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 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2626 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2627 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2630 ll_vfs_dq_init(inode);
2632 if (attr->la_valid & LA_UID)
2633 iattr.ia_valid |= ATTR_UID;
2634 if (attr->la_valid & LA_GID)
2635 iattr.ia_valid |= ATTR_GID;
2636 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2637 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2639 rc = ll_vfs_dq_transfer(inode, &iattr);
2641 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2642 "enforcement enabled on the ldiskfs "
2643 "filesystem?\n", inode->i_sb->s_id, rc);
2648 /* Handle project id transfer here properly */
2649 if (attr->la_valid & LA_PROJID &&
2650 attr->la_projid != i_projid_read(inode)) {
2651 #ifdef HAVE_PROJECT_QUOTA
2652 rc = ldiskfs_transfer_project(inode, attr->la_projid);
2657 CERROR("%s: quota transfer failed: rc = %d. Is project "
2658 "enforcement enabled on the ldiskfs "
2659 "filesystem?\n", inode->i_sb->s_id, rc);
2666 static int osd_attr_set(const struct lu_env *env,
2667 struct dt_object *dt,
2668 const struct lu_attr *attr,
2669 struct thandle *handle)
2671 struct osd_object *obj = osd_dt_obj(dt);
2672 struct inode *inode;
2675 if (!dt_object_exists(dt))
2678 LASSERT(handle != NULL);
2679 LASSERT(!dt_object_remote(dt));
2680 LASSERT(osd_invariant(obj));
2682 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2684 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2685 struct osd_thread_info *oti = osd_oti_get(env);
2686 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2687 struct lu_fid *fid1 = &oti->oti_fid;
2688 struct osd_inode_id *id = &oti->oti_id;
2689 struct iam_path_descr *ipd;
2690 struct iam_container *bag;
2691 struct osd_thandle *oh;
2694 fid_cpu_to_be(fid1, fid0);
2695 memset(id, 1, sizeof(*id));
2696 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2697 fid0)->oi_dir.od_container;
2698 ipd = osd_idx_ipd_get(env, bag);
2699 if (unlikely(ipd == NULL))
2702 oh = container_of0(handle, struct osd_thandle, ot_super);
2703 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2704 (const struct iam_rec *)id, ipd);
2705 osd_ipd_put(env, bag, ipd);
2706 return(rc > 0 ? 0 : rc);
2709 inode = obj->oo_inode;
2711 rc = osd_quota_transfer(inode, attr);
2715 spin_lock(&obj->oo_guard);
2716 rc = osd_inode_setattr(env, inode, attr);
2717 spin_unlock(&obj->oo_guard);
2721 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2723 if (!(attr->la_valid & LA_FLAGS))
2726 /* Let's check if there are extra flags need to be set into LMA */
2727 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2728 struct osd_thread_info *info = osd_oti_get(env);
2729 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2731 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2732 &info->oti_ost_attrs);
2736 lma->lma_incompat |=
2737 lustre_to_lma_flags(attr->la_flags);
2738 lustre_lma_swab(lma);
2739 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2740 lma, sizeof(*lma), XATTR_REPLACE);
2742 struct osd_device *osd = osd_obj2dev(obj);
2744 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2745 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2746 lma->lma_incompat, rc);
2749 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2751 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2754 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2759 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2760 struct osd_object *obj,
2761 const char *name, const int namelen)
2763 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2766 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2767 umode_t mode, struct dt_allocation_hint *hint,
2771 struct osd_device *osd = osd_obj2dev(obj);
2772 struct osd_thandle *oth;
2773 struct dt_object *parent = NULL;
2774 struct inode *inode;
2776 LINVRNT(osd_invariant(obj));
2777 LASSERT(obj->oo_inode == NULL);
2778 LASSERT(obj->oo_hl_head == NULL);
2780 if (S_ISDIR(mode) && ldiskfs_pdo) {
2781 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2782 if (obj->oo_hl_head == NULL)
2786 oth = container_of(th, struct osd_thandle, ot_super);
2787 LASSERT(oth->ot_handle->h_transaction != NULL);
2789 if (hint != NULL && hint->dah_parent != NULL &&
2790 !dt_object_remote(hint->dah_parent))
2791 parent = hint->dah_parent;
2793 inode = ldiskfs_create_inode(oth->ot_handle,
2794 parent ? osd_dt_obj(parent)->oo_inode :
2795 osd_sb(osd)->s_root->d_inode,
2797 if (!IS_ERR(inode)) {
2798 /* Do not update file c/mtime in ldiskfs. */
2799 inode->i_flags |= S_NOCMTIME;
2801 /* For new created object, it must be consistent,
2802 * and it is unnecessary to scrub against it. */
2803 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2805 obj->oo_inode = inode;
2808 if (obj->oo_hl_head != NULL) {
2809 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2810 obj->oo_hl_head = NULL;
2812 result = PTR_ERR(inode);
2814 LINVRNT(osd_invariant(obj));
2822 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2823 struct lu_attr *attr,
2824 struct dt_allocation_hint *hint,
2825 struct dt_object_format *dof,
2829 struct osd_thandle *oth;
2830 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2832 LASSERT(S_ISDIR(attr->la_mode));
2834 oth = container_of(th, struct osd_thandle, ot_super);
2835 LASSERT(oth->ot_handle->h_transaction != NULL);
2836 result = osd_mkfile(info, obj, mode, hint, th);
2841 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2842 struct lu_attr *attr,
2843 struct dt_allocation_hint *hint,
2844 struct dt_object_format *dof,
2848 struct osd_thandle *oth;
2849 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2851 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2853 LASSERT(S_ISREG(attr->la_mode));
2855 oth = container_of(th, struct osd_thandle, ot_super);
2856 LASSERT(oth->ot_handle->h_transaction != NULL);
2858 result = osd_mkfile(info, obj, mode, hint, th);
2860 LASSERT(obj->oo_inode != NULL);
2861 if (feat->dif_flags & DT_IND_VARKEY)
2862 result = iam_lvar_create(obj->oo_inode,
2863 feat->dif_keysize_max,
2865 feat->dif_recsize_max,
2868 result = iam_lfix_create(obj->oo_inode,
2869 feat->dif_keysize_max,
2871 feat->dif_recsize_max,
2878 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2879 struct lu_attr *attr,
2880 struct dt_allocation_hint *hint,
2881 struct dt_object_format *dof,
2884 LASSERT(S_ISREG(attr->la_mode));
2885 return osd_mkfile(info, obj, (attr->la_mode &
2886 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2889 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2890 struct lu_attr *attr,
2891 struct dt_allocation_hint *hint,
2892 struct dt_object_format *dof,
2895 LASSERT(S_ISLNK(attr->la_mode));
2896 return osd_mkfile(info, obj, (attr->la_mode &
2897 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2900 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2901 struct lu_attr *attr,
2902 struct dt_allocation_hint *hint,
2903 struct dt_object_format *dof,
2906 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2909 LINVRNT(osd_invariant(obj));
2910 LASSERT(obj->oo_inode == NULL);
2911 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2912 S_ISFIFO(mode) || S_ISSOCK(mode));
2914 result = osd_mkfile(info, obj, mode, hint, th);
2916 LASSERT(obj->oo_inode != NULL);
2918 * This inode should be marked dirty for i_rdev. Currently
2919 * that is done in the osd_attr_init().
2921 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2924 LINVRNT(osd_invariant(obj));
2928 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2930 struct dt_allocation_hint *hint,
2931 struct dt_object_format *dof,
2934 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2936 osd_obj_type_f result;
2952 result = osd_mk_index;
2963 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2964 struct dt_object *parent, struct dt_object *child,
2969 ah->dah_parent = parent;
2970 ah->dah_mode = child_mode;
2972 if (parent != NULL && !dt_object_remote(parent)) {
2973 /* will help to find FID->ino at dt_insert("..") */
2974 struct osd_object *pobj = osd_dt_obj(parent);
2975 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2979 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2980 struct lu_attr *attr, struct dt_object_format *dof)
2982 struct inode *inode = obj->oo_inode;
2983 __u64 valid = attr->la_valid;
2986 attr->la_valid &= ~(LA_TYPE | LA_MODE);
2988 if (dof->dof_type != DFT_NODE)
2989 attr->la_valid &= ~LA_RDEV;
2990 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2991 attr->la_valid &= ~LA_ATIME;
2992 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2993 attr->la_valid &= ~LA_CTIME;
2994 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2995 attr->la_valid &= ~LA_MTIME;
2997 result = osd_quota_transfer(inode, attr);
3001 if (attr->la_valid != 0) {
3002 result = osd_inode_setattr(info->oti_env, inode, attr);
3004 * The osd_inode_setattr() should always succeed here. The
3005 * only error that could be returned is EDQUOT when we are
3006 * trying to change the UID or GID of the inode. However, this
3007 * should not happen since quota enforcement is no longer
3008 * enabled on ldiskfs (lquota takes care of it).
3010 LASSERTF(result == 0, "%d\n", result);
3011 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3014 attr->la_valid = valid;
3018 * Helper function for osd_create()
3020 * \retval 0, on success
3022 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3023 struct lu_attr *attr, struct dt_allocation_hint *hint,
3024 struct dt_object_format *dof, struct thandle *th)
3029 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3031 /* we drop umask so that permissions we pass are not affected */
3032 umask = current->fs->umask;
3033 current->fs->umask = 0;
3035 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3037 if (likely(obj->oo_inode != NULL)) {
3038 LASSERT(obj->oo_inode->i_state & I_NEW);
3040 /* Unlock the inode before attr initialization to avoid
3041 * unnecessary dqget operations. LU-6378 */
3042 unlock_new_inode(obj->oo_inode);
3045 if (likely(result == 0)) {
3046 osd_attr_init(info, obj, attr, dof);
3047 osd_object_init0(obj);
3050 /* restore previous umask value */
3051 current->fs->umask = umask;
3053 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3059 * Helper function for osd_create()
3061 * \retval 0, on success
3063 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3064 const struct lu_fid *fid, struct thandle *th)
3066 struct osd_thread_info *info = osd_oti_get(env);
3067 struct osd_inode_id *id = &info->oti_id;
3068 struct osd_device *osd = osd_obj2dev(obj);
3069 struct osd_thandle *oh;
3072 LASSERT(obj->oo_inode != NULL);
3074 oh = container_of0(th, struct osd_thandle, ot_super);
3075 LASSERT(oh->ot_handle);
3076 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3078 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3079 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3080 OI_CHECK_FLD, NULL);
3081 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3086 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3087 u64 seq, struct lu_seq_range *range)
3089 struct seq_server_site *ss = osd_seq_site(osd);
3091 if (fid_seq_is_idif(seq)) {
3092 fld_range_set_ost(range);
3093 range->lsr_index = idif_ost_idx(seq);
3097 if (!fid_seq_in_fldb(seq)) {
3098 fld_range_set_mdt(range);
3100 /* FIXME: If ss is NULL, it suppose not get lsr_index
3102 range->lsr_index = ss->ss_node_id;
3106 LASSERT(ss != NULL);
3107 fld_range_set_any(range);
3108 /* OSD will only do local fld lookup */
3109 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3112 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3113 struct lu_attr *attr,
3114 struct dt_allocation_hint *hint,
3115 struct dt_object_format *dof,
3116 struct thandle *handle)
3118 struct osd_thandle *oh;
3122 LASSERT(handle != NULL);
3124 oh = container_of0(handle, struct osd_thandle, ot_super);
3125 LASSERT(oh->ot_handle == NULL);
3127 /* EA object consumes more credits than regular object: osd_mk_index
3128 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3129 * leaf_node, could involves the block, block bitmap, groups, GDT
3130 * change for each block, so add 4 * 2 credits in that case. */
3131 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3132 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3133 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3134 /* Reuse idle OI block may cause additional one OI block
3136 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3137 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3142 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3143 attr->la_projid, 1, oh, osd_dt_obj(dt),
3144 NULL, OSD_QID_INODE);
3148 /* will help to find FID->ino mapping at dt_insert() */
3149 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3156 * Called to destroy on-disk representation of the object
3158 * Concurrency: must be locked
3160 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3163 struct osd_object *obj = osd_dt_obj(dt);
3164 struct inode *inode = obj->oo_inode;
3165 struct osd_thandle *oh;
3172 oh = container_of0(th, struct osd_thandle, ot_super);
3173 LASSERT(oh->ot_handle == NULL);
3175 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3176 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3177 /* Recycle idle OI leaf may cause additional three OI blocks
3179 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3180 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3181 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3182 /* one less inode */
3183 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3184 i_projid_read(inode), -1, oh, obj, NULL,
3188 /* data to be truncated */
3189 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3190 i_projid_read(inode), 0, oh, obj, NULL,
3195 /* will help to find FID->ino when this object is being
3196 * added to PENDING/ */
3197 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3202 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3205 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3206 struct osd_object *obj = osd_dt_obj(dt);
3207 struct inode *inode = obj->oo_inode;
3208 struct osd_device *osd = osd_obj2dev(obj);
3209 struct osd_thandle *oh;
3213 oh = container_of0(th, struct osd_thandle, ot_super);
3214 LASSERT(oh->ot_handle);
3216 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3218 if (unlikely(fid_is_acct(fid)))
3221 if (S_ISDIR(inode->i_mode)) {
3222 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3223 inode->i_nlink == 2);
3224 /* it will check/delete the inode from remote parent,
3225 * how to optimize it? unlink performance impaction XXX */
3226 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3228 CERROR("%s: delete inode "DFID": rc = %d\n",
3229 osd_name(osd), PFID(fid), result);
3231 spin_lock(&obj->oo_guard);
3233 spin_unlock(&obj->oo_guard);
3234 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3237 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3239 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3241 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3242 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3243 oh->ot_handle, OI_CHECK_FLD);
3245 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3246 /* XXX: add to ext3 orphan list */
3247 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3249 /* not needed in the cache anymore */
3250 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3251 obj->oo_destroyed = 1;
3257 * Put the fid into lustre_mdt_attrs, and then place the structure
3258 * inode's ea. This fid should not be altered during the life time
3261 * \retval +ve, on success
3262 * \retval -ve, on error
3264 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3266 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3267 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3269 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3270 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3274 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3277 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3280 lustre_loa_init(loa, fid, compat, incompat);
3281 lustre_loa_swab(loa, false);
3283 /* For the OST device with 256 bytes inode size by default,
3284 * the PFID EA will be stored together with LMA EA to avoid
3285 * performance trouble. Otherwise the PFID EA can be stored
3286 * independently. LU-8998 */
3287 if ((compat & LMAC_FID_ON_OST) &&
3288 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3289 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3290 sizeof(*loa), XATTR_CREATE);
3292 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3293 sizeof(*lma), XATTR_CREATE);
3294 /* LMA may already exist, but we need to check that all the
3295 * desired compat/incompat flags have been added. */
3296 if (unlikely(rc == -EEXIST)) {
3297 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3298 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3302 if (rc < sizeof(*lma))
3305 lustre_loa_swab(loa, true);
3306 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3307 ((compat == 0 && incompat == 0) ||
3308 (!(~lma->lma_compat & compat) &&
3309 !(~lma->lma_incompat & incompat))))
3312 lma->lma_self_fid = *fid;
3313 lma->lma_compat |= compat;
3314 lma->lma_incompat |= incompat;
3315 if (rc == sizeof(*lma)) {
3316 lustre_lma_swab(lma);
3317 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3318 sizeof(*lma), XATTR_REPLACE);
3320 lustre_loa_swab(loa, false);
3321 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3322 sizeof(*loa), XATTR_REPLACE);
3330 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3331 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3332 * To have compatilibility with 1.8 ldiskfs driver we need to have
3333 * magic number at start of fid data.
3334 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3337 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3338 const struct lu_fid *fid)
3340 if (!fid_is_namespace_visible(fid) ||
3341 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3342 param->edp_magic = 0;
3346 param->edp_magic = LDISKFS_LUFID_MAGIC;
3347 param->edp_len = sizeof(struct lu_fid) + 1;
3348 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3352 * Try to read the fid from inode ea into dt_rec.
3354 * \param fid object fid.
3356 * \retval 0 on success
3358 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3359 __u32 ino, struct lu_fid *fid,
3360 struct osd_inode_id *id)
3362 struct osd_thread_info *info = osd_oti_get(env);
3363 struct inode *inode;
3366 osd_id_gen(id, ino, OSD_OII_NOGEN);
3367 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3369 RETURN(PTR_ERR(inode));
3375 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3377 struct inode *parent_dir,
3378 const struct lu_fid *dot_fid,
3379 const struct lu_fid *dot_dot_fid,
3380 struct osd_thandle *oth)
3382 struct ldiskfs_dentry_param *dot_ldp;
3383 struct ldiskfs_dentry_param *dot_dot_ldp;
3384 __u32 saved_nlink = dir->i_nlink;
3387 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3388 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3390 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3391 dot_ldp->edp_magic = 0;
3393 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3394 dir, dot_ldp, dot_dot_ldp);
3395 /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3396 * the subseqent ref_add() will increase the dir->i_nlink
3397 * as 3. That is incorrect for new created directory.
3399 * It looks like hack, because we want to make the OSD API
3400 * to be order-independent for new created directory object
3401 * between dt_insert(..) and ref_add() operations.
3403 * Here, we only restore the in-RAM dir-inode's nlink attr,
3404 * becuase if the nlink attr is not 2, then there will be
3405 * ref_add() called following the dt_insert(..), such call
3406 * will make both the in-RAM and on-disk dir-inode's nlink
3407 * attr to be set as 2. LU-7447 */
3408 set_nlink(dir, saved_nlink);
3413 * Create an local agent inode for remote entry
3415 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3416 struct osd_device *osd,
3417 struct osd_object *pobj,
3418 const struct lu_fid *fid,
3422 struct osd_thread_info *info = osd_oti_get(env);
3423 struct inode *local;
3424 struct osd_thandle *oh;
3429 oh = container_of(th, struct osd_thandle, ot_super);
3430 LASSERT(oh->ot_handle->h_transaction != NULL);
3432 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3433 if (IS_ERR(local)) {
3434 CERROR("%s: create local error %d\n", osd_name(osd),
3435 (int)PTR_ERR(local));
3439 /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3440 * correct gid on remote file, not agent here */
3441 local->i_gid = current_fsgid();
3442 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3443 unlock_new_inode(local);
3445 /* Agent inode should not have project ID */
3446 #ifdef HAVE_PROJECT_QUOTA
3447 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3448 i_projid_read(pobj->oo_inode) != 0) {
3449 rc = ldiskfs_transfer_project(local, 0);
3451 CERROR("%s: quota transfer failed: rc = %d. Is project "
3452 "quota enforcement enabled on the ldiskfs "
3453 "filesystem?\n", local->i_sb->s_id, rc);
3454 RETURN(ERR_PTR(rc));
3458 /* Set special LMA flag for local agent inode */
3459 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3461 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3462 osd_name(osd), PFID(fid), rc);
3463 RETURN(ERR_PTR(rc));
3469 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3470 lu_object_fid(&pobj->oo_dt.do_lu),
3473 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3474 osd_name(osd), PFID(fid), rc);
3475 RETURN(ERR_PTR(rc));
3482 * when direntry is deleted, we have to take care of possible agent inode
3483 * referenced by that. unfortunately we can't do this at that point:
3484 * iget() within a running transaction leads to deadlock and we better do
3485 * not call that every delete declaration to save performance. so we put
3486 * a potention agent inode on a list and process that once the transaction
3487 * is over. Notice it's not any worse in terms of real orphans as regular
3488 * object destroy doesn't put inodes on the on-disk orphan list. this should
3489 * be addressed separately
3491 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3492 struct osd_thandle *oh,
3495 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3496 struct osd_obj_orphan *oor;
3503 oor->oor_env = (struct lu_env *)env;
3504 spin_lock(&osd->od_osfs_lock);
3505 list_add(&oor->oor_list, &osd->od_orphan_list);
3506 spin_unlock(&osd->od_osfs_lock);
3508 oh->ot_remove_agents = 1;
3514 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3515 struct osd_device *osd)
3517 struct osd_thread_info *info = osd_oti_get(env);
3518 struct osd_obj_orphan *oor, *tmp;
3519 struct osd_inode_id id;
3520 struct list_head list;
3521 struct inode *inode;
3526 INIT_LIST_HEAD(&list);
3528 spin_lock(&osd->od_osfs_lock);
3529 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3530 if (oor->oor_env == env) {
3531 list_del(&oor->oor_list);
3532 list_add(&oor->oor_list, &list);
3535 spin_unlock(&osd->od_osfs_lock);
3537 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3541 list_del(&oor->oor_list);
3544 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3545 inode = osd_iget_fid(info, osd, &id, &fid);
3549 if (!osd_remote_fid(env, osd, &fid)) {
3554 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3556 mark_inode_dirty(inode);
3557 ldiskfs_journal_stop(jh);
3565 * OSD layer object create function for OST objects (b=11826).
3567 * The FID is inserted into inode xattr here.
3569 * \retval 0, on success
3570 * \retval -ve, on error
3572 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3573 struct lu_attr *attr, struct dt_allocation_hint *hint,
3574 struct dt_object_format *dof, struct thandle *th)
3576 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3577 struct osd_object *obj = osd_dt_obj(dt);
3578 struct osd_thread_info *info = osd_oti_get(env);
3579 int result, on_ost = 0;
3583 if (dt_object_exists(dt))
3586 LINVRNT(osd_invariant(obj));
3587 LASSERT(!dt_object_remote(dt));
3588 LASSERT(osd_is_write_locked(env, obj));
3589 LASSERT(th != NULL);
3591 if (unlikely(fid_is_acct(fid)))
3592 /* Quota files can't be created from the kernel any more,
3593 * 'tune2fs -O quota' will take care of creating them */
3596 result = __osd_create(info, obj, attr, hint, dof, th);
3598 if (fid_is_idif(fid) &&
3599 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3600 struct lu_fid *tfid = &info->oti_fid;
3601 struct ost_id *oi = &info->oti_ostid;
3603 fid_to_ostid(fid, oi);
3604 ostid_to_fid(tfid, oi, 0);
3606 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3607 LMAC_FID_ON_OST, 0);
3609 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3611 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3612 on_ost ? LMAC_FID_ON_OST : 0,
3615 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3616 obj->oo_dt.do_body_ops = &osd_body_ops;
3620 result = __osd_oi_insert(env, obj, fid, th);
3622 /* a small optimization - dt_insert() isn't usually applied
3623 * to OST objects, so we don't need to cache OI mapping for
3625 if (result == 0 && on_ost == 0) {
3626 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3627 result = osd_idc_find_and_init(env, osd, obj);
3628 LASSERT(result == 0);
3631 LASSERT(ergo(result == 0,
3632 dt_object_exists(dt) && !dt_object_remote(dt)));
3633 LINVRNT(osd_invariant(obj));
3637 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3638 struct thandle *handle)
3640 struct osd_thandle *oh;
3642 /* it's possible that object doesn't exist yet */
3643 LASSERT(handle != NULL);
3645 oh = container_of0(handle, struct osd_thandle, ot_super);
3646 LASSERT(oh->ot_handle == NULL);
3648 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3649 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3651 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
3657 * Concurrency: @dt is write locked.
3659 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3662 struct osd_object *obj = osd_dt_obj(dt);
3663 struct inode *inode = obj->oo_inode;
3664 struct osd_thandle *oh;
3667 if (!dt_object_exists(dt) || obj->oo_destroyed)
3670 LINVRNT(osd_invariant(obj));
3671 LASSERT(!dt_object_remote(dt));
3672 LASSERT(osd_is_write_locked(env, obj));
3673 LASSERT(th != NULL);
3675 oh = container_of0(th, struct osd_thandle, ot_super);
3676 LASSERT(oh->ot_handle != NULL);
3678 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3680 CDEBUG(D_INODE, DFID" increase nlink %d\n",
3681 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3683 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3684 * (65000) subdirectories by storing "1" in i_nlink if the link count
3685 * would otherwise overflow. Directory tranversal tools understand
3686 * that (st_nlink == 1) indicates that the filesystem dose not track
3687 * hard links count on the directory, and will not abort subdirectory
3688 * scanning early once (st_nlink - 2) subdirs have been found.
3690 * This also has to properly handle the case of inodes with nlink == 0
3691 * in case they are being linked into the PENDING directory
3693 spin_lock(&obj->oo_guard);
3694 if (unlikely(inode->i_nlink == 0))
3695 /* inc_nlink from 0 may cause WARN_ON */
3696 set_nlink(inode, 1);
3698 ldiskfs_inc_count(oh->ot_handle, inode);
3699 if (!S_ISDIR(inode->i_mode))
3700 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3702 spin_unlock(&obj->oo_guard);
3704 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3705 LINVRNT(osd_invariant(obj));
3707 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3712 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
3713 struct thandle *handle)
3715 struct osd_thandle *oh;
3717 if (!dt_object_exists(dt))
3720 LASSERT(!dt_object_remote(dt));
3721 LASSERT(handle != NULL);
3723 oh = container_of0(handle, struct osd_thandle, ot_super);
3724 LASSERT(oh->ot_handle == NULL);
3726 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3727 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3733 * Concurrency: @dt is write locked.
3735 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
3738 struct osd_object *obj = osd_dt_obj(dt);
3739 struct inode *inode = obj->oo_inode;
3740 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3741 struct osd_thandle *oh;
3743 if (!dt_object_exists(dt))
3746 LINVRNT(osd_invariant(obj));
3747 LASSERT(!dt_object_remote(dt));
3748 LASSERT(osd_is_write_locked(env, obj));
3749 LASSERT(th != NULL);
3751 oh = container_of0(th, struct osd_thandle, ot_super);
3752 LASSERT(oh->ot_handle != NULL);
3754 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3756 spin_lock(&obj->oo_guard);
3757 /* That can be result of upgrade from old Lustre version and
3758 * applied only to local files. Just skip this ref_del call.
3759 * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3760 if (inode->i_nlink == 0) {
3761 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3762 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3763 ", maybe an upgraded file? (LU-3915)\n",
3764 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3765 spin_unlock(&obj->oo_guard);
3769 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3770 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3772 ldiskfs_dec_count(oh->ot_handle, inode);
3773 spin_unlock(&obj->oo_guard);
3775 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3776 LINVRNT(osd_invariant(obj));
3778 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3784 * Concurrency: @dt is read locked.
3786 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3787 struct lu_buf *buf, const char *name)
3789 struct osd_object *obj = osd_dt_obj(dt);
3790 struct inode *inode = obj->oo_inode;
3791 struct osd_thread_info *info = osd_oti_get(env);
3792 struct dentry *dentry = &info->oti_obj_dentry;
3793 bool cache_xattr = false;
3798 /* version get is not real XATTR but uses xattr API */
3799 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3800 dt_obj_version_t *ver = buf->lb_buf;
3802 /* for version we are just using xattr API but change inode
3804 if (buf->lb_len == 0)
3805 return sizeof(dt_obj_version_t);
3807 if (buf->lb_len < sizeof(dt_obj_version_t))
3810 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
3811 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3813 *ver = LDISKFS_I(inode)->i_fs_version;
3815 return sizeof(dt_obj_version_t);
3818 if (!dt_object_exists(dt))
3821 LASSERT(!dt_object_remote(dt));
3822 LASSERT(inode->i_op != NULL);
3823 #ifdef HAVE_IOP_XATTR
3824 LASSERT(inode->i_op->getxattr != NULL);
3827 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
3828 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
3832 rc = osd_oxc_get(obj, name, buf);
3837 rc = __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3838 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
3839 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3840 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3841 struct filter_fid *ff;
3842 struct ost_layout *ol;
3844 if (!osd_dev(dt->do_lu.lo_dev)->od_is_ost)
3847 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
3851 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
3857 if (buf->lb_len == 0 || !buf->lb_buf)
3860 if (buf->lb_len < rc)
3864 ol = &ff->ff_layout;
3865 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
3866 PFID_STRIPE_IDX_BITS);
3867 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
3868 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
3869 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
3870 if (lma->lma_compat & LMAC_COMP_INFO) {
3871 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
3872 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
3873 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
3875 ol->ol_comp_start = 0;
3876 ol->ol_comp_end = 0;
3883 if (rc == -ENOENT || rc == -ENODATA)
3884 osd_oxc_add(obj, name, NULL, 0);
3885 else if (rc > 0 && buf->lb_buf != NULL)
3886 osd_oxc_add(obj, name, buf->lb_buf, rc);
3892 static int osd_declare_xattr_set(const struct lu_env *env,
3893 struct dt_object *dt,
3894 const struct lu_buf *buf, const char *name,
3895 int fl, struct thandle *handle)
3897 struct osd_thandle *oh;
3899 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3901 LASSERT(handle != NULL);
3903 oh = container_of0(handle, struct osd_thandle, ot_super);
3904 LASSERT(oh->ot_handle == NULL);
3906 if (strcmp(name, XATTR_NAME_LMA) == 0) {
3907 /* For non-upgrading case, the LMA is set first and
3908 * usually fit inode. But for upgrade case, the LMA
3909 * may be in another separated EA block. */
3910 if (dt_object_exists(dt)) {
3911 if (fl == LU_XATTR_REPLACE)
3916 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3918 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
3919 /* We may need to delete the old PFID EA. */
3920 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
3921 if (fl == LU_XATTR_REPLACE)
3928 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
3933 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3934 /* learn xattr size from osd_xattr_get if
3935 attribute has not been read yet */
3936 buflen = __osd_xattr_get(
3937 osd_dt_obj(dt)->oo_inode,
3938 &osd_oti_get(env)->oti_obj_dentry,
3943 buflen = buf->lb_len;
3946 if (buflen > sb->s_blocksize) {
3947 credits += osd_calc_bkmap_credits(
3949 (buflen + sb->s_blocksize - 1) >>
3950 sb->s_blocksize_bits);
3954 * xattr set may involve inode quota change, reserve credits for
3955 * dquot_initialize()
3957 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3960 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3966 * Concurrency: @dt is write locked.
3968 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3969 const struct lu_buf *buf, const char *name, int fl,
3970 struct thandle *handle)
3972 struct osd_object *obj = osd_dt_obj(dt);
3973 struct inode *inode = obj->oo_inode;
3974 struct osd_thread_info *info = osd_oti_get(env);
3975 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3976 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3985 /* version set is not real XATTR */
3986 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3987 dt_obj_version_t *version = buf->lb_buf;
3989 /* for version we are just using xattr API but change inode
3991 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
3993 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
3994 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3996 LDISKFS_I(inode)->i_fs_version = *version;
3997 /* Version is set after all inode operations are finished,
3998 * so we should mark it dirty here */
3999 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
4004 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4005 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4008 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4009 if (fl & LU_XATTR_REPLACE)
4010 fs_flags |= XATTR_REPLACE;
4012 if (fl & LU_XATTR_CREATE)
4013 fs_flags |= XATTR_CREATE;
4015 /* For the OST device with 256 bytes inode size by default,
4016 * the PFID EA will be stored together with LMA EA to avoid
4017 * performance trouble. Otherwise the PFID EA can be stored
4018 * independently. LU-8998 */
4019 if (strcmp(name, XATTR_NAME_FID) == 0 &&
4020 LDISKFS_INODE_SIZE(inode->i_sb) <= 256) {
4021 struct dentry *dentry = &info->oti_obj_dentry;
4022 struct filter_fid *ff;
4023 struct ost_layout *ol;
4026 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4029 ol = &ff->ff_layout;
4030 /* Old client does not send stripe information, store
4031 * the PFID EA on disk directly. */
4032 if (buf->lb_len == sizeof(struct lu_fid) ||
4033 ol->ol_stripe_size == 0) {
4034 len = sizeof(struct lu_fid);
4038 if (buf->lb_len != sizeof(*ff))
4041 rc = osd_get_lma(info, inode, dentry, loa);
4042 if (unlikely(rc == -ENODATA)) {
4043 /* Usually for upgarding from old device */
4044 lustre_loa_init(loa, lu_object_fid(&dt->do_lu),
4045 LMAC_FID_ON_OST, 0);
4053 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4054 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4055 PFID_STRIPE_IDX_BITS;
4056 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4057 lma->lma_compat |= LMAC_STRIPE_INFO;
4058 if (ol->ol_comp_id != 0) {
4059 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4060 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4061 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4062 lma->lma_compat |= LMAC_COMP_INFO;
4065 lustre_loa_swab(loa, false);
4067 /* Remove old PFID EA entry firstly. */
4068 ll_vfs_dq_init(inode);
4069 rc = inode->i_op->removexattr(dentry, name);
4070 if (rc && rc != -ENODATA)
4073 /* Store the PFID EA inside the LMA EA. */
4074 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4078 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4079 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4083 lma->lma_incompat |= LMAI_STRIPED;
4084 lustre_lma_swab(lma);
4085 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4086 sizeof(*lma), XATTR_REPLACE);
4092 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4093 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4096 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4097 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4098 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4104 * Concurrency: @dt is read locked.
4106 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4107 const struct lu_buf *buf)
4109 struct osd_object *obj = osd_dt_obj(dt);
4110 struct inode *inode = obj->oo_inode;
4111 struct osd_thread_info *info = osd_oti_get(env);
4112 struct dentry *dentry = &info->oti_obj_dentry;
4114 if (!dt_object_exists(dt))
4117 LASSERT(!dt_object_remote(dt));
4118 LASSERT(inode->i_op != NULL);
4119 LASSERT(inode->i_op->listxattr != NULL);
4121 dentry->d_inode = inode;
4122 dentry->d_sb = inode->i_sb;
4123 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4126 static int osd_declare_xattr_del(const struct lu_env *env,
4127 struct dt_object *dt, const char *name,
4128 struct thandle *handle)
4130 struct osd_thandle *oh;
4131 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4133 LASSERT(!dt_object_remote(dt));
4134 LASSERT(handle != NULL);
4136 oh = container_of0(handle, struct osd_thandle, ot_super);
4137 LASSERT(oh->ot_handle == NULL);
4139 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4140 osd_dto_credits_noquota[DTO_XATTR_SET]);
4142 * xattr del may involve inode quota change, reserve credits for
4143 * dquot_initialize()
4145 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4151 * Concurrency: @dt is write locked.
4153 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4154 const char *name, struct thandle *handle)
4156 struct osd_object *obj = osd_dt_obj(dt);
4157 struct inode *inode = obj->oo_inode;
4158 struct osd_thread_info *info = osd_oti_get(env);
4159 struct dentry *dentry = &info->oti_obj_dentry;
4162 if (!dt_object_exists(dt))
4165 LASSERT(!dt_object_remote(dt));
4166 LASSERT(inode->i_op != NULL);
4167 LASSERT(handle != NULL);
4168 #ifdef HAVE_IOP_XATTR
4169 LASSERT(inode->i_op->removexattr != NULL);
4172 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4174 ll_vfs_dq_init(inode);
4175 dentry->d_inode = inode;
4176 dentry->d_sb = inode->i_sb;
4177 rc = inode->i_op->removexattr(dentry, name);
4178 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
4179 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4181 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4183 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4184 &info->oti_ost_attrs);
4186 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
4191 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4192 lustre_lma_swab(lma);
4193 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4194 sizeof(*lma), XATTR_REPLACE);
4199 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4202 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4203 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4204 osd_oxc_del(obj, name);
4209 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4210 __u64 start, __u64 end)
4212 struct osd_object *obj = osd_dt_obj(dt);
4213 struct inode *inode = obj->oo_inode;
4214 struct osd_thread_info *info = osd_oti_get(env);
4215 struct dentry *dentry = &info->oti_obj_dentry;
4216 struct file *file = &info->oti_file;
4221 dentry->d_inode = inode;
4222 dentry->d_sb = inode->i_sb;
4223 file->f_path.dentry = dentry;
4224 file->f_mapping = inode->i_mapping;
4225 file->f_op = inode->i_fop;
4226 set_file_inode(file, inode);
4228 rc = ll_vfs_fsync_range(file, start, end, 0);
4233 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4242 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4243 const struct dt_index_features *feat)
4245 struct iam_descr *descr;
4247 if (osd_object_is_root(o))
4248 return feat == &dt_directory_features;
4250 LASSERT(o->oo_dir != NULL);
4252 descr = o->oo_dir->od_container.ic_descr;
4253 if (feat == &dt_directory_features) {
4254 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4260 feat->dif_keysize_min <= descr->id_key_size &&
4261 descr->id_key_size <= feat->dif_keysize_max &&
4262 feat->dif_recsize_min <= descr->id_rec_size &&
4263 descr->id_rec_size <= feat->dif_recsize_max &&
4264 !(feat->dif_flags & (DT_IND_VARKEY |
4265 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4266 ergo(feat->dif_flags & DT_IND_UPDATE,
4267 1 /* XXX check that object (and file system) is
4272 static int osd_iam_container_init(const struct lu_env *env,
4273 struct osd_object *obj,
4274 struct osd_directory *dir)
4276 struct iam_container *bag = &dir->od_container;
4279 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4283 result = iam_container_setup(bag);
4285 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4287 iam_container_fini(bag);
4294 * Concurrency: no external locking is necessary.
4296 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4297 const struct dt_index_features *feat)
4301 struct osd_object *obj = osd_dt_obj(dt);
4303 LINVRNT(osd_invariant(obj));
4305 if (osd_object_is_root(obj)) {
4306 dt->do_index_ops = &osd_index_ea_ops;
4308 } else if (feat == &dt_directory_features) {
4309 dt->do_index_ops = &osd_index_ea_ops;
4310 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4315 } else if (unlikely(feat == &dt_otable_features)) {
4316 dt->do_index_ops = &osd_otable_ops;
4318 } else if (unlikely(feat == &dt_acct_features)) {
4319 dt->do_index_ops = &osd_acct_index_ops;
4322 } else if (!osd_has_index(obj)) {
4323 struct osd_directory *dir;
4328 spin_lock(&obj->oo_guard);
4329 if (obj->oo_dir == NULL)
4333 * Concurrent thread allocated container data.
4336 spin_unlock(&obj->oo_guard);
4338 * Now, that we have container data, serialize its
4341 down_write(&obj->oo_ext_idx_sem);
4343 * recheck under lock.
4345 if (!osd_has_index(obj))
4346 result = osd_iam_container_init(env, obj,
4350 up_write(&obj->oo_ext_idx_sem);
4358 if (result == 0 && skip_iam == 0) {
4359 if (!osd_iam_index_probe(env, obj, feat))
4362 LINVRNT(osd_invariant(obj));
4367 static int osd_otable_it_attr_get(const struct lu_env *env,
4368 struct dt_object *dt,
4369 struct lu_attr *attr)
4375 static const struct dt_object_operations osd_obj_ops = {
4376 .do_read_lock = osd_read_lock,
4377 .do_write_lock = osd_write_lock,
4378 .do_read_unlock = osd_read_unlock,
4379 .do_write_unlock = osd_write_unlock,
4380 .do_write_locked = osd_write_locked,
4381 .do_attr_get = osd_attr_get,
4382 .do_declare_attr_set = osd_declare_attr_set,
4383 .do_attr_set = osd_attr_set,
4384 .do_ah_init = osd_ah_init,
4385 .do_declare_create = osd_declare_create,
4386 .do_create = osd_create,
4387 .do_declare_destroy = osd_declare_destroy,
4388 .do_destroy = osd_destroy,
4389 .do_index_try = osd_index_try,
4390 .do_declare_ref_add = osd_declare_ref_add,
4391 .do_ref_add = osd_ref_add,
4392 .do_declare_ref_del = osd_declare_ref_del,
4393 .do_ref_del = osd_ref_del,
4394 .do_xattr_get = osd_xattr_get,
4395 .do_declare_xattr_set = osd_declare_xattr_set,
4396 .do_xattr_set = osd_xattr_set,
4397 .do_declare_xattr_del = osd_declare_xattr_del,
4398 .do_xattr_del = osd_xattr_del,
4399 .do_xattr_list = osd_xattr_list,
4400 .do_object_sync = osd_object_sync,
4401 .do_invalidate = osd_invalidate,
4404 static const struct dt_object_operations osd_obj_otable_it_ops = {
4405 .do_attr_get = osd_otable_it_attr_get,
4406 .do_index_try = osd_index_try,
4409 static int osd_index_declare_iam_delete(const struct lu_env *env,
4410 struct dt_object *dt,
4411 const struct dt_key *key,
4412 struct thandle *handle)
4414 struct osd_thandle *oh;
4416 oh = container_of0(handle, struct osd_thandle, ot_super);
4417 LASSERT(oh->ot_handle == NULL);
4419 /* Recycle may cause additional three blocks to be changed. */
4420 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4421 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4427 * delete a (key, value) pair from index \a dt specified by \a key
4429 * \param dt osd index object
4430 * \param key key for index
4431 * \param rec record reference
4432 * \param handle transaction handler
4435 * \retval -ve failure
4437 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4438 const struct dt_key *key,
4439 struct thandle *handle)
4441 struct osd_thread_info *oti = osd_oti_get(env);
4442 struct osd_object *obj = osd_dt_obj(dt);
4443 struct osd_thandle *oh;
4444 struct iam_path_descr *ipd;
4445 struct iam_container *bag = &obj->oo_dir->od_container;
4449 if (!dt_object_exists(dt))
4452 LINVRNT(osd_invariant(obj));
4453 LASSERT(!dt_object_remote(dt));
4454 LASSERT(bag->ic_object == obj->oo_inode);
4455 LASSERT(handle != NULL);
4457 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4459 ipd = osd_idx_ipd_get(env, bag);
4460 if (unlikely(ipd == NULL))
4463 oh = container_of0(handle, struct osd_thandle, ot_super);
4464 LASSERT(oh->ot_handle != NULL);
4465 LASSERT(oh->ot_handle->h_transaction != NULL);
4467 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4468 /* swab quota uid/gid provided by caller */
4469 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4470 key = (const struct dt_key *)&oti->oti_quota_id;
4473 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4474 osd_ipd_put(env, bag, ipd);
4475 LINVRNT(osd_invariant(obj));
4476 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4480 static int osd_index_declare_ea_delete(const struct lu_env *env,
4481 struct dt_object *dt,
4482 const struct dt_key *key,
4483 struct thandle *handle)
4485 struct osd_thandle *oh;
4486 struct inode *inode;
4490 LASSERT(!dt_object_remote(dt));
4491 LASSERT(handle != NULL);
4493 oh = container_of0(handle, struct osd_thandle, ot_super);
4494 LASSERT(oh->ot_handle == NULL);
4496 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4497 if (key != NULL && unlikely(strcmp((char *)key, dotdot) == 0)) {
4498 /* '..' to a remote object has a local representative */
4499 credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
4500 /* to reset LMAI_REMOTE_PARENT */
4503 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4505 inode = osd_dt_obj(dt)->oo_inode;
4509 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4510 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
4515 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
4518 struct osd_fid_pack *rec;
4521 if (de->file_type & LDISKFS_DIRENT_LUFID) {
4522 rec = (struct osd_fid_pack *) (de->name + de->name_len + 1);
4523 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
4524 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
4530 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
4531 const struct lu_fid *fid)
4533 struct seq_server_site *ss = osd_seq_site(osd);
4536 /* FID seqs not in FLDB, must be local seq */
4537 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
4540 /* If FLD is not being initialized yet, it only happens during the
4541 * initialization, likely during mgs initialization, and we assume
4542 * this is local FID. */
4543 if (ss == NULL || ss->ss_server_fld == NULL)
4546 /* Only check the local FLDB here */
4547 if (osd_seq_exists(env, osd, fid_seq(fid)))
4553 static void osd_take_care_of_agent(const struct lu_env *env,
4554 struct osd_device *osd,
4555 struct osd_thandle *oh,
4556 struct ldiskfs_dir_entry_2 *de)
4558 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4559 struct osd_idmap_cache *idc;
4560 int rc, schedule = 0;
4562 LASSERT(de != NULL);
4564 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
4565 if (likely(rc == 0)) {
4566 idc = osd_idc_find_or_init(env, osd, fid);
4567 if (IS_ERR(idc) || idc->oic_remote)
4569 } else if (rc == -ENODATA) {
4570 /* can't get FID, postpone to the end of the
4571 * transaction when iget() is safe */
4574 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
4577 osd_schedule_agent_inode_removal(env, oh,
4578 le32_to_cpu(de->inode));
4582 * Index delete function for interoperability mode (b11826).
4583 * It will remove the directory entry added by osd_index_ea_insert().
4584 * This entry is needed to maintain name->fid mapping.
4586 * \param key, key i.e. file entry to be deleted
4588 * \retval 0, on success
4589 * \retval -ve, on error
4591 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
4592 const struct dt_key *key, struct thandle *handle)
4594 struct osd_object *obj = osd_dt_obj(dt);
4595 struct inode *dir = obj->oo_inode;
4596 struct dentry *dentry;
4597 struct osd_thandle *oh;
4598 struct ldiskfs_dir_entry_2 *de = NULL;
4599 struct buffer_head *bh;
4600 struct htree_lock *hlock = NULL;
4601 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4602 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4606 if (!dt_object_exists(dt))
4609 LINVRNT(osd_invariant(obj));
4610 LASSERT(!dt_object_remote(dt));
4611 LASSERT(handle != NULL);
4613 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4615 oh = container_of(handle, struct osd_thandle, ot_super);
4616 LASSERT(oh->ot_handle != NULL);
4617 LASSERT(oh->ot_handle->h_transaction != NULL);
4619 ll_vfs_dq_init(dir);
4620 dentry = osd_child_dentry_get(env, obj,
4621 (char *)key, strlen((char *)key));
4623 if (obj->oo_hl_head != NULL) {
4624 hlock = osd_oti_get(env)->oti_hlock;
4625 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
4626 dir, LDISKFS_HLOCK_DEL);
4628 down_write(&obj->oo_ext_idx_sem);
4631 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
4633 /* If this is not the ".." entry, it might be a remote DNE
4634 * entry and we need to check if the FID is for a remote
4635 * MDT. If the FID is not in the directory entry (e.g.
4636 * upgraded 1.8 filesystem without dirdata enabled) then
4637 * we need to get the FID from the LMA. For a remote directory
4638 * there HAS to be an LMA, it cannot be an IGIF inode in this
4641 * Delete the entry before the agent inode in order to
4642 * simplify error handling. At worst an error after deleting
4643 * the entry first might leak the agent inode afterward. The
4644 * reverse would need filesystem abort in case of error deleting
4645 * the entry after the agent had been removed, or leave a
4646 * dangling entry pointing at a random inode. */
4647 if (strcmp((char *)key, dotdot) != 0)
4648 osd_take_care_of_agent(env, osd, oh, de);
4649 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
4655 ldiskfs_htree_unlock(hlock);
4657 up_write(&obj->oo_ext_idx_sem);
4662 /* For inode on the remote MDT, .. will point to
4663 * /Agent directory, Check whether it needs to delete
4664 * from agent directory */
4665 if (unlikely(strcmp((char *)key, dotdot) == 0)) {
4668 ret = osd_delete_from_remote_parent(env, osd_obj2dev(obj),
4671 /* Sigh, the entry has been deleted, and
4672 * it is not easy to revert it back, so
4673 * let's keep this error private, and let
4674 * LFSCK fix it. XXX */
4675 CERROR("%s: delete remote parent "DFID": rc = %d\n",
4676 osd_name(osd), PFID(fid), ret);
4679 LASSERT(osd_invariant(obj));
4680 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4685 * Lookup index for \a key and copy record to \a rec.
4687 * \param dt osd index object
4688 * \param key key for index
4689 * \param rec record reference
4691 * \retval +ve success : exact mach
4692 * \retval 0 return record with key not greater than \a key
4693 * \retval -ve failure
4695 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
4696 struct dt_rec *rec, const struct dt_key *key)
4698 struct osd_object *obj = osd_dt_obj(dt);
4699 struct iam_path_descr *ipd;
4700 struct iam_container *bag = &obj->oo_dir->od_container;
4701 struct osd_thread_info *oti = osd_oti_get(env);
4702 struct iam_iterator *it = &oti->oti_idx_it;
4703 struct iam_rec *iam_rec;
4707 if (!dt_object_exists(dt))
4710 LASSERT(osd_invariant(obj));
4711 LASSERT(!dt_object_remote(dt));
4712 LASSERT(bag->ic_object == obj->oo_inode);
4714 ipd = osd_idx_ipd_get(env, bag);
4718 /* got ipd now we can start iterator. */
4719 iam_it_init(it, bag, 0, ipd);
4721 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4722 /* swab quota uid/gid provided by caller */
4723 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4724 key = (const struct dt_key *)&oti->oti_quota_id;
4727 rc = iam_it_get(it, (struct iam_key *)key);
4729 if (S_ISDIR(obj->oo_inode->i_mode))
4730 iam_rec = (struct iam_rec *)oti->oti_ldp;
4732 iam_rec = (struct iam_rec *) rec;
4734 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
4736 if (S_ISDIR(obj->oo_inode->i_mode))
4737 osd_fid_unpack((struct lu_fid *) rec,
4738 (struct osd_fid_pack *)iam_rec);
4739 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
4740 osd_quota_unpack(obj, rec);
4745 osd_ipd_put(env, bag, ipd);
4747 LINVRNT(osd_invariant(obj));
4752 static int osd_index_declare_iam_insert(const struct lu_env *env,
4753 struct dt_object *dt,
4754 const struct dt_rec *rec,
4755 const struct dt_key *key,
4756 struct thandle *handle)
4758 struct osd_thandle *oh;
4760 LASSERT(handle != NULL);
4762 oh = container_of0(handle, struct osd_thandle, ot_super);
4763 LASSERT(oh->ot_handle == NULL);
4765 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
4766 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
4772 * Inserts (key, value) pair in \a dt index object.
4774 * \param dt osd index object
4775 * \param key key for index
4776 * \param rec record reference
4777 * \param th transaction handler
4780 * \retval -ve failure
4782 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
4783 const struct dt_rec *rec,
4784 const struct dt_key *key, struct thandle *th,
4787 struct osd_object *obj = osd_dt_obj(dt);
4788 struct iam_path_descr *ipd;
4789 struct osd_thandle *oh;
4790 struct iam_container *bag;
4791 struct osd_thread_info *oti = osd_oti_get(env);
4792 struct iam_rec *iam_rec;
4796 if (!dt_object_exists(dt))
4799 LINVRNT(osd_invariant(obj));
4800 LASSERT(!dt_object_remote(dt));
4802 bag = &obj->oo_dir->od_container;
4803 LASSERT(bag->ic_object == obj->oo_inode);
4804 LASSERT(th != NULL);
4806 osd_trans_exec_op(env, th, OSD_OT_INSERT);
4808 ipd = osd_idx_ipd_get(env, bag);
4809 if (unlikely(ipd == NULL))
4812 oh = container_of0(th, struct osd_thandle, ot_super);
4813 LASSERT(oh->ot_handle != NULL);
4814 LASSERT(oh->ot_handle->h_transaction != NULL);
4815 if (S_ISDIR(obj->oo_inode->i_mode)) {
4816 iam_rec = (struct iam_rec *)oti->oti_ldp;
4817 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec, &oti->oti_fid);
4818 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4819 /* pack quota uid/gid */
4820 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4821 key = (const struct dt_key *)&oti->oti_quota_id;
4822 /* pack quota record */
4823 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
4824 iam_rec = (struct iam_rec *)rec;
4826 iam_rec = (struct iam_rec *)rec;
4829 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
4831 osd_ipd_put(env, bag, ipd);
4832 LINVRNT(osd_invariant(obj));
4833 osd_trans_exec_check(env, th, OSD_OT_INSERT);
4838 * Calls ldiskfs_add_entry() to add directory entry
4839 * into the directory. This is required for
4840 * interoperability mode (b11826)
4842 * \retval 0, on success
4843 * \retval -ve, on error
4845 static int __osd_ea_add_rec(struct osd_thread_info *info,
4846 struct osd_object *pobj, struct inode *cinode,
4847 const char *name, const struct lu_fid *fid,
4848 struct htree_lock *hlock, struct thandle *th)
4850 struct ldiskfs_dentry_param *ldp;
4851 struct dentry *child;
4852 struct osd_thandle *oth;
4855 oth = container_of(th, struct osd_thandle, ot_super);
4856 LASSERT(oth->ot_handle != NULL);
4857 LASSERT(oth->ot_handle->h_transaction != NULL);
4858 LASSERT(pobj->oo_inode);
4860 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4861 if (unlikely(pobj->oo_inode ==
4862 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
4865 osd_get_ldiskfs_dirent_param(ldp, fid);
4866 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
4867 child->d_fsdata = (void *)ldp;
4868 ll_vfs_dq_init(pobj->oo_inode);
4869 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
4870 child, cinode, hlock);
4871 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
4872 struct ldiskfs_dir_entry_2 *de;
4873 struct buffer_head *bh;
4876 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
4879 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
4882 if (S_ISDIR(cinode->i_mode))
4883 de->file_type = LDISKFS_DIRENT_LUFID |
4884 LDISKFS_FT_REG_FILE;
4886 de->file_type = LDISKFS_DIRENT_LUFID |
4888 ldiskfs_handle_dirty_metadata(oth->ot_handle,
4899 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
4900 * into the directory.Also sets flags into osd object to
4901 * indicate dot and dotdot are created. This is required for
4902 * interoperability mode (b11826)
4904 * \param dir directory for dot and dotdot fixup.
4905 * \param obj child object for linking
4907 * \retval 0, on success
4908 * \retval -ve, on error
4910 static int osd_add_dot_dotdot(struct osd_thread_info *info,
4911 struct osd_object *dir,
4912 struct inode *parent_dir, const char *name,
4913 const struct lu_fid *dot_fid,
4914 const struct lu_fid *dot_dot_fid,
4917 struct inode *inode = dir->oo_inode;
4918 struct osd_thandle *oth;
4921 oth = container_of(th, struct osd_thandle, ot_super);
4922 LASSERT(oth->ot_handle->h_transaction != NULL);
4923 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
4925 if (strcmp(name, dot) == 0) {
4926 if (dir->oo_compat_dot_created) {
4929 LASSERT(inode->i_ino == parent_dir->i_ino);
4930 dir->oo_compat_dot_created = 1;
4933 } else if (strcmp(name, dotdot) == 0) {
4934 if (!dir->oo_compat_dot_created)
4936 /* in case of rename, dotdot is already created */
4937 if (dir->oo_compat_dotdot_created) {
4938 return __osd_ea_add_rec(info, dir, parent_dir, name,
4939 dot_dot_fid, NULL, th);
4942 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
4943 struct lu_fid tfid = *dot_dot_fid;
4946 result = osd_add_dot_dotdot_internal(info,
4947 dir->oo_inode, parent_dir, dot_fid,
4950 result = osd_add_dot_dotdot_internal(info,
4951 dir->oo_inode, parent_dir, dot_fid,
4956 dir->oo_compat_dotdot_created = 1;
4964 * It will call the appropriate osd_add* function and return the
4965 * value, return by respective functions.
4967 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
4968 struct inode *cinode, const char *name,
4969 const struct lu_fid *fid, struct thandle *th)
4971 struct osd_thread_info *info = osd_oti_get(env);
4972 struct htree_lock *hlock;
4975 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
4977 if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' &&
4979 if (hlock != NULL) {
4980 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4983 down_write(&pobj->oo_ext_idx_sem);
4986 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
4987 lu_object_fid(&pobj->oo_dt.do_lu),
4990 if (hlock != NULL) {
4991 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4992 pobj->oo_inode, LDISKFS_HLOCK_ADD);
4994 down_write(&pobj->oo_ext_idx_sem);
4997 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
4998 struct lu_fid *tfid = &info->oti_fid;
5002 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5005 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5010 ldiskfs_htree_unlock(hlock);
5012 up_write(&pobj->oo_ext_idx_sem);
5018 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5019 struct osd_idmap_cache *oic)
5021 struct osd_scrub *scrub = &dev->od_scrub;
5022 struct lu_fid *fid = &oic->oic_fid;
5023 struct osd_inode_id *id = &oic->oic_lid;
5024 struct inode *inode = NULL;
5030 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5033 if (scrub->os_pos_current > id->oii_ino)
5037 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5038 if (rc == -ENOENT) {
5039 __u32 gen = id->oii_gen;
5045 inode = osd_iget(oti, dev, id);
5046 /* The inode has been removed (by race maybe). */
5047 if (IS_ERR(inode)) {
5048 rc = PTR_ERR(inode);
5050 RETURN(rc == -ESTALE ? -ENOENT : rc);
5053 /* The OI mapping is lost. */
5054 if (gen != OSD_OII_NOGEN)
5057 /* The inode may has been reused by others, we do not know,
5058 * leave it to be handled by subsequent osd_fid_lookup(). */
5060 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5067 if (thread_is_running(&scrub->os_thread)) {
5068 if (inode == NULL) {
5069 inode = osd_iget(oti, dev, id);
5070 /* The inode has been removed (by race maybe). */
5071 if (IS_ERR(inode)) {
5072 rc = PTR_ERR(inode);
5074 RETURN(rc == -ESTALE ? -ENOENT : rc);
5078 rc = osd_oii_insert(dev, oic, insert);
5079 /* There is race condition between osd_oi_lookup and OI scrub.
5080 * The OI scrub finished just after osd_oi_lookup() failure.
5081 * Under such case, it is unnecessary to trigger OI scrub again,
5082 * but try to call osd_oi_lookup() again. */
5083 if (unlikely(rc == -EAGAIN))
5086 if (!S_ISDIR(inode->i_mode))
5089 rc = osd_check_lmv(oti, dev, inode, oic);
5094 if (!dev->od_noscrub && ++once == 1) {
5095 rc = osd_scrub_start(dev, SS_AUTO_PARTIAL | SS_CLEAR_DRYRUN |
5097 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5098 "%s: trigger partial OI scrub for RPC inconsistency "
5099 "checking FID "DFID": rc = %d\n",
5100 osd_dev2name(dev), PFID(fid), rc);
5101 if (rc == 0 || rc == -EALREADY)
5114 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5115 struct osd_device *dev,
5116 struct osd_idmap_cache *oic,
5117 struct lu_fid *fid, __u32 ino)
5119 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5120 struct inode *inode;
5123 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5124 inode = osd_iget(oti, dev, &oic->oic_lid);
5125 if (IS_ERR(inode)) {
5126 fid_zero(&oic->oic_fid);
5127 return PTR_ERR(inode);
5130 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5133 fid_zero(&oic->oic_fid);
5135 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5139 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5140 struct osd_inode_id *id, const struct lu_fid *fid)
5142 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5143 id->oii_ino, id->oii_gen, info);
5144 info->oti_cache.oic_lid = *id;
5145 info->oti_cache.oic_fid = *fid;
5146 info->oti_cache.oic_dev = osd;
5150 * Get parent FID from the linkEA.
5152 * For a directory which parent resides on remote MDT, to satisfy the
5153 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5154 * the other hand, to make the lookup(..) on the directory can return
5155 * the real parent FID, we append the real parent FID after its ".."
5156 * name entry in the /REMOTE_PARENT_DIR.
5158 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5159 * backup. So after the restore, we cannot get the right parent FID from
5160 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5161 * we have stored the real parent FID in the directory object's linkEA,
5162 * we can parse the linkEA for the real parent FID.
5164 * \param[in] env pointer to the thread context
5165 * \param[in] obj pointer to the object to be handled
5166 * \param[out]fid pointer to the buffer to hold the parent FID
5168 * \retval 0 for getting the real parent FID successfully
5169 * \retval negative error number on failure
5171 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5172 struct osd_object *obj,
5175 struct osd_thread_info *oti = osd_oti_get(env);
5176 struct lu_buf *buf = &oti->oti_big_buf;
5177 struct dentry *dentry = &oti->oti_obj_dentry;
5178 struct inode *inode = obj->oo_inode;
5179 struct linkea_data ldata = { NULL };
5184 if (!S_ISDIR(inode->i_mode))
5188 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5189 buf->lb_buf, buf->lb_len);
5190 if (rc == -ERANGE) {
5191 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5194 lu_buf_realloc(buf, rc);
5195 if (buf->lb_buf == NULL)
5202 if (unlikely(rc == 0))
5208 if (unlikely(buf->lb_buf == NULL)) {
5209 lu_buf_realloc(buf, rc);
5210 if (buf->lb_buf == NULL)
5217 rc = linkea_init_with_rec(&ldata);
5219 linkea_first_entry(&ldata);
5220 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5226 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5227 struct inode *inode,
5228 const struct lu_fid *pfid,
5229 const char *name, const int namelen)
5231 struct osd_thread_info *oti = osd_oti_get(env);
5232 struct lu_buf *buf = &oti->oti_big_buf;
5233 struct dentry *dentry = &oti->oti_obj_dentry;
5234 struct linkea_data ldata = { NULL };
5235 struct lu_name cname = { .ln_name = name,
5236 .ln_namelen = namelen };
5241 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5242 buf->lb_buf, buf->lb_len);
5244 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5249 if (unlikely(rc == 0))
5252 if (buf->lb_len < rc) {
5253 lu_buf_realloc(buf, rc);
5254 if (buf->lb_buf == NULL)
5261 rc = linkea_init_with_rec(&ldata);
5263 rc = linkea_links_find(&ldata, &cname, pfid);
5269 * Calls ->lookup() to find dentry. From dentry get inode and
5270 * read inode's ea to get fid. This is required for interoperability
5273 * \retval 0, on success
5274 * \retval -ve, on error
5276 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5277 struct dt_rec *rec, const struct dt_key *key)
5279 struct inode *dir = obj->oo_inode;
5280 struct dentry *dentry;
5281 struct ldiskfs_dir_entry_2 *de;
5282 struct buffer_head *bh;
5283 struct lu_fid *fid = (struct lu_fid *) rec;
5284 struct htree_lock *hlock = NULL;
5289 LASSERT(dir->i_op != NULL);
5290 LASSERT(dir->i_op->lookup != NULL);
5292 dentry = osd_child_dentry_get(env, obj,
5293 (char *)key, strlen((char *)key));
5295 if (obj->oo_hl_head != NULL) {
5296 hlock = osd_oti_get(env)->oti_hlock;
5297 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5298 dir, LDISKFS_HLOCK_LOOKUP);
5300 down_read(&obj->oo_ext_idx_sem);
5303 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5305 struct osd_thread_info *oti = osd_oti_get(env);
5306 struct osd_inode_id *id = &oti->oti_id;
5307 struct osd_idmap_cache *oic = &oti->oti_cache;
5308 struct osd_device *dev = osd_obj2dev(obj);
5310 ino = le32_to_cpu(de->inode);
5311 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5313 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5317 rc = osd_get_fid_from_dentry(de, rec);
5319 /* done with de, release bh */
5322 if (unlikely(is_remote_parent_ino(dev, ino))) {
5323 const char *name = (const char *)key;
5325 /* If the parent is on remote MDT, and there
5326 * is no FID-in-dirent, then we have to get
5327 * the parent FID from the linkEA. */
5328 if (likely(strlen(name) == 2 &&
5329 name[0] == '.' && name[1] == '.'))
5330 rc = osd_get_pfid_from_linkea(env, obj,
5333 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5336 osd_id_gen(id, ino, OSD_OII_NOGEN);
5339 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5340 fid_zero(&oic->oic_fid);
5345 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5346 rc = osd_consistency_check(oti, dev, oic);
5348 fid_zero(&oic->oic_fid);
5357 ldiskfs_htree_unlock(hlock);
5359 up_read(&obj->oo_ext_idx_sem);
5363 static int osd_index_declare_ea_insert(const struct lu_env *env,
5364 struct dt_object *dt,
5365 const struct dt_rec *rec,
5366 const struct dt_key *key,
5367 struct thandle *handle)
5369 struct osd_thandle *oh;
5370 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5371 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5372 const struct lu_fid *fid = rec1->rec_fid;
5373 int credits, rc = 0;
5374 struct osd_idmap_cache *idc;
5377 LASSERT(!dt_object_remote(dt));
5378 LASSERT(handle != NULL);
5379 LASSERT(fid != NULL);
5380 LASSERT(rec1->rec_type != 0);
5382 oh = container_of0(handle, struct osd_thandle, ot_super);
5383 LASSERT(oh->ot_handle == NULL);
5385 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5387 /* we can't call iget() while a transactions is running
5388 * (this can lead to a deadlock), but we need to know
5389 * inum and object type. so we find this information at
5390 * declaration and cache in per-thread info */
5391 idc = osd_idc_find_or_init(env, osd, fid);
5393 RETURN(PTR_ERR(idc));
5394 if (idc->oic_remote) {
5395 /* a reference to remote inode is represented by an
5396 * agent inode which we have to create */
5397 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5398 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5401 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5403 if (osd_dt_obj(dt)->oo_inode != NULL) {
5404 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5406 /* We ignore block quota on meta pool (MDTs), so needn't
5407 * calculate how many blocks will be consumed by this index
5409 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5411 i_projid_read(inode), 0,
5412 oh, osd_dt_obj(dt), NULL,
5417 #ifdef HAVE_PROJECT_QUOTA
5418 /* Reserve credits for local agent inode to transfer
5419 * to 0, quota enforcement is ignored in this case.
5421 if (idc->oic_remote &&
5422 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
5423 i_projid_read(inode) != 0)
5424 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
5425 0, i_projid_read(inode),
5426 0, false, PRJQUOTA);
5435 * Index add function for interoperability mode (b11826).
5436 * It will add the directory entry.This entry is needed to
5437 * maintain name->fid mapping.
5439 * \param key it is key i.e. file entry to be inserted
5440 * \param rec it is value of given key i.e. fid
5442 * \retval 0, on success
5443 * \retval -ve, on error
5445 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5446 const struct dt_rec *rec,
5447 const struct dt_key *key, struct thandle *th,
5450 struct osd_object *obj = osd_dt_obj(dt);
5451 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5452 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5453 const struct lu_fid *fid = rec1->rec_fid;
5454 const char *name = (const char *)key;
5455 struct osd_thread_info *oti = osd_oti_get(env);
5456 struct inode *child_inode = NULL;
5457 struct osd_idmap_cache *idc;
5461 if (!dt_object_exists(dt))
5464 LASSERT(osd_invariant(obj));
5465 LASSERT(!dt_object_remote(dt));
5466 LASSERT(th != NULL);
5468 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5470 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5472 idc = osd_idc_find(env, osd, fid);
5473 if (unlikely(idc == NULL)) {
5474 /* this dt_insert() wasn't declared properly, so
5475 * FID is missing in OI cache. we better do not
5476 * lookup FID in FLDB/OI and don't risk to deadlock,
5477 * but in some special cases (lfsck testing, etc)
5478 * it's much simpler than fixing a caller */
5479 CERROR("%s: "DFID" wasn't declared for insert\n",
5480 osd_name(osd), PFID(fid));
5482 idc = osd_idc_find_or_init(env, osd, fid);
5484 RETURN(PTR_ERR(idc));
5487 if (idc->oic_remote) {
5488 /* Insert remote entry */
5489 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
5490 struct osd_mdobj_map *omm = osd->od_mdt_map;
5491 struct osd_thandle *oh;
5493 /* If parent on remote MDT, we need put this object
5495 oh = container_of(th, typeof(*oh), ot_super);
5496 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5498 CERROR("%s: add "DFID" error: rc = %d\n",
5500 PFID(lu_object_fid(&dt->do_lu)), rc);
5504 child_inode = igrab(omm->omm_remote_parent->d_inode);
5506 child_inode = osd_create_local_agent_inode(env, osd,
5507 obj, fid, rec1->rec_type & S_IFMT, th);
5508 if (IS_ERR(child_inode))
5509 RETURN(PTR_ERR(child_inode));
5512 /* Insert local entry */
5513 if (unlikely(idc->oic_lid.oii_ino == 0)) {
5514 /* for a reason OI cache wasn't filled properly */
5515 CERROR("%s: OIC for "DFID" isn't filled\n",
5516 osd_name(osd), PFID(fid));
5519 child_inode = oti->oti_inode;
5520 if (unlikely(child_inode == NULL)) {
5521 struct ldiskfs_inode_info *lii;
5525 child_inode = oti->oti_inode = &lii->vfs_inode;
5527 child_inode->i_sb = osd_sb(osd);
5528 child_inode->i_ino = idc->oic_lid.oii_ino;
5529 child_inode->i_mode = rec1->rec_type & S_IFMT;
5532 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
5534 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
5535 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
5537 if (child_inode && child_inode != oti->oti_inode)
5539 LASSERT(osd_invariant(obj));
5540 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5545 * Initialize osd Iterator for given osd index object.
5547 * \param dt osd index object
5550 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
5551 struct dt_object *dt,
5554 struct osd_it_iam *it;
5555 struct osd_object *obj = osd_dt_obj(dt);
5556 struct lu_object *lo = &dt->do_lu;
5557 struct iam_path_descr *ipd;
5558 struct iam_container *bag = &obj->oo_dir->od_container;
5560 if (!dt_object_exists(dt))
5561 return ERR_PTR(-ENOENT);
5565 return ERR_PTR(-ENOMEM);
5567 ipd = osd_it_ipd_get(env, bag);
5568 if (likely(ipd != NULL)) {
5572 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
5573 return (struct dt_it *)it;
5576 return ERR_PTR(-ENOMEM);
5581 * free given Iterator.
5583 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
5585 struct osd_it_iam *it = (struct osd_it_iam *)di;
5586 struct osd_object *obj = it->oi_obj;
5588 iam_it_fini(&it->oi_it);
5589 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
5590 osd_object_put(env, obj);
5595 * Move Iterator to record specified by \a key
5597 * \param di osd iterator
5598 * \param key key for index
5600 * \retval +ve di points to record with least key not larger than key
5601 * \retval 0 di points to exact matched key
5602 * \retval -ve failure
5605 static int osd_it_iam_get(const struct lu_env *env,
5606 struct dt_it *di, const struct dt_key *key)
5608 struct osd_thread_info *oti = osd_oti_get(env);
5609 struct osd_it_iam *it = (struct osd_it_iam *)di;
5611 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5612 /* swab quota uid/gid */
5613 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5614 key = (struct dt_key *)&oti->oti_quota_id;
5617 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
5623 * \param di osd iterator
5625 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
5627 struct osd_it_iam *it = (struct osd_it_iam *)di;
5629 iam_it_put(&it->oi_it);
5633 * Move iterator by one record
5635 * \param di osd iterator
5637 * \retval +1 end of container reached
5639 * \retval -ve failure
5642 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
5644 struct osd_it_iam *it = (struct osd_it_iam *)di;
5646 return iam_it_next(&it->oi_it);
5650 * Return pointer to the key under iterator.
5653 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
5654 const struct dt_it *di)
5656 struct osd_thread_info *oti = osd_oti_get(env);
5657 struct osd_it_iam *it = (struct osd_it_iam *)di;
5658 struct osd_object *obj = it->oi_obj;
5661 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
5663 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
5664 /* swab quota uid/gid */
5665 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
5666 key = (struct dt_key *)&oti->oti_quota_id;
5673 * Return size of key under iterator (in bytes)
5676 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
5678 struct osd_it_iam *it = (struct osd_it_iam *)di;
5680 return iam_it_key_size(&it->oi_it);
5684 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
5686 /* check if file type is required */
5687 if (ent->lde_attrs & LUDA_TYPE) {
5688 struct luda_type *lt;
5689 int align = sizeof(*lt) - 1;
5691 len = (len + align) & ~align;
5692 lt = (struct luda_type *)(ent->lde_name + len);
5693 lt->lt_type = cpu_to_le16(DTTOIF(type));
5696 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
5700 * build lu direct from backend fs dirent.
5704 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
5705 char *name, __u16 namelen, __u16 type, __u32 attr)
5707 ent->lde_attrs = attr | LUDA_FID;
5708 fid_cpu_to_le(&ent->lde_fid, fid);
5710 ent->lde_hash = cpu_to_le64(offset);
5711 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
5713 strncpy(ent->lde_name, name, namelen);
5714 ent->lde_name[namelen] = '\0';
5715 ent->lde_namelen = cpu_to_le16(namelen);
5717 /* append lustre attributes */
5718 osd_it_append_attrs(ent, namelen, type);
5722 * Return pointer to the record under iterator.
5724 static int osd_it_iam_rec(const struct lu_env *env,
5725 const struct dt_it *di,
5726 struct dt_rec *dtrec, __u32 attr)
5728 struct osd_it_iam *it = (struct osd_it_iam *)di;
5729 struct osd_thread_info *info = osd_oti_get(env);
5732 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
5733 const struct osd_fid_pack *rec;
5734 struct lu_fid *fid = &info->oti_fid;
5735 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
5741 name = (char *)iam_it_key_get(&it->oi_it);
5743 RETURN(PTR_ERR(name));
5745 namelen = iam_it_key_size(&it->oi_it);
5747 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
5749 RETURN(PTR_ERR(rec));
5751 rc = osd_fid_unpack(fid, rec);
5755 hash = iam_it_store(&it->oi_it);
5757 /* IAM does not store object type in IAM index (dir) */
5758 osd_it_pack_dirent(lde, fid, hash, name, namelen,
5760 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5761 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5762 (struct iam_rec *)dtrec);
5763 osd_quota_unpack(it->oi_obj, dtrec);
5765 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5766 (struct iam_rec *)dtrec);
5773 * Returns cookie for current Iterator position.
5775 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
5777 struct osd_it_iam *it = (struct osd_it_iam *)di;
5779 return iam_it_store(&it->oi_it);
5783 * Restore iterator from cookie.
5785 * \param di osd iterator
5786 * \param hash Iterator location cookie
5788 * \retval +ve di points to record with least key not larger than key.
5789 * \retval 0 di points to exact matched key
5790 * \retval -ve failure
5793 static int osd_it_iam_load(const struct lu_env *env,
5794 const struct dt_it *di, __u64 hash)
5796 struct osd_it_iam *it = (struct osd_it_iam *)di;
5798 return iam_it_load(&it->oi_it, hash);
5801 static const struct dt_index_operations osd_index_iam_ops = {
5802 .dio_lookup = osd_index_iam_lookup,
5803 .dio_declare_insert = osd_index_declare_iam_insert,
5804 .dio_insert = osd_index_iam_insert,
5805 .dio_declare_delete = osd_index_declare_iam_delete,
5806 .dio_delete = osd_index_iam_delete,
5808 .init = osd_it_iam_init,
5809 .fini = osd_it_iam_fini,
5810 .get = osd_it_iam_get,
5811 .put = osd_it_iam_put,
5812 .next = osd_it_iam_next,
5813 .key = osd_it_iam_key,
5814 .key_size = osd_it_iam_key_size,
5815 .rec = osd_it_iam_rec,
5816 .store = osd_it_iam_store,
5817 .load = osd_it_iam_load
5823 * Creates or initializes iterator context.
5825 * \retval struct osd_it_ea, iterator structure on success
5828 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
5829 struct dt_object *dt,
5832 struct osd_object *obj = osd_dt_obj(dt);
5833 struct osd_thread_info *info = osd_oti_get(env);
5834 struct osd_it_ea *oie;
5836 struct lu_object *lo = &dt->do_lu;
5837 struct dentry *obj_dentry;
5840 if (!dt_object_exists(dt) || obj->oo_destroyed)
5841 RETURN(ERR_PTR(-ENOENT));
5843 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
5845 RETURN(ERR_PTR(-ENOMEM));
5846 obj_dentry = &oie->oie_dentry;
5848 obj_dentry->d_inode = obj->oo_inode;
5849 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
5850 obj_dentry->d_name.hash = 0;
5852 oie->oie_rd_dirent = 0;
5853 oie->oie_it_dirent = 0;
5854 oie->oie_dirent = NULL;
5855 if (unlikely(!info->oti_it_ea_buf_used)) {
5856 oie->oie_buf = info->oti_it_ea_buf;
5857 info->oti_it_ea_buf_used = 1;
5859 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5860 if (oie->oie_buf == NULL)
5861 RETURN(ERR_PTR(-ENOMEM));
5865 file = &oie->oie_file;
5867 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
5868 if (attr & LUDA_64BITHASH)
5869 file->f_mode = FMODE_64BITHASH;
5871 file->f_mode = FMODE_32BITHASH;
5872 file->f_path.dentry = obj_dentry;
5873 file->f_mapping = obj->oo_inode->i_mapping;
5874 file->f_op = obj->oo_inode->i_fop;
5875 set_file_inode(file, obj->oo_inode);
5878 RETURN((struct dt_it *) oie);
5882 * Destroy or finishes iterator context.
5884 * \param di iterator structure to be destroyed
5886 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
5888 struct osd_thread_info *info = osd_oti_get(env);
5889 struct osd_it_ea *oie = (struct osd_it_ea *)di;
5890 struct osd_object *obj = oie->oie_obj;
5891 struct inode *inode = obj->oo_inode;
5894 oie->oie_file.f_op->release(inode, &oie->oie_file);
5895 osd_object_put(env, obj);
5896 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
5897 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5899 info->oti_it_ea_buf_used = 0;
5900 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
5905 * It position the iterator at given key, so that next lookup continues from
5906 * that key Or it is similar to dio_it->load() but based on a key,
5907 * rather than file position.
5909 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
5912 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
5914 static int osd_it_ea_get(const struct lu_env *env,
5915 struct dt_it *di, const struct dt_key *key)
5917 struct osd_it_ea *it = (struct osd_it_ea *)di;
5920 LASSERT(((const char *)key)[0] == '\0');
5921 it->oie_file.f_pos = 0;
5922 it->oie_rd_dirent = 0;
5923 it->oie_it_dirent = 0;
5924 it->oie_dirent = NULL;
5932 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
5936 struct osd_filldir_cbs {
5937 #ifdef HAVE_DIR_CONTEXT
5938 struct dir_context ctx;
5940 struct osd_it_ea *it;
5943 * It is called internally by ->readdir(). It fills the
5944 * iterator's in-memory data structure with required
5945 * information i.e. name, namelen, rec_size etc.
5947 * \param buf in which information to be filled in.
5948 * \param name name of the file in given dir
5950 * \retval 0 on success
5951 * \retval 1 on buffer full
5953 #ifdef HAVE_FILLDIR_USE_CTX
5954 static int osd_ldiskfs_filldir(struct dir_context *buf,
5956 static int osd_ldiskfs_filldir(void *buf,
5958 const char *name, int namelen,
5959 loff_t offset, __u64 ino, unsigned d_type)
5961 struct osd_it_ea *it =
5962 ((struct osd_filldir_cbs *)buf)->it;
5963 struct osd_object *obj = it->oie_obj;
5964 struct osd_it_ea_dirent *ent = it->oie_dirent;
5965 struct lu_fid *fid = &ent->oied_fid;
5966 struct osd_fid_pack *rec;
5969 /* this should never happen */
5970 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
5971 CERROR("ldiskfs return invalid namelen %d\n", namelen);
5975 if ((void *) ent - it->oie_buf + sizeof(*ent) + namelen >
5979 /* "." is just the object itself. */
5980 if (namelen == 1 && name[0] == '.') {
5981 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5982 } else if (d_type & LDISKFS_DIRENT_LUFID) {
5983 rec = (struct osd_fid_pack*) (name + namelen + 1);
5984 if (osd_fid_unpack(fid, rec) != 0)
5989 d_type &= ~LDISKFS_DIRENT_LUFID;
5991 /* NOT export local root. */
5992 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
5993 ino = obj->oo_inode->i_ino;
5994 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5997 ent->oied_ino = ino;
5998 ent->oied_off = offset;
5999 ent->oied_namelen = namelen;
6000 ent->oied_type = d_type;
6002 memcpy(ent->oied_name, name, namelen);
6004 it->oie_rd_dirent++;
6005 it->oie_dirent = (void *) ent + cfs_size_round(sizeof(*ent) + namelen);
6010 * Calls ->readdir() to load a directory entry at a time
6011 * and stored it in iterator's in-memory data structure.
6013 * \param di iterator's in memory structure
6015 * \retval 0 on success
6016 * \retval -ve on error
6017 * \retval +1 reach the end of entry
6019 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6020 const struct dt_it *di)
6022 struct osd_it_ea *it = (struct osd_it_ea *)di;
6023 struct osd_object *obj = it->oie_obj;
6024 struct inode *inode = obj->oo_inode;
6025 struct htree_lock *hlock = NULL;
6026 struct file *filp = &it->oie_file;
6028 struct osd_filldir_cbs buf = {
6029 #ifdef HAVE_DIR_CONTEXT
6030 .ctx.actor = osd_ldiskfs_filldir,
6036 it->oie_dirent = it->oie_buf;
6037 it->oie_rd_dirent = 0;
6039 if (obj->oo_hl_head != NULL) {
6040 hlock = osd_oti_get(env)->oti_hlock;
6041 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6042 inode, LDISKFS_HLOCK_READDIR);
6044 down_read(&obj->oo_ext_idx_sem);
6047 #ifdef HAVE_DIR_CONTEXT
6048 buf.ctx.pos = filp->f_pos;
6049 rc = inode->i_fop->iterate(filp, &buf.ctx);
6050 filp->f_pos = buf.ctx.pos;
6052 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6056 ldiskfs_htree_unlock(hlock);
6058 up_read(&obj->oo_ext_idx_sem);
6060 if (it->oie_rd_dirent == 0) {
6061 /*If it does not get any dirent, it means it has been reached
6062 *to the end of the dir */
6063 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6067 it->oie_dirent = it->oie_buf;
6068 it->oie_it_dirent = 1;
6075 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6076 * to load a directory entry at a time and stored it in
6077 * iterator's in-memory data structure.
6079 * \param di iterator's in memory structure
6081 * \retval +ve iterator reached to end
6082 * \retval 0 iterator not reached to end
6083 * \retval -ve on error
6085 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6087 struct osd_it_ea *it = (struct osd_it_ea *)di;
6092 if (it->oie_it_dirent < it->oie_rd_dirent) {
6094 (void *) it->oie_dirent +
6095 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6096 it->oie_dirent->oied_namelen);
6097 it->oie_it_dirent++;
6100 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6103 rc = osd_ldiskfs_it_fill(env, di);
6110 * Returns the key at current position from iterator's in memory structure.
6112 * \param di iterator's in memory structure
6114 * \retval key i.e. struct dt_key on success
6116 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6117 const struct dt_it *di)
6119 struct osd_it_ea *it = (struct osd_it_ea *)di;
6121 return (struct dt_key *)it->oie_dirent->oied_name;
6125 * Returns the key's size at current position from iterator's in memory structure.
6127 * \param di iterator's in memory structure
6129 * \retval key_size i.e. struct dt_key on success
6131 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6133 struct osd_it_ea *it = (struct osd_it_ea *)di;
6135 return it->oie_dirent->oied_namelen;
6138 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6140 if (LDISKFS_DIR_REC_LEN(de) >=
6141 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6148 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6149 unsigned blocksize, bool dotdot)
6152 return osd_dotdot_has_space(de);
6154 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6155 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6162 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6163 handle_t *jh, struct dentry *dentry,
6164 const struct lu_fid *fid, struct buffer_head *bh,
6165 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6168 struct inode *dir = dentry->d_parent->d_inode;
6169 struct inode *inode = dentry->d_inode;
6170 struct osd_fid_pack *rec;
6171 struct ldiskfs_dentry_param *ldp;
6172 int namelen = dentry->d_name.len;
6174 struct osd_thread_info *info = osd_oti_get(env);
6177 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6180 /* There is enough space to hold the FID-in-dirent. */
6181 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6182 rc = ldiskfs_journal_get_write_access(jh, bh);
6186 de->name[namelen] = 0;
6187 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6188 rec->fp_len = sizeof(struct lu_fid) + 1;
6189 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6190 de->file_type |= LDISKFS_DIRENT_LUFID;
6191 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6198 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6202 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6203 osd_get_ldiskfs_dirent_param(ldp, fid);
6204 dentry->d_fsdata = (void *)ldp;
6205 ll_vfs_dq_init(dir);
6206 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6207 /* It is too bad, we cannot reinsert the name entry back.
6208 * That means we lose it! */
6210 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6211 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6212 osd_ino2name(inode),
6213 dir->i_ino, dir->i_generation, namelen,
6214 dentry->d_name.name, PFID(fid), rc);
6220 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6221 struct osd_it_ea *it, struct lu_fid *fid,
6222 struct osd_inode_id *id, __u32 *attr)
6224 struct osd_thread_info *info = osd_oti_get(env);
6225 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6226 struct osd_device *dev = osd_obj2dev(obj);
6227 struct super_block *sb = osd_sb(dev);
6228 const char *devname = osd_name(dev);
6229 struct osd_it_ea_dirent *ent = it->oie_dirent;
6230 struct inode *dir = obj->oo_inode;
6231 struct htree_lock *hlock = NULL;
6232 struct buffer_head *bh = NULL;
6233 handle_t *jh = NULL;
6234 struct ldiskfs_dir_entry_2 *de;
6235 struct dentry *dentry;
6236 struct inode *inode;
6237 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6240 bool dotdot = false;
6244 if (ent->oied_name[0] == '.') {
6245 if (ent->oied_namelen == 1)
6248 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6252 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6253 inode = osd_iget(info, dev, id);
6254 if (IS_ERR(inode)) {
6255 rc = PTR_ERR(inode);
6256 if (rc == -ENOENT || rc == -ESTALE) {
6257 /* Maybe dangling name entry, or
6258 * corrupted directory entry. */
6259 *attr |= LUDA_UNKNOWN;
6262 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6263 "check_repair, dir = %lu/%u, name = %.*s, "
6264 "ino = %llu, rc = %d\n",
6265 devname, dir->i_ino, dir->i_generation,
6266 ent->oied_namelen, ent->oied_name,
6273 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6275 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6276 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6281 /* We need to ensure that the name entry is still valid.
6282 * Because it may be removed or renamed by other already.
6284 * The unlink or rename operation will start journal before PDO lock,
6285 * so to avoid deadlock, here we need to start journal handle before
6286 * related PDO lock also. But because we do not know whether there
6287 * will be something to be repaired before PDO lock, we just start
6288 * journal without conditions.
6290 * We may need to remove the name entry firstly, then insert back.
6291 * One credit is for user quota file update.
6292 * One credit is for group quota file update.
6293 * Two credits are for dirty inode. */
6294 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6295 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6297 if (dev->od_dirent_journal != 0) {
6300 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6303 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6304 "check_repair, dir = %lu/%u, credits = %d, "
6305 "name = %.*s, ino = %llu: rc = %d\n",
6306 devname, dir->i_ino, dir->i_generation, credits,
6307 ent->oied_namelen, ent->oied_name,
6310 GOTO(out_inode, rc);
6313 if (obj->oo_hl_head != NULL) {
6314 hlock = osd_oti_get(env)->oti_hlock;
6315 /* "0" means exclusive lock for the whole directory.
6316 * We need to prevent others access such name entry
6317 * during the delete + insert. Neither HLOCK_ADD nor
6318 * HLOCK_DEL cannot guarantee the atomicity. */
6319 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6321 down_write(&obj->oo_ext_idx_sem);
6324 if (obj->oo_hl_head != NULL) {
6325 hlock = osd_oti_get(env)->oti_hlock;
6326 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6327 LDISKFS_HLOCK_LOOKUP);
6329 down_read(&obj->oo_ext_idx_sem);
6333 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6334 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6335 *attr |= LUDA_IGNORE;
6340 /* For dotdot entry, if there is not enough space to hold the
6341 * FID-in-dirent, just keep them there. It only happens when the
6342 * device upgraded from 1.8 or restored from MDT file-level backup.
6343 * For the whole directory, only dotdot entry have no FID-in-dirent
6344 * and needs to get FID from LMA when readdir, it will not affect the
6345 * performance much. */
6346 if (dotdot && !osd_dotdot_has_space(de)) {
6347 *attr |= LUDA_UNKNOWN;
6353 if (lu_fid_eq(fid, &lma->lma_self_fid))
6356 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6357 struct lu_fid *tfid = &lma->lma_self_fid;
6359 if (likely(dotdot &&
6360 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6361 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6362 /* It must be REMOTE_PARENT_DIR and as the
6363 * 'dotdot' entry of remote directory */
6364 *attr |= LUDA_IGNORE;
6366 CDEBUG(D_LFSCK, "%s: expect remote agent "
6367 "parent directory, but got %.*s under "
6368 "dir = %lu/%u with the FID "DFID"\n",
6369 devname, ent->oied_namelen,
6370 ent->oied_name, dir->i_ino,
6371 dir->i_generation, PFID(tfid));
6373 *attr |= LUDA_UNKNOWN;
6380 if (!fid_is_zero(fid)) {
6381 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6383 if (rc == -ENOENT ||
6385 !(dev->od_scrub.os_file.sf_flags & SF_UPGRADE))) {
6386 /* linkEA does not recognize the dirent entry,
6387 * it may because the dirent entry corruption
6388 * and points to other's inode. */
6389 CDEBUG(D_LFSCK, "%s: the target inode does not "
6390 "recognize the dirent, dir = %lu/%u, "
6391 " name = %.*s, ino = %llu, "
6392 DFID": rc = %d\n", devname, dir->i_ino,
6393 dir->i_generation, ent->oied_namelen,
6394 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6395 *attr |= LUDA_UNKNOWN;
6400 if (rc && rc != -ENODATA) {
6401 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6402 "dir = %lu/%u, name = %.*s, ino = %llu, "
6403 DFID": rc = %d\n", devname, dir->i_ino,
6404 dir->i_generation, ent->oied_namelen,
6405 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6406 *attr |= LUDA_UNKNOWN;
6413 /* linkEA recognizes the dirent entry, the FID-in-LMA is
6414 * valid, trusted, in spite of fid_is_sane(fid) or not. */
6415 if (*attr & LUDA_VERIFY_DRYRUN) {
6416 *fid = lma->lma_self_fid;
6417 *attr |= LUDA_REPAIR;
6424 dev->od_dirent_journal = 1;
6425 if (hlock != NULL) {
6426 ldiskfs_htree_unlock(hlock);
6429 up_read(&obj->oo_ext_idx_sem);
6435 *fid = lma->lma_self_fid;
6437 /* Update or append the FID-in-dirent. */
6438 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6439 bh, de, hlock, dotdot);
6441 *attr |= LUDA_REPAIR;
6443 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6444 "the dirent, dir = %lu/%u, name = %.*s, "
6445 "ino = %llu, "DFID": rc = %d\n",
6446 devname, dir->i_ino, dir->i_generation,
6447 ent->oied_namelen, ent->oied_name,
6448 ent->oied_ino, PFID(fid), rc);
6450 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6451 if (*attr & LUDA_VERIFY_DRYRUN) {
6452 if (fid_is_sane(fid)) {
6453 *attr |= LUDA_REPAIR;
6454 } else if (dev->od_index == 0) {
6455 lu_igif_build(fid, inode->i_ino,
6456 inode->i_generation);
6457 *attr |= LUDA_UPGRADE;
6465 dev->od_dirent_journal = 1;
6466 if (hlock != NULL) {
6467 ldiskfs_htree_unlock(hlock);
6470 up_read(&obj->oo_ext_idx_sem);
6477 if (unlikely(fid_is_sane(fid))) {
6478 /* FID-in-dirent exists, but FID-in-LMA is lost.
6479 * Trust the FID-in-dirent, and add FID-in-LMA. */
6480 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
6482 *attr |= LUDA_REPAIR;
6484 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
6485 "update dirent, dir = %lu/%u, "
6486 "name = %.*s, ino = %llu, "
6488 devname, dir->i_ino, dir->i_generation,
6489 ent->oied_namelen, ent->oied_name,
6490 ent->oied_ino, PFID(fid), rc);
6491 } else if (dev->od_index == 0) {
6492 lu_igif_build(fid, inode->i_ino, inode->i_generation);
6493 /* It is probably IGIF object. Only aappend the
6494 * FID-in-dirent. OI scrub will process FID-in-LMA. */
6495 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6496 bh, de, hlock, dotdot);
6498 *attr |= LUDA_UPGRADE;
6500 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
6501 "after the dirent, dir = %lu/%u, "
6502 "name = %.*s, ino = %llu, "
6504 devname, dir->i_ino, dir->i_generation,
6505 ent->oied_namelen, ent->oied_name,
6506 ent->oied_ino, PFID(fid), rc);
6515 if (hlock != NULL) {
6516 ldiskfs_htree_unlock(hlock);
6518 if (dev->od_dirent_journal != 0)
6519 up_write(&obj->oo_ext_idx_sem);
6521 up_read(&obj->oo_ext_idx_sem);
6525 ldiskfs_journal_stop(jh);
6529 if (rc >= 0 && !dirty)
6530 dev->od_dirent_journal = 0;
6536 * Returns the value at current position from iterator's in memory structure.
6538 * \param di struct osd_it_ea, iterator's in memory structure
6539 * \param attr attr requested for dirent.
6540 * \param lde lustre dirent
6542 * \retval 0 no error and \param lde has correct lustre dirent.
6543 * \retval -ve on error
6545 static inline int osd_it_ea_rec(const struct lu_env *env,
6546 const struct dt_it *di,
6547 struct dt_rec *dtrec, __u32 attr)
6549 struct osd_it_ea *it = (struct osd_it_ea *)di;
6550 struct osd_object *obj = it->oie_obj;
6551 struct osd_device *dev = osd_obj2dev(obj);
6552 struct osd_thread_info *oti = osd_oti_get(env);
6553 struct osd_inode_id *id = &oti->oti_id;
6554 struct lu_fid *fid = &it->oie_dirent->oied_fid;
6555 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6556 __u32 ino = it->oie_dirent->oied_ino;
6560 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
6562 if (attr & LUDA_VERIFY) {
6563 if (unlikely(is_remote_parent_ino(dev, ino))) {
6564 attr |= LUDA_IGNORE;
6565 /* If the parent is on remote MDT, and there
6566 * is no FID-in-dirent, then we have to get
6567 * the parent FID from the linkEA. */
6568 if (!fid_is_sane(fid) &&
6569 it->oie_dirent->oied_namelen == 2 &&
6570 it->oie_dirent->oied_name[0] == '.' &&
6571 it->oie_dirent->oied_name[1] == '.')
6572 osd_get_pfid_from_linkea(env, obj, fid);
6574 rc = osd_dirent_check_repair(env, obj, it, fid, id,
6578 if (!fid_is_sane(fid))
6579 attr |= LUDA_UNKNOWN;
6581 attr &= ~LU_DIRENT_ATTRS_MASK;
6582 if (!fid_is_sane(fid)) {
6583 bool is_dotdot = false;
6584 if (it->oie_dirent->oied_namelen == 2 &&
6585 it->oie_dirent->oied_name[0] == '.' &&
6586 it->oie_dirent->oied_name[1] == '.')
6588 /* If the parent is on remote MDT, and there
6589 * is no FID-in-dirent, then we have to get
6590 * the parent FID from the linkEA. */
6591 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
6592 rc = osd_get_pfid_from_linkea(env, obj, fid);
6594 if (is_dotdot == false &&
6595 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
6598 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6601 osd_id_gen(id, ino, OSD_OII_NOGEN);
6605 /* Pack the entry anyway, at least the offset is right. */
6606 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
6607 it->oie_dirent->oied_name,
6608 it->oie_dirent->oied_namelen,
6609 it->oie_dirent->oied_type, attr);
6614 if (osd_remote_fid(env, dev, fid))
6617 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
6618 osd_add_oi_cache(oti, dev, id, fid);
6620 RETURN(rc > 0 ? 0 : rc);
6624 * Returns the record size size at current position.
6626 * This function will return record(lu_dirent) size in bytes.
6628 * \param[in] env execution environment
6629 * \param[in] di iterator's in memory structure
6630 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
6631 * calculate the lu_dirent size.
6633 * \retval record size(in bytes & in memory) of the current lu_dirent
6636 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
6639 struct osd_it_ea *it = (struct osd_it_ea *)di;
6641 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
6645 * Returns a cookie for current position of the iterator head, so that
6646 * user can use this cookie to load/start the iterator next time.
6648 * \param di iterator's in memory structure
6650 * \retval cookie for current position, on success
6652 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
6654 struct osd_it_ea *it = (struct osd_it_ea *)di;
6656 return it->oie_dirent->oied_off;
6660 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6661 * to load a directory entry at a time and stored it i inn,
6662 * in iterator's in-memory data structure.
6664 * \param di struct osd_it_ea, iterator's in memory structure
6666 * \retval +ve on success
6667 * \retval -ve on error
6669 static int osd_it_ea_load(const struct lu_env *env,
6670 const struct dt_it *di, __u64 hash)
6672 struct osd_it_ea *it = (struct osd_it_ea *)di;
6676 it->oie_file.f_pos = hash;
6678 rc = osd_ldiskfs_it_fill(env, di);
6689 * Index lookup function for interoperability mode (b11826).
6691 * \param key, key i.e. file name to be searched
6693 * \retval +ve, on success
6694 * \retval -ve, on error
6696 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
6697 struct dt_rec *rec, const struct dt_key *key)
6699 struct osd_object *obj = osd_dt_obj(dt);
6704 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
6705 LINVRNT(osd_invariant(obj));
6707 rc = osd_ea_lookup_rec(env, obj, rec, key);
6714 * Index and Iterator operations for interoperability
6715 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
6717 static const struct dt_index_operations osd_index_ea_ops = {
6718 .dio_lookup = osd_index_ea_lookup,
6719 .dio_declare_insert = osd_index_declare_ea_insert,
6720 .dio_insert = osd_index_ea_insert,
6721 .dio_declare_delete = osd_index_declare_ea_delete,
6722 .dio_delete = osd_index_ea_delete,
6724 .init = osd_it_ea_init,
6725 .fini = osd_it_ea_fini,
6726 .get = osd_it_ea_get,
6727 .put = osd_it_ea_put,
6728 .next = osd_it_ea_next,
6729 .key = osd_it_ea_key,
6730 .key_size = osd_it_ea_key_size,
6731 .rec = osd_it_ea_rec,
6732 .rec_size = osd_it_ea_rec_size,
6733 .store = osd_it_ea_store,
6734 .load = osd_it_ea_load
6738 static void *osd_key_init(const struct lu_context *ctx,
6739 struct lu_context_key *key)
6741 struct osd_thread_info *info;
6743 OBD_ALLOC_PTR(info);
6745 return ERR_PTR(-ENOMEM);
6747 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6748 if (info->oti_it_ea_buf == NULL)
6751 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
6753 info->oti_hlock = ldiskfs_htree_lock_alloc();
6754 if (info->oti_hlock == NULL)
6760 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6763 return ERR_PTR(-ENOMEM);
6766 static void osd_key_fini(const struct lu_context *ctx,
6767 struct lu_context_key *key, void* data)
6769 struct osd_thread_info *info = data;
6770 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
6771 struct osd_idmap_cache *idc = info->oti_ins_cache;
6773 if (info->oti_inode != NULL)
6775 if (info->oti_hlock != NULL)
6776 ldiskfs_htree_lock_free(info->oti_hlock);
6777 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6778 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
6779 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
6780 lu_buf_free(&info->oti_big_buf);
6782 LASSERT(info->oti_ins_cache_size > 0);
6783 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
6784 info->oti_ins_cache = NULL;
6785 info->oti_ins_cache_size = 0;
6790 static void osd_key_exit(const struct lu_context *ctx,
6791 struct lu_context_key *key, void *data)
6793 struct osd_thread_info *info = data;
6795 LASSERT(info->oti_r_locks == 0);
6796 LASSERT(info->oti_w_locks == 0);
6797 LASSERT(info->oti_txns == 0);
6800 /* type constructor/destructor: osd_type_init, osd_type_fini */
6801 LU_TYPE_INIT_FINI(osd, &osd_key);
6803 struct lu_context_key osd_key = {
6804 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
6805 .lct_init = osd_key_init,
6806 .lct_fini = osd_key_fini,
6807 .lct_exit = osd_key_exit
6811 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
6812 const char *name, struct lu_device *next)
6814 struct osd_device *osd = osd_dev(d);
6816 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname))
6817 >= sizeof(osd->od_svname))
6819 return osd_procfs_init(osd, name);
6822 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
6824 struct seq_server_site *ss = osd_seq_site(osd);
6828 if (osd->od_is_ost || osd->od_cl_seq != NULL)
6831 if (unlikely(ss == NULL))
6834 OBD_ALLOC_PTR(osd->od_cl_seq);
6835 if (osd->od_cl_seq == NULL)
6838 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
6839 osd->od_svname, ss->ss_server_seq);
6841 OBD_FREE_PTR(osd->od_cl_seq);
6842 osd->od_cl_seq = NULL;
6846 if (ss->ss_node_id == 0) {
6847 /* If the OSD on the sequence controller(MDT0), then allocate
6848 * sequence here, otherwise allocate sequence after connected
6849 * to MDT0 (see mdt_register_lwp_callback()). */
6850 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
6851 &osd->od_cl_seq->lcs_space, env);
6857 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
6859 if (osd->od_cl_seq == NULL)
6862 seq_client_fini(osd->od_cl_seq);
6863 OBD_FREE_PTR(osd->od_cl_seq);
6864 osd->od_cl_seq = NULL;
6867 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
6871 /* shutdown quota slave instance associated with the device */
6872 if (o->od_quota_slave != NULL) {
6873 qsd_fini(env, o->od_quota_slave);
6874 o->od_quota_slave = NULL;
6877 osd_fid_fini(env, o);
6882 static void osd_umount(const struct lu_env *env, struct osd_device *o)
6886 if (o->od_mnt != NULL) {
6887 shrink_dcache_sb(osd_sb(o));
6888 osd_sync(env, &o->od_dt_dev);
6897 static int osd_mount(const struct lu_env *env,
6898 struct osd_device *o, struct lustre_cfg *cfg)
6900 const char *name = lustre_cfg_string(cfg, 0);
6901 const char *dev = lustre_cfg_string(cfg, 1);
6903 unsigned long page, s_flags, lmd_flags = 0;
6904 struct page *__page;
6905 struct file_system_type *type;
6906 char *options = NULL;
6908 struct osd_thread_info *info = osd_oti_get(env);
6909 struct lu_fid *fid = &info->oti_fid;
6910 struct inode *inode;
6911 int rc = 0, force_over_512tb = 0;
6914 if (o->od_mnt != NULL)
6917 if (strlen(dev) >= sizeof(o->od_mntdev))
6919 strcpy(o->od_mntdev, dev);
6921 str = lustre_cfg_string(cfg, 2);
6922 s_flags = simple_strtoul(str, NULL, 0);
6923 str = strstr(str, ":");
6925 lmd_flags = simple_strtoul(str + 1, NULL, 0);
6926 opts = lustre_cfg_string(cfg, 3);
6928 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
6929 CERROR("%s: device %s extents feature is not guaranteed to "
6930 "work on big-endian systems. Use \"bigendian_extents\" "
6931 "mount option to override.\n", name, dev);
6935 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
6936 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
6937 CWARN("force_over_128tb option is deprecated. "
6938 "Filesystems less than 512TB can be created without any "
6939 "force options. Use force_over_512tb option for "
6940 "filesystems greater than 512TB.\n");
6943 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
6944 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
6945 CWARN("force_over_256tb option is deprecated. "
6946 "Filesystems less than 512TB can be created without any "
6947 "force options. Use force_over_512tb option for "
6948 "filesystems greater than 512TB.\n");
6952 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
6953 force_over_512tb = 1;
6955 __page = alloc_page(GFP_KERNEL);
6957 GOTO(out, rc = -ENOMEM);
6958 page = (unsigned long)page_address(__page);
6959 options = (char *)page;
6962 /* strip out the options for back compatiblity */
6963 static char *sout[] = {
6970 /* strip out option we processed in osd */
6971 "bigendian_extents",
6977 strcat(options, opts);
6978 for (rc = 0, str = options; sout[rc]; ) {
6979 char *op = strstr(str, sout[rc]);
6985 if (op == options || *(op - 1) == ',') {
6986 str = op + strlen(sout[rc]);
6987 if (*str == ',' || *str == '\0') {
6988 *str == ',' ? str++ : str;
6989 memmove(op, str, strlen(str) + 1);
6992 for (str = op; *str != ',' && *str != '\0'; str++)
6996 strncat(options, "user_xattr,acl", 14);
6999 /* Glom up mount options */
7000 if (*options != '\0')
7001 strcat(options, ",");
7002 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7004 type = get_fs_type("ldiskfs");
7006 CERROR("%s: cannot find ldiskfs module\n", name);
7007 GOTO(out, rc = -ENODEV);
7010 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7011 module_put(type->owner);
7013 if (IS_ERR(o->od_mnt)) {
7014 rc = PTR_ERR(o->od_mnt);
7016 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7020 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7021 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7022 force_over_512tb == 0) {
7023 CERROR("%s: device %s LDISKFS does not support filesystems "
7024 "greater than 512TB and can cause data corruption. "
7025 "Use \"force_over_512tb\" mount option to override.\n",
7027 GOTO(out_mnt, rc = -EINVAL);
7030 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7031 if (priv_dev_set_rdonly) {
7032 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
7033 o->od_dt_dev.dd_rdonly = 1;
7034 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
7037 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7040 GOTO(out_mnt, rc = -EOPNOTSUPP);
7042 } else if (priv_dev_check_rdonly &&
7043 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
7044 CERROR("%s: underlying device %s is marked as "
7045 "read-only. Setup failed\n", name, dev);
7047 GOTO(out_mnt, rc = -EROFS);
7050 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7051 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7052 GOTO(out_mnt, rc = -EINVAL);
7055 #ifdef LDISKFS_MOUNT_DIRDATA
7056 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7057 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7058 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7059 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7060 "enabling the dirdata feature. If you do not want to "
7061 "downgrade to Lustre-1.x again, you can enable it via "
7062 "'tune2fs -O dirdata device'\n", name, dev);
7064 inode = osd_sb(o)->s_root->d_inode;
7065 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7066 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7068 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7072 if (lmd_flags & LMD_FLG_NOSCRUB)
7083 __free_page(__page);
7088 static struct lu_device *osd_device_fini(const struct lu_env *env,
7089 struct lu_device *d)
7091 struct osd_device *o = osd_dev(d);
7094 osd_shutdown(env, o);
7096 osd_scrub_cleanup(env, o);
7097 osd_obj_map_fini(o);
7103 static int osd_device_init0(const struct lu_env *env,
7104 struct osd_device *o,
7105 struct lustre_cfg *cfg)
7107 struct lu_device *l = osd2lu_dev(o);
7108 struct osd_thread_info *info;
7112 /* if the module was re-loaded, env can loose its keys */
7113 rc = lu_env_refill((struct lu_env *) env);
7116 info = osd_oti_get(env);
7119 l->ld_ops = &osd_lu_ops;
7120 o->od_dt_dev.dd_ops = &osd_dt_ops;
7122 spin_lock_init(&o->od_osfs_lock);
7123 mutex_init(&o->od_otable_mutex);
7124 INIT_LIST_HEAD(&o->od_orphan_list);
7126 o->od_read_cache = 1;
7127 o->od_writethrough_cache = 1;
7128 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7130 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7131 sizeof(o->od_svname));
7132 if (cplen >= sizeof(o->od_svname)) {
7137 o->od_index = -1; /* -1 means index is invalid */
7138 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7139 if (rc == LDD_F_SV_TYPE_OST)
7142 o->od_full_scrub_ratio = OFSR_DEFAULT;
7143 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7144 rc = osd_mount(env, o, cfg);
7148 rc = osd_obj_map_init(env, o);
7152 rc = lu_site_init(&o->od_site, l);
7154 GOTO(out_compat, rc);
7155 o->od_site.ls_bottom_dev = l;
7157 rc = lu_site_init_finish(&o->od_site);
7161 INIT_LIST_HEAD(&o->od_ios_list);
7162 /* setup scrub, including OI files initialization */
7163 rc = osd_scrub_setup(env, o);
7167 rc = osd_procfs_init(o, o->od_svname);
7169 CERROR("%s: can't initialize procfs: rc = %d\n",
7171 GOTO(out_scrub, rc);
7174 LASSERT(l->ld_site->ls_linkage.next != NULL);
7175 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7177 /* initialize quota slave instance */
7178 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7180 if (IS_ERR(o->od_quota_slave)) {
7181 rc = PTR_ERR(o->od_quota_slave);
7182 o->od_quota_slave = NULL;
7183 GOTO(out_procfs, rc);
7191 osd_scrub_cleanup(env, o);
7193 lu_site_fini(&o->od_site);
7195 osd_obj_map_fini(o);
7202 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7203 struct lu_device_type *t,
7204 struct lustre_cfg *cfg)
7206 struct osd_device *o;
7211 return ERR_PTR(-ENOMEM);
7213 rc = dt_device_init(&o->od_dt_dev, t);
7215 /* Because the ctx might be revived in dt_device_init,
7216 * refill the env here */
7217 lu_env_refill((struct lu_env *)env);
7218 rc = osd_device_init0(env, o, cfg);
7220 dt_device_fini(&o->od_dt_dev);
7223 if (unlikely(rc != 0))
7226 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7229 static struct lu_device *osd_device_free(const struct lu_env *env,
7230 struct lu_device *d)
7232 struct osd_device *o = osd_dev(d);
7235 /* XXX: make osd top device in order to release reference */
7236 d->ld_site->ls_top_dev = d;
7237 lu_site_purge(env, d->ld_site, -1);
7238 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7239 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7240 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7242 lu_site_fini(&o->od_site);
7243 dt_device_fini(&o->od_dt_dev);
7248 static int osd_process_config(const struct lu_env *env,
7249 struct lu_device *d, struct lustre_cfg *cfg)
7251 struct osd_device *o = osd_dev(d);
7255 switch (cfg->lcfg_command) {
7257 rc = osd_mount(env, o, cfg);
7260 lu_dev_del_linkage(d->ld_site, d);
7261 rc = osd_shutdown(env, o);
7264 LASSERT(&o->od_dt_dev);
7265 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7266 cfg, &o->od_dt_dev);
7267 if (rc > 0 || rc == -ENOSYS) {
7268 rc = class_process_proc_param(PARAM_OST,
7269 lprocfs_osd_obd_vars,
7270 cfg, &o->od_dt_dev);
7282 static int osd_recovery_complete(const struct lu_env *env,
7283 struct lu_device *d)
7285 struct osd_device *osd = osd_dev(d);
7289 if (osd->od_quota_slave == NULL)
7292 /* start qsd instance on recovery completion, this notifies the quota
7293 * slave code that we are about to process new requests now */
7294 rc = qsd_start(env, osd->od_quota_slave);
7299 * we use exports to track all osd users
7301 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7302 struct obd_device *obd, struct obd_uuid *cluuid,
7303 struct obd_connect_data *data, void *localdata)
7305 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7306 struct lustre_handle conn;
7310 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7312 rc = class_connect(&conn, obd, cluuid);
7316 *exp = class_conn2export(&conn);
7318 spin_lock(&osd->od_osfs_lock);
7320 spin_unlock(&osd->od_osfs_lock);
7326 * once last export (we don't count self-export) disappeared
7327 * osd can be released
7329 static int osd_obd_disconnect(struct obd_export *exp)
7331 struct obd_device *obd = exp->exp_obd;
7332 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7333 int rc, release = 0;
7336 /* Only disconnect the underlying layers on the final disconnect. */
7337 spin_lock(&osd->od_osfs_lock);
7339 if (osd->od_connects == 0)
7341 spin_unlock(&osd->od_osfs_lock);
7343 rc = class_disconnect(exp); /* bz 9811 */
7345 if (rc == 0 && release)
7346 class_manual_cleanup(obd);
7350 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7351 struct lu_device *dev)
7353 struct osd_device *osd = osd_dev(dev);
7354 struct lr_server_data *lsd =
7355 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7359 if (osd->od_quota_slave != NULL) {
7360 /* set up quota slave objects */
7361 result = qsd_prepare(env, osd->od_quota_slave);
7366 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7367 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7368 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7369 osd->od_index_in_idif = 1;
7371 osd->od_index_in_idif = 0;
7372 result = osd_register_proc_index_in_idif(osd);
7377 osd->od_index_in_idif = 1;
7381 result = osd_fid_init(env, osd);
7386 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7387 struct lu_fid *fid, struct md_op_data *op_data)
7389 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7391 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7394 static const struct lu_object_operations osd_lu_obj_ops = {
7395 .loo_object_init = osd_object_init,
7396 .loo_object_delete = osd_object_delete,
7397 .loo_object_release = osd_object_release,
7398 .loo_object_free = osd_object_free,
7399 .loo_object_print = osd_object_print,
7400 .loo_object_invariant = osd_object_invariant
7403 const struct lu_device_operations osd_lu_ops = {
7404 .ldo_object_alloc = osd_object_alloc,
7405 .ldo_process_config = osd_process_config,
7406 .ldo_recovery_complete = osd_recovery_complete,
7407 .ldo_prepare = osd_prepare,
7410 static const struct lu_device_type_operations osd_device_type_ops = {
7411 .ldto_init = osd_type_init,
7412 .ldto_fini = osd_type_fini,
7414 .ldto_start = osd_type_start,
7415 .ldto_stop = osd_type_stop,
7417 .ldto_device_alloc = osd_device_alloc,
7418 .ldto_device_free = osd_device_free,
7420 .ldto_device_init = osd_device_init,
7421 .ldto_device_fini = osd_device_fini
7424 static struct lu_device_type osd_device_type = {
7425 .ldt_tags = LU_DEVICE_DT,
7426 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
7427 .ldt_ops = &osd_device_type_ops,
7428 .ldt_ctx_tags = LCT_LOCAL,
7431 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
7433 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7434 struct super_block *sb = osd_sb(osd);
7436 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
7440 * lprocfs legacy support.
7442 static struct obd_ops osd_obd_device_ops = {
7443 .o_owner = THIS_MODULE,
7444 .o_connect = osd_obd_connect,
7445 .o_disconnect = osd_obd_disconnect,
7446 .o_fid_alloc = osd_fid_alloc,
7447 .o_health_check = osd_health_check,
7450 static int __init osd_init(void)
7454 LASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
7455 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
7456 /* please, try to keep osd_thread_info smaller than a page */
7457 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
7462 rc = lu_kmem_init(ldiskfs_caches);
7466 #ifdef CONFIG_KALLSYMS
7467 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
7468 priv_dev_check_rdonly = (void *)kallsyms_lookup_name("dev_check_rdonly");
7469 /* Clear readonly is unused at this time */
7470 /*priv_dev_clear_rdonly = (void *)kallsyms_lookup_name("dev_clear_rdonly");*/
7473 rc = class_register_type(&osd_obd_device_ops, NULL, true,
7474 lprocfs_osd_module_vars,
7475 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
7477 lu_kmem_fini(ldiskfs_caches);
7481 static void __exit osd_exit(void)
7483 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
7484 lu_kmem_fini(ldiskfs_caches);
7487 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7488 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
7489 MODULE_VERSION(LUSTRE_VERSION_STRING);
7490 MODULE_LICENSE("GPL");
7492 module_init(osd_init);
7493 module_exit(osd_exit);