Top-level entry points into osd module
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <linux/fs_struct.h>
43 #include <linux/kallsyms.h>
44 #include <linux/module.h>
45 #include <linux/user_namespace.h>
46 #include <linux/uidgid.h>
48 /* prerequisite for linux/xattr.h */
49 #include <linux/types.h>
50 /* prerequisite for linux/xattr.h */
52 /* XATTR_{REPLACE,CREATE} */
53 #include <linux/xattr.h>
55 #include <ldiskfs/ldiskfs.h>
56 #include <ldiskfs/xattr.h>
57 #include <ldiskfs/ldiskfs_extents.h>
60 * struct OBD_{ALLOC,FREE}*()
63 #include <obd_support.h>
64 /* struct ptlrpc_thread */
65 #include <lustre_net.h>
66 #include <lustre_fid.h>
68 #include <uapi/linux/lustre/lustre_param.h>
70 #include "osd_internal.h"
71 #include "osd_dynlocks.h"
73 /* llo_* api support */
74 #include <md_object.h>
75 #include <lustre_quota.h>
77 #include <lustre_linkea.h>
79 /* Maximum EA size is limited by LNET_MTU for remote objects */
80 #define OSD_MAX_EA_SIZE 1048364
83 module_param(ldiskfs_pdo, int, 0644);
84 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
86 int ldiskfs_track_declares_assert;
87 module_param(ldiskfs_track_declares_assert, int, 0644);
88 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
90 /* Slab to allocate dynlocks */
91 struct kmem_cache *dynlock_cachep;
93 /* Slab to allocate osd_it_ea */
94 struct kmem_cache *osd_itea_cachep;
96 static struct lu_kmem_descr ldiskfs_caches[] = {
98 .ckd_cache = &dynlock_cachep,
99 .ckd_name = "dynlock_cache",
100 .ckd_size = sizeof(struct dynlock_handle)
103 .ckd_cache = &osd_itea_cachep,
104 .ckd_name = "osd_itea_cache",
105 .ckd_size = sizeof(struct osd_it_ea)
112 static const char dot[] = ".";
113 static const char dotdot[] = "..";
115 static const struct lu_object_operations osd_lu_obj_ops;
116 static const struct dt_object_operations osd_obj_ops;
117 static const struct dt_object_operations osd_obj_otable_it_ops;
118 static const struct dt_index_operations osd_index_iam_ops;
119 static const struct dt_index_operations osd_index_ea_ops;
121 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
122 const struct lu_fid *fid);
123 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
124 struct osd_device *osd);
126 int osd_trans_declare_op2rb[] = {
127 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
128 [OSD_OT_PUNCH] = OSD_OT_MAX,
129 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
130 [OSD_OT_CREATE] = OSD_OT_DESTROY,
131 [OSD_OT_DESTROY] = OSD_OT_CREATE,
132 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
133 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
134 [OSD_OT_WRITE] = OSD_OT_WRITE,
135 [OSD_OT_INSERT] = OSD_OT_DELETE,
136 [OSD_OT_DELETE] = OSD_OT_INSERT,
137 [OSD_OT_QUOTA] = OSD_OT_MAX,
140 static int osd_has_index(const struct osd_object *obj)
142 return obj->oo_dt.do_index_ops != NULL;
145 static int osd_object_invariant(const struct lu_object *l)
147 return osd_invariant(osd_obj(l));
151 * Concurrency: doesn't matter
153 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
155 struct osd_thread_info *oti = osd_oti_get(env);
157 return oti->oti_w_locks > 0 && o->oo_owner == env;
161 * Concurrency: doesn't access mutable data
163 static int osd_root_get(const struct lu_env *env,
164 struct dt_device *dev, struct lu_fid *f)
166 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
171 * the following set of functions are used to maintain per-thread
172 * cache of FID->ino mapping. this mechanism is needed to resolve
173 * FID to inode at dt_insert() which in turn stores ino in the
174 * directory entries to keep ldiskfs compatible with ext[34].
175 * due to locking-originated restrictions we can't lookup ino
176 * using LU cache (deadlock is possible). lookup using OI is quite
177 * expensive. so instead we maintain this cache and methods like
178 * dt_create() fill it. so in the majority of cases dt_insert() is
179 * able to find needed mapping in lockless manner.
181 static struct osd_idmap_cache *
182 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
183 const struct lu_fid *fid)
185 struct osd_thread_info *oti = osd_oti_get(env);
186 struct osd_idmap_cache *idc = oti->oti_ins_cache;
189 for (i = 0; i < oti->oti_ins_cache_used; i++) {
190 if (!lu_fid_eq(&idc[i].oic_fid, fid))
192 if (idc[i].oic_dev != osd)
201 static struct osd_idmap_cache *
202 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
203 const struct lu_fid *fid)
205 struct osd_thread_info *oti = osd_oti_get(env);
206 struct osd_idmap_cache *idc;
209 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
210 i = oti->oti_ins_cache_size * 2;
212 i = OSD_INS_CACHE_SIZE;
213 OBD_ALLOC_PTR_ARRAY(idc, i);
215 return ERR_PTR(-ENOMEM);
216 if (oti->oti_ins_cache != NULL) {
217 memcpy(idc, oti->oti_ins_cache,
218 oti->oti_ins_cache_used * sizeof(*idc));
219 OBD_FREE_PTR_ARRAY(oti->oti_ins_cache,
220 oti->oti_ins_cache_used);
222 oti->oti_ins_cache = idc;
223 oti->oti_ins_cache_size = i;
226 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
229 idc->oic_lid.oii_ino = 0;
230 idc->oic_lid.oii_gen = 0;
237 * lookup mapping for the given fid in the cache, initialize a
238 * new one if not found. the initialization checks whether the
239 * object is local or remote. for local objects, OI is used to
240 * learn ino/generation. the function is used when the caller
241 * has no information about the object, e.g. at dt_insert().
243 static struct osd_idmap_cache *
244 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
245 const struct lu_fid *fid)
247 struct osd_idmap_cache *idc;
250 idc = osd_idc_find(env, osd, fid);
251 LASSERT(!IS_ERR(idc));
255 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
256 osd->od_svname, PFID(fid));
258 /* new mapping is needed */
259 idc = osd_idc_add(env, osd, fid);
261 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
262 osd->od_svname, PFID(fid), PTR_ERR(idc));
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 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
313 osd->od_svname, PFID(fid));
315 /* new mapping is needed */
316 idc = osd_idc_add(env, osd, fid);
318 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
319 osd->od_svname, PFID(fid), PTR_ERR(idc));
323 if (obj->oo_inode != NULL) {
324 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
325 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
331 * OSD object methods.
335 * Concurrency: no concurrent access is possible that early in object
338 static struct lu_object *osd_object_alloc(const struct lu_env *env,
339 const struct lu_object_header *hdr,
342 struct osd_object *mo;
347 struct lu_object_header *h;
348 struct osd_device *o = osd_dev(d);
350 l = &mo->oo_dt.do_lu;
351 if (unlikely(o->od_in_init)) {
358 lu_object_header_init(h);
359 lu_object_init(l, h, d);
360 lu_object_add_top(h, l);
363 dt_object_init(&mo->oo_dt, NULL, d);
364 mo->oo_header = NULL;
367 mo->oo_dt.do_ops = &osd_obj_ops;
368 l->lo_ops = &osd_lu_obj_ops;
369 init_rwsem(&mo->oo_sem);
370 init_rwsem(&mo->oo_ext_idx_sem);
371 spin_lock_init(&mo->oo_guard);
372 INIT_LIST_HEAD(&mo->oo_xattr_list);
378 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
379 struct dentry *dentry, struct lustre_ost_attrs *loa)
383 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
384 (void *)loa, sizeof(*loa));
386 struct lustre_mdt_attrs *lma = &loa->loa_lma;
388 if (rc < sizeof(*lma))
392 lustre_loa_swab(loa, true);
393 /* Check LMA compatibility */
394 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
395 CWARN("%s: unsupported incompat LMA feature(s) %#x "
396 "for fid = "DFID", ino = %lu\n",
398 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
399 PFID(&lma->lma_self_fid), inode->i_ino);
402 } else if (rc == 0) {
410 * retrieve object from backend ext fs.
412 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
413 struct osd_inode_id *id)
416 struct inode *inode = NULL;
419 * if we look for an inode withing a running
420 * transaction, then we risk to deadlock
421 * osd_dirent_check_repair() breaks this
423 /* LASSERT(current->journal_info == NULL); */
425 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
427 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
428 id->oii_ino, PTR_ERR(inode));
429 } else if (id->oii_gen != OSD_OII_NOGEN &&
430 inode->i_generation != id->oii_gen) {
431 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
432 "i_generation = %u\n",
433 id->oii_ino, id->oii_gen, inode->i_generation);
435 inode = ERR_PTR(-ESTALE);
436 } else if (inode->i_nlink == 0) {
438 * due to parallel readdir and unlink,
439 * we can have dead inode here.
441 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
443 inode = ERR_PTR(-ESTALE);
444 } else if (is_bad_inode(inode)) {
445 CWARN("%s: bad inode: ino = %u\n",
446 osd_dev2name(dev), id->oii_ino);
448 inode = ERR_PTR(-ENOENT);
449 } else if ((rc = osd_attach_jinode(inode))) {
453 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
454 if (id->oii_gen == OSD_OII_NOGEN)
455 osd_id_gen(id, inode->i_ino, inode->i_generation);
458 * Do not update file c/mtime in ldiskfs.
459 * NB: we don't have any lock to protect this because we don't
460 * have reference on osd_object now, but contention with
461 * another lookup + attr_set can't happen in the tiny window
462 * between if (...) and set S_NOCMTIME.
464 if (!(inode->i_flags & S_NOCMTIME))
465 inode->i_flags |= S_NOCMTIME;
470 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
471 handle_t *handle, struct dentry *child,
472 struct inode *inode, struct htree_lock *hlock)
476 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
477 if (rc == -ENOBUFS || rc == -ENOSPC) {
478 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
479 struct inode *parent = child->d_parent->d_inode;
480 struct lu_fid *fid = NULL;
481 char fidstr[FID_LEN + 1] = "unknown";
483 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
485 fid = &loa->loa_lma.lma_self_fid;
486 } else if (rc2 == -ENODATA) {
487 if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
488 fid = &info->oti_fid3;
489 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
490 } else if (!osd->od_is_ost && osd->od_index == 0) {
491 fid = &info->oti_fid3;
492 lu_igif_build(fid, parent->i_ino,
493 parent->i_generation);
498 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
500 /* below message is checked in sanity.sh test_129 */
502 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
503 osd_name(osd), parent->i_ino, fidstr);
505 rc = 0; /* ignore such error now */
506 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
507 osd_name(osd), parent->i_ino, fidstr);
517 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
518 struct osd_inode_id *id, struct lu_fid *fid)
520 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
524 inode = osd_iget(info, dev, id);
528 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
530 *fid = loa->loa_lma.lma_self_fid;
531 } else if (rc == -ENODATA) {
532 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
533 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
535 lu_igif_build(fid, inode->i_ino, inode->i_generation);
543 static struct inode *osd_iget_check(struct osd_thread_info *info,
544 struct osd_device *dev,
545 const struct lu_fid *fid,
546 struct osd_inode_id *id,
555 * The cached OI mapping is trustable. If we cannot locate the inode
556 * via the cached OI mapping, then return the failure to the caller
557 * directly without further OI checking.
561 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
564 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
567 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
568 PFID(fid), id->oii_ino, rc);
572 if (is_bad_inode(inode)) {
577 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
578 PFID(fid), id->oii_ino);
582 if (id->oii_gen != OSD_OII_NOGEN &&
583 inode->i_generation != id->oii_gen) {
588 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
589 "oii_gen = %u, i_generation = %u\n", PFID(fid),
590 id->oii_ino, id->oii_gen, inode->i_generation);
594 if (inode->i_nlink == 0) {
599 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
600 PFID(fid), id->oii_ino);
604 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
608 __u32 saved_ino = id->oii_ino;
609 __u32 saved_gen = id->oii_gen;
612 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
614 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
616 * XXX: There are four possible cases:
618 * Backup/restore caused the OI invalid.
620 * Someone unlinked the object but NOT removed
621 * the OI mapping, such as mount target device
622 * as ldiskfs, and modify something directly.
624 * Someone just removed the object between the
625 * former oi_lookup and the iget. It is normal.
626 * 4. Other failure cases.
628 * Generally, when the device is mounted, it will
629 * auto check whether the system is restored from
630 * file-level backup or not. We trust such detect
631 * to distinguish the 1st case from the 2nd case:
632 * if the OI files are consistent but may contain
633 * stale OI mappings because of case 2, if iget()
634 * returns -ENOENT or -ESTALE, then it should be
639 * If the OI mapping was in OI file before the
640 * osd_iget_check(), but now, it is disappear,
641 * then it must be removed by race. That is a
647 * It is the OI scrub updated the OI mapping by race.
648 * The new OI mapping must be valid.
650 if (saved_ino != id->oii_ino ||
651 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
660 if (dev->od_scrub.os_scrub.os_file.sf_flags &
663 * It still can be the case 2, but we cannot
664 * distinguish it from the case 1. So return
665 * -EREMCHG to block current operation until
666 * OI scrub rebuilt the OI mappings.
675 if (inode->i_generation == id->oii_gen)
680 if (id->oii_gen == OSD_OII_NOGEN)
681 osd_id_gen(id, inode->i_ino, inode->i_generation);
684 * Do not update file c/mtime in ldiskfs.
685 * NB: we don't have any lock to protect this because we don't
686 * have reference on osd_object now, but contention with
687 * another lookup + attr_set can't happen in the tiny window
688 * between if (...) and set S_NOCMTIME.
690 if (!(inode->i_flags & S_NOCMTIME))
691 inode->i_flags |= S_NOCMTIME;
708 * \retval +v: new filter_fid does not contain self-fid
709 * \retval 0: filter_fid_18_23, contains self-fid
710 * \retval -v: other failure cases
712 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
713 struct dentry *dentry, struct lu_fid *fid)
715 struct filter_fid *ff = &info->oti_ff;
716 struct ost_id *ostid = &info->oti_ostid;
719 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
720 if (rc == sizeof(struct filter_fid_18_23)) {
721 struct filter_fid_18_23 *ff_old = (void *)ff;
723 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
724 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
726 * XXX: use 0 as the index for compatibility, the caller will
727 * handle index related issues when necessary.
730 ostid_to_fid(fid, ostid, 0);
731 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
733 } else if (rc >= 0) {
740 static int osd_lma_self_repair(struct osd_thread_info *info,
741 struct osd_device *osd, struct inode *inode,
742 const struct lu_fid *fid, __u32 compat)
747 LASSERT(current->journal_info == NULL);
749 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
750 osd_dto_credits_noquota[DTO_XATTR_SET]);
753 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
758 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
760 CWARN("%s: cannot self repair the LMA: rc = %d\n",
762 ldiskfs_journal_stop(jh);
766 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
768 struct osd_thread_info *info = osd_oti_get(env);
769 struct osd_device *osd = osd_obj2dev(obj);
770 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
771 struct lustre_mdt_attrs *lma = &loa->loa_lma;
772 struct inode *inode = obj->oo_inode;
773 struct dentry *dentry = &info->oti_obj_dentry;
774 struct lu_fid *fid = NULL;
775 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
780 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
781 (void *)loa, sizeof(*loa));
782 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
783 fid = &lma->lma_self_fid;
784 rc = osd_get_idif(info, inode, dentry, fid);
785 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
787 * For the given OST-object, if it has neither LMA nor
788 * FID in XATTR_NAME_FID, then the given FID (which is
789 * contained in the @obj, from client RPC for locating
790 * the OST-object) is trusted. We use it to generate
793 osd_lma_self_repair(info, osd, inode, rfid,
804 lustre_lma_swab(lma);
805 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
806 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
807 S_ISREG(inode->i_mode)))) {
808 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
809 "fid = "DFID", ino = %lu\n", osd_name(osd),
810 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
811 PFID(rfid), inode->i_ino);
814 fid = &lma->lma_self_fid;
815 if (lma->lma_compat & LMAC_STRIPE_INFO &&
817 obj->oo_pfid_in_lma = 1;
818 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
820 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
824 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
825 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
826 struct ost_id *oi = &info->oti_ostid;
827 struct lu_fid *fid1 = &info->oti_fid3;
828 __u32 idx = fid_idif_ost_idx(rfid);
831 * For old IDIF, the OST index is not part of the IDIF,
832 * Means that different OSTs may have the same IDIFs.
833 * Under such case, we need to make some compatible
834 * check to make sure to trigger OI scrub properly.
836 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
837 /* Given @rfid is new, LMA is old. */
838 fid_to_ostid(fid, oi);
839 ostid_to_fid(fid1, oi, idx);
840 if (lu_fid_eq(fid1, rfid)) {
841 if (osd->od_index_in_idif)
842 osd_lma_self_repair(info, osd,
856 struct osd_check_lmv_buf {
857 /* please keep it as first member */
858 struct dir_context ctx;
859 struct osd_thread_info *oclb_info;
860 struct osd_device *oclb_dev;
861 struct osd_idmap_cache *oclb_oic;
867 * It is called internally by ->iterate*() to filter out the
868 * local slave object's FID of the striped directory.
870 * \retval 1 found the local slave's FID
871 * \retval 0 continue to check next item
872 * \retval -ve for failure
874 #ifdef HAVE_FILLDIR_USE_CTX
875 static int osd_stripe_dir_filldir(struct dir_context *buf,
877 static int osd_stripe_dir_filldir(void *buf,
879 const char *name, int namelen,
880 loff_t offset, __u64 ino, unsigned int d_type)
882 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
883 struct osd_thread_info *oti = oclb->oclb_info;
884 struct lu_fid *fid = &oti->oti_fid3;
885 struct osd_inode_id *id = &oti->oti_id3;
886 struct osd_device *dev = oclb->oclb_dev;
887 struct osd_idmap_cache *oic = oclb->oclb_oic;
896 sscanf(name + 1, SFID, RFID(fid));
897 if (!fid_is_sane(fid))
900 if (osd_remote_fid(oti->oti_env, dev, fid))
903 osd_id_gen(id, ino, OSD_OII_NOGEN);
904 inode = osd_iget(oti, dev, id);
906 return PTR_ERR(inode);
909 osd_add_oi_cache(oti, dev, id, fid);
913 osd_oii_insert(dev, oic, true);
914 oclb->oclb_found = true;
920 * When lookup item under striped directory, we need to locate the master
921 * MDT-object of the striped directory firstly, then the client will send
922 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
923 * and the item's name. If the system is restored from MDT file level backup,
924 * then before the OI scrub completely built the OI files, the OI mappings of
925 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
926 * not a problem for the master MDT-object. Because when locate the master
927 * MDT-object, we will do name based lookup (for the striped directory itself)
928 * firstly, during such process we can setup the correct OI mapping for the
929 * master MDT-object. But it will be trouble for the slave MDT-object. Because
930 * the client will not trigger name based lookup on the MDT to locate the slave
931 * MDT-object before locating item under the striped directory, then when
932 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
933 * is invalid and does not know what the right OI mapping is, then the MDT has
934 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
935 * the OI file, related OI mapping is unknown yet, please try again later. And
936 * then client will re-try the RPC again and again until related OI mapping has
937 * been updated. That is quite inefficient.
939 * To resolve above trouble, we will handle it as the following two cases:
941 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
942 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
943 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
944 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
945 * directly. Please check osd_fid_lookup().
947 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
948 * Under such case, during lookup the master MDT-object, we will lookup the
949 * slave MDT-object via readdir against the master MDT-object, because the
950 * slave MDT-objects information are stored as sub-directories with the name
951 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
952 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
953 * the correct OI mapping for the slave MDT-object.
955 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
956 struct inode *inode, struct osd_idmap_cache *oic)
958 struct lu_buf *buf = &oti->oti_big_buf;
959 struct dentry *dentry = &oti->oti_obj_dentry;
961 struct lmv_mds_md_v1 *lmv1;
962 struct osd_check_lmv_buf oclb = {
963 .ctx.actor = osd_stripe_dir_filldir,
974 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
977 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
979 lu_buf_realloc(buf, rc);
980 if (buf->lb_buf == NULL)
981 GOTO(out, rc = -ENOMEM);
987 if (unlikely(rc == 0 || rc == -ENODATA))
993 if (unlikely(buf->lb_buf == NULL)) {
994 lu_buf_realloc(buf, rc);
995 if (buf->lb_buf == NULL)
996 GOTO(out, rc = -ENOMEM);
1002 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1005 filp = osd_quasi_file(oti->oti_env, inode);
1006 rc = osd_security_file_alloc(filp);
1011 oclb.oclb_items = 0;
1012 rc = iterate_dir(filp, &oclb.ctx);
1013 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1014 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1015 inode->i_fop->release(inode, filp);
1019 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
1020 DFID": rc = %d\n", osd_ino2name(inode),
1021 inode->i_ino, inode->i_generation,
1022 PFID(&oic->oic_fid), rc);
1029 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1030 const struct lu_fid *fid,
1031 const struct lu_object_conf *conf)
1033 struct osd_thread_info *info;
1034 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1035 struct osd_device *dev;
1036 struct osd_idmap_cache *oic;
1037 struct osd_inode_id *id;
1038 struct inode *inode = NULL;
1039 struct lustre_scrub *scrub;
1040 struct scrub_file *sf;
1041 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1046 bool remote = false;
1047 bool trusted = true;
1048 bool updated = false;
1049 bool checked = false;
1053 LINVRNT(osd_invariant(obj));
1054 LASSERT(obj->oo_inode == NULL);
1055 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1057 dev = osd_dev(ldev);
1058 scrub = &dev->od_scrub.os_scrub;
1059 sf = &scrub->os_file;
1060 info = osd_oti_get(env);
1062 oic = &info->oti_cache;
1064 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1068 * For the object is created as locking anchor, or for the object to
1069 * be created on disk. No need to osd_oi_lookup() at here because FID
1070 * shouldn't never be re-used, if it's really a duplicate FID from
1071 * unexpected reason, we should be able to detect it later by calling
1072 * do_create->osd_oi_insert().
1074 if (conf && conf->loc_flags & LOC_F_NEW)
1075 GOTO(out, result = 0);
1077 /* Search order: 1. per-thread cache. */
1078 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1084 if (!list_empty(&scrub->os_inconsistent_items)) {
1085 /* Search order: 2. OI scrub pending list. */
1086 result = osd_oii_lookup(dev, fid, id);
1092 * The OI mapping in the OI file can be updated by the OI scrub
1093 * when we locate the inode via FID. So it may be not trustable.
1097 /* Search order: 3. OI files. */
1098 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1099 if (result == -ENOENT) {
1100 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1101 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1102 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1104 GOTO(out, result = 0);
1109 /* -ESTALE is returned if inode of OST object doesn't exist */
1110 if (result == -ESTALE &&
1111 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1112 GOTO(out, result = 0);
1119 obj->oo_inode = NULL;
1120 /* for later passes through checks, not true on first pass */
1121 if (!IS_ERR_OR_NULL(inode))
1124 inode = osd_iget_check(info, dev, fid, id, trusted);
1125 if (!IS_ERR(inode)) {
1126 obj->oo_inode = inode;
1134 result = PTR_ERR(inode);
1135 if (result == -ENOENT || result == -ESTALE)
1136 GOTO(out, result = 0);
1138 if (result != -EREMCHG)
1143 * We still have chance to get the valid inode: for the
1144 * object which is referenced by remote name entry, the
1145 * object on the local MDT will be linked under the dir
1146 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1148 * We do not know whether the object for the given FID
1149 * is referenced by some remote name entry or not, and
1150 * especially for DNE II, a multiple-linked object may
1151 * have many name entries reside on many MDTs.
1153 * To simplify the operation, OSD will not distinguish
1154 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1155 * only happened for the RPC from other MDT during the
1156 * OI scrub, or for the client side RPC with FID only,
1157 * such as FID to path, or from old connected client.
1160 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1164 flags |= SS_AUTO_PARTIAL;
1165 flags &= ~SS_AUTO_FULL;
1170 if (scrub->os_running) {
1171 if (scrub->os_partial_scan && !scrub->os_in_join)
1174 osd_add_oi_cache(info, dev, id, fid);
1175 if (IS_ERR_OR_NULL(inode) || result) {
1176 osd_oii_insert(dev, oic, result == -ENOENT);
1177 GOTO(out, result = -EINPROGRESS);
1181 LASSERT(obj->oo_inode == inode);
1183 osd_oii_insert(dev, oic, true);
1187 if (dev->od_auto_scrub_interval == AS_NEVER) {
1189 GOTO(out, result = -EREMCHG);
1192 LASSERT(obj->oo_inode == inode);
1194 osd_add_oi_cache(info, dev, id, fid);
1199 rc1 = osd_scrub_start(env, dev, flags);
1200 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1201 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1202 if (rc1 && rc1 != -EALREADY)
1203 GOTO(out, result = -EREMCHG);
1205 osd_add_oi_cache(info, dev, id, fid);
1206 if (IS_ERR_OR_NULL(inode) || result) {
1207 osd_oii_insert(dev, oic, result == -ENOENT);
1208 GOTO(out, result = -EINPROGRESS);
1212 LASSERT(obj->oo_inode == inode);
1214 osd_oii_insert(dev, oic, true);
1219 if (unlikely(obj->oo_header))
1222 result = osd_check_lma(env, obj);
1226 LASSERTF(id->oii_ino == inode->i_ino &&
1227 id->oii_gen == inode->i_generation,
1228 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1229 PFID(fid), id->oii_ino, id->oii_gen,
1230 inode->i_ino, inode->i_generation);
1232 saved_ino = inode->i_ino;
1233 saved_gen = inode->i_generation;
1235 if (unlikely(result == -ENODATA)) {
1237 * If the OI scrub updated the OI mapping by race, it
1238 * must be valid. Trust the inode that has no LMA EA.
1243 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1246 * The OI mapping is still there, the inode is still
1247 * valid. It is just becaues the inode has no LMA EA.
1249 if (saved_ino == id->oii_ino &&
1250 saved_gen == id->oii_gen)
1254 * It is the OI scrub updated the OI mapping by race.
1255 * The new OI mapping must be valid.
1263 * "result == -ENOENT" means that the OI mappinghas been
1264 * removed by race, so the inode belongs to other object.
1266 * Others error can be returned directly.
1268 if (result == -ENOENT) {
1271 obj->oo_inode = NULL;
1276 if (result != -EREMCHG)
1282 * if two OST objects map to the same inode, and inode mode is
1283 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1284 * reserved by precreate, and not written yet, in this case, don't
1285 * set inode for the object whose FID mismatch, so that it can create
1286 * inode and not block precreate.
1288 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1289 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1290 obj->oo_inode = NULL;
1291 GOTO(out, result = 0);
1294 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1296 * "result == -ENOENT" means the cached OI mapping has been removed
1297 * from the OI file by race, above inode belongs to other object.
1299 if (result == -ENOENT) {
1302 obj->oo_inode = NULL;
1303 GOTO(out, result = 0);
1309 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1315 * It is the OI scrub updated the OI mapping by race.
1316 * The new OI mapping must be valid.
1324 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1325 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1327 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1329 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1331 obj->oo_pfid_in_lma = 1;
1332 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1334 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1335 } else if (result != -ENODATA) {
1340 obj->oo_compat_dot_created = 1;
1341 obj->oo_compat_dotdot_created = 1;
1343 if (S_ISDIR(inode->i_mode) &&
1344 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1345 osd_check_lmv(info, dev, inode, oic);
1347 result = osd_attach_jinode(inode);
1352 GOTO(out, result = 0);
1354 LASSERT(!obj->oo_hl_head);
1355 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1357 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1360 if (result || !obj->oo_inode) {
1361 if (!IS_ERR_OR_NULL(inode))
1364 obj->oo_inode = NULL;
1366 fid_zero(&oic->oic_fid);
1369 LINVRNT(osd_invariant(obj));
1374 * Concurrency: shouldn't matter.
1376 static void osd_object_init0(struct osd_object *obj)
1378 LASSERT(obj->oo_inode != NULL);
1379 obj->oo_dt.do_body_ops = &osd_body_ops;
1380 obj->oo_dt.do_lu.lo_header->loh_attr |=
1381 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1385 * Concurrency: no concurrent access is possible that early in object
1388 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1389 const struct lu_object_conf *conf)
1391 struct osd_object *obj = osd_obj(l);
1394 LINVRNT(osd_invariant(obj));
1396 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1397 cfs_fail_val == 2) {
1398 struct osd_thread_info *info = osd_oti_get(env);
1399 struct osd_idmap_cache *oic = &info->oti_cache;
1400 /* invalidate thread cache */
1401 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1403 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1404 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1405 l->lo_header->loh_attr |= LOHA_EXISTS;
1409 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1410 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1411 if (result == 0 && obj->oo_inode != NULL) {
1412 struct osd_thread_info *oti = osd_oti_get(env);
1413 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1415 osd_object_init0(obj);
1416 if (unlikely(obj->oo_header))
1419 result = osd_get_lma(oti, obj->oo_inode,
1420 &oti->oti_obj_dentry, loa);
1423 * Convert LMAI flags to lustre LMA flags
1424 * and cache it to oo_lma_flags
1427 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1428 } else if (result == -ENODATA) {
1432 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
1434 LINVRNT(osd_invariant(obj));
1439 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1440 * The left part is for value, binary mode.
1442 struct osd_xattr_entry {
1443 struct list_head oxe_list;
1447 struct rcu_head oxe_rcu;
1451 static int osd_oxc_get(struct osd_object *obj, const char *name,
1454 struct osd_xattr_entry *tmp;
1455 struct osd_xattr_entry *oxe = NULL;
1456 size_t namelen = strlen(name);
1460 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1461 if (namelen == tmp->oxe_namelen &&
1462 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1469 GOTO(out, rc = -ENOENT);
1471 if (!oxe->oxe_exist)
1472 GOTO(out, rc = -ENODATA);
1475 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1478 if (buf->lb_buf == NULL)
1481 if (buf->lb_len < rc)
1482 GOTO(out, rc = -ERANGE);
1484 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1491 static void osd_oxc_free(struct rcu_head *head)
1493 struct osd_xattr_entry *oxe;
1495 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1496 OBD_FREE(oxe, oxe->oxe_len);
1499 static void osd_oxc_add(struct osd_object *obj, const char *name,
1500 const char *buf, int buflen)
1502 struct osd_xattr_entry *oxe;
1503 struct osd_xattr_entry *old = NULL;
1504 struct osd_xattr_entry *tmp;
1505 size_t namelen = strlen(name);
1506 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1508 OBD_ALLOC(oxe, len);
1512 INIT_LIST_HEAD(&oxe->oxe_list);
1514 oxe->oxe_namelen = namelen;
1515 memcpy(oxe->oxe_buf, name, namelen);
1517 LASSERT(buf != NULL);
1518 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1519 oxe->oxe_exist = true;
1521 oxe->oxe_exist = false;
1524 /* this should be rarely called, just remove old and add new */
1525 spin_lock(&obj->oo_guard);
1526 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1527 if (namelen == tmp->oxe_namelen &&
1528 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1534 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1535 call_rcu(&old->oxe_rcu, osd_oxc_free);
1537 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1539 spin_unlock(&obj->oo_guard);
1542 static void osd_oxc_del(struct osd_object *obj, const char *name)
1544 struct osd_xattr_entry *oxe;
1545 size_t namelen = strlen(name);
1547 spin_lock(&obj->oo_guard);
1548 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1549 if (namelen == oxe->oxe_namelen &&
1550 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1551 list_del_rcu(&oxe->oxe_list);
1552 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1556 spin_unlock(&obj->oo_guard);
1559 static void osd_oxc_fini(struct osd_object *obj)
1561 struct osd_xattr_entry *oxe, *next;
1563 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1564 list_del(&oxe->oxe_list);
1565 OBD_FREE(oxe, oxe->oxe_len);
1570 * Concurrency: no concurrent access is possible that late in object
1573 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1575 struct osd_object *obj = osd_obj(l);
1576 struct lu_object_header *h = obj->oo_header;
1578 LINVRNT(osd_invariant(obj));
1581 dt_object_fini(&obj->oo_dt);
1582 if (obj->oo_hl_head != NULL)
1583 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1584 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1587 lu_object_header_fini(h);
1588 OBD_FREE_PRE(h, sizeof(*h), "kfreed");
1589 kfree_rcu(h, loh_rcu);
1594 * Concurrency: no concurrent access is possible that late in object
1597 static void osd_index_fini(struct osd_object *o)
1599 struct iam_container *bag;
1601 if (o->oo_dir != NULL) {
1602 bag = &o->oo_dir->od_container;
1603 if (o->oo_inode != NULL) {
1604 if (bag->ic_object == o->oo_inode)
1605 iam_container_fini(bag);
1607 OBD_FREE_PTR(o->oo_dir);
1613 OSD_TXN_OI_DELETE_CREDITS = 20,
1614 OSD_TXN_INODE_DELETE_CREDITS = 20
1621 #if OSD_THANDLE_STATS
1623 * Set time when the handle is allocated
1625 static void osd_th_alloced(struct osd_thandle *oth)
1627 oth->oth_alloced = ktime_get();
1631 * Set time when the handle started
1633 static void osd_th_started(struct osd_thandle *oth)
1635 oth->oth_started = ktime_get();
1639 * Check whether the we deal with this handle for too long.
1641 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1642 ktime_t alloced, ktime_t started,
1645 ktime_t now = ktime_get();
1647 LASSERT(dev != NULL);
1649 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1650 ktime_us_delta(started, alloced));
1651 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1652 ktime_us_delta(closed, started));
1653 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1654 ktime_us_delta(now, closed));
1656 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1657 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1658 oth, now, alloced, started, closed);
1659 libcfs_debug_dumpstack(NULL);
1663 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1665 ktime_t __closed = ktime_get(); \
1666 ktime_t __alloced = oth->oth_alloced; \
1667 ktime_t __started = oth->oth_started; \
1670 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1673 #else /* OSD_THANDLE_STATS */
1675 #define osd_th_alloced(h) do {} while(0)
1676 #define osd_th_started(h) do {} while(0)
1677 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1679 #endif /* OSD_THANDLE_STATS */
1682 * Concurrency: doesn't access mutable data.
1684 static int osd_param_is_not_sane(const struct osd_device *dev,
1685 const struct thandle *th)
1687 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1689 return oh->ot_credits > osd_transaction_size(dev);
1693 * Concurrency: shouldn't matter.
1695 static void osd_trans_commit_cb(struct super_block *sb,
1696 struct ldiskfs_journal_cb_entry *jcb, int error)
1698 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1699 struct thandle *th = &oh->ot_super;
1700 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1701 struct dt_txn_commit_cb *dcb, *tmp;
1703 LASSERT(oh->ot_handle == NULL);
1706 CERROR("transaction @0x%p commit error: %d\n", th, error);
1708 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1709 /* call per-transaction callbacks if any */
1710 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1712 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1713 "commit callback entry: magic=%x name='%s'\n",
1714 dcb->dcb_magic, dcb->dcb_name);
1715 list_del_init(&dcb->dcb_linkage);
1716 dcb->dcb_func(NULL, th, dcb, error);
1719 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1726 static struct thandle *osd_trans_create(const struct lu_env *env,
1727 struct dt_device *d)
1729 struct osd_thread_info *oti = osd_oti_get(env);
1730 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1731 struct osd_thandle *oh;
1737 CERROR("%s: someone try to start transaction under "
1738 "readonly mode, should be disabled.\n",
1739 osd_name(osd_dt_dev(d)));
1741 RETURN(ERR_PTR(-EROFS));
1744 /* on pending IO in this thread should left from prev. request */
1745 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1747 sb_start_write(osd_sb(osd_dt_dev(d)));
1749 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1751 sb_end_write(osd_sb(osd_dt_dev(d)));
1752 RETURN(ERR_PTR(-ENOMEM));
1755 oh->ot_quota_trans = &oti->oti_quota_trans;
1756 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1761 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1762 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1763 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1766 memset(oti->oti_declare_ops, 0,
1767 sizeof(oti->oti_declare_ops));
1768 memset(oti->oti_declare_ops_cred, 0,
1769 sizeof(oti->oti_declare_ops_cred));
1770 memset(oti->oti_declare_ops_used, 0,
1771 sizeof(oti->oti_declare_ops_used));
1773 oti->oti_ins_cache_depth++;
1778 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1780 struct osd_thread_info *oti = osd_oti_get(env);
1781 struct osd_thandle *oh;
1783 oh = container_of(th, struct osd_thandle, ot_super);
1784 LASSERT(oh != NULL);
1786 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1787 oti->oti_declare_ops[OSD_OT_CREATE],
1788 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1789 oti->oti_declare_ops_used[OSD_OT_CREATE],
1790 oti->oti_declare_ops[OSD_OT_DESTROY],
1791 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1792 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1793 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1794 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1795 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1796 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1797 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1798 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1799 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1800 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1801 oti->oti_declare_ops[OSD_OT_WRITE],
1802 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1803 oti->oti_declare_ops_used[OSD_OT_WRITE],
1804 oti->oti_declare_ops[OSD_OT_PUNCH],
1805 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1806 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1807 oti->oti_declare_ops[OSD_OT_QUOTA],
1808 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1809 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1810 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1811 oti->oti_declare_ops[OSD_OT_INSERT],
1812 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1813 oti->oti_declare_ops_used[OSD_OT_INSERT],
1814 oti->oti_declare_ops[OSD_OT_DELETE],
1815 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1816 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1817 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1818 oti->oti_declare_ops[OSD_OT_REF_ADD],
1819 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1820 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1821 oti->oti_declare_ops[OSD_OT_REF_DEL],
1822 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1823 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1827 * Concurrency: shouldn't matter.
1829 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1832 struct osd_thread_info *oti = osd_oti_get(env);
1833 struct osd_device *dev = osd_dt_dev(d);
1835 struct osd_thandle *oh;
1840 LASSERT(current->journal_info == NULL);
1842 oh = container_of(th, struct osd_thandle, ot_super);
1843 LASSERT(oh != NULL);
1844 LASSERT(oh->ot_handle == NULL);
1846 rc = dt_txn_hook_start(env, d, th);
1850 if (unlikely(osd_param_is_not_sane(dev, th))) {
1851 static unsigned long last_printed;
1852 static int last_credits;
1855 * don't make noise on a tiny testing systems
1856 * actual credits misuse will be caught anyway
1858 if (last_credits != oh->ot_credits &&
1859 time_after(jiffies, last_printed +
1860 cfs_time_seconds(60)) &&
1861 osd_transaction_size(dev) > 512) {
1862 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1863 oh->ot_credits, osd_transaction_size(dev));
1864 osd_trans_dump_creds(env, th);
1865 libcfs_debug_dumpstack(NULL);
1866 last_credits = oh->ot_credits;
1867 last_printed = jiffies;
1870 * XXX Limit the credits to 'max_transaction_buffers', and
1871 * let the underlying filesystem to catch the error if
1872 * we really need so many credits.
1874 * This should be removed when we can calculate the
1875 * credits precisely.
1877 oh->ot_credits = osd_transaction_size(dev);
1878 } else if (ldiskfs_track_declares_assert != 0) {
1880 * reserve few credits to prevent an assertion in JBD
1881 * our debugging mechanism will be able to detected
1882 * overuse. this can help to debug single-update
1885 oh->ot_credits += 10;
1886 if (unlikely(osd_param_is_not_sane(dev, th)))
1887 oh->ot_credits = osd_transaction_size(dev);
1890 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1891 GOTO(out, rc = -EIO);
1894 * XXX temporary stuff. Some abstraction layer should
1897 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1901 LASSERT(oti->oti_txns == 0);
1903 lu_device_get(&d->dd_lu_dev);
1904 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1915 static int osd_seq_exists(const struct lu_env *env,
1916 struct osd_device *osd, u64 seq)
1918 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1919 struct seq_server_site *ss = osd_seq_site(osd);
1924 LASSERT(ss != NULL);
1925 LASSERT(ss->ss_server_fld != NULL);
1927 rc = osd_fld_lookup(env, osd, seq, range);
1930 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1931 osd_name(osd), seq, rc);
1935 RETURN(ss->ss_node_id == range->lsr_index);
1938 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1940 struct dt_txn_commit_cb *dcb;
1941 struct dt_txn_commit_cb *tmp;
1943 /* call per-transaction stop callbacks if any */
1944 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1946 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1947 "commit callback entry: magic=%x name='%s'\n",
1948 dcb->dcb_magic, dcb->dcb_name);
1949 list_del_init(&dcb->dcb_linkage);
1950 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1955 * Concurrency: shouldn't matter.
1957 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1960 struct osd_thread_info *oti = osd_oti_get(env);
1961 struct osd_thandle *oh;
1962 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1963 struct osd_device *osd = osd_dt_dev(th->th_dev);
1964 struct qsd_instance *qsd = osd_def_qsd(osd);
1965 struct lquota_trans *qtrans;
1966 LIST_HEAD(truncates);
1967 int rc = 0, remove_agents = 0;
1971 oh = container_of(th, struct osd_thandle, ot_super);
1973 remove_agents = oh->ot_remove_agents;
1975 qtrans = oh->ot_quota_trans;
1976 oh->ot_quota_trans = NULL;
1978 /* move locks to local list, stop tx, execute truncates */
1979 list_splice(&oh->ot_trunc_locks, &truncates);
1981 if (oh->ot_handle != NULL) {
1984 handle_t *hdl = oh->ot_handle;
1987 * add commit callback
1988 * notice we don't do this in osd_trans_start()
1989 * as underlying transaction can change during truncate
1991 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1994 LASSERT(oti->oti_txns == 1);
1997 rc = dt_txn_hook_stop(env, th);
1999 CERROR("%s: failed in transaction hook: rc = %d\n",
2002 osd_trans_stop_cb(oh, rc);
2003 /* hook functions might modify th_sync */
2004 hdl->h_sync = th->th_sync;
2006 oh->ot_handle = NULL;
2007 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2009 CERROR("%s: failed to stop transaction: rc = %d\n",
2010 osd_name(osd), rc2);
2014 osd_process_truncates(&truncates);
2016 osd_trans_stop_cb(oh, th->th_result);
2020 osd_trunc_unlock_all(env, &truncates);
2022 /* inform the quota slave device that the transaction is stopping */
2023 qsd_op_end(env, qsd, qtrans);
2026 * as we want IO to journal and data IO be concurrent, we don't block
2027 * awaiting data IO completion in osd_do_bio(), instead we wait here
2028 * once transaction is submitted to the journal. all reqular requests
2029 * don't do direct IO (except read/write), thus this wait_event becomes
2032 * IMPORTANT: we have to wait till any IO submited by the thread is
2033 * completed otherwise iobuf may be corrupted by different request
2035 wait_event(iobuf->dr_wait,
2036 atomic_read(&iobuf->dr_numreqs) == 0);
2037 osd_fini_iobuf(osd, iobuf);
2039 rc = iobuf->dr_error;
2041 if (unlikely(remove_agents != 0))
2042 osd_process_scheduled_agent_removals(env, osd);
2044 oti->oti_ins_cache_depth--;
2045 /* reset OI cache for safety */
2046 if (oti->oti_ins_cache_depth == 0)
2047 oti->oti_ins_cache_used = 0;
2049 sb_end_write(osd_sb(osd));
2054 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2056 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2059 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2060 LASSERT(&dcb->dcb_func != NULL);
2061 if (dcb->dcb_flags & DCB_TRANS_STOP)
2062 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2064 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2070 * Called just before object is freed. Releases all resources except for
2071 * object itself (that is released by osd_object_free()).
2073 * Concurrency: no concurrent access is possible that late in object
2076 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2078 struct osd_object *obj = osd_obj(l);
2079 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2080 struct inode *inode = obj->oo_inode;
2085 LINVRNT(osd_invariant(obj));
2088 * If object is unlinked remove fid->ino mapping from object index.
2091 osd_index_fini(obj);
2096 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2097 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2099 uid = i_uid_read(inode);
2100 gid = i_gid_read(inode);
2101 projid = i_projid_read(inode);
2103 obj->oo_inode = NULL;
2106 /* do not rebalance quota if the caller needs to release memory
2107 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2108 * transaction and risk to deadlock - LU-12178 */
2109 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2112 if (!obj->oo_header && qsd) {
2113 struct osd_thread_info *info = osd_oti_get(env);
2114 struct lquota_id_info *qi = &info->oti_qi;
2116 /* Release granted quota to master if necessary */
2117 qi->lqi_id.qid_uid = uid;
2118 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2120 qi->lqi_id.qid_uid = gid;
2121 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2123 qi->lqi_id.qid_uid = projid;
2124 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2129 * Concurrency: ->loo_object_release() is called under site spin-lock.
2131 static void osd_object_release(const struct lu_env *env,
2132 struct lu_object *l)
2134 struct osd_object *o = osd_obj(l);
2137 * nobody should be releasing a non-destroyed object with nlink=0
2138 * the API allows this, but ldiskfs doesn't like and then report
2139 * this inode as deleted
2141 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2142 o->oo_inode->i_nlink == 0));
2146 * Concurrency: shouldn't matter.
2148 static int osd_object_print(const struct lu_env *env, void *cookie,
2149 lu_printer_t p, const struct lu_object *l)
2151 struct osd_object *o = osd_obj(l);
2152 struct iam_descr *d;
2154 if (o->oo_dir != NULL)
2155 d = o->oo_dir->od_container.ic_descr;
2158 return (*p)(env, cookie,
2159 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2161 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2162 o->oo_inode ? o->oo_inode->i_generation : 0,
2163 d ? d->id_ops->id_name : "plain");
2167 * Concurrency: shouldn't matter.
2169 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2170 struct obd_statfs *sfs, struct obd_statfs_info *info)
2172 struct osd_device *osd = osd_dt_dev(d);
2173 struct super_block *sb = osd_sb(osd);
2174 struct kstatfs *ksfs;
2178 if (unlikely(osd->od_mnt == NULL))
2179 return -EINPROGRESS;
2181 /* osd_lproc.c call this without env, allocate ksfs for that case */
2182 if (unlikely(env == NULL)) {
2183 OBD_ALLOC_PTR(ksfs);
2187 ksfs = &osd_oti_get(env)->oti_ksfs;
2190 result = sb->s_op->statfs(sb->s_root, ksfs);
2194 statfs_pack(sfs, ksfs);
2195 if (unlikely(sb->s_flags & SB_RDONLY))
2196 sfs->os_state |= OS_STATFS_READONLY;
2198 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2200 if (ldiskfs_has_feature_extents(sb))
2201 sfs->os_maxbytes = sb->s_maxbytes;
2203 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2206 * Reserve some space so to avoid fragmenting the filesystem too much.
2207 * Fragmentation not only impacts performance, but can also increase
2208 * metadata overhead significantly, causing grant calculation to be
2211 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2213 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2214 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2215 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2216 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2218 sfs->os_blocks -= reserved;
2219 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2220 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2223 if (unlikely(env == NULL))
2229 * Estimate space needed for file creations. We assume the largest filename
2230 * which is 2^64 - 1, hence a filename of 20 chars.
2231 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2233 #ifdef __LDISKFS_DIR_REC_LEN
2234 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2236 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2240 * Concurrency: doesn't access mutable data.
2242 static void osd_conf_get(const struct lu_env *env,
2243 const struct dt_device *dev,
2244 struct dt_device_param *param)
2246 struct osd_device *d = osd_dt_dev(dev);
2247 struct super_block *sb = osd_sb(d);
2248 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2253 * XXX should be taken from not-yet-existing fs abstraction layer.
2255 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2256 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2257 param->ddp_symlink_max = sb->s_blocksize;
2258 param->ddp_mount_type = LDD_MT_LDISKFS;
2259 if (ldiskfs_has_feature_extents(sb))
2260 param->ddp_maxbytes = sb->s_maxbytes;
2262 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2264 * inode are statically allocated, so per-inode space consumption
2265 * is the space consumed by the directory entry
2267 param->ddp_inodespace = PER_OBJ_USAGE;
2269 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2270 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2271 * maximum length to not under-count the actual number of extents
2272 * needed for writing a file if there are sub-optimal block allocations.
2274 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2275 /* worst-case extent insertion metadata overhead */
2276 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2277 param->ddp_mntopts = 0;
2278 if (test_opt(sb, XATTR_USER))
2279 param->ddp_mntopts |= MNTOPT_USERXATTR;
2280 if (test_opt(sb, POSIX_ACL))
2281 param->ddp_mntopts |= MNTOPT_ACL;
2284 * LOD might calculate the max stripe count based on max_ea_size,
2285 * so we need take account in the overhead as well,
2286 * xattr_header + magic + xattr_entry_head
2288 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2289 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2291 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2292 if (ldiskfs_has_feature_ea_inode(sb))
2293 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2297 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2299 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2300 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2303 * Preferred RPC size for efficient disk IO. 4MB shows good
2304 * all-around performance for ldiskfs, but use bigalloc chunk size
2305 * by default if larger.
2307 #if defined(LDISKFS_CLUSTER_SIZE)
2308 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2309 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2312 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2314 param->ddp_t10_cksum_type = 0;
2316 unsigned short interval = blk_integrity_interval(bi);
2317 name = blk_integrity_name(bi);
2325 if (strncmp(name, "T10-DIF-TYPE",
2326 sizeof("T10-DIF-TYPE") - 1) == 0) {
2327 /* also skip "1/3-" at end */
2328 const int type_off = sizeof("T10-DIF-TYPE.");
2329 char type_number = name[type_off - 2];
2331 if (interval != 512 && interval != 4096) {
2332 CERROR("%s: unsupported T10PI sector size %u\n",
2333 d->od_svname, interval);
2334 } else if (type_number != '1' && type_number != '3') {
2335 CERROR("%s: unsupported T10PI type %s\n",
2336 d->od_svname, name);
2337 } else if (strcmp(name + type_off, "CRC") == 0) {
2338 d->od_t10_type = type_number == '1' ?
2339 OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2340 param->ddp_t10_cksum_type = interval == 512 ?
2341 OBD_CKSUM_T10CRC512 :
2343 } else if (strcmp(name + type_off, "IP") == 0) {
2344 d->od_t10_type = type_number == '1' ?
2345 OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2346 param->ddp_t10_cksum_type = interval == 512 ?
2347 OBD_CKSUM_T10IP512 :
2350 CERROR("%s: unsupported checksum type of "
2352 d->od_svname, name);
2356 CERROR("%s: unsupported T10PI type '%s'",
2357 d->od_svname, name);
2362 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2364 return osd_sb(osd_dt_dev(d));
2368 * Concurrency: shouldn't matter.
2370 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2373 struct super_block *s = osd_sb(osd_dt_dev(d));
2376 down_read(&s->s_umount);
2377 rc = s->s_op->sync_fs(s, 1);
2378 up_read(&s->s_umount);
2380 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2387 * Start commit for OSD device.
2389 * An implementation of dt_commit_async method for OSD device.
2390 * Asychronously starts underlayng fs sync and thereby a transaction
2393 * \param env environment
2394 * \param d dt device
2396 * \see dt_device_operations
2398 static int osd_commit_async(const struct lu_env *env,
2399 struct dt_device *d)
2401 struct super_block *s = osd_sb(osd_dt_dev(d));
2406 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2407 down_read(&s->s_umount);
2408 rc = s->s_op->sync_fs(s, 0);
2409 up_read(&s->s_umount);
2414 static int (*priv_security_file_alloc)(struct file *file);
2416 int osd_security_file_alloc(struct file *file)
2418 if (priv_security_file_alloc)
2419 return priv_security_file_alloc(file);
2424 * Concurrency: shouldn't matter.
2426 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2428 struct super_block *sb = osd_sb(osd_dt_dev(d));
2429 struct block_device *dev = sb->s_bdev;
2430 int rc = -EOPNOTSUPP;
2434 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2435 osd_dt_dev(d)->od_svname, (long)dev, rc);
2441 * Note: we do not count into QUOTA here.
2442 * If we mount with --data_journal we may need more.
2444 const int osd_dto_credits_noquota[DTO_NR] = {
2447 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2448 * SINGLEDATA_TRANS_BLOCKS(8)
2449 * XXX Note: maybe iam need more, since iam have more level than
2452 [DTO_INDEX_INSERT] = 16,
2455 * just modify a single entry, probably merge few within a block
2457 [DTO_INDEX_DELETE] = 1,
2461 [DTO_INDEX_UPDATE] = 16,
2463 * 4(inode, inode bits, groups, GDT)
2464 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2466 [DTO_OBJECT_CREATE] = 4,
2468 * 4(inode, inode bits, groups, GDT)
2469 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2471 [DTO_OBJECT_DELETE] = 4,
2473 * Attr set credits (inode)
2475 [DTO_ATTR_SET_BASE] = 1,
2477 * Xattr set. The same as xattr of EXT3.
2478 * DATA_TRANS_BLOCKS(14)
2479 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2480 * are also counted in. Do not know why?
2482 [DTO_XATTR_SET] = 14,
2484 * credits for inode change during write.
2486 [DTO_WRITE_BASE] = 3,
2488 * credits for single block write.
2490 [DTO_WRITE_BLOCK] = 14,
2492 * Attr set credits for chown.
2493 * This is extra credits for setattr, and it is null without quota
2495 [DTO_ATTR_SET_CHOWN] = 0
2498 static const struct dt_device_operations osd_dt_ops = {
2499 .dt_root_get = osd_root_get,
2500 .dt_statfs = osd_statfs,
2501 .dt_trans_create = osd_trans_create,
2502 .dt_trans_start = osd_trans_start,
2503 .dt_trans_stop = osd_trans_stop,
2504 .dt_trans_cb_add = osd_trans_cb_add,
2505 .dt_conf_get = osd_conf_get,
2506 .dt_mnt_sb_get = osd_mnt_sb_get,
2507 .dt_sync = osd_sync,
2509 .dt_commit_async = osd_commit_async,
2512 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2515 struct osd_object *obj = osd_dt_obj(dt);
2516 struct osd_thread_info *oti = osd_oti_get(env);
2518 LINVRNT(osd_invariant(obj));
2520 LASSERT(obj->oo_owner != env);
2521 down_read_nested(&obj->oo_sem, role);
2523 LASSERT(obj->oo_owner == NULL);
2527 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2530 struct osd_object *obj = osd_dt_obj(dt);
2531 struct osd_thread_info *oti = osd_oti_get(env);
2533 LINVRNT(osd_invariant(obj));
2535 LASSERT(obj->oo_owner != env);
2536 down_write_nested(&obj->oo_sem, role);
2538 LASSERT(obj->oo_owner == NULL);
2539 obj->oo_owner = env;
2543 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2545 struct osd_object *obj = osd_dt_obj(dt);
2546 struct osd_thread_info *oti = osd_oti_get(env);
2548 LINVRNT(osd_invariant(obj));
2550 LASSERT(oti->oti_r_locks > 0);
2552 up_read(&obj->oo_sem);
2555 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2557 struct osd_object *obj = osd_dt_obj(dt);
2558 struct osd_thread_info *oti = osd_oti_get(env);
2560 LINVRNT(osd_invariant(obj));
2562 LASSERT(obj->oo_owner == env);
2563 LASSERT(oti->oti_w_locks > 0);
2565 obj->oo_owner = NULL;
2566 up_write(&obj->oo_sem);
2569 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2571 struct osd_object *obj = osd_dt_obj(dt);
2573 LINVRNT(osd_invariant(obj));
2575 return obj->oo_owner == env;
2578 static void osd_inode_getattr(const struct lu_env *env,
2579 struct inode *inode, struct lu_attr *attr)
2581 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2582 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2583 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2584 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2586 attr->la_atime = inode->i_atime.tv_sec;
2587 attr->la_mtime = inode->i_mtime.tv_sec;
2588 attr->la_ctime = inode->i_ctime.tv_sec;
2589 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2590 attr->la_mode = inode->i_mode;
2591 attr->la_size = i_size_read(inode);
2592 attr->la_blocks = inode->i_blocks;
2593 attr->la_uid = i_uid_read(inode);
2594 attr->la_gid = i_gid_read(inode);
2595 attr->la_projid = i_projid_read(inode);
2596 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2597 attr->la_nlink = inode->i_nlink;
2598 attr->la_rdev = inode->i_rdev;
2599 attr->la_blksize = 1 << inode->i_blkbits;
2600 attr->la_blkbits = inode->i_blkbits;
2602 * Ext4 did not transfer inherit flags from raw inode
2603 * to inode flags, and ext4 internally test raw inode
2604 * @i_flags directly. Instead of patching ext4, we do it here.
2606 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2607 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2610 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2613 struct osd_object *obj = osd_dt_obj(dt);
2614 const struct dt_it_ops *iops;
2620 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2621 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2623 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2624 *count = obj->oo_dirent_count;
2628 /* directory not initialized yet */
2629 if (!dt->do_index_ops) {
2634 iops = &dt->do_index_ops->dio_it;
2635 it = iops->init(env, dt, LUDA_64BITHASH);
2637 RETURN(PTR_ERR(it));
2639 rc = iops->load(env, it, 0);
2641 if (rc == -ENODATA) {
2648 rc = iops->next(env, it);
2650 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2654 if (iops->key_size(env, it) == 0)
2660 obj->oo_dirent_count = *count;
2665 iops->fini(env, it);
2670 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2671 struct lu_attr *attr)
2673 struct osd_object *obj = osd_dt_obj(dt);
2676 if (unlikely(!dt_object_exists(dt)))
2678 if (unlikely(obj->oo_destroyed))
2681 LASSERT(!dt_object_remote(dt));
2682 LINVRNT(osd_invariant(obj));
2684 spin_lock(&obj->oo_guard);
2685 osd_inode_getattr(env, obj->oo_inode, attr);
2686 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2687 attr->la_valid |= LA_FLAGS;
2688 attr->la_flags |= LUSTRE_ORPHAN_FL;
2690 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2691 attr->la_valid |= LA_FLAGS;
2692 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2694 spin_unlock(&obj->oo_guard);
2696 if (S_ISDIR(obj->oo_inode->i_mode) &&
2697 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2698 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2703 static int osd_declare_attr_qid(const struct lu_env *env,
2704 struct osd_object *obj,
2705 struct osd_thandle *oh, long long bspace,
2706 qid_t old_id, qid_t new_id, bool enforce,
2707 unsigned int type, bool ignore_edquot)
2710 struct osd_thread_info *info = osd_oti_get(env);
2711 struct lquota_id_info *qi = &info->oti_qi;
2713 qi->lqi_type = type;
2714 /* inode accounting */
2715 qi->lqi_is_blk = false;
2717 /* one more inode for the new id ... */
2718 qi->lqi_id.qid_uid = new_id;
2720 /* Reserve credits for the new id */
2721 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2722 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2727 /* and one less inode for the current id */
2728 qi->lqi_id.qid_uid = old_id;
2730 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2731 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2736 /* block accounting */
2737 qi->lqi_is_blk = true;
2739 /* more blocks for the new id ... */
2740 qi->lqi_id.qid_uid = new_id;
2741 qi->lqi_space = bspace;
2743 * Credits for the new uid has been reserved, re-use "obj"
2744 * to save credit reservation.
2746 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2747 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2752 /* and finally less blocks for the current uid */
2753 qi->lqi_id.qid_uid = old_id;
2754 qi->lqi_space = -bspace;
2755 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2756 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2762 static int osd_declare_attr_set(const struct lu_env *env,
2763 struct dt_object *dt,
2764 const struct lu_attr *attr,
2765 struct thandle *handle)
2767 struct osd_thandle *oh;
2768 struct osd_object *obj;
2777 LASSERT(dt != NULL);
2778 LASSERT(handle != NULL);
2780 obj = osd_dt_obj(dt);
2781 LASSERT(osd_invariant(obj));
2783 oh = container_of(handle, struct osd_thandle, ot_super);
2784 LASSERT(oh->ot_handle == NULL);
2786 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2787 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2789 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2790 osd_dto_credits_noquota[DTO_XATTR_SET]);
2792 if (attr == NULL || obj->oo_inode == NULL)
2795 bspace = obj->oo_inode->i_blocks << 9;
2796 bspace = toqb(bspace);
2799 * Changing ownership is always preformed by super user, it should not
2800 * fail with EDQUOT unless required explicitly.
2802 * We still need to call the osd_declare_qid() to calculate the journal
2803 * credits for updating quota accounting files and to trigger quota
2804 * space adjustment once the operation is completed.
2806 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2807 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2810 CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2811 "(the quota space is %lld)\n",
2812 obj->oo_inode->i_sb->s_id, attr->la_uid,
2813 attr->la_gid, bspace);
2816 uid = i_uid_read(obj->oo_inode);
2817 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2818 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2819 attr->la_uid, enforce, USRQUOTA,
2824 gid = i_gid_read(obj->oo_inode);
2825 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
2826 attr->la_uid, gid, attr->la_gid);
2827 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2828 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
2829 attr->la_gid, enforce, GRPQUOTA,
2835 #ifdef HAVE_PROJECT_QUOTA
2836 if (attr->la_valid & LA_PROJID) {
2837 __u32 projid = i_projid_read(obj->oo_inode);
2839 enforce = (attr->la_valid & LA_PROJID) &&
2840 (attr->la_projid != projid);
2841 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2842 (qid_t)projid, (qid_t)attr->la_projid,
2843 enforce, PRJQUOTA, true);
2851 static int osd_inode_setattr(const struct lu_env *env,
2852 struct inode *inode, const struct lu_attr *attr)
2854 __u64 bits = attr->la_valid;
2856 /* Only allow set size for regular file */
2857 if (!S_ISREG(inode->i_mode))
2858 bits &= ~(LA_SIZE | LA_BLOCKS);
2863 if (bits & LA_ATIME)
2864 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2865 if (bits & LA_CTIME)
2866 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2867 if (bits & LA_MTIME)
2868 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2869 if (bits & LA_SIZE) {
2870 spin_lock(&inode->i_lock);
2871 LDISKFS_I(inode)->i_disksize = attr->la_size;
2872 i_size_write(inode, attr->la_size);
2873 spin_unlock(&inode->i_lock);
2877 * OSD should not change "i_blocks" which is used by quota.
2878 * "i_blocks" should be changed by ldiskfs only.
2881 inode->i_mode = (inode->i_mode & S_IFMT) |
2882 (attr->la_mode & ~S_IFMT);
2884 i_uid_write(inode, attr->la_uid);
2886 i_gid_write(inode, attr->la_gid);
2887 if (bits & LA_PROJID)
2888 i_projid_write(inode, attr->la_projid);
2889 if (bits & LA_NLINK)
2890 set_nlink(inode, attr->la_nlink);
2892 inode->i_rdev = attr->la_rdev;
2894 if (bits & LA_FLAGS) {
2895 /* always keep S_NOCMTIME */
2896 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2898 #if defined(S_ENCRYPTED)
2899 /* Always remove S_ENCRYPTED, because ldiskfs must not be
2900 * aware of encryption status. It is just stored into LMA
2901 * so that it can be forwared to client side.
2903 inode->i_flags &= ~S_ENCRYPTED;
2906 * Ext4 did not transfer inherit flags from
2907 * @inode->i_flags to raw inode i_flags when writing
2908 * flags, we do it explictly here.
2910 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2911 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2913 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2918 #ifdef HAVE_PROJECT_QUOTA
2919 static int osd_transfer_project(struct inode *inode, __u32 projid,
2920 struct thandle *handle)
2922 struct super_block *sb = inode->i_sb;
2923 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2926 struct ldiskfs_iloc iloc;
2927 struct ldiskfs_inode *raw_inode;
2928 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2930 if (!ldiskfs_has_feature_project(sb)) {
2931 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2932 == LDISKFS_DEF_PROJID);
2933 if (projid != LDISKFS_DEF_PROJID)
2939 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2942 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2943 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2946 err = ldiskfs_get_inode_loc(inode, &iloc);
2950 raw_inode = ldiskfs_raw_inode(&iloc);
2951 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2952 struct osd_thandle *oh = container_of(handle,
2956 * try to expand inode size automatically.
2958 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
2959 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2967 dquot_initialize(inode);
2968 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2969 if (transfer_to[PRJQUOTA]) {
2970 err = __dquot_transfer(inode, transfer_to);
2971 dqput(transfer_to[PRJQUOTA]);
2980 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
2981 struct thandle *handle)
2985 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2986 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2990 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
2991 inode->i_ino, i_uid_read(inode), attr->la_uid,
2992 i_gid_read(inode), attr->la_gid);
2994 dquot_initialize(inode);
2996 if (attr->la_valid & LA_UID)
2997 iattr.ia_valid |= ATTR_UID;
2998 if (attr->la_valid & LA_GID)
2999 iattr.ia_valid |= ATTR_GID;
3000 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3001 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3003 rc = dquot_transfer(inode, &iattr);
3005 CERROR("%s: quota transfer failed: rc = %d. Is quota "
3006 "enforcement enabled on the ldiskfs "
3007 "filesystem?\n", inode->i_sb->s_id, rc);
3012 /* Handle project id transfer here properly */
3013 if (attr->la_valid & LA_PROJID &&
3014 attr->la_projid != i_projid_read(inode)) {
3015 #ifdef HAVE_PROJECT_QUOTA
3016 rc = osd_transfer_project(inode, attr->la_projid, handle);
3021 CERROR("%s: quota transfer failed: rc = %d. Is project "
3022 "enforcement enabled on the ldiskfs "
3023 "filesystem?\n", inode->i_sb->s_id, rc);
3030 static int osd_attr_set(const struct lu_env *env,
3031 struct dt_object *dt,
3032 const struct lu_attr *attr,
3033 struct thandle *handle)
3035 struct osd_object *obj = osd_dt_obj(dt);
3036 struct inode *inode;
3039 if (!dt_object_exists(dt))
3042 LASSERT(handle != NULL);
3043 LASSERT(!dt_object_remote(dt));
3044 LASSERT(osd_invariant(obj));
3046 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3048 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3049 !osd_obj2dev(obj)->od_is_ost) {
3050 struct osd_thread_info *oti = osd_oti_get(env);
3051 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3052 struct lu_fid *fid1 = &oti->oti_fid;
3053 struct osd_inode_id *id = &oti->oti_id;
3054 struct iam_path_descr *ipd;
3055 struct iam_container *bag;
3056 struct osd_thandle *oh;
3059 fid_cpu_to_be(fid1, fid0);
3060 memset(id, 1, sizeof(*id));
3061 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3062 fid0)->oi_dir.od_container;
3063 ipd = osd_idx_ipd_get(env, bag);
3064 if (unlikely(ipd == NULL))
3067 oh = container_of(handle, struct osd_thandle, ot_super);
3068 rc = iam_update(oh->ot_handle, bag,
3069 (const struct iam_key *)fid1,
3070 (const struct iam_rec *)id, ipd);
3071 osd_ipd_put(env, bag, ipd);
3072 return(rc > 0 ? 0 : rc);
3075 inode = obj->oo_inode;
3077 rc = osd_quota_transfer(inode, attr, handle);
3081 spin_lock(&obj->oo_guard);
3082 rc = osd_inode_setattr(env, inode, attr);
3083 spin_unlock(&obj->oo_guard);
3087 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3089 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3091 if (!(attr->la_valid & LA_FLAGS))
3094 /* Let's check if there are extra flags need to be set into LMA */
3095 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3096 struct osd_thread_info *info = osd_oti_get(env);
3097 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3099 LASSERT(!obj->oo_pfid_in_lma);
3101 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3102 &info->oti_ost_attrs);
3106 lma->lma_incompat |=
3107 lustre_to_lma_flags(attr->la_flags);
3108 lustre_lma_swab(lma);
3110 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3112 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3113 lma, sizeof(*lma), XATTR_REPLACE);
3115 struct osd_device *osd = osd_obj2dev(obj);
3117 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3118 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3119 lma->lma_incompat, rc);
3122 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3124 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3131 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3132 struct osd_object *obj,
3133 const char *name, const int namelen)
3135 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3138 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3139 umode_t mode, struct dt_allocation_hint *hint,
3140 struct thandle *th, struct lu_attr *attr)
3143 struct osd_device *osd = osd_obj2dev(obj);
3144 struct osd_thandle *oth;
3145 struct dt_object *parent = NULL;
3146 struct inode *inode;
3147 uid_t owner[2] = {0, 0};
3149 if (attr->la_valid & LA_UID)
3150 owner[0] = attr->la_uid;
3151 if (attr->la_valid & LA_GID)
3152 owner[1] = attr->la_gid;
3154 LINVRNT(osd_invariant(obj));
3155 LASSERT(obj->oo_inode == NULL);
3156 LASSERT(obj->oo_hl_head == NULL);
3158 if (S_ISDIR(mode) && ldiskfs_pdo) {
3160 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3161 if (obj->oo_hl_head == NULL)
3165 oth = container_of(th, struct osd_thandle, ot_super);
3166 LASSERT(oth->ot_handle->h_transaction != NULL);
3168 if (hint != NULL && hint->dah_parent != NULL &&
3169 !dt_object_remote(hint->dah_parent))
3170 parent = hint->dah_parent;
3172 inode = ldiskfs_create_inode(oth->ot_handle,
3173 parent ? osd_dt_obj(parent)->oo_inode :
3174 osd_sb(osd)->s_root->d_inode,
3176 if (!IS_ERR(inode)) {
3177 /* Do not update file c/mtime in ldiskfs. */
3178 inode->i_flags |= S_NOCMTIME;
3181 * For new created object, it must be consistent,
3182 * and it is unnecessary to scrub against it.
3184 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3186 obj->oo_inode = inode;
3189 if (obj->oo_hl_head != NULL) {
3190 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3191 obj->oo_hl_head = NULL;
3193 result = PTR_ERR(inode);
3195 LINVRNT(osd_invariant(obj));
3203 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3204 struct lu_attr *attr,
3205 struct dt_allocation_hint *hint,
3206 struct dt_object_format *dof,
3210 struct osd_thandle *oth;
3211 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3213 LASSERT(S_ISDIR(attr->la_mode));
3215 oth = container_of(th, struct osd_thandle, ot_super);
3216 LASSERT(oth->ot_handle->h_transaction != NULL);
3217 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3218 obj->oo_dirent_count = 0;
3219 result = osd_mkfile(info, obj, mode, hint, th, attr);
3224 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3225 struct lu_attr *attr,
3226 struct dt_allocation_hint *hint,
3227 struct dt_object_format *dof,
3231 struct osd_thandle *oth;
3232 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3234 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3236 LASSERT(S_ISREG(attr->la_mode));
3238 oth = container_of(th, struct osd_thandle, ot_super);
3239 LASSERT(oth->ot_handle->h_transaction != NULL);
3241 result = osd_mkfile(info, obj, mode, hint, th, attr);
3243 LASSERT(obj->oo_inode != NULL);
3244 if (feat->dif_flags & DT_IND_VARKEY)
3245 result = iam_lvar_create(obj->oo_inode,
3246 feat->dif_keysize_max,
3248 feat->dif_recsize_max,
3251 result = iam_lfix_create(obj->oo_inode,
3252 feat->dif_keysize_max,
3254 feat->dif_recsize_max,
3260 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3261 struct lu_attr *attr,
3262 struct dt_allocation_hint *hint,
3263 struct dt_object_format *dof,
3266 LASSERT(S_ISREG(attr->la_mode));
3267 return osd_mkfile(info, obj, (attr->la_mode &
3268 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3272 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3273 struct lu_attr *attr,
3274 struct dt_allocation_hint *hint,
3275 struct dt_object_format *dof,
3278 LASSERT(S_ISLNK(attr->la_mode));
3279 return osd_mkfile(info, obj, (attr->la_mode &
3280 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3284 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3285 struct lu_attr *attr,
3286 struct dt_allocation_hint *hint,
3287 struct dt_object_format *dof,
3290 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3293 LINVRNT(osd_invariant(obj));
3294 LASSERT(obj->oo_inode == NULL);
3295 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3296 S_ISFIFO(mode) || S_ISSOCK(mode));
3298 result = osd_mkfile(info, obj, mode, hint, th, attr);
3300 LASSERT(obj->oo_inode != NULL);
3302 * This inode should be marked dirty for i_rdev. Currently
3303 * that is done in the osd_attr_init().
3305 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3308 LINVRNT(osd_invariant(obj));
3312 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3314 struct dt_allocation_hint *hint,
3315 struct dt_object_format *dof,
3318 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3320 osd_obj_type_f result;
3336 result = osd_mk_index;
3347 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3348 struct dt_object *parent, struct dt_object *child,
3353 ah->dah_parent = parent;
3354 ah->dah_mode = child_mode;
3356 if (parent != NULL && !dt_object_remote(parent)) {
3357 /* will help to find FID->ino at dt_insert("..") */
3358 struct osd_object *pobj = osd_dt_obj(parent);
3360 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3364 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3365 struct lu_attr *attr, struct dt_object_format *dof,
3366 struct thandle *handle)
3368 struct inode *inode = obj->oo_inode;
3369 __u64 valid = attr->la_valid;
3372 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3374 if (dof->dof_type != DFT_NODE)
3375 attr->la_valid &= ~LA_RDEV;
3376 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3377 attr->la_valid &= ~LA_ATIME;
3378 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3379 attr->la_valid &= ~LA_CTIME;
3380 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3381 attr->la_valid &= ~LA_MTIME;
3383 result = osd_quota_transfer(inode, attr, handle);
3387 if (attr->la_valid != 0) {
3388 result = osd_inode_setattr(info->oti_env, inode, attr);
3390 * The osd_inode_setattr() should always succeed here. The
3391 * only error that could be returned is EDQUOT when we are
3392 * trying to change the UID or GID of the inode. However, this
3393 * should not happen since quota enforcement is no longer
3394 * enabled on ldiskfs (lquota takes care of it).
3396 LASSERTF(result == 0, "%d\n", result);
3397 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3400 attr->la_valid = valid;
3404 * Helper function for osd_create()
3406 * \retval 0, on success
3408 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3409 struct lu_attr *attr, struct dt_allocation_hint *hint,
3410 struct dt_object_format *dof, struct thandle *th)
3415 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3417 /* we drop umask so that permissions we pass are not affected */
3418 umask = current->fs->umask;
3419 current->fs->umask = 0;
3421 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3423 if (likely(obj->oo_inode != NULL)) {
3424 LASSERT(obj->oo_inode->i_state & I_NEW);
3427 * Unlock the inode before attr initialization to avoid
3428 * unnecessary dqget operations. LU-6378
3430 unlock_new_inode(obj->oo_inode);
3433 if (likely(result == 0)) {
3434 osd_attr_init(info, obj, attr, dof, th);
3435 osd_object_init0(obj);
3438 /* restore previous umask value */
3439 current->fs->umask = umask;
3441 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3447 * Helper function for osd_create()
3449 * \retval 0, on success
3451 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3452 const struct lu_fid *fid, struct thandle *th)
3454 struct osd_thread_info *info = osd_oti_get(env);
3455 struct osd_inode_id *id = &info->oti_id;
3456 struct osd_device *osd = osd_obj2dev(obj);
3457 struct osd_thandle *oh;
3460 LASSERT(obj->oo_inode != NULL);
3462 oh = container_of(th, struct osd_thandle, ot_super);
3463 LASSERT(oh->ot_handle);
3464 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3466 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3467 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3468 OI_CHECK_FLD, NULL);
3469 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3470 struct lu_fid next_fid = *fid;
3472 /* insert next object in advance, and map to the same inode */
3474 if (next_fid.f_oid != 0) {
3475 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3476 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3477 OI_CHECK_FLD, NULL);
3478 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3482 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3487 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3488 u64 seq, struct lu_seq_range *range)
3490 struct seq_server_site *ss = osd_seq_site(osd);
3492 if (fid_seq_is_idif(seq)) {
3493 fld_range_set_ost(range);
3494 range->lsr_index = idif_ost_idx(seq);
3498 if (!fid_seq_in_fldb(seq)) {
3499 fld_range_set_mdt(range);
3502 * FIXME: If ss is NULL, it suppose not get lsr_index
3505 range->lsr_index = ss->ss_node_id;
3509 LASSERT(ss != NULL);
3510 fld_range_set_any(range);
3511 /* OSD will only do local fld lookup */
3512 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3515 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3516 struct lu_attr *attr,
3517 struct dt_allocation_hint *hint,
3518 struct dt_object_format *dof,
3519 struct thandle *handle)
3521 struct osd_thandle *oh;
3526 LASSERT(handle != NULL);
3528 oh = container_of(handle, struct osd_thandle, ot_super);
3529 LASSERT(oh->ot_handle == NULL);
3532 * EA object consumes more credits than regular object: osd_mk_index
3533 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3534 * leaf_node, could involves the block, block bitmap, groups, GDT
3535 * change for each block, so add 4 * 2 credits in that case.
3537 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3538 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3539 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3541 * Reuse idle OI block may cause additional one OI block
3544 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3545 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3546 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3547 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3548 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3550 /* will help to find FID->ino mapping at dt_insert() */
3551 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3559 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3560 attr->la_projid, 1, oh, osd_dt_obj(dt),
3561 NULL, OSD_QID_INODE);
3569 * Called to destroy on-disk representation of the object
3571 * Concurrency: must be locked
3573 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3576 struct osd_object *obj = osd_dt_obj(dt);
3577 struct inode *inode = obj->oo_inode;
3578 struct osd_thandle *oh;
3586 oh = container_of(th, struct osd_thandle, ot_super);
3587 LASSERT(oh->ot_handle == NULL);
3589 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3590 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3592 /* For removing agent entry */
3593 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3594 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3597 * Recycle idle OI leaf may cause additional three OI blocks
3600 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3601 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3602 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3603 /* one less inode */
3604 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3605 i_projid_read(inode), -1, oh, obj, NULL,
3609 /* data to be truncated */
3610 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3611 i_projid_read(inode), 0, oh, obj, NULL,
3617 * will help to find FID->ino when this object is being
3620 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3625 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3628 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3629 struct osd_object *obj = osd_dt_obj(dt);
3630 struct inode *inode = obj->oo_inode;
3631 struct osd_device *osd = osd_obj2dev(obj);
3632 struct osd_thandle *oh;
3637 oh = container_of(th, struct osd_thandle, ot_super);
3638 LASSERT(oh->ot_handle);
3640 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3642 if (unlikely(fid_is_acct(fid)))
3645 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3646 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3648 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3649 osd_name(osd), PFID(fid), result);
3652 if (S_ISDIR(inode->i_mode)) {
3653 if (inode->i_nlink > 2)
3654 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3655 osd_name(osd), PFID(fid), inode->i_ino,
3658 spin_lock(&obj->oo_guard);
3660 spin_unlock(&obj->oo_guard);
3661 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3664 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3666 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3668 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3669 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3670 oh->ot_handle, OI_CHECK_FLD);
3672 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3673 /* XXX: add to ext3 orphan list */
3674 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3676 /* not needed in the cache anymore */
3677 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3678 obj->oo_destroyed = 1;
3684 * Put the fid into lustre_mdt_attrs, and then place the structure
3685 * inode's ea. This fid should not be altered during the life time
3688 * \retval +ve, on success
3689 * \retval -ve, on error
3691 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3693 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3694 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3696 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3697 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3702 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3705 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3708 lustre_loa_init(loa, fid, compat, incompat);
3709 lustre_loa_swab(loa, false);
3712 * For the OST device with 256 bytes inode size by default,
3713 * the PFID EA will be stored together with LMA EA to avoid
3714 * performance trouble. Otherwise the PFID EA can be stored
3715 * independently. LU-8998
3717 if ((compat & LMAC_FID_ON_OST) &&
3718 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3719 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3720 sizeof(*loa), XATTR_CREATE);
3722 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3723 sizeof(*lma), XATTR_CREATE);
3725 * LMA may already exist, but we need to check that all the
3726 * desired compat/incompat flags have been added.
3728 if (unlikely(rc == -EEXIST)) {
3729 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3730 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3734 if (rc < sizeof(*lma))
3737 lustre_loa_swab(loa, true);
3738 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3739 ((compat == 0 && incompat == 0) ||
3740 (!(~lma->lma_compat & compat) &&
3741 !(~lma->lma_incompat & incompat))))
3744 lma->lma_self_fid = *fid;
3745 lma->lma_compat |= compat;
3746 lma->lma_incompat |= incompat;
3747 if (rc == sizeof(*lma)) {
3748 lustre_lma_swab(lma);
3749 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3750 sizeof(*lma), XATTR_REPLACE);
3752 lustre_loa_swab(loa, false);
3753 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3754 sizeof(*loa), XATTR_REPLACE);
3762 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3763 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3764 * To have compatilibility with 1.8 ldiskfs driver we need to have
3765 * magic number at start of fid data.
3766 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3769 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3770 const struct lu_fid *fid)
3772 if (!fid_is_namespace_visible(fid) ||
3773 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3774 param->edp_magic = 0;
3778 param->edp_magic = LDISKFS_LUFID_MAGIC;
3779 param->edp_len = sizeof(struct lu_fid) + 1;
3780 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3784 * Try to read the fid from inode ea into dt_rec.
3786 * \param fid object fid.
3788 * \retval 0 on success
3790 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3791 __u32 ino, struct lu_fid *fid,
3792 struct osd_inode_id *id)
3794 struct osd_thread_info *info = osd_oti_get(env);
3795 struct inode *inode;
3799 osd_id_gen(id, ino, OSD_OII_NOGEN);
3800 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3802 RETURN(PTR_ERR(inode));
3808 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3810 struct inode *parent_dir,
3811 const struct lu_fid *dot_fid,
3812 const struct lu_fid *dot_dot_fid,
3813 struct osd_thandle *oth)
3815 struct ldiskfs_dentry_param *dot_ldp;
3816 struct ldiskfs_dentry_param *dot_dot_ldp;
3817 __u32 saved_nlink = dir->i_nlink;
3820 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3821 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3823 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3824 dot_ldp->edp_magic = 0;
3826 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3827 dir, dot_ldp, dot_dot_ldp);
3829 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3830 * the subseqent ref_add() will increase the dir->i_nlink
3831 * as 3. That is incorrect for new created directory.
3833 * It looks like hack, because we want to make the OSD API
3834 * to be order-independent for new created directory object
3835 * between dt_insert(..) and ref_add() operations.
3837 * Here, we only restore the in-RAM dir-inode's nlink attr,
3838 * becuase if the nlink attr is not 2, then there will be
3839 * ref_add() called following the dt_insert(..), such call
3840 * will make both the in-RAM and on-disk dir-inode's nlink
3841 * attr to be set as 2. LU-7447
3843 set_nlink(dir, saved_nlink);
3848 * Create an local agent inode for remote entry
3850 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3851 struct osd_device *osd,
3852 struct osd_object *pobj,
3853 const struct lu_fid *fid,
3857 struct osd_thread_info *info = osd_oti_get(env);
3858 struct inode *local;
3859 struct osd_thandle *oh;
3865 oh = container_of(th, struct osd_thandle, ot_super);
3866 LASSERT(oh->ot_handle->h_transaction != NULL);
3868 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type,
3870 if (IS_ERR(local)) {
3871 CERROR("%s: create local error %d\n", osd_name(osd),
3872 (int)PTR_ERR(local));
3877 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3878 * correct gid on remote file, not agent here
3880 local->i_gid = current_fsgid();
3881 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3883 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
3884 * That gives e2fsck something to look at and be happy, and allows
3885 * debugging if we need to determine where this symlink came from.
3887 if (S_ISLNK(type)) {
3888 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
3889 rc = snprintf((char *)LDISKFS_I(local)->i_data,
3890 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
3892 i_size_write(local, rc);
3893 LDISKFS_I(local)->i_disksize = rc;
3895 unlock_new_inode(local);
3897 /* Agent inode should not have project ID */
3898 #ifdef HAVE_PROJECT_QUOTA
3899 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3900 i_projid_read(pobj->oo_inode) != 0) {
3901 rc = osd_transfer_project(local, 0, th);
3903 CERROR("%s: quota transfer failed: rc = %d. Is project "
3904 "quota enforcement enabled on the ldiskfs "
3905 "filesystem?\n", local->i_sb->s_id, rc);
3906 RETURN(ERR_PTR(rc));
3910 /* Set special LMA flag for local agent inode */
3911 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3913 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3914 osd_name(osd), PFID(fid), rc);
3915 RETURN(ERR_PTR(rc));
3921 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3922 lu_object_fid(&pobj->oo_dt.do_lu),
3925 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3926 osd_name(osd), PFID(fid), rc);
3927 RETURN(ERR_PTR(rc));
3934 * when direntry is deleted, we have to take care of possible agent inode
3935 * referenced by that. unfortunately we can't do this at that point:
3936 * iget() within a running transaction leads to deadlock and we better do
3937 * not call that every delete declaration to save performance. so we put
3938 * a potention agent inode on a list and process that once the transaction
3939 * is over. Notice it's not any worse in terms of real orphans as regular
3940 * object destroy doesn't put inodes on the on-disk orphan list. this should
3941 * be addressed separately
3943 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3944 struct osd_thandle *oh,
3947 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3948 struct osd_obj_orphan *oor;
3955 oor->oor_env = (struct lu_env *)env;
3956 spin_lock(&osd->od_osfs_lock);
3957 list_add(&oor->oor_list, &osd->od_orphan_list);
3958 spin_unlock(&osd->od_osfs_lock);
3960 oh->ot_remove_agents = 1;
3966 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3967 struct osd_device *osd)
3969 struct osd_thread_info *info = osd_oti_get(env);
3970 struct osd_obj_orphan *oor, *tmp;
3971 struct osd_inode_id id;
3973 struct inode *inode;
3978 spin_lock(&osd->od_osfs_lock);
3979 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3980 if (oor->oor_env == env)
3981 list_move(&oor->oor_list, &list);
3983 spin_unlock(&osd->od_osfs_lock);
3985 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3989 list_del(&oor->oor_list);
3992 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3993 inode = osd_iget_fid(info, osd, &id, &fid);
3997 if (!osd_remote_fid(env, osd, &fid)) {
4002 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4004 mark_inode_dirty(inode);
4005 ldiskfs_journal_stop(jh);
4013 * OSD layer object create function for OST objects (b=11826).
4015 * The FID is inserted into inode xattr here.
4017 * \retval 0, on success
4018 * \retval -ve, on error
4020 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4021 struct lu_attr *attr, struct dt_allocation_hint *hint,
4022 struct dt_object_format *dof, struct thandle *th)
4024 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4025 struct osd_object *obj = osd_dt_obj(dt);
4026 struct osd_thread_info *info = osd_oti_get(env);
4027 int result, on_ost = 0;
4031 if (dt_object_exists(dt))
4034 LINVRNT(osd_invariant(obj));
4035 LASSERT(!dt_object_remote(dt));
4036 LASSERT(osd_is_write_locked(env, obj));
4037 LASSERT(th != NULL);
4039 if (unlikely(fid_is_acct(fid)))
4041 * Quota files can't be created from the kernel any more,
4042 * 'tune2fs -O quota' will take care of creating them
4046 result = __osd_create(info, obj, attr, hint, dof, th);
4048 if (fid_is_idif(fid) &&
4049 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4050 struct lu_fid *tfid = &info->oti_fid;
4051 struct ost_id *oi = &info->oti_ostid;
4053 fid_to_ostid(fid, oi);
4054 ostid_to_fid(tfid, oi, 0);
4056 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4057 LMAC_FID_ON_OST, 0);
4059 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4061 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4062 on_ost ? LMAC_FID_ON_OST : 0,
4065 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4066 obj->oo_dt.do_body_ops = &osd_body_ops;
4069 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY))
4070 result = __osd_oi_insert(env, obj, fid, th);
4073 * a small optimization - dt_insert() isn't usually applied
4074 * to OST objects, so we don't need to cache OI mapping for
4077 if (result == 0 && on_ost == 0) {
4078 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4080 result = osd_idc_find_and_init(env, osd, obj);
4081 LASSERT(result == 0);
4084 LASSERT(ergo(result == 0,
4085 dt_object_exists(dt) && !dt_object_remote(dt)));
4086 LINVRNT(osd_invariant(obj));
4090 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4091 struct thandle *handle)
4093 struct osd_thandle *oh;
4095 /* it's possible that object doesn't exist yet */
4096 LASSERT(handle != NULL);
4098 oh = container_of(handle, struct osd_thandle, ot_super);
4099 LASSERT(oh->ot_handle == NULL);
4101 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4102 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4104 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
4110 * Concurrency: @dt is write locked.
4112 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4115 struct osd_object *obj = osd_dt_obj(dt);
4116 struct inode *inode = obj->oo_inode;
4117 struct osd_thandle *oh;
4120 if (!dt_object_exists(dt) || obj->oo_destroyed)
4123 LINVRNT(osd_invariant(obj));
4124 LASSERT(!dt_object_remote(dt));
4125 LASSERT(osd_is_write_locked(env, obj));
4126 LASSERT(th != NULL);
4128 oh = container_of(th, struct osd_thandle, ot_super);
4129 LASSERT(oh->ot_handle != NULL);
4131 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4133 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4134 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4136 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4137 * (65000) subdirectories by storing "1" in i_nlink if the link count
4138 * would otherwise overflow. Directory tranversal tools understand
4139 * that (st_nlink == 1) indicates that the filesystem dose not track
4140 * hard links count on the directory, and will not abort subdirectory
4141 * scanning early once (st_nlink - 2) subdirs have been found.
4143 * This also has to properly handle the case of inodes with nlink == 0
4144 * in case they are being linked into the PENDING directory
4146 spin_lock(&obj->oo_guard);
4147 if (unlikely(inode->i_nlink == 0))
4148 /* inc_nlink from 0 may cause WARN_ON */
4149 set_nlink(inode, 1);
4151 ldiskfs_inc_count(oh->ot_handle, inode);
4152 if (!S_ISDIR(inode->i_mode))
4153 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4155 spin_unlock(&obj->oo_guard);
4157 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4158 LINVRNT(osd_invariant(obj));
4160 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4165 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4166 struct thandle *handle)
4168 struct osd_thandle *oh;
4170 if (!dt_object_exists(dt))
4173 LASSERT(!dt_object_remote(dt));
4174 LASSERT(handle != NULL);
4176 oh = container_of(handle, struct osd_thandle, ot_super);
4177 LASSERT(oh->ot_handle == NULL);
4179 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4180 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4186 * Concurrency: @dt is write locked.
4188 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4191 struct osd_object *obj = osd_dt_obj(dt);
4192 struct inode *inode = obj->oo_inode;
4193 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4194 struct osd_thandle *oh;
4196 if (!dt_object_exists(dt))
4199 LINVRNT(osd_invariant(obj));
4200 LASSERT(!dt_object_remote(dt));
4201 LASSERT(osd_is_write_locked(env, obj));
4202 LASSERT(th != NULL);
4204 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4207 oh = container_of(th, struct osd_thandle, ot_super);
4208 LASSERT(oh->ot_handle != NULL);
4210 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4212 spin_lock(&obj->oo_guard);
4214 * That can be result of upgrade from old Lustre version and
4215 * applied only to local files. Just skip this ref_del call.
4216 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4218 if (inode->i_nlink == 0) {
4219 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4220 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4221 ", maybe an upgraded file? (LU-3915)\n",
4222 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4223 spin_unlock(&obj->oo_guard);
4227 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4228 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4230 ldiskfs_dec_count(oh->ot_handle, inode);
4231 spin_unlock(&obj->oo_guard);
4233 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4234 LINVRNT(osd_invariant(obj));
4236 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4242 * Concurrency: @dt is read locked.
4244 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4245 struct lu_buf *buf, const char *name)
4247 struct osd_object *obj = osd_dt_obj(dt);
4248 struct inode *inode = obj->oo_inode;
4249 struct osd_thread_info *info = osd_oti_get(env);
4250 struct dentry *dentry = &info->oti_obj_dentry;
4251 bool cache_xattr = false;
4256 /* version get is not real XATTR but uses xattr API */
4257 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4258 dt_obj_version_t *ver = buf->lb_buf;
4261 * for version we are just using xattr API but change inode
4264 if (buf->lb_len == 0)
4265 return sizeof(dt_obj_version_t);
4267 if (buf->lb_len < sizeof(dt_obj_version_t))
4270 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4271 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4273 *ver = LDISKFS_I(inode)->i_fs_version;
4275 return sizeof(dt_obj_version_t);
4278 if (!dt_object_exists(dt))
4281 LASSERT(!dt_object_remote(dt));
4282 LASSERT(inode->i_op != NULL);
4283 #ifdef HAVE_IOP_XATTR
4284 LASSERT(inode->i_op->getxattr != NULL);
4287 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4288 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4292 rc = osd_oxc_get(obj, name, buf);
4297 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4298 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4299 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4300 struct filter_fid *ff;
4301 struct ost_layout *ol;
4303 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4307 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4310 if (buf->lb_len == 0 || !buf->lb_buf)
4313 if (buf->lb_len < rc)
4317 ol = &ff->ff_layout;
4318 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4319 PFID_STRIPE_IDX_BITS);
4320 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4321 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4322 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4323 if (lma->lma_compat & LMAC_COMP_INFO) {
4324 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4325 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4326 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4328 ol->ol_comp_start = 0;
4329 ol->ol_comp_end = 0;
4333 rc = __osd_xattr_get(inode, dentry, name,
4334 buf->lb_buf, buf->lb_len);
4338 if (rc == -ENOENT || rc == -ENODATA)
4339 osd_oxc_add(obj, name, NULL, 0);
4340 else if (rc > 0 && buf->lb_buf != NULL)
4341 osd_oxc_add(obj, name, buf->lb_buf, rc);
4347 static int osd_declare_xattr_set(const struct lu_env *env,
4348 struct dt_object *dt,
4349 const struct lu_buf *buf, const char *name,
4350 int fl, struct thandle *handle)
4352 struct osd_thandle *oh;
4354 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4356 LASSERT(handle != NULL);
4358 oh = container_of(handle, struct osd_thandle, ot_super);
4359 LASSERT(oh->ot_handle == NULL);
4361 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4363 * For non-upgrading case, the LMA is set first and
4364 * usually fit inode. But for upgrade case, the LMA
4365 * may be in another separated EA block.
4367 if (dt_object_exists(dt)) {
4368 if (fl == LU_XATTR_REPLACE)
4373 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4375 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4376 /* We may need to delete the old PFID EA. */
4377 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4378 if (fl == LU_XATTR_REPLACE)
4384 * If some name entry resides on remote MDT, then will create
4385 * agent entry under remote parent. On the other hand, if the
4386 * remote entry will be removed, then related agent entry may
4387 * need to be removed from the remote parent. So there may be
4388 * kinds of cases, let's declare enough credits. The credits
4389 * for create agent entry is enough for remove case.
4391 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4392 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4393 if (dt_object_exists(dt))
4394 credits += 1; /* For updating LMA */
4398 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4403 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4405 * learn xattr size from osd_xattr_get if
4406 * attribute has not been read yet
4408 buflen = __osd_xattr_get(
4409 osd_dt_obj(dt)->oo_inode,
4410 &osd_oti_get(env)->oti_obj_dentry,
4415 buflen = buf->lb_len;
4418 if (buflen > sb->s_blocksize) {
4419 credits += osd_calc_bkmap_credits(
4421 (buflen + sb->s_blocksize - 1) >>
4422 sb->s_blocksize_bits);
4426 * xattr set may involve inode quota change, reserve credits for
4427 * dquot_initialize()
4429 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4432 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4437 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4438 const struct lu_buf *buf, int fl,
4439 struct thandle *handle)
4441 struct osd_thread_info *info = osd_oti_get(env);
4442 struct dentry *dentry = &info->oti_obj_dentry;
4443 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4444 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4445 struct inode *inode = obj->oo_inode;
4446 struct filter_fid *ff = buf->lb_buf;
4447 struct ost_layout *ol = &ff->ff_layout;
4448 int flags = XATTR_REPLACE;
4453 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4456 rc = osd_get_lma(info, inode, dentry, loa);
4457 if (rc == -ENODATA) {
4458 /* Usually for upgarding from old device */
4459 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4460 LMAC_FID_ON_OST, 0);
4461 flags = XATTR_CREATE;
4466 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4467 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4470 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4471 /* Separate PFID EA from LMA */
4472 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4473 lustre_lma_swab(lma);
4474 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4475 sizeof(*lma), XATTR_REPLACE);
4477 obj->oo_pfid_in_lma = 0;
4478 rc = LU_XATTR_CREATE;
4484 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4488 * Old client does not send stripe information,
4489 * then store the PFID EA on disk separatedly.
4491 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4492 ol->ol_stripe_size == 0))
4495 /* Remove old PFID EA entry firstly. */
4496 dquot_initialize(inode);
4497 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4498 if (rc == -ENODATA) {
4499 if ((fl & LU_XATTR_REPLACE) && !(fl & LU_XATTR_CREATE))
4506 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4507 if (likely(ol->ol_stripe_size != 0)) {
4508 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4509 PFID_STRIPE_IDX_BITS;
4510 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4511 lma->lma_compat |= LMAC_STRIPE_INFO;
4512 if (ol->ol_comp_id != 0) {
4513 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4514 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4515 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4516 lma->lma_compat |= LMAC_COMP_INFO;
4520 lustre_loa_swab(loa, false);
4522 /* Store the PFID EA inside LMA. */
4523 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4526 obj->oo_pfid_in_lma = 1;
4532 * In DNE environment, the object (in spite of regular file or directory)
4533 * and its name entry may reside on different MDTs. Under such case, we will
4534 * create an agent entry on the MDT where the object resides. The agent entry
4535 * references the object locally, that makes the object to be visible to the
4536 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4537 * such as 'tar' can handle the object properly.
4539 * We handle the agent entry during set linkEA that is the common interface
4540 * for both regular file and directroy, can handle kinds of cases, such as
4541 * create/link/unlink/rename, and so on.
4543 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4545 * XXX: There are two known issues:
4546 * 1. For one object, we will create at most one agent entry even if there
4547 * may be more than one cross-MDTs hard links on the object. So the local
4548 * e2fsck may claim that the object's nlink is larger than the name entries
4549 * that reference such inode. And in further, the e2fsck will fix the nlink
4550 * attribute to match the local references. Then it will cause the object's
4551 * nlink attribute to be inconsistent with the global references. it is bad
4552 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4553 * other hand, the global namespace LFSCK can repair the object's attribute
4554 * according to the linkEA.
4555 * 2. There may be too many hard links on the object as to its linkEA overflow,
4556 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4557 * case happened, then at this point, we do not know whether there are some
4558 * cross-MDTs reference. But there are local references, it guarantees that
4559 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4561 static int osd_xattr_handle_linkea(const struct lu_env *env,
4562 struct osd_device *osd,
4563 struct osd_object *obj,
4564 const struct lu_buf *buf,
4565 struct thandle *handle)
4567 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4568 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4569 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4570 struct lu_name tmpname;
4571 struct osd_thandle *oh;
4573 bool remote = false;
4577 oh = container_of(handle, struct osd_thandle, ot_super);
4578 LASSERT(oh->ot_handle != NULL);
4580 rc = linkea_init_with_rec(&ldata);
4582 linkea_first_entry(&ldata);
4583 while (ldata.ld_lee != NULL && !remote) {
4584 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4586 if (osd_remote_fid(env, osd, tfid) > 0)
4589 linkea_next_entry(&ldata);
4591 } else if (rc == -ENODATA) {
4597 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4598 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4600 CERROR("%s: failed to remove agent entry for "DFID
4601 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4602 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4603 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4605 CERROR("%s: failed to create agent entry for "DFID
4606 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4613 * Concurrency: @dt is write locked.
4615 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4616 const struct lu_buf *buf, const char *name, int fl,
4617 struct thandle *handle)
4619 struct osd_object *obj = osd_dt_obj(dt);
4620 struct osd_device *osd = osd_obj2dev(obj);
4621 struct inode *inode = obj->oo_inode;
4622 struct osd_thread_info *info = osd_oti_get(env);
4632 /* version set is not real XATTR */
4633 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4634 dt_obj_version_t *version = buf->lb_buf;
4637 * for version we are just using xattr API but change inode
4640 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4642 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4643 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4645 LDISKFS_I(inode)->i_fs_version = *version;
4647 * Version is set after all inode operations are finished,
4648 * so we should mark it dirty here
4650 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4655 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4656 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4659 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4662 * For the OST device with 256 bytes inode size by default,
4663 * the PFID EA will be stored together with LMA EA to avoid
4664 * performance trouble. Otherwise the PFID EA can be stored
4665 * independently. LU-8998
4667 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4668 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4669 LASSERT(buf->lb_buf);
4671 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4674 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4675 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4676 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4678 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4682 lma->lma_incompat |= LMAI_STRIPED;
4683 lustre_lma_swab(lma);
4684 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4685 sizeof(*lma), XATTR_REPLACE);
4688 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4689 LASSERT(!osd->od_is_ost);
4691 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4696 if (fl & LU_XATTR_REPLACE)
4697 fs_flags |= XATTR_REPLACE;
4699 if (fl & LU_XATTR_CREATE)
4700 fs_flags |= XATTR_CREATE;
4702 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4703 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4706 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4707 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4708 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4714 * Concurrency: @dt is read locked.
4716 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4717 const struct lu_buf *buf)
4719 struct osd_object *obj = osd_dt_obj(dt);
4720 struct inode *inode = obj->oo_inode;
4721 struct osd_thread_info *info = osd_oti_get(env);
4722 struct dentry *dentry = &info->oti_obj_dentry;
4724 if (!dt_object_exists(dt))
4727 LASSERT(!dt_object_remote(dt));
4728 LASSERT(inode->i_op != NULL);
4729 LASSERT(inode->i_op->listxattr != NULL);
4731 dentry->d_inode = inode;
4732 dentry->d_sb = inode->i_sb;
4733 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4736 static int osd_declare_xattr_del(const struct lu_env *env,
4737 struct dt_object *dt, const char *name,
4738 struct thandle *handle)
4740 struct osd_thandle *oh;
4741 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4743 LASSERT(!dt_object_remote(dt));
4744 LASSERT(handle != NULL);
4746 oh = container_of(handle, struct osd_thandle, ot_super);
4747 LASSERT(oh->ot_handle == NULL);
4749 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4750 osd_dto_credits_noquota[DTO_XATTR_SET]);
4752 * xattr del may involve inode quota change, reserve credits for
4753 * dquot_initialize()
4755 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4761 * Concurrency: @dt is write locked.
4763 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4764 const char *name, struct thandle *handle)
4766 struct osd_object *obj = osd_dt_obj(dt);
4767 struct inode *inode = obj->oo_inode;
4768 struct osd_thread_info *info = osd_oti_get(env);
4769 struct dentry *dentry = &info->oti_obj_dentry;
4772 if (!dt_object_exists(dt))
4775 LASSERT(!dt_object_remote(dt));
4776 LASSERT(inode->i_op != NULL);
4777 LASSERT(handle != NULL);
4778 #ifdef HAVE_IOP_XATTR
4779 LASSERT(inode->i_op->removexattr != NULL);
4782 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4784 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4785 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4787 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4788 &info->oti_ost_attrs);
4790 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4792 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4793 lustre_lma_swab(lma);
4794 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4795 sizeof(*lma), XATTR_REPLACE);
4797 obj->oo_pfid_in_lma = 0;
4800 dquot_initialize(inode);
4801 dentry->d_inode = inode;
4802 dentry->d_sb = inode->i_sb;
4803 rc = ll_vfs_removexattr(dentry, inode, name);
4806 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4809 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4810 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4811 osd_oxc_del(obj, name);
4816 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4817 __u64 start, __u64 end)
4819 struct osd_object *obj = osd_dt_obj(dt);
4820 struct inode *inode = obj->oo_inode;
4821 struct file *file = osd_quasi_file(env, inode);
4826 rc = vfs_fsync_range(file, start, end, 0);
4831 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4840 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4841 const struct dt_index_features *feat)
4843 struct iam_descr *descr;
4845 if (osd_object_is_root(o))
4846 return feat == &dt_directory_features;
4848 LASSERT(o->oo_dir != NULL);
4850 descr = o->oo_dir->od_container.ic_descr;
4851 if (feat == &dt_directory_features) {
4852 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4857 return feat->dif_keysize_min <= descr->id_key_size &&
4858 descr->id_key_size <= feat->dif_keysize_max &&
4859 feat->dif_recsize_min <= descr->id_rec_size &&
4860 descr->id_rec_size <= feat->dif_recsize_max &&
4861 !(feat->dif_flags & (DT_IND_VARKEY |
4862 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4863 ergo(feat->dif_flags & DT_IND_UPDATE,
4864 1 /* XXX check that object (and fs) is writable */);
4868 static int osd_iam_container_init(const struct lu_env *env,
4869 struct osd_object *obj,
4870 struct osd_directory *dir)
4872 struct iam_container *bag = &dir->od_container;
4875 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4879 result = iam_container_setup(bag);
4881 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4883 iam_container_fini(bag);
4890 * Concurrency: no external locking is necessary.
4892 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4893 const struct dt_index_features *feat)
4897 struct osd_object *obj = osd_dt_obj(dt);
4899 LINVRNT(osd_invariant(obj));
4901 if (osd_object_is_root(obj)) {
4902 dt->do_index_ops = &osd_index_ea_ops;
4904 } else if (feat == &dt_directory_features) {
4905 dt->do_index_ops = &osd_index_ea_ops;
4906 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4911 } else if (unlikely(feat == &dt_otable_features)) {
4912 dt->do_index_ops = &osd_otable_ops;
4914 } else if (unlikely(feat == &dt_acct_features)) {
4915 dt->do_index_ops = &osd_acct_index_ops;
4918 } else if (!osd_has_index(obj)) {
4919 struct osd_directory *dir;
4920 struct osd_device *osd = osd_obj2dev(obj);
4921 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4926 spin_lock(&obj->oo_guard);
4927 if (obj->oo_dir == NULL)
4931 * Concurrent thread allocated container data.
4934 spin_unlock(&obj->oo_guard);
4936 * Now, that we have container data, serialize its
4939 down_write(&obj->oo_ext_idx_sem);
4941 * recheck under lock.
4944 if (osd_has_index(obj)) {
4949 result = osd_iam_container_init(env, obj, obj->oo_dir);
4950 if (result || feat == &dt_lfsck_namespace_features ||
4951 feat == &dt_lfsck_layout_orphan_features ||
4952 feat == &dt_lfsck_layout_dangling_features)
4955 result = osd_index_register(osd, fid,
4956 feat->dif_keysize_max,
4957 feat->dif_recsize_max);
4959 CWARN("%s: failed to register index "
4961 osd_name(osd), PFID(fid), result);
4962 else if (result > 0)
4965 CDEBUG(D_LFSCK, "%s: index object "DFID
4966 " (%d/%d) registered\n",
4967 osd_name(osd), PFID(fid),
4968 (int)feat->dif_keysize_max,
4969 (int)feat->dif_recsize_max);
4972 up_write(&obj->oo_ext_idx_sem);
4980 if (result == 0 && skip_iam == 0) {
4981 if (!osd_iam_index_probe(env, obj, feat))
4984 LINVRNT(osd_invariant(obj));
4989 static int osd_otable_it_attr_get(const struct lu_env *env,
4990 struct dt_object *dt,
4991 struct lu_attr *attr)
4997 static const struct dt_object_operations osd_obj_ops = {
4998 .do_read_lock = osd_read_lock,
4999 .do_write_lock = osd_write_lock,
5000 .do_read_unlock = osd_read_unlock,
5001 .do_write_unlock = osd_write_unlock,
5002 .do_write_locked = osd_write_locked,
5003 .do_attr_get = osd_attr_get,
5004 .do_declare_attr_set = osd_declare_attr_set,
5005 .do_attr_set = osd_attr_set,
5006 .do_ah_init = osd_ah_init,
5007 .do_declare_create = osd_declare_create,
5008 .do_create = osd_create,
5009 .do_declare_destroy = osd_declare_destroy,
5010 .do_destroy = osd_destroy,
5011 .do_index_try = osd_index_try,
5012 .do_declare_ref_add = osd_declare_ref_add,
5013 .do_ref_add = osd_ref_add,
5014 .do_declare_ref_del = osd_declare_ref_del,
5015 .do_ref_del = osd_ref_del,
5016 .do_xattr_get = osd_xattr_get,
5017 .do_declare_xattr_set = osd_declare_xattr_set,
5018 .do_xattr_set = osd_xattr_set,
5019 .do_declare_xattr_del = osd_declare_xattr_del,
5020 .do_xattr_del = osd_xattr_del,
5021 .do_xattr_list = osd_xattr_list,
5022 .do_object_sync = osd_object_sync,
5023 .do_invalidate = osd_invalidate,
5026 static const struct dt_object_operations osd_obj_otable_it_ops = {
5027 .do_attr_get = osd_otable_it_attr_get,
5028 .do_index_try = osd_index_try,
5031 static int osd_index_declare_iam_delete(const struct lu_env *env,
5032 struct dt_object *dt,
5033 const struct dt_key *key,
5034 struct thandle *handle)
5036 struct osd_thandle *oh;
5038 oh = container_of(handle, struct osd_thandle, ot_super);
5039 LASSERT(oh->ot_handle == NULL);
5041 /* Recycle may cause additional three blocks to be changed. */
5042 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5043 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5049 * delete a (key, value) pair from index \a dt specified by \a key
5051 * \param dt osd index object
5052 * \param key key for index
5053 * \param rec record reference
5054 * \param handle transaction handler
5057 * \retval -ve failure
5059 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5060 const struct dt_key *key,
5061 struct thandle *handle)
5063 struct osd_thread_info *oti = osd_oti_get(env);
5064 struct osd_object *obj = osd_dt_obj(dt);
5065 struct osd_thandle *oh;
5066 struct iam_path_descr *ipd;
5067 struct iam_container *bag = &obj->oo_dir->od_container;
5072 if (!dt_object_exists(dt))
5075 LINVRNT(osd_invariant(obj));
5076 LASSERT(!dt_object_remote(dt));
5077 LASSERT(bag->ic_object == obj->oo_inode);
5078 LASSERT(handle != NULL);
5080 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5082 ipd = osd_idx_ipd_get(env, bag);
5083 if (unlikely(ipd == NULL))
5086 oh = container_of(handle, struct osd_thandle, ot_super);
5087 LASSERT(oh->ot_handle != NULL);
5088 LASSERT(oh->ot_handle->h_transaction != NULL);
5090 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5091 /* swab quota uid/gid provided by caller */
5092 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5093 key = (const struct dt_key *)&oti->oti_quota_id;
5096 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5097 osd_ipd_put(env, bag, ipd);
5098 LINVRNT(osd_invariant(obj));
5099 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5103 static int osd_index_declare_ea_delete(const struct lu_env *env,
5104 struct dt_object *dt,
5105 const struct dt_key *key,
5106 struct thandle *handle)
5108 struct osd_thandle *oh;
5109 struct inode *inode;
5114 LASSERT(!dt_object_remote(dt));
5115 LASSERT(handle != NULL);
5117 oh = container_of(handle, struct osd_thandle, ot_super);
5118 LASSERT(oh->ot_handle == NULL);
5120 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5121 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5123 inode = osd_dt_obj(dt)->oo_inode;
5127 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5128 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5133 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5136 struct osd_fid_pack *rec;
5139 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5140 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5141 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5142 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5148 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5149 const struct lu_fid *fid)
5151 struct seq_server_site *ss = osd_seq_site(osd);
5155 /* FID seqs not in FLDB, must be local seq */
5156 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5160 * If FLD is not being initialized yet, it only happens during the
5161 * initialization, likely during mgs initialization, and we assume
5162 * this is local FID.
5164 if (ss == NULL || ss->ss_server_fld == NULL)
5167 /* Only check the local FLDB here */
5168 if (osd_seq_exists(env, osd, fid_seq(fid)))
5174 static void osd_take_care_of_agent(const struct lu_env *env,
5175 struct osd_device *osd,
5176 struct osd_thandle *oh,
5177 struct ldiskfs_dir_entry_2 *de)
5179 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5180 struct osd_idmap_cache *idc;
5181 int rc, schedule = 0;
5183 LASSERT(de != NULL);
5185 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5186 if (likely(rc == 0)) {
5187 idc = osd_idc_find_or_init(env, osd, fid);
5188 if (IS_ERR(idc) || idc->oic_remote)
5190 } else if (rc == -ENODATA) {
5192 * can't get FID, postpone to the end of the
5193 * transaction when iget() is safe
5197 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5200 osd_schedule_agent_inode_removal(env, oh,
5201 le32_to_cpu(de->inode));
5205 * Index delete function for interoperability mode (b11826).
5206 * It will remove the directory entry added by osd_index_ea_insert().
5207 * This entry is needed to maintain name->fid mapping.
5209 * \param key, key i.e. file entry to be deleted
5211 * \retval 0, on success
5212 * \retval -ve, on error
5214 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5215 const struct dt_key *key, struct thandle *handle)
5217 struct osd_object *obj = osd_dt_obj(dt);
5218 struct inode *dir = obj->oo_inode;
5219 struct dentry *dentry;
5220 struct osd_thandle *oh;
5221 struct ldiskfs_dir_entry_2 *de = NULL;
5222 struct buffer_head *bh;
5223 struct htree_lock *hlock = NULL;
5224 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5229 if (!dt_object_exists(dt))
5232 LINVRNT(osd_invariant(obj));
5233 LASSERT(!dt_object_remote(dt));
5234 LASSERT(handle != NULL);
5236 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5238 oh = container_of(handle, struct osd_thandle, ot_super);
5239 LASSERT(oh->ot_handle != NULL);
5240 LASSERT(oh->ot_handle->h_transaction != NULL);
5242 dquot_initialize(dir);
5243 dentry = osd_child_dentry_get(env, obj,
5244 (char *)key, strlen((char *)key));
5246 if (obj->oo_hl_head != NULL) {
5247 hlock = osd_oti_get(env)->oti_hlock;
5248 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5249 dir, LDISKFS_HLOCK_DEL);
5251 down_write(&obj->oo_ext_idx_sem);
5254 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5257 * If this is not the ".." entry, it might be a remote DNE
5258 * entry and we need to check if the FID is for a remote
5259 * MDT. If the FID is not in the directory entry (e.g.
5260 * upgraded 1.8 filesystem without dirdata enabled) then
5261 * we need to get the FID from the LMA. For a remote directory
5262 * there HAS to be an LMA, it cannot be an IGIF inode in this
5265 * Delete the entry before the agent inode in order to
5266 * simplify error handling. At worst an error after deleting
5267 * the entry first might leak the agent inode afterward. The
5268 * reverse would need filesystem abort in case of error deleting
5269 * the entry after the agent had been removed, or leave a
5270 * dangling entry pointing at a random inode.
5272 if (strcmp((char *)key, dotdot) != 0)
5273 osd_take_care_of_agent(env, osd, oh, de);
5274 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5280 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5281 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5282 /* NB, dirent count may not be accurate, because it's counted
5285 if (obj->oo_dirent_count)
5286 obj->oo_dirent_count--;
5288 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5291 ldiskfs_htree_unlock(hlock);
5293 up_write(&obj->oo_ext_idx_sem);
5296 LASSERT(osd_invariant(obj));
5297 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5302 * Lookup index for \a key and copy record to \a rec.
5304 * \param dt osd index object
5305 * \param key key for index
5306 * \param rec record reference
5308 * \retval +ve success : exact mach
5309 * \retval 0 return record with key not greater than \a key
5310 * \retval -ve failure
5312 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5313 struct dt_rec *rec, const struct dt_key *key)
5315 struct osd_object *obj = osd_dt_obj(dt);
5316 struct iam_path_descr *ipd;
5317 struct iam_container *bag = &obj->oo_dir->od_container;
5318 struct osd_thread_info *oti = osd_oti_get(env);
5319 struct iam_iterator *it = &oti->oti_idx_it;
5320 struct iam_rec *iam_rec;
5325 if (!dt_object_exists(dt))
5328 LASSERT(osd_invariant(obj));
5329 LASSERT(!dt_object_remote(dt));
5330 LASSERT(bag->ic_object == obj->oo_inode);
5332 ipd = osd_idx_ipd_get(env, bag);
5336 /* got ipd now we can start iterator. */
5337 iam_it_init(it, bag, 0, ipd);
5339 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5340 /* swab quota uid/gid provided by caller */
5341 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5342 key = (const struct dt_key *)&oti->oti_quota_id;
5345 rc = iam_it_get(it, (struct iam_key *)key);
5347 if (S_ISDIR(obj->oo_inode->i_mode))
5348 iam_rec = (struct iam_rec *)oti->oti_ldp;
5350 iam_rec = (struct iam_rec *)rec;
5352 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5354 if (S_ISDIR(obj->oo_inode->i_mode))
5355 osd_fid_unpack((struct lu_fid *)rec,
5356 (struct osd_fid_pack *)iam_rec);
5357 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5358 osd_quota_unpack(obj, rec);
5363 osd_ipd_put(env, bag, ipd);
5365 LINVRNT(osd_invariant(obj));
5370 static int osd_index_declare_iam_insert(const struct lu_env *env,
5371 struct dt_object *dt,
5372 const struct dt_rec *rec,
5373 const struct dt_key *key,
5374 struct thandle *handle)
5376 struct osd_thandle *oh;
5378 LASSERT(handle != NULL);
5380 oh = container_of(handle, struct osd_thandle, ot_super);
5381 LASSERT(oh->ot_handle == NULL);
5383 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5384 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5390 * Inserts (key, value) pair in \a dt index object.
5392 * \param dt osd index object
5393 * \param key key for index
5394 * \param rec record reference
5395 * \param th transaction handler
5398 * \retval -ve failure
5400 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5401 const struct dt_rec *rec,
5402 const struct dt_key *key, struct thandle *th)
5404 struct osd_object *obj = osd_dt_obj(dt);
5405 struct iam_path_descr *ipd;
5406 struct osd_thandle *oh;
5407 struct iam_container *bag;
5408 struct osd_thread_info *oti = osd_oti_get(env);
5409 struct iam_rec *iam_rec;
5414 if (!dt_object_exists(dt))
5417 LINVRNT(osd_invariant(obj));
5418 LASSERT(!dt_object_remote(dt));
5420 bag = &obj->oo_dir->od_container;
5421 LASSERT(bag->ic_object == obj->oo_inode);
5422 LASSERT(th != NULL);
5424 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5426 ipd = osd_idx_ipd_get(env, bag);
5427 if (unlikely(ipd == NULL))
5430 oh = container_of(th, struct osd_thandle, ot_super);
5431 LASSERT(oh->ot_handle != NULL);
5432 LASSERT(oh->ot_handle->h_transaction != NULL);
5433 if (S_ISDIR(obj->oo_inode->i_mode)) {
5434 iam_rec = (struct iam_rec *)oti->oti_ldp;
5435 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5437 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5438 /* pack quota uid/gid */
5439 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5440 key = (const struct dt_key *)&oti->oti_quota_id;
5441 /* pack quota record */
5442 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5443 iam_rec = (struct iam_rec *)rec;
5445 iam_rec = (struct iam_rec *)rec;
5448 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5450 osd_ipd_put(env, bag, ipd);
5451 LINVRNT(osd_invariant(obj));
5452 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5457 * Calls ldiskfs_add_entry() to add directory entry
5458 * into the directory. This is required for
5459 * interoperability mode (b11826)
5461 * \retval 0, on success
5462 * \retval -ve, on error
5464 static int __osd_ea_add_rec(struct osd_thread_info *info,
5465 struct osd_object *pobj, struct inode *cinode,
5466 const char *name, const struct lu_fid *fid,
5467 struct htree_lock *hlock, struct thandle *th)
5469 struct ldiskfs_dentry_param *ldp;
5470 struct dentry *child;
5471 struct osd_thandle *oth;
5474 oth = container_of(th, struct osd_thandle, ot_super);
5475 LASSERT(oth->ot_handle != NULL);
5476 LASSERT(oth->ot_handle->h_transaction != NULL);
5477 LASSERT(pobj->oo_inode);
5479 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5480 if (unlikely(pobj->oo_inode ==
5481 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
5484 osd_get_ldiskfs_dirent_param(ldp, fid);
5485 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5486 child->d_fsdata = (void *)ldp;
5487 dquot_initialize(pobj->oo_inode);
5488 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5489 child, cinode, hlock);
5490 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5491 struct ldiskfs_dir_entry_2 *de;
5492 struct buffer_head *bh;
5495 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5498 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5501 if (S_ISDIR(cinode->i_mode))
5502 de->file_type = LDISKFS_DIRENT_LUFID |
5503 LDISKFS_FT_REG_FILE;
5505 de->file_type = LDISKFS_DIRENT_LUFID |
5507 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5518 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5519 * into the directory.Also sets flags into osd object to
5520 * indicate dot and dotdot are created. This is required for
5521 * interoperability mode (b11826)
5523 * \param dir directory for dot and dotdot fixup.
5524 * \param obj child object for linking
5526 * \retval 0, on success
5527 * \retval -ve, on error
5529 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5530 struct osd_object *dir,
5531 struct inode *parent_dir, const char *name,
5532 const struct lu_fid *dot_fid,
5533 const struct lu_fid *dot_dot_fid,
5536 struct inode *inode = dir->oo_inode;
5537 struct osd_thandle *oth;
5540 oth = container_of(th, struct osd_thandle, ot_super);
5541 LASSERT(oth->ot_handle->h_transaction != NULL);
5542 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5544 if (strcmp(name, dot) == 0) {
5545 if (dir->oo_compat_dot_created) {
5548 LASSERT(inode->i_ino == parent_dir->i_ino);
5549 dir->oo_compat_dot_created = 1;
5552 } else if (strcmp(name, dotdot) == 0) {
5553 if (!dir->oo_compat_dot_created)
5555 /* in case of rename, dotdot is already created */
5556 if (dir->oo_compat_dotdot_created) {
5557 return __osd_ea_add_rec(info, dir, parent_dir, name,
5558 dot_dot_fid, NULL, th);
5561 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5562 struct lu_fid tfid = *dot_dot_fid;
5565 result = osd_add_dot_dotdot_internal(info,
5566 dir->oo_inode, parent_dir, dot_fid,
5569 result = osd_add_dot_dotdot_internal(info,
5570 dir->oo_inode, parent_dir, dot_fid,
5575 dir->oo_compat_dotdot_created = 1;
5583 * It will call the appropriate osd_add* function and return the
5584 * value, return by respective functions.
5586 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5587 struct inode *cinode, const char *name,
5588 const struct lu_fid *fid, struct thandle *th)
5590 struct osd_thread_info *info = osd_oti_get(env);
5591 struct htree_lock *hlock;
5594 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5596 if (name[0] == '.' && (name[1] == '\0' ||
5597 (name[1] == '.' && name[2] == '\0'))) {
5598 if (hlock != NULL) {
5599 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5602 down_write(&pobj->oo_ext_idx_sem);
5605 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5606 lu_object_fid(&pobj->oo_dt.do_lu),
5609 if (hlock != NULL) {
5610 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5611 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5613 down_write(&pobj->oo_ext_idx_sem);
5616 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5617 struct lu_fid *tfid = &info->oti_fid;
5621 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5624 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5628 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
5629 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
5630 pobj->oo_dirent_count++;
5633 ldiskfs_htree_unlock(hlock);
5635 up_write(&pobj->oo_ext_idx_sem);
5641 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5642 struct osd_idmap_cache *oic)
5644 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5645 struct lu_fid *fid = &oic->oic_fid;
5646 struct osd_inode_id *id = &oic->oic_lid;
5647 struct inode *inode = NULL;
5654 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5657 if (scrub->os_running && scrub->os_pos_current > id->oii_ino)
5660 if (dev->od_auto_scrub_interval == AS_NEVER ||
5661 ktime_get_real_seconds() <
5662 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5666 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5667 if (rc == -ENOENT) {
5668 __u32 gen = id->oii_gen;
5674 inode = osd_iget(oti, dev, id);
5675 /* The inode has been removed (by race maybe). */
5676 if (IS_ERR(inode)) {
5677 rc = PTR_ERR(inode);
5679 RETURN(rc == -ESTALE ? -ENOENT : rc);
5682 /* The OI mapping is lost. */
5683 if (gen != OSD_OII_NOGEN)
5687 * The inode may has been reused by others, we do not know,
5688 * leave it to be handled by subsequent osd_fid_lookup().
5691 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5698 if (scrub->os_running) {
5699 if (inode == NULL) {
5700 inode = osd_iget(oti, dev, id);
5701 /* The inode has been removed (by race maybe). */
5702 if (IS_ERR(inode)) {
5703 rc = PTR_ERR(inode);
5705 RETURN(rc == -ESTALE ? -ENOENT : rc);
5709 rc = osd_oii_insert(dev, oic, insert);
5711 * There is race condition between osd_oi_lookup and OI scrub.
5712 * The OI scrub finished just after osd_oi_lookup() failure.
5713 * Under such case, it is unnecessary to trigger OI scrub again,
5714 * but try to call osd_oi_lookup() again.
5716 if (unlikely(rc == -EAGAIN))
5719 if (!S_ISDIR(inode->i_mode))
5722 rc = osd_check_lmv(oti, dev, inode, oic);
5727 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5728 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5729 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5730 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5731 "%s: trigger partial OI scrub for RPC inconsistency "
5732 "checking FID "DFID": rc = %d\n",
5733 osd_dev2name(dev), PFID(fid), rc);
5734 if (rc == 0 || rc == -EALREADY)
5747 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5748 struct osd_device *dev,
5749 struct osd_idmap_cache *oic,
5750 struct lu_fid *fid, __u32 ino)
5752 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5753 struct inode *inode;
5756 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5757 inode = osd_iget(oti, dev, &oic->oic_lid);
5758 if (IS_ERR(inode)) {
5759 fid_zero(&oic->oic_fid);
5760 return PTR_ERR(inode);
5763 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5766 fid_zero(&oic->oic_fid);
5768 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5772 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5773 struct osd_inode_id *id, const struct lu_fid *fid)
5775 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5776 id->oii_ino, id->oii_gen, info);
5777 info->oti_cache.oic_lid = *id;
5778 info->oti_cache.oic_fid = *fid;
5779 info->oti_cache.oic_dev = osd;
5783 * Get parent FID from the linkEA.
5785 * For a directory which parent resides on remote MDT, to satisfy the
5786 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5787 * the other hand, to make the lookup(..) on the directory can return
5788 * the real parent FID, we append the real parent FID after its ".."
5789 * name entry in the /REMOTE_PARENT_DIR.
5791 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5792 * backup. So after the restore, we cannot get the right parent FID from
5793 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5794 * we have stored the real parent FID in the directory object's linkEA,
5795 * we can parse the linkEA for the real parent FID.
5797 * \param[in] env pointer to the thread context
5798 * \param[in] obj pointer to the object to be handled
5799 * \param[out]fid pointer to the buffer to hold the parent FID
5801 * \retval 0 for getting the real parent FID successfully
5802 * \retval negative error number on failure
5804 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5805 struct osd_object *obj,
5808 struct osd_thread_info *oti = osd_oti_get(env);
5809 struct lu_buf *buf = &oti->oti_big_buf;
5810 struct dentry *dentry = &oti->oti_obj_dentry;
5811 struct inode *inode = obj->oo_inode;
5812 struct linkea_data ldata = { NULL };
5818 if (!S_ISDIR(inode->i_mode))
5822 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5823 buf->lb_buf, buf->lb_len);
5824 if (rc == -ERANGE) {
5825 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5828 lu_buf_realloc(buf, rc);
5829 if (buf->lb_buf == NULL)
5836 if (unlikely(rc == 0))
5842 if (unlikely(buf->lb_buf == NULL)) {
5843 lu_buf_realloc(buf, rc);
5844 if (buf->lb_buf == NULL)
5851 rc = linkea_init_with_rec(&ldata);
5853 linkea_first_entry(&ldata);
5854 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5860 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5861 struct inode *inode,
5862 const struct lu_fid *pfid,
5863 const char *name, const int namelen)
5865 struct osd_thread_info *oti = osd_oti_get(env);
5866 struct lu_buf *buf = &oti->oti_big_buf;
5867 struct dentry *dentry = &oti->oti_obj_dentry;
5868 struct linkea_data ldata = { NULL };
5869 struct lu_name cname = { .ln_name = name,
5870 .ln_namelen = namelen };
5876 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5877 buf->lb_buf, buf->lb_len);
5879 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5884 if (unlikely(rc == 0))
5887 if (buf->lb_len < rc) {
5888 lu_buf_realloc(buf, rc);
5889 if (buf->lb_buf == NULL)
5896 rc = linkea_init_with_rec(&ldata);
5898 rc = linkea_links_find(&ldata, &cname, pfid);
5904 * Calls ->lookup() to find dentry. From dentry get inode and
5905 * read inode's ea to get fid. This is required for interoperability
5908 * \retval 0, on success
5909 * \retval -ve, on error
5911 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5912 struct dt_rec *rec, const struct dt_key *key)
5914 struct inode *dir = obj->oo_inode;
5915 struct dentry *dentry;
5916 struct ldiskfs_dir_entry_2 *de;
5917 struct buffer_head *bh;
5918 struct lu_fid *fid = (struct lu_fid *)rec;
5919 struct htree_lock *hlock = NULL;
5925 LASSERT(dir->i_op != NULL);
5926 LASSERT(dir->i_op->lookup != NULL);
5928 dentry = osd_child_dentry_get(env, obj,
5929 (char *)key, strlen((char *)key));
5931 if (obj->oo_hl_head != NULL) {
5932 hlock = osd_oti_get(env)->oti_hlock;
5933 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5934 dir, LDISKFS_HLOCK_LOOKUP);
5936 down_read(&obj->oo_ext_idx_sem);
5939 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5941 struct osd_thread_info *oti = osd_oti_get(env);
5942 struct osd_inode_id *id = &oti->oti_id;
5943 struct osd_idmap_cache *oic = &oti->oti_cache;
5944 struct osd_device *dev = osd_obj2dev(obj);
5946 ino = le32_to_cpu(de->inode);
5947 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5949 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5953 rc = osd_get_fid_from_dentry(de, rec);
5955 /* done with de, release bh */
5958 if (unlikely(is_remote_parent_ino(dev, ino))) {
5959 const char *name = (const char *)key;
5962 * If the parent is on remote MDT, and there
5963 * is no FID-in-dirent, then we have to get
5964 * the parent FID from the linkEA.
5966 if (likely(strlen(name) == 2 &&
5967 name[0] == '.' && name[1] == '.'))
5968 rc = osd_get_pfid_from_linkea(env, obj,
5971 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5974 osd_id_gen(id, ino, OSD_OII_NOGEN);
5977 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5978 fid_zero(&oic->oic_fid);
5983 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5984 rc = osd_consistency_check(oti, dev, oic);
5986 fid_zero(&oic->oic_fid);
5988 /* Other error should not affect lookup result. */
5998 ldiskfs_htree_unlock(hlock);
6000 up_read(&obj->oo_ext_idx_sem);
6004 static int osd_index_declare_ea_insert(const struct lu_env *env,
6005 struct dt_object *dt,
6006 const struct dt_rec *rec,
6007 const struct dt_key *key,
6008 struct thandle *handle)
6010 struct osd_thandle *oh;
6011 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6012 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6013 const struct lu_fid *fid = rec1->rec_fid;
6014 int credits, rc = 0;
6015 struct osd_idmap_cache *idc;
6019 LASSERT(!dt_object_remote(dt));
6020 LASSERT(handle != NULL);
6021 LASSERT(fid != NULL);
6022 LASSERT(rec1->rec_type != 0);
6024 oh = container_of(handle, struct osd_thandle, ot_super);
6025 LASSERT(oh->ot_handle == NULL);
6027 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6030 * we can't call iget() while a transactions is running
6031 * (this can lead to a deadlock), but we need to know
6032 * inum and object type. so we find this information at
6033 * declaration and cache in per-thread info
6035 idc = osd_idc_find_or_init(env, osd, fid);
6037 RETURN(PTR_ERR(idc));
6038 if (idc->oic_remote) {
6040 * a reference to remote inode is represented by an
6041 * agent inode which we have to create
6043 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6044 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6047 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6049 if (osd_dt_obj(dt)->oo_inode != NULL) {
6050 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6053 * We ignore block quota on meta pool (MDTs), so needn't
6054 * calculate how many blocks will be consumed by this index
6057 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6059 i_projid_read(inode), 0,
6060 oh, osd_dt_obj(dt), NULL,
6065 #ifdef HAVE_PROJECT_QUOTA
6067 * Reserve credits for local agent inode to transfer
6068 * to 0, quota enforcement is ignored in this case.
6070 if (idc->oic_remote &&
6071 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6072 i_projid_read(inode) != 0)
6073 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6074 0, i_projid_read(inode),
6075 0, false, PRJQUOTA, true);
6083 * Index add function for interoperability mode (b11826).
6084 * It will add the directory entry.This entry is needed to
6085 * maintain name->fid mapping.
6087 * \param key it is key i.e. file entry to be inserted
6088 * \param rec it is value of given key i.e. fid
6090 * \retval 0, on success
6091 * \retval -ve, on error
6093 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6094 const struct dt_rec *rec,
6095 const struct dt_key *key, struct thandle *th)
6097 struct osd_object *obj = osd_dt_obj(dt);
6098 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6099 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6100 const struct lu_fid *fid = rec1->rec_fid;
6101 const char *name = (const char *)key;
6102 struct osd_thread_info *oti = osd_oti_get(env);
6103 struct inode *child_inode = NULL;
6104 struct osd_idmap_cache *idc;
6109 if (!dt_object_exists(dt))
6112 LASSERT(osd_invariant(obj));
6113 LASSERT(!dt_object_remote(dt));
6114 LASSERT(th != NULL);
6116 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6118 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6120 idc = osd_idc_find(env, osd, fid);
6121 if (unlikely(idc == NULL)) {
6122 idc = osd_idc_find_or_init(env, osd, fid);
6125 * this dt_insert() wasn't declared properly, so
6126 * FID is missing in OI cache. we better do not
6127 * lookup FID in FLDB/OI and don't risk to deadlock,
6128 * but in some special cases (lfsck testing, etc)
6129 * it's much simpler than fixing a caller.
6131 * normally this error should be placed after the first
6132 * find, but migrate may attach source stripes to
6133 * target, which doesn't create stripes.
6135 CERROR("%s: "DFID" wasn't declared for insert\n",
6136 osd_name(osd), PFID(fid));
6138 RETURN(PTR_ERR(idc));
6142 if (idc->oic_remote) {
6143 /* Insert remote entry */
6144 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6146 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6148 child_inode = osd_create_local_agent_inode(env, osd,
6149 obj, fid, rec1->rec_type & S_IFMT, th);
6150 if (IS_ERR(child_inode))
6151 RETURN(PTR_ERR(child_inode));
6154 /* Insert local entry */
6155 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6156 /* for a reason OI cache wasn't filled properly */
6157 CERROR("%s: OIC for "DFID" isn't filled\n",
6158 osd_name(osd), PFID(fid));
6161 child_inode = oti->oti_inode;
6162 if (unlikely(child_inode == NULL)) {
6163 struct ldiskfs_inode_info *lii;
6168 child_inode = oti->oti_inode = &lii->vfs_inode;
6170 child_inode->i_sb = osd_sb(osd);
6171 child_inode->i_ino = idc->oic_lid.oii_ino;
6172 child_inode->i_mode = rec1->rec_type & S_IFMT;
6175 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6177 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6178 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6180 if (child_inode && child_inode != oti->oti_inode)
6182 LASSERT(osd_invariant(obj));
6183 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6189 * Initialize osd Iterator for given osd index object.
6191 * \param dt osd index object
6194 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6195 struct dt_object *dt,
6198 struct osd_it_iam *it;
6199 struct osd_object *obj = osd_dt_obj(dt);
6200 struct lu_object *lo = &dt->do_lu;
6201 struct iam_path_descr *ipd;
6202 struct iam_container *bag = &obj->oo_dir->od_container;
6204 if (!dt_object_exists(dt))
6205 return ERR_PTR(-ENOENT);
6209 return ERR_PTR(-ENOMEM);
6211 ipd = osd_it_ipd_get(env, bag);
6212 if (likely(ipd != NULL)) {
6216 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6217 return (struct dt_it *)it;
6220 return ERR_PTR(-ENOMEM);
6225 * free given Iterator.
6227 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6229 struct osd_it_iam *it = (struct osd_it_iam *)di;
6230 struct osd_object *obj = it->oi_obj;
6232 iam_it_fini(&it->oi_it);
6233 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6234 osd_object_put(env, obj);
6239 * Move Iterator to record specified by \a key
6241 * \param di osd iterator
6242 * \param key key for index
6244 * \retval +ve di points to record with least key not larger than key
6245 * \retval 0 di points to exact matched key
6246 * \retval -ve failure
6249 static int osd_it_iam_get(const struct lu_env *env,
6250 struct dt_it *di, const struct dt_key *key)
6252 struct osd_thread_info *oti = osd_oti_get(env);
6253 struct osd_it_iam *it = (struct osd_it_iam *)di;
6255 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6256 /* swab quota uid/gid */
6257 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6258 key = (struct dt_key *)&oti->oti_quota_id;
6261 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6267 * \param di osd iterator
6269 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6271 struct osd_it_iam *it = (struct osd_it_iam *)di;
6273 iam_it_put(&it->oi_it);
6277 * Move iterator by one record
6279 * \param di osd iterator
6281 * \retval +1 end of container reached
6283 * \retval -ve failure
6286 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6288 struct osd_it_iam *it = (struct osd_it_iam *)di;
6290 return iam_it_next(&it->oi_it);
6294 * Return pointer to the key under iterator.
6297 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6298 const struct dt_it *di)
6300 struct osd_thread_info *oti = osd_oti_get(env);
6301 struct osd_it_iam *it = (struct osd_it_iam *)di;
6302 struct osd_object *obj = it->oi_obj;
6305 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6307 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6308 /* swab quota uid/gid */
6309 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6310 key = (struct dt_key *)&oti->oti_quota_id;
6317 * Return size of key under iterator (in bytes)
6320 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6322 struct osd_it_iam *it = (struct osd_it_iam *)di;
6324 return iam_it_key_size(&it->oi_it);
6328 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6330 /* check if file type is required */
6331 if (ent->lde_attrs & LUDA_TYPE) {
6332 struct luda_type *lt;
6333 int align = sizeof(*lt) - 1;
6335 len = (len + align) & ~align;
6336 lt = (struct luda_type *)(ent->lde_name + len);
6337 lt->lt_type = cpu_to_le16(DTTOIF(type));
6340 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6344 * build lu direct from backend fs dirent.
6348 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6349 char *name, __u16 namelen, __u16 type, __u32 attr)
6351 ent->lde_attrs = attr | LUDA_FID;
6352 fid_cpu_to_le(&ent->lde_fid, fid);
6354 ent->lde_hash = cpu_to_le64(offset);
6355 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6357 strncpy(ent->lde_name, name, namelen);
6358 ent->lde_name[namelen] = '\0';
6359 ent->lde_namelen = cpu_to_le16(namelen);
6361 /* append lustre attributes */
6362 osd_it_append_attrs(ent, namelen, type);
6366 * Return pointer to the record under iterator.
6368 static int osd_it_iam_rec(const struct lu_env *env,
6369 const struct dt_it *di,
6370 struct dt_rec *dtrec, __u32 attr)
6372 struct osd_it_iam *it = (struct osd_it_iam *)di;
6373 struct osd_thread_info *info = osd_oti_get(env);
6377 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6378 const struct osd_fid_pack *rec;
6379 struct lu_fid *fid = &info->oti_fid;
6380 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6386 name = (char *)iam_it_key_get(&it->oi_it);
6388 RETURN(PTR_ERR(name));
6390 namelen = iam_it_key_size(&it->oi_it);
6392 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6394 RETURN(PTR_ERR(rec));
6396 rc = osd_fid_unpack(fid, rec);
6400 hash = iam_it_store(&it->oi_it);
6402 /* IAM does not store object type in IAM index (dir) */
6403 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6405 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6406 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6407 (struct iam_rec *)dtrec);
6408 osd_quota_unpack(it->oi_obj, dtrec);
6410 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6411 (struct iam_rec *)dtrec);
6418 * Returns cookie for current Iterator position.
6420 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6422 struct osd_it_iam *it = (struct osd_it_iam *)di;
6424 return iam_it_store(&it->oi_it);
6428 * Restore iterator from cookie.
6430 * \param di osd iterator
6431 * \param hash Iterator location cookie
6433 * \retval +ve di points to record with least key not larger than key.
6434 * \retval 0 di points to exact matched key
6435 * \retval -ve failure
6438 static int osd_it_iam_load(const struct lu_env *env,
6439 const struct dt_it *di, __u64 hash)
6441 struct osd_it_iam *it = (struct osd_it_iam *)di;
6443 return iam_it_load(&it->oi_it, hash);
6446 static const struct dt_index_operations osd_index_iam_ops = {
6447 .dio_lookup = osd_index_iam_lookup,
6448 .dio_declare_insert = osd_index_declare_iam_insert,
6449 .dio_insert = osd_index_iam_insert,
6450 .dio_declare_delete = osd_index_declare_iam_delete,
6451 .dio_delete = osd_index_iam_delete,
6453 .init = osd_it_iam_init,
6454 .fini = osd_it_iam_fini,
6455 .get = osd_it_iam_get,
6456 .put = osd_it_iam_put,
6457 .next = osd_it_iam_next,
6458 .key = osd_it_iam_key,
6459 .key_size = osd_it_iam_key_size,
6460 .rec = osd_it_iam_rec,
6461 .store = osd_it_iam_store,
6462 .load = osd_it_iam_load
6468 * Creates or initializes iterator context.
6470 * \retval struct osd_it_ea, iterator structure on success
6473 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6474 struct dt_object *dt,
6477 struct osd_object *obj = osd_dt_obj(dt);
6478 struct osd_thread_info *info = osd_oti_get(env);
6479 struct osd_it_ea *oie;
6481 struct lu_object *lo = &dt->do_lu;
6482 struct dentry *obj_dentry;
6486 if (!dt_object_exists(dt) || obj->oo_destroyed)
6487 RETURN(ERR_PTR(-ENOENT));
6489 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6491 RETURN(ERR_PTR(-ENOMEM));
6492 obj_dentry = &oie->oie_dentry;
6494 obj_dentry->d_inode = obj->oo_inode;
6495 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
6496 obj_dentry->d_name.hash = 0;
6498 oie->oie_rd_dirent = 0;
6499 oie->oie_it_dirent = 0;
6500 oie->oie_dirent = NULL;
6501 if (unlikely(!info->oti_it_ea_buf_used)) {
6502 oie->oie_buf = info->oti_it_ea_buf;
6503 info->oti_it_ea_buf_used = 1;
6505 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6506 if (oie->oie_buf == NULL)
6507 RETURN(ERR_PTR(-ENOMEM));
6511 file = &oie->oie_file;
6513 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6514 if (attr & LUDA_64BITHASH)
6515 file->f_mode = FMODE_64BITHASH;
6517 file->f_mode = FMODE_32BITHASH;
6518 file->f_path.dentry = obj_dentry;
6519 file->f_mapping = obj->oo_inode->i_mapping;
6520 file->f_op = obj->oo_inode->i_fop;
6521 file->f_inode = obj->oo_inode;
6524 RETURN((struct dt_it *)oie);
6528 * Destroy or finishes iterator context.
6530 * \param di iterator structure to be destroyed
6532 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6534 struct osd_thread_info *info = osd_oti_get(env);
6535 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6536 struct osd_object *obj = oie->oie_obj;
6537 struct inode *inode = obj->oo_inode;
6540 oie->oie_file.f_op->release(inode, &oie->oie_file);
6541 osd_object_put(env, obj);
6542 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6543 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6545 info->oti_it_ea_buf_used = 0;
6546 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6551 * It position the iterator at given key, so that next lookup continues from
6552 * that key Or it is similar to dio_it->load() but based on a key,
6553 * rather than file position.
6555 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6558 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6560 static int osd_it_ea_get(const struct lu_env *env,
6561 struct dt_it *di, const struct dt_key *key)
6563 struct osd_it_ea *it = (struct osd_it_ea *)di;
6566 LASSERT(((const char *)key)[0] == '\0');
6567 it->oie_file.f_pos = 0;
6568 it->oie_rd_dirent = 0;
6569 it->oie_it_dirent = 0;
6570 it->oie_dirent = NULL;
6578 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6582 struct osd_filldir_cbs {
6583 struct dir_context ctx;
6584 struct osd_it_ea *it;
6587 * It is called internally by ->iterate*(). It fills the
6588 * iterator's in-memory data structure with required
6589 * information i.e. name, namelen, rec_size etc.
6591 * \param buf in which information to be filled in.
6592 * \param name name of the file in given dir
6594 * \retval 0 on success
6595 * \retval 1 on buffer full
6597 #ifdef HAVE_FILLDIR_USE_CTX
6598 static int osd_ldiskfs_filldir(struct dir_context *buf,
6600 static int osd_ldiskfs_filldir(void *buf,
6602 const char *name, int namelen,
6603 loff_t offset, __u64 ino, unsigned int d_type)
6605 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6606 struct osd_object *obj = it->oie_obj;
6607 struct osd_it_ea_dirent *ent = it->oie_dirent;
6608 struct lu_fid *fid = &ent->oied_fid;
6609 struct osd_fid_pack *rec;
6613 /* this should never happen */
6614 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6615 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6619 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6623 /* "." is just the object itself. */
6624 if (namelen == 1 && name[0] == '.') {
6625 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6626 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6627 rec = (struct osd_fid_pack *)(name + namelen + 1);
6628 if (osd_fid_unpack(fid, rec) != 0)
6633 d_type &= ~LDISKFS_DIRENT_LUFID;
6635 /* NOT export local root. */
6636 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6637 ino = obj->oo_inode->i_ino;
6638 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6641 ent->oied_ino = ino;
6642 ent->oied_off = offset;
6643 ent->oied_namelen = namelen;
6644 ent->oied_type = d_type;
6646 memcpy(ent->oied_name, name, namelen);
6648 it->oie_rd_dirent++;
6649 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6654 * Calls ->iterate*() to load a directory entry at a time
6655 * and stored it in iterator's in-memory data structure.
6657 * \param di iterator's in memory structure
6659 * \retval 0 on success
6660 * \retval -ve on error
6661 * \retval +1 reach the end of entry
6663 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6664 const struct dt_it *di)
6666 struct osd_it_ea *it = (struct osd_it_ea *)di;
6667 struct osd_object *obj = it->oie_obj;
6668 struct inode *inode = obj->oo_inode;
6669 struct htree_lock *hlock = NULL;
6670 struct file *filp = &it->oie_file;
6672 struct osd_filldir_cbs buf = {
6673 .ctx.actor = osd_ldiskfs_filldir,
6678 it->oie_dirent = it->oie_buf;
6679 it->oie_rd_dirent = 0;
6681 if (obj->oo_hl_head != NULL) {
6682 hlock = osd_oti_get(env)->oti_hlock;
6683 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6684 inode, LDISKFS_HLOCK_READDIR);
6686 down_read(&obj->oo_ext_idx_sem);
6689 filp->f_cred = current_cred();
6690 rc = osd_security_file_alloc(filp);
6694 filp->f_flags |= O_NOATIME;
6695 filp->f_mode |= FMODE_NONOTIFY;
6696 rc = iterate_dir(filp, &buf.ctx);
6701 ldiskfs_htree_unlock(hlock);
6703 up_read(&obj->oo_ext_idx_sem);
6705 if (it->oie_rd_dirent == 0) {
6707 * If it does not get any dirent, it means it has been reached
6708 * to the end of the dir
6710 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6714 it->oie_dirent = it->oie_buf;
6715 it->oie_it_dirent = 1;
6722 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
6723 * to load a directory entry at a time and stored it in
6724 * iterator's in-memory data structure.
6726 * \param di iterator's in memory structure
6728 * \retval +ve iterator reached to end
6729 * \retval 0 iterator not reached to end
6730 * \retval -ve on error
6732 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6734 struct osd_it_ea *it = (struct osd_it_ea *)di;
6739 if (it->oie_it_dirent < it->oie_rd_dirent) {
6741 (void *)it->oie_dirent +
6742 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6743 it->oie_dirent->oied_namelen);
6744 it->oie_it_dirent++;
6747 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6750 rc = osd_ldiskfs_it_fill(env, di);
6757 * Returns the key at current position from iterator's in memory structure.
6759 * \param di iterator's in memory structure
6761 * \retval key i.e. struct dt_key on success
6763 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6764 const struct dt_it *di)
6766 struct osd_it_ea *it = (struct osd_it_ea *)di;
6768 return (struct dt_key *)it->oie_dirent->oied_name;
6772 * Returns key's size at current position from iterator's in memory structure.
6774 * \param di iterator's in memory structure
6776 * \retval key_size i.e. struct dt_key on success
6778 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6780 struct osd_it_ea *it = (struct osd_it_ea *)di;
6782 return it->oie_dirent->oied_namelen;
6785 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
6786 #undef LDISKFS_DIR_REC_LEN
6787 #define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
6790 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6792 if (LDISKFS_DIR_REC_LEN(de) >=
6793 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6800 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6801 unsigned int blocksize, bool dotdot)
6804 return osd_dotdot_has_space(de);
6806 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6807 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6814 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6815 handle_t *jh, struct dentry *dentry,
6816 const struct lu_fid *fid, struct buffer_head *bh,
6817 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6820 struct inode *dir = dentry->d_parent->d_inode;
6821 struct inode *inode = dentry->d_inode;
6822 struct osd_fid_pack *rec;
6823 struct ldiskfs_dentry_param *ldp;
6824 int namelen = dentry->d_name.len;
6826 struct osd_thread_info *info = osd_oti_get(env);
6830 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6833 /* There is enough space to hold the FID-in-dirent. */
6834 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6835 rc = ldiskfs_journal_get_write_access(jh, bh);
6839 de->name[namelen] = 0;
6840 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6841 rec->fp_len = sizeof(struct lu_fid) + 1;
6842 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6843 de->file_type |= LDISKFS_DIRENT_LUFID;
6844 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6851 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6855 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6856 osd_get_ldiskfs_dirent_param(ldp, fid);
6857 dentry->d_fsdata = (void *)ldp;
6858 dquot_initialize(dir);
6859 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6861 * It is too bad, we cannot reinsert the name entry back.
6862 * That means we lose it!
6865 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6866 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6867 osd_ino2name(inode),
6868 dir->i_ino, dir->i_generation, namelen,
6869 dentry->d_name.name, PFID(fid), rc);
6875 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6876 struct osd_it_ea *it, struct lu_fid *fid,
6877 struct osd_inode_id *id, __u32 *attr)
6879 struct osd_thread_info *info = osd_oti_get(env);
6880 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6881 struct osd_device *dev = osd_obj2dev(obj);
6882 struct super_block *sb = osd_sb(dev);
6883 const char *devname = osd_name(dev);
6884 struct osd_it_ea_dirent *ent = it->oie_dirent;
6885 struct inode *dir = obj->oo_inode;
6886 struct htree_lock *hlock = NULL;
6887 struct buffer_head *bh = NULL;
6888 handle_t *jh = NULL;
6889 struct ldiskfs_dir_entry_2 *de;
6890 struct dentry *dentry;
6891 struct inode *inode;
6892 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6895 bool dotdot = false;
6900 if (ent->oied_name[0] == '.') {
6901 if (ent->oied_namelen == 1)
6904 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6908 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6909 inode = osd_iget(info, dev, id);
6910 if (IS_ERR(inode)) {
6911 rc = PTR_ERR(inode);
6912 if (rc == -ENOENT || rc == -ESTALE) {
6914 * Maybe dangling name entry, or
6915 * corrupted directory entry.
6917 *attr |= LUDA_UNKNOWN;
6920 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6921 "check_repair, dir = %lu/%u, name = %.*s, "
6922 "ino = %llu, rc = %d\n",
6923 devname, dir->i_ino, dir->i_generation,
6924 ent->oied_namelen, ent->oied_name,
6931 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6933 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6934 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6940 * We need to ensure that the name entry is still valid.
6941 * Because it may be removed or renamed by other already.
6943 * The unlink or rename operation will start journal before PDO lock,
6944 * so to avoid deadlock, here we need to start journal handle before
6945 * related PDO lock also. But because we do not know whether there
6946 * will be something to be repaired before PDO lock, we just start
6947 * journal without conditions.
6949 * We may need to remove the name entry firstly, then insert back.
6950 * One credit is for user quota file update.
6951 * One credit is for group quota file update.
6952 * Two credits are for dirty inode.
6954 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6955 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6957 if (dev->od_dirent_journal != 0) {
6960 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6963 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6964 "check_repair, dir = %lu/%u, credits = %d, "
6965 "name = %.*s, ino = %llu: rc = %d\n",
6966 devname, dir->i_ino, dir->i_generation, credits,
6967 ent->oied_namelen, ent->oied_name,
6970 GOTO(out_inode, rc);
6973 if (obj->oo_hl_head != NULL) {
6974 hlock = osd_oti_get(env)->oti_hlock;
6976 * "0" means exclusive lock for the whole directory.
6977 * We need to prevent others access such name entry
6978 * during the delete + insert. Neither HLOCK_ADD nor
6979 * HLOCK_DEL cannot guarantee the atomicity.
6981 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6983 down_write(&obj->oo_ext_idx_sem);
6986 if (obj->oo_hl_head != NULL) {
6987 hlock = osd_oti_get(env)->oti_hlock;
6988 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6989 LDISKFS_HLOCK_LOOKUP);
6991 down_read(&obj->oo_ext_idx_sem);
6995 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6996 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6997 *attr |= LUDA_IGNORE;
7003 * For dotdot entry, if there is not enough space to hold the
7004 * FID-in-dirent, just keep them there. It only happens when the
7005 * device upgraded from 1.8 or restored from MDT file-level backup.
7006 * For the whole directory, only dotdot entry have no FID-in-dirent
7007 * and needs to get FID from LMA when readdir, it will not affect the
7010 if (dotdot && !osd_dotdot_has_space(de)) {
7011 *attr |= LUDA_UNKNOWN;
7017 if (lu_fid_eq(fid, &lma->lma_self_fid))
7020 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7021 struct lu_fid *tfid = &lma->lma_self_fid;
7023 if (likely(dotdot &&
7024 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7025 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7027 * It must be REMOTE_PARENT_DIR and as the
7028 * 'dotdot' entry of remote directory
7030 *attr |= LUDA_IGNORE;
7032 CDEBUG(D_LFSCK, "%s: expect remote agent "
7033 "parent directory, but got %.*s under "
7034 "dir = %lu/%u with the FID "DFID"\n",
7035 devname, ent->oied_namelen,
7036 ent->oied_name, dir->i_ino,
7037 dir->i_generation, PFID(tfid));
7039 *attr |= LUDA_UNKNOWN;
7046 if (!fid_is_zero(fid)) {
7047 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7049 if (rc == -ENOENT ||
7051 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7053 * linkEA does not recognize the dirent entry,
7054 * it may because the dirent entry corruption
7055 * and points to other's inode.
7057 CDEBUG(D_LFSCK, "%s: the target inode does not "
7058 "recognize the dirent, dir = %lu/%u, "
7059 " name = %.*s, ino = %llu, "
7060 DFID": rc = %d\n", devname, dir->i_ino,
7061 dir->i_generation, ent->oied_namelen,
7062 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7063 *attr |= LUDA_UNKNOWN;
7068 if (rc && rc != -ENODATA) {
7069 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7070 "dir = %lu/%u, name = %.*s, ino = %llu, "
7071 DFID": rc = %d\n", devname, dir->i_ino,
7072 dir->i_generation, ent->oied_namelen,
7073 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7074 *attr |= LUDA_UNKNOWN;
7082 * linkEA recognizes the dirent entry, the FID-in-LMA is
7083 * valid, trusted, in spite of fid_is_sane(fid) or not.
7085 if (*attr & LUDA_VERIFY_DRYRUN) {
7086 *fid = lma->lma_self_fid;
7087 *attr |= LUDA_REPAIR;
7094 dev->od_dirent_journal = 1;
7095 if (hlock != NULL) {
7096 ldiskfs_htree_unlock(hlock);
7099 up_read(&obj->oo_ext_idx_sem);
7105 *fid = lma->lma_self_fid;
7107 /* Update or append the FID-in-dirent. */
7108 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7109 bh, de, hlock, dotdot);
7111 *attr |= LUDA_REPAIR;
7113 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7114 "the dirent, dir = %lu/%u, name = %.*s, "
7115 "ino = %llu, "DFID": rc = %d\n",
7116 devname, dir->i_ino, dir->i_generation,
7117 ent->oied_namelen, ent->oied_name,
7118 ent->oied_ino, PFID(fid), rc);
7120 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7121 if (*attr & LUDA_VERIFY_DRYRUN) {
7122 if (fid_is_sane(fid)) {
7123 *attr |= LUDA_REPAIR;
7124 } else if (dev->od_index == 0) {
7125 lu_igif_build(fid, inode->i_ino,
7126 inode->i_generation);
7127 *attr |= LUDA_UPGRADE;
7135 dev->od_dirent_journal = 1;
7136 if (hlock != NULL) {
7137 ldiskfs_htree_unlock(hlock);
7140 up_read(&obj->oo_ext_idx_sem);
7147 if (unlikely(fid_is_sane(fid))) {
7149 * FID-in-dirent exists, but FID-in-LMA is lost.
7150 * Trust the FID-in-dirent, and add FID-in-LMA.
7152 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7154 *attr |= LUDA_REPAIR;
7156 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7157 "update dirent, dir = %lu/%u, "
7158 "name = %.*s, ino = %llu, "
7160 devname, dir->i_ino, dir->i_generation,
7161 ent->oied_namelen, ent->oied_name,
7162 ent->oied_ino, PFID(fid), rc);
7163 } else if (dev->od_index == 0) {
7164 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7166 * It is probably IGIF object. Only aappend the
7167 * FID-in-dirent. OI scrub will process FID-in-LMA.
7169 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7170 bh, de, hlock, dotdot);
7172 *attr |= LUDA_UPGRADE;
7174 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7175 "after the dirent, dir = %lu/%u, "
7176 "name = %.*s, ino = %llu, "
7178 devname, dir->i_ino, dir->i_generation,
7179 ent->oied_namelen, ent->oied_name,
7180 ent->oied_ino, PFID(fid), rc);
7189 if (hlock != NULL) {
7190 ldiskfs_htree_unlock(hlock);
7192 if (dev->od_dirent_journal != 0)
7193 up_write(&obj->oo_ext_idx_sem);
7195 up_read(&obj->oo_ext_idx_sem);
7199 ldiskfs_journal_stop(jh);
7203 if (rc >= 0 && !dirty)
7204 dev->od_dirent_journal = 0;
7210 * Returns the value at current position from iterator's in memory structure.
7212 * \param di struct osd_it_ea, iterator's in memory structure
7213 * \param attr attr requested for dirent.
7214 * \param lde lustre dirent
7216 * \retval 0 no error and \param lde has correct lustre dirent.
7217 * \retval -ve on error
7219 static inline int osd_it_ea_rec(const struct lu_env *env,
7220 const struct dt_it *di,
7221 struct dt_rec *dtrec, __u32 attr)
7223 struct osd_it_ea *it = (struct osd_it_ea *)di;
7224 struct osd_object *obj = it->oie_obj;
7225 struct osd_device *dev = osd_obj2dev(obj);
7226 struct osd_thread_info *oti = osd_oti_get(env);
7227 struct osd_inode_id *id = &oti->oti_id;
7228 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7229 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7230 __u32 ino = it->oie_dirent->oied_ino;
7235 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7237 if (attr & LUDA_VERIFY) {
7238 if (unlikely(is_remote_parent_ino(dev, ino))) {
7239 attr |= LUDA_IGNORE;
7241 * If the parent is on remote MDT, and there
7242 * is no FID-in-dirent, then we have to get
7243 * the parent FID from the linkEA.
7245 if (!fid_is_sane(fid) &&
7246 it->oie_dirent->oied_namelen == 2 &&
7247 it->oie_dirent->oied_name[0] == '.' &&
7248 it->oie_dirent->oied_name[1] == '.')
7249 osd_get_pfid_from_linkea(env, obj, fid);
7251 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7255 if (!fid_is_sane(fid))
7256 attr |= LUDA_UNKNOWN;
7258 attr &= ~LU_DIRENT_ATTRS_MASK;
7259 if (!fid_is_sane(fid)) {
7260 bool is_dotdot = false;
7262 if (it->oie_dirent->oied_namelen == 2 &&
7263 it->oie_dirent->oied_name[0] == '.' &&
7264 it->oie_dirent->oied_name[1] == '.')
7267 * If the parent is on remote MDT, and there
7268 * is no FID-in-dirent, then we have to get
7269 * the parent FID from the linkEA.
7271 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7272 rc = osd_get_pfid_from_linkea(env, obj, fid);
7274 if (is_dotdot == false &&
7275 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7278 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7281 osd_id_gen(id, ino, OSD_OII_NOGEN);
7285 /* Pack the entry anyway, at least the offset is right. */
7286 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7287 it->oie_dirent->oied_name,
7288 it->oie_dirent->oied_namelen,
7289 it->oie_dirent->oied_type, attr);
7294 if (osd_remote_fid(env, dev, fid))
7297 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
7298 osd_add_oi_cache(oti, dev, id, fid);
7300 RETURN(rc > 0 ? 0 : rc);
7304 * Returns the record size size at current position.
7306 * This function will return record(lu_dirent) size in bytes.
7308 * \param[in] env execution environment
7309 * \param[in] di iterator's in memory structure
7310 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7311 * calculate the lu_dirent size.
7313 * \retval record size(in bytes & in memory) of the current lu_dirent
7316 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7319 struct osd_it_ea *it = (struct osd_it_ea *)di;
7321 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7325 * Returns a cookie for current position of the iterator head, so that
7326 * user can use this cookie to load/start the iterator next time.
7328 * \param di iterator's in memory structure
7330 * \retval cookie for current position, on success
7332 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7334 struct osd_it_ea *it = (struct osd_it_ea *)di;
7336 return it->oie_dirent->oied_off;
7340 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7341 * to load a directory entry at a time and stored it i inn,
7342 * in iterator's in-memory data structure.
7344 * \param di struct osd_it_ea, iterator's in memory structure
7346 * \retval +ve on success
7347 * \retval -ve on error
7349 static int osd_it_ea_load(const struct lu_env *env,
7350 const struct dt_it *di, __u64 hash)
7352 struct osd_it_ea *it = (struct osd_it_ea *)di;
7356 it->oie_file.f_pos = hash;
7358 rc = osd_ldiskfs_it_fill(env, di);
7369 * Index lookup function for interoperability mode (b11826).
7371 * \param key, key i.e. file name to be searched
7373 * \retval +ve, on success
7374 * \retval -ve, on error
7376 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7377 struct dt_rec *rec, const struct dt_key *key)
7379 struct osd_object *obj = osd_dt_obj(dt);
7384 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7385 LINVRNT(osd_invariant(obj));
7387 rc = osd_ea_lookup_rec(env, obj, rec, key);
7394 * Index and Iterator operations for interoperability
7395 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7397 static const struct dt_index_operations osd_index_ea_ops = {
7398 .dio_lookup = osd_index_ea_lookup,
7399 .dio_declare_insert = osd_index_declare_ea_insert,
7400 .dio_insert = osd_index_ea_insert,
7401 .dio_declare_delete = osd_index_declare_ea_delete,
7402 .dio_delete = osd_index_ea_delete,
7404 .init = osd_it_ea_init,
7405 .fini = osd_it_ea_fini,
7406 .get = osd_it_ea_get,
7407 .put = osd_it_ea_put,
7408 .next = osd_it_ea_next,
7409 .key = osd_it_ea_key,
7410 .key_size = osd_it_ea_key_size,
7411 .rec = osd_it_ea_rec,
7412 .rec_size = osd_it_ea_rec_size,
7413 .store = osd_it_ea_store,
7414 .load = osd_it_ea_load
7418 static void *osd_key_init(const struct lu_context *ctx,
7419 struct lu_context_key *key)
7421 struct osd_thread_info *info;
7423 OBD_ALLOC_PTR(info);
7425 return ERR_PTR(-ENOMEM);
7427 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7428 if (info->oti_it_ea_buf == NULL)
7431 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7433 info->oti_hlock = ldiskfs_htree_lock_alloc();
7434 if (info->oti_hlock == NULL)
7440 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7443 return ERR_PTR(-ENOMEM);
7446 static void osd_key_fini(const struct lu_context *ctx,
7447 struct lu_context_key *key, void *data)
7449 struct osd_thread_info *info = data;
7450 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7451 struct osd_idmap_cache *idc = info->oti_ins_cache;
7453 if (info->oti_dio_pages) {
7455 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7456 struct page *page = info->oti_dio_pages[i];
7458 LASSERT(PagePrivate2(page));
7459 LASSERT(PageLocked(page));
7460 ClearPagePrivate2(page);
7465 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7466 PTLRPC_MAX_BRW_PAGES);
7469 if (info->oti_inode != NULL)
7471 if (info->oti_hlock != NULL)
7472 ldiskfs_htree_lock_free(info->oti_hlock);
7473 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7474 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7475 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7476 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7477 lu_buf_free(&info->oti_big_buf);
7479 LASSERT(info->oti_ins_cache_size > 0);
7480 OBD_FREE_PTR_ARRAY(idc, info->oti_ins_cache_size);
7481 info->oti_ins_cache = NULL;
7482 info->oti_ins_cache_size = 0;
7487 static void osd_key_exit(const struct lu_context *ctx,
7488 struct lu_context_key *key, void *data)
7490 struct osd_thread_info *info = data;
7492 LASSERT(info->oti_r_locks == 0);
7493 LASSERT(info->oti_w_locks == 0);
7494 LASSERT(info->oti_txns == 0);
7497 /* type constructor/destructor: osd_type_init, osd_type_fini */
7498 LU_TYPE_INIT_FINI(osd, &osd_key);
7500 struct lu_context_key osd_key = {
7501 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7502 .lct_init = osd_key_init,
7503 .lct_fini = osd_key_fini,
7504 .lct_exit = osd_key_exit
7508 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7509 const char *name, struct lu_device *next)
7511 struct osd_device *osd = osd_dev(d);
7513 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7514 sizeof(osd->od_svname))
7516 return osd_procfs_init(osd, name);
7519 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7521 struct seq_server_site *ss = osd_seq_site(osd);
7526 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7529 if (unlikely(ss == NULL))
7532 OBD_ALLOC_PTR(osd->od_cl_seq);
7533 if (osd->od_cl_seq == NULL)
7536 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7537 osd->od_svname, ss->ss_server_seq);
7539 if (ss->ss_node_id == 0) {
7541 * If the OSD on the sequence controller(MDT0), then allocate
7542 * sequence here, otherwise allocate sequence after connected
7543 * to MDT0 (see mdt_register_lwp_callback()).
7545 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7546 &osd->od_cl_seq->lcs_space, env);
7552 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7554 if (osd->od_cl_seq == NULL)
7557 seq_client_fini(osd->od_cl_seq);
7558 OBD_FREE_PTR(osd->od_cl_seq);
7559 osd->od_cl_seq = NULL;
7562 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7566 /* shutdown quota slave instance associated with the device */
7567 if (o->od_quota_slave_md != NULL) {
7568 struct qsd_instance *qsd = o->od_quota_slave_md;
7570 o->od_quota_slave_md = NULL;
7574 if (o->od_quota_slave_dt != NULL) {
7575 struct qsd_instance *qsd = o->od_quota_slave_dt;
7577 o->od_quota_slave_dt = NULL;
7581 osd_fid_fini(env, o);
7582 osd_scrub_cleanup(env, o);
7587 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7591 if (o->od_mnt != NULL) {
7592 shrink_dcache_sb(osd_sb(o));
7593 osd_sync(env, &o->od_dt_dev);
7602 static int osd_mount(const struct lu_env *env,
7603 struct osd_device *o, struct lustre_cfg *cfg)
7605 const char *name = lustre_cfg_string(cfg, 0);
7606 const char *dev = lustre_cfg_string(cfg, 1);
7608 unsigned long page, s_flags, lmd_flags = 0;
7609 struct page *__page;
7610 struct file_system_type *type;
7611 char *options = NULL;
7613 struct osd_thread_info *info = osd_oti_get(env);
7614 struct lu_fid *fid = &info->oti_fid;
7615 struct inode *inode;
7616 int rc = 0, force_over_1024tb = 0;
7620 if (o->od_mnt != NULL)
7623 if (strlen(dev) >= sizeof(o->od_mntdev))
7625 strcpy(o->od_mntdev, dev);
7627 str = lustre_cfg_string(cfg, 2);
7628 s_flags = simple_strtoul(str, NULL, 0);
7629 str = strstr(str, ":");
7631 lmd_flags = simple_strtoul(str + 1, NULL, 0);
7632 opts = lustre_cfg_string(cfg, 3);
7634 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7635 CERROR("%s: device %s extents feature is not guaranteed to "
7636 "work on big-endian systems. Use \"bigendian_extents\" "
7637 "mount option to override.\n", name, dev);
7641 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7642 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7643 CWARN("force_over_128tb option is deprecated. Filesystems smaller than 1024TB can be created without any force option. Use force_over_1024tb option for filesystems larger than 1024TB.\n");
7646 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7647 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7648 CWARN("force_over_256tb option is deprecated. Filesystems smaller than 1024TB can be created without any force options. Use force_over_1024tb option for filesystems larger than 1024TB.\n");
7651 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7652 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
7653 CWARN("force_over_512tb option is deprecated. Filesystems smaller than 1024TB can be created without any force options. Use force_over_1024tb option for filesystems larger than 1024TB.\n");
7657 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
7658 force_over_1024tb = 1;
7660 __page = alloc_page(GFP_KERNEL);
7662 GOTO(out, rc = -ENOMEM);
7663 page = (unsigned long)page_address(__page);
7664 options = (char *)page;
7667 /* strip out the options for back compatiblity */
7668 static const char * const sout[] = {
7675 /* strip out option we processed in osd */
7676 "bigendian_extents",
7680 "force_over_1024tb",
7683 strncat(options, opts, PAGE_SIZE);
7684 for (rc = 0, str = options; sout[rc]; ) {
7685 char *op = strstr(str, sout[rc]);
7692 if (op == options || *(op - 1) == ',') {
7693 str = op + strlen(sout[rc]);
7694 if (*str == ',' || *str == '\0') {
7695 *str == ',' ? str++ : str;
7696 memmove(op, str, strlen(str) + 1);
7699 for (str = op; *str != ',' && *str != '\0'; str++)
7703 strncat(options, "user_xattr,acl", PAGE_SIZE);
7706 /* Glom up mount options */
7707 if (*options != '\0')
7708 strncat(options, ",", PAGE_SIZE);
7709 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7711 type = get_fs_type("ldiskfs");
7713 CERROR("%s: cannot find ldiskfs module\n", name);
7714 GOTO(out, rc = -ENODEV);
7717 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7718 module_put(type->owner);
7720 if (IS_ERR(o->od_mnt)) {
7721 rc = PTR_ERR(o->od_mnt);
7723 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7727 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7728 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
7729 force_over_1024tb == 0) {
7730 CERROR("%s: device %s LDISKFS has not been tested on filesystems larger than 1024TB and may cause data corruption. Use 'force_over_1024tb' mount option to override.\n",
7732 GOTO(out_mnt, rc = -EINVAL);
7735 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7736 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7739 GOTO(out_mnt, rc = -EOPNOTSUPP);
7742 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7743 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7744 GOTO(out_mnt, rc = -EINVAL);
7747 #ifdef LDISKFS_MOUNT_DIRDATA
7748 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7749 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7750 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7751 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7752 "enabling the dirdata feature. If you do not want to "
7753 "downgrade to Lustre-1.x again, you can enable it via "
7754 "'tune2fs -O dirdata device'\n", name, dev);
7756 inode = osd_sb(o)->s_root->d_inode;
7757 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7758 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7760 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7764 if (lmd_flags & LMD_FLG_NOSCRUB)
7765 o->od_auto_scrub_interval = AS_NEVER;
7767 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7768 /* do not use pagecache with flash-backed storage */
7769 o->od_writethrough_cache = 0;
7770 o->od_read_cache = 0;
7781 __free_page(__page);
7786 static struct lu_device *osd_device_fini(const struct lu_env *env,
7787 struct lu_device *d)
7789 struct osd_device *o = osd_dev(d);
7793 osd_index_backup(env, o, false);
7794 osd_shutdown(env, o);
7796 if (o->od_oi_table != NULL)
7797 osd_oi_fini(osd_oti_get(env), o);
7798 osd_obj_map_fini(o);
7804 static int osd_device_init0(const struct lu_env *env,
7805 struct osd_device *o,
7806 struct lustre_cfg *cfg)
7808 struct lu_device *l = osd2lu_dev(o);
7809 struct osd_thread_info *info;
7813 /* if the module was re-loaded, env can loose its keys */
7814 rc = lu_env_refill((struct lu_env *)env);
7817 info = osd_oti_get(env);
7820 l->ld_ops = &osd_lu_ops;
7821 o->od_dt_dev.dd_ops = &osd_dt_ops;
7823 spin_lock_init(&o->od_osfs_lock);
7824 mutex_init(&o->od_otable_mutex);
7825 INIT_LIST_HEAD(&o->od_orphan_list);
7826 INIT_LIST_HEAD(&o->od_index_backup_list);
7827 INIT_LIST_HEAD(&o->od_index_restore_list);
7828 spin_lock_init(&o->od_lock);
7829 o->od_index_backup_policy = LIBP_NONE;
7832 o->od_read_cache = 1;
7833 o->od_writethrough_cache = 1;
7834 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7835 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
7836 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
7837 o->od_auto_scrub_interval = AS_DEFAULT;
7839 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7840 sizeof(o->od_svname));
7841 if (cplen >= sizeof(o->od_svname)) {
7846 o->od_index_backup_stop = 0;
7847 o->od_index = -1; /* -1 means index is invalid */
7848 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7849 if (rc == LDD_F_SV_TYPE_OST)
7852 o->od_full_scrub_ratio = OFSR_DEFAULT;
7853 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7854 rc = osd_mount(env, o, cfg);
7858 /* Can only check block device after mount */
7859 o->od_nonrotational =
7860 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
7862 rc = osd_obj_map_init(env, o);
7866 rc = lu_site_init(&o->od_site, l);
7868 GOTO(out_compat, rc);
7869 o->od_site.ls_bottom_dev = l;
7871 rc = lu_site_init_finish(&o->od_site);
7875 INIT_LIST_HEAD(&o->od_ios_list);
7876 /* setup scrub, including OI files initialization */
7878 rc = osd_scrub_setup(env, o);
7883 rc = osd_procfs_init(o, o->od_svname);
7885 CERROR("%s: can't initialize procfs: rc = %d\n",
7887 GOTO(out_scrub, rc);
7890 LASSERT(l->ld_site->ls_linkage.next != NULL);
7891 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7893 /* initialize quota slave instance */
7894 /* currently it's no need to prepare qsd_instance_md for OST */
7895 if (!o->od_is_ost) {
7896 o->od_quota_slave_md = qsd_init(env, o->od_svname,
7898 o->od_proc_entry, true);
7899 if (IS_ERR(o->od_quota_slave_md)) {
7900 rc = PTR_ERR(o->od_quota_slave_md);
7901 o->od_quota_slave_md = NULL;
7902 GOTO(out_procfs, rc);
7906 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
7907 o->od_proc_entry, false);
7909 if (IS_ERR(o->od_quota_slave_dt)) {
7910 if (o->od_quota_slave_md != NULL) {
7911 qsd_fini(env, o->od_quota_slave_md);
7912 o->od_quota_slave_md = NULL;
7915 rc = PTR_ERR(o->od_quota_slave_dt);
7916 o->od_quota_slave_dt = NULL;
7917 GOTO(out_procfs, rc);
7925 osd_scrub_cleanup(env, o);
7927 lu_site_fini(&o->od_site);
7929 osd_obj_map_fini(o);
7936 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7937 struct lu_device_type *t,
7938 struct lustre_cfg *cfg)
7940 struct osd_device *o;
7945 return ERR_PTR(-ENOMEM);
7947 rc = dt_device_init(&o->od_dt_dev, t);
7950 * Because the ctx might be revived in dt_device_init,
7951 * refill the env here
7953 lu_env_refill((struct lu_env *)env);
7954 rc = osd_device_init0(env, o, cfg);
7956 dt_device_fini(&o->od_dt_dev);
7959 if (unlikely(rc != 0))
7962 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7965 static struct lu_device *osd_device_free(const struct lu_env *env,
7966 struct lu_device *d)
7968 struct osd_device *o = osd_dev(d);
7972 /* XXX: make osd top device in order to release reference */
7973 d->ld_site->ls_top_dev = d;
7974 lu_site_purge(env, d->ld_site, -1);
7975 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
7976 D_ERROR, lu_cdebug_printer);
7977 lu_site_fini(&o->od_site);
7978 dt_device_fini(&o->od_dt_dev);
7983 static int osd_process_config(const struct lu_env *env,
7984 struct lu_device *d, struct lustre_cfg *cfg)
7986 struct osd_device *o = osd_dev(d);
7992 switch (cfg->lcfg_command) {
7994 rc = osd_mount(env, o, cfg);
7998 * For the case LCFG_PRE_CLEANUP is not called in advance,
7999 * that may happend if hit failure during mount process.
8001 osd_index_backup(env, o, false);
8002 lu_dev_del_linkage(d->ld_site, d);
8003 rc = osd_shutdown(env, o);
8006 LASSERT(&o->od_dt_dev);
8007 count = class_modify_config(cfg, PARAM_OSD,
8008 &o->od_dt_dev.dd_kobj);
8010 count = class_modify_config(cfg, PARAM_OST,
8011 &o->od_dt_dev.dd_kobj);
8012 rc = count > 0 ? 0 : count;
8014 case LCFG_PRE_CLEANUP:
8016 osd_index_backup(env, o,
8017 o->od_index_backup_policy != LIBP_NONE);
8027 static int osd_recovery_complete(const struct lu_env *env,
8028 struct lu_device *d)
8030 struct osd_device *osd = osd_dev(d);
8035 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8039 * start qsd instance on recovery completion, this notifies the quota
8040 * slave code that we are about to process new requests now
8042 rc = qsd_start(env, osd->od_quota_slave_dt);
8043 if (rc == 0 && osd->od_quota_slave_md != NULL)
8044 rc = qsd_start(env, osd->od_quota_slave_md);
8050 * we use exports to track all osd users
8052 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8053 struct obd_device *obd, struct obd_uuid *cluuid,
8054 struct obd_connect_data *data, void *localdata)
8056 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8057 struct lustre_handle conn;
8062 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8064 rc = class_connect(&conn, obd, cluuid);
8068 *exp = class_conn2export(&conn);
8070 spin_lock(&osd->od_osfs_lock);
8072 spin_unlock(&osd->od_osfs_lock);
8078 * once last export (we don't count self-export) disappeared
8079 * osd can be released
8081 static int osd_obd_disconnect(struct obd_export *exp)
8083 struct obd_device *obd = exp->exp_obd;
8084 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8085 int rc, release = 0;
8089 /* Only disconnect the underlying layers on the final disconnect. */
8090 spin_lock(&osd->od_osfs_lock);
8092 if (osd->od_connects == 0)
8094 spin_unlock(&osd->od_osfs_lock);
8096 rc = class_disconnect(exp); /* bz 9811 */
8098 if (rc == 0 && release)
8099 class_manual_cleanup(obd);
8103 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8104 struct lu_device *dev)
8106 struct osd_device *osd = osd_dev(dev);
8107 struct lr_server_data *lsd =
8108 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8113 if (osd->od_quota_slave_md != NULL) {
8114 /* set up quota slave objects for inode */
8115 result = qsd_prepare(env, osd->od_quota_slave_md);
8120 if (osd->od_quota_slave_dt != NULL) {
8121 /* set up quota slave objects for block */
8122 result = qsd_prepare(env, osd->od_quota_slave_dt);
8128 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8129 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8130 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8131 osd->od_index_in_idif = 1;
8133 osd->od_index_in_idif = 0;
8134 result = osd_register_proc_index_in_idif(osd);
8139 osd->od_index_in_idif = 1;
8143 result = osd_fid_init(env, osd);
8149 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8153 * see include/lu_object.h for the details.
8155 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8156 struct lu_fid *fid, struct lu_object *parent,
8157 const struct lu_name *name)
8159 struct osd_device *osd = osd_dev(d);
8161 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8164 static const struct lu_object_operations osd_lu_obj_ops = {
8165 .loo_object_init = osd_object_init,
8166 .loo_object_delete = osd_object_delete,
8167 .loo_object_release = osd_object_release,
8168 .loo_object_free = osd_object_free,
8169 .loo_object_print = osd_object_print,
8170 .loo_object_invariant = osd_object_invariant
8173 const struct lu_device_operations osd_lu_ops = {
8174 .ldo_object_alloc = osd_object_alloc,
8175 .ldo_process_config = osd_process_config,
8176 .ldo_recovery_complete = osd_recovery_complete,
8177 .ldo_prepare = osd_prepare,
8178 .ldo_fid_alloc = osd_fid_alloc,
8181 static const struct lu_device_type_operations osd_device_type_ops = {
8182 .ldto_init = osd_type_init,
8183 .ldto_fini = osd_type_fini,
8185 .ldto_start = osd_type_start,
8186 .ldto_stop = osd_type_stop,
8188 .ldto_device_alloc = osd_device_alloc,
8189 .ldto_device_free = osd_device_free,
8191 .ldto_device_init = osd_device_init,
8192 .ldto_device_fini = osd_device_fini
8195 static struct lu_device_type osd_device_type = {
8196 .ldt_tags = LU_DEVICE_DT,
8197 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8198 .ldt_ops = &osd_device_type_ops,
8199 .ldt_ctx_tags = LCT_LOCAL,
8202 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8204 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8205 struct super_block *sb = osd_sb(osd);
8207 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8211 * lprocfs legacy support.
8213 static const struct obd_ops osd_obd_device_ops = {
8214 .o_owner = THIS_MODULE,
8215 .o_connect = osd_obd_connect,
8216 .o_disconnect = osd_obd_disconnect,
8217 .o_health_check = osd_health_check,
8220 static ssize_t track_declares_assert_show(struct kobject *kobj,
8221 struct attribute *attr,
8224 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8227 static ssize_t track_declares_assert_store(struct kobject *kobj,
8228 struct attribute *attr,
8229 const char *buffer, size_t count)
8231 bool track_declares_assert;
8234 rc = kstrtobool(buffer, &track_declares_assert);
8238 ldiskfs_track_declares_assert = track_declares_assert;
8242 LUSTRE_RW_ATTR(track_declares_assert);
8244 static int __init osd_init(void)
8246 struct kobject *kobj;
8249 BUILD_BUG_ON(BH_DXLock >=
8250 sizeof(((struct buffer_head *)0)->b_state) * 8);
8251 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8252 /* please, try to keep osd_thread_info smaller than a page */
8253 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8258 rc = lu_kmem_init(ldiskfs_caches);
8262 #ifdef CONFIG_KALLSYMS
8263 priv_security_file_alloc =
8264 (void *)kallsyms_lookup_name("security_file_alloc");
8267 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
8268 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8270 lu_kmem_fini(ldiskfs_caches);
8274 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8276 rc = sysfs_create_file(kobj,
8277 &lustre_attr_track_declares_assert.attr);
8280 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8287 static void __exit osd_exit(void)
8289 struct kobject *kobj;
8291 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8293 sysfs_remove_file(kobj,
8294 &lustre_attr_track_declares_assert.attr);
8297 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8298 lu_kmem_fini(ldiskfs_caches);
8301 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8302 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8303 MODULE_VERSION(LUSTRE_VERSION_STRING);
8304 MODULE_LICENSE("GPL");
8306 module_init(osd_init);
8307 module_exit(osd_exit);