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_18_23, 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 *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(struct filter_fid_18_23)) {
673 struct filter_fid_18_23 *ff_old = (void *)ff;
675 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
676 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
678 * XXX: use 0 as the index for compatibility, the caller will
679 * handle index related issues when necessary.
682 ostid_to_fid(fid, ostid, 0);
683 /* cast to (int) so "rc" comparison is signed instead of unsigned */
684 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
686 } else if (rc >= 0) {
693 static int osd_lma_self_repair(struct osd_thread_info *info,
694 struct osd_device *osd, struct inode *inode,
695 const struct lu_fid *fid, __u32 compat)
700 LASSERT(current->journal_info == NULL);
702 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
703 osd_dto_credits_noquota[DTO_XATTR_SET]);
706 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
711 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
713 CWARN("%s: cannot self repair the LMA: rc = %d\n",
715 ldiskfs_journal_stop(jh);
719 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
721 struct osd_thread_info *info = osd_oti_get(env);
722 struct osd_device *osd = osd_obj2dev(obj);
723 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
724 struct lustre_mdt_attrs *lma = &loa->loa_lma;
725 struct inode *inode = obj->oo_inode;
726 struct dentry *dentry = &info->oti_obj_dentry;
727 struct lu_fid *fid = NULL;
728 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
732 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
733 (void *)loa, sizeof(*loa));
734 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
735 fid = &lma->lma_self_fid;
736 rc = osd_get_idif(info, inode, dentry, fid);
737 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
738 /* For the given OST-object, if it has neither LMA nor
739 * FID in XATTR_NAME_FID, then the given FID (which is
740 * contained in the @obj, from client RPC for locating
741 * the OST-object) is trusted. We use it to generate
743 osd_lma_self_repair(info, osd, inode, rfid,
754 lustre_lma_swab(lma);
755 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
756 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
757 S_ISREG(inode->i_mode)))) {
758 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
759 "fid = "DFID", ino = %lu\n", osd_name(osd),
760 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
761 PFID(rfid), inode->i_ino);
764 fid = &lma->lma_self_fid;
768 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
769 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
770 struct ost_id *oi = &info->oti_ostid;
771 struct lu_fid *fid1 = &info->oti_fid3;
772 __u32 idx = fid_idif_ost_idx(rfid);
774 /* For old IDIF, the OST index is not part of the IDIF,
775 * Means that different OSTs may have the same IDIFs.
776 * Under such case, we need to make some compatible
777 * check to make sure to trigger OI scrub properly. */
778 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
779 /* Given @rfid is new, LMA is old. */
780 fid_to_ostid(fid, oi);
781 ostid_to_fid(fid1, oi, idx);
782 if (lu_fid_eq(fid1, rfid)) {
783 if (osd->od_index_in_idif)
784 osd_lma_self_repair(info, osd,
798 struct osd_check_lmv_buf {
799 #ifdef HAVE_DIR_CONTEXT
800 /* please keep it as first member */
801 struct dir_context ctx;
803 struct osd_thread_info *oclb_info;
804 struct osd_device *oclb_dev;
805 struct osd_idmap_cache *oclb_oic;
811 * It is called internally by ->readdir() to filter out the
812 * local slave object's FID of the striped directory.
814 * \retval 1 found the local slave's FID
815 * \retval 0 continue to check next item
816 * \retval -ve for failure
818 #ifdef HAVE_FILLDIR_USE_CTX
819 static int osd_stripe_dir_filldir(struct dir_context *buf,
821 static int osd_stripe_dir_filldir(void *buf,
823 const char *name, int namelen,
824 loff_t offset, __u64 ino, unsigned d_type)
826 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
827 struct osd_thread_info *oti = oclb->oclb_info;
828 struct lu_fid *fid = &oti->oti_fid3;
829 struct osd_inode_id *id = &oti->oti_id3;
830 struct osd_device *dev = oclb->oclb_dev;
831 struct osd_idmap_cache *oic = oclb->oclb_oic;
840 sscanf(name + 1, SFID, RFID(fid));
841 if (!fid_is_sane(fid))
844 if (osd_remote_fid(oti->oti_env, dev, fid))
847 osd_id_gen(id, ino, OSD_OII_NOGEN);
848 inode = osd_iget(oti, dev, id);
850 return PTR_ERR(inode);
853 osd_add_oi_cache(oti, dev, id, fid);
857 osd_oii_insert(dev, oic, true);
858 oclb->oclb_found = true;
863 /* When lookup item under striped directory, we need to locate the master
864 * MDT-object of the striped directory firstly, then the client will send
865 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
866 * and the item's name. If the system is restored from MDT file level backup,
867 * then before the OI scrub completely built the OI files, the OI mappings of
868 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
869 * not a problem for the master MDT-object. Because when locate the master
870 * MDT-object, we will do name based lookup (for the striped directory itself)
871 * firstly, during such process we can setup the correct OI mapping for the
872 * master MDT-object. But it will be trouble for the slave MDT-object. Because
873 * the client will not trigger name based lookup on the MDT to locate the slave
874 * MDT-object before locating item under the striped directory, then when
875 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
876 * is invalid and does not know what the right OI mapping is, then the MDT has
877 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
878 * the OI file, related OI mapping is unknown yet, please try again later. And
879 * then client will re-try the RPC again and again until related OI mapping has
880 * been updated. That is quite inefficient.
882 * To resolve above trouble, we will handle it as the following two cases:
884 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
885 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
886 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
887 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
888 * directly. Please check osd_fid_lookup().
890 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
891 * Under such case, during lookup the master MDT-object, we will lookup the
892 * slave MDT-object via readdir against the master MDT-object, because the
893 * slave MDT-objects information are stored as sub-directories with the name
894 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
895 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
896 * the correct OI mapping for the slave MDT-object. */
897 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
898 struct inode *inode, struct osd_idmap_cache *oic)
900 struct lu_buf *buf = &oti->oti_big_buf;
901 struct dentry *dentry = &oti->oti_obj_dentry;
902 struct file *filp = &oti->oti_file;
903 const struct file_operations *fops;
904 struct lmv_mds_md_v1 *lmv1;
905 struct osd_check_lmv_buf oclb = {
906 #ifdef HAVE_DIR_CONTEXT
907 .ctx.actor = osd_stripe_dir_filldir,
918 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
921 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
923 lu_buf_realloc(buf, rc);
924 if (buf->lb_buf == NULL)
925 GOTO(out, rc = -ENOMEM);
931 if (unlikely(rc == 0 || rc == -ENODATA))
937 if (unlikely(buf->lb_buf == NULL)) {
938 lu_buf_realloc(buf, rc);
939 if (buf->lb_buf == NULL)
940 GOTO(out, rc = -ENOMEM);
946 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
950 dentry->d_inode = inode;
951 dentry->d_sb = inode->i_sb;
953 filp->f_path.dentry = dentry;
954 filp->f_mode = FMODE_64BITHASH;
955 filp->f_mapping = inode->i_mapping;
957 filp->private_data = NULL;
958 set_file_inode(filp, inode);
962 #ifdef HAVE_DIR_CONTEXT
963 oclb.ctx.pos = filp->f_pos;
964 rc = fops->iterate(filp, &oclb.ctx);
965 filp->f_pos = oclb.ctx.pos;
967 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
969 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
970 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
971 fops->release(inode, filp);
975 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
976 DFID": rc = %d\n", osd_ino2name(inode),
977 inode->i_ino, inode->i_generation,
978 PFID(&oic->oic_fid), rc);
985 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
986 const struct lu_fid *fid,
987 const struct lu_object_conf *conf)
989 struct osd_thread_info *info;
990 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
991 struct osd_device *dev;
992 struct osd_idmap_cache *oic;
993 struct osd_inode_id *id;
994 struct inode *inode = NULL;
995 struct osd_scrub *scrub;
996 struct scrub_file *sf;
997 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1002 bool remote = false;
1003 bool trusted = true;
1004 bool updated = false;
1007 LINVRNT(osd_invariant(obj));
1008 LASSERT(obj->oo_inode == NULL);
1009 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1011 dev = osd_dev(ldev);
1012 scrub = &dev->od_scrub;
1013 sf = &scrub->os_file;
1014 info = osd_oti_get(env);
1016 oic = &info->oti_cache;
1018 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1021 /* For the object is created as locking anchor, or for the object to
1022 * be created on disk. No need to osd_oi_lookup() at here because FID
1023 * shouldn't never be re-used, if it's really a duplicate FID from
1024 * unexpected reason, we should be able to detect it later by calling
1025 * do_create->osd_oi_insert(). */
1026 if (conf && conf->loc_flags & LOC_F_NEW)
1027 GOTO(out, result = 0);
1029 /* Search order: 1. per-thread cache. */
1030 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1036 if (!list_empty(&scrub->os_inconsistent_items)) {
1037 /* Search order: 2. OI scrub pending list. */
1038 result = osd_oii_lookup(dev, fid, id);
1043 /* The OI mapping in the OI file can be updated by the OI scrub
1044 * when we locate the inode via FID. So it may be not trustable. */
1047 /* Search order: 3. OI files. */
1048 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1049 if (result == -ENOENT) {
1050 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1051 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1052 !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1054 GOTO(out, result = 0);
1063 obj->oo_inode = NULL;
1064 /* for later passes through checks, not true on first pass */
1065 if (!IS_ERR_OR_NULL(inode))
1068 inode = osd_iget_check(info, dev, fid, id, trusted);
1069 if (!IS_ERR(inode)) {
1070 obj->oo_inode = inode;
1078 result = PTR_ERR(inode);
1079 if (result == -ENOENT || result == -ESTALE)
1080 GOTO(out, result = 0);
1082 if (result != -EREMCHG)
1086 /* We still have chance to get the valid inode: for the
1087 * object which is referenced by remote name entry, the
1088 * object on the local MDT will be linked under the dir
1089 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1091 * We do not know whether the object for the given FID
1092 * is referenced by some remote name entry or not, and
1093 * especially for DNE II, a multiple-linked object may
1094 * have many name entries reside on many MDTs.
1096 * To simplify the operation, OSD will not distinguish
1097 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1098 * only happened for the RPC from other MDT during the
1099 * OI scrub, or for the client side RPC with FID only,
1100 * such as FID to path, or from old connected client. */
1102 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1106 flags |= SS_AUTO_PARTIAL;
1107 flags &= ~SS_AUTO_FULL;
1112 if (thread_is_running(&scrub->os_thread)) {
1113 if (scrub->os_partial_scan && !scrub->os_in_join)
1116 if (IS_ERR_OR_NULL(inode) || result)
1117 GOTO(out, result = -EINPROGRESS);
1120 LASSERT(obj->oo_inode == inode);
1122 osd_add_oi_cache(info, dev, id, fid);
1123 osd_oii_insert(dev, oic, true);
1127 if (dev->od_noscrub) {
1129 GOTO(out, result = -EREMCHG);
1132 LASSERT(obj->oo_inode == inode);
1134 osd_add_oi_cache(info, dev, id, fid);
1139 rc1 = osd_scrub_start(dev, flags);
1140 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1141 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1142 if (rc1 && rc1 != -EALREADY)
1143 GOTO(out, result = -EREMCHG);
1145 if (IS_ERR_OR_NULL(inode) || result)
1146 GOTO(out, result = -EINPROGRESS);
1149 LASSERT(obj->oo_inode == inode);
1151 osd_add_oi_cache(info, dev, id, fid);
1152 osd_oii_insert(dev, oic, true);
1156 result = osd_check_lma(env, obj);
1160 LASSERTF(id->oii_ino == inode->i_ino &&
1161 id->oii_gen == inode->i_generation,
1162 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1163 PFID(fid), id->oii_ino, id->oii_gen,
1164 inode->i_ino, inode->i_generation);
1166 saved_ino = inode->i_ino;
1167 saved_gen = inode->i_generation;
1169 if (unlikely(result == -ENODATA)) {
1170 /* If the OI scrub updated the OI mapping by race, it
1171 * must be valid. Trust the inode that has no LMA EA. */
1175 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1177 /* The OI mapping is still there, the inode is still
1178 * valid. It is just becaues the inode has no LMA EA. */
1179 if (saved_ino == id->oii_ino &&
1180 saved_gen == id->oii_gen)
1183 /* It is the OI scrub updated the OI mapping by race.
1184 * The new OI mapping must be valid. */
1190 /* "result == -ENOENT" means that the OI mappinghas been
1191 * removed by race, so the inode belongs to other object.
1193 * Others error can be returned directly. */
1194 if (result == -ENOENT) {
1197 obj->oo_inode = NULL;
1202 if (result != -EREMCHG)
1207 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1208 /* "result == -ENOENT" means the cached OI mapping has been removed
1209 * from the OI file by race, above inode belongs to other object. */
1210 if (result == -ENOENT) {
1213 obj->oo_inode = NULL;
1214 GOTO(out, result = 0);
1220 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1225 /* It is the OI scrub updated the OI mapping by race.
1226 * The new OI mapping must be valid. */
1232 obj->oo_compat_dot_created = 1;
1233 obj->oo_compat_dotdot_created = 1;
1235 if (S_ISDIR(inode->i_mode) &&
1236 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1237 osd_check_lmv(info, dev, inode, oic);
1239 result = osd_attach_jinode(inode);
1244 GOTO(out, result = 0);
1246 LASSERT(!obj->oo_hl_head);
1247 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1249 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1252 if (result || !obj->oo_inode) {
1253 if (!IS_ERR_OR_NULL(inode))
1256 obj->oo_inode = NULL;
1258 fid_zero(&oic->oic_fid);
1261 LINVRNT(osd_invariant(obj));
1266 * Concurrency: shouldn't matter.
1268 static void osd_object_init0(struct osd_object *obj)
1270 LASSERT(obj->oo_inode != NULL);
1271 obj->oo_dt.do_body_ops = &osd_body_ops;
1272 obj->oo_dt.do_lu.lo_header->loh_attr |=
1273 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1277 * Concurrency: no concurrent access is possible that early in object
1280 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1281 const struct lu_object_conf *conf)
1283 struct osd_object *obj = osd_obj(l);
1286 LINVRNT(osd_invariant(obj));
1288 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1289 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1290 l->lo_header->loh_attr |= LOHA_EXISTS;
1294 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1295 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1296 if (result == 0 && obj->oo_inode != NULL) {
1297 struct osd_thread_info *oti = osd_oti_get(env);
1298 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1300 osd_object_init0(obj);
1301 result = osd_get_lma(oti, obj->oo_inode,
1302 &oti->oti_obj_dentry, loa);
1304 /* Convert LMAI flags to lustre LMA flags
1305 * and cache it to oo_lma_flags */
1307 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1308 } else if (result == -ENODATA) {
1313 LINVRNT(osd_invariant(obj));
1317 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1318 * The left part is for value, binary mode. */
1319 struct osd_xattr_entry {
1320 struct list_head oxe_list;
1324 struct rcu_head oxe_rcu;
1328 static int osd_oxc_get(struct osd_object *obj, const char *name,
1331 struct osd_xattr_entry *tmp;
1332 struct osd_xattr_entry *oxe = NULL;
1333 size_t namelen = strlen(name);
1338 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1339 if (namelen == tmp->oxe_namelen &&
1340 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1347 GOTO(out, rc = -ENOENT);
1349 if (!oxe->oxe_exist)
1350 GOTO(out, rc = -ENODATA);
1353 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1356 if (buf->lb_buf == NULL)
1359 if (buf->lb_len < rc)
1360 GOTO(out, rc = -ERANGE);
1362 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1370 static void osd_oxc_free(struct rcu_head *head)
1372 struct osd_xattr_entry *oxe;
1374 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1375 OBD_FREE(oxe, oxe->oxe_len);
1378 static void osd_oxc_add(struct osd_object *obj, const char *name,
1379 const char *buf, int buflen)
1381 struct osd_xattr_entry *oxe;
1382 struct osd_xattr_entry *old = NULL;
1383 struct osd_xattr_entry *tmp;
1384 size_t namelen = strlen(name);
1385 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1387 OBD_ALLOC(oxe, len);
1391 INIT_LIST_HEAD(&oxe->oxe_list);
1393 oxe->oxe_namelen = namelen;
1394 memcpy(oxe->oxe_buf, name, namelen);
1396 LASSERT(buf != NULL);
1397 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1398 oxe->oxe_exist = true;
1400 oxe->oxe_exist = false;
1403 /* this should be rarely called, just remove old and add new */
1404 spin_lock(&obj->oo_guard);
1405 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1406 if (namelen == tmp->oxe_namelen &&
1407 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1413 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1414 call_rcu(&old->oxe_rcu, osd_oxc_free);
1416 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1418 spin_unlock(&obj->oo_guard);
1421 static void osd_oxc_del(struct osd_object *obj, const char *name)
1423 struct osd_xattr_entry *oxe;
1424 size_t namelen = strlen(name);
1426 spin_lock(&obj->oo_guard);
1427 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1428 if (namelen == oxe->oxe_namelen &&
1429 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1430 list_del_rcu(&oxe->oxe_list);
1431 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1435 spin_unlock(&obj->oo_guard);
1438 static void osd_oxc_fini(struct osd_object *obj)
1440 struct osd_xattr_entry *oxe, *next;
1442 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1443 list_del(&oxe->oxe_list);
1444 OBD_FREE(oxe, oxe->oxe_len);
1449 * Concurrency: no concurrent access is possible that late in object
1452 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1454 struct osd_object *obj = osd_obj(l);
1456 LINVRNT(osd_invariant(obj));
1459 dt_object_fini(&obj->oo_dt);
1460 if (obj->oo_hl_head != NULL)
1461 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1466 * Concurrency: no concurrent access is possible that late in object
1469 static void osd_index_fini(struct osd_object *o)
1471 struct iam_container *bag;
1473 if (o->oo_dir != NULL) {
1474 bag = &o->oo_dir->od_container;
1475 if (o->oo_inode != NULL) {
1476 if (bag->ic_object == o->oo_inode)
1477 iam_container_fini(bag);
1479 OBD_FREE_PTR(o->oo_dir);
1485 * Concurrency: no concurrent access is possible that late in object
1486 * life-cycle (for all existing callers, that is. New callers have to provide
1487 * their own locking.)
1489 static int osd_inode_unlinked(const struct inode *inode)
1491 return inode->i_nlink == 0;
1495 OSD_TXN_OI_DELETE_CREDITS = 20,
1496 OSD_TXN_INODE_DELETE_CREDITS = 20
1503 #if OSD_THANDLE_STATS
1505 * Set time when the handle is allocated
1507 static void osd_th_alloced(struct osd_thandle *oth)
1509 oth->oth_alloced = ktime_get();
1513 * Set time when the handle started
1515 static void osd_th_started(struct osd_thandle *oth)
1517 oth->oth_started = ktime_get();
1521 * Check whether the we deal with this handle for too long.
1523 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1524 ktime_t alloced, ktime_t started,
1527 ktime_t now = ktime_get();
1529 LASSERT(dev != NULL);
1531 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1532 ktime_us_delta(started, alloced));
1533 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1534 ktime_us_delta(closed, started));
1535 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1536 ktime_us_delta(now, closed));
1538 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1539 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1540 oth, now, alloced, started, closed);
1541 libcfs_debug_dumpstack(NULL);
1545 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1547 ktime_t __closed = ktime_get(); \
1548 ktime_t __alloced = oth->oth_alloced; \
1549 ktime_t __started = oth->oth_started; \
1552 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1555 #else /* OSD_THANDLE_STATS */
1557 #define osd_th_alloced(h) do {} while(0)
1558 #define osd_th_started(h) do {} while(0)
1559 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1561 #endif /* OSD_THANDLE_STATS */
1564 * Concurrency: doesn't access mutable data.
1566 static int osd_param_is_not_sane(const struct osd_device *dev,
1567 const struct thandle *th)
1569 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1571 return oh->ot_credits > osd_transaction_size(dev);
1575 * Concurrency: shouldn't matter.
1577 static void osd_trans_commit_cb(struct super_block *sb,
1578 struct ldiskfs_journal_cb_entry *jcb, int error)
1580 struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1581 struct thandle *th = &oh->ot_super;
1582 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1583 struct dt_txn_commit_cb *dcb, *tmp;
1585 LASSERT(oh->ot_handle == NULL);
1588 CERROR("transaction @0x%p commit error: %d\n", th, error);
1590 dt_txn_hook_commit(th);
1592 /* call per-transaction callbacks if any */
1593 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1595 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1596 "commit callback entry: magic=%x name='%s'\n",
1597 dcb->dcb_magic, dcb->dcb_name);
1598 list_del_init(&dcb->dcb_linkage);
1599 dcb->dcb_func(NULL, th, dcb, error);
1602 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1609 #ifndef HAVE_SB_START_WRITE
1610 # define sb_start_write(sb) do {} while (0)
1611 # define sb_end_write(sb) do {} while (0)
1614 static struct thandle *osd_trans_create(const struct lu_env *env,
1615 struct dt_device *d)
1617 struct osd_thread_info *oti = osd_oti_get(env);
1618 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1619 struct osd_thandle *oh;
1624 CERROR("%s: someone try to start transaction under "
1625 "readonly mode, should be disabled.\n",
1626 osd_name(osd_dt_dev(d)));
1628 RETURN(ERR_PTR(-EROFS));
1631 /* on pending IO in this thread should left from prev. request */
1632 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1634 sb_start_write(osd_sb(osd_dt_dev(d)));
1636 OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1638 oh->ot_quota_trans = &oti->oti_quota_trans;
1639 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1644 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1645 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1648 memset(oti->oti_declare_ops, 0,
1649 sizeof(oti->oti_declare_ops));
1650 memset(oti->oti_declare_ops_cred, 0,
1651 sizeof(oti->oti_declare_ops_cred));
1652 memset(oti->oti_declare_ops_used, 0,
1653 sizeof(oti->oti_declare_ops_used));
1655 sb_end_write(osd_sb(osd_dt_dev(d)));
1656 th = ERR_PTR(-ENOMEM);
1661 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1663 struct osd_thread_info *oti = osd_oti_get(env);
1664 struct osd_thandle *oh;
1666 oh = container_of0(th, struct osd_thandle, ot_super);
1667 LASSERT(oh != NULL);
1669 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1670 oti->oti_declare_ops[OSD_OT_CREATE],
1671 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1672 oti->oti_declare_ops_used[OSD_OT_CREATE],
1673 oti->oti_declare_ops[OSD_OT_DESTROY],
1674 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1675 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1676 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1677 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1678 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1679 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1680 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1681 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1682 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1683 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1684 oti->oti_declare_ops[OSD_OT_WRITE],
1685 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1686 oti->oti_declare_ops_used[OSD_OT_WRITE],
1687 oti->oti_declare_ops[OSD_OT_PUNCH],
1688 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1689 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1690 oti->oti_declare_ops[OSD_OT_QUOTA],
1691 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1692 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1693 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1694 oti->oti_declare_ops[OSD_OT_INSERT],
1695 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1696 oti->oti_declare_ops_used[OSD_OT_INSERT],
1697 oti->oti_declare_ops[OSD_OT_DELETE],
1698 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1699 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1700 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1701 oti->oti_declare_ops[OSD_OT_REF_ADD],
1702 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1703 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1704 oti->oti_declare_ops[OSD_OT_REF_DEL],
1705 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1706 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1710 * Concurrency: shouldn't matter.
1712 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1715 struct osd_thread_info *oti = osd_oti_get(env);
1716 struct osd_device *dev = osd_dt_dev(d);
1718 struct osd_thandle *oh;
1723 LASSERT(current->journal_info == NULL);
1725 oh = container_of0(th, struct osd_thandle, ot_super);
1726 LASSERT(oh != NULL);
1727 LASSERT(oh->ot_handle == NULL);
1729 rc = dt_txn_hook_start(env, d, th);
1733 if (unlikely(osd_param_is_not_sane(dev, th))) {
1734 static unsigned long last_printed;
1735 static int last_credits;
1737 /* don't make noise on a tiny testing systems
1738 * actual credits misuse will be caught anyway */
1739 if (last_credits != oh->ot_credits &&
1740 time_after(jiffies, last_printed +
1741 msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1742 osd_transaction_size(dev) > 512) {
1743 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1744 oh->ot_credits, osd_transaction_size(dev));
1745 osd_trans_dump_creds(env, th);
1746 libcfs_debug_dumpstack(NULL);
1747 last_credits = oh->ot_credits;
1748 last_printed = jiffies;
1750 /* XXX Limit the credits to 'max_transaction_buffers', and
1751 * let the underlying filesystem to catch the error if
1752 * we really need so many credits.
1754 * This should be removed when we can calculate the
1755 * credits precisely. */
1756 oh->ot_credits = osd_transaction_size(dev);
1757 } else if (ldiskfs_track_declares_assert != 0) {
1758 /* reserve few credits to prevent an assertion in JBD
1759 * our debugging mechanism will be able to detected
1760 * overuse. this can help to debug single-update
1762 oh->ot_credits += 10;
1763 if (unlikely(osd_param_is_not_sane(dev, th)))
1764 oh->ot_credits = osd_transaction_size(dev);
1768 * XXX temporary stuff. Some abstraction layer should
1771 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1775 LASSERT(oti->oti_txns == 0);
1777 lu_device_get(&d->dd_lu_dev);
1778 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1789 static int osd_seq_exists(const struct lu_env *env,
1790 struct osd_device *osd, u64 seq)
1792 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1793 struct seq_server_site *ss = osd_seq_site(osd);
1797 LASSERT(ss != NULL);
1798 LASSERT(ss->ss_server_fld != NULL);
1800 rc = osd_fld_lookup(env, osd, seq, range);
1803 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1804 osd_name(osd), seq, rc);
1808 RETURN(ss->ss_node_id == range->lsr_index);
1811 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1813 struct dt_txn_commit_cb *dcb;
1814 struct dt_txn_commit_cb *tmp;
1816 /* call per-transaction stop callbacks if any */
1817 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1819 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1820 "commit callback entry: magic=%x name='%s'\n",
1821 dcb->dcb_magic, dcb->dcb_name);
1822 list_del_init(&dcb->dcb_linkage);
1823 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1828 * Concurrency: shouldn't matter.
1830 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1833 int rc = 0, remove_agents = 0;
1834 struct osd_thandle *oh;
1835 struct osd_thread_info *oti = osd_oti_get(env);
1836 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1837 struct osd_device *osd = osd_dt_dev(th->th_dev);
1838 struct qsd_instance *qsd = osd->od_quota_slave;
1839 struct lquota_trans *qtrans;
1842 oh = container_of0(th, struct osd_thandle, ot_super);
1844 remove_agents = oh->ot_remove_agents;
1846 qtrans = oh->ot_quota_trans;
1847 oh->ot_quota_trans = NULL;
1849 if (oh->ot_handle != NULL) {
1851 handle_t *hdl = oh->ot_handle;
1854 * add commit callback
1855 * notice we don't do this in osd_trans_start()
1856 * as underlying transaction can change during truncate
1858 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1861 LASSERT(oti->oti_txns == 1);
1864 rc = dt_txn_hook_stop(env, th);
1866 CERROR("%s: failed in transaction hook: rc = %d\n",
1869 osd_trans_stop_cb(oh, rc);
1870 /* hook functions might modify th_sync */
1871 hdl->h_sync = th->th_sync;
1873 oh->ot_handle = NULL;
1874 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1876 CERROR("%s: failed to stop transaction: rc = %d\n",
1877 osd_name(osd), rc2);
1881 osd_trans_stop_cb(oh, th->th_result);
1885 /* inform the quota slave device that the transaction is stopping */
1886 qsd_op_end(env, qsd, qtrans);
1888 /* as we want IO to journal and data IO be concurrent, we don't block
1889 * awaiting data IO completion in osd_do_bio(), instead we wait here
1890 * once transaction is submitted to the journal. all reqular requests
1891 * don't do direct IO (except read/write), thus this wait_event becomes
1894 * IMPORTANT: we have to wait till any IO submited by the thread is
1895 * completed otherwise iobuf may be corrupted by different request
1897 wait_event(iobuf->dr_wait,
1898 atomic_read(&iobuf->dr_numreqs) == 0);
1899 osd_fini_iobuf(osd, iobuf);
1901 rc = iobuf->dr_error;
1903 if (unlikely(remove_agents != 0))
1904 osd_process_scheduled_agent_removals(env, osd);
1906 /* reset OI cache for safety */
1907 oti->oti_ins_cache_used = 0;
1909 sb_end_write(osd_sb(osd));
1914 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1916 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1919 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1920 LASSERT(&dcb->dcb_func != NULL);
1921 if (dcb->dcb_flags & DCB_TRANS_STOP)
1922 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1924 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1930 * Called just before object is freed. Releases all resources except for
1931 * object itself (that is released by osd_object_free()).
1933 * Concurrency: no concurrent access is possible that late in object
1936 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1938 struct osd_object *obj = osd_obj(l);
1939 struct inode *inode = obj->oo_inode;
1941 LINVRNT(osd_invariant(obj));
1944 * If object is unlinked remove fid->ino mapping from object index.
1947 osd_index_fini(obj);
1948 if (inode != NULL) {
1949 struct qsd_instance *qsd = osd_obj2dev(obj)->od_quota_slave;
1950 qid_t uid = i_uid_read(inode);
1951 qid_t gid = i_gid_read(inode);
1954 obj->oo_inode = NULL;
1957 struct osd_thread_info *info = osd_oti_get(env);
1958 struct lquota_id_info *qi = &info->oti_qi;
1960 /* Release granted quota to master if necessary */
1961 qi->lqi_id.qid_uid = uid;
1962 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1964 qi->lqi_id.qid_uid = gid;
1965 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1967 qi->lqi_id.qid_uid = i_projid_read(inode);
1968 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1974 * Concurrency: ->loo_object_release() is called under site spin-lock.
1976 static void osd_object_release(const struct lu_env *env,
1977 struct lu_object *l)
1979 struct osd_object *o = osd_obj(l);
1980 /* nobody should be releasing a non-destroyed object with nlink=0
1981 * the API allows this, but ldiskfs doesn't like and then report
1982 * this inode as deleted */
1983 if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1988 * Concurrency: shouldn't matter.
1990 static int osd_object_print(const struct lu_env *env, void *cookie,
1991 lu_printer_t p, const struct lu_object *l)
1993 struct osd_object *o = osd_obj(l);
1994 struct iam_descr *d;
1996 if (o->oo_dir != NULL)
1997 d = o->oo_dir->od_container.ic_descr;
2000 return (*p)(env, cookie,
2001 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2003 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2004 o->oo_inode ? o->oo_inode->i_generation : 0,
2005 d ? d->id_ops->id_name : "plain");
2009 * Concurrency: shouldn't matter.
2011 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2012 struct obd_statfs *sfs)
2014 struct osd_device *osd = osd_dt_dev(d);
2015 struct super_block *sb = osd_sb(osd);
2016 struct kstatfs *ksfs;
2020 if (unlikely(osd->od_mnt == NULL))
2021 return -EINPROGRESS;
2023 /* osd_lproc.c call this without env, allocate ksfs for that case */
2024 if (unlikely(env == NULL)) {
2025 OBD_ALLOC_PTR(ksfs);
2029 ksfs = &osd_oti_get(env)->oti_ksfs;
2032 result = sb->s_op->statfs(sb->s_root, ksfs);
2036 statfs_pack(sfs, ksfs);
2037 if (unlikely(sb->s_flags & MS_RDONLY))
2038 sfs->os_state |= OS_STATE_READONLY;
2039 if (ldiskfs_has_feature_extents(sb))
2040 sfs->os_maxbytes = sb->s_maxbytes;
2042 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2045 * Reserve some space so to avoid fragmenting the filesystem too much.
2046 * Fragmentation not only impacts performance, but can also increase
2047 * metadata overhead significantly, causing grant calculation to be
2050 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2052 CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2053 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2054 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2055 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2057 sfs->os_blocks -= reserved;
2058 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2059 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2062 if (unlikely(env == NULL))
2068 * Estimate space needed for file creations. We assume the largest filename
2069 * which is 2^64 - 1, hence a filename of 20 chars.
2070 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2072 #ifdef __LDISKFS_DIR_REC_LEN
2073 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2075 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2079 * Concurrency: doesn't access mutable data.
2081 static void osd_conf_get(const struct lu_env *env,
2082 const struct dt_device *dev,
2083 struct dt_device_param *param)
2085 struct super_block *sb = osd_sb(osd_dt_dev(dev));
2089 * XXX should be taken from not-yet-existing fs abstraction layer.
2091 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2092 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2093 param->ddp_symlink_max = sb->s_blocksize;
2094 param->ddp_mount_type = LDD_MT_LDISKFS;
2095 if (ldiskfs_has_feature_extents(sb))
2096 param->ddp_maxbytes = sb->s_maxbytes;
2098 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2099 /* inode are statically allocated, so per-inode space consumption
2100 * is the space consumed by the directory entry */
2101 param->ddp_inodespace = PER_OBJ_USAGE;
2102 /* EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2103 * = 128MB) which is unlikely to be hit in real life. Report a smaller
2104 * maximum length to not under count the actual number of extents
2105 * needed for writing a file. */
2106 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2107 /* worst-case extent insertion metadata overhead */
2108 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2109 param->ddp_mntopts = 0;
2110 if (test_opt(sb, XATTR_USER))
2111 param->ddp_mntopts |= MNTOPT_USERXATTR;
2112 if (test_opt(sb, POSIX_ACL))
2113 param->ddp_mntopts |= MNTOPT_ACL;
2115 /* LOD might calculate the max stripe count based on max_ea_size,
2116 * so we need take account in the overhead as well,
2117 * xattr_header + magic + xattr_entry_head */
2118 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2119 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2121 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2122 if (ldiskfs_has_feature_ea_inode(sb))
2123 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2127 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2129 /* Preferred RPC size for efficient disk IO. 4MB shows good
2130 * all-around performance for ldiskfs, but use bigalloc chunk size
2131 * by default if larger. */
2132 #if defined(LDISKFS_CLUSTER_SIZE)
2133 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2134 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2137 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2141 * Concurrency: shouldn't matter.
2143 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2147 CDEBUG(D_CACHE, "%s: syncing OSD\n", osd_dt_dev(d)->od_svname);
2149 rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2151 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2158 * Start commit for OSD device.
2160 * An implementation of dt_commit_async method for OSD device.
2161 * Asychronously starts underlayng fs sync and thereby a transaction
2164 * \param env environment
2165 * \param d dt device
2167 * \see dt_device_operations
2169 static int osd_commit_async(const struct lu_env *env,
2170 struct dt_device *d)
2172 struct super_block *s = osd_sb(osd_dt_dev(d));
2175 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2176 RETURN(s->s_op->sync_fs(s, 0));
2179 /* Our own copy of the set readonly functions if present, or NU if not. */
2180 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2181 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2182 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2185 * Concurrency: shouldn't matter.
2187 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2189 struct super_block *sb = osd_sb(osd_dt_dev(d));
2190 struct block_device *dev = sb->s_bdev;
2191 int rc = -EOPNOTSUPP;
2194 if (priv_dev_set_rdonly) {
2195 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2198 CERROR("*** setting %s read-only ***\n",
2199 osd_dt_dev(d)->od_svname);
2201 if (sb->s_op->freeze_fs) {
2202 rc = sb->s_op->freeze_fs(sb);
2207 if (jdev && (jdev != dev)) {
2208 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2210 priv_dev_set_rdonly(jdev);
2212 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2213 priv_dev_set_rdonly(dev);
2215 if (sb->s_op->unfreeze_fs)
2216 sb->s_op->unfreeze_fs(sb);
2221 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2222 osd_dt_dev(d)->od_svname, (long)dev, rc);
2228 * Note: we do not count into QUOTA here.
2229 * If we mount with --data_journal we may need more.
2231 const int osd_dto_credits_noquota[DTO_NR] = {
2234 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2235 * SINGLEDATA_TRANS_BLOCKS(8)
2236 * XXX Note: maybe iam need more, since iam have more level than
2239 [DTO_INDEX_INSERT] = 16,
2242 * just modify a single entry, probably merge few within a block
2244 [DTO_INDEX_DELETE] = 1,
2248 [DTO_INDEX_UPDATE] = 16,
2250 * 4(inode, inode bits, groups, GDT)
2251 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2253 [DTO_OBJECT_CREATE] = 4,
2255 * 4(inode, inode bits, groups, GDT)
2256 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2258 [DTO_OBJECT_DELETE] = 4,
2260 * Attr set credits (inode)
2262 [DTO_ATTR_SET_BASE] = 1,
2264 * Xattr set. The same as xattr of EXT3.
2265 * DATA_TRANS_BLOCKS(14)
2266 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2267 * are also counted in. Do not know why?
2269 [DTO_XATTR_SET] = 14,
2271 * credits for inode change during write.
2273 [DTO_WRITE_BASE] = 3,
2275 * credits for single block write.
2277 [DTO_WRITE_BLOCK] = 14,
2279 * Attr set credits for chown.
2280 * This is extra credits for setattr, and it is null without quota
2282 [DTO_ATTR_SET_CHOWN] = 0
2285 static const struct dt_device_operations osd_dt_ops = {
2286 .dt_root_get = osd_root_get,
2287 .dt_statfs = osd_statfs,
2288 .dt_trans_create = osd_trans_create,
2289 .dt_trans_start = osd_trans_start,
2290 .dt_trans_stop = osd_trans_stop,
2291 .dt_trans_cb_add = osd_trans_cb_add,
2292 .dt_conf_get = osd_conf_get,
2293 .dt_sync = osd_sync,
2295 .dt_commit_async = osd_commit_async,
2298 static void osd_read_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_read_nested(&obj->oo_sem, role);
2309 LASSERT(obj->oo_owner == NULL);
2313 static void osd_write_lock(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(obj->oo_owner != env);
2322 down_write_nested(&obj->oo_sem, role);
2324 LASSERT(obj->oo_owner == NULL);
2325 obj->oo_owner = env;
2329 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2331 struct osd_object *obj = osd_dt_obj(dt);
2332 struct osd_thread_info *oti = osd_oti_get(env);
2334 LINVRNT(osd_invariant(obj));
2336 LASSERT(oti->oti_r_locks > 0);
2338 up_read(&obj->oo_sem);
2341 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2343 struct osd_object *obj = osd_dt_obj(dt);
2344 struct osd_thread_info *oti = osd_oti_get(env);
2346 LINVRNT(osd_invariant(obj));
2348 LASSERT(obj->oo_owner == env);
2349 LASSERT(oti->oti_w_locks > 0);
2351 obj->oo_owner = NULL;
2352 up_write(&obj->oo_sem);
2355 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2357 struct osd_object *obj = osd_dt_obj(dt);
2359 LINVRNT(osd_invariant(obj));
2361 return obj->oo_owner == env;
2364 static struct timespec *osd_inode_time(const struct lu_env *env,
2365 struct inode *inode, __u64 seconds)
2367 struct osd_thread_info *oti = osd_oti_get(env);
2368 struct timespec *t = &oti->oti_time;
2370 t->tv_sec = seconds;
2372 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2376 static void osd_inode_getattr(const struct lu_env *env,
2377 struct inode *inode, struct lu_attr *attr)
2379 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2380 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2381 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2382 LA_BLKSIZE | LA_TYPE;
2384 attr->la_atime = LTIME_S(inode->i_atime);
2385 attr->la_mtime = LTIME_S(inode->i_mtime);
2386 attr->la_ctime = LTIME_S(inode->i_ctime);
2387 attr->la_mode = inode->i_mode;
2388 attr->la_size = i_size_read(inode);
2389 attr->la_blocks = inode->i_blocks;
2390 attr->la_uid = i_uid_read(inode);
2391 attr->la_gid = i_gid_read(inode);
2392 attr->la_projid = i_projid_read(inode);
2393 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2394 attr->la_nlink = inode->i_nlink;
2395 attr->la_rdev = inode->i_rdev;
2396 attr->la_blksize = 1 << inode->i_blkbits;
2397 attr->la_blkbits = inode->i_blkbits;
2399 * Ext4 did not transfer inherit flags from raw inode
2400 * to inode flags, and ext4 internally test raw inode
2401 * @i_flags directly. Instead of patching ext4, we do it here.
2403 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2404 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2407 static int osd_attr_get(const struct lu_env *env,
2408 struct dt_object *dt,
2409 struct lu_attr *attr)
2411 struct osd_object *obj = osd_dt_obj(dt);
2413 if (unlikely(!dt_object_exists(dt)))
2415 if (unlikely(obj->oo_destroyed))
2418 LASSERT(!dt_object_remote(dt));
2419 LINVRNT(osd_invariant(obj));
2421 spin_lock(&obj->oo_guard);
2422 osd_inode_getattr(env, obj->oo_inode, attr);
2423 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2424 attr->la_flags |= LUSTRE_ORPHAN_FL;
2425 spin_unlock(&obj->oo_guard);
2430 static int osd_declare_attr_qid(const struct lu_env *env,
2431 struct osd_object *obj,
2432 struct osd_thandle *oh, long long bspace,
2433 qid_t old_id, qid_t new_id, bool enforce,
2437 struct osd_thread_info *info = osd_oti_get(env);
2438 struct lquota_id_info *qi = &info->oti_qi;
2440 qi->lqi_type = type;
2441 /* inode accounting */
2442 qi->lqi_is_blk = false;
2444 /* one more inode for the new id ... */
2445 qi->lqi_id.qid_uid = new_id;
2447 /* Reserve credits for the new id */
2448 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2449 if (rc == -EDQUOT || rc == -EINPROGRESS)
2454 /* and one less inode for the current id */
2455 qi->lqi_id.qid_uid = old_id;
2457 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2458 if (rc == -EDQUOT || rc == -EINPROGRESS)
2463 /* block accounting */
2464 qi->lqi_is_blk = true;
2466 /* more blocks for the new id ... */
2467 qi->lqi_id.qid_uid = new_id;
2468 qi->lqi_space = bspace;
2470 * Credits for the new uid has been reserved, re-use "obj"
2471 * to save credit reservation.
2473 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2474 if (rc == -EDQUOT || rc == -EINPROGRESS)
2479 /* and finally less blocks for the current uid */
2480 qi->lqi_id.qid_uid = old_id;
2481 qi->lqi_space = -bspace;
2482 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2483 if (rc == -EDQUOT || rc == -EINPROGRESS)
2489 static int osd_declare_attr_set(const struct lu_env *env,
2490 struct dt_object *dt,
2491 const struct lu_attr *attr,
2492 struct thandle *handle)
2494 struct osd_thandle *oh;
2495 struct osd_object *obj;
2503 LASSERT(dt != NULL);
2504 LASSERT(handle != NULL);
2506 obj = osd_dt_obj(dt);
2507 LASSERT(osd_invariant(obj));
2509 oh = container_of0(handle, struct osd_thandle, ot_super);
2510 LASSERT(oh->ot_handle == NULL);
2512 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2513 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2515 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2516 osd_dto_credits_noquota[DTO_XATTR_SET]);
2518 if (attr == NULL || obj->oo_inode == NULL)
2521 bspace = obj->oo_inode->i_blocks;
2522 bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2523 bspace = toqb(bspace);
2525 /* Changing ownership is always preformed by super user, it should not
2528 * We still need to call the osd_declare_qid() to calculate the journal
2529 * credits for updating quota accounting files and to trigger quota
2530 * space adjustment once the operation is completed.*/
2531 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2533 uid = i_uid_read(obj->oo_inode);
2534 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2535 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2536 attr->la_uid, enforce, USRQUOTA);
2540 gid = i_gid_read(obj->oo_inode);
2541 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2542 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2543 i_gid_read(obj->oo_inode),
2544 attr->la_gid, enforce, GRPQUOTA);
2549 #ifdef HAVE_PROJECT_QUOTA
2550 if (attr->la_valid & LA_PROJID) {
2551 __u32 projid = i_projid_read(obj->oo_inode);
2552 enforce = (attr->la_valid & LA_PROJID) &&
2553 (attr->la_projid != projid);
2554 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2555 (qid_t)projid, (qid_t)attr->la_projid,
2564 static int osd_inode_setattr(const struct lu_env *env,
2565 struct inode *inode, const struct lu_attr *attr)
2567 __u64 bits = attr->la_valid;
2569 /* Only allow set size for regular file */
2570 if (!S_ISREG(inode->i_mode))
2571 bits &= ~(LA_SIZE | LA_BLOCKS);
2576 if (bits & LA_ATIME)
2577 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2578 if (bits & LA_CTIME)
2579 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2580 if (bits & LA_MTIME)
2581 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2582 if (bits & LA_SIZE) {
2583 spin_lock(&inode->i_lock);
2584 LDISKFS_I(inode)->i_disksize = attr->la_size;
2585 i_size_write(inode, attr->la_size);
2586 spin_unlock(&inode->i_lock);
2589 /* OSD should not change "i_blocks" which is used by quota.
2590 * "i_blocks" should be changed by ldiskfs only. */
2592 inode->i_mode = (inode->i_mode & S_IFMT) |
2593 (attr->la_mode & ~S_IFMT);
2595 i_uid_write(inode, attr->la_uid);
2597 i_gid_write(inode, attr->la_gid);
2598 if (bits & LA_PROJID)
2599 i_projid_write(inode, attr->la_projid);
2600 if (bits & LA_NLINK)
2601 set_nlink(inode, attr->la_nlink);
2603 inode->i_rdev = attr->la_rdev;
2605 if (bits & LA_FLAGS) {
2606 /* always keep S_NOCMTIME */
2607 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2610 * Ext4 did not transfer inherit flags from
2611 * @inode->i_flags to raw inode i_flags when writing
2612 * flags, we do it explictly here.
2614 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2615 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2617 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2622 #ifdef HAVE_PROJECT_QUOTA
2623 static int osd_transfer_project(struct inode *inode, __u32 projid)
2625 struct super_block *sb = inode->i_sb;
2626 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2629 struct ldiskfs_iloc iloc;
2630 struct ldiskfs_inode *raw_inode;
2631 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
2633 if (!ldiskfs_has_feature_project(sb)) {
2634 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
2635 != LDISKFS_DEF_PROJID);
2636 if (projid != LDISKFS_DEF_PROJID)
2642 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
2645 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
2646 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
2649 err = ldiskfs_get_inode_loc(inode, &iloc);
2653 raw_inode = ldiskfs_raw_inode(&iloc);
2654 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
2661 dquot_initialize(inode);
2662 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
2663 if (transfer_to[PRJQUOTA]) {
2664 err = __dquot_transfer(inode, transfer_to);
2665 dqput(transfer_to[PRJQUOTA]);
2674 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2678 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2679 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2682 ll_vfs_dq_init(inode);
2684 if (attr->la_valid & LA_UID)
2685 iattr.ia_valid |= ATTR_UID;
2686 if (attr->la_valid & LA_GID)
2687 iattr.ia_valid |= ATTR_GID;
2688 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2689 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2691 rc = ll_vfs_dq_transfer(inode, &iattr);
2693 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2694 "enforcement enabled on the ldiskfs "
2695 "filesystem?\n", inode->i_sb->s_id, rc);
2700 /* Handle project id transfer here properly */
2701 if (attr->la_valid & LA_PROJID &&
2702 attr->la_projid != i_projid_read(inode)) {
2703 #ifdef HAVE_PROJECT_QUOTA
2704 rc = osd_transfer_project(inode, attr->la_projid);
2709 CERROR("%s: quota transfer failed: rc = %d. Is project "
2710 "enforcement enabled on the ldiskfs "
2711 "filesystem?\n", inode->i_sb->s_id, rc);
2718 static int osd_attr_set(const struct lu_env *env,
2719 struct dt_object *dt,
2720 const struct lu_attr *attr,
2721 struct thandle *handle)
2723 struct osd_object *obj = osd_dt_obj(dt);
2724 struct inode *inode;
2727 if (!dt_object_exists(dt))
2730 LASSERT(handle != NULL);
2731 LASSERT(!dt_object_remote(dt));
2732 LASSERT(osd_invariant(obj));
2734 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2736 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2737 struct osd_thread_info *oti = osd_oti_get(env);
2738 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2739 struct lu_fid *fid1 = &oti->oti_fid;
2740 struct osd_inode_id *id = &oti->oti_id;
2741 struct iam_path_descr *ipd;
2742 struct iam_container *bag;
2743 struct osd_thandle *oh;
2746 fid_cpu_to_be(fid1, fid0);
2747 memset(id, 1, sizeof(*id));
2748 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2749 fid0)->oi_dir.od_container;
2750 ipd = osd_idx_ipd_get(env, bag);
2751 if (unlikely(ipd == NULL))
2754 oh = container_of0(handle, struct osd_thandle, ot_super);
2755 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2756 (const struct iam_rec *)id, ipd);
2757 osd_ipd_put(env, bag, ipd);
2758 return(rc > 0 ? 0 : rc);
2761 inode = obj->oo_inode;
2763 rc = osd_quota_transfer(inode, attr);
2767 spin_lock(&obj->oo_guard);
2768 rc = osd_inode_setattr(env, inode, attr);
2769 spin_unlock(&obj->oo_guard);
2773 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2775 if (!(attr->la_valid & LA_FLAGS))
2778 /* Let's check if there are extra flags need to be set into LMA */
2779 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2780 struct osd_thread_info *info = osd_oti_get(env);
2781 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2783 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2784 &info->oti_ost_attrs);
2788 lma->lma_incompat |=
2789 lustre_to_lma_flags(attr->la_flags);
2790 lustre_lma_swab(lma);
2791 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2792 lma, sizeof(*lma), XATTR_REPLACE);
2794 struct osd_device *osd = osd_obj2dev(obj);
2796 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2797 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2798 lma->lma_incompat, rc);
2801 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2803 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2806 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2811 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2812 struct osd_object *obj,
2813 const char *name, const int namelen)
2815 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2818 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2819 umode_t mode, struct dt_allocation_hint *hint,
2823 struct osd_device *osd = osd_obj2dev(obj);
2824 struct osd_thandle *oth;
2825 struct dt_object *parent = NULL;
2826 struct inode *inode;
2828 LINVRNT(osd_invariant(obj));
2829 LASSERT(obj->oo_inode == NULL);
2830 LASSERT(obj->oo_hl_head == NULL);
2832 if (S_ISDIR(mode) && ldiskfs_pdo) {
2833 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2834 if (obj->oo_hl_head == NULL)
2838 oth = container_of(th, struct osd_thandle, ot_super);
2839 LASSERT(oth->ot_handle->h_transaction != NULL);
2841 if (hint != NULL && hint->dah_parent != NULL &&
2842 !dt_object_remote(hint->dah_parent))
2843 parent = hint->dah_parent;
2845 inode = ldiskfs_create_inode(oth->ot_handle,
2846 parent ? osd_dt_obj(parent)->oo_inode :
2847 osd_sb(osd)->s_root->d_inode,
2849 if (!IS_ERR(inode)) {
2850 /* Do not update file c/mtime in ldiskfs. */
2851 inode->i_flags |= S_NOCMTIME;
2853 /* For new created object, it must be consistent,
2854 * and it is unnecessary to scrub against it. */
2855 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2857 obj->oo_inode = inode;
2860 if (obj->oo_hl_head != NULL) {
2861 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2862 obj->oo_hl_head = NULL;
2864 result = PTR_ERR(inode);
2866 LINVRNT(osd_invariant(obj));
2874 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2875 struct lu_attr *attr,
2876 struct dt_allocation_hint *hint,
2877 struct dt_object_format *dof,
2881 struct osd_thandle *oth;
2882 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2884 LASSERT(S_ISDIR(attr->la_mode));
2886 oth = container_of(th, struct osd_thandle, ot_super);
2887 LASSERT(oth->ot_handle->h_transaction != NULL);
2888 result = osd_mkfile(info, obj, mode, hint, th);
2893 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2894 struct lu_attr *attr,
2895 struct dt_allocation_hint *hint,
2896 struct dt_object_format *dof,
2900 struct osd_thandle *oth;
2901 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2903 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2905 LASSERT(S_ISREG(attr->la_mode));
2907 oth = container_of(th, struct osd_thandle, ot_super);
2908 LASSERT(oth->ot_handle->h_transaction != NULL);
2910 result = osd_mkfile(info, obj, mode, hint, th);
2912 LASSERT(obj->oo_inode != NULL);
2913 if (feat->dif_flags & DT_IND_VARKEY)
2914 result = iam_lvar_create(obj->oo_inode,
2915 feat->dif_keysize_max,
2917 feat->dif_recsize_max,
2920 result = iam_lfix_create(obj->oo_inode,
2921 feat->dif_keysize_max,
2923 feat->dif_recsize_max,
2930 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2931 struct lu_attr *attr,
2932 struct dt_allocation_hint *hint,
2933 struct dt_object_format *dof,
2936 LASSERT(S_ISREG(attr->la_mode));
2937 return osd_mkfile(info, obj, (attr->la_mode &
2938 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2941 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2942 struct lu_attr *attr,
2943 struct dt_allocation_hint *hint,
2944 struct dt_object_format *dof,
2947 LASSERT(S_ISLNK(attr->la_mode));
2948 return osd_mkfile(info, obj, (attr->la_mode &
2949 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2952 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2953 struct lu_attr *attr,
2954 struct dt_allocation_hint *hint,
2955 struct dt_object_format *dof,
2958 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2961 LINVRNT(osd_invariant(obj));
2962 LASSERT(obj->oo_inode == NULL);
2963 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2964 S_ISFIFO(mode) || S_ISSOCK(mode));
2966 result = osd_mkfile(info, obj, mode, hint, th);
2968 LASSERT(obj->oo_inode != NULL);
2970 * This inode should be marked dirty for i_rdev. Currently
2971 * that is done in the osd_attr_init().
2973 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2976 LINVRNT(osd_invariant(obj));
2980 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2982 struct dt_allocation_hint *hint,
2983 struct dt_object_format *dof,
2986 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2988 osd_obj_type_f result;
3004 result = osd_mk_index;
3015 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3016 struct dt_object *parent, struct dt_object *child,
3021 ah->dah_parent = parent;
3022 ah->dah_mode = child_mode;
3024 if (parent != NULL && !dt_object_remote(parent)) {
3025 /* will help to find FID->ino at dt_insert("..") */
3026 struct osd_object *pobj = osd_dt_obj(parent);
3027 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3031 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3032 struct lu_attr *attr, struct dt_object_format *dof)
3034 struct inode *inode = obj->oo_inode;
3035 __u64 valid = attr->la_valid;
3038 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3040 if (dof->dof_type != DFT_NODE)
3041 attr->la_valid &= ~LA_RDEV;
3042 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
3043 attr->la_valid &= ~LA_ATIME;
3044 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
3045 attr->la_valid &= ~LA_CTIME;
3046 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
3047 attr->la_valid &= ~LA_MTIME;
3049 result = osd_quota_transfer(inode, attr);
3053 if (attr->la_valid != 0) {
3054 result = osd_inode_setattr(info->oti_env, inode, attr);
3056 * The osd_inode_setattr() should always succeed here. The
3057 * only error that could be returned is EDQUOT when we are
3058 * trying to change the UID or GID of the inode. However, this
3059 * should not happen since quota enforcement is no longer
3060 * enabled on ldiskfs (lquota takes care of it).
3062 LASSERTF(result == 0, "%d\n", result);
3063 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3066 attr->la_valid = valid;
3070 * Helper function for osd_create()
3072 * \retval 0, on success
3074 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3075 struct lu_attr *attr, struct dt_allocation_hint *hint,
3076 struct dt_object_format *dof, struct thandle *th)
3081 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3083 /* we drop umask so that permissions we pass are not affected */
3084 umask = current->fs->umask;
3085 current->fs->umask = 0;
3087 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3089 if (likely(obj->oo_inode != NULL)) {
3090 LASSERT(obj->oo_inode->i_state & I_NEW);
3092 /* Unlock the inode before attr initialization to avoid
3093 * unnecessary dqget operations. LU-6378 */
3094 unlock_new_inode(obj->oo_inode);
3097 if (likely(result == 0)) {
3098 osd_attr_init(info, obj, attr, dof);
3099 osd_object_init0(obj);
3102 /* restore previous umask value */
3103 current->fs->umask = umask;
3105 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3111 * Helper function for osd_create()
3113 * \retval 0, on success
3115 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3116 const struct lu_fid *fid, struct thandle *th)
3118 struct osd_thread_info *info = osd_oti_get(env);
3119 struct osd_inode_id *id = &info->oti_id;
3120 struct osd_device *osd = osd_obj2dev(obj);
3121 struct osd_thandle *oh;
3124 LASSERT(obj->oo_inode != NULL);
3126 oh = container_of0(th, struct osd_thandle, ot_super);
3127 LASSERT(oh->ot_handle);
3128 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3130 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3131 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3132 OI_CHECK_FLD, NULL);
3133 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3138 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3139 u64 seq, struct lu_seq_range *range)
3141 struct seq_server_site *ss = osd_seq_site(osd);
3143 if (fid_seq_is_idif(seq)) {
3144 fld_range_set_ost(range);
3145 range->lsr_index = idif_ost_idx(seq);
3149 if (!fid_seq_in_fldb(seq)) {
3150 fld_range_set_mdt(range);
3152 /* FIXME: If ss is NULL, it suppose not get lsr_index
3154 range->lsr_index = ss->ss_node_id;
3158 LASSERT(ss != NULL);
3159 fld_range_set_any(range);
3160 /* OSD will only do local fld lookup */
3161 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3164 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3165 struct lu_attr *attr,
3166 struct dt_allocation_hint *hint,
3167 struct dt_object_format *dof,
3168 struct thandle *handle)
3170 struct osd_thandle *oh;
3174 LASSERT(handle != NULL);
3176 oh = container_of0(handle, struct osd_thandle, ot_super);
3177 LASSERT(oh->ot_handle == NULL);
3179 /* EA object consumes more credits than regular object: osd_mk_index
3180 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3181 * leaf_node, could involves the block, block bitmap, groups, GDT
3182 * change for each block, so add 4 * 2 credits in that case. */
3183 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3184 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3185 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3186 /* Reuse idle OI block may cause additional one OI block
3188 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3189 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3194 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3195 attr->la_projid, 1, oh, osd_dt_obj(dt),
3196 NULL, OSD_QID_INODE);
3200 /* will help to find FID->ino mapping at dt_insert() */
3201 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3208 * Called to destroy on-disk representation of the object
3210 * Concurrency: must be locked
3212 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3215 struct osd_object *obj = osd_dt_obj(dt);
3216 struct inode *inode = obj->oo_inode;
3217 struct osd_thandle *oh;
3224 oh = container_of0(th, struct osd_thandle, ot_super);
3225 LASSERT(oh->ot_handle == NULL);
3227 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3228 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3229 /* Recycle idle OI leaf may cause additional three OI blocks
3231 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3232 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3233 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3234 /* one less inode */
3235 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3236 i_projid_read(inode), -1, oh, obj, NULL,
3240 /* data to be truncated */
3241 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3242 i_projid_read(inode), 0, oh, obj, NULL,
3247 /* will help to find FID->ino when this object is being
3248 * added to PENDING/ */
3249 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3254 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3257 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3258 struct osd_object *obj = osd_dt_obj(dt);
3259 struct inode *inode = obj->oo_inode;
3260 struct osd_device *osd = osd_obj2dev(obj);
3261 struct osd_thandle *oh;
3265 oh = container_of0(th, struct osd_thandle, ot_super);
3266 LASSERT(oh->ot_handle);
3268 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3270 if (unlikely(fid_is_acct(fid)))
3273 if (S_ISDIR(inode->i_mode)) {
3274 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3275 inode->i_nlink == 2);
3276 /* it will check/delete the inode from remote parent,
3277 * how to optimize it? unlink performance impaction XXX */
3278 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3280 CERROR("%s: delete inode "DFID": rc = %d\n",
3281 osd_name(osd), PFID(fid), result);
3283 spin_lock(&obj->oo_guard);
3285 spin_unlock(&obj->oo_guard);
3286 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3289 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3291 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3293 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3294 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3295 oh->ot_handle, OI_CHECK_FLD);
3297 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3298 /* XXX: add to ext3 orphan list */
3299 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3301 /* not needed in the cache anymore */
3302 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3303 obj->oo_destroyed = 1;
3309 * Put the fid into lustre_mdt_attrs, and then place the structure
3310 * inode's ea. This fid should not be altered during the life time
3313 * \retval +ve, on success
3314 * \retval -ve, on error
3316 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3318 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3319 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3321 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3322 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3326 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3329 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3332 lustre_loa_init(loa, fid, compat, incompat);
3333 lustre_loa_swab(loa, false);
3335 /* For the OST device with 256 bytes inode size by default,
3336 * the PFID EA will be stored together with LMA EA to avoid
3337 * performance trouble. Otherwise the PFID EA can be stored
3338 * independently. LU-8998 */
3339 if ((compat & LMAC_FID_ON_OST) &&
3340 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3341 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3342 sizeof(*loa), XATTR_CREATE);
3344 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3345 sizeof(*lma), XATTR_CREATE);
3346 /* LMA may already exist, but we need to check that all the
3347 * desired compat/incompat flags have been added. */
3348 if (unlikely(rc == -EEXIST)) {
3349 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3350 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3354 if (rc < sizeof(*lma))
3357 lustre_loa_swab(loa, true);
3358 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3359 ((compat == 0 && incompat == 0) ||
3360 (!(~lma->lma_compat & compat) &&
3361 !(~lma->lma_incompat & incompat))))
3364 lma->lma_self_fid = *fid;
3365 lma->lma_compat |= compat;
3366 lma->lma_incompat |= incompat;
3367 if (rc == sizeof(*lma)) {
3368 lustre_lma_swab(lma);
3369 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3370 sizeof(*lma), XATTR_REPLACE);
3372 lustre_loa_swab(loa, false);
3373 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3374 sizeof(*loa), XATTR_REPLACE);
3382 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3383 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3384 * To have compatilibility with 1.8 ldiskfs driver we need to have
3385 * magic number at start of fid data.
3386 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3389 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3390 const struct lu_fid *fid)
3392 if (!fid_is_namespace_visible(fid) ||
3393 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3394 param->edp_magic = 0;
3398 param->edp_magic = LDISKFS_LUFID_MAGIC;
3399 param->edp_len = sizeof(struct lu_fid) + 1;
3400 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3404 * Try to read the fid from inode ea into dt_rec.
3406 * \param fid object fid.
3408 * \retval 0 on success
3410 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3411 __u32 ino, struct lu_fid *fid,
3412 struct osd_inode_id *id)
3414 struct osd_thread_info *info = osd_oti_get(env);
3415 struct inode *inode;
3418 osd_id_gen(id, ino, OSD_OII_NOGEN);
3419 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3421 RETURN(PTR_ERR(inode));
3427 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3429 struct inode *parent_dir,
3430 const struct lu_fid *dot_fid,
3431 const struct lu_fid *dot_dot_fid,
3432 struct osd_thandle *oth)
3434 struct ldiskfs_dentry_param *dot_ldp;
3435 struct ldiskfs_dentry_param *dot_dot_ldp;
3436 __u32 saved_nlink = dir->i_nlink;
3439 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3440 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3442 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3443 dot_ldp->edp_magic = 0;
3445 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3446 dir, dot_ldp, dot_dot_ldp);
3447 /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3448 * the subseqent ref_add() will increase the dir->i_nlink
3449 * as 3. That is incorrect for new created directory.
3451 * It looks like hack, because we want to make the OSD API
3452 * to be order-independent for new created directory object
3453 * between dt_insert(..) and ref_add() operations.
3455 * Here, we only restore the in-RAM dir-inode's nlink attr,
3456 * becuase if the nlink attr is not 2, then there will be
3457 * ref_add() called following the dt_insert(..), such call
3458 * will make both the in-RAM and on-disk dir-inode's nlink
3459 * attr to be set as 2. LU-7447 */
3460 set_nlink(dir, saved_nlink);
3465 * Create an local agent inode for remote entry
3467 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3468 struct osd_device *osd,
3469 struct osd_object *pobj,
3470 const struct lu_fid *fid,
3474 struct osd_thread_info *info = osd_oti_get(env);
3475 struct inode *local;
3476 struct osd_thandle *oh;
3481 oh = container_of(th, struct osd_thandle, ot_super);
3482 LASSERT(oh->ot_handle->h_transaction != NULL);
3484 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3485 if (IS_ERR(local)) {
3486 CERROR("%s: create local error %d\n", osd_name(osd),
3487 (int)PTR_ERR(local));
3491 /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3492 * correct gid on remote file, not agent here */
3493 local->i_gid = current_fsgid();
3494 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3495 unlock_new_inode(local);
3497 /* Agent inode should not have project ID */
3498 #ifdef HAVE_PROJECT_QUOTA
3499 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3500 i_projid_read(pobj->oo_inode) != 0) {
3501 rc = osd_transfer_project(local, 0);
3503 CERROR("%s: quota transfer failed: rc = %d. Is project "
3504 "quota enforcement enabled on the ldiskfs "
3505 "filesystem?\n", local->i_sb->s_id, rc);
3506 RETURN(ERR_PTR(rc));
3510 /* Set special LMA flag for local agent inode */
3511 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3513 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3514 osd_name(osd), PFID(fid), rc);
3515 RETURN(ERR_PTR(rc));
3521 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3522 lu_object_fid(&pobj->oo_dt.do_lu),
3525 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3526 osd_name(osd), PFID(fid), rc);
3527 RETURN(ERR_PTR(rc));
3534 * when direntry is deleted, we have to take care of possible agent inode
3535 * referenced by that. unfortunately we can't do this at that point:
3536 * iget() within a running transaction leads to deadlock and we better do
3537 * not call that every delete declaration to save performance. so we put
3538 * a potention agent inode on a list and process that once the transaction
3539 * is over. Notice it's not any worse in terms of real orphans as regular
3540 * object destroy doesn't put inodes on the on-disk orphan list. this should
3541 * be addressed separately
3543 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3544 struct osd_thandle *oh,
3547 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3548 struct osd_obj_orphan *oor;
3555 oor->oor_env = (struct lu_env *)env;
3556 spin_lock(&osd->od_osfs_lock);
3557 list_add(&oor->oor_list, &osd->od_orphan_list);
3558 spin_unlock(&osd->od_osfs_lock);
3560 oh->ot_remove_agents = 1;
3566 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3567 struct osd_device *osd)
3569 struct osd_thread_info *info = osd_oti_get(env);
3570 struct osd_obj_orphan *oor, *tmp;
3571 struct osd_inode_id id;
3572 struct list_head list;
3573 struct inode *inode;
3578 INIT_LIST_HEAD(&list);
3580 spin_lock(&osd->od_osfs_lock);
3581 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3582 if (oor->oor_env == env) {
3583 list_del(&oor->oor_list);
3584 list_add(&oor->oor_list, &list);
3587 spin_unlock(&osd->od_osfs_lock);
3589 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3593 list_del(&oor->oor_list);
3596 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3597 inode = osd_iget_fid(info, osd, &id, &fid);
3601 if (!osd_remote_fid(env, osd, &fid)) {
3606 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3608 mark_inode_dirty(inode);
3609 ldiskfs_journal_stop(jh);
3617 * OSD layer object create function for OST objects (b=11826).
3619 * The FID is inserted into inode xattr here.
3621 * \retval 0, on success
3622 * \retval -ve, on error
3624 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3625 struct lu_attr *attr, struct dt_allocation_hint *hint,
3626 struct dt_object_format *dof, struct thandle *th)
3628 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3629 struct osd_object *obj = osd_dt_obj(dt);
3630 struct osd_thread_info *info = osd_oti_get(env);
3631 int result, on_ost = 0;
3635 if (dt_object_exists(dt))
3638 LINVRNT(osd_invariant(obj));
3639 LASSERT(!dt_object_remote(dt));
3640 LASSERT(osd_is_write_locked(env, obj));
3641 LASSERT(th != NULL);
3643 if (unlikely(fid_is_acct(fid)))
3644 /* Quota files can't be created from the kernel any more,
3645 * 'tune2fs -O quota' will take care of creating them */
3648 result = __osd_create(info, obj, attr, hint, dof, th);
3650 if (fid_is_idif(fid) &&
3651 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3652 struct lu_fid *tfid = &info->oti_fid;
3653 struct ost_id *oi = &info->oti_ostid;
3655 fid_to_ostid(fid, oi);
3656 ostid_to_fid(tfid, oi, 0);
3658 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3659 LMAC_FID_ON_OST, 0);
3661 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3663 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3664 on_ost ? LMAC_FID_ON_OST : 0,
3667 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3668 obj->oo_dt.do_body_ops = &osd_body_ops;
3672 result = __osd_oi_insert(env, obj, fid, th);
3674 /* a small optimization - dt_insert() isn't usually applied
3675 * to OST objects, so we don't need to cache OI mapping for
3677 if (result == 0 && on_ost == 0) {
3678 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3679 result = osd_idc_find_and_init(env, osd, obj);
3680 LASSERT(result == 0);
3683 LASSERT(ergo(result == 0,
3684 dt_object_exists(dt) && !dt_object_remote(dt)));
3685 LINVRNT(osd_invariant(obj));
3689 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3690 struct thandle *handle)
3692 struct osd_thandle *oh;
3694 /* it's possible that object doesn't exist yet */
3695 LASSERT(handle != NULL);
3697 oh = container_of0(handle, struct osd_thandle, ot_super);
3698 LASSERT(oh->ot_handle == NULL);
3700 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3701 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3703 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
3709 * Concurrency: @dt is write locked.
3711 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3714 struct osd_object *obj = osd_dt_obj(dt);
3715 struct inode *inode = obj->oo_inode;
3716 struct osd_thandle *oh;
3719 if (!dt_object_exists(dt) || obj->oo_destroyed)
3722 LINVRNT(osd_invariant(obj));
3723 LASSERT(!dt_object_remote(dt));
3724 LASSERT(osd_is_write_locked(env, obj));
3725 LASSERT(th != NULL);
3727 oh = container_of0(th, struct osd_thandle, ot_super);
3728 LASSERT(oh->ot_handle != NULL);
3730 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3732 CDEBUG(D_INODE, DFID" increase nlink %d\n",
3733 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3735 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3736 * (65000) subdirectories by storing "1" in i_nlink if the link count
3737 * would otherwise overflow. Directory tranversal tools understand
3738 * that (st_nlink == 1) indicates that the filesystem dose not track
3739 * hard links count on the directory, and will not abort subdirectory
3740 * scanning early once (st_nlink - 2) subdirs have been found.
3742 * This also has to properly handle the case of inodes with nlink == 0
3743 * in case they are being linked into the PENDING directory
3745 spin_lock(&obj->oo_guard);
3746 if (unlikely(inode->i_nlink == 0))
3747 /* inc_nlink from 0 may cause WARN_ON */
3748 set_nlink(inode, 1);
3750 ldiskfs_inc_count(oh->ot_handle, inode);
3751 if (!S_ISDIR(inode->i_mode))
3752 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3754 spin_unlock(&obj->oo_guard);
3756 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3757 LINVRNT(osd_invariant(obj));
3759 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3764 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
3765 struct thandle *handle)
3767 struct osd_thandle *oh;
3769 if (!dt_object_exists(dt))
3772 LASSERT(!dt_object_remote(dt));
3773 LASSERT(handle != NULL);
3775 oh = container_of0(handle, struct osd_thandle, ot_super);
3776 LASSERT(oh->ot_handle == NULL);
3778 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3779 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3785 * Concurrency: @dt is write locked.
3787 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
3790 struct osd_object *obj = osd_dt_obj(dt);
3791 struct inode *inode = obj->oo_inode;
3792 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3793 struct osd_thandle *oh;
3795 if (!dt_object_exists(dt))
3798 LINVRNT(osd_invariant(obj));
3799 LASSERT(!dt_object_remote(dt));
3800 LASSERT(osd_is_write_locked(env, obj));
3801 LASSERT(th != NULL);
3803 oh = container_of0(th, struct osd_thandle, ot_super);
3804 LASSERT(oh->ot_handle != NULL);
3806 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3808 spin_lock(&obj->oo_guard);
3809 /* That can be result of upgrade from old Lustre version and
3810 * applied only to local files. Just skip this ref_del call.
3811 * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3812 if (inode->i_nlink == 0) {
3813 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3814 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3815 ", maybe an upgraded file? (LU-3915)\n",
3816 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3817 spin_unlock(&obj->oo_guard);
3821 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3822 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3824 ldiskfs_dec_count(oh->ot_handle, inode);
3825 spin_unlock(&obj->oo_guard);
3827 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3828 LINVRNT(osd_invariant(obj));
3830 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3836 * Concurrency: @dt is read locked.
3838 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3839 struct lu_buf *buf, const char *name)
3841 struct osd_object *obj = osd_dt_obj(dt);
3842 struct inode *inode = obj->oo_inode;
3843 struct osd_thread_info *info = osd_oti_get(env);
3844 struct dentry *dentry = &info->oti_obj_dentry;
3845 bool cache_xattr = false;
3850 /* version get is not real XATTR but uses xattr API */
3851 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3852 dt_obj_version_t *ver = buf->lb_buf;
3854 /* for version we are just using xattr API but change inode
3856 if (buf->lb_len == 0)
3857 return sizeof(dt_obj_version_t);
3859 if (buf->lb_len < sizeof(dt_obj_version_t))
3862 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
3863 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3865 *ver = LDISKFS_I(inode)->i_fs_version;
3867 return sizeof(dt_obj_version_t);
3870 if (!dt_object_exists(dt))
3873 LASSERT(!dt_object_remote(dt));
3874 LASSERT(inode->i_op != NULL);
3875 #ifdef HAVE_IOP_XATTR
3876 LASSERT(inode->i_op->getxattr != NULL);
3879 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
3880 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
3884 rc = osd_oxc_get(obj, name, buf);
3889 rc = __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3890 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
3891 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3892 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3893 struct filter_fid *ff;
3894 struct ost_layout *ol;
3896 if (!osd_dev(dt->do_lu.lo_dev)->od_is_ost)
3899 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
3903 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
3909 if (buf->lb_len == 0 || !buf->lb_buf)
3912 if (buf->lb_len < rc)
3916 ol = &ff->ff_layout;
3917 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
3918 PFID_STRIPE_IDX_BITS);
3919 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
3920 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
3921 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
3922 if (lma->lma_compat & LMAC_COMP_INFO) {
3923 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
3924 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
3925 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
3927 ol->ol_comp_start = 0;
3928 ol->ol_comp_end = 0;
3935 if (rc == -ENOENT || rc == -ENODATA)
3936 osd_oxc_add(obj, name, NULL, 0);
3937 else if (rc > 0 && buf->lb_buf != NULL)
3938 osd_oxc_add(obj, name, buf->lb_buf, rc);
3944 static int osd_declare_xattr_set(const struct lu_env *env,
3945 struct dt_object *dt,
3946 const struct lu_buf *buf, const char *name,
3947 int fl, struct thandle *handle)
3949 struct osd_thandle *oh;
3951 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3953 LASSERT(handle != NULL);
3955 oh = container_of0(handle, struct osd_thandle, ot_super);
3956 LASSERT(oh->ot_handle == NULL);
3958 if (strcmp(name, XATTR_NAME_LMA) == 0) {
3959 /* For non-upgrading case, the LMA is set first and
3960 * usually fit inode. But for upgrade case, the LMA
3961 * may be in another separated EA block. */
3962 if (dt_object_exists(dt)) {
3963 if (fl == LU_XATTR_REPLACE)
3968 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3970 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
3971 /* We may need to delete the old PFID EA. */
3972 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
3973 if (fl == LU_XATTR_REPLACE)
3980 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
3985 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3986 /* learn xattr size from osd_xattr_get if
3987 attribute has not been read yet */
3988 buflen = __osd_xattr_get(
3989 osd_dt_obj(dt)->oo_inode,
3990 &osd_oti_get(env)->oti_obj_dentry,
3995 buflen = buf->lb_len;
3998 if (buflen > sb->s_blocksize) {
3999 credits += osd_calc_bkmap_credits(
4001 (buflen + sb->s_blocksize - 1) >>
4002 sb->s_blocksize_bits);
4006 * xattr set may involve inode quota change, reserve credits for
4007 * dquot_initialize()
4009 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4012 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4018 * Concurrency: @dt is write locked.
4020 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4021 const struct lu_buf *buf, const char *name, int fl,
4022 struct thandle *handle)
4024 struct osd_object *obj = osd_dt_obj(dt);
4025 struct inode *inode = obj->oo_inode;
4026 struct osd_thread_info *info = osd_oti_get(env);
4027 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4028 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4037 /* version set is not real XATTR */
4038 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4039 dt_obj_version_t *version = buf->lb_buf;
4041 /* for version we are just using xattr API but change inode
4043 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4045 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4046 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4048 LDISKFS_I(inode)->i_fs_version = *version;
4049 /* Version is set after all inode operations are finished,
4050 * so we should mark it dirty here */
4051 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4056 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4057 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4060 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4061 if (fl & LU_XATTR_REPLACE)
4062 fs_flags |= XATTR_REPLACE;
4064 if (fl & LU_XATTR_CREATE)
4065 fs_flags |= XATTR_CREATE;
4067 /* For the OST device with 256 bytes inode size by default,
4068 * the PFID EA will be stored together with LMA EA to avoid
4069 * performance trouble. Otherwise the PFID EA can be stored
4070 * independently. LU-8998 */
4071 if (strcmp(name, XATTR_NAME_FID) == 0 &&
4072 LDISKFS_INODE_SIZE(inode->i_sb) <= 256) {
4073 struct dentry *dentry = &info->oti_obj_dentry;
4074 struct filter_fid *ff;
4075 struct ost_layout *ol;
4078 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4081 ol = &ff->ff_layout;
4082 /* Old client does not send stripe information, store
4083 * the PFID EA on disk directly. */
4084 if (buf->lb_len == sizeof(struct lu_fid) ||
4085 ol->ol_stripe_size == 0) {
4086 len = sizeof(struct lu_fid);
4090 if (buf->lb_len != sizeof(*ff))
4093 rc = osd_get_lma(info, inode, dentry, loa);
4094 if (unlikely(rc == -ENODATA)) {
4095 /* Usually for upgarding from old device */
4096 lustre_loa_init(loa, lu_object_fid(&dt->do_lu),
4097 LMAC_FID_ON_OST, 0);
4105 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4106 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4107 PFID_STRIPE_IDX_BITS;
4108 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4109 lma->lma_compat |= LMAC_STRIPE_INFO;
4110 if (ol->ol_comp_id != 0) {
4111 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4112 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4113 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4114 lma->lma_compat |= LMAC_COMP_INFO;
4117 lustre_loa_swab(loa, false);
4119 /* Remove old PFID EA entry firstly. */
4120 ll_vfs_dq_init(inode);
4121 rc = inode->i_op->removexattr(dentry, name);
4122 if (rc && rc != -ENODATA)
4125 /* Store the PFID EA inside the LMA EA. */
4126 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4130 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4131 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4135 lma->lma_incompat |= LMAI_STRIPED;
4136 lustre_lma_swab(lma);
4137 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4138 sizeof(*lma), XATTR_REPLACE);
4144 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4145 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4148 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4149 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4150 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4156 * Concurrency: @dt is read locked.
4158 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4159 const struct lu_buf *buf)
4161 struct osd_object *obj = osd_dt_obj(dt);
4162 struct inode *inode = obj->oo_inode;
4163 struct osd_thread_info *info = osd_oti_get(env);
4164 struct dentry *dentry = &info->oti_obj_dentry;
4166 if (!dt_object_exists(dt))
4169 LASSERT(!dt_object_remote(dt));
4170 LASSERT(inode->i_op != NULL);
4171 LASSERT(inode->i_op->listxattr != NULL);
4173 dentry->d_inode = inode;
4174 dentry->d_sb = inode->i_sb;
4175 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4178 static int osd_declare_xattr_del(const struct lu_env *env,
4179 struct dt_object *dt, const char *name,
4180 struct thandle *handle)
4182 struct osd_thandle *oh;
4183 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4185 LASSERT(!dt_object_remote(dt));
4186 LASSERT(handle != NULL);
4188 oh = container_of0(handle, struct osd_thandle, ot_super);
4189 LASSERT(oh->ot_handle == NULL);
4191 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4192 osd_dto_credits_noquota[DTO_XATTR_SET]);
4194 * xattr del may involve inode quota change, reserve credits for
4195 * dquot_initialize()
4197 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4203 * Concurrency: @dt is write locked.
4205 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4206 const char *name, struct thandle *handle)
4208 struct osd_object *obj = osd_dt_obj(dt);
4209 struct inode *inode = obj->oo_inode;
4210 struct osd_thread_info *info = osd_oti_get(env);
4211 struct dentry *dentry = &info->oti_obj_dentry;
4214 if (!dt_object_exists(dt))
4217 LASSERT(!dt_object_remote(dt));
4218 LASSERT(inode->i_op != NULL);
4219 LASSERT(handle != NULL);
4220 #ifdef HAVE_IOP_XATTR
4221 LASSERT(inode->i_op->removexattr != NULL);
4224 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4226 ll_vfs_dq_init(inode);
4227 dentry->d_inode = inode;
4228 dentry->d_sb = inode->i_sb;
4229 rc = inode->i_op->removexattr(dentry, name);
4230 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
4231 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4233 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4235 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4236 &info->oti_ost_attrs);
4238 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
4243 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4244 lustre_lma_swab(lma);
4245 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4246 sizeof(*lma), XATTR_REPLACE);
4251 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4254 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4255 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4256 osd_oxc_del(obj, name);
4261 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4262 __u64 start, __u64 end)
4264 struct osd_object *obj = osd_dt_obj(dt);
4265 struct inode *inode = obj->oo_inode;
4266 struct osd_thread_info *info = osd_oti_get(env);
4267 struct dentry *dentry = &info->oti_obj_dentry;
4268 struct file *file = &info->oti_file;
4273 dentry->d_inode = inode;
4274 dentry->d_sb = inode->i_sb;
4275 file->f_path.dentry = dentry;
4276 file->f_mapping = inode->i_mapping;
4277 file->f_op = inode->i_fop;
4278 set_file_inode(file, inode);
4280 rc = ll_vfs_fsync_range(file, start, end, 0);
4285 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4294 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4295 const struct dt_index_features *feat)
4297 struct iam_descr *descr;
4299 if (osd_object_is_root(o))
4300 return feat == &dt_directory_features;
4302 LASSERT(o->oo_dir != NULL);
4304 descr = o->oo_dir->od_container.ic_descr;
4305 if (feat == &dt_directory_features) {
4306 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4312 feat->dif_keysize_min <= descr->id_key_size &&
4313 descr->id_key_size <= feat->dif_keysize_max &&
4314 feat->dif_recsize_min <= descr->id_rec_size &&
4315 descr->id_rec_size <= feat->dif_recsize_max &&
4316 !(feat->dif_flags & (DT_IND_VARKEY |
4317 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4318 ergo(feat->dif_flags & DT_IND_UPDATE,
4319 1 /* XXX check that object (and file system) is
4324 static int osd_iam_container_init(const struct lu_env *env,
4325 struct osd_object *obj,
4326 struct osd_directory *dir)
4328 struct iam_container *bag = &dir->od_container;
4331 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4335 result = iam_container_setup(bag);
4337 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4339 iam_container_fini(bag);
4346 * Concurrency: no external locking is necessary.
4348 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4349 const struct dt_index_features *feat)
4353 struct osd_object *obj = osd_dt_obj(dt);
4355 LINVRNT(osd_invariant(obj));
4357 if (osd_object_is_root(obj)) {
4358 dt->do_index_ops = &osd_index_ea_ops;
4360 } else if (feat == &dt_directory_features) {
4361 dt->do_index_ops = &osd_index_ea_ops;
4362 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4367 } else if (unlikely(feat == &dt_otable_features)) {
4368 dt->do_index_ops = &osd_otable_ops;
4370 } else if (unlikely(feat == &dt_acct_features)) {
4371 dt->do_index_ops = &osd_acct_index_ops;
4374 } else if (!osd_has_index(obj)) {
4375 struct osd_directory *dir;
4380 spin_lock(&obj->oo_guard);
4381 if (obj->oo_dir == NULL)
4385 * Concurrent thread allocated container data.
4388 spin_unlock(&obj->oo_guard);
4390 * Now, that we have container data, serialize its
4393 down_write(&obj->oo_ext_idx_sem);
4395 * recheck under lock.
4397 if (!osd_has_index(obj))
4398 result = osd_iam_container_init(env, obj,
4402 up_write(&obj->oo_ext_idx_sem);
4410 if (result == 0 && skip_iam == 0) {
4411 if (!osd_iam_index_probe(env, obj, feat))
4414 LINVRNT(osd_invariant(obj));
4419 static int osd_otable_it_attr_get(const struct lu_env *env,
4420 struct dt_object *dt,
4421 struct lu_attr *attr)
4427 static const struct dt_object_operations osd_obj_ops = {
4428 .do_read_lock = osd_read_lock,
4429 .do_write_lock = osd_write_lock,
4430 .do_read_unlock = osd_read_unlock,
4431 .do_write_unlock = osd_write_unlock,
4432 .do_write_locked = osd_write_locked,
4433 .do_attr_get = osd_attr_get,
4434 .do_declare_attr_set = osd_declare_attr_set,
4435 .do_attr_set = osd_attr_set,
4436 .do_ah_init = osd_ah_init,
4437 .do_declare_create = osd_declare_create,
4438 .do_create = osd_create,
4439 .do_declare_destroy = osd_declare_destroy,
4440 .do_destroy = osd_destroy,
4441 .do_index_try = osd_index_try,
4442 .do_declare_ref_add = osd_declare_ref_add,
4443 .do_ref_add = osd_ref_add,
4444 .do_declare_ref_del = osd_declare_ref_del,
4445 .do_ref_del = osd_ref_del,
4446 .do_xattr_get = osd_xattr_get,
4447 .do_declare_xattr_set = osd_declare_xattr_set,
4448 .do_xattr_set = osd_xattr_set,
4449 .do_declare_xattr_del = osd_declare_xattr_del,
4450 .do_xattr_del = osd_xattr_del,
4451 .do_xattr_list = osd_xattr_list,
4452 .do_object_sync = osd_object_sync,
4453 .do_invalidate = osd_invalidate,
4456 static const struct dt_object_operations osd_obj_otable_it_ops = {
4457 .do_attr_get = osd_otable_it_attr_get,
4458 .do_index_try = osd_index_try,
4461 static int osd_index_declare_iam_delete(const struct lu_env *env,
4462 struct dt_object *dt,
4463 const struct dt_key *key,
4464 struct thandle *handle)
4466 struct osd_thandle *oh;
4468 oh = container_of0(handle, struct osd_thandle, ot_super);
4469 LASSERT(oh->ot_handle == NULL);
4471 /* Recycle may cause additional three blocks to be changed. */
4472 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4473 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4479 * delete a (key, value) pair from index \a dt specified by \a key
4481 * \param dt osd index object
4482 * \param key key for index
4483 * \param rec record reference
4484 * \param handle transaction handler
4487 * \retval -ve failure
4489 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4490 const struct dt_key *key,
4491 struct thandle *handle)
4493 struct osd_thread_info *oti = osd_oti_get(env);
4494 struct osd_object *obj = osd_dt_obj(dt);
4495 struct osd_thandle *oh;
4496 struct iam_path_descr *ipd;
4497 struct iam_container *bag = &obj->oo_dir->od_container;
4501 if (!dt_object_exists(dt))
4504 LINVRNT(osd_invariant(obj));
4505 LASSERT(!dt_object_remote(dt));
4506 LASSERT(bag->ic_object == obj->oo_inode);
4507 LASSERT(handle != NULL);
4509 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4511 ipd = osd_idx_ipd_get(env, bag);
4512 if (unlikely(ipd == NULL))
4515 oh = container_of0(handle, struct osd_thandle, ot_super);
4516 LASSERT(oh->ot_handle != NULL);
4517 LASSERT(oh->ot_handle->h_transaction != NULL);
4519 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4520 /* swab quota uid/gid provided by caller */
4521 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4522 key = (const struct dt_key *)&oti->oti_quota_id;
4525 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4526 osd_ipd_put(env, bag, ipd);
4527 LINVRNT(osd_invariant(obj));
4528 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4532 static int osd_index_declare_ea_delete(const struct lu_env *env,
4533 struct dt_object *dt,
4534 const struct dt_key *key,
4535 struct thandle *handle)
4537 struct osd_thandle *oh;
4538 struct inode *inode;
4542 LASSERT(!dt_object_remote(dt));
4543 LASSERT(handle != NULL);
4545 oh = container_of0(handle, struct osd_thandle, ot_super);
4546 LASSERT(oh->ot_handle == NULL);
4548 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4549 if (key != NULL && unlikely(strcmp((char *)key, dotdot) == 0)) {
4550 /* '..' to a remote object has a local representative */
4551 credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
4552 /* to reset LMAI_REMOTE_PARENT */
4555 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4557 inode = osd_dt_obj(dt)->oo_inode;
4561 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4562 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
4567 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
4570 struct osd_fid_pack *rec;
4573 if (de->file_type & LDISKFS_DIRENT_LUFID) {
4574 rec = (struct osd_fid_pack *) (de->name + de->name_len + 1);
4575 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
4576 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
4582 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
4583 const struct lu_fid *fid)
4585 struct seq_server_site *ss = osd_seq_site(osd);
4588 /* FID seqs not in FLDB, must be local seq */
4589 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
4592 /* If FLD is not being initialized yet, it only happens during the
4593 * initialization, likely during mgs initialization, and we assume
4594 * this is local FID. */
4595 if (ss == NULL || ss->ss_server_fld == NULL)
4598 /* Only check the local FLDB here */
4599 if (osd_seq_exists(env, osd, fid_seq(fid)))
4605 static void osd_take_care_of_agent(const struct lu_env *env,
4606 struct osd_device *osd,
4607 struct osd_thandle *oh,
4608 struct ldiskfs_dir_entry_2 *de)
4610 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4611 struct osd_idmap_cache *idc;
4612 int rc, schedule = 0;
4614 LASSERT(de != NULL);
4616 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
4617 if (likely(rc == 0)) {
4618 idc = osd_idc_find_or_init(env, osd, fid);
4619 if (IS_ERR(idc) || idc->oic_remote)
4621 } else if (rc == -ENODATA) {
4622 /* can't get FID, postpone to the end of the
4623 * transaction when iget() is safe */
4626 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
4629 osd_schedule_agent_inode_removal(env, oh,
4630 le32_to_cpu(de->inode));
4634 * Index delete function for interoperability mode (b11826).
4635 * It will remove the directory entry added by osd_index_ea_insert().
4636 * This entry is needed to maintain name->fid mapping.
4638 * \param key, key i.e. file entry to be deleted
4640 * \retval 0, on success
4641 * \retval -ve, on error
4643 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
4644 const struct dt_key *key, struct thandle *handle)
4646 struct osd_object *obj = osd_dt_obj(dt);
4647 struct inode *dir = obj->oo_inode;
4648 struct dentry *dentry;
4649 struct osd_thandle *oh;
4650 struct ldiskfs_dir_entry_2 *de = NULL;
4651 struct buffer_head *bh;
4652 struct htree_lock *hlock = NULL;
4653 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4654 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4658 if (!dt_object_exists(dt))
4661 LINVRNT(osd_invariant(obj));
4662 LASSERT(!dt_object_remote(dt));
4663 LASSERT(handle != NULL);
4665 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4667 oh = container_of(handle, struct osd_thandle, ot_super);
4668 LASSERT(oh->ot_handle != NULL);
4669 LASSERT(oh->ot_handle->h_transaction != NULL);
4671 ll_vfs_dq_init(dir);
4672 dentry = osd_child_dentry_get(env, obj,
4673 (char *)key, strlen((char *)key));
4675 if (obj->oo_hl_head != NULL) {
4676 hlock = osd_oti_get(env)->oti_hlock;
4677 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
4678 dir, LDISKFS_HLOCK_DEL);
4680 down_write(&obj->oo_ext_idx_sem);
4683 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
4685 /* If this is not the ".." entry, it might be a remote DNE
4686 * entry and we need to check if the FID is for a remote
4687 * MDT. If the FID is not in the directory entry (e.g.
4688 * upgraded 1.8 filesystem without dirdata enabled) then
4689 * we need to get the FID from the LMA. For a remote directory
4690 * there HAS to be an LMA, it cannot be an IGIF inode in this
4693 * Delete the entry before the agent inode in order to
4694 * simplify error handling. At worst an error after deleting
4695 * the entry first might leak the agent inode afterward. The
4696 * reverse would need filesystem abort in case of error deleting
4697 * the entry after the agent had been removed, or leave a
4698 * dangling entry pointing at a random inode. */
4699 if (strcmp((char *)key, dotdot) != 0)
4700 osd_take_care_of_agent(env, osd, oh, de);
4701 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
4707 ldiskfs_htree_unlock(hlock);
4709 up_write(&obj->oo_ext_idx_sem);
4714 /* For inode on the remote MDT, .. will point to
4715 * /Agent directory, Check whether it needs to delete
4716 * from agent directory */
4717 if (unlikely(strcmp((char *)key, dotdot) == 0)) {
4720 ret = osd_delete_from_remote_parent(env, osd_obj2dev(obj),
4723 /* Sigh, the entry has been deleted, and
4724 * it is not easy to revert it back, so
4725 * let's keep this error private, and let
4726 * LFSCK fix it. XXX */
4727 CERROR("%s: delete remote parent "DFID": rc = %d\n",
4728 osd_name(osd), PFID(fid), ret);
4731 LASSERT(osd_invariant(obj));
4732 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4737 * Lookup index for \a key and copy record to \a rec.
4739 * \param dt osd index object
4740 * \param key key for index
4741 * \param rec record reference
4743 * \retval +ve success : exact mach
4744 * \retval 0 return record with key not greater than \a key
4745 * \retval -ve failure
4747 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
4748 struct dt_rec *rec, const struct dt_key *key)
4750 struct osd_object *obj = osd_dt_obj(dt);
4751 struct iam_path_descr *ipd;
4752 struct iam_container *bag = &obj->oo_dir->od_container;
4753 struct osd_thread_info *oti = osd_oti_get(env);
4754 struct iam_iterator *it = &oti->oti_idx_it;
4755 struct iam_rec *iam_rec;
4759 if (!dt_object_exists(dt))
4762 LASSERT(osd_invariant(obj));
4763 LASSERT(!dt_object_remote(dt));
4764 LASSERT(bag->ic_object == obj->oo_inode);
4766 ipd = osd_idx_ipd_get(env, bag);
4770 /* got ipd now we can start iterator. */
4771 iam_it_init(it, bag, 0, ipd);
4773 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4774 /* swab quota uid/gid provided by caller */
4775 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4776 key = (const struct dt_key *)&oti->oti_quota_id;
4779 rc = iam_it_get(it, (struct iam_key *)key);
4781 if (S_ISDIR(obj->oo_inode->i_mode))
4782 iam_rec = (struct iam_rec *)oti->oti_ldp;
4784 iam_rec = (struct iam_rec *) rec;
4786 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
4788 if (S_ISDIR(obj->oo_inode->i_mode))
4789 osd_fid_unpack((struct lu_fid *) rec,
4790 (struct osd_fid_pack *)iam_rec);
4791 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
4792 osd_quota_unpack(obj, rec);
4797 osd_ipd_put(env, bag, ipd);
4799 LINVRNT(osd_invariant(obj));
4804 static int osd_index_declare_iam_insert(const struct lu_env *env,
4805 struct dt_object *dt,
4806 const struct dt_rec *rec,
4807 const struct dt_key *key,
4808 struct thandle *handle)
4810 struct osd_thandle *oh;
4812 LASSERT(handle != NULL);
4814 oh = container_of0(handle, struct osd_thandle, ot_super);
4815 LASSERT(oh->ot_handle == NULL);
4817 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
4818 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
4824 * Inserts (key, value) pair in \a dt index object.
4826 * \param dt osd index object
4827 * \param key key for index
4828 * \param rec record reference
4829 * \param th transaction handler
4832 * \retval -ve failure
4834 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
4835 const struct dt_rec *rec,
4836 const struct dt_key *key, struct thandle *th,
4839 struct osd_object *obj = osd_dt_obj(dt);
4840 struct iam_path_descr *ipd;
4841 struct osd_thandle *oh;
4842 struct iam_container *bag;
4843 struct osd_thread_info *oti = osd_oti_get(env);
4844 struct iam_rec *iam_rec;
4848 if (!dt_object_exists(dt))
4851 LINVRNT(osd_invariant(obj));
4852 LASSERT(!dt_object_remote(dt));
4854 bag = &obj->oo_dir->od_container;
4855 LASSERT(bag->ic_object == obj->oo_inode);
4856 LASSERT(th != NULL);
4858 osd_trans_exec_op(env, th, OSD_OT_INSERT);
4860 ipd = osd_idx_ipd_get(env, bag);
4861 if (unlikely(ipd == NULL))
4864 oh = container_of0(th, struct osd_thandle, ot_super);
4865 LASSERT(oh->ot_handle != NULL);
4866 LASSERT(oh->ot_handle->h_transaction != NULL);
4867 if (S_ISDIR(obj->oo_inode->i_mode)) {
4868 iam_rec = (struct iam_rec *)oti->oti_ldp;
4869 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec, &oti->oti_fid);
4870 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4871 /* pack quota uid/gid */
4872 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4873 key = (const struct dt_key *)&oti->oti_quota_id;
4874 /* pack quota record */
4875 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
4876 iam_rec = (struct iam_rec *)rec;
4878 iam_rec = (struct iam_rec *)rec;
4881 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
4883 osd_ipd_put(env, bag, ipd);
4884 LINVRNT(osd_invariant(obj));
4885 osd_trans_exec_check(env, th, OSD_OT_INSERT);
4890 * Calls ldiskfs_add_entry() to add directory entry
4891 * into the directory. This is required for
4892 * interoperability mode (b11826)
4894 * \retval 0, on success
4895 * \retval -ve, on error
4897 static int __osd_ea_add_rec(struct osd_thread_info *info,
4898 struct osd_object *pobj, struct inode *cinode,
4899 const char *name, const struct lu_fid *fid,
4900 struct htree_lock *hlock, struct thandle *th)
4902 struct ldiskfs_dentry_param *ldp;
4903 struct dentry *child;
4904 struct osd_thandle *oth;
4907 oth = container_of(th, struct osd_thandle, ot_super);
4908 LASSERT(oth->ot_handle != NULL);
4909 LASSERT(oth->ot_handle->h_transaction != NULL);
4910 LASSERT(pobj->oo_inode);
4912 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4913 if (unlikely(pobj->oo_inode ==
4914 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
4917 osd_get_ldiskfs_dirent_param(ldp, fid);
4918 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
4919 child->d_fsdata = (void *)ldp;
4920 ll_vfs_dq_init(pobj->oo_inode);
4921 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
4922 child, cinode, hlock);
4923 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
4924 struct ldiskfs_dir_entry_2 *de;
4925 struct buffer_head *bh;
4928 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
4931 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
4934 if (S_ISDIR(cinode->i_mode))
4935 de->file_type = LDISKFS_DIRENT_LUFID |
4936 LDISKFS_FT_REG_FILE;
4938 de->file_type = LDISKFS_DIRENT_LUFID |
4940 ldiskfs_handle_dirty_metadata(oth->ot_handle,
4951 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
4952 * into the directory.Also sets flags into osd object to
4953 * indicate dot and dotdot are created. This is required for
4954 * interoperability mode (b11826)
4956 * \param dir directory for dot and dotdot fixup.
4957 * \param obj child object for linking
4959 * \retval 0, on success
4960 * \retval -ve, on error
4962 static int osd_add_dot_dotdot(struct osd_thread_info *info,
4963 struct osd_object *dir,
4964 struct inode *parent_dir, const char *name,
4965 const struct lu_fid *dot_fid,
4966 const struct lu_fid *dot_dot_fid,
4969 struct inode *inode = dir->oo_inode;
4970 struct osd_thandle *oth;
4973 oth = container_of(th, struct osd_thandle, ot_super);
4974 LASSERT(oth->ot_handle->h_transaction != NULL);
4975 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
4977 if (strcmp(name, dot) == 0) {
4978 if (dir->oo_compat_dot_created) {
4981 LASSERT(inode->i_ino == parent_dir->i_ino);
4982 dir->oo_compat_dot_created = 1;
4985 } else if (strcmp(name, dotdot) == 0) {
4986 if (!dir->oo_compat_dot_created)
4988 /* in case of rename, dotdot is already created */
4989 if (dir->oo_compat_dotdot_created) {
4990 return __osd_ea_add_rec(info, dir, parent_dir, name,
4991 dot_dot_fid, NULL, th);
4994 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
4995 struct lu_fid tfid = *dot_dot_fid;
4998 result = osd_add_dot_dotdot_internal(info,
4999 dir->oo_inode, parent_dir, dot_fid,
5002 result = osd_add_dot_dotdot_internal(info,
5003 dir->oo_inode, parent_dir, dot_fid,
5008 dir->oo_compat_dotdot_created = 1;
5016 * It will call the appropriate osd_add* function and return the
5017 * value, return by respective functions.
5019 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5020 struct inode *cinode, const char *name,
5021 const struct lu_fid *fid, struct thandle *th)
5023 struct osd_thread_info *info = osd_oti_get(env);
5024 struct htree_lock *hlock;
5027 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5029 if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' &&
5031 if (hlock != NULL) {
5032 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5035 down_write(&pobj->oo_ext_idx_sem);
5038 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5039 lu_object_fid(&pobj->oo_dt.do_lu),
5042 if (hlock != NULL) {
5043 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5044 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5046 down_write(&pobj->oo_ext_idx_sem);
5049 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5050 struct lu_fid *tfid = &info->oti_fid;
5054 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5057 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5062 ldiskfs_htree_unlock(hlock);
5064 up_write(&pobj->oo_ext_idx_sem);
5070 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5071 struct osd_idmap_cache *oic)
5073 struct osd_scrub *scrub = &dev->od_scrub;
5074 struct lu_fid *fid = &oic->oic_fid;
5075 struct osd_inode_id *id = &oic->oic_lid;
5076 struct inode *inode = NULL;
5082 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5085 if (scrub->os_pos_current > id->oii_ino)
5089 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5090 if (rc == -ENOENT) {
5091 __u32 gen = id->oii_gen;
5097 inode = osd_iget(oti, dev, id);
5098 /* The inode has been removed (by race maybe). */
5099 if (IS_ERR(inode)) {
5100 rc = PTR_ERR(inode);
5102 RETURN(rc == -ESTALE ? -ENOENT : rc);
5105 /* The OI mapping is lost. */
5106 if (gen != OSD_OII_NOGEN)
5109 /* The inode may has been reused by others, we do not know,
5110 * leave it to be handled by subsequent osd_fid_lookup(). */
5112 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5119 if (thread_is_running(&scrub->os_thread)) {
5120 if (inode == NULL) {
5121 inode = osd_iget(oti, dev, id);
5122 /* The inode has been removed (by race maybe). */
5123 if (IS_ERR(inode)) {
5124 rc = PTR_ERR(inode);
5126 RETURN(rc == -ESTALE ? -ENOENT : rc);
5130 rc = osd_oii_insert(dev, oic, insert);
5131 /* There is race condition between osd_oi_lookup and OI scrub.
5132 * The OI scrub finished just after osd_oi_lookup() failure.
5133 * Under such case, it is unnecessary to trigger OI scrub again,
5134 * but try to call osd_oi_lookup() again. */
5135 if (unlikely(rc == -EAGAIN))
5138 if (!S_ISDIR(inode->i_mode))
5141 rc = osd_check_lmv(oti, dev, inode, oic);
5146 if (!dev->od_noscrub && ++once == 1) {
5147 rc = osd_scrub_start(dev, SS_AUTO_PARTIAL | SS_CLEAR_DRYRUN |
5149 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5150 "%s: trigger partial OI scrub for RPC inconsistency "
5151 "checking FID "DFID": rc = %d\n",
5152 osd_dev2name(dev), PFID(fid), rc);
5153 if (rc == 0 || rc == -EALREADY)
5166 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5167 struct osd_device *dev,
5168 struct osd_idmap_cache *oic,
5169 struct lu_fid *fid, __u32 ino)
5171 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5172 struct inode *inode;
5175 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5176 inode = osd_iget(oti, dev, &oic->oic_lid);
5177 if (IS_ERR(inode)) {
5178 fid_zero(&oic->oic_fid);
5179 return PTR_ERR(inode);
5182 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5185 fid_zero(&oic->oic_fid);
5187 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5191 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5192 struct osd_inode_id *id, const struct lu_fid *fid)
5194 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5195 id->oii_ino, id->oii_gen, info);
5196 info->oti_cache.oic_lid = *id;
5197 info->oti_cache.oic_fid = *fid;
5198 info->oti_cache.oic_dev = osd;
5202 * Get parent FID from the linkEA.
5204 * For a directory which parent resides on remote MDT, to satisfy the
5205 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5206 * the other hand, to make the lookup(..) on the directory can return
5207 * the real parent FID, we append the real parent FID after its ".."
5208 * name entry in the /REMOTE_PARENT_DIR.
5210 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5211 * backup. So after the restore, we cannot get the right parent FID from
5212 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5213 * we have stored the real parent FID in the directory object's linkEA,
5214 * we can parse the linkEA for the real parent FID.
5216 * \param[in] env pointer to the thread context
5217 * \param[in] obj pointer to the object to be handled
5218 * \param[out]fid pointer to the buffer to hold the parent FID
5220 * \retval 0 for getting the real parent FID successfully
5221 * \retval negative error number on failure
5223 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5224 struct osd_object *obj,
5227 struct osd_thread_info *oti = osd_oti_get(env);
5228 struct lu_buf *buf = &oti->oti_big_buf;
5229 struct dentry *dentry = &oti->oti_obj_dentry;
5230 struct inode *inode = obj->oo_inode;
5231 struct linkea_data ldata = { NULL };
5236 if (!S_ISDIR(inode->i_mode))
5240 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5241 buf->lb_buf, buf->lb_len);
5242 if (rc == -ERANGE) {
5243 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5246 lu_buf_realloc(buf, rc);
5247 if (buf->lb_buf == NULL)
5254 if (unlikely(rc == 0))
5260 if (unlikely(buf->lb_buf == NULL)) {
5261 lu_buf_realloc(buf, rc);
5262 if (buf->lb_buf == NULL)
5269 rc = linkea_init_with_rec(&ldata);
5271 linkea_first_entry(&ldata);
5272 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5278 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5279 struct inode *inode,
5280 const struct lu_fid *pfid,
5281 const char *name, const int namelen)
5283 struct osd_thread_info *oti = osd_oti_get(env);
5284 struct lu_buf *buf = &oti->oti_big_buf;
5285 struct dentry *dentry = &oti->oti_obj_dentry;
5286 struct linkea_data ldata = { NULL };
5287 struct lu_name cname = { .ln_name = name,
5288 .ln_namelen = namelen };
5293 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5294 buf->lb_buf, buf->lb_len);
5296 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5301 if (unlikely(rc == 0))
5304 if (buf->lb_len < rc) {
5305 lu_buf_realloc(buf, rc);
5306 if (buf->lb_buf == NULL)
5313 rc = linkea_init_with_rec(&ldata);
5315 rc = linkea_links_find(&ldata, &cname, pfid);
5321 * Calls ->lookup() to find dentry. From dentry get inode and
5322 * read inode's ea to get fid. This is required for interoperability
5325 * \retval 0, on success
5326 * \retval -ve, on error
5328 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5329 struct dt_rec *rec, const struct dt_key *key)
5331 struct inode *dir = obj->oo_inode;
5332 struct dentry *dentry;
5333 struct ldiskfs_dir_entry_2 *de;
5334 struct buffer_head *bh;
5335 struct lu_fid *fid = (struct lu_fid *) rec;
5336 struct htree_lock *hlock = NULL;
5341 LASSERT(dir->i_op != NULL);
5342 LASSERT(dir->i_op->lookup != NULL);
5344 dentry = osd_child_dentry_get(env, obj,
5345 (char *)key, strlen((char *)key));
5347 if (obj->oo_hl_head != NULL) {
5348 hlock = osd_oti_get(env)->oti_hlock;
5349 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5350 dir, LDISKFS_HLOCK_LOOKUP);
5352 down_read(&obj->oo_ext_idx_sem);
5355 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5357 struct osd_thread_info *oti = osd_oti_get(env);
5358 struct osd_inode_id *id = &oti->oti_id;
5359 struct osd_idmap_cache *oic = &oti->oti_cache;
5360 struct osd_device *dev = osd_obj2dev(obj);
5362 ino = le32_to_cpu(de->inode);
5363 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5365 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5369 rc = osd_get_fid_from_dentry(de, rec);
5371 /* done with de, release bh */
5374 if (unlikely(is_remote_parent_ino(dev, ino))) {
5375 const char *name = (const char *)key;
5377 /* If the parent is on remote MDT, and there
5378 * is no FID-in-dirent, then we have to get
5379 * the parent FID from the linkEA. */
5380 if (likely(strlen(name) == 2 &&
5381 name[0] == '.' && name[1] == '.'))
5382 rc = osd_get_pfid_from_linkea(env, obj,
5385 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5388 osd_id_gen(id, ino, OSD_OII_NOGEN);
5391 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5392 fid_zero(&oic->oic_fid);
5397 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5398 rc = osd_consistency_check(oti, dev, oic);
5400 fid_zero(&oic->oic_fid);
5409 ldiskfs_htree_unlock(hlock);
5411 up_read(&obj->oo_ext_idx_sem);
5415 static int osd_index_declare_ea_insert(const struct lu_env *env,
5416 struct dt_object *dt,
5417 const struct dt_rec *rec,
5418 const struct dt_key *key,
5419 struct thandle *handle)
5421 struct osd_thandle *oh;
5422 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5423 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5424 const struct lu_fid *fid = rec1->rec_fid;
5425 int credits, rc = 0;
5426 struct osd_idmap_cache *idc;
5429 LASSERT(!dt_object_remote(dt));
5430 LASSERT(handle != NULL);
5431 LASSERT(fid != NULL);
5432 LASSERT(rec1->rec_type != 0);
5434 oh = container_of0(handle, struct osd_thandle, ot_super);
5435 LASSERT(oh->ot_handle == NULL);
5437 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5439 /* we can't call iget() while a transactions is running
5440 * (this can lead to a deadlock), but we need to know
5441 * inum and object type. so we find this information at
5442 * declaration and cache in per-thread info */
5443 idc = osd_idc_find_or_init(env, osd, fid);
5445 RETURN(PTR_ERR(idc));
5446 if (idc->oic_remote) {
5447 /* a reference to remote inode is represented by an
5448 * agent inode which we have to create */
5449 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5450 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5453 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5455 if (osd_dt_obj(dt)->oo_inode != NULL) {
5456 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5458 /* We ignore block quota on meta pool (MDTs), so needn't
5459 * calculate how many blocks will be consumed by this index
5461 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5463 i_projid_read(inode), 0,
5464 oh, osd_dt_obj(dt), NULL,
5469 #ifdef HAVE_PROJECT_QUOTA
5470 /* Reserve credits for local agent inode to transfer
5471 * to 0, quota enforcement is ignored in this case.
5473 if (idc->oic_remote &&
5474 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
5475 i_projid_read(inode) != 0)
5476 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
5477 0, i_projid_read(inode),
5478 0, false, PRJQUOTA);
5487 * Index add function for interoperability mode (b11826).
5488 * It will add the directory entry.This entry is needed to
5489 * maintain name->fid mapping.
5491 * \param key it is key i.e. file entry to be inserted
5492 * \param rec it is value of given key i.e. fid
5494 * \retval 0, on success
5495 * \retval -ve, on error
5497 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5498 const struct dt_rec *rec,
5499 const struct dt_key *key, struct thandle *th,
5502 struct osd_object *obj = osd_dt_obj(dt);
5503 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5504 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5505 const struct lu_fid *fid = rec1->rec_fid;
5506 const char *name = (const char *)key;
5507 struct osd_thread_info *oti = osd_oti_get(env);
5508 struct inode *child_inode = NULL;
5509 struct osd_idmap_cache *idc;
5513 if (!dt_object_exists(dt))
5516 LASSERT(osd_invariant(obj));
5517 LASSERT(!dt_object_remote(dt));
5518 LASSERT(th != NULL);
5520 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5522 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5524 idc = osd_idc_find(env, osd, fid);
5525 if (unlikely(idc == NULL)) {
5526 /* this dt_insert() wasn't declared properly, so
5527 * FID is missing in OI cache. we better do not
5528 * lookup FID in FLDB/OI and don't risk to deadlock,
5529 * but in some special cases (lfsck testing, etc)
5530 * it's much simpler than fixing a caller */
5531 CERROR("%s: "DFID" wasn't declared for insert\n",
5532 osd_name(osd), PFID(fid));
5534 idc = osd_idc_find_or_init(env, osd, fid);
5536 RETURN(PTR_ERR(idc));
5539 if (idc->oic_remote) {
5540 /* Insert remote entry */
5541 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
5542 struct osd_mdobj_map *omm = osd->od_mdt_map;
5543 struct osd_thandle *oh;
5545 /* If parent on remote MDT, we need put this object
5547 oh = container_of(th, typeof(*oh), ot_super);
5548 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5550 CERROR("%s: add "DFID" error: rc = %d\n",
5552 PFID(lu_object_fid(&dt->do_lu)), rc);
5556 child_inode = igrab(omm->omm_remote_parent->d_inode);
5558 child_inode = osd_create_local_agent_inode(env, osd,
5559 obj, fid, rec1->rec_type & S_IFMT, th);
5560 if (IS_ERR(child_inode))
5561 RETURN(PTR_ERR(child_inode));
5564 /* Insert local entry */
5565 if (unlikely(idc->oic_lid.oii_ino == 0)) {
5566 /* for a reason OI cache wasn't filled properly */
5567 CERROR("%s: OIC for "DFID" isn't filled\n",
5568 osd_name(osd), PFID(fid));
5571 child_inode = oti->oti_inode;
5572 if (unlikely(child_inode == NULL)) {
5573 struct ldiskfs_inode_info *lii;
5577 child_inode = oti->oti_inode = &lii->vfs_inode;
5579 child_inode->i_sb = osd_sb(osd);
5580 child_inode->i_ino = idc->oic_lid.oii_ino;
5581 child_inode->i_mode = rec1->rec_type & S_IFMT;
5584 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
5586 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
5587 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
5589 if (child_inode && child_inode != oti->oti_inode)
5591 LASSERT(osd_invariant(obj));
5592 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5597 * Initialize osd Iterator for given osd index object.
5599 * \param dt osd index object
5602 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
5603 struct dt_object *dt,
5606 struct osd_it_iam *it;
5607 struct osd_object *obj = osd_dt_obj(dt);
5608 struct lu_object *lo = &dt->do_lu;
5609 struct iam_path_descr *ipd;
5610 struct iam_container *bag = &obj->oo_dir->od_container;
5612 if (!dt_object_exists(dt))
5613 return ERR_PTR(-ENOENT);
5617 return ERR_PTR(-ENOMEM);
5619 ipd = osd_it_ipd_get(env, bag);
5620 if (likely(ipd != NULL)) {
5624 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
5625 return (struct dt_it *)it;
5628 return ERR_PTR(-ENOMEM);
5633 * free given Iterator.
5635 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
5637 struct osd_it_iam *it = (struct osd_it_iam *)di;
5638 struct osd_object *obj = it->oi_obj;
5640 iam_it_fini(&it->oi_it);
5641 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
5642 osd_object_put(env, obj);
5647 * Move Iterator to record specified by \a key
5649 * \param di osd iterator
5650 * \param key key for index
5652 * \retval +ve di points to record with least key not larger than key
5653 * \retval 0 di points to exact matched key
5654 * \retval -ve failure
5657 static int osd_it_iam_get(const struct lu_env *env,
5658 struct dt_it *di, const struct dt_key *key)
5660 struct osd_thread_info *oti = osd_oti_get(env);
5661 struct osd_it_iam *it = (struct osd_it_iam *)di;
5663 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5664 /* swab quota uid/gid */
5665 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5666 key = (struct dt_key *)&oti->oti_quota_id;
5669 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
5675 * \param di osd iterator
5677 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
5679 struct osd_it_iam *it = (struct osd_it_iam *)di;
5681 iam_it_put(&it->oi_it);
5685 * Move iterator by one record
5687 * \param di osd iterator
5689 * \retval +1 end of container reached
5691 * \retval -ve failure
5694 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
5696 struct osd_it_iam *it = (struct osd_it_iam *)di;
5698 return iam_it_next(&it->oi_it);
5702 * Return pointer to the key under iterator.
5705 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
5706 const struct dt_it *di)
5708 struct osd_thread_info *oti = osd_oti_get(env);
5709 struct osd_it_iam *it = (struct osd_it_iam *)di;
5710 struct osd_object *obj = it->oi_obj;
5713 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
5715 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
5716 /* swab quota uid/gid */
5717 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
5718 key = (struct dt_key *)&oti->oti_quota_id;
5725 * Return size of key under iterator (in bytes)
5728 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
5730 struct osd_it_iam *it = (struct osd_it_iam *)di;
5732 return iam_it_key_size(&it->oi_it);
5736 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
5738 /* check if file type is required */
5739 if (ent->lde_attrs & LUDA_TYPE) {
5740 struct luda_type *lt;
5741 int align = sizeof(*lt) - 1;
5743 len = (len + align) & ~align;
5744 lt = (struct luda_type *)(ent->lde_name + len);
5745 lt->lt_type = cpu_to_le16(DTTOIF(type));
5748 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
5752 * build lu direct from backend fs dirent.
5756 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
5757 char *name, __u16 namelen, __u16 type, __u32 attr)
5759 ent->lde_attrs = attr | LUDA_FID;
5760 fid_cpu_to_le(&ent->lde_fid, fid);
5762 ent->lde_hash = cpu_to_le64(offset);
5763 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
5765 strncpy(ent->lde_name, name, namelen);
5766 ent->lde_name[namelen] = '\0';
5767 ent->lde_namelen = cpu_to_le16(namelen);
5769 /* append lustre attributes */
5770 osd_it_append_attrs(ent, namelen, type);
5774 * Return pointer to the record under iterator.
5776 static int osd_it_iam_rec(const struct lu_env *env,
5777 const struct dt_it *di,
5778 struct dt_rec *dtrec, __u32 attr)
5780 struct osd_it_iam *it = (struct osd_it_iam *)di;
5781 struct osd_thread_info *info = osd_oti_get(env);
5784 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
5785 const struct osd_fid_pack *rec;
5786 struct lu_fid *fid = &info->oti_fid;
5787 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
5793 name = (char *)iam_it_key_get(&it->oi_it);
5795 RETURN(PTR_ERR(name));
5797 namelen = iam_it_key_size(&it->oi_it);
5799 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
5801 RETURN(PTR_ERR(rec));
5803 rc = osd_fid_unpack(fid, rec);
5807 hash = iam_it_store(&it->oi_it);
5809 /* IAM does not store object type in IAM index (dir) */
5810 osd_it_pack_dirent(lde, fid, hash, name, namelen,
5812 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5813 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5814 (struct iam_rec *)dtrec);
5815 osd_quota_unpack(it->oi_obj, dtrec);
5817 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5818 (struct iam_rec *)dtrec);
5825 * Returns cookie for current Iterator position.
5827 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
5829 struct osd_it_iam *it = (struct osd_it_iam *)di;
5831 return iam_it_store(&it->oi_it);
5835 * Restore iterator from cookie.
5837 * \param di osd iterator
5838 * \param hash Iterator location cookie
5840 * \retval +ve di points to record with least key not larger than key.
5841 * \retval 0 di points to exact matched key
5842 * \retval -ve failure
5845 static int osd_it_iam_load(const struct lu_env *env,
5846 const struct dt_it *di, __u64 hash)
5848 struct osd_it_iam *it = (struct osd_it_iam *)di;
5850 return iam_it_load(&it->oi_it, hash);
5853 static const struct dt_index_operations osd_index_iam_ops = {
5854 .dio_lookup = osd_index_iam_lookup,
5855 .dio_declare_insert = osd_index_declare_iam_insert,
5856 .dio_insert = osd_index_iam_insert,
5857 .dio_declare_delete = osd_index_declare_iam_delete,
5858 .dio_delete = osd_index_iam_delete,
5860 .init = osd_it_iam_init,
5861 .fini = osd_it_iam_fini,
5862 .get = osd_it_iam_get,
5863 .put = osd_it_iam_put,
5864 .next = osd_it_iam_next,
5865 .key = osd_it_iam_key,
5866 .key_size = osd_it_iam_key_size,
5867 .rec = osd_it_iam_rec,
5868 .store = osd_it_iam_store,
5869 .load = osd_it_iam_load
5875 * Creates or initializes iterator context.
5877 * \retval struct osd_it_ea, iterator structure on success
5880 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
5881 struct dt_object *dt,
5884 struct osd_object *obj = osd_dt_obj(dt);
5885 struct osd_thread_info *info = osd_oti_get(env);
5886 struct osd_it_ea *oie;
5888 struct lu_object *lo = &dt->do_lu;
5889 struct dentry *obj_dentry;
5892 if (!dt_object_exists(dt) || obj->oo_destroyed)
5893 RETURN(ERR_PTR(-ENOENT));
5895 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
5897 RETURN(ERR_PTR(-ENOMEM));
5898 obj_dentry = &oie->oie_dentry;
5900 obj_dentry->d_inode = obj->oo_inode;
5901 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
5902 obj_dentry->d_name.hash = 0;
5904 oie->oie_rd_dirent = 0;
5905 oie->oie_it_dirent = 0;
5906 oie->oie_dirent = NULL;
5907 if (unlikely(!info->oti_it_ea_buf_used)) {
5908 oie->oie_buf = info->oti_it_ea_buf;
5909 info->oti_it_ea_buf_used = 1;
5911 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5912 if (oie->oie_buf == NULL)
5913 RETURN(ERR_PTR(-ENOMEM));
5917 file = &oie->oie_file;
5919 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
5920 if (attr & LUDA_64BITHASH)
5921 file->f_mode = FMODE_64BITHASH;
5923 file->f_mode = FMODE_32BITHASH;
5924 file->f_path.dentry = obj_dentry;
5925 file->f_mapping = obj->oo_inode->i_mapping;
5926 file->f_op = obj->oo_inode->i_fop;
5927 set_file_inode(file, obj->oo_inode);
5930 RETURN((struct dt_it *) oie);
5934 * Destroy or finishes iterator context.
5936 * \param di iterator structure to be destroyed
5938 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
5940 struct osd_thread_info *info = osd_oti_get(env);
5941 struct osd_it_ea *oie = (struct osd_it_ea *)di;
5942 struct osd_object *obj = oie->oie_obj;
5943 struct inode *inode = obj->oo_inode;
5946 oie->oie_file.f_op->release(inode, &oie->oie_file);
5947 osd_object_put(env, obj);
5948 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
5949 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5951 info->oti_it_ea_buf_used = 0;
5952 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
5957 * It position the iterator at given key, so that next lookup continues from
5958 * that key Or it is similar to dio_it->load() but based on a key,
5959 * rather than file position.
5961 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
5964 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
5966 static int osd_it_ea_get(const struct lu_env *env,
5967 struct dt_it *di, const struct dt_key *key)
5969 struct osd_it_ea *it = (struct osd_it_ea *)di;
5972 LASSERT(((const char *)key)[0] == '\0');
5973 it->oie_file.f_pos = 0;
5974 it->oie_rd_dirent = 0;
5975 it->oie_it_dirent = 0;
5976 it->oie_dirent = NULL;
5984 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
5988 struct osd_filldir_cbs {
5989 #ifdef HAVE_DIR_CONTEXT
5990 struct dir_context ctx;
5992 struct osd_it_ea *it;
5995 * It is called internally by ->readdir(). It fills the
5996 * iterator's in-memory data structure with required
5997 * information i.e. name, namelen, rec_size etc.
5999 * \param buf in which information to be filled in.
6000 * \param name name of the file in given dir
6002 * \retval 0 on success
6003 * \retval 1 on buffer full
6005 #ifdef HAVE_FILLDIR_USE_CTX
6006 static int osd_ldiskfs_filldir(struct dir_context *buf,
6008 static int osd_ldiskfs_filldir(void *buf,
6010 const char *name, int namelen,
6011 loff_t offset, __u64 ino, unsigned d_type)
6013 struct osd_it_ea *it =
6014 ((struct osd_filldir_cbs *)buf)->it;
6015 struct osd_object *obj = it->oie_obj;
6016 struct osd_it_ea_dirent *ent = it->oie_dirent;
6017 struct lu_fid *fid = &ent->oied_fid;
6018 struct osd_fid_pack *rec;
6021 /* this should never happen */
6022 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6023 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6027 if ((void *) ent - it->oie_buf + sizeof(*ent) + namelen >
6031 /* "." is just the object itself. */
6032 if (namelen == 1 && name[0] == '.') {
6033 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6034 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6035 rec = (struct osd_fid_pack*) (name + namelen + 1);
6036 if (osd_fid_unpack(fid, rec) != 0)
6041 d_type &= ~LDISKFS_DIRENT_LUFID;
6043 /* NOT export local root. */
6044 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6045 ino = obj->oo_inode->i_ino;
6046 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6049 ent->oied_ino = ino;
6050 ent->oied_off = offset;
6051 ent->oied_namelen = namelen;
6052 ent->oied_type = d_type;
6054 memcpy(ent->oied_name, name, namelen);
6056 it->oie_rd_dirent++;
6057 it->oie_dirent = (void *) ent + cfs_size_round(sizeof(*ent) + namelen);
6062 * Calls ->readdir() to load a directory entry at a time
6063 * and stored it in iterator's in-memory data structure.
6065 * \param di iterator's in memory structure
6067 * \retval 0 on success
6068 * \retval -ve on error
6069 * \retval +1 reach the end of entry
6071 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6072 const struct dt_it *di)
6074 struct osd_it_ea *it = (struct osd_it_ea *)di;
6075 struct osd_object *obj = it->oie_obj;
6076 struct inode *inode = obj->oo_inode;
6077 struct htree_lock *hlock = NULL;
6078 struct file *filp = &it->oie_file;
6080 struct osd_filldir_cbs buf = {
6081 #ifdef HAVE_DIR_CONTEXT
6082 .ctx.actor = osd_ldiskfs_filldir,
6088 it->oie_dirent = it->oie_buf;
6089 it->oie_rd_dirent = 0;
6091 if (obj->oo_hl_head != NULL) {
6092 hlock = osd_oti_get(env)->oti_hlock;
6093 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6094 inode, LDISKFS_HLOCK_READDIR);
6096 down_read(&obj->oo_ext_idx_sem);
6099 #ifdef HAVE_DIR_CONTEXT
6100 buf.ctx.pos = filp->f_pos;
6101 rc = inode->i_fop->iterate(filp, &buf.ctx);
6102 filp->f_pos = buf.ctx.pos;
6104 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6108 ldiskfs_htree_unlock(hlock);
6110 up_read(&obj->oo_ext_idx_sem);
6112 if (it->oie_rd_dirent == 0) {
6113 /*If it does not get any dirent, it means it has been reached
6114 *to the end of the dir */
6115 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6119 it->oie_dirent = it->oie_buf;
6120 it->oie_it_dirent = 1;
6127 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6128 * to load a directory entry at a time and stored it in
6129 * iterator's in-memory data structure.
6131 * \param di iterator's in memory structure
6133 * \retval +ve iterator reached to end
6134 * \retval 0 iterator not reached to end
6135 * \retval -ve on error
6137 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6139 struct osd_it_ea *it = (struct osd_it_ea *)di;
6144 if (it->oie_it_dirent < it->oie_rd_dirent) {
6146 (void *) it->oie_dirent +
6147 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6148 it->oie_dirent->oied_namelen);
6149 it->oie_it_dirent++;
6152 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6155 rc = osd_ldiskfs_it_fill(env, di);
6162 * Returns the key at current position from iterator's in memory structure.
6164 * \param di iterator's in memory structure
6166 * \retval key i.e. struct dt_key on success
6168 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6169 const struct dt_it *di)
6171 struct osd_it_ea *it = (struct osd_it_ea *)di;
6173 return (struct dt_key *)it->oie_dirent->oied_name;
6177 * Returns the key's size at current position from iterator's in memory structure.
6179 * \param di iterator's in memory structure
6181 * \retval key_size i.e. struct dt_key on success
6183 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6185 struct osd_it_ea *it = (struct osd_it_ea *)di;
6187 return it->oie_dirent->oied_namelen;
6190 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6192 if (LDISKFS_DIR_REC_LEN(de) >=
6193 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6200 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6201 unsigned blocksize, bool dotdot)
6204 return osd_dotdot_has_space(de);
6206 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6207 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6214 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6215 handle_t *jh, struct dentry *dentry,
6216 const struct lu_fid *fid, struct buffer_head *bh,
6217 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6220 struct inode *dir = dentry->d_parent->d_inode;
6221 struct inode *inode = dentry->d_inode;
6222 struct osd_fid_pack *rec;
6223 struct ldiskfs_dentry_param *ldp;
6224 int namelen = dentry->d_name.len;
6226 struct osd_thread_info *info = osd_oti_get(env);
6229 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6232 /* There is enough space to hold the FID-in-dirent. */
6233 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6234 rc = ldiskfs_journal_get_write_access(jh, bh);
6238 de->name[namelen] = 0;
6239 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6240 rec->fp_len = sizeof(struct lu_fid) + 1;
6241 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6242 de->file_type |= LDISKFS_DIRENT_LUFID;
6243 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6250 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6254 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6255 osd_get_ldiskfs_dirent_param(ldp, fid);
6256 dentry->d_fsdata = (void *)ldp;
6257 ll_vfs_dq_init(dir);
6258 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6259 /* It is too bad, we cannot reinsert the name entry back.
6260 * That means we lose it! */
6262 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6263 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6264 osd_ino2name(inode),
6265 dir->i_ino, dir->i_generation, namelen,
6266 dentry->d_name.name, PFID(fid), rc);
6272 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6273 struct osd_it_ea *it, struct lu_fid *fid,
6274 struct osd_inode_id *id, __u32 *attr)
6276 struct osd_thread_info *info = osd_oti_get(env);
6277 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6278 struct osd_device *dev = osd_obj2dev(obj);
6279 struct super_block *sb = osd_sb(dev);
6280 const char *devname = osd_name(dev);
6281 struct osd_it_ea_dirent *ent = it->oie_dirent;
6282 struct inode *dir = obj->oo_inode;
6283 struct htree_lock *hlock = NULL;
6284 struct buffer_head *bh = NULL;
6285 handle_t *jh = NULL;
6286 struct ldiskfs_dir_entry_2 *de;
6287 struct dentry *dentry;
6288 struct inode *inode;
6289 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6292 bool dotdot = false;
6296 if (ent->oied_name[0] == '.') {
6297 if (ent->oied_namelen == 1)
6300 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6304 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6305 inode = osd_iget(info, dev, id);
6306 if (IS_ERR(inode)) {
6307 rc = PTR_ERR(inode);
6308 if (rc == -ENOENT || rc == -ESTALE) {
6309 /* Maybe dangling name entry, or
6310 * corrupted directory entry. */
6311 *attr |= LUDA_UNKNOWN;
6314 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6315 "check_repair, dir = %lu/%u, name = %.*s, "
6316 "ino = %llu, rc = %d\n",
6317 devname, dir->i_ino, dir->i_generation,
6318 ent->oied_namelen, ent->oied_name,
6325 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6327 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6328 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6333 /* We need to ensure that the name entry is still valid.
6334 * Because it may be removed or renamed by other already.
6336 * The unlink or rename operation will start journal before PDO lock,
6337 * so to avoid deadlock, here we need to start journal handle before
6338 * related PDO lock also. But because we do not know whether there
6339 * will be something to be repaired before PDO lock, we just start
6340 * journal without conditions.
6342 * We may need to remove the name entry firstly, then insert back.
6343 * One credit is for user quota file update.
6344 * One credit is for group quota file update.
6345 * Two credits are for dirty inode. */
6346 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6347 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6349 if (dev->od_dirent_journal != 0) {
6352 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6355 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6356 "check_repair, dir = %lu/%u, credits = %d, "
6357 "name = %.*s, ino = %llu: rc = %d\n",
6358 devname, dir->i_ino, dir->i_generation, credits,
6359 ent->oied_namelen, ent->oied_name,
6362 GOTO(out_inode, rc);
6365 if (obj->oo_hl_head != NULL) {
6366 hlock = osd_oti_get(env)->oti_hlock;
6367 /* "0" means exclusive lock for the whole directory.
6368 * We need to prevent others access such name entry
6369 * during the delete + insert. Neither HLOCK_ADD nor
6370 * HLOCK_DEL cannot guarantee the atomicity. */
6371 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6373 down_write(&obj->oo_ext_idx_sem);
6376 if (obj->oo_hl_head != NULL) {
6377 hlock = osd_oti_get(env)->oti_hlock;
6378 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6379 LDISKFS_HLOCK_LOOKUP);
6381 down_read(&obj->oo_ext_idx_sem);
6385 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6386 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6387 *attr |= LUDA_IGNORE;
6392 /* For dotdot entry, if there is not enough space to hold the
6393 * FID-in-dirent, just keep them there. It only happens when the
6394 * device upgraded from 1.8 or restored from MDT file-level backup.
6395 * For the whole directory, only dotdot entry have no FID-in-dirent
6396 * and needs to get FID from LMA when readdir, it will not affect the
6397 * performance much. */
6398 if (dotdot && !osd_dotdot_has_space(de)) {
6399 *attr |= LUDA_UNKNOWN;
6405 if (lu_fid_eq(fid, &lma->lma_self_fid))
6408 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6409 struct lu_fid *tfid = &lma->lma_self_fid;
6411 if (likely(dotdot &&
6412 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6413 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6414 /* It must be REMOTE_PARENT_DIR and as the
6415 * 'dotdot' entry of remote directory */
6416 *attr |= LUDA_IGNORE;
6418 CDEBUG(D_LFSCK, "%s: expect remote agent "
6419 "parent directory, but got %.*s under "
6420 "dir = %lu/%u with the FID "DFID"\n",
6421 devname, ent->oied_namelen,
6422 ent->oied_name, dir->i_ino,
6423 dir->i_generation, PFID(tfid));
6425 *attr |= LUDA_UNKNOWN;
6432 if (!fid_is_zero(fid)) {
6433 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6435 if (rc == -ENOENT ||
6437 !(dev->od_scrub.os_file.sf_flags & SF_UPGRADE))) {
6438 /* linkEA does not recognize the dirent entry,
6439 * it may because the dirent entry corruption
6440 * and points to other's inode. */
6441 CDEBUG(D_LFSCK, "%s: the target inode does not "
6442 "recognize the dirent, dir = %lu/%u, "
6443 " name = %.*s, ino = %llu, "
6444 DFID": rc = %d\n", devname, dir->i_ino,
6445 dir->i_generation, ent->oied_namelen,
6446 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6447 *attr |= LUDA_UNKNOWN;
6452 if (rc && rc != -ENODATA) {
6453 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6454 "dir = %lu/%u, name = %.*s, ino = %llu, "
6455 DFID": rc = %d\n", devname, dir->i_ino,
6456 dir->i_generation, ent->oied_namelen,
6457 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6458 *attr |= LUDA_UNKNOWN;
6465 /* linkEA recognizes the dirent entry, the FID-in-LMA is
6466 * valid, trusted, in spite of fid_is_sane(fid) or not. */
6467 if (*attr & LUDA_VERIFY_DRYRUN) {
6468 *fid = lma->lma_self_fid;
6469 *attr |= LUDA_REPAIR;
6476 dev->od_dirent_journal = 1;
6477 if (hlock != NULL) {
6478 ldiskfs_htree_unlock(hlock);
6481 up_read(&obj->oo_ext_idx_sem);
6487 *fid = lma->lma_self_fid;
6489 /* Update or append the FID-in-dirent. */
6490 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6491 bh, de, hlock, dotdot);
6493 *attr |= LUDA_REPAIR;
6495 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6496 "the dirent, dir = %lu/%u, name = %.*s, "
6497 "ino = %llu, "DFID": rc = %d\n",
6498 devname, dir->i_ino, dir->i_generation,
6499 ent->oied_namelen, ent->oied_name,
6500 ent->oied_ino, PFID(fid), rc);
6502 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6503 if (*attr & LUDA_VERIFY_DRYRUN) {
6504 if (fid_is_sane(fid)) {
6505 *attr |= LUDA_REPAIR;
6506 } else if (dev->od_index == 0) {
6507 lu_igif_build(fid, inode->i_ino,
6508 inode->i_generation);
6509 *attr |= LUDA_UPGRADE;
6517 dev->od_dirent_journal = 1;
6518 if (hlock != NULL) {
6519 ldiskfs_htree_unlock(hlock);
6522 up_read(&obj->oo_ext_idx_sem);
6529 if (unlikely(fid_is_sane(fid))) {
6530 /* FID-in-dirent exists, but FID-in-LMA is lost.
6531 * Trust the FID-in-dirent, and add FID-in-LMA. */
6532 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
6534 *attr |= LUDA_REPAIR;
6536 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
6537 "update dirent, dir = %lu/%u, "
6538 "name = %.*s, ino = %llu, "
6540 devname, dir->i_ino, dir->i_generation,
6541 ent->oied_namelen, ent->oied_name,
6542 ent->oied_ino, PFID(fid), rc);
6543 } else if (dev->od_index == 0) {
6544 lu_igif_build(fid, inode->i_ino, inode->i_generation);
6545 /* It is probably IGIF object. Only aappend the
6546 * FID-in-dirent. OI scrub will process FID-in-LMA. */
6547 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6548 bh, de, hlock, dotdot);
6550 *attr |= LUDA_UPGRADE;
6552 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
6553 "after the dirent, dir = %lu/%u, "
6554 "name = %.*s, ino = %llu, "
6556 devname, dir->i_ino, dir->i_generation,
6557 ent->oied_namelen, ent->oied_name,
6558 ent->oied_ino, PFID(fid), rc);
6567 if (hlock != NULL) {
6568 ldiskfs_htree_unlock(hlock);
6570 if (dev->od_dirent_journal != 0)
6571 up_write(&obj->oo_ext_idx_sem);
6573 up_read(&obj->oo_ext_idx_sem);
6577 ldiskfs_journal_stop(jh);
6581 if (rc >= 0 && !dirty)
6582 dev->od_dirent_journal = 0;
6588 * Returns the value at current position from iterator's in memory structure.
6590 * \param di struct osd_it_ea, iterator's in memory structure
6591 * \param attr attr requested for dirent.
6592 * \param lde lustre dirent
6594 * \retval 0 no error and \param lde has correct lustre dirent.
6595 * \retval -ve on error
6597 static inline int osd_it_ea_rec(const struct lu_env *env,
6598 const struct dt_it *di,
6599 struct dt_rec *dtrec, __u32 attr)
6601 struct osd_it_ea *it = (struct osd_it_ea *)di;
6602 struct osd_object *obj = it->oie_obj;
6603 struct osd_device *dev = osd_obj2dev(obj);
6604 struct osd_thread_info *oti = osd_oti_get(env);
6605 struct osd_inode_id *id = &oti->oti_id;
6606 struct lu_fid *fid = &it->oie_dirent->oied_fid;
6607 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6608 __u32 ino = it->oie_dirent->oied_ino;
6612 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
6614 if (attr & LUDA_VERIFY) {
6615 if (unlikely(is_remote_parent_ino(dev, ino))) {
6616 attr |= LUDA_IGNORE;
6617 /* If the parent is on remote MDT, and there
6618 * is no FID-in-dirent, then we have to get
6619 * the parent FID from the linkEA. */
6620 if (!fid_is_sane(fid) &&
6621 it->oie_dirent->oied_namelen == 2 &&
6622 it->oie_dirent->oied_name[0] == '.' &&
6623 it->oie_dirent->oied_name[1] == '.')
6624 osd_get_pfid_from_linkea(env, obj, fid);
6626 rc = osd_dirent_check_repair(env, obj, it, fid, id,
6630 if (!fid_is_sane(fid))
6631 attr |= LUDA_UNKNOWN;
6633 attr &= ~LU_DIRENT_ATTRS_MASK;
6634 if (!fid_is_sane(fid)) {
6635 bool is_dotdot = false;
6636 if (it->oie_dirent->oied_namelen == 2 &&
6637 it->oie_dirent->oied_name[0] == '.' &&
6638 it->oie_dirent->oied_name[1] == '.')
6640 /* If the parent is on remote MDT, and there
6641 * is no FID-in-dirent, then we have to get
6642 * the parent FID from the linkEA. */
6643 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
6644 rc = osd_get_pfid_from_linkea(env, obj, fid);
6646 if (is_dotdot == false &&
6647 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
6650 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6653 osd_id_gen(id, ino, OSD_OII_NOGEN);
6657 /* Pack the entry anyway, at least the offset is right. */
6658 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
6659 it->oie_dirent->oied_name,
6660 it->oie_dirent->oied_namelen,
6661 it->oie_dirent->oied_type, attr);
6666 if (osd_remote_fid(env, dev, fid))
6669 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
6670 osd_add_oi_cache(oti, dev, id, fid);
6672 RETURN(rc > 0 ? 0 : rc);
6676 * Returns the record size size at current position.
6678 * This function will return record(lu_dirent) size in bytes.
6680 * \param[in] env execution environment
6681 * \param[in] di iterator's in memory structure
6682 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
6683 * calculate the lu_dirent size.
6685 * \retval record size(in bytes & in memory) of the current lu_dirent
6688 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
6691 struct osd_it_ea *it = (struct osd_it_ea *)di;
6693 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
6697 * Returns a cookie for current position of the iterator head, so that
6698 * user can use this cookie to load/start the iterator next time.
6700 * \param di iterator's in memory structure
6702 * \retval cookie for current position, on success
6704 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
6706 struct osd_it_ea *it = (struct osd_it_ea *)di;
6708 return it->oie_dirent->oied_off;
6712 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6713 * to load a directory entry at a time and stored it i inn,
6714 * in iterator's in-memory data structure.
6716 * \param di struct osd_it_ea, iterator's in memory structure
6718 * \retval +ve on success
6719 * \retval -ve on error
6721 static int osd_it_ea_load(const struct lu_env *env,
6722 const struct dt_it *di, __u64 hash)
6724 struct osd_it_ea *it = (struct osd_it_ea *)di;
6728 it->oie_file.f_pos = hash;
6730 rc = osd_ldiskfs_it_fill(env, di);
6741 * Index lookup function for interoperability mode (b11826).
6743 * \param key, key i.e. file name to be searched
6745 * \retval +ve, on success
6746 * \retval -ve, on error
6748 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
6749 struct dt_rec *rec, const struct dt_key *key)
6751 struct osd_object *obj = osd_dt_obj(dt);
6756 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
6757 LINVRNT(osd_invariant(obj));
6759 rc = osd_ea_lookup_rec(env, obj, rec, key);
6766 * Index and Iterator operations for interoperability
6767 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
6769 static const struct dt_index_operations osd_index_ea_ops = {
6770 .dio_lookup = osd_index_ea_lookup,
6771 .dio_declare_insert = osd_index_declare_ea_insert,
6772 .dio_insert = osd_index_ea_insert,
6773 .dio_declare_delete = osd_index_declare_ea_delete,
6774 .dio_delete = osd_index_ea_delete,
6776 .init = osd_it_ea_init,
6777 .fini = osd_it_ea_fini,
6778 .get = osd_it_ea_get,
6779 .put = osd_it_ea_put,
6780 .next = osd_it_ea_next,
6781 .key = osd_it_ea_key,
6782 .key_size = osd_it_ea_key_size,
6783 .rec = osd_it_ea_rec,
6784 .rec_size = osd_it_ea_rec_size,
6785 .store = osd_it_ea_store,
6786 .load = osd_it_ea_load
6790 static void *osd_key_init(const struct lu_context *ctx,
6791 struct lu_context_key *key)
6793 struct osd_thread_info *info;
6795 OBD_ALLOC_PTR(info);
6797 return ERR_PTR(-ENOMEM);
6799 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6800 if (info->oti_it_ea_buf == NULL)
6803 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
6805 info->oti_hlock = ldiskfs_htree_lock_alloc();
6806 if (info->oti_hlock == NULL)
6812 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6815 return ERR_PTR(-ENOMEM);
6818 static void osd_key_fini(const struct lu_context *ctx,
6819 struct lu_context_key *key, void* data)
6821 struct osd_thread_info *info = data;
6822 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
6823 struct osd_idmap_cache *idc = info->oti_ins_cache;
6825 if (info->oti_inode != NULL)
6827 if (info->oti_hlock != NULL)
6828 ldiskfs_htree_lock_free(info->oti_hlock);
6829 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6830 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
6831 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
6832 lu_buf_free(&info->oti_big_buf);
6834 LASSERT(info->oti_ins_cache_size > 0);
6835 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
6836 info->oti_ins_cache = NULL;
6837 info->oti_ins_cache_size = 0;
6842 static void osd_key_exit(const struct lu_context *ctx,
6843 struct lu_context_key *key, void *data)
6845 struct osd_thread_info *info = data;
6847 LASSERT(info->oti_r_locks == 0);
6848 LASSERT(info->oti_w_locks == 0);
6849 LASSERT(info->oti_txns == 0);
6852 /* type constructor/destructor: osd_type_init, osd_type_fini */
6853 LU_TYPE_INIT_FINI(osd, &osd_key);
6855 struct lu_context_key osd_key = {
6856 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
6857 .lct_init = osd_key_init,
6858 .lct_fini = osd_key_fini,
6859 .lct_exit = osd_key_exit
6863 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
6864 const char *name, struct lu_device *next)
6866 struct osd_device *osd = osd_dev(d);
6868 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname))
6869 >= sizeof(osd->od_svname))
6871 return osd_procfs_init(osd, name);
6874 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
6876 struct seq_server_site *ss = osd_seq_site(osd);
6880 if (osd->od_is_ost || osd->od_cl_seq != NULL)
6883 if (unlikely(ss == NULL))
6886 OBD_ALLOC_PTR(osd->od_cl_seq);
6887 if (osd->od_cl_seq == NULL)
6890 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
6891 osd->od_svname, ss->ss_server_seq);
6893 OBD_FREE_PTR(osd->od_cl_seq);
6894 osd->od_cl_seq = NULL;
6898 if (ss->ss_node_id == 0) {
6899 /* If the OSD on the sequence controller(MDT0), then allocate
6900 * sequence here, otherwise allocate sequence after connected
6901 * to MDT0 (see mdt_register_lwp_callback()). */
6902 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
6903 &osd->od_cl_seq->lcs_space, env);
6909 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
6911 if (osd->od_cl_seq == NULL)
6914 seq_client_fini(osd->od_cl_seq);
6915 OBD_FREE_PTR(osd->od_cl_seq);
6916 osd->od_cl_seq = NULL;
6919 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
6923 /* shutdown quota slave instance associated with the device */
6924 if (o->od_quota_slave != NULL) {
6925 qsd_fini(env, o->od_quota_slave);
6926 o->od_quota_slave = NULL;
6929 osd_fid_fini(env, o);
6934 static void osd_umount(const struct lu_env *env, struct osd_device *o)
6938 if (o->od_mnt != NULL) {
6939 shrink_dcache_sb(osd_sb(o));
6940 osd_sync(env, &o->od_dt_dev);
6949 static int osd_mount(const struct lu_env *env,
6950 struct osd_device *o, struct lustre_cfg *cfg)
6952 const char *name = lustre_cfg_string(cfg, 0);
6953 const char *dev = lustre_cfg_string(cfg, 1);
6955 unsigned long page, s_flags, lmd_flags = 0;
6956 struct page *__page;
6957 struct file_system_type *type;
6958 char *options = NULL;
6960 struct osd_thread_info *info = osd_oti_get(env);
6961 struct lu_fid *fid = &info->oti_fid;
6962 struct inode *inode;
6963 int rc = 0, force_over_512tb = 0;
6966 if (o->od_mnt != NULL)
6969 if (strlen(dev) >= sizeof(o->od_mntdev))
6971 strcpy(o->od_mntdev, dev);
6973 str = lustre_cfg_string(cfg, 2);
6974 s_flags = simple_strtoul(str, NULL, 0);
6975 str = strstr(str, ":");
6977 lmd_flags = simple_strtoul(str + 1, NULL, 0);
6978 opts = lustre_cfg_string(cfg, 3);
6980 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
6981 CERROR("%s: device %s extents feature is not guaranteed to "
6982 "work on big-endian systems. Use \"bigendian_extents\" "
6983 "mount option to override.\n", name, dev);
6987 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
6988 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
6989 CWARN("force_over_128tb option is deprecated. "
6990 "Filesystems less than 512TB can be created without any "
6991 "force options. Use force_over_512tb option for "
6992 "filesystems greater than 512TB.\n");
6995 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
6996 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
6997 CWARN("force_over_256tb option is deprecated. "
6998 "Filesystems less than 512TB can be created without any "
6999 "force options. Use force_over_512tb option for "
7000 "filesystems greater than 512TB.\n");
7004 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
7005 force_over_512tb = 1;
7007 __page = alloc_page(GFP_KERNEL);
7009 GOTO(out, rc = -ENOMEM);
7010 page = (unsigned long)page_address(__page);
7011 options = (char *)page;
7014 /* strip out the options for back compatiblity */
7015 static char *sout[] = {
7022 /* strip out option we processed in osd */
7023 "bigendian_extents",
7029 strcat(options, opts);
7030 for (rc = 0, str = options; sout[rc]; ) {
7031 char *op = strstr(str, sout[rc]);
7037 if (op == options || *(op - 1) == ',') {
7038 str = op + strlen(sout[rc]);
7039 if (*str == ',' || *str == '\0') {
7040 *str == ',' ? str++ : str;
7041 memmove(op, str, strlen(str) + 1);
7044 for (str = op; *str != ',' && *str != '\0'; str++)
7048 strncat(options, "user_xattr,acl", 14);
7051 /* Glom up mount options */
7052 if (*options != '\0')
7053 strcat(options, ",");
7054 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7056 type = get_fs_type("ldiskfs");
7058 CERROR("%s: cannot find ldiskfs module\n", name);
7059 GOTO(out, rc = -ENODEV);
7062 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7063 module_put(type->owner);
7065 if (IS_ERR(o->od_mnt)) {
7066 rc = PTR_ERR(o->od_mnt);
7068 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7072 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7073 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7074 force_over_512tb == 0) {
7075 CERROR("%s: device %s LDISKFS does not support filesystems "
7076 "greater than 512TB and can cause data corruption. "
7077 "Use \"force_over_512tb\" mount option to override.\n",
7079 GOTO(out_mnt, rc = -EINVAL);
7082 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7083 if (priv_dev_set_rdonly) {
7084 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
7085 o->od_dt_dev.dd_rdonly = 1;
7086 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
7089 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7092 GOTO(out_mnt, rc = -EOPNOTSUPP);
7094 } else if (priv_dev_check_rdonly &&
7095 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
7096 CERROR("%s: underlying device %s is marked as "
7097 "read-only. Setup failed\n", name, dev);
7099 GOTO(out_mnt, rc = -EROFS);
7102 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7103 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7104 GOTO(out_mnt, rc = -EINVAL);
7107 #ifdef LDISKFS_MOUNT_DIRDATA
7108 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7109 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7110 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7111 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7112 "enabling the dirdata feature. If you do not want to "
7113 "downgrade to Lustre-1.x again, you can enable it via "
7114 "'tune2fs -O dirdata device'\n", name, dev);
7116 inode = osd_sb(o)->s_root->d_inode;
7117 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7118 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7120 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7124 if (lmd_flags & LMD_FLG_NOSCRUB)
7135 __free_page(__page);
7140 static struct lu_device *osd_device_fini(const struct lu_env *env,
7141 struct lu_device *d)
7143 struct osd_device *o = osd_dev(d);
7146 osd_shutdown(env, o);
7148 osd_scrub_cleanup(env, o);
7149 osd_obj_map_fini(o);
7155 static int osd_device_init0(const struct lu_env *env,
7156 struct osd_device *o,
7157 struct lustre_cfg *cfg)
7159 struct lu_device *l = osd2lu_dev(o);
7160 struct osd_thread_info *info;
7164 /* if the module was re-loaded, env can loose its keys */
7165 rc = lu_env_refill((struct lu_env *) env);
7168 info = osd_oti_get(env);
7171 l->ld_ops = &osd_lu_ops;
7172 o->od_dt_dev.dd_ops = &osd_dt_ops;
7174 spin_lock_init(&o->od_osfs_lock);
7175 mutex_init(&o->od_otable_mutex);
7176 INIT_LIST_HEAD(&o->od_orphan_list);
7178 o->od_read_cache = 1;
7179 o->od_writethrough_cache = 1;
7180 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7182 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7183 sizeof(o->od_svname));
7184 if (cplen >= sizeof(o->od_svname)) {
7189 o->od_index = -1; /* -1 means index is invalid */
7190 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7191 if (rc == LDD_F_SV_TYPE_OST)
7194 o->od_full_scrub_ratio = OFSR_DEFAULT;
7195 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7196 rc = osd_mount(env, o, cfg);
7200 rc = osd_obj_map_init(env, o);
7204 rc = lu_site_init(&o->od_site, l);
7206 GOTO(out_compat, rc);
7207 o->od_site.ls_bottom_dev = l;
7209 rc = lu_site_init_finish(&o->od_site);
7213 INIT_LIST_HEAD(&o->od_ios_list);
7214 /* setup scrub, including OI files initialization */
7215 rc = osd_scrub_setup(env, o);
7219 rc = osd_procfs_init(o, o->od_svname);
7221 CERROR("%s: can't initialize procfs: rc = %d\n",
7223 GOTO(out_scrub, rc);
7226 LASSERT(l->ld_site->ls_linkage.next != NULL);
7227 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7229 /* initialize quota slave instance */
7230 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7232 if (IS_ERR(o->od_quota_slave)) {
7233 rc = PTR_ERR(o->od_quota_slave);
7234 o->od_quota_slave = NULL;
7235 GOTO(out_procfs, rc);
7243 osd_scrub_cleanup(env, o);
7245 lu_site_fini(&o->od_site);
7247 osd_obj_map_fini(o);
7254 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7255 struct lu_device_type *t,
7256 struct lustre_cfg *cfg)
7258 struct osd_device *o;
7263 return ERR_PTR(-ENOMEM);
7265 rc = dt_device_init(&o->od_dt_dev, t);
7267 /* Because the ctx might be revived in dt_device_init,
7268 * refill the env here */
7269 lu_env_refill((struct lu_env *)env);
7270 rc = osd_device_init0(env, o, cfg);
7272 dt_device_fini(&o->od_dt_dev);
7275 if (unlikely(rc != 0))
7278 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7281 static struct lu_device *osd_device_free(const struct lu_env *env,
7282 struct lu_device *d)
7284 struct osd_device *o = osd_dev(d);
7287 /* XXX: make osd top device in order to release reference */
7288 d->ld_site->ls_top_dev = d;
7289 lu_site_purge(env, d->ld_site, -1);
7290 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7291 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7292 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7294 lu_site_fini(&o->od_site);
7295 dt_device_fini(&o->od_dt_dev);
7300 static int osd_process_config(const struct lu_env *env,
7301 struct lu_device *d, struct lustre_cfg *cfg)
7303 struct osd_device *o = osd_dev(d);
7307 switch (cfg->lcfg_command) {
7309 rc = osd_mount(env, o, cfg);
7312 lu_dev_del_linkage(d->ld_site, d);
7313 rc = osd_shutdown(env, o);
7316 LASSERT(&o->od_dt_dev);
7317 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7318 cfg, &o->od_dt_dev);
7319 if (rc > 0 || rc == -ENOSYS) {
7320 rc = class_process_proc_param(PARAM_OST,
7321 lprocfs_osd_obd_vars,
7322 cfg, &o->od_dt_dev);
7334 static int osd_recovery_complete(const struct lu_env *env,
7335 struct lu_device *d)
7337 struct osd_device *osd = osd_dev(d);
7341 if (osd->od_quota_slave == NULL)
7344 /* start qsd instance on recovery completion, this notifies the quota
7345 * slave code that we are about to process new requests now */
7346 rc = qsd_start(env, osd->od_quota_slave);
7351 * we use exports to track all osd users
7353 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7354 struct obd_device *obd, struct obd_uuid *cluuid,
7355 struct obd_connect_data *data, void *localdata)
7357 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7358 struct lustre_handle conn;
7362 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7364 rc = class_connect(&conn, obd, cluuid);
7368 *exp = class_conn2export(&conn);
7370 spin_lock(&osd->od_osfs_lock);
7372 spin_unlock(&osd->od_osfs_lock);
7378 * once last export (we don't count self-export) disappeared
7379 * osd can be released
7381 static int osd_obd_disconnect(struct obd_export *exp)
7383 struct obd_device *obd = exp->exp_obd;
7384 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7385 int rc, release = 0;
7388 /* Only disconnect the underlying layers on the final disconnect. */
7389 spin_lock(&osd->od_osfs_lock);
7391 if (osd->od_connects == 0)
7393 spin_unlock(&osd->od_osfs_lock);
7395 rc = class_disconnect(exp); /* bz 9811 */
7397 if (rc == 0 && release)
7398 class_manual_cleanup(obd);
7402 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7403 struct lu_device *dev)
7405 struct osd_device *osd = osd_dev(dev);
7406 struct lr_server_data *lsd =
7407 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7411 if (osd->od_quota_slave != NULL) {
7412 /* set up quota slave objects */
7413 result = qsd_prepare(env, osd->od_quota_slave);
7418 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7419 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7420 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7421 osd->od_index_in_idif = 1;
7423 osd->od_index_in_idif = 0;
7424 result = osd_register_proc_index_in_idif(osd);
7429 osd->od_index_in_idif = 1;
7433 result = osd_fid_init(env, osd);
7438 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7439 struct lu_fid *fid, struct md_op_data *op_data)
7441 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7443 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7446 static const struct lu_object_operations osd_lu_obj_ops = {
7447 .loo_object_init = osd_object_init,
7448 .loo_object_delete = osd_object_delete,
7449 .loo_object_release = osd_object_release,
7450 .loo_object_free = osd_object_free,
7451 .loo_object_print = osd_object_print,
7452 .loo_object_invariant = osd_object_invariant
7455 const struct lu_device_operations osd_lu_ops = {
7456 .ldo_object_alloc = osd_object_alloc,
7457 .ldo_process_config = osd_process_config,
7458 .ldo_recovery_complete = osd_recovery_complete,
7459 .ldo_prepare = osd_prepare,
7462 static const struct lu_device_type_operations osd_device_type_ops = {
7463 .ldto_init = osd_type_init,
7464 .ldto_fini = osd_type_fini,
7466 .ldto_start = osd_type_start,
7467 .ldto_stop = osd_type_stop,
7469 .ldto_device_alloc = osd_device_alloc,
7470 .ldto_device_free = osd_device_free,
7472 .ldto_device_init = osd_device_init,
7473 .ldto_device_fini = osd_device_fini
7476 static struct lu_device_type osd_device_type = {
7477 .ldt_tags = LU_DEVICE_DT,
7478 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
7479 .ldt_ops = &osd_device_type_ops,
7480 .ldt_ctx_tags = LCT_LOCAL,
7483 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
7485 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7486 struct super_block *sb = osd_sb(osd);
7488 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
7492 * lprocfs legacy support.
7494 static struct obd_ops osd_obd_device_ops = {
7495 .o_owner = THIS_MODULE,
7496 .o_connect = osd_obd_connect,
7497 .o_disconnect = osd_obd_disconnect,
7498 .o_fid_alloc = osd_fid_alloc,
7499 .o_health_check = osd_health_check,
7502 static int __init osd_init(void)
7506 LASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
7507 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
7508 /* please, try to keep osd_thread_info smaller than a page */
7509 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
7514 rc = lu_kmem_init(ldiskfs_caches);
7518 #ifdef CONFIG_KALLSYMS
7519 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
7520 priv_dev_check_rdonly = (void *)kallsyms_lookup_name("dev_check_rdonly");
7521 /* Clear readonly is unused at this time */
7522 /*priv_dev_clear_rdonly = (void *)kallsyms_lookup_name("dev_clear_rdonly");*/
7525 rc = class_register_type(&osd_obd_device_ops, NULL, true,
7526 lprocfs_osd_module_vars,
7527 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
7529 lu_kmem_fini(ldiskfs_caches);
7533 static void __exit osd_exit(void)
7535 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
7536 lu_kmem_fini(ldiskfs_caches);
7539 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7540 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
7541 MODULE_VERSION(LUSTRE_VERSION_STRING);
7542 MODULE_LICENSE("GPL");
7544 module_init(osd_init);
7545 module_exit(osd_exit);