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, 2017, 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/fs_struct.h>
43 #include <linux/kallsyms.h>
44 #include <linux/module.h>
45 #include <linux/user_namespace.h>
46 #include <linux/uidgid.h>
48 /* prerequisite for linux/xattr.h */
49 #include <linux/types.h>
50 /* prerequisite for linux/xattr.h */
52 /* XATTR_{REPLACE,CREATE} */
53 #include <linux/xattr.h>
55 #include <ldiskfs/ldiskfs.h>
56 #include <ldiskfs/xattr.h>
57 #include <ldiskfs/ldiskfs_extents.h>
60 * struct OBD_{ALLOC,FREE}*()
63 #include <obd_support.h>
64 /* struct ptlrpc_thread */
65 #include <lustre_net.h>
66 #include <lustre_fid.h>
68 #include <uapi/linux/lustre/lustre_param.h>
70 #include "osd_internal.h"
71 #include "osd_dynlocks.h"
73 /* llo_* api support */
74 #include <md_object.h>
75 #include <lustre_quota.h>
77 #include <lustre_linkea.h>
79 /* Maximum EA size is limited by LNET_MTU for remote objects */
80 #define OSD_MAX_EA_SIZE 1048364
83 module_param(ldiskfs_pdo, int, 0644);
84 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
86 int ldiskfs_track_declares_assert;
87 module_param(ldiskfs_track_declares_assert, int, 0644);
88 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
90 /* Slab to allocate dynlocks */
91 struct kmem_cache *dynlock_cachep;
93 /* Slab to allocate osd_it_ea */
94 struct kmem_cache *osd_itea_cachep;
96 static struct lu_kmem_descr ldiskfs_caches[] = {
98 .ckd_cache = &dynlock_cachep,
99 .ckd_name = "dynlock_cache",
100 .ckd_size = sizeof(struct dynlock_handle)
103 .ckd_cache = &osd_itea_cachep,
104 .ckd_name = "osd_itea_cache",
105 .ckd_size = sizeof(struct osd_it_ea)
112 static const char dot[] = ".";
113 static const char dotdot[] = "..";
115 static const struct lu_object_operations osd_lu_obj_ops;
116 static const struct dt_object_operations osd_obj_ops;
117 static const struct dt_object_operations osd_obj_otable_it_ops;
118 static const struct dt_index_operations osd_index_iam_ops;
119 static const struct dt_index_operations osd_index_ea_ops;
121 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
122 const struct lu_fid *fid);
123 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
124 struct osd_device *osd);
126 int osd_trans_declare_op2rb[] = {
127 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
128 [OSD_OT_PUNCH] = OSD_OT_MAX,
129 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
130 [OSD_OT_CREATE] = OSD_OT_DESTROY,
131 [OSD_OT_DESTROY] = OSD_OT_CREATE,
132 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
133 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
134 [OSD_OT_WRITE] = OSD_OT_WRITE,
135 [OSD_OT_INSERT] = OSD_OT_DELETE,
136 [OSD_OT_DELETE] = OSD_OT_INSERT,
137 [OSD_OT_QUOTA] = OSD_OT_MAX,
140 static int osd_has_index(const struct osd_object *obj)
142 return obj->oo_dt.do_index_ops != NULL;
145 static int osd_object_invariant(const struct lu_object *l)
147 return osd_invariant(osd_obj(l));
151 * Concurrency: doesn't matter
153 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
155 struct osd_thread_info *oti = osd_oti_get(env);
157 return oti->oti_w_locks > 0 && o->oo_owner == env;
161 * Concurrency: doesn't access mutable data
163 static int osd_root_get(const struct lu_env *env,
164 struct dt_device *dev, struct lu_fid *f)
166 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
171 * the following set of functions are used to maintain per-thread
172 * cache of FID->ino mapping. this mechanism is needed to resolve
173 * FID to inode at dt_insert() which in turn stores ino in the
174 * directory entries to keep ldiskfs compatible with ext[34].
175 * due to locking-originated restrictions we can't lookup ino
176 * using LU cache (deadlock is possible). lookup using OI is quite
177 * expensive. so instead we maintain this cache and methods like
178 * dt_create() fill it. so in the majority of cases dt_insert() is
179 * able to find needed mapping in lockless manner.
181 static struct osd_idmap_cache *
182 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
183 const struct lu_fid *fid)
185 struct osd_thread_info *oti = osd_oti_get(env);
186 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_PTR_ARRAY(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_PTR_ARRAY(oti->oti_ins_cache,
220 oti->oti_ins_cache_used);
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 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
256 osd->od_svname, PFID(fid));
258 /* new mapping is needed */
259 idc = osd_idc_add(env, osd, fid);
261 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
262 osd->od_svname, PFID(fid), PTR_ERR(idc));
267 rc = osd_remote_fid(env, osd, fid);
268 if (unlikely(rc < 0))
272 /* the object is local, lookup in OI */
273 /* XXX: probably cheaper to lookup in LU first? */
274 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
276 if (unlikely(rc < 0)) {
277 CERROR("can't lookup: rc = %d\n", rc);
281 /* the object is remote */
288 static void osd_idc_dump_lma(const struct lu_env *env,
289 struct osd_device *osd,
293 struct osd_thread_info *info = osd_oti_get(env);
294 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
295 const struct lu_fid *fid;
296 struct osd_inode_id lid;
300 inode = osd_ldiskfs_iget(osd_sb(osd), ino);
302 CERROR("%s: can't get inode %lu: rc = %d\n",
303 osd->od_svname, ino, (int)PTR_ERR(inode));
306 if (is_bad_inode(inode)) {
307 CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
310 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
312 CERROR("%s: can't get LMA for %lu: rc = %d\n",
313 osd->od_svname, ino, rc);
316 fid = &loa->loa_lma.lma_self_fid;
317 LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
318 PFID(fid), ino, (unsigned)inode->i_generation);
321 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
323 CERROR("%s: can't lookup "DFID": rc = %d\n",
324 osd->od_svname, PFID(fid), rc);
327 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
328 PFID(fid), lid.oii_ino, lid.oii_gen);
333 static void osd_idc_dump_debug(const struct lu_env *env,
334 struct osd_device *osd,
335 const struct lu_fid *fid,
339 struct osd_inode_id lid;
343 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
345 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
346 osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
347 osd_idc_dump_lma(env, osd, lid.oii_ino, false);
349 CERROR("%s: can't lookup "DFID": rc = %d\n",
350 osd->od_svname, PFID(fid), rc);
353 osd_idc_dump_lma(env, osd, ino1, true);
355 osd_idc_dump_lma(env, osd, ino2, true);
359 * lookup mapping for given FID and fill it from the given object.
360 * the object is lolcal by definition.
362 static int osd_idc_find_and_init(const struct lu_env *env,
363 struct osd_device *osd,
364 struct osd_object *obj)
366 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
367 struct osd_idmap_cache *idc;
369 idc = osd_idc_find(env, osd, fid);
370 LASSERT(!IS_ERR(idc));
372 if (obj->oo_inode == NULL)
374 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
375 if (idc->oic_lid.oii_ino) {
376 osd_idc_dump_debug(env, osd, fid,
377 idc->oic_lid.oii_ino,
378 obj->oo_inode->i_ino);
381 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
382 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
387 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
388 osd->od_svname, PFID(fid));
390 /* new mapping is needed */
391 idc = osd_idc_add(env, osd, fid);
393 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
394 osd->od_svname, PFID(fid), PTR_ERR(idc));
398 if (obj->oo_inode != NULL) {
399 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
400 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
406 * OSD object methods.
410 * Concurrency: no concurrent access is possible that early in object
413 static struct lu_object *osd_object_alloc(const struct lu_env *env,
414 const struct lu_object_header *hdr,
417 struct osd_object *mo;
422 struct lu_object_header *h;
423 struct osd_device *o = osd_dev(d);
425 l = &mo->oo_dt.do_lu;
426 if (unlikely(o->od_in_init)) {
433 lu_object_header_init(h);
434 lu_object_init(l, h, d);
435 lu_object_add_top(h, l);
438 dt_object_init(&mo->oo_dt, NULL, d);
439 mo->oo_header = NULL;
442 mo->oo_dt.do_ops = &osd_obj_ops;
443 l->lo_ops = &osd_lu_obj_ops;
444 init_rwsem(&mo->oo_sem);
445 init_rwsem(&mo->oo_ext_idx_sem);
446 spin_lock_init(&mo->oo_guard);
447 INIT_LIST_HEAD(&mo->oo_xattr_list);
453 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
454 struct dentry *dentry, struct lustre_ost_attrs *loa)
458 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
459 (void *)loa, sizeof(*loa));
461 struct lustre_mdt_attrs *lma = &loa->loa_lma;
463 if (rc < sizeof(*lma))
467 lustre_loa_swab(loa, true);
468 /* Check LMA compatibility */
469 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
471 CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
473 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
474 PFID(&lma->lma_self_fid), inode->i_ino, rc);
476 } else if (rc == 0) {
484 * retrieve object from backend ext fs.
486 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
487 struct osd_inode_id *id)
490 struct inode *inode = NULL;
493 * if we look for an inode withing a running
494 * transaction, then we risk to deadlock
495 * osd_dirent_check_repair() breaks this
497 /* LASSERT(current->journal_info == NULL); */
499 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
501 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
502 id->oii_ino, PTR_ERR(inode));
503 } else if (id->oii_gen != OSD_OII_NOGEN &&
504 inode->i_generation != id->oii_gen) {
505 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
506 "i_generation = %u\n",
507 id->oii_ino, id->oii_gen, inode->i_generation);
509 inode = ERR_PTR(-ESTALE);
510 } else if (inode->i_nlink == 0) {
512 * due to parallel readdir and unlink,
513 * we can have dead inode here.
515 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
517 inode = ERR_PTR(-ESTALE);
518 } else if (is_bad_inode(inode)) {
520 CWARN("%s: bad inode: ino = %u: rc = %d\n",
521 osd_dev2name(dev), id->oii_ino, rc);
524 } else if ((rc = osd_attach_jinode(inode))) {
528 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
529 if (id->oii_gen == OSD_OII_NOGEN)
530 osd_id_gen(id, inode->i_ino, inode->i_generation);
533 * Do not update file c/mtime in ldiskfs.
534 * NB: we don't have any lock to protect this because we don't
535 * have reference on osd_object now, but contention with
536 * another lookup + attr_set can't happen in the tiny window
537 * between if (...) and set S_NOCMTIME.
539 if (!(inode->i_flags & S_NOCMTIME))
540 inode->i_flags |= S_NOCMTIME;
545 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
546 handle_t *handle, struct dentry *child,
547 struct inode *inode, struct htree_lock *hlock)
551 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
552 if (rc == -ENOBUFS || rc == -ENOSPC) {
553 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
554 struct inode *parent = child->d_parent->d_inode;
555 struct lu_fid *fid = NULL;
556 char fidstr[FID_LEN + 1] = "unknown";
558 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
560 fid = &loa->loa_lma.lma_self_fid;
561 } else if (rc2 == -ENODATA) {
562 if (unlikely(parent == inode->i_sb->s_root->d_inode)) {
563 fid = &info->oti_fid3;
564 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
565 } else if (!osd->od_is_ost && osd->od_index == 0) {
566 fid = &info->oti_fid3;
567 lu_igif_build(fid, parent->i_ino,
568 parent->i_generation);
573 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
575 /* below message is checked in sanity.sh test_129 */
577 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
578 osd_name(osd), parent->i_ino, fidstr);
580 rc = 0; /* ignore such error now */
581 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
582 osd_name(osd), parent->i_ino, fidstr);
592 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
593 struct osd_inode_id *id, struct lu_fid *fid)
595 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
599 inode = osd_iget(info, dev, id);
603 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
605 *fid = loa->loa_lma.lma_self_fid;
606 } else if (rc == -ENODATA) {
607 if (unlikely(inode == osd_sb(dev)->s_root->d_inode))
608 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
610 lu_igif_build(fid, inode->i_ino, inode->i_generation);
618 static struct inode *osd_iget_check(struct osd_thread_info *info,
619 struct osd_device *dev,
620 const struct lu_fid *fid,
621 struct osd_inode_id *id,
630 * The cached OI mapping is trustable. If we cannot locate the inode
631 * via the cached OI mapping, then return the failure to the caller
632 * directly without further OI checking.
636 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
639 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
642 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
643 PFID(fid), id->oii_ino, rc);
647 if (is_bad_inode(inode)) {
652 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
653 PFID(fid), id->oii_ino);
657 if (id->oii_gen != OSD_OII_NOGEN &&
658 inode->i_generation != id->oii_gen) {
663 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
664 "oii_gen = %u, i_generation = %u\n", PFID(fid),
665 id->oii_ino, id->oii_gen, inode->i_generation);
669 if (inode->i_nlink == 0) {
674 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
675 PFID(fid), id->oii_ino);
679 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
683 __u32 saved_ino = id->oii_ino;
684 __u32 saved_gen = id->oii_gen;
687 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
689 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
691 * XXX: There are four possible cases:
693 * Backup/restore caused the OI invalid.
695 * Someone unlinked the object but NOT removed
696 * the OI mapping, such as mount target device
697 * as ldiskfs, and modify something directly.
699 * Someone just removed the object between the
700 * former oi_lookup and the iget. It is normal.
701 * 4. Other failure cases.
703 * Generally, when the device is mounted, it will
704 * auto check whether the system is restored from
705 * file-level backup or not. We trust such detect
706 * to distinguish the 1st case from the 2nd case:
707 * if the OI files are consistent but may contain
708 * stale OI mappings because of case 2, if iget()
709 * returns -ENOENT or -ESTALE, then it should be
714 * If the OI mapping was in OI file before the
715 * osd_iget_check(), but now, it is disappear,
716 * then it must be removed by race. That is a
722 * It is the OI scrub updated the OI mapping by race.
723 * The new OI mapping must be valid.
725 if (saved_ino != id->oii_ino ||
726 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
735 if (dev->od_scrub.os_scrub.os_file.sf_flags &
738 * It still can be the case 2, but we cannot
739 * distinguish it from the case 1. So return
740 * -EREMCHG to block current operation until
741 * OI scrub rebuilt the OI mappings.
750 if (inode->i_generation == id->oii_gen)
755 if (id->oii_gen == OSD_OII_NOGEN)
756 osd_id_gen(id, inode->i_ino, inode->i_generation);
759 * Do not update file c/mtime in ldiskfs.
760 * NB: we don't have any lock to protect this because we don't
761 * have reference on osd_object now, but contention with
762 * another lookup + attr_set can't happen in the tiny window
763 * between if (...) and set S_NOCMTIME.
765 if (!(inode->i_flags & S_NOCMTIME))
766 inode->i_flags |= S_NOCMTIME;
783 * \retval +v: new filter_fid does not contain self-fid
784 * \retval 0: filter_fid_18_23, contains self-fid
785 * \retval -v: other failure cases
787 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
788 struct dentry *dentry, struct lu_fid *fid)
790 struct filter_fid *ff = &info->oti_ff;
791 struct ost_id *ostid = &info->oti_ostid;
794 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
795 if (rc == sizeof(struct filter_fid_18_23)) {
796 struct filter_fid_18_23 *ff_old = (void *)ff;
798 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
799 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
801 * XXX: use 0 as the index for compatibility, the caller will
802 * handle index related issues when necessary.
805 ostid_to_fid(fid, ostid, 0);
806 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
808 } else if (rc >= 0) {
815 static int osd_lma_self_repair(struct osd_thread_info *info,
816 struct osd_device *osd, struct inode *inode,
817 const struct lu_fid *fid, __u32 compat)
822 LASSERT(current->journal_info == NULL);
824 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
825 osd_dto_credits_noquota[DTO_XATTR_SET]);
828 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
833 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
835 CWARN("%s: cannot self repair the LMA: rc = %d\n",
837 ldiskfs_journal_stop(jh);
841 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
843 struct osd_thread_info *info = osd_oti_get(env);
844 struct osd_device *osd = osd_obj2dev(obj);
845 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
846 struct lustre_mdt_attrs *lma = &loa->loa_lma;
847 struct inode *inode = obj->oo_inode;
848 struct dentry *dentry = &info->oti_obj_dentry;
849 struct lu_fid *fid = NULL;
850 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
855 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
856 (void *)loa, sizeof(*loa));
857 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
858 fid = &lma->lma_self_fid;
859 rc = osd_get_idif(info, inode, dentry, fid);
860 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
862 * For the given OST-object, if it has neither LMA nor
863 * FID in XATTR_NAME_FID, then the given FID (which is
864 * contained in the @obj, from client RPC for locating
865 * the OST-object) is trusted. We use it to generate
868 osd_lma_self_repair(info, osd, inode, rfid,
879 lustre_lma_swab(lma);
880 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
881 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
882 S_ISREG(inode->i_mode)))) {
883 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
884 "fid = "DFID", ino = %lu\n", osd_name(osd),
885 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
886 PFID(rfid), inode->i_ino);
889 fid = &lma->lma_self_fid;
890 if (lma->lma_compat & LMAC_STRIPE_INFO &&
892 obj->oo_pfid_in_lma = 1;
893 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
895 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
899 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
900 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
901 struct ost_id *oi = &info->oti_ostid;
902 struct lu_fid *fid1 = &info->oti_fid3;
903 __u32 idx = fid_idif_ost_idx(rfid);
906 * For old IDIF, the OST index is not part of the IDIF,
907 * Means that different OSTs may have the same IDIFs.
908 * Under such case, we need to make some compatible
909 * check to make sure to trigger OI scrub properly.
911 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
912 /* Given @rfid is new, LMA is old. */
913 fid_to_ostid(fid, oi);
914 ostid_to_fid(fid1, oi, idx);
915 if (lu_fid_eq(fid1, rfid)) {
916 if (osd->od_index_in_idif)
917 osd_lma_self_repair(info, osd,
931 struct osd_check_lmv_buf {
932 /* please keep it as first member */
933 struct dir_context ctx;
934 struct osd_thread_info *oclb_info;
935 struct osd_device *oclb_dev;
936 struct osd_idmap_cache *oclb_oic;
942 * It is called internally by ->iterate*() to filter out the
943 * local slave object's FID of the striped directory.
945 * \retval 1 found the local slave's FID
946 * \retval 0 continue to check next item
947 * \retval -ve for failure
949 #ifdef HAVE_FILLDIR_USE_CTX
950 static int osd_stripe_dir_filldir(struct dir_context *buf,
952 static int osd_stripe_dir_filldir(void *buf,
954 const char *name, int namelen,
955 loff_t offset, __u64 ino, unsigned int d_type)
957 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
958 struct osd_thread_info *oti = oclb->oclb_info;
959 struct lu_fid *fid = &oti->oti_fid3;
960 struct osd_inode_id *id = &oti->oti_id3;
961 struct osd_device *dev = oclb->oclb_dev;
962 struct osd_idmap_cache *oic = oclb->oclb_oic;
971 sscanf(name + 1, SFID, RFID(fid));
972 if (!fid_is_sane(fid))
975 if (osd_remote_fid(oti->oti_env, dev, fid))
978 osd_id_gen(id, ino, OSD_OII_NOGEN);
979 inode = osd_iget(oti, dev, id);
981 return PTR_ERR(inode);
984 osd_add_oi_cache(oti, dev, id, fid);
988 osd_oii_insert(dev, oic, true);
989 oclb->oclb_found = true;
995 * When lookup item under striped directory, we need to locate the master
996 * MDT-object of the striped directory firstly, then the client will send
997 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
998 * and the item's name. If the system is restored from MDT file level backup,
999 * then before the OI scrub completely built the OI files, the OI mappings of
1000 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
1001 * not a problem for the master MDT-object. Because when locate the master
1002 * MDT-object, we will do name based lookup (for the striped directory itself)
1003 * firstly, during such process we can setup the correct OI mapping for the
1004 * master MDT-object. But it will be trouble for the slave MDT-object. Because
1005 * the client will not trigger name based lookup on the MDT to locate the slave
1006 * MDT-object before locating item under the striped directory, then when
1007 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
1008 * is invalid and does not know what the right OI mapping is, then the MDT has
1009 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
1010 * the OI file, related OI mapping is unknown yet, please try again later. And
1011 * then client will re-try the RPC again and again until related OI mapping has
1012 * been updated. That is quite inefficient.
1014 * To resolve above trouble, we will handle it as the following two cases:
1016 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
1017 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
1018 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
1019 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
1020 * directly. Please check osd_fid_lookup().
1022 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
1023 * Under such case, during lookup the master MDT-object, we will lookup the
1024 * slave MDT-object via readdir against the master MDT-object, because the
1025 * slave MDT-objects information are stored as sub-directories with the name
1026 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
1027 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
1028 * the correct OI mapping for the slave MDT-object.
1030 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
1031 struct inode *inode, struct osd_idmap_cache *oic)
1033 struct lu_buf *buf = &oti->oti_big_buf;
1034 struct dentry *dentry = &oti->oti_obj_dentry;
1036 struct lmv_mds_md_v1 *lmv1;
1037 struct osd_check_lmv_buf oclb = {
1038 .ctx.actor = osd_stripe_dir_filldir,
1042 .oclb_found = false,
1049 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
1051 if (rc == -ERANGE) {
1052 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
1054 lu_buf_realloc(buf, rc);
1055 if (buf->lb_buf == NULL)
1056 GOTO(out, rc = -ENOMEM);
1062 if (unlikely(rc == 0 || rc == -ENODATA))
1068 if (unlikely(buf->lb_buf == NULL)) {
1069 lu_buf_realloc(buf, rc);
1070 if (buf->lb_buf == NULL)
1071 GOTO(out, rc = -ENOMEM);
1077 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1080 filp = osd_quasi_file(oti->oti_env, inode);
1081 rc = osd_security_file_alloc(filp);
1086 oclb.oclb_items = 0;
1087 rc = iterate_dir(filp, &oclb.ctx);
1088 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1089 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1090 inode->i_fop->release(inode, filp);
1095 "%s: cannot check LMV, ino = %lu/%u "DFID": rc = %d\n",
1096 osd_ino2name(inode), inode->i_ino, inode->i_generation,
1097 PFID(&oic->oic_fid), rc);
1104 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1105 const struct lu_fid *fid,
1106 const struct lu_object_conf *conf)
1108 struct osd_thread_info *info;
1109 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1110 struct osd_device *dev;
1111 struct osd_idmap_cache *oic;
1112 struct osd_inode_id *id;
1113 struct inode *inode = NULL;
1114 struct lustre_scrub *scrub;
1115 struct scrub_file *sf;
1116 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1121 bool remote = false;
1122 bool trusted = true;
1123 bool updated = false;
1124 bool checked = false;
1128 LINVRNT(osd_invariant(obj));
1129 LASSERT(obj->oo_inode == NULL);
1130 LASSERTF(fid_is_sane(fid) || fid_is_idif(fid), DFID"\n", PFID(fid));
1132 dev = osd_dev(ldev);
1133 scrub = &dev->od_scrub.os_scrub;
1134 sf = &scrub->os_file;
1135 info = osd_oti_get(env);
1137 oic = &info->oti_cache;
1139 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1143 * For the object is created as locking anchor, or for the object to
1144 * be created on disk. No need to osd_oi_lookup() at here because FID
1145 * shouldn't never be re-used, if it's really a duplicate FID from
1146 * unexpected reason, we should be able to detect it later by calling
1147 * do_create->osd_oi_insert().
1149 if (conf && conf->loc_flags & LOC_F_NEW)
1150 GOTO(out, result = 0);
1152 /* Search order: 1. per-thread cache. */
1153 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1159 if (!list_empty(&scrub->os_inconsistent_items)) {
1160 /* Search order: 2. OI scrub pending list. */
1161 result = osd_oii_lookup(dev, fid, id);
1167 * The OI mapping in the OI file can be updated by the OI scrub
1168 * when we locate the inode via FID. So it may be not trustable.
1172 /* Search order: 3. OI files. */
1173 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1174 if (result == -ENOENT) {
1175 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1176 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1177 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1179 GOTO(out, result = 0);
1184 /* -ESTALE is returned if inode of OST object doesn't exist */
1185 if (result == -ESTALE &&
1186 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1187 GOTO(out, result = 0);
1194 obj->oo_inode = NULL;
1195 /* for later passes through checks, not true on first pass */
1196 if (!IS_ERR_OR_NULL(inode))
1199 inode = osd_iget_check(info, dev, fid, id, trusted);
1200 if (!IS_ERR(inode)) {
1201 obj->oo_inode = inode;
1209 result = PTR_ERR(inode);
1210 if (result == -ENOENT || result == -ESTALE)
1211 GOTO(out, result = 0);
1213 if (result != -EREMCHG)
1218 * We still have chance to get the valid inode: for the
1219 * object which is referenced by remote name entry, the
1220 * object on the local MDT will be linked under the dir
1221 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1223 * We do not know whether the object for the given FID
1224 * is referenced by some remote name entry or not, and
1225 * especially for DNE II, a multiple-linked object may
1226 * have many name entries reside on many MDTs.
1228 * To simplify the operation, OSD will not distinguish
1229 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1230 * only happened for the RPC from other MDT during the
1231 * OI scrub, or for the client side RPC with FID only,
1232 * such as FID to path, or from old connected client.
1235 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1239 flags |= SS_AUTO_PARTIAL;
1240 flags &= ~SS_AUTO_FULL;
1245 if (scrub->os_running) {
1246 if (scrub->os_partial_scan && !scrub->os_in_join)
1249 osd_add_oi_cache(info, dev, id, fid);
1250 if (IS_ERR_OR_NULL(inode) || result) {
1251 osd_oii_insert(dev, oic, result == -ENOENT);
1252 GOTO(out, result = -EINPROGRESS);
1256 LASSERT(obj->oo_inode == inode);
1258 osd_oii_insert(dev, oic, true);
1262 if (dev->od_auto_scrub_interval == AS_NEVER) {
1264 GOTO(out, result = -EREMCHG);
1267 LASSERT(obj->oo_inode == inode);
1269 osd_add_oi_cache(info, dev, id, fid);
1274 rc1 = osd_scrub_start(env, dev, flags);
1275 LCONSOLE_WARN("%s: trigger OI scrub by RPC for the " DFID" with flags "
1276 "0x%x, rc = %d\n", osd_name(dev), PFID(fid), flags, rc1);
1277 if (rc1 && rc1 != -EALREADY)
1278 GOTO(out, result = -EREMCHG);
1280 osd_add_oi_cache(info, dev, id, fid);
1281 if (IS_ERR_OR_NULL(inode) || result) {
1282 osd_oii_insert(dev, oic, result == -ENOENT);
1283 GOTO(out, result = -EINPROGRESS);
1287 LASSERT(obj->oo_inode == inode);
1289 osd_oii_insert(dev, oic, true);
1294 if (unlikely(obj->oo_header))
1297 result = osd_check_lma(env, obj);
1301 LASSERTF(id->oii_ino == inode->i_ino &&
1302 id->oii_gen == inode->i_generation,
1303 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1304 PFID(fid), id->oii_ino, id->oii_gen,
1305 inode->i_ino, inode->i_generation);
1307 saved_ino = inode->i_ino;
1308 saved_gen = inode->i_generation;
1310 if (unlikely(result == -ENODATA)) {
1312 * If the OI scrub updated the OI mapping by race, it
1313 * must be valid. Trust the inode that has no LMA EA.
1318 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1321 * The OI mapping is still there, the inode is still
1322 * valid. It is just becaues the inode has no LMA EA.
1324 if (saved_ino == id->oii_ino &&
1325 saved_gen == id->oii_gen)
1329 * It is the OI scrub updated the OI mapping by race.
1330 * The new OI mapping must be valid.
1338 * "result == -ENOENT" means that the OI mappinghas been
1339 * removed by race, so the inode belongs to other object.
1341 * Others error can be returned directly.
1343 if (result == -ENOENT) {
1346 obj->oo_inode = NULL;
1351 if (result != -EREMCHG)
1357 * if two OST objects map to the same inode, and inode mode is
1358 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1359 * reserved by precreate, and not written yet, in this case, don't
1360 * set inode for the object whose FID mismatch, so that it can create
1361 * inode and not block precreate.
1363 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1364 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1365 obj->oo_inode = NULL;
1366 GOTO(out, result = 0);
1369 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1371 * "result == -ENOENT" means the cached OI mapping has been removed
1372 * from the OI file by race, above inode belongs to other object.
1374 if (result == -ENOENT) {
1377 obj->oo_inode = NULL;
1378 GOTO(out, result = 0);
1384 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1390 * It is the OI scrub updated the OI mapping by race.
1391 * The new OI mapping must be valid.
1399 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1400 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1402 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1404 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1406 obj->oo_pfid_in_lma = 1;
1407 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1409 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1410 } else if (result != -ENODATA) {
1415 obj->oo_compat_dot_created = 1;
1416 obj->oo_compat_dotdot_created = 1;
1418 if (S_ISDIR(inode->i_mode) &&
1419 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1420 osd_check_lmv(info, dev, inode, oic);
1422 result = osd_attach_jinode(inode);
1427 GOTO(out, result = 0);
1429 LASSERT(!obj->oo_hl_head);
1430 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1432 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1435 if (result || !obj->oo_inode) {
1436 if (!IS_ERR_OR_NULL(inode))
1439 obj->oo_inode = NULL;
1441 fid_zero(&oic->oic_fid);
1444 LINVRNT(osd_invariant(obj));
1449 * Concurrency: shouldn't matter.
1451 static void osd_object_init0(struct osd_object *obj)
1453 LASSERT(obj->oo_inode != NULL);
1454 obj->oo_dt.do_body_ops = &osd_body_ops;
1455 obj->oo_dt.do_lu.lo_header->loh_attr |=
1456 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1460 * Concurrency: no concurrent access is possible that early in object
1463 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1464 const struct lu_object_conf *conf)
1466 struct osd_object *obj = osd_obj(l);
1469 LINVRNT(osd_invariant(obj));
1471 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1472 cfs_fail_val == 2) {
1473 struct osd_thread_info *info = osd_oti_get(env);
1474 struct osd_idmap_cache *oic = &info->oti_cache;
1475 /* invalidate thread cache */
1476 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1478 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1479 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1480 l->lo_header->loh_attr |= LOHA_EXISTS;
1484 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1485 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1486 if (result == 0 && obj->oo_inode != NULL) {
1487 struct osd_thread_info *oti = osd_oti_get(env);
1488 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1490 osd_object_init0(obj);
1491 if (unlikely(obj->oo_header))
1494 result = osd_get_lma(oti, obj->oo_inode,
1495 &oti->oti_obj_dentry, loa);
1498 * Convert LMAI flags to lustre LMA flags
1499 * and cache it to oo_lma_flags
1502 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1503 } else if (result == -ENODATA) {
1507 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
1509 LINVRNT(osd_invariant(obj));
1514 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1515 * The left part is for value, binary mode.
1517 struct osd_xattr_entry {
1518 struct list_head oxe_list;
1522 struct rcu_head oxe_rcu;
1526 static int osd_oxc_get(struct osd_object *obj, const char *name,
1529 struct osd_xattr_entry *tmp;
1530 struct osd_xattr_entry *oxe = NULL;
1531 size_t namelen = strlen(name);
1535 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1536 if (namelen == tmp->oxe_namelen &&
1537 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1544 GOTO(out, rc = -ENOENT);
1546 if (!oxe->oxe_exist)
1547 GOTO(out, rc = -ENODATA);
1550 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1553 if (buf->lb_buf == NULL)
1556 if (buf->lb_len < rc)
1557 GOTO(out, rc = -ERANGE);
1559 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1566 static void osd_oxc_free(struct rcu_head *head)
1568 struct osd_xattr_entry *oxe;
1570 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1571 OBD_FREE(oxe, oxe->oxe_len);
1574 static void osd_oxc_add(struct osd_object *obj, const char *name,
1575 const char *buf, int buflen)
1577 struct osd_xattr_entry *oxe;
1578 struct osd_xattr_entry *old = NULL;
1579 struct osd_xattr_entry *tmp;
1580 size_t namelen = strlen(name);
1581 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1583 OBD_ALLOC(oxe, len);
1587 INIT_LIST_HEAD(&oxe->oxe_list);
1589 oxe->oxe_namelen = namelen;
1590 memcpy(oxe->oxe_buf, name, namelen);
1592 LASSERT(buf != NULL);
1593 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1594 oxe->oxe_exist = true;
1596 oxe->oxe_exist = false;
1599 /* this should be rarely called, just remove old and add new */
1600 spin_lock(&obj->oo_guard);
1601 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1602 if (namelen == tmp->oxe_namelen &&
1603 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1609 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1610 call_rcu(&old->oxe_rcu, osd_oxc_free);
1612 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1614 spin_unlock(&obj->oo_guard);
1617 static void osd_oxc_del(struct osd_object *obj, const char *name)
1619 struct osd_xattr_entry *oxe;
1620 size_t namelen = strlen(name);
1622 spin_lock(&obj->oo_guard);
1623 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1624 if (namelen == oxe->oxe_namelen &&
1625 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1626 list_del_rcu(&oxe->oxe_list);
1627 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1631 spin_unlock(&obj->oo_guard);
1634 static void osd_oxc_fini(struct osd_object *obj)
1636 struct osd_xattr_entry *oxe, *next;
1638 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1639 list_del(&oxe->oxe_list);
1640 OBD_FREE(oxe, oxe->oxe_len);
1645 * Concurrency: no concurrent access is possible that late in object
1648 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1650 struct osd_object *obj = osd_obj(l);
1651 struct lu_object_header *h = obj->oo_header;
1653 LINVRNT(osd_invariant(obj));
1656 dt_object_fini(&obj->oo_dt);
1657 if (obj->oo_hl_head != NULL)
1658 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1659 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1662 lu_object_header_free(h);
1666 * Concurrency: no concurrent access is possible that late in object
1669 static void osd_index_fini(struct osd_object *o)
1671 struct iam_container *bag;
1673 if (o->oo_dir != NULL) {
1674 bag = &o->oo_dir->od_container;
1675 if (o->oo_inode != NULL) {
1676 if (bag->ic_object == o->oo_inode)
1677 iam_container_fini(bag);
1679 OBD_FREE_PTR(o->oo_dir);
1685 OSD_TXN_OI_DELETE_CREDITS = 20,
1686 OSD_TXN_INODE_DELETE_CREDITS = 20
1693 #if OSD_THANDLE_STATS
1695 * Set time when the handle is allocated
1697 static void osd_th_alloced(struct osd_thandle *oth)
1699 oth->oth_alloced = ktime_get();
1703 * Set time when the handle started
1705 static void osd_th_started(struct osd_thandle *oth)
1707 oth->oth_started = ktime_get();
1711 * Check whether the we deal with this handle for too long.
1713 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1714 ktime_t alloced, ktime_t started,
1717 ktime_t now = ktime_get();
1719 LASSERT(dev != NULL);
1721 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1722 ktime_us_delta(started, alloced));
1723 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1724 ktime_us_delta(closed, started));
1725 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1726 ktime_us_delta(now, closed));
1728 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1729 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1730 oth, now, alloced, started, closed);
1731 libcfs_debug_dumpstack(NULL);
1735 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1737 ktime_t __closed = ktime_get(); \
1738 ktime_t __alloced = oth->oth_alloced; \
1739 ktime_t __started = oth->oth_started; \
1742 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1745 #else /* OSD_THANDLE_STATS */
1747 #define osd_th_alloced(h) do {} while(0)
1748 #define osd_th_started(h) do {} while(0)
1749 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1751 #endif /* OSD_THANDLE_STATS */
1754 * Concurrency: doesn't access mutable data.
1756 static int osd_param_is_not_sane(const struct osd_device *dev,
1757 const struct thandle *th)
1759 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1761 return oh->ot_credits > osd_transaction_size(dev);
1765 * Concurrency: shouldn't matter.
1767 static void osd_trans_commit_cb(struct super_block *sb,
1768 struct ldiskfs_journal_cb_entry *jcb, int error)
1770 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1771 struct thandle *th = &oh->ot_super;
1772 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1773 struct osd_device *osd = osd_dev(lud);
1774 struct dt_txn_commit_cb *dcb, *tmp;
1776 LASSERT(oh->ot_handle == NULL);
1779 CERROR("transaction @0x%p commit error: %d\n", th, error);
1781 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1782 /* call per-transaction callbacks if any */
1783 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1785 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1786 "commit callback entry: magic=%x name='%s'\n",
1787 dcb->dcb_magic, dcb->dcb_name);
1788 list_del_init(&dcb->dcb_linkage);
1789 dcb->dcb_func(NULL, th, dcb, error);
1792 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1793 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1794 wake_up(&osd->od_commit_cb_done);
1800 static struct thandle *osd_trans_create(const struct lu_env *env,
1801 struct dt_device *d)
1803 struct osd_thread_info *oti = osd_oti_get(env);
1804 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1805 struct osd_thandle *oh;
1811 CERROR("%s: someone try to start transaction under "
1812 "readonly mode, should be disabled.\n",
1813 osd_name(osd_dt_dev(d)));
1815 RETURN(ERR_PTR(-EROFS));
1818 /* on pending IO in this thread should left from prev. request */
1819 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1821 sb_start_write(osd_sb(osd_dt_dev(d)));
1823 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1825 sb_end_write(osd_sb(osd_dt_dev(d)));
1826 RETURN(ERR_PTR(-ENOMEM));
1829 oh->ot_quota_trans = &oti->oti_quota_trans;
1830 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1835 oh->oh_declared_ext = 0;
1836 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1837 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1838 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1841 memset(oti->oti_declare_ops, 0,
1842 sizeof(oti->oti_declare_ops));
1843 memset(oti->oti_declare_ops_cred, 0,
1844 sizeof(oti->oti_declare_ops_cred));
1845 memset(oti->oti_declare_ops_used, 0,
1846 sizeof(oti->oti_declare_ops_used));
1848 oti->oti_ins_cache_depth++;
1853 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1855 struct osd_thread_info *oti = osd_oti_get(env);
1856 struct osd_thandle *oh;
1858 oh = container_of(th, struct osd_thandle, ot_super);
1859 LASSERT(oh != NULL);
1861 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1862 oti->oti_declare_ops[OSD_OT_CREATE],
1863 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1864 oti->oti_declare_ops_used[OSD_OT_CREATE],
1865 oti->oti_declare_ops[OSD_OT_DESTROY],
1866 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1867 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1868 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1869 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1870 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1871 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1872 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1873 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1874 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1875 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1876 oti->oti_declare_ops[OSD_OT_WRITE],
1877 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1878 oti->oti_declare_ops_used[OSD_OT_WRITE],
1879 oti->oti_declare_ops[OSD_OT_PUNCH],
1880 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1881 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1882 oti->oti_declare_ops[OSD_OT_QUOTA],
1883 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1884 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1885 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1886 oti->oti_declare_ops[OSD_OT_INSERT],
1887 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1888 oti->oti_declare_ops_used[OSD_OT_INSERT],
1889 oti->oti_declare_ops[OSD_OT_DELETE],
1890 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1891 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1892 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1893 oti->oti_declare_ops[OSD_OT_REF_ADD],
1894 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1895 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1896 oti->oti_declare_ops[OSD_OT_REF_DEL],
1897 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1898 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1902 * Concurrency: shouldn't matter.
1904 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1907 struct osd_thread_info *oti = osd_oti_get(env);
1908 struct osd_device *dev = osd_dt_dev(d);
1910 struct osd_thandle *oh;
1915 LASSERT(current->journal_info == NULL);
1917 oh = container_of(th, struct osd_thandle, ot_super);
1918 LASSERT(oh != NULL);
1919 LASSERT(oh->ot_handle == NULL);
1921 rc = dt_txn_hook_start(env, d, th);
1925 if (unlikely(osd_param_is_not_sane(dev, th))) {
1926 static unsigned long last_printed;
1927 static int last_credits;
1930 * don't make noise on a tiny testing systems
1931 * actual credits misuse will be caught anyway
1933 if (last_credits != oh->ot_credits &&
1934 time_after(jiffies, last_printed +
1935 cfs_time_seconds(60)) &&
1936 osd_transaction_size(dev) > 512) {
1937 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1938 oh->ot_credits, osd_transaction_size(dev));
1939 osd_trans_dump_creds(env, th);
1940 libcfs_debug_dumpstack(NULL);
1941 last_credits = oh->ot_credits;
1942 last_printed = jiffies;
1945 * XXX Limit the credits to 'max_transaction_buffers', and
1946 * let the underlying filesystem to catch the error if
1947 * we really need so many credits.
1949 * This should be removed when we can calculate the
1950 * credits precisely.
1952 oh->ot_credits = osd_transaction_size(dev);
1953 } else if (ldiskfs_track_declares_assert != 0) {
1955 * reserve few credits to prevent an assertion in JBD
1956 * our debugging mechanism will be able to detected
1957 * overuse. this can help to debug single-update
1960 oh->ot_credits += 10;
1961 if (unlikely(osd_param_is_not_sane(dev, th)))
1962 oh->ot_credits = osd_transaction_size(dev);
1965 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1966 GOTO(out, rc = -EIO);
1969 * XXX temporary stuff. Some abstraction layer should
1972 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1976 LASSERT(oti->oti_txns == 0);
1978 atomic_inc(&dev->od_commit_cb_in_flight);
1979 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1990 static int osd_seq_exists(const struct lu_env *env,
1991 struct osd_device *osd, u64 seq)
1993 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1994 struct seq_server_site *ss = osd_seq_site(osd);
1999 LASSERT(ss != NULL);
2000 LASSERT(ss->ss_server_fld != NULL);
2002 rc = osd_fld_lookup(env, osd, seq, range);
2005 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2006 osd_name(osd), seq, rc);
2010 RETURN(ss->ss_node_id == range->lsr_index);
2013 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2015 struct dt_txn_commit_cb *dcb;
2016 struct dt_txn_commit_cb *tmp;
2018 /* call per-transaction stop callbacks if any */
2019 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2021 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2022 "commit callback entry: magic=%x name='%s'\n",
2023 dcb->dcb_magic, dcb->dcb_name);
2024 list_del_init(&dcb->dcb_linkage);
2025 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2030 * Concurrency: shouldn't matter.
2032 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2035 struct osd_thread_info *oti = osd_oti_get(env);
2036 struct osd_thandle *oh;
2037 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2038 struct osd_device *osd = osd_dt_dev(th->th_dev);
2039 struct qsd_instance *qsd = osd_def_qsd(osd);
2040 struct lquota_trans *qtrans;
2041 LIST_HEAD(truncates);
2042 int rc = 0, remove_agents = 0;
2046 oh = container_of(th, struct osd_thandle, ot_super);
2048 remove_agents = oh->ot_remove_agents;
2050 qtrans = oh->ot_quota_trans;
2051 oh->ot_quota_trans = NULL;
2053 /* move locks to local list, stop tx, execute truncates */
2054 list_splice(&oh->ot_trunc_locks, &truncates);
2056 if (oh->ot_handle != NULL) {
2059 handle_t *hdl = oh->ot_handle;
2062 * add commit callback
2063 * notice we don't do this in osd_trans_start()
2064 * as underlying transaction can change during truncate
2066 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2069 LASSERT(oti->oti_txns == 1);
2072 rc = dt_txn_hook_stop(env, th);
2074 CERROR("%s: failed in transaction hook: rc = %d\n",
2077 osd_trans_stop_cb(oh, rc);
2078 /* hook functions might modify th_sync */
2079 hdl->h_sync = th->th_sync;
2081 oh->ot_handle = NULL;
2082 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2084 CERROR("%s: failed to stop transaction: rc = %d\n",
2085 osd_name(osd), rc2);
2089 osd_process_truncates(&truncates);
2091 osd_trans_stop_cb(oh, th->th_result);
2095 osd_trunc_unlock_all(env, &truncates);
2097 /* inform the quota slave device that the transaction is stopping */
2098 qsd_op_end(env, qsd, qtrans);
2101 * as we want IO to journal and data IO be concurrent, we don't block
2102 * awaiting data IO completion in osd_do_bio(), instead we wait here
2103 * once transaction is submitted to the journal. all reqular requests
2104 * don't do direct IO (except read/write), thus this wait_event becomes
2107 * IMPORTANT: we have to wait till any IO submited by the thread is
2108 * completed otherwise iobuf may be corrupted by different request
2110 wait_event(iobuf->dr_wait,
2111 atomic_read(&iobuf->dr_numreqs) == 0);
2112 osd_fini_iobuf(osd, iobuf);
2114 rc = iobuf->dr_error;
2116 if (unlikely(remove_agents != 0))
2117 osd_process_scheduled_agent_removals(env, osd);
2119 LASSERT(oti->oti_ins_cache_depth > 0);
2120 oti->oti_ins_cache_depth--;
2121 /* reset OI cache for safety */
2122 if (oti->oti_ins_cache_depth == 0)
2123 oti->oti_ins_cache_used = 0;
2125 sb_end_write(osd_sb(osd));
2130 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2132 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2135 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2136 LASSERT(&dcb->dcb_func != NULL);
2137 if (dcb->dcb_flags & DCB_TRANS_STOP)
2138 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2140 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2146 * Called just before object is freed. Releases all resources except for
2147 * object itself (that is released by osd_object_free()).
2149 * Concurrency: no concurrent access is possible that late in object
2152 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2154 struct osd_object *obj = osd_obj(l);
2155 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2156 struct inode *inode = obj->oo_inode;
2161 LINVRNT(osd_invariant(obj));
2164 * If object is unlinked remove fid->ino mapping from object index.
2167 osd_index_fini(obj);
2172 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2173 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2175 uid = i_uid_read(inode);
2176 gid = i_gid_read(inode);
2177 projid = i_projid_read(inode);
2179 obj->oo_inode = NULL;
2182 /* do not rebalance quota if the caller needs to release memory
2183 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2184 * transaction and risk to deadlock - LU-12178 */
2185 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2188 if (!obj->oo_header && qsd) {
2189 struct osd_thread_info *info = osd_oti_get(env);
2190 struct lquota_id_info *qi = &info->oti_qi;
2192 /* Release granted quota to master if necessary */
2193 qi->lqi_id.qid_uid = uid;
2194 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2196 qi->lqi_id.qid_uid = gid;
2197 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2199 qi->lqi_id.qid_uid = projid;
2200 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2205 * Concurrency: ->loo_object_release() is called under site spin-lock.
2207 static void osd_object_release(const struct lu_env *env,
2208 struct lu_object *l)
2210 struct osd_object *o = osd_obj(l);
2213 * nobody should be releasing a non-destroyed object with nlink=0
2214 * the API allows this, but ldiskfs doesn't like and then report
2215 * this inode as deleted
2217 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2218 o->oo_inode->i_nlink == 0));
2222 * Concurrency: shouldn't matter.
2224 static int osd_object_print(const struct lu_env *env, void *cookie,
2225 lu_printer_t p, const struct lu_object *l)
2227 struct osd_object *o = osd_obj(l);
2228 struct iam_descr *d;
2230 if (o->oo_dir != NULL)
2231 d = o->oo_dir->od_container.ic_descr;
2234 return (*p)(env, cookie,
2235 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2237 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2238 o->oo_inode ? o->oo_inode->i_generation : 0,
2239 d ? d->id_ops->id_name : "plain");
2243 * Concurrency: shouldn't matter.
2245 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2246 struct obd_statfs *sfs, struct obd_statfs_info *info)
2248 struct osd_device *osd = osd_dt_dev(d);
2249 struct super_block *sb = osd_sb(osd);
2250 struct kstatfs *ksfs;
2254 if (unlikely(osd->od_mnt == NULL))
2255 return -EINPROGRESS;
2257 /* osd_lproc.c call this without env, allocate ksfs for that case */
2258 if (unlikely(env == NULL)) {
2259 OBD_ALLOC_PTR(ksfs);
2263 ksfs = &osd_oti_get(env)->oti_ksfs;
2266 result = sb->s_op->statfs(sb->s_root, ksfs);
2270 statfs_pack(sfs, ksfs);
2271 if (unlikely(sb->s_flags & SB_RDONLY))
2272 sfs->os_state |= OS_STATFS_READONLY;
2274 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2276 if (ldiskfs_has_feature_extents(sb))
2277 sfs->os_maxbytes = sb->s_maxbytes;
2279 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2282 * Reserve some space so to avoid fragmenting the filesystem too much.
2283 * Fragmentation not only impacts performance, but can also increase
2284 * metadata overhead significantly, causing grant calculation to be
2287 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2289 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2290 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2291 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2292 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2294 sfs->os_blocks -= reserved;
2295 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2296 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2299 if (unlikely(env == NULL))
2305 * Estimate space needed for file creations. We assume the largest filename
2306 * which is 2^64 - 1, hence a filename of 20 chars.
2307 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2309 #ifdef __LDISKFS_DIR_REC_LEN
2310 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2312 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2316 * Concurrency: doesn't access mutable data.
2318 static void osd_conf_get(const struct lu_env *env,
2319 const struct dt_device *dev,
2320 struct dt_device_param *param)
2322 struct osd_device *d = osd_dt_dev(dev);
2323 struct super_block *sb = osd_sb(d);
2324 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2329 * XXX should be taken from not-yet-existing fs abstraction layer.
2331 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2332 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2333 param->ddp_symlink_max = sb->s_blocksize;
2334 param->ddp_mount_type = LDD_MT_LDISKFS;
2335 if (ldiskfs_has_feature_extents(sb))
2336 param->ddp_maxbytes = sb->s_maxbytes;
2338 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2340 * inode are statically allocated, so per-inode space consumption
2341 * is the space consumed by the directory entry
2343 param->ddp_inodespace = PER_OBJ_USAGE;
2345 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2346 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2347 * maximum length to not under-count the actual number of extents
2348 * needed for writing a file if there are sub-optimal block allocations.
2350 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2351 /* worst-case extent insertion metadata overhead */
2352 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2353 param->ddp_mntopts = 0;
2354 if (test_opt(sb, XATTR_USER))
2355 param->ddp_mntopts |= MNTOPT_USERXATTR;
2356 if (test_opt(sb, POSIX_ACL))
2357 param->ddp_mntopts |= MNTOPT_ACL;
2360 * LOD might calculate the max stripe count based on max_ea_size,
2361 * so we need take account in the overhead as well,
2362 * xattr_header + magic + xattr_entry_head
2364 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2365 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2367 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2368 if (ldiskfs_has_feature_ea_inode(sb))
2369 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2373 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2375 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2376 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2379 * Preferred RPC size for efficient disk IO. 4MB shows good
2380 * all-around performance for ldiskfs, but use bigalloc chunk size
2381 * by default if larger.
2383 #if defined(LDISKFS_CLUSTER_SIZE)
2384 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2385 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2388 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2390 param->ddp_t10_cksum_type = 0;
2392 unsigned short interval = blk_integrity_interval(bi);
2393 name = blk_integrity_name(bi);
2401 if (strncmp(name, "T10-DIF-TYPE",
2402 sizeof("T10-DIF-TYPE") - 1) == 0) {
2403 /* also skip "1/3-" at end */
2404 const int type_off = sizeof("T10-DIF-TYPE.");
2405 char type_number = name[type_off - 2];
2407 if (interval != 512 && interval != 4096) {
2408 CERROR("%s: unsupported T10PI sector size %u\n",
2409 d->od_svname, interval);
2410 } else if (type_number != '1' && type_number != '3') {
2411 CERROR("%s: unsupported T10PI type %s\n",
2412 d->od_svname, name);
2413 } else if (strcmp(name + type_off, "CRC") == 0) {
2414 d->od_t10_type = type_number == '1' ?
2415 OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2416 param->ddp_t10_cksum_type = interval == 512 ?
2417 OBD_CKSUM_T10CRC512 :
2419 } else if (strcmp(name + type_off, "IP") == 0) {
2420 d->od_t10_type = type_number == '1' ?
2421 OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2422 param->ddp_t10_cksum_type = interval == 512 ?
2423 OBD_CKSUM_T10IP512 :
2426 CERROR("%s: unsupported checksum type of "
2428 d->od_svname, name);
2432 CERROR("%s: unsupported T10PI type '%s'",
2433 d->od_svname, name);
2437 param->ddp_has_lseek_data_hole = true;
2440 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2442 return osd_sb(osd_dt_dev(d));
2446 * Concurrency: shouldn't matter.
2448 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2451 struct super_block *s = osd_sb(osd_dt_dev(d));
2454 down_read(&s->s_umount);
2455 rc = s->s_op->sync_fs(s, 1);
2456 up_read(&s->s_umount);
2458 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2465 * Start commit for OSD device.
2467 * An implementation of dt_commit_async method for OSD device.
2468 * Asychronously starts underlayng fs sync and thereby a transaction
2471 * \param env environment
2472 * \param d dt device
2474 * \see dt_device_operations
2476 static int osd_commit_async(const struct lu_env *env,
2477 struct dt_device *d)
2479 struct super_block *s = osd_sb(osd_dt_dev(d));
2484 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2485 down_read(&s->s_umount);
2486 rc = s->s_op->sync_fs(s, 0);
2487 up_read(&s->s_umount);
2492 static int (*priv_security_file_alloc)(struct file *file);
2494 int osd_security_file_alloc(struct file *file)
2496 if (priv_security_file_alloc)
2497 return priv_security_file_alloc(file);
2502 * Concurrency: shouldn't matter.
2504 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2506 struct super_block *sb = osd_sb(osd_dt_dev(d));
2507 struct block_device *dev = sb->s_bdev;
2508 int rc = -EOPNOTSUPP;
2512 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2513 osd_dt_dev(d)->od_svname, (long)dev, rc);
2519 * Note: we do not count into QUOTA here.
2520 * If we mount with --data_journal we may need more.
2522 const int osd_dto_credits_noquota[DTO_NR] = {
2525 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2526 * SINGLEDATA_TRANS_BLOCKS(8)
2527 * XXX Note: maybe iam need more, since iam have more level than
2530 [DTO_INDEX_INSERT] = 16,
2533 * just modify a single entry, probably merge few within a block
2535 [DTO_INDEX_DELETE] = 1,
2539 [DTO_INDEX_UPDATE] = 16,
2541 * 4(inode, inode bits, groups, GDT)
2542 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2544 [DTO_OBJECT_CREATE] = 4,
2546 * 4(inode, inode bits, groups, GDT)
2547 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2549 [DTO_OBJECT_DELETE] = 4,
2551 * Attr set credits (inode)
2553 [DTO_ATTR_SET_BASE] = 1,
2555 * Xattr set. The same as xattr of EXT3.
2556 * DATA_TRANS_BLOCKS(14)
2557 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2558 * are also counted in. Do not know why?
2560 [DTO_XATTR_SET] = 14,
2562 * credits for inode change during write.
2564 [DTO_WRITE_BASE] = 3,
2566 * credits for single block write.
2568 [DTO_WRITE_BLOCK] = 14,
2570 * Attr set credits for chown.
2571 * This is extra credits for setattr, and it is null without quota
2573 [DTO_ATTR_SET_CHOWN] = 0
2576 static const struct dt_device_operations osd_dt_ops = {
2577 .dt_root_get = osd_root_get,
2578 .dt_statfs = osd_statfs,
2579 .dt_trans_create = osd_trans_create,
2580 .dt_trans_start = osd_trans_start,
2581 .dt_trans_stop = osd_trans_stop,
2582 .dt_trans_cb_add = osd_trans_cb_add,
2583 .dt_conf_get = osd_conf_get,
2584 .dt_mnt_sb_get = osd_mnt_sb_get,
2585 .dt_sync = osd_sync,
2587 .dt_commit_async = osd_commit_async,
2590 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2593 struct osd_object *obj = osd_dt_obj(dt);
2594 struct osd_thread_info *oti = osd_oti_get(env);
2596 LINVRNT(osd_invariant(obj));
2598 LASSERT(obj->oo_owner != env);
2599 down_read_nested(&obj->oo_sem, role);
2601 LASSERT(obj->oo_owner == NULL);
2605 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2608 struct osd_object *obj = osd_dt_obj(dt);
2609 struct osd_thread_info *oti = osd_oti_get(env);
2611 LINVRNT(osd_invariant(obj));
2613 LASSERT(obj->oo_owner != env);
2614 down_write_nested(&obj->oo_sem, role);
2616 LASSERT(obj->oo_owner == NULL);
2617 obj->oo_owner = env;
2621 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2623 struct osd_object *obj = osd_dt_obj(dt);
2624 struct osd_thread_info *oti = osd_oti_get(env);
2626 LINVRNT(osd_invariant(obj));
2628 LASSERT(oti->oti_r_locks > 0);
2630 up_read(&obj->oo_sem);
2633 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2635 struct osd_object *obj = osd_dt_obj(dt);
2636 struct osd_thread_info *oti = osd_oti_get(env);
2638 LINVRNT(osd_invariant(obj));
2640 LASSERT(obj->oo_owner == env);
2641 LASSERT(oti->oti_w_locks > 0);
2643 obj->oo_owner = NULL;
2644 up_write(&obj->oo_sem);
2647 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2649 struct osd_object *obj = osd_dt_obj(dt);
2651 LINVRNT(osd_invariant(obj));
2653 return obj->oo_owner == env;
2656 static void osd_inode_getattr(const struct lu_env *env,
2657 struct inode *inode, struct lu_attr *attr)
2659 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2660 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2661 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2662 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2664 attr->la_atime = inode->i_atime.tv_sec;
2665 attr->la_mtime = inode->i_mtime.tv_sec;
2666 attr->la_ctime = inode->i_ctime.tv_sec;
2667 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2668 attr->la_mode = inode->i_mode;
2669 attr->la_size = i_size_read(inode);
2670 attr->la_blocks = inode->i_blocks;
2671 attr->la_uid = i_uid_read(inode);
2672 attr->la_gid = i_gid_read(inode);
2673 attr->la_projid = i_projid_read(inode);
2674 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2675 attr->la_nlink = inode->i_nlink;
2676 attr->la_rdev = inode->i_rdev;
2677 attr->la_blksize = 1 << inode->i_blkbits;
2678 attr->la_blkbits = inode->i_blkbits;
2680 * Ext4 did not transfer inherit flags from raw inode
2681 * to inode flags, and ext4 internally test raw inode
2682 * @i_flags directly. Instead of patching ext4, we do it here.
2684 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2685 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2688 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2691 struct osd_object *obj = osd_dt_obj(dt);
2692 const struct dt_it_ops *iops;
2698 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2699 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2701 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2702 *count = obj->oo_dirent_count;
2706 /* directory not initialized yet */
2707 if (!dt->do_index_ops) {
2712 iops = &dt->do_index_ops->dio_it;
2713 it = iops->init(env, dt, LUDA_64BITHASH);
2715 RETURN(PTR_ERR(it));
2717 rc = iops->load(env, it, 0);
2719 if (rc == -ENODATA) {
2726 rc = iops->next(env, it);
2728 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2732 if (iops->key_size(env, it) == 0)
2738 obj->oo_dirent_count = *count;
2743 iops->fini(env, it);
2748 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2749 struct lu_attr *attr)
2751 struct osd_object *obj = osd_dt_obj(dt);
2754 if (unlikely(!dt_object_exists(dt)))
2756 if (unlikely(obj->oo_destroyed))
2759 LASSERT(!dt_object_remote(dt));
2760 LINVRNT(osd_invariant(obj));
2762 spin_lock(&obj->oo_guard);
2763 osd_inode_getattr(env, obj->oo_inode, attr);
2764 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2765 attr->la_valid |= LA_FLAGS;
2766 attr->la_flags |= LUSTRE_ORPHAN_FL;
2768 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2769 attr->la_valid |= LA_FLAGS;
2770 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2772 spin_unlock(&obj->oo_guard);
2774 if (S_ISDIR(obj->oo_inode->i_mode) &&
2775 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2776 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2781 static int osd_declare_attr_qid(const struct lu_env *env,
2782 struct osd_object *obj,
2783 struct osd_thandle *oh, long long bspace,
2784 qid_t old_id, qid_t new_id, bool enforce,
2785 unsigned int type, bool ignore_edquot)
2788 struct osd_thread_info *info = osd_oti_get(env);
2789 struct lquota_id_info *qi = &info->oti_qi;
2791 qi->lqi_type = type;
2792 /* inode accounting */
2793 qi->lqi_is_blk = false;
2795 /* one more inode for the new id ... */
2796 qi->lqi_id.qid_uid = new_id;
2798 /* Reserve credits for the new id */
2799 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2800 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2805 /* and one less inode for the current id */
2806 qi->lqi_id.qid_uid = old_id;
2808 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2809 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2814 /* block accounting */
2815 qi->lqi_is_blk = true;
2817 /* more blocks for the new id ... */
2818 qi->lqi_id.qid_uid = new_id;
2819 qi->lqi_space = bspace;
2821 * Credits for the new uid has been reserved, re-use "obj"
2822 * to save credit reservation.
2824 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2825 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2830 /* and finally less blocks for the current uid */
2831 qi->lqi_id.qid_uid = old_id;
2832 qi->lqi_space = -bspace;
2833 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2834 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2840 static int osd_declare_attr_set(const struct lu_env *env,
2841 struct dt_object *dt,
2842 const struct lu_attr *attr,
2843 struct thandle *handle)
2845 struct osd_thandle *oh;
2846 struct osd_object *obj;
2855 LASSERT(dt != NULL);
2856 LASSERT(handle != NULL);
2858 obj = osd_dt_obj(dt);
2859 LASSERT(osd_invariant(obj));
2861 oh = container_of(handle, struct osd_thandle, ot_super);
2862 LASSERT(oh->ot_handle == NULL);
2864 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2865 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2867 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2868 osd_dto_credits_noquota[DTO_XATTR_SET]);
2870 if (attr == NULL || obj->oo_inode == NULL)
2873 bspace = obj->oo_inode->i_blocks << 9;
2874 bspace = toqb(bspace);
2877 * Changing ownership is always preformed by super user, it should not
2878 * fail with EDQUOT unless required explicitly.
2880 * We still need to call the osd_declare_qid() to calculate the journal
2881 * credits for updating quota accounting files and to trigger quota
2882 * space adjustment once the operation is completed.
2884 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2885 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2889 "%s: enforce quota on UID %u, GID %u (quota space is %lld)\n",
2890 osd_ino2name(obj->oo_inode), attr->la_uid,
2891 attr->la_gid, bspace);
2894 uid = i_uid_read(obj->oo_inode);
2895 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2896 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2897 attr->la_uid, enforce, USRQUOTA,
2902 gid = i_gid_read(obj->oo_inode);
2903 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
2904 attr->la_uid, gid, attr->la_gid);
2905 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2906 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
2907 attr->la_gid, enforce, GRPQUOTA,
2913 #ifdef HAVE_PROJECT_QUOTA
2914 if (attr->la_valid & LA_PROJID) {
2915 __u32 projid = i_projid_read(obj->oo_inode);
2917 enforce = (attr->la_valid & LA_PROJID) &&
2918 (attr->la_projid != projid);
2919 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2920 (qid_t)projid, (qid_t)attr->la_projid,
2921 enforce, PRJQUOTA, true);
2929 static int osd_inode_setattr(const struct lu_env *env,
2930 struct inode *inode, const struct lu_attr *attr)
2932 __u64 bits = attr->la_valid;
2934 /* Only allow set size for regular file */
2935 if (!S_ISREG(inode->i_mode))
2936 bits &= ~(LA_SIZE | LA_BLOCKS);
2941 if (bits & LA_ATIME)
2942 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2943 if (bits & LA_CTIME)
2944 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2945 if (bits & LA_MTIME)
2946 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2947 if (bits & LA_SIZE) {
2948 spin_lock(&inode->i_lock);
2949 LDISKFS_I(inode)->i_disksize = attr->la_size;
2950 i_size_write(inode, attr->la_size);
2951 spin_unlock(&inode->i_lock);
2955 * OSD should not change "i_blocks" which is used by quota.
2956 * "i_blocks" should be changed by ldiskfs only.
2959 inode->i_mode = (inode->i_mode & S_IFMT) |
2960 (attr->la_mode & ~S_IFMT);
2962 i_uid_write(inode, attr->la_uid);
2964 i_gid_write(inode, attr->la_gid);
2965 if (bits & LA_PROJID)
2966 i_projid_write(inode, attr->la_projid);
2967 if (bits & LA_NLINK)
2968 set_nlink(inode, attr->la_nlink);
2970 inode->i_rdev = attr->la_rdev;
2972 if (bits & LA_FLAGS) {
2973 /* always keep S_NOCMTIME */
2974 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2976 #if defined(S_ENCRYPTED)
2977 /* Always remove S_ENCRYPTED, because ldiskfs must not be
2978 * aware of encryption status. It is just stored into LMA
2979 * so that it can be forwared to client side.
2981 inode->i_flags &= ~S_ENCRYPTED;
2984 * Ext4 did not transfer inherit flags from
2985 * @inode->i_flags to raw inode i_flags when writing
2986 * flags, we do it explictly here.
2988 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2989 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2991 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2996 #ifdef HAVE_PROJECT_QUOTA
2997 static int osd_transfer_project(struct inode *inode, __u32 projid,
2998 struct thandle *handle)
3000 struct super_block *sb = inode->i_sb;
3001 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3004 struct ldiskfs_iloc iloc;
3005 struct ldiskfs_inode *raw_inode;
3006 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3008 if (!ldiskfs_has_feature_project(sb)) {
3009 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3010 == LDISKFS_DEF_PROJID);
3011 if (projid != LDISKFS_DEF_PROJID)
3017 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3020 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3021 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3024 err = ldiskfs_get_inode_loc(inode, &iloc);
3028 raw_inode = ldiskfs_raw_inode(&iloc);
3029 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3030 struct osd_thandle *oh = container_of(handle,
3034 * try to expand inode size automatically.
3036 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3037 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3045 dquot_initialize(inode);
3046 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3047 if (transfer_to[PRJQUOTA]) {
3048 err = __dquot_transfer(inode, transfer_to);
3049 dqput(transfer_to[PRJQUOTA]);
3058 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3059 struct thandle *handle)
3063 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3064 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3068 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3069 inode->i_ino, i_uid_read(inode), attr->la_uid,
3070 i_gid_read(inode), attr->la_gid);
3072 dquot_initialize(inode);
3074 if (attr->la_valid & LA_UID)
3075 iattr.ia_valid |= ATTR_UID;
3076 if (attr->la_valid & LA_GID)
3077 iattr.ia_valid |= ATTR_GID;
3078 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3079 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3081 rc = dquot_transfer(inode, &iattr);
3083 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3084 osd_ino2name(inode), rc);
3089 /* Handle project id transfer here properly */
3090 if (attr->la_valid & LA_PROJID &&
3091 attr->la_projid != i_projid_read(inode)) {
3092 #ifdef HAVE_PROJECT_QUOTA
3093 rc = osd_transfer_project(inode, attr->la_projid, handle);
3098 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3099 osd_ino2name(inode), rc);
3106 static int osd_attr_set(const struct lu_env *env,
3107 struct dt_object *dt,
3108 const struct lu_attr *attr,
3109 struct thandle *handle)
3111 struct osd_object *obj = osd_dt_obj(dt);
3112 struct inode *inode;
3115 if (!dt_object_exists(dt))
3118 LASSERT(handle != NULL);
3119 LASSERT(!dt_object_remote(dt));
3120 LASSERT(osd_invariant(obj));
3122 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3124 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3125 !osd_obj2dev(obj)->od_is_ost) {
3126 struct osd_thread_info *oti = osd_oti_get(env);
3127 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3128 struct lu_fid *fid1 = &oti->oti_fid;
3129 struct osd_inode_id *id = &oti->oti_id;
3130 struct iam_path_descr *ipd;
3131 struct iam_container *bag;
3132 struct osd_thandle *oh;
3135 fid_cpu_to_be(fid1, fid0);
3136 memset(id, 1, sizeof(*id));
3137 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3138 fid0)->oi_dir.od_container;
3139 ipd = osd_idx_ipd_get(env, bag);
3140 if (unlikely(ipd == NULL))
3143 oh = container_of(handle, struct osd_thandle, ot_super);
3144 rc = iam_update(oh->ot_handle, bag,
3145 (const struct iam_key *)fid1,
3146 (const struct iam_rec *)id, ipd);
3147 osd_ipd_put(env, bag, ipd);
3148 return(rc > 0 ? 0 : rc);
3151 inode = obj->oo_inode;
3153 rc = osd_quota_transfer(inode, attr, handle);
3157 spin_lock(&obj->oo_guard);
3158 rc = osd_inode_setattr(env, inode, attr);
3159 spin_unlock(&obj->oo_guard);
3163 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3165 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3167 if (!(attr->la_valid & LA_FLAGS))
3170 /* Let's check if there are extra flags need to be set into LMA */
3171 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3172 struct osd_thread_info *info = osd_oti_get(env);
3173 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3175 LASSERT(!obj->oo_pfid_in_lma);
3177 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3178 &info->oti_ost_attrs);
3182 lma->lma_incompat |=
3183 lustre_to_lma_flags(attr->la_flags);
3184 lustre_lma_swab(lma);
3186 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3188 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3189 lma, sizeof(*lma), XATTR_REPLACE);
3191 struct osd_device *osd = osd_obj2dev(obj);
3193 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3194 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3195 lma->lma_incompat, rc);
3198 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3200 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3207 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3208 struct osd_object *obj,
3209 const char *name, const int namelen)
3211 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3214 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3215 umode_t mode, struct dt_allocation_hint *hint,
3216 struct thandle *th, struct lu_attr *attr)
3219 struct osd_device *osd = osd_obj2dev(obj);
3220 struct osd_thandle *oth;
3221 struct dt_object *parent = NULL;
3222 struct inode *inode;
3223 uid_t owner[2] = {0, 0};
3225 if (attr->la_valid & LA_UID)
3226 owner[0] = attr->la_uid;
3227 if (attr->la_valid & LA_GID)
3228 owner[1] = attr->la_gid;
3230 LINVRNT(osd_invariant(obj));
3231 LASSERT(obj->oo_inode == NULL);
3232 LASSERT(obj->oo_hl_head == NULL);
3234 if (S_ISDIR(mode) && ldiskfs_pdo) {
3236 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3237 if (obj->oo_hl_head == NULL)
3241 oth = container_of(th, struct osd_thandle, ot_super);
3242 LASSERT(oth->ot_handle->h_transaction != NULL);
3244 if (hint != NULL && hint->dah_parent != NULL &&
3245 !dt_object_remote(hint->dah_parent))
3246 parent = hint->dah_parent;
3248 inode = ldiskfs_create_inode(oth->ot_handle,
3249 parent ? osd_dt_obj(parent)->oo_inode :
3250 osd_sb(osd)->s_root->d_inode,
3252 if (!IS_ERR(inode)) {
3253 /* Do not update file c/mtime in ldiskfs. */
3254 inode->i_flags |= S_NOCMTIME;
3257 * For new created object, it must be consistent,
3258 * and it is unnecessary to scrub against it.
3260 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3262 obj->oo_inode = inode;
3265 if (obj->oo_hl_head != NULL) {
3266 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3267 obj->oo_hl_head = NULL;
3269 result = PTR_ERR(inode);
3271 LINVRNT(osd_invariant(obj));
3279 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3280 struct lu_attr *attr,
3281 struct dt_allocation_hint *hint,
3282 struct dt_object_format *dof,
3286 struct osd_thandle *oth;
3287 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3289 LASSERT(S_ISDIR(attr->la_mode));
3291 oth = container_of(th, struct osd_thandle, ot_super);
3292 LASSERT(oth->ot_handle->h_transaction != NULL);
3293 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3294 obj->oo_dirent_count = 0;
3295 result = osd_mkfile(info, obj, mode, hint, th, attr);
3300 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3301 struct lu_attr *attr,
3302 struct dt_allocation_hint *hint,
3303 struct dt_object_format *dof,
3307 struct osd_thandle *oth;
3308 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3310 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3312 LASSERT(S_ISREG(attr->la_mode));
3314 oth = container_of(th, struct osd_thandle, ot_super);
3315 LASSERT(oth->ot_handle->h_transaction != NULL);
3317 result = osd_mkfile(info, obj, mode, hint, th, attr);
3319 LASSERT(obj->oo_inode != NULL);
3320 if (feat->dif_flags & DT_IND_VARKEY)
3321 result = iam_lvar_create(obj->oo_inode,
3322 feat->dif_keysize_max,
3324 feat->dif_recsize_max,
3327 result = iam_lfix_create(obj->oo_inode,
3328 feat->dif_keysize_max,
3330 feat->dif_recsize_max,
3336 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3337 struct lu_attr *attr,
3338 struct dt_allocation_hint *hint,
3339 struct dt_object_format *dof,
3342 LASSERT(S_ISREG(attr->la_mode));
3343 return osd_mkfile(info, obj, (attr->la_mode &
3344 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3348 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3349 struct lu_attr *attr,
3350 struct dt_allocation_hint *hint,
3351 struct dt_object_format *dof,
3354 LASSERT(S_ISLNK(attr->la_mode));
3355 return osd_mkfile(info, obj, (attr->la_mode &
3356 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3360 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3361 struct lu_attr *attr,
3362 struct dt_allocation_hint *hint,
3363 struct dt_object_format *dof,
3366 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3369 LINVRNT(osd_invariant(obj));
3370 LASSERT(obj->oo_inode == NULL);
3371 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3372 S_ISFIFO(mode) || S_ISSOCK(mode));
3374 result = osd_mkfile(info, obj, mode, hint, th, attr);
3376 LASSERT(obj->oo_inode != NULL);
3378 * This inode should be marked dirty for i_rdev. Currently
3379 * that is done in the osd_attr_init().
3381 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3384 LINVRNT(osd_invariant(obj));
3388 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3390 struct dt_allocation_hint *hint,
3391 struct dt_object_format *dof,
3394 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3396 osd_obj_type_f result;
3412 result = osd_mk_index;
3423 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3424 struct dt_object *parent, struct dt_object *child,
3429 ah->dah_parent = parent;
3430 ah->dah_mode = child_mode;
3432 if (parent != NULL && !dt_object_remote(parent)) {
3433 /* will help to find FID->ino at dt_insert("..") */
3434 struct osd_object *pobj = osd_dt_obj(parent);
3436 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3440 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3441 struct lu_attr *attr, struct dt_object_format *dof,
3442 struct thandle *handle)
3444 struct inode *inode = obj->oo_inode;
3445 __u64 valid = attr->la_valid;
3448 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3450 if (dof->dof_type != DFT_NODE)
3451 attr->la_valid &= ~LA_RDEV;
3452 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3453 attr->la_valid &= ~LA_ATIME;
3454 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3455 attr->la_valid &= ~LA_CTIME;
3456 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3457 attr->la_valid &= ~LA_MTIME;
3459 result = osd_quota_transfer(inode, attr, handle);
3463 if (attr->la_valid != 0) {
3464 result = osd_inode_setattr(info->oti_env, inode, attr);
3466 * The osd_inode_setattr() should always succeed here. The
3467 * only error that could be returned is EDQUOT when we are
3468 * trying to change the UID or GID of the inode. However, this
3469 * should not happen since quota enforcement is no longer
3470 * enabled on ldiskfs (lquota takes care of it).
3472 LASSERTF(result == 0, "%d\n", result);
3473 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3476 attr->la_valid = valid;
3480 * Helper function for osd_create()
3482 * \retval 0, on success
3484 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3485 struct lu_attr *attr, struct dt_allocation_hint *hint,
3486 struct dt_object_format *dof, struct thandle *th)
3491 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3493 /* we drop umask so that permissions we pass are not affected */
3494 umask = current->fs->umask;
3495 current->fs->umask = 0;
3497 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3499 if (likely(obj->oo_inode != NULL)) {
3500 LASSERT(obj->oo_inode->i_state & I_NEW);
3503 * Unlock the inode before attr initialization to avoid
3504 * unnecessary dqget operations. LU-6378
3506 unlock_new_inode(obj->oo_inode);
3509 if (likely(result == 0)) {
3510 osd_attr_init(info, obj, attr, dof, th);
3511 osd_object_init0(obj);
3514 /* restore previous umask value */
3515 current->fs->umask = umask;
3517 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3523 * Helper function for osd_create()
3525 * \retval 0, on success
3527 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3528 const struct lu_fid *fid, struct thandle *th)
3530 struct osd_thread_info *info = osd_oti_get(env);
3531 struct osd_inode_id *id = &info->oti_id;
3532 struct osd_device *osd = osd_obj2dev(obj);
3533 struct osd_thandle *oh;
3536 LASSERT(obj->oo_inode != NULL);
3538 oh = container_of(th, struct osd_thandle, ot_super);
3539 LASSERT(oh->ot_handle);
3540 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3542 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3543 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3544 OI_CHECK_FLD, NULL);
3545 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3546 struct lu_fid next_fid = *fid;
3548 /* insert next object in advance, and map to the same inode */
3550 if (next_fid.f_oid != 0) {
3551 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3552 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3553 OI_CHECK_FLD, NULL);
3554 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3558 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3563 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3564 u64 seq, struct lu_seq_range *range)
3566 struct seq_server_site *ss = osd_seq_site(osd);
3568 if (fid_seq_is_idif(seq)) {
3569 fld_range_set_ost(range);
3570 range->lsr_index = idif_ost_idx(seq);
3574 if (!fid_seq_in_fldb(seq)) {
3575 fld_range_set_mdt(range);
3578 * FIXME: If ss is NULL, it suppose not get lsr_index
3581 range->lsr_index = ss->ss_node_id;
3585 LASSERT(ss != NULL);
3586 fld_range_set_any(range);
3587 /* OSD will only do local fld lookup */
3588 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3591 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3592 struct lu_attr *attr,
3593 struct dt_allocation_hint *hint,
3594 struct dt_object_format *dof,
3595 struct thandle *handle)
3597 struct osd_thandle *oh;
3602 LASSERT(handle != NULL);
3604 oh = container_of(handle, struct osd_thandle, ot_super);
3605 LASSERT(oh->ot_handle == NULL);
3608 * EA object consumes more credits than regular object: osd_mk_index
3609 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3610 * leaf_node, could involves the block, block bitmap, groups, GDT
3611 * change for each block, so add 4 * 2 credits in that case.
3613 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3614 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3615 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3617 * Reuse idle OI block may cause additional one OI block
3620 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3621 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3622 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3623 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3624 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3626 /* will help to find FID->ino mapping at dt_insert() */
3627 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3635 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3636 attr->la_projid, 1, oh, osd_dt_obj(dt),
3637 NULL, OSD_QID_INODE);
3645 * Called to destroy on-disk representation of the object
3647 * Concurrency: must be locked
3649 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3652 struct osd_object *obj = osd_dt_obj(dt);
3653 struct inode *inode = obj->oo_inode;
3654 struct osd_thandle *oh;
3662 oh = container_of(th, struct osd_thandle, ot_super);
3663 LASSERT(oh->ot_handle == NULL);
3665 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3666 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3668 /* For removing agent entry */
3669 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3670 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3673 * Recycle idle OI leaf may cause additional three OI blocks
3676 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3677 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3678 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3679 /* one less inode */
3680 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3681 i_projid_read(inode), -1, oh, obj, NULL,
3685 /* data to be truncated */
3686 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3687 i_projid_read(inode), 0, oh, obj, NULL,
3693 * will help to find FID->ino when this object is being
3696 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3701 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3704 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3705 struct osd_object *obj = osd_dt_obj(dt);
3706 struct inode *inode = obj->oo_inode;
3707 struct osd_device *osd = osd_obj2dev(obj);
3708 struct osd_thandle *oh;
3713 oh = container_of(th, struct osd_thandle, ot_super);
3714 LASSERT(oh->ot_handle);
3716 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3718 if (unlikely(fid_is_acct(fid)))
3721 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3722 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3724 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3725 osd_name(osd), PFID(fid), result);
3728 if (S_ISDIR(inode->i_mode)) {
3729 if (inode->i_nlink > 2)
3730 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3731 osd_name(osd), PFID(fid), inode->i_ino,
3734 spin_lock(&obj->oo_guard);
3736 spin_unlock(&obj->oo_guard);
3737 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3740 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3742 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3744 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3745 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3746 oh->ot_handle, OI_CHECK_FLD);
3748 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3749 /* XXX: add to ext3 orphan list */
3750 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3752 /* not needed in the cache anymore */
3753 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3754 obj->oo_destroyed = 1;
3760 * Put the fid into lustre_mdt_attrs, and then place the structure
3761 * inode's ea. This fid should not be altered during the life time
3764 * \retval +ve, on success
3765 * \retval -ve, on error
3767 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3769 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3770 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3772 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3773 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3778 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3781 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3784 lustre_loa_init(loa, fid, compat, incompat);
3785 lustre_loa_swab(loa, false);
3788 * For the OST device with 256 bytes inode size by default,
3789 * the PFID EA will be stored together with LMA EA to avoid
3790 * performance trouble. Otherwise the PFID EA can be stored
3791 * independently. LU-8998
3793 if ((compat & LMAC_FID_ON_OST) &&
3794 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3795 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3796 sizeof(*loa), XATTR_CREATE);
3798 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3799 sizeof(*lma), XATTR_CREATE);
3801 * LMA may already exist, but we need to check that all the
3802 * desired compat/incompat flags have been added.
3804 if (unlikely(rc == -EEXIST)) {
3805 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3806 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3810 if (rc < sizeof(*lma))
3813 lustre_loa_swab(loa, true);
3814 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3815 ((compat == 0 && incompat == 0) ||
3816 (!(~lma->lma_compat & compat) &&
3817 !(~lma->lma_incompat & incompat))))
3820 lma->lma_self_fid = *fid;
3821 lma->lma_compat |= compat;
3822 lma->lma_incompat |= incompat;
3823 if (rc == sizeof(*lma)) {
3824 lustre_lma_swab(lma);
3825 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3826 sizeof(*lma), XATTR_REPLACE);
3828 lustre_loa_swab(loa, false);
3829 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3830 sizeof(*loa), XATTR_REPLACE);
3838 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3839 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3840 * To have compatilibility with 1.8 ldiskfs driver we need to have
3841 * magic number at start of fid data.
3842 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3845 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3846 const struct lu_fid *fid)
3848 if (!fid_is_namespace_visible(fid) ||
3849 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3850 param->edp_magic = 0;
3854 param->edp_magic = LDISKFS_LUFID_MAGIC;
3855 param->edp_len = sizeof(struct lu_fid) + 1;
3856 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3860 * Try to read the fid from inode ea into dt_rec.
3862 * \param fid object fid.
3864 * \retval 0 on success
3866 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3867 __u32 ino, struct lu_fid *fid,
3868 struct osd_inode_id *id)
3870 struct osd_thread_info *info = osd_oti_get(env);
3871 struct inode *inode;
3875 osd_id_gen(id, ino, OSD_OII_NOGEN);
3876 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3878 RETURN(PTR_ERR(inode));
3884 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3886 struct inode *parent_dir,
3887 const struct lu_fid *dot_fid,
3888 const struct lu_fid *dot_dot_fid,
3889 struct osd_thandle *oth)
3891 struct ldiskfs_dentry_param *dot_ldp;
3892 struct ldiskfs_dentry_param *dot_dot_ldp;
3893 __u32 saved_nlink = dir->i_nlink;
3896 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3897 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3899 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3900 dot_ldp->edp_magic = 0;
3902 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3903 dir, dot_ldp, dot_dot_ldp);
3905 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3906 * the subseqent ref_add() will increase the dir->i_nlink
3907 * as 3. That is incorrect for new created directory.
3909 * It looks like hack, because we want to make the OSD API
3910 * to be order-independent for new created directory object
3911 * between dt_insert(..) and ref_add() operations.
3913 * Here, we only restore the in-RAM dir-inode's nlink attr,
3914 * becuase if the nlink attr is not 2, then there will be
3915 * ref_add() called following the dt_insert(..), such call
3916 * will make both the in-RAM and on-disk dir-inode's nlink
3917 * attr to be set as 2. LU-7447
3919 set_nlink(dir, saved_nlink);
3924 * Create an local agent inode for remote entry
3926 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3927 struct osd_device *osd,
3928 struct osd_object *pobj,
3929 const struct lu_fid *fid,
3933 struct osd_thread_info *info = osd_oti_get(env);
3934 struct inode *local;
3935 struct osd_thandle *oh;
3936 uid_t own[2] = {0, 0};
3942 oh = container_of(th, struct osd_thandle, ot_super);
3943 LASSERT(oh->ot_handle->h_transaction != NULL);
3945 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode, type, own);
3946 if (IS_ERR(local)) {
3947 CERROR("%s: create local error %d\n", osd_name(osd),
3948 (int)PTR_ERR(local));
3953 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3954 * correct gid on remote file, not agent here
3956 local->i_gid = current_fsgid();
3957 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3959 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
3960 * That gives e2fsck something to look at and be happy, and allows
3961 * debugging if we need to determine where this symlink came from.
3963 if (S_ISLNK(type)) {
3964 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
3965 rc = snprintf((char *)LDISKFS_I(local)->i_data,
3966 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
3968 i_size_write(local, rc);
3969 LDISKFS_I(local)->i_disksize = rc;
3971 unlock_new_inode(local);
3973 /* Agent inode should not have project ID */
3974 #ifdef HAVE_PROJECT_QUOTA
3975 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
3976 i_projid_read(pobj->oo_inode) != 0) {
3977 rc = osd_transfer_project(local, 0, th);
3979 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3980 osd_ino2name(local), rc);
3981 RETURN(ERR_PTR(rc));
3985 /* Set special LMA flag for local agent inode */
3986 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
3988 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
3989 osd_name(osd), PFID(fid), rc);
3990 RETURN(ERR_PTR(rc));
3996 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
3997 lu_object_fid(&pobj->oo_dt.do_lu),
4000 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4001 osd_name(osd), PFID(fid), rc);
4002 RETURN(ERR_PTR(rc));
4009 * when direntry is deleted, we have to take care of possible agent inode
4010 * referenced by that. unfortunately we can't do this at that point:
4011 * iget() within a running transaction leads to deadlock and we better do
4012 * not call that every delete declaration to save performance. so we put
4013 * a potention agent inode on a list and process that once the transaction
4014 * is over. Notice it's not any worse in terms of real orphans as regular
4015 * object destroy doesn't put inodes on the on-disk orphan list. this should
4016 * be addressed separately
4018 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4019 struct osd_thandle *oh,
4022 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4023 struct osd_obj_orphan *oor;
4030 oor->oor_env = (struct lu_env *)env;
4031 spin_lock(&osd->od_osfs_lock);
4032 list_add(&oor->oor_list, &osd->od_orphan_list);
4033 spin_unlock(&osd->od_osfs_lock);
4035 oh->ot_remove_agents = 1;
4041 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4042 struct osd_device *osd)
4044 struct osd_thread_info *info = osd_oti_get(env);
4045 struct osd_obj_orphan *oor, *tmp;
4046 struct osd_inode_id id;
4048 struct inode *inode;
4053 spin_lock(&osd->od_osfs_lock);
4054 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4055 if (oor->oor_env == env)
4056 list_move(&oor->oor_list, &list);
4058 spin_unlock(&osd->od_osfs_lock);
4060 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4064 list_del(&oor->oor_list);
4067 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4068 inode = osd_iget_fid(info, osd, &id, &fid);
4072 if (!osd_remote_fid(env, osd, &fid)) {
4077 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4079 mark_inode_dirty(inode);
4080 ldiskfs_journal_stop(jh);
4088 * OSD layer object create function for OST objects (b=11826).
4090 * The FID is inserted into inode xattr here.
4092 * \retval 0, on success
4093 * \retval -ve, on error
4095 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4096 struct lu_attr *attr, struct dt_allocation_hint *hint,
4097 struct dt_object_format *dof, struct thandle *th)
4099 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4100 struct osd_object *obj = osd_dt_obj(dt);
4101 struct osd_thread_info *info = osd_oti_get(env);
4102 int result, on_ost = 0;
4106 if (dt_object_exists(dt))
4109 LINVRNT(osd_invariant(obj));
4110 LASSERT(!dt_object_remote(dt));
4111 LASSERT(osd_is_write_locked(env, obj));
4112 LASSERT(th != NULL);
4114 if (unlikely(fid_is_acct(fid)))
4116 * Quota files can't be created from the kernel any more,
4117 * 'tune2fs -O quota' will take care of creating them
4121 result = __osd_create(info, obj, attr, hint, dof, th);
4123 if (fid_is_idif(fid) &&
4124 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4125 struct lu_fid *tfid = &info->oti_fid;
4126 struct ost_id *oi = &info->oti_ostid;
4128 fid_to_ostid(fid, oi);
4129 ostid_to_fid(tfid, oi, 0);
4131 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4132 LMAC_FID_ON_OST, 0);
4134 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4136 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4137 on_ost ? LMAC_FID_ON_OST : 0,
4140 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4141 obj->oo_dt.do_body_ops = &osd_body_ops;
4144 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY))
4145 result = __osd_oi_insert(env, obj, fid, th);
4148 * a small optimization - dt_insert() isn't usually applied
4149 * to OST objects, so we don't need to cache OI mapping for
4152 if (result == 0 && on_ost == 0) {
4153 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4155 result = osd_idc_find_and_init(env, osd, obj);
4156 LASSERT(result == 0);
4159 LASSERT(ergo(result == 0,
4160 dt_object_exists(dt) && !dt_object_remote(dt)));
4161 LINVRNT(osd_invariant(obj));
4165 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4166 struct thandle *handle)
4168 struct osd_thandle *oh;
4171 /* it's possible that object doesn't exist yet */
4172 LASSERT(handle != NULL);
4174 oh = container_of(handle, struct osd_thandle, ot_super);
4175 LASSERT(oh->ot_handle == NULL);
4177 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4178 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4180 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4187 * Concurrency: @dt is write locked.
4189 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4192 struct osd_object *obj = osd_dt_obj(dt);
4193 struct inode *inode = obj->oo_inode;
4194 struct osd_thandle *oh;
4197 if (!dt_object_exists(dt) || obj->oo_destroyed)
4200 LINVRNT(osd_invariant(obj));
4201 LASSERT(!dt_object_remote(dt));
4202 LASSERT(osd_is_write_locked(env, obj));
4203 LASSERT(th != NULL);
4205 oh = container_of(th, struct osd_thandle, ot_super);
4206 LASSERT(oh->ot_handle != NULL);
4208 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4210 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4211 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4213 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4214 * (65000) subdirectories by storing "1" in i_nlink if the link count
4215 * would otherwise overflow. Directory tranversal tools understand
4216 * that (st_nlink == 1) indicates that the filesystem dose not track
4217 * hard links count on the directory, and will not abort subdirectory
4218 * scanning early once (st_nlink - 2) subdirs have been found.
4220 * This also has to properly handle the case of inodes with nlink == 0
4221 * in case they are being linked into the PENDING directory
4223 spin_lock(&obj->oo_guard);
4224 if (unlikely(inode->i_nlink == 0))
4225 /* inc_nlink from 0 may cause WARN_ON */
4226 set_nlink(inode, 1);
4228 ldiskfs_inc_count(oh->ot_handle, inode);
4229 if (!S_ISDIR(inode->i_mode))
4230 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4232 spin_unlock(&obj->oo_guard);
4234 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4235 LINVRNT(osd_invariant(obj));
4237 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4242 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4243 struct thandle *handle)
4245 struct osd_thandle *oh;
4247 if (!dt_object_exists(dt))
4250 LASSERT(!dt_object_remote(dt));
4251 LASSERT(handle != NULL);
4253 oh = container_of(handle, struct osd_thandle, ot_super);
4254 LASSERT(oh->ot_handle == NULL);
4256 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4257 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4263 * Concurrency: @dt is write locked.
4265 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4268 struct osd_object *obj = osd_dt_obj(dt);
4269 struct inode *inode = obj->oo_inode;
4270 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4271 struct osd_thandle *oh;
4273 if (!dt_object_exists(dt))
4276 LINVRNT(osd_invariant(obj));
4277 LASSERT(!dt_object_remote(dt));
4278 LASSERT(osd_is_write_locked(env, obj));
4279 LASSERT(th != NULL);
4281 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4284 oh = container_of(th, struct osd_thandle, ot_super);
4285 LASSERT(oh->ot_handle != NULL);
4287 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4289 spin_lock(&obj->oo_guard);
4291 * That can be result of upgrade from old Lustre version and
4292 * applied only to local files. Just skip this ref_del call.
4293 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4295 if (inode->i_nlink == 0) {
4296 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4297 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4298 ", maybe an upgraded file? (LU-3915)\n",
4299 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4300 spin_unlock(&obj->oo_guard);
4304 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4305 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4307 ldiskfs_dec_count(oh->ot_handle, inode);
4308 spin_unlock(&obj->oo_guard);
4310 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4311 LINVRNT(osd_invariant(obj));
4313 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4319 * Concurrency: @dt is read locked.
4321 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4322 struct lu_buf *buf, const char *name)
4324 struct osd_object *obj = osd_dt_obj(dt);
4325 struct inode *inode = obj->oo_inode;
4326 struct osd_thread_info *info = osd_oti_get(env);
4327 struct dentry *dentry = &info->oti_obj_dentry;
4328 bool cache_xattr = false;
4333 /* version get is not real XATTR but uses xattr API */
4334 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4335 dt_obj_version_t *ver = buf->lb_buf;
4338 * for version we are just using xattr API but change inode
4341 if (buf->lb_len == 0)
4342 return sizeof(dt_obj_version_t);
4344 if (buf->lb_len < sizeof(dt_obj_version_t))
4347 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4348 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4350 *ver = LDISKFS_I(inode)->i_fs_version;
4352 return sizeof(dt_obj_version_t);
4355 if (!dt_object_exists(dt))
4358 LASSERT(!dt_object_remote(dt));
4359 LASSERT(inode->i_op != NULL);
4360 #ifdef HAVE_IOP_XATTR
4361 LASSERT(inode->i_op->getxattr != NULL);
4364 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4365 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4369 rc = osd_oxc_get(obj, name, buf);
4374 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4375 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4376 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4377 struct filter_fid *ff;
4378 struct ost_layout *ol;
4380 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4384 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4387 if (buf->lb_len == 0 || !buf->lb_buf)
4390 if (buf->lb_len < rc)
4394 ol = &ff->ff_layout;
4395 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4396 PFID_STRIPE_IDX_BITS);
4397 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4398 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4399 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4400 if (lma->lma_compat & LMAC_COMP_INFO) {
4401 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4402 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4403 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4405 ol->ol_comp_start = 0;
4406 ol->ol_comp_end = 0;
4410 rc = __osd_xattr_get(inode, dentry, name,
4411 buf->lb_buf, buf->lb_len);
4415 if (rc == -ENOENT || rc == -ENODATA)
4416 osd_oxc_add(obj, name, NULL, 0);
4417 else if (rc > 0 && buf->lb_buf != NULL)
4418 osd_oxc_add(obj, name, buf->lb_buf, rc);
4424 static int osd_declare_xattr_set(const struct lu_env *env,
4425 struct dt_object *dt,
4426 const struct lu_buf *buf, const char *name,
4427 int fl, struct thandle *handle)
4429 struct osd_thandle *oh;
4431 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4433 LASSERT(handle != NULL);
4435 oh = container_of(handle, struct osd_thandle, ot_super);
4436 LASSERT(oh->ot_handle == NULL);
4438 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4440 * For non-upgrading case, the LMA is set first and
4441 * usually fit inode. But for upgrade case, the LMA
4442 * may be in another separated EA block.
4444 if (dt_object_exists(dt)) {
4445 if (fl == LU_XATTR_REPLACE)
4450 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4452 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4453 /* We may need to delete the old PFID EA. */
4454 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4455 if (fl == LU_XATTR_REPLACE)
4461 * If some name entry resides on remote MDT, then will create
4462 * agent entry under remote parent. On the other hand, if the
4463 * remote entry will be removed, then related agent entry may
4464 * need to be removed from the remote parent. So there may be
4465 * kinds of cases, let's declare enough credits. The credits
4466 * for create agent entry is enough for remove case.
4468 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4469 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4470 if (dt_object_exists(dt))
4471 credits += 1; /* For updating LMA */
4475 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4480 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4482 * learn xattr size from osd_xattr_get if
4483 * attribute has not been read yet
4485 buflen = __osd_xattr_get(
4486 osd_dt_obj(dt)->oo_inode,
4487 &osd_oti_get(env)->oti_obj_dentry,
4492 buflen = buf->lb_len;
4495 if (buflen > sb->s_blocksize) {
4496 credits += osd_calc_bkmap_credits(
4498 (buflen + sb->s_blocksize - 1) >>
4499 sb->s_blocksize_bits);
4503 * xattr set may involve inode quota change, reserve credits for
4504 * dquot_initialize()
4506 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4509 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4514 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4515 const struct lu_buf *buf, int fl,
4516 struct thandle *handle)
4518 struct osd_thread_info *info = osd_oti_get(env);
4519 struct dentry *dentry = &info->oti_obj_dentry;
4520 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4521 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4522 struct inode *inode = obj->oo_inode;
4523 struct filter_fid *ff = buf->lb_buf;
4524 struct ost_layout *ol = &ff->ff_layout;
4525 int flags = XATTR_REPLACE;
4530 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4533 rc = osd_get_lma(info, inode, dentry, loa);
4534 if (rc == -ENODATA) {
4535 /* Usually for upgarding from old device */
4536 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4537 LMAC_FID_ON_OST, 0);
4538 flags = XATTR_CREATE;
4543 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4544 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4547 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4548 /* Separate PFID EA from LMA */
4549 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4550 lustre_lma_swab(lma);
4551 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4552 sizeof(*lma), XATTR_REPLACE);
4554 obj->oo_pfid_in_lma = 0;
4555 rc = LU_XATTR_CREATE;
4561 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4565 * Old client does not send stripe information,
4566 * then store the PFID EA on disk separatedly.
4568 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4569 ol->ol_stripe_size == 0))
4572 /* Remove old PFID EA entry firstly. */
4573 dquot_initialize(inode);
4574 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4575 if (rc == -ENODATA) {
4576 if ((fl & LU_XATTR_REPLACE) && !(fl & LU_XATTR_CREATE))
4583 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4584 if (likely(ol->ol_stripe_size != 0)) {
4585 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4586 PFID_STRIPE_IDX_BITS;
4587 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4588 lma->lma_compat |= LMAC_STRIPE_INFO;
4589 if (ol->ol_comp_id != 0) {
4590 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4591 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4592 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4593 lma->lma_compat |= LMAC_COMP_INFO;
4597 lustre_loa_swab(loa, false);
4599 /* Store the PFID EA inside LMA. */
4600 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4603 obj->oo_pfid_in_lma = 1;
4609 * In DNE environment, the object (in spite of regular file or directory)
4610 * and its name entry may reside on different MDTs. Under such case, we will
4611 * create an agent entry on the MDT where the object resides. The agent entry
4612 * references the object locally, that makes the object to be visible to the
4613 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4614 * such as 'tar' can handle the object properly.
4616 * We handle the agent entry during set linkEA that is the common interface
4617 * for both regular file and directroy, can handle kinds of cases, such as
4618 * create/link/unlink/rename, and so on.
4620 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4622 * XXX: There are two known issues:
4623 * 1. For one object, we will create at most one agent entry even if there
4624 * may be more than one cross-MDTs hard links on the object. So the local
4625 * e2fsck may claim that the object's nlink is larger than the name entries
4626 * that reference such inode. And in further, the e2fsck will fix the nlink
4627 * attribute to match the local references. Then it will cause the object's
4628 * nlink attribute to be inconsistent with the global references. it is bad
4629 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4630 * other hand, the global namespace LFSCK can repair the object's attribute
4631 * according to the linkEA.
4632 * 2. There may be too many hard links on the object as to its linkEA overflow,
4633 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4634 * case happened, then at this point, we do not know whether there are some
4635 * cross-MDTs reference. But there are local references, it guarantees that
4636 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4638 static int osd_xattr_handle_linkea(const struct lu_env *env,
4639 struct osd_device *osd,
4640 struct osd_object *obj,
4641 const struct lu_buf *buf,
4642 struct thandle *handle)
4644 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4645 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4646 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4647 struct lu_name tmpname;
4648 struct osd_thandle *oh;
4650 bool remote = false;
4654 oh = container_of(handle, struct osd_thandle, ot_super);
4655 LASSERT(oh->ot_handle != NULL);
4657 rc = linkea_init_with_rec(&ldata);
4659 linkea_first_entry(&ldata);
4660 while (ldata.ld_lee != NULL && !remote) {
4661 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4663 if (osd_remote_fid(env, osd, tfid) > 0)
4666 linkea_next_entry(&ldata);
4668 } else if (rc == -ENODATA) {
4674 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4675 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4677 CERROR("%s: failed to remove agent entry for "DFID
4678 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4679 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4680 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4682 CERROR("%s: failed to create agent entry for "DFID
4683 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4690 * Concurrency: @dt is write locked.
4692 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4693 const struct lu_buf *buf, const char *name, int fl,
4694 struct thandle *handle)
4696 struct osd_object *obj = osd_dt_obj(dt);
4697 struct osd_device *osd = osd_obj2dev(obj);
4698 struct inode *inode = obj->oo_inode;
4699 struct osd_thread_info *info = osd_oti_get(env);
4709 /* version set is not real XATTR */
4710 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4711 dt_obj_version_t *version = buf->lb_buf;
4714 * for version we are just using xattr API but change inode
4717 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4719 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4720 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4722 LDISKFS_I(inode)->i_fs_version = *version;
4724 * Version is set after all inode operations are finished,
4725 * so we should mark it dirty here
4727 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4732 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4733 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4736 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4739 * For the OST device with 256 bytes inode size by default,
4740 * the PFID EA will be stored together with LMA EA to avoid
4741 * performance trouble. Otherwise the PFID EA can be stored
4742 * independently. LU-8998
4744 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4745 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4746 LASSERT(buf->lb_buf);
4748 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4751 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4752 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4753 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4755 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4759 lma->lma_incompat |= LMAI_STRIPED;
4760 lustre_lma_swab(lma);
4761 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4762 sizeof(*lma), XATTR_REPLACE);
4765 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4766 LASSERT(!osd->od_is_ost);
4768 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4773 if (fl & LU_XATTR_REPLACE)
4774 fs_flags |= XATTR_REPLACE;
4776 if (fl & LU_XATTR_CREATE)
4777 fs_flags |= XATTR_CREATE;
4779 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4780 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4783 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4784 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4785 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4791 * Concurrency: @dt is read locked.
4793 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4794 const struct lu_buf *buf)
4796 struct osd_object *obj = osd_dt_obj(dt);
4797 struct inode *inode = obj->oo_inode;
4798 struct osd_thread_info *info = osd_oti_get(env);
4799 struct dentry *dentry = &info->oti_obj_dentry;
4801 if (!dt_object_exists(dt))
4804 LASSERT(!dt_object_remote(dt));
4805 LASSERT(inode->i_op != NULL);
4806 LASSERT(inode->i_op->listxattr != NULL);
4808 dentry->d_inode = inode;
4809 dentry->d_sb = inode->i_sb;
4810 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4813 static int osd_declare_xattr_del(const struct lu_env *env,
4814 struct dt_object *dt, const char *name,
4815 struct thandle *handle)
4817 struct osd_thandle *oh;
4818 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4820 LASSERT(!dt_object_remote(dt));
4821 LASSERT(handle != NULL);
4823 oh = container_of(handle, struct osd_thandle, ot_super);
4824 LASSERT(oh->ot_handle == NULL);
4826 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4827 osd_dto_credits_noquota[DTO_XATTR_SET]);
4829 * xattr del may involve inode quota change, reserve credits for
4830 * dquot_initialize()
4832 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4838 * Concurrency: @dt is write locked.
4840 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4841 const char *name, struct thandle *handle)
4843 struct osd_object *obj = osd_dt_obj(dt);
4844 struct inode *inode = obj->oo_inode;
4845 struct osd_thread_info *info = osd_oti_get(env);
4846 struct dentry *dentry = &info->oti_obj_dentry;
4849 if (!dt_object_exists(dt))
4852 LASSERT(!dt_object_remote(dt));
4853 LASSERT(inode->i_op != NULL);
4854 LASSERT(handle != NULL);
4855 #ifdef HAVE_IOP_XATTR
4856 LASSERT(inode->i_op->removexattr != NULL);
4859 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4861 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4862 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4864 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4865 &info->oti_ost_attrs);
4867 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4869 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4870 lustre_lma_swab(lma);
4871 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4872 sizeof(*lma), XATTR_REPLACE);
4874 obj->oo_pfid_in_lma = 0;
4877 dquot_initialize(inode);
4878 dentry->d_inode = inode;
4879 dentry->d_sb = inode->i_sb;
4880 rc = ll_vfs_removexattr(dentry, inode, name);
4883 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4886 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4887 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4888 osd_oxc_del(obj, name);
4893 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4894 __u64 start, __u64 end)
4896 struct osd_object *obj = osd_dt_obj(dt);
4897 struct inode *inode = obj->oo_inode;
4898 struct file *file = osd_quasi_file(env, inode);
4903 rc = vfs_fsync_range(file, start, end, 0);
4908 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4917 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4918 const struct dt_index_features *feat)
4920 struct iam_descr *descr;
4922 if (osd_object_is_root(o))
4923 return feat == &dt_directory_features;
4925 LASSERT(o->oo_dir != NULL);
4927 descr = o->oo_dir->od_container.ic_descr;
4928 if (feat == &dt_directory_features) {
4929 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4934 return feat->dif_keysize_min <= descr->id_key_size &&
4935 descr->id_key_size <= feat->dif_keysize_max &&
4936 feat->dif_recsize_min <= descr->id_rec_size &&
4937 descr->id_rec_size <= feat->dif_recsize_max &&
4938 !(feat->dif_flags & (DT_IND_VARKEY |
4939 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4940 ergo(feat->dif_flags & DT_IND_UPDATE,
4941 1 /* XXX check that object (and fs) is writable */);
4945 static int osd_iam_container_init(const struct lu_env *env,
4946 struct osd_object *obj,
4947 struct osd_directory *dir)
4949 struct iam_container *bag = &dir->od_container;
4952 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
4956 result = iam_container_setup(bag);
4958 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
4960 iam_container_fini(bag);
4967 * Concurrency: no external locking is necessary.
4969 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
4970 const struct dt_index_features *feat)
4974 struct osd_object *obj = osd_dt_obj(dt);
4976 LINVRNT(osd_invariant(obj));
4978 if (osd_object_is_root(obj)) {
4979 dt->do_index_ops = &osd_index_ea_ops;
4981 } else if (feat == &dt_directory_features) {
4982 dt->do_index_ops = &osd_index_ea_ops;
4983 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
4988 } else if (unlikely(feat == &dt_otable_features)) {
4989 dt->do_index_ops = &osd_otable_ops;
4991 } else if (unlikely(feat == &dt_acct_features)) {
4992 dt->do_index_ops = &osd_acct_index_ops;
4995 } else if (!osd_has_index(obj)) {
4996 struct osd_directory *dir;
4997 struct osd_device *osd = osd_obj2dev(obj);
4998 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5003 spin_lock(&obj->oo_guard);
5004 if (obj->oo_dir == NULL)
5008 * Concurrent thread allocated container data.
5011 spin_unlock(&obj->oo_guard);
5013 * Now, that we have container data, serialize its
5016 down_write(&obj->oo_ext_idx_sem);
5018 * recheck under lock.
5021 if (osd_has_index(obj)) {
5026 result = osd_iam_container_init(env, obj, obj->oo_dir);
5027 if (result || feat == &dt_lfsck_namespace_features ||
5028 feat == &dt_lfsck_layout_orphan_features ||
5029 feat == &dt_lfsck_layout_dangling_features)
5032 result = osd_index_register(osd, fid,
5033 feat->dif_keysize_max,
5034 feat->dif_recsize_max);
5036 CWARN("%s: failed to register index "
5038 osd_name(osd), PFID(fid), result);
5039 else if (result > 0)
5042 CDEBUG(D_LFSCK, "%s: index object "DFID
5043 " (%d/%d) registered\n",
5044 osd_name(osd), PFID(fid),
5045 (int)feat->dif_keysize_max,
5046 (int)feat->dif_recsize_max);
5049 up_write(&obj->oo_ext_idx_sem);
5057 if (result == 0 && skip_iam == 0) {
5058 if (!osd_iam_index_probe(env, obj, feat))
5061 LINVRNT(osd_invariant(obj));
5066 static int osd_otable_it_attr_get(const struct lu_env *env,
5067 struct dt_object *dt,
5068 struct lu_attr *attr)
5074 static const struct dt_object_operations osd_obj_ops = {
5075 .do_read_lock = osd_read_lock,
5076 .do_write_lock = osd_write_lock,
5077 .do_read_unlock = osd_read_unlock,
5078 .do_write_unlock = osd_write_unlock,
5079 .do_write_locked = osd_write_locked,
5080 .do_attr_get = osd_attr_get,
5081 .do_declare_attr_set = osd_declare_attr_set,
5082 .do_attr_set = osd_attr_set,
5083 .do_ah_init = osd_ah_init,
5084 .do_declare_create = osd_declare_create,
5085 .do_create = osd_create,
5086 .do_declare_destroy = osd_declare_destroy,
5087 .do_destroy = osd_destroy,
5088 .do_index_try = osd_index_try,
5089 .do_declare_ref_add = osd_declare_ref_add,
5090 .do_ref_add = osd_ref_add,
5091 .do_declare_ref_del = osd_declare_ref_del,
5092 .do_ref_del = osd_ref_del,
5093 .do_xattr_get = osd_xattr_get,
5094 .do_declare_xattr_set = osd_declare_xattr_set,
5095 .do_xattr_set = osd_xattr_set,
5096 .do_declare_xattr_del = osd_declare_xattr_del,
5097 .do_xattr_del = osd_xattr_del,
5098 .do_xattr_list = osd_xattr_list,
5099 .do_object_sync = osd_object_sync,
5100 .do_invalidate = osd_invalidate,
5103 static const struct dt_object_operations osd_obj_otable_it_ops = {
5104 .do_attr_get = osd_otable_it_attr_get,
5105 .do_index_try = osd_index_try,
5108 static int osd_index_declare_iam_delete(const struct lu_env *env,
5109 struct dt_object *dt,
5110 const struct dt_key *key,
5111 struct thandle *handle)
5113 struct osd_thandle *oh;
5115 oh = container_of(handle, struct osd_thandle, ot_super);
5116 LASSERT(oh->ot_handle == NULL);
5118 /* Recycle may cause additional three blocks to be changed. */
5119 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5120 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5126 * delete a (key, value) pair from index \a dt specified by \a key
5128 * \param dt osd index object
5129 * \param key key for index
5130 * \param rec record reference
5131 * \param handle transaction handler
5134 * \retval -ve failure
5136 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5137 const struct dt_key *key,
5138 struct thandle *handle)
5140 struct osd_thread_info *oti = osd_oti_get(env);
5141 struct osd_object *obj = osd_dt_obj(dt);
5142 struct osd_thandle *oh;
5143 struct iam_path_descr *ipd;
5144 struct iam_container *bag = &obj->oo_dir->od_container;
5149 if (!dt_object_exists(dt))
5152 LINVRNT(osd_invariant(obj));
5153 LASSERT(!dt_object_remote(dt));
5154 LASSERT(bag->ic_object == obj->oo_inode);
5155 LASSERT(handle != NULL);
5157 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5159 ipd = osd_idx_ipd_get(env, bag);
5160 if (unlikely(ipd == NULL))
5163 oh = container_of(handle, struct osd_thandle, ot_super);
5164 LASSERT(oh->ot_handle != NULL);
5165 LASSERT(oh->ot_handle->h_transaction != NULL);
5167 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5168 /* swab quota uid/gid provided by caller */
5169 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5170 key = (const struct dt_key *)&oti->oti_quota_id;
5173 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5174 osd_ipd_put(env, bag, ipd);
5175 LINVRNT(osd_invariant(obj));
5176 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5180 static int osd_index_declare_ea_delete(const struct lu_env *env,
5181 struct dt_object *dt,
5182 const struct dt_key *key,
5183 struct thandle *handle)
5185 struct osd_thandle *oh;
5186 struct inode *inode;
5191 LASSERT(!dt_object_remote(dt));
5192 LASSERT(handle != NULL);
5194 oh = container_of(handle, struct osd_thandle, ot_super);
5195 LASSERT(oh->ot_handle == NULL);
5197 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5198 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5200 inode = osd_dt_obj(dt)->oo_inode;
5204 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5205 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5210 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5213 struct osd_fid_pack *rec;
5216 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5217 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5218 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5219 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5225 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5226 const struct lu_fid *fid)
5228 struct seq_server_site *ss = osd_seq_site(osd);
5232 /* FID seqs not in FLDB, must be local seq */
5233 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5237 * If FLD is not being initialized yet, it only happens during the
5238 * initialization, likely during mgs initialization, and we assume
5239 * this is local FID.
5241 if (ss == NULL || ss->ss_server_fld == NULL)
5244 /* Only check the local FLDB here */
5245 if (osd_seq_exists(env, osd, fid_seq(fid)))
5251 static void osd_take_care_of_agent(const struct lu_env *env,
5252 struct osd_device *osd,
5253 struct osd_thandle *oh,
5254 struct ldiskfs_dir_entry_2 *de)
5256 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5257 struct osd_idmap_cache *idc;
5258 int rc, schedule = 0;
5260 LASSERT(de != NULL);
5262 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5263 if (likely(rc == 0)) {
5264 idc = osd_idc_find_or_init(env, osd, fid);
5265 if (IS_ERR(idc) || idc->oic_remote)
5267 } else if (rc == -ENODATA) {
5269 * can't get FID, postpone to the end of the
5270 * transaction when iget() is safe
5274 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5277 osd_schedule_agent_inode_removal(env, oh,
5278 le32_to_cpu(de->inode));
5282 * Index delete function for interoperability mode (b11826).
5283 * It will remove the directory entry added by osd_index_ea_insert().
5284 * This entry is needed to maintain name->fid mapping.
5286 * \param key, key i.e. file entry to be deleted
5288 * \retval 0, on success
5289 * \retval -ve, on error
5291 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5292 const struct dt_key *key, struct thandle *handle)
5294 struct osd_object *obj = osd_dt_obj(dt);
5295 struct inode *dir = obj->oo_inode;
5296 struct dentry *dentry;
5297 struct osd_thandle *oh;
5298 struct ldiskfs_dir_entry_2 *de = NULL;
5299 struct buffer_head *bh;
5300 struct htree_lock *hlock = NULL;
5301 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5306 if (!dt_object_exists(dt))
5309 LINVRNT(osd_invariant(obj));
5310 LASSERT(!dt_object_remote(dt));
5311 LASSERT(handle != NULL);
5313 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5315 oh = container_of(handle, struct osd_thandle, ot_super);
5316 LASSERT(oh->ot_handle != NULL);
5317 LASSERT(oh->ot_handle->h_transaction != NULL);
5319 dquot_initialize(dir);
5320 dentry = osd_child_dentry_get(env, obj,
5321 (char *)key, strlen((char *)key));
5323 if (obj->oo_hl_head != NULL) {
5324 hlock = osd_oti_get(env)->oti_hlock;
5325 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5326 dir, LDISKFS_HLOCK_DEL);
5328 down_write(&obj->oo_ext_idx_sem);
5331 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5334 * If this is not the ".." entry, it might be a remote DNE
5335 * entry and we need to check if the FID is for a remote
5336 * MDT. If the FID is not in the directory entry (e.g.
5337 * upgraded 1.8 filesystem without dirdata enabled) then
5338 * we need to get the FID from the LMA. For a remote directory
5339 * there HAS to be an LMA, it cannot be an IGIF inode in this
5342 * Delete the entry before the agent inode in order to
5343 * simplify error handling. At worst an error after deleting
5344 * the entry first might leak the agent inode afterward. The
5345 * reverse would need filesystem abort in case of error deleting
5346 * the entry after the agent had been removed, or leave a
5347 * dangling entry pointing at a random inode.
5349 if (strcmp((char *)key, dotdot) != 0)
5350 osd_take_care_of_agent(env, osd, oh, de);
5351 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5357 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5358 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5359 /* NB, dirent count may not be accurate, because it's counted
5362 if (obj->oo_dirent_count)
5363 obj->oo_dirent_count--;
5365 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5368 ldiskfs_htree_unlock(hlock);
5370 up_write(&obj->oo_ext_idx_sem);
5373 LASSERT(osd_invariant(obj));
5374 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5379 * Lookup index for \a key and copy record to \a rec.
5381 * \param dt osd index object
5382 * \param key key for index
5383 * \param rec record reference
5385 * \retval +ve success : exact mach
5386 * \retval 0 return record with key not greater than \a key
5387 * \retval -ve failure
5389 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5390 struct dt_rec *rec, const struct dt_key *key)
5392 struct osd_object *obj = osd_dt_obj(dt);
5393 struct iam_path_descr *ipd;
5394 struct iam_container *bag = &obj->oo_dir->od_container;
5395 struct osd_thread_info *oti = osd_oti_get(env);
5396 struct iam_iterator *it = &oti->oti_idx_it;
5397 struct iam_rec *iam_rec;
5402 if (!dt_object_exists(dt))
5405 LASSERT(osd_invariant(obj));
5406 LASSERT(!dt_object_remote(dt));
5407 LASSERT(bag->ic_object == obj->oo_inode);
5409 ipd = osd_idx_ipd_get(env, bag);
5413 /* got ipd now we can start iterator. */
5414 iam_it_init(it, bag, 0, ipd);
5416 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5417 /* swab quota uid/gid provided by caller */
5418 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5419 key = (const struct dt_key *)&oti->oti_quota_id;
5422 rc = iam_it_get(it, (struct iam_key *)key);
5424 if (S_ISDIR(obj->oo_inode->i_mode))
5425 iam_rec = (struct iam_rec *)oti->oti_ldp;
5427 iam_rec = (struct iam_rec *)rec;
5429 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5431 if (S_ISDIR(obj->oo_inode->i_mode))
5432 osd_fid_unpack((struct lu_fid *)rec,
5433 (struct osd_fid_pack *)iam_rec);
5434 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5435 osd_quota_unpack(obj, rec);
5440 osd_ipd_put(env, bag, ipd);
5442 LINVRNT(osd_invariant(obj));
5447 static int osd_index_declare_iam_insert(const struct lu_env *env,
5448 struct dt_object *dt,
5449 const struct dt_rec *rec,
5450 const struct dt_key *key,
5451 struct thandle *handle)
5453 struct osd_thandle *oh;
5455 LASSERT(handle != NULL);
5457 oh = container_of(handle, struct osd_thandle, ot_super);
5458 LASSERT(oh->ot_handle == NULL);
5460 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5461 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5467 * Inserts (key, value) pair in \a dt index object.
5469 * \param dt osd index object
5470 * \param key key for index
5471 * \param rec record reference
5472 * \param th transaction handler
5475 * \retval -ve failure
5477 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5478 const struct dt_rec *rec,
5479 const struct dt_key *key, struct thandle *th)
5481 struct osd_object *obj = osd_dt_obj(dt);
5482 struct iam_path_descr *ipd;
5483 struct osd_thandle *oh;
5484 struct iam_container *bag;
5485 struct osd_thread_info *oti = osd_oti_get(env);
5486 struct iam_rec *iam_rec;
5491 if (!dt_object_exists(dt))
5494 LINVRNT(osd_invariant(obj));
5495 LASSERT(!dt_object_remote(dt));
5497 bag = &obj->oo_dir->od_container;
5498 LASSERT(bag->ic_object == obj->oo_inode);
5499 LASSERT(th != NULL);
5501 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5503 ipd = osd_idx_ipd_get(env, bag);
5504 if (unlikely(ipd == NULL))
5507 oh = container_of(th, struct osd_thandle, ot_super);
5508 LASSERT(oh->ot_handle != NULL);
5509 LASSERT(oh->ot_handle->h_transaction != NULL);
5510 if (S_ISDIR(obj->oo_inode->i_mode)) {
5511 iam_rec = (struct iam_rec *)oti->oti_ldp;
5512 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5514 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5515 /* pack quota uid/gid */
5516 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5517 key = (const struct dt_key *)&oti->oti_quota_id;
5518 /* pack quota record */
5519 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5520 iam_rec = (struct iam_rec *)rec;
5522 iam_rec = (struct iam_rec *)rec;
5525 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5527 osd_ipd_put(env, bag, ipd);
5528 LINVRNT(osd_invariant(obj));
5529 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5534 * Calls ldiskfs_add_entry() to add directory entry
5535 * into the directory. This is required for
5536 * interoperability mode (b11826)
5538 * \retval 0, on success
5539 * \retval -ve, on error
5541 static int __osd_ea_add_rec(struct osd_thread_info *info,
5542 struct osd_object *pobj, struct inode *cinode,
5543 const char *name, const struct lu_fid *fid,
5544 struct htree_lock *hlock, struct thandle *th)
5546 struct ldiskfs_dentry_param *ldp;
5547 struct dentry *child;
5548 struct osd_thandle *oth;
5551 oth = container_of(th, struct osd_thandle, ot_super);
5552 LASSERT(oth->ot_handle != NULL);
5553 LASSERT(oth->ot_handle->h_transaction != NULL);
5554 LASSERT(pobj->oo_inode);
5556 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5557 if (unlikely(pobj->oo_inode ==
5558 osd_sb(osd_obj2dev(pobj))->s_root->d_inode))
5561 osd_get_ldiskfs_dirent_param(ldp, fid);
5562 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5563 child->d_fsdata = (void *)ldp;
5564 dquot_initialize(pobj->oo_inode);
5565 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5566 child, cinode, hlock);
5567 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5568 struct ldiskfs_dir_entry_2 *de;
5569 struct buffer_head *bh;
5572 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5575 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5578 if (S_ISDIR(cinode->i_mode))
5579 de->file_type = LDISKFS_DIRENT_LUFID |
5580 LDISKFS_FT_REG_FILE;
5582 de->file_type = LDISKFS_DIRENT_LUFID |
5584 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5595 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5596 * into the directory.Also sets flags into osd object to
5597 * indicate dot and dotdot are created. This is required for
5598 * interoperability mode (b11826)
5600 * \param dir directory for dot and dotdot fixup.
5601 * \param obj child object for linking
5603 * \retval 0, on success
5604 * \retval -ve, on error
5606 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5607 struct osd_object *dir,
5608 struct inode *parent_dir, const char *name,
5609 const struct lu_fid *dot_fid,
5610 const struct lu_fid *dot_dot_fid,
5613 struct inode *inode = dir->oo_inode;
5614 struct osd_thandle *oth;
5617 oth = container_of(th, struct osd_thandle, ot_super);
5618 LASSERT(oth->ot_handle->h_transaction != NULL);
5619 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5621 if (strcmp(name, dot) == 0) {
5622 if (dir->oo_compat_dot_created) {
5625 LASSERT(inode->i_ino == parent_dir->i_ino);
5626 dir->oo_compat_dot_created = 1;
5629 } else if (strcmp(name, dotdot) == 0) {
5630 if (!dir->oo_compat_dot_created)
5632 /* in case of rename, dotdot is already created */
5633 if (dir->oo_compat_dotdot_created) {
5634 return __osd_ea_add_rec(info, dir, parent_dir, name,
5635 dot_dot_fid, NULL, th);
5638 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5639 struct lu_fid tfid = *dot_dot_fid;
5642 result = osd_add_dot_dotdot_internal(info,
5643 dir->oo_inode, parent_dir, dot_fid,
5646 result = osd_add_dot_dotdot_internal(info,
5647 dir->oo_inode, parent_dir, dot_fid,
5652 dir->oo_compat_dotdot_created = 1;
5660 * It will call the appropriate osd_add* function and return the
5661 * value, return by respective functions.
5663 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5664 struct inode *cinode, const char *name,
5665 const struct lu_fid *fid, struct thandle *th)
5667 struct osd_thread_info *info = osd_oti_get(env);
5668 struct htree_lock *hlock;
5671 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5673 if (name[0] == '.' && (name[1] == '\0' ||
5674 (name[1] == '.' && name[2] == '\0'))) {
5675 if (hlock != NULL) {
5676 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5679 down_write(&pobj->oo_ext_idx_sem);
5682 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5683 lu_object_fid(&pobj->oo_dt.do_lu),
5686 if (hlock != NULL) {
5687 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5688 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5690 down_write(&pobj->oo_ext_idx_sem);
5693 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5694 struct lu_fid *tfid = &info->oti_fid;
5698 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5701 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5705 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
5706 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
5707 pobj->oo_dirent_count++;
5710 ldiskfs_htree_unlock(hlock);
5712 up_write(&pobj->oo_ext_idx_sem);
5718 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5719 struct osd_idmap_cache *oic)
5721 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5722 struct lu_fid *fid = &oic->oic_fid;
5723 struct osd_inode_id *id = &oic->oic_lid;
5724 struct inode *inode = NULL;
5731 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5734 if (scrub->os_running && scrub->os_pos_current > id->oii_ino)
5737 if (dev->od_auto_scrub_interval == AS_NEVER ||
5738 ktime_get_real_seconds() <
5739 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5743 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5744 if (rc == -ENOENT) {
5745 __u32 gen = id->oii_gen;
5751 inode = osd_iget(oti, dev, id);
5752 /* The inode has been removed (by race maybe). */
5753 if (IS_ERR(inode)) {
5754 rc = PTR_ERR(inode);
5756 RETURN(rc == -ESTALE ? -ENOENT : rc);
5759 /* The OI mapping is lost. */
5760 if (gen != OSD_OII_NOGEN)
5764 * The inode may has been reused by others, we do not know,
5765 * leave it to be handled by subsequent osd_fid_lookup().
5768 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5775 if (scrub->os_running) {
5776 if (inode == NULL) {
5777 inode = osd_iget(oti, dev, id);
5778 /* The inode has been removed (by race maybe). */
5779 if (IS_ERR(inode)) {
5780 rc = PTR_ERR(inode);
5782 RETURN(rc == -ESTALE ? -ENOENT : rc);
5786 rc = osd_oii_insert(dev, oic, insert);
5788 * There is race condition between osd_oi_lookup and OI scrub.
5789 * The OI scrub finished just after osd_oi_lookup() failure.
5790 * Under such case, it is unnecessary to trigger OI scrub again,
5791 * but try to call osd_oi_lookup() again.
5793 if (unlikely(rc == -EAGAIN))
5796 if (!S_ISDIR(inode->i_mode))
5799 rc = osd_check_lmv(oti, dev, inode, oic);
5804 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5805 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5806 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5807 CDEBUG(D_LFSCK | D_CONSOLE | D_WARNING,
5808 "%s: trigger partial OI scrub for RPC inconsistency checking FID "DFID": rc = %d\n",
5809 osd_dev2name(dev), PFID(fid), rc);
5810 if (rc == 0 || rc == -EALREADY)
5823 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5824 struct osd_device *dev,
5825 struct osd_idmap_cache *oic,
5826 struct lu_fid *fid, __u32 ino)
5828 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5829 struct inode *inode;
5832 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5833 inode = osd_iget(oti, dev, &oic->oic_lid);
5834 if (IS_ERR(inode)) {
5835 fid_zero(&oic->oic_fid);
5836 return PTR_ERR(inode);
5839 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5842 fid_zero(&oic->oic_fid);
5844 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5848 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5849 struct osd_inode_id *id, const struct lu_fid *fid)
5851 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5852 id->oii_ino, id->oii_gen, info);
5853 info->oti_cache.oic_lid = *id;
5854 info->oti_cache.oic_fid = *fid;
5855 info->oti_cache.oic_dev = osd;
5859 * Get parent FID from the linkEA.
5861 * For a directory which parent resides on remote MDT, to satisfy the
5862 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5863 * the other hand, to make the lookup(..) on the directory can return
5864 * the real parent FID, we append the real parent FID after its ".."
5865 * name entry in the /REMOTE_PARENT_DIR.
5867 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5868 * backup. So after the restore, we cannot get the right parent FID from
5869 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5870 * we have stored the real parent FID in the directory object's linkEA,
5871 * we can parse the linkEA for the real parent FID.
5873 * \param[in] env pointer to the thread context
5874 * \param[in] obj pointer to the object to be handled
5875 * \param[out]fid pointer to the buffer to hold the parent FID
5877 * \retval 0 for getting the real parent FID successfully
5878 * \retval negative error number on failure
5880 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5881 struct osd_object *obj,
5884 struct osd_thread_info *oti = osd_oti_get(env);
5885 struct lu_buf *buf = &oti->oti_big_buf;
5886 struct dentry *dentry = &oti->oti_obj_dentry;
5887 struct inode *inode = obj->oo_inode;
5888 struct linkea_data ldata = { NULL };
5894 if (!S_ISDIR(inode->i_mode))
5898 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5899 buf->lb_buf, buf->lb_len);
5900 if (rc == -ERANGE) {
5901 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5904 lu_buf_realloc(buf, rc);
5905 if (buf->lb_buf == NULL)
5912 if (unlikely(rc == 0))
5918 if (unlikely(buf->lb_buf == NULL)) {
5919 lu_buf_realloc(buf, rc);
5920 if (buf->lb_buf == NULL)
5927 rc = linkea_init_with_rec(&ldata);
5929 linkea_first_entry(&ldata);
5930 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5936 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5937 struct inode *inode,
5938 const struct lu_fid *pfid,
5939 const char *name, const int namelen)
5941 struct osd_thread_info *oti = osd_oti_get(env);
5942 struct lu_buf *buf = &oti->oti_big_buf;
5943 struct dentry *dentry = &oti->oti_obj_dentry;
5944 struct linkea_data ldata = { NULL };
5945 struct lu_name cname = { .ln_name = name,
5946 .ln_namelen = namelen };
5952 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5953 buf->lb_buf, buf->lb_len);
5955 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
5960 if (unlikely(rc == 0))
5963 if (buf->lb_len < rc) {
5964 lu_buf_realloc(buf, rc);
5965 if (buf->lb_buf == NULL)
5972 rc = linkea_init_with_rec(&ldata);
5974 rc = linkea_links_find(&ldata, &cname, pfid);
5980 * Calls ->lookup() to find dentry. From dentry get inode and
5981 * read inode's ea to get fid. This is required for interoperability
5984 * \retval 0, on success
5985 * \retval -ve, on error
5987 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
5988 struct dt_rec *rec, const struct dt_key *key)
5990 struct inode *dir = obj->oo_inode;
5991 struct dentry *dentry;
5992 struct ldiskfs_dir_entry_2 *de;
5993 struct buffer_head *bh;
5994 struct lu_fid *fid = (struct lu_fid *)rec;
5995 struct htree_lock *hlock = NULL;
6001 LASSERT(dir->i_op != NULL);
6002 LASSERT(dir->i_op->lookup != NULL);
6004 dentry = osd_child_dentry_get(env, obj,
6005 (char *)key, strlen((char *)key));
6007 if (obj->oo_hl_head != NULL) {
6008 hlock = osd_oti_get(env)->oti_hlock;
6009 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6010 dir, LDISKFS_HLOCK_LOOKUP);
6012 down_read(&obj->oo_ext_idx_sem);
6015 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6017 struct osd_thread_info *oti = osd_oti_get(env);
6018 struct osd_inode_id *id = &oti->oti_id;
6019 struct osd_idmap_cache *oic = &oti->oti_cache;
6020 struct osd_device *dev = osd_obj2dev(obj);
6022 ino = le32_to_cpu(de->inode);
6023 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6025 rc = osd_fail_fid_lookup(oti, dev, oic, fid, ino);
6029 rc = osd_get_fid_from_dentry(de, rec);
6031 /* done with de, release bh */
6034 if (unlikely(is_remote_parent_ino(dev, ino))) {
6035 const char *name = (const char *)key;
6038 * If the parent is on remote MDT, and there
6039 * is no FID-in-dirent, then we have to get
6040 * the parent FID from the linkEA.
6042 if (likely(strlen(name) == 2 &&
6043 name[0] == '.' && name[1] == '.'))
6044 rc = osd_get_pfid_from_linkea(env, obj,
6047 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6050 osd_id_gen(id, ino, OSD_OII_NOGEN);
6053 if (rc != 0 || osd_remote_fid(env, dev, fid)) {
6054 fid_zero(&oic->oic_fid);
6059 osd_add_oi_cache(osd_oti_get(env), osd_obj2dev(obj), id, fid);
6060 rc = osd_consistency_check(oti, dev, oic);
6062 fid_zero(&oic->oic_fid);
6064 /* Other error should not affect lookup result. */
6074 ldiskfs_htree_unlock(hlock);
6076 up_read(&obj->oo_ext_idx_sem);
6080 static int osd_index_declare_ea_insert(const struct lu_env *env,
6081 struct dt_object *dt,
6082 const struct dt_rec *rec,
6083 const struct dt_key *key,
6084 struct thandle *handle)
6086 struct osd_thandle *oh;
6087 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6088 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6089 const struct lu_fid *fid = rec1->rec_fid;
6090 int credits, rc = 0;
6091 struct osd_idmap_cache *idc;
6095 LASSERT(!dt_object_remote(dt));
6096 LASSERT(handle != NULL);
6097 LASSERT(fid != NULL);
6098 LASSERT(rec1->rec_type != 0);
6100 oh = container_of(handle, struct osd_thandle, ot_super);
6101 LASSERT(oh->ot_handle == NULL);
6103 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6106 * we can't call iget() while a transactions is running
6107 * (this can lead to a deadlock), but we need to know
6108 * inum and object type. so we find this information at
6109 * declaration and cache in per-thread info
6111 idc = osd_idc_find_or_init(env, osd, fid);
6113 RETURN(PTR_ERR(idc));
6114 if (idc->oic_remote) {
6116 * a reference to remote inode is represented by an
6117 * agent inode which we have to create
6119 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6120 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6123 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6125 if (osd_dt_obj(dt)->oo_inode != NULL) {
6126 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6129 * We ignore block quota on meta pool (MDTs), so needn't
6130 * calculate how many blocks will be consumed by this index
6133 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6135 i_projid_read(inode), 0,
6136 oh, osd_dt_obj(dt), NULL,
6141 #ifdef HAVE_PROJECT_QUOTA
6143 * Reserve credits for local agent inode to transfer
6144 * to 0, quota enforcement is ignored in this case.
6146 if (idc->oic_remote &&
6147 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6148 i_projid_read(inode) != 0)
6149 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6150 0, i_projid_read(inode),
6151 0, false, PRJQUOTA, true);
6159 * Index add function for interoperability mode (b11826).
6160 * It will add the directory entry.This entry is needed to
6161 * maintain name->fid mapping.
6163 * \param key it is key i.e. file entry to be inserted
6164 * \param rec it is value of given key i.e. fid
6166 * \retval 0, on success
6167 * \retval -ve, on error
6169 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6170 const struct dt_rec *rec,
6171 const struct dt_key *key, struct thandle *th)
6173 struct osd_object *obj = osd_dt_obj(dt);
6174 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6175 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6176 const struct lu_fid *fid = rec1->rec_fid;
6177 const char *name = (const char *)key;
6178 struct osd_thread_info *oti = osd_oti_get(env);
6179 struct inode *child_inode = NULL;
6180 struct osd_idmap_cache *idc;
6185 if (!dt_object_exists(dt))
6188 LASSERT(osd_invariant(obj));
6189 LASSERT(!dt_object_remote(dt));
6190 LASSERT(th != NULL);
6192 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6194 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6196 idc = osd_idc_find(env, osd, fid);
6197 if (unlikely(idc == NULL)) {
6198 idc = osd_idc_find_or_init(env, osd, fid);
6201 * this dt_insert() wasn't declared properly, so
6202 * FID is missing in OI cache. we better do not
6203 * lookup FID in FLDB/OI and don't risk to deadlock,
6204 * but in some special cases (lfsck testing, etc)
6205 * it's much simpler than fixing a caller.
6207 * normally this error should be placed after the first
6208 * find, but migrate may attach source stripes to
6209 * target, which doesn't create stripes.
6211 CERROR("%s: "DFID" wasn't declared for insert\n",
6212 osd_name(osd), PFID(fid));
6214 RETURN(PTR_ERR(idc));
6218 if (idc->oic_remote) {
6219 /* Insert remote entry */
6220 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6222 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6224 child_inode = osd_create_local_agent_inode(env, osd,
6225 obj, fid, rec1->rec_type & S_IFMT, th);
6226 if (IS_ERR(child_inode))
6227 RETURN(PTR_ERR(child_inode));
6230 /* Insert local entry */
6231 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6232 /* for a reason OI cache wasn't filled properly */
6233 CERROR("%s: OIC for "DFID" isn't filled\n",
6234 osd_name(osd), PFID(fid));
6237 child_inode = oti->oti_inode;
6238 if (unlikely(child_inode == NULL)) {
6239 struct ldiskfs_inode_info *lii;
6244 child_inode = oti->oti_inode = &lii->vfs_inode;
6246 child_inode->i_sb = osd_sb(osd);
6247 child_inode->i_ino = idc->oic_lid.oii_ino;
6248 child_inode->i_mode = rec1->rec_type & S_IFMT;
6251 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6253 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6254 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6256 if (child_inode && child_inode != oti->oti_inode)
6258 LASSERT(osd_invariant(obj));
6259 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6265 * Initialize osd Iterator for given osd index object.
6267 * \param dt osd index object
6270 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6271 struct dt_object *dt,
6274 struct osd_it_iam *it;
6275 struct osd_object *obj = osd_dt_obj(dt);
6276 struct lu_object *lo = &dt->do_lu;
6277 struct iam_path_descr *ipd;
6278 struct iam_container *bag = &obj->oo_dir->od_container;
6280 if (!dt_object_exists(dt))
6281 return ERR_PTR(-ENOENT);
6285 return ERR_PTR(-ENOMEM);
6287 ipd = osd_it_ipd_get(env, bag);
6288 if (likely(ipd != NULL)) {
6292 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6293 return (struct dt_it *)it;
6296 return ERR_PTR(-ENOMEM);
6301 * free given Iterator.
6303 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6305 struct osd_it_iam *it = (struct osd_it_iam *)di;
6306 struct osd_object *obj = it->oi_obj;
6308 iam_it_fini(&it->oi_it);
6309 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6310 osd_object_put(env, obj);
6315 * Move Iterator to record specified by \a key
6317 * \param di osd iterator
6318 * \param key key for index
6320 * \retval +ve di points to record with least key not larger than key
6321 * \retval 0 di points to exact matched key
6322 * \retval -ve failure
6325 static int osd_it_iam_get(const struct lu_env *env,
6326 struct dt_it *di, const struct dt_key *key)
6328 struct osd_thread_info *oti = osd_oti_get(env);
6329 struct osd_it_iam *it = (struct osd_it_iam *)di;
6331 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6332 /* swab quota uid/gid */
6333 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6334 key = (struct dt_key *)&oti->oti_quota_id;
6337 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6343 * \param di osd iterator
6345 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6347 struct osd_it_iam *it = (struct osd_it_iam *)di;
6349 iam_it_put(&it->oi_it);
6353 * Move iterator by one record
6355 * \param di osd iterator
6357 * \retval +1 end of container reached
6359 * \retval -ve failure
6362 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6364 struct osd_it_iam *it = (struct osd_it_iam *)di;
6366 return iam_it_next(&it->oi_it);
6370 * Return pointer to the key under iterator.
6373 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6374 const struct dt_it *di)
6376 struct osd_thread_info *oti = osd_oti_get(env);
6377 struct osd_it_iam *it = (struct osd_it_iam *)di;
6378 struct osd_object *obj = it->oi_obj;
6381 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6383 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6384 /* swab quota uid/gid */
6385 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6386 key = (struct dt_key *)&oti->oti_quota_id;
6393 * Return size of key under iterator (in bytes)
6396 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6398 struct osd_it_iam *it = (struct osd_it_iam *)di;
6400 return iam_it_key_size(&it->oi_it);
6404 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6406 /* check if file type is required */
6407 if (ent->lde_attrs & LUDA_TYPE) {
6408 struct luda_type *lt;
6409 int align = sizeof(*lt) - 1;
6411 len = (len + align) & ~align;
6412 lt = (struct luda_type *)(ent->lde_name + len);
6413 lt->lt_type = cpu_to_le16(DTTOIF(type));
6416 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6420 * build lu direct from backend fs dirent.
6424 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6425 char *name, __u16 namelen, __u16 type, __u32 attr)
6427 ent->lde_attrs = attr | LUDA_FID;
6428 fid_cpu_to_le(&ent->lde_fid, fid);
6430 ent->lde_hash = cpu_to_le64(offset);
6431 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6433 strncpy(ent->lde_name, name, namelen);
6434 ent->lde_name[namelen] = '\0';
6435 ent->lde_namelen = cpu_to_le16(namelen);
6437 /* append lustre attributes */
6438 osd_it_append_attrs(ent, namelen, type);
6442 * Return pointer to the record under iterator.
6444 static int osd_it_iam_rec(const struct lu_env *env,
6445 const struct dt_it *di,
6446 struct dt_rec *dtrec, __u32 attr)
6448 struct osd_it_iam *it = (struct osd_it_iam *)di;
6449 struct osd_thread_info *info = osd_oti_get(env);
6453 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6454 const struct osd_fid_pack *rec;
6455 struct lu_fid *fid = &info->oti_fid;
6456 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6462 name = (char *)iam_it_key_get(&it->oi_it);
6464 RETURN(PTR_ERR(name));
6466 namelen = iam_it_key_size(&it->oi_it);
6468 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6470 RETURN(PTR_ERR(rec));
6472 rc = osd_fid_unpack(fid, rec);
6476 hash = iam_it_store(&it->oi_it);
6478 /* IAM does not store object type in IAM index (dir) */
6479 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6481 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6482 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6483 (struct iam_rec *)dtrec);
6484 osd_quota_unpack(it->oi_obj, dtrec);
6486 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6487 (struct iam_rec *)dtrec);
6494 * Returns cookie for current Iterator position.
6496 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6498 struct osd_it_iam *it = (struct osd_it_iam *)di;
6500 return iam_it_store(&it->oi_it);
6504 * Restore iterator from cookie.
6506 * \param di osd iterator
6507 * \param hash Iterator location cookie
6509 * \retval +ve di points to record with least key not larger than key.
6510 * \retval 0 di points to exact matched key
6511 * \retval -ve failure
6514 static int osd_it_iam_load(const struct lu_env *env,
6515 const struct dt_it *di, __u64 hash)
6517 struct osd_it_iam *it = (struct osd_it_iam *)di;
6519 return iam_it_load(&it->oi_it, hash);
6522 static const struct dt_index_operations osd_index_iam_ops = {
6523 .dio_lookup = osd_index_iam_lookup,
6524 .dio_declare_insert = osd_index_declare_iam_insert,
6525 .dio_insert = osd_index_iam_insert,
6526 .dio_declare_delete = osd_index_declare_iam_delete,
6527 .dio_delete = osd_index_iam_delete,
6529 .init = osd_it_iam_init,
6530 .fini = osd_it_iam_fini,
6531 .get = osd_it_iam_get,
6532 .put = osd_it_iam_put,
6533 .next = osd_it_iam_next,
6534 .key = osd_it_iam_key,
6535 .key_size = osd_it_iam_key_size,
6536 .rec = osd_it_iam_rec,
6537 .store = osd_it_iam_store,
6538 .load = osd_it_iam_load
6544 * Creates or initializes iterator context.
6546 * \retval struct osd_it_ea, iterator structure on success
6549 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6550 struct dt_object *dt,
6553 struct osd_object *obj = osd_dt_obj(dt);
6554 struct osd_thread_info *info = osd_oti_get(env);
6555 struct osd_it_ea *oie;
6557 struct lu_object *lo = &dt->do_lu;
6558 struct dentry *obj_dentry;
6562 if (!dt_object_exists(dt) || obj->oo_destroyed)
6563 RETURN(ERR_PTR(-ENOENT));
6565 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6567 RETURN(ERR_PTR(-ENOMEM));
6568 obj_dentry = &oie->oie_dentry;
6570 obj_dentry->d_inode = obj->oo_inode;
6571 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
6572 obj_dentry->d_name.hash = 0;
6574 oie->oie_rd_dirent = 0;
6575 oie->oie_it_dirent = 0;
6576 oie->oie_dirent = NULL;
6577 if (unlikely(!info->oti_it_ea_buf_used)) {
6578 oie->oie_buf = info->oti_it_ea_buf;
6579 info->oti_it_ea_buf_used = 1;
6581 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6582 if (oie->oie_buf == NULL)
6583 RETURN(ERR_PTR(-ENOMEM));
6587 file = &oie->oie_file;
6589 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6590 if (attr & LUDA_64BITHASH)
6591 file->f_mode = FMODE_64BITHASH;
6593 file->f_mode = FMODE_32BITHASH;
6594 file->f_path.dentry = obj_dentry;
6595 file->f_mapping = obj->oo_inode->i_mapping;
6596 file->f_op = obj->oo_inode->i_fop;
6597 file->f_inode = obj->oo_inode;
6600 RETURN((struct dt_it *)oie);
6604 * Destroy or finishes iterator context.
6606 * \param di iterator structure to be destroyed
6608 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6610 struct osd_thread_info *info = osd_oti_get(env);
6611 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6612 struct osd_object *obj = oie->oie_obj;
6613 struct inode *inode = obj->oo_inode;
6616 oie->oie_file.f_op->release(inode, &oie->oie_file);
6617 osd_object_put(env, obj);
6618 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6619 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6621 info->oti_it_ea_buf_used = 0;
6622 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6627 * It position the iterator at given key, so that next lookup continues from
6628 * that key Or it is similar to dio_it->load() but based on a key,
6629 * rather than file position.
6631 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6634 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6636 static int osd_it_ea_get(const struct lu_env *env,
6637 struct dt_it *di, const struct dt_key *key)
6639 struct osd_it_ea *it = (struct osd_it_ea *)di;
6642 LASSERT(((const char *)key)[0] == '\0');
6643 it->oie_file.f_pos = 0;
6644 it->oie_rd_dirent = 0;
6645 it->oie_it_dirent = 0;
6646 it->oie_dirent = NULL;
6654 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6658 struct osd_filldir_cbs {
6659 struct dir_context ctx;
6660 struct osd_it_ea *it;
6663 * It is called internally by ->iterate*(). It fills the
6664 * iterator's in-memory data structure with required
6665 * information i.e. name, namelen, rec_size etc.
6667 * \param buf in which information to be filled in.
6668 * \param name name of the file in given dir
6670 * \retval 0 on success
6671 * \retval 1 on buffer full
6673 #ifdef HAVE_FILLDIR_USE_CTX
6674 static int osd_ldiskfs_filldir(struct dir_context *buf,
6676 static int osd_ldiskfs_filldir(void *buf,
6678 const char *name, int namelen,
6679 loff_t offset, __u64 ino, unsigned int d_type)
6681 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6682 struct osd_object *obj = it->oie_obj;
6683 struct osd_it_ea_dirent *ent = it->oie_dirent;
6684 struct lu_fid *fid = &ent->oied_fid;
6685 struct osd_fid_pack *rec;
6689 /* this should never happen */
6690 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6691 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6695 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6699 /* "." is just the object itself. */
6700 if (namelen == 1 && name[0] == '.') {
6701 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6702 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6703 rec = (struct osd_fid_pack *)(name + namelen + 1);
6704 if (osd_fid_unpack(fid, rec) != 0)
6709 d_type &= ~LDISKFS_DIRENT_LUFID;
6711 /* NOT export local root. */
6712 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6713 ino = obj->oo_inode->i_ino;
6714 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6717 ent->oied_ino = ino;
6718 ent->oied_off = offset;
6719 ent->oied_namelen = namelen;
6720 ent->oied_type = d_type;
6722 memcpy(ent->oied_name, name, namelen);
6724 it->oie_rd_dirent++;
6725 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6730 * Calls ->iterate*() to load a directory entry at a time
6731 * and stored it in iterator's in-memory data structure.
6733 * \param di iterator's in memory structure
6735 * \retval 0 on success
6736 * \retval -ve on error
6737 * \retval +1 reach the end of entry
6739 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6740 const struct dt_it *di)
6742 struct osd_it_ea *it = (struct osd_it_ea *)di;
6743 struct osd_object *obj = it->oie_obj;
6744 struct inode *inode = obj->oo_inode;
6745 struct htree_lock *hlock = NULL;
6746 struct file *filp = &it->oie_file;
6748 struct osd_filldir_cbs buf = {
6749 .ctx.actor = osd_ldiskfs_filldir,
6754 it->oie_dirent = it->oie_buf;
6755 it->oie_rd_dirent = 0;
6757 if (obj->oo_hl_head != NULL) {
6758 hlock = osd_oti_get(env)->oti_hlock;
6759 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6760 inode, LDISKFS_HLOCK_READDIR);
6762 down_read(&obj->oo_ext_idx_sem);
6765 filp->f_cred = current_cred();
6766 rc = osd_security_file_alloc(filp);
6770 filp->f_flags |= O_NOATIME;
6771 filp->f_mode |= FMODE_NONOTIFY;
6772 rc = iterate_dir(filp, &buf.ctx);
6776 if (it->oie_rd_dirent == 0) {
6778 * If it does not get any dirent, it means it has been reached
6779 * to the end of the dir
6781 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6785 it->oie_dirent = it->oie_buf;
6786 it->oie_it_dirent = 1;
6790 ldiskfs_htree_unlock(hlock);
6792 up_read(&obj->oo_ext_idx_sem);
6798 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
6799 * to load a directory entry at a time and stored it in
6800 * iterator's in-memory data structure.
6802 * \param di iterator's in memory structure
6804 * \retval +ve iterator reached to end
6805 * \retval 0 iterator not reached to end
6806 * \retval -ve on error
6808 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6810 struct osd_it_ea *it = (struct osd_it_ea *)di;
6815 if (it->oie_it_dirent < it->oie_rd_dirent) {
6817 (void *)it->oie_dirent +
6818 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6819 it->oie_dirent->oied_namelen);
6820 it->oie_it_dirent++;
6823 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6826 rc = osd_ldiskfs_it_fill(env, di);
6833 * Returns the key at current position from iterator's in memory structure.
6835 * \param di iterator's in memory structure
6837 * \retval key i.e. struct dt_key on success
6839 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6840 const struct dt_it *di)
6842 struct osd_it_ea *it = (struct osd_it_ea *)di;
6844 return (struct dt_key *)it->oie_dirent->oied_name;
6848 * Returns key's size at current position from iterator's in memory structure.
6850 * \param di iterator's in memory structure
6852 * \retval key_size i.e. struct dt_key on success
6854 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6856 struct osd_it_ea *it = (struct osd_it_ea *)di;
6858 return it->oie_dirent->oied_namelen;
6861 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
6862 #undef LDISKFS_DIR_REC_LEN
6863 #define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
6866 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6868 if (LDISKFS_DIR_REC_LEN(de) >=
6869 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6876 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6877 unsigned int blocksize, bool dotdot)
6880 return osd_dotdot_has_space(de);
6882 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6883 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6890 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6891 handle_t *jh, struct dentry *dentry,
6892 const struct lu_fid *fid, struct buffer_head *bh,
6893 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6896 struct inode *dir = dentry->d_parent->d_inode;
6897 struct inode *inode = dentry->d_inode;
6898 struct osd_fid_pack *rec;
6899 struct ldiskfs_dentry_param *ldp;
6900 int namelen = dentry->d_name.len;
6902 struct osd_thread_info *info = osd_oti_get(env);
6906 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6909 /* There is enough space to hold the FID-in-dirent. */
6910 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6911 rc = ldiskfs_journal_get_write_access(jh, bh);
6915 de->name[namelen] = 0;
6916 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6917 rec->fp_len = sizeof(struct lu_fid) + 1;
6918 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6919 de->file_type |= LDISKFS_DIRENT_LUFID;
6920 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6927 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6931 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6932 osd_get_ldiskfs_dirent_param(ldp, fid);
6933 dentry->d_fsdata = (void *)ldp;
6934 dquot_initialize(dir);
6935 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6937 * It is too bad, we cannot reinsert the name entry back.
6938 * That means we lose it!
6942 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6943 osd_ino2name(inode), dir->i_ino, dir->i_generation,
6944 namelen, dentry->d_name.name, PFID(fid), rc);
6950 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
6951 struct osd_it_ea *it, struct lu_fid *fid,
6952 struct osd_inode_id *id, __u32 *attr)
6954 struct osd_thread_info *info = osd_oti_get(env);
6955 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
6956 struct osd_device *dev = osd_obj2dev(obj);
6957 struct super_block *sb = osd_sb(dev);
6958 const char *devname = osd_name(dev);
6959 struct osd_it_ea_dirent *ent = it->oie_dirent;
6960 struct inode *dir = obj->oo_inode;
6961 struct htree_lock *hlock = NULL;
6962 struct buffer_head *bh = NULL;
6963 handle_t *jh = NULL;
6964 struct ldiskfs_dir_entry_2 *de;
6965 struct dentry *dentry;
6966 struct inode *inode;
6967 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
6970 bool dotdot = false;
6975 if (ent->oied_name[0] == '.') {
6976 if (ent->oied_namelen == 1)
6979 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
6983 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
6984 inode = osd_iget(info, dev, id);
6985 if (IS_ERR(inode)) {
6986 rc = PTR_ERR(inode);
6987 if (rc == -ENOENT || rc == -ESTALE) {
6989 * Maybe dangling name entry, or
6990 * corrupted directory entry.
6992 *attr |= LUDA_UNKNOWN;
6995 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
6996 "check_repair, dir = %lu/%u, name = %.*s, "
6997 "ino = %llu, rc = %d\n",
6998 devname, dir->i_ino, dir->i_generation,
6999 ent->oied_namelen, ent->oied_name,
7006 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
7008 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7009 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7015 * We need to ensure that the name entry is still valid.
7016 * Because it may be removed or renamed by other already.
7018 * The unlink or rename operation will start journal before PDO lock,
7019 * so to avoid deadlock, here we need to start journal handle before
7020 * related PDO lock also. But because we do not know whether there
7021 * will be something to be repaired before PDO lock, we just start
7022 * journal without conditions.
7024 * We may need to remove the name entry firstly, then insert back.
7025 * One credit is for user quota file update.
7026 * One credit is for group quota file update.
7027 * Two credits are for dirty inode.
7029 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7030 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7032 if (dev->od_dirent_journal != 0) {
7035 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7038 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7039 "check_repair, dir = %lu/%u, credits = %d, "
7040 "name = %.*s, ino = %llu: rc = %d\n",
7041 devname, dir->i_ino, dir->i_generation, credits,
7042 ent->oied_namelen, ent->oied_name,
7045 GOTO(out_inode, rc);
7048 if (obj->oo_hl_head != NULL) {
7049 hlock = osd_oti_get(env)->oti_hlock;
7051 * "0" means exclusive lock for the whole directory.
7052 * We need to prevent others access such name entry
7053 * during the delete + insert. Neither HLOCK_ADD nor
7054 * HLOCK_DEL cannot guarantee the atomicity.
7056 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7058 down_write(&obj->oo_ext_idx_sem);
7061 if (obj->oo_hl_head != NULL) {
7062 hlock = osd_oti_get(env)->oti_hlock;
7063 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7064 LDISKFS_HLOCK_LOOKUP);
7066 down_read(&obj->oo_ext_idx_sem);
7070 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7071 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7072 *attr |= LUDA_IGNORE;
7078 * For dotdot entry, if there is not enough space to hold the
7079 * FID-in-dirent, just keep them there. It only happens when the
7080 * device upgraded from 1.8 or restored from MDT file-level backup.
7081 * For the whole directory, only dotdot entry have no FID-in-dirent
7082 * and needs to get FID from LMA when readdir, it will not affect the
7085 if (dotdot && !osd_dotdot_has_space(de)) {
7086 *attr |= LUDA_UNKNOWN;
7092 if (lu_fid_eq(fid, &lma->lma_self_fid))
7095 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7096 struct lu_fid *tfid = &lma->lma_self_fid;
7098 if (likely(dotdot &&
7099 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7100 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7102 * It must be REMOTE_PARENT_DIR and as the
7103 * 'dotdot' entry of remote directory
7105 *attr |= LUDA_IGNORE;
7107 CDEBUG(D_LFSCK, "%s: expect remote agent "
7108 "parent directory, but got %.*s under "
7109 "dir = %lu/%u with the FID "DFID"\n",
7110 devname, ent->oied_namelen,
7111 ent->oied_name, dir->i_ino,
7112 dir->i_generation, PFID(tfid));
7114 *attr |= LUDA_UNKNOWN;
7121 if (!fid_is_zero(fid)) {
7122 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7124 if (rc == -ENOENT ||
7126 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7128 * linkEA does not recognize the dirent entry,
7129 * it may because the dirent entry corruption
7130 * and points to other's inode.
7132 CDEBUG(D_LFSCK, "%s: the target inode does not "
7133 "recognize the dirent, dir = %lu/%u, "
7134 " name = %.*s, ino = %llu, "
7135 DFID": rc = %d\n", devname, dir->i_ino,
7136 dir->i_generation, ent->oied_namelen,
7137 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7138 *attr |= LUDA_UNKNOWN;
7143 if (rc && rc != -ENODATA) {
7144 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7145 "dir = %lu/%u, name = %.*s, ino = %llu, "
7146 DFID": rc = %d\n", devname, dir->i_ino,
7147 dir->i_generation, ent->oied_namelen,
7148 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7149 *attr |= LUDA_UNKNOWN;
7157 * linkEA recognizes the dirent entry, the FID-in-LMA is
7158 * valid, trusted, in spite of fid_is_sane(fid) or not.
7160 if (*attr & LUDA_VERIFY_DRYRUN) {
7161 *fid = lma->lma_self_fid;
7162 *attr |= LUDA_REPAIR;
7169 dev->od_dirent_journal = 1;
7170 if (hlock != NULL) {
7171 ldiskfs_htree_unlock(hlock);
7174 up_read(&obj->oo_ext_idx_sem);
7180 *fid = lma->lma_self_fid;
7182 /* Update or append the FID-in-dirent. */
7183 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7184 bh, de, hlock, dotdot);
7186 *attr |= LUDA_REPAIR;
7188 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7189 "the dirent, dir = %lu/%u, name = %.*s, "
7190 "ino = %llu, "DFID": rc = %d\n",
7191 devname, dir->i_ino, dir->i_generation,
7192 ent->oied_namelen, ent->oied_name,
7193 ent->oied_ino, PFID(fid), rc);
7195 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7196 if (*attr & LUDA_VERIFY_DRYRUN) {
7197 if (fid_is_sane(fid)) {
7198 *attr |= LUDA_REPAIR;
7199 } else if (dev->od_index == 0) {
7200 lu_igif_build(fid, inode->i_ino,
7201 inode->i_generation);
7202 *attr |= LUDA_UPGRADE;
7210 dev->od_dirent_journal = 1;
7211 if (hlock != NULL) {
7212 ldiskfs_htree_unlock(hlock);
7215 up_read(&obj->oo_ext_idx_sem);
7222 if (unlikely(fid_is_sane(fid))) {
7224 * FID-in-dirent exists, but FID-in-LMA is lost.
7225 * Trust the FID-in-dirent, and add FID-in-LMA.
7227 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7229 *attr |= LUDA_REPAIR;
7231 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7232 "update dirent, dir = %lu/%u, "
7233 "name = %.*s, ino = %llu, "
7235 devname, dir->i_ino, dir->i_generation,
7236 ent->oied_namelen, ent->oied_name,
7237 ent->oied_ino, PFID(fid), rc);
7238 } else if (dev->od_index == 0) {
7239 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7241 * It is probably IGIF object. Only aappend the
7242 * FID-in-dirent. OI scrub will process FID-in-LMA.
7244 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7245 bh, de, hlock, dotdot);
7247 *attr |= LUDA_UPGRADE;
7249 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7250 "after the dirent, dir = %lu/%u, "
7251 "name = %.*s, ino = %llu, "
7253 devname, dir->i_ino, dir->i_generation,
7254 ent->oied_namelen, ent->oied_name,
7255 ent->oied_ino, PFID(fid), rc);
7264 if (hlock != NULL) {
7265 ldiskfs_htree_unlock(hlock);
7267 if (dev->od_dirent_journal != 0)
7268 up_write(&obj->oo_ext_idx_sem);
7270 up_read(&obj->oo_ext_idx_sem);
7274 ldiskfs_journal_stop(jh);
7278 if (rc >= 0 && !dirty)
7279 dev->od_dirent_journal = 0;
7285 * Returns the value at current position from iterator's in memory structure.
7287 * \param di struct osd_it_ea, iterator's in memory structure
7288 * \param attr attr requested for dirent.
7289 * \param lde lustre dirent
7291 * \retval 0 no error and \param lde has correct lustre dirent.
7292 * \retval -ve on error
7294 static inline int osd_it_ea_rec(const struct lu_env *env,
7295 const struct dt_it *di,
7296 struct dt_rec *dtrec, __u32 attr)
7298 struct osd_it_ea *it = (struct osd_it_ea *)di;
7299 struct osd_object *obj = it->oie_obj;
7300 struct osd_device *dev = osd_obj2dev(obj);
7301 struct osd_thread_info *oti = osd_oti_get(env);
7302 struct osd_inode_id *id = &oti->oti_id;
7303 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7304 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7305 __u32 ino = it->oie_dirent->oied_ino;
7310 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7312 if (attr & LUDA_VERIFY) {
7313 if (unlikely(is_remote_parent_ino(dev, ino))) {
7314 attr |= LUDA_IGNORE;
7316 * If the parent is on remote MDT, and there
7317 * is no FID-in-dirent, then we have to get
7318 * the parent FID from the linkEA.
7320 if (!fid_is_sane(fid) &&
7321 it->oie_dirent->oied_namelen == 2 &&
7322 it->oie_dirent->oied_name[0] == '.' &&
7323 it->oie_dirent->oied_name[1] == '.')
7324 osd_get_pfid_from_linkea(env, obj, fid);
7326 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7330 if (!fid_is_sane(fid))
7331 attr |= LUDA_UNKNOWN;
7333 attr &= ~LU_DIRENT_ATTRS_MASK;
7334 if (!fid_is_sane(fid)) {
7335 bool is_dotdot = false;
7337 if (it->oie_dirent->oied_namelen == 2 &&
7338 it->oie_dirent->oied_name[0] == '.' &&
7339 it->oie_dirent->oied_name[1] == '.')
7342 * If the parent is on remote MDT, and there
7343 * is no FID-in-dirent, then we have to get
7344 * the parent FID from the linkEA.
7346 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7347 rc = osd_get_pfid_from_linkea(env, obj, fid);
7349 if (is_dotdot == false &&
7350 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7353 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7358 /* Pack the entry anyway, at least the offset is right. */
7359 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7360 it->oie_dirent->oied_name,
7361 it->oie_dirent->oied_namelen,
7362 it->oie_dirent->oied_type, attr);
7364 RETURN(rc > 0 ? 0 : rc);
7368 * Returns the record size size at current position.
7370 * This function will return record(lu_dirent) size in bytes.
7372 * \param[in] env execution environment
7373 * \param[in] di iterator's in memory structure
7374 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7375 * calculate the lu_dirent size.
7377 * \retval record size(in bytes & in memory) of the current lu_dirent
7380 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7383 struct osd_it_ea *it = (struct osd_it_ea *)di;
7385 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7389 * Returns a cookie for current position of the iterator head, so that
7390 * user can use this cookie to load/start the iterator next time.
7392 * \param di iterator's in memory structure
7394 * \retval cookie for current position, on success
7396 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7398 struct osd_it_ea *it = (struct osd_it_ea *)di;
7400 return it->oie_dirent->oied_off;
7404 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7405 * to load a directory entry at a time and stored it i inn,
7406 * in iterator's in-memory data structure.
7408 * \param di struct osd_it_ea, iterator's in memory structure
7410 * \retval +ve on success
7411 * \retval -ve on error
7413 static int osd_it_ea_load(const struct lu_env *env,
7414 const struct dt_it *di, __u64 hash)
7416 struct osd_it_ea *it = (struct osd_it_ea *)di;
7420 it->oie_file.f_pos = hash;
7422 rc = osd_ldiskfs_it_fill(env, di);
7433 * Index lookup function for interoperability mode (b11826).
7435 * \param key, key i.e. file name to be searched
7437 * \retval +ve, on success
7438 * \retval -ve, on error
7440 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7441 struct dt_rec *rec, const struct dt_key *key)
7443 struct osd_object *obj = osd_dt_obj(dt);
7448 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7449 LINVRNT(osd_invariant(obj));
7451 rc = osd_ea_lookup_rec(env, obj, rec, key);
7458 * Index and Iterator operations for interoperability
7459 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7461 static const struct dt_index_operations osd_index_ea_ops = {
7462 .dio_lookup = osd_index_ea_lookup,
7463 .dio_declare_insert = osd_index_declare_ea_insert,
7464 .dio_insert = osd_index_ea_insert,
7465 .dio_declare_delete = osd_index_declare_ea_delete,
7466 .dio_delete = osd_index_ea_delete,
7468 .init = osd_it_ea_init,
7469 .fini = osd_it_ea_fini,
7470 .get = osd_it_ea_get,
7471 .put = osd_it_ea_put,
7472 .next = osd_it_ea_next,
7473 .key = osd_it_ea_key,
7474 .key_size = osd_it_ea_key_size,
7475 .rec = osd_it_ea_rec,
7476 .rec_size = osd_it_ea_rec_size,
7477 .store = osd_it_ea_store,
7478 .load = osd_it_ea_load
7482 static void *osd_key_init(const struct lu_context *ctx,
7483 struct lu_context_key *key)
7485 struct osd_thread_info *info;
7487 OBD_ALLOC_PTR(info);
7489 return ERR_PTR(-ENOMEM);
7491 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7492 if (info->oti_it_ea_buf == NULL)
7495 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7497 info->oti_hlock = ldiskfs_htree_lock_alloc();
7498 if (info->oti_hlock == NULL)
7504 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7507 return ERR_PTR(-ENOMEM);
7510 static void osd_key_fini(const struct lu_context *ctx,
7511 struct lu_context_key *key, void *data)
7513 struct osd_thread_info *info = data;
7514 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7515 struct osd_idmap_cache *idc = info->oti_ins_cache;
7517 if (info->oti_dio_pages) {
7519 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7520 struct page *page = info->oti_dio_pages[i];
7522 LASSERT(PagePrivate2(page));
7523 LASSERT(PageLocked(page));
7524 ClearPagePrivate2(page);
7529 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7530 PTLRPC_MAX_BRW_PAGES);
7533 if (info->oti_inode != NULL)
7535 if (info->oti_hlock != NULL)
7536 ldiskfs_htree_lock_free(info->oti_hlock);
7537 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7538 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7539 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7540 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7541 lu_buf_free(&info->oti_big_buf);
7543 LASSERT(info->oti_ins_cache_size > 0);
7544 OBD_FREE_PTR_ARRAY(idc, info->oti_ins_cache_size);
7545 info->oti_ins_cache = NULL;
7546 info->oti_ins_cache_size = 0;
7551 static void osd_key_exit(const struct lu_context *ctx,
7552 struct lu_context_key *key, void *data)
7554 struct osd_thread_info *info = data;
7556 LASSERT(info->oti_r_locks == 0);
7557 LASSERT(info->oti_w_locks == 0);
7558 LASSERT(info->oti_txns == 0);
7561 /* type constructor/destructor: osd_type_init, osd_type_fini */
7562 LU_TYPE_INIT_FINI(osd, &osd_key);
7564 struct lu_context_key osd_key = {
7565 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7566 .lct_init = osd_key_init,
7567 .lct_fini = osd_key_fini,
7568 .lct_exit = osd_key_exit
7572 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7573 const char *name, struct lu_device *next)
7575 struct osd_device *osd = osd_dev(d);
7577 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7578 sizeof(osd->od_svname))
7580 return osd_procfs_init(osd, name);
7583 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7585 struct seq_server_site *ss = osd_seq_site(osd);
7590 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7593 if (unlikely(ss == NULL))
7596 OBD_ALLOC_PTR(osd->od_cl_seq);
7597 if (osd->od_cl_seq == NULL)
7600 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7601 osd->od_svname, ss->ss_server_seq);
7603 if (ss->ss_node_id == 0) {
7605 * If the OSD on the sequence controller(MDT0), then allocate
7606 * sequence here, otherwise allocate sequence after connected
7607 * to MDT0 (see mdt_register_lwp_callback()).
7609 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7610 &osd->od_cl_seq->lcs_space, env);
7616 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7618 if (osd->od_cl_seq == NULL)
7621 seq_client_fini(osd->od_cl_seq);
7622 OBD_FREE_PTR(osd->od_cl_seq);
7623 osd->od_cl_seq = NULL;
7626 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7630 /* shutdown quota slave instance associated with the device */
7631 if (o->od_quota_slave_md != NULL) {
7632 struct qsd_instance *qsd = o->od_quota_slave_md;
7634 o->od_quota_slave_md = NULL;
7638 if (o->od_quota_slave_dt != NULL) {
7639 struct qsd_instance *qsd = o->od_quota_slave_dt;
7641 o->od_quota_slave_dt = NULL;
7645 osd_fid_fini(env, o);
7646 osd_scrub_cleanup(env, o);
7651 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7655 if (o->od_mnt != NULL) {
7656 shrink_dcache_sb(osd_sb(o));
7657 osd_sync(env, &o->od_dt_dev);
7658 wait_event(o->od_commit_cb_done,
7659 !atomic_read(&o->od_commit_cb_in_flight));
7668 static int osd_mount(const struct lu_env *env,
7669 struct osd_device *o, struct lustre_cfg *cfg)
7671 const char *name = lustre_cfg_string(cfg, 0);
7672 const char *dev = lustre_cfg_string(cfg, 1);
7674 unsigned long page, s_flags, lmd_flags = 0;
7675 struct page *__page;
7676 struct file_system_type *type;
7677 char *options = NULL;
7679 struct osd_thread_info *info = osd_oti_get(env);
7680 struct lu_fid *fid = &info->oti_fid;
7681 struct inode *inode;
7682 int rc = 0, force_over_1024tb = 0;
7686 if (o->od_mnt != NULL)
7689 if (strlen(dev) >= sizeof(o->od_mntdev))
7691 strcpy(o->od_mntdev, dev);
7693 str = lustre_cfg_string(cfg, 2);
7694 s_flags = simple_strtoul(str, NULL, 0);
7695 str = strstr(str, ":");
7697 lmd_flags = simple_strtoul(str + 1, NULL, 0);
7698 opts = lustre_cfg_string(cfg, 3);
7700 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7701 CERROR("%s: device %s extents feature is not guaranteed to "
7702 "work on big-endian systems. Use \"bigendian_extents\" "
7703 "mount option to override.\n", name, dev);
7707 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7708 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7709 CWARN("force_over_128tb option is deprecated. Filesystems smaller than 1024TB can be created without any force option. Use force_over_1024tb option for filesystems larger than 1024TB.\n");
7712 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7713 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7714 CWARN("force_over_256tb option is deprecated. Filesystems smaller than 1024TB can be created without any force options. Use force_over_1024tb option for filesystems larger than 1024TB.\n");
7717 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7718 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
7719 CWARN("force_over_512tb option is deprecated. Filesystems smaller than 1024TB can be created without any force options. Use force_over_1024tb option for filesystems larger than 1024TB.\n");
7723 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
7724 force_over_1024tb = 1;
7726 __page = alloc_page(GFP_KERNEL);
7728 GOTO(out, rc = -ENOMEM);
7729 page = (unsigned long)page_address(__page);
7730 options = (char *)page;
7733 /* strip out the options for back compatiblity */
7734 static const char * const sout[] = {
7741 /* strip out option we processed in osd */
7742 "bigendian_extents",
7746 "force_over_1024tb",
7749 strncat(options, opts, PAGE_SIZE);
7750 for (rc = 0, str = options; sout[rc]; ) {
7751 char *op = strstr(str, sout[rc]);
7758 if (op == options || *(op - 1) == ',') {
7759 str = op + strlen(sout[rc]);
7760 if (*str == ',' || *str == '\0') {
7761 *str == ',' ? str++ : str;
7762 memmove(op, str, strlen(str) + 1);
7765 for (str = op; *str != ',' && *str != '\0'; str++)
7769 strncat(options, "user_xattr,acl", PAGE_SIZE);
7772 /* Glom up mount options */
7773 if (*options != '\0')
7774 strncat(options, ",", PAGE_SIZE);
7775 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7777 type = get_fs_type("ldiskfs");
7779 CERROR("%s: cannot find ldiskfs module\n", name);
7780 GOTO(out, rc = -ENODEV);
7783 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7784 module_put(type->owner);
7786 if (IS_ERR(o->od_mnt)) {
7787 rc = PTR_ERR(o->od_mnt);
7789 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7793 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7794 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
7795 force_over_1024tb == 0) {
7796 CERROR("%s: device %s LDISKFS has not been tested on filesystems larger than 1024TB and may cause data corruption. Use 'force_over_1024tb' mount option to override.\n",
7798 GOTO(out_mnt, rc = -EINVAL);
7801 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7802 LCONSOLE_WARN("%s: not support dev_rdonly on this device",
7805 GOTO(out_mnt, rc = -EOPNOTSUPP);
7808 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7809 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7810 GOTO(out_mnt, rc = -EINVAL);
7813 #ifdef LDISKFS_MOUNT_DIRDATA
7814 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7815 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7816 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7817 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7818 "enabling the dirdata feature. If you do not want to "
7819 "downgrade to Lustre-1.x again, you can enable it via "
7820 "'tune2fs -O dirdata device'\n", name, dev);
7822 inode = osd_sb(o)->s_root->d_inode;
7823 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7824 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7826 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7830 if (lmd_flags & LMD_FLG_NOSCRUB)
7831 o->od_auto_scrub_interval = AS_NEVER;
7833 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7834 /* do not use pagecache with flash-backed storage */
7835 o->od_writethrough_cache = 0;
7836 o->od_read_cache = 0;
7847 __free_page(__page);
7852 static struct lu_device *osd_device_fini(const struct lu_env *env,
7853 struct lu_device *d)
7855 struct osd_device *o = osd_dev(d);
7859 osd_index_backup(env, o, false);
7860 osd_shutdown(env, o);
7862 if (o->od_oi_table != NULL)
7863 osd_oi_fini(osd_oti_get(env), o);
7864 if (o->od_extent_bytes_percpu)
7865 free_percpu(o->od_extent_bytes_percpu);
7866 osd_obj_map_fini(o);
7872 static int osd_device_init0(const struct lu_env *env,
7873 struct osd_device *o,
7874 struct lustre_cfg *cfg)
7876 struct lu_device *l = osd2lu_dev(o);
7877 struct osd_thread_info *info;
7881 /* if the module was re-loaded, env can loose its keys */
7882 rc = lu_env_refill((struct lu_env *)env);
7885 info = osd_oti_get(env);
7888 l->ld_ops = &osd_lu_ops;
7889 o->od_dt_dev.dd_ops = &osd_dt_ops;
7891 spin_lock_init(&o->od_osfs_lock);
7892 mutex_init(&o->od_otable_mutex);
7893 INIT_LIST_HEAD(&o->od_orphan_list);
7894 INIT_LIST_HEAD(&o->od_index_backup_list);
7895 INIT_LIST_HEAD(&o->od_index_restore_list);
7896 spin_lock_init(&o->od_lock);
7897 o->od_index_backup_policy = LIBP_NONE;
7899 init_waitqueue_head(&o->od_commit_cb_done);
7901 o->od_read_cache = 1;
7902 o->od_writethrough_cache = 1;
7903 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7904 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
7905 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
7906 o->od_auto_scrub_interval = AS_DEFAULT;
7907 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
7908 o->od_fallocate_zero_blocks = 0;
7910 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7911 sizeof(o->od_svname));
7912 if (cplen >= sizeof(o->od_svname)) {
7917 o->od_index_backup_stop = 0;
7918 o->od_index = -1; /* -1 means index is invalid */
7919 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7920 if (rc == LDD_F_SV_TYPE_OST)
7923 o->od_full_scrub_ratio = OFSR_DEFAULT;
7924 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7925 rc = osd_mount(env, o, cfg);
7929 /* Can only check block device after mount */
7930 o->od_nonrotational =
7931 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
7933 rc = osd_obj_map_init(env, o);
7937 rc = lu_site_init(&o->od_site, l);
7939 GOTO(out_compat, rc);
7940 o->od_site.ls_bottom_dev = l;
7942 rc = lu_site_init_finish(&o->od_site);
7946 INIT_LIST_HEAD(&o->od_ios_list);
7947 /* setup scrub, including OI files initialization */
7949 rc = osd_scrub_setup(env, o);
7954 rc = osd_procfs_init(o, o->od_svname);
7956 CERROR("%s: can't initialize procfs: rc = %d\n",
7958 GOTO(out_scrub, rc);
7961 LASSERT(l->ld_site->ls_linkage.next != NULL);
7962 LASSERT(l->ld_site->ls_linkage.prev != NULL);
7964 /* initialize quota slave instance */
7965 /* currently it's no need to prepare qsd_instance_md for OST */
7966 if (!o->od_is_ost) {
7967 o->od_quota_slave_md = qsd_init(env, o->od_svname,
7969 o->od_proc_entry, true);
7970 if (IS_ERR(o->od_quota_slave_md)) {
7971 rc = PTR_ERR(o->od_quota_slave_md);
7972 o->od_quota_slave_md = NULL;
7973 GOTO(out_procfs, rc);
7977 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
7978 o->od_proc_entry, false);
7980 if (IS_ERR(o->od_quota_slave_dt)) {
7981 if (o->od_quota_slave_md != NULL) {
7982 qsd_fini(env, o->od_quota_slave_md);
7983 o->od_quota_slave_md = NULL;
7986 rc = PTR_ERR(o->od_quota_slave_dt);
7987 o->od_quota_slave_dt = NULL;
7988 GOTO(out_procfs, rc);
7991 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
7992 if (!o->od_extent_bytes_percpu) {
7994 GOTO(out_procfs, rc);
8002 osd_scrub_cleanup(env, o);
8004 lu_site_fini(&o->od_site);
8006 osd_obj_map_fini(o);
8013 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8014 struct lu_device_type *t,
8015 struct lustre_cfg *cfg)
8017 struct osd_device *o;
8022 return ERR_PTR(-ENOMEM);
8024 rc = dt_device_init(&o->od_dt_dev, t);
8027 * Because the ctx might be revived in dt_device_init,
8028 * refill the env here
8030 lu_env_refill((struct lu_env *)env);
8031 rc = osd_device_init0(env, o, cfg);
8033 dt_device_fini(&o->od_dt_dev);
8036 if (unlikely(rc != 0))
8039 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8042 static struct lu_device *osd_device_free(const struct lu_env *env,
8043 struct lu_device *d)
8045 struct osd_device *o = osd_dev(d);
8049 /* XXX: make osd top device in order to release reference */
8050 d->ld_site->ls_top_dev = d;
8051 lu_site_purge(env, d->ld_site, -1);
8052 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8053 D_ERROR, lu_cdebug_printer);
8054 lu_site_fini(&o->od_site);
8055 dt_device_fini(&o->od_dt_dev);
8060 static int osd_process_config(const struct lu_env *env,
8061 struct lu_device *d, struct lustre_cfg *cfg)
8063 struct osd_device *o = osd_dev(d);
8069 switch (cfg->lcfg_command) {
8071 rc = osd_mount(env, o, cfg);
8075 * For the case LCFG_PRE_CLEANUP is not called in advance,
8076 * that may happend if hit failure during mount process.
8078 osd_index_backup(env, o, false);
8079 lu_dev_del_linkage(d->ld_site, d);
8080 rc = osd_shutdown(env, o);
8083 LASSERT(&o->od_dt_dev);
8084 count = class_modify_config(cfg, PARAM_OSD,
8085 &o->od_dt_dev.dd_kobj);
8087 count = class_modify_config(cfg, PARAM_OST,
8088 &o->od_dt_dev.dd_kobj);
8089 rc = count > 0 ? 0 : count;
8091 case LCFG_PRE_CLEANUP:
8093 osd_index_backup(env, o,
8094 o->od_index_backup_policy != LIBP_NONE);
8104 static int osd_recovery_complete(const struct lu_env *env,
8105 struct lu_device *d)
8107 struct osd_device *osd = osd_dev(d);
8112 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8116 * start qsd instance on recovery completion, this notifies the quota
8117 * slave code that we are about to process new requests now
8119 rc = qsd_start(env, osd->od_quota_slave_dt);
8120 if (rc == 0 && osd->od_quota_slave_md != NULL)
8121 rc = qsd_start(env, osd->od_quota_slave_md);
8127 * we use exports to track all osd users
8129 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8130 struct obd_device *obd, struct obd_uuid *cluuid,
8131 struct obd_connect_data *data, void *localdata)
8133 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8134 struct lustre_handle conn;
8139 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8141 rc = class_connect(&conn, obd, cluuid);
8145 *exp = class_conn2export(&conn);
8147 spin_lock(&osd->od_osfs_lock);
8149 spin_unlock(&osd->od_osfs_lock);
8155 * once last export (we don't count self-export) disappeared
8156 * osd can be released
8158 static int osd_obd_disconnect(struct obd_export *exp)
8160 struct obd_device *obd = exp->exp_obd;
8161 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8162 int rc, release = 0;
8166 /* Only disconnect the underlying layers on the final disconnect. */
8167 spin_lock(&osd->od_osfs_lock);
8169 if (osd->od_connects == 0)
8171 spin_unlock(&osd->od_osfs_lock);
8173 rc = class_disconnect(exp); /* bz 9811 */
8175 if (rc == 0 && release)
8176 class_manual_cleanup(obd);
8180 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8181 struct lu_device *dev)
8183 struct osd_device *osd = osd_dev(dev);
8184 struct lr_server_data *lsd =
8185 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8190 if (osd->od_quota_slave_md != NULL) {
8191 /* set up quota slave objects for inode */
8192 result = qsd_prepare(env, osd->od_quota_slave_md);
8197 if (osd->od_quota_slave_dt != NULL) {
8198 /* set up quota slave objects for block */
8199 result = qsd_prepare(env, osd->od_quota_slave_dt);
8205 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8206 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8207 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8208 osd->od_index_in_idif = 1;
8210 osd->od_index_in_idif = 0;
8211 result = osd_register_proc_index_in_idif(osd);
8216 osd->od_index_in_idif = 1;
8220 result = osd_fid_init(env, osd);
8226 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8230 * see include/lu_object.h for the details.
8232 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8233 struct lu_fid *fid, struct lu_object *parent,
8234 const struct lu_name *name)
8236 struct osd_device *osd = osd_dev(d);
8238 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8241 static const struct lu_object_operations osd_lu_obj_ops = {
8242 .loo_object_init = osd_object_init,
8243 .loo_object_delete = osd_object_delete,
8244 .loo_object_release = osd_object_release,
8245 .loo_object_free = osd_object_free,
8246 .loo_object_print = osd_object_print,
8247 .loo_object_invariant = osd_object_invariant
8250 const struct lu_device_operations osd_lu_ops = {
8251 .ldo_object_alloc = osd_object_alloc,
8252 .ldo_process_config = osd_process_config,
8253 .ldo_recovery_complete = osd_recovery_complete,
8254 .ldo_prepare = osd_prepare,
8255 .ldo_fid_alloc = osd_fid_alloc,
8258 static const struct lu_device_type_operations osd_device_type_ops = {
8259 .ldto_init = osd_type_init,
8260 .ldto_fini = osd_type_fini,
8262 .ldto_start = osd_type_start,
8263 .ldto_stop = osd_type_stop,
8265 .ldto_device_alloc = osd_device_alloc,
8266 .ldto_device_free = osd_device_free,
8268 .ldto_device_init = osd_device_init,
8269 .ldto_device_fini = osd_device_fini
8272 static struct lu_device_type osd_device_type = {
8273 .ldt_tags = LU_DEVICE_DT,
8274 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8275 .ldt_ops = &osd_device_type_ops,
8276 .ldt_ctx_tags = LCT_LOCAL,
8279 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8281 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8282 struct super_block *sb = osd_sb(osd);
8284 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8288 * lprocfs legacy support.
8290 static const struct obd_ops osd_obd_device_ops = {
8291 .o_owner = THIS_MODULE,
8292 .o_connect = osd_obd_connect,
8293 .o_disconnect = osd_obd_disconnect,
8294 .o_health_check = osd_health_check,
8297 static ssize_t track_declares_assert_show(struct kobject *kobj,
8298 struct attribute *attr,
8301 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8304 static ssize_t track_declares_assert_store(struct kobject *kobj,
8305 struct attribute *attr,
8306 const char *buffer, size_t count)
8308 bool track_declares_assert;
8311 rc = kstrtobool(buffer, &track_declares_assert);
8315 ldiskfs_track_declares_assert = track_declares_assert;
8319 LUSTRE_RW_ATTR(track_declares_assert);
8321 static int __init osd_init(void)
8323 struct kobject *kobj;
8326 BUILD_BUG_ON(BH_DXLock >=
8327 sizeof(((struct buffer_head *)0)->b_state) * 8);
8328 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8329 /* please, try to keep osd_thread_info smaller than a page */
8330 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8335 rc = lu_kmem_init(ldiskfs_caches);
8339 #ifdef CONFIG_KALLSYMS
8340 priv_security_file_alloc =
8341 (void *)kallsyms_lookup_name("security_file_alloc");
8344 rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
8345 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8347 lu_kmem_fini(ldiskfs_caches);
8351 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8353 rc = sysfs_create_file(kobj,
8354 &lustre_attr_track_declares_assert.attr);
8357 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8364 static void __exit osd_exit(void)
8366 struct kobject *kobj;
8368 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8370 sysfs_remove_file(kobj,
8371 &lustre_attr_track_declares_assert.attr);
8374 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8375 lu_kmem_fini(ldiskfs_caches);
8378 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8379 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8380 MODULE_VERSION(LUSTRE_VERSION_STRING);
8381 MODULE_LICENSE("GPL");
8383 module_init(osd_init);
8384 module_exit(osd_exit);