4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd/osd_handler.c
34 * Top-level entry points into osd module
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #ifdef HAVE_UIDGID_HEADER
45 # include <linux/uidgid.h>
48 /* LUSTRE_VERSION_CODE */
49 #include <lustre_ver.h>
50 /* prerequisite for linux/xattr.h */
51 #include <linux/types.h>
52 /* prerequisite for linux/xattr.h */
54 /* XATTR_{REPLACE,CREATE} */
55 #include <linux/xattr.h>
57 #include <ldiskfs/ldiskfs.h>
58 #include <ldiskfs/xattr.h>
59 #include <ldiskfs/ldiskfs_extents.h>
62 * struct OBD_{ALLOC,FREE}*()
65 #include <obd_support.h>
66 /* struct ptlrpc_thread */
67 #include <lustre_net.h>
68 #include <lustre_fid.h>
70 #include <uapi/linux/lustre_param.h>
72 #include "osd_internal.h"
73 #include "osd_dynlocks.h"
75 /* llo_* api support */
76 #include <md_object.h>
77 #include <lustre_quota.h>
79 #include <lustre_linkea.h>
81 #define PFID_STRIPE_IDX_BITS 16
82 #define PFID_STRIPE_COUNT_MASK ((1 << PFID_STRIPE_IDX_BITS) - 1)
85 module_param(ldiskfs_pdo, int, 0644);
86 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
88 int ldiskfs_track_declares_assert;
89 module_param(ldiskfs_track_declares_assert, int, 0644);
90 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
92 /* Slab to allocate dynlocks */
93 struct kmem_cache *dynlock_cachep;
95 /* Slab to allocate osd_it_ea */
96 struct kmem_cache *osd_itea_cachep;
98 static struct lu_kmem_descr ldiskfs_caches[] = {
100 .ckd_cache = &dynlock_cachep,
101 .ckd_name = "dynlock_cache",
102 .ckd_size = sizeof(struct dynlock_handle)
105 .ckd_cache = &osd_itea_cachep,
106 .ckd_name = "osd_itea_cache",
107 .ckd_size = sizeof(struct osd_it_ea)
114 static const char dot[] = ".";
115 static const char dotdot[] = "..";
117 static const struct lu_object_operations osd_lu_obj_ops;
118 static const struct dt_object_operations osd_obj_ops;
119 static const struct dt_object_operations osd_obj_otable_it_ops;
120 static const struct dt_index_operations osd_index_iam_ops;
121 static const struct dt_index_operations osd_index_ea_ops;
123 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
124 const struct lu_fid *fid);
125 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
126 struct osd_device *osd);
128 int osd_trans_declare_op2rb[] = {
129 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
130 [OSD_OT_PUNCH] = OSD_OT_MAX,
131 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
132 [OSD_OT_CREATE] = OSD_OT_DESTROY,
133 [OSD_OT_DESTROY] = OSD_OT_CREATE,
134 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
135 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
136 [OSD_OT_WRITE] = OSD_OT_WRITE,
137 [OSD_OT_INSERT] = OSD_OT_DELETE,
138 [OSD_OT_DELETE] = OSD_OT_INSERT,
139 [OSD_OT_QUOTA] = OSD_OT_MAX,
142 static int osd_has_index(const struct osd_object *obj)
144 return obj->oo_dt.do_index_ops != NULL;
147 static int osd_object_invariant(const struct lu_object *l)
149 return osd_invariant(osd_obj(l));
153 * Concurrency: doesn't matter
155 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
157 struct osd_thread_info *oti = osd_oti_get(env);
158 return oti->oti_w_locks > 0 && o->oo_owner == env;
162 * Concurrency: doesn't access mutable data
164 static int osd_root_get(const struct lu_env *env,
165 struct dt_device *dev, struct lu_fid *f)
167 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
172 * the following set of functions are used to maintain per-thread
173 * cache of FID->ino mapping. this mechanism is needed to resolve
174 * FID to inode at dt_insert() which in turn stores ino in the
175 * directory entries to keep ldiskfs compatible with ext[34].
176 * due to locking-originated restrictions we can't lookup ino
177 * using LU cache (deadlock is possible). lookup using OI is quite
178 * expensive. so instead we maintain this cache and methods like
179 * dt_create() fill it. so in the majority of cases dt_insert() is
180 * able to find needed mapping in lockless manner.
182 static struct osd_idmap_cache *
183 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
184 const struct lu_fid *fid)
186 struct osd_thread_info *oti = osd_oti_get(env);
187 struct osd_idmap_cache *idc = oti->oti_ins_cache;
189 for (i = 0; i < oti->oti_ins_cache_used; i++) {
190 if (!lu_fid_eq(&idc[i].oic_fid, fid))
192 if (idc[i].oic_dev != osd)
201 static struct osd_idmap_cache *
202 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
203 const struct lu_fid *fid)
205 struct osd_thread_info *oti = osd_oti_get(env);
206 struct osd_idmap_cache *idc;
209 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
210 i = oti->oti_ins_cache_size * 2;
212 i = OSD_INS_CACHE_SIZE;
213 OBD_ALLOC(idc, sizeof(*idc) * i);
215 return ERR_PTR(-ENOMEM);
216 if (oti->oti_ins_cache != NULL) {
217 memcpy(idc, oti->oti_ins_cache,
218 oti->oti_ins_cache_used * sizeof(*idc));
219 OBD_FREE(oti->oti_ins_cache,
220 oti->oti_ins_cache_used * sizeof(*idc));
222 oti->oti_ins_cache = idc;
223 oti->oti_ins_cache_size = i;
226 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
229 idc->oic_lid.oii_ino = 0;
230 idc->oic_lid.oii_gen = 0;
237 * lookup mapping for the given fid in the cache, initialize a
238 * new one if not found. the initialization checks whether the
239 * object is local or remote. for local objects, OI is used to
240 * learn ino/generation. the function is used when the caller
241 * has no information about the object, e.g. at dt_insert().
243 static struct osd_idmap_cache *
244 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
245 const struct lu_fid *fid)
247 struct osd_idmap_cache *idc;
250 idc = osd_idc_find(env, osd, fid);
251 LASSERT(!IS_ERR(idc));
255 /* new mapping is needed */
256 idc = osd_idc_add(env, osd, fid);
261 rc = osd_remote_fid(env, osd, fid);
262 if (unlikely(rc < 0))
266 /* the object is local, lookup in OI */
267 /* XXX: probably cheaper to lookup in LU first? */
268 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
270 if (unlikely(rc < 0)) {
271 CERROR("can't lookup: rc = %d\n", rc);
275 /* the object is remote */
283 * lookup mapping for given FID and fill it from the given object.
284 * the object is lolcal by definition.
286 static int osd_idc_find_and_init(const struct lu_env *env,
287 struct osd_device *osd,
288 struct osd_object *obj)
290 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
291 struct osd_idmap_cache *idc;
293 idc = osd_idc_find(env, osd, fid);
294 LASSERT(!IS_ERR(idc));
296 if (obj->oo_inode == NULL)
298 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
299 LASSERT(idc->oic_lid.oii_ino == 0);
300 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
301 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
306 /* new mapping is needed */
307 idc = osd_idc_add(env, osd, fid);
311 if (obj->oo_inode != NULL) {
312 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
313 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
319 * OSD object methods.
323 * Concurrency: no concurrent access is possible that early in object
326 static struct lu_object *osd_object_alloc(const struct lu_env *env,
327 const struct lu_object_header *hdr,
330 struct osd_object *mo;
336 l = &mo->oo_dt.do_lu;
337 dt_object_init(&mo->oo_dt, NULL, d);
338 mo->oo_dt.do_ops = &osd_obj_ops;
339 l->lo_ops = &osd_lu_obj_ops;
340 init_rwsem(&mo->oo_sem);
341 init_rwsem(&mo->oo_ext_idx_sem);
342 spin_lock_init(&mo->oo_guard);
343 INIT_LIST_HEAD(&mo->oo_xattr_list);
350 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
351 struct dentry *dentry, struct lustre_ost_attrs *loa)
355 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
356 (void *)loa, sizeof(*loa));
358 struct lustre_mdt_attrs *lma = &loa->loa_lma;
360 if (rc < sizeof(*lma))
364 lustre_loa_swab(loa, true);
365 /* Check LMA compatibility */
366 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
367 CWARN("%s: unsupported incompat LMA feature(s) %#x "
368 "for fid = "DFID", ino = %lu\n",
370 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
371 PFID(&lma->lma_self_fid), inode->i_ino);
374 } else if (rc == 0) {
382 * retrieve object from backend ext fs.
384 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
385 struct osd_inode_id *id)
388 struct inode *inode = NULL;
390 /* if we look for an inode withing a running
391 * transaction, then we risk to deadlock */
392 /* osd_dirent_check_repair() breaks this */
393 /*LASSERT(current->journal_info == NULL);*/
395 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
397 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
398 id->oii_ino, PTR_ERR(inode));
399 } else if (id->oii_gen != OSD_OII_NOGEN &&
400 inode->i_generation != id->oii_gen) {
401 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
402 "i_generation = %u\n",
403 id->oii_ino, id->oii_gen, inode->i_generation);
405 inode = ERR_PTR(-ESTALE);
406 } else if (inode->i_nlink == 0) {
407 /* due to parallel readdir and unlink,
408 * we can have dead inode here. */
409 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
411 inode = ERR_PTR(-ESTALE);
412 } else if (is_bad_inode(inode)) {
413 CWARN("%s: bad inode: ino = %u\n",
414 osd_dev2name(dev), id->oii_ino);
416 inode = ERR_PTR(-ENOENT);
417 } else if ((rc = osd_attach_jinode(inode))) {
421 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
422 if (id->oii_gen == OSD_OII_NOGEN)
423 osd_id_gen(id, inode->i_ino, inode->i_generation);
425 /* Do not update file c/mtime in ldiskfs.
426 * NB: we don't have any lock to protect this because we don't
427 * have reference on osd_object now, but contention with
428 * another lookup + attr_set can't happen in the tiny window
429 * between if (...) and set S_NOCMTIME. */
430 if (!(inode->i_flags & S_NOCMTIME))
431 inode->i_flags |= S_NOCMTIME;
436 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
437 handle_t *handle, struct dentry *child,
438 struct inode *inode, struct htree_lock *hlock)
442 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
443 if (rc == -ENOBUFS || rc == -ENOSPC) {
444 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
445 struct inode *parent = child->d_parent->d_inode;
446 struct lu_fid *fid = NULL;
448 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
450 fid = &loa->loa_lma.lma_self_fid;
451 } else if (rc2 == -ENODATA) {
452 if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
453 fid = &info->oti_fid3;
454 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
455 } else if (!osd->od_is_ost && osd->od_index == 0) {
456 fid = &info->oti_fid3;
457 lu_igif_build(fid, parent->i_ino,
458 parent->i_generation);
463 CWARN("%s: directory (inode: %lu, FID: "DFID") %s "
464 "maximum entry limit\n",
465 osd_name(osd), parent->i_ino, PFID(fid),
466 rc == -ENOSPC ? "has reached" : "is approaching");
468 CWARN("%s: directory (inode: %lu, FID: unknown) %s "
469 "maximum entry limit\n",
470 osd_name(osd), parent->i_ino,
471 rc == -ENOSPC ? "has reached" : "is approaching");
473 /* ignore such error now */
482 static struct inode *
483 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
484 struct osd_inode_id *id, struct lu_fid *fid)
486 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
490 inode = osd_iget(info, dev, id);
494 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
496 *fid = loa->loa_lma.lma_self_fid;
497 } else if (rc == -ENODATA) {
498 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
499 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
501 lu_igif_build(fid, inode->i_ino, inode->i_generation);
509 static struct inode *osd_iget_check(struct osd_thread_info *info,
510 struct osd_device *dev,
511 const struct lu_fid *fid,
512 struct osd_inode_id *id,
519 /* The cached OI mapping is trustable. If we cannot locate the inode
520 * via the cached OI mapping, then return the failure to the caller
521 * directly without further OI checking. */
524 inode = ldiskfs_iget(osd_sb(dev), id->oii_ino);
527 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
530 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
531 PFID(fid), id->oii_ino, rc);
535 if (is_bad_inode(inode)) {
540 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
541 PFID(fid), id->oii_ino);
545 if (id->oii_gen != OSD_OII_NOGEN &&
546 inode->i_generation != id->oii_gen) {
551 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
552 "oii_gen = %u, i_generation = %u\n", PFID(fid),
553 id->oii_ino, id->oii_gen, inode->i_generation);
557 if (inode->i_nlink == 0) {
562 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
563 PFID(fid), id->oii_ino);
567 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
571 __u32 saved_ino = id->oii_ino;
572 __u32 saved_gen = id->oii_gen;
575 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
577 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
578 /* XXX: There are four possible cases:
580 * Backup/restore caused the OI invalid.
582 * Someone unlinked the object but NOT removed
583 * the OI mapping, such as mount target device
584 * as ldiskfs, and modify something directly.
586 * Someone just removed the object between the
587 * former oi_lookup and the iget. It is normal.
588 * 4. Other failure cases.
590 * Generally, when the device is mounted, it will
591 * auto check whether the system is restored from
592 * file-level backup or not. We trust such detect
593 * to distinguish the 1st case from the 2nd case:
594 * if the OI files are consistent but may contain
595 * stale OI mappings because of case 2, if iget()
596 * returns -ENOENT or -ESTALE, then it should be
599 /* If the OI mapping was in OI file before the
600 * osd_iget_check(), but now, it is disappear,
601 * then it must be removed by race. That is a
602 * normal race case. */
605 /* It is the OI scrub updated the OI mapping by race.
606 * The new OI mapping must be valid. */
607 if (saved_ino != id->oii_ino ||
608 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
617 if (dev->od_scrub.os_file.sf_flags & SF_INCONSISTENT)
618 /* It still can be the case 2, but we cannot
619 * distinguish it from the case 1. So return
620 * -EREMCHG to block current operation until
621 * OI scrub rebuilt the OI mappings. */
629 if (inode->i_generation == id->oii_gen)
634 if (id->oii_gen == OSD_OII_NOGEN)
635 osd_id_gen(id, inode->i_ino, inode->i_generation);
637 /* Do not update file c/mtime in ldiskfs.
638 * NB: we don't have any lock to protect this because we don't
639 * have reference on osd_object now, but contention with
640 * another lookup + attr_set can't happen in the tiny window
641 * between if (...) and set S_NOCMTIME. */
642 if (!(inode->i_flags & S_NOCMTIME))
643 inode->i_flags |= S_NOCMTIME;
660 * \retval +v: new filter_fid, does not contain self-fid
661 * \retval 0: filter_fid_old, contains self-fid
662 * \retval -v: other failure cases
664 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
665 struct dentry *dentry, struct lu_fid *fid)
667 struct filter_fid_old *ff = &info->oti_ff;
668 struct ost_id *ostid = &info->oti_ostid;
671 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
672 if (rc == sizeof(*ff)) {
674 ostid_set_seq(ostid, le64_to_cpu(ff->ff_seq));
675 rc = ostid_set_id(ostid, le64_to_cpu(ff->ff_objid));
677 * XXX: use 0 as the index for compatibility, the caller will
678 * handle index related issues when necessary.
681 ostid_to_fid(fid, ostid, 0);
682 } else if (rc == sizeof(struct filter_fid)) {
684 } else if (rc >= 0) {
691 static int osd_lma_self_repair(struct osd_thread_info *info,
692 struct osd_device *osd, struct inode *inode,
693 const struct lu_fid *fid, __u32 compat)
698 LASSERT(current->journal_info == NULL);
700 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
701 osd_dto_credits_noquota[DTO_XATTR_SET]);
704 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
709 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
711 CWARN("%s: cannot self repair the LMA: rc = %d\n",
713 ldiskfs_journal_stop(jh);
717 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
719 struct osd_thread_info *info = osd_oti_get(env);
720 struct osd_device *osd = osd_obj2dev(obj);
721 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
722 struct lustre_mdt_attrs *lma = &loa->loa_lma;
723 struct inode *inode = obj->oo_inode;
724 struct dentry *dentry = &info->oti_obj_dentry;
725 struct lu_fid *fid = NULL;
726 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
730 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
731 (void *)loa, sizeof(*loa));
732 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
733 fid = &lma->lma_self_fid;
734 rc = osd_get_idif(info, inode, dentry, fid);
735 if ((rc > 0) || (rc == -ENODATA && osd->od_index_in_idif)) {
736 /* For the given OST-object, if it has neither LMA nor
737 * FID in XATTR_NAME_FID, then the given FID (which is
738 * contained in the @obj, from client RPC for locating
739 * the OST-object) is trusted. We use it to generate
741 osd_lma_self_repair(info, osd, inode, rfid,
752 lustre_lma_swab(lma);
753 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
754 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
755 S_ISREG(inode->i_mode)))) {
756 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
757 "fid = "DFID", ino = %lu\n", osd_name(osd),
758 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
759 PFID(rfid), inode->i_ino);
762 fid = &lma->lma_self_fid;
766 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
767 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
768 struct ost_id *oi = &info->oti_ostid;
769 struct lu_fid *fid1 = &info->oti_fid3;
770 __u32 idx = fid_idif_ost_idx(rfid);
772 /* For old IDIF, the OST index is not part of the IDIF,
773 * Means that different OSTs may have the same IDIFs.
774 * Under such case, we need to make some compatible
775 * check to make sure to trigger OI scrub properly. */
776 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
777 /* Given @rfid is new, LMA is old. */
778 fid_to_ostid(fid, oi);
779 ostid_to_fid(fid1, oi, idx);
780 if (lu_fid_eq(fid1, rfid)) {
781 if (osd->od_index_in_idif)
782 osd_lma_self_repair(info, osd,
796 struct osd_check_lmv_buf {
797 #ifdef HAVE_DIR_CONTEXT
798 /* please keep it as first member */
799 struct dir_context ctx;
801 struct osd_thread_info *oclb_info;
802 struct osd_device *oclb_dev;
803 struct osd_idmap_cache *oclb_oic;
809 * It is called internally by ->readdir() to filter out the
810 * local slave object's FID of the striped directory.
812 * \retval 1 found the local slave's FID
813 * \retval 0 continue to check next item
814 * \retval -ve for failure
816 #ifdef HAVE_FILLDIR_USE_CTX
817 static int osd_stripe_dir_filldir(struct dir_context *buf,
819 static int osd_stripe_dir_filldir(void *buf,
821 const char *name, int namelen,
822 loff_t offset, __u64 ino, unsigned d_type)
824 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
825 struct osd_thread_info *oti = oclb->oclb_info;
826 struct lu_fid *fid = &oti->oti_fid3;
827 struct osd_inode_id *id = &oti->oti_id3;
828 struct osd_device *dev = oclb->oclb_dev;
829 struct osd_idmap_cache *oic = oclb->oclb_oic;
838 sscanf(name + 1, SFID, RFID(fid));
839 if (!fid_is_sane(fid))
842 if (osd_remote_fid(oti->oti_env, dev, fid))
845 osd_id_gen(id, ino, OSD_OII_NOGEN);
846 inode = osd_iget(oti, dev, id);
848 return PTR_ERR(inode);
851 osd_add_oi_cache(oti, dev, id, fid);
855 osd_oii_insert(dev, oic, true);
856 oclb->oclb_found = true;
861 /* When lookup item under striped directory, we need to locate the master
862 * MDT-object of the striped directory firstly, then the client will send
863 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
864 * and the item's name. If the system is restored from MDT file level backup,
865 * then before the OI scrub completely built the OI files, the OI mappings of
866 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
867 * not a problem for the master MDT-object. Because when locate the master
868 * MDT-object, we will do name based lookup (for the striped directory itself)
869 * firstly, during such process we can setup the correct OI mapping for the
870 * master MDT-object. But it will be trouble for the slave MDT-object. Because
871 * the client will not trigger name based lookup on the MDT to locate the slave
872 * MDT-object before locating item under the striped directory, then when
873 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
874 * is invalid and does not know what the right OI mapping is, then the MDT has
875 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
876 * the OI file, related OI mapping is unknown yet, please try again later. And
877 * then client will re-try the RPC again and again until related OI mapping has
878 * been updated. That is quite inefficient.
880 * To resolve above trouble, we will handle it as the following two cases:
882 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
883 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
884 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
885 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
886 * directly. Please check osd_fid_lookup().
888 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
889 * Under such case, during lookup the master MDT-object, we will lookup the
890 * slave MDT-object via readdir against the master MDT-object, because the
891 * slave MDT-objects information are stored as sub-directories with the name
892 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
893 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
894 * the correct OI mapping for the slave MDT-object. */
895 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
896 struct inode *inode, struct osd_idmap_cache *oic)
898 struct lu_buf *buf = &oti->oti_big_buf;
899 struct dentry *dentry = &oti->oti_obj_dentry;
900 struct file *filp = &oti->oti_file;
901 const struct file_operations *fops;
902 struct lmv_mds_md_v1 *lmv1;
903 struct osd_check_lmv_buf oclb = {
904 #ifdef HAVE_DIR_CONTEXT
905 .ctx.actor = osd_stripe_dir_filldir,
916 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
919 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
921 lu_buf_realloc(buf, rc);
922 if (buf->lb_buf == NULL)
923 GOTO(out, rc = -ENOMEM);
929 if (unlikely(rc == 0 || rc == -ENODATA))
935 if (unlikely(buf->lb_buf == NULL)) {
936 lu_buf_realloc(buf, rc);
937 if (buf->lb_buf == NULL)
938 GOTO(out, rc = -ENOMEM);
944 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
948 dentry->d_inode = inode;
949 dentry->d_sb = inode->i_sb;
951 filp->f_path.dentry = dentry;
952 filp->f_mode = FMODE_64BITHASH;
953 filp->f_mapping = inode->i_mapping;
955 filp->private_data = NULL;
956 set_file_inode(filp, inode);
960 #ifdef HAVE_DIR_CONTEXT
961 oclb.ctx.pos = filp->f_pos;
962 rc = fops->iterate(filp, &oclb.ctx);
963 filp->f_pos = oclb.ctx.pos;
965 rc = fops->readdir(filp, &oclb, osd_stripe_dir_filldir);
967 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
968 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
969 fops->release(inode, filp);
973 CDEBUG(D_LFSCK, "%s: fail to check LMV EA, inode = %lu/%u,"
974 DFID": rc = %d\n", osd_ino2name(inode),
975 inode->i_ino, inode->i_generation,
976 PFID(&oic->oic_fid), rc);
983 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
984 const struct lu_fid *fid,
985 const struct lu_object_conf *conf)
987 struct osd_thread_info *info;
988 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
989 struct osd_device *dev;
990 struct osd_idmap_cache *oic;
991 struct osd_inode_id *id;
992 struct inode *inode = NULL;
993 struct osd_scrub *scrub;
994 struct scrub_file *sf;
995 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1000 bool remote = false;
1001 bool trusted = true;
1002 bool updated = false;
1005 LINVRNT(osd_invariant(obj));
1006 LASSERT(obj->oo_inode == NULL);
1007 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1009 dev = osd_dev(ldev);
1010 scrub = &dev->od_scrub;
1011 sf = &scrub->os_file;
1012 info = osd_oti_get(env);
1014 oic = &info->oti_cache;
1016 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1019 /* For the object is created as locking anchor, or for the object to
1020 * be created on disk. No need to osd_oi_lookup() at here because FID
1021 * shouldn't never be re-used, if it's really a duplicate FID from
1022 * unexpected reason, we should be able to detect it later by calling
1023 * do_create->osd_oi_insert(). */
1024 if (conf && conf->loc_flags & LOC_F_NEW)
1025 GOTO(out, result = 0);
1027 /* Search order: 1. per-thread cache. */
1028 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1034 if (!list_empty(&scrub->os_inconsistent_items)) {
1035 /* Search order: 2. OI scrub pending list. */
1036 result = osd_oii_lookup(dev, fid, id);
1041 /* The OI mapping in the OI file can be updated by the OI scrub
1042 * when we locate the inode via FID. So it may be not trustable. */
1045 /* Search order: 3. OI files. */
1046 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1047 if (result == -ENOENT) {
1048 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1049 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1050 !ldiskfs_test_bit(osd_oi_fid2idx(dev,fid),
1052 GOTO(out, result = 0);
1061 obj->oo_inode = NULL;
1062 /* for later passes through checks, not true on first pass */
1063 if (!IS_ERR_OR_NULL(inode))
1066 inode = osd_iget_check(info, dev, fid, id, trusted);
1067 if (!IS_ERR(inode)) {
1068 obj->oo_inode = inode;
1076 result = PTR_ERR(inode);
1077 if (result == -ENOENT || result == -ESTALE)
1078 GOTO(out, result = 0);
1080 if (result != -EREMCHG)
1084 /* We still have chance to get the valid inode: for the
1085 * object which is referenced by remote name entry, the
1086 * object on the local MDT will be linked under the dir
1087 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1089 * We do not know whether the object for the given FID
1090 * is referenced by some remote name entry or not, and
1091 * especially for DNE II, a multiple-linked object may
1092 * have many name entries reside on many MDTs.
1094 * To simplify the operation, OSD will not distinguish
1095 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1096 * only happened for the RPC from other MDT during the
1097 * OI scrub, or for the client side RPC with FID only,
1098 * such as FID to path, or from old connected client. */
1100 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1104 flags |= SS_AUTO_PARTIAL;
1105 flags &= ~SS_AUTO_FULL;
1110 if (thread_is_running(&scrub->os_thread)) {
1111 if (scrub->os_partial_scan && !scrub->os_in_join)
1114 if (IS_ERR_OR_NULL(inode) || result)
1115 GOTO(out, result = -EINPROGRESS);
1118 LASSERT(obj->oo_inode == inode);
1120 osd_add_oi_cache(info, dev, id, fid);
1121 osd_oii_insert(dev, oic, true);
1125 if (dev->od_noscrub) {
1127 GOTO(out, result = -EREMCHG);
1130 LASSERT(obj->oo_inode == inode);
1132 osd_add_oi_cache(info, dev, id, fid);
1137 rc1 = osd_scrub_start(dev, flags);
1138 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1139 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1140 if (rc1 && rc1 != -EALREADY)
1141 GOTO(out, result = -EREMCHG);
1143 if (IS_ERR_OR_NULL(inode) || result)
1144 GOTO(out, result = -EINPROGRESS);
1147 LASSERT(obj->oo_inode == inode);
1149 osd_add_oi_cache(info, dev, id, fid);
1150 osd_oii_insert(dev, oic, true);
1154 result = osd_check_lma(env, obj);
1158 LASSERTF(id->oii_ino == inode->i_ino &&
1159 id->oii_gen == inode->i_generation,
1160 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1161 PFID(fid), id->oii_ino, id->oii_gen,
1162 inode->i_ino, inode->i_generation);
1164 saved_ino = inode->i_ino;
1165 saved_gen = inode->i_generation;
1167 if (unlikely(result == -ENODATA)) {
1168 /* If the OI scrub updated the OI mapping by race, it
1169 * must be valid. Trust the inode that has no LMA EA. */
1173 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1175 /* The OI mapping is still there, the inode is still
1176 * valid. It is just becaues the inode has no LMA EA. */
1177 if (saved_ino == id->oii_ino &&
1178 saved_gen == id->oii_gen)
1181 /* It is the OI scrub updated the OI mapping by race.
1182 * The new OI mapping must be valid. */
1188 /* "result == -ENOENT" means that the OI mappinghas been
1189 * removed by race, so the inode belongs to other object.
1191 * Others error can be returned directly. */
1192 if (result == -ENOENT) {
1195 obj->oo_inode = NULL;
1200 if (result != -EREMCHG)
1205 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1206 /* "result == -ENOENT" means the cached OI mapping has been removed
1207 * from the OI file by race, above inode belongs to other object. */
1208 if (result == -ENOENT) {
1211 obj->oo_inode = NULL;
1212 GOTO(out, result = 0);
1218 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1223 /* It is the OI scrub updated the OI mapping by race.
1224 * The new OI mapping must be valid. */
1230 obj->oo_compat_dot_created = 1;
1231 obj->oo_compat_dotdot_created = 1;
1233 if (S_ISDIR(inode->i_mode) &&
1234 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1235 osd_check_lmv(info, dev, inode, oic);
1237 result = osd_attach_jinode(inode);
1242 GOTO(out, result = 0);
1244 LASSERT(!obj->oo_hl_head);
1245 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1247 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1250 if (result || !obj->oo_inode) {
1251 if (!IS_ERR_OR_NULL(inode))
1254 obj->oo_inode = NULL;
1256 fid_zero(&oic->oic_fid);
1259 LINVRNT(osd_invariant(obj));
1264 * Concurrency: shouldn't matter.
1266 static void osd_object_init0(struct osd_object *obj)
1268 LASSERT(obj->oo_inode != NULL);
1269 obj->oo_dt.do_body_ops = &osd_body_ops;
1270 obj->oo_dt.do_lu.lo_header->loh_attr |=
1271 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1275 * Concurrency: no concurrent access is possible that early in object
1278 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1279 const struct lu_object_conf *conf)
1281 struct osd_object *obj = osd_obj(l);
1284 LINVRNT(osd_invariant(obj));
1286 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1287 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1288 l->lo_header->loh_attr |= LOHA_EXISTS;
1292 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1293 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1294 if (result == 0 && obj->oo_inode != NULL) {
1295 struct osd_thread_info *oti = osd_oti_get(env);
1296 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1298 osd_object_init0(obj);
1299 result = osd_get_lma(oti, obj->oo_inode,
1300 &oti->oti_obj_dentry, loa);
1302 /* Convert LMAI flags to lustre LMA flags
1303 * and cache it to oo_lma_flags */
1305 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1306 } else if (result == -ENODATA) {
1311 LINVRNT(osd_invariant(obj));
1315 /* The first part of oxe_buf is xattr name, and is '\0' terminated.
1316 * The left part is for value, binary mode. */
1317 struct osd_xattr_entry {
1318 struct list_head oxe_list;
1322 struct rcu_head oxe_rcu;
1326 static int osd_oxc_get(struct osd_object *obj, const char *name,
1329 struct osd_xattr_entry *tmp;
1330 struct osd_xattr_entry *oxe = NULL;
1331 size_t namelen = strlen(name);
1336 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1337 if (namelen == tmp->oxe_namelen &&
1338 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1345 GOTO(out, rc = -ENOENT);
1347 if (!oxe->oxe_exist)
1348 GOTO(out, rc = -ENODATA);
1351 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1354 if (buf->lb_buf == NULL)
1357 if (buf->lb_len < rc)
1358 GOTO(out, rc = -ERANGE);
1360 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1368 static void osd_oxc_free(struct rcu_head *head)
1370 struct osd_xattr_entry *oxe;
1372 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1373 OBD_FREE(oxe, oxe->oxe_len);
1376 static void osd_oxc_add(struct osd_object *obj, const char *name,
1377 const char *buf, int buflen)
1379 struct osd_xattr_entry *oxe;
1380 struct osd_xattr_entry *old = NULL;
1381 struct osd_xattr_entry *tmp;
1382 size_t namelen = strlen(name);
1383 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1385 OBD_ALLOC(oxe, len);
1389 INIT_LIST_HEAD(&oxe->oxe_list);
1391 oxe->oxe_namelen = namelen;
1392 memcpy(oxe->oxe_buf, name, namelen);
1394 LASSERT(buf != NULL);
1395 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1396 oxe->oxe_exist = true;
1398 oxe->oxe_exist = false;
1401 /* this should be rarely called, just remove old and add new */
1402 spin_lock(&obj->oo_guard);
1403 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1404 if (namelen == tmp->oxe_namelen &&
1405 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1411 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1412 call_rcu(&old->oxe_rcu, osd_oxc_free);
1414 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1416 spin_unlock(&obj->oo_guard);
1419 static void osd_oxc_del(struct osd_object *obj, const char *name)
1421 struct osd_xattr_entry *oxe;
1422 size_t namelen = strlen(name);
1424 spin_lock(&obj->oo_guard);
1425 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1426 if (namelen == oxe->oxe_namelen &&
1427 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1428 list_del_rcu(&oxe->oxe_list);
1429 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1433 spin_unlock(&obj->oo_guard);
1436 static void osd_oxc_fini(struct osd_object *obj)
1438 struct osd_xattr_entry *oxe, *next;
1440 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1441 list_del(&oxe->oxe_list);
1442 OBD_FREE(oxe, oxe->oxe_len);
1447 * Concurrency: no concurrent access is possible that late in object
1450 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1452 struct osd_object *obj = osd_obj(l);
1454 LINVRNT(osd_invariant(obj));
1457 dt_object_fini(&obj->oo_dt);
1458 if (obj->oo_hl_head != NULL)
1459 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1464 * Concurrency: no concurrent access is possible that late in object
1467 static void osd_index_fini(struct osd_object *o)
1469 struct iam_container *bag;
1471 if (o->oo_dir != NULL) {
1472 bag = &o->oo_dir->od_container;
1473 if (o->oo_inode != NULL) {
1474 if (bag->ic_object == o->oo_inode)
1475 iam_container_fini(bag);
1477 OBD_FREE_PTR(o->oo_dir);
1483 * Concurrency: no concurrent access is possible that late in object
1484 * life-cycle (for all existing callers, that is. New callers have to provide
1485 * their own locking.)
1487 static int osd_inode_unlinked(const struct inode *inode)
1489 return inode->i_nlink == 0;
1493 OSD_TXN_OI_DELETE_CREDITS = 20,
1494 OSD_TXN_INODE_DELETE_CREDITS = 20
1501 #if OSD_THANDLE_STATS
1503 * Set time when the handle is allocated
1505 static void osd_th_alloced(struct osd_thandle *oth)
1507 oth->oth_alloced = ktime_get();
1511 * Set time when the handle started
1513 static void osd_th_started(struct osd_thandle *oth)
1515 oth->oth_started = ktime_get();
1519 * Check whether the we deal with this handle for too long.
1521 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1522 ktime_t alloced, ktime_t started,
1525 ktime_t now = ktime_get();
1527 LASSERT(dev != NULL);
1529 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1530 ktime_us_delta(started, alloced));
1531 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1532 ktime_us_delta(closed, started));
1533 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1534 ktime_us_delta(now, closed));
1536 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1537 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1538 oth, now, alloced, started, closed);
1539 libcfs_debug_dumpstack(NULL);
1543 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1545 ktime_t __closed = ktime_get(); \
1546 ktime_t __alloced = oth->oth_alloced; \
1547 ktime_t __started = oth->oth_started; \
1550 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1553 #else /* OSD_THANDLE_STATS */
1555 #define osd_th_alloced(h) do {} while(0)
1556 #define osd_th_started(h) do {} while(0)
1557 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1559 #endif /* OSD_THANDLE_STATS */
1562 * Concurrency: doesn't access mutable data.
1564 static int osd_param_is_not_sane(const struct osd_device *dev,
1565 const struct thandle *th)
1567 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1569 return oh->ot_credits > osd_transaction_size(dev);
1573 * Concurrency: shouldn't matter.
1575 static void osd_trans_commit_cb(struct super_block *sb,
1576 struct ldiskfs_journal_cb_entry *jcb, int error)
1578 struct osd_thandle *oh = container_of0(jcb, struct osd_thandle, ot_jcb);
1579 struct thandle *th = &oh->ot_super;
1580 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1581 struct dt_txn_commit_cb *dcb, *tmp;
1583 LASSERT(oh->ot_handle == NULL);
1586 CERROR("transaction @0x%p commit error: %d\n", th, error);
1588 dt_txn_hook_commit(th);
1590 /* call per-transaction callbacks if any */
1591 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1593 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1594 "commit callback entry: magic=%x name='%s'\n",
1595 dcb->dcb_magic, dcb->dcb_name);
1596 list_del_init(&dcb->dcb_linkage);
1597 dcb->dcb_func(NULL, th, dcb, error);
1600 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1607 #ifndef HAVE_SB_START_WRITE
1608 # define sb_start_write(sb) do {} while (0)
1609 # define sb_end_write(sb) do {} while (0)
1612 static struct thandle *osd_trans_create(const struct lu_env *env,
1613 struct dt_device *d)
1615 struct osd_thread_info *oti = osd_oti_get(env);
1616 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1617 struct osd_thandle *oh;
1622 CERROR("%s: someone try to start transaction under "
1623 "readonly mode, should be disabled.\n",
1624 osd_name(osd_dt_dev(d)));
1626 RETURN(ERR_PTR(-EROFS));
1629 /* on pending IO in this thread should left from prev. request */
1630 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1632 sb_start_write(osd_sb(osd_dt_dev(d)));
1634 OBD_ALLOC_GFP(oh, sizeof *oh, GFP_NOFS);
1636 oh->ot_quota_trans = &oti->oti_quota_trans;
1637 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1642 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1643 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1646 memset(oti->oti_declare_ops, 0,
1647 sizeof(oti->oti_declare_ops));
1648 memset(oti->oti_declare_ops_cred, 0,
1649 sizeof(oti->oti_declare_ops_cred));
1650 memset(oti->oti_declare_ops_used, 0,
1651 sizeof(oti->oti_declare_ops_used));
1653 sb_end_write(osd_sb(osd_dt_dev(d)));
1654 th = ERR_PTR(-ENOMEM);
1659 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1661 struct osd_thread_info *oti = osd_oti_get(env);
1662 struct osd_thandle *oh;
1664 oh = container_of0(th, struct osd_thandle, ot_super);
1665 LASSERT(oh != NULL);
1667 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1668 oti->oti_declare_ops[OSD_OT_CREATE],
1669 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1670 oti->oti_declare_ops_used[OSD_OT_CREATE],
1671 oti->oti_declare_ops[OSD_OT_DESTROY],
1672 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1673 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1674 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1675 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1676 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1677 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1678 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1679 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1680 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1681 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1682 oti->oti_declare_ops[OSD_OT_WRITE],
1683 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1684 oti->oti_declare_ops_used[OSD_OT_WRITE],
1685 oti->oti_declare_ops[OSD_OT_PUNCH],
1686 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1687 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1688 oti->oti_declare_ops[OSD_OT_QUOTA],
1689 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1690 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1691 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1692 oti->oti_declare_ops[OSD_OT_INSERT],
1693 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1694 oti->oti_declare_ops_used[OSD_OT_INSERT],
1695 oti->oti_declare_ops[OSD_OT_DELETE],
1696 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1697 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1698 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1699 oti->oti_declare_ops[OSD_OT_REF_ADD],
1700 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1701 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1702 oti->oti_declare_ops[OSD_OT_REF_DEL],
1703 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1704 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1708 * Concurrency: shouldn't matter.
1710 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1713 struct osd_thread_info *oti = osd_oti_get(env);
1714 struct osd_device *dev = osd_dt_dev(d);
1716 struct osd_thandle *oh;
1721 LASSERT(current->journal_info == NULL);
1723 oh = container_of0(th, struct osd_thandle, ot_super);
1724 LASSERT(oh != NULL);
1725 LASSERT(oh->ot_handle == NULL);
1727 rc = dt_txn_hook_start(env, d, th);
1731 if (unlikely(osd_param_is_not_sane(dev, th))) {
1732 static unsigned long last_printed;
1733 static int last_credits;
1735 /* don't make noise on a tiny testing systems
1736 * actual credits misuse will be caught anyway */
1737 if (last_credits != oh->ot_credits &&
1738 time_after(jiffies, last_printed +
1739 msecs_to_jiffies(60 * MSEC_PER_SEC)) &&
1740 osd_transaction_size(dev) > 512) {
1741 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1742 oh->ot_credits, osd_transaction_size(dev));
1743 osd_trans_dump_creds(env, th);
1744 libcfs_debug_dumpstack(NULL);
1745 last_credits = oh->ot_credits;
1746 last_printed = jiffies;
1748 /* XXX Limit the credits to 'max_transaction_buffers', and
1749 * let the underlying filesystem to catch the error if
1750 * we really need so many credits.
1752 * This should be removed when we can calculate the
1753 * credits precisely. */
1754 oh->ot_credits = osd_transaction_size(dev);
1755 } else if (ldiskfs_track_declares_assert != 0) {
1756 /* reserve few credits to prevent an assertion in JBD
1757 * our debugging mechanism will be able to detected
1758 * overuse. this can help to debug single-update
1760 oh->ot_credits += 10;
1761 if (unlikely(osd_param_is_not_sane(dev, th)))
1762 oh->ot_credits = osd_transaction_size(dev);
1766 * XXX temporary stuff. Some abstraction layer should
1769 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1773 LASSERT(oti->oti_txns == 0);
1775 lu_device_get(&d->dd_lu_dev);
1776 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1787 static int osd_seq_exists(const struct lu_env *env,
1788 struct osd_device *osd, u64 seq)
1790 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1791 struct seq_server_site *ss = osd_seq_site(osd);
1795 LASSERT(ss != NULL);
1796 LASSERT(ss->ss_server_fld != NULL);
1798 rc = osd_fld_lookup(env, osd, seq, range);
1801 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
1802 osd_name(osd), seq, rc);
1806 RETURN(ss->ss_node_id == range->lsr_index);
1809 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
1811 struct dt_txn_commit_cb *dcb;
1812 struct dt_txn_commit_cb *tmp;
1814 /* call per-transaction stop callbacks if any */
1815 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
1817 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1818 "commit callback entry: magic=%x name='%s'\n",
1819 dcb->dcb_magic, dcb->dcb_name);
1820 list_del_init(&dcb->dcb_linkage);
1821 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
1826 * Concurrency: shouldn't matter.
1828 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
1831 int rc = 0, remove_agents = 0;
1832 struct osd_thandle *oh;
1833 struct osd_thread_info *oti = osd_oti_get(env);
1834 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1835 struct osd_device *osd = osd_dt_dev(th->th_dev);
1836 struct qsd_instance *qsd = osd->od_quota_slave;
1837 struct lquota_trans *qtrans;
1840 oh = container_of0(th, struct osd_thandle, ot_super);
1842 remove_agents = oh->ot_remove_agents;
1844 qtrans = oh->ot_quota_trans;
1845 oh->ot_quota_trans = NULL;
1847 if (oh->ot_handle != NULL) {
1849 handle_t *hdl = oh->ot_handle;
1852 * add commit callback
1853 * notice we don't do this in osd_trans_start()
1854 * as underlying transaction can change during truncate
1856 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
1859 LASSERT(oti->oti_txns == 1);
1862 rc = dt_txn_hook_stop(env, th);
1864 CERROR("%s: failed in transaction hook: rc = %d\n",
1867 osd_trans_stop_cb(oh, rc);
1868 /* hook functions might modify th_sync */
1869 hdl->h_sync = th->th_sync;
1871 oh->ot_handle = NULL;
1872 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
1874 CERROR("%s: failed to stop transaction: rc = %d\n",
1875 osd_name(osd), rc2);
1879 osd_trans_stop_cb(oh, th->th_result);
1883 /* inform the quota slave device that the transaction is stopping */
1884 qsd_op_end(env, qsd, qtrans);
1886 /* as we want IO to journal and data IO be concurrent, we don't block
1887 * awaiting data IO completion in osd_do_bio(), instead we wait here
1888 * once transaction is submitted to the journal. all reqular requests
1889 * don't do direct IO (except read/write), thus this wait_event becomes
1892 * IMPORTANT: we have to wait till any IO submited by the thread is
1893 * completed otherwise iobuf may be corrupted by different request
1895 wait_event(iobuf->dr_wait,
1896 atomic_read(&iobuf->dr_numreqs) == 0);
1897 osd_fini_iobuf(osd, iobuf);
1899 rc = iobuf->dr_error;
1901 if (unlikely(remove_agents != 0))
1902 osd_process_scheduled_agent_removals(env, osd);
1904 /* reset OI cache for safety */
1905 oti->oti_ins_cache_used = 0;
1907 sb_end_write(osd_sb(osd));
1912 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
1914 struct osd_thandle *oh = container_of0(th, struct osd_thandle,
1917 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
1918 LASSERT(&dcb->dcb_func != NULL);
1919 if (dcb->dcb_flags & DCB_TRANS_STOP)
1920 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
1922 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
1928 * Called just before object is freed. Releases all resources except for
1929 * object itself (that is released by osd_object_free()).
1931 * Concurrency: no concurrent access is possible that late in object
1934 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
1936 struct osd_object *obj = osd_obj(l);
1937 struct inode *inode = obj->oo_inode;
1939 LINVRNT(osd_invariant(obj));
1942 * If object is unlinked remove fid->ino mapping from object index.
1945 osd_index_fini(obj);
1946 if (inode != NULL) {
1947 struct qsd_instance *qsd = osd_obj2dev(obj)->od_quota_slave;
1948 qid_t uid = i_uid_read(inode);
1949 qid_t gid = i_gid_read(inode);
1952 obj->oo_inode = NULL;
1955 struct osd_thread_info *info = osd_oti_get(env);
1956 struct lquota_id_info *qi = &info->oti_qi;
1958 /* Release granted quota to master if necessary */
1959 qi->lqi_id.qid_uid = uid;
1960 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
1962 qi->lqi_id.qid_uid = gid;
1963 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
1965 qi->lqi_id.qid_uid = i_projid_read(inode);
1966 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
1972 * Concurrency: ->loo_object_release() is called under site spin-lock.
1974 static void osd_object_release(const struct lu_env *env,
1975 struct lu_object *l)
1977 struct osd_object *o = osd_obj(l);
1978 /* nobody should be releasing a non-destroyed object with nlink=0
1979 * the API allows this, but ldiskfs doesn't like and then report
1980 * this inode as deleted */
1981 if (unlikely(!o->oo_destroyed && o->oo_inode && o->oo_inode->i_nlink == 0))
1986 * Concurrency: shouldn't matter.
1988 static int osd_object_print(const struct lu_env *env, void *cookie,
1989 lu_printer_t p, const struct lu_object *l)
1991 struct osd_object *o = osd_obj(l);
1992 struct iam_descr *d;
1994 if (o->oo_dir != NULL)
1995 d = o->oo_dir->od_container.ic_descr;
1998 return (*p)(env, cookie,
1999 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2001 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2002 o->oo_inode ? o->oo_inode->i_generation : 0,
2003 d ? d->id_ops->id_name : "plain");
2007 * Concurrency: shouldn't matter.
2009 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2010 struct obd_statfs *sfs)
2012 struct osd_device *osd = osd_dt_dev(d);
2013 struct super_block *sb = osd_sb(osd);
2014 struct kstatfs *ksfs;
2018 if (unlikely(osd->od_mnt == NULL))
2019 return -EINPROGRESS;
2021 /* osd_lproc.c call this without env, allocate ksfs for that case */
2022 if (unlikely(env == NULL)) {
2023 OBD_ALLOC_PTR(ksfs);
2027 ksfs = &osd_oti_get(env)->oti_ksfs;
2030 result = sb->s_op->statfs(sb->s_root, ksfs);
2034 statfs_pack(sfs, ksfs);
2035 if (unlikely(sb->s_flags & MS_RDONLY))
2036 sfs->os_state |= OS_STATE_READONLY;
2037 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb,
2038 LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2039 sfs->os_maxbytes = sb->s_maxbytes;
2041 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2044 * Reserve some space so to avoid fragmenting the filesystem too much.
2045 * Fragmentation not only impacts performance, but can also increase
2046 * metadata overhead significantly, causing grant calculation to be
2049 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2051 CLASSERT(OSD_STATFS_RESERVED > LDISKFS_MAX_BLOCK_SIZE);
2052 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2053 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2054 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2056 sfs->os_blocks -= reserved;
2057 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2058 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2061 if (unlikely(env == NULL))
2067 * Estimate space needed for file creations. We assume the largest filename
2068 * which is 2^64 - 1, hence a filename of 20 chars.
2069 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2071 #ifdef __LDISKFS_DIR_REC_LEN
2072 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2074 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2078 * Concurrency: doesn't access mutable data.
2080 static void osd_conf_get(const struct lu_env *env,
2081 const struct dt_device *dev,
2082 struct dt_device_param *param)
2084 struct super_block *sb = osd_sb(osd_dt_dev(dev));
2088 * XXX should be taken from not-yet-existing fs abstraction layer.
2090 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2091 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2092 param->ddp_symlink_max = sb->s_blocksize;
2093 param->ddp_mount_type = LDD_MT_LDISKFS;
2094 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EXTENTS))
2095 param->ddp_maxbytes = sb->s_maxbytes;
2097 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2098 /* inode are statically allocated, so per-inode space consumption
2099 * is the space consumed by the directory entry */
2100 param->ddp_inodespace = PER_OBJ_USAGE;
2101 /* EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2102 * = 128MB) which is unlikely to be hit in real life. Report a smaller
2103 * maximum length to not under count the actual number of extents
2104 * needed for writing a file. */
2105 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 2;
2106 /* worst-case extent insertion metadata overhead */
2107 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2108 param->ddp_mntopts = 0;
2109 if (test_opt(sb, XATTR_USER))
2110 param->ddp_mntopts |= MNTOPT_USERXATTR;
2111 if (test_opt(sb, POSIX_ACL))
2112 param->ddp_mntopts |= MNTOPT_ACL;
2114 /* LOD might calculate the max stripe count based on max_ea_size,
2115 * so we need take account in the overhead as well,
2116 * xattr_header + magic + xattr_entry_head */
2117 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2118 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2120 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2121 if (LDISKFS_HAS_INCOMPAT_FEATURE(sb, LDISKFS_FEATURE_INCOMPAT_EA_INODE))
2122 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2126 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2128 /* Preferred RPC size for efficient disk IO. 4MB shows good
2129 * all-around performance for ldiskfs, but use bigalloc chunk size
2130 * by default if larger. */
2131 #if defined(LDISKFS_CLUSTER_SIZE)
2132 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2133 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2136 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2140 * Concurrency: shouldn't matter.
2142 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2146 CDEBUG(D_CACHE, "%s: syncing OSD\n", osd_dt_dev(d)->od_svname);
2148 rc = ldiskfs_force_commit(osd_sb(osd_dt_dev(d)));
2150 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2157 * Start commit for OSD device.
2159 * An implementation of dt_commit_async method for OSD device.
2160 * Asychronously starts underlayng fs sync and thereby a transaction
2163 * \param env environment
2164 * \param d dt device
2166 * \see dt_device_operations
2168 static int osd_commit_async(const struct lu_env *env,
2169 struct dt_device *d)
2171 struct super_block *s = osd_sb(osd_dt_dev(d));
2174 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2175 RETURN(s->s_op->sync_fs(s, 0));
2178 /* Our own copy of the set readonly functions if present, or NU if not. */
2179 static int (*priv_dev_set_rdonly)(struct block_device *bdev);
2180 static int (*priv_dev_check_rdonly)(struct block_device *bdev);
2181 /* static int (*priv_dev_clear_rdonly)(struct block_device *bdev); */
2184 * Concurrency: shouldn't matter.
2186 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2188 struct super_block *sb = osd_sb(osd_dt_dev(d));
2189 struct block_device *dev = sb->s_bdev;
2190 int rc = -EOPNOTSUPP;
2193 if (priv_dev_set_rdonly) {
2194 struct block_device *jdev = LDISKFS_SB(sb)->journal_bdev;
2197 CERROR("*** setting %s read-only ***\n",
2198 osd_dt_dev(d)->od_svname);
2200 if (sb->s_op->freeze_fs) {
2201 rc = sb->s_op->freeze_fs(sb);
2206 if (jdev && (jdev != dev)) {
2207 CDEBUG(D_IOCTL | D_HA, "set journal dev %lx rdonly\n",
2209 priv_dev_set_rdonly(jdev);
2211 CDEBUG(D_IOCTL | D_HA, "set dev %lx rdonly\n", (long)dev);
2212 priv_dev_set_rdonly(dev);
2214 if (sb->s_op->unfreeze_fs)
2215 sb->s_op->unfreeze_fs(sb);
2220 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2221 osd_dt_dev(d)->od_svname, (long)dev, rc);
2227 * Note: we do not count into QUOTA here.
2228 * If we mount with --data_journal we may need more.
2230 const int osd_dto_credits_noquota[DTO_NR] = {
2233 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2234 * SINGLEDATA_TRANS_BLOCKS(8)
2235 * XXX Note: maybe iam need more, since iam have more level than
2238 [DTO_INDEX_INSERT] = 16,
2241 * just modify a single entry, probably merge few within a block
2243 [DTO_INDEX_DELETE] = 1,
2247 [DTO_INDEX_UPDATE] = 16,
2249 * 4(inode, inode bits, groups, GDT)
2250 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2252 [DTO_OBJECT_CREATE] = 4,
2254 * 4(inode, inode bits, groups, GDT)
2255 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2257 [DTO_OBJECT_DELETE] = 4,
2259 * Attr set credits (inode)
2261 [DTO_ATTR_SET_BASE] = 1,
2263 * Xattr set. The same as xattr of EXT3.
2264 * DATA_TRANS_BLOCKS(14)
2265 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2266 * are also counted in. Do not know why?
2268 [DTO_XATTR_SET] = 14,
2270 * credits for inode change during write.
2272 [DTO_WRITE_BASE] = 3,
2274 * credits for single block write.
2276 [DTO_WRITE_BLOCK] = 14,
2278 * Attr set credits for chown.
2279 * This is extra credits for setattr, and it is null without quota
2281 [DTO_ATTR_SET_CHOWN] = 0
2284 static const struct dt_device_operations osd_dt_ops = {
2285 .dt_root_get = osd_root_get,
2286 .dt_statfs = osd_statfs,
2287 .dt_trans_create = osd_trans_create,
2288 .dt_trans_start = osd_trans_start,
2289 .dt_trans_stop = osd_trans_stop,
2290 .dt_trans_cb_add = osd_trans_cb_add,
2291 .dt_conf_get = osd_conf_get,
2292 .dt_sync = osd_sync,
2294 .dt_commit_async = osd_commit_async,
2297 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2300 struct osd_object *obj = osd_dt_obj(dt);
2301 struct osd_thread_info *oti = osd_oti_get(env);
2303 LINVRNT(osd_invariant(obj));
2305 LASSERT(obj->oo_owner != env);
2306 down_read_nested(&obj->oo_sem, role);
2308 LASSERT(obj->oo_owner == NULL);
2312 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2315 struct osd_object *obj = osd_dt_obj(dt);
2316 struct osd_thread_info *oti = osd_oti_get(env);
2318 LINVRNT(osd_invariant(obj));
2320 LASSERT(obj->oo_owner != env);
2321 down_write_nested(&obj->oo_sem, role);
2323 LASSERT(obj->oo_owner == NULL);
2324 obj->oo_owner = env;
2328 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2330 struct osd_object *obj = osd_dt_obj(dt);
2331 struct osd_thread_info *oti = osd_oti_get(env);
2333 LINVRNT(osd_invariant(obj));
2335 LASSERT(oti->oti_r_locks > 0);
2337 up_read(&obj->oo_sem);
2340 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2342 struct osd_object *obj = osd_dt_obj(dt);
2343 struct osd_thread_info *oti = osd_oti_get(env);
2345 LINVRNT(osd_invariant(obj));
2347 LASSERT(obj->oo_owner == env);
2348 LASSERT(oti->oti_w_locks > 0);
2350 obj->oo_owner = NULL;
2351 up_write(&obj->oo_sem);
2354 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2356 struct osd_object *obj = osd_dt_obj(dt);
2358 LINVRNT(osd_invariant(obj));
2360 return obj->oo_owner == env;
2363 static struct timespec *osd_inode_time(const struct lu_env *env,
2364 struct inode *inode, __u64 seconds)
2366 struct osd_thread_info *oti = osd_oti_get(env);
2367 struct timespec *t = &oti->oti_time;
2369 t->tv_sec = seconds;
2371 *t = timespec_trunc(*t, inode->i_sb->s_time_gran);
2375 static void osd_inode_getattr(const struct lu_env *env,
2376 struct inode *inode, struct lu_attr *attr)
2378 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2379 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2380 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2381 LA_BLKSIZE | LA_TYPE;
2383 attr->la_atime = LTIME_S(inode->i_atime);
2384 attr->la_mtime = LTIME_S(inode->i_mtime);
2385 attr->la_ctime = LTIME_S(inode->i_ctime);
2386 attr->la_mode = inode->i_mode;
2387 attr->la_size = i_size_read(inode);
2388 attr->la_blocks = inode->i_blocks;
2389 attr->la_uid = i_uid_read(inode);
2390 attr->la_gid = i_gid_read(inode);
2391 attr->la_projid = i_projid_read(inode);
2392 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2393 attr->la_nlink = inode->i_nlink;
2394 attr->la_rdev = inode->i_rdev;
2395 attr->la_blksize = 1 << inode->i_blkbits;
2396 attr->la_blkbits = inode->i_blkbits;
2398 * Ext4 did not transfer inherit flags from raw inode
2399 * to inode flags, and ext4 internally test raw inode
2400 * @i_flags directly. Instead of patching ext4, we do it here.
2402 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2403 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2406 static int osd_attr_get(const struct lu_env *env,
2407 struct dt_object *dt,
2408 struct lu_attr *attr)
2410 struct osd_object *obj = osd_dt_obj(dt);
2412 if (unlikely(!dt_object_exists(dt)))
2414 if (unlikely(obj->oo_destroyed))
2417 LASSERT(!dt_object_remote(dt));
2418 LINVRNT(osd_invariant(obj));
2420 spin_lock(&obj->oo_guard);
2421 osd_inode_getattr(env, obj->oo_inode, attr);
2422 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
2423 attr->la_flags |= LUSTRE_ORPHAN_FL;
2424 spin_unlock(&obj->oo_guard);
2429 static int osd_declare_attr_qid(const struct lu_env *env,
2430 struct osd_object *obj,
2431 struct osd_thandle *oh, long long bspace,
2432 qid_t old_id, qid_t new_id, bool enforce,
2436 struct osd_thread_info *info = osd_oti_get(env);
2437 struct lquota_id_info *qi = &info->oti_qi;
2439 qi->lqi_type = type;
2440 /* inode accounting */
2441 qi->lqi_is_blk = false;
2443 /* one more inode for the new id ... */
2444 qi->lqi_id.qid_uid = new_id;
2446 /* Reserve credits for the new id */
2447 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2448 if (rc == -EDQUOT || rc == -EINPROGRESS)
2453 /* and one less inode for the current id */
2454 qi->lqi_id.qid_uid = old_id;
2456 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2457 if (rc == -EDQUOT || rc == -EINPROGRESS)
2462 /* block accounting */
2463 qi->lqi_is_blk = true;
2465 /* more blocks for the new id ... */
2466 qi->lqi_id.qid_uid = new_id;
2467 qi->lqi_space = bspace;
2469 * Credits for the new uid has been reserved, re-use "obj"
2470 * to save credit reservation.
2472 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2473 if (rc == -EDQUOT || rc == -EINPROGRESS)
2478 /* and finally less blocks for the current uid */
2479 qi->lqi_id.qid_uid = old_id;
2480 qi->lqi_space = -bspace;
2481 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2482 if (rc == -EDQUOT || rc == -EINPROGRESS)
2488 static int osd_declare_attr_set(const struct lu_env *env,
2489 struct dt_object *dt,
2490 const struct lu_attr *attr,
2491 struct thandle *handle)
2493 struct osd_thandle *oh;
2494 struct osd_object *obj;
2502 LASSERT(dt != NULL);
2503 LASSERT(handle != NULL);
2505 obj = osd_dt_obj(dt);
2506 LASSERT(osd_invariant(obj));
2508 oh = container_of0(handle, struct osd_thandle, ot_super);
2509 LASSERT(oh->ot_handle == NULL);
2511 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2512 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2514 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2515 osd_dto_credits_noquota[DTO_XATTR_SET]);
2517 if (attr == NULL || obj->oo_inode == NULL)
2520 bspace = obj->oo_inode->i_blocks;
2521 bspace <<= obj->oo_inode->i_sb->s_blocksize_bits;
2522 bspace = toqb(bspace);
2524 /* Changing ownership is always preformed by super user, it should not
2527 * We still need to call the osd_declare_qid() to calculate the journal
2528 * credits for updating quota accounting files and to trigger quota
2529 * space adjustment once the operation is completed.*/
2530 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2532 uid = i_uid_read(obj->oo_inode);
2533 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2534 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2535 attr->la_uid, enforce, USRQUOTA);
2539 gid = i_gid_read(obj->oo_inode);
2540 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2541 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2542 i_gid_read(obj->oo_inode),
2543 attr->la_gid, enforce, GRPQUOTA);
2548 #ifdef HAVE_PROJECT_QUOTA
2549 if (attr->la_valid & LA_PROJID) {
2550 __u32 projid = i_projid_read(obj->oo_inode);
2551 enforce = (attr->la_valid & LA_PROJID) &&
2552 (attr->la_projid != projid);
2553 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2554 (qid_t)projid, (qid_t)attr->la_projid,
2563 static int osd_inode_setattr(const struct lu_env *env,
2564 struct inode *inode, const struct lu_attr *attr)
2566 __u64 bits = attr->la_valid;
2568 /* Only allow set size for regular file */
2569 if (!S_ISREG(inode->i_mode))
2570 bits &= ~(LA_SIZE | LA_BLOCKS);
2575 if (bits & LA_ATIME)
2576 inode->i_atime = *osd_inode_time(env, inode, attr->la_atime);
2577 if (bits & LA_CTIME)
2578 inode->i_ctime = *osd_inode_time(env, inode, attr->la_ctime);
2579 if (bits & LA_MTIME)
2580 inode->i_mtime = *osd_inode_time(env, inode, attr->la_mtime);
2581 if (bits & LA_SIZE) {
2582 spin_lock(&inode->i_lock);
2583 LDISKFS_I(inode)->i_disksize = attr->la_size;
2584 i_size_write(inode, attr->la_size);
2585 spin_unlock(&inode->i_lock);
2588 /* OSD should not change "i_blocks" which is used by quota.
2589 * "i_blocks" should be changed by ldiskfs only. */
2591 inode->i_mode = (inode->i_mode & S_IFMT) |
2592 (attr->la_mode & ~S_IFMT);
2594 i_uid_write(inode, attr->la_uid);
2596 i_gid_write(inode, attr->la_gid);
2597 if (bits & LA_PROJID)
2598 i_projid_write(inode, attr->la_projid);
2599 if (bits & LA_NLINK)
2600 set_nlink(inode, attr->la_nlink);
2602 inode->i_rdev = attr->la_rdev;
2604 if (bits & LA_FLAGS) {
2605 /* always keep S_NOCMTIME */
2606 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2609 * Ext4 did not transfer inherit flags from
2610 * @inode->i_flags to raw inode i_flags when writing
2611 * flags, we do it explictly here.
2613 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2614 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2616 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2621 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr)
2625 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
2626 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
2629 ll_vfs_dq_init(inode);
2631 if (attr->la_valid & LA_UID)
2632 iattr.ia_valid |= ATTR_UID;
2633 if (attr->la_valid & LA_GID)
2634 iattr.ia_valid |= ATTR_GID;
2635 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
2636 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
2638 rc = ll_vfs_dq_transfer(inode, &iattr);
2640 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2641 "enforcement enabled on the ldiskfs "
2642 "filesystem?\n", inode->i_sb->s_id, rc);
2647 #ifdef HAVE_PROJECT_QUOTA
2648 /* Handle project id transfer here properly */
2649 if (attr->la_valid & LA_PROJID &&
2650 attr->la_projid != i_projid_read(inode)) {
2651 rc = __ldiskfs_ioctl_setproject(inode, attr->la_projid);
2653 CERROR("%s: quota transfer failed: rc = %d. Is quota "
2654 "enforcement enabled on the ldiskfs "
2655 "filesystem?\n", inode->i_sb->s_id, rc);
2663 static int osd_attr_set(const struct lu_env *env,
2664 struct dt_object *dt,
2665 const struct lu_attr *attr,
2666 struct thandle *handle)
2668 struct osd_object *obj = osd_dt_obj(dt);
2669 struct inode *inode;
2672 if (!dt_object_exists(dt))
2675 LASSERT(handle != NULL);
2676 LASSERT(!dt_object_remote(dt));
2677 LASSERT(osd_invariant(obj));
2679 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
2681 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING)) {
2682 struct osd_thread_info *oti = osd_oti_get(env);
2683 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
2684 struct lu_fid *fid1 = &oti->oti_fid;
2685 struct osd_inode_id *id = &oti->oti_id;
2686 struct iam_path_descr *ipd;
2687 struct iam_container *bag;
2688 struct osd_thandle *oh;
2691 fid_cpu_to_be(fid1, fid0);
2692 memset(id, 1, sizeof(*id));
2693 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
2694 fid0)->oi_dir.od_container;
2695 ipd = osd_idx_ipd_get(env, bag);
2696 if (unlikely(ipd == NULL))
2699 oh = container_of0(handle, struct osd_thandle, ot_super);
2700 rc = iam_update(oh->ot_handle, bag, (const struct iam_key *)fid1,
2701 (const struct iam_rec *)id, ipd);
2702 osd_ipd_put(env, bag, ipd);
2703 return(rc > 0 ? 0 : rc);
2706 inode = obj->oo_inode;
2708 rc = osd_quota_transfer(inode, attr);
2712 spin_lock(&obj->oo_guard);
2713 rc = osd_inode_setattr(env, inode, attr);
2714 spin_unlock(&obj->oo_guard);
2718 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2720 if (!(attr->la_valid & LA_FLAGS))
2723 /* Let's check if there are extra flags need to be set into LMA */
2724 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
2725 struct osd_thread_info *info = osd_oti_get(env);
2726 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
2728 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
2729 &info->oti_ost_attrs);
2733 lma->lma_incompat |=
2734 lustre_to_lma_flags(attr->la_flags);
2735 lustre_lma_swab(lma);
2736 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
2737 lma, sizeof(*lma), XATTR_REPLACE);
2739 struct osd_device *osd = osd_obj2dev(obj);
2741 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
2742 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
2743 lma->lma_incompat, rc);
2746 attr->la_flags & LUSTRE_LMA_FL_MASKS;
2748 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
2751 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
2756 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
2757 struct osd_object *obj,
2758 const char *name, const int namelen)
2760 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
2763 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
2764 umode_t mode, struct dt_allocation_hint *hint,
2768 struct osd_device *osd = osd_obj2dev(obj);
2769 struct osd_thandle *oth;
2770 struct dt_object *parent = NULL;
2771 struct inode *inode;
2773 LINVRNT(osd_invariant(obj));
2774 LASSERT(obj->oo_inode == NULL);
2775 LASSERT(obj->oo_hl_head == NULL);
2777 if (S_ISDIR(mode) && ldiskfs_pdo) {
2778 obj->oo_hl_head =ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
2779 if (obj->oo_hl_head == NULL)
2783 oth = container_of(th, struct osd_thandle, ot_super);
2784 LASSERT(oth->ot_handle->h_transaction != NULL);
2786 if (hint != NULL && hint->dah_parent != NULL &&
2787 !dt_object_remote(hint->dah_parent))
2788 parent = hint->dah_parent;
2790 inode = ldiskfs_create_inode(oth->ot_handle,
2791 parent ? osd_dt_obj(parent)->oo_inode :
2792 osd_sb(osd)->s_root->d_inode,
2794 if (!IS_ERR(inode)) {
2795 /* Do not update file c/mtime in ldiskfs. */
2796 inode->i_flags |= S_NOCMTIME;
2798 /* For new created object, it must be consistent,
2799 * and it is unnecessary to scrub against it. */
2800 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
2802 obj->oo_inode = inode;
2805 if (obj->oo_hl_head != NULL) {
2806 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
2807 obj->oo_hl_head = NULL;
2809 result = PTR_ERR(inode);
2811 LINVRNT(osd_invariant(obj));
2819 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
2820 struct lu_attr *attr,
2821 struct dt_allocation_hint *hint,
2822 struct dt_object_format *dof,
2826 struct osd_thandle *oth;
2827 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
2829 LASSERT(S_ISDIR(attr->la_mode));
2831 oth = container_of(th, struct osd_thandle, ot_super);
2832 LASSERT(oth->ot_handle->h_transaction != NULL);
2833 result = osd_mkfile(info, obj, mode, hint, th);
2838 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
2839 struct lu_attr *attr,
2840 struct dt_allocation_hint *hint,
2841 struct dt_object_format *dof,
2845 struct osd_thandle *oth;
2846 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
2848 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
2850 LASSERT(S_ISREG(attr->la_mode));
2852 oth = container_of(th, struct osd_thandle, ot_super);
2853 LASSERT(oth->ot_handle->h_transaction != NULL);
2855 result = osd_mkfile(info, obj, mode, hint, th);
2857 LASSERT(obj->oo_inode != NULL);
2858 if (feat->dif_flags & DT_IND_VARKEY)
2859 result = iam_lvar_create(obj->oo_inode,
2860 feat->dif_keysize_max,
2862 feat->dif_recsize_max,
2865 result = iam_lfix_create(obj->oo_inode,
2866 feat->dif_keysize_max,
2868 feat->dif_recsize_max,
2875 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
2876 struct lu_attr *attr,
2877 struct dt_allocation_hint *hint,
2878 struct dt_object_format *dof,
2881 LASSERT(S_ISREG(attr->la_mode));
2882 return osd_mkfile(info, obj, (attr->la_mode &
2883 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2886 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
2887 struct lu_attr *attr,
2888 struct dt_allocation_hint *hint,
2889 struct dt_object_format *dof,
2892 LASSERT(S_ISLNK(attr->la_mode));
2893 return osd_mkfile(info, obj, (attr->la_mode &
2894 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th);
2897 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
2898 struct lu_attr *attr,
2899 struct dt_allocation_hint *hint,
2900 struct dt_object_format *dof,
2903 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
2906 LINVRNT(osd_invariant(obj));
2907 LASSERT(obj->oo_inode == NULL);
2908 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
2909 S_ISFIFO(mode) || S_ISSOCK(mode));
2911 result = osd_mkfile(info, obj, mode, hint, th);
2913 LASSERT(obj->oo_inode != NULL);
2915 * This inode should be marked dirty for i_rdev. Currently
2916 * that is done in the osd_attr_init().
2918 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
2921 LINVRNT(osd_invariant(obj));
2925 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
2927 struct dt_allocation_hint *hint,
2928 struct dt_object_format *dof,
2931 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
2933 osd_obj_type_f result;
2949 result = osd_mk_index;
2960 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
2961 struct dt_object *parent, struct dt_object *child,
2966 ah->dah_parent = parent;
2967 ah->dah_mode = child_mode;
2969 if (parent != NULL && !dt_object_remote(parent)) {
2970 /* will help to find FID->ino at dt_insert("..") */
2971 struct osd_object *pobj = osd_dt_obj(parent);
2972 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
2976 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
2977 struct lu_attr *attr, struct dt_object_format *dof)
2979 struct inode *inode = obj->oo_inode;
2980 __u64 valid = attr->la_valid;
2983 attr->la_valid &= ~(LA_TYPE | LA_MODE);
2985 if (dof->dof_type != DFT_NODE)
2986 attr->la_valid &= ~LA_RDEV;
2987 if ((valid & LA_ATIME) && (attr->la_atime == LTIME_S(inode->i_atime)))
2988 attr->la_valid &= ~LA_ATIME;
2989 if ((valid & LA_CTIME) && (attr->la_ctime == LTIME_S(inode->i_ctime)))
2990 attr->la_valid &= ~LA_CTIME;
2991 if ((valid & LA_MTIME) && (attr->la_mtime == LTIME_S(inode->i_mtime)))
2992 attr->la_valid &= ~LA_MTIME;
2994 result = osd_quota_transfer(inode, attr);
2998 if (attr->la_valid != 0) {
2999 result = osd_inode_setattr(info->oti_env, inode, attr);
3001 * The osd_inode_setattr() should always succeed here. The
3002 * only error that could be returned is EDQUOT when we are
3003 * trying to change the UID or GID of the inode. However, this
3004 * should not happen since quota enforcement is no longer
3005 * enabled on ldiskfs (lquota takes care of it).
3007 LASSERTF(result == 0, "%d\n", result);
3008 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3011 attr->la_valid = valid;
3015 * Helper function for osd_create()
3017 * \retval 0, on success
3019 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3020 struct lu_attr *attr, struct dt_allocation_hint *hint,
3021 struct dt_object_format *dof, struct thandle *th)
3026 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3028 /* we drop umask so that permissions we pass are not affected */
3029 umask = current->fs->umask;
3030 current->fs->umask = 0;
3032 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3034 if (likely(obj->oo_inode != NULL)) {
3035 LASSERT(obj->oo_inode->i_state & I_NEW);
3037 /* Unlock the inode before attr initialization to avoid
3038 * unnecessary dqget operations. LU-6378 */
3039 unlock_new_inode(obj->oo_inode);
3042 if (likely(result == 0)) {
3043 osd_attr_init(info, obj, attr, dof);
3044 osd_object_init0(obj);
3047 /* restore previous umask value */
3048 current->fs->umask = umask;
3050 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3056 * Helper function for osd_create()
3058 * \retval 0, on success
3060 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3061 const struct lu_fid *fid, struct thandle *th)
3063 struct osd_thread_info *info = osd_oti_get(env);
3064 struct osd_inode_id *id = &info->oti_id;
3065 struct osd_device *osd = osd_obj2dev(obj);
3066 struct osd_thandle *oh;
3069 LASSERT(obj->oo_inode != NULL);
3071 oh = container_of0(th, struct osd_thandle, ot_super);
3072 LASSERT(oh->ot_handle);
3073 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3075 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3076 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3077 OI_CHECK_FLD, NULL);
3078 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3083 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3084 u64 seq, struct lu_seq_range *range)
3086 struct seq_server_site *ss = osd_seq_site(osd);
3088 if (fid_seq_is_idif(seq)) {
3089 fld_range_set_ost(range);
3090 range->lsr_index = idif_ost_idx(seq);
3094 if (!fid_seq_in_fldb(seq)) {
3095 fld_range_set_mdt(range);
3097 /* FIXME: If ss is NULL, it suppose not get lsr_index
3099 range->lsr_index = ss->ss_node_id;
3103 LASSERT(ss != NULL);
3104 fld_range_set_any(range);
3105 /* OSD will only do local fld lookup */
3106 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3109 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3110 struct lu_attr *attr,
3111 struct dt_allocation_hint *hint,
3112 struct dt_object_format *dof,
3113 struct thandle *handle)
3115 struct osd_thandle *oh;
3119 LASSERT(handle != NULL);
3121 oh = container_of0(handle, struct osd_thandle, ot_super);
3122 LASSERT(oh->ot_handle == NULL);
3124 /* EA object consumes more credits than regular object: osd_mk_index
3125 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3126 * leaf_node, could involves the block, block bitmap, groups, GDT
3127 * change for each block, so add 4 * 2 credits in that case. */
3128 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3129 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3130 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3131 /* Reuse idle OI block may cause additional one OI block
3133 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3134 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3139 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3140 attr->la_projid, 1, oh, osd_dt_obj(dt),
3141 NULL, OSD_QID_INODE);
3145 /* will help to find FID->ino mapping at dt_insert() */
3146 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3153 * Called to destroy on-disk representation of the object
3155 * Concurrency: must be locked
3157 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3160 struct osd_object *obj = osd_dt_obj(dt);
3161 struct inode *inode = obj->oo_inode;
3162 struct osd_thandle *oh;
3169 oh = container_of0(th, struct osd_thandle, ot_super);
3170 LASSERT(oh->ot_handle == NULL);
3172 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3173 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3174 /* Recycle idle OI leaf may cause additional three OI blocks
3176 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3177 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3178 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3179 /* one less inode */
3180 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3181 i_projid_read(inode), -1, oh, obj, NULL,
3185 /* data to be truncated */
3186 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3187 i_projid_read(inode), 0, oh, obj, NULL,
3192 /* will help to find FID->ino when this object is being
3193 * added to PENDING/ */
3194 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3199 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3202 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3203 struct osd_object *obj = osd_dt_obj(dt);
3204 struct inode *inode = obj->oo_inode;
3205 struct osd_device *osd = osd_obj2dev(obj);
3206 struct osd_thandle *oh;
3210 oh = container_of0(th, struct osd_thandle, ot_super);
3211 LASSERT(oh->ot_handle);
3213 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3215 if (unlikely(fid_is_acct(fid)))
3218 if (S_ISDIR(inode->i_mode)) {
3219 LASSERT(osd_inode_unlinked(inode) || inode->i_nlink == 1 ||
3220 inode->i_nlink == 2);
3221 /* it will check/delete the inode from remote parent,
3222 * how to optimize it? unlink performance impaction XXX */
3223 result = osd_delete_from_remote_parent(env, osd, obj, oh);
3225 CERROR("%s: delete inode "DFID": rc = %d\n",
3226 osd_name(osd), PFID(fid), result);
3228 spin_lock(&obj->oo_guard);
3230 spin_unlock(&obj->oo_guard);
3231 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3234 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3236 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3238 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3239 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3240 oh->ot_handle, OI_CHECK_FLD);
3242 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3243 /* XXX: add to ext3 orphan list */
3244 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3246 /* not needed in the cache anymore */
3247 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3248 obj->oo_destroyed = 1;
3254 * Put the fid into lustre_mdt_attrs, and then place the structure
3255 * inode's ea. This fid should not be altered during the life time
3258 * \retval +ve, on success
3259 * \retval -ve, on error
3261 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3263 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3264 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3266 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3267 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3271 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3274 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3277 lustre_loa_init(loa, fid, compat, incompat);
3278 lustre_loa_swab(loa, false);
3280 /* For the OST device with 256 bytes inode size by default,
3281 * the PFID EA will be stored together with LMA EA to avoid
3282 * performance trouble. Otherwise the PFID EA can be stored
3283 * independently. LU-8998 */
3284 if ((compat & LMAC_FID_ON_OST) &&
3285 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3286 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3287 sizeof(*loa), XATTR_CREATE);
3289 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3290 sizeof(*lma), XATTR_CREATE);
3291 /* LMA may already exist, but we need to check that all the
3292 * desired compat/incompat flags have been added. */
3293 if (unlikely(rc == -EEXIST)) {
3294 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3295 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3299 if (rc < sizeof(*lma))
3302 lustre_loa_swab(loa, true);
3303 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3304 ((compat == 0 && incompat == 0) ||
3305 (!(~lma->lma_compat & compat) &&
3306 !(~lma->lma_incompat & incompat))))
3309 lma->lma_self_fid = *fid;
3310 lma->lma_compat |= compat;
3311 lma->lma_incompat |= incompat;
3312 if (rc == sizeof(*lma)) {
3313 lustre_lma_swab(lma);
3314 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3315 sizeof(*lma), XATTR_REPLACE);
3317 lustre_loa_swab(loa, false);
3318 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3319 sizeof(*loa), XATTR_REPLACE);
3327 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3328 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3329 * To have compatilibility with 1.8 ldiskfs driver we need to have
3330 * magic number at start of fid data.
3331 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3334 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3335 const struct lu_fid *fid)
3337 if (!fid_is_namespace_visible(fid) ||
3338 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3339 param->edp_magic = 0;
3343 param->edp_magic = LDISKFS_LUFID_MAGIC;
3344 param->edp_len = sizeof(struct lu_fid) + 1;
3345 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3349 * Try to read the fid from inode ea into dt_rec.
3351 * \param fid object fid.
3353 * \retval 0 on success
3355 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3356 __u32 ino, struct lu_fid *fid,
3357 struct osd_inode_id *id)
3359 struct osd_thread_info *info = osd_oti_get(env);
3360 struct inode *inode;
3363 osd_id_gen(id, ino, OSD_OII_NOGEN);
3364 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3366 RETURN(PTR_ERR(inode));
3372 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3374 struct inode *parent_dir,
3375 const struct lu_fid *dot_fid,
3376 const struct lu_fid *dot_dot_fid,
3377 struct osd_thandle *oth)
3379 struct ldiskfs_dentry_param *dot_ldp;
3380 struct ldiskfs_dentry_param *dot_dot_ldp;
3381 __u32 saved_nlink = dir->i_nlink;
3384 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3385 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3387 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3388 dot_ldp->edp_magic = 0;
3390 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3391 dir, dot_ldp, dot_dot_ldp);
3392 /* The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3393 * the subseqent ref_add() will increase the dir->i_nlink
3394 * as 3. That is incorrect for new created directory.
3396 * It looks like hack, because we want to make the OSD API
3397 * to be order-independent for new created directory object
3398 * between dt_insert(..) and ref_add() operations.
3400 * Here, we only restore the in-RAM dir-inode's nlink attr,
3401 * becuase if the nlink attr is not 2, then there will be
3402 * ref_add() called following the dt_insert(..), such call
3403 * will make both the in-RAM and on-disk dir-inode's nlink
3404 * attr to be set as 2. LU-7447 */
3405 set_nlink(dir, saved_nlink);
3410 * Create an local agent inode for remote entry
3412 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3413 struct osd_device *osd,
3414 struct osd_object *pobj,
3415 const struct lu_fid *fid,
3419 struct osd_thread_info *info = osd_oti_get(env);
3420 struct inode *local;
3421 struct osd_thandle *oh;
3426 oh = container_of(th, struct osd_thandle, ot_super);
3427 LASSERT(oh->ot_handle->h_transaction != NULL);
3429 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type);
3430 if (IS_ERR(local)) {
3431 CERROR("%s: create local error %d\n", osd_name(osd),
3432 (int)PTR_ERR(local));
3436 /* restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3437 * correct gid on remote file, not agent here */
3438 local->i_gid = current_fsgid();
3439 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3440 unlock_new_inode(local);
3442 /* Agent inode should not have project ID */
3443 #ifdef HAVE_PROJECT_QUOTA
3444 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL) {
3445 rc = __ldiskfs_ioctl_setproject(local, 0);
3447 CERROR("%s: quota transfer failed: rc = %d. Is project "
3448 "quota enforcement enabled on the ldiskfs "
3449 "filesystem?\n", local->i_sb->s_id, rc);
3450 RETURN(ERR_PTR(rc));
3454 /* Set special LMA flag for local agent inode */
3455 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3457 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3458 osd_name(osd), PFID(fid), rc);
3459 RETURN(ERR_PTR(rc));
3465 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3466 lu_object_fid(&pobj->oo_dt.do_lu),
3469 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
3470 osd_name(osd), PFID(fid), rc);
3471 RETURN(ERR_PTR(rc));
3478 * when direntry is deleted, we have to take care of possible agent inode
3479 * referenced by that. unfortunately we can't do this at that point:
3480 * iget() within a running transaction leads to deadlock and we better do
3481 * not call that every delete declaration to save performance. so we put
3482 * a potention agent inode on a list and process that once the transaction
3483 * is over. Notice it's not any worse in terms of real orphans as regular
3484 * object destroy doesn't put inodes on the on-disk orphan list. this should
3485 * be addressed separately
3487 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
3488 struct osd_thandle *oh,
3491 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
3492 struct osd_obj_orphan *oor;
3499 oor->oor_env = (struct lu_env *)env;
3500 spin_lock(&osd->od_osfs_lock);
3501 list_add(&oor->oor_list, &osd->od_orphan_list);
3502 spin_unlock(&osd->od_osfs_lock);
3504 oh->ot_remove_agents = 1;
3510 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
3511 struct osd_device *osd)
3513 struct osd_thread_info *info = osd_oti_get(env);
3514 struct osd_obj_orphan *oor, *tmp;
3515 struct osd_inode_id id;
3516 struct list_head list;
3517 struct inode *inode;
3522 INIT_LIST_HEAD(&list);
3524 spin_lock(&osd->od_osfs_lock);
3525 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
3526 if (oor->oor_env == env) {
3527 list_del(&oor->oor_list);
3528 list_add(&oor->oor_list, &list);
3531 spin_unlock(&osd->od_osfs_lock);
3533 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
3537 list_del(&oor->oor_list);
3540 osd_id_gen(&id, ino, OSD_OII_NOGEN);
3541 inode = osd_iget_fid(info, osd, &id, &fid);
3545 if (!osd_remote_fid(env, osd, &fid)) {
3550 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
3552 mark_inode_dirty(inode);
3553 ldiskfs_journal_stop(jh);
3561 * OSD layer object create function for OST objects (b=11826).
3563 * The FID is inserted into inode xattr here.
3565 * \retval 0, on success
3566 * \retval -ve, on error
3568 static int osd_create(const struct lu_env *env, struct dt_object *dt,
3569 struct lu_attr *attr, struct dt_allocation_hint *hint,
3570 struct dt_object_format *dof, struct thandle *th)
3572 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3573 struct osd_object *obj = osd_dt_obj(dt);
3574 struct osd_thread_info *info = osd_oti_get(env);
3575 int result, on_ost = 0;
3579 if (dt_object_exists(dt))
3582 LINVRNT(osd_invariant(obj));
3583 LASSERT(!dt_object_remote(dt));
3584 LASSERT(osd_is_write_locked(env, obj));
3585 LASSERT(th != NULL);
3587 if (unlikely(fid_is_acct(fid)))
3588 /* Quota files can't be created from the kernel any more,
3589 * 'tune2fs -O quota' will take care of creating them */
3592 result = __osd_create(info, obj, attr, hint, dof, th);
3594 if (fid_is_idif(fid) &&
3595 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
3596 struct lu_fid *tfid = &info->oti_fid;
3597 struct ost_id *oi = &info->oti_ostid;
3599 fid_to_ostid(fid, oi);
3600 ostid_to_fid(tfid, oi, 0);
3602 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
3603 LMAC_FID_ON_OST, 0);
3605 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
3607 result = osd_ea_fid_set(info, obj->oo_inode, fid,
3608 on_ost ? LMAC_FID_ON_OST : 0,
3611 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
3612 obj->oo_dt.do_body_ops = &osd_body_ops;
3616 result = __osd_oi_insert(env, obj, fid, th);
3618 /* a small optimization - dt_insert() isn't usually applied
3619 * to OST objects, so we don't need to cache OI mapping for
3621 if (result == 0 && on_ost == 0) {
3622 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3623 result = osd_idc_find_and_init(env, osd, obj);
3624 LASSERT(result == 0);
3627 LASSERT(ergo(result == 0,
3628 dt_object_exists(dt) && !dt_object_remote(dt)));
3629 LINVRNT(osd_invariant(obj));
3633 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
3634 struct thandle *handle)
3636 struct osd_thandle *oh;
3638 /* it's possible that object doesn't exist yet */
3639 LASSERT(handle != NULL);
3641 oh = container_of0(handle, struct osd_thandle, ot_super);
3642 LASSERT(oh->ot_handle == NULL);
3644 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
3645 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3647 osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev), osd_dt_obj(dt));
3653 * Concurrency: @dt is write locked.
3655 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
3658 struct osd_object *obj = osd_dt_obj(dt);
3659 struct inode *inode = obj->oo_inode;
3660 struct osd_thandle *oh;
3663 if (!dt_object_exists(dt) || obj->oo_destroyed)
3666 LINVRNT(osd_invariant(obj));
3667 LASSERT(!dt_object_remote(dt));
3668 LASSERT(osd_is_write_locked(env, obj));
3669 LASSERT(th != NULL);
3671 oh = container_of0(th, struct osd_thandle, ot_super);
3672 LASSERT(oh->ot_handle != NULL);
3674 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
3676 CDEBUG(D_INODE, DFID" increase nlink %d\n",
3677 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3679 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
3680 * (65000) subdirectories by storing "1" in i_nlink if the link count
3681 * would otherwise overflow. Directory tranversal tools understand
3682 * that (st_nlink == 1) indicates that the filesystem dose not track
3683 * hard links count on the directory, and will not abort subdirectory
3684 * scanning early once (st_nlink - 2) subdirs have been found.
3686 * This also has to properly handle the case of inodes with nlink == 0
3687 * in case they are being linked into the PENDING directory
3689 spin_lock(&obj->oo_guard);
3690 if (unlikely(inode->i_nlink == 0))
3691 /* inc_nlink from 0 may cause WARN_ON */
3692 set_nlink(inode, 1);
3694 ldiskfs_inc_count(oh->ot_handle, inode);
3695 if (!S_ISDIR(inode->i_mode))
3696 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
3698 spin_unlock(&obj->oo_guard);
3700 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3701 LINVRNT(osd_invariant(obj));
3703 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
3708 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
3709 struct thandle *handle)
3711 struct osd_thandle *oh;
3713 if (!dt_object_exists(dt))
3716 LASSERT(!dt_object_remote(dt));
3717 LASSERT(handle != NULL);
3719 oh = container_of0(handle, struct osd_thandle, ot_super);
3720 LASSERT(oh->ot_handle == NULL);
3722 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
3723 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3729 * Concurrency: @dt is write locked.
3731 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
3734 struct osd_object *obj = osd_dt_obj(dt);
3735 struct inode *inode = obj->oo_inode;
3736 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
3737 struct osd_thandle *oh;
3739 if (!dt_object_exists(dt))
3742 LINVRNT(osd_invariant(obj));
3743 LASSERT(!dt_object_remote(dt));
3744 LASSERT(osd_is_write_locked(env, obj));
3745 LASSERT(th != NULL);
3747 oh = container_of0(th, struct osd_thandle, ot_super);
3748 LASSERT(oh->ot_handle != NULL);
3750 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
3752 spin_lock(&obj->oo_guard);
3753 /* That can be result of upgrade from old Lustre version and
3754 * applied only to local files. Just skip this ref_del call.
3755 * ext4_unlink() only treats this as a warning, don't LASSERT here.*/
3756 if (inode->i_nlink == 0) {
3757 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
3758 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
3759 ", maybe an upgraded file? (LU-3915)\n",
3760 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
3761 spin_unlock(&obj->oo_guard);
3765 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
3766 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
3768 ldiskfs_dec_count(oh->ot_handle, inode);
3769 spin_unlock(&obj->oo_guard);
3771 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3772 LINVRNT(osd_invariant(obj));
3774 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
3780 * Concurrency: @dt is read locked.
3782 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
3783 struct lu_buf *buf, const char *name)
3785 struct osd_object *obj = osd_dt_obj(dt);
3786 struct inode *inode = obj->oo_inode;
3787 struct osd_thread_info *info = osd_oti_get(env);
3788 struct dentry *dentry = &info->oti_obj_dentry;
3789 bool cache_xattr = false;
3794 /* version get is not real XATTR but uses xattr API */
3795 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3796 dt_obj_version_t *ver = buf->lb_buf;
3798 /* for version we are just using xattr API but change inode
3800 if (buf->lb_len == 0)
3801 return sizeof(dt_obj_version_t);
3803 if (buf->lb_len < sizeof(dt_obj_version_t))
3806 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
3807 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3809 *ver = LDISKFS_I(inode)->i_fs_version;
3811 return sizeof(dt_obj_version_t);
3814 if (!dt_object_exists(dt))
3817 LASSERT(!dt_object_remote(dt));
3818 LASSERT(inode->i_op != NULL);
3819 LASSERT(inode->i_op->getxattr != NULL);
3821 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
3822 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
3826 rc = osd_oxc_get(obj, name, buf);
3831 rc = __osd_xattr_get(inode, dentry, name, buf->lb_buf, buf->lb_len);
3832 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
3833 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3834 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3835 struct filter_fid *ff;
3836 struct ost_layout *ol;
3838 if (!osd_dev(dt->do_lu.lo_dev)->od_is_ost)
3841 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
3845 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
3851 if (buf->lb_len == 0 || !buf->lb_buf)
3854 if (buf->lb_len < rc)
3858 ol = &ff->ff_layout;
3859 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
3860 PFID_STRIPE_IDX_BITS);
3861 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
3862 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
3863 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
3864 if (lma->lma_compat & LMAC_COMP_INFO) {
3865 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
3866 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
3867 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
3869 ol->ol_comp_start = 0;
3870 ol->ol_comp_end = 0;
3877 if (rc == -ENOENT || rc == -ENODATA)
3878 osd_oxc_add(obj, name, NULL, 0);
3879 else if (rc > 0 && buf->lb_buf != NULL)
3880 osd_oxc_add(obj, name, buf->lb_buf, rc);
3886 static int osd_declare_xattr_set(const struct lu_env *env,
3887 struct dt_object *dt,
3888 const struct lu_buf *buf, const char *name,
3889 int fl, struct thandle *handle)
3891 struct osd_thandle *oh;
3893 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3895 LASSERT(handle != NULL);
3897 oh = container_of0(handle, struct osd_thandle, ot_super);
3898 LASSERT(oh->ot_handle == NULL);
3900 if (strcmp(name, XATTR_NAME_LMA) == 0) {
3901 /* For non-upgrading case, the LMA is set first and
3902 * usually fit inode. But for upgrade case, the LMA
3903 * may be in another separated EA block. */
3904 if (dt_object_exists(dt)) {
3905 if (fl == LU_XATTR_REPLACE)
3910 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3912 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
3913 /* We may need to delete the old PFID EA. */
3914 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
3915 if (fl == LU_XATTR_REPLACE)
3922 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
3927 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
3928 /* learn xattr size from osd_xattr_get if
3929 attribute has not been read yet */
3930 buflen = __osd_xattr_get(
3931 osd_dt_obj(dt)->oo_inode,
3932 &osd_oti_get(env)->oti_obj_dentry,
3937 buflen = buf->lb_len;
3940 if (buflen > sb->s_blocksize) {
3941 credits += osd_calc_bkmap_credits(
3943 (buflen + sb->s_blocksize - 1) >>
3944 sb->s_blocksize_bits);
3948 * xattr set may involve inode quota change, reserve credits for
3949 * dquot_initialize()
3951 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
3954 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
3960 * Concurrency: @dt is write locked.
3962 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
3963 const struct lu_buf *buf, const char *name, int fl,
3964 struct thandle *handle)
3966 struct osd_object *obj = osd_dt_obj(dt);
3967 struct inode *inode = obj->oo_inode;
3968 struct osd_thread_info *info = osd_oti_get(env);
3969 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3970 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3979 /* version set is not real XATTR */
3980 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
3981 dt_obj_version_t *version = buf->lb_buf;
3983 /* for version we are just using xattr API but change inode
3985 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
3987 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
3988 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
3990 LDISKFS_I(inode)->i_fs_version = *version;
3991 /* Version is set after all inode operations are finished,
3992 * so we should mark it dirty here */
3993 ll_dirty_inode(inode, I_DIRTY_DATASYNC);
3998 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
3999 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4002 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4003 if (fl & LU_XATTR_REPLACE)
4004 fs_flags |= XATTR_REPLACE;
4006 if (fl & LU_XATTR_CREATE)
4007 fs_flags |= XATTR_CREATE;
4009 /* For the OST device with 256 bytes inode size by default,
4010 * the PFID EA will be stored together with LMA EA to avoid
4011 * performance trouble. Otherwise the PFID EA can be stored
4012 * independently. LU-8998 */
4013 if (strcmp(name, XATTR_NAME_FID) == 0 &&
4014 LDISKFS_INODE_SIZE(inode->i_sb) <= 256) {
4015 struct dentry *dentry = &info->oti_obj_dentry;
4016 struct filter_fid *ff;
4017 struct ost_layout *ol;
4020 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4023 ol = &ff->ff_layout;
4024 /* Old client does not send stripe information, store
4025 * the PFID EA on disk directly. */
4026 if (buf->lb_len == sizeof(struct lu_fid) ||
4027 ol->ol_stripe_size == 0) {
4028 len = sizeof(struct lu_fid);
4032 if (buf->lb_len != sizeof(*ff))
4035 rc = osd_get_lma(info, inode, dentry, loa);
4036 if (unlikely(rc == -ENODATA)) {
4037 /* Usually for upgarding from old device */
4038 lustre_loa_init(loa, lu_object_fid(&dt->do_lu),
4039 LMAC_FID_ON_OST, 0);
4047 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4048 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4049 PFID_STRIPE_IDX_BITS;
4050 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4051 lma->lma_compat |= LMAC_STRIPE_INFO;
4052 if (ol->ol_comp_id != 0) {
4053 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4054 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4055 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4056 lma->lma_compat |= LMAC_COMP_INFO;
4059 lustre_loa_swab(loa, false);
4061 /* Remove old PFID EA entry firstly. */
4062 ll_vfs_dq_init(inode);
4063 rc = inode->i_op->removexattr(dentry, name);
4064 if (rc && rc != -ENODATA)
4067 /* Store the PFID EA inside the LMA EA. */
4068 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4072 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4073 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4077 lma->lma_incompat |= LMAI_STRIPED;
4078 lustre_lma_swab(lma);
4079 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4080 sizeof(*lma), XATTR_REPLACE);
4086 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4087 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4090 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4091 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4092 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4098 * Concurrency: @dt is read locked.
4100 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4101 const struct lu_buf *buf)
4103 struct osd_object *obj = osd_dt_obj(dt);
4104 struct inode *inode = obj->oo_inode;
4105 struct osd_thread_info *info = osd_oti_get(env);
4106 struct dentry *dentry = &info->oti_obj_dentry;
4108 if (!dt_object_exists(dt))
4111 LASSERT(!dt_object_remote(dt));
4112 LASSERT(inode->i_op != NULL);
4113 LASSERT(inode->i_op->listxattr != NULL);
4115 dentry->d_inode = inode;
4116 dentry->d_sb = inode->i_sb;
4117 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4120 static int osd_declare_xattr_del(const struct lu_env *env,
4121 struct dt_object *dt, const char *name,
4122 struct thandle *handle)
4124 struct osd_thandle *oh;
4125 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4127 LASSERT(!dt_object_remote(dt));
4128 LASSERT(handle != NULL);
4130 oh = container_of0(handle, struct osd_thandle, ot_super);
4131 LASSERT(oh->ot_handle == NULL);
4133 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4134 osd_dto_credits_noquota[DTO_XATTR_SET]);
4136 * xattr del may involve inode quota change, reserve credits for
4137 * dquot_initialize()
4139 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4145 * Concurrency: @dt is write locked.
4147 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4148 const char *name, struct thandle *handle)
4150 struct osd_object *obj = osd_dt_obj(dt);
4151 struct inode *inode = obj->oo_inode;
4152 struct osd_thread_info *info = osd_oti_get(env);
4153 struct dentry *dentry = &info->oti_obj_dentry;
4156 if (!dt_object_exists(dt))
4159 LASSERT(!dt_object_remote(dt));
4160 LASSERT(inode->i_op != NULL);
4161 LASSERT(inode->i_op->removexattr != NULL);
4162 LASSERT(handle != NULL);
4164 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4166 ll_vfs_dq_init(inode);
4167 dentry->d_inode = inode;
4168 dentry->d_sb = inode->i_sb;
4169 rc = inode->i_op->removexattr(dentry, name);
4170 if (rc == -ENODATA && strcmp(name, XATTR_NAME_FID) == 0) {
4171 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4173 LASSERT(osd_dev(dt->do_lu.lo_dev)->od_is_ost);
4175 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4176 &info->oti_ost_attrs);
4178 if (!(lma->lma_compat & LMAC_STRIPE_INFO)) {
4183 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4184 lustre_lma_swab(lma);
4185 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4186 sizeof(*lma), XATTR_REPLACE);
4191 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4194 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4195 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4196 osd_oxc_del(obj, name);
4201 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4202 __u64 start, __u64 end)
4204 struct osd_object *obj = osd_dt_obj(dt);
4205 struct inode *inode = obj->oo_inode;
4206 struct osd_thread_info *info = osd_oti_get(env);
4207 struct dentry *dentry = &info->oti_obj_dentry;
4208 struct file *file = &info->oti_file;
4213 dentry->d_inode = inode;
4214 dentry->d_sb = inode->i_sb;
4215 file->f_path.dentry = dentry;
4216 file->f_mapping = inode->i_mapping;
4217 file->f_op = inode->i_fop;
4218 set_file_inode(file, inode);
4220 rc = ll_vfs_fsync_range(file, start, end, 0);
4225 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4234 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4235 const struct dt_index_features *feat)
4237 struct iam_descr *descr;
4239 if (osd_object_is_root(o))
4240 return feat == &dt_directory_features;
4242 LASSERT(o->oo_dir != NULL);
4244 descr = o->oo_dir->od_container.ic_descr;
4245 if (feat == &dt_directory_features) {
4246 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4252 feat->dif_keysize_min <= descr->id_key_size &&
4253 descr->id_key_size <= feat->dif_keysize_max &&
4254 feat->dif_recsize_min <= descr->id_rec_size &&
4255 descr->id_rec_size <= feat->dif_recsize_max &&
4256 !(feat->dif_flags & (DT_IND_VARKEY |
4257 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4258 ergo(feat->dif_flags & DT_IND_UPDATE,
4259 1 /* XXX check that object (and file system) is
4264 static int osd_iam_container_init(const struct lu_env *env,
4265 struct osd_object *obj,
4266 struct osd_directory *dir)
4268 struct iam_container *bag = &dir->od_container;
4271 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4275 result = iam_container_setup(bag);
4277 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4279 iam_container_fini(bag);
4286 * Concurrency: no external locking is necessary.
4288 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4289 const struct dt_index_features *feat)
4293 struct osd_object *obj = osd_dt_obj(dt);
4295 LINVRNT(osd_invariant(obj));
4297 if (osd_object_is_root(obj)) {
4298 dt->do_index_ops = &osd_index_ea_ops;
4300 } else if (feat == &dt_directory_features) {
4301 dt->do_index_ops = &osd_index_ea_ops;
4302 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4307 } else if (unlikely(feat == &dt_otable_features)) {
4308 dt->do_index_ops = &osd_otable_ops;
4310 } else if (unlikely(feat == &dt_acct_features)) {
4311 dt->do_index_ops = &osd_acct_index_ops;
4314 } else if (!osd_has_index(obj)) {
4315 struct osd_directory *dir;
4320 spin_lock(&obj->oo_guard);
4321 if (obj->oo_dir == NULL)
4325 * Concurrent thread allocated container data.
4328 spin_unlock(&obj->oo_guard);
4330 * Now, that we have container data, serialize its
4333 down_write(&obj->oo_ext_idx_sem);
4335 * recheck under lock.
4337 if (!osd_has_index(obj))
4338 result = osd_iam_container_init(env, obj,
4342 up_write(&obj->oo_ext_idx_sem);
4350 if (result == 0 && skip_iam == 0) {
4351 if (!osd_iam_index_probe(env, obj, feat))
4354 LINVRNT(osd_invariant(obj));
4359 static int osd_otable_it_attr_get(const struct lu_env *env,
4360 struct dt_object *dt,
4361 struct lu_attr *attr)
4367 static const struct dt_object_operations osd_obj_ops = {
4368 .do_read_lock = osd_read_lock,
4369 .do_write_lock = osd_write_lock,
4370 .do_read_unlock = osd_read_unlock,
4371 .do_write_unlock = osd_write_unlock,
4372 .do_write_locked = osd_write_locked,
4373 .do_attr_get = osd_attr_get,
4374 .do_declare_attr_set = osd_declare_attr_set,
4375 .do_attr_set = osd_attr_set,
4376 .do_ah_init = osd_ah_init,
4377 .do_declare_create = osd_declare_create,
4378 .do_create = osd_create,
4379 .do_declare_destroy = osd_declare_destroy,
4380 .do_destroy = osd_destroy,
4381 .do_index_try = osd_index_try,
4382 .do_declare_ref_add = osd_declare_ref_add,
4383 .do_ref_add = osd_ref_add,
4384 .do_declare_ref_del = osd_declare_ref_del,
4385 .do_ref_del = osd_ref_del,
4386 .do_xattr_get = osd_xattr_get,
4387 .do_declare_xattr_set = osd_declare_xattr_set,
4388 .do_xattr_set = osd_xattr_set,
4389 .do_declare_xattr_del = osd_declare_xattr_del,
4390 .do_xattr_del = osd_xattr_del,
4391 .do_xattr_list = osd_xattr_list,
4392 .do_object_sync = osd_object_sync,
4393 .do_invalidate = osd_invalidate,
4396 static const struct dt_object_operations osd_obj_otable_it_ops = {
4397 .do_attr_get = osd_otable_it_attr_get,
4398 .do_index_try = osd_index_try,
4401 static int osd_index_declare_iam_delete(const struct lu_env *env,
4402 struct dt_object *dt,
4403 const struct dt_key *key,
4404 struct thandle *handle)
4406 struct osd_thandle *oh;
4408 oh = container_of0(handle, struct osd_thandle, ot_super);
4409 LASSERT(oh->ot_handle == NULL);
4411 /* Recycle may cause additional three blocks to be changed. */
4412 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4413 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4419 * delete a (key, value) pair from index \a dt specified by \a key
4421 * \param dt osd index object
4422 * \param key key for index
4423 * \param rec record reference
4424 * \param handle transaction handler
4427 * \retval -ve failure
4429 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
4430 const struct dt_key *key,
4431 struct thandle *handle)
4433 struct osd_thread_info *oti = osd_oti_get(env);
4434 struct osd_object *obj = osd_dt_obj(dt);
4435 struct osd_thandle *oh;
4436 struct iam_path_descr *ipd;
4437 struct iam_container *bag = &obj->oo_dir->od_container;
4441 if (!dt_object_exists(dt))
4444 LINVRNT(osd_invariant(obj));
4445 LASSERT(!dt_object_remote(dt));
4446 LASSERT(bag->ic_object == obj->oo_inode);
4447 LASSERT(handle != NULL);
4449 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4451 ipd = osd_idx_ipd_get(env, bag);
4452 if (unlikely(ipd == NULL))
4455 oh = container_of0(handle, struct osd_thandle, ot_super);
4456 LASSERT(oh->ot_handle != NULL);
4457 LASSERT(oh->ot_handle->h_transaction != NULL);
4459 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4460 /* swab quota uid/gid provided by caller */
4461 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4462 key = (const struct dt_key *)&oti->oti_quota_id;
4465 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
4466 osd_ipd_put(env, bag, ipd);
4467 LINVRNT(osd_invariant(obj));
4468 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4472 static int osd_index_declare_ea_delete(const struct lu_env *env,
4473 struct dt_object *dt,
4474 const struct dt_key *key,
4475 struct thandle *handle)
4477 struct osd_thandle *oh;
4478 struct inode *inode;
4482 LASSERT(!dt_object_remote(dt));
4483 LASSERT(handle != NULL);
4485 oh = container_of0(handle, struct osd_thandle, ot_super);
4486 LASSERT(oh->ot_handle == NULL);
4488 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
4489 if (key != NULL && unlikely(strcmp((char *)key, dotdot) == 0)) {
4490 /* '..' to a remote object has a local representative */
4491 credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
4492 /* to reset LMAI_REMOTE_PARENT */
4495 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
4497 inode = osd_dt_obj(dt)->oo_inode;
4501 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4502 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
4507 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
4510 struct osd_fid_pack *rec;
4513 if (de->file_type & LDISKFS_DIRENT_LUFID) {
4514 rec = (struct osd_fid_pack *) (de->name + de->name_len + 1);
4515 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
4516 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
4522 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
4523 const struct lu_fid *fid)
4525 struct seq_server_site *ss = osd_seq_site(osd);
4528 /* FID seqs not in FLDB, must be local seq */
4529 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
4532 /* If FLD is not being initialized yet, it only happens during the
4533 * initialization, likely during mgs initialization, and we assume
4534 * this is local FID. */
4535 if (ss == NULL || ss->ss_server_fld == NULL)
4538 /* Only check the local FLDB here */
4539 if (osd_seq_exists(env, osd, fid_seq(fid)))
4545 static void osd_take_care_of_agent(const struct lu_env *env,
4546 struct osd_device *osd,
4547 struct osd_thandle *oh,
4548 struct ldiskfs_dir_entry_2 *de)
4550 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4551 struct osd_idmap_cache *idc;
4552 int rc, schedule = 0;
4554 LASSERT(de != NULL);
4556 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
4557 if (likely(rc == 0)) {
4558 idc = osd_idc_find_or_init(env, osd, fid);
4559 if (IS_ERR(idc) || idc->oic_remote)
4561 } else if (rc == -ENODATA) {
4562 /* can't get FID, postpone to the end of the
4563 * transaction when iget() is safe */
4566 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
4569 osd_schedule_agent_inode_removal(env, oh,
4570 le32_to_cpu(de->inode));
4574 * Index delete function for interoperability mode (b11826).
4575 * It will remove the directory entry added by osd_index_ea_insert().
4576 * This entry is needed to maintain name->fid mapping.
4578 * \param key, key i.e. file entry to be deleted
4580 * \retval 0, on success
4581 * \retval -ve, on error
4583 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
4584 const struct dt_key *key, struct thandle *handle)
4586 struct osd_object *obj = osd_dt_obj(dt);
4587 struct inode *dir = obj->oo_inode;
4588 struct dentry *dentry;
4589 struct osd_thandle *oh;
4590 struct ldiskfs_dir_entry_2 *de = NULL;
4591 struct buffer_head *bh;
4592 struct htree_lock *hlock = NULL;
4593 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
4594 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4598 if (!dt_object_exists(dt))
4601 LINVRNT(osd_invariant(obj));
4602 LASSERT(!dt_object_remote(dt));
4603 LASSERT(handle != NULL);
4605 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
4607 oh = container_of(handle, struct osd_thandle, ot_super);
4608 LASSERT(oh->ot_handle != NULL);
4609 LASSERT(oh->ot_handle->h_transaction != NULL);
4611 ll_vfs_dq_init(dir);
4612 dentry = osd_child_dentry_get(env, obj,
4613 (char *)key, strlen((char *)key));
4615 if (obj->oo_hl_head != NULL) {
4616 hlock = osd_oti_get(env)->oti_hlock;
4617 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
4618 dir, LDISKFS_HLOCK_DEL);
4620 down_write(&obj->oo_ext_idx_sem);
4623 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
4625 /* If this is not the ".." entry, it might be a remote DNE
4626 * entry and we need to check if the FID is for a remote
4627 * MDT. If the FID is not in the directory entry (e.g.
4628 * upgraded 1.8 filesystem without dirdata enabled) then
4629 * we need to get the FID from the LMA. For a remote directory
4630 * there HAS to be an LMA, it cannot be an IGIF inode in this
4633 * Delete the entry before the agent inode in order to
4634 * simplify error handling. At worst an error after deleting
4635 * the entry first might leak the agent inode afterward. The
4636 * reverse would need filesystem abort in case of error deleting
4637 * the entry after the agent had been removed, or leave a
4638 * dangling entry pointing at a random inode. */
4639 if (strcmp((char *)key, dotdot) != 0)
4640 osd_take_care_of_agent(env, osd, oh, de);
4641 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
4647 ldiskfs_htree_unlock(hlock);
4649 up_write(&obj->oo_ext_idx_sem);
4654 /* For inode on the remote MDT, .. will point to
4655 * /Agent directory, Check whether it needs to delete
4656 * from agent directory */
4657 if (unlikely(strcmp((char *)key, dotdot) == 0)) {
4660 ret = osd_delete_from_remote_parent(env, osd_obj2dev(obj),
4663 /* Sigh, the entry has been deleted, and
4664 * it is not easy to revert it back, so
4665 * let's keep this error private, and let
4666 * LFSCK fix it. XXX */
4667 CERROR("%s: delete remote parent "DFID": rc = %d\n",
4668 osd_name(osd), PFID(fid), ret);
4671 LASSERT(osd_invariant(obj));
4672 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
4677 * Lookup index for \a key and copy record to \a rec.
4679 * \param dt osd index object
4680 * \param key key for index
4681 * \param rec record reference
4683 * \retval +ve success : exact mach
4684 * \retval 0 return record with key not greater than \a key
4685 * \retval -ve failure
4687 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
4688 struct dt_rec *rec, const struct dt_key *key)
4690 struct osd_object *obj = osd_dt_obj(dt);
4691 struct iam_path_descr *ipd;
4692 struct iam_container *bag = &obj->oo_dir->od_container;
4693 struct osd_thread_info *oti = osd_oti_get(env);
4694 struct iam_iterator *it = &oti->oti_idx_it;
4695 struct iam_rec *iam_rec;
4699 if (!dt_object_exists(dt))
4702 LASSERT(osd_invariant(obj));
4703 LASSERT(!dt_object_remote(dt));
4704 LASSERT(bag->ic_object == obj->oo_inode);
4706 ipd = osd_idx_ipd_get(env, bag);
4710 /* got ipd now we can start iterator. */
4711 iam_it_init(it, bag, 0, ipd);
4713 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4714 /* swab quota uid/gid provided by caller */
4715 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4716 key = (const struct dt_key *)&oti->oti_quota_id;
4719 rc = iam_it_get(it, (struct iam_key *)key);
4721 if (S_ISDIR(obj->oo_inode->i_mode))
4722 iam_rec = (struct iam_rec *)oti->oti_ldp;
4724 iam_rec = (struct iam_rec *) rec;
4726 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
4728 if (S_ISDIR(obj->oo_inode->i_mode))
4729 osd_fid_unpack((struct lu_fid *) rec,
4730 (struct osd_fid_pack *)iam_rec);
4731 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
4732 osd_quota_unpack(obj, rec);
4737 osd_ipd_put(env, bag, ipd);
4739 LINVRNT(osd_invariant(obj));
4744 static int osd_index_declare_iam_insert(const struct lu_env *env,
4745 struct dt_object *dt,
4746 const struct dt_rec *rec,
4747 const struct dt_key *key,
4748 struct thandle *handle)
4750 struct osd_thandle *oh;
4752 LASSERT(handle != NULL);
4754 oh = container_of0(handle, struct osd_thandle, ot_super);
4755 LASSERT(oh->ot_handle == NULL);
4757 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
4758 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
4764 * Inserts (key, value) pair in \a dt index object.
4766 * \param dt osd index object
4767 * \param key key for index
4768 * \param rec record reference
4769 * \param th transaction handler
4772 * \retval -ve failure
4774 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
4775 const struct dt_rec *rec,
4776 const struct dt_key *key, struct thandle *th,
4779 struct osd_object *obj = osd_dt_obj(dt);
4780 struct iam_path_descr *ipd;
4781 struct osd_thandle *oh;
4782 struct iam_container *bag;
4783 struct osd_thread_info *oti = osd_oti_get(env);
4784 struct iam_rec *iam_rec;
4788 if (!dt_object_exists(dt))
4791 LINVRNT(osd_invariant(obj));
4792 LASSERT(!dt_object_remote(dt));
4794 bag = &obj->oo_dir->od_container;
4795 LASSERT(bag->ic_object == obj->oo_inode);
4796 LASSERT(th != NULL);
4798 osd_trans_exec_op(env, th, OSD_OT_INSERT);
4800 ipd = osd_idx_ipd_get(env, bag);
4801 if (unlikely(ipd == NULL))
4804 oh = container_of0(th, struct osd_thandle, ot_super);
4805 LASSERT(oh->ot_handle != NULL);
4806 LASSERT(oh->ot_handle->h_transaction != NULL);
4807 if (S_ISDIR(obj->oo_inode->i_mode)) {
4808 iam_rec = (struct iam_rec *)oti->oti_ldp;
4809 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec, &oti->oti_fid);
4810 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
4811 /* pack quota uid/gid */
4812 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
4813 key = (const struct dt_key *)&oti->oti_quota_id;
4814 /* pack quota record */
4815 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
4816 iam_rec = (struct iam_rec *)rec;
4818 iam_rec = (struct iam_rec *)rec;
4821 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
4823 osd_ipd_put(env, bag, ipd);
4824 LINVRNT(osd_invariant(obj));
4825 osd_trans_exec_check(env, th, OSD_OT_INSERT);
4830 * Calls ldiskfs_add_entry() to add directory entry
4831 * into the directory. This is required for
4832 * interoperability mode (b11826)
4834 * \retval 0, on success
4835 * \retval -ve, on error
4837 static int __osd_ea_add_rec(struct osd_thread_info *info,
4838 struct osd_object *pobj, struct inode *cinode,
4839 const char *name, const struct lu_fid *fid,
4840 struct htree_lock *hlock, struct thandle *th)
4842 struct ldiskfs_dentry_param *ldp;
4843 struct dentry *child;
4844 struct osd_thandle *oth;
4847 oth = container_of(th, struct osd_thandle, ot_super);
4848 LASSERT(oth->ot_handle != NULL);
4849 LASSERT(oth->ot_handle->h_transaction != NULL);
4850 LASSERT(pobj->oo_inode);
4852 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4853 if (unlikely(pobj->oo_inode ==
4854 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
4857 osd_get_ldiskfs_dirent_param(ldp, fid);
4858 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
4859 child->d_fsdata = (void *)ldp;
4860 ll_vfs_dq_init(pobj->oo_inode);
4861 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
4862 child, cinode, hlock);
4863 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
4864 struct ldiskfs_dir_entry_2 *de;
4865 struct buffer_head *bh;
4868 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
4871 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
4874 if (S_ISDIR(cinode->i_mode))
4875 de->file_type = LDISKFS_DIRENT_LUFID |
4876 LDISKFS_FT_REG_FILE;
4878 de->file_type = LDISKFS_DIRENT_LUFID |
4880 ldiskfs_handle_dirty_metadata(oth->ot_handle,
4891 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
4892 * into the directory.Also sets flags into osd object to
4893 * indicate dot and dotdot are created. This is required for
4894 * interoperability mode (b11826)
4896 * \param dir directory for dot and dotdot fixup.
4897 * \param obj child object for linking
4899 * \retval 0, on success
4900 * \retval -ve, on error
4902 static int osd_add_dot_dotdot(struct osd_thread_info *info,
4903 struct osd_object *dir,
4904 struct inode *parent_dir, const char *name,
4905 const struct lu_fid *dot_fid,
4906 const struct lu_fid *dot_dot_fid,
4909 struct inode *inode = dir->oo_inode;
4910 struct osd_thandle *oth;
4913 oth = container_of(th, struct osd_thandle, ot_super);
4914 LASSERT(oth->ot_handle->h_transaction != NULL);
4915 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
4917 if (strcmp(name, dot) == 0) {
4918 if (dir->oo_compat_dot_created) {
4921 LASSERT(inode->i_ino == parent_dir->i_ino);
4922 dir->oo_compat_dot_created = 1;
4925 } else if (strcmp(name, dotdot) == 0) {
4926 if (!dir->oo_compat_dot_created)
4928 /* in case of rename, dotdot is already created */
4929 if (dir->oo_compat_dotdot_created) {
4930 return __osd_ea_add_rec(info, dir, parent_dir, name,
4931 dot_dot_fid, NULL, th);
4934 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
4935 struct lu_fid tfid = *dot_dot_fid;
4938 result = osd_add_dot_dotdot_internal(info,
4939 dir->oo_inode, parent_dir, dot_fid,
4942 result = osd_add_dot_dotdot_internal(info,
4943 dir->oo_inode, parent_dir, dot_fid,
4948 dir->oo_compat_dotdot_created = 1;
4956 * It will call the appropriate osd_add* function and return the
4957 * value, return by respective functions.
4959 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
4960 struct inode *cinode, const char *name,
4961 const struct lu_fid *fid, struct thandle *th)
4963 struct osd_thread_info *info = osd_oti_get(env);
4964 struct htree_lock *hlock;
4967 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
4969 if (name[0] == '.' && (name[1] == '\0' || (name[1] == '.' &&
4971 if (hlock != NULL) {
4972 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4975 down_write(&pobj->oo_ext_idx_sem);
4978 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
4979 lu_object_fid(&pobj->oo_dt.do_lu),
4982 if (hlock != NULL) {
4983 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
4984 pobj->oo_inode, LDISKFS_HLOCK_ADD);
4986 down_write(&pobj->oo_ext_idx_sem);
4989 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
4990 struct lu_fid *tfid = &info->oti_fid;
4994 rc = __osd_ea_add_rec(info, pobj, cinode, name,
4997 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5002 ldiskfs_htree_unlock(hlock);
5004 up_write(&pobj->oo_ext_idx_sem);
5010 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5011 struct osd_idmap_cache *oic)
5013 struct osd_scrub *scrub = &dev->od_scrub;
5014 struct lu_fid *fid = &oic->oic_fid;
5015 struct osd_inode_id *id = &oic->oic_lid;
5016 struct inode *inode = NULL;
5022 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5025 if (scrub->os_pos_current > id->oii_ino)
5029 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5030 if (rc == -ENOENT) {
5031 __u32 gen = id->oii_gen;
5037 inode = osd_iget(oti, dev, id);
5038 /* The inode has been removed (by race maybe). */
5039 if (IS_ERR(inode)) {
5040 rc = PTR_ERR(inode);
5042 RETURN(rc == -ESTALE ? -ENOENT : rc);
5045 /* The OI mapping is lost. */
5046 if (gen != OSD_OII_NOGEN)
5049 /* The inode may has been reused by others, we do not know,
5050 * leave it to be handled by subsequent osd_fid_lookup(). */
5052 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5059 if (thread_is_running(&scrub->os_thread)) {
5060 if (inode == NULL) {
5061 inode = osd_iget(oti, dev, id);
5062 /* The inode has been removed (by race maybe). */
5063 if (IS_ERR(inode)) {
5064 rc = PTR_ERR(inode);
5066 RETURN(rc == -ESTALE ? -ENOENT : rc);
5070 rc = osd_oii_insert(dev, oic, insert);
5071 /* There is race condition between osd_oi_lookup and OI scrub.
5072 * The OI scrub finished just after osd_oi_lookup() failure.
5073 * Under such case, it is unnecessary to trigger OI scrub again,
5074 * but try to call osd_oi_lookup() again. */
5075 if (unlikely(rc == -EAGAIN))
5078 if (!S_ISDIR(inode->i_mode))
5081 rc = osd_check_lmv(oti, dev, inode, oic);
5086 if (!dev->od_noscrub && ++once == 1) {
5087 rc = osd_scrub_start(dev, SS_AUTO_PARTIAL | SS_CLEAR_DRYRUN |
5089 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5090 "%s: trigger partial OI scrub for RPC inconsistency "
5091 "checking FID "DFID": rc = %d\n",
5092 osd_dev2name(dev), PFID(fid), rc);
5093 if (rc == 0 || rc == -EALREADY)
5106 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5107 struct osd_device *dev,
5108 struct osd_idmap_cache *oic,
5109 struct lu_fid *fid, __u32 ino)
5111 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5112 struct inode *inode;
5115 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5116 inode = osd_iget(oti, dev, &oic->oic_lid);
5117 if (IS_ERR(inode)) {
5118 fid_zero(&oic->oic_fid);
5119 return PTR_ERR(inode);
5122 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5125 fid_zero(&oic->oic_fid);
5127 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5131 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5132 struct osd_inode_id *id, const struct lu_fid *fid)
5134 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5135 id->oii_ino, id->oii_gen, info);
5136 info->oti_cache.oic_lid = *id;
5137 info->oti_cache.oic_fid = *fid;
5138 info->oti_cache.oic_dev = osd;
5142 * Get parent FID from the linkEA.
5144 * For a directory which parent resides on remote MDT, to satisfy the
5145 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5146 * the other hand, to make the lookup(..) on the directory can return
5147 * the real parent FID, we append the real parent FID after its ".."
5148 * name entry in the /REMOTE_PARENT_DIR.
5150 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5151 * backup. So after the restore, we cannot get the right parent FID from
5152 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5153 * we have stored the real parent FID in the directory object's linkEA,
5154 * we can parse the linkEA for the real parent FID.
5156 * \param[in] env pointer to the thread context
5157 * \param[in] obj pointer to the object to be handled
5158 * \param[out]fid pointer to the buffer to hold the parent FID
5160 * \retval 0 for getting the real parent FID successfully
5161 * \retval negative error number on failure
5163 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5164 struct osd_object *obj,
5167 struct osd_thread_info *oti = osd_oti_get(env);
5168 struct lu_buf *buf = &oti->oti_big_buf;
5169 struct dentry *dentry = &oti->oti_obj_dentry;
5170 struct inode *inode = obj->oo_inode;
5171 struct linkea_data ldata = { NULL };
5176 if (!S_ISDIR(inode->i_mode))
5180 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5181 buf->lb_buf, buf->lb_len);
5182 if (rc == -ERANGE) {
5183 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5186 lu_buf_realloc(buf, rc);
5187 if (buf->lb_buf == NULL)
5194 if (unlikely(rc == 0))
5200 if (unlikely(buf->lb_buf == NULL)) {
5201 lu_buf_realloc(buf, rc);
5202 if (buf->lb_buf == NULL)
5209 rc = linkea_init_with_rec(&ldata);
5211 linkea_first_entry(&ldata);
5212 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5218 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5219 struct inode *inode,
5220 const struct lu_fid *pfid,
5221 const char *name, const int namelen)
5223 struct osd_thread_info *oti = osd_oti_get(env);
5224 struct lu_buf *buf = &oti->oti_big_buf;
5225 struct dentry *dentry = &oti->oti_obj_dentry;
5226 struct linkea_data ldata = { NULL };
5227 struct lu_name cname = { .ln_name = name,
5228 .ln_namelen = namelen };
5233 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5234 buf->lb_buf, buf->lb_len);
5236 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5241 if (unlikely(rc == 0))
5244 if (buf->lb_len < rc) {
5245 lu_buf_realloc(buf, rc);
5246 if (buf->lb_buf == NULL)
5253 rc = linkea_init_with_rec(&ldata);
5255 rc = linkea_links_find(&ldata, &cname, pfid);
5261 * Calls ->lookup() to find dentry. From dentry get inode and
5262 * read inode's ea to get fid. This is required for interoperability
5265 * \retval 0, on success
5266 * \retval -ve, on error
5268 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5269 struct dt_rec *rec, const struct dt_key *key)
5271 struct inode *dir = obj->oo_inode;
5272 struct dentry *dentry;
5273 struct ldiskfs_dir_entry_2 *de;
5274 struct buffer_head *bh;
5275 struct lu_fid *fid = (struct lu_fid *) rec;
5276 struct htree_lock *hlock = NULL;
5281 LASSERT(dir->i_op != NULL);
5282 LASSERT(dir->i_op->lookup != NULL);
5284 dentry = osd_child_dentry_get(env, obj,
5285 (char *)key, strlen((char *)key));
5287 if (obj->oo_hl_head != NULL) {
5288 hlock = osd_oti_get(env)->oti_hlock;
5289 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5290 dir, LDISKFS_HLOCK_LOOKUP);
5292 down_read(&obj->oo_ext_idx_sem);
5295 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5297 struct osd_thread_info *oti = osd_oti_get(env);
5298 struct osd_inode_id *id = &oti->oti_id;
5299 struct osd_idmap_cache *oic = &oti->oti_cache;
5300 struct osd_device *dev = osd_obj2dev(obj);
5302 ino = le32_to_cpu(de->inode);
5303 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
5305 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
5309 rc = osd_get_fid_from_dentry(de, rec);
5311 /* done with de, release bh */
5314 if (unlikely(is_remote_parent_ino(dev, ino))) {
5315 const char *name = (const char *)key;
5317 /* If the parent is on remote MDT, and there
5318 * is no FID-in-dirent, then we have to get
5319 * the parent FID from the linkEA. */
5320 if (likely(strlen(name) == 2 &&
5321 name[0] == '.' && name[1] == '.'))
5322 rc = osd_get_pfid_from_linkea(env, obj,
5325 rc = osd_ea_fid_get(env, obj, ino, fid, id);
5328 osd_id_gen(id, ino, OSD_OII_NOGEN);
5331 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
5332 fid_zero(&oic->oic_fid);
5337 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
5338 rc = osd_consistency_check(oti, dev, oic);
5340 fid_zero(&oic->oic_fid);
5349 ldiskfs_htree_unlock(hlock);
5351 up_read(&obj->oo_ext_idx_sem);
5355 static int osd_index_declare_ea_insert(const struct lu_env *env,
5356 struct dt_object *dt,
5357 const struct dt_rec *rec,
5358 const struct dt_key *key,
5359 struct thandle *handle)
5361 struct osd_thandle *oh;
5362 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5363 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5364 const struct lu_fid *fid = rec1->rec_fid;
5365 int credits, rc = 0;
5366 struct osd_idmap_cache *idc;
5369 LASSERT(!dt_object_remote(dt));
5370 LASSERT(handle != NULL);
5371 LASSERT(fid != NULL);
5372 LASSERT(rec1->rec_type != 0);
5374 oh = container_of0(handle, struct osd_thandle, ot_super);
5375 LASSERT(oh->ot_handle == NULL);
5377 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
5379 /* we can't call iget() while a transactions is running
5380 * (this can lead to a deadlock), but we need to know
5381 * inum and object type. so we find this information at
5382 * declaration and cache in per-thread info */
5383 idc = osd_idc_find_or_init(env, osd, fid);
5385 RETURN(PTR_ERR(idc));
5386 if (idc->oic_remote) {
5387 /* a reference to remote inode is represented by an
5388 * agent inode which we have to create */
5389 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
5390 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
5393 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
5395 if (osd_dt_obj(dt)->oo_inode != NULL) {
5396 struct inode *inode = osd_dt_obj(dt)->oo_inode;
5398 /* We ignore block quota on meta pool (MDTs), so needn't
5399 * calculate how many blocks will be consumed by this index
5401 rc = osd_declare_inode_qid(env, i_uid_read(inode),
5403 i_projid_read(inode), 0,
5404 oh, osd_dt_obj(dt), NULL,
5412 * Index add function for interoperability mode (b11826).
5413 * It will add the directory entry.This entry is needed to
5414 * maintain name->fid mapping.
5416 * \param key it is key i.e. file entry to be inserted
5417 * \param rec it is value of given key i.e. fid
5419 * \retval 0, on success
5420 * \retval -ve, on error
5422 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
5423 const struct dt_rec *rec,
5424 const struct dt_key *key, struct thandle *th,
5427 struct osd_object *obj = osd_dt_obj(dt);
5428 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5429 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
5430 const struct lu_fid *fid = rec1->rec_fid;
5431 const char *name = (const char *)key;
5432 struct osd_thread_info *oti = osd_oti_get(env);
5433 struct inode *child_inode = NULL;
5434 struct osd_idmap_cache *idc;
5438 if (!dt_object_exists(dt))
5441 LASSERT(osd_invariant(obj));
5442 LASSERT(!dt_object_remote(dt));
5443 LASSERT(th != NULL);
5445 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5447 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
5449 idc = osd_idc_find(env, osd, fid);
5450 if (unlikely(idc == NULL)) {
5451 /* this dt_insert() wasn't declared properly, so
5452 * FID is missing in OI cache. we better do not
5453 * lookup FID in FLDB/OI and don't risk to deadlock,
5454 * but in some special cases (lfsck testing, etc)
5455 * it's much simpler than fixing a caller */
5456 CERROR("%s: "DFID" wasn't declared for insert\n",
5457 osd_name(osd), PFID(fid));
5459 idc = osd_idc_find_or_init(env, osd, fid);
5461 RETURN(PTR_ERR(idc));
5464 if (idc->oic_remote) {
5465 /* Insert remote entry */
5466 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
5467 struct osd_mdobj_map *omm = osd->od_mdt_map;
5468 struct osd_thandle *oh;
5470 /* If parent on remote MDT, we need put this object
5472 oh = container_of(th, typeof(*oh), ot_super);
5473 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5475 CERROR("%s: add "DFID" error: rc = %d\n",
5477 PFID(lu_object_fid(&dt->do_lu)), rc);
5481 child_inode = igrab(omm->omm_remote_parent->d_inode);
5483 child_inode = osd_create_local_agent_inode(env, osd,
5484 obj, fid, rec1->rec_type & S_IFMT, th);
5485 if (IS_ERR(child_inode))
5486 RETURN(PTR_ERR(child_inode));
5489 /* Insert local entry */
5490 if (unlikely(idc->oic_lid.oii_ino == 0)) {
5491 /* for a reason OI cache wasn't filled properly */
5492 CERROR("%s: OIC for "DFID" isn't filled\n",
5493 osd_name(osd), PFID(fid));
5496 child_inode = oti->oti_inode;
5497 if (unlikely(child_inode == NULL)) {
5498 struct ldiskfs_inode_info *lii;
5502 child_inode = oti->oti_inode = &lii->vfs_inode;
5504 child_inode->i_sb = osd_sb(osd);
5505 child_inode->i_ino = idc->oic_lid.oii_ino;
5506 child_inode->i_mode = rec1->rec_type & S_IFMT;
5509 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
5511 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
5512 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
5514 if (child_inode && child_inode != oti->oti_inode)
5516 LASSERT(osd_invariant(obj));
5517 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5522 * Initialize osd Iterator for given osd index object.
5524 * \param dt osd index object
5527 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
5528 struct dt_object *dt,
5531 struct osd_it_iam *it;
5532 struct osd_object *obj = osd_dt_obj(dt);
5533 struct lu_object *lo = &dt->do_lu;
5534 struct iam_path_descr *ipd;
5535 struct iam_container *bag = &obj->oo_dir->od_container;
5537 if (!dt_object_exists(dt))
5538 return ERR_PTR(-ENOENT);
5542 return ERR_PTR(-ENOMEM);
5544 ipd = osd_it_ipd_get(env, bag);
5545 if (likely(ipd != NULL)) {
5549 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
5550 return (struct dt_it *)it;
5553 return ERR_PTR(-ENOMEM);
5558 * free given Iterator.
5560 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
5562 struct osd_it_iam *it = (struct osd_it_iam *)di;
5563 struct osd_object *obj = it->oi_obj;
5565 iam_it_fini(&it->oi_it);
5566 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
5567 osd_object_put(env, obj);
5572 * Move Iterator to record specified by \a key
5574 * \param di osd iterator
5575 * \param key key for index
5577 * \retval +ve di points to record with least key not larger than key
5578 * \retval 0 di points to exact matched key
5579 * \retval -ve failure
5582 static int osd_it_iam_get(const struct lu_env *env,
5583 struct dt_it *di, const struct dt_key *key)
5585 struct osd_thread_info *oti = osd_oti_get(env);
5586 struct osd_it_iam *it = (struct osd_it_iam *)di;
5588 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5589 /* swab quota uid/gid */
5590 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5591 key = (struct dt_key *)&oti->oti_quota_id;
5594 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
5600 * \param di osd iterator
5602 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
5604 struct osd_it_iam *it = (struct osd_it_iam *)di;
5606 iam_it_put(&it->oi_it);
5610 * Move iterator by one record
5612 * \param di osd iterator
5614 * \retval +1 end of container reached
5616 * \retval -ve failure
5619 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
5621 struct osd_it_iam *it = (struct osd_it_iam *)di;
5623 return iam_it_next(&it->oi_it);
5627 * Return pointer to the key under iterator.
5630 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
5631 const struct dt_it *di)
5633 struct osd_thread_info *oti = osd_oti_get(env);
5634 struct osd_it_iam *it = (struct osd_it_iam *)di;
5635 struct osd_object *obj = it->oi_obj;
5638 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
5640 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
5641 /* swab quota uid/gid */
5642 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
5643 key = (struct dt_key *)&oti->oti_quota_id;
5650 * Return size of key under iterator (in bytes)
5653 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
5655 struct osd_it_iam *it = (struct osd_it_iam *)di;
5657 return iam_it_key_size(&it->oi_it);
5661 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
5663 /* check if file type is required */
5664 if (ent->lde_attrs & LUDA_TYPE) {
5665 struct luda_type *lt;
5666 int align = sizeof(*lt) - 1;
5668 len = (len + align) & ~align;
5669 lt = (struct luda_type *)(ent->lde_name + len);
5670 lt->lt_type = cpu_to_le16(DTTOIF(type));
5673 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
5677 * build lu direct from backend fs dirent.
5681 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
5682 char *name, __u16 namelen, __u16 type, __u32 attr)
5684 ent->lde_attrs = attr | LUDA_FID;
5685 fid_cpu_to_le(&ent->lde_fid, fid);
5687 ent->lde_hash = cpu_to_le64(offset);
5688 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
5690 strncpy(ent->lde_name, name, namelen);
5691 ent->lde_name[namelen] = '\0';
5692 ent->lde_namelen = cpu_to_le16(namelen);
5694 /* append lustre attributes */
5695 osd_it_append_attrs(ent, namelen, type);
5699 * Return pointer to the record under iterator.
5701 static int osd_it_iam_rec(const struct lu_env *env,
5702 const struct dt_it *di,
5703 struct dt_rec *dtrec, __u32 attr)
5705 struct osd_it_iam *it = (struct osd_it_iam *)di;
5706 struct osd_thread_info *info = osd_oti_get(env);
5709 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
5710 const struct osd_fid_pack *rec;
5711 struct lu_fid *fid = &info->oti_fid;
5712 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
5718 name = (char *)iam_it_key_get(&it->oi_it);
5720 RETURN(PTR_ERR(name));
5722 namelen = iam_it_key_size(&it->oi_it);
5724 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
5726 RETURN(PTR_ERR(rec));
5728 rc = osd_fid_unpack(fid, rec);
5732 hash = iam_it_store(&it->oi_it);
5734 /* IAM does not store object type in IAM index (dir) */
5735 osd_it_pack_dirent(lde, fid, hash, name, namelen,
5737 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
5738 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5739 (struct iam_rec *)dtrec);
5740 osd_quota_unpack(it->oi_obj, dtrec);
5742 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
5743 (struct iam_rec *)dtrec);
5750 * Returns cookie for current Iterator position.
5752 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
5754 struct osd_it_iam *it = (struct osd_it_iam *)di;
5756 return iam_it_store(&it->oi_it);
5760 * Restore iterator from cookie.
5762 * \param di osd iterator
5763 * \param hash Iterator location cookie
5765 * \retval +ve di points to record with least key not larger than key.
5766 * \retval 0 di points to exact matched key
5767 * \retval -ve failure
5770 static int osd_it_iam_load(const struct lu_env *env,
5771 const struct dt_it *di, __u64 hash)
5773 struct osd_it_iam *it = (struct osd_it_iam *)di;
5775 return iam_it_load(&it->oi_it, hash);
5778 static const struct dt_index_operations osd_index_iam_ops = {
5779 .dio_lookup = osd_index_iam_lookup,
5780 .dio_declare_insert = osd_index_declare_iam_insert,
5781 .dio_insert = osd_index_iam_insert,
5782 .dio_declare_delete = osd_index_declare_iam_delete,
5783 .dio_delete = osd_index_iam_delete,
5785 .init = osd_it_iam_init,
5786 .fini = osd_it_iam_fini,
5787 .get = osd_it_iam_get,
5788 .put = osd_it_iam_put,
5789 .next = osd_it_iam_next,
5790 .key = osd_it_iam_key,
5791 .key_size = osd_it_iam_key_size,
5792 .rec = osd_it_iam_rec,
5793 .store = osd_it_iam_store,
5794 .load = osd_it_iam_load
5800 * Creates or initializes iterator context.
5802 * \retval struct osd_it_ea, iterator structure on success
5805 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
5806 struct dt_object *dt,
5809 struct osd_object *obj = osd_dt_obj(dt);
5810 struct osd_thread_info *info = osd_oti_get(env);
5811 struct osd_it_ea *oie;
5813 struct lu_object *lo = &dt->do_lu;
5814 struct dentry *obj_dentry;
5817 if (!dt_object_exists(dt) || obj->oo_destroyed)
5818 RETURN(ERR_PTR(-ENOENT));
5820 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
5822 RETURN(ERR_PTR(-ENOMEM));
5823 obj_dentry = &oie->oie_dentry;
5825 obj_dentry->d_inode = obj->oo_inode;
5826 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
5827 obj_dentry->d_name.hash = 0;
5829 oie->oie_rd_dirent = 0;
5830 oie->oie_it_dirent = 0;
5831 oie->oie_dirent = NULL;
5832 if (unlikely(!info->oti_it_ea_buf_used)) {
5833 oie->oie_buf = info->oti_it_ea_buf;
5834 info->oti_it_ea_buf_used = 1;
5836 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5837 if (oie->oie_buf == NULL)
5838 RETURN(ERR_PTR(-ENOMEM));
5842 file = &oie->oie_file;
5844 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
5845 if (attr & LUDA_64BITHASH)
5846 file->f_mode = FMODE_64BITHASH;
5848 file->f_mode = FMODE_32BITHASH;
5849 file->f_path.dentry = obj_dentry;
5850 file->f_mapping = obj->oo_inode->i_mapping;
5851 file->f_op = obj->oo_inode->i_fop;
5852 set_file_inode(file, obj->oo_inode);
5855 RETURN((struct dt_it *) oie);
5859 * Destroy or finishes iterator context.
5861 * \param di iterator structure to be destroyed
5863 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
5865 struct osd_thread_info *info = osd_oti_get(env);
5866 struct osd_it_ea *oie = (struct osd_it_ea *)di;
5867 struct osd_object *obj = oie->oie_obj;
5868 struct inode *inode = obj->oo_inode;
5871 oie->oie_file.f_op->release(inode, &oie->oie_file);
5872 osd_object_put(env, obj);
5873 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
5874 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
5876 info->oti_it_ea_buf_used = 0;
5877 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
5882 * It position the iterator at given key, so that next lookup continues from
5883 * that key Or it is similar to dio_it->load() but based on a key,
5884 * rather than file position.
5886 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
5889 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
5891 static int osd_it_ea_get(const struct lu_env *env,
5892 struct dt_it *di, const struct dt_key *key)
5894 struct osd_it_ea *it = (struct osd_it_ea *)di;
5897 LASSERT(((const char *)key)[0] == '\0');
5898 it->oie_file.f_pos = 0;
5899 it->oie_rd_dirent = 0;
5900 it->oie_it_dirent = 0;
5901 it->oie_dirent = NULL;
5909 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
5913 struct osd_filldir_cbs {
5914 #ifdef HAVE_DIR_CONTEXT
5915 struct dir_context ctx;
5917 struct osd_it_ea *it;
5920 * It is called internally by ->readdir(). It fills the
5921 * iterator's in-memory data structure with required
5922 * information i.e. name, namelen, rec_size etc.
5924 * \param buf in which information to be filled in.
5925 * \param name name of the file in given dir
5927 * \retval 0 on success
5928 * \retval 1 on buffer full
5930 #ifdef HAVE_FILLDIR_USE_CTX
5931 static int osd_ldiskfs_filldir(struct dir_context *buf,
5933 static int osd_ldiskfs_filldir(void *buf,
5935 const char *name, int namelen,
5936 loff_t offset, __u64 ino, unsigned d_type)
5938 struct osd_it_ea *it =
5939 ((struct osd_filldir_cbs *)buf)->it;
5940 struct osd_object *obj = it->oie_obj;
5941 struct osd_it_ea_dirent *ent = it->oie_dirent;
5942 struct lu_fid *fid = &ent->oied_fid;
5943 struct osd_fid_pack *rec;
5946 /* this should never happen */
5947 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
5948 CERROR("ldiskfs return invalid namelen %d\n", namelen);
5952 if ((void *) ent - it->oie_buf + sizeof(*ent) + namelen >
5956 /* "." is just the object itself. */
5957 if (namelen == 1 && name[0] == '.') {
5958 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5959 } else if (d_type & LDISKFS_DIRENT_LUFID) {
5960 rec = (struct osd_fid_pack*) (name + namelen + 1);
5961 if (osd_fid_unpack(fid, rec) != 0)
5966 d_type &= ~LDISKFS_DIRENT_LUFID;
5968 /* NOT export local root. */
5969 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
5970 ino = obj->oo_inode->i_ino;
5971 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
5974 ent->oied_ino = ino;
5975 ent->oied_off = offset;
5976 ent->oied_namelen = namelen;
5977 ent->oied_type = d_type;
5979 memcpy(ent->oied_name, name, namelen);
5981 it->oie_rd_dirent++;
5982 it->oie_dirent = (void *) ent + cfs_size_round(sizeof(*ent) + namelen);
5987 * Calls ->readdir() to load a directory entry at a time
5988 * and stored it in iterator's in-memory data structure.
5990 * \param di iterator's in memory structure
5992 * \retval 0 on success
5993 * \retval -ve on error
5994 * \retval +1 reach the end of entry
5996 static int osd_ldiskfs_it_fill(const struct lu_env *env,
5997 const struct dt_it *di)
5999 struct osd_it_ea *it = (struct osd_it_ea *)di;
6000 struct osd_object *obj = it->oie_obj;
6001 struct inode *inode = obj->oo_inode;
6002 struct htree_lock *hlock = NULL;
6003 struct file *filp = &it->oie_file;
6005 struct osd_filldir_cbs buf = {
6006 #ifdef HAVE_DIR_CONTEXT
6007 .ctx.actor = osd_ldiskfs_filldir,
6013 it->oie_dirent = it->oie_buf;
6014 it->oie_rd_dirent = 0;
6016 if (obj->oo_hl_head != NULL) {
6017 hlock = osd_oti_get(env)->oti_hlock;
6018 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6019 inode, LDISKFS_HLOCK_READDIR);
6021 down_read(&obj->oo_ext_idx_sem);
6024 #ifdef HAVE_DIR_CONTEXT
6025 buf.ctx.pos = filp->f_pos;
6026 rc = inode->i_fop->iterate(filp, &buf.ctx);
6027 filp->f_pos = buf.ctx.pos;
6029 rc = inode->i_fop->readdir(filp, &buf, osd_ldiskfs_filldir);
6033 ldiskfs_htree_unlock(hlock);
6035 up_read(&obj->oo_ext_idx_sem);
6037 if (it->oie_rd_dirent == 0) {
6038 /*If it does not get any dirent, it means it has been reached
6039 *to the end of the dir */
6040 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6044 it->oie_dirent = it->oie_buf;
6045 it->oie_it_dirent = 1;
6052 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6053 * to load a directory entry at a time and stored it in
6054 * iterator's in-memory data structure.
6056 * \param di iterator's in memory structure
6058 * \retval +ve iterator reached to end
6059 * \retval 0 iterator not reached to end
6060 * \retval -ve on error
6062 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6064 struct osd_it_ea *it = (struct osd_it_ea *)di;
6069 if (it->oie_it_dirent < it->oie_rd_dirent) {
6071 (void *) it->oie_dirent +
6072 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6073 it->oie_dirent->oied_namelen);
6074 it->oie_it_dirent++;
6077 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6080 rc = osd_ldiskfs_it_fill(env, di);
6087 * Returns the key at current position from iterator's in memory structure.
6089 * \param di iterator's in memory structure
6091 * \retval key i.e. struct dt_key on success
6093 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6094 const struct dt_it *di)
6096 struct osd_it_ea *it = (struct osd_it_ea *)di;
6098 return (struct dt_key *)it->oie_dirent->oied_name;
6102 * Returns the key's size at current position from iterator's in memory structure.
6104 * \param di iterator's in memory structure
6106 * \retval key_size i.e. struct dt_key on success
6108 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6110 struct osd_it_ea *it = (struct osd_it_ea *)di;
6112 return it->oie_dirent->oied_namelen;
6115 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6117 if (LDISKFS_DIR_REC_LEN(de) >=
6118 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6125 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6126 unsigned blocksize, bool dotdot)
6129 return osd_dotdot_has_space(de);
6131 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6132 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6139 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6140 handle_t *jh, struct dentry *dentry,
6141 const struct lu_fid *fid, struct buffer_head *bh,
6142 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6145 struct inode *dir = dentry->d_parent->d_inode;
6146 struct inode *inode = dentry->d_inode;
6147 struct osd_fid_pack *rec;
6148 struct ldiskfs_dentry_param *ldp;
6149 int namelen = dentry->d_name.len;
6151 struct osd_thread_info *info = osd_oti_get(env);
6154 if (!LDISKFS_HAS_INCOMPAT_FEATURE(inode->i_sb,
6155 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
6158 /* There is enough space to hold the FID-in-dirent. */
6159 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6160 rc = ldiskfs_journal_get_write_access(jh, bh);
6164 de->name[namelen] = 0;
6165 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6166 rec->fp_len = sizeof(struct lu_fid) + 1;
6167 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6168 de->file_type |= LDISKFS_DIRENT_LUFID;
6169 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6176 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6180 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6181 osd_get_ldiskfs_dirent_param(ldp, fid);
6182 dentry->d_fsdata = (void *)ldp;
6183 ll_vfs_dq_init(dir);
6184 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6185 /* It is too bad, we cannot reinsert the name entry back.
6186 * That means we lose it! */
6188 CDEBUG(D_LFSCK, "%s: fail to reinsert the dirent, "
6189 "dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6190 osd_ino2name(inode),
6191 dir->i_ino, dir->i_generation, namelen,
6192 dentry->d_name.name, PFID(fid), rc);
6198 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6199 struct osd_it_ea *it, struct lu_fid *fid,
6200 struct osd_inode_id *id, __u32 *attr)
6202 struct osd_thread_info *info = osd_oti_get(env);
6203 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6204 struct osd_device *dev = osd_obj2dev(obj);
6205 struct super_block *sb = osd_sb(dev);
6206 const char *devname = osd_name(dev);
6207 struct osd_it_ea_dirent *ent = it->oie_dirent;
6208 struct inode *dir = obj->oo_inode;
6209 struct htree_lock *hlock = NULL;
6210 struct buffer_head *bh = NULL;
6211 handle_t *jh = NULL;
6212 struct ldiskfs_dir_entry_2 *de;
6213 struct dentry *dentry;
6214 struct inode *inode;
6215 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6218 bool dotdot = false;
6222 if (ent->oied_name[0] == '.') {
6223 if (ent->oied_namelen == 1)
6226 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6230 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6231 inode = osd_iget(info, dev, id);
6232 if (IS_ERR(inode)) {
6233 rc = PTR_ERR(inode);
6234 if (rc == -ENOENT || rc == -ESTALE) {
6235 /* Maybe dangling name entry, or
6236 * corrupted directory entry. */
6237 *attr |= LUDA_UNKNOWN;
6240 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6241 "check_repair, dir = %lu/%u, name = %.*s, "
6242 "ino = %llu, rc = %d\n",
6243 devname, dir->i_ino, dir->i_generation,
6244 ent->oied_namelen, ent->oied_name,
6251 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
6253 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
6254 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
6259 /* We need to ensure that the name entry is still valid.
6260 * Because it may be removed or renamed by other already.
6262 * The unlink or rename operation will start journal before PDO lock,
6263 * so to avoid deadlock, here we need to start journal handle before
6264 * related PDO lock also. But because we do not know whether there
6265 * will be something to be repaired before PDO lock, we just start
6266 * journal without conditions.
6268 * We may need to remove the name entry firstly, then insert back.
6269 * One credit is for user quota file update.
6270 * One credit is for group quota file update.
6271 * Two credits are for dirty inode. */
6272 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
6273 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
6275 if (dev->od_dirent_journal != 0) {
6278 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
6281 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
6282 "check_repair, dir = %lu/%u, credits = %d, "
6283 "name = %.*s, ino = %llu: rc = %d\n",
6284 devname, dir->i_ino, dir->i_generation, credits,
6285 ent->oied_namelen, ent->oied_name,
6288 GOTO(out_inode, rc);
6291 if (obj->oo_hl_head != NULL) {
6292 hlock = osd_oti_get(env)->oti_hlock;
6293 /* "0" means exclusive lock for the whole directory.
6294 * We need to prevent others access such name entry
6295 * during the delete + insert. Neither HLOCK_ADD nor
6296 * HLOCK_DEL cannot guarantee the atomicity. */
6297 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
6299 down_write(&obj->oo_ext_idx_sem);
6302 if (obj->oo_hl_head != NULL) {
6303 hlock = osd_oti_get(env)->oti_hlock;
6304 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
6305 LDISKFS_HLOCK_LOOKUP);
6307 down_read(&obj->oo_ext_idx_sem);
6311 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6312 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
6313 *attr |= LUDA_IGNORE;
6318 /* For dotdot entry, if there is not enough space to hold the
6319 * FID-in-dirent, just keep them there. It only happens when the
6320 * device upgraded from 1.8 or restored from MDT file-level backup.
6321 * For the whole directory, only dotdot entry have no FID-in-dirent
6322 * and needs to get FID from LMA when readdir, it will not affect the
6323 * performance much. */
6324 if (dotdot && !osd_dotdot_has_space(de)) {
6325 *attr |= LUDA_UNKNOWN;
6331 if (lu_fid_eq(fid, &lma->lma_self_fid))
6334 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
6335 struct lu_fid *tfid = &lma->lma_self_fid;
6337 if (likely(dotdot &&
6338 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
6339 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
6340 /* It must be REMOTE_PARENT_DIR and as the
6341 * 'dotdot' entry of remote directory */
6342 *attr |= LUDA_IGNORE;
6344 CDEBUG(D_LFSCK, "%s: expect remote agent "
6345 "parent directory, but got %.*s under "
6346 "dir = %lu/%u with the FID "DFID"\n",
6347 devname, ent->oied_namelen,
6348 ent->oied_name, dir->i_ino,
6349 dir->i_generation, PFID(tfid));
6351 *attr |= LUDA_UNKNOWN;
6358 if (!fid_is_zero(fid)) {
6359 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
6361 if (rc == -ENOENT ||
6363 !(dev->od_scrub.os_file.sf_flags & SF_UPGRADE))) {
6364 /* linkEA does not recognize the dirent entry,
6365 * it may because the dirent entry corruption
6366 * and points to other's inode. */
6367 CDEBUG(D_LFSCK, "%s: the target inode does not "
6368 "recognize the dirent, dir = %lu/%u, "
6369 " name = %.*s, ino = %llu, "
6370 DFID": rc = %d\n", devname, dir->i_ino,
6371 dir->i_generation, ent->oied_namelen,
6372 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6373 *attr |= LUDA_UNKNOWN;
6378 if (rc && rc != -ENODATA) {
6379 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
6380 "dir = %lu/%u, name = %.*s, ino = %llu, "
6381 DFID": rc = %d\n", devname, dir->i_ino,
6382 dir->i_generation, ent->oied_namelen,
6383 ent->oied_name, ent->oied_ino, PFID(fid), rc);
6384 *attr |= LUDA_UNKNOWN;
6391 /* linkEA recognizes the dirent entry, the FID-in-LMA is
6392 * valid, trusted, in spite of fid_is_sane(fid) or not. */
6393 if (*attr & LUDA_VERIFY_DRYRUN) {
6394 *fid = lma->lma_self_fid;
6395 *attr |= LUDA_REPAIR;
6402 dev->od_dirent_journal = 1;
6403 if (hlock != NULL) {
6404 ldiskfs_htree_unlock(hlock);
6407 up_read(&obj->oo_ext_idx_sem);
6413 *fid = lma->lma_self_fid;
6415 /* Update or append the FID-in-dirent. */
6416 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6417 bh, de, hlock, dotdot);
6419 *attr |= LUDA_REPAIR;
6421 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
6422 "the dirent, dir = %lu/%u, name = %.*s, "
6423 "ino = %llu, "DFID": rc = %d\n",
6424 devname, dir->i_ino, dir->i_generation,
6425 ent->oied_namelen, ent->oied_name,
6426 ent->oied_ino, PFID(fid), rc);
6428 /* lma is NULL, trust the FID-in-dirent if it is valid. */
6429 if (*attr & LUDA_VERIFY_DRYRUN) {
6430 if (fid_is_sane(fid)) {
6431 *attr |= LUDA_REPAIR;
6432 } else if (dev->od_index == 0) {
6433 lu_igif_build(fid, inode->i_ino,
6434 inode->i_generation);
6435 *attr |= LUDA_UPGRADE;
6443 dev->od_dirent_journal = 1;
6444 if (hlock != NULL) {
6445 ldiskfs_htree_unlock(hlock);
6448 up_read(&obj->oo_ext_idx_sem);
6455 if (unlikely(fid_is_sane(fid))) {
6456 /* FID-in-dirent exists, but FID-in-LMA is lost.
6457 * Trust the FID-in-dirent, and add FID-in-LMA. */
6458 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
6460 *attr |= LUDA_REPAIR;
6462 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
6463 "update dirent, dir = %lu/%u, "
6464 "name = %.*s, ino = %llu, "
6466 devname, dir->i_ino, dir->i_generation,
6467 ent->oied_namelen, ent->oied_name,
6468 ent->oied_ino, PFID(fid), rc);
6469 } else if (dev->od_index == 0) {
6470 lu_igif_build(fid, inode->i_ino, inode->i_generation);
6471 /* It is probably IGIF object. Only aappend the
6472 * FID-in-dirent. OI scrub will process FID-in-LMA. */
6473 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
6474 bh, de, hlock, dotdot);
6476 *attr |= LUDA_UPGRADE;
6478 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
6479 "after the dirent, dir = %lu/%u, "
6480 "name = %.*s, ino = %llu, "
6482 devname, dir->i_ino, dir->i_generation,
6483 ent->oied_namelen, ent->oied_name,
6484 ent->oied_ino, PFID(fid), rc);
6493 if (hlock != NULL) {
6494 ldiskfs_htree_unlock(hlock);
6496 if (dev->od_dirent_journal != 0)
6497 up_write(&obj->oo_ext_idx_sem);
6499 up_read(&obj->oo_ext_idx_sem);
6503 ldiskfs_journal_stop(jh);
6507 if (rc >= 0 && !dirty)
6508 dev->od_dirent_journal = 0;
6514 * Returns the value at current position from iterator's in memory structure.
6516 * \param di struct osd_it_ea, iterator's in memory structure
6517 * \param attr attr requested for dirent.
6518 * \param lde lustre dirent
6520 * \retval 0 no error and \param lde has correct lustre dirent.
6521 * \retval -ve on error
6523 static inline int osd_it_ea_rec(const struct lu_env *env,
6524 const struct dt_it *di,
6525 struct dt_rec *dtrec, __u32 attr)
6527 struct osd_it_ea *it = (struct osd_it_ea *)di;
6528 struct osd_object *obj = it->oie_obj;
6529 struct osd_device *dev = osd_obj2dev(obj);
6530 struct osd_thread_info *oti = osd_oti_get(env);
6531 struct osd_inode_id *id = &oti->oti_id;
6532 struct lu_fid *fid = &it->oie_dirent->oied_fid;
6533 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6534 __u32 ino = it->oie_dirent->oied_ino;
6538 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
6540 if (attr & LUDA_VERIFY) {
6541 if (unlikely(is_remote_parent_ino(dev, ino))) {
6542 attr |= LUDA_IGNORE;
6543 /* If the parent is on remote MDT, and there
6544 * is no FID-in-dirent, then we have to get
6545 * the parent FID from the linkEA. */
6546 if (!fid_is_sane(fid) &&
6547 it->oie_dirent->oied_namelen == 2 &&
6548 it->oie_dirent->oied_name[0] == '.' &&
6549 it->oie_dirent->oied_name[1] == '.')
6550 osd_get_pfid_from_linkea(env, obj, fid);
6552 rc = osd_dirent_check_repair(env, obj, it, fid, id,
6556 if (!fid_is_sane(fid))
6557 attr |= LUDA_UNKNOWN;
6559 attr &= ~LU_DIRENT_ATTRS_MASK;
6560 if (!fid_is_sane(fid)) {
6561 bool is_dotdot = false;
6562 if (it->oie_dirent->oied_namelen == 2 &&
6563 it->oie_dirent->oied_name[0] == '.' &&
6564 it->oie_dirent->oied_name[1] == '.')
6566 /* If the parent is on remote MDT, and there
6567 * is no FID-in-dirent, then we have to get
6568 * the parent FID from the linkEA. */
6569 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
6570 rc = osd_get_pfid_from_linkea(env, obj, fid);
6572 if (is_dotdot == false &&
6573 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
6576 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6579 osd_id_gen(id, ino, OSD_OII_NOGEN);
6583 /* Pack the entry anyway, at least the offset is right. */
6584 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
6585 it->oie_dirent->oied_name,
6586 it->oie_dirent->oied_namelen,
6587 it->oie_dirent->oied_type, attr);
6592 if (osd_remote_fid(env, dev, fid))
6595 if (likely(!(attr & (LUDA_IGNORE | LUDA_UNKNOWN)) && rc == 0))
6596 osd_add_oi_cache(oti, dev, id, fid);
6598 RETURN(rc > 0 ? 0 : rc);
6602 * Returns the record size size at current position.
6604 * This function will return record(lu_dirent) size in bytes.
6606 * \param[in] env execution environment
6607 * \param[in] di iterator's in memory structure
6608 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
6609 * calculate the lu_dirent size.
6611 * \retval record size(in bytes & in memory) of the current lu_dirent
6614 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
6617 struct osd_it_ea *it = (struct osd_it_ea *)di;
6619 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
6623 * Returns a cookie for current position of the iterator head, so that
6624 * user can use this cookie to load/start the iterator next time.
6626 * \param di iterator's in memory structure
6628 * \retval cookie for current position, on success
6630 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
6632 struct osd_it_ea *it = (struct osd_it_ea *)di;
6634 return it->oie_dirent->oied_off;
6638 * It calls osd_ldiskfs_it_fill() which will use ->readdir()
6639 * to load a directory entry at a time and stored it i inn,
6640 * in iterator's in-memory data structure.
6642 * \param di struct osd_it_ea, iterator's in memory structure
6644 * \retval +ve on success
6645 * \retval -ve on error
6647 static int osd_it_ea_load(const struct lu_env *env,
6648 const struct dt_it *di, __u64 hash)
6650 struct osd_it_ea *it = (struct osd_it_ea *)di;
6654 it->oie_file.f_pos = hash;
6656 rc = osd_ldiskfs_it_fill(env, di);
6667 * Index lookup function for interoperability mode (b11826).
6669 * \param key, key i.e. file name to be searched
6671 * \retval +ve, on success
6672 * \retval -ve, on error
6674 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
6675 struct dt_rec *rec, const struct dt_key *key)
6677 struct osd_object *obj = osd_dt_obj(dt);
6682 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
6683 LINVRNT(osd_invariant(obj));
6685 rc = osd_ea_lookup_rec(env, obj, rec, key);
6692 * Index and Iterator operations for interoperability
6693 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
6695 static const struct dt_index_operations osd_index_ea_ops = {
6696 .dio_lookup = osd_index_ea_lookup,
6697 .dio_declare_insert = osd_index_declare_ea_insert,
6698 .dio_insert = osd_index_ea_insert,
6699 .dio_declare_delete = osd_index_declare_ea_delete,
6700 .dio_delete = osd_index_ea_delete,
6702 .init = osd_it_ea_init,
6703 .fini = osd_it_ea_fini,
6704 .get = osd_it_ea_get,
6705 .put = osd_it_ea_put,
6706 .next = osd_it_ea_next,
6707 .key = osd_it_ea_key,
6708 .key_size = osd_it_ea_key_size,
6709 .rec = osd_it_ea_rec,
6710 .rec_size = osd_it_ea_rec_size,
6711 .store = osd_it_ea_store,
6712 .load = osd_it_ea_load
6716 static void *osd_key_init(const struct lu_context *ctx,
6717 struct lu_context_key *key)
6719 struct osd_thread_info *info;
6721 OBD_ALLOC_PTR(info);
6723 return ERR_PTR(-ENOMEM);
6725 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6726 if (info->oti_it_ea_buf == NULL)
6729 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
6731 info->oti_hlock = ldiskfs_htree_lock_alloc();
6732 if (info->oti_hlock == NULL)
6738 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6741 return ERR_PTR(-ENOMEM);
6744 static void osd_key_fini(const struct lu_context *ctx,
6745 struct lu_context_key *key, void* data)
6747 struct osd_thread_info *info = data;
6748 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
6749 struct osd_idmap_cache *idc = info->oti_ins_cache;
6751 if (info->oti_inode != NULL)
6753 if (info->oti_hlock != NULL)
6754 ldiskfs_htree_lock_free(info->oti_hlock);
6755 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
6756 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
6757 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
6758 lu_buf_free(&info->oti_big_buf);
6760 LASSERT(info->oti_ins_cache_size > 0);
6761 OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size);
6762 info->oti_ins_cache = NULL;
6763 info->oti_ins_cache_size = 0;
6768 static void osd_key_exit(const struct lu_context *ctx,
6769 struct lu_context_key *key, void *data)
6771 struct osd_thread_info *info = data;
6773 LASSERT(info->oti_r_locks == 0);
6774 LASSERT(info->oti_w_locks == 0);
6775 LASSERT(info->oti_txns == 0);
6778 /* type constructor/destructor: osd_type_init, osd_type_fini */
6779 LU_TYPE_INIT_FINI(osd, &osd_key);
6781 struct lu_context_key osd_key = {
6782 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
6783 .lct_init = osd_key_init,
6784 .lct_fini = osd_key_fini,
6785 .lct_exit = osd_key_exit
6789 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
6790 const char *name, struct lu_device *next)
6792 struct osd_device *osd = osd_dev(d);
6794 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname))
6795 >= sizeof(osd->od_svname))
6797 return osd_procfs_init(osd, name);
6800 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
6802 struct seq_server_site *ss = osd_seq_site(osd);
6806 if (osd->od_is_ost || osd->od_cl_seq != NULL)
6809 if (unlikely(ss == NULL))
6812 OBD_ALLOC_PTR(osd->od_cl_seq);
6813 if (osd->od_cl_seq == NULL)
6816 rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
6817 osd->od_svname, ss->ss_server_seq);
6819 OBD_FREE_PTR(osd->od_cl_seq);
6820 osd->od_cl_seq = NULL;
6824 if (ss->ss_node_id == 0) {
6825 /* If the OSD on the sequence controller(MDT0), then allocate
6826 * sequence here, otherwise allocate sequence after connected
6827 * to MDT0 (see mdt_register_lwp_callback()). */
6828 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
6829 &osd->od_cl_seq->lcs_space, env);
6835 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
6837 if (osd->od_cl_seq == NULL)
6840 seq_client_fini(osd->od_cl_seq);
6841 OBD_FREE_PTR(osd->od_cl_seq);
6842 osd->od_cl_seq = NULL;
6845 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
6849 /* shutdown quota slave instance associated with the device */
6850 if (o->od_quota_slave != NULL) {
6851 qsd_fini(env, o->od_quota_slave);
6852 o->od_quota_slave = NULL;
6855 osd_fid_fini(env, o);
6860 static void osd_umount(const struct lu_env *env, struct osd_device *o)
6864 if (o->od_mnt != NULL) {
6865 shrink_dcache_sb(osd_sb(o));
6866 osd_sync(env, &o->od_dt_dev);
6875 static int osd_mount(const struct lu_env *env,
6876 struct osd_device *o, struct lustre_cfg *cfg)
6878 const char *name = lustre_cfg_string(cfg, 0);
6879 const char *dev = lustre_cfg_string(cfg, 1);
6881 unsigned long page, s_flags, lmd_flags = 0;
6882 struct page *__page;
6883 struct file_system_type *type;
6884 char *options = NULL;
6886 struct osd_thread_info *info = osd_oti_get(env);
6887 struct lu_fid *fid = &info->oti_fid;
6888 struct inode *inode;
6889 int rc = 0, force_over_512tb = 0;
6892 if (o->od_mnt != NULL)
6895 if (strlen(dev) >= sizeof(o->od_mntdev))
6897 strcpy(o->od_mntdev, dev);
6899 str = lustre_cfg_string(cfg, 2);
6900 s_flags = simple_strtoul(str, NULL, 0);
6901 str = strstr(str, ":");
6903 lmd_flags = simple_strtoul(str + 1, NULL, 0);
6904 opts = lustre_cfg_string(cfg, 3);
6906 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
6907 CERROR("%s: device %s extents feature is not guaranteed to "
6908 "work on big-endian systems. Use \"bigendian_extents\" "
6909 "mount option to override.\n", name, dev);
6913 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
6914 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
6915 CWARN("force_over_128tb option is deprecated. "
6916 "Filesystems less than 512TB can be created without any "
6917 "force options. Use force_over_512tb option for "
6918 "filesystems greater than 512TB.\n");
6921 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
6922 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
6923 CWARN("force_over_256tb option is deprecated. "
6924 "Filesystems less than 512TB can be created without any "
6925 "force options. Use force_over_512tb option for "
6926 "filesystems greater than 512TB.\n");
6930 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL)
6931 force_over_512tb = 1;
6933 __page = alloc_page(GFP_KERNEL);
6935 GOTO(out, rc = -ENOMEM);
6936 page = (unsigned long)page_address(__page);
6937 options = (char *)page;
6940 /* strip out the options for back compatiblity */
6941 static char *sout[] = {
6948 /* strip out option we processed in osd */
6949 "bigendian_extents",
6955 strcat(options, opts);
6956 for (rc = 0, str = options; sout[rc]; ) {
6957 char *op = strstr(str, sout[rc]);
6963 if (op == options || *(op - 1) == ',') {
6964 str = op + strlen(sout[rc]);
6965 if (*str == ',' || *str == '\0') {
6966 *str == ',' ? str++ : str;
6967 memmove(op, str, strlen(str) + 1);
6970 for (str = op; *str != ',' && *str != '\0'; str++)
6974 strncat(options, "user_xattr,acl", 14);
6977 /* Glom up mount options */
6978 if (*options != '\0')
6979 strcat(options, ",");
6980 strlcat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
6982 type = get_fs_type("ldiskfs");
6984 CERROR("%s: cannot find ldiskfs module\n", name);
6985 GOTO(out, rc = -ENODEV);
6988 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
6989 module_put(type->owner);
6991 if (IS_ERR(o->od_mnt)) {
6992 rc = PTR_ERR(o->od_mnt);
6994 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
6998 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
6999 osd_sb(o)->s_blocksize_bits > 512ULL << 40 &&
7000 force_over_512tb == 0) {
7001 CERROR("%s: device %s LDISKFS does not support filesystems "
7002 "greater than 512TB and can cause data corruption. "
7003 "Use \"force_over_512tb\" mount option to override.\n",
7005 GOTO(out_mnt, rc = -EINVAL);
7008 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7009 if (priv_dev_set_rdonly) {
7010 priv_dev_set_rdonly(osd_sb(o)->s_bdev);
7011 o->od_dt_dev.dd_rdonly = 1;
7012 LCONSOLE_WARN("%s: set dev_rdonly on this device\n",
7015 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7018 GOTO(out_mnt, rc = -EOPNOTSUPP);
7020 } else if (priv_dev_check_rdonly &&
7021 priv_dev_check_rdonly(osd_sb(o)->s_bdev)) {
7022 CERROR("%s: underlying device %s is marked as "
7023 "read-only. Setup failed\n", name, dev);
7025 GOTO(out_mnt, rc = -EROFS);
7028 if (!LDISKFS_HAS_COMPAT_FEATURE(o->od_mnt->mnt_sb,
7029 LDISKFS_FEATURE_COMPAT_HAS_JOURNAL)) {
7030 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7031 GOTO(out_mnt, rc = -EINVAL);
7034 #ifdef LDISKFS_MOUNT_DIRDATA
7035 if (LDISKFS_HAS_INCOMPAT_FEATURE(o->od_mnt->mnt_sb,
7036 LDISKFS_FEATURE_INCOMPAT_DIRDATA))
7037 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7038 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7039 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7040 "enabling the dirdata feature. If you do not want to "
7041 "downgrade to Lustre-1.x again, you can enable it via "
7042 "'tune2fs -O dirdata device'\n", name, dev);
7044 inode = osd_sb(o)->s_root->d_inode;
7045 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7046 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7048 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7052 if (lmd_flags & LMD_FLG_NOSCRUB)
7063 __free_page(__page);
7068 static struct lu_device *osd_device_fini(const struct lu_env *env,
7069 struct lu_device *d)
7071 struct osd_device *o = osd_dev(d);
7074 osd_shutdown(env, o);
7076 osd_scrub_cleanup(env, o);
7077 osd_obj_map_fini(o);
7083 static int osd_device_init0(const struct lu_env *env,
7084 struct osd_device *o,
7085 struct lustre_cfg *cfg)
7087 struct lu_device *l = osd2lu_dev(o);
7088 struct osd_thread_info *info;
7092 /* if the module was re-loaded, env can loose its keys */
7093 rc = lu_env_refill((struct lu_env *) env);
7096 info = osd_oti_get(env);
7099 l->ld_ops = &osd_lu_ops;
7100 o->od_dt_dev.dd_ops = &osd_dt_ops;
7102 spin_lock_init(&o->od_osfs_lock);
7103 mutex_init(&o->od_otable_mutex);
7104 INIT_LIST_HEAD(&o->od_orphan_list);
7106 o->od_read_cache = 1;
7107 o->od_writethrough_cache = 1;
7108 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7110 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7111 sizeof(o->od_svname));
7112 if (cplen >= sizeof(o->od_svname)) {
7117 o->od_index = -1; /* -1 means index is invalid */
7118 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7119 if (rc == LDD_F_SV_TYPE_OST)
7122 o->od_full_scrub_ratio = OFSR_DEFAULT;
7123 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7124 rc = osd_mount(env, o, cfg);
7128 rc = osd_obj_map_init(env, o);
7132 rc = lu_site_init(&o->od_site, l);
7134 GOTO(out_compat, rc);
7135 o->od_site.ls_bottom_dev = l;
7137 rc = lu_site_init_finish(&o->od_site);
7141 INIT_LIST_HEAD(&o->od_ios_list);
7142 /* setup scrub, including OI files initialization */
7143 rc = osd_scrub_setup(env, o);
7147 rc = osd_procfs_init(o, o->od_svname);
7149 CERROR("%s: can't initialize procfs: rc = %d\n",
7151 GOTO(out_scrub, rc);
7154 LASSERT(l->ld_site->ls_linkage.next != NULL);
7155 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7157 /* initialize quota slave instance */
7158 o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
7160 if (IS_ERR(o->od_quota_slave)) {
7161 rc = PTR_ERR(o->od_quota_slave);
7162 o->od_quota_slave = NULL;
7163 GOTO(out_procfs, rc);
7171 osd_scrub_cleanup(env, o);
7173 lu_site_fini(&o->od_site);
7175 osd_obj_map_fini(o);
7182 static struct lu_device *osd_device_alloc(const struct lu_env *env,
7183 struct lu_device_type *t,
7184 struct lustre_cfg *cfg)
7186 struct osd_device *o;
7191 return ERR_PTR(-ENOMEM);
7193 rc = dt_device_init(&o->od_dt_dev, t);
7195 /* Because the ctx might be revived in dt_device_init,
7196 * refill the env here */
7197 lu_env_refill((struct lu_env *)env);
7198 rc = osd_device_init0(env, o, cfg);
7200 dt_device_fini(&o->od_dt_dev);
7203 if (unlikely(rc != 0))
7206 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
7209 static struct lu_device *osd_device_free(const struct lu_env *env,
7210 struct lu_device *d)
7212 struct osd_device *o = osd_dev(d);
7215 /* XXX: make osd top device in order to release reference */
7216 d->ld_site->ls_top_dev = d;
7217 lu_site_purge(env, d->ld_site, -1);
7218 if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
7219 LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
7220 lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
7222 lu_site_fini(&o->od_site);
7223 dt_device_fini(&o->od_dt_dev);
7228 static int osd_process_config(const struct lu_env *env,
7229 struct lu_device *d, struct lustre_cfg *cfg)
7231 struct osd_device *o = osd_dev(d);
7235 switch (cfg->lcfg_command) {
7237 rc = osd_mount(env, o, cfg);
7240 lu_dev_del_linkage(d->ld_site, d);
7241 rc = osd_shutdown(env, o);
7244 LASSERT(&o->od_dt_dev);
7245 rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
7246 cfg, &o->od_dt_dev);
7247 if (rc > 0 || rc == -ENOSYS) {
7248 rc = class_process_proc_param(PARAM_OST,
7249 lprocfs_osd_obd_vars,
7250 cfg, &o->od_dt_dev);
7262 static int osd_recovery_complete(const struct lu_env *env,
7263 struct lu_device *d)
7265 struct osd_device *osd = osd_dev(d);
7269 if (osd->od_quota_slave == NULL)
7272 /* start qsd instance on recovery completion, this notifies the quota
7273 * slave code that we are about to process new requests now */
7274 rc = qsd_start(env, osd->od_quota_slave);
7279 * we use exports to track all osd users
7281 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
7282 struct obd_device *obd, struct obd_uuid *cluuid,
7283 struct obd_connect_data *data, void *localdata)
7285 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7286 struct lustre_handle conn;
7290 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
7292 rc = class_connect(&conn, obd, cluuid);
7296 *exp = class_conn2export(&conn);
7298 spin_lock(&osd->od_osfs_lock);
7300 spin_unlock(&osd->od_osfs_lock);
7306 * once last export (we don't count self-export) disappeared
7307 * osd can be released
7309 static int osd_obd_disconnect(struct obd_export *exp)
7311 struct obd_device *obd = exp->exp_obd;
7312 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7313 int rc, release = 0;
7316 /* Only disconnect the underlying layers on the final disconnect. */
7317 spin_lock(&osd->od_osfs_lock);
7319 if (osd->od_connects == 0)
7321 spin_unlock(&osd->od_osfs_lock);
7323 rc = class_disconnect(exp); /* bz 9811 */
7325 if (rc == 0 && release)
7326 class_manual_cleanup(obd);
7330 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
7331 struct lu_device *dev)
7333 struct osd_device *osd = osd_dev(dev);
7334 struct lr_server_data *lsd =
7335 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
7339 if (osd->od_quota_slave != NULL) {
7340 /* set up quota slave objects */
7341 result = qsd_prepare(env, osd->od_quota_slave);
7346 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
7347 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
7348 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
7349 osd->od_index_in_idif = 1;
7351 osd->od_index_in_idif = 0;
7352 result = osd_register_proc_index_in_idif(osd);
7357 osd->od_index_in_idif = 1;
7361 result = osd_fid_init(env, osd);
7366 static int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
7367 struct lu_fid *fid, struct md_op_data *op_data)
7369 struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);
7371 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
7374 static const struct lu_object_operations osd_lu_obj_ops = {
7375 .loo_object_init = osd_object_init,
7376 .loo_object_delete = osd_object_delete,
7377 .loo_object_release = osd_object_release,
7378 .loo_object_free = osd_object_free,
7379 .loo_object_print = osd_object_print,
7380 .loo_object_invariant = osd_object_invariant
7383 const struct lu_device_operations osd_lu_ops = {
7384 .ldo_object_alloc = osd_object_alloc,
7385 .ldo_process_config = osd_process_config,
7386 .ldo_recovery_complete = osd_recovery_complete,
7387 .ldo_prepare = osd_prepare,
7390 static const struct lu_device_type_operations osd_device_type_ops = {
7391 .ldto_init = osd_type_init,
7392 .ldto_fini = osd_type_fini,
7394 .ldto_start = osd_type_start,
7395 .ldto_stop = osd_type_stop,
7397 .ldto_device_alloc = osd_device_alloc,
7398 .ldto_device_free = osd_device_free,
7400 .ldto_device_init = osd_device_init,
7401 .ldto_device_fini = osd_device_fini
7404 static struct lu_device_type osd_device_type = {
7405 .ldt_tags = LU_DEVICE_DT,
7406 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
7407 .ldt_ops = &osd_device_type_ops,
7408 .ldt_ctx_tags = LCT_LOCAL,
7411 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
7413 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
7414 struct super_block *sb = osd_sb(osd);
7416 return (osd->od_mnt == NULL || sb->s_flags & MS_RDONLY);
7420 * lprocfs legacy support.
7422 static struct obd_ops osd_obd_device_ops = {
7423 .o_owner = THIS_MODULE,
7424 .o_connect = osd_obd_connect,
7425 .o_disconnect = osd_obd_disconnect,
7426 .o_fid_alloc = osd_fid_alloc,
7427 .o_health_check = osd_health_check,
7430 static int __init osd_init(void)
7434 LASSERT(BH_DXLock < sizeof(((struct buffer_head *)0)->b_state) * 8);
7435 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
7436 /* please, try to keep osd_thread_info smaller than a page */
7437 CLASSERT(sizeof(struct osd_thread_info) <= PAGE_SIZE);
7442 rc = lu_kmem_init(ldiskfs_caches);
7446 #ifdef CONFIG_KALLSYMS
7447 priv_dev_set_rdonly = (void *)kallsyms_lookup_name("dev_set_rdonly");
7448 priv_dev_check_rdonly = (void *)kallsyms_lookup_name("dev_check_rdonly");
7449 /* Clear readonly is unused at this time */
7450 /*priv_dev_clear_rdonly = (void *)kallsyms_lookup_name("dev_clear_rdonly");*/
7453 rc = class_register_type(&osd_obd_device_ops, NULL, true,
7454 lprocfs_osd_module_vars,
7455 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
7457 lu_kmem_fini(ldiskfs_caches);
7461 static void __exit osd_exit(void)
7463 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
7464 lu_kmem_fini(ldiskfs_caches);
7467 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
7468 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
7469 MODULE_VERSION(LUSTRE_VERSION_STRING);
7470 MODULE_LICENSE("GPL");
7472 module_init(osd_init);
7473 module_exit(osd_exit);