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/
31 * lustre/osd/osd_handler.c
33 * Top-level entry points into osd module
35 * Author: Nikita Danilov <nikita@clusterfs.com>
36 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
39 #define DEBUG_SUBSYSTEM S_OSD
41 #include <linux/fs_struct.h>
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #include <linux/uidgid.h>
45 #ifdef HAVE_INODE_IVERSION
46 #include <linux/iversion.h>
48 #define inode_peek_iversion(__inode) ((__inode)->i_version)
51 /* prerequisite for linux/xattr.h */
52 #include <linux/types.h>
53 /* prerequisite for linux/xattr.h */
55 /* XATTR_{REPLACE,CREATE} */
56 #include <linux/xattr.h>
58 #include <ldiskfs/ldiskfs.h>
59 #include <ldiskfs/xattr.h>
60 #include <ldiskfs/ldiskfs_extents.h>
63 * struct OBD_{ALLOC,FREE}*()
65 #include <obd_support.h>
66 #include <libcfs/libcfs.h>
67 /* struct ptlrpc_thread */
68 #include <lustre_net.h>
69 #include <lustre_fid.h>
71 #include <uapi/linux/lustre/lustre_param.h>
72 #include <uapi/linux/lustre/lustre_disk.h>
74 #include "osd_internal.h"
75 #include "osd_dynlocks.h"
77 /* llo_* api support */
78 #include <md_object.h>
79 #include <lustre_quota.h>
81 #include <lustre_linkea.h>
83 /* encoding routines */
84 #include <lustre_crypto.h>
86 /* Maximum EA size is limited by LNET_MTU for remote objects */
87 #define OSD_MAX_EA_SIZE 1048364
90 module_param(ldiskfs_pdo, int, 0644);
91 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
93 int ldiskfs_track_declares_assert;
94 module_param(ldiskfs_track_declares_assert, int, 0644);
95 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
97 /* 1 GiB in 512-byte sectors */
98 int ldiskfs_delayed_unlink_blocks = (1 << (30 - 9));
100 /* Slab to allocate dynlocks */
101 struct kmem_cache *dynlock_cachep;
103 /* Slab to allocate osd_it_ea */
104 struct kmem_cache *osd_itea_cachep;
106 static struct lu_kmem_descr ldiskfs_caches[] = {
108 .ckd_cache = &biop_cachep,
109 .ckd_name = "biop_cache",
110 .ckd_size = sizeof(struct osd_bio_private)
113 .ckd_cache = &dynlock_cachep,
114 .ckd_name = "dynlock_cache",
115 .ckd_size = sizeof(struct dynlock_handle)
118 .ckd_cache = &osd_itea_cachep,
119 .ckd_name = "osd_itea_cache",
120 .ckd_size = sizeof(struct osd_it_ea)
127 static atomic_t osd_mount_seq;
129 static const char dot[] = ".";
130 static const char dotdot[] = "..";
132 static const struct lu_object_operations osd_lu_obj_ops;
133 static const struct dt_object_operations osd_obj_ops;
134 static const struct dt_object_operations osd_obj_otable_it_ops;
135 static const struct dt_index_operations osd_index_iam_ops;
136 static const struct dt_index_operations osd_index_ea_ops;
138 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
139 const struct lu_fid *fid);
140 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
141 struct osd_device *osd);
142 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
143 const struct lu_buf *buf, const char *name, int fl,
144 struct thandle *handle);
146 int osd_trans_declare_op2rb[] = {
147 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
148 [OSD_OT_PUNCH] = OSD_OT_MAX,
149 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
150 [OSD_OT_CREATE] = OSD_OT_DESTROY,
151 [OSD_OT_DESTROY] = OSD_OT_CREATE,
152 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
153 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
154 [OSD_OT_WRITE] = OSD_OT_WRITE,
155 [OSD_OT_INSERT] = OSD_OT_DELETE,
156 [OSD_OT_DELETE] = OSD_OT_INSERT,
157 [OSD_OT_QUOTA] = OSD_OT_MAX,
160 static int osd_has_index(const struct osd_object *obj)
162 return obj->oo_dt.do_index_ops != NULL;
165 static int osd_object_invariant(const struct lu_object *l)
167 return osd_invariant(osd_obj(l));
171 * Concurrency: doesn't matter
173 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
175 struct osd_thread_info *oti = osd_oti_get(env);
177 return oti->oti_w_locks > 0 && o->oo_owner == env;
181 * Concurrency: doesn't access mutable data
183 static int osd_root_get(const struct lu_env *env,
184 struct dt_device *dev, struct lu_fid *f)
186 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
191 * the following set of functions are used to maintain per-thread
192 * cache of FID->ino mapping. this mechanism is needed to resolve
193 * FID to inode at dt_insert() which in turn stores ino in the
194 * directory entries to keep ldiskfs compatible with ext[34].
195 * due to locking-originated restrictions we can't lookup ino
196 * using LU cache (deadlock is possible). lookup using OI is quite
197 * expensive. so instead we maintain this cache and methods like
198 * dt_create() fill it. so in the majority of cases dt_insert() is
199 * able to find needed mapping in lockless manner.
201 static struct osd_idmap_cache *
202 osd_idc_find(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 = oti->oti_ins_cache;
209 for (i = 0; i < oti->oti_ins_cache_used; i++) {
210 if (!lu_fid_eq(&idc[i].oic_fid, fid))
212 if (idc[i].oic_dev != osd)
221 static struct osd_idmap_cache *
222 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
223 const struct lu_fid *fid)
225 struct osd_thread_info *oti = osd_oti_get(env);
226 struct osd_idmap_cache *idc;
229 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
230 i = oti->oti_ins_cache_size * 2;
232 i = OSD_INS_CACHE_SIZE;
233 OBD_ALLOC_PTR_ARRAY_LARGE(idc, i);
235 return ERR_PTR(-ENOMEM);
236 if (oti->oti_ins_cache != NULL) {
237 memcpy(idc, oti->oti_ins_cache,
238 oti->oti_ins_cache_used * sizeof(*idc));
239 OBD_FREE_PTR_ARRAY_LARGE(oti->oti_ins_cache,
240 oti->oti_ins_cache_used);
242 oti->oti_ins_cache = idc;
243 oti->oti_ins_cache_size = i;
246 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
249 idc->oic_lid.oii_ino = 0;
250 idc->oic_lid.oii_gen = 0;
257 * lookup mapping for the given fid in the cache, initialize a
258 * new one if not found. the initialization checks whether the
259 * object is local or remote. for local objects, OI is used to
260 * learn ino/generation. the function is used when the caller
261 * has no information about the object, e.g. at dt_insert().
263 static struct osd_idmap_cache *
264 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
265 const struct lu_fid *fid)
267 struct osd_idmap_cache *idc;
270 idc = osd_idc_find(env, osd, fid);
271 LASSERT(!IS_ERR(idc));
275 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
276 osd->od_svname, PFID(fid));
278 /* new mapping is needed */
279 idc = osd_idc_add(env, osd, fid);
281 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
282 osd->od_svname, PFID(fid), PTR_ERR(idc));
287 rc = osd_remote_fid(env, osd, fid);
288 if (unlikely(rc < 0))
292 /* the object is local, lookup in OI */
293 /* XXX: probably cheaper to lookup in LU first? */
294 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
296 if (unlikely(rc < 0)) {
297 CERROR("can't lookup: rc = %d\n", rc);
301 /* the object is remote */
308 static void osd_idc_dump_lma(const struct lu_env *env,
309 struct osd_device *osd,
313 struct osd_thread_info *info = osd_oti_get(env);
314 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
315 const struct lu_fid *fid;
316 struct osd_inode_id lid;
320 inode = osd_ldiskfs_iget(osd_sb(osd), ino);
322 CERROR("%s: can't get inode %lu: rc = %d\n",
323 osd->od_svname, ino, (int)PTR_ERR(inode));
326 if (is_bad_inode(inode)) {
327 CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
330 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
332 CERROR("%s: can't get LMA for %lu: rc = %d\n",
333 osd->od_svname, ino, rc);
336 fid = &loa->loa_lma.lma_self_fid;
337 LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
338 PFID(fid), ino, (unsigned)inode->i_generation);
341 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
343 CERROR("%s: can't lookup "DFID": rc = %d\n",
344 osd->od_svname, PFID(fid), rc);
347 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
348 PFID(fid), lid.oii_ino, lid.oii_gen);
353 static void osd_idc_dump_debug(const struct lu_env *env,
354 struct osd_device *osd,
355 const struct lu_fid *fid,
359 struct osd_inode_id lid;
363 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
365 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
366 osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
367 osd_idc_dump_lma(env, osd, lid.oii_ino, false);
369 CERROR("%s: can't lookup "DFID": rc = %d\n",
370 osd->od_svname, PFID(fid), rc);
373 osd_idc_dump_lma(env, osd, ino1, true);
375 osd_idc_dump_lma(env, osd, ino2, true);
379 * lookup mapping for given FID and fill it from the given object.
380 * the object is lolcal by definition.
382 static int osd_idc_find_and_init(const struct lu_env *env,
383 struct osd_device *osd,
384 struct osd_object *obj)
386 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
387 struct osd_idmap_cache *idc;
389 idc = osd_idc_find(env, osd, fid);
390 LASSERT(!IS_ERR(idc));
392 if (obj->oo_inode == NULL)
394 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
395 if (idc->oic_lid.oii_ino) {
396 osd_idc_dump_debug(env, osd, fid,
397 idc->oic_lid.oii_ino,
398 obj->oo_inode->i_ino);
401 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
402 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
407 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
408 osd->od_svname, PFID(fid));
410 /* new mapping is needed */
411 idc = osd_idc_add(env, osd, fid);
413 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
414 osd->od_svname, PFID(fid), PTR_ERR(idc));
418 if (obj->oo_inode != NULL) {
419 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
420 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
426 * OSD object methods.
430 * Concurrency: no concurrent access is possible that early in object
433 static struct lu_object *osd_object_alloc(const struct lu_env *env,
434 const struct lu_object_header *hdr,
437 struct osd_object *mo;
442 struct lu_object_header *h;
443 struct osd_device *o = osd_dev(d);
445 l = &mo->oo_dt.do_lu;
446 if (unlikely(o->od_in_init)) {
453 lu_object_header_init(h);
454 lu_object_init(l, h, d);
455 lu_object_add_top(h, l);
458 dt_object_init(&mo->oo_dt, NULL, d);
459 mo->oo_header = NULL;
462 mo->oo_dt.do_ops = &osd_obj_ops;
463 l->lo_ops = &osd_lu_obj_ops;
464 init_rwsem(&mo->oo_sem);
465 init_rwsem(&mo->oo_ext_idx_sem);
466 spin_lock_init(&mo->oo_guard);
467 INIT_LIST_HEAD(&mo->oo_xattr_list);
473 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
474 struct dentry *dentry, struct lustre_ost_attrs *loa)
478 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
479 (void *)loa, sizeof(*loa));
481 struct lustre_mdt_attrs *lma = &loa->loa_lma;
483 if (rc < sizeof(*lma))
487 lustre_loa_swab(loa, true);
488 /* Check LMA compatibility */
489 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
491 CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
493 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
494 PFID(&lma->lma_self_fid), inode->i_ino, rc);
496 } else if (rc == 0) {
504 * retrieve object from backend ext fs.
506 static struct inode *osd_iget2(struct osd_thread_info *info,
507 struct osd_device *dev, struct osd_inode_id *id,
508 bool special, int *err)
510 struct inode *inode = NULL;
514 * if we look for an inode withing a running
515 * transaction, then we risk to deadlock
516 * osd_dirent_check_repair() breaks this
518 /* LASSERT(current->journal_info == NULL); */
520 inode = osd_ldiskfs_iget_special(osd_sb(dev), id->oii_ino, special);
522 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
523 id->oii_ino, PTR_ERR(inode));
524 } else if (id->oii_gen != OSD_OII_NOGEN &&
525 inode->i_generation != id->oii_gen) {
526 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
527 "i_generation = %u\n",
528 id->oii_ino, id->oii_gen, inode->i_generation);
530 } else if (inode->i_nlink == 0) {
532 * due to parallel readdir and unlink,
533 * we can have dead inode here.
535 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
537 } else if (is_bad_inode(inode)) {
538 CWARN("%s: bad inode: ino = %u: rc = %d\n",
539 osd_dev2name(dev), id->oii_ino, -ENOENT);
541 } else if (osd_is_ea_inode(inode)) {
543 * EA inode is internal ldiskfs object, should don't visible
546 CDEBUG(D_INODE, "EA inode: ino = %u\n", id->oii_ino);
548 } else if ((rc = osd_attach_jinode(inode))) {
549 CDEBUG(D_INODE, "jbd: ino = %u rc = %d\n", id->oii_ino, rc);
551 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
552 if (id->oii_gen == OSD_OII_NOGEN)
553 osd_id_gen(id, inode->i_ino, inode->i_generation);
556 * Do not update file c/mtime in ldiskfs.
557 * NB: we don't have any lock to protect this because we don't
558 * have reference on osd_object now, but contention with
559 * another lookup + attr_set can't happen in the tiny window
560 * between if (...) and set S_NOCMTIME.
562 if (!(inode->i_flags & S_NOCMTIME))
563 inode->i_flags |= S_NOCMTIME;
571 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
572 struct osd_inode_id *id)
577 inode = osd_iget2(info, dev, id, 0, &rc);
587 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
588 handle_t *handle, struct dentry *child,
589 struct inode *inode, struct htree_lock *hlock)
593 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
594 if (rc == -ENOBUFS || rc == -ENOSPC) {
595 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
596 struct inode *parent = child->d_parent->d_inode;
597 struct lu_fid *fid = NULL;
598 char fidstr[FID_LEN + 1] = "unknown";
600 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
602 fid = &loa->loa_lma.lma_self_fid;
603 } else if (rc2 == -ENODATA) {
604 if (unlikely(is_root_inode(parent))) {
605 fid = &info->oti_fid3;
606 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
607 } else if (!osd->od_is_ost && osd->od_index == 0) {
608 fid = &info->oti_fid3;
609 lu_igif_build(fid, parent->i_ino,
610 parent->i_generation);
615 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
617 /* below message is checked in sanity.sh test_129 */
619 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
620 osd_name(osd), parent->i_ino, fidstr);
622 rc = 0; /* ignore such error now */
623 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
624 osd_name(osd), parent->i_ino, fidstr);
634 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
635 struct osd_inode_id *id, struct lu_fid *fid)
637 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
641 inode = osd_iget(info, dev, id);
645 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
647 *fid = loa->loa_lma.lma_self_fid;
648 } else if (rc == -ENODATA) {
649 if (unlikely(is_root_inode(inode)))
650 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
652 lu_igif_build(fid, inode->i_ino, inode->i_generation);
660 static struct inode *osd_iget_check(struct osd_thread_info *info,
661 struct osd_device *dev,
662 const struct lu_fid *fid,
663 struct osd_inode_id *id,
667 bool special = false;
673 * The cached OI mapping is trustable. If we cannot locate the inode
674 * via the cached OI mapping, then return the failure to the caller
675 * directly without further OI checking.
678 if (unlikely(fid_is_acct(fid)))
681 inode = osd_iget2(info, dev, id, special, &rc);
683 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
686 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
687 PFID(fid), id->oii_ino, rc);
693 __u32 saved_ino = id->oii_ino;
694 __u32 saved_gen = id->oii_gen;
697 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
699 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
701 * XXX: There are four possible cases:
703 * Backup/restore caused the OI invalid.
705 * Someone unlinked the object but NOT removed
706 * the OI mapping, such as mount target device
707 * as ldiskfs, and modify something directly.
709 * Someone just removed the object between the
710 * former oi_lookup and the iget. It is normal.
711 * 4. Other failure cases.
713 * Generally, when the device is mounted, it will
714 * auto check whether the system is restored from
715 * file-level backup or not. We trust such detect
716 * to distinguish the 1st case from the 2nd case:
717 * if the OI files are consistent but may contain
718 * stale OI mappings because of case 2, if iget()
719 * returns -ENOENT or -ESTALE, then it should be
724 * If the OI mapping was in OI file before the
725 * osd_iget_check(), but now, it is disappear,
726 * then it must be removed by race. That is a
732 * It is the OI scrub updated the OI mapping by race.
733 * The new OI mapping must be valid.
735 if (saved_ino != id->oii_ino ||
736 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
745 if (dev->od_scrub.os_scrub.os_file.sf_flags &
748 * It still can be the case 2, but we cannot
749 * distinguish it from the case 1. So return
750 * -EREMCHG to block current operation until
751 * OI scrub rebuilt the OI mappings.
760 if (inode->i_generation == id->oii_gen)
780 * \retval +v: new filter_fid does not contain self-fid
781 * \retval 0: filter_fid_18_23, contains self-fid
782 * \retval -v: other failure cases
784 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
785 struct dentry *dentry, struct lu_fid *fid)
787 struct filter_fid *ff = &info->oti_ff;
788 struct ost_id *ostid = &info->oti_ostid;
791 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
792 if (rc == sizeof(struct filter_fid_18_23)) {
793 struct filter_fid_18_23 *ff_old = (void *)ff;
795 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
796 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
798 * XXX: use 0 as the index for compatibility, the caller will
799 * handle index related issues when necessary.
802 ostid_to_fid(fid, ostid, 0);
803 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
805 } else if (rc >= 0) {
812 static int osd_lma_self_repair(struct osd_thread_info *info,
813 struct osd_device *osd, struct inode *inode,
814 const struct lu_fid *fid, __u32 compat)
819 LASSERT(current->journal_info == NULL);
821 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
822 osd_dto_credits_noquota[DTO_XATTR_SET]);
825 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
830 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
832 CWARN("%s: cannot self repair the LMA: rc = %d\n",
834 ldiskfs_journal_stop(jh);
838 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
840 struct osd_thread_info *info = osd_oti_get(env);
841 struct osd_device *osd = osd_obj2dev(obj);
842 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
843 struct lustre_mdt_attrs *lma = &loa->loa_lma;
844 struct inode *inode = obj->oo_inode;
845 struct dentry *dentry = &info->oti_obj_dentry;
846 struct lu_fid *fid = NULL;
847 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
852 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
853 (void *)loa, sizeof(*loa));
854 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
855 fid = &lma->lma_self_fid;
856 rc = osd_get_idif(info, inode, dentry, fid);
857 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
859 * For the given OST-object, if it has neither LMA nor
860 * FID in XATTR_NAME_FID, then the given FID (which is
861 * contained in the @obj, from client RPC for locating
862 * the OST-object) is trusted. We use it to generate
865 osd_lma_self_repair(info, osd, inode, rfid,
876 lustre_lma_swab(lma);
877 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
878 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
879 S_ISREG(inode->i_mode)))) {
880 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
881 "fid = "DFID", ino = %lu\n", osd_name(osd),
882 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
883 PFID(rfid), inode->i_ino);
886 fid = &lma->lma_self_fid;
887 if (lma->lma_compat & LMAC_STRIPE_INFO &&
889 obj->oo_pfid_in_lma = 1;
890 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
892 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
896 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
897 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
898 struct ost_id *oi = &info->oti_ostid;
899 struct lu_fid *fid1 = &info->oti_fid3;
900 __u32 idx = fid_idif_ost_idx(rfid);
903 * For old IDIF, the OST index is not part of the IDIF,
904 * Means that different OSTs may have the same IDIFs.
905 * Under such case, we need to make some compatible
906 * check to make sure to trigger OI scrub properly.
908 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
909 /* Given @rfid is new, LMA is old. */
910 fid_to_ostid(fid, oi);
911 ostid_to_fid(fid1, oi, idx);
912 if (lu_fid_eq(fid1, rfid)) {
913 if (osd->od_index_in_idif)
914 osd_lma_self_repair(info, osd,
928 struct osd_check_lmv_buf {
929 /* please keep it as first member */
930 struct dir_context ctx;
931 struct osd_thread_info *oclb_info;
932 struct osd_device *oclb_dev;
938 * It is called internally by ->iterate*() to filter out the
939 * local slave object's FID of the striped directory.
941 * \retval 1 found the local slave's FID
942 * \retval 0 continue to check next item
943 * \retval -ve for failure
945 #ifdef HAVE_FILLDIR_USE_CTX
946 static FILLDIR_TYPE do_osd_stripe_dir_filldir(struct dir_context *buf,
948 static int osd_stripe_dir_filldir(void *buf,
950 const char *name, int namelen,
951 loff_t offset, __u64 ino, unsigned int d_type)
953 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
954 struct osd_thread_info *oti = oclb->oclb_info;
955 struct lu_fid *fid = &oti->oti_fid3;
956 struct osd_inode_id *id = &oti->oti_id3;
957 struct osd_inode_id id2;
958 struct osd_device *dev = oclb->oclb_dev;
967 sscanf(name + 1, SFID, RFID(fid));
968 if (!fid_is_sane(fid))
971 if (osd_remote_fid(oti->oti_env, dev, fid))
974 osd_id_gen(id, ino, OSD_OII_NOGEN);
975 inode = osd_iget(oti, dev, id);
977 return PTR_ERR(inode);
980 osd_add_oi_cache(oti, dev, id, fid);
981 /* Check shard by scrub only if it has a problem with OI */
982 if (osd_oi_lookup(oti, dev, fid, &id2, 0) || !osd_id_eq(id, &id2))
983 osd_scrub_oi_insert(dev, fid, id, true);
984 oclb->oclb_found = true;
989 WRAP_FILLDIR_FN(do_, osd_stripe_dir_filldir)
992 * When lookup item under striped directory, we need to locate the master
993 * MDT-object of the striped directory firstly, then the client will send
994 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
995 * and the item's name. If the system is restored from MDT file level backup,
996 * then before the OI scrub completely built the OI files, the OI mappings of
997 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
998 * not a problem for the master MDT-object. Because when locate the master
999 * MDT-object, we will do name based lookup (for the striped directory itself)
1000 * firstly, during such process we can setup the correct OI mapping for the
1001 * master MDT-object. But it will be trouble for the slave MDT-object. Because
1002 * the client will not trigger name based lookup on the MDT to locate the slave
1003 * MDT-object before locating item under the striped directory, then when
1004 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
1005 * is invalid and does not know what the right OI mapping is, then the MDT has
1006 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
1007 * the OI file, related OI mapping is unknown yet, please try again later. And
1008 * then client will re-try the RPC again and again until related OI mapping has
1009 * been updated. That is quite inefficient.
1011 * To resolve above trouble, we will handle it as the following two cases:
1013 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
1014 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
1015 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
1016 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
1017 * directly. Please check osd_fid_lookup().
1019 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
1020 * Under such case, during lookup the master MDT-object, we will lookup the
1021 * slave MDT-object via readdir against the master MDT-object, because the
1022 * slave MDT-objects information are stored as sub-directories with the name
1023 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
1024 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
1025 * the correct OI mapping for the slave MDT-object.
1027 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
1028 struct inode *inode)
1030 struct lu_buf *buf = &oti->oti_big_buf;
1032 struct lmv_mds_md_v1 *lmv1;
1033 struct osd_check_lmv_buf oclb = {
1034 .ctx.actor = osd_stripe_dir_filldir,
1037 .oclb_found = false,
1042 /* We should use the VFS layer to create a real dentry. */
1043 oti->oti_obj_dentry.d_inode = inode;
1044 oti->oti_obj_dentry.d_sb = inode->i_sb;
1046 filp = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
1051 filp->f_mode |= FMODE_64BITHASH;
1055 rc = __osd_xattr_get(inode, filp->f_path.dentry, XATTR_NAME_LMV,
1056 buf->lb_buf, buf->lb_len);
1057 if (rc == -ERANGE) {
1058 rc = __osd_xattr_get(inode, filp->f_path.dentry,
1059 XATTR_NAME_LMV, NULL, 0);
1061 lu_buf_realloc(buf, rc);
1062 if (buf->lb_buf == NULL)
1063 GOTO(out, rc = -ENOMEM);
1069 if (unlikely(rc == 0 || rc == -ENODATA))
1075 if (unlikely(buf->lb_buf == NULL)) {
1076 lu_buf_realloc(buf, rc);
1077 if (buf->lb_buf == NULL)
1078 GOTO(out, rc = -ENOMEM);
1084 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1088 oclb.oclb_items = 0;
1089 rc = iterate_dir(filp, &oclb.ctx);
1090 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1091 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1096 "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
1097 osd_ino2name(inode), inode->i_ino, inode->i_generation,
1106 * Is object in scrub inconsistent/stale list.
1108 * \a scrub has two lists, os_inconsistent_items contains mappings to fix, while
1109 * os_stale_items contains mappings failed to fix.
1111 static bool fid_in_scrub_list(struct lustre_scrub *scrub,
1112 const struct list_head *list,
1113 const struct lu_fid *fid)
1115 struct osd_inconsistent_item *oii;
1117 if (list_empty(list))
1120 spin_lock(&scrub->os_lock);
1121 list_for_each_entry(oii, list, oii_list) {
1122 if (lu_fid_eq(fid, &oii->oii_cache.oic_fid)) {
1123 spin_unlock(&scrub->os_lock);
1127 spin_unlock(&scrub->os_lock);
1132 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1133 const struct lu_fid *fid,
1134 const struct lu_object_conf *conf)
1136 struct osd_thread_info *info;
1137 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1138 struct osd_device *dev;
1139 struct osd_idmap_cache *oic;
1140 struct osd_inode_id *id;
1141 struct inode *inode = NULL;
1142 struct lustre_scrub *scrub;
1143 struct scrub_file *sf;
1144 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1149 bool remote = false;
1150 bool trusted = true;
1151 bool updated = false;
1152 bool checked = false;
1157 LINVRNT(osd_invariant(obj));
1158 LASSERT(obj->oo_inode == NULL);
1160 if (fid_is_sane(fid) == 0) {
1161 CERROR("%s: invalid FID "DFID"\n", ldev->ld_obd->obd_name,
1167 dev = osd_dev(ldev);
1168 scrub = &dev->od_scrub.os_scrub;
1169 sf = &scrub->os_file;
1170 info = osd_oti_get(env);
1172 oic = &info->oti_cache;
1174 if (CFS_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1178 * For the object is created as locking anchor, or for the object to
1179 * be created on disk. No need to osd_oi_lookup() at here because FID
1180 * shouldn't never be re-used, if it's really a duplicate FID from
1181 * unexpected reason, we should be able to detect it later by calling
1182 * do_create->osd_oi_insert().
1184 if (conf && conf->loc_flags & LOC_F_NEW)
1187 /* Search order: 1. per-thread cache. */
1188 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1193 /* Search order: 2. OI scrub pending list. */
1195 memset(id, 0, sizeof(struct osd_inode_id));
1196 if (fid_in_scrub_list(scrub, &scrub->os_inconsistent_items, fid) &&
1198 RETURN(-EINPROGRESS);
1200 stale = fid_in_scrub_list(scrub, &scrub->os_stale_items, fid);
1201 if (stale && CFS_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1205 * The OI mapping in the OI file can be updated by the OI scrub
1206 * when we locate the inode via FID. So it may be not trustable.
1210 /* Search order: 3. OI files. */
1211 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1212 if (result == -ENOENT) {
1213 if (!fid_is_norm(fid) ||
1214 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1215 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1217 GOTO(out, result = 0);
1222 /* -ESTALE is returned if inode of OST object doesn't exist */
1223 if (result == -ESTALE &&
1224 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1225 GOTO(out, result = 0);
1232 obj->oo_inode = NULL;
1233 /* for later passes through checks, not true on first pass */
1234 if (!IS_ERR_OR_NULL(inode))
1237 inode = osd_iget_check(info, dev, fid, id, trusted);
1238 if (!IS_ERR(inode)) {
1239 obj->oo_inode = inode;
1247 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1250 result = PTR_ERR(inode);
1251 if (result == -ENOENT || result == -ESTALE)
1252 GOTO(out, result = 0);
1254 if (result != -EREMCHG)
1258 /* don't trigger repeatedly for stale mapping */
1260 GOTO(out, result = -ESTALE);
1263 * We still have chance to get the valid inode: for the
1264 * object which is referenced by remote name entry, the
1265 * object on the local MDT will be linked under the dir
1266 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1268 * We do not know whether the object for the given FID
1269 * is referenced by some remote name entry or not, and
1270 * especially for DNE II, a multiple-linked object may
1271 * have many name entries reside on many MDTs.
1273 * To simplify the operation, OSD will not distinguish
1274 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1275 * only happened for the RPC from other MDT during the
1276 * OI scrub, or for the client side RPC with FID only,
1277 * such as FID to path, or from old connected client.
1280 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1284 flags |= SS_AUTO_PARTIAL;
1285 flags &= ~SS_AUTO_FULL;
1290 if (scrub->os_running) {
1291 if (scrub->os_partial_scan && !scrub->os_in_join)
1294 if (IS_ERR_OR_NULL(inode) || result) {
1295 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1296 GOTO(out, result = -EINPROGRESS);
1300 LASSERT(obj->oo_inode == inode);
1302 osd_scrub_oi_insert(dev, fid, id, true);
1306 if (dev->od_scrub.os_scrub.os_auto_scrub_interval == AS_NEVER) {
1308 GOTO(out, result = -EREMCHG);
1311 LASSERT(obj->oo_inode == inode);
1313 osd_add_oi_cache(info, dev, id, fid);
1318 if (IS_ERR_OR_NULL(inode) || result)
1319 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1321 rc1 = osd_scrub_start(env, dev, flags);
1322 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
1323 "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
1324 osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
1325 if (rc1 && rc1 != -EALREADY)
1326 GOTO(out, result = -EREMCHG);
1328 if (IS_ERR_OR_NULL(inode) || result)
1329 GOTO(out, result = -EINPROGRESS);
1332 LASSERT(obj->oo_inode == inode);
1337 if (unlikely(obj->oo_header))
1340 result = osd_check_lma(env, obj);
1344 LASSERTF(id->oii_ino == inode->i_ino &&
1345 id->oii_gen == inode->i_generation,
1346 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1347 PFID(fid), id->oii_ino, id->oii_gen,
1348 inode->i_ino, inode->i_generation);
1350 saved_ino = inode->i_ino;
1351 saved_gen = inode->i_generation;
1353 if (unlikely(result == -ENODATA)) {
1355 * If the OI scrub updated the OI mapping by race, it
1356 * must be valid. Trust the inode that has no LMA EA.
1361 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1364 * The OI mapping is still there, the inode is still
1365 * valid. It is just becaues the inode has no LMA EA.
1367 if (saved_ino == id->oii_ino &&
1368 saved_gen == id->oii_gen)
1372 * It is the OI scrub updated the OI mapping by race.
1373 * The new OI mapping must be valid.
1381 * "result == -ENOENT" means that the OI mappinghas been
1382 * removed by race, so the inode belongs to other object.
1384 * Others error can be returned directly.
1386 if (result == -ENOENT) {
1387 obj->oo_inode = NULL;
1392 if (result != -EREMCHG)
1398 * if two OST objects map to the same inode, and inode mode is
1399 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1400 * reserved by precreate, and not written yet, in this case, don't
1401 * set inode for the object whose FID mismatch, so that it can create
1402 * inode and not block precreate.
1404 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1405 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1406 obj->oo_inode = NULL;
1407 GOTO(out, result = 0);
1410 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1412 * "result == -ENOENT" means the cached OI mapping has been removed
1413 * from the OI file by race, above inode belongs to other object.
1415 if (result == -ENOENT) {
1416 obj->oo_inode = NULL;
1417 GOTO(out, result = 0);
1423 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1425 osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
1431 * It is the OI scrub updated the OI mapping by race.
1432 * The new OI mapping must be valid.
1440 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1441 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1443 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1445 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1447 obj->oo_pfid_in_lma = 1;
1448 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1450 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1451 } else if (result != -ENODATA) {
1456 obj->oo_compat_dot_created = 1;
1457 obj->oo_compat_dotdot_created = 1;
1459 if (S_ISDIR(inode->i_mode) &&
1460 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1461 osd_check_lmv(info, dev, inode);
1463 result = osd_attach_jinode(inode);
1468 GOTO(out, result = 0);
1470 LASSERT(!obj->oo_hl_head);
1471 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1473 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1476 if (!result && stale)
1477 osd_scrub_oi_resurrect(scrub, fid);
1479 if (result || !obj->oo_inode) {
1480 if (!IS_ERR_OR_NULL(inode))
1483 obj->oo_inode = NULL;
1485 fid_zero(&oic->oic_fid);
1488 LINVRNT(osd_invariant(obj));
1493 * Concurrency: shouldn't matter.
1495 static void osd_object_init0(struct osd_object *obj)
1497 LASSERT(obj->oo_inode != NULL);
1498 obj->oo_dt.do_body_ops = &osd_body_ops;
1499 obj->oo_dt.do_lu.lo_header->loh_attr |=
1500 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1504 * Concurrency: no concurrent access is possible that early in object
1507 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1508 const struct lu_object_conf *conf)
1510 struct osd_object *obj = osd_obj(l);
1513 LINVRNT(osd_invariant(obj));
1515 if (CFS_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1516 cfs_fail_val == 2) {
1517 struct osd_thread_info *info = osd_oti_get(env);
1518 struct osd_idmap_cache *oic = &info->oti_cache;
1519 /* invalidate thread cache */
1520 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1522 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1523 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1524 l->lo_header->loh_attr |= LOHA_EXISTS;
1528 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1529 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1530 if (result == 0 && obj->oo_inode != NULL) {
1531 struct osd_thread_info *oti = osd_oti_get(env);
1532 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1534 osd_object_init0(obj);
1535 if (unlikely(obj->oo_header))
1538 result = osd_get_lma(oti, obj->oo_inode,
1539 &oti->oti_obj_dentry, loa);
1542 * Convert LMAI flags to lustre LMA flags
1543 * and cache it to oo_lma_flags
1546 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1547 } else if (result == -ENODATA) {
1551 atomic_set(&obj->oo_dirent_count, LU_DIRENT_COUNT_UNSET);
1553 LINVRNT(osd_invariant(obj));
1558 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1559 * The left part is for value, binary mode.
1561 struct osd_xattr_entry {
1562 struct list_head oxe_list;
1566 struct rcu_head oxe_rcu;
1570 static int osd_oxc_get(struct osd_object *obj, const char *name,
1573 struct osd_xattr_entry *tmp;
1574 struct osd_xattr_entry *oxe = NULL;
1575 size_t namelen = strlen(name);
1579 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1580 if (namelen == tmp->oxe_namelen &&
1581 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1588 GOTO(out, rc = -ENOENT);
1590 if (!oxe->oxe_exist)
1591 GOTO(out, rc = -ENODATA);
1594 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1597 if (buf->lb_buf == NULL)
1600 if (buf->lb_len < rc)
1601 GOTO(out, rc = -ERANGE);
1603 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1610 static void osd_oxc_free(struct rcu_head *head)
1612 struct osd_xattr_entry *oxe;
1614 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1615 OBD_FREE(oxe, oxe->oxe_len);
1618 static void osd_oxc_add(struct osd_object *obj, const char *name,
1619 const char *buf, int buflen)
1621 struct osd_xattr_entry *oxe;
1622 struct osd_xattr_entry *old = NULL;
1623 struct osd_xattr_entry *tmp;
1624 size_t namelen = strlen(name);
1625 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1627 OBD_ALLOC(oxe, len);
1631 INIT_LIST_HEAD(&oxe->oxe_list);
1633 oxe->oxe_namelen = namelen;
1634 memcpy(oxe->oxe_buf, name, namelen);
1636 LASSERT(buf != NULL);
1637 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1638 oxe->oxe_exist = true;
1640 oxe->oxe_exist = false;
1643 /* this should be rarely called, just remove old and add new */
1644 spin_lock(&obj->oo_guard);
1645 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1646 if (namelen == tmp->oxe_namelen &&
1647 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1653 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1654 call_rcu(&old->oxe_rcu, osd_oxc_free);
1656 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1658 spin_unlock(&obj->oo_guard);
1661 static void osd_oxc_del(struct osd_object *obj, const char *name)
1663 struct osd_xattr_entry *oxe;
1664 size_t namelen = strlen(name);
1666 spin_lock(&obj->oo_guard);
1667 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1668 if (namelen == oxe->oxe_namelen &&
1669 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1670 list_del_rcu(&oxe->oxe_list);
1671 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1675 spin_unlock(&obj->oo_guard);
1678 static void osd_oxc_fini(struct osd_object *obj)
1680 struct osd_xattr_entry *oxe, *next;
1682 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1683 list_del(&oxe->oxe_list);
1684 OBD_FREE(oxe, oxe->oxe_len);
1689 * Concurrency: no concurrent access is possible that late in object
1692 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1694 struct osd_object *obj = osd_obj(l);
1695 struct lu_object_header *h = obj->oo_header;
1697 LINVRNT(osd_invariant(obj));
1700 dt_object_fini(&obj->oo_dt);
1701 if (obj->oo_hl_head != NULL)
1702 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1703 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1706 lu_object_header_free(h);
1710 * Concurrency: no concurrent access is possible that late in object
1713 static void osd_index_fini(struct osd_object *o)
1715 struct iam_container *bag;
1717 if (o->oo_dir != NULL) {
1718 bag = &o->oo_dir->od_container;
1719 if (o->oo_inode != NULL) {
1720 if (bag->ic_object == o->oo_inode)
1721 iam_container_fini(bag);
1723 OBD_FREE_PTR(o->oo_dir);
1729 OSD_TXN_OI_DELETE_CREDITS = 20,
1730 OSD_TXN_INODE_DELETE_CREDITS = 20
1737 #if OSD_THANDLE_STATS
1739 * Set time when the handle is allocated
1741 static void osd_th_alloced(struct osd_thandle *oth)
1743 oth->oth_alloced = ktime_get();
1747 * Set time when the handle started
1749 static void osd_th_started(struct osd_thandle *oth)
1751 oth->oth_started = ktime_get();
1755 * Check whether the we deal with this handle for too long.
1757 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1758 ktime_t alloced, ktime_t started,
1761 ktime_t now = ktime_get();
1763 LASSERT(dev != NULL);
1765 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1766 ktime_us_delta(started, alloced));
1767 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1768 ktime_us_delta(closed, started));
1769 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1770 ktime_us_delta(now, closed));
1772 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1773 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1774 oth, now, alloced, started, closed);
1779 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1781 ktime_t __closed = ktime_get(); \
1782 ktime_t __alloced = oth->oth_alloced; \
1783 ktime_t __started = oth->oth_started; \
1786 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1789 #else /* OSD_THANDLE_STATS */
1791 #define osd_th_alloced(h) do {} while(0)
1792 #define osd_th_started(h) do {} while(0)
1793 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1795 #endif /* OSD_THANDLE_STATS */
1798 * in some cases (like overstriped files) the same operations on the same
1799 * objects are declared many times and this may lead to huge number of
1800 * credits which can be a problem and/or cause performance degradation.
1801 * this function is to remember what declarations have been made within
1802 * a given thandle and then skip duplications.
1803 * limit it's scope so that regular small transactions don't need all
1804 * this overhead with allocations, lists.
1805 * also, limit scope to the specific objects like llogs, etc.
1807 static inline bool osd_check_special_fid(const struct lu_fid *f)
1809 if (fid_seq_is_llog(f->f_seq))
1811 if (f->f_seq == FID_SEQ_LOCAL_FILE &&
1812 f->f_oid == MDD_LOV_OBJ_OID)
1817 bool osd_tx_was_declared(const struct lu_env *env, struct osd_thandle *oth,
1818 struct dt_object *dt, enum dt_txn_op op, loff_t pos)
1820 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1821 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1822 struct osd_thread_info *oti = osd_oti_get(env);
1823 struct osd_obj_declare *old;
1828 /* small transactions don't need this overhead */
1829 if (oti->oti_declare_ops[DTO_OBJECT_CREATE] < 10 &&
1830 oti->oti_declare_ops[DTO_WRITE_BASE] < 10)
1833 if (osd_check_special_fid(fid) == 0)
1836 list_for_each_entry(old, &oth->ot_declare_list, old_list) {
1837 if (old->old_op == op && old->old_pos == pos &&
1838 lu_fid_eq(&old->old_fid, fid))
1842 if (unlikely(old == NULL))
1844 old->old_fid = *lu_object_fid(&dt->do_lu);
1847 list_add(&old->old_list, &oth->ot_declare_list);
1851 static void osd_tx_declaration_free(struct osd_thandle *oth)
1853 struct osd_obj_declare *old, *tmp;
1855 list_for_each_entry_safe(old, tmp, &oth->ot_declare_list, old_list) {
1856 list_del_init(&old->old_list);
1862 * Concurrency: doesn't access mutable data.
1864 static int osd_param_is_not_sane(const struct osd_device *dev,
1865 const struct thandle *th)
1867 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1869 return oh->ot_credits > osd_transaction_size(dev);
1873 * Concurrency: shouldn't matter.
1875 static void osd_trans_commit_cb(struct super_block *sb,
1876 struct ldiskfs_journal_cb_entry *jcb, int error)
1878 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1879 struct thandle *th = &oh->ot_super;
1880 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1881 struct osd_device *osd = osd_dev(lud);
1882 struct dt_txn_commit_cb *dcb, *tmp;
1884 LASSERT(oh->ot_handle == NULL);
1887 CERROR("transaction @0x%p commit error: %d\n", th, error);
1889 /* call per-transaction callbacks if any */
1890 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1892 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1893 "commit callback entry: magic=%x name='%s'\n",
1894 dcb->dcb_magic, dcb->dcb_name);
1895 list_del_init(&dcb->dcb_linkage);
1896 dcb->dcb_func(NULL, th, dcb, error);
1899 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1900 wake_up(&osd->od_commit_cb_done);
1906 static struct thandle *osd_trans_create(const struct lu_env *env,
1907 struct dt_device *d)
1909 struct osd_thread_info *oti = osd_oti_get(env);
1910 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1911 struct osd_thandle *oh;
1917 CERROR("%s: someone try to start transaction under "
1918 "readonly mode, should be disabled.\n",
1919 osd_name(osd_dt_dev(d)));
1921 RETURN(ERR_PTR(-EROFS));
1924 /* on pending IO in this thread should left from prev. request */
1925 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1927 sb_start_write(osd_sb(osd_dt_dev(d)));
1929 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1931 sb_end_write(osd_sb(osd_dt_dev(d)));
1932 RETURN(ERR_PTR(-ENOMEM));
1935 oh->ot_quota_trans = &oti->oti_quota_trans;
1936 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1941 oh->oh_declared_ext = 0;
1942 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1943 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1944 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1945 INIT_LIST_HEAD(&oh->ot_declare_list);
1948 memset(oti->oti_declare_ops, 0,
1949 sizeof(oti->oti_declare_ops));
1950 memset(oti->oti_declare_ops_cred, 0,
1951 sizeof(oti->oti_declare_ops_cred));
1952 memset(oti->oti_declare_ops_used, 0,
1953 sizeof(oti->oti_declare_ops_used));
1955 oti->oti_ins_cache_depth++;
1960 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1962 struct osd_thread_info *oti = osd_oti_get(env);
1963 struct osd_thandle *oh;
1965 oh = container_of(th, struct osd_thandle, ot_super);
1966 LASSERT(oh != NULL);
1968 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1969 oti->oti_declare_ops[OSD_OT_CREATE],
1970 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1971 oti->oti_declare_ops_used[OSD_OT_CREATE],
1972 oti->oti_declare_ops[OSD_OT_DESTROY],
1973 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1974 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1975 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1976 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1977 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1978 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1979 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1980 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1981 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1982 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1983 oti->oti_declare_ops[OSD_OT_WRITE],
1984 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1985 oti->oti_declare_ops_used[OSD_OT_WRITE],
1986 oti->oti_declare_ops[OSD_OT_PUNCH],
1987 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1988 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1989 oti->oti_declare_ops[OSD_OT_QUOTA],
1990 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1991 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1992 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1993 oti->oti_declare_ops[OSD_OT_INSERT],
1994 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1995 oti->oti_declare_ops_used[OSD_OT_INSERT],
1996 oti->oti_declare_ops[OSD_OT_DELETE],
1997 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1998 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1999 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
2000 oti->oti_declare_ops[OSD_OT_REF_ADD],
2001 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
2002 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
2003 oti->oti_declare_ops[OSD_OT_REF_DEL],
2004 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
2005 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
2008 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
2009 static void osd_ldiskfs_credits_for_revoke(struct osd_device *osd,
2010 struct osd_thandle *oh,
2011 int *credits, int *revoke)
2013 int blocks = LDISKFS_MAX_EXTENT_DEPTH * oh->oh_declared_ext;
2014 *revoke += ldiskfs_trans_default_revoke_credits(osd_sb(osd)) + blocks;
2017 static void osd_ldiskfs_credits_for_revoke(struct osd_device *osd,
2018 struct osd_thandle *oh,
2019 int *credits, int *revoke)
2021 struct journal_s *journal = LDISKFS_SB(osd_sb(osd))->s_journal;
2022 int blocks, jbsize, records_per_block;
2024 blocks = LDISKFS_MAX_EXTENT_DEPTH * oh->oh_declared_ext;
2025 jbsize = journal->j_blocksize;
2026 jbsize -= sizeof(struct jbd2_journal_block_tail) +
2027 sizeof(jbd2_journal_revoke_header_t);
2028 records_per_block = jbsize / 8;
2029 *credits += (blocks + records_per_block - 1) / records_per_block;
2034 * Concurrency: shouldn't matter.
2036 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
2039 struct osd_thread_info *oti = osd_oti_get(env);
2040 struct osd_device *dev = osd_dt_dev(d);
2042 struct osd_thandle *oh;
2047 LASSERT(current->journal_info == NULL);
2049 oh = container_of(th, struct osd_thandle, ot_super);
2050 LASSERT(oh != NULL);
2051 LASSERT(oh->ot_handle == NULL);
2052 if (unlikely(ldiskfs_track_declares_assert != 0)) {
2053 LASSERT(oti->oti_r_locks == 0);
2054 LASSERT(oti->oti_w_locks == 0);
2057 rc = dt_txn_hook_start(env, d, th);
2061 if (unlikely(osd_param_is_not_sane(dev, th))) {
2062 static unsigned long last_printed;
2063 static int last_credits;
2065 lprocfs_counter_add(dev->od_stats,
2066 LPROC_OSD_TOO_MANY_CREDITS, 1);
2069 * don't make noise on a tiny testing systems
2070 * actual credits misuse will be caught anyway
2072 if (last_credits != oh->ot_credits &&
2073 time_after(jiffies, last_printed +
2074 cfs_time_seconds(60)) &&
2075 osd_transaction_size(dev) > 512) {
2076 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
2077 oh->ot_credits, osd_transaction_size(dev));
2078 osd_trans_dump_creds(env, th);
2080 last_credits = oh->ot_credits;
2081 last_printed = jiffies;
2084 * XXX Limit the credits to 'max_transaction_buffers', and
2085 * let the underlying filesystem to catch the error if
2086 * we really need so many credits.
2088 * This should be removed when we can calculate the
2089 * credits precisely.
2091 oh->ot_credits = osd_transaction_size(dev);
2092 } else if (ldiskfs_track_declares_assert != 0) {
2094 * reserve few credits to prevent an assertion in JBD
2095 * our debugging mechanism will be able to detected
2096 * overuse. this can help to debug single-update
2099 oh->ot_credits += 10;
2100 if (unlikely(osd_param_is_not_sane(dev, th)))
2101 oh->ot_credits = osd_transaction_size(dev);
2104 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
2105 GOTO(out, rc = -EIO);
2108 * we ignore quota checks for system-owned files, but still
2109 * need to count blocks for uid/gid/projid
2111 osd_trans_declare_op(env, oh, OSD_OT_QUOTA, 3);
2114 * XXX temporary stuff. Some abstraction layer should
2117 osd_ldiskfs_credits_for_revoke(dev, oh, &oh->ot_credits, &revoke);
2118 jh = osd_journal_start_with_revoke(osd_sb(dev), LDISKFS_HT_MISC,
2119 oh->ot_credits, revoke);
2123 LASSERT(oti->oti_txns == 0);
2125 atomic_inc(&dev->od_commit_cb_in_flight);
2135 static int osd_seq_exists(const struct lu_env *env,
2136 struct osd_device *osd, u64 seq)
2138 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
2139 struct seq_server_site *ss = osd_seq_site(osd);
2144 LASSERT(ss != NULL);
2145 LASSERT(ss->ss_server_fld != NULL);
2147 rc = osd_fld_lookup(env, osd, seq, range);
2150 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2151 osd_name(osd), seq, rc);
2155 RETURN(ss->ss_node_id == range->lsr_index);
2158 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2160 struct dt_txn_commit_cb *dcb;
2161 struct dt_txn_commit_cb *tmp;
2163 /* call per-transaction stop callbacks if any */
2164 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2166 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2167 "commit callback entry: magic=%x name='%s'\n",
2168 dcb->dcb_magic, dcb->dcb_name);
2169 list_del_init(&dcb->dcb_linkage);
2170 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2175 * Concurrency: shouldn't matter.
2177 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2180 struct osd_thread_info *oti = osd_oti_get(env);
2181 struct osd_thandle *oh;
2182 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2183 struct osd_device *osd = osd_dt_dev(th->th_dev);
2184 struct qsd_instance *qsd = osd_def_qsd(osd);
2185 struct lquota_trans *qtrans;
2186 LIST_HEAD(truncates);
2187 int rc = 0, remove_agents = 0;
2191 oh = container_of(th, struct osd_thandle, ot_super);
2193 remove_agents = oh->ot_remove_agents;
2195 qtrans = oh->ot_quota_trans;
2196 oh->ot_quota_trans = NULL;
2198 osd_tx_declaration_free(oh);
2200 /* move locks to local list, stop tx, execute truncates */
2201 list_splice(&oh->ot_trunc_locks, &truncates);
2203 if (oh->ot_handle != NULL) {
2206 handle_t *hdl = oh->ot_handle;
2209 * add commit callback
2210 * notice we don't do this in osd_trans_start()
2211 * as underlying transaction can change during truncate
2213 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2216 LASSERT(oti->oti_txns == 1);
2219 rc = dt_txn_hook_stop(env, th);
2221 CERROR("%s: failed in transaction hook: rc = %d\n",
2224 osd_trans_stop_cb(oh, rc);
2225 /* hook functions might modify th_sync */
2226 hdl->h_sync = th->th_sync;
2228 oh->ot_handle = NULL;
2229 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2231 CERROR("%s: failed to stop transaction: rc = %d\n",
2232 osd_name(osd), rc2);
2236 /* We preserve the origin behavior of ignoring any
2237 * failures with the underlying punch / truncate
2238 * operation. We do record for debugging if an error
2239 * does occur in the lctl dk logs.
2241 rc2 = osd_process_truncates(env, &truncates);
2243 CERROR("%s: failed truncate process: rc = %d\n",
2244 osd_name(osd), rc2);
2246 osd_trans_stop_cb(oh, th->th_result);
2250 osd_trunc_unlock_all(env, &truncates);
2252 /* inform the quota slave device that the transaction is stopping */
2253 qsd_op_end(env, qsd, qtrans);
2256 * as we want IO to journal and data IO be concurrent, we don't block
2257 * awaiting data IO completion in osd_do_bio(), instead we wait here
2258 * once transaction is submitted to the journal. all reqular requests
2259 * don't do direct IO (except read/write), thus this wait_event becomes
2262 * IMPORTANT: we have to wait till any IO submited by the thread is
2263 * completed otherwise iobuf may be corrupted by different request
2265 wait_event(iobuf->dr_wait,
2266 atomic_read(&iobuf->dr_numreqs) == 0);
2269 rc = iobuf->dr_error;
2271 osd_fini_iobuf(osd, iobuf);
2273 if (unlikely(remove_agents != 0))
2274 osd_process_scheduled_agent_removals(env, osd);
2276 LASSERT(oti->oti_ins_cache_depth > 0);
2277 oti->oti_ins_cache_depth--;
2278 /* reset OI cache for safety */
2279 if (oti->oti_ins_cache_depth == 0)
2280 oti->oti_ins_cache_used = 0;
2282 sb_end_write(osd_sb(osd));
2287 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2289 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2292 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2293 LASSERT(&dcb->dcb_func != NULL);
2294 if (dcb->dcb_flags & DCB_TRANS_STOP)
2295 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2297 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2302 struct osd_delayed_iput_work {
2303 struct work_struct diw_work;
2304 struct inode *diw_inode;
2307 static void osd_delayed_iput_fn(struct work_struct *work)
2309 struct osd_delayed_iput_work *diwork;
2310 struct inode *inode;
2312 diwork = container_of(work, struct osd_delayed_iput_work, diw_work);
2313 inode = diwork->diw_inode;
2314 CDEBUG(D_INODE, "%s: delayed iput (ino=%lu)\n",
2315 inode->i_sb->s_id, inode->i_ino);
2317 OBD_FREE_PTR(diwork);
2320 noinline static void osd_delayed_iput(struct inode *inode,
2321 struct osd_delayed_iput_work *diwork)
2326 INIT_WORK(&diwork->diw_work, osd_delayed_iput_fn);
2327 diwork->diw_inode = inode;
2328 queue_work(LDISKFS_SB(inode->i_sb)->s_misc_wq,
2334 * Called just before object is freed. Releases all resources except for
2335 * object itself (that is released by osd_object_free()).
2337 * Concurrency: no concurrent access is possible that late in object
2340 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2342 struct osd_object *obj = osd_obj(l);
2343 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2344 struct inode *inode = obj->oo_inode;
2345 struct osd_delayed_iput_work *diwork = NULL;
2350 LINVRNT(osd_invariant(obj));
2353 * If object is unlinked remove fid->ino mapping from object index.
2356 osd_index_fini(obj);
2361 if (inode->i_blocks > ldiskfs_delayed_unlink_blocks)
2362 OBD_ALLOC(diwork, sizeof(*diwork));
2364 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2365 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2367 uid = i_uid_read(inode);
2368 gid = i_gid_read(inode);
2369 projid = i_projid_read(inode);
2371 obj->oo_inode = NULL;
2372 osd_delayed_iput(inode, diwork);
2374 /* do not rebalance quota if the caller needs to release memory
2375 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2376 * transaction and risk to deadlock - LU-12178 */
2377 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2380 if (!obj->oo_header && qsd) {
2381 struct osd_thread_info *info = osd_oti_get(env);
2382 struct lquota_id_info *qi = &info->oti_qi;
2384 /* Release granted quota to master if necessary */
2385 qi->lqi_id.qid_uid = uid;
2386 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2388 qi->lqi_id.qid_uid = gid;
2389 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2391 qi->lqi_id.qid_uid = projid;
2392 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2397 * Concurrency: ->loo_object_release() is called under site spin-lock.
2399 static void osd_object_release(const struct lu_env *env,
2400 struct lu_object *l)
2402 struct osd_object *o = osd_obj(l);
2405 * nobody should be releasing a non-destroyed object with nlink=0
2406 * the API allows this, but ldiskfs doesn't like and then report
2407 * this inode as deleted
2409 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2410 o->oo_inode->i_nlink == 0));
2414 * Concurrency: shouldn't matter.
2416 static int osd_object_print(const struct lu_env *env, void *cookie,
2417 lu_printer_t p, const struct lu_object *l)
2419 struct osd_object *o = osd_obj(l);
2420 struct iam_descr *d;
2422 if (o->oo_dir != NULL)
2423 d = o->oo_dir->od_container.ic_descr;
2426 return (*p)(env, cookie,
2427 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2429 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2430 o->oo_inode ? o->oo_inode->i_generation : 0,
2431 d ? d->id_ops->id_name : "plain");
2435 * Concurrency: shouldn't matter.
2437 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2438 struct obd_statfs *sfs, struct obd_statfs_info *info)
2440 struct osd_device *osd = osd_dt_dev(d);
2441 struct super_block *sb = osd_sb(osd);
2442 struct kstatfs *ksfs;
2446 if (unlikely(osd->od_mnt == NULL))
2447 return -EINPROGRESS;
2449 /* osd_lproc.c call this without env, allocate ksfs for that case */
2450 if (unlikely(env == NULL)) {
2451 OBD_ALLOC_PTR(ksfs);
2455 ksfs = &osd_oti_get(env)->oti_ksfs;
2458 result = sb->s_op->statfs(sb->s_root, ksfs);
2462 statfs_pack(sfs, ksfs);
2463 if (unlikely(sb->s_flags & SB_RDONLY))
2464 sfs->os_state |= OS_STATFS_READONLY;
2466 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2468 if (ldiskfs_has_feature_extents(sb))
2469 sfs->os_maxbytes = sb->s_maxbytes;
2471 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2474 * Reserve some space so to avoid fragmenting the filesystem too much.
2475 * Fragmentation not only impacts performance, but can also increase
2476 * metadata overhead significantly, causing grant calculation to be
2479 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2481 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2482 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2483 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2484 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2486 sfs->os_blocks -= reserved;
2487 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2488 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2491 if (unlikely(env == NULL))
2497 * Estimate space needed for file creations. We assume the largest filename
2498 * which is 2^64 - 1, hence a filename of 20 chars.
2499 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2501 #ifdef __LDISKFS_DIR_REC_LEN
2502 # define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2503 #elif defined LDISKFS_DIR_REC_LEN_WITH_DIR
2504 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20, NULL)
2506 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2510 * Concurrency: doesn't access mutable data.
2512 static void osd_conf_get(const struct lu_env *env,
2513 const struct dt_device *dev,
2514 struct dt_device_param *param)
2516 struct osd_device *d = osd_dt_dev(dev);
2517 struct super_block *sb = osd_sb(d);
2518 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2523 * XXX should be taken from not-yet-existing fs abstraction layer.
2525 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2526 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2527 param->ddp_symlink_max = sb->s_blocksize;
2528 param->ddp_mount_type = LDD_MT_LDISKFS;
2529 if (ldiskfs_has_feature_extents(sb))
2530 param->ddp_maxbytes = sb->s_maxbytes;
2532 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2534 * inode are statically allocated, so per-inode space consumption
2535 * is the space consumed by the directory entry
2537 param->ddp_inodespace = PER_OBJ_USAGE;
2539 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2540 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2541 * maximum length to not under-count the actual number of extents
2542 * needed for writing a file if there are sub-optimal block allocations.
2544 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2545 /* worst-case extent insertion metadata overhead */
2546 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2547 param->ddp_mntopts = 0;
2548 if (test_opt(sb, XATTR_USER))
2549 param->ddp_mntopts |= MNTOPT_USERXATTR;
2550 if (test_opt(sb, POSIX_ACL))
2551 param->ddp_mntopts |= MNTOPT_ACL;
2554 * LOD might calculate the max stripe count based on max_ea_size,
2555 * so we need take account in the overhead as well,
2556 * xattr_header + magic + xattr_entry_head
2558 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2559 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2561 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2562 if (ldiskfs_has_feature_ea_inode(sb))
2563 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2567 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2569 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2570 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2573 * Preferred RPC size for efficient disk IO. 4MB shows good
2574 * all-around performance for ldiskfs, but use bigalloc chunk size
2575 * by default if larger.
2577 #if defined(LDISKFS_CLUSTER_SIZE)
2578 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2579 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2582 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2584 param->ddp_t10_cksum_type = 0;
2586 unsigned short interval = blk_integrity_interval(bi);
2587 name = blk_integrity_name(bi);
2597 if (strncmp(name, "T10-DIF-TYPE",
2598 sizeof("T10-DIF-TYPE") - 1) == 0) {
2599 /* also skip "1/2/3-" at end */
2600 const int type_off = sizeof("T10-DIF-TYPE.");
2601 char type_number = name[type_off - 2];
2603 if (interval != 512 && interval != 4096) {
2604 CERROR("%s: unsupported T10PI sector size %u\n",
2605 d->od_svname, interval);
2608 switch (type_number) {
2610 d->od_t10_type = OSD_T10_TYPE1;
2613 d->od_t10_type = OSD_T10_TYPE2;
2616 d->od_t10_type = OSD_T10_TYPE3;
2619 CERROR("%s: unsupported T10PI type %s\n",
2620 d->od_svname, name);
2623 if (strcmp(name + type_off, "CRC") == 0) {
2624 d->od_t10_type |= OSD_T10_TYPE_CRC;
2625 param->ddp_t10_cksum_type = interval == 512 ?
2626 OBD_CKSUM_T10CRC512 :
2628 } else if (strcmp(name + type_off, "IP") == 0) {
2629 d->od_t10_type |= OSD_T10_TYPE_IP;
2630 param->ddp_t10_cksum_type = interval == 512 ?
2631 OBD_CKSUM_T10IP512 :
2634 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2635 d->od_svname, name);
2640 CERROR("%s: unsupported T10PI type '%s'\n",
2641 d->od_svname, name);
2646 param->ddp_has_lseek_data_hole = true;
2649 static struct vfsmount *osd_mnt_get(const struct dt_device *d)
2651 return osd_dt_dev(d)->od_mnt;
2655 * Concurrency: shouldn't matter.
2657 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2660 struct super_block *s = osd_sb(osd_dt_dev(d));
2663 down_read(&s->s_umount);
2664 rc = s->s_op->sync_fs(s, 1);
2665 up_read(&s->s_umount);
2667 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2674 * Start commit for OSD device.
2676 * An implementation of dt_commit_async method for OSD device.
2677 * Asychronously starts underlayng fs sync and thereby a transaction
2680 * \param env environment
2681 * \param d dt device
2683 * \see dt_device_operations
2685 static int osd_commit_async(const struct lu_env *env,
2686 struct dt_device *d)
2688 struct super_block *s = osd_sb(osd_dt_dev(d));
2693 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2694 down_read(&s->s_umount);
2695 rc = s->s_op->sync_fs(s, 0);
2696 up_read(&s->s_umount);
2702 * Concurrency: shouldn't matter.
2704 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2706 struct super_block *sb = osd_sb(osd_dt_dev(d));
2707 struct block_device *dev = sb->s_bdev;
2708 int rc = -EOPNOTSUPP;
2712 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2713 osd_dt_dev(d)->od_svname, (long)dev, rc);
2719 * Note: we do not count into QUOTA here.
2720 * If we mount with --data_journal we may need more.
2722 const int osd_dto_credits_noquota[DTO_NR] = {
2725 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2726 * SINGLEDATA_TRANS_BLOCKS(8)
2727 * XXX Note: maybe iam need more, since iam have more level than
2730 [DTO_INDEX_INSERT] = 16,
2733 * just modify a single entry, probably merge few within a block
2735 [DTO_INDEX_DELETE] = 1,
2739 [DTO_INDEX_UPDATE] = 16,
2741 * 4(inode, inode bits, groups, GDT)
2742 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2744 [DTO_OBJECT_CREATE] = 4,
2746 * 4(inode, inode bits, groups, GDT)
2747 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2749 [DTO_OBJECT_DELETE] = 4,
2751 * Attr set credits (inode)
2753 [DTO_ATTR_SET_BASE] = 1,
2755 * Xattr set. The same as xattr of EXT3.
2756 * DATA_TRANS_BLOCKS(14)
2757 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2758 * are also counted in. Do not know why?
2760 [DTO_XATTR_SET] = 14,
2762 * credits for inode change during write.
2764 [DTO_WRITE_BASE] = 3,
2766 * credits for single block write.
2768 [DTO_WRITE_BLOCK] = 14,
2770 * Attr set credits for chown.
2771 * This is extra credits for setattr, and it is null without quota
2773 [DTO_ATTR_SET_CHOWN] = 0
2776 /* reserve or free quota for some operation */
2777 static int osd_reserve_or_free_quota(const struct lu_env *env,
2778 struct dt_device *dev,
2779 struct lquota_id_info *qi)
2781 struct osd_device *osd = osd_dt_dev(dev);
2782 struct qsd_instance *qsd = NULL;
2788 qsd = osd->od_quota_slave_dt;
2790 qsd = osd->od_quota_slave_md;
2792 rc = qsd_reserve_or_free_quota(env, qsd, qi);
2796 static const struct dt_device_operations osd_dt_ops = {
2797 .dt_root_get = osd_root_get,
2798 .dt_statfs = osd_statfs,
2799 .dt_trans_create = osd_trans_create,
2800 .dt_trans_start = osd_trans_start,
2801 .dt_trans_stop = osd_trans_stop,
2802 .dt_trans_cb_add = osd_trans_cb_add,
2803 .dt_conf_get = osd_conf_get,
2804 .dt_mnt_get = osd_mnt_get,
2805 .dt_sync = osd_sync,
2807 .dt_commit_async = osd_commit_async,
2808 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
2811 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2814 struct osd_object *obj = osd_dt_obj(dt);
2815 struct osd_thread_info *oti = osd_oti_get(env);
2817 LINVRNT(osd_invariant(obj));
2819 LASSERT(obj->oo_owner != env);
2820 down_read_nested(&obj->oo_sem, role);
2822 LASSERT(obj->oo_owner == NULL);
2826 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2829 struct osd_object *obj = osd_dt_obj(dt);
2830 struct osd_thread_info *oti = osd_oti_get(env);
2832 LINVRNT(osd_invariant(obj));
2834 LASSERT(obj->oo_owner != env);
2835 down_write_nested(&obj->oo_sem, role);
2837 LASSERT(obj->oo_owner == NULL);
2838 obj->oo_owner = env;
2842 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2844 struct osd_object *obj = osd_dt_obj(dt);
2845 struct osd_thread_info *oti = osd_oti_get(env);
2847 LINVRNT(osd_invariant(obj));
2849 LASSERT(oti->oti_r_locks > 0);
2851 up_read(&obj->oo_sem);
2854 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2856 struct osd_object *obj = osd_dt_obj(dt);
2857 struct osd_thread_info *oti = osd_oti_get(env);
2859 LINVRNT(osd_invariant(obj));
2861 LASSERT(obj->oo_owner == env);
2862 LASSERT(oti->oti_w_locks > 0);
2864 obj->oo_owner = NULL;
2865 up_write(&obj->oo_sem);
2868 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2870 struct osd_object *obj = osd_dt_obj(dt);
2872 LINVRNT(osd_invariant(obj));
2874 return obj->oo_owner == env;
2877 static void osd_inode_getattr(const struct lu_env *env,
2878 struct inode *inode, struct lu_attr *attr)
2880 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2881 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2882 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2883 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2885 attr->la_atime = inode_get_atime_sec(inode);
2886 attr->la_mtime = inode_get_mtime_sec(inode);
2887 attr->la_ctime = inode_get_ctime_sec(inode);
2888 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2889 attr->la_mode = inode->i_mode;
2890 attr->la_size = i_size_read(inode);
2891 attr->la_blocks = inode->i_blocks;
2892 attr->la_uid = i_uid_read(inode);
2893 attr->la_gid = i_gid_read(inode);
2894 attr->la_projid = i_projid_read(inode);
2895 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2896 attr->la_nlink = inode->i_nlink;
2897 attr->la_rdev = inode->i_rdev;
2898 attr->la_blksize = 1 << inode->i_blkbits;
2899 attr->la_blkbits = inode->i_blkbits;
2901 * Ext4 did not transfer inherit flags from raw inode
2902 * to inode flags, and ext4 internally test raw inode
2903 * @i_flags directly. Instead of patching ext4, we do it here.
2905 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2906 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2909 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2912 struct osd_object *obj = osd_dt_obj(dt);
2913 const struct dt_it_ops *iops;
2919 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2920 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2922 /* directory not initialized yet */
2923 if (!dt->do_index_ops) {
2928 spin_lock(&obj->oo_guard);
2929 *count = atomic_read(&obj->oo_dirent_count);
2930 if (*count == LU_DIRENT_COUNT_UNSET)
2931 atomic_set(&obj->oo_dirent_count, 0);
2932 spin_unlock(&obj->oo_guard);
2933 if (*count != LU_DIRENT_COUNT_UNSET)
2937 iops = &dt->do_index_ops->dio_it;
2938 it = iops->init(env, dt, LUDA_64BITHASH);
2940 GOTO(out, rc = PTR_ERR(it));
2942 rc = iops->load(env, it, 0);
2949 rc = iops->next(env, it);
2951 for (; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2955 if (iops->key_size(env, it) == 0)
2960 if (rc == 1 || rc == -ESTALE)
2964 iops->fini(env, it);
2966 /* If counting dirents failed, use the current count (if any).
2968 * At worst this means the directory will not be split until the
2969 * count can be completed successfully (remount or oo_dirent_count
2970 * incremented by adding new entries). This avoids re-walking
2971 * the whole directory on each access and hitting the same error.
2973 if (rc && *count == 0)
2974 *count = LU_DIRENT_COUNT_UNSET;
2975 atomic_set(&obj->oo_dirent_count, *count);
2979 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2980 struct lu_attr *attr)
2982 struct osd_object *obj = osd_dt_obj(dt);
2985 if (unlikely(!dt_object_exists(dt)))
2987 if (unlikely(obj->oo_destroyed))
2990 LASSERT(!dt_object_remote(dt));
2991 LINVRNT(osd_invariant(obj));
2993 spin_lock(&obj->oo_guard);
2994 osd_inode_getattr(env, obj->oo_inode, attr);
2995 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2996 attr->la_valid |= LA_FLAGS;
2997 attr->la_flags |= LUSTRE_ORPHAN_FL;
2999 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
3000 attr->la_valid |= LA_FLAGS;
3001 attr->la_flags |= LUSTRE_ENCRYPT_FL;
3003 spin_unlock(&obj->oo_guard);
3005 if (S_ISDIR(obj->oo_inode->i_mode) &&
3006 (attr->la_valid & LA_DIRENT_CNT) &&
3007 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
3008 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
3010 attr->la_valid &= ~LA_DIRENT_CNT;
3015 static int osd_declare_attr_qid(const struct lu_env *env,
3016 struct osd_object *obj,
3017 struct osd_thandle *oh, long long bspace,
3018 qid_t old_id, qid_t new_id, bool enforce,
3022 struct osd_thread_info *info = osd_oti_get(env);
3023 struct lquota_id_info *qi = &info->oti_qi;
3025 qi->lqi_type = type;
3026 /* inode accounting */
3027 qi->lqi_is_blk = false;
3029 /* one more inode for the new id ... */
3030 qi->lqi_id.qid_uid = new_id;
3032 /* Reserve credits for the new id */
3033 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
3034 if (rc == -EDQUOT || rc == -EINPROGRESS)
3039 /* and one less inode for the current id */
3040 qi->lqi_id.qid_uid = old_id;
3042 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3043 if (rc == -EDQUOT || rc == -EINPROGRESS)
3048 /* block accounting */
3049 qi->lqi_is_blk = true;
3051 /* more blocks for the new id ... */
3052 qi->lqi_id.qid_uid = new_id;
3053 qi->lqi_space = bspace;
3055 * Credits for the new uid has been reserved, re-use "obj"
3056 * to save credit reservation.
3058 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3059 if (rc == -EDQUOT || rc == -EINPROGRESS)
3064 /* and finally less blocks for the current uid */
3065 qi->lqi_id.qid_uid = old_id;
3066 qi->lqi_space = -bspace;
3067 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3068 if (rc == -EDQUOT || rc == -EINPROGRESS)
3074 static int osd_declare_attr_set(const struct lu_env *env,
3075 struct dt_object *dt,
3076 const struct lu_attr *attr,
3077 struct thandle *handle)
3079 struct osd_thandle *oh;
3080 struct osd_object *obj;
3089 LASSERT(dt != NULL);
3090 LASSERT(handle != NULL);
3092 obj = osd_dt_obj(dt);
3093 LASSERT(osd_invariant(obj));
3095 oh = container_of(handle, struct osd_thandle, ot_super);
3096 LASSERT(oh->ot_handle == NULL);
3098 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
3099 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3101 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
3102 osd_dto_credits_noquota[DTO_XATTR_SET]);
3104 if (attr == NULL || obj->oo_inode == NULL)
3107 bspace = obj->oo_inode->i_blocks << 9;
3108 bspace = toqb(bspace);
3111 * Changing ownership is always preformed by super user, it should not
3112 * fail with EDQUOT unless required explicitly.
3114 * We still need to call the osd_declare_qid() to calculate the journal
3115 * credits for updating quota accounting files and to trigger quota
3116 * space adjustment once the operation is completed.
3118 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
3120 uid = i_uid_read(obj->oo_inode);
3121 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
3122 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
3123 attr->la_uid, enforce, USRQUOTA);
3127 gid = i_gid_read(obj->oo_inode);
3128 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
3129 attr->la_uid, gid, attr->la_gid);
3130 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
3131 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
3132 attr->la_gid, enforce, GRPQUOTA);
3137 #ifdef HAVE_PROJECT_QUOTA
3138 if (attr->la_valid & LA_PROJID) {
3139 __u32 projid = i_projid_read(obj->oo_inode);
3141 enforce = (attr->la_valid & LA_PROJID) &&
3142 (attr->la_projid != projid);
3143 rc = osd_declare_attr_qid(env, obj, oh, bspace,
3144 (qid_t)projid, (qid_t)attr->la_projid,
3150 /* punch must be aware we are dealing with an encrypted file */
3151 if (attr->la_valid & LA_FLAGS && attr->la_flags & LUSTRE_ENCRYPT_FL)
3152 obj->oo_lma_flags |= LUSTRE_ENCRYPT_FL;
3157 static int osd_inode_setattr(const struct lu_env *env,
3158 struct inode *inode, const struct lu_attr *attr)
3160 __u64 bits = attr->la_valid;
3162 /* Only allow set size for regular file */
3163 if (!S_ISREG(inode->i_mode))
3164 bits &= ~(LA_SIZE | LA_BLOCKS);
3169 if (bits & LA_ATIME)
3170 inode_set_atime_to_ts(inode,
3171 osd_inode_time(inode, attr->la_atime));
3172 if (bits & LA_CTIME)
3173 inode_set_ctime_to_ts(inode,
3174 osd_inode_time(inode, attr->la_ctime));
3175 if (bits & LA_MTIME)
3176 inode_set_mtime_to_ts(inode,
3177 osd_inode_time(inode, attr->la_mtime));
3178 if (bits & LA_SIZE) {
3179 spin_lock(&inode->i_lock);
3180 LDISKFS_I(inode)->i_disksize = attr->la_size;
3181 i_size_write(inode, attr->la_size);
3182 spin_unlock(&inode->i_lock);
3186 * OSD should not change "i_blocks" which is used by quota.
3187 * "i_blocks" should be changed by ldiskfs only.
3190 inode->i_mode = (inode->i_mode & S_IFMT) |
3191 (attr->la_mode & ~S_IFMT);
3193 i_uid_write(inode, attr->la_uid);
3195 i_gid_write(inode, attr->la_gid);
3196 if (bits & LA_PROJID)
3197 i_projid_write(inode, attr->la_projid);
3198 if (bits & LA_NLINK)
3199 set_nlink(inode, attr->la_nlink);
3201 inode->i_rdev = attr->la_rdev;
3203 if (bits & LA_FLAGS) {
3204 /* always keep S_NOCMTIME */
3205 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
3207 #if defined(S_ENCRYPTED)
3208 /* Always remove S_ENCRYPTED, because ldiskfs must not be
3209 * aware of encryption status. It is just stored into LMA
3210 * so that it can be forwared to client side.
3212 inode->i_flags &= ~S_ENCRYPTED;
3215 * Ext4 did not transfer inherit flags from
3216 * @inode->i_flags to raw inode i_flags when writing
3217 * flags, we do it explictly here.
3219 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
3220 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
3222 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
3227 #ifdef HAVE_PROJECT_QUOTA
3228 static int osd_transfer_project(struct inode *inode, __u32 projid,
3229 struct thandle *handle)
3231 struct super_block *sb = inode->i_sb;
3232 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3235 struct ldiskfs_iloc iloc;
3236 struct ldiskfs_inode *raw_inode;
3237 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3239 if (!ldiskfs_has_feature_project(sb)) {
3240 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3241 == LDISKFS_DEF_PROJID);
3242 if (projid != LDISKFS_DEF_PROJID)
3248 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3251 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3252 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3255 err = ldiskfs_get_inode_loc(inode, &iloc);
3259 raw_inode = ldiskfs_raw_inode(&iloc);
3260 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3261 struct osd_thandle *oh = container_of(handle,
3265 * try to expand inode size automatically.
3267 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3268 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3276 dquot_initialize(inode);
3277 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3278 if (transfer_to[PRJQUOTA]) {
3279 lock_dquot_transfer(inode);
3280 err = __dquot_transfer(inode, transfer_to);
3281 unlock_dquot_transfer(inode);
3282 dqput(transfer_to[PRJQUOTA]);
3291 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3292 struct thandle *handle)
3296 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3297 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3301 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3302 inode->i_ino, i_uid_read(inode), attr->la_uid,
3303 i_gid_read(inode), attr->la_gid);
3305 dquot_initialize(inode);
3307 if (attr->la_valid & LA_UID)
3308 iattr.ia_valid |= ATTR_UID;
3309 if (attr->la_valid & LA_GID)
3310 iattr.ia_valid |= ATTR_GID;
3311 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3312 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3314 lock_dquot_transfer(inode);
3315 rc = osd_dquot_transfer(&nop_mnt_idmap, inode, &iattr);
3316 unlock_dquot_transfer(inode);
3318 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3319 osd_ino2name(inode), rc);
3324 /* Handle project id transfer here properly */
3325 if (attr->la_valid & LA_PROJID &&
3326 attr->la_projid != i_projid_read(inode)) {
3327 if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
3329 #ifdef HAVE_PROJECT_QUOTA
3330 rc = osd_transfer_project(inode, attr->la_projid, handle);
3335 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3336 osd_ino2name(inode), rc);
3343 static int osd_attr_set(const struct lu_env *env,
3344 struct dt_object *dt,
3345 const struct lu_attr *attr,
3346 struct thandle *handle)
3348 struct osd_thread_info *info = osd_oti_get(env);
3349 struct osd_object *obj = osd_dt_obj(dt);
3350 struct osd_device *osd = osd_obj2dev(obj);
3351 struct inode *inode;
3354 if (!dt_object_exists(dt))
3357 LASSERT(handle != NULL);
3358 LASSERT(!dt_object_remote(dt));
3359 LASSERT(osd_invariant(obj));
3361 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3363 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3364 !osd_obj2dev(obj)->od_is_ost) {
3365 struct osd_thread_info *oti = osd_oti_get(env);
3366 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3367 struct lu_fid *fid1 = &oti->oti_fid;
3368 struct osd_inode_id *id = &oti->oti_id;
3369 struct iam_path_descr *ipd;
3370 struct iam_container *bag;
3371 struct osd_thandle *oh;
3374 fid_cpu_to_be(fid1, fid0);
3375 memset(id, 1, sizeof(*id));
3376 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3377 fid0)->oi_dir.od_container;
3378 ipd = osd_idx_ipd_get(env, bag);
3379 if (unlikely(ipd == NULL))
3382 oh = container_of(handle, struct osd_thandle, ot_super);
3383 rc = iam_update(oh->ot_handle, bag,
3384 (const struct iam_key *)fid1,
3385 (const struct iam_rec *)id, ipd);
3386 osd_ipd_put(env, bag, ipd);
3387 return(rc > 0 ? 0 : rc);
3390 inode = obj->oo_inode;
3392 rc = osd_quota_transfer(inode, attr, handle);
3396 spin_lock(&obj->oo_guard);
3397 rc = osd_inode_setattr(env, inode, attr);
3398 spin_unlock(&obj->oo_guard);
3402 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3404 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3406 if (!(attr->la_valid & LA_FLAGS))
3409 /* If setting LUSTRE_ENCRYPT_FL on an OST object, also set a dummy
3410 * enc ctx xattr, with 2 benefits:
3411 * - setting the LL_XATTR_NAME_ENCRYPTION_CONTEXT xattr internally sets
3412 * the LDISKFS_ENCRYPT_FL flag on the on-disk inode;
3413 * - it makes e2fsprogs happy to see an enc ctx for an inode that has
3414 * the LDISKFS_ENCRYPT_FL flag
3415 * We do not need the actual encryption context on OST objects, it is
3416 * only stored on MDT inodes, at file creation time.
3418 if (!(LDISKFS_I(obj->oo_inode)->i_flags & LDISKFS_ENCRYPT_FL) &&
3419 attr->la_flags & LUSTRE_ENCRYPT_FL && osd->od_is_ost &&
3420 !CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) {
3423 /* use a dummy enc ctx, fine with e2fsprogs */
3424 buf.lb_buf = "\xFF";
3426 rc = osd_xattr_set(env, dt, &buf,
3427 LL_XATTR_NAME_ENCRYPTION_CONTEXT,
3430 CWARN("%s: set "DFID" enc ctx failed: rc = %d\n",
3431 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3435 /* Let's check if there are extra flags need to be set into LMA */
3436 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3437 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3439 LASSERT(!obj->oo_pfid_in_lma);
3441 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3442 &info->oti_ost_attrs);
3446 if ((lma->lma_incompat & lustre_to_lma_flags(attr->la_flags)) ==
3447 lustre_to_lma_flags(attr->la_flags))
3448 /* if lma incompat already has the flags,
3449 * save a useless call to xattr_set
3453 lma->lma_incompat |=
3454 lustre_to_lma_flags(attr->la_flags);
3455 lustre_lma_swab(lma);
3457 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3459 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3460 lma, sizeof(*lma), XATTR_REPLACE);
3462 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3463 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3464 lma->lma_incompat, rc);
3467 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3468 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3475 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3476 struct osd_object *obj,
3477 const char *name, const int namelen)
3479 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3482 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3483 umode_t mode, struct dt_allocation_hint *hint,
3484 struct thandle *th, struct lu_attr *attr)
3487 struct osd_device *osd = osd_obj2dev(obj);
3488 struct osd_thandle *oth;
3489 struct dt_object *parent = NULL;
3490 struct inode *inode;
3491 struct iattr iattr = {
3492 .ia_valid = ATTR_UID | ATTR_GID |
3493 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3494 .ia_ctime.tv_sec = attr->la_ctime,
3495 .ia_mtime.tv_sec = attr->la_mtime,
3496 .ia_atime.tv_sec = attr->la_atime,
3497 .ia_uid = GLOBAL_ROOT_UID,
3498 .ia_gid = GLOBAL_ROOT_GID,
3500 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3502 if (attr->la_valid & LA_UID)
3503 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3504 if (attr->la_valid & LA_GID)
3505 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3507 LINVRNT(osd_invariant(obj));
3508 LASSERT(obj->oo_inode == NULL);
3509 LASSERT(obj->oo_hl_head == NULL);
3511 if (S_ISDIR(mode) && ldiskfs_pdo) {
3513 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3514 if (obj->oo_hl_head == NULL)
3518 oth = container_of(th, struct osd_thandle, ot_super);
3519 LASSERT(oth->ot_handle->h_transaction != NULL);
3521 if (hint != NULL && hint->dah_parent != NULL &&
3522 !dt_object_remote(hint->dah_parent))
3523 parent = hint->dah_parent;
3525 /* if a time component is not valid set it to UTIME_OMIT */
3526 if (!(attr->la_valid & LA_CTIME))
3527 iattr.ia_ctime = omit;
3528 if (!(attr->la_valid & LA_MTIME))
3529 iattr.ia_mtime = omit;
3530 if (!(attr->la_valid & LA_ATIME))
3531 iattr.ia_atime = omit;
3533 inode = ldiskfs_create_inode(oth->ot_handle,
3534 parent ? osd_dt_obj(parent)->oo_inode :
3535 osd_sb(osd)->s_root->d_inode,
3537 if (!IS_ERR(inode)) {
3538 /* Do not update file c/mtime in ldiskfs. */
3539 inode->i_flags |= S_NOCMTIME;
3542 * For new created object, it must be consistent,
3543 * and it is unnecessary to scrub against it.
3545 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3547 obj->oo_inode = inode;
3550 if (obj->oo_hl_head != NULL) {
3551 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3552 obj->oo_hl_head = NULL;
3554 result = PTR_ERR(inode);
3556 LINVRNT(osd_invariant(obj));
3564 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3565 struct lu_attr *attr,
3566 struct dt_allocation_hint *hint,
3567 struct dt_object_format *dof,
3571 struct osd_thandle *oth;
3572 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3574 LASSERT(S_ISDIR(attr->la_mode));
3576 oth = container_of(th, struct osd_thandle, ot_super);
3577 LASSERT(oth->ot_handle->h_transaction != NULL);
3578 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3579 atomic_set(&obj->oo_dirent_count, 0);
3580 result = osd_mkfile(info, obj, mode, hint, th, attr);
3585 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3586 struct lu_attr *attr,
3587 struct dt_allocation_hint *hint,
3588 struct dt_object_format *dof,
3592 struct osd_thandle *oth;
3593 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3595 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3597 LASSERT(S_ISREG(attr->la_mode));
3599 oth = container_of(th, struct osd_thandle, ot_super);
3600 LASSERT(oth->ot_handle->h_transaction != NULL);
3602 result = osd_mkfile(info, obj, mode, hint, th, attr);
3604 LASSERT(obj->oo_inode != NULL);
3605 if (feat->dif_flags & DT_IND_VARKEY)
3606 result = iam_lvar_create(obj->oo_inode,
3607 feat->dif_keysize_max,
3609 feat->dif_recsize_max,
3612 result = iam_lfix_create(obj->oo_inode,
3613 feat->dif_keysize_max,
3615 feat->dif_recsize_max,
3621 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3622 struct lu_attr *attr,
3623 struct dt_allocation_hint *hint,
3624 struct dt_object_format *dof,
3627 LASSERT(S_ISREG(attr->la_mode));
3628 return osd_mkfile(info, obj, (attr->la_mode &
3629 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3633 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3634 struct lu_attr *attr,
3635 struct dt_allocation_hint *hint,
3636 struct dt_object_format *dof,
3639 LASSERT(S_ISLNK(attr->la_mode));
3640 return osd_mkfile(info, obj, (attr->la_mode &
3641 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3645 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3646 struct lu_attr *attr,
3647 struct dt_allocation_hint *hint,
3648 struct dt_object_format *dof,
3651 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3654 LINVRNT(osd_invariant(obj));
3655 LASSERT(obj->oo_inode == NULL);
3656 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3657 S_ISFIFO(mode) || S_ISSOCK(mode));
3659 result = osd_mkfile(info, obj, mode, hint, th, attr);
3661 LASSERT(obj->oo_inode != NULL);
3663 * This inode should be marked dirty for i_rdev. Currently
3664 * that is done in the osd_attr_init().
3666 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3669 LINVRNT(osd_invariant(obj));
3673 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3675 struct dt_allocation_hint *hint,
3676 struct dt_object_format *dof,
3679 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3681 osd_obj_type_f result;
3697 result = osd_mk_index;
3707 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3708 struct dt_object *parent, struct dt_object *child,
3713 ah->dah_parent = parent;
3715 if (parent != NULL && !dt_object_remote(parent)) {
3716 /* will help to find FID->ino at dt_insert("..") */
3717 struct osd_object *pobj = osd_dt_obj(parent);
3719 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3723 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3724 struct lu_attr *attr, struct dt_object_format *dof,
3725 struct thandle *handle)
3727 struct inode *inode = obj->oo_inode;
3728 __u64 valid = attr->la_valid;
3731 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3733 if (dof->dof_type != DFT_NODE)
3734 attr->la_valid &= ~LA_RDEV;
3735 if ((valid & LA_ATIME) &&
3736 (attr->la_atime == inode_get_atime_sec(inode)))
3737 attr->la_valid &= ~LA_ATIME;
3738 if ((valid & LA_CTIME) &&
3739 (attr->la_ctime == inode_get_ctime_sec(inode)))
3740 attr->la_valid &= ~LA_CTIME;
3741 if ((valid & LA_MTIME) &&
3742 (attr->la_mtime == inode_get_mtime_sec(inode)))
3743 attr->la_valid &= ~LA_MTIME;
3745 result = osd_quota_transfer(inode, attr, handle);
3749 if (attr->la_valid != 0) {
3750 result = osd_inode_setattr(info->oti_env, inode, attr);
3752 * The osd_inode_setattr() should always succeed here. The
3753 * only error that could be returned is EDQUOT when we are
3754 * trying to change the UID or GID of the inode. However, this
3755 * should not happen since quota enforcement is no longer
3756 * enabled on ldiskfs (lquota takes care of it).
3758 LASSERTF(result == 0, "%d\n", result);
3759 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3762 attr->la_valid = valid;
3766 * Helper function for osd_create()
3768 * \retval 0, on success
3770 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3771 struct lu_attr *attr, struct dt_allocation_hint *hint,
3772 struct dt_object_format *dof, struct thandle *th)
3777 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3779 /* we drop umask so that permissions we pass are not affected */
3780 umask = current->fs->umask;
3781 current->fs->umask = 0;
3783 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3785 if (likely(obj->oo_inode != NULL)) {
3786 LASSERT(obj->oo_inode->i_state & I_NEW);
3789 * Unlock the inode before attr initialization to avoid
3790 * unnecessary dqget operations. LU-6378
3792 unlock_new_inode(obj->oo_inode);
3795 if (likely(result == 0)) {
3796 osd_attr_init(info, obj, attr, dof, th);
3797 osd_object_init0(obj);
3800 /* restore previous umask value */
3801 current->fs->umask = umask;
3803 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3809 * Helper function for osd_create()
3811 * \retval 0, on success
3813 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3814 const struct lu_fid *fid, struct thandle *th)
3816 struct osd_thread_info *info = osd_oti_get(env);
3817 struct osd_inode_id *id = &info->oti_id;
3818 struct osd_device *osd = osd_obj2dev(obj);
3819 struct osd_thandle *oh;
3822 LASSERT(obj->oo_inode != NULL);
3824 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3827 oh = container_of(th, struct osd_thandle, ot_super);
3828 LASSERT(oh->ot_handle);
3829 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3831 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3832 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_REUSE) && osd->od_is_ost &&
3834 struct lu_fid tfid = *fid;
3837 osd_oi_insert(info, osd, &tfid, id, oh->ot_handle,
3838 OI_CHECK_FLD, NULL);
3839 /* clear NOSCRUB flag so that it can be scrubbed immediately */
3840 ldiskfs_clear_inode_state(obj->oo_inode,
3841 LDISKFS_STATE_LUSTRE_NOSCRUB);
3844 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3845 OI_CHECK_FLD, NULL);
3847 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3848 struct lu_fid next_fid = *fid;
3850 /* insert next object in advance, and map to the same inode */
3852 if (next_fid.f_oid != 0) {
3853 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3854 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3855 OI_CHECK_FLD, NULL);
3856 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3860 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3865 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3866 u64 seq, struct lu_seq_range *range)
3868 struct seq_server_site *ss = osd_seq_site(osd);
3870 if (fid_seq_is_idif(seq)) {
3871 fld_range_set_ost(range);
3872 range->lsr_index = idif_ost_idx(seq);
3876 if (!fid_seq_in_fldb(seq)) {
3877 fld_range_set_mdt(range);
3880 * FIXME: If ss is NULL, it suppose not get lsr_index
3883 range->lsr_index = ss->ss_node_id;
3887 LASSERT(ss != NULL);
3888 fld_range_set_any(range);
3889 /* OSD will only do local fld lookup */
3890 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3893 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3894 struct lu_attr *attr,
3895 struct dt_allocation_hint *hint,
3896 struct dt_object_format *dof,
3897 struct thandle *handle)
3899 struct osd_thandle *oh;
3900 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3906 LASSERT(handle != NULL);
3908 oh = container_of(handle, struct osd_thandle, ot_super);
3909 LASSERT(oh->ot_handle == NULL);
3911 if (osd_tx_was_declared(env, oh, dt, DTO_OBJECT_CREATE, 0))
3915 * EA object consumes more credits than regular object: osd_mk_index
3916 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3917 * leaf_node, could involves the block, block bitmap, groups, GDT
3918 * change for each block, so add 4 * 2 credits in that case.
3920 * The default ACL initialization may consume an additional 16 blocks
3922 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3923 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3926 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3927 * credits for possible default ACL creation in new inode
3929 if (hint && hint->dah_acl_len)
3930 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3931 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3932 sb->s_blocksize_bits);
3934 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3937 * Reuse idle OI block may cause additional one OI block
3940 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3941 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3942 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3943 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3944 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3946 /* will help to find FID->ino mapping at dt_insert() */
3947 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3955 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3956 attr->la_projid, 1, oh, osd_dt_obj(dt),
3957 NULL, OSD_QID_INODE);
3965 * Called to destroy on-disk representation of the object
3967 * Concurrency: must be locked
3969 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3972 struct osd_object *obj = osd_dt_obj(dt);
3973 struct inode *inode = obj->oo_inode;
3974 struct osd_thandle *oh;
3982 oh = container_of(th, struct osd_thandle, ot_super);
3983 LASSERT(oh->ot_handle == NULL);
3985 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3986 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3988 /* For removing agent entry */
3989 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3990 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3993 * Recycle idle OI leaf may cause additional three OI blocks
3996 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3997 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3998 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3999 /* one less inode */
4000 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4001 i_projid_read(inode), -1, oh, obj, NULL,
4005 /* data to be truncated */
4006 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4007 i_projid_read(inode), 0, oh, obj, NULL,
4013 * will help to find FID->ino when this object is being
4016 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
4021 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
4024 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4025 struct osd_object *obj = osd_dt_obj(dt);
4026 struct inode *inode = obj->oo_inode;
4027 struct osd_device *osd = osd_obj2dev(obj);
4028 struct osd_thandle *oh;
4033 oh = container_of(th, struct osd_thandle, ot_super);
4034 LASSERT(oh->ot_handle);
4036 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
4038 if (unlikely(fid_is_acct(fid)))
4041 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
4042 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
4044 CERROR("%s: remove agent entry "DFID": rc = %d\n",
4045 osd_name(osd), PFID(fid), result);
4048 if (S_ISDIR(inode->i_mode)) {
4049 if (inode->i_nlink > 2)
4050 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
4051 osd_name(osd), PFID(fid), inode->i_ino,
4054 spin_lock(&obj->oo_guard);
4056 spin_unlock(&obj->oo_guard);
4057 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4060 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
4062 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
4064 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
4065 result = osd_oi_delete(osd_oti_get(env), osd, fid,
4066 oh->ot_handle, OI_CHECK_FLD);
4068 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
4069 /* XXX: add to ext3 orphan list */
4070 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
4072 /* not needed in the cache anymore */
4073 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
4074 obj->oo_destroyed = 1;
4080 * Put the fid into lustre_mdt_attrs, and then place the structure
4081 * inode's ea. This fid should not be altered during the life time
4084 * \retval +ve, on success
4085 * \retval -ve, on error
4087 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
4089 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
4090 const struct lu_fid *fid, __u32 compat, __u32 incompat)
4092 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4093 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4098 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INLMA))
4101 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
4104 lustre_loa_init(loa, fid, compat, incompat);
4105 lustre_loa_swab(loa, false);
4108 * For the OST device with 256 bytes inode size by default,
4109 * the PFID EA will be stored together with LMA EA to avoid
4110 * performance trouble. Otherwise the PFID EA can be stored
4111 * independently. LU-8998
4113 if ((compat & LMAC_FID_ON_OST) &&
4114 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
4115 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4116 sizeof(*loa), XATTR_CREATE);
4118 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4119 sizeof(*lma), XATTR_CREATE);
4121 * LMA may already exist, but we need to check that all the
4122 * desired compat/incompat flags have been added.
4124 if (unlikely(rc == -EEXIST)) {
4125 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
4126 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
4130 if (rc < sizeof(*lma))
4133 lustre_loa_swab(loa, true);
4134 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
4135 ((compat == 0 && incompat == 0) ||
4136 (!(~lma->lma_compat & compat) &&
4137 !(~lma->lma_incompat & incompat))))
4140 lma->lma_self_fid = *fid;
4141 lma->lma_compat |= compat;
4142 lma->lma_incompat |= incompat;
4143 if (rc == sizeof(*lma)) {
4144 lustre_lma_swab(lma);
4145 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4146 sizeof(*lma), XATTR_REPLACE);
4148 lustre_loa_swab(loa, false);
4149 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4150 sizeof(*loa), XATTR_REPLACE);
4158 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
4159 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
4160 * To have compatilibility with 1.8 ldiskfs driver we need to have
4161 * magic number at start of fid data.
4162 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
4165 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
4166 const struct lu_fid *fid)
4168 if (!fid_is_namespace_visible(fid) ||
4169 CFS_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
4170 param->edp_magic = 0;
4174 param->edp_magic = LDISKFS_LUFID_MAGIC;
4175 param->edp_len = sizeof(struct lu_fid) + 1;
4176 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
4180 * Try to read the fid from inode ea into dt_rec.
4182 * \param fid object fid.
4184 * \retval 0 on success
4186 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
4187 __u32 ino, struct lu_fid *fid,
4188 struct osd_inode_id *id)
4190 struct osd_thread_info *info = osd_oti_get(env);
4191 struct inode *inode;
4195 osd_id_gen(id, ino, OSD_OII_NOGEN);
4196 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
4198 RETURN(PTR_ERR(inode));
4204 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
4206 struct inode *parent_dir,
4207 const struct lu_fid *dot_fid,
4208 const struct lu_fid *dot_dot_fid,
4209 struct osd_thandle *oth)
4211 struct ldiskfs_dentry_param *dot_ldp;
4212 struct ldiskfs_dentry_param *dot_dot_ldp;
4213 __u32 saved_nlink = dir->i_nlink;
4216 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
4219 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
4220 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
4222 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4223 dot_ldp->edp_magic = 0;
4225 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
4226 dir, dot_ldp, dot_dot_ldp);
4228 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
4229 * the subseqent ref_add() will increase the dir->i_nlink
4230 * as 3. That is incorrect for new created directory.
4232 * It looks like hack, because we want to make the OSD API
4233 * to be order-independent for new created directory object
4234 * between dt_insert(..) and ref_add() operations.
4236 * Here, we only restore the in-RAM dir-inode's nlink attr,
4237 * becuase if the nlink attr is not 2, then there will be
4238 * ref_add() called following the dt_insert(..), such call
4239 * will make both the in-RAM and on-disk dir-inode's nlink
4240 * attr to be set as 2. LU-7447
4242 set_nlink(dir, saved_nlink);
4247 * Create an local agent inode for remote entry
4249 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
4250 struct osd_device *osd,
4251 struct osd_object *pobj,
4252 const struct lu_fid *fid,
4256 struct osd_thread_info *info = osd_oti_get(env);
4257 struct inode *local;
4258 struct osd_thandle *oh;
4259 struct iattr iattr = {
4260 .ia_valid = ATTR_UID | ATTR_GID |
4261 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
4262 .ia_ctime.tv_nsec = UTIME_OMIT,
4263 .ia_mtime.tv_nsec = UTIME_OMIT,
4264 .ia_atime.tv_nsec = UTIME_OMIT,
4265 .ia_uid = GLOBAL_ROOT_UID,
4266 .ia_gid = GLOBAL_ROOT_GID,
4273 oh = container_of(th, struct osd_thandle, ot_super);
4274 LASSERT(oh->ot_handle->h_transaction != NULL);
4276 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
4278 if (IS_ERR(local)) {
4279 CERROR("%s: create local error %d\n", osd_name(osd),
4280 (int)PTR_ERR(local));
4285 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
4286 * correct gid on remote file, not agent here
4288 local->i_gid = current_fsgid();
4289 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4291 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4292 * That gives e2fsck something to look at and be happy, and allows
4293 * debugging if we need to determine where this symlink came from.
4295 if (S_ISLNK(type)) {
4296 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4297 ldiskfs_clear_inode_flag(local, LDISKFS_INODE_EXTENTS);
4298 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4299 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4301 i_size_write(local, rc);
4302 LDISKFS_I(local)->i_disksize = rc;
4304 unlock_new_inode(local);
4306 /* Agent inode should not have project ID */
4307 #ifdef HAVE_PROJECT_QUOTA
4308 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4309 i_projid_read(pobj->oo_inode) != 0) {
4310 rc = osd_transfer_project(local, 0, th);
4312 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4313 osd_ino2name(local), rc);
4314 RETURN(ERR_PTR(rc));
4318 /* Set special LMA flag for local agent inode */
4319 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4321 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4322 osd_name(osd), PFID(fid), rc);
4323 RETURN(ERR_PTR(rc));
4329 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode, fid,
4330 lu_object_fid(&pobj->oo_dt.do_lu),
4333 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4334 osd_name(osd), PFID(fid), rc);
4335 RETURN(ERR_PTR(rc));
4342 * when direntry is deleted, we have to take care of possible agent inode
4343 * referenced by that. unfortunately we can't do this at that point:
4344 * iget() within a running transaction leads to deadlock and we better do
4345 * not call that every delete declaration to save performance. so we put
4346 * a potention agent inode on a list and process that once the transaction
4347 * is over. Notice it's not any worse in terms of real orphans as regular
4348 * object destroy doesn't put inodes on the on-disk orphan list. this should
4349 * be addressed separately
4351 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4352 struct osd_thandle *oh,
4355 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4356 struct osd_obj_orphan *oor;
4363 oor->oor_env = (struct lu_env *)env;
4364 spin_lock(&osd->od_osfs_lock);
4365 list_add(&oor->oor_list, &osd->od_orphan_list);
4366 spin_unlock(&osd->od_osfs_lock);
4368 oh->ot_remove_agents = 1;
4374 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4375 struct osd_device *osd)
4377 struct osd_thread_info *info = osd_oti_get(env);
4378 struct osd_obj_orphan *oor, *tmp;
4379 struct osd_inode_id id;
4381 struct inode *inode;
4386 spin_lock(&osd->od_osfs_lock);
4387 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4388 if (oor->oor_env == env)
4389 list_move(&oor->oor_list, &list);
4391 spin_unlock(&osd->od_osfs_lock);
4393 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4397 list_del(&oor->oor_list);
4400 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4401 inode = osd_iget_fid(info, osd, &id, &fid);
4405 if (!osd_remote_fid(env, osd, &fid)) {
4410 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4412 mark_inode_dirty(inode);
4413 ldiskfs_journal_stop(jh);
4421 * OSD layer object create function for OST objects (b=11826).
4423 * The FID is inserted into inode xattr here.
4425 * \retval 0, on success
4426 * \retval -ve, on error
4428 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4429 struct lu_attr *attr, struct dt_allocation_hint *hint,
4430 struct dt_object_format *dof, struct thandle *th)
4432 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4433 struct osd_object *obj = osd_dt_obj(dt);
4434 struct osd_thread_info *info = osd_oti_get(env);
4435 int result, on_ost = 0;
4439 if (dt_object_exists(dt))
4442 LINVRNT(osd_invariant(obj));
4443 LASSERT(!dt_object_remote(dt));
4444 LASSERT(osd_is_write_locked(env, obj));
4445 LASSERT(th != NULL);
4447 if (unlikely(fid_is_acct(fid)))
4449 * Quota files can't be created from the kernel any more,
4450 * 'tune2fs -O quota' will take care of creating them
4454 result = __osd_create(info, obj, attr, hint, dof, th);
4456 if (fid_is_idif(fid) &&
4457 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4458 struct lu_fid *tfid = &info->oti_fid;
4459 struct ost_id *oi = &info->oti_ostid;
4461 fid_to_ostid(fid, oi);
4462 ostid_to_fid(tfid, oi, 0);
4464 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4465 LMAC_FID_ON_OST, 0);
4467 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4469 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4470 on_ost ? LMAC_FID_ON_OST : 0,
4473 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4474 obj->oo_dt.do_body_ops = &osd_body_ops;
4477 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4478 struct inode *inode = obj->oo_inode;
4480 result = __osd_oi_insert(env, obj, fid, th);
4481 if (result && inode) {
4482 spin_lock(&obj->oo_guard);
4484 spin_unlock(&obj->oo_guard);
4485 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4486 ldiskfs_set_inode_state(inode,
4487 LDISKFS_STATE_LUSTRE_DESTROY);
4489 obj->oo_inode = NULL;
4494 * a small optimization - dt_insert() isn't usually applied
4495 * to OST objects, so we don't need to cache OI mapping for
4498 if (result == 0 && on_ost == 0) {
4499 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4501 result = osd_idc_find_and_init(env, osd, obj);
4502 LASSERT(result == 0);
4505 LASSERT(ergo(result == 0,
4506 dt_object_exists(dt) && !dt_object_remote(dt)));
4507 LINVRNT(osd_invariant(obj));
4511 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4512 struct thandle *handle)
4514 struct osd_thandle *oh;
4517 /* it's possible that object doesn't exist yet */
4518 LASSERT(handle != NULL);
4520 oh = container_of(handle, struct osd_thandle, ot_super);
4521 LASSERT(oh->ot_handle == NULL);
4523 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4524 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4526 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4533 * Concurrency: @dt is write locked.
4535 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4538 struct osd_object *obj = osd_dt_obj(dt);
4539 struct inode *inode = obj->oo_inode;
4540 struct osd_thandle *oh;
4543 if (!dt_object_exists(dt) || obj->oo_destroyed)
4546 LINVRNT(osd_invariant(obj));
4547 LASSERT(!dt_object_remote(dt));
4548 LASSERT(osd_is_write_locked(env, obj));
4549 LASSERT(th != NULL);
4551 oh = container_of(th, struct osd_thandle, ot_super);
4552 LASSERT(oh->ot_handle != NULL);
4554 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4556 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4557 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4559 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4560 * (65000) subdirectories by storing "1" in i_nlink if the link count
4561 * would otherwise overflow. Directory tranversal tools understand
4562 * that (st_nlink == 1) indicates that the filesystem dose not track
4563 * hard links count on the directory, and will not abort subdirectory
4564 * scanning early once (st_nlink - 2) subdirs have been found.
4566 * This also has to properly handle the case of inodes with nlink == 0
4567 * in case they are being linked into the PENDING directory
4569 spin_lock(&obj->oo_guard);
4570 if (unlikely(inode->i_nlink == 0))
4571 /* inc_nlink from 0 may cause WARN_ON */
4572 set_nlink(inode, 1);
4574 osd_ldiskfs_inc_count(oh->ot_handle, inode);
4575 if (!S_ISDIR(inode->i_mode))
4576 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4578 spin_unlock(&obj->oo_guard);
4580 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4581 LINVRNT(osd_invariant(obj));
4583 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4588 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4589 struct thandle *handle)
4591 struct osd_thandle *oh;
4593 if (!dt_object_exists(dt))
4596 LASSERT(!dt_object_remote(dt));
4597 LASSERT(handle != NULL);
4599 oh = container_of(handle, struct osd_thandle, ot_super);
4600 LASSERT(oh->ot_handle == NULL);
4602 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4603 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4609 * Concurrency: @dt is write locked.
4611 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4614 struct osd_object *obj = osd_dt_obj(dt);
4615 struct inode *inode = obj->oo_inode;
4616 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4617 struct osd_thandle *oh;
4619 if (!dt_object_exists(dt))
4622 LINVRNT(osd_invariant(obj));
4623 LASSERT(!dt_object_remote(dt));
4624 LASSERT(osd_is_write_locked(env, obj));
4625 LASSERT(th != NULL);
4627 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4630 oh = container_of(th, struct osd_thandle, ot_super);
4631 LASSERT(oh->ot_handle != NULL);
4633 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4635 spin_lock(&obj->oo_guard);
4637 * That can be result of upgrade from old Lustre version and
4638 * applied only to local files. Just skip this ref_del call.
4639 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4641 if (inode->i_nlink == 0) {
4642 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4643 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4644 ", maybe an upgraded file? (LU-3915)\n",
4645 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4646 spin_unlock(&obj->oo_guard);
4650 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4651 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4653 osd_ldiskfs_dec_count(oh->ot_handle, inode);
4654 spin_unlock(&obj->oo_guard);
4656 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4657 LINVRNT(osd_invariant(obj));
4659 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4665 * Concurrency: @dt is read locked.
4667 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4668 struct lu_buf *buf, const char *name)
4670 struct osd_object *obj = osd_dt_obj(dt);
4671 struct inode *inode = obj->oo_inode;
4672 struct osd_thread_info *info = osd_oti_get(env);
4673 struct dentry *dentry = &info->oti_obj_dentry;
4674 bool cache_xattr = false;
4679 /* version get is not real XATTR but uses xattr API */
4680 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4681 dt_obj_version_t *ver = buf->lb_buf;
4684 * for version we are just using xattr API but change inode
4687 if (buf->lb_len == 0)
4688 return sizeof(dt_obj_version_t);
4690 if (buf->lb_len < sizeof(dt_obj_version_t))
4693 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4694 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4696 *ver = LDISKFS_I(inode)->i_fs_version;
4698 return sizeof(dt_obj_version_t);
4701 if (!dt_object_exists(dt))
4704 LASSERT(!dt_object_remote(dt));
4705 LASSERT(inode->i_op != NULL);
4706 #ifdef HAVE_IOP_XATTR
4707 LASSERT(inode->i_op->getxattr != NULL);
4710 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4711 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4715 rc = osd_oxc_get(obj, name, buf);
4720 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4721 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4722 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4723 struct filter_fid *ff;
4724 struct ost_layout *ol;
4726 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4730 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4733 if (buf->lb_len == 0 || !buf->lb_buf)
4736 if (buf->lb_len < rc)
4740 ol = &ff->ff_layout;
4741 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4742 PFID_STRIPE_IDX_BITS);
4743 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4744 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4745 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4746 if (lma->lma_compat & LMAC_COMP_INFO) {
4747 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4748 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4749 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4751 ol->ol_comp_start = 0;
4752 ol->ol_comp_end = 0;
4756 /* Get enc context xattr directly from ldiskfs instead of going
4757 * through the VFS, as there is no xattr handler for
4760 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0)
4761 rc = ldiskfs_xattr_get(inode,
4762 LDISKFS_XATTR_INDEX_ENCRYPTION,
4763 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
4764 buf->lb_buf, buf->lb_len);
4766 rc = __osd_xattr_get(inode, dentry, name,
4767 buf->lb_buf, buf->lb_len);
4771 if (rc == -ENOENT || rc == -ENODATA)
4772 osd_oxc_add(obj, name, NULL, 0);
4773 else if (rc > 0 && buf->lb_buf != NULL)
4774 osd_oxc_add(obj, name, buf->lb_buf, rc);
4780 static int osd_declare_xattr_set(const struct lu_env *env,
4781 struct dt_object *dt,
4782 const struct lu_buf *buf, const char *name,
4783 int fl, struct thandle *handle)
4785 struct osd_thandle *oh;
4787 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4789 LASSERT(handle != NULL);
4791 oh = container_of(handle, struct osd_thandle, ot_super);
4792 LASSERT(oh->ot_handle == NULL);
4794 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4796 * For non-upgrading case, the LMA is set first and
4797 * usually fit inode. But for upgrade case, the LMA
4798 * may be in another separated EA block.
4800 if (dt_object_exists(dt)) {
4801 if (fl == LU_XATTR_REPLACE)
4806 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4808 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4809 /* We may need to delete the old PFID EA. */
4810 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4811 if (fl == LU_XATTR_REPLACE)
4817 * If some name entry resides on remote MDT, then will create
4818 * agent entry under remote parent. On the other hand, if the
4819 * remote entry will be removed, then related agent entry may
4820 * need to be removed from the remote parent. So there may be
4821 * kinds of cases, let's declare enough credits. The credits
4822 * for create agent entry is enough for remove case.
4824 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4825 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4826 if (dt_object_exists(dt))
4827 credits += 1; /* For updating LMA */
4831 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4836 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4838 * learn xattr size from osd_xattr_get if
4839 * attribute has not been read yet
4841 buflen = __osd_xattr_get(
4842 osd_dt_obj(dt)->oo_inode,
4843 &osd_oti_get(env)->oti_obj_dentry,
4848 buflen = buf->lb_len;
4851 if (buflen > sb->s_blocksize) {
4852 credits += osd_calc_bkmap_credits(
4854 (buflen + sb->s_blocksize - 1) >>
4855 sb->s_blocksize_bits);
4859 * xattr set may involve inode quota change, reserve credits for
4860 * dquot_initialize()
4862 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4865 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4870 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4871 const struct lu_buf *buf, int fl,
4872 struct thandle *handle)
4874 struct osd_thread_info *info = osd_oti_get(env);
4875 struct dentry *dentry = &info->oti_obj_dentry;
4876 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4877 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4878 struct inode *inode = obj->oo_inode;
4879 struct filter_fid *ff = buf->lb_buf;
4880 struct ost_layout *ol = &ff->ff_layout;
4881 int flags = XATTR_REPLACE;
4886 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4889 rc = osd_get_lma(info, inode, dentry, loa);
4890 if (rc == -ENODATA) {
4891 /* Usually for upgarding from old device */
4892 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4893 LMAC_FID_ON_OST, 0);
4894 flags = XATTR_CREATE;
4899 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4900 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4903 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4904 /* Separate PFID EA from LMA */
4905 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4906 lustre_lma_swab(lma);
4907 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4908 sizeof(*lma), XATTR_REPLACE);
4910 obj->oo_pfid_in_lma = 0;
4911 rc = LU_XATTR_CREATE;
4917 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4921 * Old client does not send stripe information,
4922 * then store the PFID EA on disk separatedly.
4924 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4925 ol->ol_stripe_size == 0))
4928 /* Remove old PFID EA entry firstly. */
4929 dquot_initialize(inode);
4930 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4931 if (rc == -ENODATA) {
4932 /* XATTR_NAME_FID is already absent */
4939 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4940 if (likely(ol->ol_stripe_size != 0)) {
4941 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4942 PFID_STRIPE_IDX_BITS;
4943 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4944 lma->lma_compat |= LMAC_STRIPE_INFO;
4945 if (ol->ol_comp_id != 0) {
4946 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4947 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4948 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4949 lma->lma_compat |= LMAC_COMP_INFO;
4953 lustre_loa_swab(loa, false);
4955 /* Store the PFID EA inside LMA. */
4956 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4959 obj->oo_pfid_in_lma = 1;
4965 * In DNE environment, the object (in spite of regular file or directory)
4966 * and its name entry may reside on different MDTs. Under such case, we will
4967 * create an agent entry on the MDT where the object resides. The agent entry
4968 * references the object locally, that makes the object to be visible to the
4969 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4970 * such as 'tar' can handle the object properly.
4972 * We handle the agent entry during set linkEA that is the common interface
4973 * for both regular file and directroy, can handle kinds of cases, such as
4974 * create/link/unlink/rename, and so on.
4976 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4978 * XXX: There are two known issues:
4979 * 1. For one object, we will create at most one agent entry even if there
4980 * may be more than one cross-MDTs hard links on the object. So the local
4981 * e2fsck may claim that the object's nlink is larger than the name entries
4982 * that reference such inode. And in further, the e2fsck will fix the nlink
4983 * attribute to match the local references. Then it will cause the object's
4984 * nlink attribute to be inconsistent with the global references. it is bad
4985 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4986 * other hand, the global namespace LFSCK can repair the object's attribute
4987 * according to the linkEA.
4988 * 2. There may be too many hard links on the object as to its linkEA overflow,
4989 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4990 * case happened, then at this point, we do not know whether there are some
4991 * cross-MDTs reference. But there are local references, it guarantees that
4992 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4994 static int osd_xattr_handle_linkea(const struct lu_env *env,
4995 struct osd_device *osd,
4996 struct osd_object *obj,
4997 const struct lu_buf *buf,
4998 struct thandle *handle)
5000 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
5001 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
5002 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
5003 struct lu_name tmpname;
5004 struct osd_thandle *oh;
5006 bool remote = false;
5010 oh = container_of(handle, struct osd_thandle, ot_super);
5011 LASSERT(oh->ot_handle != NULL);
5013 rc = linkea_init_with_rec(&ldata);
5015 linkea_first_entry(&ldata);
5016 while (ldata.ld_lee != NULL && !remote) {
5017 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
5019 if (osd_remote_fid(env, osd, tfid) > 0)
5022 linkea_next_entry(&ldata);
5024 } else if (rc == -ENODATA) {
5030 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
5031 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
5033 CERROR("%s: failed to remove agent entry for "DFID
5034 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5035 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
5036 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5038 CERROR("%s: failed to create agent entry for "DFID
5039 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5046 * Concurrency: @dt is write locked.
5048 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
5049 const struct lu_buf *buf, const char *name, int fl,
5050 struct thandle *handle)
5052 struct osd_object *obj = osd_dt_obj(dt);
5053 struct osd_device *osd = osd_obj2dev(obj);
5054 struct inode *inode = obj->oo_inode;
5055 struct osd_thread_info *info = osd_oti_get(env);
5065 /* version set is not real XATTR */
5066 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
5067 dt_obj_version_t *version = buf->lb_buf;
5070 * for version we are just using xattr API but change inode
5073 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
5076 DFID" set version %#llx (old %#llx) for inode %lu\n",
5077 PFID(lu_object_fid(&dt->do_lu)), *version,
5078 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
5080 LDISKFS_I(inode)->i_fs_version = *version;
5082 * Version is set after all inode operations are finished,
5083 * so we should mark it dirty here
5085 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
5090 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
5091 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
5094 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5097 * For the OST device with 256 bytes inode size by default,
5098 * the PFID EA will be stored together with LMA EA to avoid
5099 * performance trouble. Otherwise the PFID EA can be stored
5100 * independently. LU-8998
5102 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
5103 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
5104 LASSERT(buf->lb_buf);
5106 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
5109 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
5110 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
5111 struct lustre_mdt_attrs *lma = &loa->loa_lma;
5113 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
5117 lma->lma_incompat |= LMAI_STRIPED;
5118 lustre_lma_swab(lma);
5119 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5120 sizeof(*lma), XATTR_REPLACE);
5123 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
5124 LASSERT(!osd->od_is_ost);
5126 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
5131 if (fl & LU_XATTR_REPLACE)
5132 fs_flags |= XATTR_REPLACE;
5134 if (fl & LU_XATTR_CREATE)
5135 fs_flags |= XATTR_CREATE;
5137 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0) {
5138 /* Set enc context xattr directly in ldiskfs instead of going
5139 * through the VFS, as there is no xattr handler for
5142 struct osd_thandle *oth = container_of(handle,
5146 if (!oth->ot_handle)
5147 /* this should be already part of a transaction */
5150 rc = ldiskfs_xattr_set_handle(oth->ot_handle, inode,
5151 LDISKFS_XATTR_INDEX_ENCRYPTION,
5152 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
5153 buf->lb_buf, len, fs_flags);
5155 rc = __osd_xattr_set(info, inode, name,
5156 buf->lb_buf, len, fs_flags);
5158 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5161 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5162 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5163 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
5169 * Concurrency: @dt is read locked.
5171 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
5172 const struct lu_buf *buf)
5174 struct osd_object *obj = osd_dt_obj(dt);
5175 struct osd_device *dev = osd_obj2dev(obj);
5176 struct inode *inode = obj->oo_inode;
5177 struct osd_thread_info *info = osd_oti_get(env);
5178 struct dentry *dentry = &info->oti_obj_dentry;
5181 if (!dt_object_exists(dt))
5184 LASSERT(!dt_object_remote(dt));
5185 LASSERT(inode->i_op != NULL);
5186 LASSERT(inode->i_op->listxattr != NULL);
5188 dentry->d_inode = inode;
5189 dentry->d_sb = inode->i_sb;
5190 rc = inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
5192 if (rc < 0 || buf->lb_buf == NULL)
5195 /* Hide virtual project ID xattr from list if disabled */
5196 if (!dev->od_enable_projid_xattr) {
5197 char *end = (char *)buf->lb_buf + rc;
5198 char *p = buf->lb_buf;
5201 char *next = p + strlen(p) + 1;
5203 if (strcmp(p, XATTR_NAME_PROJID) == 0) {
5205 memmove(p, next, end - next);
5217 static int osd_declare_xattr_del(const struct lu_env *env,
5218 struct dt_object *dt, const char *name,
5219 struct thandle *handle)
5221 struct osd_thandle *oh;
5222 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
5224 LASSERT(!dt_object_remote(dt));
5225 LASSERT(handle != NULL);
5227 oh = container_of(handle, struct osd_thandle, ot_super);
5228 LASSERT(oh->ot_handle == NULL);
5230 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
5231 osd_dto_credits_noquota[DTO_XATTR_SET]);
5233 * xattr del may involve inode quota change, reserve credits for
5234 * dquot_initialize()
5236 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
5242 * Concurrency: @dt is write locked.
5244 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
5245 const char *name, struct thandle *handle)
5247 struct osd_object *obj = osd_dt_obj(dt);
5248 struct inode *inode = obj->oo_inode;
5249 struct osd_thread_info *info = osd_oti_get(env);
5250 struct dentry *dentry = &info->oti_obj_dentry;
5253 if (!dt_object_exists(dt))
5256 LASSERT(!dt_object_remote(dt));
5257 LASSERT(inode->i_op != NULL);
5258 LASSERT(handle != NULL);
5259 #ifdef HAVE_IOP_XATTR
5260 LASSERT(inode->i_op->removexattr != NULL);
5263 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5265 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
5266 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
5268 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
5269 &info->oti_ost_attrs);
5271 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
5273 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
5274 lustre_lma_swab(lma);
5275 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5276 sizeof(*lma), XATTR_REPLACE);
5278 obj->oo_pfid_in_lma = 0;
5281 dquot_initialize(inode);
5282 dentry->d_inode = inode;
5283 dentry->d_sb = inode->i_sb;
5284 rc = ll_vfs_removexattr(dentry, inode, name);
5287 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5290 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5291 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5292 osd_oxc_del(obj, name);
5297 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
5298 __u64 start, __u64 end)
5300 struct osd_object *obj = osd_dt_obj(dt);
5301 struct osd_device *dev = osd_obj2dev(obj);
5302 struct inode *inode = obj->oo_inode;
5307 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
5310 RETURN(PTR_ERR(file));
5312 file->f_mode |= FMODE_64BITHASH;
5313 rc = vfs_fsync_range(file, start, end, 0);
5320 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
5325 static bool osd_check_stale(struct dt_object *dt)
5333 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
5334 const struct dt_index_features *feat)
5336 struct iam_descr *descr;
5338 if (osd_object_is_root(o))
5339 return feat == &dt_directory_features;
5341 LASSERT(o->oo_dir != NULL);
5343 descr = o->oo_dir->od_container.ic_descr;
5344 if (feat == &dt_directory_features) {
5345 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
5350 return feat->dif_keysize_min <= descr->id_key_size &&
5351 descr->id_key_size <= feat->dif_keysize_max &&
5352 feat->dif_recsize_min <= descr->id_rec_size &&
5353 descr->id_rec_size <= feat->dif_recsize_max &&
5354 !(feat->dif_flags & (DT_IND_VARKEY |
5355 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5356 ergo(feat->dif_flags & DT_IND_UPDATE,
5357 1 /* XXX check that object (and fs) is writable */);
5361 static int osd_iam_container_init(const struct lu_env *env,
5362 struct osd_object *obj,
5363 struct osd_directory *dir)
5365 struct iam_container *bag = &dir->od_container;
5368 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5372 result = iam_container_setup(bag);
5374 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5376 iam_container_fini(bag);
5383 * Concurrency: no external locking is necessary.
5385 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5386 const struct dt_index_features *feat)
5390 struct osd_object *obj = osd_dt_obj(dt);
5392 LINVRNT(osd_invariant(obj));
5394 if (osd_object_is_root(obj)) {
5395 dt->do_index_ops = &osd_index_ea_ops;
5397 } else if (feat == &dt_directory_features) {
5398 dt->do_index_ops = &osd_index_ea_ops;
5399 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5404 } else if (unlikely(feat == &dt_otable_features)) {
5405 dt->do_index_ops = &osd_otable_ops;
5407 } else if (unlikely(feat == &dt_acct_features)) {
5408 dt->do_index_ops = &osd_acct_index_ops;
5411 } else if (!osd_has_index(obj)) {
5412 struct osd_directory *dir;
5413 struct osd_device *osd = osd_obj2dev(obj);
5414 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5419 spin_lock(&obj->oo_guard);
5420 if (obj->oo_dir == NULL)
5424 * Concurrent thread allocated container data.
5427 spin_unlock(&obj->oo_guard);
5429 * Now, that we have container data, serialize its
5432 down_write(&obj->oo_ext_idx_sem);
5434 * recheck under lock.
5437 if (osd_has_index(obj)) {
5442 result = osd_iam_container_init(env, obj, obj->oo_dir);
5443 if (result || feat == &dt_lfsck_namespace_features ||
5444 feat == &dt_lfsck_layout_orphan_features ||
5445 feat == &dt_lfsck_layout_dangling_features)
5448 result = osd_index_register(osd, fid,
5449 feat->dif_keysize_max,
5450 feat->dif_recsize_max);
5452 CWARN("%s: failed to register index "
5454 osd_name(osd), PFID(fid), result);
5455 else if (result > 0)
5458 CDEBUG(D_LFSCK, "%s: index object "DFID
5459 " (%d/%d) registered\n",
5460 osd_name(osd), PFID(fid),
5461 (int)feat->dif_keysize_max,
5462 (int)feat->dif_recsize_max);
5465 up_write(&obj->oo_ext_idx_sem);
5473 if (result == 0 && skip_iam == 0) {
5474 if (!osd_iam_index_probe(env, obj, feat))
5477 LINVRNT(osd_invariant(obj));
5482 static int osd_otable_it_attr_get(const struct lu_env *env,
5483 struct dt_object *dt,
5484 struct lu_attr *attr)
5490 static const struct dt_object_operations osd_obj_ops = {
5491 .do_read_lock = osd_read_lock,
5492 .do_write_lock = osd_write_lock,
5493 .do_read_unlock = osd_read_unlock,
5494 .do_write_unlock = osd_write_unlock,
5495 .do_write_locked = osd_write_locked,
5496 .do_attr_get = osd_attr_get,
5497 .do_declare_attr_set = osd_declare_attr_set,
5498 .do_attr_set = osd_attr_set,
5499 .do_ah_init = osd_ah_init,
5500 .do_declare_create = osd_declare_create,
5501 .do_create = osd_create,
5502 .do_declare_destroy = osd_declare_destroy,
5503 .do_destroy = osd_destroy,
5504 .do_index_try = osd_index_try,
5505 .do_declare_ref_add = osd_declare_ref_add,
5506 .do_ref_add = osd_ref_add,
5507 .do_declare_ref_del = osd_declare_ref_del,
5508 .do_ref_del = osd_ref_del,
5509 .do_xattr_get = osd_xattr_get,
5510 .do_declare_xattr_set = osd_declare_xattr_set,
5511 .do_xattr_set = osd_xattr_set,
5512 .do_declare_xattr_del = osd_declare_xattr_del,
5513 .do_xattr_del = osd_xattr_del,
5514 .do_xattr_list = osd_xattr_list,
5515 .do_object_sync = osd_object_sync,
5516 .do_invalidate = osd_invalidate,
5517 .do_check_stale = osd_check_stale,
5520 static const struct dt_object_operations osd_obj_otable_it_ops = {
5521 .do_attr_get = osd_otable_it_attr_get,
5522 .do_index_try = osd_index_try,
5525 static int osd_index_declare_iam_delete(const struct lu_env *env,
5526 struct dt_object *dt,
5527 const struct dt_key *key,
5528 struct thandle *handle)
5530 struct osd_thandle *oh;
5532 oh = container_of(handle, struct osd_thandle, ot_super);
5533 LASSERT(oh->ot_handle == NULL);
5535 /* Recycle may cause additional three blocks to be changed. */
5536 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5537 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5543 * delete a (key, value) pair from index \a dt specified by \a key
5545 * \param dt osd index object
5546 * \param key key for index
5547 * \param rec record reference
5548 * \param handle transaction handler
5551 * \retval -ve failure
5553 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5554 const struct dt_key *key,
5555 struct thandle *handle)
5557 struct osd_thread_info *oti = osd_oti_get(env);
5558 struct osd_object *obj = osd_dt_obj(dt);
5559 struct osd_thandle *oh;
5560 struct iam_path_descr *ipd;
5561 struct iam_container *bag = &obj->oo_dir->od_container;
5566 if (!dt_object_exists(dt))
5569 LINVRNT(osd_invariant(obj));
5570 LASSERT(!dt_object_remote(dt));
5571 LASSERT(bag->ic_object == obj->oo_inode);
5572 LASSERT(handle != NULL);
5574 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5576 ipd = osd_idx_ipd_get(env, bag);
5577 if (unlikely(ipd == NULL))
5580 oh = container_of(handle, struct osd_thandle, ot_super);
5581 LASSERT(oh->ot_handle != NULL);
5582 LASSERT(oh->ot_handle->h_transaction != NULL);
5584 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5585 /* swab quota uid/gid provided by caller */
5586 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5587 key = (const struct dt_key *)&oti->oti_quota_id;
5590 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5591 osd_ipd_put(env, bag, ipd);
5592 LINVRNT(osd_invariant(obj));
5593 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5597 static int osd_index_declare_ea_delete(const struct lu_env *env,
5598 struct dt_object *dt,
5599 const struct dt_key *key,
5600 struct thandle *handle)
5602 struct osd_thandle *oh;
5603 struct inode *inode;
5608 LASSERT(!dt_object_remote(dt));
5609 LASSERT(handle != NULL);
5611 oh = container_of(handle, struct osd_thandle, ot_super);
5612 LASSERT(oh->ot_handle == NULL);
5614 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5615 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5617 inode = osd_dt_obj(dt)->oo_inode;
5621 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5622 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5627 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5630 struct osd_fid_pack *rec;
5633 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5634 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5635 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5636 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5642 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5643 const struct lu_fid *fid)
5645 struct seq_server_site *ss = osd_seq_site(osd);
5649 /* FID seqs not in FLDB, must be local seq */
5650 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5654 * If FLD is not being initialized yet, it only happens during the
5655 * initialization, likely during mgs initialization, and we assume
5656 * this is local FID.
5658 if (ss == NULL || ss->ss_server_fld == NULL)
5661 /* Only check the local FLDB here */
5662 if (osd_seq_exists(env, osd, fid_seq(fid)))
5668 static void osd_take_care_of_agent(const struct lu_env *env,
5669 struct osd_device *osd,
5670 struct osd_thandle *oh,
5671 struct ldiskfs_dir_entry_2 *de)
5673 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5674 struct osd_idmap_cache *idc;
5675 int rc, schedule = 0;
5677 LASSERT(de != NULL);
5679 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5680 if (likely(rc == 0)) {
5681 idc = osd_idc_find_or_init(env, osd, fid);
5682 if (IS_ERR(idc) || idc->oic_remote)
5684 } else if (rc == -ENODATA) {
5686 * can't get FID, postpone to the end of the
5687 * transaction when iget() is safe
5691 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5694 osd_schedule_agent_inode_removal(env, oh,
5695 le32_to_cpu(de->inode));
5699 * Utility function to get real name from object name
5701 * \param[in] obj pointer to the object to be handled
5702 * \param[in] name object name
5703 * \param[in] len object name len
5704 * \param[out]ln pointer to the struct lu_name to hold the real name
5706 * If file is not encrypted, real name is just the object name.
5707 * If file is encrypted, object name needs to be decoded. In
5708 * this case a new buffer is allocated, and ln->ln_name needs to be freed by
5711 * \retval 0, on success
5712 * \retval -ve, on error
5714 static int obj_name2lu_name(struct osd_object *obj, const char *name,
5715 int len, struct lu_name *ln)
5717 if (!(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
5719 ln->ln_namelen = len;
5721 char *buf = kmalloc(len, GFP_NOFS);
5726 len = critical_decode(name, len, buf);
5728 ln->ln_namelen = len;
5735 * Index delete function for interoperability mode (b11826).
5736 * It will remove the directory entry added by osd_index_ea_insert().
5737 * This entry is needed to maintain name->fid mapping.
5739 * \param key, key i.e. file entry to be deleted
5741 * \retval 0, on success
5742 * \retval -ve, on error
5744 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5745 const struct dt_key *key, struct thandle *handle)
5747 struct osd_object *obj = osd_dt_obj(dt);
5748 struct inode *dir = obj->oo_inode;
5749 struct dentry *dentry;
5750 struct osd_thandle *oh;
5751 struct ldiskfs_dir_entry_2 *de = NULL;
5752 struct buffer_head *bh;
5753 struct htree_lock *hlock = NULL;
5754 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5760 if (!dt_object_exists(dt))
5763 LINVRNT(osd_invariant(obj));
5764 LASSERT(!dt_object_remote(dt));
5765 LASSERT(handle != NULL);
5767 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
5771 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5773 oh = container_of(handle, struct osd_thandle, ot_super);
5774 LASSERT(oh->ot_handle != NULL);
5775 LASSERT(oh->ot_handle->h_transaction != NULL);
5777 dquot_initialize(dir);
5778 dentry = osd_child_dentry_get(env, obj, ln.ln_name, ln.ln_namelen);
5780 if (obj->oo_hl_head != NULL) {
5781 hlock = osd_oti_get(env)->oti_hlock;
5782 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5783 dir, LDISKFS_HLOCK_DEL);
5785 down_write(&obj->oo_ext_idx_sem);
5788 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5791 * If this is not the ".." entry, it might be a remote DNE
5792 * entry and we need to check if the FID is for a remote
5793 * MDT. If the FID is not in the directory entry (e.g.
5794 * upgraded 1.8 filesystem without dirdata enabled) then
5795 * we need to get the FID from the LMA. For a remote directory
5796 * there HAS to be an LMA, it cannot be an IGIF inode in this
5799 * Delete the entry before the agent inode in order to
5800 * simplify error handling. At worst an error after deleting
5801 * the entry first might leak the agent inode afterward. The
5802 * reverse would need filesystem abort in case of error deleting
5803 * the entry after the agent had been removed, or leave a
5804 * dangling entry pointing at a random inode.
5806 if (strcmp((char *)key, dotdot) != 0) {
5807 osd_take_care_of_agent(env, osd, oh, de);
5808 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5815 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
5816 atomic_dec_if_positive(&obj->oo_dirent_count);
5818 ldiskfs_htree_unlock(hlock);
5820 up_write(&obj->oo_ext_idx_sem);
5823 LASSERT(osd_invariant(obj));
5824 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5825 if (ln.ln_name != (char *)key)
5831 * Lookup index for \a key and copy record to \a rec.
5833 * \param dt osd index object
5834 * \param key key for index
5835 * \param rec record reference
5837 * \retval +ve success : exact mach
5838 * \retval 0 return record with key not greater than \a key
5839 * \retval -ve failure
5841 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5842 struct dt_rec *rec, const struct dt_key *key)
5844 struct osd_object *obj = osd_dt_obj(dt);
5845 struct iam_path_descr *ipd;
5846 struct iam_container *bag = &obj->oo_dir->od_container;
5847 struct osd_thread_info *oti = osd_oti_get(env);
5848 struct iam_iterator *it = &oti->oti_idx_it;
5849 struct iam_rec *iam_rec;
5854 if (!dt_object_exists(dt))
5857 LASSERT(osd_invariant(obj));
5858 LASSERT(!dt_object_remote(dt));
5859 LASSERT(bag->ic_object == obj->oo_inode);
5861 ipd = osd_idx_ipd_get(env, bag);
5865 /* got ipd now we can start iterator. */
5866 iam_it_init(it, bag, 0, ipd);
5868 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5869 /* swab quota uid/gid provided by caller */
5870 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5871 key = (const struct dt_key *)&oti->oti_quota_id;
5874 rc = iam_it_get(it, (struct iam_key *)key);
5876 if (S_ISDIR(obj->oo_inode->i_mode))
5877 iam_rec = (struct iam_rec *)oti->oti_ldp;
5879 iam_rec = (struct iam_rec *)rec;
5881 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5883 if (S_ISDIR(obj->oo_inode->i_mode))
5884 osd_fid_unpack((struct lu_fid *)rec,
5885 (struct osd_fid_pack *)iam_rec);
5886 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5887 osd_quota_unpack(obj, rec);
5892 osd_ipd_put(env, bag, ipd);
5894 LINVRNT(osd_invariant(obj));
5899 static int osd_index_declare_iam_insert(const struct lu_env *env,
5900 struct dt_object *dt,
5901 const struct dt_rec *rec,
5902 const struct dt_key *key,
5903 struct thandle *handle)
5905 struct osd_thandle *oh;
5907 LASSERT(handle != NULL);
5909 oh = container_of(handle, struct osd_thandle, ot_super);
5910 LASSERT(oh->ot_handle == NULL);
5912 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5913 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5919 * Inserts (key, value) pair in \a dt index object.
5921 * \param dt osd index object
5922 * \param key key for index
5923 * \param rec record reference
5924 * \param th transaction handler
5927 * \retval -ve failure
5929 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5930 const struct dt_rec *rec,
5931 const struct dt_key *key, struct thandle *th)
5933 struct osd_object *obj = osd_dt_obj(dt);
5934 struct iam_path_descr *ipd;
5935 struct osd_thandle *oh;
5936 struct iam_container *bag;
5937 struct osd_thread_info *oti = osd_oti_get(env);
5938 struct iam_rec *iam_rec;
5943 if (!dt_object_exists(dt))
5946 LINVRNT(osd_invariant(obj));
5947 LASSERT(!dt_object_remote(dt));
5949 bag = &obj->oo_dir->od_container;
5950 LASSERT(bag->ic_object == obj->oo_inode);
5951 LASSERT(th != NULL);
5953 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5955 ipd = osd_idx_ipd_get(env, bag);
5956 if (unlikely(ipd == NULL))
5959 oh = container_of(th, struct osd_thandle, ot_super);
5960 LASSERT(oh->ot_handle != NULL);
5961 LASSERT(oh->ot_handle->h_transaction != NULL);
5962 if (S_ISDIR(obj->oo_inode->i_mode)) {
5963 iam_rec = (struct iam_rec *)oti->oti_ldp;
5964 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5966 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5967 /* pack quota uid/gid */
5968 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5969 key = (const struct dt_key *)&oti->oti_quota_id;
5970 /* pack quota record */
5971 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5972 iam_rec = (struct iam_rec *)rec;
5974 iam_rec = (struct iam_rec *)rec;
5977 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5979 osd_ipd_put(env, bag, ipd);
5980 LINVRNT(osd_invariant(obj));
5981 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5986 * Calls ldiskfs_add_entry() to add directory entry
5987 * into the directory. This is required for
5988 * interoperability mode (b11826)
5990 * \retval 0, on success
5991 * \retval -ve, on error
5993 static int __osd_ea_add_rec(struct osd_thread_info *info,
5994 struct osd_object *pobj, struct inode *cinode,
5995 const char *name, const struct lu_fid *fid,
5996 struct htree_lock *hlock, struct thandle *th)
5998 struct ldiskfs_dentry_param *ldp;
5999 struct dentry *child;
6000 struct osd_thandle *oth;
6004 oth = container_of(th, struct osd_thandle, ot_super);
6005 LASSERT(oth->ot_handle != NULL);
6006 LASSERT(oth->ot_handle->h_transaction != NULL);
6007 LASSERT(pobj->oo_inode);
6009 rc = obj_name2lu_name(pobj, name, strlen(name), &ln);
6013 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
6014 if (unlikely(osd_object_is_root(pobj)))
6017 osd_get_ldiskfs_dirent_param(ldp, fid);
6018 child = osd_child_dentry_get(info->oti_env, pobj,
6019 ln.ln_name, ln.ln_namelen);
6020 child->d_fsdata = (void *)ldp;
6021 dquot_initialize(pobj->oo_inode);
6022 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
6023 child, cinode, hlock);
6024 if (rc == 0 && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
6025 struct ldiskfs_dir_entry_2 *de;
6026 struct buffer_head *bh;
6029 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
6032 rc1 = osd_ldiskfs_journal_get_write_access(
6033 oth->ot_handle, pobj->oo_inode->i_sb, bh,
6036 if (S_ISDIR(cinode->i_mode))
6037 de->file_type = LDISKFS_DIRENT_LUFID |
6038 LDISKFS_FT_REG_FILE;
6040 de->file_type = LDISKFS_DIRENT_LUFID |
6042 ldiskfs_handle_dirty_metadata(oth->ot_handle,
6049 if (ln.ln_name != name)
6055 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
6056 * into the directory.Also sets flags into osd object to
6057 * indicate dot and dotdot are created. This is required for
6058 * interoperability mode (b11826)
6060 * \param dir directory for dot and dotdot fixup.
6061 * \param obj child object for linking
6063 * \retval 0, on success
6064 * \retval -ve, on error
6066 static int osd_add_dot_dotdot(struct osd_thread_info *info,
6067 struct osd_object *dir,
6068 struct inode *parent_dir, const char *name,
6069 const struct lu_fid *dot_fid,
6070 const struct lu_fid *dot_dot_fid,
6073 struct inode *inode = dir->oo_inode;
6074 struct osd_thandle *oth;
6077 oth = container_of(th, struct osd_thandle, ot_super);
6078 LASSERT(oth->ot_handle->h_transaction != NULL);
6079 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
6081 if (strcmp(name, dot) == 0) {
6082 if (dir->oo_compat_dot_created) {
6085 LASSERT(inode->i_ino == parent_dir->i_ino);
6086 dir->oo_compat_dot_created = 1;
6089 } else if (strcmp(name, dotdot) == 0) {
6090 if (!dir->oo_compat_dot_created)
6092 /* in case of rename, dotdot is already created */
6093 if (dir->oo_compat_dotdot_created) {
6094 return __osd_ea_add_rec(info, dir, parent_dir, name,
6095 dot_dot_fid, NULL, th);
6098 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
6099 struct lu_fid tfid = *dot_dot_fid;
6102 result = osd_add_dot_dotdot_internal(info,
6103 dir->oo_inode, parent_dir, dot_fid,
6106 result = osd_add_dot_dotdot_internal(info,
6107 dir->oo_inode, parent_dir, dot_fid,
6112 dir->oo_compat_dotdot_created = 1;
6120 * It will call the appropriate osd_add* function and return the
6121 * value, return by respective functions.
6123 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
6124 struct inode *cinode, const char *name,
6125 const struct lu_fid *fid, struct thandle *th)
6127 struct osd_thread_info *info = osd_oti_get(env);
6128 struct htree_lock *hlock;
6131 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
6133 if (name[0] == '.' && (name[1] == '\0' ||
6134 (name[1] == '.' && name[2] == '\0'))) {
6135 if (hlock != NULL) {
6136 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6139 down_write(&pobj->oo_ext_idx_sem);
6142 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
6143 lu_object_fid(&pobj->oo_dt.do_lu),
6146 if (hlock != NULL) {
6147 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6148 pobj->oo_inode, LDISKFS_HLOCK_ADD);
6150 down_write(&pobj->oo_ext_idx_sem);
6153 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
6154 struct lu_fid *tfid = &info->oti_fid;
6158 rc = __osd_ea_add_rec(info, pobj, cinode, name,
6161 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
6165 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))){
6166 int dirent_count = atomic_read(&pobj->oo_dirent_count);
6168 /* avoid extremely unlikely 2B-entry directory overflow case */
6169 if (dirent_count != LU_DIRENT_COUNT_UNSET &&
6170 likely(dirent_count < INT_MAX - NR_CPUS))
6171 atomic_inc(&pobj->oo_dirent_count);
6175 ldiskfs_htree_unlock(hlock);
6177 up_write(&pobj->oo_ext_idx_sem);
6183 osd_ldiskfs_consistency_check(struct osd_thread_info *oti,
6184 struct osd_device *dev,
6185 const struct lu_fid *fid,
6186 struct osd_inode_id *id)
6188 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
6189 struct inode *inode = NULL;
6195 if (!scrub_needs_check(scrub, fid, id->oii_ino))
6198 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
6199 if (rc == -ENOENT) {
6200 __u32 gen = id->oii_gen;
6206 inode = osd_iget(oti, dev, id);
6207 /* The inode has been removed (by race maybe). */
6208 if (IS_ERR(inode)) {
6209 rc = PTR_ERR(inode);
6211 RETURN(rc == -ESTALE ? -ENOENT : rc);
6214 /* The OI mapping is lost. */
6215 if (gen != OSD_OII_NOGEN)
6219 * The inode may has been reused by others, we do not know,
6220 * leave it to be handled by subsequent osd_fid_lookup().
6223 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
6230 if (scrub->os_running) {
6231 if (inode == NULL) {
6232 inode = osd_iget(oti, dev, id);
6233 /* The inode has been removed (by race maybe). */
6234 if (IS_ERR(inode)) {
6235 rc = PTR_ERR(inode);
6237 RETURN(rc == -ESTALE ? -ENOENT : rc);
6241 rc = osd_scrub_oi_insert(dev, fid, id, insert);
6243 * There is race condition between osd_oi_lookup and OI scrub.
6244 * The OI scrub finished just after osd_oi_lookup() failure.
6245 * Under such case, it is unnecessary to trigger OI scrub again,
6246 * but try to call osd_oi_lookup() again.
6248 if (unlikely(rc == -EAGAIN))
6251 if (!S_ISDIR(inode->i_mode))
6254 rc = osd_check_lmv(oti, dev, inode);
6259 if (dev->od_scrub.os_scrub.os_auto_scrub_interval != AS_NEVER &&
6261 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
6262 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
6263 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
6264 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
6265 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
6266 if (rc == 0 || rc == -EALREADY)
6278 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
6279 struct osd_device *dev,
6280 struct lu_fid *fid, __u32 ino)
6282 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
6283 struct osd_idmap_cache *oic = &oti->oti_cache;
6284 struct inode *inode;
6287 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
6288 inode = osd_iget(oti, dev, &oic->oic_lid);
6289 if (IS_ERR(inode)) {
6290 fid_zero(&oic->oic_fid);
6291 return PTR_ERR(inode);
6294 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
6297 fid_zero(&oic->oic_fid);
6299 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
6303 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
6304 struct osd_inode_id *id, const struct lu_fid *fid)
6306 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
6307 id->oii_ino, id->oii_gen, info);
6308 info->oti_cache.oic_lid = *id;
6309 info->oti_cache.oic_fid = *fid;
6310 info->oti_cache.oic_dev = osd;
6314 * Get parent FID from the linkEA.
6316 * For a directory which parent resides on remote MDT, to satisfy the
6317 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
6318 * the other hand, to make the lookup(..) on the directory can return
6319 * the real parent FID, we append the real parent FID after its ".."
6320 * name entry in the /REMOTE_PARENT_DIR.
6322 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
6323 * backup. So after the restore, we cannot get the right parent FID from
6324 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
6325 * we have stored the real parent FID in the directory object's linkEA,
6326 * we can parse the linkEA for the real parent FID.
6328 * \param[in] env pointer to the thread context
6329 * \param[in] obj pointer to the object to be handled
6330 * \param[out]fid pointer to the buffer to hold the parent FID
6332 * \retval 0 for getting the real parent FID successfully
6333 * \retval negative error number on failure
6335 static int osd_get_pfid_from_linkea(const struct lu_env *env,
6336 struct osd_object *obj,
6339 struct osd_thread_info *oti = osd_oti_get(env);
6340 struct lu_buf *buf = &oti->oti_big_buf;
6341 struct dentry *dentry = &oti->oti_obj_dentry;
6342 struct inode *inode = obj->oo_inode;
6343 struct linkea_data ldata = { NULL };
6349 if (!S_ISDIR(inode->i_mode))
6353 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6354 buf->lb_buf, buf->lb_len);
6355 if (rc == -ERANGE) {
6356 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6359 lu_buf_realloc(buf, rc);
6360 if (buf->lb_buf == NULL)
6367 if (unlikely(rc == 0))
6373 if (unlikely(buf->lb_buf == NULL)) {
6374 lu_buf_realloc(buf, rc);
6375 if (buf->lb_buf == NULL)
6382 rc = linkea_init_with_rec(&ldata);
6384 linkea_first_entry(&ldata);
6385 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
6391 static int osd_verify_ent_by_linkea(const struct lu_env *env,
6392 struct inode *inode,
6393 const struct lu_fid *pfid,
6394 const char *name, const int namelen)
6396 struct osd_thread_info *oti = osd_oti_get(env);
6397 struct lu_buf *buf = &oti->oti_big_buf;
6398 struct dentry *dentry = &oti->oti_obj_dentry;
6399 struct linkea_data ldata = { NULL };
6400 struct lu_name cname = { .ln_name = name,
6401 .ln_namelen = namelen };
6407 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6408 buf->lb_buf, buf->lb_len);
6410 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6415 if (unlikely(rc == 0))
6418 if (buf->lb_len < rc) {
6419 lu_buf_realloc(buf, rc);
6420 if (buf->lb_buf == NULL)
6427 rc = linkea_init_with_rec(&ldata);
6429 rc = linkea_links_find(&ldata, &cname, pfid);
6435 * Calls ->lookup() to find dentry. From dentry get inode and
6436 * read inode's ea to get fid. This is required for interoperability
6439 * \retval 0, on success
6440 * \retval -ve, on error
6442 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6443 struct dt_rec *rec, const struct lu_name *ln)
6445 struct inode *dir = obj->oo_inode;
6446 struct dentry *dentry;
6447 struct ldiskfs_dir_entry_2 *de;
6448 struct buffer_head *bh;
6449 struct lu_fid *fid = (struct lu_fid *)rec;
6450 struct htree_lock *hlock = NULL;
6456 LASSERT(dir->i_op != NULL);
6457 LASSERT(dir->i_op->lookup != NULL);
6459 dentry = osd_child_dentry_get(env, obj, ln->ln_name, ln->ln_namelen);
6461 if (obj->oo_hl_head != NULL) {
6462 hlock = osd_oti_get(env)->oti_hlock;
6463 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6464 dir, LDISKFS_HLOCK_LOOKUP);
6466 down_read(&obj->oo_ext_idx_sem);
6469 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6471 struct osd_thread_info *oti = osd_oti_get(env);
6472 struct osd_inode_id *id = &oti->oti_id;
6473 struct osd_device *dev = osd_obj2dev(obj);
6475 ino = le32_to_cpu(de->inode);
6476 if (CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6478 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6482 rc = osd_get_fid_from_dentry(de, rec);
6484 /* done with de, release bh */
6487 if (unlikely(is_remote_parent_ino(dev, ino))) {
6489 * If the parent is on remote MDT, and there
6490 * is no FID-in-dirent, then we have to get
6491 * the parent FID from the linkEA.
6493 if (likely(ln->ln_namelen == 2 &&
6494 ln->ln_name[0] == '.' && ln->ln_name[1] == '.'))
6495 rc = osd_get_pfid_from_linkea(env, obj,
6498 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6501 osd_id_gen(id, ino, OSD_OII_NOGEN);
6504 if (rc != 0 || osd_remote_fid(env, dev, fid))
6507 rc = osd_ldiskfs_consistency_check(oti, dev, fid, id);
6508 if (rc != -ENOENT) {
6509 /* Other error should not affect lookup result. */
6512 /* Normal file mapping should be added into OI cache
6513 * after FID in LMA check, but for local files like
6514 * hsm_actions, their FIDs are not stored in OI files,
6515 * see osd_initial_OI_scrub(), and here is the only
6516 * place to load mapping into OI cache.
6518 if (!fid_is_namespace_visible(fid))
6519 osd_add_oi_cache(osd_oti_get(env),
6520 osd_obj2dev(obj), id, fid);
6522 CDEBUG(D_INODE, DFID"/"DNAME" => "DFID"\n",
6523 PFID(lu_object_fid(&obj->oo_dt.do_lu)), PNAME(ln),
6533 ldiskfs_htree_unlock(hlock);
6535 up_read(&obj->oo_ext_idx_sem);
6539 static int osd_index_declare_ea_insert(const struct lu_env *env,
6540 struct dt_object *dt,
6541 const struct dt_rec *rec,
6542 const struct dt_key *key,
6543 struct thandle *handle)
6545 struct osd_thandle *oh;
6546 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6547 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6548 const struct lu_fid *fid = rec1->rec_fid;
6549 int credits, rc = 0;
6550 struct osd_idmap_cache *idc;
6554 LASSERT(!dt_object_remote(dt));
6555 LASSERT(handle != NULL);
6556 LASSERT(fid != NULL);
6557 LASSERT(rec1->rec_type != 0);
6559 oh = container_of(handle, struct osd_thandle, ot_super);
6560 LASSERT(oh->ot_handle == NULL);
6562 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6565 * we can't call iget() while a transactions is running
6566 * (this can lead to a deadlock), but we need to know
6567 * inum and object type. so we find this information at
6568 * declaration and cache in per-thread info
6570 idc = osd_idc_find_or_init(env, osd, fid);
6572 RETURN(PTR_ERR(idc));
6573 if (idc->oic_remote) {
6575 * a reference to remote inode is represented by an
6576 * agent inode which we have to create
6578 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6579 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6582 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6584 if (osd_dt_obj(dt)->oo_inode != NULL) {
6585 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6588 * We ignore block quota on meta pool (MDTs), so needn't
6589 * calculate how many blocks will be consumed by this index
6592 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6594 i_projid_read(inode), 0,
6595 oh, osd_dt_obj(dt), NULL,
6600 #ifdef HAVE_PROJECT_QUOTA
6602 * Reserve credits for local agent inode to transfer
6603 * to 0, quota enforcement is ignored in this case.
6605 if (idc->oic_remote &&
6606 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6607 i_projid_read(inode) != 0)
6608 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6609 0, i_projid_read(inode),
6610 0, false, PRJQUOTA);
6618 * Index add function for interoperability mode (b11826).
6619 * It will add the directory entry.This entry is needed to
6620 * maintain name->fid mapping.
6622 * \param key it is key i.e. file entry to be inserted
6623 * \param rec it is value of given key i.e. fid
6625 * \retval 0, on success
6626 * \retval -ve, on error
6628 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6629 const struct dt_rec *rec,
6630 const struct dt_key *key, struct thandle *th)
6632 struct osd_object *obj = osd_dt_obj(dt);
6633 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6634 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6635 const struct lu_fid *fid = rec1->rec_fid;
6636 const char *name = (const char *)key;
6637 struct osd_thread_info *oti = osd_oti_get(env);
6638 struct inode *child_inode = NULL;
6639 struct osd_idmap_cache *idc;
6644 if (!dt_object_exists(dt))
6647 LASSERT(osd_invariant(obj));
6648 LASSERT(!dt_object_remote(dt));
6649 LASSERT(th != NULL);
6651 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6653 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6655 idc = osd_idc_find(env, osd, fid);
6656 if (unlikely(idc == NULL)) {
6657 idc = osd_idc_find_or_init(env, osd, fid);
6660 * this dt_insert() wasn't declared properly, so
6661 * FID is missing in OI cache. we better do not
6662 * lookup FID in FLDB/OI and don't risk to deadlock,
6663 * but in some special cases (lfsck testing, etc)
6664 * it's much simpler than fixing a caller.
6666 * normally this error should be placed after the first
6667 * find, but migrate may attach source stripes to
6668 * target, which doesn't create stripes.
6670 CERROR("%s: "DFID" wasn't declared for insert\n",
6671 osd_name(osd), PFID(fid));
6673 RETURN(PTR_ERR(idc));
6677 if (idc->oic_remote) {
6678 /* Insert remote entry */
6679 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6681 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6683 child_inode = osd_create_local_agent_inode(env, osd,
6684 obj, fid, rec1->rec_type & S_IFMT, th);
6685 if (IS_ERR(child_inode))
6686 RETURN(PTR_ERR(child_inode));
6689 /* Insert local entry */
6690 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6691 /* for a reason OI cache wasn't filled properly */
6692 CERROR("%s: OIC for "DFID" isn't filled\n",
6693 osd_name(osd), PFID(fid));
6696 child_inode = oti->oti_inode;
6697 if (unlikely(child_inode == NULL)) {
6698 struct ldiskfs_inode_info *lii;
6703 child_inode = oti->oti_inode = &lii->vfs_inode;
6705 child_inode->i_sb = osd_sb(osd);
6706 child_inode->i_ino = idc->oic_lid.oii_ino;
6707 child_inode->i_mode = rec1->rec_type & S_IFMT;
6710 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6712 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6713 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6715 if (child_inode && child_inode != oti->oti_inode)
6717 LASSERT(osd_invariant(obj));
6718 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6724 * Initialize osd Iterator for given osd index object.
6726 * \param dt osd index object
6729 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6730 struct dt_object *dt,
6733 struct osd_it_iam *it;
6734 struct osd_object *obj = osd_dt_obj(dt);
6735 struct lu_object *lo = &dt->do_lu;
6736 struct iam_path_descr *ipd;
6737 struct iam_container *bag = &obj->oo_dir->od_container;
6739 if (!dt_object_exists(dt))
6740 return ERR_PTR(-ENOENT);
6744 return ERR_PTR(-ENOMEM);
6746 ipd = osd_it_ipd_get(env, bag);
6747 if (likely(ipd != NULL)) {
6751 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6752 return (struct dt_it *)it;
6755 return ERR_PTR(-ENOMEM);
6760 * free given Iterator.
6762 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6764 struct osd_it_iam *it = (struct osd_it_iam *)di;
6765 struct osd_object *obj = it->oi_obj;
6767 iam_it_fini(&it->oi_it);
6768 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6769 osd_object_put(env, obj);
6774 * Move Iterator to record specified by \a key
6776 * \param di osd iterator
6777 * \param key key for index
6779 * \retval +ve di points to record with least key not larger than key
6780 * \retval 0 di points to exact matched key
6781 * \retval -ve failure
6784 static int osd_it_iam_get(const struct lu_env *env,
6785 struct dt_it *di, const struct dt_key *key)
6787 struct osd_thread_info *oti = osd_oti_get(env);
6788 struct osd_it_iam *it = (struct osd_it_iam *)di;
6790 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6791 /* swab quota uid/gid */
6792 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6793 key = (struct dt_key *)&oti->oti_quota_id;
6796 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6802 * \param di osd iterator
6804 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6806 struct osd_it_iam *it = (struct osd_it_iam *)di;
6808 iam_it_put(&it->oi_it);
6812 * Move iterator by one record
6814 * \param di osd iterator
6816 * \retval +1 end of container reached
6818 * \retval -ve failure
6821 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6823 struct osd_it_iam *it = (struct osd_it_iam *)di;
6825 return iam_it_next(&it->oi_it);
6829 * Return pointer to the key under iterator.
6832 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6833 const struct dt_it *di)
6835 struct osd_thread_info *oti = osd_oti_get(env);
6836 struct osd_it_iam *it = (struct osd_it_iam *)di;
6837 struct osd_object *obj = it->oi_obj;
6840 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6842 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6843 /* swab quota uid/gid */
6844 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6845 key = (struct dt_key *)&oti->oti_quota_id;
6852 * Return size of key under iterator (in bytes)
6855 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6857 struct osd_it_iam *it = (struct osd_it_iam *)di;
6859 return iam_it_key_size(&it->oi_it);
6863 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6865 /* check if file type is required */
6866 if (ent->lde_attrs & LUDA_TYPE) {
6867 struct luda_type *lt;
6868 int align = sizeof(*lt) - 1;
6870 len = (len + align) & ~align;
6871 lt = (struct luda_type *)(ent->lde_name + len);
6872 lt->lt_type = cpu_to_le16(DTTOIF(type));
6875 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6879 * build lu direct from backend fs dirent.
6883 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6884 char *name, __u16 namelen, __u16 type, __u32 attr)
6886 ent->lde_attrs = attr | LUDA_FID;
6887 fid_cpu_to_le(&ent->lde_fid, fid);
6889 ent->lde_hash = cpu_to_le64(offset);
6890 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6892 strncpy(ent->lde_name, name, namelen);
6893 ent->lde_name[namelen] = '\0';
6894 ent->lde_namelen = cpu_to_le16(namelen);
6896 /* append lustre attributes */
6897 osd_it_append_attrs(ent, namelen, type);
6901 * Return pointer to the record under iterator.
6903 static int osd_it_iam_rec(const struct lu_env *env,
6904 const struct dt_it *di,
6905 struct dt_rec *dtrec, __u32 attr)
6907 struct osd_it_iam *it = (struct osd_it_iam *)di;
6908 struct osd_thread_info *info = osd_oti_get(env);
6912 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6913 const struct osd_fid_pack *rec;
6914 struct lu_fid *fid = &info->oti_fid;
6915 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6921 name = (char *)iam_it_key_get(&it->oi_it);
6923 RETURN(PTR_ERR(name));
6925 namelen = iam_it_key_size(&it->oi_it);
6927 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6929 RETURN(PTR_ERR(rec));
6931 rc = osd_fid_unpack(fid, rec);
6935 hash = iam_it_store(&it->oi_it);
6937 /* IAM does not store object type in IAM index (dir) */
6938 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6940 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6941 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6942 (struct iam_rec *)dtrec);
6943 osd_quota_unpack(it->oi_obj, dtrec);
6945 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6946 (struct iam_rec *)dtrec);
6953 * Returns cookie for current Iterator position.
6955 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6957 struct osd_it_iam *it = (struct osd_it_iam *)di;
6959 return iam_it_store(&it->oi_it);
6963 * Restore iterator from cookie.
6965 * \param di osd iterator
6966 * \param hash Iterator location cookie
6968 * \retval +ve di points to record with least key not larger than key.
6969 * \retval 0 di points to exact matched key
6970 * \retval -ve failure
6973 static int osd_it_iam_load(const struct lu_env *env,
6974 const struct dt_it *di, __u64 hash)
6976 struct osd_it_iam *it = (struct osd_it_iam *)di;
6978 return iam_it_load(&it->oi_it, hash);
6981 static const struct dt_index_operations osd_index_iam_ops = {
6982 .dio_lookup = osd_index_iam_lookup,
6983 .dio_declare_insert = osd_index_declare_iam_insert,
6984 .dio_insert = osd_index_iam_insert,
6985 .dio_declare_delete = osd_index_declare_iam_delete,
6986 .dio_delete = osd_index_iam_delete,
6988 .init = osd_it_iam_init,
6989 .fini = osd_it_iam_fini,
6990 .get = osd_it_iam_get,
6991 .put = osd_it_iam_put,
6992 .next = osd_it_iam_next,
6993 .key = osd_it_iam_key,
6994 .key_size = osd_it_iam_key_size,
6995 .rec = osd_it_iam_rec,
6996 .store = osd_it_iam_store,
6997 .load = osd_it_iam_load
7001 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
7002 struct osd_device *dev,
7003 struct inode *inode, u32 attr)
7005 struct osd_thread_info *info = osd_oti_get(env);
7006 struct osd_it_ea *oie;
7010 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
7013 RETURN(ERR_CAST(file));
7015 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
7016 if (attr & LUDA_64BITHASH)
7017 file->f_mode |= FMODE_64BITHASH;
7019 file->f_mode |= FMODE_32BITHASH;
7022 OBD_SLAB_ALLOC_PTR(oie, osd_itea_cachep);
7026 oie->oie_rd_dirent = 0;
7027 oie->oie_it_dirent = 0;
7028 oie->oie_dirent = NULL;
7029 if (unlikely(!info->oti_it_ea_buf_used)) {
7030 oie->oie_buf = info->oti_it_ea_buf;
7031 info->oti_it_ea_buf_used = 1;
7033 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7037 oie->oie_obj = NULL;
7038 oie->oie_file = file;
7043 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7047 return ERR_PTR(-ENOMEM);
7051 * Creates or initializes iterator context.
7053 * \retval struct osd_it_ea, iterator structure on success
7056 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
7057 struct dt_object *dt,
7060 struct osd_object *obj = osd_dt_obj(dt);
7061 struct osd_device *dev = osd_obj2dev(obj);
7062 struct lu_object *lo = &dt->do_lu;
7063 struct osd_it_ea *oie;
7067 if (!dt_object_exists(dt) || obj->oo_destroyed)
7068 RETURN(ERR_PTR(-ENOENT));
7070 oie = osd_it_dir_init(env, dev, obj->oo_inode, attr);
7072 RETURN(ERR_CAST(oie));
7076 RETURN((struct dt_it *)oie);
7079 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
7080 struct inode *inode)
7082 struct osd_thread_info *info = osd_oti_get(env);
7085 fput(oie->oie_file);
7086 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
7087 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7089 info->oti_it_ea_buf_used = 0;
7090 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7095 * Destroy or finishes iterator context.
7097 * \param di iterator structure to be destroyed
7099 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
7101 struct osd_it_ea *oie = (struct osd_it_ea *)di;
7102 struct osd_object *obj = oie->oie_obj;
7103 struct inode *inode = obj->oo_inode;
7106 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
7107 osd_object_put(env, obj);
7112 * It position the iterator at given key, so that next lookup continues from
7113 * that key Or it is similar to dio_it->load() but based on a key,
7114 * rather than file position.
7116 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
7119 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
7121 static int osd_it_ea_get(const struct lu_env *env,
7122 struct dt_it *di, const struct dt_key *key)
7124 struct osd_it_ea *it = (struct osd_it_ea *)di;
7127 LASSERT(((const char *)key)[0] == '\0');
7128 it->oie_file->f_pos = 0;
7129 it->oie_rd_dirent = 0;
7130 it->oie_it_dirent = 0;
7131 it->oie_dirent = NULL;
7139 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
7143 struct osd_filldir_cbs {
7144 struct dir_context ctx;
7145 struct osd_it_ea *it;
7148 * It is called internally by ->iterate*(). It fills the
7149 * iterator's in-memory data structure with required
7150 * information i.e. name, namelen, rec_size etc.
7152 * \param buf in which information to be filled in.
7153 * \param name name of the file in given dir
7155 * \retval 0 on success
7156 * \retval 1 on buffer full
7158 #ifdef HAVE_FILLDIR_USE_CTX
7159 static FILLDIR_TYPE do_osd_ldiskfs_filldir(struct dir_context *ctx,
7161 static int osd_ldiskfs_filldir(void *ctx,
7163 const char *name, int namelen,
7164 loff_t offset, __u64 ino, unsigned int d_type)
7166 struct osd_it_ea *it = ((struct osd_filldir_cbs *)ctx)->it;
7167 struct osd_object *obj = it->oie_obj;
7168 struct osd_it_ea_dirent *ent = it->oie_dirent;
7169 struct lu_fid *fid = &ent->oied_fid;
7170 char *buf = it->oie_buf;
7171 struct osd_fid_pack *rec;
7174 /* this should never happen */
7175 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
7176 CERROR("ldiskfs return invalid namelen %d\n", namelen);
7180 /* Check for enough space. Note oied_name is not NUL terminated. */
7181 if (&ent->oied_name[namelen] > buf + OSD_IT_EA_BUFSIZE)
7184 /* "." is just the object itself. */
7185 if (namelen == 1 && name[0] == '.') {
7187 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7188 } else if (d_type & LDISKFS_DIRENT_LUFID) {
7189 rec = (struct osd_fid_pack *)(name + namelen + 1);
7190 if (osd_fid_unpack(fid, rec) != 0)
7195 d_type &= ~LDISKFS_DIRENT_LUFID;
7197 /* NOT export local root. */
7199 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
7200 ino = obj->oo_inode->i_ino;
7201 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7204 if (obj == NULL || !(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
7205 ent->oied_namelen = namelen;
7206 memcpy(ent->oied_name, name, namelen);
7208 int encoded_namelen = critical_chars(name, namelen);
7210 /* Check again for enough space. */
7211 if (&ent->oied_name[encoded_namelen] > buf + OSD_IT_EA_BUFSIZE)
7214 ent->oied_namelen = encoded_namelen;
7216 if (encoded_namelen == namelen)
7217 memcpy(ent->oied_name, name, namelen);
7219 critical_encode(name, namelen, ent->oied_name);
7222 ent->oied_ino = ino;
7223 ent->oied_off = offset;
7224 ent->oied_type = d_type;
7226 it->oie_rd_dirent++;
7227 it->oie_dirent = (void *)ent + round_up(sizeof(*ent) + ent->oied_namelen, 8);
7231 WRAP_FILLDIR_FN(do_, osd_ldiskfs_filldir)
7234 * Calls ->iterate*() to load a directory entry at a time
7235 * and stored it in iterator's in-memory data structure.
7237 * \param di iterator's in memory structure
7239 * \retval 0 on success
7240 * \retval -ve on error
7241 * \retval +1 reach the end of entry
7243 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
7245 struct osd_it_ea *it = (struct osd_it_ea *)di;
7246 struct osd_object *obj = it->oie_obj;
7247 struct htree_lock *hlock = NULL;
7248 struct file *filp = it->oie_file;
7250 struct osd_filldir_cbs buf = {
7251 .ctx.actor = osd_ldiskfs_filldir,
7256 it->oie_dirent = it->oie_buf;
7257 it->oie_rd_dirent = 0;
7260 if (obj->oo_hl_head != NULL) {
7261 hlock = osd_oti_get(env)->oti_hlock;
7262 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
7264 LDISKFS_HLOCK_READDIR);
7266 down_read(&obj->oo_ext_idx_sem);
7270 rc = iterate_dir(filp, &buf.ctx);
7274 if (it->oie_rd_dirent == 0) {
7276 * If it does not get any dirent, it means it has been reached
7277 * to the end of the dir
7279 it->oie_file->f_pos = ldiskfs_get_htree_eof(it->oie_file);
7283 it->oie_dirent = it->oie_buf;
7284 it->oie_it_dirent = 1;
7289 ldiskfs_htree_unlock(hlock);
7291 up_read(&obj->oo_ext_idx_sem);
7298 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7299 * to load a directory entry at a time and stored it in
7300 * iterator's in-memory data structure.
7302 * \param di iterator's in memory structure
7304 * \retval +ve iterator reached to end
7305 * \retval 0 iterator not reached to end
7306 * \retval -ve on error
7308 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
7310 struct osd_it_ea *it = (struct osd_it_ea *)di;
7315 if (it->oie_it_dirent < it->oie_rd_dirent) {
7317 (void *)it->oie_dirent +
7318 round_up(sizeof(struct osd_it_ea_dirent) +
7319 it->oie_dirent->oied_namelen, 8);
7320 it->oie_it_dirent++;
7323 if (it->oie_file->f_pos == ldiskfs_get_htree_eof(it->oie_file))
7326 rc = osd_ldiskfs_it_fill(env, di);
7333 * Returns the key at current position from iterator's in memory structure.
7335 * \param di iterator's in memory structure
7337 * \retval key i.e. struct dt_key on success
7339 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
7340 const struct dt_it *di)
7342 struct osd_it_ea *it = (struct osd_it_ea *)di;
7344 return (struct dt_key *)it->oie_dirent->oied_name;
7348 * Returns key's size at current position from iterator's in memory structure.
7350 * \param di iterator's in memory structure
7352 * \retval key_size i.e. struct dt_key on success
7354 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
7356 struct osd_it_ea *it = (struct osd_it_ea *)di;
7358 return it->oie_dirent->oied_namelen;
7361 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
7362 #undef LDISKFS_DIR_REC_LEN
7363 # if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7364 # define LDISKFS_DIR_REC_LEN(de, dir) LDISKFS_DIR_ENTRY_LEN_((de), (dir))
7366 # define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
7370 #if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7371 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de), NULL)
7373 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de))
7376 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
7378 if (LDISKFS_DIR_REC_LEN_DIR(de) >=
7379 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
7386 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
7387 unsigned int blocksize, bool dotdot)
7390 return osd_dotdot_has_space(de);
7392 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
7393 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
7400 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
7401 handle_t *jh, struct dentry *dentry,
7402 const struct lu_fid *fid, struct buffer_head *bh,
7403 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
7406 struct inode *dir = dentry->d_parent->d_inode;
7407 struct inode *inode = dentry->d_inode;
7408 struct osd_fid_pack *rec;
7409 struct ldiskfs_dentry_param *ldp;
7410 int namelen = dentry->d_name.len;
7412 struct osd_thread_info *info = osd_oti_get(env);
7416 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7419 /* There is enough space to hold the FID-in-dirent. */
7420 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7421 rc = osd_ldiskfs_journal_get_write_access(jh, dir->i_sb, bh,
7426 de->name[namelen] = 0;
7427 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7428 rec->fp_len = sizeof(struct lu_fid) + 1;
7429 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7430 de->file_type |= LDISKFS_DIRENT_LUFID;
7431 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7438 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7442 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7443 osd_get_ldiskfs_dirent_param(ldp, fid);
7444 dentry->d_fsdata = (void *)ldp;
7445 dquot_initialize(dir);
7446 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7448 * It is too bad, we cannot reinsert the name entry back.
7449 * That means we lose it!
7453 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7454 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7455 namelen, dentry->d_name.name, PFID(fid), rc);
7461 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7462 struct osd_it_ea *it, struct lu_fid *fid,
7463 struct osd_inode_id *id, __u32 *attr)
7465 struct osd_thread_info *info = osd_oti_get(env);
7466 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7467 struct osd_device *dev = osd_obj2dev(obj);
7468 struct super_block *sb = osd_sb(dev);
7469 const char *devname = osd_name(dev);
7470 struct osd_it_ea_dirent *ent = it->oie_dirent;
7471 struct inode *dir = obj->oo_inode;
7472 struct htree_lock *hlock = NULL;
7473 struct buffer_head *bh = NULL;
7474 handle_t *jh = NULL;
7475 struct ldiskfs_dir_entry_2 *de;
7476 struct dentry *dentry;
7477 struct inode *inode;
7478 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7481 bool dotdot = false;
7487 if (ent->oied_name[0] == '.') {
7488 if (ent->oied_namelen == 1)
7491 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7495 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7496 inode = osd_iget(info, dev, id);
7497 if (IS_ERR(inode)) {
7498 rc = PTR_ERR(inode);
7499 if (rc == -ENOENT || rc == -ESTALE) {
7501 * Maybe dangling name entry, or
7502 * corrupted directory entry.
7504 *attr |= LUDA_UNKNOWN;
7507 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7508 "check_repair, dir = %lu/%u, name = %.*s, "
7509 "ino = %llu, rc = %d\n",
7510 devname, dir->i_ino, dir->i_generation,
7511 ent->oied_namelen, ent->oied_name,
7518 rc = obj_name2lu_name(obj, ent->oied_name, ent->oied_namelen, &ln);
7522 dentry = osd_child_dentry_by_inode(env, dir, ln.ln_name, ln.ln_namelen);
7523 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7524 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7530 * We need to ensure that the name entry is still valid.
7531 * Because it may be removed or renamed by other already.
7533 * The unlink or rename operation will start journal before PDO lock,
7534 * so to avoid deadlock, here we need to start journal handle before
7535 * related PDO lock also. But because we do not know whether there
7536 * will be something to be repaired before PDO lock, we just start
7537 * journal without conditions.
7539 * We may need to remove the name entry firstly, then insert back.
7540 * One credit is for user quota file update.
7541 * One credit is for group quota file update.
7542 * Two credits are for dirty inode.
7544 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7545 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7547 if (dev->od_dirent_journal != 0) {
7550 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7553 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7554 "check_repair, dir = %lu/%u, credits = %d, "
7555 "name = %.*s, ino = %llu: rc = %d\n",
7556 devname, dir->i_ino, dir->i_generation, credits,
7557 ent->oied_namelen, ent->oied_name,
7560 GOTO(out_inode, rc);
7563 if (obj->oo_hl_head != NULL) {
7564 hlock = osd_oti_get(env)->oti_hlock;
7566 * "0" means exclusive lock for the whole directory.
7567 * We need to prevent others access such name entry
7568 * during the delete + insert. Neither HLOCK_ADD nor
7569 * HLOCK_DEL cannot guarantee the atomicity.
7571 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7573 down_write(&obj->oo_ext_idx_sem);
7576 if (obj->oo_hl_head != NULL) {
7577 hlock = osd_oti_get(env)->oti_hlock;
7578 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7579 LDISKFS_HLOCK_LOOKUP);
7581 down_read(&obj->oo_ext_idx_sem);
7585 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7586 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7587 *attr |= LUDA_IGNORE;
7593 * For dotdot entry, if there is not enough space to hold the
7594 * FID-in-dirent, just keep them there. It only happens when the
7595 * device upgraded from 1.8 or restored from MDT file-level backup.
7596 * For the whole directory, only dotdot entry have no FID-in-dirent
7597 * and needs to get FID from LMA when readdir, it will not affect the
7600 if (dotdot && !osd_dotdot_has_space(de)) {
7601 *attr |= LUDA_UNKNOWN;
7607 if (lu_fid_eq(fid, &lma->lma_self_fid))
7610 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7611 struct lu_fid *tfid = &lma->lma_self_fid;
7613 if (likely(dotdot &&
7614 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7615 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7617 * It must be REMOTE_PARENT_DIR and as the
7618 * 'dotdot' entry of remote directory
7620 *attr |= LUDA_IGNORE;
7622 CDEBUG(D_LFSCK, "%s: expect remote agent "
7623 "parent directory, but got %.*s under "
7624 "dir = %lu/%u with the FID "DFID"\n",
7625 devname, ent->oied_namelen,
7626 ent->oied_name, dir->i_ino,
7627 dir->i_generation, PFID(tfid));
7629 *attr |= LUDA_UNKNOWN;
7636 if (!fid_is_zero(fid)) {
7637 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7639 if (rc == -ENOENT ||
7641 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7643 * linkEA does not recognize the dirent entry,
7644 * it may because the dirent entry corruption
7645 * and points to other's inode.
7647 CDEBUG(D_LFSCK, "%s: the target inode does not "
7648 "recognize the dirent, dir = %lu/%u, "
7649 " name = %.*s, ino = %llu, "
7650 DFID": rc = %d\n", devname, dir->i_ino,
7651 dir->i_generation, ent->oied_namelen,
7652 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7653 *attr |= LUDA_UNKNOWN;
7658 if (rc && rc != -ENODATA) {
7659 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7660 "dir = %lu/%u, name = %.*s, ino = %llu, "
7661 DFID": rc = %d\n", devname, dir->i_ino,
7662 dir->i_generation, ent->oied_namelen,
7663 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7664 *attr |= LUDA_UNKNOWN;
7672 * linkEA recognizes the dirent entry, the FID-in-LMA is
7673 * valid, trusted, in spite of fid_is_sane(fid) or not.
7675 if (*attr & LUDA_VERIFY_DRYRUN) {
7676 *fid = lma->lma_self_fid;
7677 *attr |= LUDA_REPAIR;
7684 dev->od_dirent_journal = 1;
7685 if (hlock != NULL) {
7686 ldiskfs_htree_unlock(hlock);
7689 up_read(&obj->oo_ext_idx_sem);
7695 *fid = lma->lma_self_fid;
7697 /* Update or append the FID-in-dirent. */
7698 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7699 bh, de, hlock, dotdot);
7701 *attr |= LUDA_REPAIR;
7703 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7704 "the dirent, dir = %lu/%u, name = %.*s, "
7705 "ino = %llu, "DFID": rc = %d\n",
7706 devname, dir->i_ino, dir->i_generation,
7707 ent->oied_namelen, ent->oied_name,
7708 ent->oied_ino, PFID(fid), rc);
7710 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7711 if (*attr & LUDA_VERIFY_DRYRUN) {
7712 if (fid_is_sane(fid)) {
7713 *attr |= LUDA_REPAIR;
7714 } else if (dev->od_index == 0) {
7715 lu_igif_build(fid, inode->i_ino,
7716 inode->i_generation);
7717 *attr |= LUDA_UPGRADE;
7725 dev->od_dirent_journal = 1;
7726 if (hlock != NULL) {
7727 ldiskfs_htree_unlock(hlock);
7730 up_read(&obj->oo_ext_idx_sem);
7737 if (unlikely(fid_is_sane(fid))) {
7739 * FID-in-dirent exists, but FID-in-LMA is lost.
7740 * Trust the FID-in-dirent, and add FID-in-LMA.
7742 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7744 *attr |= LUDA_REPAIR;
7746 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7747 "update dirent, dir = %lu/%u, "
7748 "name = %.*s, ino = %llu, "
7750 devname, dir->i_ino, dir->i_generation,
7751 ent->oied_namelen, ent->oied_name,
7752 ent->oied_ino, PFID(fid), rc);
7753 } else if (dev->od_index == 0) {
7754 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7756 * It is probably IGIF object. Only aappend the
7757 * FID-in-dirent. OI scrub will process FID-in-LMA.
7759 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7760 bh, de, hlock, dotdot);
7762 *attr |= LUDA_UPGRADE;
7764 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7765 "after the dirent, dir = %lu/%u, "
7766 "name = %.*s, ino = %llu, "
7768 devname, dir->i_ino, dir->i_generation,
7769 ent->oied_namelen, ent->oied_name,
7770 ent->oied_ino, PFID(fid), rc);
7779 if (hlock != NULL) {
7780 ldiskfs_htree_unlock(hlock);
7782 if (dev->od_dirent_journal != 0)
7783 up_write(&obj->oo_ext_idx_sem);
7785 up_read(&obj->oo_ext_idx_sem);
7789 ldiskfs_journal_stop(jh);
7793 if (rc >= 0 && !dirty)
7794 dev->od_dirent_journal = 0;
7795 if (ln.ln_name != ent->oied_name)
7802 * Returns the value at current position from iterator's in memory structure.
7804 * \param di struct osd_it_ea, iterator's in memory structure
7805 * \param attr attr requested for dirent.
7806 * \param lde lustre dirent
7808 * \retval 0 no error and \param lde has correct lustre dirent.
7809 * \retval -ve on error
7811 static inline int osd_it_ea_rec(const struct lu_env *env,
7812 const struct dt_it *di,
7813 struct dt_rec *dtrec, __u32 attr)
7815 struct osd_it_ea *it = (struct osd_it_ea *)di;
7816 struct osd_object *obj = it->oie_obj;
7817 struct osd_device *dev = osd_obj2dev(obj);
7818 struct osd_thread_info *oti = osd_oti_get(env);
7819 struct osd_inode_id *id = &oti->oti_id;
7820 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7821 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7822 __u32 ino = it->oie_dirent->oied_ino;
7827 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7829 if (attr & LUDA_VERIFY) {
7830 if (unlikely(is_remote_parent_ino(dev, ino))) {
7831 attr |= LUDA_IGNORE;
7833 * If the parent is on remote MDT, and there
7834 * is no FID-in-dirent, then we have to get
7835 * the parent FID from the linkEA.
7837 if (!fid_is_sane(fid) &&
7838 it->oie_dirent->oied_namelen == 2 &&
7839 it->oie_dirent->oied_name[0] == '.' &&
7840 it->oie_dirent->oied_name[1] == '.')
7841 osd_get_pfid_from_linkea(env, obj, fid);
7843 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7847 if (!fid_is_sane(fid))
7848 attr |= LUDA_UNKNOWN;
7850 attr &= ~LU_DIRENT_ATTRS_MASK;
7851 if (!fid_is_sane(fid)) {
7852 bool is_dotdot = false;
7854 if (it->oie_dirent->oied_namelen == 2 &&
7855 it->oie_dirent->oied_name[0] == '.' &&
7856 it->oie_dirent->oied_name[1] == '.')
7859 * If the parent is on remote MDT, and there
7860 * is no FID-in-dirent, then we have to get
7861 * the parent FID from the linkEA.
7863 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7864 rc = osd_get_pfid_from_linkea(env, obj, fid);
7866 if (is_dotdot == false &&
7867 CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7870 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7875 /* Pack the entry anyway, at least the offset is right. */
7876 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7877 it->oie_dirent->oied_name,
7878 it->oie_dirent->oied_namelen,
7879 it->oie_dirent->oied_type, attr);
7881 RETURN(rc > 0 ? 0 : rc);
7885 * Returns the record size size at current position.
7887 * This function will return record(lu_dirent) size in bytes.
7889 * \param[in] env execution environment
7890 * \param[in] di iterator's in memory structure
7891 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7892 * calculate the lu_dirent size.
7894 * \retval record size(in bytes & in memory) of the current lu_dirent
7897 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7900 struct osd_it_ea *it = (struct osd_it_ea *)di;
7902 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7906 * Returns a cookie for current position of the iterator head, so that
7907 * user can use this cookie to load/start the iterator next time.
7909 * \param di iterator's in memory structure
7911 * \retval cookie for current position, on success
7913 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7915 struct osd_it_ea *it = (struct osd_it_ea *)di;
7917 return it->oie_dirent->oied_off;
7921 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7922 * to load a directory entry at a time and stored it i inn,
7923 * in iterator's in-memory data structure.
7925 * \param di struct osd_it_ea, iterator's in memory structure
7927 * \retval +ve on success
7928 * \retval -ve on error
7930 static int osd_it_ea_load(const struct lu_env *env,
7931 const struct dt_it *di, __u64 hash)
7933 struct osd_it_ea *it = (struct osd_it_ea *)di;
7937 it->oie_file->f_pos = hash;
7939 rc = osd_ldiskfs_it_fill(env, di);
7949 static int osd_olc_lookup(const struct lu_env *env, struct osd_object *obj,
7950 u64 iversion, struct dt_rec *rec,
7951 const struct lu_name *ln, int *result)
7953 struct osd_thread_info *oti = osd_oti_get(env);
7954 struct osd_lookup_cache *olc = oti->oti_lookup_cache;
7955 struct osd_device *osd = osd_obj2dev(obj);
7956 struct osd_lookup_cache_object *cobj = &oti->oti_cobj;
7959 if (unlikely(olc == NULL))
7962 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
7964 * umount has happened, a new OSD could land to the previous
7965 * address so we can't use it any more, invalidate our cache
7967 memset(olc, 0, sizeof(*olc));
7968 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
7972 memset(cobj, 0, sizeof(*cobj));
7973 cobj->lco_osd = osd;
7974 cobj->lco_ino = obj->oo_inode->i_ino;
7975 cobj->lco_gen = obj->oo_inode->i_generation;
7976 cobj->lco_version = iversion;
7978 for (i = 0; i < OSD_LOOKUP_CACHE_MAX; i++) {
7979 struct osd_lookup_cache_entry *entry;
7981 entry = &olc->olc_entry[i];
7982 /* compare if osd/ino/generation/version match */
7983 if (memcmp(&entry->lce_obj, cobj, sizeof(*cobj)) != 0)
7985 if (entry->lce_namelen != ln->ln_namelen)
7987 if (memcmp(entry->lce_name, ln->ln_name, ln->ln_namelen) != 0)
7990 memcpy(rec, &entry->lce_fid, sizeof(entry->lce_fid));
7991 *result = entry->lce_rc;
7997 static void osd_olc_save(const struct lu_env *env, struct osd_object *obj,
7998 struct dt_rec *rec, const struct lu_name *ln,
7999 const int result, u64 iversion)
8001 struct osd_thread_info *oti = osd_oti_get(env);
8002 struct osd_lookup_cache_entry *entry;
8003 struct osd_lookup_cache *olc;
8005 if (unlikely(oti->oti_lookup_cache == NULL)) {
8006 OBD_ALLOC_PTR(oti->oti_lookup_cache);
8007 if (oti->oti_lookup_cache == NULL)
8011 olc = oti->oti_lookup_cache;
8012 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
8013 memset(olc, 0, sizeof(*olc));
8014 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
8017 entry = &olc->olc_entry[olc->olc_cur];
8019 /* invaliate cache slot if needed */
8020 if (entry->lce_obj.lco_osd)
8021 memset(&entry->lce_obj, 0, sizeof(entry->lce_obj));
8023 /* XXX: some kind of LRU */
8024 entry->lce_obj.lco_osd = osd_obj2dev(obj);
8025 entry->lce_obj.lco_ino = obj->oo_inode->i_ino;
8026 entry->lce_obj.lco_gen = obj->oo_inode->i_generation;
8027 entry->lce_obj.lco_version = iversion;
8029 LASSERT(ln->ln_namelen <= LDISKFS_NAME_LEN + 1);
8030 entry->lce_namelen = ln->ln_namelen;
8031 memcpy(entry->lce_name, ln->ln_name, ln->ln_namelen);
8032 memcpy(&entry->lce_fid, rec, sizeof(entry->lce_fid));
8033 entry->lce_rc = result;
8035 if (++olc->olc_cur == OSD_LOOKUP_CACHE_MAX)
8040 * Index lookup function for interoperability mode (b11826).
8042 * \param key, key i.e. file name to be searched
8044 * \retval +ve, on success
8045 * \retval -ve, on error
8047 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
8048 struct dt_rec *rec, const struct dt_key *key)
8050 struct osd_object *obj = osd_dt_obj(dt);
8057 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
8058 LINVRNT(osd_invariant(obj));
8060 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
8065 * grab version before actual lookup, so that we recognize potential
8066 * insert between osd_ea_lookup_rec() and osd_olc_save()
8068 iversion = inode_peek_iversion(obj->oo_inode);
8070 if (osd_olc_lookup(env, obj, iversion, rec, &ln, &result))
8071 GOTO(out, rc = result);
8073 rc = osd_ea_lookup_rec(env, obj, rec, &ln);
8077 osd_olc_save(env, obj, rec, &ln, rc, iversion);
8080 if (ln.ln_name != (char *)key)
8086 * Index and Iterator operations for interoperability
8087 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
8089 static const struct dt_index_operations osd_index_ea_ops = {
8090 .dio_lookup = osd_index_ea_lookup,
8091 .dio_declare_insert = osd_index_declare_ea_insert,
8092 .dio_insert = osd_index_ea_insert,
8093 .dio_declare_delete = osd_index_declare_ea_delete,
8094 .dio_delete = osd_index_ea_delete,
8096 .init = osd_it_ea_init,
8097 .fini = osd_it_ea_fini,
8098 .get = osd_it_ea_get,
8099 .put = osd_it_ea_put,
8100 .next = osd_it_ea_next,
8101 .key = osd_it_ea_key,
8102 .key_size = osd_it_ea_key_size,
8103 .rec = osd_it_ea_rec,
8104 .rec_size = osd_it_ea_rec_size,
8105 .store = osd_it_ea_store,
8106 .load = osd_it_ea_load
8110 static void *osd_key_init(const struct lu_context *ctx,
8111 struct lu_context_key *key)
8113 struct osd_thread_info *info;
8115 OBD_ALLOC_PTR(info);
8117 return ERR_PTR(-ENOMEM);
8119 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8120 if (info->oti_it_ea_buf == NULL)
8123 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
8125 info->oti_hlock = ldiskfs_htree_lock_alloc();
8126 if (info->oti_hlock == NULL)
8132 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8135 return ERR_PTR(-ENOMEM);
8138 static void osd_key_fini(const struct lu_context *ctx,
8139 struct lu_context_key *key, void *data)
8141 struct osd_thread_info *info = data;
8142 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
8143 struct osd_idmap_cache *idc = info->oti_ins_cache;
8145 if (info->oti_dio_pages) {
8147 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
8148 struct page *page = info->oti_dio_pages[i];
8150 LASSERT(PagePrivate2(page));
8151 LASSERT(PageLocked(page));
8152 ClearPagePrivate2(page);
8157 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
8158 PTLRPC_MAX_BRW_PAGES);
8161 if (info->oti_inode != NULL)
8163 if (info->oti_hlock != NULL)
8164 ldiskfs_htree_lock_free(info->oti_hlock);
8165 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8166 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
8167 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
8168 lu_buf_free(&info->oti_big_buf);
8170 LASSERT(info->oti_ins_cache_size > 0);
8171 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
8172 info->oti_ins_cache = NULL;
8173 info->oti_ins_cache_size = 0;
8175 if (info->oti_lookup_cache)
8176 OBD_FREE_PTR(info->oti_lookup_cache);
8180 static void osd_key_exit(const struct lu_context *ctx,
8181 struct lu_context_key *key, void *data)
8183 struct osd_thread_info *info = data;
8184 struct osd_lookup_cache *olc = info->oti_lookup_cache;
8187 memset(olc, 0, sizeof(*olc));
8188 LASSERT(info->oti_r_locks == 0);
8189 LASSERT(info->oti_w_locks == 0);
8190 LASSERT(info->oti_txns == 0);
8191 LASSERTF(info->oti_dio_pages_used == 0, "%d\n",
8192 info->oti_dio_pages_used);
8195 /* type constructor/destructor: osd_type_init, osd_type_fini */
8196 LU_TYPE_INIT_FINI(osd, &osd_key);
8198 struct lu_context_key osd_key = {
8199 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
8200 .lct_init = osd_key_init,
8201 .lct_fini = osd_key_fini,
8202 .lct_exit = osd_key_exit
8206 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
8207 const char *name, struct lu_device *next)
8209 struct osd_device *osd = osd_dev(d);
8210 int cplen = strscpy(osd->od_svname, name, sizeof(osd->od_svname));
8214 return osd_procfs_init(osd, name);
8217 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
8221 /* shutdown quota slave instance associated with the device */
8222 if (o->od_quota_slave_md != NULL) {
8223 struct qsd_instance *qsd = o->od_quota_slave_md;
8225 o->od_quota_slave_md = NULL;
8229 if (o->od_quota_slave_dt != NULL) {
8230 struct qsd_instance *qsd = o->od_quota_slave_dt;
8232 o->od_quota_slave_dt = NULL;
8236 seq_target_fini(env, &o->od_dt_dev);
8237 osd_scrub_cleanup(env, o);
8242 #ifdef HAVE_FLUSH_DELAYED_FPUT
8243 # define cfs_flush_delayed_fput() flush_delayed_fput()
8245 void (*cfs_flush_delayed_fput)(void);
8246 #endif /* HAVE_FLUSH_DELAYED_FPUT */
8248 static void osd_umount(const struct lu_env *env, struct osd_device *o)
8252 atomic_inc(&osd_mount_seq);
8254 if (o->od_mnt != NULL) {
8255 shrink_dcache_sb(osd_sb(o));
8256 osd_sync(env, &o->od_dt_dev);
8257 wait_event(o->od_commit_cb_done,
8258 !atomic_read(&o->od_commit_cb_in_flight));
8264 /* to be sure all delayed fput are finished */
8265 cfs_flush_delayed_fput();
8270 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8271 # ifndef LDISKFS_HAS_INCOMPAT_FEATURE
8272 /* Newer kernels provide the ldiskfs_set_feature_largedir() wrapper already,
8273 * which calls ldiskfs_update_dynamic_rev() to update ancient filesystems.
8274 * All ldiskfs filesystems are already v2, so it is a no-op and unnecessary.
8275 * This avoids maintaining patches to export this otherwise-useless function.
8277 void ldiskfs_update_dynamic_rev(struct super_block *sb)
8284 static int osd_mount(const struct lu_env *env,
8285 struct osd_device *o, struct lustre_cfg *cfg)
8287 const char *name = lustre_cfg_string(cfg, 0);
8288 const char *dev = lustre_cfg_string(cfg, 1);
8290 unsigned long page, s_flags = 0, lmd_flags = 0;
8291 struct page *__page;
8292 struct file_system_type *type;
8293 char *options = NULL;
8295 struct osd_thread_info *info = osd_oti_get(env);
8296 struct lu_fid *fid = &info->oti_fid;
8297 struct inode *inode;
8298 int rc = 0, force_over_1024tb = 0;
8302 if (o->od_mnt != NULL)
8305 if (strlen(dev) >= sizeof(o->od_mntdev))
8307 strcpy(o->od_mntdev, dev);
8309 str = lustre_cfg_buf(cfg, 2);
8310 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
8312 opts = lustre_cfg_string(cfg, 3);
8314 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
8315 CERROR("%s: device %s extents feature is not guaranteed to "
8316 "work on big-endian systems. Use \"bigendian_extents\" "
8317 "mount option to override.\n", name, dev);
8321 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
8322 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
8323 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");
8326 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
8327 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
8328 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");
8331 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8332 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
8333 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");
8337 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
8338 force_over_1024tb = 1;
8340 __page = alloc_page(GFP_KERNEL);
8342 GOTO(out, rc = -ENOMEM);
8343 page = (unsigned long)page_address(__page);
8344 options = (char *)page;
8347 /* strip out the options for back compatiblity */
8348 static const char * const sout[] = {
8355 /* strip out option we processed in osd */
8356 "bigendian_extents",
8360 "force_over_1024tb",
8365 strncat(options, opts, PAGE_SIZE);
8366 for (rc = 0, str = options; sout[rc]; ) {
8367 char *op = strstr(str, sout[rc]);
8374 if (op == options || *(op - 1) == ',') {
8375 str = op + strlen(sout[rc]);
8376 if (*str == ',' || *str == '\0') {
8377 *str == ',' ? str++ : str;
8378 memmove(op, str, strlen(str) + 1);
8381 for (str = op; *str != ',' && *str != '\0'; str++)
8385 strncat(options, "user_xattr,acl", PAGE_SIZE);
8388 /* Glom up mount options */
8389 if (*options != '\0')
8390 strncat(options, ",", PAGE_SIZE);
8391 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
8393 type = get_fs_type("ldiskfs");
8395 CERROR("%s: cannot find ldiskfs module\n", name);
8396 GOTO(out, rc = -ENODEV);
8399 s_flags |= SB_KERNMOUNT;
8400 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
8401 module_put(type->owner);
8403 if (IS_ERR(o->od_mnt)) {
8404 rc = PTR_ERR(o->od_mnt);
8406 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
8410 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
8411 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
8412 force_over_1024tb == 0) {
8413 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",
8415 GOTO(out_mnt, rc = -EINVAL);
8418 if (test_bit(LMD_FLG_DEV_RDONLY, &lmd_flags)) {
8419 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
8422 GOTO(out_mnt, rc = -EOPNOTSUPP);
8425 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
8426 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
8427 GOTO(out_mnt, rc = -EINVAL);
8430 if (ldiskfs_has_feature_fast_commit(o->od_mnt->mnt_sb)) {
8431 CERROR("%s: device %s is mounted with fast_commit that breaks recovery\n",
8433 GOTO(out_mnt, rc = -EOPNOTSUPP);
8436 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8437 #ifdef LDISKFS_MOUNT_DIRDATA
8438 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
8439 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
8440 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
8441 CWARN("%s: device %s was upgraded from Lustre-1.x without "
8442 "enabling the dirdata feature. If you do not want to "
8443 "downgrade to Lustre-1.x again, you can enable it via "
8444 "'tune2fs -O dirdata device'\n", name, dev);
8446 /* enable large_dir on MDTs to avoid REMOTE_PARENT_DIR overflow,
8447 * and on very large OSTs to avoid object directory overflow */
8448 if (unlikely(!ldiskfs_has_feature_largedir(o->od_mnt->mnt_sb) &&
8449 !strstr(name, "MGS"))) {
8450 ldiskfs_set_feature_largedir(o->od_mnt->mnt_sb);
8451 LCONSOLE_INFO("%s: enabled 'large_dir' feature on device %s\n",
8455 inode = osd_sb(o)->s_root->d_inode;
8456 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
8457 if (!o->od_dt_dev.dd_rdonly) {
8458 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
8460 CERROR("%s: failed to set lma on %s root inode\n",
8466 if (test_bit(LMD_FLG_NOSCRUB, &lmd_flags))
8467 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_NEVER;
8469 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
8470 /* do not use pagecache with flash-backed storage */
8471 o->od_writethrough_cache = 0;
8472 o->od_read_cache = 0;
8483 __free_page(__page);
8488 static struct lu_device *osd_device_fini(const struct lu_env *env,
8489 struct lu_device *d)
8491 struct osd_device *o = osd_dev(d);
8495 osd_index_backup(env, o, false);
8496 osd_shutdown(env, o);
8498 if (o->od_oi_table != NULL)
8499 osd_oi_fini(osd_oti_get(env), o);
8500 if (o->od_extent_bytes_percpu)
8501 free_percpu(o->od_extent_bytes_percpu);
8502 osd_obj_map_fini(o);
8508 static int osd_device_init0(const struct lu_env *env,
8509 struct osd_device *o,
8510 struct lustre_cfg *cfg)
8512 struct lu_device *l = osd2lu_dev(o);
8513 struct osd_thread_info *info;
8516 bool restored = false;
8519 /* if the module was re-loaded, env can loose its keys */
8520 rc = lu_env_refill((struct lu_env *)env);
8523 info = osd_oti_get(env);
8526 l->ld_ops = &osd_lu_ops;
8527 o->od_dt_dev.dd_ops = &osd_dt_ops;
8529 spin_lock_init(&o->od_osfs_lock);
8530 mutex_init(&o->od_otable_mutex);
8531 INIT_LIST_HEAD(&o->od_orphan_list);
8532 INIT_LIST_HEAD(&o->od_index_backup_list);
8533 INIT_LIST_HEAD(&o->od_index_restore_list);
8534 spin_lock_init(&o->od_lock);
8535 o->od_index_backup_policy = LIBP_NONE;
8537 init_waitqueue_head(&o->od_commit_cb_done);
8539 o->od_read_cache = 1;
8540 o->od_writethrough_cache = 1;
8541 o->od_enable_projid_xattr = 0;
8542 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
8543 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
8544 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
8545 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_DEFAULT;
8546 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8547 o->od_fallocate_zero_blocks = 0;
8549 cplen = strscpy(o->od_svname, lustre_cfg_string(cfg, 4),
8550 sizeof(o->od_svname));
8552 GOTO(out, rc = cplen);
8554 o->od_index_backup_stop = 0;
8555 o->od_index = -1; /* -1 means index is invalid */
8556 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8557 if (rc == LDD_F_SV_TYPE_OST)
8560 o->od_full_scrub_ratio = OFSR_DEFAULT;
8561 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8562 rc = osd_mount(env, o, cfg);
8566 /* Can only check block device after mount */
8567 o->od_nonrotational =
8568 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8570 rc = osd_obj_map_init(env, o);
8574 rc = lu_site_init(&o->od_site, l);
8576 GOTO(out_compat, rc);
8577 o->od_site.ls_bottom_dev = l;
8579 rc = lu_site_init_finish(&o->od_site);
8583 opts = lustre_cfg_string(cfg, 3);
8584 if (opts && strstr(opts, "resetoi"))
8587 INIT_LIST_HEAD(&o->od_ios_list);
8589 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
8591 GOTO(out_brw_stats, rc);
8593 /* setup scrub, including OI files initialization */
8595 rc = osd_scrub_setup(env, o, restored);
8598 GOTO(out_brw_stats, rc);
8600 rc = osd_procfs_init(o, o->od_svname);
8602 CERROR("%s: can't initialize procfs: rc = %d\n",
8604 GOTO(out_scrub, rc);
8607 LASSERT(l->ld_site->ls_linkage.next != NULL);
8608 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8610 /* initialize quota slave instance */
8611 /* currently it's no need to prepare qsd_instance_md for OST */
8612 if (!o->od_is_ost) {
8613 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8614 &o->od_dt_dev, o->od_proc_entry,
8616 if (IS_ERR(o->od_quota_slave_md)) {
8617 rc = PTR_ERR(o->od_quota_slave_md);
8618 o->od_quota_slave_md = NULL;
8619 GOTO(out_procfs, rc);
8623 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8624 o->od_proc_entry, false, true);
8626 if (IS_ERR(o->od_quota_slave_dt)) {
8627 if (o->od_quota_slave_md != NULL) {
8628 qsd_fini(env, o->od_quota_slave_md);
8629 o->od_quota_slave_md = NULL;
8632 rc = PTR_ERR(o->od_quota_slave_dt);
8633 o->od_quota_slave_dt = NULL;
8634 GOTO(out_procfs, rc);
8637 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8638 if (!o->od_extent_bytes_percpu) {
8640 GOTO(out_procfs, rc);
8648 osd_scrub_cleanup(env, o);
8650 lprocfs_fini_brw_stats(&o->od_brw_stats);
8652 lu_site_fini(&o->od_site);
8654 osd_obj_map_fini(o);
8661 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8662 struct lu_device_type *t,
8663 struct lustre_cfg *cfg)
8665 struct osd_device *o;
8670 return ERR_PTR(-ENOMEM);
8672 rc = dt_device_init(&o->od_dt_dev, t);
8675 * Because the ctx might be revived in dt_device_init,
8676 * refill the env here
8678 lu_env_refill((struct lu_env *)env);
8679 rc = osd_device_init0(env, o, cfg);
8681 dt_device_fini(&o->od_dt_dev);
8684 if (unlikely(rc != 0))
8687 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8690 static struct lu_device *osd_device_free(const struct lu_env *env,
8691 struct lu_device *d)
8693 struct osd_device *o = osd_dev(d);
8697 /* XXX: make osd top device in order to release reference */
8698 d->ld_site->ls_top_dev = d;
8699 lu_site_purge(env, d->ld_site, -1);
8700 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8701 D_ERROR, lu_cdebug_printer);
8702 lu_site_fini(&o->od_site);
8703 dt_device_fini(&o->od_dt_dev);
8708 static int osd_process_config(const struct lu_env *env,
8709 struct lu_device *d, struct lustre_cfg *cfg)
8711 struct osd_device *o = osd_dev(d);
8717 switch (cfg->lcfg_command) {
8719 rc = osd_mount(env, o, cfg);
8723 * For the case LCFG_PRE_CLEANUP is not called in advance,
8724 * that may happend if hit failure during mount process.
8726 osd_index_backup(env, o, false);
8727 lu_dev_del_linkage(d->ld_site, d);
8728 rc = osd_shutdown(env, o);
8731 LASSERT(&o->od_dt_dev);
8732 count = class_modify_config(cfg, PARAM_OSD,
8733 &o->od_dt_dev.dd_kobj);
8735 count = class_modify_config(cfg, PARAM_OST,
8736 &o->od_dt_dev.dd_kobj);
8737 rc = count > 0 ? 0 : count;
8739 case LCFG_PRE_CLEANUP:
8741 osd_index_backup(env, o,
8742 o->od_index_backup_policy != LIBP_NONE);
8752 static int osd_recovery_complete(const struct lu_env *env,
8753 struct lu_device *d)
8755 struct osd_device *osd = osd_dev(d);
8760 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8764 * start qsd instance on recovery completion, this notifies the quota
8765 * slave code that we are about to process new requests now
8767 rc = qsd_start(env, osd->od_quota_slave_dt);
8768 if (rc == 0 && osd->od_quota_slave_md != NULL)
8769 rc = qsd_start(env, osd->od_quota_slave_md);
8775 * we use exports to track all osd users
8777 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8778 struct obd_device *obd, struct obd_uuid *cluuid,
8779 struct obd_connect_data *data, void *localdata)
8781 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8782 struct lustre_handle conn;
8787 CDEBUG(D_CONFIG, "connect #%d\n", atomic_read(&osd->od_connects));
8789 rc = class_connect(&conn, obd, cluuid);
8793 *exp = class_conn2export(&conn);
8795 atomic_inc(&osd->od_connects);
8801 * once last export (we don't count self-export) disappeared
8802 * osd can be released
8804 static int osd_obd_disconnect(struct obd_export *exp)
8806 struct obd_device *obd = exp->exp_obd;
8807 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8808 int rc, release = 0;
8812 /* Only disconnect the underlying layers on the final disconnect. */
8813 release = atomic_dec_and_test(&osd->od_connects);
8815 rc = class_disconnect(exp); /* bz 9811 */
8817 if (rc == 0 && release)
8818 class_manual_cleanup(obd);
8822 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8823 struct lu_device *dev)
8825 struct osd_device *osd = osd_dev(dev);
8826 struct lr_server_data *lsd =
8827 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8832 if (osd->od_quota_slave_md != NULL) {
8833 /* set up quota slave objects for inode */
8834 result = qsd_prepare(env, osd->od_quota_slave_md);
8839 if (osd->od_quota_slave_dt != NULL) {
8840 /* set up quota slave objects for block */
8841 result = qsd_prepare(env, osd->od_quota_slave_dt);
8847 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8848 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8849 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8850 osd->od_index_in_idif = 1;
8852 osd->od_index_in_idif = 0;
8853 result = osd_register_proc_index_in_idif(osd);
8858 osd->od_index_in_idif = 1;
8862 result = seq_target_init(env, &osd->od_dt_dev,
8869 static const struct lu_object_operations osd_lu_obj_ops = {
8870 .loo_object_init = osd_object_init,
8871 .loo_object_delete = osd_object_delete,
8872 .loo_object_release = osd_object_release,
8873 .loo_object_free = osd_object_free,
8874 .loo_object_print = osd_object_print,
8875 .loo_object_invariant = osd_object_invariant
8878 const struct lu_device_operations osd_lu_ops = {
8879 .ldo_object_alloc = osd_object_alloc,
8880 .ldo_process_config = osd_process_config,
8881 .ldo_recovery_complete = osd_recovery_complete,
8882 .ldo_prepare = osd_prepare,
8883 .ldo_fid_alloc = fid_alloc_generic,
8886 static const struct lu_device_type_operations osd_device_type_ops = {
8887 .ldto_init = osd_type_init,
8888 .ldto_fini = osd_type_fini,
8890 .ldto_start = osd_type_start,
8891 .ldto_stop = osd_type_stop,
8893 .ldto_device_alloc = osd_device_alloc,
8894 .ldto_device_free = osd_device_free,
8896 .ldto_device_init = osd_device_init,
8897 .ldto_device_fini = osd_device_fini
8900 static struct lu_device_type osd_device_type = {
8901 .ldt_tags = LU_DEVICE_DT,
8902 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8903 .ldt_ops = &osd_device_type_ops,
8904 .ldt_ctx_tags = LCT_LOCAL,
8907 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8909 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8910 struct super_block *sb = osd_sb(osd);
8912 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8916 * lprocfs legacy support.
8918 static const struct obd_ops osd_obd_device_ops = {
8919 .o_owner = THIS_MODULE,
8920 .o_connect = osd_obd_connect,
8921 .o_disconnect = osd_obd_disconnect,
8922 .o_health_check = osd_health_check,
8925 static ssize_t delayed_unlink_mb_show(struct kobject *kobj,
8926 struct attribute *attr, char *buf)
8928 return snprintf(buf, PAGE_SIZE, "%d\n",
8929 ldiskfs_delayed_unlink_blocks >> 11);
8932 static ssize_t delayed_unlink_mb_store(struct kobject *kobj,
8933 struct attribute *attr,
8934 const char *buffer, size_t count)
8936 u64 delayed_unlink_bytes;
8939 rc = sysfs_memparse(buffer, count, &delayed_unlink_bytes, "MiB");
8943 ldiskfs_delayed_unlink_blocks = delayed_unlink_bytes >> 9;
8947 LUSTRE_RW_ATTR(delayed_unlink_mb);
8950 static ssize_t track_declares_assert_show(struct kobject *kobj,
8951 struct attribute *attr,
8954 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8957 static ssize_t track_declares_assert_store(struct kobject *kobj,
8958 struct attribute *attr,
8959 const char *buffer, size_t count)
8961 bool track_declares_assert;
8964 rc = kstrtobool(buffer, &track_declares_assert);
8968 ldiskfs_track_declares_assert = track_declares_assert;
8972 LUSTRE_RW_ATTR(track_declares_assert);
8974 static int __init osd_init(void)
8976 struct kobject *kobj;
8979 BUILD_BUG_ON(BH_DXLock >=
8980 sizeof(((struct buffer_head *)0)->b_state) * 8);
8981 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8982 /* please, try to keep osd_thread_info smaller than a page */
8983 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8986 rc = libcfs_setup();
8992 rc = lu_kmem_init(ldiskfs_caches);
8996 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8997 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8999 lu_kmem_fini(ldiskfs_caches);
9003 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9005 rc = sysfs_create_file(kobj,
9006 &lustre_attr_track_declares_assert.attr);
9008 CWARN("%s: track_declares_assert sysfs registration failed: rc = %d\n",
9013 rc = sysfs_create_file(kobj,
9014 &lustre_attr_delayed_unlink_mb.attr);
9016 CWARN("%s: delayed_unlink_mb registration failed: rc = %d\n",
9024 #ifndef HAVE_FLUSH_DELAYED_FPUT
9025 if (unlikely(cfs_flush_delayed_fput == NULL))
9026 cfs_flush_delayed_fput =
9027 cfs_kallsyms_lookup_name("flush_delayed_fput");
9033 static void __exit osd_exit(void)
9035 struct kobject *kobj;
9037 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9039 sysfs_remove_file(kobj,
9040 &lustre_attr_track_declares_assert.attr);
9043 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
9044 lu_kmem_fini(ldiskfs_caches);
9047 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
9048 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
9049 MODULE_VERSION(LUSTRE_VERSION_STRING);
9050 MODULE_LICENSE("GPL");
9052 module_init(osd_init);
9053 module_exit(osd_exit);