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,
511 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(osd_sb(dev), id->oii_ino);
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, &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,
672 * The cached OI mapping is trustable. If we cannot locate the inode
673 * via the cached OI mapping, then return the failure to the caller
674 * directly without further OI checking.
678 inode = osd_iget2(info, dev, id, &rc);
680 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
683 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
684 PFID(fid), id->oii_ino, rc);
690 __u32 saved_ino = id->oii_ino;
691 __u32 saved_gen = id->oii_gen;
694 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
696 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
698 * XXX: There are four possible cases:
700 * Backup/restore caused the OI invalid.
702 * Someone unlinked the object but NOT removed
703 * the OI mapping, such as mount target device
704 * as ldiskfs, and modify something directly.
706 * Someone just removed the object between the
707 * former oi_lookup and the iget. It is normal.
708 * 4. Other failure cases.
710 * Generally, when the device is mounted, it will
711 * auto check whether the system is restored from
712 * file-level backup or not. We trust such detect
713 * to distinguish the 1st case from the 2nd case:
714 * if the OI files are consistent but may contain
715 * stale OI mappings because of case 2, if iget()
716 * returns -ENOENT or -ESTALE, then it should be
721 * If the OI mapping was in OI file before the
722 * osd_iget_check(), but now, it is disappear,
723 * then it must be removed by race. That is a
729 * It is the OI scrub updated the OI mapping by race.
730 * The new OI mapping must be valid.
732 if (saved_ino != id->oii_ino ||
733 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
742 if (dev->od_scrub.os_scrub.os_file.sf_flags &
745 * It still can be the case 2, but we cannot
746 * distinguish it from the case 1. So return
747 * -EREMCHG to block current operation until
748 * OI scrub rebuilt the OI mappings.
757 if (inode->i_generation == id->oii_gen)
777 * \retval +v: new filter_fid does not contain self-fid
778 * \retval 0: filter_fid_18_23, contains self-fid
779 * \retval -v: other failure cases
781 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
782 struct dentry *dentry, struct lu_fid *fid)
784 struct filter_fid *ff = &info->oti_ff;
785 struct ost_id *ostid = &info->oti_ostid;
788 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
789 if (rc == sizeof(struct filter_fid_18_23)) {
790 struct filter_fid_18_23 *ff_old = (void *)ff;
792 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
793 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
795 * XXX: use 0 as the index for compatibility, the caller will
796 * handle index related issues when necessary.
799 ostid_to_fid(fid, ostid, 0);
800 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
802 } else if (rc >= 0) {
809 static int osd_lma_self_repair(struct osd_thread_info *info,
810 struct osd_device *osd, struct inode *inode,
811 const struct lu_fid *fid, __u32 compat)
816 LASSERT(current->journal_info == NULL);
818 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
819 osd_dto_credits_noquota[DTO_XATTR_SET]);
822 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
827 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
829 CWARN("%s: cannot self repair the LMA: rc = %d\n",
831 ldiskfs_journal_stop(jh);
835 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
837 struct osd_thread_info *info = osd_oti_get(env);
838 struct osd_device *osd = osd_obj2dev(obj);
839 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
840 struct lustre_mdt_attrs *lma = &loa->loa_lma;
841 struct inode *inode = obj->oo_inode;
842 struct dentry *dentry = &info->oti_obj_dentry;
843 struct lu_fid *fid = NULL;
844 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
849 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
850 (void *)loa, sizeof(*loa));
851 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
852 fid = &lma->lma_self_fid;
853 rc = osd_get_idif(info, inode, dentry, fid);
854 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
856 * For the given OST-object, if it has neither LMA nor
857 * FID in XATTR_NAME_FID, then the given FID (which is
858 * contained in the @obj, from client RPC for locating
859 * the OST-object) is trusted. We use it to generate
862 osd_lma_self_repair(info, osd, inode, rfid,
873 lustre_lma_swab(lma);
874 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
875 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
876 S_ISREG(inode->i_mode)))) {
877 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
878 "fid = "DFID", ino = %lu\n", osd_name(osd),
879 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
880 PFID(rfid), inode->i_ino);
883 fid = &lma->lma_self_fid;
884 if (lma->lma_compat & LMAC_STRIPE_INFO &&
886 obj->oo_pfid_in_lma = 1;
887 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
889 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
893 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
894 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
895 struct ost_id *oi = &info->oti_ostid;
896 struct lu_fid *fid1 = &info->oti_fid3;
897 __u32 idx = fid_idif_ost_idx(rfid);
900 * For old IDIF, the OST index is not part of the IDIF,
901 * Means that different OSTs may have the same IDIFs.
902 * Under such case, we need to make some compatible
903 * check to make sure to trigger OI scrub properly.
905 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
906 /* Given @rfid is new, LMA is old. */
907 fid_to_ostid(fid, oi);
908 ostid_to_fid(fid1, oi, idx);
909 if (lu_fid_eq(fid1, rfid)) {
910 if (osd->od_index_in_idif)
911 osd_lma_self_repair(info, osd,
925 struct osd_check_lmv_buf {
926 /* please keep it as first member */
927 struct dir_context ctx;
928 struct osd_thread_info *oclb_info;
929 struct osd_device *oclb_dev;
935 * It is called internally by ->iterate*() to filter out the
936 * local slave object's FID of the striped directory.
938 * \retval 1 found the local slave's FID
939 * \retval 0 continue to check next item
940 * \retval -ve for failure
942 #ifdef HAVE_FILLDIR_USE_CTX
943 static FILLDIR_TYPE do_osd_stripe_dir_filldir(struct dir_context *buf,
945 static int osd_stripe_dir_filldir(void *buf,
947 const char *name, int namelen,
948 loff_t offset, __u64 ino, unsigned int d_type)
950 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
951 struct osd_thread_info *oti = oclb->oclb_info;
952 struct lu_fid *fid = &oti->oti_fid3;
953 struct osd_inode_id *id = &oti->oti_id3;
954 struct osd_inode_id id2;
955 struct osd_device *dev = oclb->oclb_dev;
964 sscanf(name + 1, SFID, RFID(fid));
965 if (!fid_is_sane(fid))
968 if (osd_remote_fid(oti->oti_env, dev, fid))
971 osd_id_gen(id, ino, OSD_OII_NOGEN);
972 inode = osd_iget(oti, dev, id);
974 return PTR_ERR(inode);
977 osd_add_oi_cache(oti, dev, id, fid);
978 /* Check shard by scrub only if it has a problem with OI */
979 if (osd_oi_lookup(oti, dev, fid, &id2, 0) || !osd_id_eq(id, &id2))
980 osd_scrub_oi_insert(dev, fid, id, true);
981 oclb->oclb_found = true;
986 WRAP_FILLDIR_FN(do_, osd_stripe_dir_filldir)
989 * When lookup item under striped directory, we need to locate the master
990 * MDT-object of the striped directory firstly, then the client will send
991 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
992 * and the item's name. If the system is restored from MDT file level backup,
993 * then before the OI scrub completely built the OI files, the OI mappings of
994 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
995 * not a problem for the master MDT-object. Because when locate the master
996 * MDT-object, we will do name based lookup (for the striped directory itself)
997 * firstly, during such process we can setup the correct OI mapping for the
998 * master MDT-object. But it will be trouble for the slave MDT-object. Because
999 * the client will not trigger name based lookup on the MDT to locate the slave
1000 * MDT-object before locating item under the striped directory, then when
1001 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
1002 * is invalid and does not know what the right OI mapping is, then the MDT has
1003 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
1004 * the OI file, related OI mapping is unknown yet, please try again later. And
1005 * then client will re-try the RPC again and again until related OI mapping has
1006 * been updated. That is quite inefficient.
1008 * To resolve above trouble, we will handle it as the following two cases:
1010 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
1011 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
1012 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
1013 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
1014 * directly. Please check osd_fid_lookup().
1016 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
1017 * Under such case, during lookup the master MDT-object, we will lookup the
1018 * slave MDT-object via readdir against the master MDT-object, because the
1019 * slave MDT-objects information are stored as sub-directories with the name
1020 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
1021 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
1022 * the correct OI mapping for the slave MDT-object.
1024 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
1025 struct inode *inode)
1027 struct lu_buf *buf = &oti->oti_big_buf;
1029 struct lmv_mds_md_v1 *lmv1;
1030 struct osd_check_lmv_buf oclb = {
1031 .ctx.actor = osd_stripe_dir_filldir,
1034 .oclb_found = false,
1039 /* We should use the VFS layer to create a real dentry. */
1040 oti->oti_obj_dentry.d_inode = inode;
1041 oti->oti_obj_dentry.d_sb = inode->i_sb;
1043 filp = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
1048 filp->f_mode |= FMODE_64BITHASH;
1052 rc = __osd_xattr_get(inode, filp->f_path.dentry, XATTR_NAME_LMV,
1053 buf->lb_buf, buf->lb_len);
1054 if (rc == -ERANGE) {
1055 rc = __osd_xattr_get(inode, filp->f_path.dentry,
1056 XATTR_NAME_LMV, NULL, 0);
1058 lu_buf_realloc(buf, rc);
1059 if (buf->lb_buf == NULL)
1060 GOTO(out, rc = -ENOMEM);
1066 if (unlikely(rc == 0 || rc == -ENODATA))
1072 if (unlikely(buf->lb_buf == NULL)) {
1073 lu_buf_realloc(buf, rc);
1074 if (buf->lb_buf == NULL)
1075 GOTO(out, rc = -ENOMEM);
1081 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1085 oclb.oclb_items = 0;
1086 rc = iterate_dir(filp, &oclb.ctx);
1087 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1088 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1093 "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
1094 osd_ino2name(inode), inode->i_ino, inode->i_generation,
1103 * Is object in scrub inconsistent/stale list.
1105 * \a scrub has two lists, os_inconsistent_items contains mappings to fix, while
1106 * os_stale_items contains mappings failed to fix.
1108 static bool fid_in_scrub_list(struct lustre_scrub *scrub,
1109 const struct list_head *list,
1110 const struct lu_fid *fid)
1112 struct osd_inconsistent_item *oii;
1114 if (list_empty(list))
1117 spin_lock(&scrub->os_lock);
1118 list_for_each_entry(oii, list, oii_list) {
1119 if (lu_fid_eq(fid, &oii->oii_cache.oic_fid)) {
1120 spin_unlock(&scrub->os_lock);
1124 spin_unlock(&scrub->os_lock);
1129 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1130 const struct lu_fid *fid,
1131 const struct lu_object_conf *conf)
1133 struct osd_thread_info *info;
1134 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1135 struct osd_device *dev;
1136 struct osd_idmap_cache *oic;
1137 struct osd_inode_id *id;
1138 struct inode *inode = NULL;
1139 struct lustre_scrub *scrub;
1140 struct scrub_file *sf;
1141 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1146 bool remote = false;
1147 bool trusted = true;
1148 bool updated = false;
1149 bool checked = false;
1154 LINVRNT(osd_invariant(obj));
1155 LASSERT(obj->oo_inode == NULL);
1157 if (fid_is_sane(fid) == 0) {
1158 CERROR("%s: invalid FID "DFID"\n", ldev->ld_obd->obd_name,
1164 dev = osd_dev(ldev);
1165 scrub = &dev->od_scrub.os_scrub;
1166 sf = &scrub->os_file;
1167 info = osd_oti_get(env);
1169 oic = &info->oti_cache;
1171 if (CFS_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1175 * For the object is created as locking anchor, or for the object to
1176 * be created on disk. No need to osd_oi_lookup() at here because FID
1177 * shouldn't never be re-used, if it's really a duplicate FID from
1178 * unexpected reason, we should be able to detect it later by calling
1179 * do_create->osd_oi_insert().
1181 if (conf && conf->loc_flags & LOC_F_NEW)
1184 /* Search order: 1. per-thread cache. */
1185 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1190 /* Search order: 2. OI scrub pending list. */
1192 memset(id, 0, sizeof(struct osd_inode_id));
1193 if (fid_in_scrub_list(scrub, &scrub->os_inconsistent_items, fid) &&
1195 RETURN(-EINPROGRESS);
1197 stale = fid_in_scrub_list(scrub, &scrub->os_stale_items, fid);
1198 if (stale && CFS_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1202 * The OI mapping in the OI file can be updated by the OI scrub
1203 * when we locate the inode via FID. So it may be not trustable.
1207 /* Search order: 3. OI files. */
1208 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1209 if (result == -ENOENT) {
1210 if (!fid_is_norm(fid) ||
1211 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1212 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1214 GOTO(out, result = 0);
1219 /* -ESTALE is returned if inode of OST object doesn't exist */
1220 if (result == -ESTALE &&
1221 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1222 GOTO(out, result = 0);
1229 obj->oo_inode = NULL;
1230 /* for later passes through checks, not true on first pass */
1231 if (!IS_ERR_OR_NULL(inode))
1234 inode = osd_iget_check(info, dev, fid, id, trusted);
1235 if (!IS_ERR(inode)) {
1236 obj->oo_inode = inode;
1244 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1247 result = PTR_ERR(inode);
1248 if (result == -ENOENT || result == -ESTALE)
1249 GOTO(out, result = 0);
1251 if (result != -EREMCHG)
1255 /* don't trigger repeatedly for stale mapping */
1257 GOTO(out, result = -ESTALE);
1260 * We still have chance to get the valid inode: for the
1261 * object which is referenced by remote name entry, the
1262 * object on the local MDT will be linked under the dir
1263 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1265 * We do not know whether the object for the given FID
1266 * is referenced by some remote name entry or not, and
1267 * especially for DNE II, a multiple-linked object may
1268 * have many name entries reside on many MDTs.
1270 * To simplify the operation, OSD will not distinguish
1271 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1272 * only happened for the RPC from other MDT during the
1273 * OI scrub, or for the client side RPC with FID only,
1274 * such as FID to path, or from old connected client.
1277 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1281 flags |= SS_AUTO_PARTIAL;
1282 flags &= ~SS_AUTO_FULL;
1287 if (scrub->os_running) {
1288 if (scrub->os_partial_scan && !scrub->os_in_join)
1291 if (IS_ERR_OR_NULL(inode) || result) {
1292 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1293 GOTO(out, result = -EINPROGRESS);
1297 LASSERT(obj->oo_inode == inode);
1299 osd_scrub_oi_insert(dev, fid, id, true);
1303 if (dev->od_scrub.os_scrub.os_auto_scrub_interval == AS_NEVER) {
1305 GOTO(out, result = -EREMCHG);
1308 LASSERT(obj->oo_inode == inode);
1310 osd_add_oi_cache(info, dev, id, fid);
1315 if (IS_ERR_OR_NULL(inode) || result)
1316 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1318 rc1 = osd_scrub_start(env, dev, flags);
1319 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
1320 "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
1321 osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
1322 if (rc1 && rc1 != -EALREADY)
1323 GOTO(out, result = -EREMCHG);
1325 if (IS_ERR_OR_NULL(inode) || result)
1326 GOTO(out, result = -EINPROGRESS);
1329 LASSERT(obj->oo_inode == inode);
1334 if (unlikely(obj->oo_header))
1337 result = osd_check_lma(env, obj);
1341 LASSERTF(id->oii_ino == inode->i_ino &&
1342 id->oii_gen == inode->i_generation,
1343 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1344 PFID(fid), id->oii_ino, id->oii_gen,
1345 inode->i_ino, inode->i_generation);
1347 saved_ino = inode->i_ino;
1348 saved_gen = inode->i_generation;
1350 if (unlikely(result == -ENODATA)) {
1352 * If the OI scrub updated the OI mapping by race, it
1353 * must be valid. Trust the inode that has no LMA EA.
1358 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1361 * The OI mapping is still there, the inode is still
1362 * valid. It is just becaues the inode has no LMA EA.
1364 if (saved_ino == id->oii_ino &&
1365 saved_gen == id->oii_gen)
1369 * It is the OI scrub updated the OI mapping by race.
1370 * The new OI mapping must be valid.
1378 * "result == -ENOENT" means that the OI mappinghas been
1379 * removed by race, so the inode belongs to other object.
1381 * Others error can be returned directly.
1383 if (result == -ENOENT) {
1384 obj->oo_inode = NULL;
1389 if (result != -EREMCHG)
1395 * if two OST objects map to the same inode, and inode mode is
1396 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1397 * reserved by precreate, and not written yet, in this case, don't
1398 * set inode for the object whose FID mismatch, so that it can create
1399 * inode and not block precreate.
1401 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1402 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1403 obj->oo_inode = NULL;
1404 GOTO(out, result = 0);
1407 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1409 * "result == -ENOENT" means the cached OI mapping has been removed
1410 * from the OI file by race, above inode belongs to other object.
1412 if (result == -ENOENT) {
1413 obj->oo_inode = NULL;
1414 GOTO(out, result = 0);
1420 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1422 osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
1428 * It is the OI scrub updated the OI mapping by race.
1429 * The new OI mapping must be valid.
1437 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1438 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1440 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1442 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1444 obj->oo_pfid_in_lma = 1;
1445 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1447 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1448 } else if (result != -ENODATA) {
1453 obj->oo_compat_dot_created = 1;
1454 obj->oo_compat_dotdot_created = 1;
1456 if (S_ISDIR(inode->i_mode) &&
1457 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1458 osd_check_lmv(info, dev, inode);
1460 result = osd_attach_jinode(inode);
1465 GOTO(out, result = 0);
1467 LASSERT(!obj->oo_hl_head);
1468 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1470 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1473 if (!result && stale)
1474 osd_scrub_oi_resurrect(scrub, fid);
1476 if (result || !obj->oo_inode) {
1477 if (!IS_ERR_OR_NULL(inode))
1480 obj->oo_inode = NULL;
1482 fid_zero(&oic->oic_fid);
1485 LINVRNT(osd_invariant(obj));
1490 * Concurrency: shouldn't matter.
1492 static void osd_object_init0(struct osd_object *obj)
1494 LASSERT(obj->oo_inode != NULL);
1495 obj->oo_dt.do_body_ops = &osd_body_ops;
1496 obj->oo_dt.do_lu.lo_header->loh_attr |=
1497 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1501 * Concurrency: no concurrent access is possible that early in object
1504 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1505 const struct lu_object_conf *conf)
1507 struct osd_object *obj = osd_obj(l);
1510 LINVRNT(osd_invariant(obj));
1512 if (CFS_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1513 cfs_fail_val == 2) {
1514 struct osd_thread_info *info = osd_oti_get(env);
1515 struct osd_idmap_cache *oic = &info->oti_cache;
1516 /* invalidate thread cache */
1517 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1519 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1520 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1521 l->lo_header->loh_attr |= LOHA_EXISTS;
1525 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1526 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1527 if (result == 0 && obj->oo_inode != NULL) {
1528 struct osd_thread_info *oti = osd_oti_get(env);
1529 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1531 osd_object_init0(obj);
1532 if (unlikely(obj->oo_header))
1535 result = osd_get_lma(oti, obj->oo_inode,
1536 &oti->oti_obj_dentry, loa);
1539 * Convert LMAI flags to lustre LMA flags
1540 * and cache it to oo_lma_flags
1543 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1544 } else if (result == -ENODATA) {
1548 atomic_set(&obj->oo_dirent_count, LU_DIRENT_COUNT_UNSET);
1550 LINVRNT(osd_invariant(obj));
1555 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1556 * The left part is for value, binary mode.
1558 struct osd_xattr_entry {
1559 struct list_head oxe_list;
1563 struct rcu_head oxe_rcu;
1567 static int osd_oxc_get(struct osd_object *obj, const char *name,
1570 struct osd_xattr_entry *tmp;
1571 struct osd_xattr_entry *oxe = NULL;
1572 size_t namelen = strlen(name);
1576 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1577 if (namelen == tmp->oxe_namelen &&
1578 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1585 GOTO(out, rc = -ENOENT);
1587 if (!oxe->oxe_exist)
1588 GOTO(out, rc = -ENODATA);
1591 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1594 if (buf->lb_buf == NULL)
1597 if (buf->lb_len < rc)
1598 GOTO(out, rc = -ERANGE);
1600 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1607 static void osd_oxc_free(struct rcu_head *head)
1609 struct osd_xattr_entry *oxe;
1611 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1612 OBD_FREE(oxe, oxe->oxe_len);
1615 static void osd_oxc_add(struct osd_object *obj, const char *name,
1616 const char *buf, int buflen)
1618 struct osd_xattr_entry *oxe;
1619 struct osd_xattr_entry *old = NULL;
1620 struct osd_xattr_entry *tmp;
1621 size_t namelen = strlen(name);
1622 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1624 OBD_ALLOC(oxe, len);
1628 INIT_LIST_HEAD(&oxe->oxe_list);
1630 oxe->oxe_namelen = namelen;
1631 memcpy(oxe->oxe_buf, name, namelen);
1633 LASSERT(buf != NULL);
1634 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1635 oxe->oxe_exist = true;
1637 oxe->oxe_exist = false;
1640 /* this should be rarely called, just remove old and add new */
1641 spin_lock(&obj->oo_guard);
1642 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1643 if (namelen == tmp->oxe_namelen &&
1644 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1650 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1651 call_rcu(&old->oxe_rcu, osd_oxc_free);
1653 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1655 spin_unlock(&obj->oo_guard);
1658 static void osd_oxc_del(struct osd_object *obj, const char *name)
1660 struct osd_xattr_entry *oxe;
1661 size_t namelen = strlen(name);
1663 spin_lock(&obj->oo_guard);
1664 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1665 if (namelen == oxe->oxe_namelen &&
1666 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1667 list_del_rcu(&oxe->oxe_list);
1668 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1672 spin_unlock(&obj->oo_guard);
1675 static void osd_oxc_fini(struct osd_object *obj)
1677 struct osd_xattr_entry *oxe, *next;
1679 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1680 list_del(&oxe->oxe_list);
1681 OBD_FREE(oxe, oxe->oxe_len);
1686 * Concurrency: no concurrent access is possible that late in object
1689 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1691 struct osd_object *obj = osd_obj(l);
1692 struct lu_object_header *h = obj->oo_header;
1694 LINVRNT(osd_invariant(obj));
1697 dt_object_fini(&obj->oo_dt);
1698 if (obj->oo_hl_head != NULL)
1699 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1700 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1703 lu_object_header_free(h);
1707 * Concurrency: no concurrent access is possible that late in object
1710 static void osd_index_fini(struct osd_object *o)
1712 struct iam_container *bag;
1714 if (o->oo_dir != NULL) {
1715 bag = &o->oo_dir->od_container;
1716 if (o->oo_inode != NULL) {
1717 if (bag->ic_object == o->oo_inode)
1718 iam_container_fini(bag);
1720 OBD_FREE_PTR(o->oo_dir);
1726 OSD_TXN_OI_DELETE_CREDITS = 20,
1727 OSD_TXN_INODE_DELETE_CREDITS = 20
1734 #if OSD_THANDLE_STATS
1736 * Set time when the handle is allocated
1738 static void osd_th_alloced(struct osd_thandle *oth)
1740 oth->oth_alloced = ktime_get();
1744 * Set time when the handle started
1746 static void osd_th_started(struct osd_thandle *oth)
1748 oth->oth_started = ktime_get();
1752 * Check whether the we deal with this handle for too long.
1754 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1755 ktime_t alloced, ktime_t started,
1758 ktime_t now = ktime_get();
1760 LASSERT(dev != NULL);
1762 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1763 ktime_us_delta(started, alloced));
1764 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1765 ktime_us_delta(closed, started));
1766 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1767 ktime_us_delta(now, closed));
1769 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1770 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1771 oth, now, alloced, started, closed);
1776 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1778 ktime_t __closed = ktime_get(); \
1779 ktime_t __alloced = oth->oth_alloced; \
1780 ktime_t __started = oth->oth_started; \
1783 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1786 #else /* OSD_THANDLE_STATS */
1788 #define osd_th_alloced(h) do {} while(0)
1789 #define osd_th_started(h) do {} while(0)
1790 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1792 #endif /* OSD_THANDLE_STATS */
1795 * in some cases (like overstriped files) the same operations on the same
1796 * objects are declared many times and this may lead to huge number of
1797 * credits which can be a problem and/or cause performance degradation.
1798 * this function is to remember what declarations have been made within
1799 * a given thandle and then skip duplications.
1800 * limit it's scope so that regular small transactions don't need all
1801 * this overhead with allocations, lists.
1802 * also, limit scope to the specific objects like llogs, etc.
1804 static inline bool osd_check_special_fid(const struct lu_fid *f)
1806 if (fid_seq_is_llog(f->f_seq))
1808 if (f->f_seq == FID_SEQ_LOCAL_FILE &&
1809 f->f_oid == MDD_LOV_OBJ_OID)
1814 bool osd_tx_was_declared(const struct lu_env *env, struct osd_thandle *oth,
1815 struct dt_object *dt, enum dt_txn_op op, loff_t pos)
1817 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1818 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1819 struct osd_thread_info *oti = osd_oti_get(env);
1820 struct osd_obj_declare *old;
1825 /* small transactions don't need this overhead */
1826 if (oti->oti_declare_ops[DTO_OBJECT_CREATE] < 10 &&
1827 oti->oti_declare_ops[DTO_WRITE_BASE] < 10)
1830 if (osd_check_special_fid(fid) == 0)
1833 list_for_each_entry(old, &oth->ot_declare_list, old_list) {
1834 if (old->old_op == op && old->old_pos == pos &&
1835 lu_fid_eq(&old->old_fid, fid))
1839 if (unlikely(old == NULL))
1841 old->old_fid = *lu_object_fid(&dt->do_lu);
1844 list_add(&old->old_list, &oth->ot_declare_list);
1848 void osd_tx_declaration_free(struct osd_thandle *oth)
1850 struct osd_obj_declare *old, *tmp;
1852 list_for_each_entry_safe(old, tmp, &oth->ot_declare_list, old_list) {
1853 list_del_init(&old->old_list);
1859 * Concurrency: doesn't access mutable data.
1861 static int osd_param_is_not_sane(const struct osd_device *dev,
1862 const struct thandle *th)
1864 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1866 return oh->ot_credits > osd_transaction_size(dev);
1870 * Concurrency: shouldn't matter.
1872 static void osd_trans_commit_cb(struct super_block *sb,
1873 struct ldiskfs_journal_cb_entry *jcb, int error)
1875 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1876 struct thandle *th = &oh->ot_super;
1877 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1878 struct osd_device *osd = osd_dev(lud);
1879 struct dt_txn_commit_cb *dcb, *tmp;
1881 LASSERT(oh->ot_handle == NULL);
1884 CERROR("transaction @0x%p commit error: %d\n", th, error);
1886 /* call per-transaction callbacks if any */
1887 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1889 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1890 "commit callback entry: magic=%x name='%s'\n",
1891 dcb->dcb_magic, dcb->dcb_name);
1892 list_del_init(&dcb->dcb_linkage);
1893 dcb->dcb_func(NULL, th, dcb, error);
1896 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1897 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1898 wake_up(&osd->od_commit_cb_done);
1904 static struct thandle *osd_trans_create(const struct lu_env *env,
1905 struct dt_device *d)
1907 struct osd_thread_info *oti = osd_oti_get(env);
1908 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1909 struct osd_thandle *oh;
1915 CERROR("%s: someone try to start transaction under "
1916 "readonly mode, should be disabled.\n",
1917 osd_name(osd_dt_dev(d)));
1919 RETURN(ERR_PTR(-EROFS));
1922 /* on pending IO in this thread should left from prev. request */
1923 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1925 sb_start_write(osd_sb(osd_dt_dev(d)));
1927 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1929 sb_end_write(osd_sb(osd_dt_dev(d)));
1930 RETURN(ERR_PTR(-ENOMEM));
1933 oh->ot_quota_trans = &oti->oti_quota_trans;
1934 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1939 oh->oh_declared_ext = 0;
1940 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1941 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1942 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1943 INIT_LIST_HEAD(&oh->ot_declare_list);
1946 memset(oti->oti_declare_ops, 0,
1947 sizeof(oti->oti_declare_ops));
1948 memset(oti->oti_declare_ops_cred, 0,
1949 sizeof(oti->oti_declare_ops_cred));
1950 memset(oti->oti_declare_ops_used, 0,
1951 sizeof(oti->oti_declare_ops_used));
1953 oti->oti_ins_cache_depth++;
1958 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1960 struct osd_thread_info *oti = osd_oti_get(env);
1961 struct osd_thandle *oh;
1963 oh = container_of(th, struct osd_thandle, ot_super);
1964 LASSERT(oh != NULL);
1966 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1967 oti->oti_declare_ops[OSD_OT_CREATE],
1968 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1969 oti->oti_declare_ops_used[OSD_OT_CREATE],
1970 oti->oti_declare_ops[OSD_OT_DESTROY],
1971 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1972 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1973 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1974 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1975 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1976 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1977 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1978 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1979 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1980 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1981 oti->oti_declare_ops[OSD_OT_WRITE],
1982 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1983 oti->oti_declare_ops_used[OSD_OT_WRITE],
1984 oti->oti_declare_ops[OSD_OT_PUNCH],
1985 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1986 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1987 oti->oti_declare_ops[OSD_OT_QUOTA],
1988 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1989 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1990 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1991 oti->oti_declare_ops[OSD_OT_INSERT],
1992 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1993 oti->oti_declare_ops_used[OSD_OT_INSERT],
1994 oti->oti_declare_ops[OSD_OT_DELETE],
1995 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1996 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1997 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1998 oti->oti_declare_ops[OSD_OT_REF_ADD],
1999 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
2000 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
2001 oti->oti_declare_ops[OSD_OT_REF_DEL],
2002 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
2003 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
2007 * Concurrency: shouldn't matter.
2009 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
2012 struct osd_thread_info *oti = osd_oti_get(env);
2013 struct osd_device *dev = osd_dt_dev(d);
2015 struct osd_thandle *oh;
2020 LASSERT(current->journal_info == NULL);
2022 oh = container_of(th, struct osd_thandle, ot_super);
2023 LASSERT(oh != NULL);
2024 LASSERT(oh->ot_handle == NULL);
2025 if (unlikely(ldiskfs_track_declares_assert != 0)) {
2026 LASSERT(oti->oti_r_locks == 0);
2027 LASSERT(oti->oti_w_locks == 0);
2030 rc = dt_txn_hook_start(env, d, th);
2034 if (unlikely(osd_param_is_not_sane(dev, th))) {
2035 static unsigned long last_printed;
2036 static int last_credits;
2038 lprocfs_counter_add(dev->od_stats,
2039 LPROC_OSD_TOO_MANY_CREDITS, 1);
2042 * don't make noise on a tiny testing systems
2043 * actual credits misuse will be caught anyway
2045 if (last_credits != oh->ot_credits &&
2046 time_after(jiffies, last_printed +
2047 cfs_time_seconds(60)) &&
2048 osd_transaction_size(dev) > 512) {
2049 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
2050 oh->ot_credits, osd_transaction_size(dev));
2051 osd_trans_dump_creds(env, th);
2053 last_credits = oh->ot_credits;
2054 last_printed = jiffies;
2057 * XXX Limit the credits to 'max_transaction_buffers', and
2058 * let the underlying filesystem to catch the error if
2059 * we really need so many credits.
2061 * This should be removed when we can calculate the
2062 * credits precisely.
2064 oh->ot_credits = osd_transaction_size(dev);
2065 } else if (ldiskfs_track_declares_assert != 0) {
2067 * reserve few credits to prevent an assertion in JBD
2068 * our debugging mechanism will be able to detected
2069 * overuse. this can help to debug single-update
2072 oh->ot_credits += 10;
2073 if (unlikely(osd_param_is_not_sane(dev, th)))
2074 oh->ot_credits = osd_transaction_size(dev);
2077 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
2078 GOTO(out, rc = -EIO);
2081 * we ignore quota checks for system-owned files, but still
2082 * need to count blocks for uid/gid/projid
2084 osd_trans_declare_op(env, oh, OSD_OT_QUOTA, 3);
2087 * XXX temporary stuff. Some abstraction layer should
2090 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
2094 LASSERT(oti->oti_txns == 0);
2096 atomic_inc(&dev->od_commit_cb_in_flight);
2097 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
2108 static int osd_seq_exists(const struct lu_env *env,
2109 struct osd_device *osd, u64 seq)
2111 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
2112 struct seq_server_site *ss = osd_seq_site(osd);
2117 LASSERT(ss != NULL);
2118 LASSERT(ss->ss_server_fld != NULL);
2120 rc = osd_fld_lookup(env, osd, seq, range);
2123 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2124 osd_name(osd), seq, rc);
2128 RETURN(ss->ss_node_id == range->lsr_index);
2131 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2133 struct dt_txn_commit_cb *dcb;
2134 struct dt_txn_commit_cb *tmp;
2136 /* call per-transaction stop callbacks if any */
2137 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2139 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2140 "commit callback entry: magic=%x name='%s'\n",
2141 dcb->dcb_magic, dcb->dcb_name);
2142 list_del_init(&dcb->dcb_linkage);
2143 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2148 * Concurrency: shouldn't matter.
2150 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2153 struct osd_thread_info *oti = osd_oti_get(env);
2154 struct osd_thandle *oh;
2155 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2156 struct osd_device *osd = osd_dt_dev(th->th_dev);
2157 struct qsd_instance *qsd = osd_def_qsd(osd);
2158 struct lquota_trans *qtrans;
2159 LIST_HEAD(truncates);
2160 int rc = 0, remove_agents = 0;
2164 oh = container_of(th, struct osd_thandle, ot_super);
2166 remove_agents = oh->ot_remove_agents;
2168 qtrans = oh->ot_quota_trans;
2169 oh->ot_quota_trans = NULL;
2171 osd_tx_declaration_free(oh);
2173 /* move locks to local list, stop tx, execute truncates */
2174 list_splice(&oh->ot_trunc_locks, &truncates);
2176 if (oh->ot_handle != NULL) {
2179 handle_t *hdl = oh->ot_handle;
2182 * add commit callback
2183 * notice we don't do this in osd_trans_start()
2184 * as underlying transaction can change during truncate
2186 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2189 LASSERT(oti->oti_txns == 1);
2192 rc = dt_txn_hook_stop(env, th);
2194 CERROR("%s: failed in transaction hook: rc = %d\n",
2197 osd_trans_stop_cb(oh, rc);
2198 /* hook functions might modify th_sync */
2199 hdl->h_sync = th->th_sync;
2201 oh->ot_handle = NULL;
2202 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2204 CERROR("%s: failed to stop transaction: rc = %d\n",
2205 osd_name(osd), rc2);
2209 /* We preserve the origin behavior of ignoring any
2210 * failures with the underlying punch / truncate
2211 * operation. We do record for debugging if an error
2212 * does occur in the lctl dk logs.
2214 rc2 = osd_process_truncates(env, &truncates);
2216 CERROR("%s: failed truncate process: rc = %d\n",
2217 osd_name(osd), rc2);
2219 osd_trans_stop_cb(oh, th->th_result);
2223 osd_trunc_unlock_all(env, &truncates);
2225 /* inform the quota slave device that the transaction is stopping */
2226 qsd_op_end(env, qsd, qtrans);
2229 * as we want IO to journal and data IO be concurrent, we don't block
2230 * awaiting data IO completion in osd_do_bio(), instead we wait here
2231 * once transaction is submitted to the journal. all reqular requests
2232 * don't do direct IO (except read/write), thus this wait_event becomes
2235 * IMPORTANT: we have to wait till any IO submited by the thread is
2236 * completed otherwise iobuf may be corrupted by different request
2238 wait_event(iobuf->dr_wait,
2239 atomic_read(&iobuf->dr_numreqs) == 0);
2242 rc = iobuf->dr_error;
2244 osd_fini_iobuf(osd, iobuf);
2246 if (unlikely(remove_agents != 0))
2247 osd_process_scheduled_agent_removals(env, osd);
2249 LASSERT(oti->oti_ins_cache_depth > 0);
2250 oti->oti_ins_cache_depth--;
2251 /* reset OI cache for safety */
2252 if (oti->oti_ins_cache_depth == 0)
2253 oti->oti_ins_cache_used = 0;
2255 sb_end_write(osd_sb(osd));
2260 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2262 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2265 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2266 LASSERT(&dcb->dcb_func != NULL);
2267 if (dcb->dcb_flags & DCB_TRANS_STOP)
2268 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2270 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2275 struct osd_delayed_iput_work {
2276 struct work_struct diw_work;
2277 struct inode *diw_inode;
2280 static void osd_delayed_iput_fn(struct work_struct *work)
2282 struct osd_delayed_iput_work *diwork;
2283 struct inode *inode;
2285 diwork = container_of(work, struct osd_delayed_iput_work, diw_work);
2286 inode = diwork->diw_inode;
2287 CDEBUG(D_INODE, "%s: delayed iput (ino=%lu)\n",
2288 inode->i_sb->s_id, inode->i_ino);
2290 OBD_FREE_PTR(diwork);
2293 noinline void osd_delayed_iput(struct inode *inode,
2294 struct osd_delayed_iput_work *diwork)
2299 INIT_WORK(&diwork->diw_work, osd_delayed_iput_fn);
2300 diwork->diw_inode = inode;
2301 queue_work(LDISKFS_SB(inode->i_sb)->s_misc_wq,
2307 * Called just before object is freed. Releases all resources except for
2308 * object itself (that is released by osd_object_free()).
2310 * Concurrency: no concurrent access is possible that late in object
2313 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2315 struct osd_object *obj = osd_obj(l);
2316 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2317 struct inode *inode = obj->oo_inode;
2318 struct osd_delayed_iput_work *diwork = NULL;
2323 LINVRNT(osd_invariant(obj));
2326 * If object is unlinked remove fid->ino mapping from object index.
2329 osd_index_fini(obj);
2334 if (inode->i_blocks > ldiskfs_delayed_unlink_blocks)
2335 OBD_ALLOC(diwork, sizeof(*diwork));
2337 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2338 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2340 uid = i_uid_read(inode);
2341 gid = i_gid_read(inode);
2342 projid = i_projid_read(inode);
2344 obj->oo_inode = NULL;
2345 osd_delayed_iput(inode, diwork);
2347 /* do not rebalance quota if the caller needs to release memory
2348 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2349 * transaction and risk to deadlock - LU-12178 */
2350 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2353 if (!obj->oo_header && qsd) {
2354 struct osd_thread_info *info = osd_oti_get(env);
2355 struct lquota_id_info *qi = &info->oti_qi;
2357 /* Release granted quota to master if necessary */
2358 qi->lqi_id.qid_uid = uid;
2359 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2361 qi->lqi_id.qid_uid = gid;
2362 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2364 qi->lqi_id.qid_uid = projid;
2365 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2370 * Concurrency: ->loo_object_release() is called under site spin-lock.
2372 static void osd_object_release(const struct lu_env *env,
2373 struct lu_object *l)
2375 struct osd_object *o = osd_obj(l);
2378 * nobody should be releasing a non-destroyed object with nlink=0
2379 * the API allows this, but ldiskfs doesn't like and then report
2380 * this inode as deleted
2382 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2383 o->oo_inode->i_nlink == 0));
2387 * Concurrency: shouldn't matter.
2389 static int osd_object_print(const struct lu_env *env, void *cookie,
2390 lu_printer_t p, const struct lu_object *l)
2392 struct osd_object *o = osd_obj(l);
2393 struct iam_descr *d;
2395 if (o->oo_dir != NULL)
2396 d = o->oo_dir->od_container.ic_descr;
2399 return (*p)(env, cookie,
2400 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2402 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2403 o->oo_inode ? o->oo_inode->i_generation : 0,
2404 d ? d->id_ops->id_name : "plain");
2408 * Concurrency: shouldn't matter.
2410 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2411 struct obd_statfs *sfs, struct obd_statfs_info *info)
2413 struct osd_device *osd = osd_dt_dev(d);
2414 struct super_block *sb = osd_sb(osd);
2415 struct kstatfs *ksfs;
2419 if (unlikely(osd->od_mnt == NULL))
2420 return -EINPROGRESS;
2422 /* osd_lproc.c call this without env, allocate ksfs for that case */
2423 if (unlikely(env == NULL)) {
2424 OBD_ALLOC_PTR(ksfs);
2428 ksfs = &osd_oti_get(env)->oti_ksfs;
2431 result = sb->s_op->statfs(sb->s_root, ksfs);
2435 statfs_pack(sfs, ksfs);
2436 if (unlikely(sb->s_flags & SB_RDONLY))
2437 sfs->os_state |= OS_STATFS_READONLY;
2439 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2441 if (ldiskfs_has_feature_extents(sb))
2442 sfs->os_maxbytes = sb->s_maxbytes;
2444 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2447 * Reserve some space so to avoid fragmenting the filesystem too much.
2448 * Fragmentation not only impacts performance, but can also increase
2449 * metadata overhead significantly, causing grant calculation to be
2452 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2454 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2455 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2456 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2457 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2459 sfs->os_blocks -= reserved;
2460 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2461 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2464 if (unlikely(env == NULL))
2470 * Estimate space needed for file creations. We assume the largest filename
2471 * which is 2^64 - 1, hence a filename of 20 chars.
2472 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2474 #ifdef __LDISKFS_DIR_REC_LEN
2475 # define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2476 #elif defined LDISKFS_DIR_REC_LEN_WITH_DIR
2477 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20, NULL)
2479 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2483 * Concurrency: doesn't access mutable data.
2485 static void osd_conf_get(const struct lu_env *env,
2486 const struct dt_device *dev,
2487 struct dt_device_param *param)
2489 struct osd_device *d = osd_dt_dev(dev);
2490 struct super_block *sb = osd_sb(d);
2491 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2496 * XXX should be taken from not-yet-existing fs abstraction layer.
2498 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2499 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2500 param->ddp_symlink_max = sb->s_blocksize;
2501 param->ddp_mount_type = LDD_MT_LDISKFS;
2502 if (ldiskfs_has_feature_extents(sb))
2503 param->ddp_maxbytes = sb->s_maxbytes;
2505 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2507 * inode are statically allocated, so per-inode space consumption
2508 * is the space consumed by the directory entry
2510 param->ddp_inodespace = PER_OBJ_USAGE;
2512 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2513 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2514 * maximum length to not under-count the actual number of extents
2515 * needed for writing a file if there are sub-optimal block allocations.
2517 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2518 /* worst-case extent insertion metadata overhead */
2519 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2520 param->ddp_mntopts = 0;
2521 if (test_opt(sb, XATTR_USER))
2522 param->ddp_mntopts |= MNTOPT_USERXATTR;
2523 if (test_opt(sb, POSIX_ACL))
2524 param->ddp_mntopts |= MNTOPT_ACL;
2527 * LOD might calculate the max stripe count based on max_ea_size,
2528 * so we need take account in the overhead as well,
2529 * xattr_header + magic + xattr_entry_head
2531 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2532 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2534 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2535 if (ldiskfs_has_feature_ea_inode(sb))
2536 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2540 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2542 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2543 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2546 * Preferred RPC size for efficient disk IO. 4MB shows good
2547 * all-around performance for ldiskfs, but use bigalloc chunk size
2548 * by default if larger.
2550 #if defined(LDISKFS_CLUSTER_SIZE)
2551 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2552 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2555 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2557 param->ddp_t10_cksum_type = 0;
2559 unsigned short interval = blk_integrity_interval(bi);
2560 name = blk_integrity_name(bi);
2570 if (strncmp(name, "T10-DIF-TYPE",
2571 sizeof("T10-DIF-TYPE") - 1) == 0) {
2572 /* also skip "1/2/3-" at end */
2573 const int type_off = sizeof("T10-DIF-TYPE.");
2574 char type_number = name[type_off - 2];
2576 if (interval != 512 && interval != 4096) {
2577 CERROR("%s: unsupported T10PI sector size %u\n",
2578 d->od_svname, interval);
2581 switch (type_number) {
2583 d->od_t10_type = OSD_T10_TYPE1;
2586 d->od_t10_type = OSD_T10_TYPE2;
2589 d->od_t10_type = OSD_T10_TYPE3;
2592 CERROR("%s: unsupported T10PI type %s\n",
2593 d->od_svname, name);
2596 if (strcmp(name + type_off, "CRC") == 0) {
2597 d->od_t10_type |= OSD_T10_TYPE_CRC;
2598 param->ddp_t10_cksum_type = interval == 512 ?
2599 OBD_CKSUM_T10CRC512 :
2601 } else if (strcmp(name + type_off, "IP") == 0) {
2602 d->od_t10_type |= OSD_T10_TYPE_IP;
2603 param->ddp_t10_cksum_type = interval == 512 ?
2604 OBD_CKSUM_T10IP512 :
2607 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2608 d->od_svname, name);
2613 CERROR("%s: unsupported T10PI type '%s'\n",
2614 d->od_svname, name);
2619 param->ddp_has_lseek_data_hole = true;
2622 static struct vfsmount *osd_mnt_get(const struct dt_device *d)
2624 return osd_dt_dev(d)->od_mnt;
2628 * Concurrency: shouldn't matter.
2630 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2633 struct super_block *s = osd_sb(osd_dt_dev(d));
2636 down_read(&s->s_umount);
2637 rc = s->s_op->sync_fs(s, 1);
2638 up_read(&s->s_umount);
2640 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2647 * Start commit for OSD device.
2649 * An implementation of dt_commit_async method for OSD device.
2650 * Asychronously starts underlayng fs sync and thereby a transaction
2653 * \param env environment
2654 * \param d dt device
2656 * \see dt_device_operations
2658 static int osd_commit_async(const struct lu_env *env,
2659 struct dt_device *d)
2661 struct super_block *s = osd_sb(osd_dt_dev(d));
2666 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2667 down_read(&s->s_umount);
2668 rc = s->s_op->sync_fs(s, 0);
2669 up_read(&s->s_umount);
2675 * Concurrency: shouldn't matter.
2677 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2679 struct super_block *sb = osd_sb(osd_dt_dev(d));
2680 struct block_device *dev = sb->s_bdev;
2681 int rc = -EOPNOTSUPP;
2685 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2686 osd_dt_dev(d)->od_svname, (long)dev, rc);
2692 * Note: we do not count into QUOTA here.
2693 * If we mount with --data_journal we may need more.
2695 const int osd_dto_credits_noquota[DTO_NR] = {
2698 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2699 * SINGLEDATA_TRANS_BLOCKS(8)
2700 * XXX Note: maybe iam need more, since iam have more level than
2703 [DTO_INDEX_INSERT] = 16,
2706 * just modify a single entry, probably merge few within a block
2708 [DTO_INDEX_DELETE] = 1,
2712 [DTO_INDEX_UPDATE] = 16,
2714 * 4(inode, inode bits, groups, GDT)
2715 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2717 [DTO_OBJECT_CREATE] = 4,
2719 * 4(inode, inode bits, groups, GDT)
2720 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2722 [DTO_OBJECT_DELETE] = 4,
2724 * Attr set credits (inode)
2726 [DTO_ATTR_SET_BASE] = 1,
2728 * Xattr set. The same as xattr of EXT3.
2729 * DATA_TRANS_BLOCKS(14)
2730 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2731 * are also counted in. Do not know why?
2733 [DTO_XATTR_SET] = 14,
2735 * credits for inode change during write.
2737 [DTO_WRITE_BASE] = 3,
2739 * credits for single block write.
2741 [DTO_WRITE_BLOCK] = 14,
2743 * Attr set credits for chown.
2744 * This is extra credits for setattr, and it is null without quota
2746 [DTO_ATTR_SET_CHOWN] = 0
2749 /* reserve or free quota for some operation */
2750 static int osd_reserve_or_free_quota(const struct lu_env *env,
2751 struct dt_device *dev,
2752 struct lquota_id_info *qi)
2754 struct osd_device *osd = osd_dt_dev(dev);
2755 struct qsd_instance *qsd = NULL;
2761 qsd = osd->od_quota_slave_dt;
2763 qsd = osd->od_quota_slave_md;
2765 rc = qsd_reserve_or_free_quota(env, qsd, qi);
2769 static const struct dt_device_operations osd_dt_ops = {
2770 .dt_root_get = osd_root_get,
2771 .dt_statfs = osd_statfs,
2772 .dt_trans_create = osd_trans_create,
2773 .dt_trans_start = osd_trans_start,
2774 .dt_trans_stop = osd_trans_stop,
2775 .dt_trans_cb_add = osd_trans_cb_add,
2776 .dt_conf_get = osd_conf_get,
2777 .dt_mnt_get = osd_mnt_get,
2778 .dt_sync = osd_sync,
2780 .dt_commit_async = osd_commit_async,
2781 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
2784 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2787 struct osd_object *obj = osd_dt_obj(dt);
2788 struct osd_thread_info *oti = osd_oti_get(env);
2790 LINVRNT(osd_invariant(obj));
2792 LASSERT(obj->oo_owner != env);
2793 down_read_nested(&obj->oo_sem, role);
2795 LASSERT(obj->oo_owner == NULL);
2799 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2802 struct osd_object *obj = osd_dt_obj(dt);
2803 struct osd_thread_info *oti = osd_oti_get(env);
2805 LINVRNT(osd_invariant(obj));
2807 LASSERT(obj->oo_owner != env);
2808 down_write_nested(&obj->oo_sem, role);
2810 LASSERT(obj->oo_owner == NULL);
2811 obj->oo_owner = env;
2815 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2817 struct osd_object *obj = osd_dt_obj(dt);
2818 struct osd_thread_info *oti = osd_oti_get(env);
2820 LINVRNT(osd_invariant(obj));
2822 LASSERT(oti->oti_r_locks > 0);
2824 up_read(&obj->oo_sem);
2827 static void osd_write_unlock(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 LASSERT(oti->oti_w_locks > 0);
2837 obj->oo_owner = NULL;
2838 up_write(&obj->oo_sem);
2841 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2843 struct osd_object *obj = osd_dt_obj(dt);
2845 LINVRNT(osd_invariant(obj));
2847 return obj->oo_owner == env;
2850 static void osd_inode_getattr(const struct lu_env *env,
2851 struct inode *inode, struct lu_attr *attr)
2853 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2854 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2855 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2856 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2858 attr->la_atime = inode->i_atime.tv_sec;
2859 attr->la_mtime = inode->i_mtime.tv_sec;
2860 attr->la_ctime = inode->i_ctime.tv_sec;
2861 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2862 attr->la_mode = inode->i_mode;
2863 attr->la_size = i_size_read(inode);
2864 attr->la_blocks = inode->i_blocks;
2865 attr->la_uid = i_uid_read(inode);
2866 attr->la_gid = i_gid_read(inode);
2867 attr->la_projid = i_projid_read(inode);
2868 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2869 attr->la_nlink = inode->i_nlink;
2870 attr->la_rdev = inode->i_rdev;
2871 attr->la_blksize = 1 << inode->i_blkbits;
2872 attr->la_blkbits = inode->i_blkbits;
2874 * Ext4 did not transfer inherit flags from raw inode
2875 * to inode flags, and ext4 internally test raw inode
2876 * @i_flags directly. Instead of patching ext4, we do it here.
2878 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2879 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2882 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2885 struct osd_object *obj = osd_dt_obj(dt);
2886 const struct dt_it_ops *iops;
2892 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2893 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2895 /* directory not initialized yet */
2896 if (!dt->do_index_ops) {
2901 spin_lock(&obj->oo_guard);
2902 *count = atomic_read(&obj->oo_dirent_count);
2903 if (*count == LU_DIRENT_COUNT_UNSET)
2904 atomic_set(&obj->oo_dirent_count, 0);
2905 spin_unlock(&obj->oo_guard);
2906 if (*count != LU_DIRENT_COUNT_UNSET)
2910 iops = &dt->do_index_ops->dio_it;
2911 it = iops->init(env, dt, LUDA_64BITHASH);
2913 GOTO(out, rc = PTR_ERR(it));
2915 rc = iops->load(env, it, 0);
2922 rc = iops->next(env, it);
2924 for (; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2928 if (iops->key_size(env, it) == 0)
2933 if (rc == 1 || rc == -ESTALE)
2937 iops->fini(env, it);
2939 /* If counting dirents failed, use the current count (if any).
2941 * At worst this means the directory will not be split until the
2942 * count can be completed successfully (remount or oo_dirent_count
2943 * incremented by adding new entries). This avoids re-walking
2944 * the whole directory on each access and hitting the same error.
2946 if (rc && *count == 0)
2947 *count = LU_DIRENT_COUNT_UNSET;
2948 atomic_set(&obj->oo_dirent_count, *count);
2952 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2953 struct lu_attr *attr)
2955 struct osd_object *obj = osd_dt_obj(dt);
2958 if (unlikely(!dt_object_exists(dt)))
2960 if (unlikely(obj->oo_destroyed))
2963 LASSERT(!dt_object_remote(dt));
2964 LINVRNT(osd_invariant(obj));
2966 spin_lock(&obj->oo_guard);
2967 osd_inode_getattr(env, obj->oo_inode, attr);
2968 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2969 attr->la_valid |= LA_FLAGS;
2970 attr->la_flags |= LUSTRE_ORPHAN_FL;
2972 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2973 attr->la_valid |= LA_FLAGS;
2974 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2976 spin_unlock(&obj->oo_guard);
2978 if (S_ISDIR(obj->oo_inode->i_mode) &&
2979 (attr->la_valid & LA_DIRENT_CNT) &&
2980 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2981 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2983 attr->la_valid &= ~LA_DIRENT_CNT;
2988 static int osd_declare_attr_qid(const struct lu_env *env,
2989 struct osd_object *obj,
2990 struct osd_thandle *oh, long long bspace,
2991 qid_t old_id, qid_t new_id, bool enforce,
2995 struct osd_thread_info *info = osd_oti_get(env);
2996 struct lquota_id_info *qi = &info->oti_qi;
2998 qi->lqi_type = type;
2999 /* inode accounting */
3000 qi->lqi_is_blk = false;
3002 /* one more inode for the new id ... */
3003 qi->lqi_id.qid_uid = new_id;
3005 /* Reserve credits for the new id */
3006 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
3007 if (rc == -EDQUOT || rc == -EINPROGRESS)
3012 /* and one less inode for the current id */
3013 qi->lqi_id.qid_uid = old_id;
3015 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3016 if (rc == -EDQUOT || rc == -EINPROGRESS)
3021 /* block accounting */
3022 qi->lqi_is_blk = true;
3024 /* more blocks for the new id ... */
3025 qi->lqi_id.qid_uid = new_id;
3026 qi->lqi_space = bspace;
3028 * Credits for the new uid has been reserved, re-use "obj"
3029 * to save credit reservation.
3031 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3032 if (rc == -EDQUOT || rc == -EINPROGRESS)
3037 /* and finally less blocks for the current uid */
3038 qi->lqi_id.qid_uid = old_id;
3039 qi->lqi_space = -bspace;
3040 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3041 if (rc == -EDQUOT || rc == -EINPROGRESS)
3047 static int osd_declare_attr_set(const struct lu_env *env,
3048 struct dt_object *dt,
3049 const struct lu_attr *attr,
3050 struct thandle *handle)
3052 struct osd_thandle *oh;
3053 struct osd_object *obj;
3062 LASSERT(dt != NULL);
3063 LASSERT(handle != NULL);
3065 obj = osd_dt_obj(dt);
3066 LASSERT(osd_invariant(obj));
3068 oh = container_of(handle, struct osd_thandle, ot_super);
3069 LASSERT(oh->ot_handle == NULL);
3071 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
3072 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3074 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
3075 osd_dto_credits_noquota[DTO_XATTR_SET]);
3077 if (attr == NULL || obj->oo_inode == NULL)
3080 bspace = obj->oo_inode->i_blocks << 9;
3081 bspace = toqb(bspace);
3084 * Changing ownership is always preformed by super user, it should not
3085 * fail with EDQUOT unless required explicitly.
3087 * We still need to call the osd_declare_qid() to calculate the journal
3088 * credits for updating quota accounting files and to trigger quota
3089 * space adjustment once the operation is completed.
3091 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
3093 uid = i_uid_read(obj->oo_inode);
3094 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
3095 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
3096 attr->la_uid, enforce, USRQUOTA);
3100 gid = i_gid_read(obj->oo_inode);
3101 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
3102 attr->la_uid, gid, attr->la_gid);
3103 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
3104 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
3105 attr->la_gid, enforce, GRPQUOTA);
3110 #ifdef HAVE_PROJECT_QUOTA
3111 if (attr->la_valid & LA_PROJID) {
3112 __u32 projid = i_projid_read(obj->oo_inode);
3114 enforce = (attr->la_valid & LA_PROJID) &&
3115 (attr->la_projid != projid);
3116 rc = osd_declare_attr_qid(env, obj, oh, bspace,
3117 (qid_t)projid, (qid_t)attr->la_projid,
3123 /* punch must be aware we are dealing with an encrypted file */
3124 if (attr->la_valid & LA_FLAGS && attr->la_flags & LUSTRE_ENCRYPT_FL)
3125 obj->oo_lma_flags |= LUSTRE_ENCRYPT_FL;
3130 static int osd_inode_setattr(const struct lu_env *env,
3131 struct inode *inode, const struct lu_attr *attr)
3133 __u64 bits = attr->la_valid;
3135 /* Only allow set size for regular file */
3136 if (!S_ISREG(inode->i_mode))
3137 bits &= ~(LA_SIZE | LA_BLOCKS);
3142 if (bits & LA_ATIME)
3143 inode->i_atime = osd_inode_time(inode, attr->la_atime);
3144 if (bits & LA_CTIME)
3145 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
3146 if (bits & LA_MTIME)
3147 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
3148 if (bits & LA_SIZE) {
3149 spin_lock(&inode->i_lock);
3150 LDISKFS_I(inode)->i_disksize = attr->la_size;
3151 i_size_write(inode, attr->la_size);
3152 spin_unlock(&inode->i_lock);
3156 * OSD should not change "i_blocks" which is used by quota.
3157 * "i_blocks" should be changed by ldiskfs only.
3160 inode->i_mode = (inode->i_mode & S_IFMT) |
3161 (attr->la_mode & ~S_IFMT);
3163 i_uid_write(inode, attr->la_uid);
3165 i_gid_write(inode, attr->la_gid);
3166 if (bits & LA_PROJID)
3167 i_projid_write(inode, attr->la_projid);
3168 if (bits & LA_NLINK)
3169 set_nlink(inode, attr->la_nlink);
3171 inode->i_rdev = attr->la_rdev;
3173 if (bits & LA_FLAGS) {
3174 /* always keep S_NOCMTIME */
3175 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
3177 #if defined(S_ENCRYPTED)
3178 /* Always remove S_ENCRYPTED, because ldiskfs must not be
3179 * aware of encryption status. It is just stored into LMA
3180 * so that it can be forwared to client side.
3182 inode->i_flags &= ~S_ENCRYPTED;
3185 * Ext4 did not transfer inherit flags from
3186 * @inode->i_flags to raw inode i_flags when writing
3187 * flags, we do it explictly here.
3189 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
3190 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
3192 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
3197 #ifdef HAVE_PROJECT_QUOTA
3198 static int osd_transfer_project(struct inode *inode, __u32 projid,
3199 struct thandle *handle)
3201 struct super_block *sb = inode->i_sb;
3202 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3205 struct ldiskfs_iloc iloc;
3206 struct ldiskfs_inode *raw_inode;
3207 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3209 if (!ldiskfs_has_feature_project(sb)) {
3210 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3211 == LDISKFS_DEF_PROJID);
3212 if (projid != LDISKFS_DEF_PROJID)
3218 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3221 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3222 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3225 err = ldiskfs_get_inode_loc(inode, &iloc);
3229 raw_inode = ldiskfs_raw_inode(&iloc);
3230 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3231 struct osd_thandle *oh = container_of(handle,
3235 * try to expand inode size automatically.
3237 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3238 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3246 dquot_initialize(inode);
3247 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3248 if (transfer_to[PRJQUOTA]) {
3249 lock_dquot_transfer(inode);
3250 err = __dquot_transfer(inode, transfer_to);
3251 unlock_dquot_transfer(inode);
3252 dqput(transfer_to[PRJQUOTA]);
3261 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3262 struct thandle *handle)
3266 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3267 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3271 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3272 inode->i_ino, i_uid_read(inode), attr->la_uid,
3273 i_gid_read(inode), attr->la_gid);
3275 dquot_initialize(inode);
3277 if (attr->la_valid & LA_UID)
3278 iattr.ia_valid |= ATTR_UID;
3279 if (attr->la_valid & LA_GID)
3280 iattr.ia_valid |= ATTR_GID;
3281 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3282 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3284 lock_dquot_transfer(inode);
3285 rc = osd_dquot_transfer(&nop_mnt_idmap, inode, &iattr);
3286 unlock_dquot_transfer(inode);
3288 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3289 osd_ino2name(inode), rc);
3294 /* Handle project id transfer here properly */
3295 if (attr->la_valid & LA_PROJID &&
3296 attr->la_projid != i_projid_read(inode)) {
3297 if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
3299 #ifdef HAVE_PROJECT_QUOTA
3300 rc = osd_transfer_project(inode, attr->la_projid, handle);
3305 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3306 osd_ino2name(inode), rc);
3313 static int osd_attr_set(const struct lu_env *env,
3314 struct dt_object *dt,
3315 const struct lu_attr *attr,
3316 struct thandle *handle)
3318 struct osd_thread_info *info = osd_oti_get(env);
3319 struct osd_object *obj = osd_dt_obj(dt);
3320 struct osd_device *osd = osd_obj2dev(obj);
3321 struct inode *inode;
3324 if (!dt_object_exists(dt))
3327 LASSERT(handle != NULL);
3328 LASSERT(!dt_object_remote(dt));
3329 LASSERT(osd_invariant(obj));
3331 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3333 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3334 !osd_obj2dev(obj)->od_is_ost) {
3335 struct osd_thread_info *oti = osd_oti_get(env);
3336 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3337 struct lu_fid *fid1 = &oti->oti_fid;
3338 struct osd_inode_id *id = &oti->oti_id;
3339 struct iam_path_descr *ipd;
3340 struct iam_container *bag;
3341 struct osd_thandle *oh;
3344 fid_cpu_to_be(fid1, fid0);
3345 memset(id, 1, sizeof(*id));
3346 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3347 fid0)->oi_dir.od_container;
3348 ipd = osd_idx_ipd_get(env, bag);
3349 if (unlikely(ipd == NULL))
3352 oh = container_of(handle, struct osd_thandle, ot_super);
3353 rc = iam_update(oh->ot_handle, bag,
3354 (const struct iam_key *)fid1,
3355 (const struct iam_rec *)id, ipd);
3356 osd_ipd_put(env, bag, ipd);
3357 return(rc > 0 ? 0 : rc);
3360 inode = obj->oo_inode;
3362 rc = osd_quota_transfer(inode, attr, handle);
3366 spin_lock(&obj->oo_guard);
3367 rc = osd_inode_setattr(env, inode, attr);
3368 spin_unlock(&obj->oo_guard);
3372 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3374 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3376 if (!(attr->la_valid & LA_FLAGS))
3379 /* If setting LUSTRE_ENCRYPT_FL on an OST object, also set a dummy
3380 * enc ctx xattr, with 2 benefits:
3381 * - setting the LL_XATTR_NAME_ENCRYPTION_CONTEXT xattr internally sets
3382 * the LDISKFS_ENCRYPT_FL flag on the on-disk inode;
3383 * - it makes e2fsprogs happy to see an enc ctx for an inode that has
3384 * the LDISKFS_ENCRYPT_FL flag
3385 * We do not need the actual encryption context on OST objects, it is
3386 * only stored on MDT inodes, at file creation time.
3388 if (!(LDISKFS_I(obj->oo_inode)->i_flags & LDISKFS_ENCRYPT_FL) &&
3389 attr->la_flags & LUSTRE_ENCRYPT_FL && osd->od_is_ost &&
3390 !CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) {
3393 /* use a dummy enc ctx, fine with e2fsprogs */
3394 buf.lb_buf = "\xFF";
3396 rc = osd_xattr_set(env, dt, &buf,
3397 LL_XATTR_NAME_ENCRYPTION_CONTEXT,
3400 CWARN("%s: set "DFID" enc ctx failed: rc = %d\n",
3401 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3405 /* Let's check if there are extra flags need to be set into LMA */
3406 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3407 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3409 LASSERT(!obj->oo_pfid_in_lma);
3411 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3412 &info->oti_ost_attrs);
3416 if ((lma->lma_incompat & lustre_to_lma_flags(attr->la_flags)) ==
3417 lustre_to_lma_flags(attr->la_flags))
3418 /* if lma incompat already has the flags,
3419 * save a useless call to xattr_set
3423 lma->lma_incompat |=
3424 lustre_to_lma_flags(attr->la_flags);
3425 lustre_lma_swab(lma);
3427 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3429 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3430 lma, sizeof(*lma), XATTR_REPLACE);
3432 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3433 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3434 lma->lma_incompat, rc);
3437 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3438 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3445 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3446 struct osd_object *obj,
3447 const char *name, const int namelen)
3449 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3452 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3453 umode_t mode, struct dt_allocation_hint *hint,
3454 struct thandle *th, struct lu_attr *attr)
3457 struct osd_device *osd = osd_obj2dev(obj);
3458 struct osd_thandle *oth;
3459 struct dt_object *parent = NULL;
3460 struct inode *inode;
3461 struct iattr iattr = {
3462 .ia_valid = ATTR_UID | ATTR_GID |
3463 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3464 .ia_ctime.tv_sec = attr->la_ctime,
3465 .ia_mtime.tv_sec = attr->la_mtime,
3466 .ia_atime.tv_sec = attr->la_atime,
3467 .ia_uid = GLOBAL_ROOT_UID,
3468 .ia_gid = GLOBAL_ROOT_GID,
3470 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3472 if (attr->la_valid & LA_UID)
3473 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3474 if (attr->la_valid & LA_GID)
3475 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3477 LINVRNT(osd_invariant(obj));
3478 LASSERT(obj->oo_inode == NULL);
3479 LASSERT(obj->oo_hl_head == NULL);
3481 if (S_ISDIR(mode) && ldiskfs_pdo) {
3483 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3484 if (obj->oo_hl_head == NULL)
3488 oth = container_of(th, struct osd_thandle, ot_super);
3489 LASSERT(oth->ot_handle->h_transaction != NULL);
3491 if (hint != NULL && hint->dah_parent != NULL &&
3492 !dt_object_remote(hint->dah_parent))
3493 parent = hint->dah_parent;
3495 /* if a time component is not valid set it to UTIME_OMIT */
3496 if (!(attr->la_valid & LA_CTIME))
3497 iattr.ia_ctime = omit;
3498 if (!(attr->la_valid & LA_MTIME))
3499 iattr.ia_mtime = omit;
3500 if (!(attr->la_valid & LA_ATIME))
3501 iattr.ia_atime = omit;
3503 inode = ldiskfs_create_inode(oth->ot_handle,
3504 parent ? osd_dt_obj(parent)->oo_inode :
3505 osd_sb(osd)->s_root->d_inode,
3507 if (!IS_ERR(inode)) {
3508 /* Do not update file c/mtime in ldiskfs. */
3509 inode->i_flags |= S_NOCMTIME;
3512 * For new created object, it must be consistent,
3513 * and it is unnecessary to scrub against it.
3515 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3517 obj->oo_inode = inode;
3520 if (obj->oo_hl_head != NULL) {
3521 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3522 obj->oo_hl_head = NULL;
3524 result = PTR_ERR(inode);
3526 LINVRNT(osd_invariant(obj));
3534 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3535 struct lu_attr *attr,
3536 struct dt_allocation_hint *hint,
3537 struct dt_object_format *dof,
3541 struct osd_thandle *oth;
3542 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3544 LASSERT(S_ISDIR(attr->la_mode));
3546 oth = container_of(th, struct osd_thandle, ot_super);
3547 LASSERT(oth->ot_handle->h_transaction != NULL);
3548 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3549 atomic_set(&obj->oo_dirent_count, 0);
3550 result = osd_mkfile(info, obj, mode, hint, th, attr);
3555 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3556 struct lu_attr *attr,
3557 struct dt_allocation_hint *hint,
3558 struct dt_object_format *dof,
3562 struct osd_thandle *oth;
3563 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3565 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3567 LASSERT(S_ISREG(attr->la_mode));
3569 oth = container_of(th, struct osd_thandle, ot_super);
3570 LASSERT(oth->ot_handle->h_transaction != NULL);
3572 result = osd_mkfile(info, obj, mode, hint, th, attr);
3574 LASSERT(obj->oo_inode != NULL);
3575 if (feat->dif_flags & DT_IND_VARKEY)
3576 result = iam_lvar_create(obj->oo_inode,
3577 feat->dif_keysize_max,
3579 feat->dif_recsize_max,
3582 result = iam_lfix_create(obj->oo_inode,
3583 feat->dif_keysize_max,
3585 feat->dif_recsize_max,
3591 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3592 struct lu_attr *attr,
3593 struct dt_allocation_hint *hint,
3594 struct dt_object_format *dof,
3597 LASSERT(S_ISREG(attr->la_mode));
3598 return osd_mkfile(info, obj, (attr->la_mode &
3599 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3603 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3604 struct lu_attr *attr,
3605 struct dt_allocation_hint *hint,
3606 struct dt_object_format *dof,
3609 LASSERT(S_ISLNK(attr->la_mode));
3610 return osd_mkfile(info, obj, (attr->la_mode &
3611 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3615 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3616 struct lu_attr *attr,
3617 struct dt_allocation_hint *hint,
3618 struct dt_object_format *dof,
3621 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3624 LINVRNT(osd_invariant(obj));
3625 LASSERT(obj->oo_inode == NULL);
3626 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3627 S_ISFIFO(mode) || S_ISSOCK(mode));
3629 result = osd_mkfile(info, obj, mode, hint, th, attr);
3631 LASSERT(obj->oo_inode != NULL);
3633 * This inode should be marked dirty for i_rdev. Currently
3634 * that is done in the osd_attr_init().
3636 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3639 LINVRNT(osd_invariant(obj));
3643 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3645 struct dt_allocation_hint *hint,
3646 struct dt_object_format *dof,
3649 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3651 osd_obj_type_f result;
3667 result = osd_mk_index;
3677 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3678 struct dt_object *parent, struct dt_object *child,
3683 ah->dah_parent = parent;
3685 if (parent != NULL && !dt_object_remote(parent)) {
3686 /* will help to find FID->ino at dt_insert("..") */
3687 struct osd_object *pobj = osd_dt_obj(parent);
3689 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3693 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3694 struct lu_attr *attr, struct dt_object_format *dof,
3695 struct thandle *handle)
3697 struct inode *inode = obj->oo_inode;
3698 __u64 valid = attr->la_valid;
3701 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3703 if (dof->dof_type != DFT_NODE)
3704 attr->la_valid &= ~LA_RDEV;
3705 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3706 attr->la_valid &= ~LA_ATIME;
3707 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3708 attr->la_valid &= ~LA_CTIME;
3709 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3710 attr->la_valid &= ~LA_MTIME;
3712 result = osd_quota_transfer(inode, attr, handle);
3716 if (attr->la_valid != 0) {
3717 result = osd_inode_setattr(info->oti_env, inode, attr);
3719 * The osd_inode_setattr() should always succeed here. The
3720 * only error that could be returned is EDQUOT when we are
3721 * trying to change the UID or GID of the inode. However, this
3722 * should not happen since quota enforcement is no longer
3723 * enabled on ldiskfs (lquota takes care of it).
3725 LASSERTF(result == 0, "%d\n", result);
3726 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3729 attr->la_valid = valid;
3733 * Helper function for osd_create()
3735 * \retval 0, on success
3737 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3738 struct lu_attr *attr, struct dt_allocation_hint *hint,
3739 struct dt_object_format *dof, struct thandle *th)
3744 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3746 /* we drop umask so that permissions we pass are not affected */
3747 umask = current->fs->umask;
3748 current->fs->umask = 0;
3750 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3752 if (likely(obj->oo_inode != NULL)) {
3753 LASSERT(obj->oo_inode->i_state & I_NEW);
3756 * Unlock the inode before attr initialization to avoid
3757 * unnecessary dqget operations. LU-6378
3759 unlock_new_inode(obj->oo_inode);
3762 if (likely(result == 0)) {
3763 osd_attr_init(info, obj, attr, dof, th);
3764 osd_object_init0(obj);
3767 /* restore previous umask value */
3768 current->fs->umask = umask;
3770 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3776 * Helper function for osd_create()
3778 * \retval 0, on success
3780 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3781 const struct lu_fid *fid, struct thandle *th)
3783 struct osd_thread_info *info = osd_oti_get(env);
3784 struct osd_inode_id *id = &info->oti_id;
3785 struct osd_device *osd = osd_obj2dev(obj);
3786 struct osd_thandle *oh;
3789 LASSERT(obj->oo_inode != NULL);
3791 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3794 oh = container_of(th, struct osd_thandle, ot_super);
3795 LASSERT(oh->ot_handle);
3796 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3798 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3799 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_REUSE) && osd->od_is_ost &&
3801 struct lu_fid tfid = *fid;
3804 osd_oi_insert(info, osd, &tfid, id, oh->ot_handle,
3805 OI_CHECK_FLD, NULL);
3806 /* clear NOSCRUB flag so that it can be scrubbed immediately */
3807 ldiskfs_clear_inode_state(obj->oo_inode,
3808 LDISKFS_STATE_LUSTRE_NOSCRUB);
3811 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3812 OI_CHECK_FLD, NULL);
3814 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3815 struct lu_fid next_fid = *fid;
3817 /* insert next object in advance, and map to the same inode */
3819 if (next_fid.f_oid != 0) {
3820 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3821 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3822 OI_CHECK_FLD, NULL);
3823 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3827 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3832 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3833 u64 seq, struct lu_seq_range *range)
3835 struct seq_server_site *ss = osd_seq_site(osd);
3837 if (fid_seq_is_idif(seq)) {
3838 fld_range_set_ost(range);
3839 range->lsr_index = idif_ost_idx(seq);
3843 if (!fid_seq_in_fldb(seq)) {
3844 fld_range_set_mdt(range);
3847 * FIXME: If ss is NULL, it suppose not get lsr_index
3850 range->lsr_index = ss->ss_node_id;
3854 LASSERT(ss != NULL);
3855 fld_range_set_any(range);
3856 /* OSD will only do local fld lookup */
3857 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3860 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3861 struct lu_attr *attr,
3862 struct dt_allocation_hint *hint,
3863 struct dt_object_format *dof,
3864 struct thandle *handle)
3866 struct osd_thandle *oh;
3867 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3873 LASSERT(handle != NULL);
3875 oh = container_of(handle, struct osd_thandle, ot_super);
3876 LASSERT(oh->ot_handle == NULL);
3878 if (osd_tx_was_declared(env, oh, dt, DTO_OBJECT_CREATE, 0))
3882 * EA object consumes more credits than regular object: osd_mk_index
3883 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3884 * leaf_node, could involves the block, block bitmap, groups, GDT
3885 * change for each block, so add 4 * 2 credits in that case.
3887 * The default ACL initialization may consume an additional 16 blocks
3889 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3890 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3893 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3894 * credits for possible default ACL creation in new inode
3896 if (hint && hint->dah_acl_len)
3897 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3898 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3899 sb->s_blocksize_bits);
3901 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3904 * Reuse idle OI block may cause additional one OI block
3907 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3908 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3909 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3910 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3911 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3913 /* will help to find FID->ino mapping at dt_insert() */
3914 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3922 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3923 attr->la_projid, 1, oh, osd_dt_obj(dt),
3924 NULL, OSD_QID_INODE);
3932 * Called to destroy on-disk representation of the object
3934 * Concurrency: must be locked
3936 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3939 struct osd_object *obj = osd_dt_obj(dt);
3940 struct inode *inode = obj->oo_inode;
3941 struct osd_thandle *oh;
3949 oh = container_of(th, struct osd_thandle, ot_super);
3950 LASSERT(oh->ot_handle == NULL);
3952 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3953 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3955 /* For removing agent entry */
3956 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3957 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3960 * Recycle idle OI leaf may cause additional three OI blocks
3963 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3964 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3965 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3966 /* one less inode */
3967 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3968 i_projid_read(inode), -1, oh, obj, NULL,
3972 /* data to be truncated */
3973 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3974 i_projid_read(inode), 0, oh, obj, NULL,
3980 * will help to find FID->ino when this object is being
3983 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3988 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3991 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3992 struct osd_object *obj = osd_dt_obj(dt);
3993 struct inode *inode = obj->oo_inode;
3994 struct osd_device *osd = osd_obj2dev(obj);
3995 struct osd_thandle *oh;
4000 oh = container_of(th, struct osd_thandle, ot_super);
4001 LASSERT(oh->ot_handle);
4003 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
4005 if (unlikely(fid_is_acct(fid)))
4008 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
4009 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
4011 CERROR("%s: remove agent entry "DFID": rc = %d\n",
4012 osd_name(osd), PFID(fid), result);
4015 if (S_ISDIR(inode->i_mode)) {
4016 if (inode->i_nlink > 2)
4017 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
4018 osd_name(osd), PFID(fid), inode->i_ino,
4021 spin_lock(&obj->oo_guard);
4023 spin_unlock(&obj->oo_guard);
4024 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4027 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
4029 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
4031 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
4032 result = osd_oi_delete(osd_oti_get(env), osd, fid,
4033 oh->ot_handle, OI_CHECK_FLD);
4035 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
4036 /* XXX: add to ext3 orphan list */
4037 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
4039 /* not needed in the cache anymore */
4040 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
4041 obj->oo_destroyed = 1;
4047 * Put the fid into lustre_mdt_attrs, and then place the structure
4048 * inode's ea. This fid should not be altered during the life time
4051 * \retval +ve, on success
4052 * \retval -ve, on error
4054 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
4056 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
4057 const struct lu_fid *fid, __u32 compat, __u32 incompat)
4059 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4060 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4065 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INLMA))
4068 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
4071 lustre_loa_init(loa, fid, compat, incompat);
4072 lustre_loa_swab(loa, false);
4075 * For the OST device with 256 bytes inode size by default,
4076 * the PFID EA will be stored together with LMA EA to avoid
4077 * performance trouble. Otherwise the PFID EA can be stored
4078 * independently. LU-8998
4080 if ((compat & LMAC_FID_ON_OST) &&
4081 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
4082 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4083 sizeof(*loa), XATTR_CREATE);
4085 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4086 sizeof(*lma), XATTR_CREATE);
4088 * LMA may already exist, but we need to check that all the
4089 * desired compat/incompat flags have been added.
4091 if (unlikely(rc == -EEXIST)) {
4092 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
4093 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
4097 if (rc < sizeof(*lma))
4100 lustre_loa_swab(loa, true);
4101 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
4102 ((compat == 0 && incompat == 0) ||
4103 (!(~lma->lma_compat & compat) &&
4104 !(~lma->lma_incompat & incompat))))
4107 lma->lma_self_fid = *fid;
4108 lma->lma_compat |= compat;
4109 lma->lma_incompat |= incompat;
4110 if (rc == sizeof(*lma)) {
4111 lustre_lma_swab(lma);
4112 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4113 sizeof(*lma), XATTR_REPLACE);
4115 lustre_loa_swab(loa, false);
4116 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4117 sizeof(*loa), XATTR_REPLACE);
4125 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
4126 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
4127 * To have compatilibility with 1.8 ldiskfs driver we need to have
4128 * magic number at start of fid data.
4129 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
4132 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
4133 const struct lu_fid *fid)
4135 if (!fid_is_namespace_visible(fid) ||
4136 CFS_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
4137 param->edp_magic = 0;
4141 param->edp_magic = LDISKFS_LUFID_MAGIC;
4142 param->edp_len = sizeof(struct lu_fid) + 1;
4143 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
4147 * Try to read the fid from inode ea into dt_rec.
4149 * \param fid object fid.
4151 * \retval 0 on success
4153 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
4154 __u32 ino, struct lu_fid *fid,
4155 struct osd_inode_id *id)
4157 struct osd_thread_info *info = osd_oti_get(env);
4158 struct inode *inode;
4162 osd_id_gen(id, ino, OSD_OII_NOGEN);
4163 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
4165 RETURN(PTR_ERR(inode));
4171 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
4173 struct inode *parent_dir,
4174 const struct lu_fid *dot_fid,
4175 const struct lu_fid *dot_dot_fid,
4176 struct osd_thandle *oth)
4178 struct ldiskfs_dentry_param *dot_ldp;
4179 struct ldiskfs_dentry_param *dot_dot_ldp;
4180 __u32 saved_nlink = dir->i_nlink;
4183 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
4186 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
4187 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
4189 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4190 dot_ldp->edp_magic = 0;
4192 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
4193 dir, dot_ldp, dot_dot_ldp);
4195 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
4196 * the subseqent ref_add() will increase the dir->i_nlink
4197 * as 3. That is incorrect for new created directory.
4199 * It looks like hack, because we want to make the OSD API
4200 * to be order-independent for new created directory object
4201 * between dt_insert(..) and ref_add() operations.
4203 * Here, we only restore the in-RAM dir-inode's nlink attr,
4204 * becuase if the nlink attr is not 2, then there will be
4205 * ref_add() called following the dt_insert(..), such call
4206 * will make both the in-RAM and on-disk dir-inode's nlink
4207 * attr to be set as 2. LU-7447
4209 set_nlink(dir, saved_nlink);
4214 * Create an local agent inode for remote entry
4216 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
4217 struct osd_device *osd,
4218 struct osd_object *pobj,
4219 const struct lu_fid *fid,
4223 struct osd_thread_info *info = osd_oti_get(env);
4224 struct inode *local;
4225 struct osd_thandle *oh;
4226 struct iattr iattr = {
4227 .ia_valid = ATTR_UID | ATTR_GID |
4228 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
4229 .ia_ctime.tv_nsec = UTIME_OMIT,
4230 .ia_mtime.tv_nsec = UTIME_OMIT,
4231 .ia_atime.tv_nsec = UTIME_OMIT,
4232 .ia_uid = GLOBAL_ROOT_UID,
4233 .ia_gid = GLOBAL_ROOT_GID,
4240 oh = container_of(th, struct osd_thandle, ot_super);
4241 LASSERT(oh->ot_handle->h_transaction != NULL);
4243 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
4245 if (IS_ERR(local)) {
4246 CERROR("%s: create local error %d\n", osd_name(osd),
4247 (int)PTR_ERR(local));
4252 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
4253 * correct gid on remote file, not agent here
4255 local->i_gid = current_fsgid();
4256 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4258 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4259 * That gives e2fsck something to look at and be happy, and allows
4260 * debugging if we need to determine where this symlink came from.
4262 if (S_ISLNK(type)) {
4263 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4264 ldiskfs_clear_inode_flag(local, LDISKFS_INODE_EXTENTS);
4265 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4266 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4268 i_size_write(local, rc);
4269 LDISKFS_I(local)->i_disksize = rc;
4271 unlock_new_inode(local);
4273 /* Agent inode should not have project ID */
4274 #ifdef HAVE_PROJECT_QUOTA
4275 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4276 i_projid_read(pobj->oo_inode) != 0) {
4277 rc = osd_transfer_project(local, 0, th);
4279 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4280 osd_ino2name(local), rc);
4281 RETURN(ERR_PTR(rc));
4285 /* Set special LMA flag for local agent inode */
4286 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4288 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4289 osd_name(osd), PFID(fid), rc);
4290 RETURN(ERR_PTR(rc));
4296 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode, fid,
4297 lu_object_fid(&pobj->oo_dt.do_lu),
4300 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4301 osd_name(osd), PFID(fid), rc);
4302 RETURN(ERR_PTR(rc));
4309 * when direntry is deleted, we have to take care of possible agent inode
4310 * referenced by that. unfortunately we can't do this at that point:
4311 * iget() within a running transaction leads to deadlock and we better do
4312 * not call that every delete declaration to save performance. so we put
4313 * a potention agent inode on a list and process that once the transaction
4314 * is over. Notice it's not any worse in terms of real orphans as regular
4315 * object destroy doesn't put inodes on the on-disk orphan list. this should
4316 * be addressed separately
4318 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4319 struct osd_thandle *oh,
4322 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4323 struct osd_obj_orphan *oor;
4330 oor->oor_env = (struct lu_env *)env;
4331 spin_lock(&osd->od_osfs_lock);
4332 list_add(&oor->oor_list, &osd->od_orphan_list);
4333 spin_unlock(&osd->od_osfs_lock);
4335 oh->ot_remove_agents = 1;
4341 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4342 struct osd_device *osd)
4344 struct osd_thread_info *info = osd_oti_get(env);
4345 struct osd_obj_orphan *oor, *tmp;
4346 struct osd_inode_id id;
4348 struct inode *inode;
4353 spin_lock(&osd->od_osfs_lock);
4354 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4355 if (oor->oor_env == env)
4356 list_move(&oor->oor_list, &list);
4358 spin_unlock(&osd->od_osfs_lock);
4360 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4364 list_del(&oor->oor_list);
4367 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4368 inode = osd_iget_fid(info, osd, &id, &fid);
4372 if (!osd_remote_fid(env, osd, &fid)) {
4377 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4379 mark_inode_dirty(inode);
4380 ldiskfs_journal_stop(jh);
4388 * OSD layer object create function for OST objects (b=11826).
4390 * The FID is inserted into inode xattr here.
4392 * \retval 0, on success
4393 * \retval -ve, on error
4395 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4396 struct lu_attr *attr, struct dt_allocation_hint *hint,
4397 struct dt_object_format *dof, struct thandle *th)
4399 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4400 struct osd_object *obj = osd_dt_obj(dt);
4401 struct osd_thread_info *info = osd_oti_get(env);
4402 int result, on_ost = 0;
4406 if (dt_object_exists(dt))
4409 LINVRNT(osd_invariant(obj));
4410 LASSERT(!dt_object_remote(dt));
4411 LASSERT(osd_is_write_locked(env, obj));
4412 LASSERT(th != NULL);
4414 if (unlikely(fid_is_acct(fid)))
4416 * Quota files can't be created from the kernel any more,
4417 * 'tune2fs -O quota' will take care of creating them
4421 result = __osd_create(info, obj, attr, hint, dof, th);
4423 if (fid_is_idif(fid) &&
4424 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4425 struct lu_fid *tfid = &info->oti_fid;
4426 struct ost_id *oi = &info->oti_ostid;
4428 fid_to_ostid(fid, oi);
4429 ostid_to_fid(tfid, oi, 0);
4431 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4432 LMAC_FID_ON_OST, 0);
4434 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4436 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4437 on_ost ? LMAC_FID_ON_OST : 0,
4440 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4441 obj->oo_dt.do_body_ops = &osd_body_ops;
4444 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4445 struct inode *inode = obj->oo_inode;
4447 result = __osd_oi_insert(env, obj, fid, th);
4448 if (result && inode) {
4449 spin_lock(&obj->oo_guard);
4451 spin_unlock(&obj->oo_guard);
4452 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4453 ldiskfs_set_inode_state(inode,
4454 LDISKFS_STATE_LUSTRE_DESTROY);
4456 obj->oo_inode = NULL;
4461 * a small optimization - dt_insert() isn't usually applied
4462 * to OST objects, so we don't need to cache OI mapping for
4465 if (result == 0 && on_ost == 0) {
4466 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4468 result = osd_idc_find_and_init(env, osd, obj);
4469 LASSERT(result == 0);
4472 LASSERT(ergo(result == 0,
4473 dt_object_exists(dt) && !dt_object_remote(dt)));
4474 LINVRNT(osd_invariant(obj));
4478 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4479 struct thandle *handle)
4481 struct osd_thandle *oh;
4484 /* it's possible that object doesn't exist yet */
4485 LASSERT(handle != NULL);
4487 oh = container_of(handle, struct osd_thandle, ot_super);
4488 LASSERT(oh->ot_handle == NULL);
4490 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4491 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4493 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4500 * Concurrency: @dt is write locked.
4502 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4505 struct osd_object *obj = osd_dt_obj(dt);
4506 struct inode *inode = obj->oo_inode;
4507 struct osd_thandle *oh;
4510 if (!dt_object_exists(dt) || obj->oo_destroyed)
4513 LINVRNT(osd_invariant(obj));
4514 LASSERT(!dt_object_remote(dt));
4515 LASSERT(osd_is_write_locked(env, obj));
4516 LASSERT(th != NULL);
4518 oh = container_of(th, struct osd_thandle, ot_super);
4519 LASSERT(oh->ot_handle != NULL);
4521 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4523 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4524 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4526 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4527 * (65000) subdirectories by storing "1" in i_nlink if the link count
4528 * would otherwise overflow. Directory tranversal tools understand
4529 * that (st_nlink == 1) indicates that the filesystem dose not track
4530 * hard links count on the directory, and will not abort subdirectory
4531 * scanning early once (st_nlink - 2) subdirs have been found.
4533 * This also has to properly handle the case of inodes with nlink == 0
4534 * in case they are being linked into the PENDING directory
4536 spin_lock(&obj->oo_guard);
4537 if (unlikely(inode->i_nlink == 0))
4538 /* inc_nlink from 0 may cause WARN_ON */
4539 set_nlink(inode, 1);
4541 osd_ldiskfs_inc_count(oh->ot_handle, inode);
4542 if (!S_ISDIR(inode->i_mode))
4543 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4545 spin_unlock(&obj->oo_guard);
4547 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4548 LINVRNT(osd_invariant(obj));
4550 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4555 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4556 struct thandle *handle)
4558 struct osd_thandle *oh;
4560 if (!dt_object_exists(dt))
4563 LASSERT(!dt_object_remote(dt));
4564 LASSERT(handle != NULL);
4566 oh = container_of(handle, struct osd_thandle, ot_super);
4567 LASSERT(oh->ot_handle == NULL);
4569 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4570 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4576 * Concurrency: @dt is write locked.
4578 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4581 struct osd_object *obj = osd_dt_obj(dt);
4582 struct inode *inode = obj->oo_inode;
4583 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4584 struct osd_thandle *oh;
4586 if (!dt_object_exists(dt))
4589 LINVRNT(osd_invariant(obj));
4590 LASSERT(!dt_object_remote(dt));
4591 LASSERT(osd_is_write_locked(env, obj));
4592 LASSERT(th != NULL);
4594 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4597 oh = container_of(th, struct osd_thandle, ot_super);
4598 LASSERT(oh->ot_handle != NULL);
4600 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4602 spin_lock(&obj->oo_guard);
4604 * That can be result of upgrade from old Lustre version and
4605 * applied only to local files. Just skip this ref_del call.
4606 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4608 if (inode->i_nlink == 0) {
4609 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4610 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4611 ", maybe an upgraded file? (LU-3915)\n",
4612 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4613 spin_unlock(&obj->oo_guard);
4617 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4618 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4620 osd_ldiskfs_dec_count(oh->ot_handle, inode);
4621 spin_unlock(&obj->oo_guard);
4623 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4624 LINVRNT(osd_invariant(obj));
4626 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4632 * Concurrency: @dt is read locked.
4634 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4635 struct lu_buf *buf, const char *name)
4637 struct osd_object *obj = osd_dt_obj(dt);
4638 struct inode *inode = obj->oo_inode;
4639 struct osd_thread_info *info = osd_oti_get(env);
4640 struct dentry *dentry = &info->oti_obj_dentry;
4641 bool cache_xattr = false;
4646 /* version get is not real XATTR but uses xattr API */
4647 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4648 dt_obj_version_t *ver = buf->lb_buf;
4651 * for version we are just using xattr API but change inode
4654 if (buf->lb_len == 0)
4655 return sizeof(dt_obj_version_t);
4657 if (buf->lb_len < sizeof(dt_obj_version_t))
4660 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4661 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4663 *ver = LDISKFS_I(inode)->i_fs_version;
4665 return sizeof(dt_obj_version_t);
4668 if (!dt_object_exists(dt))
4671 LASSERT(!dt_object_remote(dt));
4672 LASSERT(inode->i_op != NULL);
4673 #ifdef HAVE_IOP_XATTR
4674 LASSERT(inode->i_op->getxattr != NULL);
4677 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4678 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4682 rc = osd_oxc_get(obj, name, buf);
4687 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4688 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4689 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4690 struct filter_fid *ff;
4691 struct ost_layout *ol;
4693 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4697 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4700 if (buf->lb_len == 0 || !buf->lb_buf)
4703 if (buf->lb_len < rc)
4707 ol = &ff->ff_layout;
4708 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4709 PFID_STRIPE_IDX_BITS);
4710 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4711 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4712 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4713 if (lma->lma_compat & LMAC_COMP_INFO) {
4714 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4715 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4716 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4718 ol->ol_comp_start = 0;
4719 ol->ol_comp_end = 0;
4723 /* Get enc context xattr directly from ldiskfs instead of going
4724 * through the VFS, as there is no xattr handler for
4727 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0)
4728 rc = ldiskfs_xattr_get(inode,
4729 LDISKFS_XATTR_INDEX_ENCRYPTION,
4730 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
4731 buf->lb_buf, buf->lb_len);
4733 rc = __osd_xattr_get(inode, dentry, name,
4734 buf->lb_buf, buf->lb_len);
4738 if (rc == -ENOENT || rc == -ENODATA)
4739 osd_oxc_add(obj, name, NULL, 0);
4740 else if (rc > 0 && buf->lb_buf != NULL)
4741 osd_oxc_add(obj, name, buf->lb_buf, rc);
4747 static int osd_declare_xattr_set(const struct lu_env *env,
4748 struct dt_object *dt,
4749 const struct lu_buf *buf, const char *name,
4750 int fl, struct thandle *handle)
4752 struct osd_thandle *oh;
4754 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4756 LASSERT(handle != NULL);
4758 oh = container_of(handle, struct osd_thandle, ot_super);
4759 LASSERT(oh->ot_handle == NULL);
4761 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4763 * For non-upgrading case, the LMA is set first and
4764 * usually fit inode. But for upgrade case, the LMA
4765 * may be in another separated EA block.
4767 if (dt_object_exists(dt)) {
4768 if (fl == LU_XATTR_REPLACE)
4773 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4775 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4776 /* We may need to delete the old PFID EA. */
4777 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4778 if (fl == LU_XATTR_REPLACE)
4784 * If some name entry resides on remote MDT, then will create
4785 * agent entry under remote parent. On the other hand, if the
4786 * remote entry will be removed, then related agent entry may
4787 * need to be removed from the remote parent. So there may be
4788 * kinds of cases, let's declare enough credits. The credits
4789 * for create agent entry is enough for remove case.
4791 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4792 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4793 if (dt_object_exists(dt))
4794 credits += 1; /* For updating LMA */
4798 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4803 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4805 * learn xattr size from osd_xattr_get if
4806 * attribute has not been read yet
4808 buflen = __osd_xattr_get(
4809 osd_dt_obj(dt)->oo_inode,
4810 &osd_oti_get(env)->oti_obj_dentry,
4815 buflen = buf->lb_len;
4818 if (buflen > sb->s_blocksize) {
4819 credits += osd_calc_bkmap_credits(
4821 (buflen + sb->s_blocksize - 1) >>
4822 sb->s_blocksize_bits);
4826 * xattr set may involve inode quota change, reserve credits for
4827 * dquot_initialize()
4829 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4832 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4837 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4838 const struct lu_buf *buf, int fl,
4839 struct thandle *handle)
4841 struct osd_thread_info *info = osd_oti_get(env);
4842 struct dentry *dentry = &info->oti_obj_dentry;
4843 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4844 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4845 struct inode *inode = obj->oo_inode;
4846 struct filter_fid *ff = buf->lb_buf;
4847 struct ost_layout *ol = &ff->ff_layout;
4848 int flags = XATTR_REPLACE;
4853 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4856 rc = osd_get_lma(info, inode, dentry, loa);
4857 if (rc == -ENODATA) {
4858 /* Usually for upgarding from old device */
4859 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4860 LMAC_FID_ON_OST, 0);
4861 flags = XATTR_CREATE;
4866 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4867 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4870 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4871 /* Separate PFID EA from LMA */
4872 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4873 lustre_lma_swab(lma);
4874 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4875 sizeof(*lma), XATTR_REPLACE);
4877 obj->oo_pfid_in_lma = 0;
4878 rc = LU_XATTR_CREATE;
4884 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4888 * Old client does not send stripe information,
4889 * then store the PFID EA on disk separatedly.
4891 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4892 ol->ol_stripe_size == 0))
4895 /* Remove old PFID EA entry firstly. */
4896 dquot_initialize(inode);
4897 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4898 if (rc == -ENODATA) {
4899 /* XATTR_NAME_FID is already absent */
4906 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4907 if (likely(ol->ol_stripe_size != 0)) {
4908 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4909 PFID_STRIPE_IDX_BITS;
4910 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4911 lma->lma_compat |= LMAC_STRIPE_INFO;
4912 if (ol->ol_comp_id != 0) {
4913 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4914 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4915 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4916 lma->lma_compat |= LMAC_COMP_INFO;
4920 lustre_loa_swab(loa, false);
4922 /* Store the PFID EA inside LMA. */
4923 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4926 obj->oo_pfid_in_lma = 1;
4932 * In DNE environment, the object (in spite of regular file or directory)
4933 * and its name entry may reside on different MDTs. Under such case, we will
4934 * create an agent entry on the MDT where the object resides. The agent entry
4935 * references the object locally, that makes the object to be visible to the
4936 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4937 * such as 'tar' can handle the object properly.
4939 * We handle the agent entry during set linkEA that is the common interface
4940 * for both regular file and directroy, can handle kinds of cases, such as
4941 * create/link/unlink/rename, and so on.
4943 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4945 * XXX: There are two known issues:
4946 * 1. For one object, we will create at most one agent entry even if there
4947 * may be more than one cross-MDTs hard links on the object. So the local
4948 * e2fsck may claim that the object's nlink is larger than the name entries
4949 * that reference such inode. And in further, the e2fsck will fix the nlink
4950 * attribute to match the local references. Then it will cause the object's
4951 * nlink attribute to be inconsistent with the global references. it is bad
4952 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4953 * other hand, the global namespace LFSCK can repair the object's attribute
4954 * according to the linkEA.
4955 * 2. There may be too many hard links on the object as to its linkEA overflow,
4956 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4957 * case happened, then at this point, we do not know whether there are some
4958 * cross-MDTs reference. But there are local references, it guarantees that
4959 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4961 static int osd_xattr_handle_linkea(const struct lu_env *env,
4962 struct osd_device *osd,
4963 struct osd_object *obj,
4964 const struct lu_buf *buf,
4965 struct thandle *handle)
4967 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4968 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4969 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4970 struct lu_name tmpname;
4971 struct osd_thandle *oh;
4973 bool remote = false;
4977 oh = container_of(handle, struct osd_thandle, ot_super);
4978 LASSERT(oh->ot_handle != NULL);
4980 rc = linkea_init_with_rec(&ldata);
4982 linkea_first_entry(&ldata);
4983 while (ldata.ld_lee != NULL && !remote) {
4984 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4986 if (osd_remote_fid(env, osd, tfid) > 0)
4989 linkea_next_entry(&ldata);
4991 } else if (rc == -ENODATA) {
4997 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4998 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
5000 CERROR("%s: failed to remove agent entry for "DFID
5001 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5002 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
5003 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5005 CERROR("%s: failed to create agent entry for "DFID
5006 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5013 * Concurrency: @dt is write locked.
5015 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
5016 const struct lu_buf *buf, const char *name, int fl,
5017 struct thandle *handle)
5019 struct osd_object *obj = osd_dt_obj(dt);
5020 struct osd_device *osd = osd_obj2dev(obj);
5021 struct inode *inode = obj->oo_inode;
5022 struct osd_thread_info *info = osd_oti_get(env);
5032 /* version set is not real XATTR */
5033 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
5034 dt_obj_version_t *version = buf->lb_buf;
5037 * for version we are just using xattr API but change inode
5040 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
5043 DFID" set version %#llx (old %#llx) for inode %lu\n",
5044 PFID(lu_object_fid(&dt->do_lu)), *version,
5045 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
5047 LDISKFS_I(inode)->i_fs_version = *version;
5049 * Version is set after all inode operations are finished,
5050 * so we should mark it dirty here
5052 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
5057 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
5058 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
5061 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5064 * For the OST device with 256 bytes inode size by default,
5065 * the PFID EA will be stored together with LMA EA to avoid
5066 * performance trouble. Otherwise the PFID EA can be stored
5067 * independently. LU-8998
5069 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
5070 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
5071 LASSERT(buf->lb_buf);
5073 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
5076 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
5077 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
5078 struct lustre_mdt_attrs *lma = &loa->loa_lma;
5080 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
5084 lma->lma_incompat |= LMAI_STRIPED;
5085 lustre_lma_swab(lma);
5086 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5087 sizeof(*lma), XATTR_REPLACE);
5090 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
5091 LASSERT(!osd->od_is_ost);
5093 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
5098 if (fl & LU_XATTR_REPLACE)
5099 fs_flags |= XATTR_REPLACE;
5101 if (fl & LU_XATTR_CREATE)
5102 fs_flags |= XATTR_CREATE;
5104 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0) {
5105 /* Set enc context xattr directly in ldiskfs instead of going
5106 * through the VFS, as there is no xattr handler for
5109 struct osd_thandle *oth = container_of(handle,
5113 if (!oth->ot_handle)
5114 /* this should be already part of a transaction */
5117 rc = ldiskfs_xattr_set_handle(oth->ot_handle, inode,
5118 LDISKFS_XATTR_INDEX_ENCRYPTION,
5119 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
5120 buf->lb_buf, len, fs_flags);
5122 rc = __osd_xattr_set(info, inode, name,
5123 buf->lb_buf, len, fs_flags);
5125 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5128 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5129 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5130 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
5136 * Concurrency: @dt is read locked.
5138 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
5139 const struct lu_buf *buf)
5141 struct osd_object *obj = osd_dt_obj(dt);
5142 struct osd_device *dev = osd_obj2dev(obj);
5143 struct inode *inode = obj->oo_inode;
5144 struct osd_thread_info *info = osd_oti_get(env);
5145 struct dentry *dentry = &info->oti_obj_dentry;
5148 if (!dt_object_exists(dt))
5151 LASSERT(!dt_object_remote(dt));
5152 LASSERT(inode->i_op != NULL);
5153 LASSERT(inode->i_op->listxattr != NULL);
5155 dentry->d_inode = inode;
5156 dentry->d_sb = inode->i_sb;
5157 rc = inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
5159 if (rc < 0 || buf->lb_buf == NULL)
5162 /* Hide virtual project ID xattr from list if disabled */
5163 if (!dev->od_enable_projid_xattr) {
5164 char *end = (char *)buf->lb_buf + rc;
5165 char *p = buf->lb_buf;
5168 char *next = p + strlen(p) + 1;
5170 if (strcmp(p, XATTR_NAME_PROJID) == 0) {
5172 memmove(p, next, end - next);
5184 static int osd_declare_xattr_del(const struct lu_env *env,
5185 struct dt_object *dt, const char *name,
5186 struct thandle *handle)
5188 struct osd_thandle *oh;
5189 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
5191 LASSERT(!dt_object_remote(dt));
5192 LASSERT(handle != NULL);
5194 oh = container_of(handle, struct osd_thandle, ot_super);
5195 LASSERT(oh->ot_handle == NULL);
5197 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
5198 osd_dto_credits_noquota[DTO_XATTR_SET]);
5200 * xattr del may involve inode quota change, reserve credits for
5201 * dquot_initialize()
5203 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
5209 * Concurrency: @dt is write locked.
5211 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
5212 const char *name, struct thandle *handle)
5214 struct osd_object *obj = osd_dt_obj(dt);
5215 struct inode *inode = obj->oo_inode;
5216 struct osd_thread_info *info = osd_oti_get(env);
5217 struct dentry *dentry = &info->oti_obj_dentry;
5220 if (!dt_object_exists(dt))
5223 LASSERT(!dt_object_remote(dt));
5224 LASSERT(inode->i_op != NULL);
5225 LASSERT(handle != NULL);
5226 #ifdef HAVE_IOP_XATTR
5227 LASSERT(inode->i_op->removexattr != NULL);
5230 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5232 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
5233 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
5235 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
5236 &info->oti_ost_attrs);
5238 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
5240 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
5241 lustre_lma_swab(lma);
5242 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5243 sizeof(*lma), XATTR_REPLACE);
5245 obj->oo_pfid_in_lma = 0;
5248 dquot_initialize(inode);
5249 dentry->d_inode = inode;
5250 dentry->d_sb = inode->i_sb;
5251 rc = ll_vfs_removexattr(dentry, inode, name);
5254 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5257 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5258 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5259 osd_oxc_del(obj, name);
5264 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
5265 __u64 start, __u64 end)
5267 struct osd_object *obj = osd_dt_obj(dt);
5268 struct osd_device *dev = osd_obj2dev(obj);
5269 struct inode *inode = obj->oo_inode;
5274 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
5277 RETURN(PTR_ERR(file));
5279 file->f_mode |= FMODE_64BITHASH;
5280 rc = vfs_fsync_range(file, start, end, 0);
5287 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
5292 static bool osd_check_stale(struct dt_object *dt)
5300 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
5301 const struct dt_index_features *feat)
5303 struct iam_descr *descr;
5305 if (osd_object_is_root(o))
5306 return feat == &dt_directory_features;
5308 LASSERT(o->oo_dir != NULL);
5310 descr = o->oo_dir->od_container.ic_descr;
5311 if (feat == &dt_directory_features) {
5312 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
5317 return feat->dif_keysize_min <= descr->id_key_size &&
5318 descr->id_key_size <= feat->dif_keysize_max &&
5319 feat->dif_recsize_min <= descr->id_rec_size &&
5320 descr->id_rec_size <= feat->dif_recsize_max &&
5321 !(feat->dif_flags & (DT_IND_VARKEY |
5322 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5323 ergo(feat->dif_flags & DT_IND_UPDATE,
5324 1 /* XXX check that object (and fs) is writable */);
5328 static int osd_iam_container_init(const struct lu_env *env,
5329 struct osd_object *obj,
5330 struct osd_directory *dir)
5332 struct iam_container *bag = &dir->od_container;
5335 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5339 result = iam_container_setup(bag);
5341 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5343 iam_container_fini(bag);
5350 * Concurrency: no external locking is necessary.
5352 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5353 const struct dt_index_features *feat)
5357 struct osd_object *obj = osd_dt_obj(dt);
5359 LINVRNT(osd_invariant(obj));
5361 if (osd_object_is_root(obj)) {
5362 dt->do_index_ops = &osd_index_ea_ops;
5364 } else if (feat == &dt_directory_features) {
5365 dt->do_index_ops = &osd_index_ea_ops;
5366 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5371 } else if (unlikely(feat == &dt_otable_features)) {
5372 dt->do_index_ops = &osd_otable_ops;
5374 } else if (unlikely(feat == &dt_acct_features)) {
5375 dt->do_index_ops = &osd_acct_index_ops;
5378 } else if (!osd_has_index(obj)) {
5379 struct osd_directory *dir;
5380 struct osd_device *osd = osd_obj2dev(obj);
5381 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5386 spin_lock(&obj->oo_guard);
5387 if (obj->oo_dir == NULL)
5391 * Concurrent thread allocated container data.
5394 spin_unlock(&obj->oo_guard);
5396 * Now, that we have container data, serialize its
5399 down_write(&obj->oo_ext_idx_sem);
5401 * recheck under lock.
5404 if (osd_has_index(obj)) {
5409 result = osd_iam_container_init(env, obj, obj->oo_dir);
5410 if (result || feat == &dt_lfsck_namespace_features ||
5411 feat == &dt_lfsck_layout_orphan_features ||
5412 feat == &dt_lfsck_layout_dangling_features)
5415 result = osd_index_register(osd, fid,
5416 feat->dif_keysize_max,
5417 feat->dif_recsize_max);
5419 CWARN("%s: failed to register index "
5421 osd_name(osd), PFID(fid), result);
5422 else if (result > 0)
5425 CDEBUG(D_LFSCK, "%s: index object "DFID
5426 " (%d/%d) registered\n",
5427 osd_name(osd), PFID(fid),
5428 (int)feat->dif_keysize_max,
5429 (int)feat->dif_recsize_max);
5432 up_write(&obj->oo_ext_idx_sem);
5440 if (result == 0 && skip_iam == 0) {
5441 if (!osd_iam_index_probe(env, obj, feat))
5444 LINVRNT(osd_invariant(obj));
5449 static int osd_otable_it_attr_get(const struct lu_env *env,
5450 struct dt_object *dt,
5451 struct lu_attr *attr)
5457 static const struct dt_object_operations osd_obj_ops = {
5458 .do_read_lock = osd_read_lock,
5459 .do_write_lock = osd_write_lock,
5460 .do_read_unlock = osd_read_unlock,
5461 .do_write_unlock = osd_write_unlock,
5462 .do_write_locked = osd_write_locked,
5463 .do_attr_get = osd_attr_get,
5464 .do_declare_attr_set = osd_declare_attr_set,
5465 .do_attr_set = osd_attr_set,
5466 .do_ah_init = osd_ah_init,
5467 .do_declare_create = osd_declare_create,
5468 .do_create = osd_create,
5469 .do_declare_destroy = osd_declare_destroy,
5470 .do_destroy = osd_destroy,
5471 .do_index_try = osd_index_try,
5472 .do_declare_ref_add = osd_declare_ref_add,
5473 .do_ref_add = osd_ref_add,
5474 .do_declare_ref_del = osd_declare_ref_del,
5475 .do_ref_del = osd_ref_del,
5476 .do_xattr_get = osd_xattr_get,
5477 .do_declare_xattr_set = osd_declare_xattr_set,
5478 .do_xattr_set = osd_xattr_set,
5479 .do_declare_xattr_del = osd_declare_xattr_del,
5480 .do_xattr_del = osd_xattr_del,
5481 .do_xattr_list = osd_xattr_list,
5482 .do_object_sync = osd_object_sync,
5483 .do_invalidate = osd_invalidate,
5484 .do_check_stale = osd_check_stale,
5487 static const struct dt_object_operations osd_obj_otable_it_ops = {
5488 .do_attr_get = osd_otable_it_attr_get,
5489 .do_index_try = osd_index_try,
5492 static int osd_index_declare_iam_delete(const struct lu_env *env,
5493 struct dt_object *dt,
5494 const struct dt_key *key,
5495 struct thandle *handle)
5497 struct osd_thandle *oh;
5499 oh = container_of(handle, struct osd_thandle, ot_super);
5500 LASSERT(oh->ot_handle == NULL);
5502 /* Recycle may cause additional three blocks to be changed. */
5503 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5504 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5510 * delete a (key, value) pair from index \a dt specified by \a key
5512 * \param dt osd index object
5513 * \param key key for index
5514 * \param rec record reference
5515 * \param handle transaction handler
5518 * \retval -ve failure
5520 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5521 const struct dt_key *key,
5522 struct thandle *handle)
5524 struct osd_thread_info *oti = osd_oti_get(env);
5525 struct osd_object *obj = osd_dt_obj(dt);
5526 struct osd_thandle *oh;
5527 struct iam_path_descr *ipd;
5528 struct iam_container *bag = &obj->oo_dir->od_container;
5533 if (!dt_object_exists(dt))
5536 LINVRNT(osd_invariant(obj));
5537 LASSERT(!dt_object_remote(dt));
5538 LASSERT(bag->ic_object == obj->oo_inode);
5539 LASSERT(handle != NULL);
5541 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5543 ipd = osd_idx_ipd_get(env, bag);
5544 if (unlikely(ipd == NULL))
5547 oh = container_of(handle, struct osd_thandle, ot_super);
5548 LASSERT(oh->ot_handle != NULL);
5549 LASSERT(oh->ot_handle->h_transaction != NULL);
5551 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5552 /* swab quota uid/gid provided by caller */
5553 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5554 key = (const struct dt_key *)&oti->oti_quota_id;
5557 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5558 osd_ipd_put(env, bag, ipd);
5559 LINVRNT(osd_invariant(obj));
5560 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5564 static int osd_index_declare_ea_delete(const struct lu_env *env,
5565 struct dt_object *dt,
5566 const struct dt_key *key,
5567 struct thandle *handle)
5569 struct osd_thandle *oh;
5570 struct inode *inode;
5575 LASSERT(!dt_object_remote(dt));
5576 LASSERT(handle != NULL);
5578 oh = container_of(handle, struct osd_thandle, ot_super);
5579 LASSERT(oh->ot_handle == NULL);
5581 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5582 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5584 inode = osd_dt_obj(dt)->oo_inode;
5588 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5589 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5594 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5597 struct osd_fid_pack *rec;
5600 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5601 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5602 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5603 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5609 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5610 const struct lu_fid *fid)
5612 struct seq_server_site *ss = osd_seq_site(osd);
5616 /* FID seqs not in FLDB, must be local seq */
5617 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5621 * If FLD is not being initialized yet, it only happens during the
5622 * initialization, likely during mgs initialization, and we assume
5623 * this is local FID.
5625 if (ss == NULL || ss->ss_server_fld == NULL)
5628 /* Only check the local FLDB here */
5629 if (osd_seq_exists(env, osd, fid_seq(fid)))
5635 static void osd_take_care_of_agent(const struct lu_env *env,
5636 struct osd_device *osd,
5637 struct osd_thandle *oh,
5638 struct ldiskfs_dir_entry_2 *de)
5640 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5641 struct osd_idmap_cache *idc;
5642 int rc, schedule = 0;
5644 LASSERT(de != NULL);
5646 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5647 if (likely(rc == 0)) {
5648 idc = osd_idc_find_or_init(env, osd, fid);
5649 if (IS_ERR(idc) || idc->oic_remote)
5651 } else if (rc == -ENODATA) {
5653 * can't get FID, postpone to the end of the
5654 * transaction when iget() is safe
5658 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5661 osd_schedule_agent_inode_removal(env, oh,
5662 le32_to_cpu(de->inode));
5666 * Utility function to get real name from object name
5668 * \param[in] obj pointer to the object to be handled
5669 * \param[in] name object name
5670 * \param[in] len object name len
5671 * \param[out]ln pointer to the struct lu_name to hold the real name
5673 * If file is not encrypted, real name is just the object name.
5674 * If file is encrypted, object name needs to be decoded. In
5675 * this case a new buffer is allocated, and ln->ln_name needs to be freed by
5678 * \retval 0, on success
5679 * \retval -ve, on error
5681 static int obj_name2lu_name(struct osd_object *obj, const char *name,
5682 int len, struct lu_name *ln)
5684 if (!(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
5686 ln->ln_namelen = len;
5688 char *buf = kmalloc(len, GFP_NOFS);
5693 len = critical_decode(name, len, buf);
5695 ln->ln_namelen = len;
5702 * Index delete function for interoperability mode (b11826).
5703 * It will remove the directory entry added by osd_index_ea_insert().
5704 * This entry is needed to maintain name->fid mapping.
5706 * \param key, key i.e. file entry to be deleted
5708 * \retval 0, on success
5709 * \retval -ve, on error
5711 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5712 const struct dt_key *key, struct thandle *handle)
5714 struct osd_object *obj = osd_dt_obj(dt);
5715 struct inode *dir = obj->oo_inode;
5716 struct dentry *dentry;
5717 struct osd_thandle *oh;
5718 struct ldiskfs_dir_entry_2 *de = NULL;
5719 struct buffer_head *bh;
5720 struct htree_lock *hlock = NULL;
5721 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5727 if (!dt_object_exists(dt))
5730 LINVRNT(osd_invariant(obj));
5731 LASSERT(!dt_object_remote(dt));
5732 LASSERT(handle != NULL);
5734 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
5738 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5740 oh = container_of(handle, struct osd_thandle, ot_super);
5741 LASSERT(oh->ot_handle != NULL);
5742 LASSERT(oh->ot_handle->h_transaction != NULL);
5744 dquot_initialize(dir);
5745 dentry = osd_child_dentry_get(env, obj, ln.ln_name, ln.ln_namelen);
5747 if (obj->oo_hl_head != NULL) {
5748 hlock = osd_oti_get(env)->oti_hlock;
5749 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5750 dir, LDISKFS_HLOCK_DEL);
5752 down_write(&obj->oo_ext_idx_sem);
5755 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5758 * If this is not the ".." entry, it might be a remote DNE
5759 * entry and we need to check if the FID is for a remote
5760 * MDT. If the FID is not in the directory entry (e.g.
5761 * upgraded 1.8 filesystem without dirdata enabled) then
5762 * we need to get the FID from the LMA. For a remote directory
5763 * there HAS to be an LMA, it cannot be an IGIF inode in this
5766 * Delete the entry before the agent inode in order to
5767 * simplify error handling. At worst an error after deleting
5768 * the entry first might leak the agent inode afterward. The
5769 * reverse would need filesystem abort in case of error deleting
5770 * the entry after the agent had been removed, or leave a
5771 * dangling entry pointing at a random inode.
5773 if (strcmp((char *)key, dotdot) != 0)
5774 osd_take_care_of_agent(env, osd, oh, de);
5775 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5781 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
5782 atomic_dec_if_positive(&obj->oo_dirent_count);
5784 ldiskfs_htree_unlock(hlock);
5786 up_write(&obj->oo_ext_idx_sem);
5789 LASSERT(osd_invariant(obj));
5790 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5791 if (ln.ln_name != (char *)key)
5797 * Lookup index for \a key and copy record to \a rec.
5799 * \param dt osd index object
5800 * \param key key for index
5801 * \param rec record reference
5803 * \retval +ve success : exact mach
5804 * \retval 0 return record with key not greater than \a key
5805 * \retval -ve failure
5807 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5808 struct dt_rec *rec, const struct dt_key *key)
5810 struct osd_object *obj = osd_dt_obj(dt);
5811 struct iam_path_descr *ipd;
5812 struct iam_container *bag = &obj->oo_dir->od_container;
5813 struct osd_thread_info *oti = osd_oti_get(env);
5814 struct iam_iterator *it = &oti->oti_idx_it;
5815 struct iam_rec *iam_rec;
5820 if (!dt_object_exists(dt))
5823 LASSERT(osd_invariant(obj));
5824 LASSERT(!dt_object_remote(dt));
5825 LASSERT(bag->ic_object == obj->oo_inode);
5827 ipd = osd_idx_ipd_get(env, bag);
5831 /* got ipd now we can start iterator. */
5832 iam_it_init(it, bag, 0, ipd);
5834 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5835 /* swab quota uid/gid provided by caller */
5836 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5837 key = (const struct dt_key *)&oti->oti_quota_id;
5840 rc = iam_it_get(it, (struct iam_key *)key);
5842 if (S_ISDIR(obj->oo_inode->i_mode))
5843 iam_rec = (struct iam_rec *)oti->oti_ldp;
5845 iam_rec = (struct iam_rec *)rec;
5847 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5849 if (S_ISDIR(obj->oo_inode->i_mode))
5850 osd_fid_unpack((struct lu_fid *)rec,
5851 (struct osd_fid_pack *)iam_rec);
5852 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5853 osd_quota_unpack(obj, rec);
5858 osd_ipd_put(env, bag, ipd);
5860 LINVRNT(osd_invariant(obj));
5865 static int osd_index_declare_iam_insert(const struct lu_env *env,
5866 struct dt_object *dt,
5867 const struct dt_rec *rec,
5868 const struct dt_key *key,
5869 struct thandle *handle)
5871 struct osd_thandle *oh;
5873 LASSERT(handle != NULL);
5875 oh = container_of(handle, struct osd_thandle, ot_super);
5876 LASSERT(oh->ot_handle == NULL);
5878 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5879 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5885 * Inserts (key, value) pair in \a dt index object.
5887 * \param dt osd index object
5888 * \param key key for index
5889 * \param rec record reference
5890 * \param th transaction handler
5893 * \retval -ve failure
5895 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5896 const struct dt_rec *rec,
5897 const struct dt_key *key, struct thandle *th)
5899 struct osd_object *obj = osd_dt_obj(dt);
5900 struct iam_path_descr *ipd;
5901 struct osd_thandle *oh;
5902 struct iam_container *bag;
5903 struct osd_thread_info *oti = osd_oti_get(env);
5904 struct iam_rec *iam_rec;
5909 if (!dt_object_exists(dt))
5912 LINVRNT(osd_invariant(obj));
5913 LASSERT(!dt_object_remote(dt));
5915 bag = &obj->oo_dir->od_container;
5916 LASSERT(bag->ic_object == obj->oo_inode);
5917 LASSERT(th != NULL);
5919 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5921 ipd = osd_idx_ipd_get(env, bag);
5922 if (unlikely(ipd == NULL))
5925 oh = container_of(th, struct osd_thandle, ot_super);
5926 LASSERT(oh->ot_handle != NULL);
5927 LASSERT(oh->ot_handle->h_transaction != NULL);
5928 if (S_ISDIR(obj->oo_inode->i_mode)) {
5929 iam_rec = (struct iam_rec *)oti->oti_ldp;
5930 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5932 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5933 /* pack quota uid/gid */
5934 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5935 key = (const struct dt_key *)&oti->oti_quota_id;
5936 /* pack quota record */
5937 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5938 iam_rec = (struct iam_rec *)rec;
5940 iam_rec = (struct iam_rec *)rec;
5943 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5945 osd_ipd_put(env, bag, ipd);
5946 LINVRNT(osd_invariant(obj));
5947 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5952 * Calls ldiskfs_add_entry() to add directory entry
5953 * into the directory. This is required for
5954 * interoperability mode (b11826)
5956 * \retval 0, on success
5957 * \retval -ve, on error
5959 static int __osd_ea_add_rec(struct osd_thread_info *info,
5960 struct osd_object *pobj, struct inode *cinode,
5961 const char *name, const struct lu_fid *fid,
5962 struct htree_lock *hlock, struct thandle *th)
5964 struct ldiskfs_dentry_param *ldp;
5965 struct dentry *child;
5966 struct osd_thandle *oth;
5970 oth = container_of(th, struct osd_thandle, ot_super);
5971 LASSERT(oth->ot_handle != NULL);
5972 LASSERT(oth->ot_handle->h_transaction != NULL);
5973 LASSERT(pobj->oo_inode);
5975 rc = obj_name2lu_name(pobj, name, strlen(name), &ln);
5979 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5980 if (unlikely(osd_object_is_root(pobj)))
5983 osd_get_ldiskfs_dirent_param(ldp, fid);
5984 child = osd_child_dentry_get(info->oti_env, pobj,
5985 ln.ln_name, ln.ln_namelen);
5986 child->d_fsdata = (void *)ldp;
5987 dquot_initialize(pobj->oo_inode);
5988 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5989 child, cinode, hlock);
5990 if (rc == 0 && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5991 struct ldiskfs_dir_entry_2 *de;
5992 struct buffer_head *bh;
5995 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5998 rc1 = osd_ldiskfs_journal_get_write_access(
5999 oth->ot_handle, pobj->oo_inode->i_sb, bh,
6002 if (S_ISDIR(cinode->i_mode))
6003 de->file_type = LDISKFS_DIRENT_LUFID |
6004 LDISKFS_FT_REG_FILE;
6006 de->file_type = LDISKFS_DIRENT_LUFID |
6008 ldiskfs_handle_dirty_metadata(oth->ot_handle,
6015 if (ln.ln_name != name)
6021 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
6022 * into the directory.Also sets flags into osd object to
6023 * indicate dot and dotdot are created. This is required for
6024 * interoperability mode (b11826)
6026 * \param dir directory for dot and dotdot fixup.
6027 * \param obj child object for linking
6029 * \retval 0, on success
6030 * \retval -ve, on error
6032 static int osd_add_dot_dotdot(struct osd_thread_info *info,
6033 struct osd_object *dir,
6034 struct inode *parent_dir, const char *name,
6035 const struct lu_fid *dot_fid,
6036 const struct lu_fid *dot_dot_fid,
6039 struct inode *inode = dir->oo_inode;
6040 struct osd_thandle *oth;
6043 oth = container_of(th, struct osd_thandle, ot_super);
6044 LASSERT(oth->ot_handle->h_transaction != NULL);
6045 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
6047 if (strcmp(name, dot) == 0) {
6048 if (dir->oo_compat_dot_created) {
6051 LASSERT(inode->i_ino == parent_dir->i_ino);
6052 dir->oo_compat_dot_created = 1;
6055 } else if (strcmp(name, dotdot) == 0) {
6056 if (!dir->oo_compat_dot_created)
6058 /* in case of rename, dotdot is already created */
6059 if (dir->oo_compat_dotdot_created) {
6060 return __osd_ea_add_rec(info, dir, parent_dir, name,
6061 dot_dot_fid, NULL, th);
6064 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
6065 struct lu_fid tfid = *dot_dot_fid;
6068 result = osd_add_dot_dotdot_internal(info,
6069 dir->oo_inode, parent_dir, dot_fid,
6072 result = osd_add_dot_dotdot_internal(info,
6073 dir->oo_inode, parent_dir, dot_fid,
6078 dir->oo_compat_dotdot_created = 1;
6086 * It will call the appropriate osd_add* function and return the
6087 * value, return by respective functions.
6089 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
6090 struct inode *cinode, const char *name,
6091 const struct lu_fid *fid, struct thandle *th)
6093 struct osd_thread_info *info = osd_oti_get(env);
6094 struct htree_lock *hlock;
6097 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
6099 if (name[0] == '.' && (name[1] == '\0' ||
6100 (name[1] == '.' && name[2] == '\0'))) {
6101 if (hlock != NULL) {
6102 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6105 down_write(&pobj->oo_ext_idx_sem);
6108 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
6109 lu_object_fid(&pobj->oo_dt.do_lu),
6112 if (hlock != NULL) {
6113 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6114 pobj->oo_inode, LDISKFS_HLOCK_ADD);
6116 down_write(&pobj->oo_ext_idx_sem);
6119 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
6120 struct lu_fid *tfid = &info->oti_fid;
6124 rc = __osd_ea_add_rec(info, pobj, cinode, name,
6127 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
6131 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))){
6132 int dirent_count = atomic_read(&pobj->oo_dirent_count);
6134 /* avoid extremely unlikely 2B-entry directory overflow case */
6135 if (dirent_count != LU_DIRENT_COUNT_UNSET &&
6136 likely(dirent_count < INT_MAX - NR_CPUS))
6137 atomic_inc(&pobj->oo_dirent_count);
6141 ldiskfs_htree_unlock(hlock);
6143 up_write(&pobj->oo_ext_idx_sem);
6149 osd_ldiskfs_consistency_check(struct osd_thread_info *oti,
6150 struct osd_device *dev,
6151 const struct lu_fid *fid,
6152 struct osd_inode_id *id)
6154 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
6155 struct inode *inode = NULL;
6161 if (!scrub_needs_check(scrub, fid, id->oii_ino))
6164 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
6165 if (rc == -ENOENT) {
6166 __u32 gen = id->oii_gen;
6172 inode = osd_iget(oti, dev, id);
6173 /* The inode has been removed (by race maybe). */
6174 if (IS_ERR(inode)) {
6175 rc = PTR_ERR(inode);
6177 RETURN(rc == -ESTALE ? -ENOENT : rc);
6180 /* The OI mapping is lost. */
6181 if (gen != OSD_OII_NOGEN)
6185 * The inode may has been reused by others, we do not know,
6186 * leave it to be handled by subsequent osd_fid_lookup().
6189 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
6196 if (scrub->os_running) {
6197 if (inode == NULL) {
6198 inode = osd_iget(oti, dev, id);
6199 /* The inode has been removed (by race maybe). */
6200 if (IS_ERR(inode)) {
6201 rc = PTR_ERR(inode);
6203 RETURN(rc == -ESTALE ? -ENOENT : rc);
6207 rc = osd_scrub_oi_insert(dev, fid, id, insert);
6209 * There is race condition between osd_oi_lookup and OI scrub.
6210 * The OI scrub finished just after osd_oi_lookup() failure.
6211 * Under such case, it is unnecessary to trigger OI scrub again,
6212 * but try to call osd_oi_lookup() again.
6214 if (unlikely(rc == -EAGAIN))
6217 if (!S_ISDIR(inode->i_mode))
6220 rc = osd_check_lmv(oti, dev, inode);
6225 if (dev->od_scrub.os_scrub.os_auto_scrub_interval != AS_NEVER &&
6227 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
6228 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
6229 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
6230 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
6231 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
6232 if (rc == 0 || rc == -EALREADY)
6244 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
6245 struct osd_device *dev,
6246 struct lu_fid *fid, __u32 ino)
6248 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
6249 struct osd_idmap_cache *oic = &oti->oti_cache;
6250 struct inode *inode;
6253 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
6254 inode = osd_iget(oti, dev, &oic->oic_lid);
6255 if (IS_ERR(inode)) {
6256 fid_zero(&oic->oic_fid);
6257 return PTR_ERR(inode);
6260 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
6263 fid_zero(&oic->oic_fid);
6265 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
6269 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
6270 struct osd_inode_id *id, const struct lu_fid *fid)
6272 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
6273 id->oii_ino, id->oii_gen, info);
6274 info->oti_cache.oic_lid = *id;
6275 info->oti_cache.oic_fid = *fid;
6276 info->oti_cache.oic_dev = osd;
6280 * Get parent FID from the linkEA.
6282 * For a directory which parent resides on remote MDT, to satisfy the
6283 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
6284 * the other hand, to make the lookup(..) on the directory can return
6285 * the real parent FID, we append the real parent FID after its ".."
6286 * name entry in the /REMOTE_PARENT_DIR.
6288 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
6289 * backup. So after the restore, we cannot get the right parent FID from
6290 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
6291 * we have stored the real parent FID in the directory object's linkEA,
6292 * we can parse the linkEA for the real parent FID.
6294 * \param[in] env pointer to the thread context
6295 * \param[in] obj pointer to the object to be handled
6296 * \param[out]fid pointer to the buffer to hold the parent FID
6298 * \retval 0 for getting the real parent FID successfully
6299 * \retval negative error number on failure
6301 static int osd_get_pfid_from_linkea(const struct lu_env *env,
6302 struct osd_object *obj,
6305 struct osd_thread_info *oti = osd_oti_get(env);
6306 struct lu_buf *buf = &oti->oti_big_buf;
6307 struct dentry *dentry = &oti->oti_obj_dentry;
6308 struct inode *inode = obj->oo_inode;
6309 struct linkea_data ldata = { NULL };
6315 if (!S_ISDIR(inode->i_mode))
6319 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6320 buf->lb_buf, buf->lb_len);
6321 if (rc == -ERANGE) {
6322 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6325 lu_buf_realloc(buf, rc);
6326 if (buf->lb_buf == NULL)
6333 if (unlikely(rc == 0))
6339 if (unlikely(buf->lb_buf == NULL)) {
6340 lu_buf_realloc(buf, rc);
6341 if (buf->lb_buf == NULL)
6348 rc = linkea_init_with_rec(&ldata);
6350 linkea_first_entry(&ldata);
6351 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
6357 static int osd_verify_ent_by_linkea(const struct lu_env *env,
6358 struct inode *inode,
6359 const struct lu_fid *pfid,
6360 const char *name, const int namelen)
6362 struct osd_thread_info *oti = osd_oti_get(env);
6363 struct lu_buf *buf = &oti->oti_big_buf;
6364 struct dentry *dentry = &oti->oti_obj_dentry;
6365 struct linkea_data ldata = { NULL };
6366 struct lu_name cname = { .ln_name = name,
6367 .ln_namelen = namelen };
6373 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6374 buf->lb_buf, buf->lb_len);
6376 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6381 if (unlikely(rc == 0))
6384 if (buf->lb_len < rc) {
6385 lu_buf_realloc(buf, rc);
6386 if (buf->lb_buf == NULL)
6393 rc = linkea_init_with_rec(&ldata);
6395 rc = linkea_links_find(&ldata, &cname, pfid);
6401 * Calls ->lookup() to find dentry. From dentry get inode and
6402 * read inode's ea to get fid. This is required for interoperability
6405 * \retval 0, on success
6406 * \retval -ve, on error
6408 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6409 struct dt_rec *rec, const struct lu_name *ln)
6411 struct inode *dir = obj->oo_inode;
6412 struct dentry *dentry;
6413 struct ldiskfs_dir_entry_2 *de;
6414 struct buffer_head *bh;
6415 struct lu_fid *fid = (struct lu_fid *)rec;
6416 struct htree_lock *hlock = NULL;
6422 LASSERT(dir->i_op != NULL);
6423 LASSERT(dir->i_op->lookup != NULL);
6425 dentry = osd_child_dentry_get(env, obj, ln->ln_name, ln->ln_namelen);
6427 if (obj->oo_hl_head != NULL) {
6428 hlock = osd_oti_get(env)->oti_hlock;
6429 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6430 dir, LDISKFS_HLOCK_LOOKUP);
6432 down_read(&obj->oo_ext_idx_sem);
6435 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6437 struct osd_thread_info *oti = osd_oti_get(env);
6438 struct osd_inode_id *id = &oti->oti_id;
6439 struct osd_device *dev = osd_obj2dev(obj);
6441 ino = le32_to_cpu(de->inode);
6442 if (CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6444 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6448 rc = osd_get_fid_from_dentry(de, rec);
6450 /* done with de, release bh */
6453 if (unlikely(is_remote_parent_ino(dev, ino))) {
6455 * If the parent is on remote MDT, and there
6456 * is no FID-in-dirent, then we have to get
6457 * the parent FID from the linkEA.
6459 if (likely(ln->ln_namelen == 2 &&
6460 ln->ln_name[0] == '.' && ln->ln_name[1] == '.'))
6461 rc = osd_get_pfid_from_linkea(env, obj,
6464 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6467 osd_id_gen(id, ino, OSD_OII_NOGEN);
6470 if (rc != 0 || osd_remote_fid(env, dev, fid))
6473 rc = osd_ldiskfs_consistency_check(oti, dev, fid, id);
6474 if (rc != -ENOENT) {
6475 /* Other error should not affect lookup result. */
6478 /* Normal file mapping should be added into OI cache
6479 * after FID in LMA check, but for local files like
6480 * hsm_actions, their FIDs are not stored in OI files,
6481 * see osd_initial_OI_scrub(), and here is the only
6482 * place to load mapping into OI cache.
6484 if (!fid_is_namespace_visible(fid))
6485 osd_add_oi_cache(osd_oti_get(env),
6486 osd_obj2dev(obj), id, fid);
6488 CDEBUG(D_INODE, DFID"/"DNAME" => "DFID"\n",
6489 PFID(lu_object_fid(&obj->oo_dt.do_lu)), PNAME(ln),
6499 ldiskfs_htree_unlock(hlock);
6501 up_read(&obj->oo_ext_idx_sem);
6505 static int osd_index_declare_ea_insert(const struct lu_env *env,
6506 struct dt_object *dt,
6507 const struct dt_rec *rec,
6508 const struct dt_key *key,
6509 struct thandle *handle)
6511 struct osd_thandle *oh;
6512 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6513 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6514 const struct lu_fid *fid = rec1->rec_fid;
6515 int credits, rc = 0;
6516 struct osd_idmap_cache *idc;
6520 LASSERT(!dt_object_remote(dt));
6521 LASSERT(handle != NULL);
6522 LASSERT(fid != NULL);
6523 LASSERT(rec1->rec_type != 0);
6525 oh = container_of(handle, struct osd_thandle, ot_super);
6526 LASSERT(oh->ot_handle == NULL);
6528 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6531 * we can't call iget() while a transactions is running
6532 * (this can lead to a deadlock), but we need to know
6533 * inum and object type. so we find this information at
6534 * declaration and cache in per-thread info
6536 idc = osd_idc_find_or_init(env, osd, fid);
6538 RETURN(PTR_ERR(idc));
6539 if (idc->oic_remote) {
6541 * a reference to remote inode is represented by an
6542 * agent inode which we have to create
6544 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6545 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6548 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6550 if (osd_dt_obj(dt)->oo_inode != NULL) {
6551 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6554 * We ignore block quota on meta pool (MDTs), so needn't
6555 * calculate how many blocks will be consumed by this index
6558 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6560 i_projid_read(inode), 0,
6561 oh, osd_dt_obj(dt), NULL,
6566 #ifdef HAVE_PROJECT_QUOTA
6568 * Reserve credits for local agent inode to transfer
6569 * to 0, quota enforcement is ignored in this case.
6571 if (idc->oic_remote &&
6572 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6573 i_projid_read(inode) != 0)
6574 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6575 0, i_projid_read(inode),
6576 0, false, PRJQUOTA);
6584 * Index add function for interoperability mode (b11826).
6585 * It will add the directory entry.This entry is needed to
6586 * maintain name->fid mapping.
6588 * \param key it is key i.e. file entry to be inserted
6589 * \param rec it is value of given key i.e. fid
6591 * \retval 0, on success
6592 * \retval -ve, on error
6594 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6595 const struct dt_rec *rec,
6596 const struct dt_key *key, struct thandle *th)
6598 struct osd_object *obj = osd_dt_obj(dt);
6599 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6600 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6601 const struct lu_fid *fid = rec1->rec_fid;
6602 const char *name = (const char *)key;
6603 struct osd_thread_info *oti = osd_oti_get(env);
6604 struct inode *child_inode = NULL;
6605 struct osd_idmap_cache *idc;
6610 if (!dt_object_exists(dt))
6613 LASSERT(osd_invariant(obj));
6614 LASSERT(!dt_object_remote(dt));
6615 LASSERT(th != NULL);
6617 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6619 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6621 idc = osd_idc_find(env, osd, fid);
6622 if (unlikely(idc == NULL)) {
6623 idc = osd_idc_find_or_init(env, osd, fid);
6626 * this dt_insert() wasn't declared properly, so
6627 * FID is missing in OI cache. we better do not
6628 * lookup FID in FLDB/OI and don't risk to deadlock,
6629 * but in some special cases (lfsck testing, etc)
6630 * it's much simpler than fixing a caller.
6632 * normally this error should be placed after the first
6633 * find, but migrate may attach source stripes to
6634 * target, which doesn't create stripes.
6636 CERROR("%s: "DFID" wasn't declared for insert\n",
6637 osd_name(osd), PFID(fid));
6639 RETURN(PTR_ERR(idc));
6643 if (idc->oic_remote) {
6644 /* Insert remote entry */
6645 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6647 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6649 child_inode = osd_create_local_agent_inode(env, osd,
6650 obj, fid, rec1->rec_type & S_IFMT, th);
6651 if (IS_ERR(child_inode))
6652 RETURN(PTR_ERR(child_inode));
6655 /* Insert local entry */
6656 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6657 /* for a reason OI cache wasn't filled properly */
6658 CERROR("%s: OIC for "DFID" isn't filled\n",
6659 osd_name(osd), PFID(fid));
6662 child_inode = oti->oti_inode;
6663 if (unlikely(child_inode == NULL)) {
6664 struct ldiskfs_inode_info *lii;
6669 child_inode = oti->oti_inode = &lii->vfs_inode;
6671 child_inode->i_sb = osd_sb(osd);
6672 child_inode->i_ino = idc->oic_lid.oii_ino;
6673 child_inode->i_mode = rec1->rec_type & S_IFMT;
6676 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6678 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6679 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6681 if (child_inode && child_inode != oti->oti_inode)
6683 LASSERT(osd_invariant(obj));
6684 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6690 * Initialize osd Iterator for given osd index object.
6692 * \param dt osd index object
6695 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6696 struct dt_object *dt,
6699 struct osd_it_iam *it;
6700 struct osd_object *obj = osd_dt_obj(dt);
6701 struct lu_object *lo = &dt->do_lu;
6702 struct iam_path_descr *ipd;
6703 struct iam_container *bag = &obj->oo_dir->od_container;
6705 if (!dt_object_exists(dt))
6706 return ERR_PTR(-ENOENT);
6710 return ERR_PTR(-ENOMEM);
6712 ipd = osd_it_ipd_get(env, bag);
6713 if (likely(ipd != NULL)) {
6717 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6718 return (struct dt_it *)it;
6721 return ERR_PTR(-ENOMEM);
6726 * free given Iterator.
6728 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6730 struct osd_it_iam *it = (struct osd_it_iam *)di;
6731 struct osd_object *obj = it->oi_obj;
6733 iam_it_fini(&it->oi_it);
6734 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6735 osd_object_put(env, obj);
6740 * Move Iterator to record specified by \a key
6742 * \param di osd iterator
6743 * \param key key for index
6745 * \retval +ve di points to record with least key not larger than key
6746 * \retval 0 di points to exact matched key
6747 * \retval -ve failure
6750 static int osd_it_iam_get(const struct lu_env *env,
6751 struct dt_it *di, const struct dt_key *key)
6753 struct osd_thread_info *oti = osd_oti_get(env);
6754 struct osd_it_iam *it = (struct osd_it_iam *)di;
6756 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6757 /* swab quota uid/gid */
6758 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6759 key = (struct dt_key *)&oti->oti_quota_id;
6762 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6768 * \param di osd iterator
6770 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6772 struct osd_it_iam *it = (struct osd_it_iam *)di;
6774 iam_it_put(&it->oi_it);
6778 * Move iterator by one record
6780 * \param di osd iterator
6782 * \retval +1 end of container reached
6784 * \retval -ve failure
6787 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6789 struct osd_it_iam *it = (struct osd_it_iam *)di;
6791 return iam_it_next(&it->oi_it);
6795 * Return pointer to the key under iterator.
6798 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6799 const struct dt_it *di)
6801 struct osd_thread_info *oti = osd_oti_get(env);
6802 struct osd_it_iam *it = (struct osd_it_iam *)di;
6803 struct osd_object *obj = it->oi_obj;
6806 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6808 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6809 /* swab quota uid/gid */
6810 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6811 key = (struct dt_key *)&oti->oti_quota_id;
6818 * Return size of key under iterator (in bytes)
6821 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6823 struct osd_it_iam *it = (struct osd_it_iam *)di;
6825 return iam_it_key_size(&it->oi_it);
6829 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6831 /* check if file type is required */
6832 if (ent->lde_attrs & LUDA_TYPE) {
6833 struct luda_type *lt;
6834 int align = sizeof(*lt) - 1;
6836 len = (len + align) & ~align;
6837 lt = (struct luda_type *)(ent->lde_name + len);
6838 lt->lt_type = cpu_to_le16(DTTOIF(type));
6841 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6845 * build lu direct from backend fs dirent.
6849 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6850 char *name, __u16 namelen, __u16 type, __u32 attr)
6852 ent->lde_attrs = attr | LUDA_FID;
6853 fid_cpu_to_le(&ent->lde_fid, fid);
6855 ent->lde_hash = cpu_to_le64(offset);
6856 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6858 strncpy(ent->lde_name, name, namelen);
6859 ent->lde_name[namelen] = '\0';
6860 ent->lde_namelen = cpu_to_le16(namelen);
6862 /* append lustre attributes */
6863 osd_it_append_attrs(ent, namelen, type);
6867 * Return pointer to the record under iterator.
6869 static int osd_it_iam_rec(const struct lu_env *env,
6870 const struct dt_it *di,
6871 struct dt_rec *dtrec, __u32 attr)
6873 struct osd_it_iam *it = (struct osd_it_iam *)di;
6874 struct osd_thread_info *info = osd_oti_get(env);
6878 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6879 const struct osd_fid_pack *rec;
6880 struct lu_fid *fid = &info->oti_fid;
6881 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6887 name = (char *)iam_it_key_get(&it->oi_it);
6889 RETURN(PTR_ERR(name));
6891 namelen = iam_it_key_size(&it->oi_it);
6893 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6895 RETURN(PTR_ERR(rec));
6897 rc = osd_fid_unpack(fid, rec);
6901 hash = iam_it_store(&it->oi_it);
6903 /* IAM does not store object type in IAM index (dir) */
6904 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6906 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6907 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6908 (struct iam_rec *)dtrec);
6909 osd_quota_unpack(it->oi_obj, dtrec);
6911 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6912 (struct iam_rec *)dtrec);
6919 * Returns cookie for current Iterator position.
6921 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6923 struct osd_it_iam *it = (struct osd_it_iam *)di;
6925 return iam_it_store(&it->oi_it);
6929 * Restore iterator from cookie.
6931 * \param di osd iterator
6932 * \param hash Iterator location cookie
6934 * \retval +ve di points to record with least key not larger than key.
6935 * \retval 0 di points to exact matched key
6936 * \retval -ve failure
6939 static int osd_it_iam_load(const struct lu_env *env,
6940 const struct dt_it *di, __u64 hash)
6942 struct osd_it_iam *it = (struct osd_it_iam *)di;
6944 return iam_it_load(&it->oi_it, hash);
6947 static const struct dt_index_operations osd_index_iam_ops = {
6948 .dio_lookup = osd_index_iam_lookup,
6949 .dio_declare_insert = osd_index_declare_iam_insert,
6950 .dio_insert = osd_index_iam_insert,
6951 .dio_declare_delete = osd_index_declare_iam_delete,
6952 .dio_delete = osd_index_iam_delete,
6954 .init = osd_it_iam_init,
6955 .fini = osd_it_iam_fini,
6956 .get = osd_it_iam_get,
6957 .put = osd_it_iam_put,
6958 .next = osd_it_iam_next,
6959 .key = osd_it_iam_key,
6960 .key_size = osd_it_iam_key_size,
6961 .rec = osd_it_iam_rec,
6962 .store = osd_it_iam_store,
6963 .load = osd_it_iam_load
6967 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
6968 struct osd_device *dev,
6969 struct inode *inode, u32 attr)
6971 struct osd_thread_info *info = osd_oti_get(env);
6972 struct osd_it_ea *oie;
6976 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
6979 RETURN(ERR_CAST(file));
6981 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6982 if (attr & LUDA_64BITHASH)
6983 file->f_mode |= FMODE_64BITHASH;
6985 file->f_mode |= FMODE_32BITHASH;
6988 OBD_SLAB_ALLOC_PTR(oie, osd_itea_cachep);
6992 oie->oie_rd_dirent = 0;
6993 oie->oie_it_dirent = 0;
6994 oie->oie_dirent = NULL;
6995 if (unlikely(!info->oti_it_ea_buf_used)) {
6996 oie->oie_buf = info->oti_it_ea_buf;
6997 info->oti_it_ea_buf_used = 1;
6999 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7003 oie->oie_obj = NULL;
7004 oie->oie_file = file;
7009 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7013 return ERR_PTR(-ENOMEM);
7017 * Creates or initializes iterator context.
7019 * \retval struct osd_it_ea, iterator structure on success
7022 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
7023 struct dt_object *dt,
7026 struct osd_object *obj = osd_dt_obj(dt);
7027 struct osd_device *dev = osd_obj2dev(obj);
7028 struct lu_object *lo = &dt->do_lu;
7029 struct osd_it_ea *oie;
7033 if (!dt_object_exists(dt) || obj->oo_destroyed)
7034 RETURN(ERR_PTR(-ENOENT));
7036 oie = osd_it_dir_init(env, dev, obj->oo_inode, attr);
7038 RETURN(ERR_CAST(oie));
7042 RETURN((struct dt_it *)oie);
7045 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
7046 struct inode *inode)
7048 struct osd_thread_info *info = osd_oti_get(env);
7051 fput(oie->oie_file);
7052 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
7053 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7055 info->oti_it_ea_buf_used = 0;
7056 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7061 * Destroy or finishes iterator context.
7063 * \param di iterator structure to be destroyed
7065 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
7067 struct osd_it_ea *oie = (struct osd_it_ea *)di;
7068 struct osd_object *obj = oie->oie_obj;
7069 struct inode *inode = obj->oo_inode;
7072 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
7073 osd_object_put(env, obj);
7078 * It position the iterator at given key, so that next lookup continues from
7079 * that key Or it is similar to dio_it->load() but based on a key,
7080 * rather than file position.
7082 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
7085 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
7087 static int osd_it_ea_get(const struct lu_env *env,
7088 struct dt_it *di, const struct dt_key *key)
7090 struct osd_it_ea *it = (struct osd_it_ea *)di;
7093 LASSERT(((const char *)key)[0] == '\0');
7094 it->oie_file->f_pos = 0;
7095 it->oie_rd_dirent = 0;
7096 it->oie_it_dirent = 0;
7097 it->oie_dirent = NULL;
7105 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
7109 struct osd_filldir_cbs {
7110 struct dir_context ctx;
7111 struct osd_it_ea *it;
7114 * It is called internally by ->iterate*(). It fills the
7115 * iterator's in-memory data structure with required
7116 * information i.e. name, namelen, rec_size etc.
7118 * \param buf in which information to be filled in.
7119 * \param name name of the file in given dir
7121 * \retval 0 on success
7122 * \retval 1 on buffer full
7124 #ifdef HAVE_FILLDIR_USE_CTX
7125 static FILLDIR_TYPE do_osd_ldiskfs_filldir(struct dir_context *ctx,
7127 static int osd_ldiskfs_filldir(void *ctx,
7129 const char *name, int namelen,
7130 loff_t offset, __u64 ino, unsigned int d_type)
7132 struct osd_it_ea *it = ((struct osd_filldir_cbs *)ctx)->it;
7133 struct osd_object *obj = it->oie_obj;
7134 struct osd_it_ea_dirent *ent = it->oie_dirent;
7135 struct lu_fid *fid = &ent->oied_fid;
7136 char *buf = it->oie_buf;
7137 struct osd_fid_pack *rec;
7140 /* this should never happen */
7141 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
7142 CERROR("ldiskfs return invalid namelen %d\n", namelen);
7146 /* Check for enough space. Note oied_name is not NUL terminated. */
7147 if (&ent->oied_name[namelen] > buf + OSD_IT_EA_BUFSIZE)
7150 /* "." is just the object itself. */
7151 if (namelen == 1 && name[0] == '.') {
7153 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7154 } else if (d_type & LDISKFS_DIRENT_LUFID) {
7155 rec = (struct osd_fid_pack *)(name + namelen + 1);
7156 if (osd_fid_unpack(fid, rec) != 0)
7161 d_type &= ~LDISKFS_DIRENT_LUFID;
7163 /* NOT export local root. */
7165 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
7166 ino = obj->oo_inode->i_ino;
7167 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7170 if (obj == NULL || !(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
7171 ent->oied_namelen = namelen;
7172 memcpy(ent->oied_name, name, namelen);
7174 int encoded_namelen = critical_chars(name, namelen);
7176 /* Check again for enough space. */
7177 if (&ent->oied_name[encoded_namelen] > buf + OSD_IT_EA_BUFSIZE)
7180 ent->oied_namelen = encoded_namelen;
7182 if (encoded_namelen == namelen)
7183 memcpy(ent->oied_name, name, namelen);
7185 critical_encode(name, namelen, ent->oied_name);
7188 ent->oied_ino = ino;
7189 ent->oied_off = offset;
7190 ent->oied_type = d_type;
7192 it->oie_rd_dirent++;
7193 it->oie_dirent = (void *)ent + round_up(sizeof(*ent) + ent->oied_namelen, 8);
7197 WRAP_FILLDIR_FN(do_, osd_ldiskfs_filldir)
7200 * Calls ->iterate*() to load a directory entry at a time
7201 * and stored it in iterator's in-memory data structure.
7203 * \param di iterator's in memory structure
7205 * \retval 0 on success
7206 * \retval -ve on error
7207 * \retval +1 reach the end of entry
7209 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
7211 struct osd_it_ea *it = (struct osd_it_ea *)di;
7212 struct osd_object *obj = it->oie_obj;
7213 struct htree_lock *hlock = NULL;
7214 struct file *filp = it->oie_file;
7216 struct osd_filldir_cbs buf = {
7217 .ctx.actor = osd_ldiskfs_filldir,
7222 it->oie_dirent = it->oie_buf;
7223 it->oie_rd_dirent = 0;
7226 if (obj->oo_hl_head != NULL) {
7227 hlock = osd_oti_get(env)->oti_hlock;
7228 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
7230 LDISKFS_HLOCK_READDIR);
7232 down_read(&obj->oo_ext_idx_sem);
7236 rc = iterate_dir(filp, &buf.ctx);
7240 if (it->oie_rd_dirent == 0) {
7242 * If it does not get any dirent, it means it has been reached
7243 * to the end of the dir
7245 it->oie_file->f_pos = ldiskfs_get_htree_eof(it->oie_file);
7249 it->oie_dirent = it->oie_buf;
7250 it->oie_it_dirent = 1;
7255 ldiskfs_htree_unlock(hlock);
7257 up_read(&obj->oo_ext_idx_sem);
7264 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7265 * to load a directory entry at a time and stored it in
7266 * iterator's in-memory data structure.
7268 * \param di iterator's in memory structure
7270 * \retval +ve iterator reached to end
7271 * \retval 0 iterator not reached to end
7272 * \retval -ve on error
7274 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
7276 struct osd_it_ea *it = (struct osd_it_ea *)di;
7281 if (it->oie_it_dirent < it->oie_rd_dirent) {
7283 (void *)it->oie_dirent +
7284 round_up(sizeof(struct osd_it_ea_dirent) +
7285 it->oie_dirent->oied_namelen, 8);
7286 it->oie_it_dirent++;
7289 if (it->oie_file->f_pos == ldiskfs_get_htree_eof(it->oie_file))
7292 rc = osd_ldiskfs_it_fill(env, di);
7299 * Returns the key at current position from iterator's in memory structure.
7301 * \param di iterator's in memory structure
7303 * \retval key i.e. struct dt_key on success
7305 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
7306 const struct dt_it *di)
7308 struct osd_it_ea *it = (struct osd_it_ea *)di;
7310 return (struct dt_key *)it->oie_dirent->oied_name;
7314 * Returns key's size at current position from iterator's in memory structure.
7316 * \param di iterator's in memory structure
7318 * \retval key_size i.e. struct dt_key on success
7320 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
7322 struct osd_it_ea *it = (struct osd_it_ea *)di;
7324 return it->oie_dirent->oied_namelen;
7327 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
7328 #undef LDISKFS_DIR_REC_LEN
7329 # if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7330 # define LDISKFS_DIR_REC_LEN(de, dir) LDISKFS_DIR_ENTRY_LEN_((de), (dir))
7332 # define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
7336 #if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7337 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de), NULL)
7339 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de))
7342 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
7344 if (LDISKFS_DIR_REC_LEN_DIR(de) >=
7345 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
7352 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
7353 unsigned int blocksize, bool dotdot)
7356 return osd_dotdot_has_space(de);
7358 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
7359 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
7366 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
7367 handle_t *jh, struct dentry *dentry,
7368 const struct lu_fid *fid, struct buffer_head *bh,
7369 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
7372 struct inode *dir = dentry->d_parent->d_inode;
7373 struct inode *inode = dentry->d_inode;
7374 struct osd_fid_pack *rec;
7375 struct ldiskfs_dentry_param *ldp;
7376 int namelen = dentry->d_name.len;
7378 struct osd_thread_info *info = osd_oti_get(env);
7382 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7385 /* There is enough space to hold the FID-in-dirent. */
7386 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7387 rc = osd_ldiskfs_journal_get_write_access(jh, dir->i_sb, bh,
7392 de->name[namelen] = 0;
7393 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7394 rec->fp_len = sizeof(struct lu_fid) + 1;
7395 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7396 de->file_type |= LDISKFS_DIRENT_LUFID;
7397 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7404 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7408 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7409 osd_get_ldiskfs_dirent_param(ldp, fid);
7410 dentry->d_fsdata = (void *)ldp;
7411 dquot_initialize(dir);
7412 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7414 * It is too bad, we cannot reinsert the name entry back.
7415 * That means we lose it!
7419 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7420 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7421 namelen, dentry->d_name.name, PFID(fid), rc);
7427 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7428 struct osd_it_ea *it, struct lu_fid *fid,
7429 struct osd_inode_id *id, __u32 *attr)
7431 struct osd_thread_info *info = osd_oti_get(env);
7432 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7433 struct osd_device *dev = osd_obj2dev(obj);
7434 struct super_block *sb = osd_sb(dev);
7435 const char *devname = osd_name(dev);
7436 struct osd_it_ea_dirent *ent = it->oie_dirent;
7437 struct inode *dir = obj->oo_inode;
7438 struct htree_lock *hlock = NULL;
7439 struct buffer_head *bh = NULL;
7440 handle_t *jh = NULL;
7441 struct ldiskfs_dir_entry_2 *de;
7442 struct dentry *dentry;
7443 struct inode *inode;
7444 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7447 bool dotdot = false;
7453 if (ent->oied_name[0] == '.') {
7454 if (ent->oied_namelen == 1)
7457 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7461 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7462 inode = osd_iget(info, dev, id);
7463 if (IS_ERR(inode)) {
7464 rc = PTR_ERR(inode);
7465 if (rc == -ENOENT || rc == -ESTALE) {
7467 * Maybe dangling name entry, or
7468 * corrupted directory entry.
7470 *attr |= LUDA_UNKNOWN;
7473 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7474 "check_repair, dir = %lu/%u, name = %.*s, "
7475 "ino = %llu, rc = %d\n",
7476 devname, dir->i_ino, dir->i_generation,
7477 ent->oied_namelen, ent->oied_name,
7484 rc = obj_name2lu_name(obj, ent->oied_name, ent->oied_namelen, &ln);
7488 dentry = osd_child_dentry_by_inode(env, dir, ln.ln_name, ln.ln_namelen);
7489 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7490 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7496 * We need to ensure that the name entry is still valid.
7497 * Because it may be removed or renamed by other already.
7499 * The unlink or rename operation will start journal before PDO lock,
7500 * so to avoid deadlock, here we need to start journal handle before
7501 * related PDO lock also. But because we do not know whether there
7502 * will be something to be repaired before PDO lock, we just start
7503 * journal without conditions.
7505 * We may need to remove the name entry firstly, then insert back.
7506 * One credit is for user quota file update.
7507 * One credit is for group quota file update.
7508 * Two credits are for dirty inode.
7510 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7511 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7513 if (dev->od_dirent_journal != 0) {
7516 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7519 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7520 "check_repair, dir = %lu/%u, credits = %d, "
7521 "name = %.*s, ino = %llu: rc = %d\n",
7522 devname, dir->i_ino, dir->i_generation, credits,
7523 ent->oied_namelen, ent->oied_name,
7526 GOTO(out_inode, rc);
7529 if (obj->oo_hl_head != NULL) {
7530 hlock = osd_oti_get(env)->oti_hlock;
7532 * "0" means exclusive lock for the whole directory.
7533 * We need to prevent others access such name entry
7534 * during the delete + insert. Neither HLOCK_ADD nor
7535 * HLOCK_DEL cannot guarantee the atomicity.
7537 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7539 down_write(&obj->oo_ext_idx_sem);
7542 if (obj->oo_hl_head != NULL) {
7543 hlock = osd_oti_get(env)->oti_hlock;
7544 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7545 LDISKFS_HLOCK_LOOKUP);
7547 down_read(&obj->oo_ext_idx_sem);
7551 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7552 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7553 *attr |= LUDA_IGNORE;
7559 * For dotdot entry, if there is not enough space to hold the
7560 * FID-in-dirent, just keep them there. It only happens when the
7561 * device upgraded from 1.8 or restored from MDT file-level backup.
7562 * For the whole directory, only dotdot entry have no FID-in-dirent
7563 * and needs to get FID from LMA when readdir, it will not affect the
7566 if (dotdot && !osd_dotdot_has_space(de)) {
7567 *attr |= LUDA_UNKNOWN;
7573 if (lu_fid_eq(fid, &lma->lma_self_fid))
7576 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7577 struct lu_fid *tfid = &lma->lma_self_fid;
7579 if (likely(dotdot &&
7580 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7581 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7583 * It must be REMOTE_PARENT_DIR and as the
7584 * 'dotdot' entry of remote directory
7586 *attr |= LUDA_IGNORE;
7588 CDEBUG(D_LFSCK, "%s: expect remote agent "
7589 "parent directory, but got %.*s under "
7590 "dir = %lu/%u with the FID "DFID"\n",
7591 devname, ent->oied_namelen,
7592 ent->oied_name, dir->i_ino,
7593 dir->i_generation, PFID(tfid));
7595 *attr |= LUDA_UNKNOWN;
7602 if (!fid_is_zero(fid)) {
7603 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7605 if (rc == -ENOENT ||
7607 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7609 * linkEA does not recognize the dirent entry,
7610 * it may because the dirent entry corruption
7611 * and points to other's inode.
7613 CDEBUG(D_LFSCK, "%s: the target inode does not "
7614 "recognize the dirent, dir = %lu/%u, "
7615 " name = %.*s, ino = %llu, "
7616 DFID": rc = %d\n", devname, dir->i_ino,
7617 dir->i_generation, ent->oied_namelen,
7618 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7619 *attr |= LUDA_UNKNOWN;
7624 if (rc && rc != -ENODATA) {
7625 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7626 "dir = %lu/%u, name = %.*s, ino = %llu, "
7627 DFID": rc = %d\n", devname, dir->i_ino,
7628 dir->i_generation, ent->oied_namelen,
7629 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7630 *attr |= LUDA_UNKNOWN;
7638 * linkEA recognizes the dirent entry, the FID-in-LMA is
7639 * valid, trusted, in spite of fid_is_sane(fid) or not.
7641 if (*attr & LUDA_VERIFY_DRYRUN) {
7642 *fid = lma->lma_self_fid;
7643 *attr |= LUDA_REPAIR;
7650 dev->od_dirent_journal = 1;
7651 if (hlock != NULL) {
7652 ldiskfs_htree_unlock(hlock);
7655 up_read(&obj->oo_ext_idx_sem);
7661 *fid = lma->lma_self_fid;
7663 /* Update or append the FID-in-dirent. */
7664 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7665 bh, de, hlock, dotdot);
7667 *attr |= LUDA_REPAIR;
7669 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7670 "the dirent, dir = %lu/%u, name = %.*s, "
7671 "ino = %llu, "DFID": rc = %d\n",
7672 devname, dir->i_ino, dir->i_generation,
7673 ent->oied_namelen, ent->oied_name,
7674 ent->oied_ino, PFID(fid), rc);
7676 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7677 if (*attr & LUDA_VERIFY_DRYRUN) {
7678 if (fid_is_sane(fid)) {
7679 *attr |= LUDA_REPAIR;
7680 } else if (dev->od_index == 0) {
7681 lu_igif_build(fid, inode->i_ino,
7682 inode->i_generation);
7683 *attr |= LUDA_UPGRADE;
7691 dev->od_dirent_journal = 1;
7692 if (hlock != NULL) {
7693 ldiskfs_htree_unlock(hlock);
7696 up_read(&obj->oo_ext_idx_sem);
7703 if (unlikely(fid_is_sane(fid))) {
7705 * FID-in-dirent exists, but FID-in-LMA is lost.
7706 * Trust the FID-in-dirent, and add FID-in-LMA.
7708 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7710 *attr |= LUDA_REPAIR;
7712 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7713 "update dirent, dir = %lu/%u, "
7714 "name = %.*s, ino = %llu, "
7716 devname, dir->i_ino, dir->i_generation,
7717 ent->oied_namelen, ent->oied_name,
7718 ent->oied_ino, PFID(fid), rc);
7719 } else if (dev->od_index == 0) {
7720 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7722 * It is probably IGIF object. Only aappend the
7723 * FID-in-dirent. OI scrub will process FID-in-LMA.
7725 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7726 bh, de, hlock, dotdot);
7728 *attr |= LUDA_UPGRADE;
7730 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7731 "after the dirent, dir = %lu/%u, "
7732 "name = %.*s, ino = %llu, "
7734 devname, dir->i_ino, dir->i_generation,
7735 ent->oied_namelen, ent->oied_name,
7736 ent->oied_ino, PFID(fid), rc);
7745 if (hlock != NULL) {
7746 ldiskfs_htree_unlock(hlock);
7748 if (dev->od_dirent_journal != 0)
7749 up_write(&obj->oo_ext_idx_sem);
7751 up_read(&obj->oo_ext_idx_sem);
7755 ldiskfs_journal_stop(jh);
7759 if (rc >= 0 && !dirty)
7760 dev->od_dirent_journal = 0;
7761 if (ln.ln_name != ent->oied_name)
7768 * Returns the value at current position from iterator's in memory structure.
7770 * \param di struct osd_it_ea, iterator's in memory structure
7771 * \param attr attr requested for dirent.
7772 * \param lde lustre dirent
7774 * \retval 0 no error and \param lde has correct lustre dirent.
7775 * \retval -ve on error
7777 static inline int osd_it_ea_rec(const struct lu_env *env,
7778 const struct dt_it *di,
7779 struct dt_rec *dtrec, __u32 attr)
7781 struct osd_it_ea *it = (struct osd_it_ea *)di;
7782 struct osd_object *obj = it->oie_obj;
7783 struct osd_device *dev = osd_obj2dev(obj);
7784 struct osd_thread_info *oti = osd_oti_get(env);
7785 struct osd_inode_id *id = &oti->oti_id;
7786 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7787 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7788 __u32 ino = it->oie_dirent->oied_ino;
7793 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7795 if (attr & LUDA_VERIFY) {
7796 if (unlikely(is_remote_parent_ino(dev, ino))) {
7797 attr |= LUDA_IGNORE;
7799 * If the parent is on remote MDT, and there
7800 * is no FID-in-dirent, then we have to get
7801 * the parent FID from the linkEA.
7803 if (!fid_is_sane(fid) &&
7804 it->oie_dirent->oied_namelen == 2 &&
7805 it->oie_dirent->oied_name[0] == '.' &&
7806 it->oie_dirent->oied_name[1] == '.')
7807 osd_get_pfid_from_linkea(env, obj, fid);
7809 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7813 if (!fid_is_sane(fid))
7814 attr |= LUDA_UNKNOWN;
7816 attr &= ~LU_DIRENT_ATTRS_MASK;
7817 if (!fid_is_sane(fid)) {
7818 bool is_dotdot = false;
7820 if (it->oie_dirent->oied_namelen == 2 &&
7821 it->oie_dirent->oied_name[0] == '.' &&
7822 it->oie_dirent->oied_name[1] == '.')
7825 * If the parent is on remote MDT, and there
7826 * is no FID-in-dirent, then we have to get
7827 * the parent FID from the linkEA.
7829 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7830 rc = osd_get_pfid_from_linkea(env, obj, fid);
7832 if (is_dotdot == false &&
7833 CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7836 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7841 /* Pack the entry anyway, at least the offset is right. */
7842 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7843 it->oie_dirent->oied_name,
7844 it->oie_dirent->oied_namelen,
7845 it->oie_dirent->oied_type, attr);
7847 RETURN(rc > 0 ? 0 : rc);
7851 * Returns the record size size at current position.
7853 * This function will return record(lu_dirent) size in bytes.
7855 * \param[in] env execution environment
7856 * \param[in] di iterator's in memory structure
7857 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7858 * calculate the lu_dirent size.
7860 * \retval record size(in bytes & in memory) of the current lu_dirent
7863 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7866 struct osd_it_ea *it = (struct osd_it_ea *)di;
7868 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7872 * Returns a cookie for current position of the iterator head, so that
7873 * user can use this cookie to load/start the iterator next time.
7875 * \param di iterator's in memory structure
7877 * \retval cookie for current position, on success
7879 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7881 struct osd_it_ea *it = (struct osd_it_ea *)di;
7883 return it->oie_dirent->oied_off;
7887 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7888 * to load a directory entry at a time and stored it i inn,
7889 * in iterator's in-memory data structure.
7891 * \param di struct osd_it_ea, iterator's in memory structure
7893 * \retval +ve on success
7894 * \retval -ve on error
7896 static int osd_it_ea_load(const struct lu_env *env,
7897 const struct dt_it *di, __u64 hash)
7899 struct osd_it_ea *it = (struct osd_it_ea *)di;
7903 it->oie_file->f_pos = hash;
7905 rc = osd_ldiskfs_it_fill(env, di);
7915 int osd_olc_lookup(const struct lu_env *env, struct osd_object *obj,
7916 u64 iversion, struct dt_rec *rec,
7917 const struct lu_name *ln, int *result)
7919 struct osd_thread_info *oti = osd_oti_get(env);
7920 struct osd_lookup_cache *olc = oti->oti_lookup_cache;
7921 struct osd_device *osd = osd_obj2dev(obj);
7922 struct osd_lookup_cache_object *cobj = &oti->oti_cobj;
7925 if (unlikely(olc == NULL))
7928 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
7930 * umount has happened, a new OSD could land to the previous
7931 * address so we can't use it any more, invalidate our cache
7933 memset(olc, 0, sizeof(*olc));
7934 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
7938 memset(cobj, 0, sizeof(*cobj));
7939 cobj->lco_osd = osd;
7940 cobj->lco_ino = obj->oo_inode->i_ino;
7941 cobj->lco_gen = obj->oo_inode->i_generation;
7942 cobj->lco_version = iversion;
7944 for (i = 0; i < OSD_LOOKUP_CACHE_MAX; i++) {
7945 struct osd_lookup_cache_entry *entry;
7947 entry = &olc->olc_entry[i];
7948 /* compare if osd/ino/generation/version match */
7949 if (memcmp(&entry->lce_obj, cobj, sizeof(*cobj)) != 0)
7951 if (entry->lce_namelen != ln->ln_namelen)
7953 if (memcmp(entry->lce_name, ln->ln_name, ln->ln_namelen) != 0)
7956 memcpy(rec, &entry->lce_fid, sizeof(entry->lce_fid));
7957 *result = entry->lce_rc;
7963 void osd_olc_save(const struct lu_env *env, struct osd_object *obj,
7964 struct dt_rec *rec, const struct lu_name *ln,
7965 const int result, u64 iversion)
7967 struct osd_thread_info *oti = osd_oti_get(env);
7968 struct osd_lookup_cache_entry *entry;
7969 struct osd_lookup_cache *olc;
7971 if (unlikely(oti->oti_lookup_cache == NULL)) {
7972 OBD_ALLOC_PTR(oti->oti_lookup_cache);
7973 if (oti->oti_lookup_cache == NULL)
7977 olc = oti->oti_lookup_cache;
7978 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
7979 memset(olc, 0, sizeof(*olc));
7980 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
7983 entry = &olc->olc_entry[olc->olc_cur];
7985 /* invaliate cache slot if needed */
7986 if (entry->lce_obj.lco_osd)
7987 memset(&entry->lce_obj, 0, sizeof(entry->lce_obj));
7989 /* XXX: some kind of LRU */
7990 entry->lce_obj.lco_osd = osd_obj2dev(obj);
7991 entry->lce_obj.lco_ino = obj->oo_inode->i_ino;
7992 entry->lce_obj.lco_gen = obj->oo_inode->i_generation;
7993 entry->lce_obj.lco_version = iversion;
7995 LASSERT(ln->ln_namelen <= LDISKFS_NAME_LEN + 1);
7996 entry->lce_namelen = ln->ln_namelen;
7997 memcpy(entry->lce_name, ln->ln_name, ln->ln_namelen);
7998 memcpy(&entry->lce_fid, rec, sizeof(entry->lce_fid));
7999 entry->lce_rc = result;
8001 if (++olc->olc_cur == OSD_LOOKUP_CACHE_MAX)
8006 * Index lookup function for interoperability mode (b11826).
8008 * \param key, key i.e. file name to be searched
8010 * \retval +ve, on success
8011 * \retval -ve, on error
8013 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
8014 struct dt_rec *rec, const struct dt_key *key)
8016 struct osd_object *obj = osd_dt_obj(dt);
8023 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
8024 LINVRNT(osd_invariant(obj));
8026 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
8031 * grab version before actual lookup, so that we recognize potential
8032 * insert between osd_ea_lookup_rec() and osd_olc_save()
8034 iversion = inode_peek_iversion(obj->oo_inode);
8036 if (osd_olc_lookup(env, obj, iversion, rec, &ln, &result))
8037 GOTO(out, rc = result);
8039 rc = osd_ea_lookup_rec(env, obj, rec, &ln);
8043 osd_olc_save(env, obj, rec, &ln, rc, iversion);
8046 if (ln.ln_name != (char *)key)
8052 * Index and Iterator operations for interoperability
8053 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
8055 static const struct dt_index_operations osd_index_ea_ops = {
8056 .dio_lookup = osd_index_ea_lookup,
8057 .dio_declare_insert = osd_index_declare_ea_insert,
8058 .dio_insert = osd_index_ea_insert,
8059 .dio_declare_delete = osd_index_declare_ea_delete,
8060 .dio_delete = osd_index_ea_delete,
8062 .init = osd_it_ea_init,
8063 .fini = osd_it_ea_fini,
8064 .get = osd_it_ea_get,
8065 .put = osd_it_ea_put,
8066 .next = osd_it_ea_next,
8067 .key = osd_it_ea_key,
8068 .key_size = osd_it_ea_key_size,
8069 .rec = osd_it_ea_rec,
8070 .rec_size = osd_it_ea_rec_size,
8071 .store = osd_it_ea_store,
8072 .load = osd_it_ea_load
8076 static void *osd_key_init(const struct lu_context *ctx,
8077 struct lu_context_key *key)
8079 struct osd_thread_info *info;
8081 OBD_ALLOC_PTR(info);
8083 return ERR_PTR(-ENOMEM);
8085 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8086 if (info->oti_it_ea_buf == NULL)
8089 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
8091 info->oti_hlock = ldiskfs_htree_lock_alloc();
8092 if (info->oti_hlock == NULL)
8098 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8101 return ERR_PTR(-ENOMEM);
8104 static void osd_key_fini(const struct lu_context *ctx,
8105 struct lu_context_key *key, void *data)
8107 struct osd_thread_info *info = data;
8108 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
8109 struct osd_idmap_cache *idc = info->oti_ins_cache;
8111 if (info->oti_dio_pages) {
8113 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
8114 struct page *page = info->oti_dio_pages[i];
8116 LASSERT(PagePrivate2(page));
8117 LASSERT(PageLocked(page));
8118 ClearPagePrivate2(page);
8123 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
8124 PTLRPC_MAX_BRW_PAGES);
8127 if (info->oti_inode != NULL)
8129 if (info->oti_hlock != NULL)
8130 ldiskfs_htree_lock_free(info->oti_hlock);
8131 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8132 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
8133 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
8134 lu_buf_free(&info->oti_big_buf);
8136 LASSERT(info->oti_ins_cache_size > 0);
8137 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
8138 info->oti_ins_cache = NULL;
8139 info->oti_ins_cache_size = 0;
8141 if (info->oti_lookup_cache)
8142 OBD_FREE_PTR(info->oti_lookup_cache);
8146 static void osd_key_exit(const struct lu_context *ctx,
8147 struct lu_context_key *key, void *data)
8149 struct osd_thread_info *info = data;
8150 struct osd_lookup_cache *olc = info->oti_lookup_cache;
8153 memset(olc, 0, sizeof(*olc));
8154 LASSERT(info->oti_r_locks == 0);
8155 LASSERT(info->oti_w_locks == 0);
8156 LASSERT(info->oti_txns == 0);
8157 LASSERTF(info->oti_dio_pages_used == 0, "%d\n",
8158 info->oti_dio_pages_used);
8161 /* type constructor/destructor: osd_type_init, osd_type_fini */
8162 LU_TYPE_INIT_FINI(osd, &osd_key);
8164 struct lu_context_key osd_key = {
8165 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
8166 .lct_init = osd_key_init,
8167 .lct_fini = osd_key_fini,
8168 .lct_exit = osd_key_exit
8172 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
8173 const char *name, struct lu_device *next)
8175 struct osd_device *osd = osd_dev(d);
8177 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
8178 sizeof(osd->od_svname))
8180 return osd_procfs_init(osd, name);
8183 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
8185 struct seq_server_site *ss = osd_seq_site(osd);
8190 if (osd->od_is_ost || osd->od_cl_seq != NULL)
8193 if (unlikely(ss == NULL))
8196 OBD_ALLOC_PTR(osd->od_cl_seq);
8197 if (osd->od_cl_seq == NULL)
8200 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
8201 osd->od_svname, ss->ss_server_seq);
8203 if (ss->ss_node_id == 0) {
8205 * If the OSD on the sequence controller(MDT0), then allocate
8206 * sequence here, otherwise allocate sequence after connected
8207 * to MDT0 (see mdt_register_lwp_callback()).
8209 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
8210 &osd->od_cl_seq->lcs_space, env);
8216 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
8218 if (osd->od_cl_seq == NULL)
8221 seq_client_fini(osd->od_cl_seq);
8222 OBD_FREE_PTR(osd->od_cl_seq);
8223 osd->od_cl_seq = NULL;
8226 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
8230 /* shutdown quota slave instance associated with the device */
8231 if (o->od_quota_slave_md != NULL) {
8232 struct qsd_instance *qsd = o->od_quota_slave_md;
8234 o->od_quota_slave_md = NULL;
8238 if (o->od_quota_slave_dt != NULL) {
8239 struct qsd_instance *qsd = o->od_quota_slave_dt;
8241 o->od_quota_slave_dt = NULL;
8245 osd_fid_fini(env, o);
8246 osd_scrub_cleanup(env, o);
8251 #ifdef HAVE_FLUSH_DELAYED_FPUT
8252 # define cfs_flush_delayed_fput() flush_delayed_fput()
8254 void (*cfs_flush_delayed_fput)(void);
8255 #endif /* HAVE_FLUSH_DELAYED_FPUT */
8257 static void osd_umount(const struct lu_env *env, struct osd_device *o)
8261 atomic_inc(&osd_mount_seq);
8263 if (o->od_mnt != NULL) {
8264 shrink_dcache_sb(osd_sb(o));
8265 osd_sync(env, &o->od_dt_dev);
8266 wait_event(o->od_commit_cb_done,
8267 !atomic_read(&o->od_commit_cb_in_flight));
8273 /* to be sure all delayed fput are finished */
8274 cfs_flush_delayed_fput();
8279 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8280 # ifndef LDISKFS_HAS_INCOMPAT_FEATURE
8281 /* Newer kernels provide the ldiskfs_set_feature_largedir() wrapper already,
8282 * which calls ldiskfs_update_dynamic_rev() to update ancient filesystems.
8283 * All ldiskfs filesystems are already v2, so it is a no-op and unnecessary.
8284 * This avoids maintaining patches to export this otherwise-useless function.
8286 void ldiskfs_update_dynamic_rev(struct super_block *sb)
8293 static int osd_mount(const struct lu_env *env,
8294 struct osd_device *o, struct lustre_cfg *cfg)
8296 const char *name = lustre_cfg_string(cfg, 0);
8297 const char *dev = lustre_cfg_string(cfg, 1);
8299 unsigned long page, s_flags = 0, lmd_flags = 0;
8300 struct page *__page;
8301 struct file_system_type *type;
8302 char *options = NULL;
8304 struct osd_thread_info *info = osd_oti_get(env);
8305 struct lu_fid *fid = &info->oti_fid;
8306 struct inode *inode;
8307 int rc = 0, force_over_1024tb = 0;
8311 if (o->od_mnt != NULL)
8314 if (strlen(dev) >= sizeof(o->od_mntdev))
8316 strcpy(o->od_mntdev, dev);
8318 str = lustre_cfg_buf(cfg, 2);
8319 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
8321 opts = lustre_cfg_string(cfg, 3);
8323 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
8324 CERROR("%s: device %s extents feature is not guaranteed to "
8325 "work on big-endian systems. Use \"bigendian_extents\" "
8326 "mount option to override.\n", name, dev);
8330 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
8331 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
8332 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");
8335 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
8336 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
8337 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");
8340 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8341 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
8342 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");
8346 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
8347 force_over_1024tb = 1;
8349 __page = alloc_page(GFP_KERNEL);
8351 GOTO(out, rc = -ENOMEM);
8352 page = (unsigned long)page_address(__page);
8353 options = (char *)page;
8356 /* strip out the options for back compatiblity */
8357 static const char * const sout[] = {
8364 /* strip out option we processed in osd */
8365 "bigendian_extents",
8369 "force_over_1024tb",
8373 strncat(options, opts, PAGE_SIZE);
8374 for (rc = 0, str = options; sout[rc]; ) {
8375 char *op = strstr(str, sout[rc]);
8382 if (op == options || *(op - 1) == ',') {
8383 str = op + strlen(sout[rc]);
8384 if (*str == ',' || *str == '\0') {
8385 *str == ',' ? str++ : str;
8386 memmove(op, str, strlen(str) + 1);
8389 for (str = op; *str != ',' && *str != '\0'; str++)
8393 strncat(options, "user_xattr,acl", PAGE_SIZE);
8396 /* Glom up mount options */
8397 if (*options != '\0')
8398 strncat(options, ",", PAGE_SIZE);
8399 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
8401 type = get_fs_type("ldiskfs");
8403 CERROR("%s: cannot find ldiskfs module\n", name);
8404 GOTO(out, rc = -ENODEV);
8407 s_flags |= SB_KERNMOUNT;
8408 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
8409 module_put(type->owner);
8411 if (IS_ERR(o->od_mnt)) {
8412 rc = PTR_ERR(o->od_mnt);
8414 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
8418 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
8419 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
8420 force_over_1024tb == 0) {
8421 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",
8423 GOTO(out_mnt, rc = -EINVAL);
8426 if (test_bit(LMD_FLG_DEV_RDONLY, &lmd_flags)) {
8427 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
8430 GOTO(out_mnt, rc = -EOPNOTSUPP);
8433 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
8434 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
8435 GOTO(out_mnt, rc = -EINVAL);
8438 if (ldiskfs_has_feature_fast_commit(o->od_mnt->mnt_sb)) {
8439 CERROR("%s: device %s is mounted with fast_commit that breaks recovery\n",
8441 GOTO(out_mnt, rc = -EOPNOTSUPP);
8444 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8445 #ifdef LDISKFS_MOUNT_DIRDATA
8446 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
8447 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
8448 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
8449 CWARN("%s: device %s was upgraded from Lustre-1.x without "
8450 "enabling the dirdata feature. If you do not want to "
8451 "downgrade to Lustre-1.x again, you can enable it via "
8452 "'tune2fs -O dirdata device'\n", name, dev);
8454 /* enable large_dir on MDTs to avoid REMOTE_PARENT_DIR overflow,
8455 * and on very large OSTs to avoid object directory overflow */
8456 if (unlikely(!ldiskfs_has_feature_largedir(o->od_mnt->mnt_sb) &&
8457 !strstr(name, "MGS"))) {
8458 ldiskfs_set_feature_largedir(o->od_mnt->mnt_sb);
8459 LCONSOLE_INFO("%s: enabled 'large_dir' feature on device %s\n",
8463 inode = osd_sb(o)->s_root->d_inode;
8464 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
8465 if (!o->od_dt_dev.dd_rdonly) {
8466 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
8468 CERROR("%s: failed to set lma on %s root inode\n",
8474 if (test_bit(LMD_FLG_NOSCRUB, &lmd_flags))
8475 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_NEVER;
8477 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
8478 /* do not use pagecache with flash-backed storage */
8479 o->od_writethrough_cache = 0;
8480 o->od_read_cache = 0;
8491 __free_page(__page);
8496 static struct lu_device *osd_device_fini(const struct lu_env *env,
8497 struct lu_device *d)
8499 struct osd_device *o = osd_dev(d);
8503 osd_index_backup(env, o, false);
8504 osd_shutdown(env, o);
8506 if (o->od_oi_table != NULL)
8507 osd_oi_fini(osd_oti_get(env), o);
8508 if (o->od_extent_bytes_percpu)
8509 free_percpu(o->od_extent_bytes_percpu);
8510 osd_obj_map_fini(o);
8516 static int osd_device_init0(const struct lu_env *env,
8517 struct osd_device *o,
8518 struct lustre_cfg *cfg)
8520 struct lu_device *l = osd2lu_dev(o);
8521 struct osd_thread_info *info;
8524 bool restored = false;
8527 /* if the module was re-loaded, env can loose its keys */
8528 rc = lu_env_refill((struct lu_env *)env);
8531 info = osd_oti_get(env);
8534 l->ld_ops = &osd_lu_ops;
8535 o->od_dt_dev.dd_ops = &osd_dt_ops;
8537 spin_lock_init(&o->od_osfs_lock);
8538 mutex_init(&o->od_otable_mutex);
8539 INIT_LIST_HEAD(&o->od_orphan_list);
8540 INIT_LIST_HEAD(&o->od_index_backup_list);
8541 INIT_LIST_HEAD(&o->od_index_restore_list);
8542 spin_lock_init(&o->od_lock);
8543 o->od_index_backup_policy = LIBP_NONE;
8545 init_waitqueue_head(&o->od_commit_cb_done);
8547 o->od_read_cache = 1;
8548 o->od_writethrough_cache = 1;
8549 o->od_enable_projid_xattr = 0;
8550 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
8551 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
8552 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
8553 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_DEFAULT;
8554 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8555 o->od_fallocate_zero_blocks = 0;
8557 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
8558 sizeof(o->od_svname));
8559 if (cplen >= sizeof(o->od_svname)) {
8564 o->od_index_backup_stop = 0;
8565 o->od_index = -1; /* -1 means index is invalid */
8566 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8567 if (rc == LDD_F_SV_TYPE_OST)
8570 o->od_full_scrub_ratio = OFSR_DEFAULT;
8571 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8572 rc = osd_mount(env, o, cfg);
8576 /* Can only check block device after mount */
8577 o->od_nonrotational =
8578 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8580 rc = osd_obj_map_init(env, o);
8584 rc = lu_site_init(&o->od_site, l);
8586 GOTO(out_compat, rc);
8587 o->od_site.ls_bottom_dev = l;
8589 rc = lu_site_init_finish(&o->od_site);
8593 opts = lustre_cfg_string(cfg, 3);
8594 if (opts && strstr(opts, "resetoi"))
8597 INIT_LIST_HEAD(&o->od_ios_list);
8599 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
8601 GOTO(out_brw_stats, rc);
8603 /* setup scrub, including OI files initialization */
8605 rc = osd_scrub_setup(env, o, restored);
8608 GOTO(out_brw_stats, rc);
8610 rc = osd_procfs_init(o, o->od_svname);
8612 CERROR("%s: can't initialize procfs: rc = %d\n",
8614 GOTO(out_scrub, rc);
8617 LASSERT(l->ld_site->ls_linkage.next != NULL);
8618 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8620 /* initialize quota slave instance */
8621 /* currently it's no need to prepare qsd_instance_md for OST */
8622 if (!o->od_is_ost) {
8623 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8624 &o->od_dt_dev, o->od_proc_entry,
8626 if (IS_ERR(o->od_quota_slave_md)) {
8627 rc = PTR_ERR(o->od_quota_slave_md);
8628 o->od_quota_slave_md = NULL;
8629 GOTO(out_procfs, rc);
8633 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8634 o->od_proc_entry, false, true);
8636 if (IS_ERR(o->od_quota_slave_dt)) {
8637 if (o->od_quota_slave_md != NULL) {
8638 qsd_fini(env, o->od_quota_slave_md);
8639 o->od_quota_slave_md = NULL;
8642 rc = PTR_ERR(o->od_quota_slave_dt);
8643 o->od_quota_slave_dt = NULL;
8644 GOTO(out_procfs, rc);
8647 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8648 if (!o->od_extent_bytes_percpu) {
8650 GOTO(out_procfs, rc);
8658 osd_scrub_cleanup(env, o);
8660 lprocfs_fini_brw_stats(&o->od_brw_stats);
8662 lu_site_fini(&o->od_site);
8664 osd_obj_map_fini(o);
8671 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8672 struct lu_device_type *t,
8673 struct lustre_cfg *cfg)
8675 struct osd_device *o;
8680 return ERR_PTR(-ENOMEM);
8682 rc = dt_device_init(&o->od_dt_dev, t);
8685 * Because the ctx might be revived in dt_device_init,
8686 * refill the env here
8688 lu_env_refill((struct lu_env *)env);
8689 rc = osd_device_init0(env, o, cfg);
8691 dt_device_fini(&o->od_dt_dev);
8694 if (unlikely(rc != 0))
8697 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8700 static struct lu_device *osd_device_free(const struct lu_env *env,
8701 struct lu_device *d)
8703 struct osd_device *o = osd_dev(d);
8707 /* XXX: make osd top device in order to release reference */
8708 d->ld_site->ls_top_dev = d;
8709 lu_site_purge(env, d->ld_site, -1);
8710 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8711 D_ERROR, lu_cdebug_printer);
8712 lu_site_fini(&o->od_site);
8713 dt_device_fini(&o->od_dt_dev);
8718 static int osd_process_config(const struct lu_env *env,
8719 struct lu_device *d, struct lustre_cfg *cfg)
8721 struct osd_device *o = osd_dev(d);
8727 switch (cfg->lcfg_command) {
8729 rc = osd_mount(env, o, cfg);
8733 * For the case LCFG_PRE_CLEANUP is not called in advance,
8734 * that may happend if hit failure during mount process.
8736 osd_index_backup(env, o, false);
8737 lu_dev_del_linkage(d->ld_site, d);
8738 rc = osd_shutdown(env, o);
8741 LASSERT(&o->od_dt_dev);
8742 count = class_modify_config(cfg, PARAM_OSD,
8743 &o->od_dt_dev.dd_kobj);
8745 count = class_modify_config(cfg, PARAM_OST,
8746 &o->od_dt_dev.dd_kobj);
8747 rc = count > 0 ? 0 : count;
8749 case LCFG_PRE_CLEANUP:
8751 osd_index_backup(env, o,
8752 o->od_index_backup_policy != LIBP_NONE);
8762 static int osd_recovery_complete(const struct lu_env *env,
8763 struct lu_device *d)
8765 struct osd_device *osd = osd_dev(d);
8770 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8774 * start qsd instance on recovery completion, this notifies the quota
8775 * slave code that we are about to process new requests now
8777 rc = qsd_start(env, osd->od_quota_slave_dt);
8778 if (rc == 0 && osd->od_quota_slave_md != NULL)
8779 rc = qsd_start(env, osd->od_quota_slave_md);
8785 * we use exports to track all osd users
8787 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8788 struct obd_device *obd, struct obd_uuid *cluuid,
8789 struct obd_connect_data *data, void *localdata)
8791 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8792 struct lustre_handle conn;
8797 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8799 rc = class_connect(&conn, obd, cluuid);
8803 *exp = class_conn2export(&conn);
8805 spin_lock(&osd->od_osfs_lock);
8807 spin_unlock(&osd->od_osfs_lock);
8813 * once last export (we don't count self-export) disappeared
8814 * osd can be released
8816 static int osd_obd_disconnect(struct obd_export *exp)
8818 struct obd_device *obd = exp->exp_obd;
8819 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8820 int rc, release = 0;
8824 /* Only disconnect the underlying layers on the final disconnect. */
8825 spin_lock(&osd->od_osfs_lock);
8827 if (osd->od_connects == 0)
8829 spin_unlock(&osd->od_osfs_lock);
8831 rc = class_disconnect(exp); /* bz 9811 */
8833 if (rc == 0 && release)
8834 class_manual_cleanup(obd);
8838 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8839 struct lu_device *dev)
8841 struct osd_device *osd = osd_dev(dev);
8842 struct lr_server_data *lsd =
8843 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8848 if (osd->od_quota_slave_md != NULL) {
8849 /* set up quota slave objects for inode */
8850 result = qsd_prepare(env, osd->od_quota_slave_md);
8855 if (osd->od_quota_slave_dt != NULL) {
8856 /* set up quota slave objects for block */
8857 result = qsd_prepare(env, osd->od_quota_slave_dt);
8863 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8864 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8865 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8866 osd->od_index_in_idif = 1;
8868 osd->od_index_in_idif = 0;
8869 result = osd_register_proc_index_in_idif(osd);
8874 osd->od_index_in_idif = 1;
8878 result = osd_fid_init(env, osd);
8884 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8888 * see include/lu_object.h for the details.
8890 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8891 struct lu_fid *fid, struct lu_object *parent,
8892 const struct lu_name *name)
8894 struct osd_device *osd = osd_dev(d);
8896 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8899 static const struct lu_object_operations osd_lu_obj_ops = {
8900 .loo_object_init = osd_object_init,
8901 .loo_object_delete = osd_object_delete,
8902 .loo_object_release = osd_object_release,
8903 .loo_object_free = osd_object_free,
8904 .loo_object_print = osd_object_print,
8905 .loo_object_invariant = osd_object_invariant
8908 const struct lu_device_operations osd_lu_ops = {
8909 .ldo_object_alloc = osd_object_alloc,
8910 .ldo_process_config = osd_process_config,
8911 .ldo_recovery_complete = osd_recovery_complete,
8912 .ldo_prepare = osd_prepare,
8913 .ldo_fid_alloc = osd_fid_alloc,
8916 static const struct lu_device_type_operations osd_device_type_ops = {
8917 .ldto_init = osd_type_init,
8918 .ldto_fini = osd_type_fini,
8920 .ldto_start = osd_type_start,
8921 .ldto_stop = osd_type_stop,
8923 .ldto_device_alloc = osd_device_alloc,
8924 .ldto_device_free = osd_device_free,
8926 .ldto_device_init = osd_device_init,
8927 .ldto_device_fini = osd_device_fini
8930 static struct lu_device_type osd_device_type = {
8931 .ldt_tags = LU_DEVICE_DT,
8932 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8933 .ldt_ops = &osd_device_type_ops,
8934 .ldt_ctx_tags = LCT_LOCAL,
8937 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8939 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8940 struct super_block *sb = osd_sb(osd);
8942 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8946 * lprocfs legacy support.
8948 static const struct obd_ops osd_obd_device_ops = {
8949 .o_owner = THIS_MODULE,
8950 .o_connect = osd_obd_connect,
8951 .o_disconnect = osd_obd_disconnect,
8952 .o_health_check = osd_health_check,
8955 static ssize_t delayed_unlink_mb_show(struct kobject *kobj,
8956 struct attribute *attr, char *buf)
8958 return snprintf(buf, PAGE_SIZE, "%d\n",
8959 ldiskfs_delayed_unlink_blocks >> 11);
8962 static ssize_t delayed_unlink_mb_store(struct kobject *kobj,
8963 struct attribute *attr,
8964 const char *buffer, size_t count)
8966 u64 delayed_unlink_bytes;
8969 rc = sysfs_memparse(buffer, count, &delayed_unlink_bytes, "MiB");
8973 ldiskfs_delayed_unlink_blocks = delayed_unlink_bytes >> 9;
8977 LUSTRE_RW_ATTR(delayed_unlink_mb);
8980 static ssize_t track_declares_assert_show(struct kobject *kobj,
8981 struct attribute *attr,
8984 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8987 static ssize_t track_declares_assert_store(struct kobject *kobj,
8988 struct attribute *attr,
8989 const char *buffer, size_t count)
8991 bool track_declares_assert;
8994 rc = kstrtobool(buffer, &track_declares_assert);
8998 ldiskfs_track_declares_assert = track_declares_assert;
9002 LUSTRE_RW_ATTR(track_declares_assert);
9004 static int __init osd_init(void)
9006 struct kobject *kobj;
9009 BUILD_BUG_ON(BH_DXLock >=
9010 sizeof(((struct buffer_head *)0)->b_state) * 8);
9011 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
9012 /* please, try to keep osd_thread_info smaller than a page */
9013 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
9016 rc = libcfs_setup();
9022 rc = lu_kmem_init(ldiskfs_caches);
9026 rc = class_register_type(&osd_obd_device_ops, NULL, true,
9027 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
9029 lu_kmem_fini(ldiskfs_caches);
9033 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9035 rc = sysfs_create_file(kobj,
9036 &lustre_attr_track_declares_assert.attr);
9038 CWARN("%s: track_declares_assert sysfs registration failed: rc = %d\n",
9043 rc = sysfs_create_file(kobj,
9044 &lustre_attr_delayed_unlink_mb.attr);
9046 CWARN("%s: delayed_unlink_mb registration failed: rc = %d\n",
9054 #ifndef HAVE_FLUSH_DELAYED_FPUT
9055 if (unlikely(cfs_flush_delayed_fput == NULL))
9056 cfs_flush_delayed_fput =
9057 cfs_kallsyms_lookup_name("flush_delayed_fput");
9063 static void __exit osd_exit(void)
9065 struct kobject *kobj;
9067 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9069 sysfs_remove_file(kobj,
9070 &lustre_attr_track_declares_assert.attr);
9073 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
9074 lu_kmem_fini(ldiskfs_caches);
9077 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
9078 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
9079 MODULE_VERSION(LUSTRE_VERSION_STRING);
9080 MODULE_LICENSE("GPL");
9082 module_init(osd_init);
9083 module_exit(osd_exit);