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;
807 * It is called internally by ->readdir() to filter out the
808 * local slave object's FID of the striped directory.
810 * \retval 1 found the local slave's FID
811 * \retval 0 continue to check next item
812 * \retval -ve for failure
814 #ifdef HAVE_FILLDIR_USE_CTX
815 static int osd_stripe_dir_filldir(struct dir_context *buf,
817 static int osd_stripe_dir_filldir(void *buf,
819 const char *name, int namelen,
820 loff_t offset, __u64 ino, unsigned d_type)
822 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
823 struct osd_thread_info *oti = oclb->oclb_info;
824 struct lu_fid *fid = &oti->oti_fid3;
825 struct osd_inode_id *id = &oti->oti_id3;
826 struct osd_device *dev = oclb->oclb_dev;
827 struct osd_idmap_cache *oic = oclb->oclb_oic;
835 sscanf(name + 1, SFID, RFID(fid));
836 if (!fid_is_sane(fid))
839 if (osd_remote_fid(oti->oti_env, dev, fid))
842 osd_id_gen(id, ino, OSD_OII_NOGEN);
843 inode = osd_iget(oti, dev, id);
845 return PTR_ERR(inode);
848 osd_add_oi_cache(oti, dev, id, fid);
852 rc = osd_oii_insert(dev, oic, true);
854 return rc == 0 ? 1 : rc;
857 /* When lookup item under striped directory, we need to locate the master
858 * MDT-object of the striped directory firstly, then the client will send
859 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
860 * and the item's name. If the system is restored from MDT file level backup,
861 * then before the OI scrub completely built the OI files, the OI mappings of
862 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
863 * not a problem for the master MDT-object. Because when locate the master
864 * MDT-object, we will do name based lookup (for the striped directory itself)
865 * firstly, during such process we can setup the correct OI mapping for the
866 * master MDT-object. But it will be trouble for the slave MDT-object. Because
867 * the client will not trigger name based lookup on the MDT to locate the slave
868 * MDT-object before locating item under the striped directory, then when
869 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
870 * is invalid and does not know what the right OI mapping is, then the MDT has
871 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
872 * the OI file, related OI mapping is unknown yet, please try again later. And
873 * then client will re-try the RPC again and again until related OI mapping has
874 * been updated. That is quite inefficient.
876 * To resolve above trouble, we will handle it as the following two cases:
878 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
879 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
880 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
881 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
882 * directly. Please check osd_fid_lookup().
884 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
885 * Under such case, during lookup the master MDT-object, we will lookup the
886 * slave MDT-object via readdir against the master MDT-object, because the
887 * slave MDT-objects information are stored as sub-directories with the name
888 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
889 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
890 * the correct OI mapping for the slave MDT-object. */
891 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
892 struct inode *inode, struct osd_idmap_cache *oic)
894 struct lu_buf *buf = &oti->oti_big_buf;
895 struct dentry *dentry = &oti->oti_obj_dentry;
896 struct file *filp = &oti->oti_file;
897 const struct file_operations *fops;
898 struct lmv_mds_md_v1 *lmv1;
899 struct osd_check_lmv_buf oclb = {
900 #ifdef HAVE_DIR_CONTEXT
901 .ctx.actor = osd_stripe_dir_filldir,
911 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
914 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
916 lu_buf_realloc(buf, rc);
917 if (buf->lb_buf == NULL)
918 GOTO(out, rc = -ENOMEM);
924 if (unlikely(rc == 0 || rc == -ENODATA))
930 if (unlikely(buf->lb_buf == NULL)) {
931 lu_buf_realloc(buf, rc);
932 if (buf->lb_buf == NULL)
933 GOTO(out, rc = -ENOMEM);
939 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
943 dentry->d_inode = inode;
944 dentry->d_sb = inode->i_sb;
946 filp->f_path.dentry = dentry;
947 filp->f_mode = FMODE_64BITHASH;
948 filp->f_mapping = inode->i_mapping;
950 filp->private_data = NULL;
951 set_file_inode(filp, inode);
953 #ifdef HAVE_DIR_CONTEXT
954 oclb.ctx.pos = filp->f_pos;
955 rc = fops->iterate(filp, &oclb.ctx);
956 filp->f_pos = oclb.ctx.pos;
958 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
960 fops->release(inode, filp);
964 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
965 DFID": rc = %d\n", osd_ino2name(inode),
966 inode->i_ino, inode->i_generation,
967 PFID(&oic->oic_fid), rc);
974 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
975 const struct lu_fid *fid,
976 const struct lu_object_conf *conf)
978 struct osd_thread_info *info;
979 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
980 struct osd_device *dev;
981 struct osd_idmap_cache *oic;
982 struct osd_inode_id *id;
983 struct inode *inode = NULL;
984 struct osd_scrub *scrub;
985 struct scrub_file *sf;
986 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
993 bool updated = false;
996 LINVRNT(osd_invariant(obj));
997 LASSERT(obj->oo_inode == NULL);
998 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1000 dev = osd_dev(ldev);
1001 scrub = &dev->od_scrub;
1002 sf = &scrub->os_file;
1003 info = osd_oti_get(env);
1005 oic = &info->oti_cache;
1007 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1010 /* For the object is created as locking anchor, or for the object to
1011 * be created on disk. No need to osd_oi_lookup() at here because FID
1012 * shouldn't never be re-used, if it's really a duplicate FID from
1013 * unexpected reason, we should be able to detect it later by calling
1014 * do_create->osd_oi_insert(). */
1015 if (conf && conf->loc_flags & LOC_F_NEW)
1016 GOTO(out, result = 0);
1018 /* Search order: 1. per-thread cache. */
1019 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1025 if (!list_empty(&scrub->os_inconsistent_items)) {
1026 /* Search order: 2. OI scrub pending list. */
1027 result = osd_oii_lookup(dev, fid, id);
1032 /* The OI mapping in the OI file can be updated by the OI scrub
1033 * when we locate the inode via FID. So it may be not trustable. */
1036 /* Search order: 3. OI files. */
1037 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1038 if (result == -ENOENT) {
1039 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1040 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1041 !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1043 GOTO(out, result = 0);
1052 obj->oo_inode = NULL;
1053 /* for later passes through checks, not true on first pass */
1054 if (!IS_ERR_OR_NULL(inode))
1057 inode = osd_iget_check(info, dev, fid, id, trusted);
1058 if (!IS_ERR(inode)) {
1059 obj->oo_inode = inode;
1067 result = PTR_ERR(inode);
1068 if (result == -ENOENT || result == -ESTALE)
1069 GOTO(out, result = 0);
1071 if (result != -EREMCHG)
1075 /* We still have chance to get the valid inode: for the
1076 * object which is referenced by remote name entry, the
1077 * object on the local MDT will be linked under the dir
1078 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1080 * We do not know whether the object for the given FID
1081 * is referenced by some remote name entry or not, and
1082 * especially for DNE II, a multiple-linked object may
1083 * have many name entries reside on many MDTs.
1085 * To simplify the operation, OSD will not distinguish
1086 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1087 * only happened for the RPC from other MDT during the
1088 * OI scrub, or for the client side RPC with FID only,
1089 * such as FID to path, or from old connected client. */
1090 if (!remote && !fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1091 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1095 flags |= SS_AUTO_PARTIAL;
1096 flags &= ~SS_AUTO_FULL;
1101 if (thread_is_running(&scrub->os_thread)) {
1102 if (scrub->os_partial_scan && !scrub->os_in_join)
1105 if (IS_ERR_OR_NULL(inode) || result)
1106 GOTO(out, result = -EINPROGRESS);
1109 LASSERT(obj->oo_inode == inode);
1111 osd_add_oi_cache(info, dev, id, fid);
1112 osd_oii_insert(dev, oic, true);
1116 if (dev->od_noscrub) {
1118 GOTO(out, result = -EREMCHG);
1121 LASSERT(obj->oo_inode == inode);
1123 osd_add_oi_cache(info, dev, id, fid);
1128 rc1 = osd_scrub_start(dev, flags);
1129 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1130 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1131 if (rc1 && rc1 != -EALREADY)
1132 GOTO(out, result = -EREMCHG);
1134 if (IS_ERR_OR_NULL(inode) || result)
1135 GOTO(out, result = -EINPROGRESS);
1138 LASSERT(obj->oo_inode == inode);
1140 osd_add_oi_cache(info, dev, id, fid);
1141 osd_oii_insert(dev, oic, true);
1145 result = osd_check_lma(env, obj);
1149 LASSERTF(id->oii_ino == inode->i_ino &&
1150 id->oii_gen == inode->i_generation,
1151 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1152 PFID(fid), id->oii_ino, id->oii_gen,
1153 inode->i_ino, inode->i_generation);
1155 saved_ino = inode->i_ino;
1156 saved_gen = inode->i_generation;
1158 if (unlikely(result == -ENODATA)) {
1159 /* If the OI scrub updated the OI mapping by race, it
1160 * must be valid. Trust the inode that has no LMA EA. */
1164 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1166 /* The OI mapping is still there, the inode is still
1167 * valid. It is just becaues the inode has no LMA EA. */
1168 if (saved_ino == id->oii_ino &&
1169 saved_gen == id->oii_gen)
1172 /* It is the OI scrub updated the OI mapping by race.
1173 * The new OI mapping must be valid. */
1179 /* "result == -ENOENT" means that the OI mappinghas been
1180 * removed by race, so the inode belongs to other object.
1182 * Others error can be returned directly. */
1183 if (result == -ENOENT) {
1186 obj->oo_inode = NULL;
1191 if (result != -EREMCHG)
1196 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1197 /* "result == -ENOENT" means the cached OI mapping has been removed
1198 * from the OI file by race, above inode belongs to other object. */
1199 if (result == -ENOENT) {
1202 obj->oo_inode = NULL;
1203 GOTO(out, result = 0);
1209 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1214 /* It is the OI scrub updated the OI mapping by race.
1215 * The new OI mapping must be valid. */
1221 obj->oo_compat_dot_created = 1;
1222 obj->oo_compat_dotdot_created = 1;
1224 if (S_ISDIR(inode->i_mode) &&
1225 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1226 osd_check_lmv(info, dev, inode, oic);
1228 result = osd_attach_jinode(inode);
1233 GOTO(out, result = 0);
1235 LASSERT(!obj->oo_hl_head);
1236 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1238 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1241 if (result || !obj->oo_inode) {
1242 if (!IS_ERR_OR_NULL(inode))
1245 obj->oo_inode = NULL;
1247 fid_zero(&oic->oic_fid);
1250 LINVRNT(osd_invariant(obj));
1255 * Concurrency: shouldn't matter.
1257 static void osd_object_init0(struct osd_object *obj)
1259 LASSERT(obj->oo_inode != NULL);
1260 obj->oo_dt.do_body_ops = &osd_body_ops;
1261 obj->oo_dt.do_lu.lo_header->loh_attr |=
1262 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1266 * Concurrency: no concurrent access is possible that early in object
1269 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1270 const struct lu_object_conf *conf)
1272 struct osd_object *obj = osd_obj(l);
1275 LINVRNT(osd_invariant(obj));
1277 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1278 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1279 l->lo_header->loh_attr |= LOHA_EXISTS;
1283 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1284 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1285 if (result == 0 && obj->oo_inode != NULL) {
1286 struct osd_thread_info *oti = osd_oti_get(env);
1287 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1289 osd_object_init0(obj);
1290 result = osd_get_lma(oti, obj->oo_inode,
1291 &oti->oti_obj_dentry, loa);
1293 /* Convert LMAI flags to lustre LMA flags
1294 * and cache it to oo_lma_flags */
1296 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1297 } else if (result == -ENODATA) {
1302 LINVRNT(osd_invariant(obj));
1306 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1307 * The left part is for value, binary mode. */
1308 struct osd_xattr_entry {
1309 struct list_head oxe_list;
1313 struct rcu_head oxe_rcu;
1317 static int osd_oxc_get(struct osd_object *obj, const char *name,
1320 struct osd_xattr_entry *tmp;
1321 struct osd_xattr_entry *oxe = NULL;
1322 size_t namelen = strlen(name);
1327 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1328 if (namelen == tmp->oxe_namelen &&
1329 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1336 GOTO(out, rc = -ENOENT);
1338 if (!oxe->oxe_exist)
1339 GOTO(out, rc = -ENODATA);
1342 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1345 if (buf->lb_buf == NULL)
1348 if (buf->lb_len < rc)
1349 GOTO(out, rc = -ERANGE);
1351 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1359 static void osd_oxc_free(struct rcu_head *head)
1361 struct osd_xattr_entry *oxe;
1363 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1364 OBD_FREE(oxe, oxe->oxe_len);
1367 static void osd_oxc_add(struct osd_object *obj, const char *name,
1368 const char *buf, int buflen)
1370 struct osd_xattr_entry *oxe;
1371 struct osd_xattr_entry *old = NULL;
1372 struct osd_xattr_entry *tmp;
1373 size_t namelen = strlen(name);
1374 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1376 OBD_ALLOC(oxe, len);
1380 INIT_LIST_HEAD(&oxe->oxe_list);
1382 oxe->oxe_namelen = namelen;
1383 memcpy(oxe->oxe_buf, name, namelen);
1385 LASSERT(buf != NULL);
1386 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1387 oxe->oxe_exist = true;
1389 oxe->oxe_exist = false;
1392 /* this should be rarely called, just remove old and add new */
1393 spin_lock(&obj->oo_guard);
1394 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1395 if (namelen == tmp->oxe_namelen &&
1396 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1402 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1403 call_rcu(&old->oxe_rcu, osd_oxc_free);
1405 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1407 spin_unlock(&obj->oo_guard);
1410 static void osd_oxc_del(struct osd_object *obj, const char *name)
1412 struct osd_xattr_entry *oxe;
1413 size_t namelen = strlen(name);
1415 spin_lock(&obj->oo_guard);
1416 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1417 if (namelen == oxe->oxe_namelen &&
1418 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1419 list_del_rcu(&oxe->oxe_list);
1420 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1424 spin_unlock(&obj->oo_guard);
1427 static void osd_oxc_fini(struct osd_object *obj)
1429 struct osd_xattr_entry *oxe, *next;
1431 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1432 list_del(&oxe->oxe_list);
1433 OBD_FREE(oxe, oxe->oxe_len);
1438 * Concurrency: no concurrent access is possible that late in object
1441 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1443 struct osd_object *obj = osd_obj(l);
1445 LINVRNT(osd_invariant(obj));
1448 dt_object_fini(&obj->oo_dt);
1449 if (obj->oo_hl_head != NULL)
1450 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1455 * Concurrency: no concurrent access is possible that late in object
1458 static void osd_index_fini(struct osd_object *o)
1460 struct iam_container *bag;
1462 if (o->oo_dir != NULL) {
1463 bag = &o->oo_dir->od_container;
1464 if (o->oo_inode != NULL) {
1465 if (bag->ic_object == o->oo_inode)
1466 iam_container_fini(bag);
1468 OBD_FREE_PTR(o->oo_dir);
1474 * Concurrency: no concurrent access is possible that late in object
1475 * life-cycle (for all existing callers, that is. New callers have to provide
1476 * their own locking.)
1478 static int osd_inode_unlinked(const struct inode *inode)
1480 return inode->i_nlink == 0;
1484 OSD_TXN_OI_DELETE_CREDITS = 20,
1485 OSD_TXN_INODE_DELETE_CREDITS = 20
1492 #if OSD_THANDLE_STATS
1494 * Set time when the handle is allocated
1496 static void osd_th_alloced(struct osd_thandle *oth)
1498 oth->oth_alloced = ktime_get();
1502 * Set time when the handle started
1504 static void osd_th_started(struct osd_thandle *oth)
1506 oth->oth_started = ktime_get();
1510 * Check whether the we deal with this handle for too long.
1512 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1513 ktime_t alloced, ktime_t started,
1516 ktime_t now = ktime_get();
1518 LASSERT(dev != NULL);
1520 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1521 ktime_us_delta(started, alloced));
1522 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1523 ktime_us_delta(closed, started));
1524 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1525 ktime_us_delta(now, closed));
1527 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1528 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1529 oth, now, alloced, started, closed);
1530 libcfs_debug_dumpstack(NULL);
1534 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1536 ktime_t __closed = ktime_get(); \
1537 ktime_t __alloced = oth->oth_alloced; \
1538 ktime_t __started = oth->oth_started; \
1541 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1544 #else /* OSD_THANDLE_STATS */
1546 #define osd_th_alloced(h) do {} while(0)
1547 #define osd_th_started(h) do {} while(0)
1548 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1550 #endif /* OSD_THANDLE_STATS */
1553 * Concurrency: doesn't access mutable data.
1555 static int osd_param_is_not_sane(const struct osd_device *dev,
1556 const struct thandle *th)
1558 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1560 return oh->ot_credits > osd_transaction_size(dev);
1564 * Concurrency: shouldn't matter.
1566 static void osd_trans_commit_cb(struct super_block *sb,
1567 struct ldiskfs_journal_cb_entry *jcb, int error)
1569 struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1570 struct thandle *th = &oh->ot_super;
1571 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1572 struct dt_txn_commit_cb *dcb, *tmp;
1574 LASSERT(oh->ot_handle == NULL);
1577 CERROR("transaction @0x%p commit error: %d\n", th, error);
1579 dt_txn_hook_commit(th);
1581 /* call per-transaction callbacks if any */
1582 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1584 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1585 "commit callback entry: magic=%x name='%s'\n",
1586 dcb->dcb_magic, dcb->dcb_name);
1587 list_del_init(&dcb->dcb_linkage);
1588 dcb->dcb_func(NULL, th, dcb, error);
1591 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1595 lu_context_exit(&th->th_ctx);
1596 lu_context_fini(&th->th_ctx);
1600 #ifndef HAVE_SB_START_WRITE
1601 # define sb_start_write(sb) do {} while (0)
1602 # define sb_end_write(sb) do {} while (0)
1605 static struct thandle *osd_trans_create(const struct lu_env *env,
1606 struct dt_device *d)
1608 struct osd_thread_info *oti = osd_oti_get(env);
1609 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1610 struct osd_thandle *oh;
1615 CERROR("%s: someone try to start transaction under "
1616 "readonly mode, should be disabled.\n",
1617 osd_name(osd_dt_dev(d)));
1619 RETURN(ERR_PTR(-EROFS));
1622 /* on pending IO in this thread should left from prev. request */
1623 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1625 sb_start_write(osd_sb(osd_dt_dev(d)));
1627 OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1629 oh->ot_quota_trans = &oti->oti_quota_trans;
1630 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1634 th->th_tags = LCT_TX_HANDLE;
1636 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1637 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1640 memset(oti->oti_declare_ops, 0,
1641 sizeof(oti->oti_declare_ops));
1642 memset(oti->oti_declare_ops_cred, 0,
1643 sizeof(oti->oti_declare_ops_cred));
1644 memset(oti->oti_declare_ops_used, 0,
1645 sizeof(oti->oti_declare_ops_used));
1647 sb_end_write(osd_sb(osd_dt_dev(d)));
1648 th = ERR_PTR(-ENOMEM);
1653 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1655 struct osd_thread_info *oti = osd_oti_get(env);
1656 struct osd_thandle *oh;
1658 oh = container_of0(th, struct osd_thandle, ot_super);
1659 LASSERT(oh != NULL);
1661 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1662 oti->oti_declare_ops[OSD_OT_CREATE],
1663 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1664 oti->oti_declare_ops_used[OSD_OT_CREATE],
1665 oti->oti_declare_ops[OSD_OT_DESTROY],
1666 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1667 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1668 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1669 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1670 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1671 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1672 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1673 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1674 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1675 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1676 oti->oti_declare_ops[OSD_OT_WRITE],
1677 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1678 oti->oti_declare_ops_used[OSD_OT_WRITE],
1679 oti->oti_declare_ops[OSD_OT_PUNCH],
1680 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1681 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1682 oti->oti_declare_ops[OSD_OT_QUOTA],
1683 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1684 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1685 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1686 oti->oti_declare_ops[OSD_OT_INSERT],
1687 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1688 oti->oti_declare_ops_used[OSD_OT_INSERT],
1689 oti->oti_declare_ops[OSD_OT_DELETE],
1690 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1691 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1692 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1693 oti->oti_declare_ops[OSD_OT_REF_ADD],
1694 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1695 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1696 oti->oti_declare_ops[OSD_OT_REF_DEL],
1697 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1698 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1702 * Concurrency: shouldn't matter.
1704 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1707 struct osd_thread_info *oti = osd_oti_get(env);
1708 struct osd_device *dev = osd_dt_dev(d);
1710 struct osd_thandle *oh;
1715 LASSERT(current->journal_info == NULL);
1717 oh = container_of0(th, struct osd_thandle, ot_super);
1718 LASSERT(oh != NULL);
1719 LASSERT(oh->ot_handle == NULL);
1721 rc = dt_txn_hook_start(env, d, th);
1725 if (unlikely(osd_param_is_not_sane(dev, th))) {
1726 static unsigned long last_printed;
1727 static int last_credits;
1729 /* don't make noise on a tiny testing systems
1730 * actual credits misuse will be caught anyway */
1731 if (last_credits != oh->ot_credits &&
1732 time_after(jiffies, last_printed +
1733 msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1734 osd_transaction_size(dev) > 512) {
1735 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1736 oh->ot_credits, osd_transaction_size(dev));
1737 osd_trans_dump_creds(env, th);
1738 libcfs_debug_dumpstack(NULL);
1739 last_credits = oh->ot_credits;
1740 last_printed = jiffies;
1742 /* XXX Limit the credits to 'max_transaction_buffers', and
1743 * let the underlying filesystem to catch the error if
1744 * we really need so many credits.
1746 * This should be removed when we can calculate the
1747 * credits precisely. */
1748 oh->ot_credits = osd_transaction_size(dev);
1749 } else if (ldiskfs_track_declares_assert != 0) {
1750 /* reserve few credits to prevent an assertion in JBD
1751 * our debugging mechanism will be able to detected
1752 * overuse. this can help to debug single-update
1754 oh->ot_credits += 10;
1755 if (unlikely(osd_param_is_not_sane(dev, th)))
1756 oh->ot_credits = osd_transaction_size(dev);
1760 * XXX temporary stuff. Some abstraction layer should
1763 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1767 LASSERT(oti->oti_txns == 0);
1768 lu_context_init(&th->th_ctx, th->th_tags);
1769 lu_context_enter(&th->th_ctx);
1771 lu_device_get(&d->dd_lu_dev);
1772 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1783 static int osd_seq_exists(const struct lu_env *env,
1784 struct osd_device *osd, u64 seq)
1786 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1787 struct seq_server_site *ss = osd_seq_site(osd);
1791 LASSERT(ss != NULL);
1792 LASSERT(ss->ss_server_fld != NULL);
1794 rc = osd_fld_lookup(env, osd, seq, range);
1797 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1798 osd_name(osd), seq, rc);
1802 RETURN(ss->ss_node_id == range->lsr_index);
1805 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1807 struct dt_txn_commit_cb *dcb;
1808 struct dt_txn_commit_cb *tmp;
1810 /* call per-transaction stop callbacks if any */
1811 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1813 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1814 "commit callback entry: magic=%x name='%s'\n",
1815 dcb->dcb_magic, dcb->dcb_name);
1816 list_del_init(&dcb->dcb_linkage);
1817 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1822 * Concurrency: shouldn't matter.
1824 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1827 int rc = 0, remove_agents = 0;
1828 struct osd_thandle *oh;
1829 struct osd_thread_info *oti = osd_oti_get(env);
1830 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1831 struct osd_device *osd = osd_dt_dev(th->th_dev);
1832 struct qsd_instance *qsd = osd->od_quota_slave;
1833 struct lquota_trans *qtrans;
1836 oh = container_of0(th, struct osd_thandle, ot_super);
1838 /* reset OI cache for safety */
1839 oti->oti_ins_cache_used = 0;
1841 remove_agents = oh->ot_remove_agents;
1843 qtrans = oh->ot_quota_trans;
1844 oh->ot_quota_trans = NULL;
1846 if (oh->ot_handle != NULL) {
1848 handle_t *hdl = oh->ot_handle;
1851 * add commit callback
1852 * notice we don't do this in osd_trans_start()
1853 * as underlying transaction can change during truncate
1855 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1858 LASSERT(oti->oti_txns == 1);
1861 rc = dt_txn_hook_stop(env, th);
1863 CERROR("%s: failed in transaction hook: rc = %d\n",
1866 osd_trans_stop_cb(oh, rc);
1867 /* hook functions might modify th_sync */
1868 hdl->h_sync = th->th_sync;
1870 oh->ot_handle = NULL;
1871 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1873 CERROR("%s: failed to stop transaction: rc = %d\n",
1874 osd_name(osd), rc2);
1878 osd_trans_stop_cb(oh, th->th_result);
1882 /* inform the quota slave device that the transaction is stopping */
1883 qsd_op_end(env, qsd, qtrans);
1885 /* as we want IO to journal and data IO be concurrent, we don't block
1886 * awaiting data IO completion in osd_do_bio(), instead we wait here
1887 * once transaction is submitted to the journal. all reqular requests
1888 * don't do direct IO (except read/write), thus this wait_event becomes
1891 * IMPORTANT: we have to wait till any IO submited by the thread is
1892 * completed otherwise iobuf may be corrupted by different request
1894 wait_event(iobuf->dr_wait,
1895 atomic_read(&iobuf->dr_numreqs) == 0);
1896 osd_fini_iobuf(osd, iobuf);
1898 rc = iobuf->dr_error;
1900 if (unlikely(remove_agents != 0))
1901 osd_process_scheduled_agent_removals(env, osd);
1903 sb_end_write(osd_sb(osd));
1908 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1910 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1913 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1914 LASSERT(&dcb->dcb_func != NULL);
1915 if (dcb->dcb_flags & DCB_TRANS_STOP)
1916 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1918 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1924 * Called just before object is freed. Releases all resources except for
1925 * object itself (that is released by osd_object_free()).
1927 * Concurrency: no concurrent access is possible that late in object
1930 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1932 struct osd_object *obj = osd_obj(l);
1933 struct inode *inode = obj->oo_inode;
1935 LINVRNT(osd_invariant(obj));
1938 * If object is unlinked remove fid->ino mapping from object index.
1941 osd_index_fini(obj);
1942 if (inode != NULL) {
1943 struct qsd_instance *qsd = osd_obj2dev(obj)->od_quota_slave;
1944 qid_t uid = i_uid_read(inode);
1945 qid_t gid = i_gid_read(inode);
1948 obj->oo_inode = NULL;
1951 struct osd_thread_info *info = osd_oti_get(env);
1952 struct lquota_id_info *qi = &info->oti_qi;
1954 /* Release granted quota to master if necessary */
1955 qi->lqi_id.qid_uid = uid;
1956 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1958 qi->lqi_id.qid_uid = gid;
1959 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1961 qi->lqi_id.qid_uid = i_projid_read(inode);
1962 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1968 * Concurrency: ->loo_object_release() is called under site spin-lock.
1970 static void osd_object_release(const struct lu_env *env,
1971 struct lu_object *l)
1973 struct osd_object *o = osd_obj(l);
1974 /* nobody should be releasing a non-destroyed object with nlink=0
1975 * the API allows this, but ldiskfs doesn't like and then report
1976 * this inode as deleted */
1977 if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1982 * Concurrency: shouldn't matter.
1984 static int osd_object_print(const struct lu_env *env, void *cookie,
1985 lu_printer_t p, const struct lu_object *l)
1987 struct osd_object *o = osd_obj(l);
1988 struct iam_descr *d;
1990 if (o->oo_dir != NULL)
1991 d = o->oo_dir->od_container.ic_descr;
1994 return (*p)(env, cookie,
1995 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
1997 o->oo_inode ? o->oo_inode->i_ino : 0UL,
1998 o->oo_inode ? o->oo_inode->i_generation : 0,
1999 d ? d->id_ops->id_name : "plain");
2003 * Concurrency: shouldn't matter.
2005 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2006 struct obd_statfs *sfs)
2008 struct osd_device *osd = osd_dt_dev(d);
2009 struct super_block *sb = osd_sb(osd);
2010 struct kstatfs *ksfs;
2014 if (unlikely(osd->od_mnt == NULL))
2015 return -EINPROGRESS;
2017 /* osd_lproc.c call this without env, allocate ksfs for that case */
2018 if (unlikely(env == NULL)) {
2019 OBD_ALLOC_PTR(ksfs);
2023 ksfs = &osd_oti_get(env)->oti_ksfs;
2026 result = sb->s_op->statfs(sb->s_root, ksfs);
2030 statfs_pack(sfs, ksfs);
2031 if (unlikely(sb->s_flags & MS_RDONLY))
2032 sfs->os_state |= OS_STATE_READONLY;
2033 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2034 LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2035 sfs->os_maxbytes = sb->s_maxbytes;
2037 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2040 * Reserve some space so to avoid fragmenting the filesystem too much.
2041 * Fragmentation not only impacts performance, but can also increase
2042 * metadata overhead significantly, causing grant calculation to be
2045 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2047 CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2048 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2049 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2050 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2052 sfs->os_blocks -= reserved;
2053 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2054 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2057 if (unlikely(env == NULL))
2063 * Estimate space needed for file creations. We assume the largest filename
2064 * which is 2^64 - 1, hence a filename of 20 chars.
2065 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2067 #ifdef __LDISKFS_DIR_REC_LEN
2068 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2070 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2074 * Concurrency: doesn't access mutable data.
2076 static void osd_conf_get(const struct lu_env *env,
2077 const struct dt_device *dev,
2078 struct dt_device_param *param)
2080 struct super_block *sb = osd_sb(osd_dt_dev(dev));
2084 * XXX should be taken from not-yet-existing fs abstraction layer.
2086 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2087 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2088 param->ddp_symlink_max = sb->s_blocksize;
2089 param->ddp_mount_type = LDD_MT_LDISKFS;
2090 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2091 param->ddp_maxbytes = sb->s_maxbytes;
2093 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2094 /* inode are statically allocated, so per-inode space consumption
2095 * is the space consumed by the directory entry */
2096 param->ddp_inodespace = PER_OBJ_USAGE;
2097 /* EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2098 * = 128MB) which is unlikely to be hit in real life. Report a smaller
2099 * maximum length to not under count the actual number of extents
2100 * needed for writing a file. */
2101 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2102 /* worst-case extent insertion metadata overhead */
2103 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2104 param->ddp_mntopts = 0;
2105 if (test_opt(sb, XATTR_USER))
2106 param->ddp_mntopts |= MNTOPT_USERXATTR;
2107 if (test_opt(sb, POSIX_ACL))
2108 param->ddp_mntopts |= MNTOPT_ACL;
2110 /* LOD might calculate the max stripe count based on max_ea_size,
2111 * so we need take account in the overhead as well,
2112 * xattr_header + magic + xattr_entry_head */
2113 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2114 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2116 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2117 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2118 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2122 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2126 * Concurrency: shouldn't matter.
2128 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2132 CDEBUG(D_CACHE, "%s: syncing OSD\n", osd_dt_dev(d)->od_svname);
2134 rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2136 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2143 * Start commit for OSD device.
2145 * An implementation of dt_commit_async method for OSD device.
2146 * Asychronously starts underlayng fs sync and thereby a transaction
2149 * \param env environment
2150 * \param d dt device
2152 * \see dt_device_operations
2154 static int osd_commit_async(const struct lu_env *env,
2155 struct dt_device *d)
2157 struct super_block *s = osd_sb(osd_dt_dev(d));
2160 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2161 RETURN(s->s_op->sync_fs(s, 0));
2164 /* Our own copy of the set readonly functions if present, or NU if not. */
2165 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2166 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2167 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2170 * Concurrency: shouldn't matter.
2172 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2174 struct super_block *sb = osd_sb(osd_dt_dev(d));
2175 struct block_device *dev = sb->s_bdev;
2176 int rc = -EOPNOTSUPP;
2179 if (priv_dev_set_rdonly) {
2180 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2183 CERROR("*** setting %s read-only ***\n",
2184 osd_dt_dev(d)->od_svname);
2186 if (sb->s_op->freeze_fs) {
2187 rc = sb->s_op->freeze_fs(sb);
2192 if (jdev && (jdev != dev)) {
2193 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2195 priv_dev_set_rdonly(jdev);
2197 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2198 priv_dev_set_rdonly(dev);
2200 if (sb->s_op->unfreeze_fs)
2201 sb->s_op->unfreeze_fs(sb);
2206 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2207 osd_dt_dev(d)->od_svname, (long)dev, rc);
2213 * Note: we do not count into QUOTA here.
2214 * If we mount with --data_journal we may need more.
2216 const int osd_dto_credits_noquota[DTO_NR] = {
2219 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2220 * SINGLEDATA_TRANS_BLOCKS(8)
2221 * XXX Note: maybe iam need more, since iam have more level than
2224 [DTO_INDEX_INSERT] = 16,
2227 * just modify a single entry, probably merge few within a block
2229 [DTO_INDEX_DELETE] = 1,
2233 [DTO_INDEX_UPDATE] = 16,
2235 * 4(inode, inode bits, groups, GDT)
2236 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2238 [DTO_OBJECT_CREATE] = 4,
2240 * 4(inode, inode bits, groups, GDT)
2241 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2243 [DTO_OBJECT_DELETE] = 4,
2245 * Attr set credits (inode)
2247 [DTO_ATTR_SET_BASE] = 1,
2249 * Xattr set. The same as xattr of EXT3.
2250 * DATA_TRANS_BLOCKS(14)
2251 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2252 * are also counted in. Do not know why?
2254 [DTO_XATTR_SET] = 14,
2256 * credits for inode change during write.
2258 [DTO_WRITE_BASE] = 3,
2260 * credits for single block write.
2262 [DTO_WRITE_BLOCK] = 14,
2264 * Attr set credits for chown.
2265 * This is extra credits for setattr, and it is null without quota
2267 [DTO_ATTR_SET_CHOWN] = 0
2270 static const struct dt_device_operations osd_dt_ops = {
2271 .dt_root_get = osd_root_get,
2272 .dt_statfs = osd_statfs,
2273 .dt_trans_create = osd_trans_create,
2274 .dt_trans_start = osd_trans_start,
2275 .dt_trans_stop = osd_trans_stop,
2276 .dt_trans_cb_add = osd_trans_cb_add,
2277 .dt_conf_get = osd_conf_get,
2278 .dt_sync = osd_sync,
2280 .dt_commit_async = osd_commit_async,
2283 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2286 struct osd_object *obj = osd_dt_obj(dt);
2287 struct osd_thread_info *oti = osd_oti_get(env);
2289 LINVRNT(osd_invariant(obj));
2291 LASSERT(obj->oo_owner != env);
2292 down_read_nested(&obj->oo_sem, role);
2294 LASSERT(obj->oo_owner == NULL);
2298 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2301 struct osd_object *obj = osd_dt_obj(dt);
2302 struct osd_thread_info *oti = osd_oti_get(env);
2304 LINVRNT(osd_invariant(obj));
2306 LASSERT(obj->oo_owner != env);
2307 down_write_nested(&obj->oo_sem, role);
2309 LASSERT(obj->oo_owner == NULL);
2310 obj->oo_owner = env;
2314 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2316 struct osd_object *obj = osd_dt_obj(dt);
2317 struct osd_thread_info *oti = osd_oti_get(env);
2319 LINVRNT(osd_invariant(obj));
2321 LASSERT(oti->oti_r_locks > 0);
2323 up_read(&obj->oo_sem);
2326 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2328 struct osd_object *obj = osd_dt_obj(dt);
2329 struct osd_thread_info *oti = osd_oti_get(env);
2331 LINVRNT(osd_invariant(obj));
2333 LASSERT(obj->oo_owner == env);
2334 LASSERT(oti->oti_w_locks > 0);
2336 obj->oo_owner = NULL;
2337 up_write(&obj->oo_sem);
2340 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2342 struct osd_object *obj = osd_dt_obj(dt);
2344 LINVRNT(osd_invariant(obj));
2346 return obj->oo_owner == env;
2349 static struct timespec *osd_inode_time(const struct lu_env *env,
2350 struct inode *inode, __u64 seconds)
2352 struct osd_thread_info *oti = osd_oti_get(env);
2353 struct timespec *t = &oti->oti_time;
2355 t->tv_sec = seconds;
2357 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2361 static void osd_inode_getattr(const struct lu_env *env,
2362 struct inode *inode, struct lu_attr *attr)
2364 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2365 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2366 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2367 LA_BLKSIZE | LA_TYPE;
2369 attr->la_atime = LTIME_S(inode->i_atime);
2370 attr->la_mtime = LTIME_S(inode->i_mtime);
2371 attr->la_ctime = LTIME_S(inode->i_ctime);
2372 attr->la_mode = inode->i_mode;
2373 attr->la_size = i_size_read(inode);
2374 attr->la_blocks = inode->i_blocks;
2375 attr->la_uid = i_uid_read(inode);
2376 attr->la_gid = i_gid_read(inode);
2377 attr->la_projid = i_projid_read(inode);
2378 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2379 attr->la_nlink = inode->i_nlink;
2380 attr->la_rdev = inode->i_rdev;
2381 attr->la_blksize = 1 << inode->i_blkbits;
2382 attr->la_blkbits = inode->i_blkbits;
2384 * Ext4 did not transfer inherit flags from raw inode
2385 * to inode flags, and ext4 internally test raw inode
2386 * @i_flags directly. Instead of patching ext4, we do it here.
2388 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2389 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2392 static int osd_attr_get(const struct lu_env *env,
2393 struct dt_object *dt,
2394 struct lu_attr *attr)
2396 struct osd_object *obj = osd_dt_obj(dt);
2398 if (unlikely(!dt_object_exists(dt)))
2400 if (unlikely(obj->oo_destroyed))
2403 LASSERT(!dt_object_remote(dt));
2404 LINVRNT(osd_invariant(obj));
2406 spin_lock(&obj->oo_guard);
2407 osd_inode_getattr(env, obj->oo_inode, attr);
2408 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2409 attr->la_flags |= LUSTRE_ORPHAN_FL;
2410 spin_unlock(&obj->oo_guard);
2415 static int osd_declare_attr_qid(const struct lu_env *env,
2416 struct osd_object *obj,
2417 struct osd_thandle *oh, long long bspace,
2418 qid_t old_id, qid_t new_id, bool enforce,
2422 struct osd_thread_info *info = osd_oti_get(env);
2423 struct lquota_id_info *qi = &info->oti_qi;
2425 qi->lqi_type = type;
2426 /* inode accounting */
2427 qi->lqi_is_blk = false;
2429 /* one more inode for the new id ... */
2430 qi->lqi_id.qid_uid = new_id;
2432 /* Reserve credits for the new id */
2433 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2434 if (rc == -EDQUOT || rc == -EINPROGRESS)
2439 /* and one less inode for the current id */
2440 qi->lqi_id.qid_uid = old_id;
2442 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2443 if (rc == -EDQUOT || rc == -EINPROGRESS)
2448 /* block accounting */
2449 qi->lqi_is_blk = true;
2451 /* more blocks for the new id ... */
2452 qi->lqi_id.qid_uid = new_id;
2453 qi->lqi_space = bspace;
2455 * Credits for the new uid has been reserved, re-use "obj"
2456 * to save credit reservation.
2458 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2459 if (rc == -EDQUOT || rc == -EINPROGRESS)
2464 /* and finally less blocks for the current uid */
2465 qi->lqi_id.qid_uid = old_id;
2466 qi->lqi_space = -bspace;
2467 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2468 if (rc == -EDQUOT || rc == -EINPROGRESS)
2474 static int osd_declare_attr_set(const struct lu_env *env,
2475 struct dt_object *dt,
2476 const struct lu_attr *attr,
2477 struct thandle *handle)
2479 struct osd_thandle *oh;
2480 struct osd_object *obj;
2488 LASSERT(dt != NULL);
2489 LASSERT(handle != NULL);
2491 obj = osd_dt_obj(dt);
2492 LASSERT(osd_invariant(obj));
2494 oh = container_of0(handle, struct osd_thandle, ot_super);
2495 LASSERT(oh->ot_handle == NULL);
2497 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2498 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2500 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2501 osd_dto_credits_noquota[DTO_XATTR_SET]);
2503 if (attr == NULL || obj->oo_inode == NULL)
2506 bspace = obj->oo_inode->i_blocks;
2507 bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2508 bspace = toqb(bspace);
2510 /* Changing ownership is always preformed by super user, it should not
2513 * We still need to call the osd_declare_qid() to calculate the journal
2514 * credits for updating quota accounting files and to trigger quota
2515 * space adjustment once the operation is completed.*/
2516 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2518 uid = i_uid_read(obj->oo_inode);
2519 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2520 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2521 attr->la_uid, enforce, USRQUOTA);
2525 gid = i_gid_read(obj->oo_inode);
2526 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2527 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2528 i_gid_read(obj->oo_inode),
2529 attr->la_gid, enforce, GRPQUOTA);
2534 #ifdef HAVE_PROJECT_QUOTA
2535 if (attr->la_valid & LA_PROJID) {
2536 __u32 projid = i_projid_read(obj->oo_inode);
2537 enforce = (attr->la_valid & LA_PROJID) &&
2538 (attr->la_projid != projid);
2539 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2540 (qid_t)projid, (qid_t)attr->la_projid,
2549 static int osd_inode_setattr(const struct lu_env *env,
2550 struct inode *inode, const struct lu_attr *attr)
2552 __u64 bits = attr->la_valid;
2554 /* Only allow set size for regular file */
2555 if (!S_ISREG(inode->i_mode))
2556 bits &= ~(LA_SIZE | LA_BLOCKS);
2561 if (bits & LA_ATIME)
2562 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2563 if (bits & LA_CTIME)
2564 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2565 if (bits & LA_MTIME)
2566 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2567 if (bits & LA_SIZE) {
2568 spin_lock(&inode->i_lock);
2569 LDISKFS_I(inode)->i_disksize = attr->la_size;
2570 i_size_write(inode, attr->la_size);
2571 spin_unlock(&inode->i_lock);
2574 /* OSD should not change "i_blocks" which is used by quota.
2575 * "i_blocks" should be changed by ldiskfs only. */
2577 inode->i_mode = (inode->i_mode & S_IFMT) |
2578 (attr->la_mode & ~S_IFMT);
2580 i_uid_write(inode, attr->la_uid);
2582 i_gid_write(inode, attr->la_gid);
2583 if (bits & LA_PROJID)
2584 i_projid_write(inode, attr->la_projid);
2585 if (bits & LA_NLINK)
2586 set_nlink(inode, attr->la_nlink);
2588 inode->i_rdev = attr->la_rdev;
2590 if (bits & LA_FLAGS) {
2591 /* always keep S_NOCMTIME */
2592 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2595 * Ext4 did not transfer inherit flags from
2596 * @inode->i_flags to raw inode i_flags when writing
2597 * flags, we do it explictly here.
2599 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2600 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2605 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2609 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2610 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2613 ll_vfs_dq_init(inode);
2615 if (attr->la_valid & LA_UID)
2616 iattr.ia_valid |= ATTR_UID;
2617 if (attr->la_valid & LA_GID)
2618 iattr.ia_valid |= ATTR_GID;
2619 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2620 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2622 rc = ll_vfs_dq_transfer(inode, &iattr);
2624 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2625 "enforcement enabled on the ldiskfs "
2626 "filesystem?\n", inode->i_sb->s_id, rc);
2631 #ifdef HAVE_PROJECT_QUOTA
2632 /* Handle project id transfer here properly */
2633 if (attr->la_valid & LA_PROJID &&
2634 attr->la_projid != i_projid_read(inode)) {
2635 rc = __ldiskfs_ioctl_setproject(inode, attr->la_projid);
2637 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2638 "enforcement enabled on the ldiskfs "
2639 "filesystem?\n", inode->i_sb->s_id, rc);
2647 static int osd_attr_set(const struct lu_env *env,
2648 struct dt_object *dt,
2649 const struct lu_attr *attr,
2650 struct thandle *handle)
2652 struct osd_object *obj = osd_dt_obj(dt);
2653 struct inode *inode;
2656 if (!dt_object_exists(dt))
2659 LASSERT(handle != NULL);
2660 LASSERT(!dt_object_remote(dt));
2661 LASSERT(osd_invariant(obj));
2663 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2665 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2666 struct osd_thread_info *oti = osd_oti_get(env);
2667 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2668 struct lu_fid *fid1 = &oti->oti_fid;
2669 struct osd_inode_id *id = &oti->oti_id;
2670 struct iam_path_descr *ipd;
2671 struct iam_container *bag;
2672 struct osd_thandle *oh;
2675 fid_cpu_to_be(fid1, fid0);
2676 memset(id, 1, sizeof(*id));
2677 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2678 fid0)->oi_dir.od_container;
2679 ipd = osd_idx_ipd_get(env, bag);
2680 if (unlikely(ipd == NULL))
2683 oh = container_of0(handle, struct osd_thandle, ot_super);
2684 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2685 (const struct iam_rec *)id, ipd);
2686 osd_ipd_put(env, bag, ipd);
2687 return(rc > 0 ? 0 : rc);
2690 inode = obj->oo_inode;
2692 rc = osd_quota_transfer(inode, attr);
2696 spin_lock(&obj->oo_guard);
2697 rc = osd_inode_setattr(env, inode, attr);
2698 spin_unlock(&obj->oo_guard);
2702 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2704 if (!(attr->la_valid & LA_FLAGS))
2707 /* Let's check if there are extra flags need to be set into LMA */
2708 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2709 struct osd_thread_info *info = osd_oti_get(env);
2710 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2712 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2713 &info->oti_ost_attrs);
2717 lma->lma_incompat |=
2718 lustre_to_lma_flags(attr->la_flags);
2719 lustre_lma_swab(lma);
2720 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2721 lma, sizeof(*lma), XATTR_REPLACE);
2723 struct osd_device *osd = osd_obj2dev(obj);
2725 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2726 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2727 lma->lma_incompat, rc);
2730 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2732 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2735 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2740 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2741 struct osd_object *obj,
2742 const char *name, const int namelen)
2744 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2747 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2748 umode_t mode, struct dt_allocation_hint *hint,
2752 struct osd_device *osd = osd_obj2dev(obj);
2753 struct osd_thandle *oth;
2754 struct dt_object *parent = NULL;
2755 struct inode *inode;
2757 LINVRNT(osd_invariant(obj));
2758 LASSERT(obj->oo_inode == NULL);
2759 LASSERT(obj->oo_hl_head == NULL);
2761 if (S_ISDIR(mode) && ldiskfs_pdo) {
2762 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2763 if (obj->oo_hl_head == NULL)
2767 oth = container_of(th, struct osd_thandle, ot_super);
2768 LASSERT(oth->ot_handle->h_transaction != NULL);
2770 if (hint != NULL && hint->dah_parent != NULL &&
2771 !dt_object_remote(hint->dah_parent))
2772 parent = hint->dah_parent;
2774 inode = ldiskfs_create_inode(oth->ot_handle,
2775 parent ? osd_dt_obj(parent)->oo_inode :
2776 osd_sb(osd)->s_root->d_inode,
2778 if (!IS_ERR(inode)) {
2779 /* Do not update file c/mtime in ldiskfs. */
2780 inode->i_flags |= S_NOCMTIME;
2782 /* For new created object, it must be consistent,
2783 * and it is unnecessary to scrub against it. */
2784 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2786 obj->oo_inode = inode;
2789 if (obj->oo_hl_head != NULL) {
2790 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2791 obj->oo_hl_head = NULL;
2793 result = PTR_ERR(inode);
2795 LINVRNT(osd_invariant(obj));
2803 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2804 struct lu_attr *attr,
2805 struct dt_allocation_hint *hint,
2806 struct dt_object_format *dof,
2810 struct osd_thandle *oth;
2811 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2813 LASSERT(S_ISDIR(attr->la_mode));
2815 oth = container_of(th, struct osd_thandle, ot_super);
2816 LASSERT(oth->ot_handle->h_transaction != NULL);
2817 result = osd_mkfile(info, obj, mode, hint, th);
2822 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2823 struct lu_attr *attr,
2824 struct dt_allocation_hint *hint,
2825 struct dt_object_format *dof,
2829 struct osd_thandle *oth;
2830 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2832 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2834 LASSERT(S_ISREG(attr->la_mode));
2836 oth = container_of(th, struct osd_thandle, ot_super);
2837 LASSERT(oth->ot_handle->h_transaction != NULL);
2839 result = osd_mkfile(info, obj, mode, hint, th);
2841 LASSERT(obj->oo_inode != NULL);
2842 if (feat->dif_flags & DT_IND_VARKEY)
2843 result = iam_lvar_create(obj->oo_inode,
2844 feat->dif_keysize_max,
2846 feat->dif_recsize_max,
2849 result = iam_lfix_create(obj->oo_inode,
2850 feat->dif_keysize_max,
2852 feat->dif_recsize_max,
2859 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2860 struct lu_attr *attr,
2861 struct dt_allocation_hint *hint,
2862 struct dt_object_format *dof,
2865 LASSERT(S_ISREG(attr->la_mode));
2866 return osd_mkfile(info, obj, (attr->la_mode &
2867 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2870 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2871 struct lu_attr *attr,
2872 struct dt_allocation_hint *hint,
2873 struct dt_object_format *dof,
2876 LASSERT(S_ISLNK(attr->la_mode));
2877 return osd_mkfile(info, obj, (attr->la_mode &
2878 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2881 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2882 struct lu_attr *attr,
2883 struct dt_allocation_hint *hint,
2884 struct dt_object_format *dof,
2887 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2890 LINVRNT(osd_invariant(obj));
2891 LASSERT(obj->oo_inode == NULL);
2892 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2893 S_ISFIFO(mode) || S_ISSOCK(mode));
2895 result = osd_mkfile(info, obj, mode, hint, th);
2897 LASSERT(obj->oo_inode != NULL);
2899 * This inode should be marked dirty for i_rdev. Currently
2900 * that is done in the osd_attr_init().
2902 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2905 LINVRNT(osd_invariant(obj));
2909 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2911 struct dt_allocation_hint *hint,
2912 struct dt_object_format *dof,
2915 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2917 osd_obj_type_f result;
2933 result = osd_mk_index;
2944 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2945 struct dt_object *parent, struct dt_object *child,
2950 ah->dah_parent = parent;
2951 ah->dah_mode = child_mode;
2953 if (parent != NULL && !dt_object_remote(parent)) {
2954 /* will help to find FID->ino at dt_insert("..") */
2955 struct osd_object *pobj = osd_dt_obj(parent);
2956 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2960 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2961 struct lu_attr *attr, struct dt_object_format *dof)
2963 struct inode *inode = obj->oo_inode;
2964 __u64 valid = attr->la_valid;
2967 attr->la_valid &= ~(LA_TYPE | LA_MODE);
2969 if (dof->dof_type != DFT_NODE)
2970 attr->la_valid &= ~LA_RDEV;
2971 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2972 attr->la_valid &= ~LA_ATIME;
2973 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2974 attr->la_valid &= ~LA_CTIME;
2975 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2976 attr->la_valid &= ~LA_MTIME;
2978 result = osd_quota_transfer(inode, attr);
2982 if (attr->la_valid != 0) {
2983 result = osd_inode_setattr(info->oti_env, inode, attr);
2985 * The osd_inode_setattr() should always succeed here. The
2986 * only error that could be returned is EDQUOT when we are
2987 * trying to change the UID or GID of the inode. However, this
2988 * should not happen since quota enforcement is no longer
2989 * enabled on ldiskfs (lquota takes care of it).
2991 LASSERTF(result == 0, "%d\n", result);
2992 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2995 attr->la_valid = valid;
2999 * Helper function for osd_create()
3001 * \retval 0, on success
3003 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3004 struct lu_attr *attr, struct dt_allocation_hint *hint,
3005 struct dt_object_format *dof, struct thandle *th)
3010 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3012 /* we drop umask so that permissions we pass are not affected */
3013 umask = current->fs->umask;
3014 current->fs->umask = 0;
3016 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3018 if (likely(obj->oo_inode != NULL)) {
3019 LASSERT(obj->oo_inode->i_state & I_NEW);
3021 /* Unlock the inode before attr initialization to avoid
3022 * unnecessary dqget operations. LU-6378 */
3023 unlock_new_inode(obj->oo_inode);
3026 if (likely(result == 0)) {
3027 osd_attr_init(info, obj, attr, dof);
3028 osd_object_init0(obj);
3031 /* restore previous umask value */
3032 current->fs->umask = umask;
3034 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3040 * Helper function for osd_create()
3042 * \retval 0, on success
3044 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3045 const struct lu_fid *fid, struct thandle *th)
3047 struct osd_thread_info *info = osd_oti_get(env);
3048 struct osd_inode_id *id = &info->oti_id;
3049 struct osd_device *osd = osd_obj2dev(obj);
3050 struct osd_thandle *oh;
3053 LASSERT(obj->oo_inode != NULL);
3055 oh = container_of0(th, struct osd_thandle, ot_super);
3056 LASSERT(oh->ot_handle);
3057 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3059 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3060 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3061 OI_CHECK_FLD, NULL);
3062 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3067 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3068 u64 seq, struct lu_seq_range *range)
3070 struct seq_server_site *ss = osd_seq_site(osd);
3072 if (fid_seq_is_idif(seq)) {
3073 fld_range_set_ost(range);
3074 range->lsr_index = idif_ost_idx(seq);
3078 if (!fid_seq_in_fldb(seq)) {
3079 fld_range_set_mdt(range);
3081 /* FIXME: If ss is NULL, it suppose not get lsr_index
3083 range->lsr_index = ss->ss_node_id;
3087 LASSERT(ss != NULL);
3088 fld_range_set_any(range);
3089 /* OSD will only do local fld lookup */
3090 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3093 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3094 struct lu_attr *attr,
3095 struct dt_allocation_hint *hint,
3096 struct dt_object_format *dof,
3097 struct thandle *handle)
3099 struct osd_thandle *oh;
3103 LASSERT(handle != NULL);
3105 oh = container_of0(handle, struct osd_thandle, ot_super);
3106 LASSERT(oh->ot_handle == NULL);
3108 /* EA object consumes more credits than regular object: osd_mk_index
3109 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3110 * leaf_node, could involves the block, block bitmap, groups, GDT
3111 * change for each block, so add 4 * 2 credits in that case. */
3112 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3113 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3114 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3115 /* Reuse idle OI block may cause additional one OI block
3117 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3118 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3123 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3124 attr->la_projid, 1, oh, osd_dt_obj(dt),
3125 NULL, OSD_QID_INODE);
3129 /* will help to find FID->ino mapping at dt_insert() */
3130 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3137 * Called to destroy on-disk representation of the object
3139 * Concurrency: must be locked
3141 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3144 struct osd_object *obj = osd_dt_obj(dt);
3145 struct inode *inode = obj->oo_inode;
3146 struct osd_thandle *oh;
3153 oh = container_of0(th, struct osd_thandle, ot_super);
3154 LASSERT(oh->ot_handle == NULL);
3156 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3157 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3158 /* Recycle idle OI leaf may cause additional three OI blocks
3160 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3161 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3162 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3163 /* one less inode */
3164 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3165 i_projid_read(inode), -1, oh, obj, NULL,
3169 /* data to be truncated */
3170 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3171 i_projid_read(inode), 0, oh, obj, NULL,
3176 /* will help to find FID->ino when this object is being
3177 * added to PENDING/ */
3178 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3183 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3186 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3187 struct osd_object *obj = osd_dt_obj(dt);
3188 struct inode *inode = obj->oo_inode;
3189 struct osd_device *osd = osd_obj2dev(obj);
3190 struct osd_thandle *oh;
3194 oh = container_of0(th, struct osd_thandle, ot_super);
3195 LASSERT(oh->ot_handle);
3197 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3199 if (unlikely(fid_is_acct(fid)))
3202 if (S_ISDIR(inode->i_mode)) {
3203 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3204 inode->i_nlink == 2);
3205 /* it will check/delete the inode from remote parent,
3206 * how to optimize it? unlink performance impaction XXX */
3207 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3209 CERROR("%s: delete inode "DFID": rc = %d\n",
3210 osd_name(osd), PFID(fid), result);
3212 spin_lock(&obj->oo_guard);
3214 spin_unlock(&obj->oo_guard);
3215 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3218 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3220 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3222 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3223 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3224 oh->ot_handle, OI_CHECK_FLD);
3226 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3227 /* XXX: add to ext3 orphan list */
3228 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3230 /* not needed in the cache anymore */
3231 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3232 obj->oo_destroyed = 1;
3238 * Put the fid into lustre_mdt_attrs, and then place the structure
3239 * inode's ea. This fid should not be altered during the life time
3242 * \retval +ve, on success
3243 * \retval -ve, on error
3245 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3247 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3248 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3250 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3251 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3255 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3258 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3261 lustre_loa_init(loa, fid, compat, incompat);
3262 lustre_loa_swab(loa, false);
3264 /* For the OST device with 256 bytes inode size by default,
3265 * the PFID EA will be stored together with LMA EA to avoid
3266 * performance trouble. Otherwise the PFID EA can be stored
3267 * independently. LU-8998 */
3268 if ((compat & LMAC_FID_ON_OST) &&
3269 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3270 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3271 sizeof(*loa), XATTR_CREATE);
3273 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3274 sizeof(*lma), XATTR_CREATE);
3275 /* LMA may already exist, but we need to check that all the
3276 * desired compat/incompat flags have been added. */
3277 if (unlikely(rc == -EEXIST)) {
3278 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3279 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3283 if (rc < sizeof(*lma))
3286 lustre_loa_swab(loa, true);
3287 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3288 ((compat == 0 && incompat == 0) ||
3289 (!(~lma->lma_compat & compat) &&
3290 !(~lma->lma_incompat & incompat))))
3293 lma->lma_self_fid = *fid;
3294 lma->lma_compat |= compat;
3295 lma->lma_incompat |= incompat;
3296 if (rc == sizeof(*lma)) {
3297 lustre_lma_swab(lma);
3298 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3299 sizeof(*lma), XATTR_REPLACE);
3301 lustre_loa_swab(loa, false);
3302 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3303 sizeof(*loa), XATTR_REPLACE);
3311 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3312 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3313 * To have compatilibility with 1.8 ldiskfs driver we need to have
3314 * magic number at start of fid data.
3315 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3318 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3319 const struct lu_fid *fid)
3321 if (!fid_is_namespace_visible(fid) ||
3322 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3323 param->edp_magic = 0;
3327 param->edp_magic = LDISKFS_LUFID_MAGIC;
3328 param->edp_len = sizeof(struct lu_fid) + 1;
3329 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3333 * Try to read the fid from inode ea into dt_rec.
3335 * \param fid object fid.
3337 * \retval 0 on success
3339 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3340 __u32 ino, struct lu_fid *fid,
3341 struct osd_inode_id *id)
3343 struct osd_thread_info *info = osd_oti_get(env);
3344 struct inode *inode;
3347 osd_id_gen(id, ino, OSD_OII_NOGEN);
3348 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3350 RETURN(PTR_ERR(inode));
3356 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3358 struct inode *parent_dir,
3359 const struct lu_fid *dot_fid,
3360 const struct lu_fid *dot_dot_fid,
3361 struct osd_thandle *oth)
3363 struct ldiskfs_dentry_param *dot_ldp;
3364 struct ldiskfs_dentry_param *dot_dot_ldp;
3365 __u32 saved_nlink = dir->i_nlink;
3368 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3369 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3371 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3372 dot_ldp->edp_magic = 0;
3374 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3375 dir, dot_ldp, dot_dot_ldp);
3376 /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3377 * the subseqent ref_add() will increase the dir->i_nlink
3378 * as 3. That is incorrect for new created directory.
3380 * It looks like hack, because we want to make the OSD API
3381 * to be order-independent for new created directory object
3382 * between dt_insert(..) and ref_add() operations.
3384 * Here, we only restore the in-RAM dir-inode's nlink attr,
3385 * becuase if the nlink attr is not 2, then there will be
3386 * ref_add() called following the dt_insert(..), such call
3387 * will make both the in-RAM and on-disk dir-inode's nlink
3388 * attr to be set as 2. LU-7447 */
3389 set_nlink(dir, saved_nlink);
3394 * Create an local agent inode for remote entry
3396 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3397 struct osd_device *osd,
3398 struct osd_object *pobj,
3399 const struct lu_fid *fid,
3403 struct osd_thread_info *info = osd_oti_get(env);
3404 struct inode *local;
3405 struct osd_thandle *oh;
3410 oh = container_of(th, struct osd_thandle, ot_super);
3411 LASSERT(oh->ot_handle->h_transaction != NULL);
3413 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3414 if (IS_ERR(local)) {
3415 CERROR("%s: create local error %d\n", osd_name(osd),
3416 (int)PTR_ERR(local));
3420 /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3421 * correct gid on remote file, not agent here */
3422 local->i_gid = current_fsgid();
3423 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3424 unlock_new_inode(local);
3426 /* Agent inode should not have project ID */
3427 #ifdef HAVE_PROJECT_QUOTA
3428 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL) {
3429 rc = __ldiskfs_ioctl_setproject(local, 0);
3431 CERROR("%s: quota transfer failed: rc = %d. Is project "
3432 "quota enforcement enabled on the ldiskfs "
3433 "filesystem?\n", local->i_sb->s_id, rc);
3434 RETURN(ERR_PTR(rc));
3438 /* Set special LMA flag for local agent inode */
3439 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3441 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3442 osd_name(osd), PFID(fid), rc);
3443 RETURN(ERR_PTR(rc));
3449 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3450 lu_object_fid(&pobj->oo_dt.do_lu),
3453 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3454 osd_name(osd), PFID(fid), rc);
3455 RETURN(ERR_PTR(rc));
3462 * when direntry is deleted, we have to take care of possible agent inode
3463 * referenced by that. unfortunately we can't do this at that point:
3464 * iget() within a running transaction leads to deadlock and we better do
3465 * not call that every delete declaration to save performance. so we put
3466 * a potention agent inode on a list and process that once the transaction
3467 * is over. Notice it's not any worse in terms of real orphans as regular
3468 * object destroy doesn't put inodes on the on-disk orphan list. this should
3469 * be addressed separately
3471 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3472 struct osd_thandle *oh,
3475 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3476 struct osd_obj_orphan *oor;
3483 oor->oor_env = (struct lu_env *)env;
3484 spin_lock(&osd->od_osfs_lock);
3485 list_add(&oor->oor_list, &osd->od_orphan_list);
3486 spin_unlock(&osd->od_osfs_lock);
3488 oh->ot_remove_agents = 1;
3494 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3495 struct osd_device *osd)
3497 struct osd_thread_info *info = osd_oti_get(env);
3498 struct osd_obj_orphan *oor, *tmp;
3499 struct osd_inode_id id;
3500 struct list_head list;
3501 struct inode *inode;
3506 INIT_LIST_HEAD(&list);
3508 spin_lock(&osd->od_osfs_lock);
3509 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3510 if (oor->oor_env == env) {
3511 list_del(&oor->oor_list);
3512 list_add(&oor->oor_list, &list);
3515 spin_unlock(&osd->od_osfs_lock);
3517 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3521 list_del(&oor->oor_list);
3524 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3525 inode = osd_iget_fid(info, osd, &id, &fid);
3529 if (!osd_remote_fid(env, osd, &fid)) {
3534 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3536 mark_inode_dirty(inode);
3537 ldiskfs_journal_stop(jh);
3545 * OSD layer object create function for OST objects (b=11826).
3547 * The FID is inserted into inode xattr here.
3549 * \retval 0, on success
3550 * \retval -ve, on error
3552 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3553 struct lu_attr *attr, struct dt_allocation_hint *hint,
3554 struct dt_object_format *dof, struct thandle *th)
3556 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3557 struct osd_object *obj = osd_dt_obj(dt);
3558 struct osd_thread_info *info = osd_oti_get(env);
3559 int result, on_ost = 0;
3563 if (dt_object_exists(dt))
3566 LINVRNT(osd_invariant(obj));
3567 LASSERT(!dt_object_remote(dt));
3568 LASSERT(osd_is_write_locked(env, obj));
3569 LASSERT(th != NULL);
3571 if (unlikely(fid_is_acct(fid)))
3572 /* Quota files can't be created from the kernel any more,
3573 * 'tune2fs -O quota' will take care of creating them */
3576 result = __osd_create(info, obj, attr, hint, dof, th);
3578 if (fid_is_idif(fid) &&
3579 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3580 struct lu_fid *tfid = &info->oti_fid;
3581 struct ost_id *oi = &info->oti_ostid;
3583 fid_to_ostid(fid, oi);
3584 ostid_to_fid(tfid, oi, 0);
3586 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3587 LMAC_FID_ON_OST, 0);
3589 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3591 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3592 on_ost ? LMAC_FID_ON_OST : 0,
3595 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3596 obj->oo_dt.do_body_ops = &osd_body_ops;
3600 result = __osd_oi_insert(env, obj, fid, th);
3602 /* a small optimization - dt_insert() isn't usually applied
3603 * to OST objects, so we don't need to cache OI mapping for
3605 if (result == 0 && on_ost == 0) {
3606 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3607 result = osd_idc_find_and_init(env, osd, obj);
3608 LASSERT(result == 0);
3611 LASSERT(ergo(result == 0,
3612 dt_object_exists(dt) && !dt_object_remote(dt)));
3613 LINVRNT(osd_invariant(obj));
3617 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3618 struct thandle *handle)
3620 struct osd_thandle *oh;
3622 /* it's possible that object doesn't exist yet */
3623 LASSERT(handle != NULL);
3625 oh = container_of0(handle, struct osd_thandle, ot_super);
3626 LASSERT(oh->ot_handle == NULL);
3628 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3629 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3635 * Concurrency: @dt is write locked.
3637 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3640 struct osd_object *obj = osd_dt_obj(dt);
3641 struct inode *inode = obj->oo_inode;
3642 struct osd_thandle *oh;
3645 if (!dt_object_exists(dt) || obj->oo_destroyed)
3648 LINVRNT(osd_invariant(obj));
3649 LASSERT(!dt_object_remote(dt));
3650 LASSERT(osd_is_write_locked(env, obj));
3651 LASSERT(th != NULL);
3653 oh = container_of0(th, struct osd_thandle, ot_super);
3654 LASSERT(oh->ot_handle != NULL);
3656 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3658 CDEBUG(D_INODE, DFID" increase nlink %d\n",
3659 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3661 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3662 * (65000) subdirectories by storing "1" in i_nlink if the link count
3663 * would otherwise overflow. Directory tranversal tools understand
3664 * that (st_nlink == 1) indicates that the filesystem dose not track
3665 * hard links count on the directory, and will not abort subdirectory
3666 * scanning early once (st_nlink - 2) subdirs have been found.
3668 * This also has to properly handle the case of inodes with nlink == 0
3669 * in case they are being linked into the PENDING directory
3671 spin_lock(&obj->oo_guard);
3672 if (unlikely(inode->i_nlink == 0))
3673 /* inc_nlink from 0 may cause WARN_ON */
3674 set_nlink(inode, 1);
3676 ldiskfs_inc_count(oh->ot_handle, inode);
3677 if (!S_ISDIR(inode->i_mode))
3678 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3680 spin_unlock(&obj->oo_guard);
3682 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3683 LINVRNT(osd_invariant(obj));
3685 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3690 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
3691 struct thandle *handle)
3693 struct osd_thandle *oh;
3695 if (!dt_object_exists(dt))
3698 LASSERT(!dt_object_remote(dt));
3699 LASSERT(handle != NULL);
3701 oh = container_of0(handle, struct osd_thandle, ot_super);
3702 LASSERT(oh->ot_handle == NULL);
3704 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3705 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3711 * Concurrency: @dt is write locked.
3713 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
3716 struct osd_object *obj = osd_dt_obj(dt);
3717 struct inode *inode = obj->oo_inode;
3718 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3719 struct osd_thandle *oh;
3721 if (!dt_object_exists(dt))
3724 LINVRNT(osd_invariant(obj));
3725 LASSERT(!dt_object_remote(dt));
3726 LASSERT(osd_is_write_locked(env, obj));
3727 LASSERT(th != NULL);
3729 oh = container_of0(th, struct osd_thandle, ot_super);
3730 LASSERT(oh->ot_handle != NULL);
3732 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3734 spin_lock(&obj->oo_guard);
3735 /* That can be result of upgrade from old Lustre version and
3736 * applied only to local files. Just skip this ref_del call.
3737 * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3738 if (inode->i_nlink == 0) {
3739 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3740 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3741 ", maybe an upgraded file? (LU-3915)\n",
3742 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3743 spin_unlock(&obj->oo_guard);
3747 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3748 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3750 ldiskfs_dec_count(oh->ot_handle, inode);
3751 spin_unlock(&obj->oo_guard);
3753 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3754 LINVRNT(osd_invariant(obj));
3756 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3762 * Concurrency: @dt is read locked.
3764 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3765 struct lu_buf *buf, const char *name)
3767 struct osd_object *obj = osd_dt_obj(dt);
3768 struct inode *inode = obj->oo_inode;
3769 struct osd_thread_info *info = osd_oti_get(env);
3770 struct dentry *dentry = &info->oti_obj_dentry;
3771 bool cache_xattr = false;
3776 /* version get is not real XATTR but uses xattr API */
3777 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3778 dt_obj_version_t *ver = buf->lb_buf;
3780 /* for version we are just using xattr API but change inode
3782 if (buf->lb_len == 0)
3783 return sizeof(dt_obj_version_t);
3785 if (buf->lb_len < sizeof(dt_obj_version_t))
3788 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
3789 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3791 *ver = LDISKFS_I(inode)->i_fs_version;
3793 return sizeof(dt_obj_version_t);
3796 if (!dt_object_exists(dt))
3799 LASSERT(!dt_object_remote(dt));
3800 LASSERT(inode->i_op != NULL);
3801 LASSERT(inode->i_op->getxattr != NULL);
3803 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
3804 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
3808 rc = osd_oxc_get(obj, name, buf);
3813 rc = __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3814 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
3815 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3816 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3817 struct filter_fid *ff;
3818 struct ost_layout *ol;
3820 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
3822 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
3826 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
3832 if (buf->lb_len == 0 || !buf->lb_buf)
3835 if (buf->lb_len < rc)
3839 ol = &ff->ff_layout;
3840 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
3841 PFID_STRIPE_IDX_BITS);
3842 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
3843 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
3844 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
3845 if (lma->lma_compat & LMAC_COMP_INFO) {
3846 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
3847 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
3848 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
3850 ol->ol_comp_start = 0;
3851 ol->ol_comp_end = 0;
3858 if (rc == -ENOENT || rc == -ENODATA)
3859 osd_oxc_add(obj, name, NULL, 0);
3860 else if (rc > 0 && buf->lb_buf != NULL)
3861 osd_oxc_add(obj, name, buf->lb_buf, rc);
3867 static int osd_declare_xattr_set(const struct lu_env *env,
3868 struct dt_object *dt,
3869 const struct lu_buf *buf, const char *name,
3870 int fl, struct thandle *handle)
3872 struct osd_thandle *oh;
3874 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3876 LASSERT(handle != NULL);
3878 oh = container_of0(handle, struct osd_thandle, ot_super);
3879 LASSERT(oh->ot_handle == NULL);
3881 if (strcmp(name, XATTR_NAME_LMA) == 0) {
3882 /* For non-upgrading case, the LMA is set first and
3883 * usually fit inode. But for upgrade case, the LMA
3884 * may be in another separated EA block. */
3885 if (dt_object_exists(dt)) {
3886 if (fl == LU_XATTR_REPLACE)
3891 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3893 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
3894 /* We may need to delete the old PFID EA. */
3895 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
3896 if (fl == LU_XATTR_REPLACE)
3903 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
3908 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3909 /* learn xattr size from osd_xattr_get if
3910 attribute has not been read yet */
3911 buflen = __osd_xattr_get(
3912 osd_dt_obj(dt)->oo_inode,
3913 &osd_oti_get(env)->oti_obj_dentry,
3918 buflen = buf->lb_len;
3921 if (buflen > sb->s_blocksize) {
3922 credits += osd_calc_bkmap_credits(
3924 (buflen + sb->s_blocksize - 1) >>
3925 sb->s_blocksize_bits);
3929 * xattr set may involve inode quota change, reserve credits for
3930 * dquot_initialize()
3932 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3935 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3941 * Concurrency: @dt is write locked.
3943 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3944 const struct lu_buf *buf, const char *name, int fl,
3945 struct thandle *handle)
3947 struct osd_object *obj = osd_dt_obj(dt);
3948 struct inode *inode = obj->oo_inode;
3949 struct osd_thread_info *info = osd_oti_get(env);
3950 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3951 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3960 /* version set is not real XATTR */
3961 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3962 dt_obj_version_t *version = buf->lb_buf;
3964 /* for version we are just using xattr API but change inode
3966 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
3968 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
3969 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3971 LDISKFS_I(inode)->i_fs_version = *version;
3972 /* Version is set after all inode operations are finished,
3973 * so we should mark it dirty here */
3974 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3979 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
3980 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
3983 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3984 if (fl & LU_XATTR_REPLACE)
3985 fs_flags |= XATTR_REPLACE;
3987 if (fl & LU_XATTR_CREATE)
3988 fs_flags |= XATTR_CREATE;
3990 /* For the OST device with 256 bytes inode size by default,
3991 * the PFID EA will be stored together with LMA EA to avoid
3992 * performance trouble. Otherwise the PFID EA can be stored
3993 * independently. LU-8998 */
3994 if (strcmp(name, XATTR_NAME_FID) == 0 &&
3995 LDISKFS_INODE_SIZE(inode->i_sb) <= 256) {
3996 struct dentry *dentry = &info->oti_obj_dentry;
3997 struct filter_fid *ff;
3998 struct ost_layout *ol;
4001 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4004 ol = &ff->ff_layout;
4005 /* Old client does not send stripe information, store
4006 * the PFID EA on disk directly. */
4007 if (buf->lb_len == sizeof(struct lu_fid) ||
4008 ol->ol_stripe_size == 0) {
4009 len = sizeof(struct lu_fid);
4013 if (buf->lb_len != sizeof(*ff))
4016 rc = osd_get_lma(info, inode, dentry, loa);
4017 if (unlikely(rc == -ENODATA)) {
4018 /* Usually for upgarding from old device */
4019 lustre_loa_init(loa, lu_object_fid(&dt->do_lu),
4020 LMAC_FID_ON_OST, 0);
4028 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4029 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4030 PFID_STRIPE_IDX_BITS;
4031 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4032 lma->lma_compat |= LMAC_STRIPE_INFO;
4033 if (ol->ol_comp_id != 0) {
4034 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4035 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4036 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4037 lma->lma_compat |= LMAC_COMP_INFO;
4040 lustre_loa_swab(loa, false);
4042 /* Remove old PFID EA entry firstly. */
4043 ll_vfs_dq_init(inode);
4044 rc = inode->i_op->removexattr(dentry, name);
4045 if (rc && rc != -ENODATA)
4048 /* Store the PFID EA inside the LMA EA. */
4049 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4053 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4054 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4058 lma->lma_incompat |= LMAI_STRIPED;
4059 lustre_lma_swab(lma);
4060 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4061 sizeof(*lma), XATTR_REPLACE);
4067 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4068 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4071 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4072 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4073 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4079 * Concurrency: @dt is read locked.
4081 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4082 const struct lu_buf *buf)
4084 struct osd_object *obj = osd_dt_obj(dt);
4085 struct inode *inode = obj->oo_inode;
4086 struct osd_thread_info *info = osd_oti_get(env);
4087 struct dentry *dentry = &info->oti_obj_dentry;
4089 if (!dt_object_exists(dt))
4092 LASSERT(!dt_object_remote(dt));
4093 LASSERT(inode->i_op != NULL);
4094 LASSERT(inode->i_op->listxattr != NULL);
4096 dentry->d_inode = inode;
4097 dentry->d_sb = inode->i_sb;
4098 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4101 static int osd_declare_xattr_del(const struct lu_env *env,
4102 struct dt_object *dt, const char *name,
4103 struct thandle *handle)
4105 struct osd_thandle *oh;
4106 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4108 LASSERT(!dt_object_remote(dt));
4109 LASSERT(handle != NULL);
4111 oh = container_of0(handle, struct osd_thandle, ot_super);
4112 LASSERT(oh->ot_handle == NULL);
4114 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4115 osd_dto_credits_noquota[DTO_XATTR_SET]);
4117 * xattr del may involve inode quota change, reserve credits for
4118 * dquot_initialize()
4120 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4126 * Concurrency: @dt is write locked.
4128 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4129 const char *name, struct thandle *handle)
4131 struct osd_object *obj = osd_dt_obj(dt);
4132 struct inode *inode = obj->oo_inode;
4133 struct osd_thread_info *info = osd_oti_get(env);
4134 struct dentry *dentry = &info->oti_obj_dentry;
4137 if (!dt_object_exists(dt))
4140 LASSERT(!dt_object_remote(dt));
4141 LASSERT(inode->i_op != NULL);
4142 LASSERT(inode->i_op->removexattr != NULL);
4143 LASSERT(handle != NULL);
4145 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4147 ll_vfs_dq_init(inode);
4148 dentry->d_inode = inode;
4149 dentry->d_sb = inode->i_sb;
4150 rc = inode->i_op->removexattr(dentry, name);
4151 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
4152 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4154 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4156 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4157 &info->oti_ost_attrs);
4159 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
4164 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4165 lustre_lma_swab(lma);
4166 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4167 sizeof(*lma), XATTR_REPLACE);
4172 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4175 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4176 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4177 osd_oxc_del(obj, name);
4182 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4183 __u64 start, __u64 end)
4185 struct osd_object *obj = osd_dt_obj(dt);
4186 struct inode *inode = obj->oo_inode;
4187 struct osd_thread_info *info = osd_oti_get(env);
4188 struct dentry *dentry = &info->oti_obj_dentry;
4189 struct file *file = &info->oti_file;
4194 dentry->d_inode = inode;
4195 dentry->d_sb = inode->i_sb;
4196 file->f_path.dentry = dentry;
4197 file->f_mapping = inode->i_mapping;
4198 file->f_op = inode->i_fop;
4199 set_file_inode(file, inode);
4201 rc = ll_vfs_fsync_range(file, start, end, 0);
4206 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4215 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4216 const struct dt_index_features *feat)
4218 struct iam_descr *descr;
4220 if (osd_object_is_root(o))
4221 return feat == &dt_directory_features;
4223 LASSERT(o->oo_dir != NULL);
4225 descr = o->oo_dir->od_container.ic_descr;
4226 if (feat == &dt_directory_features) {
4227 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4233 feat->dif_keysize_min <= descr->id_key_size &&
4234 descr->id_key_size <= feat->dif_keysize_max &&
4235 feat->dif_recsize_min <= descr->id_rec_size &&
4236 descr->id_rec_size <= feat->dif_recsize_max &&
4237 !(feat->dif_flags & (DT_IND_VARKEY |
4238 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4239 ergo(feat->dif_flags & DT_IND_UPDATE,
4240 1 /* XXX check that object (and file system) is
4245 static int osd_iam_container_init(const struct lu_env *env,
4246 struct osd_object *obj,
4247 struct osd_directory *dir)
4249 struct iam_container *bag = &dir->od_container;
4252 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4256 result = iam_container_setup(bag);
4258 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4260 iam_container_fini(bag);
4267 * Concurrency: no external locking is necessary.
4269 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4270 const struct dt_index_features *feat)
4274 struct osd_object *obj = osd_dt_obj(dt);
4276 LINVRNT(osd_invariant(obj));
4278 if (osd_object_is_root(obj)) {
4279 dt->do_index_ops = &osd_index_ea_ops;
4281 } else if (feat == &dt_directory_features) {
4282 dt->do_index_ops = &osd_index_ea_ops;
4283 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4288 } else if (unlikely(feat == &dt_otable_features)) {
4289 dt->do_index_ops = &osd_otable_ops;
4291 } else if (unlikely(feat == &dt_acct_features)) {
4292 dt->do_index_ops = &osd_acct_index_ops;
4295 } else if (!osd_has_index(obj)) {
4296 struct osd_directory *dir;
4301 spin_lock(&obj->oo_guard);
4302 if (obj->oo_dir == NULL)
4306 * Concurrent thread allocated container data.
4309 spin_unlock(&obj->oo_guard);
4311 * Now, that we have container data, serialize its
4314 down_write(&obj->oo_ext_idx_sem);
4316 * recheck under lock.
4318 if (!osd_has_index(obj))
4319 result = osd_iam_container_init(env, obj,
4323 up_write(&obj->oo_ext_idx_sem);
4331 if (result == 0 && skip_iam == 0) {
4332 if (!osd_iam_index_probe(env, obj, feat))
4335 LINVRNT(osd_invariant(obj));
4340 static int osd_otable_it_attr_get(const struct lu_env *env,
4341 struct dt_object *dt,
4342 struct lu_attr *attr)
4348 static const struct dt_object_operations osd_obj_ops = {
4349 .do_read_lock = osd_read_lock,
4350 .do_write_lock = osd_write_lock,
4351 .do_read_unlock = osd_read_unlock,
4352 .do_write_unlock = osd_write_unlock,
4353 .do_write_locked = osd_write_locked,
4354 .do_attr_get = osd_attr_get,
4355 .do_declare_attr_set = osd_declare_attr_set,
4356 .do_attr_set = osd_attr_set,
4357 .do_ah_init = osd_ah_init,
4358 .do_declare_create = osd_declare_create,
4359 .do_create = osd_create,
4360 .do_declare_destroy = osd_declare_destroy,
4361 .do_destroy = osd_destroy,
4362 .do_index_try = osd_index_try,
4363 .do_declare_ref_add = osd_declare_ref_add,
4364 .do_ref_add = osd_ref_add,
4365 .do_declare_ref_del = osd_declare_ref_del,
4366 .do_ref_del = osd_ref_del,
4367 .do_xattr_get = osd_xattr_get,
4368 .do_declare_xattr_set = osd_declare_xattr_set,
4369 .do_xattr_set = osd_xattr_set,
4370 .do_declare_xattr_del = osd_declare_xattr_del,
4371 .do_xattr_del = osd_xattr_del,
4372 .do_xattr_list = osd_xattr_list,
4373 .do_object_sync = osd_object_sync,
4374 .do_invalidate = osd_invalidate,
4377 static const struct dt_object_operations osd_obj_otable_it_ops = {
4378 .do_attr_get = osd_otable_it_attr_get,
4379 .do_index_try = osd_index_try,
4382 static int osd_index_declare_iam_delete(const struct lu_env *env,
4383 struct dt_object *dt,
4384 const struct dt_key *key,
4385 struct thandle *handle)
4387 struct osd_thandle *oh;
4389 oh = container_of0(handle, struct osd_thandle, ot_super);
4390 LASSERT(oh->ot_handle == NULL);
4392 /* Recycle may cause additional three blocks to be changed. */
4393 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4394 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4400 * delete a (key, value) pair from index \a dt specified by \a key
4402 * \param dt osd index object
4403 * \param key key for index
4404 * \param rec record reference
4405 * \param handle transaction handler
4408 * \retval -ve failure
4410 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4411 const struct dt_key *key,
4412 struct thandle *handle)
4414 struct osd_thread_info *oti = osd_oti_get(env);
4415 struct osd_object *obj = osd_dt_obj(dt);
4416 struct osd_thandle *oh;
4417 struct iam_path_descr *ipd;
4418 struct iam_container *bag = &obj->oo_dir->od_container;
4422 if (!dt_object_exists(dt))
4425 LINVRNT(osd_invariant(obj));
4426 LASSERT(!dt_object_remote(dt));
4427 LASSERT(bag->ic_object == obj->oo_inode);
4428 LASSERT(handle != NULL);
4430 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4432 ipd = osd_idx_ipd_get(env, bag);
4433 if (unlikely(ipd == NULL))
4436 oh = container_of0(handle, struct osd_thandle, ot_super);
4437 LASSERT(oh->ot_handle != NULL);
4438 LASSERT(oh->ot_handle->h_transaction != NULL);
4440 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4441 /* swab quota uid/gid provided by caller */
4442 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4443 key = (const struct dt_key *)&oti->oti_quota_id;
4446 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4447 osd_ipd_put(env, bag, ipd);
4448 LINVRNT(osd_invariant(obj));
4449 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4453 static int osd_index_declare_ea_delete(const struct lu_env *env,
4454 struct dt_object *dt,
4455 const struct dt_key *key,
4456 struct thandle *handle)
4458 struct osd_thandle *oh;
4459 struct inode *inode;
4463 LASSERT(!dt_object_remote(dt));
4464 LASSERT(handle != NULL);
4466 oh = container_of0(handle, struct osd_thandle, ot_super);
4467 LASSERT(oh->ot_handle == NULL);
4469 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4470 if (key != NULL && unlikely(strcmp((char *)key, dotdot) == 0)) {
4471 /* '..' to a remote object has a local representative */
4472 credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
4473 /* to reset LMAI_REMOTE_PARENT */
4476 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4478 inode = osd_dt_obj(dt)->oo_inode;
4482 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4483 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
4488 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
4491 struct osd_fid_pack *rec;
4494 if (de->file_type & LDISKFS_DIRENT_LUFID) {
4495 rec = (struct osd_fid_pack *) (de->name + de->name_len + 1);
4496 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
4497 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
4503 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
4504 const struct lu_fid *fid)
4506 struct seq_server_site *ss = osd_seq_site(osd);
4509 /* FID seqs not in FLDB, must be local seq */
4510 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
4513 /* If FLD is not being initialized yet, it only happens during the
4514 * initialization, likely during mgs initialization, and we assume
4515 * this is local FID. */
4516 if (ss == NULL || ss->ss_server_fld == NULL)
4519 /* Only check the local FLDB here */
4520 if (osd_seq_exists(env, osd, fid_seq(fid)))
4527 * Index delete function for interoperability mode (b11826).
4528 * It will remove the directory entry added by osd_index_ea_insert().
4529 * This entry is needed to maintain name->fid mapping.
4531 * \param key, key i.e. file entry to be deleted
4533 * \retval 0, on success
4534 * \retval -ve, on error
4536 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
4537 const struct dt_key *key, struct thandle *handle)
4539 struct osd_object *obj = osd_dt_obj(dt);
4540 struct inode *dir = obj->oo_inode;
4541 struct dentry *dentry;
4542 struct osd_thandle *oh;
4543 struct ldiskfs_dir_entry_2 *de = NULL;
4544 struct buffer_head *bh;
4545 struct htree_lock *hlock = NULL;
4546 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4547 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4551 if (!dt_object_exists(dt))
4554 LINVRNT(osd_invariant(obj));
4555 LASSERT(!dt_object_remote(dt));
4556 LASSERT(handle != NULL);
4558 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4560 oh = container_of(handle, struct osd_thandle, ot_super);
4561 LASSERT(oh->ot_handle != NULL);
4562 LASSERT(oh->ot_handle->h_transaction != NULL);
4564 ll_vfs_dq_init(dir);
4565 dentry = osd_child_dentry_get(env, obj,
4566 (char *)key, strlen((char *)key));
4568 if (obj->oo_hl_head != NULL) {
4569 hlock = osd_oti_get(env)->oti_hlock;
4570 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
4571 dir, LDISKFS_HLOCK_DEL);
4573 down_write(&obj->oo_ext_idx_sem);
4576 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
4578 /* If this is not the ".." entry, it might be a remote DNE
4579 * entry and we need to check if the FID is for a remote
4580 * MDT. If the FID is not in the directory entry (e.g.
4581 * upgraded 1.8 filesystem without dirdata enabled) then
4582 * we need to get the FID from the LMA. For a remote directory
4583 * there HAS to be an LMA, it cannot be an IGIF inode in this
4586 * Delete the entry before the agent inode in order to
4587 * simplify error handling. At worst an error after deleting
4588 * the entry first might leak the agent inode afterward. The
4589 * reverse would need filesystem abort in case of error deleting
4590 * the entry after the agent had been removed, or leave a
4591 * dangling entry pointing at a random inode. */
4592 if (strcmp((char *)key, dotdot) != 0) {
4593 LASSERT(de != NULL);
4594 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
4595 if (rc == -ENODATA) {
4596 /* can't get FID, postpone to the end of the
4597 * transaction when iget() is safe */
4598 osd_schedule_agent_inode_removal(env, oh,
4599 le32_to_cpu(de->inode));
4600 } else if (rc == 0 &&
4601 unlikely(osd_remote_fid(env, osd, fid))) {
4602 osd_schedule_agent_inode_removal(env, oh,
4603 le32_to_cpu(de->inode));
4606 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
4612 ldiskfs_htree_unlock(hlock);
4614 up_write(&obj->oo_ext_idx_sem);
4619 /* For inode on the remote MDT, .. will point to
4620 * /Agent directory, Check whether it needs to delete
4621 * from agent directory */
4622 if (unlikely(strcmp((char *)key, dotdot) == 0)) {
4625 ret = osd_delete_from_remote_parent(env, osd_obj2dev(obj),
4628 /* Sigh, the entry has been deleted, and
4629 * it is not easy to revert it back, so
4630 * let's keep this error private, and let
4631 * LFSCK fix it. XXX */
4632 CERROR("%s: delete remote parent "DFID": rc = %d\n",
4633 osd_name(osd), PFID(fid), ret);
4636 LASSERT(osd_invariant(obj));
4637 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4642 * Lookup index for \a key and copy record to \a rec.
4644 * \param dt osd index object
4645 * \param key key for index
4646 * \param rec record reference
4648 * \retval +ve success : exact mach
4649 * \retval 0 return record with key not greater than \a key
4650 * \retval -ve failure
4652 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
4653 struct dt_rec *rec, const struct dt_key *key)
4655 struct osd_object *obj = osd_dt_obj(dt);
4656 struct iam_path_descr *ipd;
4657 struct iam_container *bag = &obj->oo_dir->od_container;
4658 struct osd_thread_info *oti = osd_oti_get(env);
4659 struct iam_iterator *it = &oti->oti_idx_it;
4660 struct iam_rec *iam_rec;
4664 if (!dt_object_exists(dt))
4667 LASSERT(osd_invariant(obj));
4668 LASSERT(!dt_object_remote(dt));
4669 LASSERT(bag->ic_object == obj->oo_inode);
4671 ipd = osd_idx_ipd_get(env, bag);
4675 /* got ipd now we can start iterator. */
4676 iam_it_init(it, bag, 0, ipd);
4678 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4679 /* swab quota uid/gid provided by caller */
4680 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4681 key = (const struct dt_key *)&oti->oti_quota_id;
4684 rc = iam_it_get(it, (struct iam_key *)key);
4686 if (S_ISDIR(obj->oo_inode->i_mode))
4687 iam_rec = (struct iam_rec *)oti->oti_ldp;
4689 iam_rec = (struct iam_rec *) rec;
4691 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
4693 if (S_ISDIR(obj->oo_inode->i_mode))
4694 osd_fid_unpack((struct lu_fid *) rec,
4695 (struct osd_fid_pack *)iam_rec);
4696 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
4697 osd_quota_unpack(obj, rec);
4702 osd_ipd_put(env, bag, ipd);
4704 LINVRNT(osd_invariant(obj));
4709 static int osd_index_declare_iam_insert(const struct lu_env *env,
4710 struct dt_object *dt,
4711 const struct dt_rec *rec,
4712 const struct dt_key *key,
4713 struct thandle *handle)
4715 struct osd_thandle *oh;
4717 LASSERT(handle != NULL);
4719 oh = container_of0(handle, struct osd_thandle, ot_super);
4720 LASSERT(oh->ot_handle == NULL);
4722 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
4723 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
4729 * Inserts (key, value) pair in \a dt index object.
4731 * \param dt osd index object
4732 * \param key key for index
4733 * \param rec record reference
4734 * \param th transaction handler
4737 * \retval -ve failure
4739 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
4740 const struct dt_rec *rec,
4741 const struct dt_key *key, struct thandle *th,
4744 struct osd_object *obj = osd_dt_obj(dt);
4745 struct iam_path_descr *ipd;
4746 struct osd_thandle *oh;
4747 struct iam_container *bag;
4748 struct osd_thread_info *oti = osd_oti_get(env);
4749 struct iam_rec *iam_rec;
4753 if (!dt_object_exists(dt))
4756 LINVRNT(osd_invariant(obj));
4757 LASSERT(!dt_object_remote(dt));
4759 bag = &obj->oo_dir->od_container;
4760 LASSERT(bag->ic_object == obj->oo_inode);
4761 LASSERT(th != NULL);
4763 osd_trans_exec_op(env, th, OSD_OT_INSERT);
4765 ipd = osd_idx_ipd_get(env, bag);
4766 if (unlikely(ipd == NULL))
4769 oh = container_of0(th, struct osd_thandle, ot_super);
4770 LASSERT(oh->ot_handle != NULL);
4771 LASSERT(oh->ot_handle->h_transaction != NULL);
4772 if (S_ISDIR(obj->oo_inode->i_mode)) {
4773 iam_rec = (struct iam_rec *)oti->oti_ldp;
4774 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec, &oti->oti_fid);
4775 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4776 /* pack quota uid/gid */
4777 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4778 key = (const struct dt_key *)&oti->oti_quota_id;
4779 /* pack quota record */
4780 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
4781 iam_rec = (struct iam_rec *)rec;
4783 iam_rec = (struct iam_rec *)rec;
4786 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
4788 osd_ipd_put(env, bag, ipd);
4789 LINVRNT(osd_invariant(obj));
4790 osd_trans_exec_check(env, th, OSD_OT_INSERT);
4795 * Calls ldiskfs_add_entry() to add directory entry
4796 * into the directory. This is required for
4797 * interoperability mode (b11826)
4799 * \retval 0, on success
4800 * \retval -ve, on error
4802 static int __osd_ea_add_rec(struct osd_thread_info *info,
4803 struct osd_object *pobj, struct inode *cinode,
4804 const char *name, const struct lu_fid *fid,
4805 struct htree_lock *hlock, struct thandle *th)
4807 struct ldiskfs_dentry_param *ldp;
4808 struct dentry *child;
4809 struct osd_thandle *oth;
4812 oth = container_of(th, struct osd_thandle, ot_super);
4813 LASSERT(oth->ot_handle != NULL);
4814 LASSERT(oth->ot_handle->h_transaction != NULL);
4815 LASSERT(pobj->oo_inode);
4817 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4818 if (unlikely(pobj->oo_inode ==
4819 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
4822 osd_get_ldiskfs_dirent_param(ldp, fid);
4823 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
4824 child->d_fsdata = (void *)ldp;
4825 ll_vfs_dq_init(pobj->oo_inode);
4826 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
4827 child, cinode, hlock);
4828 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
4829 struct ldiskfs_dir_entry_2 *de;
4830 struct buffer_head *bh;
4833 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
4836 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
4839 if (S_ISDIR(cinode->i_mode))
4840 de->file_type = LDISKFS_DIRENT_LUFID |
4841 LDISKFS_FT_REG_FILE;
4843 de->file_type = LDISKFS_DIRENT_LUFID |
4845 ldiskfs_handle_dirty_metadata(oth->ot_handle,
4856 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
4857 * into the directory.Also sets flags into osd object to
4858 * indicate dot and dotdot are created. This is required for
4859 * interoperability mode (b11826)
4861 * \param dir directory for dot and dotdot fixup.
4862 * \param obj child object for linking
4864 * \retval 0, on success
4865 * \retval -ve, on error
4867 static int osd_add_dot_dotdot(struct osd_thread_info *info,
4868 struct osd_object *dir,
4869 struct inode *parent_dir, const char *name,
4870 const struct lu_fid *dot_fid,
4871 const struct lu_fid *dot_dot_fid,
4874 struct inode *inode = dir->oo_inode;
4875 struct osd_thandle *oth;
4878 oth = container_of(th, struct osd_thandle, ot_super);
4879 LASSERT(oth->ot_handle->h_transaction != NULL);
4880 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
4882 if (strcmp(name, dot) == 0) {
4883 if (dir->oo_compat_dot_created) {
4886 LASSERT(inode->i_ino == parent_dir->i_ino);
4887 dir->oo_compat_dot_created = 1;
4890 } else if (strcmp(name, dotdot) == 0) {
4891 if (!dir->oo_compat_dot_created)
4893 /* in case of rename, dotdot is already created */
4894 if (dir->oo_compat_dotdot_created) {
4895 return __osd_ea_add_rec(info, dir, parent_dir, name,
4896 dot_dot_fid, NULL, th);
4899 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
4900 struct lu_fid tfid = *dot_dot_fid;
4903 result = osd_add_dot_dotdot_internal(info,
4904 dir->oo_inode, parent_dir, dot_fid,
4907 result = osd_add_dot_dotdot_internal(info,
4908 dir->oo_inode, parent_dir, dot_fid,
4913 dir->oo_compat_dotdot_created = 1;
4921 * It will call the appropriate osd_add* function and return the
4922 * value, return by respective functions.
4924 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
4925 struct inode *cinode, const char *name,
4926 const struct lu_fid *fid, struct thandle *th)
4928 struct osd_thread_info *info = osd_oti_get(env);
4929 struct htree_lock *hlock;
4932 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
4934 if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' &&
4936 if (hlock != NULL) {
4937 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4940 down_write(&pobj->oo_ext_idx_sem);
4943 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
4944 lu_object_fid(&pobj->oo_dt.do_lu),
4947 if (hlock != NULL) {
4948 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4949 pobj->oo_inode, LDISKFS_HLOCK_ADD);
4951 down_write(&pobj->oo_ext_idx_sem);
4954 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
4955 struct lu_fid *tfid = &info->oti_fid;
4959 rc = __osd_ea_add_rec(info, pobj, cinode, name,
4962 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
4967 ldiskfs_htree_unlock(hlock);
4969 up_write(&pobj->oo_ext_idx_sem);
4975 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
4976 struct osd_idmap_cache *oic)
4978 struct osd_scrub *scrub = &dev->od_scrub;
4979 struct lu_fid *fid = &oic->oic_fid;
4980 struct osd_inode_id *id = &oic->oic_lid;
4981 struct inode *inode = NULL;
4987 if (!fid_is_norm(fid) && !fid_is_igif(fid))
4990 if (scrub->os_pos_current > id->oii_ino)
4994 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
4995 if (rc == -ENOENT) {
4996 __u32 gen = id->oii_gen;
5002 inode = osd_iget(oti, dev, id);
5003 /* The inode has been removed (by race maybe). */
5004 if (IS_ERR(inode)) {
5005 rc = PTR_ERR(inode);
5007 RETURN(rc == -ESTALE ? -ENOENT : rc);
5010 /* The OI mapping is lost. */
5011 if (gen != OSD_OII_NOGEN)
5014 /* The inode may has been reused by others, we do not know,
5015 * leave it to be handled by subsequent osd_fid_lookup(). */
5017 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5024 if (thread_is_running(&scrub->os_thread)) {
5025 if (inode == NULL) {
5026 inode = osd_iget(oti, dev, id);
5027 /* The inode has been removed (by race maybe). */
5028 if (IS_ERR(inode)) {
5029 rc = PTR_ERR(inode);
5031 RETURN(rc == -ESTALE ? -ENOENT : rc);
5035 rc = osd_oii_insert(dev, oic, insert);
5036 /* There is race condition between osd_oi_lookup and OI scrub.
5037 * The OI scrub finished just after osd_oi_lookup() failure.
5038 * Under such case, it is unnecessary to trigger OI scrub again,
5039 * but try to call osd_oi_lookup() again. */
5040 if (unlikely(rc == -EAGAIN))
5043 if (!S_ISDIR(inode->i_mode))
5046 rc = osd_check_lmv(oti, dev, inode, oic);
5051 if (!dev->od_noscrub && ++once == 1) {
5052 rc = osd_scrub_start(dev, SS_AUTO_PARTIAL | SS_CLEAR_DRYRUN |
5054 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5055 "%s: trigger partial OI scrub for RPC inconsistency "
5056 "checking FID "DFID": rc = %d\n",
5057 osd_dev2name(dev), PFID(fid), rc);
5058 if (rc == 0 || rc == -EALREADY)
5071 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5072 struct osd_device *dev,
5073 struct osd_idmap_cache *oic,
5074 struct lu_fid *fid, __u32 ino)
5076 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5077 struct inode *inode;
5080 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5081 inode = osd_iget(oti, dev, &oic->oic_lid);
5082 if (IS_ERR(inode)) {
5083 fid_zero(&oic->oic_fid);
5084 return PTR_ERR(inode);
5087 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5090 fid_zero(&oic->oic_fid);
5092 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5096 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5097 struct osd_inode_id *id, const struct lu_fid *fid)
5099 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5100 id->oii_ino, id->oii_gen, info);
5101 info->oti_cache.oic_lid = *id;
5102 info->oti_cache.oic_fid = *fid;
5103 info->oti_cache.oic_dev = osd;
5107 * Get parent FID from the linkEA.
5109 * For a directory which parent resides on remote MDT, to satisfy the
5110 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5111 * the other hand, to make the lookup(..) on the directory can return
5112 * the real parent FID, we append the real parent FID after its ".."
5113 * name entry in the /REMOTE_PARENT_DIR.
5115 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5116 * backup. So after the restore, we cannot get the right parent FID from
5117 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5118 * we have stored the real parent FID in the directory object's linkEA,
5119 * we can parse the linkEA for the real parent FID.
5121 * \param[in] env pointer to the thread context
5122 * \param[in] obj pointer to the object to be handled
5123 * \param[out]fid pointer to the buffer to hold the parent FID
5125 * \retval 0 for getting the real parent FID successfully
5126 * \retval negative error number on failure
5128 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5129 struct osd_object *obj,
5132 struct osd_thread_info *oti = osd_oti_get(env);
5133 struct lu_buf *buf = &oti->oti_big_buf;
5134 struct dentry *dentry = &oti->oti_obj_dentry;
5135 struct inode *inode = obj->oo_inode;
5136 struct linkea_data ldata = { NULL };
5141 if (!S_ISDIR(inode->i_mode))
5145 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5146 buf->lb_buf, buf->lb_len);
5147 if (rc == -ERANGE) {
5148 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5151 lu_buf_realloc(buf, rc);
5152 if (buf->lb_buf == NULL)
5159 if (unlikely(rc == 0))
5165 if (unlikely(buf->lb_buf == NULL)) {
5166 lu_buf_realloc(buf, rc);
5167 if (buf->lb_buf == NULL)
5174 rc = linkea_init_with_rec(&ldata);
5176 linkea_first_entry(&ldata);
5177 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5183 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5184 struct inode *inode,
5185 const struct lu_fid *pfid,
5186 const char *name, const int namelen)
5188 struct osd_thread_info *oti = osd_oti_get(env);
5189 struct lu_buf *buf = &oti->oti_big_buf;
5190 struct dentry *dentry = &oti->oti_obj_dentry;
5191 struct linkea_data ldata = { NULL };
5192 struct lu_name cname = { .ln_name = name,
5193 .ln_namelen = namelen };
5198 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5199 buf->lb_buf, buf->lb_len);
5201 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5206 if (unlikely(rc == 0))
5209 if (buf->lb_len < rc) {
5210 lu_buf_realloc(buf, rc);
5211 if (buf->lb_buf == NULL)
5218 rc = linkea_init_with_rec(&ldata);
5220 rc = linkea_links_find(&ldata, &cname, pfid);
5226 * Calls ->lookup() to find dentry. From dentry get inode and
5227 * read inode's ea to get fid. This is required for interoperability
5230 * \retval 0, on success
5231 * \retval -ve, on error
5233 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5234 struct dt_rec *rec, const struct dt_key *key)
5236 struct inode *dir = obj->oo_inode;
5237 struct dentry *dentry;
5238 struct ldiskfs_dir_entry_2 *de;
5239 struct buffer_head *bh;
5240 struct lu_fid *fid = (struct lu_fid *) rec;
5241 struct htree_lock *hlock = NULL;
5246 LASSERT(dir->i_op != NULL);
5247 LASSERT(dir->i_op->lookup != NULL);
5249 dentry = osd_child_dentry_get(env, obj,
5250 (char *)key, strlen((char *)key));
5252 if (obj->oo_hl_head != NULL) {
5253 hlock = osd_oti_get(env)->oti_hlock;
5254 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5255 dir, LDISKFS_HLOCK_LOOKUP);
5257 down_read(&obj->oo_ext_idx_sem);
5260 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5262 struct osd_thread_info *oti = osd_oti_get(env);
5263 struct osd_inode_id *id = &oti->oti_id;
5264 struct osd_idmap_cache *oic = &oti->oti_cache;
5265 struct osd_device *dev = osd_obj2dev(obj);
5267 ino = le32_to_cpu(de->inode);
5268 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5270 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5274 rc = osd_get_fid_from_dentry(de, rec);
5276 /* done with de, release bh */
5279 if (unlikely(ino == osd_remote_parent_ino(dev))) {
5280 const char *name = (const char *)key;
5282 /* If the parent is on remote MDT, and there
5283 * is no FID-in-dirent, then we have to get
5284 * the parent FID from the linkEA. */
5285 if (likely(strlen(name) == 2 &&
5286 name[0] == '.' && name[1] == '.'))
5287 rc = osd_get_pfid_from_linkea(env, obj,
5290 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5293 osd_id_gen(id, ino, OSD_OII_NOGEN);
5296 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5297 fid_zero(&oic->oic_fid);
5302 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5303 rc = osd_consistency_check(oti, dev, oic);
5305 fid_zero(&oic->oic_fid);
5314 ldiskfs_htree_unlock(hlock);
5316 up_read(&obj->oo_ext_idx_sem);
5320 static int osd_index_declare_ea_insert(const struct lu_env *env,
5321 struct dt_object *dt,
5322 const struct dt_rec *rec,
5323 const struct dt_key *key,
5324 struct thandle *handle)
5326 struct osd_thandle *oh;
5327 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5328 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5329 const struct lu_fid *fid = rec1->rec_fid;
5330 int credits, rc = 0;
5331 struct osd_idmap_cache *idc;
5334 LASSERT(!dt_object_remote(dt));
5335 LASSERT(handle != NULL);
5336 LASSERT(fid != NULL);
5337 LASSERT(rec1->rec_type != 0);
5339 oh = container_of0(handle, struct osd_thandle, ot_super);
5340 LASSERT(oh->ot_handle == NULL);
5342 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5344 /* we can't call iget() while a transactions is running
5345 * (this can lead to a deadlock), but we need to know
5346 * inum and object type. so we find this information at
5347 * declaration and cache in per-thread info */
5348 idc = osd_idc_find_or_init(env, osd, fid);
5350 RETURN(PTR_ERR(idc));
5351 if (idc->oic_remote) {
5352 /* a reference to remote inode is represented by an
5353 * agent inode which we have to create */
5354 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5355 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5358 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5360 if (osd_dt_obj(dt)->oo_inode != NULL) {
5361 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5363 /* We ignore block quota on meta pool (MDTs), so needn't
5364 * calculate how many blocks will be consumed by this index
5366 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5368 i_projid_read(inode), 0,
5369 oh, osd_dt_obj(dt), NULL,
5377 * Index add function for interoperability mode (b11826).
5378 * It will add the directory entry.This entry is needed to
5379 * maintain name->fid mapping.
5381 * \param key it is key i.e. file entry to be inserted
5382 * \param rec it is value of given key i.e. fid
5384 * \retval 0, on success
5385 * \retval -ve, on error
5387 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5388 const struct dt_rec *rec,
5389 const struct dt_key *key, struct thandle *th,
5392 struct osd_object *obj = osd_dt_obj(dt);
5393 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5394 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5395 const struct lu_fid *fid = rec1->rec_fid;
5396 const char *name = (const char *)key;
5397 struct osd_thread_info *oti = osd_oti_get(env);
5398 struct inode *child_inode = NULL;
5399 struct osd_idmap_cache *idc;
5403 if (!dt_object_exists(dt))
5406 LASSERT(osd_invariant(obj));
5407 LASSERT(!dt_object_remote(dt));
5408 LASSERT(th != NULL);
5410 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5412 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5414 idc = osd_idc_find(env, osd, fid);
5415 if (unlikely(idc == NULL)) {
5416 /* this dt_insert() wasn't declared properly, so
5417 * FID is missing in OI cache. we better do not
5418 * lookup FID in FLDB/OI and don't risk to deadlock,
5419 * but in some special cases (lfsck testing, etc)
5420 * it's much simpler than fixing a caller */
5421 CERROR("%s: "DFID" wasn't declared for insert\n",
5422 osd_name(osd), PFID(fid));
5424 idc = osd_idc_find_or_init(env, osd, fid);
5426 RETURN(PTR_ERR(idc));
5429 if (idc->oic_remote) {
5430 /* Insert remote entry */
5431 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
5432 struct osd_mdobj_map *omm = osd->od_mdt_map;
5433 struct osd_thandle *oh;
5435 /* If parent on remote MDT, we need put this object
5437 oh = container_of(th, typeof(*oh), ot_super);
5438 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5440 CERROR("%s: add "DFID" error: rc = %d\n",
5442 PFID(lu_object_fid(&dt->do_lu)), rc);
5446 child_inode = igrab(omm->omm_remote_parent->d_inode);
5448 child_inode = osd_create_local_agent_inode(env, osd,
5449 obj, fid, rec1->rec_type & S_IFMT, th);
5450 if (IS_ERR(child_inode))
5451 RETURN(PTR_ERR(child_inode));
5454 /* Insert local entry */
5455 if (unlikely(idc->oic_lid.oii_ino == 0)) {
5456 /* for a reason OI cache wasn't filled properly */
5457 CERROR("%s: OIC for "DFID" isn't filled\n",
5458 osd_name(osd), PFID(fid));
5461 child_inode = oti->oti_inode;
5462 if (unlikely(child_inode == NULL)) {
5463 struct ldiskfs_inode_info *lii;
5467 child_inode = oti->oti_inode = &lii->vfs_inode;
5469 child_inode->i_sb = osd_sb(osd);
5470 child_inode->i_ino = idc->oic_lid.oii_ino;
5471 child_inode->i_mode = rec1->rec_type & S_IFMT;
5474 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
5476 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
5477 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
5479 if (child_inode && child_inode != oti->oti_inode)
5481 LASSERT(osd_invariant(obj));
5482 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5487 * Initialize osd Iterator for given osd index object.
5489 * \param dt osd index object
5492 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
5493 struct dt_object *dt,
5496 struct osd_it_iam *it;
5497 struct osd_object *obj = osd_dt_obj(dt);
5498 struct lu_object *lo = &dt->do_lu;
5499 struct iam_path_descr *ipd;
5500 struct iam_container *bag = &obj->oo_dir->od_container;
5502 if (!dt_object_exists(dt))
5503 return ERR_PTR(-ENOENT);
5507 return ERR_PTR(-ENOMEM);
5509 ipd = osd_it_ipd_get(env, bag);
5510 if (likely(ipd != NULL)) {
5514 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
5515 return (struct dt_it *)it;
5518 return ERR_PTR(-ENOMEM);
5523 * free given Iterator.
5525 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
5527 struct osd_it_iam *it = (struct osd_it_iam *)di;
5528 struct osd_object *obj = it->oi_obj;
5530 iam_it_fini(&it->oi_it);
5531 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
5532 osd_object_put(env, obj);
5537 * Move Iterator to record specified by \a key
5539 * \param di osd iterator
5540 * \param key key for index
5542 * \retval +ve di points to record with least key not larger than key
5543 * \retval 0 di points to exact matched key
5544 * \retval -ve failure
5547 static int osd_it_iam_get(const struct lu_env *env,
5548 struct dt_it *di, const struct dt_key *key)
5550 struct osd_thread_info *oti = osd_oti_get(env);
5551 struct osd_it_iam *it = (struct osd_it_iam *)di;
5553 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5554 /* swab quota uid/gid */
5555 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5556 key = (struct dt_key *)&oti->oti_quota_id;
5559 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
5565 * \param di osd iterator
5567 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
5569 struct osd_it_iam *it = (struct osd_it_iam *)di;
5571 iam_it_put(&it->oi_it);
5575 * Move iterator by one record
5577 * \param di osd iterator
5579 * \retval +1 end of container reached
5581 * \retval -ve failure
5584 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
5586 struct osd_it_iam *it = (struct osd_it_iam *)di;
5588 return iam_it_next(&it->oi_it);
5592 * Return pointer to the key under iterator.
5595 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
5596 const struct dt_it *di)
5598 struct osd_thread_info *oti = osd_oti_get(env);
5599 struct osd_it_iam *it = (struct osd_it_iam *)di;
5600 struct osd_object *obj = it->oi_obj;
5603 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
5605 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
5606 /* swab quota uid/gid */
5607 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
5608 key = (struct dt_key *)&oti->oti_quota_id;
5615 * Return size of key under iterator (in bytes)
5618 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
5620 struct osd_it_iam *it = (struct osd_it_iam *)di;
5622 return iam_it_key_size(&it->oi_it);
5626 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
5628 /* check if file type is required */
5629 if (ent->lde_attrs & LUDA_TYPE) {
5630 struct luda_type *lt;
5631 int align = sizeof(*lt) - 1;
5633 len = (len + align) & ~align;
5634 lt = (struct luda_type *)(ent->lde_name + len);
5635 lt->lt_type = cpu_to_le16(DTTOIF(type));
5638 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
5642 * build lu direct from backend fs dirent.
5646 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
5647 char *name, __u16 namelen, __u16 type, __u32 attr)
5649 ent->lde_attrs = attr | LUDA_FID;
5650 fid_cpu_to_le(&ent->lde_fid, fid);
5652 ent->lde_hash = cpu_to_le64(offset);
5653 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
5655 strncpy(ent->lde_name, name, namelen);
5656 ent->lde_name[namelen] = '\0';
5657 ent->lde_namelen = cpu_to_le16(namelen);
5659 /* append lustre attributes */
5660 osd_it_append_attrs(ent, namelen, type);
5664 * Return pointer to the record under iterator.
5666 static int osd_it_iam_rec(const struct lu_env *env,
5667 const struct dt_it *di,
5668 struct dt_rec *dtrec, __u32 attr)
5670 struct osd_it_iam *it = (struct osd_it_iam *)di;
5671 struct osd_thread_info *info = osd_oti_get(env);
5674 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
5675 const struct osd_fid_pack *rec;
5676 struct lu_fid *fid = &info->oti_fid;
5677 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
5683 name = (char *)iam_it_key_get(&it->oi_it);
5685 RETURN(PTR_ERR(name));
5687 namelen = iam_it_key_size(&it->oi_it);
5689 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
5691 RETURN(PTR_ERR(rec));
5693 rc = osd_fid_unpack(fid, rec);
5697 hash = iam_it_store(&it->oi_it);
5699 /* IAM does not store object type in IAM index (dir) */
5700 osd_it_pack_dirent(lde, fid, hash, name, namelen,
5702 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5703 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5704 (struct iam_rec *)dtrec);
5705 osd_quota_unpack(it->oi_obj, dtrec);
5707 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5708 (struct iam_rec *)dtrec);
5715 * Returns cookie for current Iterator position.
5717 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
5719 struct osd_it_iam *it = (struct osd_it_iam *)di;
5721 return iam_it_store(&it->oi_it);
5725 * Restore iterator from cookie.
5727 * \param di osd iterator
5728 * \param hash Iterator location cookie
5730 * \retval +ve di points to record with least key not larger than key.
5731 * \retval 0 di points to exact matched key
5732 * \retval -ve failure
5735 static int osd_it_iam_load(const struct lu_env *env,
5736 const struct dt_it *di, __u64 hash)
5738 struct osd_it_iam *it = (struct osd_it_iam *)di;
5740 return iam_it_load(&it->oi_it, hash);
5743 static const struct dt_index_operations osd_index_iam_ops = {
5744 .dio_lookup = osd_index_iam_lookup,
5745 .dio_declare_insert = osd_index_declare_iam_insert,
5746 .dio_insert = osd_index_iam_insert,
5747 .dio_declare_delete = osd_index_declare_iam_delete,
5748 .dio_delete = osd_index_iam_delete,
5750 .init = osd_it_iam_init,
5751 .fini = osd_it_iam_fini,
5752 .get = osd_it_iam_get,
5753 .put = osd_it_iam_put,
5754 .next = osd_it_iam_next,
5755 .key = osd_it_iam_key,
5756 .key_size = osd_it_iam_key_size,
5757 .rec = osd_it_iam_rec,
5758 .store = osd_it_iam_store,
5759 .load = osd_it_iam_load
5765 * Creates or initializes iterator context.
5767 * \retval struct osd_it_ea, iterator structure on success
5770 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
5771 struct dt_object *dt,
5774 struct osd_object *obj = osd_dt_obj(dt);
5775 struct osd_thread_info *info = osd_oti_get(env);
5776 struct osd_it_ea *oie;
5778 struct lu_object *lo = &dt->do_lu;
5779 struct dentry *obj_dentry;
5782 if (!dt_object_exists(dt) || obj->oo_destroyed)
5783 RETURN(ERR_PTR(-ENOENT));
5785 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
5787 RETURN(ERR_PTR(-ENOMEM));
5788 obj_dentry = &oie->oie_dentry;
5790 obj_dentry->d_inode = obj->oo_inode;
5791 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
5792 obj_dentry->d_name.hash = 0;
5794 oie->oie_rd_dirent = 0;
5795 oie->oie_it_dirent = 0;
5796 oie->oie_dirent = NULL;
5797 if (unlikely(!info->oti_it_ea_buf_used)) {
5798 oie->oie_buf = info->oti_it_ea_buf;
5799 info->oti_it_ea_buf_used = 1;
5801 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5802 if (oie->oie_buf == NULL)
5803 RETURN(ERR_PTR(-ENOMEM));
5807 file = &oie->oie_file;
5809 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
5810 if (attr & LUDA_64BITHASH)
5811 file->f_mode = FMODE_64BITHASH;
5813 file->f_mode = FMODE_32BITHASH;
5814 file->f_path.dentry = obj_dentry;
5815 file->f_mapping = obj->oo_inode->i_mapping;
5816 file->f_op = obj->oo_inode->i_fop;
5817 set_file_inode(file, obj->oo_inode);
5820 RETURN((struct dt_it *) oie);
5824 * Destroy or finishes iterator context.
5826 * \param di iterator structure to be destroyed
5828 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
5830 struct osd_thread_info *info = osd_oti_get(env);
5831 struct osd_it_ea *oie = (struct osd_it_ea *)di;
5832 struct osd_object *obj = oie->oie_obj;
5833 struct inode *inode = obj->oo_inode;
5836 oie->oie_file.f_op->release(inode, &oie->oie_file);
5837 osd_object_put(env, obj);
5838 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
5839 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5841 info->oti_it_ea_buf_used = 0;
5842 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
5847 * It position the iterator at given key, so that next lookup continues from
5848 * that key Or it is similar to dio_it->load() but based on a key,
5849 * rather than file position.
5851 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
5854 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
5856 static int osd_it_ea_get(const struct lu_env *env,
5857 struct dt_it *di, const struct dt_key *key)
5859 struct osd_it_ea *it = (struct osd_it_ea *)di;
5862 LASSERT(((const char *)key)[0] == '\0');
5863 it->oie_file.f_pos = 0;
5864 it->oie_rd_dirent = 0;
5865 it->oie_it_dirent = 0;
5866 it->oie_dirent = NULL;
5874 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
5878 struct osd_filldir_cbs {
5879 #ifdef HAVE_DIR_CONTEXT
5880 struct dir_context ctx;
5882 struct osd_it_ea *it;
5885 * It is called internally by ->readdir(). It fills the
5886 * iterator's in-memory data structure with required
5887 * information i.e. name, namelen, rec_size etc.
5889 * \param buf in which information to be filled in.
5890 * \param name name of the file in given dir
5892 * \retval 0 on success
5893 * \retval 1 on buffer full
5895 #ifdef HAVE_FILLDIR_USE_CTX
5896 static int osd_ldiskfs_filldir(struct dir_context *buf,
5898 static int osd_ldiskfs_filldir(void *buf,
5900 const char *name, int namelen,
5901 loff_t offset, __u64 ino, unsigned d_type)
5903 struct osd_it_ea *it =
5904 ((struct osd_filldir_cbs *)buf)->it;
5905 struct osd_object *obj = it->oie_obj;
5906 struct osd_it_ea_dirent *ent = it->oie_dirent;
5907 struct lu_fid *fid = &ent->oied_fid;
5908 struct osd_fid_pack *rec;
5911 /* this should never happen */
5912 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
5913 CERROR("ldiskfs return invalid namelen %d\n", namelen);
5917 if ((void *) ent - it->oie_buf + sizeof(*ent) + namelen >
5921 /* "." is just the object itself. */
5922 if (namelen == 1 && name[0] == '.') {
5923 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5924 } else if (d_type & LDISKFS_DIRENT_LUFID) {
5925 rec = (struct osd_fid_pack*) (name + namelen + 1);
5926 if (osd_fid_unpack(fid, rec) != 0)
5931 d_type &= ~LDISKFS_DIRENT_LUFID;
5933 /* NOT export local root. */
5934 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
5935 ino = obj->oo_inode->i_ino;
5936 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5939 ent->oied_ino = ino;
5940 ent->oied_off = offset;
5941 ent->oied_namelen = namelen;
5942 ent->oied_type = d_type;
5944 memcpy(ent->oied_name, name, namelen);
5946 it->oie_rd_dirent++;
5947 it->oie_dirent = (void *) ent + cfs_size_round(sizeof(*ent) + namelen);
5952 * Calls ->readdir() to load a directory entry at a time
5953 * and stored it in iterator's in-memory data structure.
5955 * \param di iterator's in memory structure
5957 * \retval 0 on success
5958 * \retval -ve on error
5959 * \retval +1 reach the end of entry
5961 static int osd_ldiskfs_it_fill(const struct lu_env *env,
5962 const struct dt_it *di)
5964 struct osd_it_ea *it = (struct osd_it_ea *)di;
5965 struct osd_object *obj = it->oie_obj;
5966 struct inode *inode = obj->oo_inode;
5967 struct htree_lock *hlock = NULL;
5968 struct file *filp = &it->oie_file;
5970 struct osd_filldir_cbs buf = {
5971 #ifdef HAVE_DIR_CONTEXT
5972 .ctx.actor = osd_ldiskfs_filldir,
5978 it->oie_dirent = it->oie_buf;
5979 it->oie_rd_dirent = 0;
5981 if (obj->oo_hl_head != NULL) {
5982 hlock = osd_oti_get(env)->oti_hlock;
5983 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5984 inode, LDISKFS_HLOCK_READDIR);
5986 down_read(&obj->oo_ext_idx_sem);
5989 #ifdef HAVE_DIR_CONTEXT
5990 buf.ctx.pos = filp->f_pos;
5991 rc = inode->i_fop->iterate(filp, &buf.ctx);
5992 filp->f_pos = buf.ctx.pos;
5994 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
5998 ldiskfs_htree_unlock(hlock);
6000 up_read(&obj->oo_ext_idx_sem);
6002 if (it->oie_rd_dirent == 0) {
6003 /*If it does not get any dirent, it means it has been reached
6004 *to the end of the dir */
6005 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6009 it->oie_dirent = it->oie_buf;
6010 it->oie_it_dirent = 1;
6017 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6018 * to load a directory entry at a time and stored it in
6019 * iterator's in-memory data structure.
6021 * \param di iterator's in memory structure
6023 * \retval +ve iterator reached to end
6024 * \retval 0 iterator not reached to end
6025 * \retval -ve on error
6027 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6029 struct osd_it_ea *it = (struct osd_it_ea *)di;
6034 if (it->oie_it_dirent < it->oie_rd_dirent) {
6036 (void *) it->oie_dirent +
6037 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6038 it->oie_dirent->oied_namelen);
6039 it->oie_it_dirent++;
6042 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6045 rc = osd_ldiskfs_it_fill(env, di);
6052 * Returns the key at current position from iterator's in memory structure.
6054 * \param di iterator's in memory structure
6056 * \retval key i.e. struct dt_key on success
6058 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6059 const struct dt_it *di)
6061 struct osd_it_ea *it = (struct osd_it_ea *)di;
6063 return (struct dt_key *)it->oie_dirent->oied_name;
6067 * Returns the key's size at current position from iterator's in memory structure.
6069 * \param di iterator's in memory structure
6071 * \retval key_size i.e. struct dt_key on success
6073 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6075 struct osd_it_ea *it = (struct osd_it_ea *)di;
6077 return it->oie_dirent->oied_namelen;
6080 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6082 if (LDISKFS_DIR_REC_LEN(de) >=
6083 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6090 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6091 unsigned blocksize, bool dotdot)
6094 return osd_dotdot_has_space(de);
6096 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6097 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6104 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6105 handle_t *jh, struct dentry *dentry,
6106 const struct lu_fid *fid, struct buffer_head *bh,
6107 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6110 struct inode *dir = dentry->d_parent->d_inode;
6111 struct inode *inode = dentry->d_inode;
6112 struct osd_fid_pack *rec;
6113 struct ldiskfs_dentry_param *ldp;
6114 int namelen = dentry->d_name.len;
6116 struct osd_thread_info *info = osd_oti_get(env);
6119 if (!LDISKFS_HAS_INCOMPAT_FEATURE(inode->i_sb,
6120 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
6123 /* There is enough space to hold the FID-in-dirent. */
6124 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6125 rc = ldiskfs_journal_get_write_access(jh, bh);
6129 de->name[namelen] = 0;
6130 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6131 rec->fp_len = sizeof(struct lu_fid) + 1;
6132 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6133 de->file_type |= LDISKFS_DIRENT_LUFID;
6134 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6141 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6145 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6146 osd_get_ldiskfs_dirent_param(ldp, fid);
6147 dentry->d_fsdata = (void *)ldp;
6148 ll_vfs_dq_init(dir);
6149 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6150 /* It is too bad, we cannot reinsert the name entry back.
6151 * That means we lose it! */
6153 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6154 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6155 osd_ino2name(inode),
6156 dir->i_ino, dir->i_generation, namelen,
6157 dentry->d_name.name, PFID(fid), rc);
6163 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6164 struct osd_it_ea *it, struct lu_fid *fid,
6165 struct osd_inode_id *id, __u32 *attr)
6167 struct osd_thread_info *info = osd_oti_get(env);
6168 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6169 struct osd_device *dev = osd_obj2dev(obj);
6170 struct super_block *sb = osd_sb(dev);
6171 const char *devname = osd_name(dev);
6172 struct osd_it_ea_dirent *ent = it->oie_dirent;
6173 struct inode *dir = obj->oo_inode;
6174 struct htree_lock *hlock = NULL;
6175 struct buffer_head *bh = NULL;
6176 handle_t *jh = NULL;
6177 struct ldiskfs_dir_entry_2 *de;
6178 struct dentry *dentry;
6179 struct inode *inode;
6180 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6183 bool dotdot = false;
6187 if (ent->oied_name[0] == '.') {
6188 if (ent->oied_namelen == 1)
6191 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6195 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6196 inode = osd_iget(info, dev, id);
6197 if (IS_ERR(inode)) {
6198 rc = PTR_ERR(inode);
6199 if (rc == -ENOENT || rc == -ESTALE) {
6200 /* Maybe dangling name entry, or
6201 * corrupted directory entry. */
6202 *attr |= LUDA_UNKNOWN;
6205 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6206 "check_repair, dir = %lu/%u, name = %.*s, "
6207 "ino = %llu, rc = %d\n",
6208 devname, dir->i_ino, dir->i_generation,
6209 ent->oied_namelen, ent->oied_name,
6216 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6218 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6219 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6224 /* We need to ensure that the name entry is still valid.
6225 * Because it may be removed or renamed by other already.
6227 * The unlink or rename operation will start journal before PDO lock,
6228 * so to avoid deadlock, here we need to start journal handle before
6229 * related PDO lock also. But because we do not know whether there
6230 * will be something to be repaired before PDO lock, we just start
6231 * journal without conditions.
6233 * We may need to remove the name entry firstly, then insert back.
6234 * One credit is for user quota file update.
6235 * One credit is for group quota file update.
6236 * Two credits are for dirty inode. */
6237 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6238 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6240 if (dev->od_dirent_journal != 0) {
6243 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6246 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6247 "check_repair, dir = %lu/%u, credits = %d, "
6248 "name = %.*s, ino = %llu: rc = %d\n",
6249 devname, dir->i_ino, dir->i_generation, credits,
6250 ent->oied_namelen, ent->oied_name,
6253 GOTO(out_inode, rc);
6256 if (obj->oo_hl_head != NULL) {
6257 hlock = osd_oti_get(env)->oti_hlock;
6258 /* "0" means exclusive lock for the whole directory.
6259 * We need to prevent others access such name entry
6260 * during the delete + insert. Neither HLOCK_ADD nor
6261 * HLOCK_DEL cannot guarantee the atomicity. */
6262 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6264 down_write(&obj->oo_ext_idx_sem);
6267 if (obj->oo_hl_head != NULL) {
6268 hlock = osd_oti_get(env)->oti_hlock;
6269 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6270 LDISKFS_HLOCK_LOOKUP);
6272 down_read(&obj->oo_ext_idx_sem);
6276 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6277 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6278 *attr |= LUDA_IGNORE;
6283 /* For dotdot entry, if there is not enough space to hold the
6284 * FID-in-dirent, just keep them there. It only happens when the
6285 * device upgraded from 1.8 or restored from MDT file-level backup.
6286 * For the whole directory, only dotdot entry have no FID-in-dirent
6287 * and needs to get FID from LMA when readdir, it will not affect the
6288 * performance much. */
6289 if (dotdot && !osd_dotdot_has_space(de)) {
6290 *attr |= LUDA_UNKNOWN;
6296 if (lu_fid_eq(fid, &lma->lma_self_fid))
6299 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6300 struct lu_fid *tfid = &lma->lma_self_fid;
6302 if (likely(dotdot &&
6303 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6304 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6305 /* It must be REMOTE_PARENT_DIR and as the
6306 * 'dotdot' entry of remote directory */
6307 *attr |= LUDA_IGNORE;
6309 CDEBUG(D_LFSCK, "%s: expect remote agent "
6310 "parent directory, but got %.*s under "
6311 "dir = %lu/%u with the FID "DFID"\n",
6312 devname, ent->oied_namelen,
6313 ent->oied_name, dir->i_ino,
6314 dir->i_generation, PFID(tfid));
6316 *attr |= LUDA_UNKNOWN;
6323 if (!fid_is_zero(fid)) {
6324 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6326 if (rc == -ENOENT ||
6328 !(dev->od_scrub.os_file.sf_flags & SF_UPGRADE))) {
6329 /* linkEA does not recognize the dirent entry,
6330 * it may because the dirent entry corruption
6331 * and points to other's inode. */
6332 CDEBUG(D_LFSCK, "%s: the target inode does not "
6333 "recognize the dirent, dir = %lu/%u, "
6334 " name = %.*s, ino = %llu, "
6335 DFID": rc = %d\n", devname, dir->i_ino,
6336 dir->i_generation, ent->oied_namelen,
6337 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6338 *attr |= LUDA_UNKNOWN;
6343 if (rc && rc != -ENODATA) {
6344 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6345 "dir = %lu/%u, name = %.*s, ino = %llu, "
6346 DFID": rc = %d\n", devname, dir->i_ino,
6347 dir->i_generation, ent->oied_namelen,
6348 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6349 *attr |= LUDA_UNKNOWN;
6356 /* linkEA recognizes the dirent entry, the FID-in-LMA is
6357 * valid, trusted, in spite of fid_is_sane(fid) or not. */
6358 if (*attr & LUDA_VERIFY_DRYRUN) {
6359 *fid = lma->lma_self_fid;
6360 *attr |= LUDA_REPAIR;
6367 dev->od_dirent_journal = 1;
6368 if (hlock != NULL) {
6369 ldiskfs_htree_unlock(hlock);
6372 up_read(&obj->oo_ext_idx_sem);
6378 *fid = lma->lma_self_fid;
6380 /* Update or append the FID-in-dirent. */
6381 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6382 bh, de, hlock, dotdot);
6384 *attr |= LUDA_REPAIR;
6386 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6387 "the dirent, dir = %lu/%u, name = %.*s, "
6388 "ino = %llu, "DFID": rc = %d\n",
6389 devname, dir->i_ino, dir->i_generation,
6390 ent->oied_namelen, ent->oied_name,
6391 ent->oied_ino, PFID(fid), rc);
6393 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6394 if (*attr & LUDA_VERIFY_DRYRUN) {
6395 if (fid_is_sane(fid)) {
6396 *attr |= LUDA_REPAIR;
6397 } else if (dev->od_index == 0) {
6398 lu_igif_build(fid, inode->i_ino,
6399 inode->i_generation);
6400 *attr |= LUDA_UPGRADE;
6408 dev->od_dirent_journal = 1;
6409 if (hlock != NULL) {
6410 ldiskfs_htree_unlock(hlock);
6413 up_read(&obj->oo_ext_idx_sem);
6420 if (unlikely(fid_is_sane(fid))) {
6421 /* FID-in-dirent exists, but FID-in-LMA is lost.
6422 * Trust the FID-in-dirent, and add FID-in-LMA. */
6423 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
6425 *attr |= LUDA_REPAIR;
6427 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
6428 "update dirent, dir = %lu/%u, "
6429 "name = %.*s, ino = %llu, "
6431 devname, dir->i_ino, dir->i_generation,
6432 ent->oied_namelen, ent->oied_name,
6433 ent->oied_ino, PFID(fid), rc);
6434 } else if (dev->od_index == 0) {
6435 lu_igif_build(fid, inode->i_ino, inode->i_generation);
6436 /* It is probably IGIF object. Only aappend the
6437 * FID-in-dirent. OI scrub will process FID-in-LMA. */
6438 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6439 bh, de, hlock, dotdot);
6441 *attr |= LUDA_UPGRADE;
6443 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
6444 "after the dirent, dir = %lu/%u, "
6445 "name = %.*s, ino = %llu, "
6447 devname, dir->i_ino, dir->i_generation,
6448 ent->oied_namelen, ent->oied_name,
6449 ent->oied_ino, PFID(fid), rc);
6458 if (hlock != NULL) {
6459 ldiskfs_htree_unlock(hlock);
6461 if (dev->od_dirent_journal != 0)
6462 up_write(&obj->oo_ext_idx_sem);
6464 up_read(&obj->oo_ext_idx_sem);
6468 ldiskfs_journal_stop(jh);
6472 if (rc >= 0 && !dirty)
6473 dev->od_dirent_journal = 0;
6479 * Returns the value at current position from iterator's in memory structure.
6481 * \param di struct osd_it_ea, iterator's in memory structure
6482 * \param attr attr requested for dirent.
6483 * \param lde lustre dirent
6485 * \retval 0 no error and \param lde has correct lustre dirent.
6486 * \retval -ve on error
6488 static inline int osd_it_ea_rec(const struct lu_env *env,
6489 const struct dt_it *di,
6490 struct dt_rec *dtrec, __u32 attr)
6492 struct osd_it_ea *it = (struct osd_it_ea *)di;
6493 struct osd_object *obj = it->oie_obj;
6494 struct osd_device *dev = osd_obj2dev(obj);
6495 struct osd_thread_info *oti = osd_oti_get(env);
6496 struct osd_inode_id *id = &oti->oti_id;
6497 struct lu_fid *fid = &it->oie_dirent->oied_fid;
6498 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6499 __u32 ino = it->oie_dirent->oied_ino;
6503 LASSERT(obj->oo_inode != dev->od_mdt_map->omm_remote_parent->d_inode);
6505 if (attr & LUDA_VERIFY) {
6506 if (unlikely(ino == osd_remote_parent_ino(dev))) {
6507 attr |= LUDA_IGNORE;
6508 /* If the parent is on remote MDT, and there
6509 * is no FID-in-dirent, then we have to get
6510 * the parent FID from the linkEA. */
6511 if (!fid_is_sane(fid) &&
6512 it->oie_dirent->oied_namelen == 2 &&
6513 it->oie_dirent->oied_name[0] == '.' &&
6514 it->oie_dirent->oied_name[1] == '.')
6515 osd_get_pfid_from_linkea(env, obj, fid);
6517 rc = osd_dirent_check_repair(env, obj, it, fid, id,
6521 if (!fid_is_sane(fid))
6522 attr |= LUDA_UNKNOWN;
6524 attr &= ~LU_DIRENT_ATTRS_MASK;
6525 if (!fid_is_sane(fid)) {
6526 bool is_dotdot = false;
6527 if (it->oie_dirent->oied_namelen == 2 &&
6528 it->oie_dirent->oied_name[0] == '.' &&
6529 it->oie_dirent->oied_name[1] == '.')
6531 /* If the parent is on remote MDT, and there
6532 * is no FID-in-dirent, then we have to get
6533 * the parent FID from the linkEA. */
6534 if (ino == osd_remote_parent_ino(dev) && is_dotdot) {
6535 rc = osd_get_pfid_from_linkea(env, obj, fid);
6537 if (is_dotdot == false &&
6538 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
6541 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6544 osd_id_gen(id, ino, OSD_OII_NOGEN);
6548 /* Pack the entry anyway, at least the offset is right. */
6549 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
6550 it->oie_dirent->oied_name,
6551 it->oie_dirent->oied_namelen,
6552 it->oie_dirent->oied_type, attr);
6557 if (osd_remote_fid(env, dev, fid))
6560 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
6561 osd_add_oi_cache(oti, dev, id, fid);
6563 RETURN(rc > 0 ? 0 : rc);
6567 * Returns the record size size at current position.
6569 * This function will return record(lu_dirent) size in bytes.
6571 * \param[in] env execution environment
6572 * \param[in] di iterator's in memory structure
6573 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
6574 * calculate the lu_dirent size.
6576 * \retval record size(in bytes & in memory) of the current lu_dirent
6579 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
6582 struct osd_it_ea *it = (struct osd_it_ea *)di;
6584 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
6588 * Returns a cookie for current position of the iterator head, so that
6589 * user can use this cookie to load/start the iterator next time.
6591 * \param di iterator's in memory structure
6593 * \retval cookie for current position, on success
6595 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
6597 struct osd_it_ea *it = (struct osd_it_ea *)di;
6599 return it->oie_dirent->oied_off;
6603 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6604 * to load a directory entry at a time and stored it i inn,
6605 * in iterator's in-memory data structure.
6607 * \param di struct osd_it_ea, iterator's in memory structure
6609 * \retval +ve on success
6610 * \retval -ve on error
6612 static int osd_it_ea_load(const struct lu_env *env,
6613 const struct dt_it *di, __u64 hash)
6615 struct osd_it_ea *it = (struct osd_it_ea *)di;
6619 it->oie_file.f_pos = hash;
6621 rc = osd_ldiskfs_it_fill(env, di);
6632 * Index lookup function for interoperability mode (b11826).
6634 * \param key, key i.e. file name to be searched
6636 * \retval +ve, on success
6637 * \retval -ve, on error
6639 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
6640 struct dt_rec *rec, const struct dt_key *key)
6642 struct osd_object *obj = osd_dt_obj(dt);
6647 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
6648 LINVRNT(osd_invariant(obj));
6650 rc = osd_ea_lookup_rec(env, obj, rec, key);
6657 * Index and Iterator operations for interoperability
6658 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
6660 static const struct dt_index_operations osd_index_ea_ops = {
6661 .dio_lookup = osd_index_ea_lookup,
6662 .dio_declare_insert = osd_index_declare_ea_insert,
6663 .dio_insert = osd_index_ea_insert,
6664 .dio_declare_delete = osd_index_declare_ea_delete,
6665 .dio_delete = osd_index_ea_delete,
6667 .init = osd_it_ea_init,
6668 .fini = osd_it_ea_fini,
6669 .get = osd_it_ea_get,
6670 .put = osd_it_ea_put,
6671 .next = osd_it_ea_next,
6672 .key = osd_it_ea_key,
6673 .key_size = osd_it_ea_key_size,
6674 .rec = osd_it_ea_rec,
6675 .rec_size = osd_it_ea_rec_size,
6676 .store = osd_it_ea_store,
6677 .load = osd_it_ea_load
6681 static void *osd_key_init(const struct lu_context *ctx,
6682 struct lu_context_key *key)
6684 struct osd_thread_info *info;
6686 OBD_ALLOC_PTR(info);
6688 return ERR_PTR(-ENOMEM);
6690 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6691 if (info->oti_it_ea_buf == NULL)
6694 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
6696 info->oti_hlock = ldiskfs_htree_lock_alloc();
6697 if (info->oti_hlock == NULL)
6703 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6706 return ERR_PTR(-ENOMEM);
6709 static void osd_key_fini(const struct lu_context *ctx,
6710 struct lu_context_key *key, void* data)
6712 struct osd_thread_info *info = data;
6713 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
6714 struct osd_idmap_cache *idc = info->oti_ins_cache;
6716 if (info->oti_inode != NULL)
6718 if (info->oti_hlock != NULL)
6719 ldiskfs_htree_lock_free(info->oti_hlock);
6720 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6721 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
6722 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
6723 lu_buf_free(&info->oti_big_buf);
6725 LASSERT(info->oti_ins_cache_size > 0);
6726 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
6727 info->oti_ins_cache = NULL;
6728 info->oti_ins_cache_size = 0;
6733 static void osd_key_exit(const struct lu_context *ctx,
6734 struct lu_context_key *key, void *data)
6736 struct osd_thread_info *info = data;
6738 LASSERT(info->oti_r_locks == 0);
6739 LASSERT(info->oti_w_locks == 0);
6740 LASSERT(info->oti_txns == 0);
6743 /* type constructor/destructor: osd_type_init, osd_type_fini */
6744 LU_TYPE_INIT_FINI(osd, &osd_key);
6746 struct lu_context_key osd_key = {
6747 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
6748 .lct_init = osd_key_init,
6749 .lct_fini = osd_key_fini,
6750 .lct_exit = osd_key_exit
6754 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
6755 const char *name, struct lu_device *next)
6757 struct osd_device *osd = osd_dev(d);
6759 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname))
6760 >= sizeof(osd->od_svname))
6762 return osd_procfs_init(osd, name);
6765 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
6767 struct seq_server_site *ss = osd_seq_site(osd);
6771 if (osd->od_is_ost || osd->od_cl_seq != NULL)
6774 if (unlikely(ss == NULL))
6777 OBD_ALLOC_PTR(osd->od_cl_seq);
6778 if (osd->od_cl_seq == NULL)
6781 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
6782 osd->od_svname, ss->ss_server_seq);
6784 OBD_FREE_PTR(osd->od_cl_seq);
6785 osd->od_cl_seq = NULL;
6789 if (ss->ss_node_id == 0) {
6790 /* If the OSD on the sequence controller(MDT0), then allocate
6791 * sequence here, otherwise allocate sequence after connected
6792 * to MDT0 (see mdt_register_lwp_callback()). */
6793 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
6794 &osd->od_cl_seq->lcs_space, env);
6800 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
6802 if (osd->od_cl_seq == NULL)
6805 seq_client_fini(osd->od_cl_seq);
6806 OBD_FREE_PTR(osd->od_cl_seq);
6807 osd->od_cl_seq = NULL;
6810 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
6814 /* shutdown quota slave instance associated with the device */
6815 if (o->od_quota_slave != NULL) {
6816 qsd_fini(env, o->od_quota_slave);
6817 o->od_quota_slave = NULL;
6820 osd_fid_fini(env, o);
6825 static void osd_umount(const struct lu_env *env, struct osd_device *o)
6829 if (o->od_mnt != NULL) {
6830 shrink_dcache_sb(osd_sb(o));
6831 osd_sync(env, &o->od_dt_dev);
6840 static int osd_mount(const struct lu_env *env,
6841 struct osd_device *o, struct lustre_cfg *cfg)
6843 const char *name = lustre_cfg_string(cfg, 0);
6844 const char *dev = lustre_cfg_string(cfg, 1);
6846 unsigned long page, s_flags, lmd_flags = 0;
6847 struct page *__page;
6848 struct file_system_type *type;
6849 char *options = NULL;
6851 struct osd_thread_info *info = osd_oti_get(env);
6852 struct lu_fid *fid = &info->oti_fid;
6853 struct inode *inode;
6854 int rc = 0, force_over_512tb = 0;
6857 if (o->od_mnt != NULL)
6860 if (strlen(dev) >= sizeof(o->od_mntdev))
6862 strcpy(o->od_mntdev, dev);
6864 str = lustre_cfg_string(cfg, 2);
6865 s_flags = simple_strtoul(str, NULL, 0);
6866 str = strstr(str, ":");
6868 lmd_flags = simple_strtoul(str + 1, NULL, 0);
6869 opts = lustre_cfg_string(cfg, 3);
6871 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
6872 CERROR("%s: device %s extents feature is not guaranteed to "
6873 "work on big-endian systems. Use \"bigendian_extents\" "
6874 "mount option to override.\n", name, dev);
6878 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
6879 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
6880 CWARN("force_over_128tb option is deprecated. "
6881 "Filesystems less than 512TB can be created without any "
6882 "force options. Use force_over_512tb option for "
6883 "filesystems greater than 512TB.\n");
6886 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
6887 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
6888 CWARN("force_over_256tb option is deprecated. "
6889 "Filesystems less than 512TB can be created without any "
6890 "force options. Use force_over_512tb option for "
6891 "filesystems greater than 512TB.\n");
6895 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
6896 force_over_512tb = 1;
6898 __page = alloc_page(GFP_KERNEL);
6900 GOTO(out, rc = -ENOMEM);
6901 page = (unsigned long)page_address(__page);
6902 options = (char *)page;
6905 /* strip out the options for back compatiblity */
6906 static char *sout[] = {
6913 /* strip out option we processed in osd */
6914 "bigendian_extents",
6920 strcat(options, opts);
6921 for (rc = 0, str = options; sout[rc]; ) {
6922 char *op = strstr(str, sout[rc]);
6928 if (op == options || *(op - 1) == ',') {
6929 str = op + strlen(sout[rc]);
6930 if (*str == ',' || *str == '\0') {
6931 *str == ',' ? str++ : str;
6932 memmove(op, str, strlen(str) + 1);
6935 for (str = op; *str != ',' && *str != '\0'; str++)
6939 strncat(options, "user_xattr,acl", 14);
6942 /* Glom up mount options */
6943 if (*options != '\0')
6944 strcat(options, ",");
6945 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
6947 type = get_fs_type("ldiskfs");
6949 CERROR("%s: cannot find ldiskfs module\n", name);
6950 GOTO(out, rc = -ENODEV);
6953 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
6954 module_put(type->owner);
6956 if (IS_ERR(o->od_mnt)) {
6957 rc = PTR_ERR(o->od_mnt);
6959 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
6963 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
6964 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
6965 force_over_512tb == 0) {
6966 CERROR("%s: device %s LDISKFS does not support filesystems "
6967 "greater than 512TB and can cause data corruption. "
6968 "Use \"force_over_512tb\" mount option to override.\n",
6970 GOTO(out_mnt, rc = -EINVAL);
6973 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
6974 if (priv_dev_set_rdonly) {
6975 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
6976 o->od_dt_dev.dd_rdonly = 1;
6977 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
6980 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
6983 GOTO(out_mnt, rc = -EOPNOTSUPP);
6985 } else if (priv_dev_check_rdonly &&
6986 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
6987 CERROR("%s: underlying device %s is marked as "
6988 "read-only. Setup failed\n", name, dev);
6990 GOTO(out_mnt, rc = -EROFS);
6993 if (!LDISKFS_HAS_COMPAT_FEATURE(o->od_mnt->mnt_sb,
6994 LDISKFS_FEATURE_COMPAT_HAS_JOURNAL)) {
6995 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
6996 GOTO(out_mnt, rc = -EINVAL);
6999 #ifdef LDISKFS_MOUNT_DIRDATA
7000 if (LDISKFS_HAS_INCOMPAT_FEATURE(o->od_mnt->mnt_sb,
7001 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
7002 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7003 else if (!o->od_is_ost)
7004 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7005 "enabling the dirdata feature. If you do not want to "
7006 "downgrade to Lustre-1.x again, you can enable it via "
7007 "'tune2fs -O dirdata device'\n", name, dev);
7009 inode = osd_sb(o)->s_root->d_inode;
7010 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7011 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7013 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7017 if (lmd_flags & LMD_FLG_NOSCRUB)
7028 __free_page(__page);
7033 static struct lu_device *osd_device_fini(const struct lu_env *env,
7034 struct lu_device *d)
7036 struct osd_device *o = osd_dev(d);
7039 osd_shutdown(env, o);
7041 osd_scrub_cleanup(env, o);
7042 osd_obj_map_fini(o);
7048 static int osd_device_init0(const struct lu_env *env,
7049 struct osd_device *o,
7050 struct lustre_cfg *cfg)
7052 struct lu_device *l = osd2lu_dev(o);
7053 struct osd_thread_info *info;
7057 /* if the module was re-loaded, env can loose its keys */
7058 rc = lu_env_refill((struct lu_env *) env);
7061 info = osd_oti_get(env);
7064 l->ld_ops = &osd_lu_ops;
7065 o->od_dt_dev.dd_ops = &osd_dt_ops;
7067 spin_lock_init(&o->od_osfs_lock);
7068 mutex_init(&o->od_otable_mutex);
7069 INIT_LIST_HEAD(&o->od_orphan_list);
7071 o->od_read_cache = 1;
7072 o->od_writethrough_cache = 1;
7073 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7075 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7076 sizeof(o->od_svname));
7077 if (cplen >= sizeof(o->od_svname)) {
7082 o->od_index = -1; /* -1 means index is invalid */
7083 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7084 if (rc == LDD_F_SV_TYPE_OST)
7087 o->od_full_scrub_ratio = OFSR_DEFAULT;
7088 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7089 rc = osd_mount(env, o, cfg);
7093 rc = osd_obj_map_init(env, o);
7097 rc = lu_site_init(&o->od_site, l);
7099 GOTO(out_compat, rc);
7100 o->od_site.ls_bottom_dev = l;
7102 rc = lu_site_init_finish(&o->od_site);
7106 INIT_LIST_HEAD(&o->od_ios_list);
7107 /* setup scrub, including OI files initialization */
7108 rc = osd_scrub_setup(env, o);
7112 rc = osd_procfs_init(o, o->od_svname);
7114 CERROR("%s: can't initialize procfs: rc = %d\n",
7116 GOTO(out_scrub, rc);
7119 LASSERT(l->ld_site->ls_linkage.next != NULL);
7120 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7122 /* initialize quota slave instance */
7123 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7125 if (IS_ERR(o->od_quota_slave)) {
7126 rc = PTR_ERR(o->od_quota_slave);
7127 o->od_quota_slave = NULL;
7128 GOTO(out_procfs, rc);
7136 osd_scrub_cleanup(env, o);
7138 lu_site_fini(&o->od_site);
7140 osd_obj_map_fini(o);
7147 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7148 struct lu_device_type *t,
7149 struct lustre_cfg *cfg)
7151 struct osd_device *o;
7156 return ERR_PTR(-ENOMEM);
7158 rc = dt_device_init(&o->od_dt_dev, t);
7160 /* Because the ctx might be revived in dt_device_init,
7161 * refill the env here */
7162 lu_env_refill((struct lu_env *)env);
7163 rc = osd_device_init0(env, o, cfg);
7165 dt_device_fini(&o->od_dt_dev);
7168 if (unlikely(rc != 0))
7171 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7174 static struct lu_device *osd_device_free(const struct lu_env *env,
7175 struct lu_device *d)
7177 struct osd_device *o = osd_dev(d);
7180 /* XXX: make osd top device in order to release reference */
7181 d->ld_site->ls_top_dev = d;
7182 lu_site_purge(env, d->ld_site, -1);
7183 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7184 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7185 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7187 lu_site_fini(&o->od_site);
7188 dt_device_fini(&o->od_dt_dev);
7193 static int osd_process_config(const struct lu_env *env,
7194 struct lu_device *d, struct lustre_cfg *cfg)
7196 struct osd_device *o = osd_dev(d);
7200 switch (cfg->lcfg_command) {
7202 rc = osd_mount(env, o, cfg);
7205 lu_dev_del_linkage(d->ld_site, d);
7206 rc = osd_shutdown(env, o);
7209 LASSERT(&o->od_dt_dev);
7210 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7211 cfg, &o->od_dt_dev);
7212 if (rc > 0 || rc == -ENOSYS)
7213 rc = class_process_proc_param(PARAM_OST,
7214 lprocfs_osd_obd_vars,
7215 cfg, &o->od_dt_dev);
7224 static int osd_recovery_complete(const struct lu_env *env,
7225 struct lu_device *d)
7227 struct osd_device *osd = osd_dev(d);
7231 if (osd->od_quota_slave == NULL)
7234 /* start qsd instance on recovery completion, this notifies the quota
7235 * slave code that we are about to process new requests now */
7236 rc = qsd_start(env, osd->od_quota_slave);
7241 * we use exports to track all osd users
7243 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7244 struct obd_device *obd, struct obd_uuid *cluuid,
7245 struct obd_connect_data *data, void *localdata)
7247 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7248 struct lustre_handle conn;
7252 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7254 rc = class_connect(&conn, obd, cluuid);
7258 *exp = class_conn2export(&conn);
7260 spin_lock(&osd->od_osfs_lock);
7262 spin_unlock(&osd->od_osfs_lock);
7268 * once last export (we don't count self-export) disappeared
7269 * osd can be released
7271 static int osd_obd_disconnect(struct obd_export *exp)
7273 struct obd_device *obd = exp->exp_obd;
7274 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7275 int rc, release = 0;
7278 /* Only disconnect the underlying layers on the final disconnect. */
7279 spin_lock(&osd->od_osfs_lock);
7281 if (osd->od_connects == 0)
7283 spin_unlock(&osd->od_osfs_lock);
7285 rc = class_disconnect(exp); /* bz 9811 */
7287 if (rc == 0 && release)
7288 class_manual_cleanup(obd);
7292 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7293 struct lu_device *dev)
7295 struct osd_device *osd = osd_dev(dev);
7296 struct lr_server_data *lsd =
7297 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7301 if (osd->od_quota_slave != NULL) {
7302 /* set up quota slave objects */
7303 result = qsd_prepare(env, osd->od_quota_slave);
7308 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7309 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7310 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7311 osd->od_index_in_idif = 1;
7313 osd->od_index_in_idif = 0;
7314 result = osd_register_proc_index_in_idif(osd);
7319 osd->od_index_in_idif = 1;
7323 result = osd_fid_init(env, osd);
7328 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7329 struct lu_fid *fid, struct md_op_data *op_data)
7331 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7333 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7336 static const struct lu_object_operations osd_lu_obj_ops = {
7337 .loo_object_init = osd_object_init,
7338 .loo_object_delete = osd_object_delete,
7339 .loo_object_release = osd_object_release,
7340 .loo_object_free = osd_object_free,
7341 .loo_object_print = osd_object_print,
7342 .loo_object_invariant = osd_object_invariant
7345 const struct lu_device_operations osd_lu_ops = {
7346 .ldo_object_alloc = osd_object_alloc,
7347 .ldo_process_config = osd_process_config,
7348 .ldo_recovery_complete = osd_recovery_complete,
7349 .ldo_prepare = osd_prepare,
7352 static const struct lu_device_type_operations osd_device_type_ops = {
7353 .ldto_init = osd_type_init,
7354 .ldto_fini = osd_type_fini,
7356 .ldto_start = osd_type_start,
7357 .ldto_stop = osd_type_stop,
7359 .ldto_device_alloc = osd_device_alloc,
7360 .ldto_device_free = osd_device_free,
7362 .ldto_device_init = osd_device_init,
7363 .ldto_device_fini = osd_device_fini
7366 static struct lu_device_type osd_device_type = {
7367 .ldt_tags = LU_DEVICE_DT,
7368 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
7369 .ldt_ops = &osd_device_type_ops,
7370 .ldt_ctx_tags = LCT_LOCAL,
7373 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
7375 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7376 struct super_block *sb = osd_sb(osd);
7378 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
7382 * lprocfs legacy support.
7384 static struct obd_ops osd_obd_device_ops = {
7385 .o_owner = THIS_MODULE,
7386 .o_connect = osd_obd_connect,
7387 .o_disconnect = osd_obd_disconnect,
7388 .o_fid_alloc = osd_fid_alloc,
7389 .o_health_check = osd_health_check,
7392 static int __init osd_init(void)
7396 LASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
7397 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
7398 /* please, try to keep osd_thread_info smaller than a page */
7399 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
7404 rc = lu_kmem_init(ldiskfs_caches);
7408 #ifdef CONFIG_KALLSYMS
7409 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
7410 priv_dev_check_rdonly = (void *)kallsyms_lookup_name("dev_check_rdonly");
7411 /* Clear readonly is unused at this time */
7412 /*priv_dev_clear_rdonly = (void *)kallsyms_lookup_name("dev_clear_rdonly");*/
7415 rc = class_register_type(&osd_obd_device_ops, NULL, true,
7416 lprocfs_osd_module_vars,
7417 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
7419 lu_kmem_fini(ldiskfs_caches);
7423 static void __exit osd_exit(void)
7425 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
7426 lu_kmem_fini(ldiskfs_caches);
7429 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7430 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
7431 MODULE_VERSION(LUSTRE_VERSION_STRING);
7432 MODULE_LICENSE("GPL");
7434 module_init(osd_init);
7435 module_exit(osd_exit);