4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd/osd_handler.c
34 * Top-level entry points into osd module
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #ifdef HAVE_UIDGID_HEADER
45 # include <linux/uidgid.h>
48 /* LUSTRE_VERSION_CODE */
49 #include <lustre_ver.h>
50 /* prerequisite for linux/xattr.h */
51 #include <linux/types.h>
52 /* prerequisite for linux/xattr.h */
54 /* XATTR_{REPLACE,CREATE} */
55 #include <linux/xattr.h>
57 #include <ldiskfs/ldiskfs.h>
58 #include <ldiskfs/xattr.h>
59 #include <ldiskfs/ldiskfs_extents.h>
62 * struct OBD_{ALLOC,FREE}*()
65 #include <obd_support.h>
66 /* struct ptlrpc_thread */
67 #include <lustre_net.h>
68 #include <lustre_fid.h>
70 #include <uapi/linux/lustre_param.h>
72 #include "osd_internal.h"
73 #include "osd_dynlocks.h"
75 /* llo_* api support */
76 #include <md_object.h>
77 #include <lustre_quota.h>
79 #include <lustre_linkea.h>
81 #define PFID_STRIPE_IDX_BITS 16
82 #define PFID_STRIPE_COUNT_MASK ((1 << PFID_STRIPE_IDX_BITS) - 1)
85 module_param(ldiskfs_pdo, int, 0644);
86 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
88 int ldiskfs_track_declares_assert;
89 module_param(ldiskfs_track_declares_assert, int, 0644);
90 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
92 /* Slab to allocate dynlocks */
93 struct kmem_cache *dynlock_cachep;
95 /* Slab to allocate osd_it_ea */
96 struct kmem_cache *osd_itea_cachep;
98 static struct lu_kmem_descr ldiskfs_caches[] = {
100 .ckd_cache = &dynlock_cachep,
101 .ckd_name = "dynlock_cache",
102 .ckd_size = sizeof(struct dynlock_handle)
105 .ckd_cache = &osd_itea_cachep,
106 .ckd_name = "osd_itea_cache",
107 .ckd_size = sizeof(struct osd_it_ea)
114 static const char dot[] = ".";
115 static const char dotdot[] = "..";
117 static const struct lu_object_operations osd_lu_obj_ops;
118 static const struct dt_object_operations osd_obj_ops;
119 static const struct dt_object_operations osd_obj_otable_it_ops;
120 static const struct dt_index_operations osd_index_iam_ops;
121 static const struct dt_index_operations osd_index_ea_ops;
123 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
124 const struct lu_fid *fid);
125 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
126 struct osd_device *osd);
128 int osd_trans_declare_op2rb[] = {
129 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
130 [OSD_OT_PUNCH] = OSD_OT_MAX,
131 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
132 [OSD_OT_CREATE] = OSD_OT_DESTROY,
133 [OSD_OT_DESTROY] = OSD_OT_CREATE,
134 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
135 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
136 [OSD_OT_WRITE] = OSD_OT_WRITE,
137 [OSD_OT_INSERT] = OSD_OT_DELETE,
138 [OSD_OT_DELETE] = OSD_OT_INSERT,
139 [OSD_OT_QUOTA] = OSD_OT_MAX,
142 static int osd_has_index(const struct osd_object *obj)
144 return obj->oo_dt.do_index_ops != NULL;
147 static int osd_object_invariant(const struct lu_object *l)
149 return osd_invariant(osd_obj(l));
153 * Concurrency: doesn't matter
155 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
157 struct osd_thread_info *oti = osd_oti_get(env);
158 return oti->oti_w_locks > 0 && o->oo_owner == env;
162 * Concurrency: doesn't access mutable data
164 static int osd_root_get(const struct lu_env *env,
165 struct dt_device *dev, struct lu_fid *f)
167 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
172 * the following set of functions are used to maintain per-thread
173 * cache of FID->ino mapping. this mechanism is needed to resolve
174 * FID to inode at dt_insert() which in turn stores ino in the
175 * directory entries to keep ldiskfs compatible with ext[34].
176 * due to locking-originated restrictions we can't lookup ino
177 * using LU cache (deadlock is possible). lookup using OI is quite
178 * expensive. so instead we maintain this cache and methods like
179 * dt_create() fill it. so in the majority of cases dt_insert() is
180 * able to find needed mapping in lockless manner.
182 static struct osd_idmap_cache *
183 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
184 const struct lu_fid *fid)
186 struct osd_thread_info *oti = osd_oti_get(env);
187 struct osd_idmap_cache *idc = oti->oti_ins_cache;
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(idc, sizeof(*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(oti->oti_ins_cache,
220 oti->oti_ins_cache_used * sizeof(*idc));
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 /* new mapping is needed */
256 idc = osd_idc_add(env, osd, fid);
261 rc = osd_remote_fid(env, osd, fid);
262 if (unlikely(rc < 0))
266 /* the object is local, lookup in OI */
267 /* XXX: probably cheaper to lookup in LU first? */
268 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
270 if (unlikely(rc < 0)) {
271 CERROR("can't lookup: rc = %d\n", rc);
275 /* the object is remote */
283 * lookup mapping for given FID and fill it from the given object.
284 * the object is lolcal by definition.
286 static int osd_idc_find_and_init(const struct lu_env *env,
287 struct osd_device *osd,
288 struct osd_object *obj)
290 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
291 struct osd_idmap_cache *idc;
293 idc = osd_idc_find(env, osd, fid);
294 LASSERT(!IS_ERR(idc));
296 if (obj->oo_inode == NULL)
298 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
299 LASSERT(idc->oic_lid.oii_ino == 0);
300 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
301 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
306 /* new mapping is needed */
307 idc = osd_idc_add(env, osd, fid);
311 if (obj->oo_inode != NULL) {
312 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
313 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
319 * OSD object methods.
323 * Concurrency: no concurrent access is possible that early in object
326 static struct lu_object *osd_object_alloc(const struct lu_env *env,
327 const struct lu_object_header *hdr,
330 struct osd_object *mo;
336 l = &mo->oo_dt.do_lu;
337 dt_object_init(&mo->oo_dt, NULL, d);
338 mo->oo_dt.do_ops = &osd_obj_ops;
339 l->lo_ops = &osd_lu_obj_ops;
340 init_rwsem(&mo->oo_sem);
341 init_rwsem(&mo->oo_ext_idx_sem);
342 spin_lock_init(&mo->oo_guard);
343 INIT_LIST_HEAD(&mo->oo_xattr_list);
350 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
351 struct dentry *dentry, struct lustre_ost_attrs *loa)
355 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
356 (void *)loa, sizeof(*loa));
358 struct lustre_mdt_attrs *lma = &loa->loa_lma;
360 if (rc < sizeof(*lma))
364 lustre_loa_swab(loa, true);
365 /* Check LMA compatibility */
366 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
367 CWARN("%s: unsupported incompat LMA feature(s) %#x "
368 "for fid = "DFID", ino = %lu\n",
370 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
371 PFID(&lma->lma_self_fid), inode->i_ino);
374 } else if (rc == 0) {
382 * retrieve object from backend ext fs.
384 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
385 struct osd_inode_id *id)
388 struct inode *inode = NULL;
390 /* if we look for an inode withing a running
391 * transaction, then we risk to deadlock */
392 /* osd_dirent_check_repair() breaks this */
393 /*LASSERT(current->journal_info == NULL);*/
395 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
397 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
398 id->oii_ino, PTR_ERR(inode));
399 } else if (id->oii_gen != OSD_OII_NOGEN &&
400 inode->i_generation != id->oii_gen) {
401 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
402 "i_generation = %u\n",
403 id->oii_ino, id->oii_gen, inode->i_generation);
405 inode = ERR_PTR(-ESTALE);
406 } else if (inode->i_nlink == 0) {
407 /* due to parallel readdir and unlink,
408 * we can have dead inode here. */
409 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
411 inode = ERR_PTR(-ESTALE);
412 } else if (is_bad_inode(inode)) {
413 CWARN("%s: bad inode: ino = %u\n",
414 osd_dev2name(dev), id->oii_ino);
416 inode = ERR_PTR(-ENOENT);
417 } else if ((rc = osd_attach_jinode(inode))) {
421 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
422 if (id->oii_gen == OSD_OII_NOGEN)
423 osd_id_gen(id, inode->i_ino, inode->i_generation);
425 /* Do not update file c/mtime in ldiskfs.
426 * NB: we don't have any lock to protect this because we don't
427 * have reference on osd_object now, but contention with
428 * another lookup + attr_set can't happen in the tiny window
429 * between if (...) and set S_NOCMTIME. */
430 if (!(inode->i_flags & S_NOCMTIME))
431 inode->i_flags |= S_NOCMTIME;
436 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
437 handle_t *handle, struct dentry *child,
438 struct inode *inode, struct htree_lock *hlock)
442 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
443 if (rc == -ENOBUFS || rc == -ENOSPC) {
444 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
445 struct inode *parent = child->d_parent->d_inode;
446 struct lu_fid *fid = NULL;
448 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
450 fid = &loa->loa_lma.lma_self_fid;
451 } else if (rc2 == -ENODATA) {
452 if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
453 fid = &info->oti_fid3;
454 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
455 } else if (!osd->od_is_ost && osd->od_index == 0) {
456 fid = &info->oti_fid3;
457 lu_igif_build(fid, parent->i_ino,
458 parent->i_generation);
463 CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
464 "maximum entry limit\n",
465 osd_name(osd), parent->i_ino, PFID(fid),
466 rc == -ENOSPC ? "has reached" : "is approaching");
468 CWARN("%s: directory (inode: %lu, FID: unknown) %s "
469 "maximum entry limit\n",
470 osd_name(osd), parent->i_ino,
471 rc == -ENOSPC ? "has reached" : "is approaching");
473 /* ignore such error now */
482 static struct inode *
483 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
484 struct osd_inode_id *id, struct lu_fid *fid)
486 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
490 inode = osd_iget(info, dev, id);
494 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
496 *fid = loa->loa_lma.lma_self_fid;
497 } else if (rc == -ENODATA) {
498 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
499 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
501 lu_igif_build(fid, inode->i_ino, inode->i_generation);
509 static struct inode *osd_iget_check(struct osd_thread_info *info,
510 struct osd_device *dev,
511 const struct lu_fid *fid,
512 struct osd_inode_id *id,
519 /* The cached OI mapping is trustable. If we cannot locate the inode
520 * via the cached OI mapping, then return the failure to the caller
521 * directly without further OI checking. */
524 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
527 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
530 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
531 PFID(fid), id->oii_ino, rc);
535 if (is_bad_inode(inode)) {
540 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
541 PFID(fid), id->oii_ino);
545 if (id->oii_gen != OSD_OII_NOGEN &&
546 inode->i_generation != id->oii_gen) {
551 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
552 "oii_gen = %u, i_generation = %u\n", PFID(fid),
553 id->oii_ino, id->oii_gen, inode->i_generation);
557 if (inode->i_nlink == 0) {
562 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
563 PFID(fid), id->oii_ino);
567 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
571 __u32 saved_ino = id->oii_ino;
572 __u32 saved_gen = id->oii_gen;
575 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
577 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
578 /* XXX: There are four possible cases:
580 * Backup/restore caused the OI invalid.
582 * Someone unlinked the object but NOT removed
583 * the OI mapping, such as mount target device
584 * as ldiskfs, and modify something directly.
586 * Someone just removed the object between the
587 * former oi_lookup and the iget. It is normal.
588 * 4. Other failure cases.
590 * Generally, when the device is mounted, it will
591 * auto check whether the system is restored from
592 * file-level backup or not. We trust such detect
593 * to distinguish the 1st case from the 2nd case:
594 * if the OI files are consistent but may contain
595 * stale OI mappings because of case 2, if iget()
596 * returns -ENOENT or -ESTALE, then it should be
599 /* If the OI mapping was in OI file before the
600 * osd_iget_check(), but now, it is disappear,
601 * then it must be removed by race. That is a
602 * normal race case. */
605 /* It is the OI scrub updated the OI mapping by race.
606 * The new OI mapping must be valid. */
607 if (saved_ino != id->oii_ino ||
608 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
617 if (dev->od_scrub.os_file.sf_flags & SF_INCONSISTENT)
618 /* It still can be the case 2, but we cannot
619 * distinguish it from the case 1. So return
620 * -EREMCHG to block current operation until
621 * OI scrub rebuilt the OI mappings. */
629 if (inode->i_generation == id->oii_gen)
634 if (id->oii_gen == OSD_OII_NOGEN)
635 osd_id_gen(id, inode->i_ino, inode->i_generation);
637 /* Do not update file c/mtime in ldiskfs.
638 * NB: we don't have any lock to protect this because we don't
639 * have reference on osd_object now, but contention with
640 * another lookup + attr_set can't happen in the tiny window
641 * between if (...) and set S_NOCMTIME. */
642 if (!(inode->i_flags & S_NOCMTIME))
643 inode->i_flags |= S_NOCMTIME;
660 * \retval +v: new filter_fid, does not contain self-fid
661 * \retval 0: filter_fid_old, contains self-fid
662 * \retval -v: other failure cases
664 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
665 struct dentry *dentry, struct lu_fid *fid)
667 struct filter_fid_old *ff = &info->oti_ff;
668 struct ost_id *ostid = &info->oti_ostid;
671 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
672 if (rc == sizeof(*ff)) {
674 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
675 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
677 * XXX: use 0 as the index for compatibility, the caller will
678 * handle index related issues when necessary.
681 ostid_to_fid(fid, ostid, 0);
682 } else if (rc == sizeof(struct filter_fid)) {
684 } else if (rc >= 0) {
691 static int osd_lma_self_repair(struct osd_thread_info *info,
692 struct osd_device *osd, struct inode *inode,
693 const struct lu_fid *fid, __u32 compat)
698 LASSERT(current->journal_info == NULL);
700 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
701 osd_dto_credits_noquota[DTO_XATTR_SET]);
704 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
709 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
711 CWARN("%s: cannot self repair the LMA: rc = %d\n",
713 ldiskfs_journal_stop(jh);
717 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
719 struct osd_thread_info *info = osd_oti_get(env);
720 struct osd_device *osd = osd_obj2dev(obj);
721 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
722 struct lustre_mdt_attrs *lma = &loa->loa_lma;
723 struct inode *inode = obj->oo_inode;
724 struct dentry *dentry = &info->oti_obj_dentry;
725 struct lu_fid *fid = NULL;
726 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
730 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
731 (void *)loa, sizeof(*loa));
732 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
733 fid = &lma->lma_self_fid;
734 rc = osd_get_idif(info, inode, dentry, fid);
735 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
736 /* For the given OST-object, if it has neither LMA nor
737 * FID in XATTR_NAME_FID, then the given FID (which is
738 * contained in the @obj, from client RPC for locating
739 * the OST-object) is trusted. We use it to generate
741 osd_lma_self_repair(info, osd, inode, rfid,
752 lustre_lma_swab(lma);
753 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
754 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
755 S_ISREG(inode->i_mode)))) {
756 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
757 "fid = "DFID", ino = %lu\n", osd_name(osd),
758 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
759 PFID(rfid), inode->i_ino);
762 fid = &lma->lma_self_fid;
766 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
767 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
768 struct ost_id *oi = &info->oti_ostid;
769 struct lu_fid *fid1 = &info->oti_fid3;
770 __u32 idx = fid_idif_ost_idx(rfid);
772 /* For old IDIF, the OST index is not part of the IDIF,
773 * Means that different OSTs may have the same IDIFs.
774 * Under such case, we need to make some compatible
775 * check to make sure to trigger OI scrub properly. */
776 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
777 /* Given @rfid is new, LMA is old. */
778 fid_to_ostid(fid, oi);
779 ostid_to_fid(fid1, oi, idx);
780 if (lu_fid_eq(fid1, rfid)) {
781 if (osd->od_index_in_idif)
782 osd_lma_self_repair(info, osd,
796 struct osd_check_lmv_buf {
797 #ifdef HAVE_DIR_CONTEXT
798 /* please keep it as first member */
799 struct dir_context ctx;
801 struct osd_thread_info *oclb_info;
802 struct osd_device *oclb_dev;
803 struct osd_idmap_cache *oclb_oic;
809 * It is called internally by ->readdir() to filter out the
810 * local slave object's FID of the striped directory.
812 * \retval 1 found the local slave's FID
813 * \retval 0 continue to check next item
814 * \retval -ve for failure
816 #ifdef HAVE_FILLDIR_USE_CTX
817 static int osd_stripe_dir_filldir(struct dir_context *buf,
819 static int osd_stripe_dir_filldir(void *buf,
821 const char *name, int namelen,
822 loff_t offset, __u64 ino, unsigned d_type)
824 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
825 struct osd_thread_info *oti = oclb->oclb_info;
826 struct lu_fid *fid = &oti->oti_fid3;
827 struct osd_inode_id *id = &oti->oti_id3;
828 struct osd_device *dev = oclb->oclb_dev;
829 struct osd_idmap_cache *oic = oclb->oclb_oic;
838 sscanf(name + 1, SFID, RFID(fid));
839 if (!fid_is_sane(fid))
842 if (osd_remote_fid(oti->oti_env, dev, fid))
845 osd_id_gen(id, ino, OSD_OII_NOGEN);
846 inode = osd_iget(oti, dev, id);
848 return PTR_ERR(inode);
851 osd_add_oi_cache(oti, dev, id, fid);
855 osd_oii_insert(dev, oic, true);
856 oclb->oclb_found = true;
861 /* When lookup item under striped directory, we need to locate the master
862 * MDT-object of the striped directory firstly, then the client will send
863 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
864 * and the item's name. If the system is restored from MDT file level backup,
865 * then before the OI scrub completely built the OI files, the OI mappings of
866 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
867 * not a problem for the master MDT-object. Because when locate the master
868 * MDT-object, we will do name based lookup (for the striped directory itself)
869 * firstly, during such process we can setup the correct OI mapping for the
870 * master MDT-object. But it will be trouble for the slave MDT-object. Because
871 * the client will not trigger name based lookup on the MDT to locate the slave
872 * MDT-object before locating item under the striped directory, then when
873 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
874 * is invalid and does not know what the right OI mapping is, then the MDT has
875 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
876 * the OI file, related OI mapping is unknown yet, please try again later. And
877 * then client will re-try the RPC again and again until related OI mapping has
878 * been updated. That is quite inefficient.
880 * To resolve above trouble, we will handle it as the following two cases:
882 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
883 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
884 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
885 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
886 * directly. Please check osd_fid_lookup().
888 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
889 * Under such case, during lookup the master MDT-object, we will lookup the
890 * slave MDT-object via readdir against the master MDT-object, because the
891 * slave MDT-objects information are stored as sub-directories with the name
892 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
893 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
894 * the correct OI mapping for the slave MDT-object. */
895 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
896 struct inode *inode, struct osd_idmap_cache *oic)
898 struct lu_buf *buf = &oti->oti_big_buf;
899 struct dentry *dentry = &oti->oti_obj_dentry;
900 struct file *filp = &oti->oti_file;
901 const struct file_operations *fops;
902 struct lmv_mds_md_v1 *lmv1;
903 struct osd_check_lmv_buf oclb = {
904 #ifdef HAVE_DIR_CONTEXT
905 .ctx.actor = osd_stripe_dir_filldir,
916 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
919 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
921 lu_buf_realloc(buf, rc);
922 if (buf->lb_buf == NULL)
923 GOTO(out, rc = -ENOMEM);
929 if (unlikely(rc == 0 || rc == -ENODATA))
935 if (unlikely(buf->lb_buf == NULL)) {
936 lu_buf_realloc(buf, rc);
937 if (buf->lb_buf == NULL)
938 GOTO(out, rc = -ENOMEM);
944 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
948 dentry->d_inode = inode;
949 dentry->d_sb = inode->i_sb;
951 filp->f_path.dentry = dentry;
952 filp->f_mode = FMODE_64BITHASH;
953 filp->f_mapping = inode->i_mapping;
955 filp->private_data = NULL;
956 set_file_inode(filp, inode);
960 #ifdef HAVE_DIR_CONTEXT
961 oclb.ctx.pos = filp->f_pos;
962 rc = fops->iterate(filp, &oclb.ctx);
963 filp->f_pos = oclb.ctx.pos;
965 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
967 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
968 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
969 fops->release(inode, filp);
973 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
974 DFID": rc = %d\n", osd_ino2name(inode),
975 inode->i_ino, inode->i_generation,
976 PFID(&oic->oic_fid), rc);
983 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
984 const struct lu_fid *fid,
985 const struct lu_object_conf *conf)
987 struct osd_thread_info *info;
988 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
989 struct osd_device *dev;
990 struct osd_idmap_cache *oic;
991 struct osd_inode_id *id;
992 struct inode *inode = NULL;
993 struct osd_scrub *scrub;
994 struct scrub_file *sf;
995 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1000 bool remote = false;
1001 bool trusted = true;
1002 bool updated = false;
1005 LINVRNT(osd_invariant(obj));
1006 LASSERT(obj->oo_inode == NULL);
1007 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1009 dev = osd_dev(ldev);
1010 scrub = &dev->od_scrub;
1011 sf = &scrub->os_file;
1012 info = osd_oti_get(env);
1014 oic = &info->oti_cache;
1016 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1019 /* For the object is created as locking anchor, or for the object to
1020 * be created on disk. No need to osd_oi_lookup() at here because FID
1021 * shouldn't never be re-used, if it's really a duplicate FID from
1022 * unexpected reason, we should be able to detect it later by calling
1023 * do_create->osd_oi_insert(). */
1024 if (conf && conf->loc_flags & LOC_F_NEW)
1025 GOTO(out, result = 0);
1027 /* Search order: 1. per-thread cache. */
1028 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1034 if (!list_empty(&scrub->os_inconsistent_items)) {
1035 /* Search order: 2. OI scrub pending list. */
1036 result = osd_oii_lookup(dev, fid, id);
1041 /* The OI mapping in the OI file can be updated by the OI scrub
1042 * when we locate the inode via FID. So it may be not trustable. */
1045 /* Search order: 3. OI files. */
1046 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1047 if (result == -ENOENT) {
1048 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1049 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1050 !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1052 GOTO(out, result = 0);
1061 obj->oo_inode = NULL;
1062 /* for later passes through checks, not true on first pass */
1063 if (!IS_ERR_OR_NULL(inode))
1066 inode = osd_iget_check(info, dev, fid, id, trusted);
1067 if (!IS_ERR(inode)) {
1068 obj->oo_inode = inode;
1076 result = PTR_ERR(inode);
1077 if (result == -ENOENT || result == -ESTALE)
1078 GOTO(out, result = 0);
1080 if (result != -EREMCHG)
1084 /* We still have chance to get the valid inode: for the
1085 * object which is referenced by remote name entry, the
1086 * object on the local MDT will be linked under the dir
1087 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1089 * We do not know whether the object for the given FID
1090 * is referenced by some remote name entry or not, and
1091 * especially for DNE II, a multiple-linked object may
1092 * have many name entries reside on many MDTs.
1094 * To simplify the operation, OSD will not distinguish
1095 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1096 * only happened for the RPC from other MDT during the
1097 * OI scrub, or for the client side RPC with FID only,
1098 * such as FID to path, or from old connected client. */
1099 if (!remote && !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1100 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1104 flags |= SS_AUTO_PARTIAL;
1105 flags &= ~SS_AUTO_FULL;
1110 if (thread_is_running(&scrub->os_thread)) {
1111 if (scrub->os_partial_scan && !scrub->os_in_join)
1114 if (IS_ERR_OR_NULL(inode) || result)
1115 GOTO(out, result = -EINPROGRESS);
1118 LASSERT(obj->oo_inode == inode);
1120 osd_add_oi_cache(info, dev, id, fid);
1121 osd_oii_insert(dev, oic, true);
1125 if (dev->od_noscrub) {
1127 GOTO(out, result = -EREMCHG);
1130 LASSERT(obj->oo_inode == inode);
1132 osd_add_oi_cache(info, dev, id, fid);
1137 rc1 = osd_scrub_start(dev, flags);
1138 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1139 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1140 if (rc1 && rc1 != -EALREADY)
1141 GOTO(out, result = -EREMCHG);
1143 if (IS_ERR_OR_NULL(inode) || result)
1144 GOTO(out, result = -EINPROGRESS);
1147 LASSERT(obj->oo_inode == inode);
1149 osd_add_oi_cache(info, dev, id, fid);
1150 osd_oii_insert(dev, oic, true);
1154 result = osd_check_lma(env, obj);
1158 LASSERTF(id->oii_ino == inode->i_ino &&
1159 id->oii_gen == inode->i_generation,
1160 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1161 PFID(fid), id->oii_ino, id->oii_gen,
1162 inode->i_ino, inode->i_generation);
1164 saved_ino = inode->i_ino;
1165 saved_gen = inode->i_generation;
1167 if (unlikely(result == -ENODATA)) {
1168 /* If the OI scrub updated the OI mapping by race, it
1169 * must be valid. Trust the inode that has no LMA EA. */
1173 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1175 /* The OI mapping is still there, the inode is still
1176 * valid. It is just becaues the inode has no LMA EA. */
1177 if (saved_ino == id->oii_ino &&
1178 saved_gen == id->oii_gen)
1181 /* It is the OI scrub updated the OI mapping by race.
1182 * The new OI mapping must be valid. */
1188 /* "result == -ENOENT" means that the OI mappinghas been
1189 * removed by race, so the inode belongs to other object.
1191 * Others error can be returned directly. */
1192 if (result == -ENOENT) {
1195 obj->oo_inode = NULL;
1200 if (result != -EREMCHG)
1205 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1206 /* "result == -ENOENT" means the cached OI mapping has been removed
1207 * from the OI file by race, above inode belongs to other object. */
1208 if (result == -ENOENT) {
1211 obj->oo_inode = NULL;
1212 GOTO(out, result = 0);
1218 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1223 /* It is the OI scrub updated the OI mapping by race.
1224 * The new OI mapping must be valid. */
1230 obj->oo_compat_dot_created = 1;
1231 obj->oo_compat_dotdot_created = 1;
1233 if (S_ISDIR(inode->i_mode) &&
1234 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1235 osd_check_lmv(info, dev, inode, oic);
1237 result = osd_attach_jinode(inode);
1242 GOTO(out, result = 0);
1244 LASSERT(!obj->oo_hl_head);
1245 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1247 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1250 if (result || !obj->oo_inode) {
1251 if (!IS_ERR_OR_NULL(inode))
1254 obj->oo_inode = NULL;
1256 fid_zero(&oic->oic_fid);
1259 LINVRNT(osd_invariant(obj));
1264 * Concurrency: shouldn't matter.
1266 static void osd_object_init0(struct osd_object *obj)
1268 LASSERT(obj->oo_inode != NULL);
1269 obj->oo_dt.do_body_ops = &osd_body_ops;
1270 obj->oo_dt.do_lu.lo_header->loh_attr |=
1271 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1275 * Concurrency: no concurrent access is possible that early in object
1278 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1279 const struct lu_object_conf *conf)
1281 struct osd_object *obj = osd_obj(l);
1284 LINVRNT(osd_invariant(obj));
1286 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1287 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1288 l->lo_header->loh_attr |= LOHA_EXISTS;
1292 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1293 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1294 if (result == 0 && obj->oo_inode != NULL) {
1295 struct osd_thread_info *oti = osd_oti_get(env);
1296 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1298 osd_object_init0(obj);
1299 result = osd_get_lma(oti, obj->oo_inode,
1300 &oti->oti_obj_dentry, loa);
1302 /* Convert LMAI flags to lustre LMA flags
1303 * and cache it to oo_lma_flags */
1305 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1306 } else if (result == -ENODATA) {
1311 LINVRNT(osd_invariant(obj));
1315 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1316 * The left part is for value, binary mode. */
1317 struct osd_xattr_entry {
1318 struct list_head oxe_list;
1322 struct rcu_head oxe_rcu;
1326 static int osd_oxc_get(struct osd_object *obj, const char *name,
1329 struct osd_xattr_entry *tmp;
1330 struct osd_xattr_entry *oxe = NULL;
1331 size_t namelen = strlen(name);
1336 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1337 if (namelen == tmp->oxe_namelen &&
1338 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1345 GOTO(out, rc = -ENOENT);
1347 if (!oxe->oxe_exist)
1348 GOTO(out, rc = -ENODATA);
1351 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1354 if (buf->lb_buf == NULL)
1357 if (buf->lb_len < rc)
1358 GOTO(out, rc = -ERANGE);
1360 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1368 static void osd_oxc_free(struct rcu_head *head)
1370 struct osd_xattr_entry *oxe;
1372 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1373 OBD_FREE(oxe, oxe->oxe_len);
1376 static void osd_oxc_add(struct osd_object *obj, const char *name,
1377 const char *buf, int buflen)
1379 struct osd_xattr_entry *oxe;
1380 struct osd_xattr_entry *old = NULL;
1381 struct osd_xattr_entry *tmp;
1382 size_t namelen = strlen(name);
1383 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1385 OBD_ALLOC(oxe, len);
1389 INIT_LIST_HEAD(&oxe->oxe_list);
1391 oxe->oxe_namelen = namelen;
1392 memcpy(oxe->oxe_buf, name, namelen);
1394 LASSERT(buf != NULL);
1395 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1396 oxe->oxe_exist = true;
1398 oxe->oxe_exist = false;
1401 /* this should be rarely called, just remove old and add new */
1402 spin_lock(&obj->oo_guard);
1403 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1404 if (namelen == tmp->oxe_namelen &&
1405 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1411 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1412 call_rcu(&old->oxe_rcu, osd_oxc_free);
1414 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1416 spin_unlock(&obj->oo_guard);
1419 static void osd_oxc_del(struct osd_object *obj, const char *name)
1421 struct osd_xattr_entry *oxe;
1422 size_t namelen = strlen(name);
1424 spin_lock(&obj->oo_guard);
1425 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1426 if (namelen == oxe->oxe_namelen &&
1427 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1428 list_del_rcu(&oxe->oxe_list);
1429 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1433 spin_unlock(&obj->oo_guard);
1436 static void osd_oxc_fini(struct osd_object *obj)
1438 struct osd_xattr_entry *oxe, *next;
1440 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1441 list_del(&oxe->oxe_list);
1442 OBD_FREE(oxe, oxe->oxe_len);
1447 * Concurrency: no concurrent access is possible that late in object
1450 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1452 struct osd_object *obj = osd_obj(l);
1454 LINVRNT(osd_invariant(obj));
1457 dt_object_fini(&obj->oo_dt);
1458 if (obj->oo_hl_head != NULL)
1459 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1464 * Concurrency: no concurrent access is possible that late in object
1467 static void osd_index_fini(struct osd_object *o)
1469 struct iam_container *bag;
1471 if (o->oo_dir != NULL) {
1472 bag = &o->oo_dir->od_container;
1473 if (o->oo_inode != NULL) {
1474 if (bag->ic_object == o->oo_inode)
1475 iam_container_fini(bag);
1477 OBD_FREE_PTR(o->oo_dir);
1483 * Concurrency: no concurrent access is possible that late in object
1484 * life-cycle (for all existing callers, that is. New callers have to provide
1485 * their own locking.)
1487 static int osd_inode_unlinked(const struct inode *inode)
1489 return inode->i_nlink == 0;
1493 OSD_TXN_OI_DELETE_CREDITS = 20,
1494 OSD_TXN_INODE_DELETE_CREDITS = 20
1501 #if OSD_THANDLE_STATS
1503 * Set time when the handle is allocated
1505 static void osd_th_alloced(struct osd_thandle *oth)
1507 oth->oth_alloced = ktime_get();
1511 * Set time when the handle started
1513 static void osd_th_started(struct osd_thandle *oth)
1515 oth->oth_started = ktime_get();
1519 * Check whether the we deal with this handle for too long.
1521 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1522 ktime_t alloced, ktime_t started,
1525 ktime_t now = ktime_get();
1527 LASSERT(dev != NULL);
1529 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1530 ktime_us_delta(started, alloced));
1531 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1532 ktime_us_delta(closed, started));
1533 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1534 ktime_us_delta(now, closed));
1536 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1537 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1538 oth, now, alloced, started, closed);
1539 libcfs_debug_dumpstack(NULL);
1543 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1545 ktime_t __closed = ktime_get(); \
1546 ktime_t __alloced = oth->oth_alloced; \
1547 ktime_t __started = oth->oth_started; \
1550 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1553 #else /* OSD_THANDLE_STATS */
1555 #define osd_th_alloced(h) do {} while(0)
1556 #define osd_th_started(h) do {} while(0)
1557 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1559 #endif /* OSD_THANDLE_STATS */
1562 * Concurrency: doesn't access mutable data.
1564 static int osd_param_is_not_sane(const struct osd_device *dev,
1565 const struct thandle *th)
1567 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1569 return oh->ot_credits > osd_transaction_size(dev);
1573 * Concurrency: shouldn't matter.
1575 static void osd_trans_commit_cb(struct super_block *sb,
1576 struct ldiskfs_journal_cb_entry *jcb, int error)
1578 struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1579 struct thandle *th = &oh->ot_super;
1580 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1581 struct dt_txn_commit_cb *dcb, *tmp;
1583 LASSERT(oh->ot_handle == NULL);
1586 CERROR("transaction @0x%p commit error: %d\n", th, error);
1588 dt_txn_hook_commit(th);
1590 /* call per-transaction callbacks if any */
1591 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1593 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1594 "commit callback entry: magic=%x name='%s'\n",
1595 dcb->dcb_magic, dcb->dcb_name);
1596 list_del_init(&dcb->dcb_linkage);
1597 dcb->dcb_func(NULL, th, dcb, error);
1600 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1604 lu_context_exit(&th->th_ctx);
1605 lu_context_fini(&th->th_ctx);
1609 #ifndef HAVE_SB_START_WRITE
1610 # define sb_start_write(sb) do {} while (0)
1611 # define sb_end_write(sb) do {} while (0)
1614 static struct thandle *osd_trans_create(const struct lu_env *env,
1615 struct dt_device *d)
1617 struct osd_thread_info *oti = osd_oti_get(env);
1618 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1619 struct osd_thandle *oh;
1624 CERROR("%s: someone try to start transaction under "
1625 "readonly mode, should be disabled.\n",
1626 osd_name(osd_dt_dev(d)));
1628 RETURN(ERR_PTR(-EROFS));
1631 /* on pending IO in this thread should left from prev. request */
1632 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1634 sb_start_write(osd_sb(osd_dt_dev(d)));
1636 OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1638 oh->ot_quota_trans = &oti->oti_quota_trans;
1639 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1643 th->th_tags = LCT_TX_HANDLE;
1645 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1646 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1649 memset(oti->oti_declare_ops, 0,
1650 sizeof(oti->oti_declare_ops));
1651 memset(oti->oti_declare_ops_cred, 0,
1652 sizeof(oti->oti_declare_ops_cred));
1653 memset(oti->oti_declare_ops_used, 0,
1654 sizeof(oti->oti_declare_ops_used));
1656 sb_end_write(osd_sb(osd_dt_dev(d)));
1657 th = ERR_PTR(-ENOMEM);
1662 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1664 struct osd_thread_info *oti = osd_oti_get(env);
1665 struct osd_thandle *oh;
1667 oh = container_of0(th, struct osd_thandle, ot_super);
1668 LASSERT(oh != NULL);
1670 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1671 oti->oti_declare_ops[OSD_OT_CREATE],
1672 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1673 oti->oti_declare_ops_used[OSD_OT_CREATE],
1674 oti->oti_declare_ops[OSD_OT_DESTROY],
1675 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1676 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1677 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1678 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1679 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1680 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1681 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1682 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1683 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1684 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1685 oti->oti_declare_ops[OSD_OT_WRITE],
1686 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1687 oti->oti_declare_ops_used[OSD_OT_WRITE],
1688 oti->oti_declare_ops[OSD_OT_PUNCH],
1689 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1690 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1691 oti->oti_declare_ops[OSD_OT_QUOTA],
1692 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1693 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1694 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1695 oti->oti_declare_ops[OSD_OT_INSERT],
1696 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1697 oti->oti_declare_ops_used[OSD_OT_INSERT],
1698 oti->oti_declare_ops[OSD_OT_DELETE],
1699 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1700 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1701 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1702 oti->oti_declare_ops[OSD_OT_REF_ADD],
1703 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1704 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1705 oti->oti_declare_ops[OSD_OT_REF_DEL],
1706 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1707 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1711 * Concurrency: shouldn't matter.
1713 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1716 struct osd_thread_info *oti = osd_oti_get(env);
1717 struct osd_device *dev = osd_dt_dev(d);
1719 struct osd_thandle *oh;
1724 LASSERT(current->journal_info == NULL);
1726 oh = container_of0(th, struct osd_thandle, ot_super);
1727 LASSERT(oh != NULL);
1728 LASSERT(oh->ot_handle == NULL);
1730 rc = dt_txn_hook_start(env, d, th);
1734 if (unlikely(osd_param_is_not_sane(dev, th))) {
1735 static unsigned long last_printed;
1736 static int last_credits;
1738 /* don't make noise on a tiny testing systems
1739 * actual credits misuse will be caught anyway */
1740 if (last_credits != oh->ot_credits &&
1741 time_after(jiffies, last_printed +
1742 msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1743 osd_transaction_size(dev) > 512) {
1744 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1745 oh->ot_credits, osd_transaction_size(dev));
1746 osd_trans_dump_creds(env, th);
1747 libcfs_debug_dumpstack(NULL);
1748 last_credits = oh->ot_credits;
1749 last_printed = jiffies;
1751 /* XXX Limit the credits to 'max_transaction_buffers', and
1752 * let the underlying filesystem to catch the error if
1753 * we really need so many credits.
1755 * This should be removed when we can calculate the
1756 * credits precisely. */
1757 oh->ot_credits = osd_transaction_size(dev);
1758 } else if (ldiskfs_track_declares_assert != 0) {
1759 /* reserve few credits to prevent an assertion in JBD
1760 * our debugging mechanism will be able to detected
1761 * overuse. this can help to debug single-update
1763 oh->ot_credits += 10;
1764 if (unlikely(osd_param_is_not_sane(dev, th)))
1765 oh->ot_credits = osd_transaction_size(dev);
1769 * XXX temporary stuff. Some abstraction layer should
1772 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1776 LASSERT(oti->oti_txns == 0);
1777 lu_context_init(&th->th_ctx, th->th_tags);
1778 lu_context_enter(&th->th_ctx);
1780 lu_device_get(&d->dd_lu_dev);
1781 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1792 static int osd_seq_exists(const struct lu_env *env,
1793 struct osd_device *osd, u64 seq)
1795 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1796 struct seq_server_site *ss = osd_seq_site(osd);
1800 LASSERT(ss != NULL);
1801 LASSERT(ss->ss_server_fld != NULL);
1803 rc = osd_fld_lookup(env, osd, seq, range);
1806 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1807 osd_name(osd), seq, rc);
1811 RETURN(ss->ss_node_id == range->lsr_index);
1814 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1816 struct dt_txn_commit_cb *dcb;
1817 struct dt_txn_commit_cb *tmp;
1819 /* call per-transaction stop callbacks if any */
1820 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1822 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1823 "commit callback entry: magic=%x name='%s'\n",
1824 dcb->dcb_magic, dcb->dcb_name);
1825 list_del_init(&dcb->dcb_linkage);
1826 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1831 * Concurrency: shouldn't matter.
1833 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1836 int rc = 0, remove_agents = 0;
1837 struct osd_thandle *oh;
1838 struct osd_thread_info *oti = osd_oti_get(env);
1839 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1840 struct osd_device *osd = osd_dt_dev(th->th_dev);
1841 struct qsd_instance *qsd = osd->od_quota_slave;
1842 struct lquota_trans *qtrans;
1845 oh = container_of0(th, struct osd_thandle, ot_super);
1847 remove_agents = oh->ot_remove_agents;
1849 qtrans = oh->ot_quota_trans;
1850 oh->ot_quota_trans = NULL;
1852 if (oh->ot_handle != NULL) {
1854 handle_t *hdl = oh->ot_handle;
1857 * add commit callback
1858 * notice we don't do this in osd_trans_start()
1859 * as underlying transaction can change during truncate
1861 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1864 LASSERT(oti->oti_txns == 1);
1867 rc = dt_txn_hook_stop(env, th);
1869 CERROR("%s: failed in transaction hook: rc = %d\n",
1872 osd_trans_stop_cb(oh, rc);
1873 /* hook functions might modify th_sync */
1874 hdl->h_sync = th->th_sync;
1876 oh->ot_handle = NULL;
1877 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1879 CERROR("%s: failed to stop transaction: rc = %d\n",
1880 osd_name(osd), rc2);
1884 osd_trans_stop_cb(oh, th->th_result);
1888 /* inform the quota slave device that the transaction is stopping */
1889 qsd_op_end(env, qsd, qtrans);
1891 /* as we want IO to journal and data IO be concurrent, we don't block
1892 * awaiting data IO completion in osd_do_bio(), instead we wait here
1893 * once transaction is submitted to the journal. all reqular requests
1894 * don't do direct IO (except read/write), thus this wait_event becomes
1897 * IMPORTANT: we have to wait till any IO submited by the thread is
1898 * completed otherwise iobuf may be corrupted by different request
1900 wait_event(iobuf->dr_wait,
1901 atomic_read(&iobuf->dr_numreqs) == 0);
1902 osd_fini_iobuf(osd, iobuf);
1904 rc = iobuf->dr_error;
1906 if (unlikely(remove_agents != 0))
1907 osd_process_scheduled_agent_removals(env, osd);
1909 /* reset OI cache for safety */
1910 oti->oti_ins_cache_used = 0;
1912 sb_end_write(osd_sb(osd));
1917 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1919 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1922 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1923 LASSERT(&dcb->dcb_func != NULL);
1924 if (dcb->dcb_flags & DCB_TRANS_STOP)
1925 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1927 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1933 * Called just before object is freed. Releases all resources except for
1934 * object itself (that is released by osd_object_free()).
1936 * Concurrency: no concurrent access is possible that late in object
1939 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1941 struct osd_object *obj = osd_obj(l);
1942 struct inode *inode = obj->oo_inode;
1944 LINVRNT(osd_invariant(obj));
1947 * If object is unlinked remove fid->ino mapping from object index.
1950 osd_index_fini(obj);
1951 if (inode != NULL) {
1952 struct qsd_instance *qsd = osd_obj2dev(obj)->od_quota_slave;
1953 qid_t uid = i_uid_read(inode);
1954 qid_t gid = i_gid_read(inode);
1957 obj->oo_inode = NULL;
1960 struct osd_thread_info *info = osd_oti_get(env);
1961 struct lquota_id_info *qi = &info->oti_qi;
1963 /* Release granted quota to master if necessary */
1964 qi->lqi_id.qid_uid = uid;
1965 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1967 qi->lqi_id.qid_uid = gid;
1968 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1970 qi->lqi_id.qid_uid = i_projid_read(inode);
1971 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1977 * Concurrency: ->loo_object_release() is called under site spin-lock.
1979 static void osd_object_release(const struct lu_env *env,
1980 struct lu_object *l)
1982 struct osd_object *o = osd_obj(l);
1983 /* nobody should be releasing a non-destroyed object with nlink=0
1984 * the API allows this, but ldiskfs doesn't like and then report
1985 * this inode as deleted */
1986 if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1991 * Concurrency: shouldn't matter.
1993 static int osd_object_print(const struct lu_env *env, void *cookie,
1994 lu_printer_t p, const struct lu_object *l)
1996 struct osd_object *o = osd_obj(l);
1997 struct iam_descr *d;
1999 if (o->oo_dir != NULL)
2000 d = o->oo_dir->od_container.ic_descr;
2003 return (*p)(env, cookie,
2004 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2006 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2007 o->oo_inode ? o->oo_inode->i_generation : 0,
2008 d ? d->id_ops->id_name : "plain");
2012 * Concurrency: shouldn't matter.
2014 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2015 struct obd_statfs *sfs)
2017 struct osd_device *osd = osd_dt_dev(d);
2018 struct super_block *sb = osd_sb(osd);
2019 struct kstatfs *ksfs;
2023 if (unlikely(osd->od_mnt == NULL))
2024 return -EINPROGRESS;
2026 /* osd_lproc.c call this without env, allocate ksfs for that case */
2027 if (unlikely(env == NULL)) {
2028 OBD_ALLOC_PTR(ksfs);
2032 ksfs = &osd_oti_get(env)->oti_ksfs;
2035 result = sb->s_op->statfs(sb->s_root, ksfs);
2039 statfs_pack(sfs, ksfs);
2040 if (unlikely(sb->s_flags & MS_RDONLY))
2041 sfs->os_state |= OS_STATE_READONLY;
2042 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2043 LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2044 sfs->os_maxbytes = sb->s_maxbytes;
2046 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2049 * Reserve some space so to avoid fragmenting the filesystem too much.
2050 * Fragmentation not only impacts performance, but can also increase
2051 * metadata overhead significantly, causing grant calculation to be
2054 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2056 CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2057 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2058 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2059 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2061 sfs->os_blocks -= reserved;
2062 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2063 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2066 if (unlikely(env == NULL))
2072 * Estimate space needed for file creations. We assume the largest filename
2073 * which is 2^64 - 1, hence a filename of 20 chars.
2074 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2076 #ifdef __LDISKFS_DIR_REC_LEN
2077 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2079 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2083 * Concurrency: doesn't access mutable data.
2085 static void osd_conf_get(const struct lu_env *env,
2086 const struct dt_device *dev,
2087 struct dt_device_param *param)
2089 struct super_block *sb = osd_sb(osd_dt_dev(dev));
2093 * XXX should be taken from not-yet-existing fs abstraction layer.
2095 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2096 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2097 param->ddp_symlink_max = sb->s_blocksize;
2098 param->ddp_mount_type = LDD_MT_LDISKFS;
2099 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2100 param->ddp_maxbytes = sb->s_maxbytes;
2102 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2103 /* inode are statically allocated, so per-inode space consumption
2104 * is the space consumed by the directory entry */
2105 param->ddp_inodespace = PER_OBJ_USAGE;
2106 /* EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2107 * = 128MB) which is unlikely to be hit in real life. Report a smaller
2108 * maximum length to not under count the actual number of extents
2109 * needed for writing a file. */
2110 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2111 /* worst-case extent insertion metadata overhead */
2112 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2113 param->ddp_mntopts = 0;
2114 if (test_opt(sb, XATTR_USER))
2115 param->ddp_mntopts |= MNTOPT_USERXATTR;
2116 if (test_opt(sb, POSIX_ACL))
2117 param->ddp_mntopts |= MNTOPT_ACL;
2119 /* LOD might calculate the max stripe count based on max_ea_size,
2120 * so we need take account in the overhead as well,
2121 * xattr_header + magic + xattr_entry_head */
2122 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2123 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2125 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2126 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2127 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2131 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2133 /* Preferred RPC size for efficient disk IO. 4MB shows good
2134 * all-around performance for ldiskfs, but use bigalloc chunk size
2135 * by default if larger. */
2136 #if defined(LDISKFS_CLUSTER_SIZE)
2137 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2138 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2141 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2145 * Concurrency: shouldn't matter.
2147 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2151 CDEBUG(D_CACHE, "%s: syncing OSD\n", osd_dt_dev(d)->od_svname);
2153 rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2155 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2162 * Start commit for OSD device.
2164 * An implementation of dt_commit_async method for OSD device.
2165 * Asychronously starts underlayng fs sync and thereby a transaction
2168 * \param env environment
2169 * \param d dt device
2171 * \see dt_device_operations
2173 static int osd_commit_async(const struct lu_env *env,
2174 struct dt_device *d)
2176 struct super_block *s = osd_sb(osd_dt_dev(d));
2179 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2180 RETURN(s->s_op->sync_fs(s, 0));
2183 /* Our own copy of the set readonly functions if present, or NU if not. */
2184 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2185 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2186 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2189 * Concurrency: shouldn't matter.
2191 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2193 struct super_block *sb = osd_sb(osd_dt_dev(d));
2194 struct block_device *dev = sb->s_bdev;
2195 int rc = -EOPNOTSUPP;
2198 if (priv_dev_set_rdonly) {
2199 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2202 CERROR("*** setting %s read-only ***\n",
2203 osd_dt_dev(d)->od_svname);
2205 if (sb->s_op->freeze_fs) {
2206 rc = sb->s_op->freeze_fs(sb);
2211 if (jdev && (jdev != dev)) {
2212 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2214 priv_dev_set_rdonly(jdev);
2216 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2217 priv_dev_set_rdonly(dev);
2219 if (sb->s_op->unfreeze_fs)
2220 sb->s_op->unfreeze_fs(sb);
2225 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2226 osd_dt_dev(d)->od_svname, (long)dev, rc);
2232 * Note: we do not count into QUOTA here.
2233 * If we mount with --data_journal we may need more.
2235 const int osd_dto_credits_noquota[DTO_NR] = {
2238 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2239 * SINGLEDATA_TRANS_BLOCKS(8)
2240 * XXX Note: maybe iam need more, since iam have more level than
2243 [DTO_INDEX_INSERT] = 16,
2246 * just modify a single entry, probably merge few within a block
2248 [DTO_INDEX_DELETE] = 1,
2252 [DTO_INDEX_UPDATE] = 16,
2254 * 4(inode, inode bits, groups, GDT)
2255 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2257 [DTO_OBJECT_CREATE] = 4,
2259 * 4(inode, inode bits, groups, GDT)
2260 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2262 [DTO_OBJECT_DELETE] = 4,
2264 * Attr set credits (inode)
2266 [DTO_ATTR_SET_BASE] = 1,
2268 * Xattr set. The same as xattr of EXT3.
2269 * DATA_TRANS_BLOCKS(14)
2270 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2271 * are also counted in. Do not know why?
2273 [DTO_XATTR_SET] = 14,
2275 * credits for inode change during write.
2277 [DTO_WRITE_BASE] = 3,
2279 * credits for single block write.
2281 [DTO_WRITE_BLOCK] = 14,
2283 * Attr set credits for chown.
2284 * This is extra credits for setattr, and it is null without quota
2286 [DTO_ATTR_SET_CHOWN] = 0
2289 static const struct dt_device_operations osd_dt_ops = {
2290 .dt_root_get = osd_root_get,
2291 .dt_statfs = osd_statfs,
2292 .dt_trans_create = osd_trans_create,
2293 .dt_trans_start = osd_trans_start,
2294 .dt_trans_stop = osd_trans_stop,
2295 .dt_trans_cb_add = osd_trans_cb_add,
2296 .dt_conf_get = osd_conf_get,
2297 .dt_sync = osd_sync,
2299 .dt_commit_async = osd_commit_async,
2302 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2305 struct osd_object *obj = osd_dt_obj(dt);
2306 struct osd_thread_info *oti = osd_oti_get(env);
2308 LINVRNT(osd_invariant(obj));
2310 LASSERT(obj->oo_owner != env);
2311 down_read_nested(&obj->oo_sem, role);
2313 LASSERT(obj->oo_owner == NULL);
2317 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2320 struct osd_object *obj = osd_dt_obj(dt);
2321 struct osd_thread_info *oti = osd_oti_get(env);
2323 LINVRNT(osd_invariant(obj));
2325 LASSERT(obj->oo_owner != env);
2326 down_write_nested(&obj->oo_sem, role);
2328 LASSERT(obj->oo_owner == NULL);
2329 obj->oo_owner = env;
2333 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2335 struct osd_object *obj = osd_dt_obj(dt);
2336 struct osd_thread_info *oti = osd_oti_get(env);
2338 LINVRNT(osd_invariant(obj));
2340 LASSERT(oti->oti_r_locks > 0);
2342 up_read(&obj->oo_sem);
2345 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2347 struct osd_object *obj = osd_dt_obj(dt);
2348 struct osd_thread_info *oti = osd_oti_get(env);
2350 LINVRNT(osd_invariant(obj));
2352 LASSERT(obj->oo_owner == env);
2353 LASSERT(oti->oti_w_locks > 0);
2355 obj->oo_owner = NULL;
2356 up_write(&obj->oo_sem);
2359 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2361 struct osd_object *obj = osd_dt_obj(dt);
2363 LINVRNT(osd_invariant(obj));
2365 return obj->oo_owner == env;
2368 static struct timespec *osd_inode_time(const struct lu_env *env,
2369 struct inode *inode, __u64 seconds)
2371 struct osd_thread_info *oti = osd_oti_get(env);
2372 struct timespec *t = &oti->oti_time;
2374 t->tv_sec = seconds;
2376 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2380 static void osd_inode_getattr(const struct lu_env *env,
2381 struct inode *inode, struct lu_attr *attr)
2383 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2384 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2385 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2386 LA_BLKSIZE | LA_TYPE;
2388 attr->la_atime = LTIME_S(inode->i_atime);
2389 attr->la_mtime = LTIME_S(inode->i_mtime);
2390 attr->la_ctime = LTIME_S(inode->i_ctime);
2391 attr->la_mode = inode->i_mode;
2392 attr->la_size = i_size_read(inode);
2393 attr->la_blocks = inode->i_blocks;
2394 attr->la_uid = i_uid_read(inode);
2395 attr->la_gid = i_gid_read(inode);
2396 attr->la_projid = i_projid_read(inode);
2397 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2398 attr->la_nlink = inode->i_nlink;
2399 attr->la_rdev = inode->i_rdev;
2400 attr->la_blksize = 1 << inode->i_blkbits;
2401 attr->la_blkbits = inode->i_blkbits;
2403 * Ext4 did not transfer inherit flags from raw inode
2404 * to inode flags, and ext4 internally test raw inode
2405 * @i_flags directly. Instead of patching ext4, we do it here.
2407 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2408 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2411 static int osd_attr_get(const struct lu_env *env,
2412 struct dt_object *dt,
2413 struct lu_attr *attr)
2415 struct osd_object *obj = osd_dt_obj(dt);
2417 if (unlikely(!dt_object_exists(dt)))
2419 if (unlikely(obj->oo_destroyed))
2422 LASSERT(!dt_object_remote(dt));
2423 LINVRNT(osd_invariant(obj));
2425 spin_lock(&obj->oo_guard);
2426 osd_inode_getattr(env, obj->oo_inode, attr);
2427 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2428 attr->la_flags |= LUSTRE_ORPHAN_FL;
2429 spin_unlock(&obj->oo_guard);
2434 static int osd_declare_attr_qid(const struct lu_env *env,
2435 struct osd_object *obj,
2436 struct osd_thandle *oh, long long bspace,
2437 qid_t old_id, qid_t new_id, bool enforce,
2438 unsigned type, bool ignore_edquot)
2441 struct osd_thread_info *info = osd_oti_get(env);
2442 struct lquota_id_info *qi = &info->oti_qi;
2444 qi->lqi_type = type;
2445 /* inode accounting */
2446 qi->lqi_is_blk = false;
2448 /* one more inode for the new id ... */
2449 qi->lqi_id.qid_uid = new_id;
2451 /* Reserve credits for the new id */
2452 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2453 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2458 /* and one less inode for the current id */
2459 qi->lqi_id.qid_uid = old_id;
2461 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2462 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2467 /* block accounting */
2468 qi->lqi_is_blk = true;
2470 /* more blocks for the new id ... */
2471 qi->lqi_id.qid_uid = new_id;
2472 qi->lqi_space = bspace;
2474 * Credits for the new uid has been reserved, re-use "obj"
2475 * to save credit reservation.
2477 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2478 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2483 /* and finally less blocks for the current uid */
2484 qi->lqi_id.qid_uid = old_id;
2485 qi->lqi_space = -bspace;
2486 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2487 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2493 static int osd_declare_attr_set(const struct lu_env *env,
2494 struct dt_object *dt,
2495 const struct lu_attr *attr,
2496 struct thandle *handle)
2498 struct osd_thandle *oh;
2499 struct osd_object *obj;
2507 LASSERT(dt != NULL);
2508 LASSERT(handle != NULL);
2510 obj = osd_dt_obj(dt);
2511 LASSERT(osd_invariant(obj));
2513 oh = container_of0(handle, struct osd_thandle, ot_super);
2514 LASSERT(oh->ot_handle == NULL);
2516 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2517 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2519 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2520 osd_dto_credits_noquota[DTO_XATTR_SET]);
2522 if (attr == NULL || obj->oo_inode == NULL)
2525 bspace = obj->oo_inode->i_blocks << 9;
2526 bspace = toqb(bspace);
2528 /* Changing ownership is always preformed by super user, it should not
2529 * fail with EDQUOT unless required explicitly.
2531 * We still need to call the osd_declare_qid() to calculate the journal
2532 * credits for updating quota accounting files and to trigger quota
2533 * space adjustment once the operation is completed.*/
2534 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2535 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2538 CDEBUG(D_QUOTA, "%s: enforce quota on UID %u, GID %u"
2539 "(the quota space is %lld)\n",
2540 obj->oo_inode->i_sb->s_id, attr->la_uid,
2541 attr->la_gid, bspace);
2544 uid = i_uid_read(obj->oo_inode);
2545 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2546 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2547 attr->la_uid, enforce, USRQUOTA,
2552 gid = i_gid_read(obj->oo_inode);
2553 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2554 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2555 i_gid_read(obj->oo_inode),
2556 attr->la_gid, enforce, GRPQUOTA,
2562 #ifdef HAVE_PROJECT_QUOTA
2563 if (attr->la_valid & LA_PROJID) {
2564 __u32 projid = i_projid_read(obj->oo_inode);
2565 enforce = (attr->la_valid & LA_PROJID) &&
2566 (attr->la_projid != projid);
2567 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2568 (qid_t)projid, (qid_t)attr->la_projid,
2569 enforce, PRJQUOTA, true);
2577 static int osd_inode_setattr(const struct lu_env *env,
2578 struct inode *inode, const struct lu_attr *attr)
2580 __u64 bits = attr->la_valid;
2582 /* Only allow set size for regular file */
2583 if (!S_ISREG(inode->i_mode))
2584 bits &= ~(LA_SIZE | LA_BLOCKS);
2589 if (bits & LA_ATIME)
2590 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2591 if (bits & LA_CTIME)
2592 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2593 if (bits & LA_MTIME)
2594 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2595 if (bits & LA_SIZE) {
2596 spin_lock(&inode->i_lock);
2597 LDISKFS_I(inode)->i_disksize = attr->la_size;
2598 i_size_write(inode, attr->la_size);
2599 spin_unlock(&inode->i_lock);
2602 /* OSD should not change "i_blocks" which is used by quota.
2603 * "i_blocks" should be changed by ldiskfs only. */
2605 inode->i_mode = (inode->i_mode & S_IFMT) |
2606 (attr->la_mode & ~S_IFMT);
2608 i_uid_write(inode, attr->la_uid);
2610 i_gid_write(inode, attr->la_gid);
2611 if (bits & LA_PROJID)
2612 i_projid_write(inode, attr->la_projid);
2613 if (bits & LA_NLINK)
2614 set_nlink(inode, attr->la_nlink);
2616 inode->i_rdev = attr->la_rdev;
2618 if (bits & LA_FLAGS) {
2619 /* always keep S_NOCMTIME */
2620 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2623 * Ext4 did not transfer inherit flags from
2624 * @inode->i_flags to raw inode i_flags when writing
2625 * flags, we do it explictly here.
2627 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2628 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2630 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2635 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2639 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2640 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2643 ll_vfs_dq_init(inode);
2645 if (attr->la_valid & LA_UID)
2646 iattr.ia_valid |= ATTR_UID;
2647 if (attr->la_valid & LA_GID)
2648 iattr.ia_valid |= ATTR_GID;
2649 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2650 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2652 rc = ll_vfs_dq_transfer(inode, &iattr);
2654 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2655 "enforcement enabled on the ldiskfs "
2656 "filesystem?\n", inode->i_sb->s_id, rc);
2661 #ifdef HAVE_PROJECT_QUOTA
2662 /* Handle project id transfer here properly */
2663 if (attr->la_valid & LA_PROJID &&
2664 attr->la_projid != i_projid_read(inode)) {
2665 rc = __ldiskfs_ioctl_setproject(inode, attr->la_projid);
2667 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2668 "enforcement enabled on the ldiskfs "
2669 "filesystem?\n", inode->i_sb->s_id, rc);
2677 static int osd_attr_set(const struct lu_env *env,
2678 struct dt_object *dt,
2679 const struct lu_attr *attr,
2680 struct thandle *handle)
2682 struct osd_object *obj = osd_dt_obj(dt);
2683 struct inode *inode;
2686 if (!dt_object_exists(dt))
2689 LASSERT(handle != NULL);
2690 LASSERT(!dt_object_remote(dt));
2691 LASSERT(osd_invariant(obj));
2693 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2695 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2696 struct osd_thread_info *oti = osd_oti_get(env);
2697 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2698 struct lu_fid *fid1 = &oti->oti_fid;
2699 struct osd_inode_id *id = &oti->oti_id;
2700 struct iam_path_descr *ipd;
2701 struct iam_container *bag;
2702 struct osd_thandle *oh;
2705 fid_cpu_to_be(fid1, fid0);
2706 memset(id, 1, sizeof(*id));
2707 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2708 fid0)->oi_dir.od_container;
2709 ipd = osd_idx_ipd_get(env, bag);
2710 if (unlikely(ipd == NULL))
2713 oh = container_of0(handle, struct osd_thandle, ot_super);
2714 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2715 (const struct iam_rec *)id, ipd);
2716 osd_ipd_put(env, bag, ipd);
2717 return(rc > 0 ? 0 : rc);
2720 inode = obj->oo_inode;
2722 rc = osd_quota_transfer(inode, attr);
2726 spin_lock(&obj->oo_guard);
2727 rc = osd_inode_setattr(env, inode, attr);
2728 spin_unlock(&obj->oo_guard);
2732 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2734 if (!(attr->la_valid & LA_FLAGS))
2737 /* Let's check if there are extra flags need to be set into LMA */
2738 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2739 struct osd_thread_info *info = osd_oti_get(env);
2740 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2742 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2743 &info->oti_ost_attrs);
2747 lma->lma_incompat |=
2748 lustre_to_lma_flags(attr->la_flags);
2749 lustre_lma_swab(lma);
2750 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2751 lma, sizeof(*lma), XATTR_REPLACE);
2753 struct osd_device *osd = osd_obj2dev(obj);
2755 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2756 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2757 lma->lma_incompat, rc);
2760 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2762 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2765 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2770 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2771 struct osd_object *obj,
2772 const char *name, const int namelen)
2774 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2777 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2778 umode_t mode, struct dt_allocation_hint *hint,
2782 struct osd_device *osd = osd_obj2dev(obj);
2783 struct osd_thandle *oth;
2784 struct dt_object *parent = NULL;
2785 struct inode *inode;
2787 LINVRNT(osd_invariant(obj));
2788 LASSERT(obj->oo_inode == NULL);
2789 LASSERT(obj->oo_hl_head == NULL);
2791 if (S_ISDIR(mode) && ldiskfs_pdo) {
2792 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2793 if (obj->oo_hl_head == NULL)
2797 oth = container_of(th, struct osd_thandle, ot_super);
2798 LASSERT(oth->ot_handle->h_transaction != NULL);
2800 if (hint != NULL && hint->dah_parent != NULL &&
2801 !dt_object_remote(hint->dah_parent))
2802 parent = hint->dah_parent;
2804 inode = ldiskfs_create_inode(oth->ot_handle,
2805 parent ? osd_dt_obj(parent)->oo_inode :
2806 osd_sb(osd)->s_root->d_inode,
2808 if (!IS_ERR(inode)) {
2809 /* Do not update file c/mtime in ldiskfs. */
2810 inode->i_flags |= S_NOCMTIME;
2812 /* For new created object, it must be consistent,
2813 * and it is unnecessary to scrub against it. */
2814 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2816 obj->oo_inode = inode;
2819 if (obj->oo_hl_head != NULL) {
2820 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2821 obj->oo_hl_head = NULL;
2823 result = PTR_ERR(inode);
2825 LINVRNT(osd_invariant(obj));
2833 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2834 struct lu_attr *attr,
2835 struct dt_allocation_hint *hint,
2836 struct dt_object_format *dof,
2840 struct osd_thandle *oth;
2841 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2843 LASSERT(S_ISDIR(attr->la_mode));
2845 oth = container_of(th, struct osd_thandle, ot_super);
2846 LASSERT(oth->ot_handle->h_transaction != NULL);
2847 result = osd_mkfile(info, obj, mode, hint, th);
2852 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2853 struct lu_attr *attr,
2854 struct dt_allocation_hint *hint,
2855 struct dt_object_format *dof,
2859 struct osd_thandle *oth;
2860 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2862 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2864 LASSERT(S_ISREG(attr->la_mode));
2866 oth = container_of(th, struct osd_thandle, ot_super);
2867 LASSERT(oth->ot_handle->h_transaction != NULL);
2869 result = osd_mkfile(info, obj, mode, hint, th);
2871 LASSERT(obj->oo_inode != NULL);
2872 if (feat->dif_flags & DT_IND_VARKEY)
2873 result = iam_lvar_create(obj->oo_inode,
2874 feat->dif_keysize_max,
2876 feat->dif_recsize_max,
2879 result = iam_lfix_create(obj->oo_inode,
2880 feat->dif_keysize_max,
2882 feat->dif_recsize_max,
2889 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2890 struct lu_attr *attr,
2891 struct dt_allocation_hint *hint,
2892 struct dt_object_format *dof,
2895 LASSERT(S_ISREG(attr->la_mode));
2896 return osd_mkfile(info, obj, (attr->la_mode &
2897 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2900 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2901 struct lu_attr *attr,
2902 struct dt_allocation_hint *hint,
2903 struct dt_object_format *dof,
2906 LASSERT(S_ISLNK(attr->la_mode));
2907 return osd_mkfile(info, obj, (attr->la_mode &
2908 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2911 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2912 struct lu_attr *attr,
2913 struct dt_allocation_hint *hint,
2914 struct dt_object_format *dof,
2917 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2920 LINVRNT(osd_invariant(obj));
2921 LASSERT(obj->oo_inode == NULL);
2922 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2923 S_ISFIFO(mode) || S_ISSOCK(mode));
2925 result = osd_mkfile(info, obj, mode, hint, th);
2927 LASSERT(obj->oo_inode != NULL);
2929 * This inode should be marked dirty for i_rdev. Currently
2930 * that is done in the osd_attr_init().
2932 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2935 LINVRNT(osd_invariant(obj));
2939 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2941 struct dt_allocation_hint *hint,
2942 struct dt_object_format *dof,
2945 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2947 osd_obj_type_f result;
2963 result = osd_mk_index;
2974 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2975 struct dt_object *parent, struct dt_object *child,
2980 ah->dah_parent = parent;
2981 ah->dah_mode = child_mode;
2983 if (parent != NULL && !dt_object_remote(parent)) {
2984 /* will help to find FID->ino at dt_insert("..") */
2985 struct osd_object *pobj = osd_dt_obj(parent);
2986 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2990 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2991 struct lu_attr *attr, struct dt_object_format *dof)
2993 struct inode *inode = obj->oo_inode;
2994 __u64 valid = attr->la_valid;
2997 attr->la_valid &= ~(LA_TYPE | LA_MODE);
2999 if (dof->dof_type != DFT_NODE)
3000 attr->la_valid &= ~LA_RDEV;
3001 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
3002 attr->la_valid &= ~LA_ATIME;
3003 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
3004 attr->la_valid &= ~LA_CTIME;
3005 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
3006 attr->la_valid &= ~LA_MTIME;
3008 result = osd_quota_transfer(inode, attr);
3012 if (attr->la_valid != 0) {
3013 result = osd_inode_setattr(info->oti_env, inode, attr);
3015 * The osd_inode_setattr() should always succeed here. The
3016 * only error that could be returned is EDQUOT when we are
3017 * trying to change the UID or GID of the inode. However, this
3018 * should not happen since quota enforcement is no longer
3019 * enabled on ldiskfs (lquota takes care of it).
3021 LASSERTF(result == 0, "%d\n", result);
3022 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3025 attr->la_valid = valid;
3029 * Helper function for osd_create()
3031 * \retval 0, on success
3033 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3034 struct lu_attr *attr, struct dt_allocation_hint *hint,
3035 struct dt_object_format *dof, struct thandle *th)
3040 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3042 /* we drop umask so that permissions we pass are not affected */
3043 umask = current->fs->umask;
3044 current->fs->umask = 0;
3046 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3048 if (likely(obj->oo_inode != NULL)) {
3049 LASSERT(obj->oo_inode->i_state & I_NEW);
3051 /* Unlock the inode before attr initialization to avoid
3052 * unnecessary dqget operations. LU-6378 */
3053 unlock_new_inode(obj->oo_inode);
3056 if (likely(result == 0)) {
3057 osd_attr_init(info, obj, attr, dof);
3058 osd_object_init0(obj);
3061 /* restore previous umask value */
3062 current->fs->umask = umask;
3064 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3070 * Helper function for osd_create()
3072 * \retval 0, on success
3074 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3075 const struct lu_fid *fid, struct thandle *th)
3077 struct osd_thread_info *info = osd_oti_get(env);
3078 struct osd_inode_id *id = &info->oti_id;
3079 struct osd_device *osd = osd_obj2dev(obj);
3080 struct osd_thandle *oh;
3083 LASSERT(obj->oo_inode != NULL);
3085 oh = container_of0(th, struct osd_thandle, ot_super);
3086 LASSERT(oh->ot_handle);
3087 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3089 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3090 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3091 OI_CHECK_FLD, NULL);
3092 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3097 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3098 u64 seq, struct lu_seq_range *range)
3100 struct seq_server_site *ss = osd_seq_site(osd);
3102 if (fid_seq_is_idif(seq)) {
3103 fld_range_set_ost(range);
3104 range->lsr_index = idif_ost_idx(seq);
3108 if (!fid_seq_in_fldb(seq)) {
3109 fld_range_set_mdt(range);
3111 /* FIXME: If ss is NULL, it suppose not get lsr_index
3113 range->lsr_index = ss->ss_node_id;
3117 LASSERT(ss != NULL);
3118 fld_range_set_any(range);
3119 /* OSD will only do local fld lookup */
3120 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3123 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3124 struct lu_attr *attr,
3125 struct dt_allocation_hint *hint,
3126 struct dt_object_format *dof,
3127 struct thandle *handle)
3129 struct osd_thandle *oh;
3133 LASSERT(handle != NULL);
3135 oh = container_of0(handle, struct osd_thandle, ot_super);
3136 LASSERT(oh->ot_handle == NULL);
3138 /* EA object consumes more credits than regular object: osd_mk_index
3139 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3140 * leaf_node, could involves the block, block bitmap, groups, GDT
3141 * change for each block, so add 4 * 2 credits in that case. */
3142 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3143 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3144 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3145 /* Reuse idle OI block may cause additional one OI block
3147 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3148 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3153 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3154 attr->la_projid, 1, oh, osd_dt_obj(dt),
3155 NULL, OSD_QID_INODE);
3159 /* will help to find FID->ino mapping at dt_insert() */
3160 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3167 * Called to destroy on-disk representation of the object
3169 * Concurrency: must be locked
3171 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3174 struct osd_object *obj = osd_dt_obj(dt);
3175 struct inode *inode = obj->oo_inode;
3176 struct osd_thandle *oh;
3183 oh = container_of0(th, struct osd_thandle, ot_super);
3184 LASSERT(oh->ot_handle == NULL);
3186 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3187 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3188 /* Recycle idle OI leaf may cause additional three OI blocks
3190 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3191 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3192 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3193 /* one less inode */
3194 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3195 i_projid_read(inode), -1, oh, obj, NULL,
3199 /* data to be truncated */
3200 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3201 i_projid_read(inode), 0, oh, obj, NULL,
3206 /* will help to find FID->ino when this object is being
3207 * added to PENDING/ */
3208 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3213 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3216 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3217 struct osd_object *obj = osd_dt_obj(dt);
3218 struct inode *inode = obj->oo_inode;
3219 struct osd_device *osd = osd_obj2dev(obj);
3220 struct osd_thandle *oh;
3224 oh = container_of0(th, struct osd_thandle, ot_super);
3225 LASSERT(oh->ot_handle);
3227 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3229 if (unlikely(fid_is_acct(fid)))
3232 if (S_ISDIR(inode->i_mode)) {
3233 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3234 inode->i_nlink == 2);
3235 /* it will check/delete the inode from remote parent,
3236 * how to optimize it? unlink performance impaction XXX */
3237 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3239 CERROR("%s: delete inode "DFID": rc = %d\n",
3240 osd_name(osd), PFID(fid), result);
3242 spin_lock(&obj->oo_guard);
3244 spin_unlock(&obj->oo_guard);
3245 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3248 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3250 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3252 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3253 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3254 oh->ot_handle, OI_CHECK_FLD);
3256 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3257 /* XXX: add to ext3 orphan list */
3258 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3260 /* not needed in the cache anymore */
3261 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3262 obj->oo_destroyed = 1;
3268 * Put the fid into lustre_mdt_attrs, and then place the structure
3269 * inode's ea. This fid should not be altered during the life time
3272 * \retval +ve, on success
3273 * \retval -ve, on error
3275 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3277 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3278 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3280 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3281 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3285 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3288 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3291 lustre_loa_init(loa, fid, compat, incompat);
3292 lustre_loa_swab(loa, false);
3294 /* For the OST device with 256 bytes inode size by default,
3295 * the PFID EA will be stored together with LMA EA to avoid
3296 * performance trouble. Otherwise the PFID EA can be stored
3297 * independently. LU-8998 */
3298 if ((compat & LMAC_FID_ON_OST) &&
3299 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3300 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3301 sizeof(*loa), XATTR_CREATE);
3303 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3304 sizeof(*lma), XATTR_CREATE);
3305 /* LMA may already exist, but we need to check that all the
3306 * desired compat/incompat flags have been added. */
3307 if (unlikely(rc == -EEXIST)) {
3308 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3309 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3313 if (rc < sizeof(*lma))
3316 lustre_loa_swab(loa, true);
3317 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3318 ((compat == 0 && incompat == 0) ||
3319 (!(~lma->lma_compat & compat) &&
3320 !(~lma->lma_incompat & incompat))))
3323 lma->lma_self_fid = *fid;
3324 lma->lma_compat |= compat;
3325 lma->lma_incompat |= incompat;
3326 if (rc == sizeof(*lma)) {
3327 lustre_lma_swab(lma);
3328 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3329 sizeof(*lma), XATTR_REPLACE);
3331 lustre_loa_swab(loa, false);
3332 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3333 sizeof(*loa), XATTR_REPLACE);
3341 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3342 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3343 * To have compatilibility with 1.8 ldiskfs driver we need to have
3344 * magic number at start of fid data.
3345 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3348 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3349 const struct lu_fid *fid)
3351 if (!fid_is_namespace_visible(fid) ||
3352 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3353 param->edp_magic = 0;
3357 param->edp_magic = LDISKFS_LUFID_MAGIC;
3358 param->edp_len = sizeof(struct lu_fid) + 1;
3359 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3363 * Try to read the fid from inode ea into dt_rec.
3365 * \param fid object fid.
3367 * \retval 0 on success
3369 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3370 __u32 ino, struct lu_fid *fid,
3371 struct osd_inode_id *id)
3373 struct osd_thread_info *info = osd_oti_get(env);
3374 struct inode *inode;
3377 osd_id_gen(id, ino, OSD_OII_NOGEN);
3378 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3380 RETURN(PTR_ERR(inode));
3386 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3388 struct inode *parent_dir,
3389 const struct lu_fid *dot_fid,
3390 const struct lu_fid *dot_dot_fid,
3391 struct osd_thandle *oth)
3393 struct ldiskfs_dentry_param *dot_ldp;
3394 struct ldiskfs_dentry_param *dot_dot_ldp;
3395 __u32 saved_nlink = dir->i_nlink;
3398 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3399 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3401 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3402 dot_ldp->edp_magic = 0;
3404 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3405 dir, dot_ldp, dot_dot_ldp);
3406 /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3407 * the subseqent ref_add() will increase the dir->i_nlink
3408 * as 3. That is incorrect for new created directory.
3410 * It looks like hack, because we want to make the OSD API
3411 * to be order-independent for new created directory object
3412 * between dt_insert(..) and ref_add() operations.
3414 * Here, we only restore the in-RAM dir-inode's nlink attr,
3415 * becuase if the nlink attr is not 2, then there will be
3416 * ref_add() called following the dt_insert(..), such call
3417 * will make both the in-RAM and on-disk dir-inode's nlink
3418 * attr to be set as 2. LU-7447 */
3419 set_nlink(dir, saved_nlink);
3424 * Create an local agent inode for remote entry
3426 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3427 struct osd_device *osd,
3428 struct osd_object *pobj,
3429 const struct lu_fid *fid,
3433 struct osd_thread_info *info = osd_oti_get(env);
3434 struct inode *local;
3435 struct osd_thandle *oh;
3440 oh = container_of(th, struct osd_thandle, ot_super);
3441 LASSERT(oh->ot_handle->h_transaction != NULL);
3443 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3444 if (IS_ERR(local)) {
3445 CERROR("%s: create local error %d\n", osd_name(osd),
3446 (int)PTR_ERR(local));
3450 /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3451 * correct gid on remote file, not agent here */
3452 local->i_gid = current_fsgid();
3453 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3454 unlock_new_inode(local);
3456 /* Agent inode should not have project ID */
3457 #ifdef HAVE_PROJECT_QUOTA
3458 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL) {
3459 rc = __ldiskfs_ioctl_setproject(local, 0);
3461 CERROR("%s: quota transfer failed: rc = %d. Is project "
3462 "quota enforcement enabled on the ldiskfs "
3463 "filesystem?\n", local->i_sb->s_id, rc);
3464 RETURN(ERR_PTR(rc));
3468 /* Set special LMA flag for local agent inode */
3469 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3471 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3472 osd_name(osd), PFID(fid), rc);
3473 RETURN(ERR_PTR(rc));
3479 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3480 lu_object_fid(&pobj->oo_dt.do_lu),
3483 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3484 osd_name(osd), PFID(fid), rc);
3485 RETURN(ERR_PTR(rc));
3492 * when direntry is deleted, we have to take care of possible agent inode
3493 * referenced by that. unfortunately we can't do this at that point:
3494 * iget() within a running transaction leads to deadlock and we better do
3495 * not call that every delete declaration to save performance. so we put
3496 * a potention agent inode on a list and process that once the transaction
3497 * is over. Notice it's not any worse in terms of real orphans as regular
3498 * object destroy doesn't put inodes on the on-disk orphan list. this should
3499 * be addressed separately
3501 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3502 struct osd_thandle *oh,
3505 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3506 struct osd_obj_orphan *oor;
3513 oor->oor_env = (struct lu_env *)env;
3514 spin_lock(&osd->od_osfs_lock);
3515 list_add(&oor->oor_list, &osd->od_orphan_list);
3516 spin_unlock(&osd->od_osfs_lock);
3518 oh->ot_remove_agents = 1;
3524 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3525 struct osd_device *osd)
3527 struct osd_thread_info *info = osd_oti_get(env);
3528 struct osd_obj_orphan *oor, *tmp;
3529 struct osd_inode_id id;
3530 struct list_head list;
3531 struct inode *inode;
3536 INIT_LIST_HEAD(&list);
3538 spin_lock(&osd->od_osfs_lock);
3539 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3540 if (oor->oor_env == env) {
3541 list_del(&oor->oor_list);
3542 list_add(&oor->oor_list, &list);
3545 spin_unlock(&osd->od_osfs_lock);
3547 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3551 list_del(&oor->oor_list);
3554 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3555 inode = osd_iget_fid(info, osd, &id, &fid);
3559 if (!osd_remote_fid(env, osd, &fid)) {
3564 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3566 mark_inode_dirty(inode);
3567 ldiskfs_journal_stop(jh);
3575 * OSD layer object create function for OST objects (b=11826).
3577 * The FID is inserted into inode xattr here.
3579 * \retval 0, on success
3580 * \retval -ve, on error
3582 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3583 struct lu_attr *attr, struct dt_allocation_hint *hint,
3584 struct dt_object_format *dof, struct thandle *th)
3586 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3587 struct osd_object *obj = osd_dt_obj(dt);
3588 struct osd_thread_info *info = osd_oti_get(env);
3589 int result, on_ost = 0;
3593 if (dt_object_exists(dt))
3596 LINVRNT(osd_invariant(obj));
3597 LASSERT(!dt_object_remote(dt));
3598 LASSERT(osd_is_write_locked(env, obj));
3599 LASSERT(th != NULL);
3601 if (unlikely(fid_is_acct(fid)))
3602 /* Quota files can't be created from the kernel any more,
3603 * 'tune2fs -O quota' will take care of creating them */
3606 result = __osd_create(info, obj, attr, hint, dof, th);
3608 if (fid_is_idif(fid) &&
3609 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3610 struct lu_fid *tfid = &info->oti_fid;
3611 struct ost_id *oi = &info->oti_ostid;
3613 fid_to_ostid(fid, oi);
3614 ostid_to_fid(tfid, oi, 0);
3616 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3617 LMAC_FID_ON_OST, 0);
3619 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3621 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3622 on_ost ? LMAC_FID_ON_OST : 0,
3625 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3626 obj->oo_dt.do_body_ops = &osd_body_ops;
3630 result = __osd_oi_insert(env, obj, fid, th);
3632 /* a small optimization - dt_insert() isn't usually applied
3633 * to OST objects, so we don't need to cache OI mapping for
3635 if (result == 0 && on_ost == 0) {
3636 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3637 result = osd_idc_find_and_init(env, osd, obj);
3638 LASSERT(result == 0);
3641 LASSERT(ergo(result == 0,
3642 dt_object_exists(dt) && !dt_object_remote(dt)));
3643 LINVRNT(osd_invariant(obj));
3647 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3648 struct thandle *handle)
3650 struct osd_thandle *oh;
3652 /* it's possible that object doesn't exist yet */
3653 LASSERT(handle != NULL);
3655 oh = container_of0(handle, struct osd_thandle, ot_super);
3656 LASSERT(oh->ot_handle == NULL);
3658 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3659 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3661 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
3667 * Concurrency: @dt is write locked.
3669 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3672 struct osd_object *obj = osd_dt_obj(dt);
3673 struct inode *inode = obj->oo_inode;
3674 struct osd_thandle *oh;
3677 if (!dt_object_exists(dt) || obj->oo_destroyed)
3680 LINVRNT(osd_invariant(obj));
3681 LASSERT(!dt_object_remote(dt));
3682 LASSERT(osd_is_write_locked(env, obj));
3683 LASSERT(th != NULL);
3685 oh = container_of0(th, struct osd_thandle, ot_super);
3686 LASSERT(oh->ot_handle != NULL);
3688 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3690 CDEBUG(D_INODE, DFID" increase nlink %d\n",
3691 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3693 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3694 * (65000) subdirectories by storing "1" in i_nlink if the link count
3695 * would otherwise overflow. Directory tranversal tools understand
3696 * that (st_nlink == 1) indicates that the filesystem dose not track
3697 * hard links count on the directory, and will not abort subdirectory
3698 * scanning early once (st_nlink - 2) subdirs have been found.
3700 * This also has to properly handle the case of inodes with nlink == 0
3701 * in case they are being linked into the PENDING directory
3703 spin_lock(&obj->oo_guard);
3704 if (unlikely(inode->i_nlink == 0))
3705 /* inc_nlink from 0 may cause WARN_ON */
3706 set_nlink(inode, 1);
3708 ldiskfs_inc_count(oh->ot_handle, inode);
3709 if (!S_ISDIR(inode->i_mode))
3710 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3712 spin_unlock(&obj->oo_guard);
3714 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3715 LINVRNT(osd_invariant(obj));
3717 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3722 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
3723 struct thandle *handle)
3725 struct osd_thandle *oh;
3727 if (!dt_object_exists(dt))
3730 LASSERT(!dt_object_remote(dt));
3731 LASSERT(handle != NULL);
3733 oh = container_of0(handle, struct osd_thandle, ot_super);
3734 LASSERT(oh->ot_handle == NULL);
3736 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3737 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3743 * Concurrency: @dt is write locked.
3745 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
3748 struct osd_object *obj = osd_dt_obj(dt);
3749 struct inode *inode = obj->oo_inode;
3750 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3751 struct osd_thandle *oh;
3753 if (!dt_object_exists(dt))
3756 LINVRNT(osd_invariant(obj));
3757 LASSERT(!dt_object_remote(dt));
3758 LASSERT(osd_is_write_locked(env, obj));
3759 LASSERT(th != NULL);
3761 oh = container_of0(th, struct osd_thandle, ot_super);
3762 LASSERT(oh->ot_handle != NULL);
3764 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3766 spin_lock(&obj->oo_guard);
3767 /* That can be result of upgrade from old Lustre version and
3768 * applied only to local files. Just skip this ref_del call.
3769 * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3770 if (inode->i_nlink == 0) {
3771 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3772 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3773 ", maybe an upgraded file? (LU-3915)\n",
3774 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3775 spin_unlock(&obj->oo_guard);
3779 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3780 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3782 ldiskfs_dec_count(oh->ot_handle, inode);
3783 spin_unlock(&obj->oo_guard);
3785 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3786 LINVRNT(osd_invariant(obj));
3788 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3794 * Concurrency: @dt is read locked.
3796 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3797 struct lu_buf *buf, const char *name)
3799 struct osd_object *obj = osd_dt_obj(dt);
3800 struct inode *inode = obj->oo_inode;
3801 struct osd_thread_info *info = osd_oti_get(env);
3802 struct dentry *dentry = &info->oti_obj_dentry;
3803 bool cache_xattr = false;
3808 /* version get is not real XATTR but uses xattr API */
3809 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3810 dt_obj_version_t *ver = buf->lb_buf;
3812 /* for version we are just using xattr API but change inode
3814 if (buf->lb_len == 0)
3815 return sizeof(dt_obj_version_t);
3817 if (buf->lb_len < sizeof(dt_obj_version_t))
3820 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
3821 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3823 *ver = LDISKFS_I(inode)->i_fs_version;
3825 return sizeof(dt_obj_version_t);
3828 if (!dt_object_exists(dt))
3831 LASSERT(!dt_object_remote(dt));
3832 LASSERT(inode->i_op != NULL);
3833 LASSERT(inode->i_op->getxattr != NULL);
3835 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
3836 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
3840 rc = osd_oxc_get(obj, name, buf);
3845 rc = __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3846 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
3847 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3848 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3849 struct filter_fid *ff;
3850 struct ost_layout *ol;
3852 if (!osd_dev(dt->do_lu.lo_dev)->od_is_ost)
3855 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
3859 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
3865 if (buf->lb_len == 0 || !buf->lb_buf)
3868 if (buf->lb_len < rc)
3872 ol = &ff->ff_layout;
3873 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
3874 PFID_STRIPE_IDX_BITS);
3875 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
3876 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
3877 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
3878 if (lma->lma_compat & LMAC_COMP_INFO) {
3879 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
3880 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
3881 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
3883 ol->ol_comp_start = 0;
3884 ol->ol_comp_end = 0;
3891 if (rc == -ENOENT || rc == -ENODATA)
3892 osd_oxc_add(obj, name, NULL, 0);
3893 else if (rc > 0 && buf->lb_buf != NULL)
3894 osd_oxc_add(obj, name, buf->lb_buf, rc);
3900 static int osd_declare_xattr_set(const struct lu_env *env,
3901 struct dt_object *dt,
3902 const struct lu_buf *buf, const char *name,
3903 int fl, struct thandle *handle)
3905 struct osd_thandle *oh;
3907 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3909 LASSERT(handle != NULL);
3911 oh = container_of0(handle, struct osd_thandle, ot_super);
3912 LASSERT(oh->ot_handle == NULL);
3914 if (strcmp(name, XATTR_NAME_LMA) == 0) {
3915 /* For non-upgrading case, the LMA is set first and
3916 * usually fit inode. But for upgrade case, the LMA
3917 * may be in another separated EA block. */
3918 if (dt_object_exists(dt)) {
3919 if (fl == LU_XATTR_REPLACE)
3924 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3926 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
3927 /* We may need to delete the old PFID EA. */
3928 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
3929 if (fl == LU_XATTR_REPLACE)
3936 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
3941 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3942 /* learn xattr size from osd_xattr_get if
3943 attribute has not been read yet */
3944 buflen = __osd_xattr_get(
3945 osd_dt_obj(dt)->oo_inode,
3946 &osd_oti_get(env)->oti_obj_dentry,
3951 buflen = buf->lb_len;
3954 if (buflen > sb->s_blocksize) {
3955 credits += osd_calc_bkmap_credits(
3957 (buflen + sb->s_blocksize - 1) >>
3958 sb->s_blocksize_bits);
3962 * xattr set may involve inode quota change, reserve credits for
3963 * dquot_initialize()
3965 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3968 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3974 * Concurrency: @dt is write locked.
3976 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3977 const struct lu_buf *buf, const char *name, int fl,
3978 struct thandle *handle)
3980 struct osd_object *obj = osd_dt_obj(dt);
3981 struct inode *inode = obj->oo_inode;
3982 struct osd_thread_info *info = osd_oti_get(env);
3983 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3984 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3993 /* version set is not real XATTR */
3994 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3995 dt_obj_version_t *version = buf->lb_buf;
3997 /* for version we are just using xattr API but change inode
3999 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4001 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4002 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4004 LDISKFS_I(inode)->i_fs_version = *version;
4005 /* Version is set after all inode operations are finished,
4006 * so we should mark it dirty here */
4007 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4012 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4013 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4016 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4017 if (fl & LU_XATTR_REPLACE)
4018 fs_flags |= XATTR_REPLACE;
4020 if (fl & LU_XATTR_CREATE)
4021 fs_flags |= XATTR_CREATE;
4023 /* For the OST device with 256 bytes inode size by default,
4024 * the PFID EA will be stored together with LMA EA to avoid
4025 * performance trouble. Otherwise the PFID EA can be stored
4026 * independently. LU-8998 */
4027 if (strcmp(name, XATTR_NAME_FID) == 0 &&
4028 LDISKFS_INODE_SIZE(inode->i_sb) <= 256) {
4029 struct dentry *dentry = &info->oti_obj_dentry;
4030 struct filter_fid *ff;
4031 struct ost_layout *ol;
4034 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4037 ol = &ff->ff_layout;
4038 /* Old client does not send stripe information, store
4039 * the PFID EA on disk directly. */
4040 if (buf->lb_len == sizeof(struct lu_fid) ||
4041 ol->ol_stripe_size == 0) {
4042 len = sizeof(struct lu_fid);
4046 if (buf->lb_len != sizeof(*ff))
4049 rc = osd_get_lma(info, inode, dentry, loa);
4050 if (unlikely(rc == -ENODATA)) {
4051 /* Usually for upgarding from old device */
4052 lustre_loa_init(loa, lu_object_fid(&dt->do_lu),
4053 LMAC_FID_ON_OST, 0);
4061 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4062 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4063 PFID_STRIPE_IDX_BITS;
4064 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4065 lma->lma_compat |= LMAC_STRIPE_INFO;
4066 if (ol->ol_comp_id != 0) {
4067 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4068 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4069 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4070 lma->lma_compat |= LMAC_COMP_INFO;
4073 lustre_loa_swab(loa, false);
4075 /* Remove old PFID EA entry firstly. */
4076 ll_vfs_dq_init(inode);
4077 rc = inode->i_op->removexattr(dentry, name);
4078 if (rc && rc != -ENODATA)
4081 /* Store the PFID EA inside the LMA EA. */
4082 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4086 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4087 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4091 lma->lma_incompat |= LMAI_STRIPED;
4092 lustre_lma_swab(lma);
4093 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4094 sizeof(*lma), XATTR_REPLACE);
4100 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4101 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4104 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4105 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4106 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4112 * Concurrency: @dt is read locked.
4114 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4115 const struct lu_buf *buf)
4117 struct osd_object *obj = osd_dt_obj(dt);
4118 struct inode *inode = obj->oo_inode;
4119 struct osd_thread_info *info = osd_oti_get(env);
4120 struct dentry *dentry = &info->oti_obj_dentry;
4122 if (!dt_object_exists(dt))
4125 LASSERT(!dt_object_remote(dt));
4126 LASSERT(inode->i_op != NULL);
4127 LASSERT(inode->i_op->listxattr != NULL);
4129 dentry->d_inode = inode;
4130 dentry->d_sb = inode->i_sb;
4131 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4134 static int osd_declare_xattr_del(const struct lu_env *env,
4135 struct dt_object *dt, const char *name,
4136 struct thandle *handle)
4138 struct osd_thandle *oh;
4139 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4141 LASSERT(!dt_object_remote(dt));
4142 LASSERT(handle != NULL);
4144 oh = container_of0(handle, struct osd_thandle, ot_super);
4145 LASSERT(oh->ot_handle == NULL);
4147 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4148 osd_dto_credits_noquota[DTO_XATTR_SET]);
4150 * xattr del may involve inode quota change, reserve credits for
4151 * dquot_initialize()
4153 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4159 * Concurrency: @dt is write locked.
4161 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4162 const char *name, struct thandle *handle)
4164 struct osd_object *obj = osd_dt_obj(dt);
4165 struct inode *inode = obj->oo_inode;
4166 struct osd_thread_info *info = osd_oti_get(env);
4167 struct dentry *dentry = &info->oti_obj_dentry;
4170 if (!dt_object_exists(dt))
4173 LASSERT(!dt_object_remote(dt));
4174 LASSERT(inode->i_op != NULL);
4175 LASSERT(inode->i_op->removexattr != NULL);
4176 LASSERT(handle != NULL);
4178 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4180 ll_vfs_dq_init(inode);
4181 dentry->d_inode = inode;
4182 dentry->d_sb = inode->i_sb;
4183 rc = inode->i_op->removexattr(dentry, name);
4184 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
4185 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4187 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4189 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4190 &info->oti_ost_attrs);
4192 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
4197 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4198 lustre_lma_swab(lma);
4199 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4200 sizeof(*lma), XATTR_REPLACE);
4205 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4208 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4209 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4210 osd_oxc_del(obj, name);
4215 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4216 __u64 start, __u64 end)
4218 struct osd_object *obj = osd_dt_obj(dt);
4219 struct inode *inode = obj->oo_inode;
4220 struct osd_thread_info *info = osd_oti_get(env);
4221 struct dentry *dentry = &info->oti_obj_dentry;
4222 struct file *file = &info->oti_file;
4227 dentry->d_inode = inode;
4228 dentry->d_sb = inode->i_sb;
4229 file->f_path.dentry = dentry;
4230 file->f_mapping = inode->i_mapping;
4231 file->f_op = inode->i_fop;
4232 set_file_inode(file, inode);
4234 rc = ll_vfs_fsync_range(file, start, end, 0);
4239 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4248 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4249 const struct dt_index_features *feat)
4251 struct iam_descr *descr;
4253 if (osd_object_is_root(o))
4254 return feat == &dt_directory_features;
4256 LASSERT(o->oo_dir != NULL);
4258 descr = o->oo_dir->od_container.ic_descr;
4259 if (feat == &dt_directory_features) {
4260 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4266 feat->dif_keysize_min <= descr->id_key_size &&
4267 descr->id_key_size <= feat->dif_keysize_max &&
4268 feat->dif_recsize_min <= descr->id_rec_size &&
4269 descr->id_rec_size <= feat->dif_recsize_max &&
4270 !(feat->dif_flags & (DT_IND_VARKEY |
4271 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4272 ergo(feat->dif_flags & DT_IND_UPDATE,
4273 1 /* XXX check that object (and file system) is
4278 static int osd_iam_container_init(const struct lu_env *env,
4279 struct osd_object *obj,
4280 struct osd_directory *dir)
4282 struct iam_container *bag = &dir->od_container;
4285 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4289 result = iam_container_setup(bag);
4291 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4293 iam_container_fini(bag);
4300 * Concurrency: no external locking is necessary.
4302 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4303 const struct dt_index_features *feat)
4307 struct osd_object *obj = osd_dt_obj(dt);
4309 LINVRNT(osd_invariant(obj));
4311 if (osd_object_is_root(obj)) {
4312 dt->do_index_ops = &osd_index_ea_ops;
4314 } else if (feat == &dt_directory_features) {
4315 dt->do_index_ops = &osd_index_ea_ops;
4316 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4321 } else if (unlikely(feat == &dt_otable_features)) {
4322 dt->do_index_ops = &osd_otable_ops;
4324 } else if (unlikely(feat == &dt_acct_features)) {
4325 dt->do_index_ops = &osd_acct_index_ops;
4328 } else if (!osd_has_index(obj)) {
4329 struct osd_directory *dir;
4334 spin_lock(&obj->oo_guard);
4335 if (obj->oo_dir == NULL)
4339 * Concurrent thread allocated container data.
4342 spin_unlock(&obj->oo_guard);
4344 * Now, that we have container data, serialize its
4347 down_write(&obj->oo_ext_idx_sem);
4349 * recheck under lock.
4351 if (!osd_has_index(obj))
4352 result = osd_iam_container_init(env, obj,
4356 up_write(&obj->oo_ext_idx_sem);
4364 if (result == 0 && skip_iam == 0) {
4365 if (!osd_iam_index_probe(env, obj, feat))
4368 LINVRNT(osd_invariant(obj));
4373 static int osd_otable_it_attr_get(const struct lu_env *env,
4374 struct dt_object *dt,
4375 struct lu_attr *attr)
4381 static const struct dt_object_operations osd_obj_ops = {
4382 .do_read_lock = osd_read_lock,
4383 .do_write_lock = osd_write_lock,
4384 .do_read_unlock = osd_read_unlock,
4385 .do_write_unlock = osd_write_unlock,
4386 .do_write_locked = osd_write_locked,
4387 .do_attr_get = osd_attr_get,
4388 .do_declare_attr_set = osd_declare_attr_set,
4389 .do_attr_set = osd_attr_set,
4390 .do_ah_init = osd_ah_init,
4391 .do_declare_create = osd_declare_create,
4392 .do_create = osd_create,
4393 .do_declare_destroy = osd_declare_destroy,
4394 .do_destroy = osd_destroy,
4395 .do_index_try = osd_index_try,
4396 .do_declare_ref_add = osd_declare_ref_add,
4397 .do_ref_add = osd_ref_add,
4398 .do_declare_ref_del = osd_declare_ref_del,
4399 .do_ref_del = osd_ref_del,
4400 .do_xattr_get = osd_xattr_get,
4401 .do_declare_xattr_set = osd_declare_xattr_set,
4402 .do_xattr_set = osd_xattr_set,
4403 .do_declare_xattr_del = osd_declare_xattr_del,
4404 .do_xattr_del = osd_xattr_del,
4405 .do_xattr_list = osd_xattr_list,
4406 .do_object_sync = osd_object_sync,
4407 .do_invalidate = osd_invalidate,
4410 static const struct dt_object_operations osd_obj_otable_it_ops = {
4411 .do_attr_get = osd_otable_it_attr_get,
4412 .do_index_try = osd_index_try,
4415 static int osd_index_declare_iam_delete(const struct lu_env *env,
4416 struct dt_object *dt,
4417 const struct dt_key *key,
4418 struct thandle *handle)
4420 struct osd_thandle *oh;
4422 oh = container_of0(handle, struct osd_thandle, ot_super);
4423 LASSERT(oh->ot_handle == NULL);
4425 /* Recycle may cause additional three blocks to be changed. */
4426 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4427 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4433 * delete a (key, value) pair from index \a dt specified by \a key
4435 * \param dt osd index object
4436 * \param key key for index
4437 * \param rec record reference
4438 * \param handle transaction handler
4441 * \retval -ve failure
4443 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4444 const struct dt_key *key,
4445 struct thandle *handle)
4447 struct osd_thread_info *oti = osd_oti_get(env);
4448 struct osd_object *obj = osd_dt_obj(dt);
4449 struct osd_thandle *oh;
4450 struct iam_path_descr *ipd;
4451 struct iam_container *bag = &obj->oo_dir->od_container;
4455 if (!dt_object_exists(dt))
4458 LINVRNT(osd_invariant(obj));
4459 LASSERT(!dt_object_remote(dt));
4460 LASSERT(bag->ic_object == obj->oo_inode);
4461 LASSERT(handle != NULL);
4463 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4465 ipd = osd_idx_ipd_get(env, bag);
4466 if (unlikely(ipd == NULL))
4469 oh = container_of0(handle, struct osd_thandle, ot_super);
4470 LASSERT(oh->ot_handle != NULL);
4471 LASSERT(oh->ot_handle->h_transaction != NULL);
4473 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4474 /* swab quota uid/gid provided by caller */
4475 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4476 key = (const struct dt_key *)&oti->oti_quota_id;
4479 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4480 osd_ipd_put(env, bag, ipd);
4481 LINVRNT(osd_invariant(obj));
4482 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4486 static int osd_index_declare_ea_delete(const struct lu_env *env,
4487 struct dt_object *dt,
4488 const struct dt_key *key,
4489 struct thandle *handle)
4491 struct osd_thandle *oh;
4492 struct inode *inode;
4496 LASSERT(!dt_object_remote(dt));
4497 LASSERT(handle != NULL);
4499 oh = container_of0(handle, struct osd_thandle, ot_super);
4500 LASSERT(oh->ot_handle == NULL);
4502 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4503 if (key != NULL && unlikely(strcmp((char *)key, dotdot) == 0)) {
4504 /* '..' to a remote object has a local representative */
4505 credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
4506 /* to reset LMAI_REMOTE_PARENT */
4509 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4511 inode = osd_dt_obj(dt)->oo_inode;
4515 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4516 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
4521 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
4524 struct osd_fid_pack *rec;
4527 if (de->file_type & LDISKFS_DIRENT_LUFID) {
4528 rec = (struct osd_fid_pack *) (de->name + de->name_len + 1);
4529 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
4530 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
4536 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
4537 const struct lu_fid *fid)
4539 struct seq_server_site *ss = osd_seq_site(osd);
4542 /* FID seqs not in FLDB, must be local seq */
4543 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
4546 /* If FLD is not being initialized yet, it only happens during the
4547 * initialization, likely during mgs initialization, and we assume
4548 * this is local FID. */
4549 if (ss == NULL || ss->ss_server_fld == NULL)
4552 /* Only check the local FLDB here */
4553 if (osd_seq_exists(env, osd, fid_seq(fid)))
4559 static void osd_take_care_of_agent(const struct lu_env *env,
4560 struct osd_device *osd,
4561 struct osd_thandle *oh,
4562 struct ldiskfs_dir_entry_2 *de)
4564 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4565 struct osd_idmap_cache *idc;
4566 int rc, schedule = 0;
4568 LASSERT(de != NULL);
4570 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
4571 if (likely(rc == 0)) {
4572 idc = osd_idc_find_or_init(env, osd, fid);
4573 if (IS_ERR(idc) || idc->oic_remote)
4575 } else if (rc == -ENODATA) {
4576 /* can't get FID, postpone to the end of the
4577 * transaction when iget() is safe */
4580 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
4583 osd_schedule_agent_inode_removal(env, oh,
4584 le32_to_cpu(de->inode));
4588 * Index delete function for interoperability mode (b11826).
4589 * It will remove the directory entry added by osd_index_ea_insert().
4590 * This entry is needed to maintain name->fid mapping.
4592 * \param key, key i.e. file entry to be deleted
4594 * \retval 0, on success
4595 * \retval -ve, on error
4597 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
4598 const struct dt_key *key, struct thandle *handle)
4600 struct osd_object *obj = osd_dt_obj(dt);
4601 struct inode *dir = obj->oo_inode;
4602 struct dentry *dentry;
4603 struct osd_thandle *oh;
4604 struct ldiskfs_dir_entry_2 *de = NULL;
4605 struct buffer_head *bh;
4606 struct htree_lock *hlock = NULL;
4607 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4608 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4612 if (!dt_object_exists(dt))
4615 LINVRNT(osd_invariant(obj));
4616 LASSERT(!dt_object_remote(dt));
4617 LASSERT(handle != NULL);
4619 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4621 oh = container_of(handle, struct osd_thandle, ot_super);
4622 LASSERT(oh->ot_handle != NULL);
4623 LASSERT(oh->ot_handle->h_transaction != NULL);
4625 ll_vfs_dq_init(dir);
4626 dentry = osd_child_dentry_get(env, obj,
4627 (char *)key, strlen((char *)key));
4629 if (obj->oo_hl_head != NULL) {
4630 hlock = osd_oti_get(env)->oti_hlock;
4631 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
4632 dir, LDISKFS_HLOCK_DEL);
4634 down_write(&obj->oo_ext_idx_sem);
4637 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
4639 /* If this is not the ".." entry, it might be a remote DNE
4640 * entry and we need to check if the FID is for a remote
4641 * MDT. If the FID is not in the directory entry (e.g.
4642 * upgraded 1.8 filesystem without dirdata enabled) then
4643 * we need to get the FID from the LMA. For a remote directory
4644 * there HAS to be an LMA, it cannot be an IGIF inode in this
4647 * Delete the entry before the agent inode in order to
4648 * simplify error handling. At worst an error after deleting
4649 * the entry first might leak the agent inode afterward. The
4650 * reverse would need filesystem abort in case of error deleting
4651 * the entry after the agent had been removed, or leave a
4652 * dangling entry pointing at a random inode. */
4653 if (strcmp((char *)key, dotdot) != 0)
4654 osd_take_care_of_agent(env, osd, oh, de);
4655 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
4661 ldiskfs_htree_unlock(hlock);
4663 up_write(&obj->oo_ext_idx_sem);
4668 /* For inode on the remote MDT, .. will point to
4669 * /Agent directory, Check whether it needs to delete
4670 * from agent directory */
4671 if (unlikely(strcmp((char *)key, dotdot) == 0)) {
4674 ret = osd_delete_from_remote_parent(env, osd_obj2dev(obj),
4677 /* Sigh, the entry has been deleted, and
4678 * it is not easy to revert it back, so
4679 * let's keep this error private, and let
4680 * LFSCK fix it. XXX */
4681 CERROR("%s: delete remote parent "DFID": rc = %d\n",
4682 osd_name(osd), PFID(fid), ret);
4685 LASSERT(osd_invariant(obj));
4686 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4691 * Lookup index for \a key and copy record to \a rec.
4693 * \param dt osd index object
4694 * \param key key for index
4695 * \param rec record reference
4697 * \retval +ve success : exact mach
4698 * \retval 0 return record with key not greater than \a key
4699 * \retval -ve failure
4701 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
4702 struct dt_rec *rec, const struct dt_key *key)
4704 struct osd_object *obj = osd_dt_obj(dt);
4705 struct iam_path_descr *ipd;
4706 struct iam_container *bag = &obj->oo_dir->od_container;
4707 struct osd_thread_info *oti = osd_oti_get(env);
4708 struct iam_iterator *it = &oti->oti_idx_it;
4709 struct iam_rec *iam_rec;
4713 if (!dt_object_exists(dt))
4716 LASSERT(osd_invariant(obj));
4717 LASSERT(!dt_object_remote(dt));
4718 LASSERT(bag->ic_object == obj->oo_inode);
4720 ipd = osd_idx_ipd_get(env, bag);
4724 /* got ipd now we can start iterator. */
4725 iam_it_init(it, bag, 0, ipd);
4727 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4728 /* swab quota uid/gid provided by caller */
4729 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4730 key = (const struct dt_key *)&oti->oti_quota_id;
4733 rc = iam_it_get(it, (struct iam_key *)key);
4735 if (S_ISDIR(obj->oo_inode->i_mode))
4736 iam_rec = (struct iam_rec *)oti->oti_ldp;
4738 iam_rec = (struct iam_rec *) rec;
4740 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
4742 if (S_ISDIR(obj->oo_inode->i_mode))
4743 osd_fid_unpack((struct lu_fid *) rec,
4744 (struct osd_fid_pack *)iam_rec);
4745 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
4746 osd_quota_unpack(obj, rec);
4751 osd_ipd_put(env, bag, ipd);
4753 LINVRNT(osd_invariant(obj));
4758 static int osd_index_declare_iam_insert(const struct lu_env *env,
4759 struct dt_object *dt,
4760 const struct dt_rec *rec,
4761 const struct dt_key *key,
4762 struct thandle *handle)
4764 struct osd_thandle *oh;
4766 LASSERT(handle != NULL);
4768 oh = container_of0(handle, struct osd_thandle, ot_super);
4769 LASSERT(oh->ot_handle == NULL);
4771 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
4772 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
4778 * Inserts (key, value) pair in \a dt index object.
4780 * \param dt osd index object
4781 * \param key key for index
4782 * \param rec record reference
4783 * \param th transaction handler
4786 * \retval -ve failure
4788 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
4789 const struct dt_rec *rec,
4790 const struct dt_key *key, struct thandle *th,
4793 struct osd_object *obj = osd_dt_obj(dt);
4794 struct iam_path_descr *ipd;
4795 struct osd_thandle *oh;
4796 struct iam_container *bag;
4797 struct osd_thread_info *oti = osd_oti_get(env);
4798 struct iam_rec *iam_rec;
4802 if (!dt_object_exists(dt))
4805 LINVRNT(osd_invariant(obj));
4806 LASSERT(!dt_object_remote(dt));
4808 bag = &obj->oo_dir->od_container;
4809 LASSERT(bag->ic_object == obj->oo_inode);
4810 LASSERT(th != NULL);
4812 osd_trans_exec_op(env, th, OSD_OT_INSERT);
4814 ipd = osd_idx_ipd_get(env, bag);
4815 if (unlikely(ipd == NULL))
4818 oh = container_of0(th, struct osd_thandle, ot_super);
4819 LASSERT(oh->ot_handle != NULL);
4820 LASSERT(oh->ot_handle->h_transaction != NULL);
4821 if (S_ISDIR(obj->oo_inode->i_mode)) {
4822 iam_rec = (struct iam_rec *)oti->oti_ldp;
4823 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec, &oti->oti_fid);
4824 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4825 /* pack quota uid/gid */
4826 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4827 key = (const struct dt_key *)&oti->oti_quota_id;
4828 /* pack quota record */
4829 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
4830 iam_rec = (struct iam_rec *)rec;
4832 iam_rec = (struct iam_rec *)rec;
4835 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
4837 osd_ipd_put(env, bag, ipd);
4838 LINVRNT(osd_invariant(obj));
4839 osd_trans_exec_check(env, th, OSD_OT_INSERT);
4844 * Calls ldiskfs_add_entry() to add directory entry
4845 * into the directory. This is required for
4846 * interoperability mode (b11826)
4848 * \retval 0, on success
4849 * \retval -ve, on error
4851 static int __osd_ea_add_rec(struct osd_thread_info *info,
4852 struct osd_object *pobj, struct inode *cinode,
4853 const char *name, const struct lu_fid *fid,
4854 struct htree_lock *hlock, struct thandle *th)
4856 struct ldiskfs_dentry_param *ldp;
4857 struct dentry *child;
4858 struct osd_thandle *oth;
4861 oth = container_of(th, struct osd_thandle, ot_super);
4862 LASSERT(oth->ot_handle != NULL);
4863 LASSERT(oth->ot_handle->h_transaction != NULL);
4864 LASSERT(pobj->oo_inode);
4866 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4867 if (unlikely(pobj->oo_inode ==
4868 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
4871 osd_get_ldiskfs_dirent_param(ldp, fid);
4872 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
4873 child->d_fsdata = (void *)ldp;
4874 ll_vfs_dq_init(pobj->oo_inode);
4875 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
4876 child, cinode, hlock);
4877 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
4878 struct ldiskfs_dir_entry_2 *de;
4879 struct buffer_head *bh;
4882 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
4885 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
4888 if (S_ISDIR(cinode->i_mode))
4889 de->file_type = LDISKFS_DIRENT_LUFID |
4890 LDISKFS_FT_REG_FILE;
4892 de->file_type = LDISKFS_DIRENT_LUFID |
4894 ldiskfs_handle_dirty_metadata(oth->ot_handle,
4905 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
4906 * into the directory.Also sets flags into osd object to
4907 * indicate dot and dotdot are created. This is required for
4908 * interoperability mode (b11826)
4910 * \param dir directory for dot and dotdot fixup.
4911 * \param obj child object for linking
4913 * \retval 0, on success
4914 * \retval -ve, on error
4916 static int osd_add_dot_dotdot(struct osd_thread_info *info,
4917 struct osd_object *dir,
4918 struct inode *parent_dir, const char *name,
4919 const struct lu_fid *dot_fid,
4920 const struct lu_fid *dot_dot_fid,
4923 struct inode *inode = dir->oo_inode;
4924 struct osd_thandle *oth;
4927 oth = container_of(th, struct osd_thandle, ot_super);
4928 LASSERT(oth->ot_handle->h_transaction != NULL);
4929 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
4931 if (strcmp(name, dot) == 0) {
4932 if (dir->oo_compat_dot_created) {
4935 LASSERT(inode->i_ino == parent_dir->i_ino);
4936 dir->oo_compat_dot_created = 1;
4939 } else if (strcmp(name, dotdot) == 0) {
4940 if (!dir->oo_compat_dot_created)
4942 /* in case of rename, dotdot is already created */
4943 if (dir->oo_compat_dotdot_created) {
4944 return __osd_ea_add_rec(info, dir, parent_dir, name,
4945 dot_dot_fid, NULL, th);
4948 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
4949 struct lu_fid tfid = *dot_dot_fid;
4952 result = osd_add_dot_dotdot_internal(info,
4953 dir->oo_inode, parent_dir, dot_fid,
4956 result = osd_add_dot_dotdot_internal(info,
4957 dir->oo_inode, parent_dir, dot_fid,
4962 dir->oo_compat_dotdot_created = 1;
4970 * It will call the appropriate osd_add* function and return the
4971 * value, return by respective functions.
4973 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
4974 struct inode *cinode, const char *name,
4975 const struct lu_fid *fid, struct thandle *th)
4977 struct osd_thread_info *info = osd_oti_get(env);
4978 struct htree_lock *hlock;
4981 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
4983 if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' &&
4985 if (hlock != NULL) {
4986 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4989 down_write(&pobj->oo_ext_idx_sem);
4992 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
4993 lu_object_fid(&pobj->oo_dt.do_lu),
4996 if (hlock != NULL) {
4997 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4998 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5000 down_write(&pobj->oo_ext_idx_sem);
5003 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5004 struct lu_fid *tfid = &info->oti_fid;
5008 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5011 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5016 ldiskfs_htree_unlock(hlock);
5018 up_write(&pobj->oo_ext_idx_sem);
5024 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5025 struct osd_idmap_cache *oic)
5027 struct osd_scrub *scrub = &dev->od_scrub;
5028 struct lu_fid *fid = &oic->oic_fid;
5029 struct osd_inode_id *id = &oic->oic_lid;
5030 struct inode *inode = NULL;
5036 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5039 if (scrub->os_pos_current > id->oii_ino)
5043 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5044 if (rc == -ENOENT) {
5045 __u32 gen = id->oii_gen;
5051 inode = osd_iget(oti, dev, id);
5052 /* The inode has been removed (by race maybe). */
5053 if (IS_ERR(inode)) {
5054 rc = PTR_ERR(inode);
5056 RETURN(rc == -ESTALE ? -ENOENT : rc);
5059 /* The OI mapping is lost. */
5060 if (gen != OSD_OII_NOGEN)
5063 /* The inode may has been reused by others, we do not know,
5064 * leave it to be handled by subsequent osd_fid_lookup(). */
5066 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5073 if (thread_is_running(&scrub->os_thread)) {
5074 if (inode == NULL) {
5075 inode = osd_iget(oti, dev, id);
5076 /* The inode has been removed (by race maybe). */
5077 if (IS_ERR(inode)) {
5078 rc = PTR_ERR(inode);
5080 RETURN(rc == -ESTALE ? -ENOENT : rc);
5084 rc = osd_oii_insert(dev, oic, insert);
5085 /* There is race condition between osd_oi_lookup and OI scrub.
5086 * The OI scrub finished just after osd_oi_lookup() failure.
5087 * Under such case, it is unnecessary to trigger OI scrub again,
5088 * but try to call osd_oi_lookup() again. */
5089 if (unlikely(rc == -EAGAIN))
5092 if (!S_ISDIR(inode->i_mode))
5095 rc = osd_check_lmv(oti, dev, inode, oic);
5100 if (!dev->od_noscrub && ++once == 1) {
5101 rc = osd_scrub_start(dev, SS_AUTO_PARTIAL | SS_CLEAR_DRYRUN |
5103 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5104 "%s: trigger partial OI scrub for RPC inconsistency "
5105 "checking FID "DFID": rc = %d\n",
5106 osd_dev2name(dev), PFID(fid), rc);
5107 if (rc == 0 || rc == -EALREADY)
5120 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5121 struct osd_device *dev,
5122 struct osd_idmap_cache *oic,
5123 struct lu_fid *fid, __u32 ino)
5125 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5126 struct inode *inode;
5129 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5130 inode = osd_iget(oti, dev, &oic->oic_lid);
5131 if (IS_ERR(inode)) {
5132 fid_zero(&oic->oic_fid);
5133 return PTR_ERR(inode);
5136 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5139 fid_zero(&oic->oic_fid);
5141 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5145 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5146 struct osd_inode_id *id, const struct lu_fid *fid)
5148 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5149 id->oii_ino, id->oii_gen, info);
5150 info->oti_cache.oic_lid = *id;
5151 info->oti_cache.oic_fid = *fid;
5152 info->oti_cache.oic_dev = osd;
5156 * Get parent FID from the linkEA.
5158 * For a directory which parent resides on remote MDT, to satisfy the
5159 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5160 * the other hand, to make the lookup(..) on the directory can return
5161 * the real parent FID, we append the real parent FID after its ".."
5162 * name entry in the /REMOTE_PARENT_DIR.
5164 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5165 * backup. So after the restore, we cannot get the right parent FID from
5166 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5167 * we have stored the real parent FID in the directory object's linkEA,
5168 * we can parse the linkEA for the real parent FID.
5170 * \param[in] env pointer to the thread context
5171 * \param[in] obj pointer to the object to be handled
5172 * \param[out]fid pointer to the buffer to hold the parent FID
5174 * \retval 0 for getting the real parent FID successfully
5175 * \retval negative error number on failure
5177 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5178 struct osd_object *obj,
5181 struct osd_thread_info *oti = osd_oti_get(env);
5182 struct lu_buf *buf = &oti->oti_big_buf;
5183 struct dentry *dentry = &oti->oti_obj_dentry;
5184 struct inode *inode = obj->oo_inode;
5185 struct linkea_data ldata = { NULL };
5190 if (!S_ISDIR(inode->i_mode))
5194 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5195 buf->lb_buf, buf->lb_len);
5196 if (rc == -ERANGE) {
5197 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5200 lu_buf_realloc(buf, rc);
5201 if (buf->lb_buf == NULL)
5208 if (unlikely(rc == 0))
5214 if (unlikely(buf->lb_buf == NULL)) {
5215 lu_buf_realloc(buf, rc);
5216 if (buf->lb_buf == NULL)
5223 rc = linkea_init_with_rec(&ldata);
5225 linkea_first_entry(&ldata);
5226 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5232 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5233 struct inode *inode,
5234 const struct lu_fid *pfid,
5235 const char *name, const int namelen)
5237 struct osd_thread_info *oti = osd_oti_get(env);
5238 struct lu_buf *buf = &oti->oti_big_buf;
5239 struct dentry *dentry = &oti->oti_obj_dentry;
5240 struct linkea_data ldata = { NULL };
5241 struct lu_name cname = { .ln_name = name,
5242 .ln_namelen = namelen };
5247 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5248 buf->lb_buf, buf->lb_len);
5250 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5255 if (unlikely(rc == 0))
5258 if (buf->lb_len < rc) {
5259 lu_buf_realloc(buf, rc);
5260 if (buf->lb_buf == NULL)
5267 rc = linkea_init_with_rec(&ldata);
5269 rc = linkea_links_find(&ldata, &cname, pfid);
5275 * Calls ->lookup() to find dentry. From dentry get inode and
5276 * read inode's ea to get fid. This is required for interoperability
5279 * \retval 0, on success
5280 * \retval -ve, on error
5282 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5283 struct dt_rec *rec, const struct dt_key *key)
5285 struct inode *dir = obj->oo_inode;
5286 struct dentry *dentry;
5287 struct ldiskfs_dir_entry_2 *de;
5288 struct buffer_head *bh;
5289 struct lu_fid *fid = (struct lu_fid *) rec;
5290 struct htree_lock *hlock = NULL;
5295 LASSERT(dir->i_op != NULL);
5296 LASSERT(dir->i_op->lookup != NULL);
5298 dentry = osd_child_dentry_get(env, obj,
5299 (char *)key, strlen((char *)key));
5301 if (obj->oo_hl_head != NULL) {
5302 hlock = osd_oti_get(env)->oti_hlock;
5303 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5304 dir, LDISKFS_HLOCK_LOOKUP);
5306 down_read(&obj->oo_ext_idx_sem);
5309 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5311 struct osd_thread_info *oti = osd_oti_get(env);
5312 struct osd_inode_id *id = &oti->oti_id;
5313 struct osd_idmap_cache *oic = &oti->oti_cache;
5314 struct osd_device *dev = osd_obj2dev(obj);
5316 ino = le32_to_cpu(de->inode);
5317 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5319 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5323 rc = osd_get_fid_from_dentry(de, rec);
5325 /* done with de, release bh */
5328 if (unlikely(ino == osd_remote_parent_ino(dev))) {
5329 const char *name = (const char *)key;
5331 /* If the parent is on remote MDT, and there
5332 * is no FID-in-dirent, then we have to get
5333 * the parent FID from the linkEA. */
5334 if (likely(strlen(name) == 2 &&
5335 name[0] == '.' && name[1] == '.'))
5336 rc = osd_get_pfid_from_linkea(env, obj,
5339 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5342 osd_id_gen(id, ino, OSD_OII_NOGEN);
5345 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5346 fid_zero(&oic->oic_fid);
5351 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5352 rc = osd_consistency_check(oti, dev, oic);
5354 fid_zero(&oic->oic_fid);
5363 ldiskfs_htree_unlock(hlock);
5365 up_read(&obj->oo_ext_idx_sem);
5369 static int osd_index_declare_ea_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;
5376 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5377 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5378 const struct lu_fid *fid = rec1->rec_fid;
5379 int credits, rc = 0;
5380 struct osd_idmap_cache *idc;
5383 LASSERT(!dt_object_remote(dt));
5384 LASSERT(handle != NULL);
5385 LASSERT(fid != NULL);
5386 LASSERT(rec1->rec_type != 0);
5388 oh = container_of0(handle, struct osd_thandle, ot_super);
5389 LASSERT(oh->ot_handle == NULL);
5391 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5393 /* we can't call iget() while a transactions is running
5394 * (this can lead to a deadlock), but we need to know
5395 * inum and object type. so we find this information at
5396 * declaration and cache in per-thread info */
5397 idc = osd_idc_find_or_init(env, osd, fid);
5399 RETURN(PTR_ERR(idc));
5400 if (idc->oic_remote) {
5401 /* a reference to remote inode is represented by an
5402 * agent inode which we have to create */
5403 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5404 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5407 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5409 if (osd_dt_obj(dt)->oo_inode != NULL) {
5410 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5412 /* We ignore block quota on meta pool (MDTs), so needn't
5413 * calculate how many blocks will be consumed by this index
5415 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5417 i_projid_read(inode), 0,
5418 oh, osd_dt_obj(dt), NULL,
5426 * Index add function for interoperability mode (b11826).
5427 * It will add the directory entry.This entry is needed to
5428 * maintain name->fid mapping.
5430 * \param key it is key i.e. file entry to be inserted
5431 * \param rec it is value of given key i.e. fid
5433 * \retval 0, on success
5434 * \retval -ve, on error
5436 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5437 const struct dt_rec *rec,
5438 const struct dt_key *key, struct thandle *th,
5441 struct osd_object *obj = osd_dt_obj(dt);
5442 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5443 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5444 const struct lu_fid *fid = rec1->rec_fid;
5445 const char *name = (const char *)key;
5446 struct osd_thread_info *oti = osd_oti_get(env);
5447 struct inode *child_inode = NULL;
5448 struct osd_idmap_cache *idc;
5452 if (!dt_object_exists(dt))
5455 LASSERT(osd_invariant(obj));
5456 LASSERT(!dt_object_remote(dt));
5457 LASSERT(th != NULL);
5459 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5461 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5463 idc = osd_idc_find(env, osd, fid);
5464 if (unlikely(idc == NULL)) {
5465 /* this dt_insert() wasn't declared properly, so
5466 * FID is missing in OI cache. we better do not
5467 * lookup FID in FLDB/OI and don't risk to deadlock,
5468 * but in some special cases (lfsck testing, etc)
5469 * it's much simpler than fixing a caller */
5470 CERROR("%s: "DFID" wasn't declared for insert\n",
5471 osd_name(osd), PFID(fid));
5473 idc = osd_idc_find_or_init(env, osd, fid);
5475 RETURN(PTR_ERR(idc));
5478 if (idc->oic_remote) {
5479 /* Insert remote entry */
5480 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
5481 struct osd_mdobj_map *omm = osd->od_mdt_map;
5482 struct osd_thandle *oh;
5484 /* If parent on remote MDT, we need put this object
5486 oh = container_of(th, typeof(*oh), ot_super);
5487 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5489 CERROR("%s: add "DFID" error: rc = %d\n",
5491 PFID(lu_object_fid(&dt->do_lu)), rc);
5495 child_inode = igrab(omm->omm_remote_parent->d_inode);
5497 child_inode = osd_create_local_agent_inode(env, osd,
5498 obj, fid, rec1->rec_type & S_IFMT, th);
5499 if (IS_ERR(child_inode))
5500 RETURN(PTR_ERR(child_inode));
5503 /* Insert local entry */
5504 if (unlikely(idc->oic_lid.oii_ino == 0)) {
5505 /* for a reason OI cache wasn't filled properly */
5506 CERROR("%s: OIC for "DFID" isn't filled\n",
5507 osd_name(osd), PFID(fid));
5510 child_inode = oti->oti_inode;
5511 if (unlikely(child_inode == NULL)) {
5512 struct ldiskfs_inode_info *lii;
5516 child_inode = oti->oti_inode = &lii->vfs_inode;
5518 child_inode->i_sb = osd_sb(osd);
5519 child_inode->i_ino = idc->oic_lid.oii_ino;
5520 child_inode->i_mode = rec1->rec_type & S_IFMT;
5523 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
5525 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
5526 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
5528 if (child_inode && child_inode != oti->oti_inode)
5530 LASSERT(osd_invariant(obj));
5531 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5536 * Initialize osd Iterator for given osd index object.
5538 * \param dt osd index object
5541 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
5542 struct dt_object *dt,
5545 struct osd_it_iam *it;
5546 struct osd_object *obj = osd_dt_obj(dt);
5547 struct lu_object *lo = &dt->do_lu;
5548 struct iam_path_descr *ipd;
5549 struct iam_container *bag = &obj->oo_dir->od_container;
5551 if (!dt_object_exists(dt))
5552 return ERR_PTR(-ENOENT);
5556 return ERR_PTR(-ENOMEM);
5558 ipd = osd_it_ipd_get(env, bag);
5559 if (likely(ipd != NULL)) {
5563 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
5564 return (struct dt_it *)it;
5567 return ERR_PTR(-ENOMEM);
5572 * free given Iterator.
5574 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
5576 struct osd_it_iam *it = (struct osd_it_iam *)di;
5577 struct osd_object *obj = it->oi_obj;
5579 iam_it_fini(&it->oi_it);
5580 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
5581 osd_object_put(env, obj);
5586 * Move Iterator to record specified by \a key
5588 * \param di osd iterator
5589 * \param key key for index
5591 * \retval +ve di points to record with least key not larger than key
5592 * \retval 0 di points to exact matched key
5593 * \retval -ve failure
5596 static int osd_it_iam_get(const struct lu_env *env,
5597 struct dt_it *di, const struct dt_key *key)
5599 struct osd_thread_info *oti = osd_oti_get(env);
5600 struct osd_it_iam *it = (struct osd_it_iam *)di;
5602 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5603 /* swab quota uid/gid */
5604 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5605 key = (struct dt_key *)&oti->oti_quota_id;
5608 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
5614 * \param di osd iterator
5616 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
5618 struct osd_it_iam *it = (struct osd_it_iam *)di;
5620 iam_it_put(&it->oi_it);
5624 * Move iterator by one record
5626 * \param di osd iterator
5628 * \retval +1 end of container reached
5630 * \retval -ve failure
5633 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
5635 struct osd_it_iam *it = (struct osd_it_iam *)di;
5637 return iam_it_next(&it->oi_it);
5641 * Return pointer to the key under iterator.
5644 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
5645 const struct dt_it *di)
5647 struct osd_thread_info *oti = osd_oti_get(env);
5648 struct osd_it_iam *it = (struct osd_it_iam *)di;
5649 struct osd_object *obj = it->oi_obj;
5652 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
5654 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
5655 /* swab quota uid/gid */
5656 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
5657 key = (struct dt_key *)&oti->oti_quota_id;
5664 * Return size of key under iterator (in bytes)
5667 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
5669 struct osd_it_iam *it = (struct osd_it_iam *)di;
5671 return iam_it_key_size(&it->oi_it);
5675 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
5677 /* check if file type is required */
5678 if (ent->lde_attrs & LUDA_TYPE) {
5679 struct luda_type *lt;
5680 int align = sizeof(*lt) - 1;
5682 len = (len + align) & ~align;
5683 lt = (struct luda_type *)(ent->lde_name + len);
5684 lt->lt_type = cpu_to_le16(DTTOIF(type));
5687 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
5691 * build lu direct from backend fs dirent.
5695 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
5696 char *name, __u16 namelen, __u16 type, __u32 attr)
5698 ent->lde_attrs = attr | LUDA_FID;
5699 fid_cpu_to_le(&ent->lde_fid, fid);
5701 ent->lde_hash = cpu_to_le64(offset);
5702 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
5704 strncpy(ent->lde_name, name, namelen);
5705 ent->lde_name[namelen] = '\0';
5706 ent->lde_namelen = cpu_to_le16(namelen);
5708 /* append lustre attributes */
5709 osd_it_append_attrs(ent, namelen, type);
5713 * Return pointer to the record under iterator.
5715 static int osd_it_iam_rec(const struct lu_env *env,
5716 const struct dt_it *di,
5717 struct dt_rec *dtrec, __u32 attr)
5719 struct osd_it_iam *it = (struct osd_it_iam *)di;
5720 struct osd_thread_info *info = osd_oti_get(env);
5723 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
5724 const struct osd_fid_pack *rec;
5725 struct lu_fid *fid = &info->oti_fid;
5726 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
5732 name = (char *)iam_it_key_get(&it->oi_it);
5734 RETURN(PTR_ERR(name));
5736 namelen = iam_it_key_size(&it->oi_it);
5738 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
5740 RETURN(PTR_ERR(rec));
5742 rc = osd_fid_unpack(fid, rec);
5746 hash = iam_it_store(&it->oi_it);
5748 /* IAM does not store object type in IAM index (dir) */
5749 osd_it_pack_dirent(lde, fid, hash, name, namelen,
5751 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5752 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5753 (struct iam_rec *)dtrec);
5754 osd_quota_unpack(it->oi_obj, dtrec);
5756 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5757 (struct iam_rec *)dtrec);
5764 * Returns cookie for current Iterator position.
5766 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
5768 struct osd_it_iam *it = (struct osd_it_iam *)di;
5770 return iam_it_store(&it->oi_it);
5774 * Restore iterator from cookie.
5776 * \param di osd iterator
5777 * \param hash Iterator location cookie
5779 * \retval +ve di points to record with least key not larger than key.
5780 * \retval 0 di points to exact matched key
5781 * \retval -ve failure
5784 static int osd_it_iam_load(const struct lu_env *env,
5785 const struct dt_it *di, __u64 hash)
5787 struct osd_it_iam *it = (struct osd_it_iam *)di;
5789 return iam_it_load(&it->oi_it, hash);
5792 static const struct dt_index_operations osd_index_iam_ops = {
5793 .dio_lookup = osd_index_iam_lookup,
5794 .dio_declare_insert = osd_index_declare_iam_insert,
5795 .dio_insert = osd_index_iam_insert,
5796 .dio_declare_delete = osd_index_declare_iam_delete,
5797 .dio_delete = osd_index_iam_delete,
5799 .init = osd_it_iam_init,
5800 .fini = osd_it_iam_fini,
5801 .get = osd_it_iam_get,
5802 .put = osd_it_iam_put,
5803 .next = osd_it_iam_next,
5804 .key = osd_it_iam_key,
5805 .key_size = osd_it_iam_key_size,
5806 .rec = osd_it_iam_rec,
5807 .store = osd_it_iam_store,
5808 .load = osd_it_iam_load
5814 * Creates or initializes iterator context.
5816 * \retval struct osd_it_ea, iterator structure on success
5819 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
5820 struct dt_object *dt,
5823 struct osd_object *obj = osd_dt_obj(dt);
5824 struct osd_thread_info *info = osd_oti_get(env);
5825 struct osd_it_ea *oie;
5827 struct lu_object *lo = &dt->do_lu;
5828 struct dentry *obj_dentry;
5831 if (!dt_object_exists(dt) || obj->oo_destroyed)
5832 RETURN(ERR_PTR(-ENOENT));
5834 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
5836 RETURN(ERR_PTR(-ENOMEM));
5837 obj_dentry = &oie->oie_dentry;
5839 obj_dentry->d_inode = obj->oo_inode;
5840 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
5841 obj_dentry->d_name.hash = 0;
5843 oie->oie_rd_dirent = 0;
5844 oie->oie_it_dirent = 0;
5845 oie->oie_dirent = NULL;
5846 if (unlikely(!info->oti_it_ea_buf_used)) {
5847 oie->oie_buf = info->oti_it_ea_buf;
5848 info->oti_it_ea_buf_used = 1;
5850 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5851 if (oie->oie_buf == NULL)
5852 RETURN(ERR_PTR(-ENOMEM));
5856 file = &oie->oie_file;
5858 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
5859 if (attr & LUDA_64BITHASH)
5860 file->f_mode = FMODE_64BITHASH;
5862 file->f_mode = FMODE_32BITHASH;
5863 file->f_path.dentry = obj_dentry;
5864 file->f_mapping = obj->oo_inode->i_mapping;
5865 file->f_op = obj->oo_inode->i_fop;
5866 set_file_inode(file, obj->oo_inode);
5869 RETURN((struct dt_it *) oie);
5873 * Destroy or finishes iterator context.
5875 * \param di iterator structure to be destroyed
5877 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
5879 struct osd_thread_info *info = osd_oti_get(env);
5880 struct osd_it_ea *oie = (struct osd_it_ea *)di;
5881 struct osd_object *obj = oie->oie_obj;
5882 struct inode *inode = obj->oo_inode;
5885 oie->oie_file.f_op->release(inode, &oie->oie_file);
5886 osd_object_put(env, obj);
5887 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
5888 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5890 info->oti_it_ea_buf_used = 0;
5891 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
5896 * It position the iterator at given key, so that next lookup continues from
5897 * that key Or it is similar to dio_it->load() but based on a key,
5898 * rather than file position.
5900 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
5903 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
5905 static int osd_it_ea_get(const struct lu_env *env,
5906 struct dt_it *di, const struct dt_key *key)
5908 struct osd_it_ea *it = (struct osd_it_ea *)di;
5911 LASSERT(((const char *)key)[0] == '\0');
5912 it->oie_file.f_pos = 0;
5913 it->oie_rd_dirent = 0;
5914 it->oie_it_dirent = 0;
5915 it->oie_dirent = NULL;
5923 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
5927 struct osd_filldir_cbs {
5928 #ifdef HAVE_DIR_CONTEXT
5929 struct dir_context ctx;
5931 struct osd_it_ea *it;
5934 * It is called internally by ->readdir(). It fills the
5935 * iterator's in-memory data structure with required
5936 * information i.e. name, namelen, rec_size etc.
5938 * \param buf in which information to be filled in.
5939 * \param name name of the file in given dir
5941 * \retval 0 on success
5942 * \retval 1 on buffer full
5944 #ifdef HAVE_FILLDIR_USE_CTX
5945 static int osd_ldiskfs_filldir(struct dir_context *buf,
5947 static int osd_ldiskfs_filldir(void *buf,
5949 const char *name, int namelen,
5950 loff_t offset, __u64 ino, unsigned d_type)
5952 struct osd_it_ea *it =
5953 ((struct osd_filldir_cbs *)buf)->it;
5954 struct osd_object *obj = it->oie_obj;
5955 struct osd_it_ea_dirent *ent = it->oie_dirent;
5956 struct lu_fid *fid = &ent->oied_fid;
5957 struct osd_fid_pack *rec;
5960 /* this should never happen */
5961 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
5962 CERROR("ldiskfs return invalid namelen %d\n", namelen);
5966 if ((void *) ent - it->oie_buf + sizeof(*ent) + namelen >
5970 /* "." is just the object itself. */
5971 if (namelen == 1 && name[0] == '.') {
5972 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5973 } else if (d_type & LDISKFS_DIRENT_LUFID) {
5974 rec = (struct osd_fid_pack*) (name + namelen + 1);
5975 if (osd_fid_unpack(fid, rec) != 0)
5980 d_type &= ~LDISKFS_DIRENT_LUFID;
5982 /* NOT export local root. */
5983 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
5984 ino = obj->oo_inode->i_ino;
5985 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5988 ent->oied_ino = ino;
5989 ent->oied_off = offset;
5990 ent->oied_namelen = namelen;
5991 ent->oied_type = d_type;
5993 memcpy(ent->oied_name, name, namelen);
5995 it->oie_rd_dirent++;
5996 it->oie_dirent = (void *) ent + cfs_size_round(sizeof(*ent) + namelen);
6001 * Calls ->readdir() to load a directory entry at a time
6002 * and stored it in iterator's in-memory data structure.
6004 * \param di iterator's in memory structure
6006 * \retval 0 on success
6007 * \retval -ve on error
6008 * \retval +1 reach the end of entry
6010 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6011 const struct dt_it *di)
6013 struct osd_it_ea *it = (struct osd_it_ea *)di;
6014 struct osd_object *obj = it->oie_obj;
6015 struct inode *inode = obj->oo_inode;
6016 struct htree_lock *hlock = NULL;
6017 struct file *filp = &it->oie_file;
6019 struct osd_filldir_cbs buf = {
6020 #ifdef HAVE_DIR_CONTEXT
6021 .ctx.actor = osd_ldiskfs_filldir,
6027 it->oie_dirent = it->oie_buf;
6028 it->oie_rd_dirent = 0;
6030 if (obj->oo_hl_head != NULL) {
6031 hlock = osd_oti_get(env)->oti_hlock;
6032 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6033 inode, LDISKFS_HLOCK_READDIR);
6035 down_read(&obj->oo_ext_idx_sem);
6038 #ifdef HAVE_DIR_CONTEXT
6039 buf.ctx.pos = filp->f_pos;
6040 rc = inode->i_fop->iterate(filp, &buf.ctx);
6041 filp->f_pos = buf.ctx.pos;
6043 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6047 ldiskfs_htree_unlock(hlock);
6049 up_read(&obj->oo_ext_idx_sem);
6051 if (it->oie_rd_dirent == 0) {
6052 /*If it does not get any dirent, it means it has been reached
6053 *to the end of the dir */
6054 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6058 it->oie_dirent = it->oie_buf;
6059 it->oie_it_dirent = 1;
6066 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6067 * to load a directory entry at a time and stored it in
6068 * iterator's in-memory data structure.
6070 * \param di iterator's in memory structure
6072 * \retval +ve iterator reached to end
6073 * \retval 0 iterator not reached to end
6074 * \retval -ve on error
6076 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6078 struct osd_it_ea *it = (struct osd_it_ea *)di;
6083 if (it->oie_it_dirent < it->oie_rd_dirent) {
6085 (void *) it->oie_dirent +
6086 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6087 it->oie_dirent->oied_namelen);
6088 it->oie_it_dirent++;
6091 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6094 rc = osd_ldiskfs_it_fill(env, di);
6101 * Returns the key at current position from iterator's in memory structure.
6103 * \param di iterator's in memory structure
6105 * \retval key i.e. struct dt_key on success
6107 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6108 const struct dt_it *di)
6110 struct osd_it_ea *it = (struct osd_it_ea *)di;
6112 return (struct dt_key *)it->oie_dirent->oied_name;
6116 * Returns the key's size at current position from iterator's in memory structure.
6118 * \param di iterator's in memory structure
6120 * \retval key_size i.e. struct dt_key on success
6122 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6124 struct osd_it_ea *it = (struct osd_it_ea *)di;
6126 return it->oie_dirent->oied_namelen;
6129 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6131 if (LDISKFS_DIR_REC_LEN(de) >=
6132 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6139 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6140 unsigned blocksize, bool dotdot)
6143 return osd_dotdot_has_space(de);
6145 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6146 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6153 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6154 handle_t *jh, struct dentry *dentry,
6155 const struct lu_fid *fid, struct buffer_head *bh,
6156 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6159 struct inode *dir = dentry->d_parent->d_inode;
6160 struct inode *inode = dentry->d_inode;
6161 struct osd_fid_pack *rec;
6162 struct ldiskfs_dentry_param *ldp;
6163 int namelen = dentry->d_name.len;
6165 struct osd_thread_info *info = osd_oti_get(env);
6168 if (!LDISKFS_HAS_INCOMPAT_FEATURE(inode->i_sb,
6169 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
6172 /* There is enough space to hold the FID-in-dirent. */
6173 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6174 rc = ldiskfs_journal_get_write_access(jh, bh);
6178 de->name[namelen] = 0;
6179 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6180 rec->fp_len = sizeof(struct lu_fid) + 1;
6181 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6182 de->file_type |= LDISKFS_DIRENT_LUFID;
6183 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6190 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6194 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6195 osd_get_ldiskfs_dirent_param(ldp, fid);
6196 dentry->d_fsdata = (void *)ldp;
6197 ll_vfs_dq_init(dir);
6198 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6199 /* It is too bad, we cannot reinsert the name entry back.
6200 * That means we lose it! */
6202 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6203 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6204 osd_ino2name(inode),
6205 dir->i_ino, dir->i_generation, namelen,
6206 dentry->d_name.name, PFID(fid), rc);
6212 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6213 struct osd_it_ea *it, struct lu_fid *fid,
6214 struct osd_inode_id *id, __u32 *attr)
6216 struct osd_thread_info *info = osd_oti_get(env);
6217 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6218 struct osd_device *dev = osd_obj2dev(obj);
6219 struct super_block *sb = osd_sb(dev);
6220 const char *devname = osd_name(dev);
6221 struct osd_it_ea_dirent *ent = it->oie_dirent;
6222 struct inode *dir = obj->oo_inode;
6223 struct htree_lock *hlock = NULL;
6224 struct buffer_head *bh = NULL;
6225 handle_t *jh = NULL;
6226 struct ldiskfs_dir_entry_2 *de;
6227 struct dentry *dentry;
6228 struct inode *inode;
6229 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6232 bool dotdot = false;
6236 if (ent->oied_name[0] == '.') {
6237 if (ent->oied_namelen == 1)
6240 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6244 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6245 inode = osd_iget(info, dev, id);
6246 if (IS_ERR(inode)) {
6247 rc = PTR_ERR(inode);
6248 if (rc == -ENOENT || rc == -ESTALE) {
6249 /* Maybe dangling name entry, or
6250 * corrupted directory entry. */
6251 *attr |= LUDA_UNKNOWN;
6254 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6255 "check_repair, dir = %lu/%u, name = %.*s, "
6256 "ino = %llu, rc = %d\n",
6257 devname, dir->i_ino, dir->i_generation,
6258 ent->oied_namelen, ent->oied_name,
6265 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6267 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6268 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6273 /* We need to ensure that the name entry is still valid.
6274 * Because it may be removed or renamed by other already.
6276 * The unlink or rename operation will start journal before PDO lock,
6277 * so to avoid deadlock, here we need to start journal handle before
6278 * related PDO lock also. But because we do not know whether there
6279 * will be something to be repaired before PDO lock, we just start
6280 * journal without conditions.
6282 * We may need to remove the name entry firstly, then insert back.
6283 * One credit is for user quota file update.
6284 * One credit is for group quota file update.
6285 * Two credits are for dirty inode. */
6286 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6287 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6289 if (dev->od_dirent_journal != 0) {
6292 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6295 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6296 "check_repair, dir = %lu/%u, credits = %d, "
6297 "name = %.*s, ino = %llu: rc = %d\n",
6298 devname, dir->i_ino, dir->i_generation, credits,
6299 ent->oied_namelen, ent->oied_name,
6302 GOTO(out_inode, rc);
6305 if (obj->oo_hl_head != NULL) {
6306 hlock = osd_oti_get(env)->oti_hlock;
6307 /* "0" means exclusive lock for the whole directory.
6308 * We need to prevent others access such name entry
6309 * during the delete + insert. Neither HLOCK_ADD nor
6310 * HLOCK_DEL cannot guarantee the atomicity. */
6311 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6313 down_write(&obj->oo_ext_idx_sem);
6316 if (obj->oo_hl_head != NULL) {
6317 hlock = osd_oti_get(env)->oti_hlock;
6318 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6319 LDISKFS_HLOCK_LOOKUP);
6321 down_read(&obj->oo_ext_idx_sem);
6325 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6326 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6327 *attr |= LUDA_IGNORE;
6332 /* For dotdot entry, if there is not enough space to hold the
6333 * FID-in-dirent, just keep them there. It only happens when the
6334 * device upgraded from 1.8 or restored from MDT file-level backup.
6335 * For the whole directory, only dotdot entry have no FID-in-dirent
6336 * and needs to get FID from LMA when readdir, it will not affect the
6337 * performance much. */
6338 if (dotdot && !osd_dotdot_has_space(de)) {
6339 *attr |= LUDA_UNKNOWN;
6345 if (lu_fid_eq(fid, &lma->lma_self_fid))
6348 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6349 struct lu_fid *tfid = &lma->lma_self_fid;
6351 if (likely(dotdot &&
6352 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6353 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6354 /* It must be REMOTE_PARENT_DIR and as the
6355 * 'dotdot' entry of remote directory */
6356 *attr |= LUDA_IGNORE;
6358 CDEBUG(D_LFSCK, "%s: expect remote agent "
6359 "parent directory, but got %.*s under "
6360 "dir = %lu/%u with the FID "DFID"\n",
6361 devname, ent->oied_namelen,
6362 ent->oied_name, dir->i_ino,
6363 dir->i_generation, PFID(tfid));
6365 *attr |= LUDA_UNKNOWN;
6372 if (!fid_is_zero(fid)) {
6373 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6375 if (rc == -ENOENT ||
6377 !(dev->od_scrub.os_file.sf_flags & SF_UPGRADE))) {
6378 /* linkEA does not recognize the dirent entry,
6379 * it may because the dirent entry corruption
6380 * and points to other's inode. */
6381 CDEBUG(D_LFSCK, "%s: the target inode does not "
6382 "recognize the dirent, dir = %lu/%u, "
6383 " name = %.*s, ino = %llu, "
6384 DFID": rc = %d\n", devname, dir->i_ino,
6385 dir->i_generation, ent->oied_namelen,
6386 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6387 *attr |= LUDA_UNKNOWN;
6392 if (rc && rc != -ENODATA) {
6393 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6394 "dir = %lu/%u, name = %.*s, ino = %llu, "
6395 DFID": rc = %d\n", devname, dir->i_ino,
6396 dir->i_generation, ent->oied_namelen,
6397 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6398 *attr |= LUDA_UNKNOWN;
6405 /* linkEA recognizes the dirent entry, the FID-in-LMA is
6406 * valid, trusted, in spite of fid_is_sane(fid) or not. */
6407 if (*attr & LUDA_VERIFY_DRYRUN) {
6408 *fid = lma->lma_self_fid;
6409 *attr |= LUDA_REPAIR;
6416 dev->od_dirent_journal = 1;
6417 if (hlock != NULL) {
6418 ldiskfs_htree_unlock(hlock);
6421 up_read(&obj->oo_ext_idx_sem);
6427 *fid = lma->lma_self_fid;
6429 /* Update or append the FID-in-dirent. */
6430 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6431 bh, de, hlock, dotdot);
6433 *attr |= LUDA_REPAIR;
6435 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6436 "the dirent, dir = %lu/%u, name = %.*s, "
6437 "ino = %llu, "DFID": rc = %d\n",
6438 devname, dir->i_ino, dir->i_generation,
6439 ent->oied_namelen, ent->oied_name,
6440 ent->oied_ino, PFID(fid), rc);
6442 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6443 if (*attr & LUDA_VERIFY_DRYRUN) {
6444 if (fid_is_sane(fid)) {
6445 *attr |= LUDA_REPAIR;
6446 } else if (dev->od_index == 0) {
6447 lu_igif_build(fid, inode->i_ino,
6448 inode->i_generation);
6449 *attr |= LUDA_UPGRADE;
6457 dev->od_dirent_journal = 1;
6458 if (hlock != NULL) {
6459 ldiskfs_htree_unlock(hlock);
6462 up_read(&obj->oo_ext_idx_sem);
6469 if (unlikely(fid_is_sane(fid))) {
6470 /* FID-in-dirent exists, but FID-in-LMA is lost.
6471 * Trust the FID-in-dirent, and add FID-in-LMA. */
6472 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
6474 *attr |= LUDA_REPAIR;
6476 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
6477 "update dirent, dir = %lu/%u, "
6478 "name = %.*s, ino = %llu, "
6480 devname, dir->i_ino, dir->i_generation,
6481 ent->oied_namelen, ent->oied_name,
6482 ent->oied_ino, PFID(fid), rc);
6483 } else if (dev->od_index == 0) {
6484 lu_igif_build(fid, inode->i_ino, inode->i_generation);
6485 /* It is probably IGIF object. Only aappend the
6486 * FID-in-dirent. OI scrub will process FID-in-LMA. */
6487 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6488 bh, de, hlock, dotdot);
6490 *attr |= LUDA_UPGRADE;
6492 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
6493 "after the dirent, dir = %lu/%u, "
6494 "name = %.*s, ino = %llu, "
6496 devname, dir->i_ino, dir->i_generation,
6497 ent->oied_namelen, ent->oied_name,
6498 ent->oied_ino, PFID(fid), rc);
6507 if (hlock != NULL) {
6508 ldiskfs_htree_unlock(hlock);
6510 if (dev->od_dirent_journal != 0)
6511 up_write(&obj->oo_ext_idx_sem);
6513 up_read(&obj->oo_ext_idx_sem);
6517 ldiskfs_journal_stop(jh);
6521 if (rc >= 0 && !dirty)
6522 dev->od_dirent_journal = 0;
6528 * Returns the value at current position from iterator's in memory structure.
6530 * \param di struct osd_it_ea, iterator's in memory structure
6531 * \param attr attr requested for dirent.
6532 * \param lde lustre dirent
6534 * \retval 0 no error and \param lde has correct lustre dirent.
6535 * \retval -ve on error
6537 static inline int osd_it_ea_rec(const struct lu_env *env,
6538 const struct dt_it *di,
6539 struct dt_rec *dtrec, __u32 attr)
6541 struct osd_it_ea *it = (struct osd_it_ea *)di;
6542 struct osd_object *obj = it->oie_obj;
6543 struct osd_device *dev = osd_obj2dev(obj);
6544 struct osd_thread_info *oti = osd_oti_get(env);
6545 struct osd_inode_id *id = &oti->oti_id;
6546 struct lu_fid *fid = &it->oie_dirent->oied_fid;
6547 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6548 __u32 ino = it->oie_dirent->oied_ino;
6552 LASSERT(obj->oo_inode != dev->od_mdt_map->omm_remote_parent->d_inode);
6554 if (attr & LUDA_VERIFY) {
6555 if (unlikely(ino == osd_remote_parent_ino(dev))) {
6556 attr |= LUDA_IGNORE;
6557 /* If the parent is on remote MDT, and there
6558 * is no FID-in-dirent, then we have to get
6559 * the parent FID from the linkEA. */
6560 if (!fid_is_sane(fid) &&
6561 it->oie_dirent->oied_namelen == 2 &&
6562 it->oie_dirent->oied_name[0] == '.' &&
6563 it->oie_dirent->oied_name[1] == '.')
6564 osd_get_pfid_from_linkea(env, obj, fid);
6566 rc = osd_dirent_check_repair(env, obj, it, fid, id,
6570 if (!fid_is_sane(fid))
6571 attr |= LUDA_UNKNOWN;
6573 attr &= ~LU_DIRENT_ATTRS_MASK;
6574 if (!fid_is_sane(fid)) {
6575 bool is_dotdot = false;
6576 if (it->oie_dirent->oied_namelen == 2 &&
6577 it->oie_dirent->oied_name[0] == '.' &&
6578 it->oie_dirent->oied_name[1] == '.')
6580 /* If the parent is on remote MDT, and there
6581 * is no FID-in-dirent, then we have to get
6582 * the parent FID from the linkEA. */
6583 if (ino == osd_remote_parent_ino(dev) && is_dotdot) {
6584 rc = osd_get_pfid_from_linkea(env, obj, fid);
6586 if (is_dotdot == false &&
6587 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
6590 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6593 osd_id_gen(id, ino, OSD_OII_NOGEN);
6597 /* Pack the entry anyway, at least the offset is right. */
6598 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
6599 it->oie_dirent->oied_name,
6600 it->oie_dirent->oied_namelen,
6601 it->oie_dirent->oied_type, attr);
6606 if (osd_remote_fid(env, dev, fid))
6609 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
6610 osd_add_oi_cache(oti, dev, id, fid);
6612 RETURN(rc > 0 ? 0 : rc);
6616 * Returns the record size size at current position.
6618 * This function will return record(lu_dirent) size in bytes.
6620 * \param[in] env execution environment
6621 * \param[in] di iterator's in memory structure
6622 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
6623 * calculate the lu_dirent size.
6625 * \retval record size(in bytes & in memory) of the current lu_dirent
6628 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
6631 struct osd_it_ea *it = (struct osd_it_ea *)di;
6633 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
6637 * Returns a cookie for current position of the iterator head, so that
6638 * user can use this cookie to load/start the iterator next time.
6640 * \param di iterator's in memory structure
6642 * \retval cookie for current position, on success
6644 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
6646 struct osd_it_ea *it = (struct osd_it_ea *)di;
6648 return it->oie_dirent->oied_off;
6652 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6653 * to load a directory entry at a time and stored it i inn,
6654 * in iterator's in-memory data structure.
6656 * \param di struct osd_it_ea, iterator's in memory structure
6658 * \retval +ve on success
6659 * \retval -ve on error
6661 static int osd_it_ea_load(const struct lu_env *env,
6662 const struct dt_it *di, __u64 hash)
6664 struct osd_it_ea *it = (struct osd_it_ea *)di;
6668 it->oie_file.f_pos = hash;
6670 rc = osd_ldiskfs_it_fill(env, di);
6681 * Index lookup function for interoperability mode (b11826).
6683 * \param key, key i.e. file name to be searched
6685 * \retval +ve, on success
6686 * \retval -ve, on error
6688 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
6689 struct dt_rec *rec, const struct dt_key *key)
6691 struct osd_object *obj = osd_dt_obj(dt);
6696 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
6697 LINVRNT(osd_invariant(obj));
6699 rc = osd_ea_lookup_rec(env, obj, rec, key);
6706 * Index and Iterator operations for interoperability
6707 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
6709 static const struct dt_index_operations osd_index_ea_ops = {
6710 .dio_lookup = osd_index_ea_lookup,
6711 .dio_declare_insert = osd_index_declare_ea_insert,
6712 .dio_insert = osd_index_ea_insert,
6713 .dio_declare_delete = osd_index_declare_ea_delete,
6714 .dio_delete = osd_index_ea_delete,
6716 .init = osd_it_ea_init,
6717 .fini = osd_it_ea_fini,
6718 .get = osd_it_ea_get,
6719 .put = osd_it_ea_put,
6720 .next = osd_it_ea_next,
6721 .key = osd_it_ea_key,
6722 .key_size = osd_it_ea_key_size,
6723 .rec = osd_it_ea_rec,
6724 .rec_size = osd_it_ea_rec_size,
6725 .store = osd_it_ea_store,
6726 .load = osd_it_ea_load
6730 static void *osd_key_init(const struct lu_context *ctx,
6731 struct lu_context_key *key)
6733 struct osd_thread_info *info;
6735 OBD_ALLOC_PTR(info);
6737 return ERR_PTR(-ENOMEM);
6739 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6740 if (info->oti_it_ea_buf == NULL)
6743 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
6745 info->oti_hlock = ldiskfs_htree_lock_alloc();
6746 if (info->oti_hlock == NULL)
6752 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6755 return ERR_PTR(-ENOMEM);
6758 static void osd_key_fini(const struct lu_context *ctx,
6759 struct lu_context_key *key, void* data)
6761 struct osd_thread_info *info = data;
6762 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
6763 struct osd_idmap_cache *idc = info->oti_ins_cache;
6765 if (info->oti_inode != NULL)
6767 if (info->oti_hlock != NULL)
6768 ldiskfs_htree_lock_free(info->oti_hlock);
6769 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6770 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
6771 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
6772 lu_buf_free(&info->oti_big_buf);
6774 LASSERT(info->oti_ins_cache_size > 0);
6775 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
6776 info->oti_ins_cache = NULL;
6777 info->oti_ins_cache_size = 0;
6782 static void osd_key_exit(const struct lu_context *ctx,
6783 struct lu_context_key *key, void *data)
6785 struct osd_thread_info *info = data;
6787 LASSERT(info->oti_r_locks == 0);
6788 LASSERT(info->oti_w_locks == 0);
6789 LASSERT(info->oti_txns == 0);
6792 /* type constructor/destructor: osd_type_init, osd_type_fini */
6793 LU_TYPE_INIT_FINI(osd, &osd_key);
6795 struct lu_context_key osd_key = {
6796 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
6797 .lct_init = osd_key_init,
6798 .lct_fini = osd_key_fini,
6799 .lct_exit = osd_key_exit
6803 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
6804 const char *name, struct lu_device *next)
6806 struct osd_device *osd = osd_dev(d);
6808 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname))
6809 >= sizeof(osd->od_svname))
6811 return osd_procfs_init(osd, name);
6814 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
6816 struct seq_server_site *ss = osd_seq_site(osd);
6820 if (osd->od_is_ost || osd->od_cl_seq != NULL)
6823 if (unlikely(ss == NULL))
6826 OBD_ALLOC_PTR(osd->od_cl_seq);
6827 if (osd->od_cl_seq == NULL)
6830 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
6831 osd->od_svname, ss->ss_server_seq);
6833 OBD_FREE_PTR(osd->od_cl_seq);
6834 osd->od_cl_seq = NULL;
6838 if (ss->ss_node_id == 0) {
6839 /* If the OSD on the sequence controller(MDT0), then allocate
6840 * sequence here, otherwise allocate sequence after connected
6841 * to MDT0 (see mdt_register_lwp_callback()). */
6842 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
6843 &osd->od_cl_seq->lcs_space, env);
6849 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
6851 if (osd->od_cl_seq == NULL)
6854 seq_client_fini(osd->od_cl_seq);
6855 OBD_FREE_PTR(osd->od_cl_seq);
6856 osd->od_cl_seq = NULL;
6859 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
6863 /* shutdown quota slave instance associated with the device */
6864 if (o->od_quota_slave != NULL) {
6865 qsd_fini(env, o->od_quota_slave);
6866 o->od_quota_slave = NULL;
6869 osd_fid_fini(env, o);
6874 static void osd_umount(const struct lu_env *env, struct osd_device *o)
6878 if (o->od_mnt != NULL) {
6879 shrink_dcache_sb(osd_sb(o));
6880 osd_sync(env, &o->od_dt_dev);
6889 static int osd_mount(const struct lu_env *env,
6890 struct osd_device *o, struct lustre_cfg *cfg)
6892 const char *name = lustre_cfg_string(cfg, 0);
6893 const char *dev = lustre_cfg_string(cfg, 1);
6895 unsigned long page, s_flags, lmd_flags = 0;
6896 struct page *__page;
6897 struct file_system_type *type;
6898 char *options = NULL;
6900 struct osd_thread_info *info = osd_oti_get(env);
6901 struct lu_fid *fid = &info->oti_fid;
6902 struct inode *inode;
6903 int rc = 0, force_over_512tb = 0;
6906 if (o->od_mnt != NULL)
6909 if (strlen(dev) >= sizeof(o->od_mntdev))
6911 strcpy(o->od_mntdev, dev);
6913 str = lustre_cfg_string(cfg, 2);
6914 s_flags = simple_strtoul(str, NULL, 0);
6915 str = strstr(str, ":");
6917 lmd_flags = simple_strtoul(str + 1, NULL, 0);
6918 opts = lustre_cfg_string(cfg, 3);
6920 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
6921 CERROR("%s: device %s extents feature is not guaranteed to "
6922 "work on big-endian systems. Use \"bigendian_extents\" "
6923 "mount option to override.\n", name, dev);
6927 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
6928 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
6929 CWARN("force_over_128tb option is deprecated. "
6930 "Filesystems less than 512TB can be created without any "
6931 "force options. Use force_over_512tb option for "
6932 "filesystems greater than 512TB.\n");
6935 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
6936 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
6937 CWARN("force_over_256tb option is deprecated. "
6938 "Filesystems less than 512TB can be created without any "
6939 "force options. Use force_over_512tb option for "
6940 "filesystems greater than 512TB.\n");
6944 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
6945 force_over_512tb = 1;
6947 __page = alloc_page(GFP_KERNEL);
6949 GOTO(out, rc = -ENOMEM);
6950 page = (unsigned long)page_address(__page);
6951 options = (char *)page;
6954 /* strip out the options for back compatiblity */
6955 static char *sout[] = {
6962 /* strip out option we processed in osd */
6963 "bigendian_extents",
6969 strcat(options, opts);
6970 for (rc = 0, str = options; sout[rc]; ) {
6971 char *op = strstr(str, sout[rc]);
6977 if (op == options || *(op - 1) == ',') {
6978 str = op + strlen(sout[rc]);
6979 if (*str == ',' || *str == '\0') {
6980 *str == ',' ? str++ : str;
6981 memmove(op, str, strlen(str) + 1);
6984 for (str = op; *str != ',' && *str != '\0'; str++)
6988 strncat(options, "user_xattr,acl", 14);
6991 /* Glom up mount options */
6992 if (*options != '\0')
6993 strcat(options, ",");
6994 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
6996 type = get_fs_type("ldiskfs");
6998 CERROR("%s: cannot find ldiskfs module\n", name);
6999 GOTO(out, rc = -ENODEV);
7002 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7003 module_put(type->owner);
7005 if (IS_ERR(o->od_mnt)) {
7006 rc = PTR_ERR(o->od_mnt);
7008 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7012 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7013 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7014 force_over_512tb == 0) {
7015 CERROR("%s: device %s LDISKFS does not support filesystems "
7016 "greater than 512TB and can cause data corruption. "
7017 "Use \"force_over_512tb\" mount option to override.\n",
7019 GOTO(out_mnt, rc = -EINVAL);
7022 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7023 if (priv_dev_set_rdonly) {
7024 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
7025 o->od_dt_dev.dd_rdonly = 1;
7026 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
7029 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7032 GOTO(out_mnt, rc = -EOPNOTSUPP);
7034 } else if (priv_dev_check_rdonly &&
7035 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
7036 CERROR("%s: underlying device %s is marked as "
7037 "read-only. Setup failed\n", name, dev);
7039 GOTO(out_mnt, rc = -EROFS);
7042 if (!LDISKFS_HAS_COMPAT_FEATURE(o->od_mnt->mnt_sb,
7043 LDISKFS_FEATURE_COMPAT_HAS_JOURNAL)) {
7044 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7045 GOTO(out_mnt, rc = -EINVAL);
7048 #ifdef LDISKFS_MOUNT_DIRDATA
7049 if (LDISKFS_HAS_INCOMPAT_FEATURE(o->od_mnt->mnt_sb,
7050 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
7051 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7052 else if (!o->od_is_ost)
7053 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7054 "enabling the dirdata feature. If you do not want to "
7055 "downgrade to Lustre-1.x again, you can enable it via "
7056 "'tune2fs -O dirdata device'\n", name, dev);
7058 inode = osd_sb(o)->s_root->d_inode;
7059 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7060 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7062 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7066 if (lmd_flags & LMD_FLG_NOSCRUB)
7077 __free_page(__page);
7082 static struct lu_device *osd_device_fini(const struct lu_env *env,
7083 struct lu_device *d)
7085 struct osd_device *o = osd_dev(d);
7088 osd_shutdown(env, o);
7090 osd_scrub_cleanup(env, o);
7091 osd_obj_map_fini(o);
7097 static int osd_device_init0(const struct lu_env *env,
7098 struct osd_device *o,
7099 struct lustre_cfg *cfg)
7101 struct lu_device *l = osd2lu_dev(o);
7102 struct osd_thread_info *info;
7106 /* if the module was re-loaded, env can loose its keys */
7107 rc = lu_env_refill((struct lu_env *) env);
7110 info = osd_oti_get(env);
7113 l->ld_ops = &osd_lu_ops;
7114 o->od_dt_dev.dd_ops = &osd_dt_ops;
7116 spin_lock_init(&o->od_osfs_lock);
7117 mutex_init(&o->od_otable_mutex);
7118 INIT_LIST_HEAD(&o->od_orphan_list);
7120 o->od_read_cache = 1;
7121 o->od_writethrough_cache = 1;
7122 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7124 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7125 sizeof(o->od_svname));
7126 if (cplen >= sizeof(o->od_svname)) {
7131 o->od_index = -1; /* -1 means index is invalid */
7132 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7133 if (rc == LDD_F_SV_TYPE_OST)
7136 o->od_full_scrub_ratio = OFSR_DEFAULT;
7137 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7138 rc = osd_mount(env, o, cfg);
7142 rc = osd_obj_map_init(env, o);
7146 rc = lu_site_init(&o->od_site, l);
7148 GOTO(out_compat, rc);
7149 o->od_site.ls_bottom_dev = l;
7151 rc = lu_site_init_finish(&o->od_site);
7155 INIT_LIST_HEAD(&o->od_ios_list);
7156 /* setup scrub, including OI files initialization */
7157 rc = osd_scrub_setup(env, o);
7161 rc = osd_procfs_init(o, o->od_svname);
7163 CERROR("%s: can't initialize procfs: rc = %d\n",
7165 GOTO(out_scrub, rc);
7168 LASSERT(l->ld_site->ls_linkage.next != NULL);
7169 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7171 /* initialize quota slave instance */
7172 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7174 if (IS_ERR(o->od_quota_slave)) {
7175 rc = PTR_ERR(o->od_quota_slave);
7176 o->od_quota_slave = NULL;
7177 GOTO(out_procfs, rc);
7185 osd_scrub_cleanup(env, o);
7187 lu_site_fini(&o->od_site);
7189 osd_obj_map_fini(o);
7196 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7197 struct lu_device_type *t,
7198 struct lustre_cfg *cfg)
7200 struct osd_device *o;
7205 return ERR_PTR(-ENOMEM);
7207 rc = dt_device_init(&o->od_dt_dev, t);
7209 /* Because the ctx might be revived in dt_device_init,
7210 * refill the env here */
7211 lu_env_refill((struct lu_env *)env);
7212 rc = osd_device_init0(env, o, cfg);
7214 dt_device_fini(&o->od_dt_dev);
7217 if (unlikely(rc != 0))
7220 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7223 static struct lu_device *osd_device_free(const struct lu_env *env,
7224 struct lu_device *d)
7226 struct osd_device *o = osd_dev(d);
7229 /* XXX: make osd top device in order to release reference */
7230 d->ld_site->ls_top_dev = d;
7231 lu_site_purge(env, d->ld_site, -1);
7232 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7233 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7234 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7236 lu_site_fini(&o->od_site);
7237 dt_device_fini(&o->od_dt_dev);
7242 static int osd_process_config(const struct lu_env *env,
7243 struct lu_device *d, struct lustre_cfg *cfg)
7245 struct osd_device *o = osd_dev(d);
7249 switch (cfg->lcfg_command) {
7251 rc = osd_mount(env, o, cfg);
7254 lu_dev_del_linkage(d->ld_site, d);
7255 rc = osd_shutdown(env, o);
7258 LASSERT(&o->od_dt_dev);
7259 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7260 cfg, &o->od_dt_dev);
7261 if (rc > 0 || rc == -ENOSYS) {
7262 rc = class_process_proc_param(PARAM_OST,
7263 lprocfs_osd_obd_vars,
7264 cfg, &o->od_dt_dev);
7276 static int osd_recovery_complete(const struct lu_env *env,
7277 struct lu_device *d)
7279 struct osd_device *osd = osd_dev(d);
7283 if (osd->od_quota_slave == NULL)
7286 /* start qsd instance on recovery completion, this notifies the quota
7287 * slave code that we are about to process new requests now */
7288 rc = qsd_start(env, osd->od_quota_slave);
7293 * we use exports to track all osd users
7295 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7296 struct obd_device *obd, struct obd_uuid *cluuid,
7297 struct obd_connect_data *data, void *localdata)
7299 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7300 struct lustre_handle conn;
7304 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7306 rc = class_connect(&conn, obd, cluuid);
7310 *exp = class_conn2export(&conn);
7312 spin_lock(&osd->od_osfs_lock);
7314 spin_unlock(&osd->od_osfs_lock);
7320 * once last export (we don't count self-export) disappeared
7321 * osd can be released
7323 static int osd_obd_disconnect(struct obd_export *exp)
7325 struct obd_device *obd = exp->exp_obd;
7326 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7327 int rc, release = 0;
7330 /* Only disconnect the underlying layers on the final disconnect. */
7331 spin_lock(&osd->od_osfs_lock);
7333 if (osd->od_connects == 0)
7335 spin_unlock(&osd->od_osfs_lock);
7337 rc = class_disconnect(exp); /* bz 9811 */
7339 if (rc == 0 && release)
7340 class_manual_cleanup(obd);
7344 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7345 struct lu_device *dev)
7347 struct osd_device *osd = osd_dev(dev);
7348 struct lr_server_data *lsd =
7349 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7353 if (osd->od_quota_slave != NULL) {
7354 /* set up quota slave objects */
7355 result = qsd_prepare(env, osd->od_quota_slave);
7360 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7361 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7362 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7363 osd->od_index_in_idif = 1;
7365 osd->od_index_in_idif = 0;
7366 result = osd_register_proc_index_in_idif(osd);
7371 osd->od_index_in_idif = 1;
7375 result = osd_fid_init(env, osd);
7380 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7381 struct lu_fid *fid, struct md_op_data *op_data)
7383 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7385 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7388 static const struct lu_object_operations osd_lu_obj_ops = {
7389 .loo_object_init = osd_object_init,
7390 .loo_object_delete = osd_object_delete,
7391 .loo_object_release = osd_object_release,
7392 .loo_object_free = osd_object_free,
7393 .loo_object_print = osd_object_print,
7394 .loo_object_invariant = osd_object_invariant
7397 const struct lu_device_operations osd_lu_ops = {
7398 .ldo_object_alloc = osd_object_alloc,
7399 .ldo_process_config = osd_process_config,
7400 .ldo_recovery_complete = osd_recovery_complete,
7401 .ldo_prepare = osd_prepare,
7404 static const struct lu_device_type_operations osd_device_type_ops = {
7405 .ldto_init = osd_type_init,
7406 .ldto_fini = osd_type_fini,
7408 .ldto_start = osd_type_start,
7409 .ldto_stop = osd_type_stop,
7411 .ldto_device_alloc = osd_device_alloc,
7412 .ldto_device_free = osd_device_free,
7414 .ldto_device_init = osd_device_init,
7415 .ldto_device_fini = osd_device_fini
7418 static struct lu_device_type osd_device_type = {
7419 .ldt_tags = LU_DEVICE_DT,
7420 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
7421 .ldt_ops = &osd_device_type_ops,
7422 .ldt_ctx_tags = LCT_LOCAL,
7425 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
7427 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7428 struct super_block *sb = osd_sb(osd);
7430 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
7434 * lprocfs legacy support.
7436 static struct obd_ops osd_obd_device_ops = {
7437 .o_owner = THIS_MODULE,
7438 .o_connect = osd_obd_connect,
7439 .o_disconnect = osd_obd_disconnect,
7440 .o_fid_alloc = osd_fid_alloc,
7441 .o_health_check = osd_health_check,
7444 static int __init osd_init(void)
7448 LASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
7449 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
7450 /* please, try to keep osd_thread_info smaller than a page */
7451 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
7456 rc = lu_kmem_init(ldiskfs_caches);
7460 #ifdef CONFIG_KALLSYMS
7461 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
7462 priv_dev_check_rdonly = (void *)kallsyms_lookup_name("dev_check_rdonly");
7463 /* Clear readonly is unused at this time */
7464 /*priv_dev_clear_rdonly = (void *)kallsyms_lookup_name("dev_clear_rdonly");*/
7467 rc = class_register_type(&osd_obd_device_ops, NULL, true,
7468 lprocfs_osd_module_vars,
7469 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
7471 lu_kmem_fini(ldiskfs_caches);
7475 static void __exit osd_exit(void)
7477 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
7478 lu_kmem_fini(ldiskfs_caches);
7481 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7482 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
7483 MODULE_VERSION(LUSTRE_VERSION_STRING);
7484 MODULE_LICENSE("GPL");
7486 module_init(osd_init);
7487 module_exit(osd_exit);