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_free(h);
1591 * Concurrency: no concurrent access is possible that late in object
1594 static void osd_index_fini(struct osd_object *o)
1596 struct iam_container *bag;
1598 if (o->oo_dir != NULL) {
1599 bag = &o->oo_dir->od_container;
1600 if (o->oo_inode != NULL) {
1601 if (bag->ic_object == o->oo_inode)
1602 iam_container_fini(bag);
1604 OBD_FREE_PTR(o->oo_dir);
1610 OSD_TXN_OI_DELETE_CREDITS = 20,
1611 OSD_TXN_INODE_DELETE_CREDITS = 20
1618 #if OSD_THANDLE_STATS
1620 * Set time when the handle is allocated
1622 static void osd_th_alloced(struct osd_thandle *oth)
1624 oth->oth_alloced = ktime_get();
1628 * Set time when the handle started
1630 static void osd_th_started(struct osd_thandle *oth)
1632 oth->oth_started = ktime_get();
1636 * Check whether the we deal with this handle for too long.
1638 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1639 ktime_t alloced, ktime_t started,
1642 ktime_t now = ktime_get();
1644 LASSERT(dev != NULL);
1646 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1647 ktime_us_delta(started, alloced));
1648 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1649 ktime_us_delta(closed, started));
1650 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1651 ktime_us_delta(now, closed));
1653 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1654 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1655 oth, now, alloced, started, closed);
1656 libcfs_debug_dumpstack(NULL);
1660 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1662 ktime_t __closed = ktime_get(); \
1663 ktime_t __alloced = oth->oth_alloced; \
1664 ktime_t __started = oth->oth_started; \
1667 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1670 #else /* OSD_THANDLE_STATS */
1672 #define osd_th_alloced(h) do {} while(0)
1673 #define osd_th_started(h) do {} while(0)
1674 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1676 #endif /* OSD_THANDLE_STATS */
1679 * Concurrency: doesn't access mutable data.
1681 static int osd_param_is_not_sane(const struct osd_device *dev,
1682 const struct thandle *th)
1684 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1686 return oh->ot_credits > osd_transaction_size(dev);
1690 * Concurrency: shouldn't matter.
1692 static void osd_trans_commit_cb(struct super_block *sb,
1693 struct ldiskfs_journal_cb_entry *jcb, int error)
1695 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1696 struct thandle *th = &oh->ot_super;
1697 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1698 struct dt_txn_commit_cb *dcb, *tmp;
1700 LASSERT(oh->ot_handle == NULL);
1703 CERROR("transaction @0x%p commit error: %d\n", th, error);
1705 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1706 /* call per-transaction callbacks if any */
1707 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1709 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1710 "commit callback entry: magic=%x name='%s'\n",
1711 dcb->dcb_magic, dcb->dcb_name);
1712 list_del_init(&dcb->dcb_linkage);
1713 dcb->dcb_func(NULL, th, dcb, error);
1716 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1723 static struct thandle *osd_trans_create(const struct lu_env *env,
1724 struct dt_device *d)
1726 struct osd_thread_info *oti = osd_oti_get(env);
1727 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1728 struct osd_thandle *oh;
1734 CERROR("%s: someone try to start transaction under "
1735 "readonly mode, should be disabled.\n",
1736 osd_name(osd_dt_dev(d)));
1738 RETURN(ERR_PTR(-EROFS));
1741 /* on pending IO in this thread should left from prev. request */
1742 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1744 sb_start_write(osd_sb(osd_dt_dev(d)));
1746 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1748 sb_end_write(osd_sb(osd_dt_dev(d)));
1749 RETURN(ERR_PTR(-ENOMEM));
1752 oh->ot_quota_trans = &oti->oti_quota_trans;
1753 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1758 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1759 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1760 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1763 memset(oti->oti_declare_ops, 0,
1764 sizeof(oti->oti_declare_ops));
1765 memset(oti->oti_declare_ops_cred, 0,
1766 sizeof(oti->oti_declare_ops_cred));
1767 memset(oti->oti_declare_ops_used, 0,
1768 sizeof(oti->oti_declare_ops_used));
1770 oti->oti_ins_cache_depth++;
1775 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1777 struct osd_thread_info *oti = osd_oti_get(env);
1778 struct osd_thandle *oh;
1780 oh = container_of(th, struct osd_thandle, ot_super);
1781 LASSERT(oh != NULL);
1783 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1784 oti->oti_declare_ops[OSD_OT_CREATE],
1785 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1786 oti->oti_declare_ops_used[OSD_OT_CREATE],
1787 oti->oti_declare_ops[OSD_OT_DESTROY],
1788 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1789 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1790 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1791 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1792 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1793 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1794 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1795 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1796 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1797 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1798 oti->oti_declare_ops[OSD_OT_WRITE],
1799 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1800 oti->oti_declare_ops_used[OSD_OT_WRITE],
1801 oti->oti_declare_ops[OSD_OT_PUNCH],
1802 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1803 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1804 oti->oti_declare_ops[OSD_OT_QUOTA],
1805 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1806 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1807 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1808 oti->oti_declare_ops[OSD_OT_INSERT],
1809 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1810 oti->oti_declare_ops_used[OSD_OT_INSERT],
1811 oti->oti_declare_ops[OSD_OT_DELETE],
1812 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1813 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1814 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1815 oti->oti_declare_ops[OSD_OT_REF_ADD],
1816 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1817 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1818 oti->oti_declare_ops[OSD_OT_REF_DEL],
1819 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1820 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1824 * Concurrency: shouldn't matter.
1826 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1829 struct osd_thread_info *oti = osd_oti_get(env);
1830 struct osd_device *dev = osd_dt_dev(d);
1832 struct osd_thandle *oh;
1837 LASSERT(current->journal_info == NULL);
1839 oh = container_of(th, struct osd_thandle, ot_super);
1840 LASSERT(oh != NULL);
1841 LASSERT(oh->ot_handle == NULL);
1843 rc = dt_txn_hook_start(env, d, th);
1847 if (unlikely(osd_param_is_not_sane(dev, th))) {
1848 static unsigned long last_printed;
1849 static int last_credits;
1852 * don't make noise on a tiny testing systems
1853 * actual credits misuse will be caught anyway
1855 if (last_credits != oh->ot_credits &&
1856 time_after(jiffies, last_printed +
1857 cfs_time_seconds(60)) &&
1858 osd_transaction_size(dev) > 512) {
1859 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1860 oh->ot_credits, osd_transaction_size(dev));
1861 osd_trans_dump_creds(env, th);
1862 libcfs_debug_dumpstack(NULL);
1863 last_credits = oh->ot_credits;
1864 last_printed = jiffies;
1867 * XXX Limit the credits to 'max_transaction_buffers', and
1868 * let the underlying filesystem to catch the error if
1869 * we really need so many credits.
1871 * This should be removed when we can calculate the
1872 * credits precisely.
1874 oh->ot_credits = osd_transaction_size(dev);
1875 } else if (ldiskfs_track_declares_assert != 0) {
1877 * reserve few credits to prevent an assertion in JBD
1878 * our debugging mechanism will be able to detected
1879 * overuse. this can help to debug single-update
1882 oh->ot_credits += 10;
1883 if (unlikely(osd_param_is_not_sane(dev, th)))
1884 oh->ot_credits = osd_transaction_size(dev);
1887 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1888 GOTO(out, rc = -EIO);
1891 * XXX temporary stuff. Some abstraction layer should
1894 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1898 LASSERT(oti->oti_txns == 0);
1900 lu_device_get(&d->dd_lu_dev);
1901 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1912 static int osd_seq_exists(const struct lu_env *env,
1913 struct osd_device *osd, u64 seq)
1915 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1916 struct seq_server_site *ss = osd_seq_site(osd);
1921 LASSERT(ss != NULL);
1922 LASSERT(ss->ss_server_fld != NULL);
1924 rc = osd_fld_lookup(env, osd, seq, range);
1927 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1928 osd_name(osd), seq, rc);
1932 RETURN(ss->ss_node_id == range->lsr_index);
1935 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1937 struct dt_txn_commit_cb *dcb;
1938 struct dt_txn_commit_cb *tmp;
1940 /* call per-transaction stop callbacks if any */
1941 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1943 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1944 "commit callback entry: magic=%x name='%s'\n",
1945 dcb->dcb_magic, dcb->dcb_name);
1946 list_del_init(&dcb->dcb_linkage);
1947 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1952 * Concurrency: shouldn't matter.
1954 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1957 struct osd_thread_info *oti = osd_oti_get(env);
1958 struct osd_thandle *oh;
1959 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1960 struct osd_device *osd = osd_dt_dev(th->th_dev);
1961 struct qsd_instance *qsd = osd_def_qsd(osd);
1962 struct lquota_trans *qtrans;
1963 LIST_HEAD(truncates);
1964 int rc = 0, remove_agents = 0;
1968 oh = container_of(th, struct osd_thandle, ot_super);
1970 remove_agents = oh->ot_remove_agents;
1972 qtrans = oh->ot_quota_trans;
1973 oh->ot_quota_trans = NULL;
1975 /* move locks to local list, stop tx, execute truncates */
1976 list_splice(&oh->ot_trunc_locks, &truncates);
1978 if (oh->ot_handle != NULL) {
1981 handle_t *hdl = oh->ot_handle;
1984 * add commit callback
1985 * notice we don't do this in osd_trans_start()
1986 * as underlying transaction can change during truncate
1988 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1991 LASSERT(oti->oti_txns == 1);
1994 rc = dt_txn_hook_stop(env, th);
1996 CERROR("%s: failed in transaction hook: rc = %d\n",
1999 osd_trans_stop_cb(oh, rc);
2000 /* hook functions might modify th_sync */
2001 hdl->h_sync = th->th_sync;
2003 oh->ot_handle = NULL;
2004 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2006 CERROR("%s: failed to stop transaction: rc = %d\n",
2007 osd_name(osd), rc2);
2011 osd_process_truncates(&truncates);
2013 osd_trans_stop_cb(oh, th->th_result);
2017 osd_trunc_unlock_all(env, &truncates);
2019 /* inform the quota slave device that the transaction is stopping */
2020 qsd_op_end(env, qsd, qtrans);
2023 * as we want IO to journal and data IO be concurrent, we don't block
2024 * awaiting data IO completion in osd_do_bio(), instead we wait here
2025 * once transaction is submitted to the journal. all reqular requests
2026 * don't do direct IO (except read/write), thus this wait_event becomes
2029 * IMPORTANT: we have to wait till any IO submited by the thread is
2030 * completed otherwise iobuf may be corrupted by different request
2032 wait_event(iobuf->dr_wait,
2033 atomic_read(&iobuf->dr_numreqs) == 0);
2034 osd_fini_iobuf(osd, iobuf);
2036 rc = iobuf->dr_error;
2038 if (unlikely(remove_agents != 0))
2039 osd_process_scheduled_agent_removals(env, osd);
2041 oti->oti_ins_cache_depth--;
2042 /* reset OI cache for safety */
2043 if (oti->oti_ins_cache_depth == 0)
2044 oti->oti_ins_cache_used = 0;
2046 sb_end_write(osd_sb(osd));
2051 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2053 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2056 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2057 LASSERT(&dcb->dcb_func != NULL);
2058 if (dcb->dcb_flags & DCB_TRANS_STOP)
2059 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2061 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2067 * Called just before object is freed. Releases all resources except for
2068 * object itself (that is released by osd_object_free()).
2070 * Concurrency: no concurrent access is possible that late in object
2073 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2075 struct osd_object *obj = osd_obj(l);
2076 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2077 struct inode *inode = obj->oo_inode;
2082 LINVRNT(osd_invariant(obj));
2085 * If object is unlinked remove fid->ino mapping from object index.
2088 osd_index_fini(obj);
2093 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2094 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2096 uid = i_uid_read(inode);
2097 gid = i_gid_read(inode);
2098 projid = i_projid_read(inode);
2100 obj->oo_inode = NULL;
2103 /* do not rebalance quota if the caller needs to release memory
2104 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2105 * transaction and risk to deadlock - LU-12178 */
2106 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2109 if (!obj->oo_header && qsd) {
2110 struct osd_thread_info *info = osd_oti_get(env);
2111 struct lquota_id_info *qi = &info->oti_qi;
2113 /* Release granted quota to master if necessary */
2114 qi->lqi_id.qid_uid = uid;
2115 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2117 qi->lqi_id.qid_uid = gid;
2118 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2120 qi->lqi_id.qid_uid = projid;
2121 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2126 * Concurrency: ->loo_object_release() is called under site spin-lock.
2128 static void osd_object_release(const struct lu_env *env,
2129 struct lu_object *l)
2131 struct osd_object *o = osd_obj(l);
2134 * nobody should be releasing a non-destroyed object with nlink=0
2135 * the API allows this, but ldiskfs doesn't like and then report
2136 * this inode as deleted
2138 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2139 o->oo_inode->i_nlink == 0));
2143 * Concurrency: shouldn't matter.
2145 static int osd_object_print(const struct lu_env *env, void *cookie,
2146 lu_printer_t p, const struct lu_object *l)
2148 struct osd_object *o = osd_obj(l);
2149 struct iam_descr *d;
2151 if (o->oo_dir != NULL)
2152 d = o->oo_dir->od_container.ic_descr;
2155 return (*p)(env, cookie,
2156 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2158 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2159 o->oo_inode ? o->oo_inode->i_generation : 0,
2160 d ? d->id_ops->id_name : "plain");
2164 * Concurrency: shouldn't matter.
2166 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2167 struct obd_statfs *sfs, struct obd_statfs_info *info)
2169 struct osd_device *osd = osd_dt_dev(d);
2170 struct super_block *sb = osd_sb(osd);
2171 struct kstatfs *ksfs;
2175 if (unlikely(osd->od_mnt == NULL))
2176 return -EINPROGRESS;
2178 /* osd_lproc.c call this without env, allocate ksfs for that case */
2179 if (unlikely(env == NULL)) {
2180 OBD_ALLOC_PTR(ksfs);
2184 ksfs = &osd_oti_get(env)->oti_ksfs;
2187 result = sb->s_op->statfs(sb->s_root, ksfs);
2191 statfs_pack(sfs, ksfs);
2192 if (unlikely(sb->s_flags & SB_RDONLY))
2193 sfs->os_state |= OS_STATFS_READONLY;
2195 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2197 if (ldiskfs_has_feature_extents(sb))
2198 sfs->os_maxbytes = sb->s_maxbytes;
2200 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2203 * Reserve some space so to avoid fragmenting the filesystem too much.
2204 * Fragmentation not only impacts performance, but can also increase
2205 * metadata overhead significantly, causing grant calculation to be
2208 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2210 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2211 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2212 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2213 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2215 sfs->os_blocks -= reserved;
2216 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2217 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2220 if (unlikely(env == NULL))
2226 * Estimate space needed for file creations. We assume the largest filename
2227 * which is 2^64 - 1, hence a filename of 20 chars.
2228 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2230 #ifdef __LDISKFS_DIR_REC_LEN
2231 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2233 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2237 * Concurrency: doesn't access mutable data.
2239 static void osd_conf_get(const struct lu_env *env,
2240 const struct dt_device *dev,
2241 struct dt_device_param *param)
2243 struct osd_device *d = osd_dt_dev(dev);
2244 struct super_block *sb = osd_sb(d);
2245 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2250 * XXX should be taken from not-yet-existing fs abstraction layer.
2252 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2253 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2254 param->ddp_symlink_max = sb->s_blocksize;
2255 param->ddp_mount_type = LDD_MT_LDISKFS;
2256 if (ldiskfs_has_feature_extents(sb))
2257 param->ddp_maxbytes = sb->s_maxbytes;
2259 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2261 * inode are statically allocated, so per-inode space consumption
2262 * is the space consumed by the directory entry
2264 param->ddp_inodespace = PER_OBJ_USAGE;
2266 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2267 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2268 * maximum length to not under-count the actual number of extents
2269 * needed for writing a file if there are sub-optimal block allocations.
2271 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2272 /* worst-case extent insertion metadata overhead */
2273 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2274 param->ddp_mntopts = 0;
2275 if (test_opt(sb, XATTR_USER))
2276 param->ddp_mntopts |= MNTOPT_USERXATTR;
2277 if (test_opt(sb, POSIX_ACL))
2278 param->ddp_mntopts |= MNTOPT_ACL;
2281 * LOD might calculate the max stripe count based on max_ea_size,
2282 * so we need take account in the overhead as well,
2283 * xattr_header + magic + xattr_entry_head
2285 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2286 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2288 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2289 if (ldiskfs_has_feature_ea_inode(sb))
2290 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2294 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2296 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2297 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2300 * Preferred RPC size for efficient disk IO. 4MB shows good
2301 * all-around performance for ldiskfs, but use bigalloc chunk size
2302 * by default if larger.
2304 #if defined(LDISKFS_CLUSTER_SIZE)
2305 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2306 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2309 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2311 param->ddp_t10_cksum_type = 0;
2313 unsigned short interval = blk_integrity_interval(bi);
2314 name = blk_integrity_name(bi);
2322 if (strncmp(name, "T10-DIF-TYPE",
2323 sizeof("T10-DIF-TYPE") - 1) == 0) {
2324 /* also skip "1/3-" at end */
2325 const int type_off = sizeof("T10-DIF-TYPE.");
2326 char type_number = name[type_off - 2];
2328 if (interval != 512 && interval != 4096) {
2329 CERROR("%s: unsupported T10PI sector size %u\n",
2330 d->od_svname, interval);
2331 } else if (type_number != '1' && type_number != '3') {
2332 CERROR("%s: unsupported T10PI type %s\n",
2333 d->od_svname, name);
2334 } else if (strcmp(name + type_off, "CRC") == 0) {
2335 d->od_t10_type = type_number == '1' ?
2336 OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2337 param->ddp_t10_cksum_type = interval == 512 ?
2338 OBD_CKSUM_T10CRC512 :
2340 } else if (strcmp(name + type_off, "IP") == 0) {
2341 d->od_t10_type = type_number == '1' ?
2342 OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2343 param->ddp_t10_cksum_type = interval == 512 ?
2344 OBD_CKSUM_T10IP512 :
2347 CERROR("%s: unsupported checksum type of "
2349 d->od_svname, name);
2353 CERROR("%s: unsupported T10PI type '%s'",
2354 d->od_svname, name);
2358 param->ddp_has_lseek_data_hole = true;
2361 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2363 return osd_sb(osd_dt_dev(d));
2367 * Concurrency: shouldn't matter.
2369 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2372 struct super_block *s = osd_sb(osd_dt_dev(d));
2375 down_read(&s->s_umount);
2376 rc = s->s_op->sync_fs(s, 1);
2377 up_read(&s->s_umount);
2379 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2386 * Start commit for OSD device.
2388 * An implementation of dt_commit_async method for OSD device.
2389 * Asychronously starts underlayng fs sync and thereby a transaction
2392 * \param env environment
2393 * \param d dt device
2395 * \see dt_device_operations
2397 static int osd_commit_async(const struct lu_env *env,
2398 struct dt_device *d)
2400 struct super_block *s = osd_sb(osd_dt_dev(d));
2405 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2406 down_read(&s->s_umount);
2407 rc = s->s_op->sync_fs(s, 0);
2408 up_read(&s->s_umount);
2413 static int (*priv_security_file_alloc)(struct file *file);
2415 int osd_security_file_alloc(struct file *file)
2417 if (priv_security_file_alloc)
2418 return priv_security_file_alloc(file);
2423 * Concurrency: shouldn't matter.
2425 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2427 struct super_block *sb = osd_sb(osd_dt_dev(d));
2428 struct block_device *dev = sb->s_bdev;
2429 int rc = -EOPNOTSUPP;
2433 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2434 osd_dt_dev(d)->od_svname, (long)dev, rc);
2440 * Note: we do not count into QUOTA here.
2441 * If we mount with --data_journal we may need more.
2443 const int osd_dto_credits_noquota[DTO_NR] = {
2446 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2447 * SINGLEDATA_TRANS_BLOCKS(8)
2448 * XXX Note: maybe iam need more, since iam have more level than
2451 [DTO_INDEX_INSERT] = 16,
2454 * just modify a single entry, probably merge few within a block
2456 [DTO_INDEX_DELETE] = 1,
2460 [DTO_INDEX_UPDATE] = 16,
2462 * 4(inode, inode bits, groups, GDT)
2463 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2465 [DTO_OBJECT_CREATE] = 4,
2467 * 4(inode, inode bits, groups, GDT)
2468 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2470 [DTO_OBJECT_DELETE] = 4,
2472 * Attr set credits (inode)
2474 [DTO_ATTR_SET_BASE] = 1,
2476 * Xattr set. The same as xattr of EXT3.
2477 * DATA_TRANS_BLOCKS(14)
2478 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2479 * are also counted in. Do not know why?
2481 [DTO_XATTR_SET] = 14,
2483 * credits for inode change during write.
2485 [DTO_WRITE_BASE] = 3,
2487 * credits for single block write.
2489 [DTO_WRITE_BLOCK] = 14,
2491 * Attr set credits for chown.
2492 * This is extra credits for setattr, and it is null without quota
2494 [DTO_ATTR_SET_CHOWN] = 0
2497 static const struct dt_device_operations osd_dt_ops = {
2498 .dt_root_get = osd_root_get,
2499 .dt_statfs = osd_statfs,
2500 .dt_trans_create = osd_trans_create,
2501 .dt_trans_start = osd_trans_start,
2502 .dt_trans_stop = osd_trans_stop,
2503 .dt_trans_cb_add = osd_trans_cb_add,
2504 .dt_conf_get = osd_conf_get,
2505 .dt_mnt_sb_get = osd_mnt_sb_get,
2506 .dt_sync = osd_sync,
2508 .dt_commit_async = osd_commit_async,
2511 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2514 struct osd_object *obj = osd_dt_obj(dt);
2515 struct osd_thread_info *oti = osd_oti_get(env);
2517 LINVRNT(osd_invariant(obj));
2519 LASSERT(obj->oo_owner != env);
2520 down_read_nested(&obj->oo_sem, role);
2522 LASSERT(obj->oo_owner == NULL);
2526 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2529 struct osd_object *obj = osd_dt_obj(dt);
2530 struct osd_thread_info *oti = osd_oti_get(env);
2532 LINVRNT(osd_invariant(obj));
2534 LASSERT(obj->oo_owner != env);
2535 down_write_nested(&obj->oo_sem, role);
2537 LASSERT(obj->oo_owner == NULL);
2538 obj->oo_owner = env;
2542 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2544 struct osd_object *obj = osd_dt_obj(dt);
2545 struct osd_thread_info *oti = osd_oti_get(env);
2547 LINVRNT(osd_invariant(obj));
2549 LASSERT(oti->oti_r_locks > 0);
2551 up_read(&obj->oo_sem);
2554 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2556 struct osd_object *obj = osd_dt_obj(dt);
2557 struct osd_thread_info *oti = osd_oti_get(env);
2559 LINVRNT(osd_invariant(obj));
2561 LASSERT(obj->oo_owner == env);
2562 LASSERT(oti->oti_w_locks > 0);
2564 obj->oo_owner = NULL;
2565 up_write(&obj->oo_sem);
2568 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2570 struct osd_object *obj = osd_dt_obj(dt);
2572 LINVRNT(osd_invariant(obj));
2574 return obj->oo_owner == env;
2577 static void osd_inode_getattr(const struct lu_env *env,
2578 struct inode *inode, struct lu_attr *attr)
2580 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2581 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2582 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2583 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2585 attr->la_atime = inode->i_atime.tv_sec;
2586 attr->la_mtime = inode->i_mtime.tv_sec;
2587 attr->la_ctime = inode->i_ctime.tv_sec;
2588 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2589 attr->la_mode = inode->i_mode;
2590 attr->la_size = i_size_read(inode);
2591 attr->la_blocks = inode->i_blocks;
2592 attr->la_uid = i_uid_read(inode);
2593 attr->la_gid = i_gid_read(inode);
2594 attr->la_projid = i_projid_read(inode);
2595 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2596 attr->la_nlink = inode->i_nlink;
2597 attr->la_rdev = inode->i_rdev;
2598 attr->la_blksize = 1 << inode->i_blkbits;
2599 attr->la_blkbits = inode->i_blkbits;
2601 * Ext4 did not transfer inherit flags from raw inode
2602 * to inode flags, and ext4 internally test raw inode
2603 * @i_flags directly. Instead of patching ext4, we do it here.
2605 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2606 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2609 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2612 struct osd_object *obj = osd_dt_obj(dt);
2613 const struct dt_it_ops *iops;
2619 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2620 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2622 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2623 *count = obj->oo_dirent_count;
2627 /* directory not initialized yet */
2628 if (!dt->do_index_ops) {
2633 iops = &dt->do_index_ops->dio_it;
2634 it = iops->init(env, dt, LUDA_64BITHASH);
2636 RETURN(PTR_ERR(it));
2638 rc = iops->load(env, it, 0);
2640 if (rc == -ENODATA) {
2647 rc = iops->next(env, it);
2649 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2653 if (iops->key_size(env, it) == 0)
2659 obj->oo_dirent_count = *count;
2664 iops->fini(env, it);
2669 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2670 struct lu_attr *attr)
2672 struct osd_object *obj = osd_dt_obj(dt);
2675 if (unlikely(!dt_object_exists(dt)))
2677 if (unlikely(obj->oo_destroyed))
2680 LASSERT(!dt_object_remote(dt));
2681 LINVRNT(osd_invariant(obj));
2683 spin_lock(&obj->oo_guard);
2684 osd_inode_getattr(env, obj->oo_inode, attr);
2685 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2686 attr->la_valid |= LA_FLAGS;
2687 attr->la_flags |= LUSTRE_ORPHAN_FL;
2689 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2690 attr->la_valid |= LA_FLAGS;
2691 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2693 spin_unlock(&obj->oo_guard);
2695 if (S_ISDIR(obj->oo_inode->i_mode) &&
2696 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2697 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2702 static int osd_declare_attr_qid(const struct lu_env *env,
2703 struct osd_object *obj,
2704 struct osd_thandle *oh, long long bspace,
2705 qid_t old_id, qid_t new_id, bool enforce,
2706 unsigned int type, bool ignore_edquot)
2709 struct osd_thread_info *info = osd_oti_get(env);
2710 struct lquota_id_info *qi = &info->oti_qi;
2712 qi->lqi_type = type;
2713 /* inode accounting */
2714 qi->lqi_is_blk = false;
2716 /* one more inode for the new id ... */
2717 qi->lqi_id.qid_uid = new_id;
2719 /* Reserve credits for the new id */
2720 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2721 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2726 /* and one less inode for the current id */
2727 qi->lqi_id.qid_uid = old_id;
2729 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2730 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2735 /* block accounting */
2736 qi->lqi_is_blk = true;
2738 /* more blocks for the new id ... */
2739 qi->lqi_id.qid_uid = new_id;
2740 qi->lqi_space = bspace;
2742 * Credits for the new uid has been reserved, re-use "obj"
2743 * to save credit reservation.
2745 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2746 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2751 /* and finally less blocks for the current uid */
2752 qi->lqi_id.qid_uid = old_id;
2753 qi->lqi_space = -bspace;
2754 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2755 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2761 static int osd_declare_attr_set(const struct lu_env *env,
2762 struct dt_object *dt,
2763 const struct lu_attr *attr,
2764 struct thandle *handle)
2766 struct osd_thandle *oh;
2767 struct osd_object *obj;
2776 LASSERT(dt != NULL);
2777 LASSERT(handle != NULL);
2779 obj = osd_dt_obj(dt);
2780 LASSERT(osd_invariant(obj));
2782 oh = container_of(handle, struct osd_thandle, ot_super);
2783 LASSERT(oh->ot_handle == NULL);
2785 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2786 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2788 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2789 osd_dto_credits_noquota[DTO_XATTR_SET]);
2791 if (attr == NULL || obj->oo_inode == NULL)
2794 bspace = obj->oo_inode->i_blocks << 9;
2795 bspace = toqb(bspace);
2798 * Changing ownership is always preformed by super user, it should not
2799 * fail with EDQUOT unless required explicitly.
2801 * We still need to call the osd_declare_qid() to calculate the journal
2802 * credits for updating quota accounting files and to trigger quota
2803 * space adjustment once the operation is completed.
2805 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2806 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2809 CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2810 "(the quota space is %lld)\n",
2811 obj->oo_inode->i_sb->s_id, attr->la_uid,
2812 attr->la_gid, bspace);
2815 uid = i_uid_read(obj->oo_inode);
2816 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2817 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2818 attr->la_uid, enforce, USRQUOTA,
2823 gid = i_gid_read(obj->oo_inode);
2824 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
2825 attr->la_uid, gid, attr->la_gid);
2826 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2827 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
2828 attr->la_gid, enforce, GRPQUOTA,
2834 #ifdef HAVE_PROJECT_QUOTA
2835 if (attr->la_valid & LA_PROJID) {
2836 __u32 projid = i_projid_read(obj->oo_inode);
2838 enforce = (attr->la_valid & LA_PROJID) &&
2839 (attr->la_projid != projid);
2840 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2841 (qid_t)projid, (qid_t)attr->la_projid,
2842 enforce, PRJQUOTA, true);
2850 static int osd_inode_setattr(const struct lu_env *env,
2851 struct inode *inode, const struct lu_attr *attr)
2853 __u64 bits = attr->la_valid;
2855 /* Only allow set size for regular file */
2856 if (!S_ISREG(inode->i_mode))
2857 bits &= ~(LA_SIZE | LA_BLOCKS);
2862 if (bits & LA_ATIME)
2863 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2864 if (bits & LA_CTIME)
2865 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2866 if (bits & LA_MTIME)
2867 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2868 if (bits & LA_SIZE) {
2869 spin_lock(&inode->i_lock);
2870 LDISKFS_I(inode)->i_disksize = attr->la_size;
2871 i_size_write(inode, attr->la_size);
2872 spin_unlock(&inode->i_lock);
2876 * OSD should not change "i_blocks" which is used by quota.
2877 * "i_blocks" should be changed by ldiskfs only.
2880 inode->i_mode = (inode->i_mode & S_IFMT) |
2881 (attr->la_mode & ~S_IFMT);
2883 i_uid_write(inode, attr->la_uid);
2885 i_gid_write(inode, attr->la_gid);
2886 if (bits & LA_PROJID)
2887 i_projid_write(inode, attr->la_projid);
2888 if (bits & LA_NLINK)
2889 set_nlink(inode, attr->la_nlink);
2891 inode->i_rdev = attr->la_rdev;
2893 if (bits & LA_FLAGS) {
2894 /* always keep S_NOCMTIME */
2895 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2897 #if defined(S_ENCRYPTED)
2898 /* Always remove S_ENCRYPTED, because ldiskfs must not be
2899 * aware of encryption status. It is just stored into LMA
2900 * so that it can be forwared to client side.
2902 inode->i_flags &= ~S_ENCRYPTED;
2905 * Ext4 did not transfer inherit flags from
2906 * @inode->i_flags to raw inode i_flags when writing
2907 * flags, we do it explictly here.
2909 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2910 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2912 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2917 #ifdef HAVE_PROJECT_QUOTA
2918 static int osd_transfer_project(struct inode *inode, __u32 projid,
2919 struct thandle *handle)
2921 struct super_block *sb = inode->i_sb;
2922 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2925 struct ldiskfs_iloc iloc;
2926 struct ldiskfs_inode *raw_inode;
2927 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2929 if (!ldiskfs_has_feature_project(sb)) {
2930 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2931 == LDISKFS_DEF_PROJID);
2932 if (projid != LDISKFS_DEF_PROJID)
2938 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2941 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2942 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2945 err = ldiskfs_get_inode_loc(inode, &iloc);
2949 raw_inode = ldiskfs_raw_inode(&iloc);
2950 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2951 struct osd_thandle *oh = container_of(handle,
2955 * try to expand inode size automatically.
2957 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
2958 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2966 dquot_initialize(inode);
2967 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2968 if (transfer_to[PRJQUOTA]) {
2969 err = __dquot_transfer(inode, transfer_to);
2970 dqput(transfer_to[PRJQUOTA]);
2979 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
2980 struct thandle *handle)
2984 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2985 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2989 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
2990 inode->i_ino, i_uid_read(inode), attr->la_uid,
2991 i_gid_read(inode), attr->la_gid);
2993 dquot_initialize(inode);
2995 if (attr->la_valid & LA_UID)
2996 iattr.ia_valid |= ATTR_UID;
2997 if (attr->la_valid & LA_GID)
2998 iattr.ia_valid |= ATTR_GID;
2999 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3000 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3002 rc = dquot_transfer(inode, &iattr);
3004 CERROR("%s: quota transfer failed: rc = %d. Is quota "
3005 "enforcement enabled on the ldiskfs "
3006 "filesystem?\n", inode->i_sb->s_id, rc);
3011 /* Handle project id transfer here properly */
3012 if (attr->la_valid & LA_PROJID &&
3013 attr->la_projid != i_projid_read(inode)) {
3014 #ifdef HAVE_PROJECT_QUOTA
3015 rc = osd_transfer_project(inode, attr->la_projid, handle);
3020 CERROR("%s: quota transfer failed: rc = %d. Is project "
3021 "enforcement enabled on the ldiskfs "
3022 "filesystem?\n", inode->i_sb->s_id, rc);
3029 static int osd_attr_set(const struct lu_env *env,
3030 struct dt_object *dt,
3031 const struct lu_attr *attr,
3032 struct thandle *handle)
3034 struct osd_object *obj = osd_dt_obj(dt);
3035 struct inode *inode;
3038 if (!dt_object_exists(dt))
3041 LASSERT(handle != NULL);
3042 LASSERT(!dt_object_remote(dt));
3043 LASSERT(osd_invariant(obj));
3045 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3047 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3048 !osd_obj2dev(obj)->od_is_ost) {
3049 struct osd_thread_info *oti = osd_oti_get(env);
3050 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3051 struct lu_fid *fid1 = &oti->oti_fid;
3052 struct osd_inode_id *id = &oti->oti_id;
3053 struct iam_path_descr *ipd;
3054 struct iam_container *bag;
3055 struct osd_thandle *oh;
3058 fid_cpu_to_be(fid1, fid0);
3059 memset(id, 1, sizeof(*id));
3060 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3061 fid0)->oi_dir.od_container;
3062 ipd = osd_idx_ipd_get(env, bag);
3063 if (unlikely(ipd == NULL))
3066 oh = container_of(handle, struct osd_thandle, ot_super);
3067 rc = iam_update(oh->ot_handle, bag,
3068 (const struct iam_key *)fid1,
3069 (const struct iam_rec *)id, ipd);
3070 osd_ipd_put(env, bag, ipd);
3071 return(rc > 0 ? 0 : rc);
3074 inode = obj->oo_inode;
3076 rc = osd_quota_transfer(inode, attr, handle);
3080 spin_lock(&obj->oo_guard);
3081 rc = osd_inode_setattr(env, inode, attr);
3082 spin_unlock(&obj->oo_guard);
3086 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3088 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3090 if (!(attr->la_valid & LA_FLAGS))
3093 /* Let's check if there are extra flags need to be set into LMA */
3094 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3095 struct osd_thread_info *info = osd_oti_get(env);
3096 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3098 LASSERT(!obj->oo_pfid_in_lma);
3100 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3101 &info->oti_ost_attrs);
3105 lma->lma_incompat |=
3106 lustre_to_lma_flags(attr->la_flags);
3107 lustre_lma_swab(lma);
3109 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3111 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3112 lma, sizeof(*lma), XATTR_REPLACE);
3114 struct osd_device *osd = osd_obj2dev(obj);
3116 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3117 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3118 lma->lma_incompat, rc);
3121 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3123 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3130 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3131 struct osd_object *obj,
3132 const char *name, const int namelen)
3134 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3137 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3138 umode_t mode, struct dt_allocation_hint *hint,
3139 struct thandle *th, struct lu_attr *attr)
3142 struct osd_device *osd = osd_obj2dev(obj);
3143 struct osd_thandle *oth;
3144 struct dt_object *parent = NULL;
3145 struct inode *inode;
3146 uid_t owner[2] = {0, 0};
3148 if (attr->la_valid & LA_UID)
3149 owner[0] = attr->la_uid;
3150 if (attr->la_valid & LA_GID)
3151 owner[1] = attr->la_gid;
3153 LINVRNT(osd_invariant(obj));
3154 LASSERT(obj->oo_inode == NULL);
3155 LASSERT(obj->oo_hl_head == NULL);
3157 if (S_ISDIR(mode) && ldiskfs_pdo) {
3159 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3160 if (obj->oo_hl_head == NULL)
3164 oth = container_of(th, struct osd_thandle, ot_super);
3165 LASSERT(oth->ot_handle->h_transaction != NULL);
3167 if (hint != NULL && hint->dah_parent != NULL &&
3168 !dt_object_remote(hint->dah_parent))
3169 parent = hint->dah_parent;
3171 inode = ldiskfs_create_inode(oth->ot_handle,
3172 parent ? osd_dt_obj(parent)->oo_inode :
3173 osd_sb(osd)->s_root->d_inode,
3175 if (!IS_ERR(inode)) {
3176 /* Do not update file c/mtime in ldiskfs. */
3177 inode->i_flags |= S_NOCMTIME;
3180 * For new created object, it must be consistent,
3181 * and it is unnecessary to scrub against it.
3183 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3185 obj->oo_inode = inode;
3188 if (obj->oo_hl_head != NULL) {
3189 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3190 obj->oo_hl_head = NULL;
3192 result = PTR_ERR(inode);
3194 LINVRNT(osd_invariant(obj));
3202 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3203 struct lu_attr *attr,
3204 struct dt_allocation_hint *hint,
3205 struct dt_object_format *dof,
3209 struct osd_thandle *oth;
3210 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3212 LASSERT(S_ISDIR(attr->la_mode));
3214 oth = container_of(th, struct osd_thandle, ot_super);
3215 LASSERT(oth->ot_handle->h_transaction != NULL);
3216 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3217 obj->oo_dirent_count = 0;
3218 result = osd_mkfile(info, obj, mode, hint, th, attr);
3223 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3224 struct lu_attr *attr,
3225 struct dt_allocation_hint *hint,
3226 struct dt_object_format *dof,
3230 struct osd_thandle *oth;
3231 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3233 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3235 LASSERT(S_ISREG(attr->la_mode));
3237 oth = container_of(th, struct osd_thandle, ot_super);
3238 LASSERT(oth->ot_handle->h_transaction != NULL);
3240 result = osd_mkfile(info, obj, mode, hint, th, attr);
3242 LASSERT(obj->oo_inode != NULL);
3243 if (feat->dif_flags & DT_IND_VARKEY)
3244 result = iam_lvar_create(obj->oo_inode,
3245 feat->dif_keysize_max,
3247 feat->dif_recsize_max,
3250 result = iam_lfix_create(obj->oo_inode,
3251 feat->dif_keysize_max,
3253 feat->dif_recsize_max,
3259 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3260 struct lu_attr *attr,
3261 struct dt_allocation_hint *hint,
3262 struct dt_object_format *dof,
3265 LASSERT(S_ISREG(attr->la_mode));
3266 return osd_mkfile(info, obj, (attr->la_mode &
3267 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3271 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3272 struct lu_attr *attr,
3273 struct dt_allocation_hint *hint,
3274 struct dt_object_format *dof,
3277 LASSERT(S_ISLNK(attr->la_mode));
3278 return osd_mkfile(info, obj, (attr->la_mode &
3279 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3283 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3284 struct lu_attr *attr,
3285 struct dt_allocation_hint *hint,
3286 struct dt_object_format *dof,
3289 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3292 LINVRNT(osd_invariant(obj));
3293 LASSERT(obj->oo_inode == NULL);
3294 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3295 S_ISFIFO(mode) || S_ISSOCK(mode));
3297 result = osd_mkfile(info, obj, mode, hint, th, attr);
3299 LASSERT(obj->oo_inode != NULL);
3301 * This inode should be marked dirty for i_rdev. Currently
3302 * that is done in the osd_attr_init().
3304 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3307 LINVRNT(osd_invariant(obj));
3311 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3313 struct dt_allocation_hint *hint,
3314 struct dt_object_format *dof,
3317 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3319 osd_obj_type_f result;
3335 result = osd_mk_index;
3346 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3347 struct dt_object *parent, struct dt_object *child,
3352 ah->dah_parent = parent;
3353 ah->dah_mode = child_mode;
3355 if (parent != NULL && !dt_object_remote(parent)) {
3356 /* will help to find FID->ino at dt_insert("..") */
3357 struct osd_object *pobj = osd_dt_obj(parent);
3359 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3363 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3364 struct lu_attr *attr, struct dt_object_format *dof,
3365 struct thandle *handle)
3367 struct inode *inode = obj->oo_inode;
3368 __u64 valid = attr->la_valid;
3371 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3373 if (dof->dof_type != DFT_NODE)
3374 attr->la_valid &= ~LA_RDEV;
3375 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3376 attr->la_valid &= ~LA_ATIME;
3377 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3378 attr->la_valid &= ~LA_CTIME;
3379 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3380 attr->la_valid &= ~LA_MTIME;
3382 result = osd_quota_transfer(inode, attr, handle);
3386 if (attr->la_valid != 0) {
3387 result = osd_inode_setattr(info->oti_env, inode, attr);
3389 * The osd_inode_setattr() should always succeed here. The
3390 * only error that could be returned is EDQUOT when we are
3391 * trying to change the UID or GID of the inode. However, this
3392 * should not happen since quota enforcement is no longer
3393 * enabled on ldiskfs (lquota takes care of it).
3395 LASSERTF(result == 0, "%d\n", result);
3396 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3399 attr->la_valid = valid;
3403 * Helper function for osd_create()
3405 * \retval 0, on success
3407 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3408 struct lu_attr *attr, struct dt_allocation_hint *hint,
3409 struct dt_object_format *dof, struct thandle *th)
3414 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3416 /* we drop umask so that permissions we pass are not affected */
3417 umask = current->fs->umask;
3418 current->fs->umask = 0;
3420 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3422 if (likely(obj->oo_inode != NULL)) {
3423 LASSERT(obj->oo_inode->i_state & I_NEW);
3426 * Unlock the inode before attr initialization to avoid
3427 * unnecessary dqget operations. LU-6378
3429 unlock_new_inode(obj->oo_inode);
3432 if (likely(result == 0)) {
3433 osd_attr_init(info, obj, attr, dof, th);
3434 osd_object_init0(obj);
3437 /* restore previous umask value */
3438 current->fs->umask = umask;
3440 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3446 * Helper function for osd_create()
3448 * \retval 0, on success
3450 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3451 const struct lu_fid *fid, struct thandle *th)
3453 struct osd_thread_info *info = osd_oti_get(env);
3454 struct osd_inode_id *id = &info->oti_id;
3455 struct osd_device *osd = osd_obj2dev(obj);
3456 struct osd_thandle *oh;
3459 LASSERT(obj->oo_inode != NULL);
3461 oh = container_of(th, struct osd_thandle, ot_super);
3462 LASSERT(oh->ot_handle);
3463 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3465 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3466 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3467 OI_CHECK_FLD, NULL);
3468 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3469 struct lu_fid next_fid = *fid;
3471 /* insert next object in advance, and map to the same inode */
3473 if (next_fid.f_oid != 0) {
3474 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3475 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3476 OI_CHECK_FLD, NULL);
3477 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3481 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3486 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3487 u64 seq, struct lu_seq_range *range)
3489 struct seq_server_site *ss = osd_seq_site(osd);
3491 if (fid_seq_is_idif(seq)) {
3492 fld_range_set_ost(range);
3493 range->lsr_index = idif_ost_idx(seq);
3497 if (!fid_seq_in_fldb(seq)) {
3498 fld_range_set_mdt(range);
3501 * FIXME: If ss is NULL, it suppose not get lsr_index
3504 range->lsr_index = ss->ss_node_id;
3508 LASSERT(ss != NULL);
3509 fld_range_set_any(range);
3510 /* OSD will only do local fld lookup */
3511 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3514 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3515 struct lu_attr *attr,
3516 struct dt_allocation_hint *hint,
3517 struct dt_object_format *dof,
3518 struct thandle *handle)
3520 struct osd_thandle *oh;
3525 LASSERT(handle != NULL);
3527 oh = container_of(handle, struct osd_thandle, ot_super);
3528 LASSERT(oh->ot_handle == NULL);
3531 * EA object consumes more credits than regular object: osd_mk_index
3532 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3533 * leaf_node, could involves the block, block bitmap, groups, GDT
3534 * change for each block, so add 4 * 2 credits in that case.
3536 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3537 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3538 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3540 * Reuse idle OI block may cause additional one OI block
3543 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3544 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3545 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3546 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3547 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3549 /* will help to find FID->ino mapping at dt_insert() */
3550 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3558 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3559 attr->la_projid, 1, oh, osd_dt_obj(dt),
3560 NULL, OSD_QID_INODE);
3568 * Called to destroy on-disk representation of the object
3570 * Concurrency: must be locked
3572 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3575 struct osd_object *obj = osd_dt_obj(dt);
3576 struct inode *inode = obj->oo_inode;
3577 struct osd_thandle *oh;
3585 oh = container_of(th, struct osd_thandle, ot_super);
3586 LASSERT(oh->ot_handle == NULL);
3588 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3589 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3591 /* For removing agent entry */
3592 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3593 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3596 * Recycle idle OI leaf may cause additional three OI blocks
3599 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3600 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3601 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3602 /* one less inode */
3603 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3604 i_projid_read(inode), -1, oh, obj, NULL,
3608 /* data to be truncated */
3609 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3610 i_projid_read(inode), 0, oh, obj, NULL,
3616 * will help to find FID->ino when this object is being
3619 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3624 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3627 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3628 struct osd_object *obj = osd_dt_obj(dt);
3629 struct inode *inode = obj->oo_inode;
3630 struct osd_device *osd = osd_obj2dev(obj);
3631 struct osd_thandle *oh;
3636 oh = container_of(th, struct osd_thandle, ot_super);
3637 LASSERT(oh->ot_handle);
3639 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3641 if (unlikely(fid_is_acct(fid)))
3644 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3645 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3647 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3648 osd_name(osd), PFID(fid), result);
3651 if (S_ISDIR(inode->i_mode)) {
3652 if (inode->i_nlink > 2)
3653 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3654 osd_name(osd), PFID(fid), inode->i_ino,
3657 spin_lock(&obj->oo_guard);
3659 spin_unlock(&obj->oo_guard);
3660 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3663 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3665 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3667 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3668 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3669 oh->ot_handle, OI_CHECK_FLD);
3671 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3672 /* XXX: add to ext3 orphan list */
3673 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3675 /* not needed in the cache anymore */
3676 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3677 obj->oo_destroyed = 1;
3683 * Put the fid into lustre_mdt_attrs, and then place the structure
3684 * inode's ea. This fid should not be altered during the life time
3687 * \retval +ve, on success
3688 * \retval -ve, on error
3690 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3692 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3693 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3695 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3696 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3701 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3704 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3707 lustre_loa_init(loa, fid, compat, incompat);
3708 lustre_loa_swab(loa, false);
3711 * For the OST device with 256 bytes inode size by default,
3712 * the PFID EA will be stored together with LMA EA to avoid
3713 * performance trouble. Otherwise the PFID EA can be stored
3714 * independently. LU-8998
3716 if ((compat & LMAC_FID_ON_OST) &&
3717 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3718 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3719 sizeof(*loa), XATTR_CREATE);
3721 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3722 sizeof(*lma), XATTR_CREATE);
3724 * LMA may already exist, but we need to check that all the
3725 * desired compat/incompat flags have been added.
3727 if (unlikely(rc == -EEXIST)) {
3728 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3729 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3733 if (rc < sizeof(*lma))
3736 lustre_loa_swab(loa, true);
3737 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3738 ((compat == 0 && incompat == 0) ||
3739 (!(~lma->lma_compat & compat) &&
3740 !(~lma->lma_incompat & incompat))))
3743 lma->lma_self_fid = *fid;
3744 lma->lma_compat |= compat;
3745 lma->lma_incompat |= incompat;
3746 if (rc == sizeof(*lma)) {
3747 lustre_lma_swab(lma);
3748 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3749 sizeof(*lma), XATTR_REPLACE);
3751 lustre_loa_swab(loa, false);
3752 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3753 sizeof(*loa), XATTR_REPLACE);
3761 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3762 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3763 * To have compatilibility with 1.8 ldiskfs driver we need to have
3764 * magic number at start of fid data.
3765 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3768 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3769 const struct lu_fid *fid)
3771 if (!fid_is_namespace_visible(fid) ||
3772 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3773 param->edp_magic = 0;
3777 param->edp_magic = LDISKFS_LUFID_MAGIC;
3778 param->edp_len = sizeof(struct lu_fid) + 1;
3779 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3783 * Try to read the fid from inode ea into dt_rec.
3785 * \param fid object fid.
3787 * \retval 0 on success
3789 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3790 __u32 ino, struct lu_fid *fid,
3791 struct osd_inode_id *id)
3793 struct osd_thread_info *info = osd_oti_get(env);
3794 struct inode *inode;
3798 osd_id_gen(id, ino, OSD_OII_NOGEN);
3799 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3801 RETURN(PTR_ERR(inode));
3807 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3809 struct inode *parent_dir,
3810 const struct lu_fid *dot_fid,
3811 const struct lu_fid *dot_dot_fid,
3812 struct osd_thandle *oth)
3814 struct ldiskfs_dentry_param *dot_ldp;
3815 struct ldiskfs_dentry_param *dot_dot_ldp;
3816 __u32 saved_nlink = dir->i_nlink;
3819 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3820 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3822 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3823 dot_ldp->edp_magic = 0;
3825 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3826 dir, dot_ldp, dot_dot_ldp);
3828 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3829 * the subseqent ref_add() will increase the dir->i_nlink
3830 * as 3. That is incorrect for new created directory.
3832 * It looks like hack, because we want to make the OSD API
3833 * to be order-independent for new created directory object
3834 * between dt_insert(..) and ref_add() operations.
3836 * Here, we only restore the in-RAM dir-inode's nlink attr,
3837 * becuase if the nlink attr is not 2, then there will be
3838 * ref_add() called following the dt_insert(..), such call
3839 * will make both the in-RAM and on-disk dir-inode's nlink
3840 * attr to be set as 2. LU-7447
3842 set_nlink(dir, saved_nlink);
3847 * Create an local agent inode for remote entry
3849 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3850 struct osd_device *osd,
3851 struct osd_object *pobj,
3852 const struct lu_fid *fid,
3856 struct osd_thread_info *info = osd_oti_get(env);
3857 struct inode *local;
3858 struct osd_thandle *oh;
3859 uid_t own[2] = {0, 0};
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, own);
3869 if (IS_ERR(local)) {
3870 CERROR("%s: create local error %d\n", osd_name(osd),
3871 (int)PTR_ERR(local));
3876 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3877 * correct gid on remote file, not agent here
3879 local->i_gid = current_fsgid();
3880 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3882 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
3883 * That gives e2fsck something to look at and be happy, and allows
3884 * debugging if we need to determine where this symlink came from.
3886 if (S_ISLNK(type)) {
3887 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
3888 rc = snprintf((char *)LDISKFS_I(local)->i_data,
3889 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
3891 i_size_write(local, rc);
3892 LDISKFS_I(local)->i_disksize = rc;
3894 unlock_new_inode(local);
3896 /* Agent inode should not have project ID */
3897 #ifdef HAVE_PROJECT_QUOTA
3898 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3899 i_projid_read(pobj->oo_inode) != 0) {
3900 rc = osd_transfer_project(local, 0, th);
3902 CERROR("%s: quota transfer failed: rc = %d. Is project "
3903 "quota enforcement enabled on the ldiskfs "
3904 "filesystem?\n", local->i_sb->s_id, rc);
3905 RETURN(ERR_PTR(rc));
3909 /* Set special LMA flag for local agent inode */
3910 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3912 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3913 osd_name(osd), PFID(fid), rc);
3914 RETURN(ERR_PTR(rc));
3920 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3921 lu_object_fid(&pobj->oo_dt.do_lu),
3924 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3925 osd_name(osd), PFID(fid), rc);
3926 RETURN(ERR_PTR(rc));
3933 * when direntry is deleted, we have to take care of possible agent inode
3934 * referenced by that. unfortunately we can't do this at that point:
3935 * iget() within a running transaction leads to deadlock and we better do
3936 * not call that every delete declaration to save performance. so we put
3937 * a potention agent inode on a list and process that once the transaction
3938 * is over. Notice it's not any worse in terms of real orphans as regular
3939 * object destroy doesn't put inodes on the on-disk orphan list. this should
3940 * be addressed separately
3942 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3943 struct osd_thandle *oh,
3946 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3947 struct osd_obj_orphan *oor;
3954 oor->oor_env = (struct lu_env *)env;
3955 spin_lock(&osd->od_osfs_lock);
3956 list_add(&oor->oor_list, &osd->od_orphan_list);
3957 spin_unlock(&osd->od_osfs_lock);
3959 oh->ot_remove_agents = 1;
3965 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3966 struct osd_device *osd)
3968 struct osd_thread_info *info = osd_oti_get(env);
3969 struct osd_obj_orphan *oor, *tmp;
3970 struct osd_inode_id id;
3972 struct inode *inode;
3977 spin_lock(&osd->od_osfs_lock);
3978 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3979 if (oor->oor_env == env)
3980 list_move(&oor->oor_list, &list);
3982 spin_unlock(&osd->od_osfs_lock);
3984 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3988 list_del(&oor->oor_list);
3991 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3992 inode = osd_iget_fid(info, osd, &id, &fid);
3996 if (!osd_remote_fid(env, osd, &fid)) {
4001 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4003 mark_inode_dirty(inode);
4004 ldiskfs_journal_stop(jh);
4012 * OSD layer object create function for OST objects (b=11826).
4014 * The FID is inserted into inode xattr here.
4016 * \retval 0, on success
4017 * \retval -ve, on error
4019 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4020 struct lu_attr *attr, struct dt_allocation_hint *hint,
4021 struct dt_object_format *dof, struct thandle *th)
4023 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4024 struct osd_object *obj = osd_dt_obj(dt);
4025 struct osd_thread_info *info = osd_oti_get(env);
4026 int result, on_ost = 0;
4030 if (dt_object_exists(dt))
4033 LINVRNT(osd_invariant(obj));
4034 LASSERT(!dt_object_remote(dt));
4035 LASSERT(osd_is_write_locked(env, obj));
4036 LASSERT(th != NULL);
4038 if (unlikely(fid_is_acct(fid)))
4040 * Quota files can't be created from the kernel any more,
4041 * 'tune2fs -O quota' will take care of creating them
4045 result = __osd_create(info, obj, attr, hint, dof, th);
4047 if (fid_is_idif(fid) &&
4048 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4049 struct lu_fid *tfid = &info->oti_fid;
4050 struct ost_id *oi = &info->oti_ostid;
4052 fid_to_ostid(fid, oi);
4053 ostid_to_fid(tfid, oi, 0);
4055 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4056 LMAC_FID_ON_OST, 0);
4058 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4060 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4061 on_ost ? LMAC_FID_ON_OST : 0,
4064 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4065 obj->oo_dt.do_body_ops = &osd_body_ops;
4068 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY))
4069 result = __osd_oi_insert(env, obj, fid, th);
4072 * a small optimization - dt_insert() isn't usually applied
4073 * to OST objects, so we don't need to cache OI mapping for
4076 if (result == 0 && on_ost == 0) {
4077 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4079 result = osd_idc_find_and_init(env, osd, obj);
4080 LASSERT(result == 0);
4083 LASSERT(ergo(result == 0,
4084 dt_object_exists(dt) && !dt_object_remote(dt)));
4085 LINVRNT(osd_invariant(obj));
4089 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4090 struct thandle *handle)
4092 struct osd_thandle *oh;
4094 /* it's possible that object doesn't exist yet */
4095 LASSERT(handle != NULL);
4097 oh = container_of(handle, struct osd_thandle, ot_super);
4098 LASSERT(oh->ot_handle == NULL);
4100 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4101 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4103 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
4109 * Concurrency: @dt is write locked.
4111 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4114 struct osd_object *obj = osd_dt_obj(dt);
4115 struct inode *inode = obj->oo_inode;
4116 struct osd_thandle *oh;
4119 if (!dt_object_exists(dt) || obj->oo_destroyed)
4122 LINVRNT(osd_invariant(obj));
4123 LASSERT(!dt_object_remote(dt));
4124 LASSERT(osd_is_write_locked(env, obj));
4125 LASSERT(th != NULL);
4127 oh = container_of(th, struct osd_thandle, ot_super);
4128 LASSERT(oh->ot_handle != NULL);
4130 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4132 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4133 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4135 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4136 * (65000) subdirectories by storing "1" in i_nlink if the link count
4137 * would otherwise overflow. Directory tranversal tools understand
4138 * that (st_nlink == 1) indicates that the filesystem dose not track
4139 * hard links count on the directory, and will not abort subdirectory
4140 * scanning early once (st_nlink - 2) subdirs have been found.
4142 * This also has to properly handle the case of inodes with nlink == 0
4143 * in case they are being linked into the PENDING directory
4145 spin_lock(&obj->oo_guard);
4146 if (unlikely(inode->i_nlink == 0))
4147 /* inc_nlink from 0 may cause WARN_ON */
4148 set_nlink(inode, 1);
4150 ldiskfs_inc_count(oh->ot_handle, inode);
4151 if (!S_ISDIR(inode->i_mode))
4152 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4154 spin_unlock(&obj->oo_guard);
4156 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4157 LINVRNT(osd_invariant(obj));
4159 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4164 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4165 struct thandle *handle)
4167 struct osd_thandle *oh;
4169 if (!dt_object_exists(dt))
4172 LASSERT(!dt_object_remote(dt));
4173 LASSERT(handle != NULL);
4175 oh = container_of(handle, struct osd_thandle, ot_super);
4176 LASSERT(oh->ot_handle == NULL);
4178 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4179 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4185 * Concurrency: @dt is write locked.
4187 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4190 struct osd_object *obj = osd_dt_obj(dt);
4191 struct inode *inode = obj->oo_inode;
4192 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4193 struct osd_thandle *oh;
4195 if (!dt_object_exists(dt))
4198 LINVRNT(osd_invariant(obj));
4199 LASSERT(!dt_object_remote(dt));
4200 LASSERT(osd_is_write_locked(env, obj));
4201 LASSERT(th != NULL);
4203 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4206 oh = container_of(th, struct osd_thandle, ot_super);
4207 LASSERT(oh->ot_handle != NULL);
4209 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4211 spin_lock(&obj->oo_guard);
4213 * That can be result of upgrade from old Lustre version and
4214 * applied only to local files. Just skip this ref_del call.
4215 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4217 if (inode->i_nlink == 0) {
4218 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4219 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4220 ", maybe an upgraded file? (LU-3915)\n",
4221 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4222 spin_unlock(&obj->oo_guard);
4226 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4227 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4229 ldiskfs_dec_count(oh->ot_handle, inode);
4230 spin_unlock(&obj->oo_guard);
4232 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4233 LINVRNT(osd_invariant(obj));
4235 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4241 * Concurrency: @dt is read locked.
4243 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4244 struct lu_buf *buf, const char *name)
4246 struct osd_object *obj = osd_dt_obj(dt);
4247 struct inode *inode = obj->oo_inode;
4248 struct osd_thread_info *info = osd_oti_get(env);
4249 struct dentry *dentry = &info->oti_obj_dentry;
4250 bool cache_xattr = false;
4255 /* version get is not real XATTR but uses xattr API */
4256 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4257 dt_obj_version_t *ver = buf->lb_buf;
4260 * for version we are just using xattr API but change inode
4263 if (buf->lb_len == 0)
4264 return sizeof(dt_obj_version_t);
4266 if (buf->lb_len < sizeof(dt_obj_version_t))
4269 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4270 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4272 *ver = LDISKFS_I(inode)->i_fs_version;
4274 return sizeof(dt_obj_version_t);
4277 if (!dt_object_exists(dt))
4280 LASSERT(!dt_object_remote(dt));
4281 LASSERT(inode->i_op != NULL);
4282 #ifdef HAVE_IOP_XATTR
4283 LASSERT(inode->i_op->getxattr != NULL);
4286 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4287 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4291 rc = osd_oxc_get(obj, name, buf);
4296 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4297 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4298 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4299 struct filter_fid *ff;
4300 struct ost_layout *ol;
4302 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4306 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4309 if (buf->lb_len == 0 || !buf->lb_buf)
4312 if (buf->lb_len < rc)
4316 ol = &ff->ff_layout;
4317 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4318 PFID_STRIPE_IDX_BITS);
4319 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4320 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4321 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4322 if (lma->lma_compat & LMAC_COMP_INFO) {
4323 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4324 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4325 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4327 ol->ol_comp_start = 0;
4328 ol->ol_comp_end = 0;
4332 rc = __osd_xattr_get(inode, dentry, name,
4333 buf->lb_buf, buf->lb_len);
4337 if (rc == -ENOENT || rc == -ENODATA)
4338 osd_oxc_add(obj, name, NULL, 0);
4339 else if (rc > 0 && buf->lb_buf != NULL)
4340 osd_oxc_add(obj, name, buf->lb_buf, rc);
4346 static int osd_declare_xattr_set(const struct lu_env *env,
4347 struct dt_object *dt,
4348 const struct lu_buf *buf, const char *name,
4349 int fl, struct thandle *handle)
4351 struct osd_thandle *oh;
4353 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4355 LASSERT(handle != NULL);
4357 oh = container_of(handle, struct osd_thandle, ot_super);
4358 LASSERT(oh->ot_handle == NULL);
4360 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4362 * For non-upgrading case, the LMA is set first and
4363 * usually fit inode. But for upgrade case, the LMA
4364 * may be in another separated EA block.
4366 if (dt_object_exists(dt)) {
4367 if (fl == LU_XATTR_REPLACE)
4372 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4374 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4375 /* We may need to delete the old PFID EA. */
4376 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4377 if (fl == LU_XATTR_REPLACE)
4383 * If some name entry resides on remote MDT, then will create
4384 * agent entry under remote parent. On the other hand, if the
4385 * remote entry will be removed, then related agent entry may
4386 * need to be removed from the remote parent. So there may be
4387 * kinds of cases, let's declare enough credits. The credits
4388 * for create agent entry is enough for remove case.
4390 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4391 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4392 if (dt_object_exists(dt))
4393 credits += 1; /* For updating LMA */
4397 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4402 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4404 * learn xattr size from osd_xattr_get if
4405 * attribute has not been read yet
4407 buflen = __osd_xattr_get(
4408 osd_dt_obj(dt)->oo_inode,
4409 &osd_oti_get(env)->oti_obj_dentry,
4414 buflen = buf->lb_len;
4417 if (buflen > sb->s_blocksize) {
4418 credits += osd_calc_bkmap_credits(
4420 (buflen + sb->s_blocksize - 1) >>
4421 sb->s_blocksize_bits);
4425 * xattr set may involve inode quota change, reserve credits for
4426 * dquot_initialize()
4428 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4431 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4436 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4437 const struct lu_buf *buf, int fl,
4438 struct thandle *handle)
4440 struct osd_thread_info *info = osd_oti_get(env);
4441 struct dentry *dentry = &info->oti_obj_dentry;
4442 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4443 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4444 struct inode *inode = obj->oo_inode;
4445 struct filter_fid *ff = buf->lb_buf;
4446 struct ost_layout *ol = &ff->ff_layout;
4447 int flags = XATTR_REPLACE;
4452 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4455 rc = osd_get_lma(info, inode, dentry, loa);
4456 if (rc == -ENODATA) {
4457 /* Usually for upgarding from old device */
4458 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4459 LMAC_FID_ON_OST, 0);
4460 flags = XATTR_CREATE;
4465 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4466 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4469 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4470 /* Separate PFID EA from LMA */
4471 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4472 lustre_lma_swab(lma);
4473 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4474 sizeof(*lma), XATTR_REPLACE);
4476 obj->oo_pfid_in_lma = 0;
4477 rc = LU_XATTR_CREATE;
4483 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4487 * Old client does not send stripe information,
4488 * then store the PFID EA on disk separatedly.
4490 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4491 ol->ol_stripe_size == 0))
4494 /* Remove old PFID EA entry firstly. */
4495 dquot_initialize(inode);
4496 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4497 if (rc == -ENODATA) {
4498 if ((fl & LU_XATTR_REPLACE) && !(fl & LU_XATTR_CREATE))
4505 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4506 if (likely(ol->ol_stripe_size != 0)) {
4507 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4508 PFID_STRIPE_IDX_BITS;
4509 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4510 lma->lma_compat |= LMAC_STRIPE_INFO;
4511 if (ol->ol_comp_id != 0) {
4512 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4513 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4514 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4515 lma->lma_compat |= LMAC_COMP_INFO;
4519 lustre_loa_swab(loa, false);
4521 /* Store the PFID EA inside LMA. */
4522 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4525 obj->oo_pfid_in_lma = 1;
4531 * In DNE environment, the object (in spite of regular file or directory)
4532 * and its name entry may reside on different MDTs. Under such case, we will
4533 * create an agent entry on the MDT where the object resides. The agent entry
4534 * references the object locally, that makes the object to be visible to the
4535 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4536 * such as 'tar' can handle the object properly.
4538 * We handle the agent entry during set linkEA that is the common interface
4539 * for both regular file and directroy, can handle kinds of cases, such as
4540 * create/link/unlink/rename, and so on.
4542 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4544 * XXX: There are two known issues:
4545 * 1. For one object, we will create at most one agent entry even if there
4546 * may be more than one cross-MDTs hard links on the object. So the local
4547 * e2fsck may claim that the object's nlink is larger than the name entries
4548 * that reference such inode. And in further, the e2fsck will fix the nlink
4549 * attribute to match the local references. Then it will cause the object's
4550 * nlink attribute to be inconsistent with the global references. it is bad
4551 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4552 * other hand, the global namespace LFSCK can repair the object's attribute
4553 * according to the linkEA.
4554 * 2. There may be too many hard links on the object as to its linkEA overflow,
4555 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4556 * case happened, then at this point, we do not know whether there are some
4557 * cross-MDTs reference. But there are local references, it guarantees that
4558 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4560 static int osd_xattr_handle_linkea(const struct lu_env *env,
4561 struct osd_device *osd,
4562 struct osd_object *obj,
4563 const struct lu_buf *buf,
4564 struct thandle *handle)
4566 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4567 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4568 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4569 struct lu_name tmpname;
4570 struct osd_thandle *oh;
4572 bool remote = false;
4576 oh = container_of(handle, struct osd_thandle, ot_super);
4577 LASSERT(oh->ot_handle != NULL);
4579 rc = linkea_init_with_rec(&ldata);
4581 linkea_first_entry(&ldata);
4582 while (ldata.ld_lee != NULL && !remote) {
4583 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4585 if (osd_remote_fid(env, osd, tfid) > 0)
4588 linkea_next_entry(&ldata);
4590 } else if (rc == -ENODATA) {
4596 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4597 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4599 CERROR("%s: failed to remove agent entry for "DFID
4600 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4601 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4602 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4604 CERROR("%s: failed to create agent entry for "DFID
4605 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4612 * Concurrency: @dt is write locked.
4614 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4615 const struct lu_buf *buf, const char *name, int fl,
4616 struct thandle *handle)
4618 struct osd_object *obj = osd_dt_obj(dt);
4619 struct osd_device *osd = osd_obj2dev(obj);
4620 struct inode *inode = obj->oo_inode;
4621 struct osd_thread_info *info = osd_oti_get(env);
4631 /* version set is not real XATTR */
4632 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4633 dt_obj_version_t *version = buf->lb_buf;
4636 * for version we are just using xattr API but change inode
4639 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4641 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4642 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4644 LDISKFS_I(inode)->i_fs_version = *version;
4646 * Version is set after all inode operations are finished,
4647 * so we should mark it dirty here
4649 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4654 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4655 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4658 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4661 * For the OST device with 256 bytes inode size by default,
4662 * the PFID EA will be stored together with LMA EA to avoid
4663 * performance trouble. Otherwise the PFID EA can be stored
4664 * independently. LU-8998
4666 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4667 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4668 LASSERT(buf->lb_buf);
4670 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4673 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4674 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4675 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4677 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4681 lma->lma_incompat |= LMAI_STRIPED;
4682 lustre_lma_swab(lma);
4683 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4684 sizeof(*lma), XATTR_REPLACE);
4687 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4688 LASSERT(!osd->od_is_ost);
4690 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4695 if (fl & LU_XATTR_REPLACE)
4696 fs_flags |= XATTR_REPLACE;
4698 if (fl & LU_XATTR_CREATE)
4699 fs_flags |= XATTR_CREATE;
4701 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4702 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4705 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4706 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4707 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4713 * Concurrency: @dt is read locked.
4715 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4716 const struct lu_buf *buf)
4718 struct osd_object *obj = osd_dt_obj(dt);
4719 struct inode *inode = obj->oo_inode;
4720 struct osd_thread_info *info = osd_oti_get(env);
4721 struct dentry *dentry = &info->oti_obj_dentry;
4723 if (!dt_object_exists(dt))
4726 LASSERT(!dt_object_remote(dt));
4727 LASSERT(inode->i_op != NULL);
4728 LASSERT(inode->i_op->listxattr != NULL);
4730 dentry->d_inode = inode;
4731 dentry->d_sb = inode->i_sb;
4732 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4735 static int osd_declare_xattr_del(const struct lu_env *env,
4736 struct dt_object *dt, const char *name,
4737 struct thandle *handle)
4739 struct osd_thandle *oh;
4740 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4742 LASSERT(!dt_object_remote(dt));
4743 LASSERT(handle != NULL);
4745 oh = container_of(handle, struct osd_thandle, ot_super);
4746 LASSERT(oh->ot_handle == NULL);
4748 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4749 osd_dto_credits_noquota[DTO_XATTR_SET]);
4751 * xattr del may involve inode quota change, reserve credits for
4752 * dquot_initialize()
4754 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4760 * Concurrency: @dt is write locked.
4762 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4763 const char *name, struct thandle *handle)
4765 struct osd_object *obj = osd_dt_obj(dt);
4766 struct inode *inode = obj->oo_inode;
4767 struct osd_thread_info *info = osd_oti_get(env);
4768 struct dentry *dentry = &info->oti_obj_dentry;
4771 if (!dt_object_exists(dt))
4774 LASSERT(!dt_object_remote(dt));
4775 LASSERT(inode->i_op != NULL);
4776 LASSERT(handle != NULL);
4777 #ifdef HAVE_IOP_XATTR
4778 LASSERT(inode->i_op->removexattr != NULL);
4781 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4783 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4784 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4786 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4787 &info->oti_ost_attrs);
4789 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4791 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4792 lustre_lma_swab(lma);
4793 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4794 sizeof(*lma), XATTR_REPLACE);
4796 obj->oo_pfid_in_lma = 0;
4799 dquot_initialize(inode);
4800 dentry->d_inode = inode;
4801 dentry->d_sb = inode->i_sb;
4802 rc = ll_vfs_removexattr(dentry, inode, name);
4805 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4808 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4809 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4810 osd_oxc_del(obj, name);
4815 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4816 __u64 start, __u64 end)
4818 struct osd_object *obj = osd_dt_obj(dt);
4819 struct inode *inode = obj->oo_inode;
4820 struct file *file = osd_quasi_file(env, inode);
4825 rc = vfs_fsync_range(file, start, end, 0);
4830 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4839 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4840 const struct dt_index_features *feat)
4842 struct iam_descr *descr;
4844 if (osd_object_is_root(o))
4845 return feat == &dt_directory_features;
4847 LASSERT(o->oo_dir != NULL);
4849 descr = o->oo_dir->od_container.ic_descr;
4850 if (feat == &dt_directory_features) {
4851 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4856 return feat->dif_keysize_min <= descr->id_key_size &&
4857 descr->id_key_size <= feat->dif_keysize_max &&
4858 feat->dif_recsize_min <= descr->id_rec_size &&
4859 descr->id_rec_size <= feat->dif_recsize_max &&
4860 !(feat->dif_flags & (DT_IND_VARKEY |
4861 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4862 ergo(feat->dif_flags & DT_IND_UPDATE,
4863 1 /* XXX check that object (and fs) is writable */);
4867 static int osd_iam_container_init(const struct lu_env *env,
4868 struct osd_object *obj,
4869 struct osd_directory *dir)
4871 struct iam_container *bag = &dir->od_container;
4874 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4878 result = iam_container_setup(bag);
4880 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4882 iam_container_fini(bag);
4889 * Concurrency: no external locking is necessary.
4891 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4892 const struct dt_index_features *feat)
4896 struct osd_object *obj = osd_dt_obj(dt);
4898 LINVRNT(osd_invariant(obj));
4900 if (osd_object_is_root(obj)) {
4901 dt->do_index_ops = &osd_index_ea_ops;
4903 } else if (feat == &dt_directory_features) {
4904 dt->do_index_ops = &osd_index_ea_ops;
4905 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4910 } else if (unlikely(feat == &dt_otable_features)) {
4911 dt->do_index_ops = &osd_otable_ops;
4913 } else if (unlikely(feat == &dt_acct_features)) {
4914 dt->do_index_ops = &osd_acct_index_ops;
4917 } else if (!osd_has_index(obj)) {
4918 struct osd_directory *dir;
4919 struct osd_device *osd = osd_obj2dev(obj);
4920 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4925 spin_lock(&obj->oo_guard);
4926 if (obj->oo_dir == NULL)
4930 * Concurrent thread allocated container data.
4933 spin_unlock(&obj->oo_guard);
4935 * Now, that we have container data, serialize its
4938 down_write(&obj->oo_ext_idx_sem);
4940 * recheck under lock.
4943 if (osd_has_index(obj)) {
4948 result = osd_iam_container_init(env, obj, obj->oo_dir);
4949 if (result || feat == &dt_lfsck_namespace_features ||
4950 feat == &dt_lfsck_layout_orphan_features ||
4951 feat == &dt_lfsck_layout_dangling_features)
4954 result = osd_index_register(osd, fid,
4955 feat->dif_keysize_max,
4956 feat->dif_recsize_max);
4958 CWARN("%s: failed to register index "
4960 osd_name(osd), PFID(fid), result);
4961 else if (result > 0)
4964 CDEBUG(D_LFSCK, "%s: index object "DFID
4965 " (%d/%d) registered\n",
4966 osd_name(osd), PFID(fid),
4967 (int)feat->dif_keysize_max,
4968 (int)feat->dif_recsize_max);
4971 up_write(&obj->oo_ext_idx_sem);
4979 if (result == 0 && skip_iam == 0) {
4980 if (!osd_iam_index_probe(env, obj, feat))
4983 LINVRNT(osd_invariant(obj));
4988 static int osd_otable_it_attr_get(const struct lu_env *env,
4989 struct dt_object *dt,
4990 struct lu_attr *attr)
4996 static const struct dt_object_operations osd_obj_ops = {
4997 .do_read_lock = osd_read_lock,
4998 .do_write_lock = osd_write_lock,
4999 .do_read_unlock = osd_read_unlock,
5000 .do_write_unlock = osd_write_unlock,
5001 .do_write_locked = osd_write_locked,
5002 .do_attr_get = osd_attr_get,
5003 .do_declare_attr_set = osd_declare_attr_set,
5004 .do_attr_set = osd_attr_set,
5005 .do_ah_init = osd_ah_init,
5006 .do_declare_create = osd_declare_create,
5007 .do_create = osd_create,
5008 .do_declare_destroy = osd_declare_destroy,
5009 .do_destroy = osd_destroy,
5010 .do_index_try = osd_index_try,
5011 .do_declare_ref_add = osd_declare_ref_add,
5012 .do_ref_add = osd_ref_add,
5013 .do_declare_ref_del = osd_declare_ref_del,
5014 .do_ref_del = osd_ref_del,
5015 .do_xattr_get = osd_xattr_get,
5016 .do_declare_xattr_set = osd_declare_xattr_set,
5017 .do_xattr_set = osd_xattr_set,
5018 .do_declare_xattr_del = osd_declare_xattr_del,
5019 .do_xattr_del = osd_xattr_del,
5020 .do_xattr_list = osd_xattr_list,
5021 .do_object_sync = osd_object_sync,
5022 .do_invalidate = osd_invalidate,
5025 static const struct dt_object_operations osd_obj_otable_it_ops = {
5026 .do_attr_get = osd_otable_it_attr_get,
5027 .do_index_try = osd_index_try,
5030 static int osd_index_declare_iam_delete(const struct lu_env *env,
5031 struct dt_object *dt,
5032 const struct dt_key *key,
5033 struct thandle *handle)
5035 struct osd_thandle *oh;
5037 oh = container_of(handle, struct osd_thandle, ot_super);
5038 LASSERT(oh->ot_handle == NULL);
5040 /* Recycle may cause additional three blocks to be changed. */
5041 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5042 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5048 * delete a (key, value) pair from index \a dt specified by \a key
5050 * \param dt osd index object
5051 * \param key key for index
5052 * \param rec record reference
5053 * \param handle transaction handler
5056 * \retval -ve failure
5058 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5059 const struct dt_key *key,
5060 struct thandle *handle)
5062 struct osd_thread_info *oti = osd_oti_get(env);
5063 struct osd_object *obj = osd_dt_obj(dt);
5064 struct osd_thandle *oh;
5065 struct iam_path_descr *ipd;
5066 struct iam_container *bag = &obj->oo_dir->od_container;
5071 if (!dt_object_exists(dt))
5074 LINVRNT(osd_invariant(obj));
5075 LASSERT(!dt_object_remote(dt));
5076 LASSERT(bag->ic_object == obj->oo_inode);
5077 LASSERT(handle != NULL);
5079 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5081 ipd = osd_idx_ipd_get(env, bag);
5082 if (unlikely(ipd == NULL))
5085 oh = container_of(handle, struct osd_thandle, ot_super);
5086 LASSERT(oh->ot_handle != NULL);
5087 LASSERT(oh->ot_handle->h_transaction != NULL);
5089 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5090 /* swab quota uid/gid provided by caller */
5091 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5092 key = (const struct dt_key *)&oti->oti_quota_id;
5095 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5096 osd_ipd_put(env, bag, ipd);
5097 LINVRNT(osd_invariant(obj));
5098 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5102 static int osd_index_declare_ea_delete(const struct lu_env *env,
5103 struct dt_object *dt,
5104 const struct dt_key *key,
5105 struct thandle *handle)
5107 struct osd_thandle *oh;
5108 struct inode *inode;
5113 LASSERT(!dt_object_remote(dt));
5114 LASSERT(handle != NULL);
5116 oh = container_of(handle, struct osd_thandle, ot_super);
5117 LASSERT(oh->ot_handle == NULL);
5119 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5120 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5122 inode = osd_dt_obj(dt)->oo_inode;
5126 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5127 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5132 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5135 struct osd_fid_pack *rec;
5138 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5139 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5140 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5141 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5147 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5148 const struct lu_fid *fid)
5150 struct seq_server_site *ss = osd_seq_site(osd);
5154 /* FID seqs not in FLDB, must be local seq */
5155 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5159 * If FLD is not being initialized yet, it only happens during the
5160 * initialization, likely during mgs initialization, and we assume
5161 * this is local FID.
5163 if (ss == NULL || ss->ss_server_fld == NULL)
5166 /* Only check the local FLDB here */
5167 if (osd_seq_exists(env, osd, fid_seq(fid)))
5173 static void osd_take_care_of_agent(const struct lu_env *env,
5174 struct osd_device *osd,
5175 struct osd_thandle *oh,
5176 struct ldiskfs_dir_entry_2 *de)
5178 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5179 struct osd_idmap_cache *idc;
5180 int rc, schedule = 0;
5182 LASSERT(de != NULL);
5184 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5185 if (likely(rc == 0)) {
5186 idc = osd_idc_find_or_init(env, osd, fid);
5187 if (IS_ERR(idc) || idc->oic_remote)
5189 } else if (rc == -ENODATA) {
5191 * can't get FID, postpone to the end of the
5192 * transaction when iget() is safe
5196 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5199 osd_schedule_agent_inode_removal(env, oh,
5200 le32_to_cpu(de->inode));
5204 * Index delete function for interoperability mode (b11826).
5205 * It will remove the directory entry added by osd_index_ea_insert().
5206 * This entry is needed to maintain name->fid mapping.
5208 * \param key, key i.e. file entry to be deleted
5210 * \retval 0, on success
5211 * \retval -ve, on error
5213 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5214 const struct dt_key *key, struct thandle *handle)
5216 struct osd_object *obj = osd_dt_obj(dt);
5217 struct inode *dir = obj->oo_inode;
5218 struct dentry *dentry;
5219 struct osd_thandle *oh;
5220 struct ldiskfs_dir_entry_2 *de = NULL;
5221 struct buffer_head *bh;
5222 struct htree_lock *hlock = NULL;
5223 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5228 if (!dt_object_exists(dt))
5231 LINVRNT(osd_invariant(obj));
5232 LASSERT(!dt_object_remote(dt));
5233 LASSERT(handle != NULL);
5235 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5237 oh = container_of(handle, struct osd_thandle, ot_super);
5238 LASSERT(oh->ot_handle != NULL);
5239 LASSERT(oh->ot_handle->h_transaction != NULL);
5241 dquot_initialize(dir);
5242 dentry = osd_child_dentry_get(env, obj,
5243 (char *)key, strlen((char *)key));
5245 if (obj->oo_hl_head != NULL) {
5246 hlock = osd_oti_get(env)->oti_hlock;
5247 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5248 dir, LDISKFS_HLOCK_DEL);
5250 down_write(&obj->oo_ext_idx_sem);
5253 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5256 * If this is not the ".." entry, it might be a remote DNE
5257 * entry and we need to check if the FID is for a remote
5258 * MDT. If the FID is not in the directory entry (e.g.
5259 * upgraded 1.8 filesystem without dirdata enabled) then
5260 * we need to get the FID from the LMA. For a remote directory
5261 * there HAS to be an LMA, it cannot be an IGIF inode in this
5264 * Delete the entry before the agent inode in order to
5265 * simplify error handling. At worst an error after deleting
5266 * the entry first might leak the agent inode afterward. The
5267 * reverse would need filesystem abort in case of error deleting
5268 * the entry after the agent had been removed, or leave a
5269 * dangling entry pointing at a random inode.
5271 if (strcmp((char *)key, dotdot) != 0)
5272 osd_take_care_of_agent(env, osd, oh, de);
5273 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5279 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5280 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5281 /* NB, dirent count may not be accurate, because it's counted
5284 if (obj->oo_dirent_count)
5285 obj->oo_dirent_count--;
5287 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5290 ldiskfs_htree_unlock(hlock);
5292 up_write(&obj->oo_ext_idx_sem);
5295 LASSERT(osd_invariant(obj));
5296 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5301 * Lookup index for \a key and copy record to \a rec.
5303 * \param dt osd index object
5304 * \param key key for index
5305 * \param rec record reference
5307 * \retval +ve success : exact mach
5308 * \retval 0 return record with key not greater than \a key
5309 * \retval -ve failure
5311 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5312 struct dt_rec *rec, const struct dt_key *key)
5314 struct osd_object *obj = osd_dt_obj(dt);
5315 struct iam_path_descr *ipd;
5316 struct iam_container *bag = &obj->oo_dir->od_container;
5317 struct osd_thread_info *oti = osd_oti_get(env);
5318 struct iam_iterator *it = &oti->oti_idx_it;
5319 struct iam_rec *iam_rec;
5324 if (!dt_object_exists(dt))
5327 LASSERT(osd_invariant(obj));
5328 LASSERT(!dt_object_remote(dt));
5329 LASSERT(bag->ic_object == obj->oo_inode);
5331 ipd = osd_idx_ipd_get(env, bag);
5335 /* got ipd now we can start iterator. */
5336 iam_it_init(it, bag, 0, ipd);
5338 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5339 /* swab quota uid/gid provided by caller */
5340 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5341 key = (const struct dt_key *)&oti->oti_quota_id;
5344 rc = iam_it_get(it, (struct iam_key *)key);
5346 if (S_ISDIR(obj->oo_inode->i_mode))
5347 iam_rec = (struct iam_rec *)oti->oti_ldp;
5349 iam_rec = (struct iam_rec *)rec;
5351 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5353 if (S_ISDIR(obj->oo_inode->i_mode))
5354 osd_fid_unpack((struct lu_fid *)rec,
5355 (struct osd_fid_pack *)iam_rec);
5356 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5357 osd_quota_unpack(obj, rec);
5362 osd_ipd_put(env, bag, ipd);
5364 LINVRNT(osd_invariant(obj));
5369 static int osd_index_declare_iam_insert(const struct lu_env *env,
5370 struct dt_object *dt,
5371 const struct dt_rec *rec,
5372 const struct dt_key *key,
5373 struct thandle *handle)
5375 struct osd_thandle *oh;
5377 LASSERT(handle != NULL);
5379 oh = container_of(handle, struct osd_thandle, ot_super);
5380 LASSERT(oh->ot_handle == NULL);
5382 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5383 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5389 * Inserts (key, value) pair in \a dt index object.
5391 * \param dt osd index object
5392 * \param key key for index
5393 * \param rec record reference
5394 * \param th transaction handler
5397 * \retval -ve failure
5399 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5400 const struct dt_rec *rec,
5401 const struct dt_key *key, struct thandle *th)
5403 struct osd_object *obj = osd_dt_obj(dt);
5404 struct iam_path_descr *ipd;
5405 struct osd_thandle *oh;
5406 struct iam_container *bag;
5407 struct osd_thread_info *oti = osd_oti_get(env);
5408 struct iam_rec *iam_rec;
5413 if (!dt_object_exists(dt))
5416 LINVRNT(osd_invariant(obj));
5417 LASSERT(!dt_object_remote(dt));
5419 bag = &obj->oo_dir->od_container;
5420 LASSERT(bag->ic_object == obj->oo_inode);
5421 LASSERT(th != NULL);
5423 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5425 ipd = osd_idx_ipd_get(env, bag);
5426 if (unlikely(ipd == NULL))
5429 oh = container_of(th, struct osd_thandle, ot_super);
5430 LASSERT(oh->ot_handle != NULL);
5431 LASSERT(oh->ot_handle->h_transaction != NULL);
5432 if (S_ISDIR(obj->oo_inode->i_mode)) {
5433 iam_rec = (struct iam_rec *)oti->oti_ldp;
5434 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5436 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5437 /* pack quota uid/gid */
5438 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5439 key = (const struct dt_key *)&oti->oti_quota_id;
5440 /* pack quota record */
5441 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5442 iam_rec = (struct iam_rec *)rec;
5444 iam_rec = (struct iam_rec *)rec;
5447 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5449 osd_ipd_put(env, bag, ipd);
5450 LINVRNT(osd_invariant(obj));
5451 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5456 * Calls ldiskfs_add_entry() to add directory entry
5457 * into the directory. This is required for
5458 * interoperability mode (b11826)
5460 * \retval 0, on success
5461 * \retval -ve, on error
5463 static int __osd_ea_add_rec(struct osd_thread_info *info,
5464 struct osd_object *pobj, struct inode *cinode,
5465 const char *name, const struct lu_fid *fid,
5466 struct htree_lock *hlock, struct thandle *th)
5468 struct ldiskfs_dentry_param *ldp;
5469 struct dentry *child;
5470 struct osd_thandle *oth;
5473 oth = container_of(th, struct osd_thandle, ot_super);
5474 LASSERT(oth->ot_handle != NULL);
5475 LASSERT(oth->ot_handle->h_transaction != NULL);
5476 LASSERT(pobj->oo_inode);
5478 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5479 if (unlikely(pobj->oo_inode ==
5480 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
5483 osd_get_ldiskfs_dirent_param(ldp, fid);
5484 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5485 child->d_fsdata = (void *)ldp;
5486 dquot_initialize(pobj->oo_inode);
5487 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5488 child, cinode, hlock);
5489 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5490 struct ldiskfs_dir_entry_2 *de;
5491 struct buffer_head *bh;
5494 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5497 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5500 if (S_ISDIR(cinode->i_mode))
5501 de->file_type = LDISKFS_DIRENT_LUFID |
5502 LDISKFS_FT_REG_FILE;
5504 de->file_type = LDISKFS_DIRENT_LUFID |
5506 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5517 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5518 * into the directory.Also sets flags into osd object to
5519 * indicate dot and dotdot are created. This is required for
5520 * interoperability mode (b11826)
5522 * \param dir directory for dot and dotdot fixup.
5523 * \param obj child object for linking
5525 * \retval 0, on success
5526 * \retval -ve, on error
5528 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5529 struct osd_object *dir,
5530 struct inode *parent_dir, const char *name,
5531 const struct lu_fid *dot_fid,
5532 const struct lu_fid *dot_dot_fid,
5535 struct inode *inode = dir->oo_inode;
5536 struct osd_thandle *oth;
5539 oth = container_of(th, struct osd_thandle, ot_super);
5540 LASSERT(oth->ot_handle->h_transaction != NULL);
5541 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5543 if (strcmp(name, dot) == 0) {
5544 if (dir->oo_compat_dot_created) {
5547 LASSERT(inode->i_ino == parent_dir->i_ino);
5548 dir->oo_compat_dot_created = 1;
5551 } else if (strcmp(name, dotdot) == 0) {
5552 if (!dir->oo_compat_dot_created)
5554 /* in case of rename, dotdot is already created */
5555 if (dir->oo_compat_dotdot_created) {
5556 return __osd_ea_add_rec(info, dir, parent_dir, name,
5557 dot_dot_fid, NULL, th);
5560 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5561 struct lu_fid tfid = *dot_dot_fid;
5564 result = osd_add_dot_dotdot_internal(info,
5565 dir->oo_inode, parent_dir, dot_fid,
5568 result = osd_add_dot_dotdot_internal(info,
5569 dir->oo_inode, parent_dir, dot_fid,
5574 dir->oo_compat_dotdot_created = 1;
5582 * It will call the appropriate osd_add* function and return the
5583 * value, return by respective functions.
5585 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5586 struct inode *cinode, const char *name,
5587 const struct lu_fid *fid, struct thandle *th)
5589 struct osd_thread_info *info = osd_oti_get(env);
5590 struct htree_lock *hlock;
5593 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5595 if (name[0] == '.' && (name[1] == '\0' ||
5596 (name[1] == '.' && name[2] == '\0'))) {
5597 if (hlock != NULL) {
5598 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5601 down_write(&pobj->oo_ext_idx_sem);
5604 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5605 lu_object_fid(&pobj->oo_dt.do_lu),
5608 if (hlock != NULL) {
5609 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5610 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5612 down_write(&pobj->oo_ext_idx_sem);
5615 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5616 struct lu_fid *tfid = &info->oti_fid;
5620 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5623 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5627 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
5628 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
5629 pobj->oo_dirent_count++;
5632 ldiskfs_htree_unlock(hlock);
5634 up_write(&pobj->oo_ext_idx_sem);
5640 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5641 struct osd_idmap_cache *oic)
5643 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5644 struct lu_fid *fid = &oic->oic_fid;
5645 struct osd_inode_id *id = &oic->oic_lid;
5646 struct inode *inode = NULL;
5653 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5656 if (scrub->os_running && scrub->os_pos_current > id->oii_ino)
5659 if (dev->od_auto_scrub_interval == AS_NEVER ||
5660 ktime_get_real_seconds() <
5661 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5665 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5666 if (rc == -ENOENT) {
5667 __u32 gen = id->oii_gen;
5673 inode = osd_iget(oti, dev, id);
5674 /* The inode has been removed (by race maybe). */
5675 if (IS_ERR(inode)) {
5676 rc = PTR_ERR(inode);
5678 RETURN(rc == -ESTALE ? -ENOENT : rc);
5681 /* The OI mapping is lost. */
5682 if (gen != OSD_OII_NOGEN)
5686 * The inode may has been reused by others, we do not know,
5687 * leave it to be handled by subsequent osd_fid_lookup().
5690 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5697 if (scrub->os_running) {
5698 if (inode == NULL) {
5699 inode = osd_iget(oti, dev, id);
5700 /* The inode has been removed (by race maybe). */
5701 if (IS_ERR(inode)) {
5702 rc = PTR_ERR(inode);
5704 RETURN(rc == -ESTALE ? -ENOENT : rc);
5708 rc = osd_oii_insert(dev, oic, insert);
5710 * There is race condition between osd_oi_lookup and OI scrub.
5711 * The OI scrub finished just after osd_oi_lookup() failure.
5712 * Under such case, it is unnecessary to trigger OI scrub again,
5713 * but try to call osd_oi_lookup() again.
5715 if (unlikely(rc == -EAGAIN))
5718 if (!S_ISDIR(inode->i_mode))
5721 rc = osd_check_lmv(oti, dev, inode, oic);
5726 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5727 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5728 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5729 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5730 "%s: trigger partial OI scrub for RPC inconsistency "
5731 "checking FID "DFID": rc = %d\n",
5732 osd_dev2name(dev), PFID(fid), rc);
5733 if (rc == 0 || rc == -EALREADY)
5746 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5747 struct osd_device *dev,
5748 struct osd_idmap_cache *oic,
5749 struct lu_fid *fid, __u32 ino)
5751 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5752 struct inode *inode;
5755 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5756 inode = osd_iget(oti, dev, &oic->oic_lid);
5757 if (IS_ERR(inode)) {
5758 fid_zero(&oic->oic_fid);
5759 return PTR_ERR(inode);
5762 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5765 fid_zero(&oic->oic_fid);
5767 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5771 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5772 struct osd_inode_id *id, const struct lu_fid *fid)
5774 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5775 id->oii_ino, id->oii_gen, info);
5776 info->oti_cache.oic_lid = *id;
5777 info->oti_cache.oic_fid = *fid;
5778 info->oti_cache.oic_dev = osd;
5782 * Get parent FID from the linkEA.
5784 * For a directory which parent resides on remote MDT, to satisfy the
5785 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5786 * the other hand, to make the lookup(..) on the directory can return
5787 * the real parent FID, we append the real parent FID after its ".."
5788 * name entry in the /REMOTE_PARENT_DIR.
5790 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5791 * backup. So after the restore, we cannot get the right parent FID from
5792 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5793 * we have stored the real parent FID in the directory object's linkEA,
5794 * we can parse the linkEA for the real parent FID.
5796 * \param[in] env pointer to the thread context
5797 * \param[in] obj pointer to the object to be handled
5798 * \param[out]fid pointer to the buffer to hold the parent FID
5800 * \retval 0 for getting the real parent FID successfully
5801 * \retval negative error number on failure
5803 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5804 struct osd_object *obj,
5807 struct osd_thread_info *oti = osd_oti_get(env);
5808 struct lu_buf *buf = &oti->oti_big_buf;
5809 struct dentry *dentry = &oti->oti_obj_dentry;
5810 struct inode *inode = obj->oo_inode;
5811 struct linkea_data ldata = { NULL };
5817 if (!S_ISDIR(inode->i_mode))
5821 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5822 buf->lb_buf, buf->lb_len);
5823 if (rc == -ERANGE) {
5824 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5827 lu_buf_realloc(buf, rc);
5828 if (buf->lb_buf == NULL)
5835 if (unlikely(rc == 0))
5841 if (unlikely(buf->lb_buf == NULL)) {
5842 lu_buf_realloc(buf, rc);
5843 if (buf->lb_buf == NULL)
5850 rc = linkea_init_with_rec(&ldata);
5852 linkea_first_entry(&ldata);
5853 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5859 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5860 struct inode *inode,
5861 const struct lu_fid *pfid,
5862 const char *name, const int namelen)
5864 struct osd_thread_info *oti = osd_oti_get(env);
5865 struct lu_buf *buf = &oti->oti_big_buf;
5866 struct dentry *dentry = &oti->oti_obj_dentry;
5867 struct linkea_data ldata = { NULL };
5868 struct lu_name cname = { .ln_name = name,
5869 .ln_namelen = namelen };
5875 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5876 buf->lb_buf, buf->lb_len);
5878 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5883 if (unlikely(rc == 0))
5886 if (buf->lb_len < rc) {
5887 lu_buf_realloc(buf, rc);
5888 if (buf->lb_buf == NULL)
5895 rc = linkea_init_with_rec(&ldata);
5897 rc = linkea_links_find(&ldata, &cname, pfid);
5903 * Calls ->lookup() to find dentry. From dentry get inode and
5904 * read inode's ea to get fid. This is required for interoperability
5907 * \retval 0, on success
5908 * \retval -ve, on error
5910 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5911 struct dt_rec *rec, const struct dt_key *key)
5913 struct inode *dir = obj->oo_inode;
5914 struct dentry *dentry;
5915 struct ldiskfs_dir_entry_2 *de;
5916 struct buffer_head *bh;
5917 struct lu_fid *fid = (struct lu_fid *)rec;
5918 struct htree_lock *hlock = NULL;
5924 LASSERT(dir->i_op != NULL);
5925 LASSERT(dir->i_op->lookup != NULL);
5927 dentry = osd_child_dentry_get(env, obj,
5928 (char *)key, strlen((char *)key));
5930 if (obj->oo_hl_head != NULL) {
5931 hlock = osd_oti_get(env)->oti_hlock;
5932 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5933 dir, LDISKFS_HLOCK_LOOKUP);
5935 down_read(&obj->oo_ext_idx_sem);
5938 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5940 struct osd_thread_info *oti = osd_oti_get(env);
5941 struct osd_inode_id *id = &oti->oti_id;
5942 struct osd_idmap_cache *oic = &oti->oti_cache;
5943 struct osd_device *dev = osd_obj2dev(obj);
5945 ino = le32_to_cpu(de->inode);
5946 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5948 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5952 rc = osd_get_fid_from_dentry(de, rec);
5954 /* done with de, release bh */
5957 if (unlikely(is_remote_parent_ino(dev, ino))) {
5958 const char *name = (const char *)key;
5961 * If the parent is on remote MDT, and there
5962 * is no FID-in-dirent, then we have to get
5963 * the parent FID from the linkEA.
5965 if (likely(strlen(name) == 2 &&
5966 name[0] == '.' && name[1] == '.'))
5967 rc = osd_get_pfid_from_linkea(env, obj,
5970 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5973 osd_id_gen(id, ino, OSD_OII_NOGEN);
5976 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5977 fid_zero(&oic->oic_fid);
5982 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5983 rc = osd_consistency_check(oti, dev, oic);
5985 fid_zero(&oic->oic_fid);
5987 /* Other error should not affect lookup result. */
5997 ldiskfs_htree_unlock(hlock);
5999 up_read(&obj->oo_ext_idx_sem);
6003 static int osd_index_declare_ea_insert(const struct lu_env *env,
6004 struct dt_object *dt,
6005 const struct dt_rec *rec,
6006 const struct dt_key *key,
6007 struct thandle *handle)
6009 struct osd_thandle *oh;
6010 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6011 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6012 const struct lu_fid *fid = rec1->rec_fid;
6013 int credits, rc = 0;
6014 struct osd_idmap_cache *idc;
6018 LASSERT(!dt_object_remote(dt));
6019 LASSERT(handle != NULL);
6020 LASSERT(fid != NULL);
6021 LASSERT(rec1->rec_type != 0);
6023 oh = container_of(handle, struct osd_thandle, ot_super);
6024 LASSERT(oh->ot_handle == NULL);
6026 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6029 * we can't call iget() while a transactions is running
6030 * (this can lead to a deadlock), but we need to know
6031 * inum and object type. so we find this information at
6032 * declaration and cache in per-thread info
6034 idc = osd_idc_find_or_init(env, osd, fid);
6036 RETURN(PTR_ERR(idc));
6037 if (idc->oic_remote) {
6039 * a reference to remote inode is represented by an
6040 * agent inode which we have to create
6042 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6043 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6046 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6048 if (osd_dt_obj(dt)->oo_inode != NULL) {
6049 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6052 * We ignore block quota on meta pool (MDTs), so needn't
6053 * calculate how many blocks will be consumed by this index
6056 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6058 i_projid_read(inode), 0,
6059 oh, osd_dt_obj(dt), NULL,
6064 #ifdef HAVE_PROJECT_QUOTA
6066 * Reserve credits for local agent inode to transfer
6067 * to 0, quota enforcement is ignored in this case.
6069 if (idc->oic_remote &&
6070 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6071 i_projid_read(inode) != 0)
6072 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6073 0, i_projid_read(inode),
6074 0, false, PRJQUOTA, true);
6082 * Index add function for interoperability mode (b11826).
6083 * It will add the directory entry.This entry is needed to
6084 * maintain name->fid mapping.
6086 * \param key it is key i.e. file entry to be inserted
6087 * \param rec it is value of given key i.e. fid
6089 * \retval 0, on success
6090 * \retval -ve, on error
6092 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6093 const struct dt_rec *rec,
6094 const struct dt_key *key, struct thandle *th)
6096 struct osd_object *obj = osd_dt_obj(dt);
6097 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6098 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6099 const struct lu_fid *fid = rec1->rec_fid;
6100 const char *name = (const char *)key;
6101 struct osd_thread_info *oti = osd_oti_get(env);
6102 struct inode *child_inode = NULL;
6103 struct osd_idmap_cache *idc;
6108 if (!dt_object_exists(dt))
6111 LASSERT(osd_invariant(obj));
6112 LASSERT(!dt_object_remote(dt));
6113 LASSERT(th != NULL);
6115 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6117 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6119 idc = osd_idc_find(env, osd, fid);
6120 if (unlikely(idc == NULL)) {
6121 idc = osd_idc_find_or_init(env, osd, fid);
6124 * this dt_insert() wasn't declared properly, so
6125 * FID is missing in OI cache. we better do not
6126 * lookup FID in FLDB/OI and don't risk to deadlock,
6127 * but in some special cases (lfsck testing, etc)
6128 * it's much simpler than fixing a caller.
6130 * normally this error should be placed after the first
6131 * find, but migrate may attach source stripes to
6132 * target, which doesn't create stripes.
6134 CERROR("%s: "DFID" wasn't declared for insert\n",
6135 osd_name(osd), PFID(fid));
6137 RETURN(PTR_ERR(idc));
6141 if (idc->oic_remote) {
6142 /* Insert remote entry */
6143 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6145 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6147 child_inode = osd_create_local_agent_inode(env, osd,
6148 obj, fid, rec1->rec_type & S_IFMT, th);
6149 if (IS_ERR(child_inode))
6150 RETURN(PTR_ERR(child_inode));
6153 /* Insert local entry */
6154 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6155 /* for a reason OI cache wasn't filled properly */
6156 CERROR("%s: OIC for "DFID" isn't filled\n",
6157 osd_name(osd), PFID(fid));
6160 child_inode = oti->oti_inode;
6161 if (unlikely(child_inode == NULL)) {
6162 struct ldiskfs_inode_info *lii;
6167 child_inode = oti->oti_inode = &lii->vfs_inode;
6169 child_inode->i_sb = osd_sb(osd);
6170 child_inode->i_ino = idc->oic_lid.oii_ino;
6171 child_inode->i_mode = rec1->rec_type & S_IFMT;
6174 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6176 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6177 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6179 if (child_inode && child_inode != oti->oti_inode)
6181 LASSERT(osd_invariant(obj));
6182 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6188 * Initialize osd Iterator for given osd index object.
6190 * \param dt osd index object
6193 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6194 struct dt_object *dt,
6197 struct osd_it_iam *it;
6198 struct osd_object *obj = osd_dt_obj(dt);
6199 struct lu_object *lo = &dt->do_lu;
6200 struct iam_path_descr *ipd;
6201 struct iam_container *bag = &obj->oo_dir->od_container;
6203 if (!dt_object_exists(dt))
6204 return ERR_PTR(-ENOENT);
6208 return ERR_PTR(-ENOMEM);
6210 ipd = osd_it_ipd_get(env, bag);
6211 if (likely(ipd != NULL)) {
6215 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6216 return (struct dt_it *)it;
6219 return ERR_PTR(-ENOMEM);
6224 * free given Iterator.
6226 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6228 struct osd_it_iam *it = (struct osd_it_iam *)di;
6229 struct osd_object *obj = it->oi_obj;
6231 iam_it_fini(&it->oi_it);
6232 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6233 osd_object_put(env, obj);
6238 * Move Iterator to record specified by \a key
6240 * \param di osd iterator
6241 * \param key key for index
6243 * \retval +ve di points to record with least key not larger than key
6244 * \retval 0 di points to exact matched key
6245 * \retval -ve failure
6248 static int osd_it_iam_get(const struct lu_env *env,
6249 struct dt_it *di, const struct dt_key *key)
6251 struct osd_thread_info *oti = osd_oti_get(env);
6252 struct osd_it_iam *it = (struct osd_it_iam *)di;
6254 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6255 /* swab quota uid/gid */
6256 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6257 key = (struct dt_key *)&oti->oti_quota_id;
6260 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6266 * \param di osd iterator
6268 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6270 struct osd_it_iam *it = (struct osd_it_iam *)di;
6272 iam_it_put(&it->oi_it);
6276 * Move iterator by one record
6278 * \param di osd iterator
6280 * \retval +1 end of container reached
6282 * \retval -ve failure
6285 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6287 struct osd_it_iam *it = (struct osd_it_iam *)di;
6289 return iam_it_next(&it->oi_it);
6293 * Return pointer to the key under iterator.
6296 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6297 const struct dt_it *di)
6299 struct osd_thread_info *oti = osd_oti_get(env);
6300 struct osd_it_iam *it = (struct osd_it_iam *)di;
6301 struct osd_object *obj = it->oi_obj;
6304 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6306 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6307 /* swab quota uid/gid */
6308 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6309 key = (struct dt_key *)&oti->oti_quota_id;
6316 * Return size of key under iterator (in bytes)
6319 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6321 struct osd_it_iam *it = (struct osd_it_iam *)di;
6323 return iam_it_key_size(&it->oi_it);
6327 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6329 /* check if file type is required */
6330 if (ent->lde_attrs & LUDA_TYPE) {
6331 struct luda_type *lt;
6332 int align = sizeof(*lt) - 1;
6334 len = (len + align) & ~align;
6335 lt = (struct luda_type *)(ent->lde_name + len);
6336 lt->lt_type = cpu_to_le16(DTTOIF(type));
6339 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6343 * build lu direct from backend fs dirent.
6347 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6348 char *name, __u16 namelen, __u16 type, __u32 attr)
6350 ent->lde_attrs = attr | LUDA_FID;
6351 fid_cpu_to_le(&ent->lde_fid, fid);
6353 ent->lde_hash = cpu_to_le64(offset);
6354 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6356 strncpy(ent->lde_name, name, namelen);
6357 ent->lde_name[namelen] = '\0';
6358 ent->lde_namelen = cpu_to_le16(namelen);
6360 /* append lustre attributes */
6361 osd_it_append_attrs(ent, namelen, type);
6365 * Return pointer to the record under iterator.
6367 static int osd_it_iam_rec(const struct lu_env *env,
6368 const struct dt_it *di,
6369 struct dt_rec *dtrec, __u32 attr)
6371 struct osd_it_iam *it = (struct osd_it_iam *)di;
6372 struct osd_thread_info *info = osd_oti_get(env);
6376 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6377 const struct osd_fid_pack *rec;
6378 struct lu_fid *fid = &info->oti_fid;
6379 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6385 name = (char *)iam_it_key_get(&it->oi_it);
6387 RETURN(PTR_ERR(name));
6389 namelen = iam_it_key_size(&it->oi_it);
6391 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6393 RETURN(PTR_ERR(rec));
6395 rc = osd_fid_unpack(fid, rec);
6399 hash = iam_it_store(&it->oi_it);
6401 /* IAM does not store object type in IAM index (dir) */
6402 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6404 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6405 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6406 (struct iam_rec *)dtrec);
6407 osd_quota_unpack(it->oi_obj, dtrec);
6409 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6410 (struct iam_rec *)dtrec);
6417 * Returns cookie for current Iterator position.
6419 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6421 struct osd_it_iam *it = (struct osd_it_iam *)di;
6423 return iam_it_store(&it->oi_it);
6427 * Restore iterator from cookie.
6429 * \param di osd iterator
6430 * \param hash Iterator location cookie
6432 * \retval +ve di points to record with least key not larger than key.
6433 * \retval 0 di points to exact matched key
6434 * \retval -ve failure
6437 static int osd_it_iam_load(const struct lu_env *env,
6438 const struct dt_it *di, __u64 hash)
6440 struct osd_it_iam *it = (struct osd_it_iam *)di;
6442 return iam_it_load(&it->oi_it, hash);
6445 static const struct dt_index_operations osd_index_iam_ops = {
6446 .dio_lookup = osd_index_iam_lookup,
6447 .dio_declare_insert = osd_index_declare_iam_insert,
6448 .dio_insert = osd_index_iam_insert,
6449 .dio_declare_delete = osd_index_declare_iam_delete,
6450 .dio_delete = osd_index_iam_delete,
6452 .init = osd_it_iam_init,
6453 .fini = osd_it_iam_fini,
6454 .get = osd_it_iam_get,
6455 .put = osd_it_iam_put,
6456 .next = osd_it_iam_next,
6457 .key = osd_it_iam_key,
6458 .key_size = osd_it_iam_key_size,
6459 .rec = osd_it_iam_rec,
6460 .store = osd_it_iam_store,
6461 .load = osd_it_iam_load
6467 * Creates or initializes iterator context.
6469 * \retval struct osd_it_ea, iterator structure on success
6472 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6473 struct dt_object *dt,
6476 struct osd_object *obj = osd_dt_obj(dt);
6477 struct osd_thread_info *info = osd_oti_get(env);
6478 struct osd_it_ea *oie;
6480 struct lu_object *lo = &dt->do_lu;
6481 struct dentry *obj_dentry;
6485 if (!dt_object_exists(dt) || obj->oo_destroyed)
6486 RETURN(ERR_PTR(-ENOENT));
6488 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6490 RETURN(ERR_PTR(-ENOMEM));
6491 obj_dentry = &oie->oie_dentry;
6493 obj_dentry->d_inode = obj->oo_inode;
6494 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
6495 obj_dentry->d_name.hash = 0;
6497 oie->oie_rd_dirent = 0;
6498 oie->oie_it_dirent = 0;
6499 oie->oie_dirent = NULL;
6500 if (unlikely(!info->oti_it_ea_buf_used)) {
6501 oie->oie_buf = info->oti_it_ea_buf;
6502 info->oti_it_ea_buf_used = 1;
6504 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6505 if (oie->oie_buf == NULL)
6506 RETURN(ERR_PTR(-ENOMEM));
6510 file = &oie->oie_file;
6512 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6513 if (attr & LUDA_64BITHASH)
6514 file->f_mode = FMODE_64BITHASH;
6516 file->f_mode = FMODE_32BITHASH;
6517 file->f_path.dentry = obj_dentry;
6518 file->f_mapping = obj->oo_inode->i_mapping;
6519 file->f_op = obj->oo_inode->i_fop;
6520 file->f_inode = obj->oo_inode;
6523 RETURN((struct dt_it *)oie);
6527 * Destroy or finishes iterator context.
6529 * \param di iterator structure to be destroyed
6531 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6533 struct osd_thread_info *info = osd_oti_get(env);
6534 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6535 struct osd_object *obj = oie->oie_obj;
6536 struct inode *inode = obj->oo_inode;
6539 oie->oie_file.f_op->release(inode, &oie->oie_file);
6540 osd_object_put(env, obj);
6541 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6542 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6544 info->oti_it_ea_buf_used = 0;
6545 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6550 * It position the iterator at given key, so that next lookup continues from
6551 * that key Or it is similar to dio_it->load() but based on a key,
6552 * rather than file position.
6554 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6557 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6559 static int osd_it_ea_get(const struct lu_env *env,
6560 struct dt_it *di, const struct dt_key *key)
6562 struct osd_it_ea *it = (struct osd_it_ea *)di;
6565 LASSERT(((const char *)key)[0] == '\0');
6566 it->oie_file.f_pos = 0;
6567 it->oie_rd_dirent = 0;
6568 it->oie_it_dirent = 0;
6569 it->oie_dirent = NULL;
6577 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6581 struct osd_filldir_cbs {
6582 struct dir_context ctx;
6583 struct osd_it_ea *it;
6586 * It is called internally by ->iterate*(). It fills the
6587 * iterator's in-memory data structure with required
6588 * information i.e. name, namelen, rec_size etc.
6590 * \param buf in which information to be filled in.
6591 * \param name name of the file in given dir
6593 * \retval 0 on success
6594 * \retval 1 on buffer full
6596 #ifdef HAVE_FILLDIR_USE_CTX
6597 static int osd_ldiskfs_filldir(struct dir_context *buf,
6599 static int osd_ldiskfs_filldir(void *buf,
6601 const char *name, int namelen,
6602 loff_t offset, __u64 ino, unsigned int d_type)
6604 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6605 struct osd_object *obj = it->oie_obj;
6606 struct osd_it_ea_dirent *ent = it->oie_dirent;
6607 struct lu_fid *fid = &ent->oied_fid;
6608 struct osd_fid_pack *rec;
6612 /* this should never happen */
6613 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6614 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6618 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6622 /* "." is just the object itself. */
6623 if (namelen == 1 && name[0] == '.') {
6624 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6625 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6626 rec = (struct osd_fid_pack *)(name + namelen + 1);
6627 if (osd_fid_unpack(fid, rec) != 0)
6632 d_type &= ~LDISKFS_DIRENT_LUFID;
6634 /* NOT export local root. */
6635 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6636 ino = obj->oo_inode->i_ino;
6637 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6640 ent->oied_ino = ino;
6641 ent->oied_off = offset;
6642 ent->oied_namelen = namelen;
6643 ent->oied_type = d_type;
6645 memcpy(ent->oied_name, name, namelen);
6647 it->oie_rd_dirent++;
6648 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6653 * Calls ->iterate*() to load a directory entry at a time
6654 * and stored it in iterator's in-memory data structure.
6656 * \param di iterator's in memory structure
6658 * \retval 0 on success
6659 * \retval -ve on error
6660 * \retval +1 reach the end of entry
6662 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6663 const struct dt_it *di)
6665 struct osd_it_ea *it = (struct osd_it_ea *)di;
6666 struct osd_object *obj = it->oie_obj;
6667 struct inode *inode = obj->oo_inode;
6668 struct htree_lock *hlock = NULL;
6669 struct file *filp = &it->oie_file;
6671 struct osd_filldir_cbs buf = {
6672 .ctx.actor = osd_ldiskfs_filldir,
6677 it->oie_dirent = it->oie_buf;
6678 it->oie_rd_dirent = 0;
6680 if (obj->oo_hl_head != NULL) {
6681 hlock = osd_oti_get(env)->oti_hlock;
6682 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6683 inode, LDISKFS_HLOCK_READDIR);
6685 down_read(&obj->oo_ext_idx_sem);
6688 filp->f_cred = current_cred();
6689 rc = osd_security_file_alloc(filp);
6693 filp->f_flags |= O_NOATIME;
6694 filp->f_mode |= FMODE_NONOTIFY;
6695 rc = iterate_dir(filp, &buf.ctx);
6700 ldiskfs_htree_unlock(hlock);
6702 up_read(&obj->oo_ext_idx_sem);
6704 if (it->oie_rd_dirent == 0) {
6706 * If it does not get any dirent, it means it has been reached
6707 * to the end of the dir
6709 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6713 it->oie_dirent = it->oie_buf;
6714 it->oie_it_dirent = 1;
6721 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
6722 * to load a directory entry at a time and stored it in
6723 * iterator's in-memory data structure.
6725 * \param di iterator's in memory structure
6727 * \retval +ve iterator reached to end
6728 * \retval 0 iterator not reached to end
6729 * \retval -ve on error
6731 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6733 struct osd_it_ea *it = (struct osd_it_ea *)di;
6738 if (it->oie_it_dirent < it->oie_rd_dirent) {
6740 (void *)it->oie_dirent +
6741 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6742 it->oie_dirent->oied_namelen);
6743 it->oie_it_dirent++;
6746 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6749 rc = osd_ldiskfs_it_fill(env, di);
6756 * Returns the key at current position from iterator's in memory structure.
6758 * \param di iterator's in memory structure
6760 * \retval key i.e. struct dt_key on success
6762 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6763 const struct dt_it *di)
6765 struct osd_it_ea *it = (struct osd_it_ea *)di;
6767 return (struct dt_key *)it->oie_dirent->oied_name;
6771 * Returns key's size at current position from iterator's in memory structure.
6773 * \param di iterator's in memory structure
6775 * \retval key_size i.e. struct dt_key on success
6777 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6779 struct osd_it_ea *it = (struct osd_it_ea *)di;
6781 return it->oie_dirent->oied_namelen;
6784 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
6785 #undef LDISKFS_DIR_REC_LEN
6786 #define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
6789 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6791 if (LDISKFS_DIR_REC_LEN(de) >=
6792 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6799 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6800 unsigned int blocksize, bool dotdot)
6803 return osd_dotdot_has_space(de);
6805 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6806 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6813 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6814 handle_t *jh, struct dentry *dentry,
6815 const struct lu_fid *fid, struct buffer_head *bh,
6816 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6819 struct inode *dir = dentry->d_parent->d_inode;
6820 struct inode *inode = dentry->d_inode;
6821 struct osd_fid_pack *rec;
6822 struct ldiskfs_dentry_param *ldp;
6823 int namelen = dentry->d_name.len;
6825 struct osd_thread_info *info = osd_oti_get(env);
6829 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6832 /* There is enough space to hold the FID-in-dirent. */
6833 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6834 rc = ldiskfs_journal_get_write_access(jh, bh);
6838 de->name[namelen] = 0;
6839 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6840 rec->fp_len = sizeof(struct lu_fid) + 1;
6841 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6842 de->file_type |= LDISKFS_DIRENT_LUFID;
6843 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6850 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6854 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6855 osd_get_ldiskfs_dirent_param(ldp, fid);
6856 dentry->d_fsdata = (void *)ldp;
6857 dquot_initialize(dir);
6858 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6860 * It is too bad, we cannot reinsert the name entry back.
6861 * That means we lose it!
6864 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6865 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6866 osd_ino2name(inode),
6867 dir->i_ino, dir->i_generation, namelen,
6868 dentry->d_name.name, PFID(fid), rc);
6874 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6875 struct osd_it_ea *it, struct lu_fid *fid,
6876 struct osd_inode_id *id, __u32 *attr)
6878 struct osd_thread_info *info = osd_oti_get(env);
6879 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6880 struct osd_device *dev = osd_obj2dev(obj);
6881 struct super_block *sb = osd_sb(dev);
6882 const char *devname = osd_name(dev);
6883 struct osd_it_ea_dirent *ent = it->oie_dirent;
6884 struct inode *dir = obj->oo_inode;
6885 struct htree_lock *hlock = NULL;
6886 struct buffer_head *bh = NULL;
6887 handle_t *jh = NULL;
6888 struct ldiskfs_dir_entry_2 *de;
6889 struct dentry *dentry;
6890 struct inode *inode;
6891 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6894 bool dotdot = false;
6899 if (ent->oied_name[0] == '.') {
6900 if (ent->oied_namelen == 1)
6903 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6907 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6908 inode = osd_iget(info, dev, id);
6909 if (IS_ERR(inode)) {
6910 rc = PTR_ERR(inode);
6911 if (rc == -ENOENT || rc == -ESTALE) {
6913 * Maybe dangling name entry, or
6914 * corrupted directory entry.
6916 *attr |= LUDA_UNKNOWN;
6919 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6920 "check_repair, dir = %lu/%u, name = %.*s, "
6921 "ino = %llu, rc = %d\n",
6922 devname, dir->i_ino, dir->i_generation,
6923 ent->oied_namelen, ent->oied_name,
6930 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6932 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6933 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6939 * We need to ensure that the name entry is still valid.
6940 * Because it may be removed or renamed by other already.
6942 * The unlink or rename operation will start journal before PDO lock,
6943 * so to avoid deadlock, here we need to start journal handle before
6944 * related PDO lock also. But because we do not know whether there
6945 * will be something to be repaired before PDO lock, we just start
6946 * journal without conditions.
6948 * We may need to remove the name entry firstly, then insert back.
6949 * One credit is for user quota file update.
6950 * One credit is for group quota file update.
6951 * Two credits are for dirty inode.
6953 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6954 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6956 if (dev->od_dirent_journal != 0) {
6959 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6962 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6963 "check_repair, dir = %lu/%u, credits = %d, "
6964 "name = %.*s, ino = %llu: rc = %d\n",
6965 devname, dir->i_ino, dir->i_generation, credits,
6966 ent->oied_namelen, ent->oied_name,
6969 GOTO(out_inode, rc);
6972 if (obj->oo_hl_head != NULL) {
6973 hlock = osd_oti_get(env)->oti_hlock;
6975 * "0" means exclusive lock for the whole directory.
6976 * We need to prevent others access such name entry
6977 * during the delete + insert. Neither HLOCK_ADD nor
6978 * HLOCK_DEL cannot guarantee the atomicity.
6980 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6982 down_write(&obj->oo_ext_idx_sem);
6985 if (obj->oo_hl_head != NULL) {
6986 hlock = osd_oti_get(env)->oti_hlock;
6987 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6988 LDISKFS_HLOCK_LOOKUP);
6990 down_read(&obj->oo_ext_idx_sem);
6994 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6995 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6996 *attr |= LUDA_IGNORE;
7002 * For dotdot entry, if there is not enough space to hold the
7003 * FID-in-dirent, just keep them there. It only happens when the
7004 * device upgraded from 1.8 or restored from MDT file-level backup.
7005 * For the whole directory, only dotdot entry have no FID-in-dirent
7006 * and needs to get FID from LMA when readdir, it will not affect the
7009 if (dotdot && !osd_dotdot_has_space(de)) {
7010 *attr |= LUDA_UNKNOWN;
7016 if (lu_fid_eq(fid, &lma->lma_self_fid))
7019 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7020 struct lu_fid *tfid = &lma->lma_self_fid;
7022 if (likely(dotdot &&
7023 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7024 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7026 * It must be REMOTE_PARENT_DIR and as the
7027 * 'dotdot' entry of remote directory
7029 *attr |= LUDA_IGNORE;
7031 CDEBUG(D_LFSCK, "%s: expect remote agent "
7032 "parent directory, but got %.*s under "
7033 "dir = %lu/%u with the FID "DFID"\n",
7034 devname, ent->oied_namelen,
7035 ent->oied_name, dir->i_ino,
7036 dir->i_generation, PFID(tfid));
7038 *attr |= LUDA_UNKNOWN;
7045 if (!fid_is_zero(fid)) {
7046 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7048 if (rc == -ENOENT ||
7050 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7052 * linkEA does not recognize the dirent entry,
7053 * it may because the dirent entry corruption
7054 * and points to other's inode.
7056 CDEBUG(D_LFSCK, "%s: the target inode does not "
7057 "recognize the dirent, dir = %lu/%u, "
7058 " name = %.*s, ino = %llu, "
7059 DFID": rc = %d\n", devname, dir->i_ino,
7060 dir->i_generation, ent->oied_namelen,
7061 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7062 *attr |= LUDA_UNKNOWN;
7067 if (rc && rc != -ENODATA) {
7068 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7069 "dir = %lu/%u, name = %.*s, ino = %llu, "
7070 DFID": rc = %d\n", devname, dir->i_ino,
7071 dir->i_generation, ent->oied_namelen,
7072 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7073 *attr |= LUDA_UNKNOWN;
7081 * linkEA recognizes the dirent entry, the FID-in-LMA is
7082 * valid, trusted, in spite of fid_is_sane(fid) or not.
7084 if (*attr & LUDA_VERIFY_DRYRUN) {
7085 *fid = lma->lma_self_fid;
7086 *attr |= LUDA_REPAIR;
7093 dev->od_dirent_journal = 1;
7094 if (hlock != NULL) {
7095 ldiskfs_htree_unlock(hlock);
7098 up_read(&obj->oo_ext_idx_sem);
7104 *fid = lma->lma_self_fid;
7106 /* Update or append the FID-in-dirent. */
7107 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7108 bh, de, hlock, dotdot);
7110 *attr |= LUDA_REPAIR;
7112 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7113 "the dirent, dir = %lu/%u, name = %.*s, "
7114 "ino = %llu, "DFID": rc = %d\n",
7115 devname, dir->i_ino, dir->i_generation,
7116 ent->oied_namelen, ent->oied_name,
7117 ent->oied_ino, PFID(fid), rc);
7119 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7120 if (*attr & LUDA_VERIFY_DRYRUN) {
7121 if (fid_is_sane(fid)) {
7122 *attr |= LUDA_REPAIR;
7123 } else if (dev->od_index == 0) {
7124 lu_igif_build(fid, inode->i_ino,
7125 inode->i_generation);
7126 *attr |= LUDA_UPGRADE;
7134 dev->od_dirent_journal = 1;
7135 if (hlock != NULL) {
7136 ldiskfs_htree_unlock(hlock);
7139 up_read(&obj->oo_ext_idx_sem);
7146 if (unlikely(fid_is_sane(fid))) {
7148 * FID-in-dirent exists, but FID-in-LMA is lost.
7149 * Trust the FID-in-dirent, and add FID-in-LMA.
7151 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7153 *attr |= LUDA_REPAIR;
7155 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7156 "update dirent, dir = %lu/%u, "
7157 "name = %.*s, ino = %llu, "
7159 devname, dir->i_ino, dir->i_generation,
7160 ent->oied_namelen, ent->oied_name,
7161 ent->oied_ino, PFID(fid), rc);
7162 } else if (dev->od_index == 0) {
7163 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7165 * It is probably IGIF object. Only aappend the
7166 * FID-in-dirent. OI scrub will process FID-in-LMA.
7168 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7169 bh, de, hlock, dotdot);
7171 *attr |= LUDA_UPGRADE;
7173 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7174 "after the dirent, dir = %lu/%u, "
7175 "name = %.*s, ino = %llu, "
7177 devname, dir->i_ino, dir->i_generation,
7178 ent->oied_namelen, ent->oied_name,
7179 ent->oied_ino, PFID(fid), rc);
7188 if (hlock != NULL) {
7189 ldiskfs_htree_unlock(hlock);
7191 if (dev->od_dirent_journal != 0)
7192 up_write(&obj->oo_ext_idx_sem);
7194 up_read(&obj->oo_ext_idx_sem);
7198 ldiskfs_journal_stop(jh);
7202 if (rc >= 0 && !dirty)
7203 dev->od_dirent_journal = 0;
7209 * Returns the value at current position from iterator's in memory structure.
7211 * \param di struct osd_it_ea, iterator's in memory structure
7212 * \param attr attr requested for dirent.
7213 * \param lde lustre dirent
7215 * \retval 0 no error and \param lde has correct lustre dirent.
7216 * \retval -ve on error
7218 static inline int osd_it_ea_rec(const struct lu_env *env,
7219 const struct dt_it *di,
7220 struct dt_rec *dtrec, __u32 attr)
7222 struct osd_it_ea *it = (struct osd_it_ea *)di;
7223 struct osd_object *obj = it->oie_obj;
7224 struct osd_device *dev = osd_obj2dev(obj);
7225 struct osd_thread_info *oti = osd_oti_get(env);
7226 struct osd_inode_id *id = &oti->oti_id;
7227 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7228 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7229 __u32 ino = it->oie_dirent->oied_ino;
7234 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7236 if (attr & LUDA_VERIFY) {
7237 if (unlikely(is_remote_parent_ino(dev, ino))) {
7238 attr |= LUDA_IGNORE;
7240 * If the parent is on remote MDT, and there
7241 * is no FID-in-dirent, then we have to get
7242 * the parent FID from the linkEA.
7244 if (!fid_is_sane(fid) &&
7245 it->oie_dirent->oied_namelen == 2 &&
7246 it->oie_dirent->oied_name[0] == '.' &&
7247 it->oie_dirent->oied_name[1] == '.')
7248 osd_get_pfid_from_linkea(env, obj, fid);
7250 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7254 if (!fid_is_sane(fid))
7255 attr |= LUDA_UNKNOWN;
7257 attr &= ~LU_DIRENT_ATTRS_MASK;
7258 if (!fid_is_sane(fid)) {
7259 bool is_dotdot = false;
7261 if (it->oie_dirent->oied_namelen == 2 &&
7262 it->oie_dirent->oied_name[0] == '.' &&
7263 it->oie_dirent->oied_name[1] == '.')
7266 * If the parent is on remote MDT, and there
7267 * is no FID-in-dirent, then we have to get
7268 * the parent FID from the linkEA.
7270 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7271 rc = osd_get_pfid_from_linkea(env, obj, fid);
7273 if (is_dotdot == false &&
7274 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7277 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7282 /* Pack the entry anyway, at least the offset is right. */
7283 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7284 it->oie_dirent->oied_name,
7285 it->oie_dirent->oied_namelen,
7286 it->oie_dirent->oied_type, attr);
7288 RETURN(rc > 0 ? 0 : rc);
7292 * Returns the record size size at current position.
7294 * This function will return record(lu_dirent) size in bytes.
7296 * \param[in] env execution environment
7297 * \param[in] di iterator's in memory structure
7298 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7299 * calculate the lu_dirent size.
7301 * \retval record size(in bytes & in memory) of the current lu_dirent
7304 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7307 struct osd_it_ea *it = (struct osd_it_ea *)di;
7309 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7313 * Returns a cookie for current position of the iterator head, so that
7314 * user can use this cookie to load/start the iterator next time.
7316 * \param di iterator's in memory structure
7318 * \retval cookie for current position, on success
7320 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7322 struct osd_it_ea *it = (struct osd_it_ea *)di;
7324 return it->oie_dirent->oied_off;
7328 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7329 * to load a directory entry at a time and stored it i inn,
7330 * in iterator's in-memory data structure.
7332 * \param di struct osd_it_ea, iterator's in memory structure
7334 * \retval +ve on success
7335 * \retval -ve on error
7337 static int osd_it_ea_load(const struct lu_env *env,
7338 const struct dt_it *di, __u64 hash)
7340 struct osd_it_ea *it = (struct osd_it_ea *)di;
7344 it->oie_file.f_pos = hash;
7346 rc = osd_ldiskfs_it_fill(env, di);
7357 * Index lookup function for interoperability mode (b11826).
7359 * \param key, key i.e. file name to be searched
7361 * \retval +ve, on success
7362 * \retval -ve, on error
7364 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7365 struct dt_rec *rec, const struct dt_key *key)
7367 struct osd_object *obj = osd_dt_obj(dt);
7372 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7373 LINVRNT(osd_invariant(obj));
7375 rc = osd_ea_lookup_rec(env, obj, rec, key);
7382 * Index and Iterator operations for interoperability
7383 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7385 static const struct dt_index_operations osd_index_ea_ops = {
7386 .dio_lookup = osd_index_ea_lookup,
7387 .dio_declare_insert = osd_index_declare_ea_insert,
7388 .dio_insert = osd_index_ea_insert,
7389 .dio_declare_delete = osd_index_declare_ea_delete,
7390 .dio_delete = osd_index_ea_delete,
7392 .init = osd_it_ea_init,
7393 .fini = osd_it_ea_fini,
7394 .get = osd_it_ea_get,
7395 .put = osd_it_ea_put,
7396 .next = osd_it_ea_next,
7397 .key = osd_it_ea_key,
7398 .key_size = osd_it_ea_key_size,
7399 .rec = osd_it_ea_rec,
7400 .rec_size = osd_it_ea_rec_size,
7401 .store = osd_it_ea_store,
7402 .load = osd_it_ea_load
7406 static void *osd_key_init(const struct lu_context *ctx,
7407 struct lu_context_key *key)
7409 struct osd_thread_info *info;
7411 OBD_ALLOC_PTR(info);
7413 return ERR_PTR(-ENOMEM);
7415 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7416 if (info->oti_it_ea_buf == NULL)
7419 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7421 info->oti_hlock = ldiskfs_htree_lock_alloc();
7422 if (info->oti_hlock == NULL)
7428 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7431 return ERR_PTR(-ENOMEM);
7434 static void osd_key_fini(const struct lu_context *ctx,
7435 struct lu_context_key *key, void *data)
7437 struct osd_thread_info *info = data;
7438 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7439 struct osd_idmap_cache *idc = info->oti_ins_cache;
7441 if (info->oti_dio_pages) {
7443 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7444 struct page *page = info->oti_dio_pages[i];
7446 LASSERT(PagePrivate2(page));
7447 LASSERT(PageLocked(page));
7448 ClearPagePrivate2(page);
7453 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7454 PTLRPC_MAX_BRW_PAGES);
7457 if (info->oti_inode != NULL)
7459 if (info->oti_hlock != NULL)
7460 ldiskfs_htree_lock_free(info->oti_hlock);
7461 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7462 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7463 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7464 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7465 lu_buf_free(&info->oti_big_buf);
7467 LASSERT(info->oti_ins_cache_size > 0);
7468 OBD_FREE_PTR_ARRAY(idc, info->oti_ins_cache_size);
7469 info->oti_ins_cache = NULL;
7470 info->oti_ins_cache_size = 0;
7475 static void osd_key_exit(const struct lu_context *ctx,
7476 struct lu_context_key *key, void *data)
7478 struct osd_thread_info *info = data;
7480 LASSERT(info->oti_r_locks == 0);
7481 LASSERT(info->oti_w_locks == 0);
7482 LASSERT(info->oti_txns == 0);
7485 /* type constructor/destructor: osd_type_init, osd_type_fini */
7486 LU_TYPE_INIT_FINI(osd, &osd_key);
7488 struct lu_context_key osd_key = {
7489 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7490 .lct_init = osd_key_init,
7491 .lct_fini = osd_key_fini,
7492 .lct_exit = osd_key_exit
7496 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7497 const char *name, struct lu_device *next)
7499 struct osd_device *osd = osd_dev(d);
7501 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7502 sizeof(osd->od_svname))
7504 return osd_procfs_init(osd, name);
7507 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7509 struct seq_server_site *ss = osd_seq_site(osd);
7514 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7517 if (unlikely(ss == NULL))
7520 OBD_ALLOC_PTR(osd->od_cl_seq);
7521 if (osd->od_cl_seq == NULL)
7524 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7525 osd->od_svname, ss->ss_server_seq);
7527 if (ss->ss_node_id == 0) {
7529 * If the OSD on the sequence controller(MDT0), then allocate
7530 * sequence here, otherwise allocate sequence after connected
7531 * to MDT0 (see mdt_register_lwp_callback()).
7533 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7534 &osd->od_cl_seq->lcs_space, env);
7540 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7542 if (osd->od_cl_seq == NULL)
7545 seq_client_fini(osd->od_cl_seq);
7546 OBD_FREE_PTR(osd->od_cl_seq);
7547 osd->od_cl_seq = NULL;
7550 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7554 /* shutdown quota slave instance associated with the device */
7555 if (o->od_quota_slave_md != NULL) {
7556 struct qsd_instance *qsd = o->od_quota_slave_md;
7558 o->od_quota_slave_md = NULL;
7562 if (o->od_quota_slave_dt != NULL) {
7563 struct qsd_instance *qsd = o->od_quota_slave_dt;
7565 o->od_quota_slave_dt = NULL;
7569 osd_fid_fini(env, o);
7570 osd_scrub_cleanup(env, o);
7575 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7579 if (o->od_mnt != NULL) {
7580 shrink_dcache_sb(osd_sb(o));
7581 osd_sync(env, &o->od_dt_dev);
7590 static int osd_mount(const struct lu_env *env,
7591 struct osd_device *o, struct lustre_cfg *cfg)
7593 const char *name = lustre_cfg_string(cfg, 0);
7594 const char *dev = lustre_cfg_string(cfg, 1);
7596 unsigned long page, s_flags, lmd_flags = 0;
7597 struct page *__page;
7598 struct file_system_type *type;
7599 char *options = NULL;
7601 struct osd_thread_info *info = osd_oti_get(env);
7602 struct lu_fid *fid = &info->oti_fid;
7603 struct inode *inode;
7604 int rc = 0, force_over_1024tb = 0;
7608 if (o->od_mnt != NULL)
7611 if (strlen(dev) >= sizeof(o->od_mntdev))
7613 strcpy(o->od_mntdev, dev);
7615 str = lustre_cfg_string(cfg, 2);
7616 s_flags = simple_strtoul(str, NULL, 0);
7617 str = strstr(str, ":");
7619 lmd_flags = simple_strtoul(str + 1, NULL, 0);
7620 opts = lustre_cfg_string(cfg, 3);
7622 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7623 CERROR("%s: device %s extents feature is not guaranteed to "
7624 "work on big-endian systems. Use \"bigendian_extents\" "
7625 "mount option to override.\n", name, dev);
7629 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7630 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7631 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");
7634 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7635 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7636 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");
7639 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7640 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
7641 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");
7645 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
7646 force_over_1024tb = 1;
7648 __page = alloc_page(GFP_KERNEL);
7650 GOTO(out, rc = -ENOMEM);
7651 page = (unsigned long)page_address(__page);
7652 options = (char *)page;
7655 /* strip out the options for back compatiblity */
7656 static const char * const sout[] = {
7663 /* strip out option we processed in osd */
7664 "bigendian_extents",
7668 "force_over_1024tb",
7671 strncat(options, opts, PAGE_SIZE);
7672 for (rc = 0, str = options; sout[rc]; ) {
7673 char *op = strstr(str, sout[rc]);
7680 if (op == options || *(op - 1) == ',') {
7681 str = op + strlen(sout[rc]);
7682 if (*str == ',' || *str == '\0') {
7683 *str == ',' ? str++ : str;
7684 memmove(op, str, strlen(str) + 1);
7687 for (str = op; *str != ',' && *str != '\0'; str++)
7691 strncat(options, "user_xattr,acl", PAGE_SIZE);
7694 /* Glom up mount options */
7695 if (*options != '\0')
7696 strncat(options, ",", PAGE_SIZE);
7697 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7699 type = get_fs_type("ldiskfs");
7701 CERROR("%s: cannot find ldiskfs module\n", name);
7702 GOTO(out, rc = -ENODEV);
7705 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7706 module_put(type->owner);
7708 if (IS_ERR(o->od_mnt)) {
7709 rc = PTR_ERR(o->od_mnt);
7711 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7715 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7716 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
7717 force_over_1024tb == 0) {
7718 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",
7720 GOTO(out_mnt, rc = -EINVAL);
7723 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7724 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7727 GOTO(out_mnt, rc = -EOPNOTSUPP);
7730 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7731 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7732 GOTO(out_mnt, rc = -EINVAL);
7735 #ifdef LDISKFS_MOUNT_DIRDATA
7736 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7737 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7738 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7739 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7740 "enabling the dirdata feature. If you do not want to "
7741 "downgrade to Lustre-1.x again, you can enable it via "
7742 "'tune2fs -O dirdata device'\n", name, dev);
7744 inode = osd_sb(o)->s_root->d_inode;
7745 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7746 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7748 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7752 if (lmd_flags & LMD_FLG_NOSCRUB)
7753 o->od_auto_scrub_interval = AS_NEVER;
7755 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7756 /* do not use pagecache with flash-backed storage */
7757 o->od_writethrough_cache = 0;
7758 o->od_read_cache = 0;
7769 __free_page(__page);
7774 static struct lu_device *osd_device_fini(const struct lu_env *env,
7775 struct lu_device *d)
7777 struct osd_device *o = osd_dev(d);
7781 osd_index_backup(env, o, false);
7782 osd_shutdown(env, o);
7784 if (o->od_oi_table != NULL)
7785 osd_oi_fini(osd_oti_get(env), o);
7786 osd_obj_map_fini(o);
7792 static int osd_device_init0(const struct lu_env *env,
7793 struct osd_device *o,
7794 struct lustre_cfg *cfg)
7796 struct lu_device *l = osd2lu_dev(o);
7797 struct osd_thread_info *info;
7801 /* if the module was re-loaded, env can loose its keys */
7802 rc = lu_env_refill((struct lu_env *)env);
7805 info = osd_oti_get(env);
7808 l->ld_ops = &osd_lu_ops;
7809 o->od_dt_dev.dd_ops = &osd_dt_ops;
7811 spin_lock_init(&o->od_osfs_lock);
7812 mutex_init(&o->od_otable_mutex);
7813 INIT_LIST_HEAD(&o->od_orphan_list);
7814 INIT_LIST_HEAD(&o->od_index_backup_list);
7815 INIT_LIST_HEAD(&o->od_index_restore_list);
7816 spin_lock_init(&o->od_lock);
7817 o->od_index_backup_policy = LIBP_NONE;
7820 o->od_read_cache = 1;
7821 o->od_writethrough_cache = 1;
7822 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7823 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
7824 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
7825 o->od_auto_scrub_interval = AS_DEFAULT;
7827 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7828 sizeof(o->od_svname));
7829 if (cplen >= sizeof(o->od_svname)) {
7834 o->od_index_backup_stop = 0;
7835 o->od_index = -1; /* -1 means index is invalid */
7836 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7837 if (rc == LDD_F_SV_TYPE_OST)
7840 o->od_full_scrub_ratio = OFSR_DEFAULT;
7841 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7842 rc = osd_mount(env, o, cfg);
7846 /* Can only check block device after mount */
7847 o->od_nonrotational =
7848 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
7850 rc = osd_obj_map_init(env, o);
7854 rc = lu_site_init(&o->od_site, l);
7856 GOTO(out_compat, rc);
7857 o->od_site.ls_bottom_dev = l;
7859 rc = lu_site_init_finish(&o->od_site);
7863 INIT_LIST_HEAD(&o->od_ios_list);
7864 /* setup scrub, including OI files initialization */
7866 rc = osd_scrub_setup(env, o);
7871 rc = osd_procfs_init(o, o->od_svname);
7873 CERROR("%s: can't initialize procfs: rc = %d\n",
7875 GOTO(out_scrub, rc);
7878 LASSERT(l->ld_site->ls_linkage.next != NULL);
7879 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7881 /* initialize quota slave instance */
7882 /* currently it's no need to prepare qsd_instance_md for OST */
7883 if (!o->od_is_ost) {
7884 o->od_quota_slave_md = qsd_init(env, o->od_svname,
7886 o->od_proc_entry, true);
7887 if (IS_ERR(o->od_quota_slave_md)) {
7888 rc = PTR_ERR(o->od_quota_slave_md);
7889 o->od_quota_slave_md = NULL;
7890 GOTO(out_procfs, rc);
7894 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
7895 o->od_proc_entry, false);
7897 if (IS_ERR(o->od_quota_slave_dt)) {
7898 if (o->od_quota_slave_md != NULL) {
7899 qsd_fini(env, o->od_quota_slave_md);
7900 o->od_quota_slave_md = NULL;
7903 rc = PTR_ERR(o->od_quota_slave_dt);
7904 o->od_quota_slave_dt = NULL;
7905 GOTO(out_procfs, rc);
7913 osd_scrub_cleanup(env, o);
7915 lu_site_fini(&o->od_site);
7917 osd_obj_map_fini(o);
7924 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7925 struct lu_device_type *t,
7926 struct lustre_cfg *cfg)
7928 struct osd_device *o;
7933 return ERR_PTR(-ENOMEM);
7935 rc = dt_device_init(&o->od_dt_dev, t);
7938 * Because the ctx might be revived in dt_device_init,
7939 * refill the env here
7941 lu_env_refill((struct lu_env *)env);
7942 rc = osd_device_init0(env, o, cfg);
7944 dt_device_fini(&o->od_dt_dev);
7947 if (unlikely(rc != 0))
7950 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7953 static struct lu_device *osd_device_free(const struct lu_env *env,
7954 struct lu_device *d)
7956 struct osd_device *o = osd_dev(d);
7960 /* XXX: make osd top device in order to release reference */
7961 d->ld_site->ls_top_dev = d;
7962 lu_site_purge(env, d->ld_site, -1);
7963 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
7964 D_ERROR, lu_cdebug_printer);
7965 lu_site_fini(&o->od_site);
7966 dt_device_fini(&o->od_dt_dev);
7971 static int osd_process_config(const struct lu_env *env,
7972 struct lu_device *d, struct lustre_cfg *cfg)
7974 struct osd_device *o = osd_dev(d);
7980 switch (cfg->lcfg_command) {
7982 rc = osd_mount(env, o, cfg);
7986 * For the case LCFG_PRE_CLEANUP is not called in advance,
7987 * that may happend if hit failure during mount process.
7989 osd_index_backup(env, o, false);
7990 lu_dev_del_linkage(d->ld_site, d);
7991 rc = osd_shutdown(env, o);
7994 LASSERT(&o->od_dt_dev);
7995 count = class_modify_config(cfg, PARAM_OSD,
7996 &o->od_dt_dev.dd_kobj);
7998 count = class_modify_config(cfg, PARAM_OST,
7999 &o->od_dt_dev.dd_kobj);
8000 rc = count > 0 ? 0 : count;
8002 case LCFG_PRE_CLEANUP:
8004 osd_index_backup(env, o,
8005 o->od_index_backup_policy != LIBP_NONE);
8015 static int osd_recovery_complete(const struct lu_env *env,
8016 struct lu_device *d)
8018 struct osd_device *osd = osd_dev(d);
8023 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8027 * start qsd instance on recovery completion, this notifies the quota
8028 * slave code that we are about to process new requests now
8030 rc = qsd_start(env, osd->od_quota_slave_dt);
8031 if (rc == 0 && osd->od_quota_slave_md != NULL)
8032 rc = qsd_start(env, osd->od_quota_slave_md);
8038 * we use exports to track all osd users
8040 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8041 struct obd_device *obd, struct obd_uuid *cluuid,
8042 struct obd_connect_data *data, void *localdata)
8044 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8045 struct lustre_handle conn;
8050 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8052 rc = class_connect(&conn, obd, cluuid);
8056 *exp = class_conn2export(&conn);
8058 spin_lock(&osd->od_osfs_lock);
8060 spin_unlock(&osd->od_osfs_lock);
8066 * once last export (we don't count self-export) disappeared
8067 * osd can be released
8069 static int osd_obd_disconnect(struct obd_export *exp)
8071 struct obd_device *obd = exp->exp_obd;
8072 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8073 int rc, release = 0;
8077 /* Only disconnect the underlying layers on the final disconnect. */
8078 spin_lock(&osd->od_osfs_lock);
8080 if (osd->od_connects == 0)
8082 spin_unlock(&osd->od_osfs_lock);
8084 rc = class_disconnect(exp); /* bz 9811 */
8086 if (rc == 0 && release)
8087 class_manual_cleanup(obd);
8091 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8092 struct lu_device *dev)
8094 struct osd_device *osd = osd_dev(dev);
8095 struct lr_server_data *lsd =
8096 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8101 if (osd->od_quota_slave_md != NULL) {
8102 /* set up quota slave objects for inode */
8103 result = qsd_prepare(env, osd->od_quota_slave_md);
8108 if (osd->od_quota_slave_dt != NULL) {
8109 /* set up quota slave objects for block */
8110 result = qsd_prepare(env, osd->od_quota_slave_dt);
8116 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8117 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8118 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8119 osd->od_index_in_idif = 1;
8121 osd->od_index_in_idif = 0;
8122 result = osd_register_proc_index_in_idif(osd);
8127 osd->od_index_in_idif = 1;
8131 result = osd_fid_init(env, osd);
8137 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8141 * see include/lu_object.h for the details.
8143 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8144 struct lu_fid *fid, struct lu_object *parent,
8145 const struct lu_name *name)
8147 struct osd_device *osd = osd_dev(d);
8149 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8152 static const struct lu_object_operations osd_lu_obj_ops = {
8153 .loo_object_init = osd_object_init,
8154 .loo_object_delete = osd_object_delete,
8155 .loo_object_release = osd_object_release,
8156 .loo_object_free = osd_object_free,
8157 .loo_object_print = osd_object_print,
8158 .loo_object_invariant = osd_object_invariant
8161 const struct lu_device_operations osd_lu_ops = {
8162 .ldo_object_alloc = osd_object_alloc,
8163 .ldo_process_config = osd_process_config,
8164 .ldo_recovery_complete = osd_recovery_complete,
8165 .ldo_prepare = osd_prepare,
8166 .ldo_fid_alloc = osd_fid_alloc,
8169 static const struct lu_device_type_operations osd_device_type_ops = {
8170 .ldto_init = osd_type_init,
8171 .ldto_fini = osd_type_fini,
8173 .ldto_start = osd_type_start,
8174 .ldto_stop = osd_type_stop,
8176 .ldto_device_alloc = osd_device_alloc,
8177 .ldto_device_free = osd_device_free,
8179 .ldto_device_init = osd_device_init,
8180 .ldto_device_fini = osd_device_fini
8183 static struct lu_device_type osd_device_type = {
8184 .ldt_tags = LU_DEVICE_DT,
8185 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8186 .ldt_ops = &osd_device_type_ops,
8187 .ldt_ctx_tags = LCT_LOCAL,
8190 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8192 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8193 struct super_block *sb = osd_sb(osd);
8195 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8199 * lprocfs legacy support.
8201 static const struct obd_ops osd_obd_device_ops = {
8202 .o_owner = THIS_MODULE,
8203 .o_connect = osd_obd_connect,
8204 .o_disconnect = osd_obd_disconnect,
8205 .o_health_check = osd_health_check,
8208 static ssize_t track_declares_assert_show(struct kobject *kobj,
8209 struct attribute *attr,
8212 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8215 static ssize_t track_declares_assert_store(struct kobject *kobj,
8216 struct attribute *attr,
8217 const char *buffer, size_t count)
8219 bool track_declares_assert;
8222 rc = kstrtobool(buffer, &track_declares_assert);
8226 ldiskfs_track_declares_assert = track_declares_assert;
8230 LUSTRE_RW_ATTR(track_declares_assert);
8232 static int __init osd_init(void)
8234 struct kobject *kobj;
8237 BUILD_BUG_ON(BH_DXLock >=
8238 sizeof(((struct buffer_head *)0)->b_state) * 8);
8239 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8240 /* please, try to keep osd_thread_info smaller than a page */
8241 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8246 rc = lu_kmem_init(ldiskfs_caches);
8250 #ifdef CONFIG_KALLSYMS
8251 priv_security_file_alloc =
8252 (void *)kallsyms_lookup_name("security_file_alloc");
8255 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
8256 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8258 lu_kmem_fini(ldiskfs_caches);
8262 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8264 rc = sysfs_create_file(kobj,
8265 &lustre_attr_track_declares_assert.attr);
8268 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8275 static void __exit osd_exit(void)
8277 struct kobject *kobj;
8279 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8281 sysfs_remove_file(kobj,
8282 &lustre_attr_track_declares_assert.attr);
8285 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8286 lu_kmem_fini(ldiskfs_caches);
8289 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8290 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8291 MODULE_VERSION(LUSTRE_VERSION_STRING);
8292 MODULE_LICENSE("GPL");
8294 module_init(osd_init);
8295 module_exit(osd_exit);