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]);
2006 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
2007 void osd_ldiskfs_credits_for_revoke(struct osd_device *osd,
2008 struct osd_thandle *oh,
2009 int *credits, int *revoke)
2011 int blocks = LDISKFS_MAX_EXTENT_DEPTH * oh->oh_declared_ext;
2012 *revoke += ldiskfs_trans_default_revoke_credits(osd_sb(osd)) + blocks;
2015 void osd_ldiskfs_credits_for_revoke(struct osd_device *osd,
2016 struct osd_thandle *oh,
2017 int *credits, int *revoke)
2019 struct journal_s *journal = LDISKFS_SB(osd_sb(osd))->s_journal;
2020 int blocks, jbsize, records_per_block;
2022 blocks = LDISKFS_MAX_EXTENT_DEPTH * oh->oh_declared_ext;
2023 jbsize = journal->j_blocksize;
2024 jbsize -= sizeof(struct jbd2_journal_block_tail) +
2025 sizeof(jbd2_journal_revoke_header_t);
2026 records_per_block = jbsize / 8;
2027 *credits += (blocks + records_per_block - 1) / records_per_block;
2032 * Concurrency: shouldn't matter.
2034 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
2037 struct osd_thread_info *oti = osd_oti_get(env);
2038 struct osd_device *dev = osd_dt_dev(d);
2040 struct osd_thandle *oh;
2045 LASSERT(current->journal_info == NULL);
2047 oh = container_of(th, struct osd_thandle, ot_super);
2048 LASSERT(oh != NULL);
2049 LASSERT(oh->ot_handle == NULL);
2050 if (unlikely(ldiskfs_track_declares_assert != 0)) {
2051 LASSERT(oti->oti_r_locks == 0);
2052 LASSERT(oti->oti_w_locks == 0);
2055 rc = dt_txn_hook_start(env, d, th);
2059 if (unlikely(osd_param_is_not_sane(dev, th))) {
2060 static unsigned long last_printed;
2061 static int last_credits;
2063 lprocfs_counter_add(dev->od_stats,
2064 LPROC_OSD_TOO_MANY_CREDITS, 1);
2067 * don't make noise on a tiny testing systems
2068 * actual credits misuse will be caught anyway
2070 if (last_credits != oh->ot_credits &&
2071 time_after(jiffies, last_printed +
2072 cfs_time_seconds(60)) &&
2073 osd_transaction_size(dev) > 512) {
2074 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
2075 oh->ot_credits, osd_transaction_size(dev));
2076 osd_trans_dump_creds(env, th);
2078 last_credits = oh->ot_credits;
2079 last_printed = jiffies;
2082 * XXX Limit the credits to 'max_transaction_buffers', and
2083 * let the underlying filesystem to catch the error if
2084 * we really need so many credits.
2086 * This should be removed when we can calculate the
2087 * credits precisely.
2089 oh->ot_credits = osd_transaction_size(dev);
2090 } else if (ldiskfs_track_declares_assert != 0) {
2092 * reserve few credits to prevent an assertion in JBD
2093 * our debugging mechanism will be able to detected
2094 * overuse. this can help to debug single-update
2097 oh->ot_credits += 10;
2098 if (unlikely(osd_param_is_not_sane(dev, th)))
2099 oh->ot_credits = osd_transaction_size(dev);
2102 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
2103 GOTO(out, rc = -EIO);
2106 * we ignore quota checks for system-owned files, but still
2107 * need to count blocks for uid/gid/projid
2109 osd_trans_declare_op(env, oh, OSD_OT_QUOTA, 3);
2112 * XXX temporary stuff. Some abstraction layer should
2115 osd_ldiskfs_credits_for_revoke(dev, oh, &oh->ot_credits, &revoke);
2116 jh = osd_journal_start_with_revoke(osd_sb(dev), LDISKFS_HT_MISC,
2117 oh->ot_credits, revoke);
2121 LASSERT(oti->oti_txns == 0);
2123 atomic_inc(&dev->od_commit_cb_in_flight);
2124 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
2135 static int osd_seq_exists(const struct lu_env *env,
2136 struct osd_device *osd, u64 seq)
2138 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
2139 struct seq_server_site *ss = osd_seq_site(osd);
2144 LASSERT(ss != NULL);
2145 LASSERT(ss->ss_server_fld != NULL);
2147 rc = osd_fld_lookup(env, osd, seq, range);
2150 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2151 osd_name(osd), seq, rc);
2155 RETURN(ss->ss_node_id == range->lsr_index);
2158 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2160 struct dt_txn_commit_cb *dcb;
2161 struct dt_txn_commit_cb *tmp;
2163 /* call per-transaction stop callbacks if any */
2164 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2166 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2167 "commit callback entry: magic=%x name='%s'\n",
2168 dcb->dcb_magic, dcb->dcb_name);
2169 list_del_init(&dcb->dcb_linkage);
2170 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2175 * Concurrency: shouldn't matter.
2177 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2180 struct osd_thread_info *oti = osd_oti_get(env);
2181 struct osd_thandle *oh;
2182 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2183 struct osd_device *osd = osd_dt_dev(th->th_dev);
2184 struct qsd_instance *qsd = osd_def_qsd(osd);
2185 struct lquota_trans *qtrans;
2186 LIST_HEAD(truncates);
2187 int rc = 0, remove_agents = 0;
2191 oh = container_of(th, struct osd_thandle, ot_super);
2193 remove_agents = oh->ot_remove_agents;
2195 qtrans = oh->ot_quota_trans;
2196 oh->ot_quota_trans = NULL;
2198 osd_tx_declaration_free(oh);
2200 /* move locks to local list, stop tx, execute truncates */
2201 list_splice(&oh->ot_trunc_locks, &truncates);
2203 if (oh->ot_handle != NULL) {
2206 handle_t *hdl = oh->ot_handle;
2209 * add commit callback
2210 * notice we don't do this in osd_trans_start()
2211 * as underlying transaction can change during truncate
2213 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2216 LASSERT(oti->oti_txns == 1);
2219 rc = dt_txn_hook_stop(env, th);
2221 CERROR("%s: failed in transaction hook: rc = %d\n",
2224 osd_trans_stop_cb(oh, rc);
2225 /* hook functions might modify th_sync */
2226 hdl->h_sync = th->th_sync;
2228 oh->ot_handle = NULL;
2229 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2231 CERROR("%s: failed to stop transaction: rc = %d\n",
2232 osd_name(osd), rc2);
2236 /* We preserve the origin behavior of ignoring any
2237 * failures with the underlying punch / truncate
2238 * operation. We do record for debugging if an error
2239 * does occur in the lctl dk logs.
2241 rc2 = osd_process_truncates(env, &truncates);
2243 CERROR("%s: failed truncate process: rc = %d\n",
2244 osd_name(osd), rc2);
2246 osd_trans_stop_cb(oh, th->th_result);
2250 osd_trunc_unlock_all(env, &truncates);
2252 /* inform the quota slave device that the transaction is stopping */
2253 qsd_op_end(env, qsd, qtrans);
2256 * as we want IO to journal and data IO be concurrent, we don't block
2257 * awaiting data IO completion in osd_do_bio(), instead we wait here
2258 * once transaction is submitted to the journal. all reqular requests
2259 * don't do direct IO (except read/write), thus this wait_event becomes
2262 * IMPORTANT: we have to wait till any IO submited by the thread is
2263 * completed otherwise iobuf may be corrupted by different request
2265 wait_event(iobuf->dr_wait,
2266 atomic_read(&iobuf->dr_numreqs) == 0);
2269 rc = iobuf->dr_error;
2271 osd_fini_iobuf(osd, iobuf);
2273 if (unlikely(remove_agents != 0))
2274 osd_process_scheduled_agent_removals(env, osd);
2276 LASSERT(oti->oti_ins_cache_depth > 0);
2277 oti->oti_ins_cache_depth--;
2278 /* reset OI cache for safety */
2279 if (oti->oti_ins_cache_depth == 0)
2280 oti->oti_ins_cache_used = 0;
2282 sb_end_write(osd_sb(osd));
2287 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2289 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2292 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2293 LASSERT(&dcb->dcb_func != NULL);
2294 if (dcb->dcb_flags & DCB_TRANS_STOP)
2295 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2297 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2302 struct osd_delayed_iput_work {
2303 struct work_struct diw_work;
2304 struct inode *diw_inode;
2307 static void osd_delayed_iput_fn(struct work_struct *work)
2309 struct osd_delayed_iput_work *diwork;
2310 struct inode *inode;
2312 diwork = container_of(work, struct osd_delayed_iput_work, diw_work);
2313 inode = diwork->diw_inode;
2314 CDEBUG(D_INODE, "%s: delayed iput (ino=%lu)\n",
2315 inode->i_sb->s_id, inode->i_ino);
2317 OBD_FREE_PTR(diwork);
2320 noinline void osd_delayed_iput(struct inode *inode,
2321 struct osd_delayed_iput_work *diwork)
2326 INIT_WORK(&diwork->diw_work, osd_delayed_iput_fn);
2327 diwork->diw_inode = inode;
2328 queue_work(LDISKFS_SB(inode->i_sb)->s_misc_wq,
2334 * Called just before object is freed. Releases all resources except for
2335 * object itself (that is released by osd_object_free()).
2337 * Concurrency: no concurrent access is possible that late in object
2340 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2342 struct osd_object *obj = osd_obj(l);
2343 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2344 struct inode *inode = obj->oo_inode;
2345 struct osd_delayed_iput_work *diwork = NULL;
2350 LINVRNT(osd_invariant(obj));
2353 * If object is unlinked remove fid->ino mapping from object index.
2356 osd_index_fini(obj);
2361 if (inode->i_blocks > ldiskfs_delayed_unlink_blocks)
2362 OBD_ALLOC(diwork, sizeof(*diwork));
2364 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2365 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2367 uid = i_uid_read(inode);
2368 gid = i_gid_read(inode);
2369 projid = i_projid_read(inode);
2371 obj->oo_inode = NULL;
2372 osd_delayed_iput(inode, diwork);
2374 /* do not rebalance quota if the caller needs to release memory
2375 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2376 * transaction and risk to deadlock - LU-12178 */
2377 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2380 if (!obj->oo_header && qsd) {
2381 struct osd_thread_info *info = osd_oti_get(env);
2382 struct lquota_id_info *qi = &info->oti_qi;
2384 /* Release granted quota to master if necessary */
2385 qi->lqi_id.qid_uid = uid;
2386 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2388 qi->lqi_id.qid_uid = gid;
2389 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2391 qi->lqi_id.qid_uid = projid;
2392 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2397 * Concurrency: ->loo_object_release() is called under site spin-lock.
2399 static void osd_object_release(const struct lu_env *env,
2400 struct lu_object *l)
2402 struct osd_object *o = osd_obj(l);
2405 * nobody should be releasing a non-destroyed object with nlink=0
2406 * the API allows this, but ldiskfs doesn't like and then report
2407 * this inode as deleted
2409 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2410 o->oo_inode->i_nlink == 0));
2414 * Concurrency: shouldn't matter.
2416 static int osd_object_print(const struct lu_env *env, void *cookie,
2417 lu_printer_t p, const struct lu_object *l)
2419 struct osd_object *o = osd_obj(l);
2420 struct iam_descr *d;
2422 if (o->oo_dir != NULL)
2423 d = o->oo_dir->od_container.ic_descr;
2426 return (*p)(env, cookie,
2427 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2429 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2430 o->oo_inode ? o->oo_inode->i_generation : 0,
2431 d ? d->id_ops->id_name : "plain");
2435 * Concurrency: shouldn't matter.
2437 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2438 struct obd_statfs *sfs, struct obd_statfs_info *info)
2440 struct osd_device *osd = osd_dt_dev(d);
2441 struct super_block *sb = osd_sb(osd);
2442 struct kstatfs *ksfs;
2446 if (unlikely(osd->od_mnt == NULL))
2447 return -EINPROGRESS;
2449 /* osd_lproc.c call this without env, allocate ksfs for that case */
2450 if (unlikely(env == NULL)) {
2451 OBD_ALLOC_PTR(ksfs);
2455 ksfs = &osd_oti_get(env)->oti_ksfs;
2458 result = sb->s_op->statfs(sb->s_root, ksfs);
2462 statfs_pack(sfs, ksfs);
2463 if (unlikely(sb->s_flags & SB_RDONLY))
2464 sfs->os_state |= OS_STATFS_READONLY;
2466 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2468 if (ldiskfs_has_feature_extents(sb))
2469 sfs->os_maxbytes = sb->s_maxbytes;
2471 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2474 * Reserve some space so to avoid fragmenting the filesystem too much.
2475 * Fragmentation not only impacts performance, but can also increase
2476 * metadata overhead significantly, causing grant calculation to be
2479 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2481 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2482 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2483 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2484 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2486 sfs->os_blocks -= reserved;
2487 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2488 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2491 if (unlikely(env == NULL))
2497 * Estimate space needed for file creations. We assume the largest filename
2498 * which is 2^64 - 1, hence a filename of 20 chars.
2499 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2501 #ifdef __LDISKFS_DIR_REC_LEN
2502 # define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2503 #elif defined LDISKFS_DIR_REC_LEN_WITH_DIR
2504 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20, NULL)
2506 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2510 * Concurrency: doesn't access mutable data.
2512 static void osd_conf_get(const struct lu_env *env,
2513 const struct dt_device *dev,
2514 struct dt_device_param *param)
2516 struct osd_device *d = osd_dt_dev(dev);
2517 struct super_block *sb = osd_sb(d);
2518 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2523 * XXX should be taken from not-yet-existing fs abstraction layer.
2525 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2526 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2527 param->ddp_symlink_max = sb->s_blocksize;
2528 param->ddp_mount_type = LDD_MT_LDISKFS;
2529 if (ldiskfs_has_feature_extents(sb))
2530 param->ddp_maxbytes = sb->s_maxbytes;
2532 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2534 * inode are statically allocated, so per-inode space consumption
2535 * is the space consumed by the directory entry
2537 param->ddp_inodespace = PER_OBJ_USAGE;
2539 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2540 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2541 * maximum length to not under-count the actual number of extents
2542 * needed for writing a file if there are sub-optimal block allocations.
2544 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2545 /* worst-case extent insertion metadata overhead */
2546 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2547 param->ddp_mntopts = 0;
2548 if (test_opt(sb, XATTR_USER))
2549 param->ddp_mntopts |= MNTOPT_USERXATTR;
2550 if (test_opt(sb, POSIX_ACL))
2551 param->ddp_mntopts |= MNTOPT_ACL;
2554 * LOD might calculate the max stripe count based on max_ea_size,
2555 * so we need take account in the overhead as well,
2556 * xattr_header + magic + xattr_entry_head
2558 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2559 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2561 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2562 if (ldiskfs_has_feature_ea_inode(sb))
2563 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2567 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2569 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2570 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2573 * Preferred RPC size for efficient disk IO. 4MB shows good
2574 * all-around performance for ldiskfs, but use bigalloc chunk size
2575 * by default if larger.
2577 #if defined(LDISKFS_CLUSTER_SIZE)
2578 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2579 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2582 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2584 param->ddp_t10_cksum_type = 0;
2586 unsigned short interval = blk_integrity_interval(bi);
2587 name = blk_integrity_name(bi);
2597 if (strncmp(name, "T10-DIF-TYPE",
2598 sizeof("T10-DIF-TYPE") - 1) == 0) {
2599 /* also skip "1/2/3-" at end */
2600 const int type_off = sizeof("T10-DIF-TYPE.");
2601 char type_number = name[type_off - 2];
2603 if (interval != 512 && interval != 4096) {
2604 CERROR("%s: unsupported T10PI sector size %u\n",
2605 d->od_svname, interval);
2608 switch (type_number) {
2610 d->od_t10_type = OSD_T10_TYPE1;
2613 d->od_t10_type = OSD_T10_TYPE2;
2616 d->od_t10_type = OSD_T10_TYPE3;
2619 CERROR("%s: unsupported T10PI type %s\n",
2620 d->od_svname, name);
2623 if (strcmp(name + type_off, "CRC") == 0) {
2624 d->od_t10_type |= OSD_T10_TYPE_CRC;
2625 param->ddp_t10_cksum_type = interval == 512 ?
2626 OBD_CKSUM_T10CRC512 :
2628 } else if (strcmp(name + type_off, "IP") == 0) {
2629 d->od_t10_type |= OSD_T10_TYPE_IP;
2630 param->ddp_t10_cksum_type = interval == 512 ?
2631 OBD_CKSUM_T10IP512 :
2634 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2635 d->od_svname, name);
2640 CERROR("%s: unsupported T10PI type '%s'\n",
2641 d->od_svname, name);
2646 param->ddp_has_lseek_data_hole = true;
2649 static struct vfsmount *osd_mnt_get(const struct dt_device *d)
2651 return osd_dt_dev(d)->od_mnt;
2655 * Concurrency: shouldn't matter.
2657 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2660 struct super_block *s = osd_sb(osd_dt_dev(d));
2663 down_read(&s->s_umount);
2664 rc = s->s_op->sync_fs(s, 1);
2665 up_read(&s->s_umount);
2667 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2674 * Start commit for OSD device.
2676 * An implementation of dt_commit_async method for OSD device.
2677 * Asychronously starts underlayng fs sync and thereby a transaction
2680 * \param env environment
2681 * \param d dt device
2683 * \see dt_device_operations
2685 static int osd_commit_async(const struct lu_env *env,
2686 struct dt_device *d)
2688 struct super_block *s = osd_sb(osd_dt_dev(d));
2693 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2694 down_read(&s->s_umount);
2695 rc = s->s_op->sync_fs(s, 0);
2696 up_read(&s->s_umount);
2702 * Concurrency: shouldn't matter.
2704 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2706 struct super_block *sb = osd_sb(osd_dt_dev(d));
2707 struct block_device *dev = sb->s_bdev;
2708 int rc = -EOPNOTSUPP;
2712 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2713 osd_dt_dev(d)->od_svname, (long)dev, rc);
2719 * Note: we do not count into QUOTA here.
2720 * If we mount with --data_journal we may need more.
2722 const int osd_dto_credits_noquota[DTO_NR] = {
2725 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2726 * SINGLEDATA_TRANS_BLOCKS(8)
2727 * XXX Note: maybe iam need more, since iam have more level than
2730 [DTO_INDEX_INSERT] = 16,
2733 * just modify a single entry, probably merge few within a block
2735 [DTO_INDEX_DELETE] = 1,
2739 [DTO_INDEX_UPDATE] = 16,
2741 * 4(inode, inode bits, groups, GDT)
2742 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2744 [DTO_OBJECT_CREATE] = 4,
2746 * 4(inode, inode bits, groups, GDT)
2747 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2749 [DTO_OBJECT_DELETE] = 4,
2751 * Attr set credits (inode)
2753 [DTO_ATTR_SET_BASE] = 1,
2755 * Xattr set. The same as xattr of EXT3.
2756 * DATA_TRANS_BLOCKS(14)
2757 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2758 * are also counted in. Do not know why?
2760 [DTO_XATTR_SET] = 14,
2762 * credits for inode change during write.
2764 [DTO_WRITE_BASE] = 3,
2766 * credits for single block write.
2768 [DTO_WRITE_BLOCK] = 14,
2770 * Attr set credits for chown.
2771 * This is extra credits for setattr, and it is null without quota
2773 [DTO_ATTR_SET_CHOWN] = 0
2776 /* reserve or free quota for some operation */
2777 static int osd_reserve_or_free_quota(const struct lu_env *env,
2778 struct dt_device *dev,
2779 struct lquota_id_info *qi)
2781 struct osd_device *osd = osd_dt_dev(dev);
2782 struct qsd_instance *qsd = NULL;
2788 qsd = osd->od_quota_slave_dt;
2790 qsd = osd->od_quota_slave_md;
2792 rc = qsd_reserve_or_free_quota(env, qsd, qi);
2796 static const struct dt_device_operations osd_dt_ops = {
2797 .dt_root_get = osd_root_get,
2798 .dt_statfs = osd_statfs,
2799 .dt_trans_create = osd_trans_create,
2800 .dt_trans_start = osd_trans_start,
2801 .dt_trans_stop = osd_trans_stop,
2802 .dt_trans_cb_add = osd_trans_cb_add,
2803 .dt_conf_get = osd_conf_get,
2804 .dt_mnt_get = osd_mnt_get,
2805 .dt_sync = osd_sync,
2807 .dt_commit_async = osd_commit_async,
2808 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
2811 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2814 struct osd_object *obj = osd_dt_obj(dt);
2815 struct osd_thread_info *oti = osd_oti_get(env);
2817 LINVRNT(osd_invariant(obj));
2819 LASSERT(obj->oo_owner != env);
2820 down_read_nested(&obj->oo_sem, role);
2822 LASSERT(obj->oo_owner == NULL);
2826 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2829 struct osd_object *obj = osd_dt_obj(dt);
2830 struct osd_thread_info *oti = osd_oti_get(env);
2832 LINVRNT(osd_invariant(obj));
2834 LASSERT(obj->oo_owner != env);
2835 down_write_nested(&obj->oo_sem, role);
2837 LASSERT(obj->oo_owner == NULL);
2838 obj->oo_owner = env;
2842 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2844 struct osd_object *obj = osd_dt_obj(dt);
2845 struct osd_thread_info *oti = osd_oti_get(env);
2847 LINVRNT(osd_invariant(obj));
2849 LASSERT(oti->oti_r_locks > 0);
2851 up_read(&obj->oo_sem);
2854 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2856 struct osd_object *obj = osd_dt_obj(dt);
2857 struct osd_thread_info *oti = osd_oti_get(env);
2859 LINVRNT(osd_invariant(obj));
2861 LASSERT(obj->oo_owner == env);
2862 LASSERT(oti->oti_w_locks > 0);
2864 obj->oo_owner = NULL;
2865 up_write(&obj->oo_sem);
2868 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2870 struct osd_object *obj = osd_dt_obj(dt);
2872 LINVRNT(osd_invariant(obj));
2874 return obj->oo_owner == env;
2877 static void osd_inode_getattr(const struct lu_env *env,
2878 struct inode *inode, struct lu_attr *attr)
2880 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2881 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2882 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2883 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2885 attr->la_atime = inode->i_atime.tv_sec;
2886 attr->la_mtime = inode->i_mtime.tv_sec;
2887 attr->la_ctime = inode->i_ctime.tv_sec;
2888 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2889 attr->la_mode = inode->i_mode;
2890 attr->la_size = i_size_read(inode);
2891 attr->la_blocks = inode->i_blocks;
2892 attr->la_uid = i_uid_read(inode);
2893 attr->la_gid = i_gid_read(inode);
2894 attr->la_projid = i_projid_read(inode);
2895 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2896 attr->la_nlink = inode->i_nlink;
2897 attr->la_rdev = inode->i_rdev;
2898 attr->la_blksize = 1 << inode->i_blkbits;
2899 attr->la_blkbits = inode->i_blkbits;
2901 * Ext4 did not transfer inherit flags from raw inode
2902 * to inode flags, and ext4 internally test raw inode
2903 * @i_flags directly. Instead of patching ext4, we do it here.
2905 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2906 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2909 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2912 struct osd_object *obj = osd_dt_obj(dt);
2913 const struct dt_it_ops *iops;
2919 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2920 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2922 /* directory not initialized yet */
2923 if (!dt->do_index_ops) {
2928 spin_lock(&obj->oo_guard);
2929 *count = atomic_read(&obj->oo_dirent_count);
2930 if (*count == LU_DIRENT_COUNT_UNSET)
2931 atomic_set(&obj->oo_dirent_count, 0);
2932 spin_unlock(&obj->oo_guard);
2933 if (*count != LU_DIRENT_COUNT_UNSET)
2937 iops = &dt->do_index_ops->dio_it;
2938 it = iops->init(env, dt, LUDA_64BITHASH);
2940 GOTO(out, rc = PTR_ERR(it));
2942 rc = iops->load(env, it, 0);
2949 rc = iops->next(env, it);
2951 for (; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2955 if (iops->key_size(env, it) == 0)
2960 if (rc == 1 || rc == -ESTALE)
2964 iops->fini(env, it);
2966 /* If counting dirents failed, use the current count (if any).
2968 * At worst this means the directory will not be split until the
2969 * count can be completed successfully (remount or oo_dirent_count
2970 * incremented by adding new entries). This avoids re-walking
2971 * the whole directory on each access and hitting the same error.
2973 if (rc && *count == 0)
2974 *count = LU_DIRENT_COUNT_UNSET;
2975 atomic_set(&obj->oo_dirent_count, *count);
2979 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2980 struct lu_attr *attr)
2982 struct osd_object *obj = osd_dt_obj(dt);
2985 if (unlikely(!dt_object_exists(dt)))
2987 if (unlikely(obj->oo_destroyed))
2990 LASSERT(!dt_object_remote(dt));
2991 LINVRNT(osd_invariant(obj));
2993 spin_lock(&obj->oo_guard);
2994 osd_inode_getattr(env, obj->oo_inode, attr);
2995 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2996 attr->la_valid |= LA_FLAGS;
2997 attr->la_flags |= LUSTRE_ORPHAN_FL;
2999 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
3000 attr->la_valid |= LA_FLAGS;
3001 attr->la_flags |= LUSTRE_ENCRYPT_FL;
3003 spin_unlock(&obj->oo_guard);
3005 if (S_ISDIR(obj->oo_inode->i_mode) &&
3006 (attr->la_valid & LA_DIRENT_CNT) &&
3007 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
3008 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
3010 attr->la_valid &= ~LA_DIRENT_CNT;
3015 static int osd_declare_attr_qid(const struct lu_env *env,
3016 struct osd_object *obj,
3017 struct osd_thandle *oh, long long bspace,
3018 qid_t old_id, qid_t new_id, bool enforce,
3022 struct osd_thread_info *info = osd_oti_get(env);
3023 struct lquota_id_info *qi = &info->oti_qi;
3025 qi->lqi_type = type;
3026 /* inode accounting */
3027 qi->lqi_is_blk = false;
3029 /* one more inode for the new id ... */
3030 qi->lqi_id.qid_uid = new_id;
3032 /* Reserve credits for the new id */
3033 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
3034 if (rc == -EDQUOT || rc == -EINPROGRESS)
3039 /* and one less inode for the current id */
3040 qi->lqi_id.qid_uid = old_id;
3042 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3043 if (rc == -EDQUOT || rc == -EINPROGRESS)
3048 /* block accounting */
3049 qi->lqi_is_blk = true;
3051 /* more blocks for the new id ... */
3052 qi->lqi_id.qid_uid = new_id;
3053 qi->lqi_space = bspace;
3055 * Credits for the new uid has been reserved, re-use "obj"
3056 * to save credit reservation.
3058 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3059 if (rc == -EDQUOT || rc == -EINPROGRESS)
3064 /* and finally less blocks for the current uid */
3065 qi->lqi_id.qid_uid = old_id;
3066 qi->lqi_space = -bspace;
3067 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3068 if (rc == -EDQUOT || rc == -EINPROGRESS)
3074 static int osd_declare_attr_set(const struct lu_env *env,
3075 struct dt_object *dt,
3076 const struct lu_attr *attr,
3077 struct thandle *handle)
3079 struct osd_thandle *oh;
3080 struct osd_object *obj;
3089 LASSERT(dt != NULL);
3090 LASSERT(handle != NULL);
3092 obj = osd_dt_obj(dt);
3093 LASSERT(osd_invariant(obj));
3095 oh = container_of(handle, struct osd_thandle, ot_super);
3096 LASSERT(oh->ot_handle == NULL);
3098 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
3099 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3101 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
3102 osd_dto_credits_noquota[DTO_XATTR_SET]);
3104 if (attr == NULL || obj->oo_inode == NULL)
3107 bspace = obj->oo_inode->i_blocks << 9;
3108 bspace = toqb(bspace);
3111 * Changing ownership is always preformed by super user, it should not
3112 * fail with EDQUOT unless required explicitly.
3114 * We still need to call the osd_declare_qid() to calculate the journal
3115 * credits for updating quota accounting files and to trigger quota
3116 * space adjustment once the operation is completed.
3118 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
3120 uid = i_uid_read(obj->oo_inode);
3121 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
3122 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
3123 attr->la_uid, enforce, USRQUOTA);
3127 gid = i_gid_read(obj->oo_inode);
3128 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
3129 attr->la_uid, gid, attr->la_gid);
3130 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
3131 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
3132 attr->la_gid, enforce, GRPQUOTA);
3137 #ifdef HAVE_PROJECT_QUOTA
3138 if (attr->la_valid & LA_PROJID) {
3139 __u32 projid = i_projid_read(obj->oo_inode);
3141 enforce = (attr->la_valid & LA_PROJID) &&
3142 (attr->la_projid != projid);
3143 rc = osd_declare_attr_qid(env, obj, oh, bspace,
3144 (qid_t)projid, (qid_t)attr->la_projid,
3150 /* punch must be aware we are dealing with an encrypted file */
3151 if (attr->la_valid & LA_FLAGS && attr->la_flags & LUSTRE_ENCRYPT_FL)
3152 obj->oo_lma_flags |= LUSTRE_ENCRYPT_FL;
3157 static int osd_inode_setattr(const struct lu_env *env,
3158 struct inode *inode, const struct lu_attr *attr)
3160 __u64 bits = attr->la_valid;
3162 /* Only allow set size for regular file */
3163 if (!S_ISREG(inode->i_mode))
3164 bits &= ~(LA_SIZE | LA_BLOCKS);
3169 if (bits & LA_ATIME)
3170 inode->i_atime = osd_inode_time(inode, attr->la_atime);
3171 if (bits & LA_CTIME)
3172 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
3173 if (bits & LA_MTIME)
3174 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
3175 if (bits & LA_SIZE) {
3176 spin_lock(&inode->i_lock);
3177 LDISKFS_I(inode)->i_disksize = attr->la_size;
3178 i_size_write(inode, attr->la_size);
3179 spin_unlock(&inode->i_lock);
3183 * OSD should not change "i_blocks" which is used by quota.
3184 * "i_blocks" should be changed by ldiskfs only.
3187 inode->i_mode = (inode->i_mode & S_IFMT) |
3188 (attr->la_mode & ~S_IFMT);
3190 i_uid_write(inode, attr->la_uid);
3192 i_gid_write(inode, attr->la_gid);
3193 if (bits & LA_PROJID)
3194 i_projid_write(inode, attr->la_projid);
3195 if (bits & LA_NLINK)
3196 set_nlink(inode, attr->la_nlink);
3198 inode->i_rdev = attr->la_rdev;
3200 if (bits & LA_FLAGS) {
3201 /* always keep S_NOCMTIME */
3202 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
3204 #if defined(S_ENCRYPTED)
3205 /* Always remove S_ENCRYPTED, because ldiskfs must not be
3206 * aware of encryption status. It is just stored into LMA
3207 * so that it can be forwared to client side.
3209 inode->i_flags &= ~S_ENCRYPTED;
3212 * Ext4 did not transfer inherit flags from
3213 * @inode->i_flags to raw inode i_flags when writing
3214 * flags, we do it explictly here.
3216 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
3217 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
3219 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
3224 #ifdef HAVE_PROJECT_QUOTA
3225 static int osd_transfer_project(struct inode *inode, __u32 projid,
3226 struct thandle *handle)
3228 struct super_block *sb = inode->i_sb;
3229 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3232 struct ldiskfs_iloc iloc;
3233 struct ldiskfs_inode *raw_inode;
3234 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3236 if (!ldiskfs_has_feature_project(sb)) {
3237 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3238 == LDISKFS_DEF_PROJID);
3239 if (projid != LDISKFS_DEF_PROJID)
3245 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3248 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3249 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3252 err = ldiskfs_get_inode_loc(inode, &iloc);
3256 raw_inode = ldiskfs_raw_inode(&iloc);
3257 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3258 struct osd_thandle *oh = container_of(handle,
3262 * try to expand inode size automatically.
3264 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3265 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3273 dquot_initialize(inode);
3274 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3275 if (transfer_to[PRJQUOTA]) {
3276 lock_dquot_transfer(inode);
3277 err = __dquot_transfer(inode, transfer_to);
3278 unlock_dquot_transfer(inode);
3279 dqput(transfer_to[PRJQUOTA]);
3288 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3289 struct thandle *handle)
3293 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3294 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3298 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3299 inode->i_ino, i_uid_read(inode), attr->la_uid,
3300 i_gid_read(inode), attr->la_gid);
3302 dquot_initialize(inode);
3304 if (attr->la_valid & LA_UID)
3305 iattr.ia_valid |= ATTR_UID;
3306 if (attr->la_valid & LA_GID)
3307 iattr.ia_valid |= ATTR_GID;
3308 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3309 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3311 lock_dquot_transfer(inode);
3312 rc = osd_dquot_transfer(&nop_mnt_idmap, inode, &iattr);
3313 unlock_dquot_transfer(inode);
3315 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3316 osd_ino2name(inode), rc);
3321 /* Handle project id transfer here properly */
3322 if (attr->la_valid & LA_PROJID &&
3323 attr->la_projid != i_projid_read(inode)) {
3324 if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
3326 #ifdef HAVE_PROJECT_QUOTA
3327 rc = osd_transfer_project(inode, attr->la_projid, handle);
3332 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3333 osd_ino2name(inode), rc);
3340 static int osd_attr_set(const struct lu_env *env,
3341 struct dt_object *dt,
3342 const struct lu_attr *attr,
3343 struct thandle *handle)
3345 struct osd_thread_info *info = osd_oti_get(env);
3346 struct osd_object *obj = osd_dt_obj(dt);
3347 struct osd_device *osd = osd_obj2dev(obj);
3348 struct inode *inode;
3351 if (!dt_object_exists(dt))
3354 LASSERT(handle != NULL);
3355 LASSERT(!dt_object_remote(dt));
3356 LASSERT(osd_invariant(obj));
3358 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3360 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3361 !osd_obj2dev(obj)->od_is_ost) {
3362 struct osd_thread_info *oti = osd_oti_get(env);
3363 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3364 struct lu_fid *fid1 = &oti->oti_fid;
3365 struct osd_inode_id *id = &oti->oti_id;
3366 struct iam_path_descr *ipd;
3367 struct iam_container *bag;
3368 struct osd_thandle *oh;
3371 fid_cpu_to_be(fid1, fid0);
3372 memset(id, 1, sizeof(*id));
3373 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3374 fid0)->oi_dir.od_container;
3375 ipd = osd_idx_ipd_get(env, bag);
3376 if (unlikely(ipd == NULL))
3379 oh = container_of(handle, struct osd_thandle, ot_super);
3380 rc = iam_update(oh->ot_handle, bag,
3381 (const struct iam_key *)fid1,
3382 (const struct iam_rec *)id, ipd);
3383 osd_ipd_put(env, bag, ipd);
3384 return(rc > 0 ? 0 : rc);
3387 inode = obj->oo_inode;
3389 rc = osd_quota_transfer(inode, attr, handle);
3393 spin_lock(&obj->oo_guard);
3394 rc = osd_inode_setattr(env, inode, attr);
3395 spin_unlock(&obj->oo_guard);
3399 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3401 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3403 if (!(attr->la_valid & LA_FLAGS))
3406 /* If setting LUSTRE_ENCRYPT_FL on an OST object, also set a dummy
3407 * enc ctx xattr, with 2 benefits:
3408 * - setting the LL_XATTR_NAME_ENCRYPTION_CONTEXT xattr internally sets
3409 * the LDISKFS_ENCRYPT_FL flag on the on-disk inode;
3410 * - it makes e2fsprogs happy to see an enc ctx for an inode that has
3411 * the LDISKFS_ENCRYPT_FL flag
3412 * We do not need the actual encryption context on OST objects, it is
3413 * only stored on MDT inodes, at file creation time.
3415 if (!(LDISKFS_I(obj->oo_inode)->i_flags & LDISKFS_ENCRYPT_FL) &&
3416 attr->la_flags & LUSTRE_ENCRYPT_FL && osd->od_is_ost &&
3417 !CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) {
3420 /* use a dummy enc ctx, fine with e2fsprogs */
3421 buf.lb_buf = "\xFF";
3423 rc = osd_xattr_set(env, dt, &buf,
3424 LL_XATTR_NAME_ENCRYPTION_CONTEXT,
3427 CWARN("%s: set "DFID" enc ctx failed: rc = %d\n",
3428 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3432 /* Let's check if there are extra flags need to be set into LMA */
3433 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3434 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3436 LASSERT(!obj->oo_pfid_in_lma);
3438 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3439 &info->oti_ost_attrs);
3443 if ((lma->lma_incompat & lustre_to_lma_flags(attr->la_flags)) ==
3444 lustre_to_lma_flags(attr->la_flags))
3445 /* if lma incompat already has the flags,
3446 * save a useless call to xattr_set
3450 lma->lma_incompat |=
3451 lustre_to_lma_flags(attr->la_flags);
3452 lustre_lma_swab(lma);
3454 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3456 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3457 lma, sizeof(*lma), XATTR_REPLACE);
3459 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3460 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3461 lma->lma_incompat, rc);
3464 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3465 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3472 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3473 struct osd_object *obj,
3474 const char *name, const int namelen)
3476 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3479 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3480 umode_t mode, struct dt_allocation_hint *hint,
3481 struct thandle *th, struct lu_attr *attr)
3484 struct osd_device *osd = osd_obj2dev(obj);
3485 struct osd_thandle *oth;
3486 struct dt_object *parent = NULL;
3487 struct inode *inode;
3488 struct iattr iattr = {
3489 .ia_valid = ATTR_UID | ATTR_GID |
3490 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3491 .ia_ctime.tv_sec = attr->la_ctime,
3492 .ia_mtime.tv_sec = attr->la_mtime,
3493 .ia_atime.tv_sec = attr->la_atime,
3494 .ia_uid = GLOBAL_ROOT_UID,
3495 .ia_gid = GLOBAL_ROOT_GID,
3497 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3499 if (attr->la_valid & LA_UID)
3500 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3501 if (attr->la_valid & LA_GID)
3502 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3504 LINVRNT(osd_invariant(obj));
3505 LASSERT(obj->oo_inode == NULL);
3506 LASSERT(obj->oo_hl_head == NULL);
3508 if (S_ISDIR(mode) && ldiskfs_pdo) {
3510 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3511 if (obj->oo_hl_head == NULL)
3515 oth = container_of(th, struct osd_thandle, ot_super);
3516 LASSERT(oth->ot_handle->h_transaction != NULL);
3518 if (hint != NULL && hint->dah_parent != NULL &&
3519 !dt_object_remote(hint->dah_parent))
3520 parent = hint->dah_parent;
3522 /* if a time component is not valid set it to UTIME_OMIT */
3523 if (!(attr->la_valid & LA_CTIME))
3524 iattr.ia_ctime = omit;
3525 if (!(attr->la_valid & LA_MTIME))
3526 iattr.ia_mtime = omit;
3527 if (!(attr->la_valid & LA_ATIME))
3528 iattr.ia_atime = omit;
3530 inode = ldiskfs_create_inode(oth->ot_handle,
3531 parent ? osd_dt_obj(parent)->oo_inode :
3532 osd_sb(osd)->s_root->d_inode,
3534 if (!IS_ERR(inode)) {
3535 /* Do not update file c/mtime in ldiskfs. */
3536 inode->i_flags |= S_NOCMTIME;
3539 * For new created object, it must be consistent,
3540 * and it is unnecessary to scrub against it.
3542 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3544 obj->oo_inode = inode;
3547 if (obj->oo_hl_head != NULL) {
3548 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3549 obj->oo_hl_head = NULL;
3551 result = PTR_ERR(inode);
3553 LINVRNT(osd_invariant(obj));
3561 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3562 struct lu_attr *attr,
3563 struct dt_allocation_hint *hint,
3564 struct dt_object_format *dof,
3568 struct osd_thandle *oth;
3569 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3571 LASSERT(S_ISDIR(attr->la_mode));
3573 oth = container_of(th, struct osd_thandle, ot_super);
3574 LASSERT(oth->ot_handle->h_transaction != NULL);
3575 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3576 atomic_set(&obj->oo_dirent_count, 0);
3577 result = osd_mkfile(info, obj, mode, hint, th, attr);
3582 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3583 struct lu_attr *attr,
3584 struct dt_allocation_hint *hint,
3585 struct dt_object_format *dof,
3589 struct osd_thandle *oth;
3590 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3592 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3594 LASSERT(S_ISREG(attr->la_mode));
3596 oth = container_of(th, struct osd_thandle, ot_super);
3597 LASSERT(oth->ot_handle->h_transaction != NULL);
3599 result = osd_mkfile(info, obj, mode, hint, th, attr);
3601 LASSERT(obj->oo_inode != NULL);
3602 if (feat->dif_flags & DT_IND_VARKEY)
3603 result = iam_lvar_create(obj->oo_inode,
3604 feat->dif_keysize_max,
3606 feat->dif_recsize_max,
3609 result = iam_lfix_create(obj->oo_inode,
3610 feat->dif_keysize_max,
3612 feat->dif_recsize_max,
3618 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3619 struct lu_attr *attr,
3620 struct dt_allocation_hint *hint,
3621 struct dt_object_format *dof,
3624 LASSERT(S_ISREG(attr->la_mode));
3625 return osd_mkfile(info, obj, (attr->la_mode &
3626 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3630 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3631 struct lu_attr *attr,
3632 struct dt_allocation_hint *hint,
3633 struct dt_object_format *dof,
3636 LASSERT(S_ISLNK(attr->la_mode));
3637 return osd_mkfile(info, obj, (attr->la_mode &
3638 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3642 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3643 struct lu_attr *attr,
3644 struct dt_allocation_hint *hint,
3645 struct dt_object_format *dof,
3648 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3651 LINVRNT(osd_invariant(obj));
3652 LASSERT(obj->oo_inode == NULL);
3653 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3654 S_ISFIFO(mode) || S_ISSOCK(mode));
3656 result = osd_mkfile(info, obj, mode, hint, th, attr);
3658 LASSERT(obj->oo_inode != NULL);
3660 * This inode should be marked dirty for i_rdev. Currently
3661 * that is done in the osd_attr_init().
3663 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3666 LINVRNT(osd_invariant(obj));
3670 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3672 struct dt_allocation_hint *hint,
3673 struct dt_object_format *dof,
3676 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3678 osd_obj_type_f result;
3694 result = osd_mk_index;
3704 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3705 struct dt_object *parent, struct dt_object *child,
3710 ah->dah_parent = parent;
3712 if (parent != NULL && !dt_object_remote(parent)) {
3713 /* will help to find FID->ino at dt_insert("..") */
3714 struct osd_object *pobj = osd_dt_obj(parent);
3716 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3720 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3721 struct lu_attr *attr, struct dt_object_format *dof,
3722 struct thandle *handle)
3724 struct inode *inode = obj->oo_inode;
3725 __u64 valid = attr->la_valid;
3728 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3730 if (dof->dof_type != DFT_NODE)
3731 attr->la_valid &= ~LA_RDEV;
3732 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3733 attr->la_valid &= ~LA_ATIME;
3734 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3735 attr->la_valid &= ~LA_CTIME;
3736 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3737 attr->la_valid &= ~LA_MTIME;
3739 result = osd_quota_transfer(inode, attr, handle);
3743 if (attr->la_valid != 0) {
3744 result = osd_inode_setattr(info->oti_env, inode, attr);
3746 * The osd_inode_setattr() should always succeed here. The
3747 * only error that could be returned is EDQUOT when we are
3748 * trying to change the UID or GID of the inode. However, this
3749 * should not happen since quota enforcement is no longer
3750 * enabled on ldiskfs (lquota takes care of it).
3752 LASSERTF(result == 0, "%d\n", result);
3753 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3756 attr->la_valid = valid;
3760 * Helper function for osd_create()
3762 * \retval 0, on success
3764 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3765 struct lu_attr *attr, struct dt_allocation_hint *hint,
3766 struct dt_object_format *dof, struct thandle *th)
3771 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3773 /* we drop umask so that permissions we pass are not affected */
3774 umask = current->fs->umask;
3775 current->fs->umask = 0;
3777 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3779 if (likely(obj->oo_inode != NULL)) {
3780 LASSERT(obj->oo_inode->i_state & I_NEW);
3783 * Unlock the inode before attr initialization to avoid
3784 * unnecessary dqget operations. LU-6378
3786 unlock_new_inode(obj->oo_inode);
3789 if (likely(result == 0)) {
3790 osd_attr_init(info, obj, attr, dof, th);
3791 osd_object_init0(obj);
3794 /* restore previous umask value */
3795 current->fs->umask = umask;
3797 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3803 * Helper function for osd_create()
3805 * \retval 0, on success
3807 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3808 const struct lu_fid *fid, struct thandle *th)
3810 struct osd_thread_info *info = osd_oti_get(env);
3811 struct osd_inode_id *id = &info->oti_id;
3812 struct osd_device *osd = osd_obj2dev(obj);
3813 struct osd_thandle *oh;
3816 LASSERT(obj->oo_inode != NULL);
3818 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3821 oh = container_of(th, struct osd_thandle, ot_super);
3822 LASSERT(oh->ot_handle);
3823 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3825 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3826 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_REUSE) && osd->od_is_ost &&
3828 struct lu_fid tfid = *fid;
3831 osd_oi_insert(info, osd, &tfid, id, oh->ot_handle,
3832 OI_CHECK_FLD, NULL);
3833 /* clear NOSCRUB flag so that it can be scrubbed immediately */
3834 ldiskfs_clear_inode_state(obj->oo_inode,
3835 LDISKFS_STATE_LUSTRE_NOSCRUB);
3838 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3839 OI_CHECK_FLD, NULL);
3841 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3842 struct lu_fid next_fid = *fid;
3844 /* insert next object in advance, and map to the same inode */
3846 if (next_fid.f_oid != 0) {
3847 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3848 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3849 OI_CHECK_FLD, NULL);
3850 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3854 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3859 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3860 u64 seq, struct lu_seq_range *range)
3862 struct seq_server_site *ss = osd_seq_site(osd);
3864 if (fid_seq_is_idif(seq)) {
3865 fld_range_set_ost(range);
3866 range->lsr_index = idif_ost_idx(seq);
3870 if (!fid_seq_in_fldb(seq)) {
3871 fld_range_set_mdt(range);
3874 * FIXME: If ss is NULL, it suppose not get lsr_index
3877 range->lsr_index = ss->ss_node_id;
3881 LASSERT(ss != NULL);
3882 fld_range_set_any(range);
3883 /* OSD will only do local fld lookup */
3884 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3887 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3888 struct lu_attr *attr,
3889 struct dt_allocation_hint *hint,
3890 struct dt_object_format *dof,
3891 struct thandle *handle)
3893 struct osd_thandle *oh;
3894 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3900 LASSERT(handle != NULL);
3902 oh = container_of(handle, struct osd_thandle, ot_super);
3903 LASSERT(oh->ot_handle == NULL);
3905 if (osd_tx_was_declared(env, oh, dt, DTO_OBJECT_CREATE, 0))
3909 * EA object consumes more credits than regular object: osd_mk_index
3910 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3911 * leaf_node, could involves the block, block bitmap, groups, GDT
3912 * change for each block, so add 4 * 2 credits in that case.
3914 * The default ACL initialization may consume an additional 16 blocks
3916 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3917 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3920 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3921 * credits for possible default ACL creation in new inode
3923 if (hint && hint->dah_acl_len)
3924 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3925 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3926 sb->s_blocksize_bits);
3928 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3931 * Reuse idle OI block may cause additional one OI block
3934 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3935 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3936 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3937 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3938 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3940 /* will help to find FID->ino mapping at dt_insert() */
3941 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3949 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3950 attr->la_projid, 1, oh, osd_dt_obj(dt),
3951 NULL, OSD_QID_INODE);
3959 * Called to destroy on-disk representation of the object
3961 * Concurrency: must be locked
3963 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3966 struct osd_object *obj = osd_dt_obj(dt);
3967 struct inode *inode = obj->oo_inode;
3968 struct osd_thandle *oh;
3976 oh = container_of(th, struct osd_thandle, ot_super);
3977 LASSERT(oh->ot_handle == NULL);
3979 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3980 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3982 /* For removing agent entry */
3983 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3984 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3987 * Recycle idle OI leaf may cause additional three OI blocks
3990 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3991 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3992 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3993 /* one less inode */
3994 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3995 i_projid_read(inode), -1, oh, obj, NULL,
3999 /* data to be truncated */
4000 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4001 i_projid_read(inode), 0, oh, obj, NULL,
4007 * will help to find FID->ino when this object is being
4010 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
4015 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
4018 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4019 struct osd_object *obj = osd_dt_obj(dt);
4020 struct inode *inode = obj->oo_inode;
4021 struct osd_device *osd = osd_obj2dev(obj);
4022 struct osd_thandle *oh;
4027 oh = container_of(th, struct osd_thandle, ot_super);
4028 LASSERT(oh->ot_handle);
4030 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
4032 if (unlikely(fid_is_acct(fid)))
4035 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
4036 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
4038 CERROR("%s: remove agent entry "DFID": rc = %d\n",
4039 osd_name(osd), PFID(fid), result);
4042 if (S_ISDIR(inode->i_mode)) {
4043 if (inode->i_nlink > 2)
4044 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
4045 osd_name(osd), PFID(fid), inode->i_ino,
4048 spin_lock(&obj->oo_guard);
4050 spin_unlock(&obj->oo_guard);
4051 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4054 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
4056 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
4058 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
4059 result = osd_oi_delete(osd_oti_get(env), osd, fid,
4060 oh->ot_handle, OI_CHECK_FLD);
4062 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
4063 /* XXX: add to ext3 orphan list */
4064 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
4066 /* not needed in the cache anymore */
4067 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
4068 obj->oo_destroyed = 1;
4074 * Put the fid into lustre_mdt_attrs, and then place the structure
4075 * inode's ea. This fid should not be altered during the life time
4078 * \retval +ve, on success
4079 * \retval -ve, on error
4081 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
4083 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
4084 const struct lu_fid *fid, __u32 compat, __u32 incompat)
4086 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4087 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4092 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INLMA))
4095 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
4098 lustre_loa_init(loa, fid, compat, incompat);
4099 lustre_loa_swab(loa, false);
4102 * For the OST device with 256 bytes inode size by default,
4103 * the PFID EA will be stored together with LMA EA to avoid
4104 * performance trouble. Otherwise the PFID EA can be stored
4105 * independently. LU-8998
4107 if ((compat & LMAC_FID_ON_OST) &&
4108 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
4109 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4110 sizeof(*loa), XATTR_CREATE);
4112 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4113 sizeof(*lma), XATTR_CREATE);
4115 * LMA may already exist, but we need to check that all the
4116 * desired compat/incompat flags have been added.
4118 if (unlikely(rc == -EEXIST)) {
4119 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
4120 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
4124 if (rc < sizeof(*lma))
4127 lustre_loa_swab(loa, true);
4128 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
4129 ((compat == 0 && incompat == 0) ||
4130 (!(~lma->lma_compat & compat) &&
4131 !(~lma->lma_incompat & incompat))))
4134 lma->lma_self_fid = *fid;
4135 lma->lma_compat |= compat;
4136 lma->lma_incompat |= incompat;
4137 if (rc == sizeof(*lma)) {
4138 lustre_lma_swab(lma);
4139 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4140 sizeof(*lma), XATTR_REPLACE);
4142 lustre_loa_swab(loa, false);
4143 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4144 sizeof(*loa), XATTR_REPLACE);
4152 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
4153 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
4154 * To have compatilibility with 1.8 ldiskfs driver we need to have
4155 * magic number at start of fid data.
4156 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
4159 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
4160 const struct lu_fid *fid)
4162 if (!fid_is_namespace_visible(fid) ||
4163 CFS_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
4164 param->edp_magic = 0;
4168 param->edp_magic = LDISKFS_LUFID_MAGIC;
4169 param->edp_len = sizeof(struct lu_fid) + 1;
4170 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
4174 * Try to read the fid from inode ea into dt_rec.
4176 * \param fid object fid.
4178 * \retval 0 on success
4180 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
4181 __u32 ino, struct lu_fid *fid,
4182 struct osd_inode_id *id)
4184 struct osd_thread_info *info = osd_oti_get(env);
4185 struct inode *inode;
4189 osd_id_gen(id, ino, OSD_OII_NOGEN);
4190 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
4192 RETURN(PTR_ERR(inode));
4198 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
4200 struct inode *parent_dir,
4201 const struct lu_fid *dot_fid,
4202 const struct lu_fid *dot_dot_fid,
4203 struct osd_thandle *oth)
4205 struct ldiskfs_dentry_param *dot_ldp;
4206 struct ldiskfs_dentry_param *dot_dot_ldp;
4207 __u32 saved_nlink = dir->i_nlink;
4210 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
4213 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
4214 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
4216 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4217 dot_ldp->edp_magic = 0;
4219 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
4220 dir, dot_ldp, dot_dot_ldp);
4222 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
4223 * the subseqent ref_add() will increase the dir->i_nlink
4224 * as 3. That is incorrect for new created directory.
4226 * It looks like hack, because we want to make the OSD API
4227 * to be order-independent for new created directory object
4228 * between dt_insert(..) and ref_add() operations.
4230 * Here, we only restore the in-RAM dir-inode's nlink attr,
4231 * becuase if the nlink attr is not 2, then there will be
4232 * ref_add() called following the dt_insert(..), such call
4233 * will make both the in-RAM and on-disk dir-inode's nlink
4234 * attr to be set as 2. LU-7447
4236 set_nlink(dir, saved_nlink);
4241 * Create an local agent inode for remote entry
4243 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
4244 struct osd_device *osd,
4245 struct osd_object *pobj,
4246 const struct lu_fid *fid,
4250 struct osd_thread_info *info = osd_oti_get(env);
4251 struct inode *local;
4252 struct osd_thandle *oh;
4253 struct iattr iattr = {
4254 .ia_valid = ATTR_UID | ATTR_GID |
4255 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
4256 .ia_ctime.tv_nsec = UTIME_OMIT,
4257 .ia_mtime.tv_nsec = UTIME_OMIT,
4258 .ia_atime.tv_nsec = UTIME_OMIT,
4259 .ia_uid = GLOBAL_ROOT_UID,
4260 .ia_gid = GLOBAL_ROOT_GID,
4267 oh = container_of(th, struct osd_thandle, ot_super);
4268 LASSERT(oh->ot_handle->h_transaction != NULL);
4270 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
4272 if (IS_ERR(local)) {
4273 CERROR("%s: create local error %d\n", osd_name(osd),
4274 (int)PTR_ERR(local));
4279 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
4280 * correct gid on remote file, not agent here
4282 local->i_gid = current_fsgid();
4283 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4285 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4286 * That gives e2fsck something to look at and be happy, and allows
4287 * debugging if we need to determine where this symlink came from.
4289 if (S_ISLNK(type)) {
4290 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4291 ldiskfs_clear_inode_flag(local, LDISKFS_INODE_EXTENTS);
4292 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4293 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4295 i_size_write(local, rc);
4296 LDISKFS_I(local)->i_disksize = rc;
4298 unlock_new_inode(local);
4300 /* Agent inode should not have project ID */
4301 #ifdef HAVE_PROJECT_QUOTA
4302 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4303 i_projid_read(pobj->oo_inode) != 0) {
4304 rc = osd_transfer_project(local, 0, th);
4306 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4307 osd_ino2name(local), rc);
4308 RETURN(ERR_PTR(rc));
4312 /* Set special LMA flag for local agent inode */
4313 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4315 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4316 osd_name(osd), PFID(fid), rc);
4317 RETURN(ERR_PTR(rc));
4323 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode, fid,
4324 lu_object_fid(&pobj->oo_dt.do_lu),
4327 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4328 osd_name(osd), PFID(fid), rc);
4329 RETURN(ERR_PTR(rc));
4336 * when direntry is deleted, we have to take care of possible agent inode
4337 * referenced by that. unfortunately we can't do this at that point:
4338 * iget() within a running transaction leads to deadlock and we better do
4339 * not call that every delete declaration to save performance. so we put
4340 * a potention agent inode on a list and process that once the transaction
4341 * is over. Notice it's not any worse in terms of real orphans as regular
4342 * object destroy doesn't put inodes on the on-disk orphan list. this should
4343 * be addressed separately
4345 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4346 struct osd_thandle *oh,
4349 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4350 struct osd_obj_orphan *oor;
4357 oor->oor_env = (struct lu_env *)env;
4358 spin_lock(&osd->od_osfs_lock);
4359 list_add(&oor->oor_list, &osd->od_orphan_list);
4360 spin_unlock(&osd->od_osfs_lock);
4362 oh->ot_remove_agents = 1;
4368 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4369 struct osd_device *osd)
4371 struct osd_thread_info *info = osd_oti_get(env);
4372 struct osd_obj_orphan *oor, *tmp;
4373 struct osd_inode_id id;
4375 struct inode *inode;
4380 spin_lock(&osd->od_osfs_lock);
4381 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4382 if (oor->oor_env == env)
4383 list_move(&oor->oor_list, &list);
4385 spin_unlock(&osd->od_osfs_lock);
4387 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4391 list_del(&oor->oor_list);
4394 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4395 inode = osd_iget_fid(info, osd, &id, &fid);
4399 if (!osd_remote_fid(env, osd, &fid)) {
4404 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4406 mark_inode_dirty(inode);
4407 ldiskfs_journal_stop(jh);
4415 * OSD layer object create function for OST objects (b=11826).
4417 * The FID is inserted into inode xattr here.
4419 * \retval 0, on success
4420 * \retval -ve, on error
4422 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4423 struct lu_attr *attr, struct dt_allocation_hint *hint,
4424 struct dt_object_format *dof, struct thandle *th)
4426 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4427 struct osd_object *obj = osd_dt_obj(dt);
4428 struct osd_thread_info *info = osd_oti_get(env);
4429 int result, on_ost = 0;
4433 if (dt_object_exists(dt))
4436 LINVRNT(osd_invariant(obj));
4437 LASSERT(!dt_object_remote(dt));
4438 LASSERT(osd_is_write_locked(env, obj));
4439 LASSERT(th != NULL);
4441 if (unlikely(fid_is_acct(fid)))
4443 * Quota files can't be created from the kernel any more,
4444 * 'tune2fs -O quota' will take care of creating them
4448 result = __osd_create(info, obj, attr, hint, dof, th);
4450 if (fid_is_idif(fid) &&
4451 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4452 struct lu_fid *tfid = &info->oti_fid;
4453 struct ost_id *oi = &info->oti_ostid;
4455 fid_to_ostid(fid, oi);
4456 ostid_to_fid(tfid, oi, 0);
4458 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4459 LMAC_FID_ON_OST, 0);
4461 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4463 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4464 on_ost ? LMAC_FID_ON_OST : 0,
4467 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4468 obj->oo_dt.do_body_ops = &osd_body_ops;
4471 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4472 struct inode *inode = obj->oo_inode;
4474 result = __osd_oi_insert(env, obj, fid, th);
4475 if (result && inode) {
4476 spin_lock(&obj->oo_guard);
4478 spin_unlock(&obj->oo_guard);
4479 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4480 ldiskfs_set_inode_state(inode,
4481 LDISKFS_STATE_LUSTRE_DESTROY);
4483 obj->oo_inode = NULL;
4488 * a small optimization - dt_insert() isn't usually applied
4489 * to OST objects, so we don't need to cache OI mapping for
4492 if (result == 0 && on_ost == 0) {
4493 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4495 result = osd_idc_find_and_init(env, osd, obj);
4496 LASSERT(result == 0);
4499 LASSERT(ergo(result == 0,
4500 dt_object_exists(dt) && !dt_object_remote(dt)));
4501 LINVRNT(osd_invariant(obj));
4505 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4506 struct thandle *handle)
4508 struct osd_thandle *oh;
4511 /* it's possible that object doesn't exist yet */
4512 LASSERT(handle != NULL);
4514 oh = container_of(handle, struct osd_thandle, ot_super);
4515 LASSERT(oh->ot_handle == NULL);
4517 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4518 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4520 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4527 * Concurrency: @dt is write locked.
4529 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4532 struct osd_object *obj = osd_dt_obj(dt);
4533 struct inode *inode = obj->oo_inode;
4534 struct osd_thandle *oh;
4537 if (!dt_object_exists(dt) || obj->oo_destroyed)
4540 LINVRNT(osd_invariant(obj));
4541 LASSERT(!dt_object_remote(dt));
4542 LASSERT(osd_is_write_locked(env, obj));
4543 LASSERT(th != NULL);
4545 oh = container_of(th, struct osd_thandle, ot_super);
4546 LASSERT(oh->ot_handle != NULL);
4548 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4550 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4551 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4553 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4554 * (65000) subdirectories by storing "1" in i_nlink if the link count
4555 * would otherwise overflow. Directory tranversal tools understand
4556 * that (st_nlink == 1) indicates that the filesystem dose not track
4557 * hard links count on the directory, and will not abort subdirectory
4558 * scanning early once (st_nlink - 2) subdirs have been found.
4560 * This also has to properly handle the case of inodes with nlink == 0
4561 * in case they are being linked into the PENDING directory
4563 spin_lock(&obj->oo_guard);
4564 if (unlikely(inode->i_nlink == 0))
4565 /* inc_nlink from 0 may cause WARN_ON */
4566 set_nlink(inode, 1);
4568 osd_ldiskfs_inc_count(oh->ot_handle, inode);
4569 if (!S_ISDIR(inode->i_mode))
4570 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4572 spin_unlock(&obj->oo_guard);
4574 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4575 LINVRNT(osd_invariant(obj));
4577 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4582 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4583 struct thandle *handle)
4585 struct osd_thandle *oh;
4587 if (!dt_object_exists(dt))
4590 LASSERT(!dt_object_remote(dt));
4591 LASSERT(handle != NULL);
4593 oh = container_of(handle, struct osd_thandle, ot_super);
4594 LASSERT(oh->ot_handle == NULL);
4596 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4597 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4603 * Concurrency: @dt is write locked.
4605 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4608 struct osd_object *obj = osd_dt_obj(dt);
4609 struct inode *inode = obj->oo_inode;
4610 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4611 struct osd_thandle *oh;
4613 if (!dt_object_exists(dt))
4616 LINVRNT(osd_invariant(obj));
4617 LASSERT(!dt_object_remote(dt));
4618 LASSERT(osd_is_write_locked(env, obj));
4619 LASSERT(th != NULL);
4621 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4624 oh = container_of(th, struct osd_thandle, ot_super);
4625 LASSERT(oh->ot_handle != NULL);
4627 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4629 spin_lock(&obj->oo_guard);
4631 * That can be result of upgrade from old Lustre version and
4632 * applied only to local files. Just skip this ref_del call.
4633 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4635 if (inode->i_nlink == 0) {
4636 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4637 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4638 ", maybe an upgraded file? (LU-3915)\n",
4639 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4640 spin_unlock(&obj->oo_guard);
4644 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4645 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4647 osd_ldiskfs_dec_count(oh->ot_handle, inode);
4648 spin_unlock(&obj->oo_guard);
4650 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4651 LINVRNT(osd_invariant(obj));
4653 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4659 * Concurrency: @dt is read locked.
4661 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4662 struct lu_buf *buf, const char *name)
4664 struct osd_object *obj = osd_dt_obj(dt);
4665 struct inode *inode = obj->oo_inode;
4666 struct osd_thread_info *info = osd_oti_get(env);
4667 struct dentry *dentry = &info->oti_obj_dentry;
4668 bool cache_xattr = false;
4673 /* version get is not real XATTR but uses xattr API */
4674 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4675 dt_obj_version_t *ver = buf->lb_buf;
4678 * for version we are just using xattr API but change inode
4681 if (buf->lb_len == 0)
4682 return sizeof(dt_obj_version_t);
4684 if (buf->lb_len < sizeof(dt_obj_version_t))
4687 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4688 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4690 *ver = LDISKFS_I(inode)->i_fs_version;
4692 return sizeof(dt_obj_version_t);
4695 if (!dt_object_exists(dt))
4698 LASSERT(!dt_object_remote(dt));
4699 LASSERT(inode->i_op != NULL);
4700 #ifdef HAVE_IOP_XATTR
4701 LASSERT(inode->i_op->getxattr != NULL);
4704 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4705 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4709 rc = osd_oxc_get(obj, name, buf);
4714 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4715 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4716 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4717 struct filter_fid *ff;
4718 struct ost_layout *ol;
4720 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4724 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4727 if (buf->lb_len == 0 || !buf->lb_buf)
4730 if (buf->lb_len < rc)
4734 ol = &ff->ff_layout;
4735 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4736 PFID_STRIPE_IDX_BITS);
4737 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4738 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4739 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4740 if (lma->lma_compat & LMAC_COMP_INFO) {
4741 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4742 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4743 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4745 ol->ol_comp_start = 0;
4746 ol->ol_comp_end = 0;
4750 /* Get enc context xattr directly from ldiskfs instead of going
4751 * through the VFS, as there is no xattr handler for
4754 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0)
4755 rc = ldiskfs_xattr_get(inode,
4756 LDISKFS_XATTR_INDEX_ENCRYPTION,
4757 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
4758 buf->lb_buf, buf->lb_len);
4760 rc = __osd_xattr_get(inode, dentry, name,
4761 buf->lb_buf, buf->lb_len);
4765 if (rc == -ENOENT || rc == -ENODATA)
4766 osd_oxc_add(obj, name, NULL, 0);
4767 else if (rc > 0 && buf->lb_buf != NULL)
4768 osd_oxc_add(obj, name, buf->lb_buf, rc);
4774 static int osd_declare_xattr_set(const struct lu_env *env,
4775 struct dt_object *dt,
4776 const struct lu_buf *buf, const char *name,
4777 int fl, struct thandle *handle)
4779 struct osd_thandle *oh;
4781 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4783 LASSERT(handle != NULL);
4785 oh = container_of(handle, struct osd_thandle, ot_super);
4786 LASSERT(oh->ot_handle == NULL);
4788 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4790 * For non-upgrading case, the LMA is set first and
4791 * usually fit inode. But for upgrade case, the LMA
4792 * may be in another separated EA block.
4794 if (dt_object_exists(dt)) {
4795 if (fl == LU_XATTR_REPLACE)
4800 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4802 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4803 /* We may need to delete the old PFID EA. */
4804 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4805 if (fl == LU_XATTR_REPLACE)
4811 * If some name entry resides on remote MDT, then will create
4812 * agent entry under remote parent. On the other hand, if the
4813 * remote entry will be removed, then related agent entry may
4814 * need to be removed from the remote parent. So there may be
4815 * kinds of cases, let's declare enough credits. The credits
4816 * for create agent entry is enough for remove case.
4818 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4819 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4820 if (dt_object_exists(dt))
4821 credits += 1; /* For updating LMA */
4825 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4830 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4832 * learn xattr size from osd_xattr_get if
4833 * attribute has not been read yet
4835 buflen = __osd_xattr_get(
4836 osd_dt_obj(dt)->oo_inode,
4837 &osd_oti_get(env)->oti_obj_dentry,
4842 buflen = buf->lb_len;
4845 if (buflen > sb->s_blocksize) {
4846 credits += osd_calc_bkmap_credits(
4848 (buflen + sb->s_blocksize - 1) >>
4849 sb->s_blocksize_bits);
4853 * xattr set may involve inode quota change, reserve credits for
4854 * dquot_initialize()
4856 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4859 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4864 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4865 const struct lu_buf *buf, int fl,
4866 struct thandle *handle)
4868 struct osd_thread_info *info = osd_oti_get(env);
4869 struct dentry *dentry = &info->oti_obj_dentry;
4870 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4871 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4872 struct inode *inode = obj->oo_inode;
4873 struct filter_fid *ff = buf->lb_buf;
4874 struct ost_layout *ol = &ff->ff_layout;
4875 int flags = XATTR_REPLACE;
4880 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4883 rc = osd_get_lma(info, inode, dentry, loa);
4884 if (rc == -ENODATA) {
4885 /* Usually for upgarding from old device */
4886 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4887 LMAC_FID_ON_OST, 0);
4888 flags = XATTR_CREATE;
4893 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4894 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4897 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4898 /* Separate PFID EA from LMA */
4899 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4900 lustre_lma_swab(lma);
4901 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4902 sizeof(*lma), XATTR_REPLACE);
4904 obj->oo_pfid_in_lma = 0;
4905 rc = LU_XATTR_CREATE;
4911 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4915 * Old client does not send stripe information,
4916 * then store the PFID EA on disk separatedly.
4918 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4919 ol->ol_stripe_size == 0))
4922 /* Remove old PFID EA entry firstly. */
4923 dquot_initialize(inode);
4924 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4925 if (rc == -ENODATA) {
4926 /* XATTR_NAME_FID is already absent */
4933 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4934 if (likely(ol->ol_stripe_size != 0)) {
4935 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4936 PFID_STRIPE_IDX_BITS;
4937 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4938 lma->lma_compat |= LMAC_STRIPE_INFO;
4939 if (ol->ol_comp_id != 0) {
4940 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4941 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4942 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4943 lma->lma_compat |= LMAC_COMP_INFO;
4947 lustre_loa_swab(loa, false);
4949 /* Store the PFID EA inside LMA. */
4950 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4953 obj->oo_pfid_in_lma = 1;
4959 * In DNE environment, the object (in spite of regular file or directory)
4960 * and its name entry may reside on different MDTs. Under such case, we will
4961 * create an agent entry on the MDT where the object resides. The agent entry
4962 * references the object locally, that makes the object to be visible to the
4963 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4964 * such as 'tar' can handle the object properly.
4966 * We handle the agent entry during set linkEA that is the common interface
4967 * for both regular file and directroy, can handle kinds of cases, such as
4968 * create/link/unlink/rename, and so on.
4970 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4972 * XXX: There are two known issues:
4973 * 1. For one object, we will create at most one agent entry even if there
4974 * may be more than one cross-MDTs hard links on the object. So the local
4975 * e2fsck may claim that the object's nlink is larger than the name entries
4976 * that reference such inode. And in further, the e2fsck will fix the nlink
4977 * attribute to match the local references. Then it will cause the object's
4978 * nlink attribute to be inconsistent with the global references. it is bad
4979 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4980 * other hand, the global namespace LFSCK can repair the object's attribute
4981 * according to the linkEA.
4982 * 2. There may be too many hard links on the object as to its linkEA overflow,
4983 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4984 * case happened, then at this point, we do not know whether there are some
4985 * cross-MDTs reference. But there are local references, it guarantees that
4986 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4988 static int osd_xattr_handle_linkea(const struct lu_env *env,
4989 struct osd_device *osd,
4990 struct osd_object *obj,
4991 const struct lu_buf *buf,
4992 struct thandle *handle)
4994 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4995 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4996 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4997 struct lu_name tmpname;
4998 struct osd_thandle *oh;
5000 bool remote = false;
5004 oh = container_of(handle, struct osd_thandle, ot_super);
5005 LASSERT(oh->ot_handle != NULL);
5007 rc = linkea_init_with_rec(&ldata);
5009 linkea_first_entry(&ldata);
5010 while (ldata.ld_lee != NULL && !remote) {
5011 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
5013 if (osd_remote_fid(env, osd, tfid) > 0)
5016 linkea_next_entry(&ldata);
5018 } else if (rc == -ENODATA) {
5024 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
5025 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
5027 CERROR("%s: failed to remove agent entry for "DFID
5028 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5029 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
5030 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5032 CERROR("%s: failed to create agent entry for "DFID
5033 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5040 * Concurrency: @dt is write locked.
5042 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
5043 const struct lu_buf *buf, const char *name, int fl,
5044 struct thandle *handle)
5046 struct osd_object *obj = osd_dt_obj(dt);
5047 struct osd_device *osd = osd_obj2dev(obj);
5048 struct inode *inode = obj->oo_inode;
5049 struct osd_thread_info *info = osd_oti_get(env);
5059 /* version set is not real XATTR */
5060 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
5061 dt_obj_version_t *version = buf->lb_buf;
5064 * for version we are just using xattr API but change inode
5067 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
5070 DFID" set version %#llx (old %#llx) for inode %lu\n",
5071 PFID(lu_object_fid(&dt->do_lu)), *version,
5072 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
5074 LDISKFS_I(inode)->i_fs_version = *version;
5076 * Version is set after all inode operations are finished,
5077 * so we should mark it dirty here
5079 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
5084 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
5085 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
5088 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5091 * For the OST device with 256 bytes inode size by default,
5092 * the PFID EA will be stored together with LMA EA to avoid
5093 * performance trouble. Otherwise the PFID EA can be stored
5094 * independently. LU-8998
5096 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
5097 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
5098 LASSERT(buf->lb_buf);
5100 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
5103 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
5104 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
5105 struct lustre_mdt_attrs *lma = &loa->loa_lma;
5107 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
5111 lma->lma_incompat |= LMAI_STRIPED;
5112 lustre_lma_swab(lma);
5113 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5114 sizeof(*lma), XATTR_REPLACE);
5117 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
5118 LASSERT(!osd->od_is_ost);
5120 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
5125 if (fl & LU_XATTR_REPLACE)
5126 fs_flags |= XATTR_REPLACE;
5128 if (fl & LU_XATTR_CREATE)
5129 fs_flags |= XATTR_CREATE;
5131 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0) {
5132 /* Set enc context xattr directly in ldiskfs instead of going
5133 * through the VFS, as there is no xattr handler for
5136 struct osd_thandle *oth = container_of(handle,
5140 if (!oth->ot_handle)
5141 /* this should be already part of a transaction */
5144 rc = ldiskfs_xattr_set_handle(oth->ot_handle, inode,
5145 LDISKFS_XATTR_INDEX_ENCRYPTION,
5146 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
5147 buf->lb_buf, len, fs_flags);
5149 rc = __osd_xattr_set(info, inode, name,
5150 buf->lb_buf, len, fs_flags);
5152 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5155 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5156 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5157 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
5163 * Concurrency: @dt is read locked.
5165 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
5166 const struct lu_buf *buf)
5168 struct osd_object *obj = osd_dt_obj(dt);
5169 struct osd_device *dev = osd_obj2dev(obj);
5170 struct inode *inode = obj->oo_inode;
5171 struct osd_thread_info *info = osd_oti_get(env);
5172 struct dentry *dentry = &info->oti_obj_dentry;
5175 if (!dt_object_exists(dt))
5178 LASSERT(!dt_object_remote(dt));
5179 LASSERT(inode->i_op != NULL);
5180 LASSERT(inode->i_op->listxattr != NULL);
5182 dentry->d_inode = inode;
5183 dentry->d_sb = inode->i_sb;
5184 rc = inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
5186 if (rc < 0 || buf->lb_buf == NULL)
5189 /* Hide virtual project ID xattr from list if disabled */
5190 if (!dev->od_enable_projid_xattr) {
5191 char *end = (char *)buf->lb_buf + rc;
5192 char *p = buf->lb_buf;
5195 char *next = p + strlen(p) + 1;
5197 if (strcmp(p, XATTR_NAME_PROJID) == 0) {
5199 memmove(p, next, end - next);
5211 static int osd_declare_xattr_del(const struct lu_env *env,
5212 struct dt_object *dt, const char *name,
5213 struct thandle *handle)
5215 struct osd_thandle *oh;
5216 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
5218 LASSERT(!dt_object_remote(dt));
5219 LASSERT(handle != NULL);
5221 oh = container_of(handle, struct osd_thandle, ot_super);
5222 LASSERT(oh->ot_handle == NULL);
5224 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
5225 osd_dto_credits_noquota[DTO_XATTR_SET]);
5227 * xattr del may involve inode quota change, reserve credits for
5228 * dquot_initialize()
5230 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
5236 * Concurrency: @dt is write locked.
5238 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
5239 const char *name, struct thandle *handle)
5241 struct osd_object *obj = osd_dt_obj(dt);
5242 struct inode *inode = obj->oo_inode;
5243 struct osd_thread_info *info = osd_oti_get(env);
5244 struct dentry *dentry = &info->oti_obj_dentry;
5247 if (!dt_object_exists(dt))
5250 LASSERT(!dt_object_remote(dt));
5251 LASSERT(inode->i_op != NULL);
5252 LASSERT(handle != NULL);
5253 #ifdef HAVE_IOP_XATTR
5254 LASSERT(inode->i_op->removexattr != NULL);
5257 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5259 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
5260 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
5262 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
5263 &info->oti_ost_attrs);
5265 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
5267 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
5268 lustre_lma_swab(lma);
5269 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5270 sizeof(*lma), XATTR_REPLACE);
5272 obj->oo_pfid_in_lma = 0;
5275 dquot_initialize(inode);
5276 dentry->d_inode = inode;
5277 dentry->d_sb = inode->i_sb;
5278 rc = ll_vfs_removexattr(dentry, inode, name);
5281 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5284 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5285 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5286 osd_oxc_del(obj, name);
5291 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
5292 __u64 start, __u64 end)
5294 struct osd_object *obj = osd_dt_obj(dt);
5295 struct osd_device *dev = osd_obj2dev(obj);
5296 struct inode *inode = obj->oo_inode;
5301 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
5304 RETURN(PTR_ERR(file));
5306 file->f_mode |= FMODE_64BITHASH;
5307 rc = vfs_fsync_range(file, start, end, 0);
5314 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
5319 static bool osd_check_stale(struct dt_object *dt)
5327 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
5328 const struct dt_index_features *feat)
5330 struct iam_descr *descr;
5332 if (osd_object_is_root(o))
5333 return feat == &dt_directory_features;
5335 LASSERT(o->oo_dir != NULL);
5337 descr = o->oo_dir->od_container.ic_descr;
5338 if (feat == &dt_directory_features) {
5339 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
5344 return feat->dif_keysize_min <= descr->id_key_size &&
5345 descr->id_key_size <= feat->dif_keysize_max &&
5346 feat->dif_recsize_min <= descr->id_rec_size &&
5347 descr->id_rec_size <= feat->dif_recsize_max &&
5348 !(feat->dif_flags & (DT_IND_VARKEY |
5349 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5350 ergo(feat->dif_flags & DT_IND_UPDATE,
5351 1 /* XXX check that object (and fs) is writable */);
5355 static int osd_iam_container_init(const struct lu_env *env,
5356 struct osd_object *obj,
5357 struct osd_directory *dir)
5359 struct iam_container *bag = &dir->od_container;
5362 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5366 result = iam_container_setup(bag);
5368 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5370 iam_container_fini(bag);
5377 * Concurrency: no external locking is necessary.
5379 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5380 const struct dt_index_features *feat)
5384 struct osd_object *obj = osd_dt_obj(dt);
5386 LINVRNT(osd_invariant(obj));
5388 if (osd_object_is_root(obj)) {
5389 dt->do_index_ops = &osd_index_ea_ops;
5391 } else if (feat == &dt_directory_features) {
5392 dt->do_index_ops = &osd_index_ea_ops;
5393 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5398 } else if (unlikely(feat == &dt_otable_features)) {
5399 dt->do_index_ops = &osd_otable_ops;
5401 } else if (unlikely(feat == &dt_acct_features)) {
5402 dt->do_index_ops = &osd_acct_index_ops;
5405 } else if (!osd_has_index(obj)) {
5406 struct osd_directory *dir;
5407 struct osd_device *osd = osd_obj2dev(obj);
5408 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5413 spin_lock(&obj->oo_guard);
5414 if (obj->oo_dir == NULL)
5418 * Concurrent thread allocated container data.
5421 spin_unlock(&obj->oo_guard);
5423 * Now, that we have container data, serialize its
5426 down_write(&obj->oo_ext_idx_sem);
5428 * recheck under lock.
5431 if (osd_has_index(obj)) {
5436 result = osd_iam_container_init(env, obj, obj->oo_dir);
5437 if (result || feat == &dt_lfsck_namespace_features ||
5438 feat == &dt_lfsck_layout_orphan_features ||
5439 feat == &dt_lfsck_layout_dangling_features)
5442 result = osd_index_register(osd, fid,
5443 feat->dif_keysize_max,
5444 feat->dif_recsize_max);
5446 CWARN("%s: failed to register index "
5448 osd_name(osd), PFID(fid), result);
5449 else if (result > 0)
5452 CDEBUG(D_LFSCK, "%s: index object "DFID
5453 " (%d/%d) registered\n",
5454 osd_name(osd), PFID(fid),
5455 (int)feat->dif_keysize_max,
5456 (int)feat->dif_recsize_max);
5459 up_write(&obj->oo_ext_idx_sem);
5467 if (result == 0 && skip_iam == 0) {
5468 if (!osd_iam_index_probe(env, obj, feat))
5471 LINVRNT(osd_invariant(obj));
5476 static int osd_otable_it_attr_get(const struct lu_env *env,
5477 struct dt_object *dt,
5478 struct lu_attr *attr)
5484 static const struct dt_object_operations osd_obj_ops = {
5485 .do_read_lock = osd_read_lock,
5486 .do_write_lock = osd_write_lock,
5487 .do_read_unlock = osd_read_unlock,
5488 .do_write_unlock = osd_write_unlock,
5489 .do_write_locked = osd_write_locked,
5490 .do_attr_get = osd_attr_get,
5491 .do_declare_attr_set = osd_declare_attr_set,
5492 .do_attr_set = osd_attr_set,
5493 .do_ah_init = osd_ah_init,
5494 .do_declare_create = osd_declare_create,
5495 .do_create = osd_create,
5496 .do_declare_destroy = osd_declare_destroy,
5497 .do_destroy = osd_destroy,
5498 .do_index_try = osd_index_try,
5499 .do_declare_ref_add = osd_declare_ref_add,
5500 .do_ref_add = osd_ref_add,
5501 .do_declare_ref_del = osd_declare_ref_del,
5502 .do_ref_del = osd_ref_del,
5503 .do_xattr_get = osd_xattr_get,
5504 .do_declare_xattr_set = osd_declare_xattr_set,
5505 .do_xattr_set = osd_xattr_set,
5506 .do_declare_xattr_del = osd_declare_xattr_del,
5507 .do_xattr_del = osd_xattr_del,
5508 .do_xattr_list = osd_xattr_list,
5509 .do_object_sync = osd_object_sync,
5510 .do_invalidate = osd_invalidate,
5511 .do_check_stale = osd_check_stale,
5514 static const struct dt_object_operations osd_obj_otable_it_ops = {
5515 .do_attr_get = osd_otable_it_attr_get,
5516 .do_index_try = osd_index_try,
5519 static int osd_index_declare_iam_delete(const struct lu_env *env,
5520 struct dt_object *dt,
5521 const struct dt_key *key,
5522 struct thandle *handle)
5524 struct osd_thandle *oh;
5526 oh = container_of(handle, struct osd_thandle, ot_super);
5527 LASSERT(oh->ot_handle == NULL);
5529 /* Recycle may cause additional three blocks to be changed. */
5530 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5531 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5537 * delete a (key, value) pair from index \a dt specified by \a key
5539 * \param dt osd index object
5540 * \param key key for index
5541 * \param rec record reference
5542 * \param handle transaction handler
5545 * \retval -ve failure
5547 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5548 const struct dt_key *key,
5549 struct thandle *handle)
5551 struct osd_thread_info *oti = osd_oti_get(env);
5552 struct osd_object *obj = osd_dt_obj(dt);
5553 struct osd_thandle *oh;
5554 struct iam_path_descr *ipd;
5555 struct iam_container *bag = &obj->oo_dir->od_container;
5560 if (!dt_object_exists(dt))
5563 LINVRNT(osd_invariant(obj));
5564 LASSERT(!dt_object_remote(dt));
5565 LASSERT(bag->ic_object == obj->oo_inode);
5566 LASSERT(handle != NULL);
5568 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5570 ipd = osd_idx_ipd_get(env, bag);
5571 if (unlikely(ipd == NULL))
5574 oh = container_of(handle, struct osd_thandle, ot_super);
5575 LASSERT(oh->ot_handle != NULL);
5576 LASSERT(oh->ot_handle->h_transaction != NULL);
5578 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5579 /* swab quota uid/gid provided by caller */
5580 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5581 key = (const struct dt_key *)&oti->oti_quota_id;
5584 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5585 osd_ipd_put(env, bag, ipd);
5586 LINVRNT(osd_invariant(obj));
5587 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5591 static int osd_index_declare_ea_delete(const struct lu_env *env,
5592 struct dt_object *dt,
5593 const struct dt_key *key,
5594 struct thandle *handle)
5596 struct osd_thandle *oh;
5597 struct inode *inode;
5602 LASSERT(!dt_object_remote(dt));
5603 LASSERT(handle != NULL);
5605 oh = container_of(handle, struct osd_thandle, ot_super);
5606 LASSERT(oh->ot_handle == NULL);
5608 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5609 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5611 inode = osd_dt_obj(dt)->oo_inode;
5615 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5616 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5621 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5624 struct osd_fid_pack *rec;
5627 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5628 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5629 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5630 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5636 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5637 const struct lu_fid *fid)
5639 struct seq_server_site *ss = osd_seq_site(osd);
5643 /* FID seqs not in FLDB, must be local seq */
5644 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5648 * If FLD is not being initialized yet, it only happens during the
5649 * initialization, likely during mgs initialization, and we assume
5650 * this is local FID.
5652 if (ss == NULL || ss->ss_server_fld == NULL)
5655 /* Only check the local FLDB here */
5656 if (osd_seq_exists(env, osd, fid_seq(fid)))
5662 static void osd_take_care_of_agent(const struct lu_env *env,
5663 struct osd_device *osd,
5664 struct osd_thandle *oh,
5665 struct ldiskfs_dir_entry_2 *de)
5667 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5668 struct osd_idmap_cache *idc;
5669 int rc, schedule = 0;
5671 LASSERT(de != NULL);
5673 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5674 if (likely(rc == 0)) {
5675 idc = osd_idc_find_or_init(env, osd, fid);
5676 if (IS_ERR(idc) || idc->oic_remote)
5678 } else if (rc == -ENODATA) {
5680 * can't get FID, postpone to the end of the
5681 * transaction when iget() is safe
5685 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5688 osd_schedule_agent_inode_removal(env, oh,
5689 le32_to_cpu(de->inode));
5693 * Utility function to get real name from object name
5695 * \param[in] obj pointer to the object to be handled
5696 * \param[in] name object name
5697 * \param[in] len object name len
5698 * \param[out]ln pointer to the struct lu_name to hold the real name
5700 * If file is not encrypted, real name is just the object name.
5701 * If file is encrypted, object name needs to be decoded. In
5702 * this case a new buffer is allocated, and ln->ln_name needs to be freed by
5705 * \retval 0, on success
5706 * \retval -ve, on error
5708 static int obj_name2lu_name(struct osd_object *obj, const char *name,
5709 int len, struct lu_name *ln)
5711 if (!(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
5713 ln->ln_namelen = len;
5715 char *buf = kmalloc(len, GFP_NOFS);
5720 len = critical_decode(name, len, buf);
5722 ln->ln_namelen = len;
5729 * Index delete function for interoperability mode (b11826).
5730 * It will remove the directory entry added by osd_index_ea_insert().
5731 * This entry is needed to maintain name->fid mapping.
5733 * \param key, key i.e. file entry to be deleted
5735 * \retval 0, on success
5736 * \retval -ve, on error
5738 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5739 const struct dt_key *key, struct thandle *handle)
5741 struct osd_object *obj = osd_dt_obj(dt);
5742 struct inode *dir = obj->oo_inode;
5743 struct dentry *dentry;
5744 struct osd_thandle *oh;
5745 struct ldiskfs_dir_entry_2 *de = NULL;
5746 struct buffer_head *bh;
5747 struct htree_lock *hlock = NULL;
5748 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5754 if (!dt_object_exists(dt))
5757 LINVRNT(osd_invariant(obj));
5758 LASSERT(!dt_object_remote(dt));
5759 LASSERT(handle != NULL);
5761 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
5765 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5767 oh = container_of(handle, struct osd_thandle, ot_super);
5768 LASSERT(oh->ot_handle != NULL);
5769 LASSERT(oh->ot_handle->h_transaction != NULL);
5771 dquot_initialize(dir);
5772 dentry = osd_child_dentry_get(env, obj, ln.ln_name, ln.ln_namelen);
5774 if (obj->oo_hl_head != NULL) {
5775 hlock = osd_oti_get(env)->oti_hlock;
5776 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5777 dir, LDISKFS_HLOCK_DEL);
5779 down_write(&obj->oo_ext_idx_sem);
5782 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5785 * If this is not the ".." entry, it might be a remote DNE
5786 * entry and we need to check if the FID is for a remote
5787 * MDT. If the FID is not in the directory entry (e.g.
5788 * upgraded 1.8 filesystem without dirdata enabled) then
5789 * we need to get the FID from the LMA. For a remote directory
5790 * there HAS to be an LMA, it cannot be an IGIF inode in this
5793 * Delete the entry before the agent inode in order to
5794 * simplify error handling. At worst an error after deleting
5795 * the entry first might leak the agent inode afterward. The
5796 * reverse would need filesystem abort in case of error deleting
5797 * the entry after the agent had been removed, or leave a
5798 * dangling entry pointing at a random inode.
5800 if (strcmp((char *)key, dotdot) != 0)
5801 osd_take_care_of_agent(env, osd, oh, de);
5802 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5808 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
5809 atomic_dec_if_positive(&obj->oo_dirent_count);
5811 ldiskfs_htree_unlock(hlock);
5813 up_write(&obj->oo_ext_idx_sem);
5816 LASSERT(osd_invariant(obj));
5817 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5818 if (ln.ln_name != (char *)key)
5824 * Lookup index for \a key and copy record to \a rec.
5826 * \param dt osd index object
5827 * \param key key for index
5828 * \param rec record reference
5830 * \retval +ve success : exact mach
5831 * \retval 0 return record with key not greater than \a key
5832 * \retval -ve failure
5834 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5835 struct dt_rec *rec, const struct dt_key *key)
5837 struct osd_object *obj = osd_dt_obj(dt);
5838 struct iam_path_descr *ipd;
5839 struct iam_container *bag = &obj->oo_dir->od_container;
5840 struct osd_thread_info *oti = osd_oti_get(env);
5841 struct iam_iterator *it = &oti->oti_idx_it;
5842 struct iam_rec *iam_rec;
5847 if (!dt_object_exists(dt))
5850 LASSERT(osd_invariant(obj));
5851 LASSERT(!dt_object_remote(dt));
5852 LASSERT(bag->ic_object == obj->oo_inode);
5854 ipd = osd_idx_ipd_get(env, bag);
5858 /* got ipd now we can start iterator. */
5859 iam_it_init(it, bag, 0, ipd);
5861 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5862 /* swab quota uid/gid provided by caller */
5863 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5864 key = (const struct dt_key *)&oti->oti_quota_id;
5867 rc = iam_it_get(it, (struct iam_key *)key);
5869 if (S_ISDIR(obj->oo_inode->i_mode))
5870 iam_rec = (struct iam_rec *)oti->oti_ldp;
5872 iam_rec = (struct iam_rec *)rec;
5874 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5876 if (S_ISDIR(obj->oo_inode->i_mode))
5877 osd_fid_unpack((struct lu_fid *)rec,
5878 (struct osd_fid_pack *)iam_rec);
5879 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5880 osd_quota_unpack(obj, rec);
5885 osd_ipd_put(env, bag, ipd);
5887 LINVRNT(osd_invariant(obj));
5892 static int osd_index_declare_iam_insert(const struct lu_env *env,
5893 struct dt_object *dt,
5894 const struct dt_rec *rec,
5895 const struct dt_key *key,
5896 struct thandle *handle)
5898 struct osd_thandle *oh;
5900 LASSERT(handle != NULL);
5902 oh = container_of(handle, struct osd_thandle, ot_super);
5903 LASSERT(oh->ot_handle == NULL);
5905 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5906 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5912 * Inserts (key, value) pair in \a dt index object.
5914 * \param dt osd index object
5915 * \param key key for index
5916 * \param rec record reference
5917 * \param th transaction handler
5920 * \retval -ve failure
5922 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5923 const struct dt_rec *rec,
5924 const struct dt_key *key, struct thandle *th)
5926 struct osd_object *obj = osd_dt_obj(dt);
5927 struct iam_path_descr *ipd;
5928 struct osd_thandle *oh;
5929 struct iam_container *bag;
5930 struct osd_thread_info *oti = osd_oti_get(env);
5931 struct iam_rec *iam_rec;
5936 if (!dt_object_exists(dt))
5939 LINVRNT(osd_invariant(obj));
5940 LASSERT(!dt_object_remote(dt));
5942 bag = &obj->oo_dir->od_container;
5943 LASSERT(bag->ic_object == obj->oo_inode);
5944 LASSERT(th != NULL);
5946 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5948 ipd = osd_idx_ipd_get(env, bag);
5949 if (unlikely(ipd == NULL))
5952 oh = container_of(th, struct osd_thandle, ot_super);
5953 LASSERT(oh->ot_handle != NULL);
5954 LASSERT(oh->ot_handle->h_transaction != NULL);
5955 if (S_ISDIR(obj->oo_inode->i_mode)) {
5956 iam_rec = (struct iam_rec *)oti->oti_ldp;
5957 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5959 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5960 /* pack quota uid/gid */
5961 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5962 key = (const struct dt_key *)&oti->oti_quota_id;
5963 /* pack quota record */
5964 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5965 iam_rec = (struct iam_rec *)rec;
5967 iam_rec = (struct iam_rec *)rec;
5970 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5972 osd_ipd_put(env, bag, ipd);
5973 LINVRNT(osd_invariant(obj));
5974 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5979 * Calls ldiskfs_add_entry() to add directory entry
5980 * into the directory. This is required for
5981 * interoperability mode (b11826)
5983 * \retval 0, on success
5984 * \retval -ve, on error
5986 static int __osd_ea_add_rec(struct osd_thread_info *info,
5987 struct osd_object *pobj, struct inode *cinode,
5988 const char *name, const struct lu_fid *fid,
5989 struct htree_lock *hlock, struct thandle *th)
5991 struct ldiskfs_dentry_param *ldp;
5992 struct dentry *child;
5993 struct osd_thandle *oth;
5997 oth = container_of(th, struct osd_thandle, ot_super);
5998 LASSERT(oth->ot_handle != NULL);
5999 LASSERT(oth->ot_handle->h_transaction != NULL);
6000 LASSERT(pobj->oo_inode);
6002 rc = obj_name2lu_name(pobj, name, strlen(name), &ln);
6006 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
6007 if (unlikely(osd_object_is_root(pobj)))
6010 osd_get_ldiskfs_dirent_param(ldp, fid);
6011 child = osd_child_dentry_get(info->oti_env, pobj,
6012 ln.ln_name, ln.ln_namelen);
6013 child->d_fsdata = (void *)ldp;
6014 dquot_initialize(pobj->oo_inode);
6015 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
6016 child, cinode, hlock);
6017 if (rc == 0 && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
6018 struct ldiskfs_dir_entry_2 *de;
6019 struct buffer_head *bh;
6022 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
6025 rc1 = osd_ldiskfs_journal_get_write_access(
6026 oth->ot_handle, pobj->oo_inode->i_sb, bh,
6029 if (S_ISDIR(cinode->i_mode))
6030 de->file_type = LDISKFS_DIRENT_LUFID |
6031 LDISKFS_FT_REG_FILE;
6033 de->file_type = LDISKFS_DIRENT_LUFID |
6035 ldiskfs_handle_dirty_metadata(oth->ot_handle,
6042 if (ln.ln_name != name)
6048 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
6049 * into the directory.Also sets flags into osd object to
6050 * indicate dot and dotdot are created. This is required for
6051 * interoperability mode (b11826)
6053 * \param dir directory for dot and dotdot fixup.
6054 * \param obj child object for linking
6056 * \retval 0, on success
6057 * \retval -ve, on error
6059 static int osd_add_dot_dotdot(struct osd_thread_info *info,
6060 struct osd_object *dir,
6061 struct inode *parent_dir, const char *name,
6062 const struct lu_fid *dot_fid,
6063 const struct lu_fid *dot_dot_fid,
6066 struct inode *inode = dir->oo_inode;
6067 struct osd_thandle *oth;
6070 oth = container_of(th, struct osd_thandle, ot_super);
6071 LASSERT(oth->ot_handle->h_transaction != NULL);
6072 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
6074 if (strcmp(name, dot) == 0) {
6075 if (dir->oo_compat_dot_created) {
6078 LASSERT(inode->i_ino == parent_dir->i_ino);
6079 dir->oo_compat_dot_created = 1;
6082 } else if (strcmp(name, dotdot) == 0) {
6083 if (!dir->oo_compat_dot_created)
6085 /* in case of rename, dotdot is already created */
6086 if (dir->oo_compat_dotdot_created) {
6087 return __osd_ea_add_rec(info, dir, parent_dir, name,
6088 dot_dot_fid, NULL, th);
6091 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
6092 struct lu_fid tfid = *dot_dot_fid;
6095 result = osd_add_dot_dotdot_internal(info,
6096 dir->oo_inode, parent_dir, dot_fid,
6099 result = osd_add_dot_dotdot_internal(info,
6100 dir->oo_inode, parent_dir, dot_fid,
6105 dir->oo_compat_dotdot_created = 1;
6113 * It will call the appropriate osd_add* function and return the
6114 * value, return by respective functions.
6116 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
6117 struct inode *cinode, const char *name,
6118 const struct lu_fid *fid, struct thandle *th)
6120 struct osd_thread_info *info = osd_oti_get(env);
6121 struct htree_lock *hlock;
6124 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
6126 if (name[0] == '.' && (name[1] == '\0' ||
6127 (name[1] == '.' && name[2] == '\0'))) {
6128 if (hlock != NULL) {
6129 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6132 down_write(&pobj->oo_ext_idx_sem);
6135 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
6136 lu_object_fid(&pobj->oo_dt.do_lu),
6139 if (hlock != NULL) {
6140 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6141 pobj->oo_inode, LDISKFS_HLOCK_ADD);
6143 down_write(&pobj->oo_ext_idx_sem);
6146 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
6147 struct lu_fid *tfid = &info->oti_fid;
6151 rc = __osd_ea_add_rec(info, pobj, cinode, name,
6154 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
6158 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))){
6159 int dirent_count = atomic_read(&pobj->oo_dirent_count);
6161 /* avoid extremely unlikely 2B-entry directory overflow case */
6162 if (dirent_count != LU_DIRENT_COUNT_UNSET &&
6163 likely(dirent_count < INT_MAX - NR_CPUS))
6164 atomic_inc(&pobj->oo_dirent_count);
6168 ldiskfs_htree_unlock(hlock);
6170 up_write(&pobj->oo_ext_idx_sem);
6176 osd_ldiskfs_consistency_check(struct osd_thread_info *oti,
6177 struct osd_device *dev,
6178 const struct lu_fid *fid,
6179 struct osd_inode_id *id)
6181 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
6182 struct inode *inode = NULL;
6188 if (!scrub_needs_check(scrub, fid, id->oii_ino))
6191 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
6192 if (rc == -ENOENT) {
6193 __u32 gen = id->oii_gen;
6199 inode = osd_iget(oti, dev, id);
6200 /* The inode has been removed (by race maybe). */
6201 if (IS_ERR(inode)) {
6202 rc = PTR_ERR(inode);
6204 RETURN(rc == -ESTALE ? -ENOENT : rc);
6207 /* The OI mapping is lost. */
6208 if (gen != OSD_OII_NOGEN)
6212 * The inode may has been reused by others, we do not know,
6213 * leave it to be handled by subsequent osd_fid_lookup().
6216 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
6223 if (scrub->os_running) {
6224 if (inode == NULL) {
6225 inode = osd_iget(oti, dev, id);
6226 /* The inode has been removed (by race maybe). */
6227 if (IS_ERR(inode)) {
6228 rc = PTR_ERR(inode);
6230 RETURN(rc == -ESTALE ? -ENOENT : rc);
6234 rc = osd_scrub_oi_insert(dev, fid, id, insert);
6236 * There is race condition between osd_oi_lookup and OI scrub.
6237 * The OI scrub finished just after osd_oi_lookup() failure.
6238 * Under such case, it is unnecessary to trigger OI scrub again,
6239 * but try to call osd_oi_lookup() again.
6241 if (unlikely(rc == -EAGAIN))
6244 if (!S_ISDIR(inode->i_mode))
6247 rc = osd_check_lmv(oti, dev, inode);
6252 if (dev->od_scrub.os_scrub.os_auto_scrub_interval != AS_NEVER &&
6254 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
6255 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
6256 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
6257 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
6258 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
6259 if (rc == 0 || rc == -EALREADY)
6271 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
6272 struct osd_device *dev,
6273 struct lu_fid *fid, __u32 ino)
6275 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
6276 struct osd_idmap_cache *oic = &oti->oti_cache;
6277 struct inode *inode;
6280 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
6281 inode = osd_iget(oti, dev, &oic->oic_lid);
6282 if (IS_ERR(inode)) {
6283 fid_zero(&oic->oic_fid);
6284 return PTR_ERR(inode);
6287 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
6290 fid_zero(&oic->oic_fid);
6292 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
6296 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
6297 struct osd_inode_id *id, const struct lu_fid *fid)
6299 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
6300 id->oii_ino, id->oii_gen, info);
6301 info->oti_cache.oic_lid = *id;
6302 info->oti_cache.oic_fid = *fid;
6303 info->oti_cache.oic_dev = osd;
6307 * Get parent FID from the linkEA.
6309 * For a directory which parent resides on remote MDT, to satisfy the
6310 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
6311 * the other hand, to make the lookup(..) on the directory can return
6312 * the real parent FID, we append the real parent FID after its ".."
6313 * name entry in the /REMOTE_PARENT_DIR.
6315 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
6316 * backup. So after the restore, we cannot get the right parent FID from
6317 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
6318 * we have stored the real parent FID in the directory object's linkEA,
6319 * we can parse the linkEA for the real parent FID.
6321 * \param[in] env pointer to the thread context
6322 * \param[in] obj pointer to the object to be handled
6323 * \param[out]fid pointer to the buffer to hold the parent FID
6325 * \retval 0 for getting the real parent FID successfully
6326 * \retval negative error number on failure
6328 static int osd_get_pfid_from_linkea(const struct lu_env *env,
6329 struct osd_object *obj,
6332 struct osd_thread_info *oti = osd_oti_get(env);
6333 struct lu_buf *buf = &oti->oti_big_buf;
6334 struct dentry *dentry = &oti->oti_obj_dentry;
6335 struct inode *inode = obj->oo_inode;
6336 struct linkea_data ldata = { NULL };
6342 if (!S_ISDIR(inode->i_mode))
6346 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6347 buf->lb_buf, buf->lb_len);
6348 if (rc == -ERANGE) {
6349 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6352 lu_buf_realloc(buf, rc);
6353 if (buf->lb_buf == NULL)
6360 if (unlikely(rc == 0))
6366 if (unlikely(buf->lb_buf == NULL)) {
6367 lu_buf_realloc(buf, rc);
6368 if (buf->lb_buf == NULL)
6375 rc = linkea_init_with_rec(&ldata);
6377 linkea_first_entry(&ldata);
6378 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
6384 static int osd_verify_ent_by_linkea(const struct lu_env *env,
6385 struct inode *inode,
6386 const struct lu_fid *pfid,
6387 const char *name, const int namelen)
6389 struct osd_thread_info *oti = osd_oti_get(env);
6390 struct lu_buf *buf = &oti->oti_big_buf;
6391 struct dentry *dentry = &oti->oti_obj_dentry;
6392 struct linkea_data ldata = { NULL };
6393 struct lu_name cname = { .ln_name = name,
6394 .ln_namelen = namelen };
6400 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6401 buf->lb_buf, buf->lb_len);
6403 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6408 if (unlikely(rc == 0))
6411 if (buf->lb_len < rc) {
6412 lu_buf_realloc(buf, rc);
6413 if (buf->lb_buf == NULL)
6420 rc = linkea_init_with_rec(&ldata);
6422 rc = linkea_links_find(&ldata, &cname, pfid);
6428 * Calls ->lookup() to find dentry. From dentry get inode and
6429 * read inode's ea to get fid. This is required for interoperability
6432 * \retval 0, on success
6433 * \retval -ve, on error
6435 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6436 struct dt_rec *rec, const struct lu_name *ln)
6438 struct inode *dir = obj->oo_inode;
6439 struct dentry *dentry;
6440 struct ldiskfs_dir_entry_2 *de;
6441 struct buffer_head *bh;
6442 struct lu_fid *fid = (struct lu_fid *)rec;
6443 struct htree_lock *hlock = NULL;
6449 LASSERT(dir->i_op != NULL);
6450 LASSERT(dir->i_op->lookup != NULL);
6452 dentry = osd_child_dentry_get(env, obj, ln->ln_name, ln->ln_namelen);
6454 if (obj->oo_hl_head != NULL) {
6455 hlock = osd_oti_get(env)->oti_hlock;
6456 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6457 dir, LDISKFS_HLOCK_LOOKUP);
6459 down_read(&obj->oo_ext_idx_sem);
6462 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6464 struct osd_thread_info *oti = osd_oti_get(env);
6465 struct osd_inode_id *id = &oti->oti_id;
6466 struct osd_device *dev = osd_obj2dev(obj);
6468 ino = le32_to_cpu(de->inode);
6469 if (CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6471 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6475 rc = osd_get_fid_from_dentry(de, rec);
6477 /* done with de, release bh */
6480 if (unlikely(is_remote_parent_ino(dev, ino))) {
6482 * If the parent is on remote MDT, and there
6483 * is no FID-in-dirent, then we have to get
6484 * the parent FID from the linkEA.
6486 if (likely(ln->ln_namelen == 2 &&
6487 ln->ln_name[0] == '.' && ln->ln_name[1] == '.'))
6488 rc = osd_get_pfid_from_linkea(env, obj,
6491 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6494 osd_id_gen(id, ino, OSD_OII_NOGEN);
6497 if (rc != 0 || osd_remote_fid(env, dev, fid))
6500 rc = osd_ldiskfs_consistency_check(oti, dev, fid, id);
6501 if (rc != -ENOENT) {
6502 /* Other error should not affect lookup result. */
6505 /* Normal file mapping should be added into OI cache
6506 * after FID in LMA check, but for local files like
6507 * hsm_actions, their FIDs are not stored in OI files,
6508 * see osd_initial_OI_scrub(), and here is the only
6509 * place to load mapping into OI cache.
6511 if (!fid_is_namespace_visible(fid))
6512 osd_add_oi_cache(osd_oti_get(env),
6513 osd_obj2dev(obj), id, fid);
6515 CDEBUG(D_INODE, DFID"/"DNAME" => "DFID"\n",
6516 PFID(lu_object_fid(&obj->oo_dt.do_lu)), PNAME(ln),
6526 ldiskfs_htree_unlock(hlock);
6528 up_read(&obj->oo_ext_idx_sem);
6532 static int osd_index_declare_ea_insert(const struct lu_env *env,
6533 struct dt_object *dt,
6534 const struct dt_rec *rec,
6535 const struct dt_key *key,
6536 struct thandle *handle)
6538 struct osd_thandle *oh;
6539 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6540 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6541 const struct lu_fid *fid = rec1->rec_fid;
6542 int credits, rc = 0;
6543 struct osd_idmap_cache *idc;
6547 LASSERT(!dt_object_remote(dt));
6548 LASSERT(handle != NULL);
6549 LASSERT(fid != NULL);
6550 LASSERT(rec1->rec_type != 0);
6552 oh = container_of(handle, struct osd_thandle, ot_super);
6553 LASSERT(oh->ot_handle == NULL);
6555 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6558 * we can't call iget() while a transactions is running
6559 * (this can lead to a deadlock), but we need to know
6560 * inum and object type. so we find this information at
6561 * declaration and cache in per-thread info
6563 idc = osd_idc_find_or_init(env, osd, fid);
6565 RETURN(PTR_ERR(idc));
6566 if (idc->oic_remote) {
6568 * a reference to remote inode is represented by an
6569 * agent inode which we have to create
6571 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6572 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6575 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6577 if (osd_dt_obj(dt)->oo_inode != NULL) {
6578 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6581 * We ignore block quota on meta pool (MDTs), so needn't
6582 * calculate how many blocks will be consumed by this index
6585 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6587 i_projid_read(inode), 0,
6588 oh, osd_dt_obj(dt), NULL,
6593 #ifdef HAVE_PROJECT_QUOTA
6595 * Reserve credits for local agent inode to transfer
6596 * to 0, quota enforcement is ignored in this case.
6598 if (idc->oic_remote &&
6599 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6600 i_projid_read(inode) != 0)
6601 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6602 0, i_projid_read(inode),
6603 0, false, PRJQUOTA);
6611 * Index add function for interoperability mode (b11826).
6612 * It will add the directory entry.This entry is needed to
6613 * maintain name->fid mapping.
6615 * \param key it is key i.e. file entry to be inserted
6616 * \param rec it is value of given key i.e. fid
6618 * \retval 0, on success
6619 * \retval -ve, on error
6621 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6622 const struct dt_rec *rec,
6623 const struct dt_key *key, struct thandle *th)
6625 struct osd_object *obj = osd_dt_obj(dt);
6626 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6627 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6628 const struct lu_fid *fid = rec1->rec_fid;
6629 const char *name = (const char *)key;
6630 struct osd_thread_info *oti = osd_oti_get(env);
6631 struct inode *child_inode = NULL;
6632 struct osd_idmap_cache *idc;
6637 if (!dt_object_exists(dt))
6640 LASSERT(osd_invariant(obj));
6641 LASSERT(!dt_object_remote(dt));
6642 LASSERT(th != NULL);
6644 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6646 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6648 idc = osd_idc_find(env, osd, fid);
6649 if (unlikely(idc == NULL)) {
6650 idc = osd_idc_find_or_init(env, osd, fid);
6653 * this dt_insert() wasn't declared properly, so
6654 * FID is missing in OI cache. we better do not
6655 * lookup FID in FLDB/OI and don't risk to deadlock,
6656 * but in some special cases (lfsck testing, etc)
6657 * it's much simpler than fixing a caller.
6659 * normally this error should be placed after the first
6660 * find, but migrate may attach source stripes to
6661 * target, which doesn't create stripes.
6663 CERROR("%s: "DFID" wasn't declared for insert\n",
6664 osd_name(osd), PFID(fid));
6666 RETURN(PTR_ERR(idc));
6670 if (idc->oic_remote) {
6671 /* Insert remote entry */
6672 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6674 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6676 child_inode = osd_create_local_agent_inode(env, osd,
6677 obj, fid, rec1->rec_type & S_IFMT, th);
6678 if (IS_ERR(child_inode))
6679 RETURN(PTR_ERR(child_inode));
6682 /* Insert local entry */
6683 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6684 /* for a reason OI cache wasn't filled properly */
6685 CERROR("%s: OIC for "DFID" isn't filled\n",
6686 osd_name(osd), PFID(fid));
6689 child_inode = oti->oti_inode;
6690 if (unlikely(child_inode == NULL)) {
6691 struct ldiskfs_inode_info *lii;
6696 child_inode = oti->oti_inode = &lii->vfs_inode;
6698 child_inode->i_sb = osd_sb(osd);
6699 child_inode->i_ino = idc->oic_lid.oii_ino;
6700 child_inode->i_mode = rec1->rec_type & S_IFMT;
6703 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6705 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6706 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6708 if (child_inode && child_inode != oti->oti_inode)
6710 LASSERT(osd_invariant(obj));
6711 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6717 * Initialize osd Iterator for given osd index object.
6719 * \param dt osd index object
6722 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6723 struct dt_object *dt,
6726 struct osd_it_iam *it;
6727 struct osd_object *obj = osd_dt_obj(dt);
6728 struct lu_object *lo = &dt->do_lu;
6729 struct iam_path_descr *ipd;
6730 struct iam_container *bag = &obj->oo_dir->od_container;
6732 if (!dt_object_exists(dt))
6733 return ERR_PTR(-ENOENT);
6737 return ERR_PTR(-ENOMEM);
6739 ipd = osd_it_ipd_get(env, bag);
6740 if (likely(ipd != NULL)) {
6744 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6745 return (struct dt_it *)it;
6748 return ERR_PTR(-ENOMEM);
6753 * free given Iterator.
6755 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6757 struct osd_it_iam *it = (struct osd_it_iam *)di;
6758 struct osd_object *obj = it->oi_obj;
6760 iam_it_fini(&it->oi_it);
6761 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6762 osd_object_put(env, obj);
6767 * Move Iterator to record specified by \a key
6769 * \param di osd iterator
6770 * \param key key for index
6772 * \retval +ve di points to record with least key not larger than key
6773 * \retval 0 di points to exact matched key
6774 * \retval -ve failure
6777 static int osd_it_iam_get(const struct lu_env *env,
6778 struct dt_it *di, const struct dt_key *key)
6780 struct osd_thread_info *oti = osd_oti_get(env);
6781 struct osd_it_iam *it = (struct osd_it_iam *)di;
6783 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6784 /* swab quota uid/gid */
6785 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6786 key = (struct dt_key *)&oti->oti_quota_id;
6789 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6795 * \param di osd iterator
6797 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6799 struct osd_it_iam *it = (struct osd_it_iam *)di;
6801 iam_it_put(&it->oi_it);
6805 * Move iterator by one record
6807 * \param di osd iterator
6809 * \retval +1 end of container reached
6811 * \retval -ve failure
6814 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6816 struct osd_it_iam *it = (struct osd_it_iam *)di;
6818 return iam_it_next(&it->oi_it);
6822 * Return pointer to the key under iterator.
6825 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6826 const struct dt_it *di)
6828 struct osd_thread_info *oti = osd_oti_get(env);
6829 struct osd_it_iam *it = (struct osd_it_iam *)di;
6830 struct osd_object *obj = it->oi_obj;
6833 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6835 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6836 /* swab quota uid/gid */
6837 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6838 key = (struct dt_key *)&oti->oti_quota_id;
6845 * Return size of key under iterator (in bytes)
6848 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6850 struct osd_it_iam *it = (struct osd_it_iam *)di;
6852 return iam_it_key_size(&it->oi_it);
6856 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6858 /* check if file type is required */
6859 if (ent->lde_attrs & LUDA_TYPE) {
6860 struct luda_type *lt;
6861 int align = sizeof(*lt) - 1;
6863 len = (len + align) & ~align;
6864 lt = (struct luda_type *)(ent->lde_name + len);
6865 lt->lt_type = cpu_to_le16(DTTOIF(type));
6868 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6872 * build lu direct from backend fs dirent.
6876 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6877 char *name, __u16 namelen, __u16 type, __u32 attr)
6879 ent->lde_attrs = attr | LUDA_FID;
6880 fid_cpu_to_le(&ent->lde_fid, fid);
6882 ent->lde_hash = cpu_to_le64(offset);
6883 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6885 strncpy(ent->lde_name, name, namelen);
6886 ent->lde_name[namelen] = '\0';
6887 ent->lde_namelen = cpu_to_le16(namelen);
6889 /* append lustre attributes */
6890 osd_it_append_attrs(ent, namelen, type);
6894 * Return pointer to the record under iterator.
6896 static int osd_it_iam_rec(const struct lu_env *env,
6897 const struct dt_it *di,
6898 struct dt_rec *dtrec, __u32 attr)
6900 struct osd_it_iam *it = (struct osd_it_iam *)di;
6901 struct osd_thread_info *info = osd_oti_get(env);
6905 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6906 const struct osd_fid_pack *rec;
6907 struct lu_fid *fid = &info->oti_fid;
6908 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6914 name = (char *)iam_it_key_get(&it->oi_it);
6916 RETURN(PTR_ERR(name));
6918 namelen = iam_it_key_size(&it->oi_it);
6920 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6922 RETURN(PTR_ERR(rec));
6924 rc = osd_fid_unpack(fid, rec);
6928 hash = iam_it_store(&it->oi_it);
6930 /* IAM does not store object type in IAM index (dir) */
6931 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6933 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6934 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6935 (struct iam_rec *)dtrec);
6936 osd_quota_unpack(it->oi_obj, dtrec);
6938 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6939 (struct iam_rec *)dtrec);
6946 * Returns cookie for current Iterator position.
6948 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6950 struct osd_it_iam *it = (struct osd_it_iam *)di;
6952 return iam_it_store(&it->oi_it);
6956 * Restore iterator from cookie.
6958 * \param di osd iterator
6959 * \param hash Iterator location cookie
6961 * \retval +ve di points to record with least key not larger than key.
6962 * \retval 0 di points to exact matched key
6963 * \retval -ve failure
6966 static int osd_it_iam_load(const struct lu_env *env,
6967 const struct dt_it *di, __u64 hash)
6969 struct osd_it_iam *it = (struct osd_it_iam *)di;
6971 return iam_it_load(&it->oi_it, hash);
6974 static const struct dt_index_operations osd_index_iam_ops = {
6975 .dio_lookup = osd_index_iam_lookup,
6976 .dio_declare_insert = osd_index_declare_iam_insert,
6977 .dio_insert = osd_index_iam_insert,
6978 .dio_declare_delete = osd_index_declare_iam_delete,
6979 .dio_delete = osd_index_iam_delete,
6981 .init = osd_it_iam_init,
6982 .fini = osd_it_iam_fini,
6983 .get = osd_it_iam_get,
6984 .put = osd_it_iam_put,
6985 .next = osd_it_iam_next,
6986 .key = osd_it_iam_key,
6987 .key_size = osd_it_iam_key_size,
6988 .rec = osd_it_iam_rec,
6989 .store = osd_it_iam_store,
6990 .load = osd_it_iam_load
6994 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
6995 struct osd_device *dev,
6996 struct inode *inode, u32 attr)
6998 struct osd_thread_info *info = osd_oti_get(env);
6999 struct osd_it_ea *oie;
7003 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
7006 RETURN(ERR_CAST(file));
7008 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
7009 if (attr & LUDA_64BITHASH)
7010 file->f_mode |= FMODE_64BITHASH;
7012 file->f_mode |= FMODE_32BITHASH;
7015 OBD_SLAB_ALLOC_PTR(oie, osd_itea_cachep);
7019 oie->oie_rd_dirent = 0;
7020 oie->oie_it_dirent = 0;
7021 oie->oie_dirent = NULL;
7022 if (unlikely(!info->oti_it_ea_buf_used)) {
7023 oie->oie_buf = info->oti_it_ea_buf;
7024 info->oti_it_ea_buf_used = 1;
7026 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7030 oie->oie_obj = NULL;
7031 oie->oie_file = file;
7036 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7040 return ERR_PTR(-ENOMEM);
7044 * Creates or initializes iterator context.
7046 * \retval struct osd_it_ea, iterator structure on success
7049 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
7050 struct dt_object *dt,
7053 struct osd_object *obj = osd_dt_obj(dt);
7054 struct osd_device *dev = osd_obj2dev(obj);
7055 struct lu_object *lo = &dt->do_lu;
7056 struct osd_it_ea *oie;
7060 if (!dt_object_exists(dt) || obj->oo_destroyed)
7061 RETURN(ERR_PTR(-ENOENT));
7063 oie = osd_it_dir_init(env, dev, obj->oo_inode, attr);
7065 RETURN(ERR_CAST(oie));
7069 RETURN((struct dt_it *)oie);
7072 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
7073 struct inode *inode)
7075 struct osd_thread_info *info = osd_oti_get(env);
7078 fput(oie->oie_file);
7079 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
7080 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7082 info->oti_it_ea_buf_used = 0;
7083 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7088 * Destroy or finishes iterator context.
7090 * \param di iterator structure to be destroyed
7092 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
7094 struct osd_it_ea *oie = (struct osd_it_ea *)di;
7095 struct osd_object *obj = oie->oie_obj;
7096 struct inode *inode = obj->oo_inode;
7099 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
7100 osd_object_put(env, obj);
7105 * It position the iterator at given key, so that next lookup continues from
7106 * that key Or it is similar to dio_it->load() but based on a key,
7107 * rather than file position.
7109 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
7112 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
7114 static int osd_it_ea_get(const struct lu_env *env,
7115 struct dt_it *di, const struct dt_key *key)
7117 struct osd_it_ea *it = (struct osd_it_ea *)di;
7120 LASSERT(((const char *)key)[0] == '\0');
7121 it->oie_file->f_pos = 0;
7122 it->oie_rd_dirent = 0;
7123 it->oie_it_dirent = 0;
7124 it->oie_dirent = NULL;
7132 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
7136 struct osd_filldir_cbs {
7137 struct dir_context ctx;
7138 struct osd_it_ea *it;
7141 * It is called internally by ->iterate*(). It fills the
7142 * iterator's in-memory data structure with required
7143 * information i.e. name, namelen, rec_size etc.
7145 * \param buf in which information to be filled in.
7146 * \param name name of the file in given dir
7148 * \retval 0 on success
7149 * \retval 1 on buffer full
7151 #ifdef HAVE_FILLDIR_USE_CTX
7152 static FILLDIR_TYPE do_osd_ldiskfs_filldir(struct dir_context *ctx,
7154 static int osd_ldiskfs_filldir(void *ctx,
7156 const char *name, int namelen,
7157 loff_t offset, __u64 ino, unsigned int d_type)
7159 struct osd_it_ea *it = ((struct osd_filldir_cbs *)ctx)->it;
7160 struct osd_object *obj = it->oie_obj;
7161 struct osd_it_ea_dirent *ent = it->oie_dirent;
7162 struct lu_fid *fid = &ent->oied_fid;
7163 char *buf = it->oie_buf;
7164 struct osd_fid_pack *rec;
7167 /* this should never happen */
7168 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
7169 CERROR("ldiskfs return invalid namelen %d\n", namelen);
7173 /* Check for enough space. Note oied_name is not NUL terminated. */
7174 if (&ent->oied_name[namelen] > buf + OSD_IT_EA_BUFSIZE)
7177 /* "." is just the object itself. */
7178 if (namelen == 1 && name[0] == '.') {
7180 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7181 } else if (d_type & LDISKFS_DIRENT_LUFID) {
7182 rec = (struct osd_fid_pack *)(name + namelen + 1);
7183 if (osd_fid_unpack(fid, rec) != 0)
7188 d_type &= ~LDISKFS_DIRENT_LUFID;
7190 /* NOT export local root. */
7192 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
7193 ino = obj->oo_inode->i_ino;
7194 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7197 if (obj == NULL || !(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
7198 ent->oied_namelen = namelen;
7199 memcpy(ent->oied_name, name, namelen);
7201 int encoded_namelen = critical_chars(name, namelen);
7203 /* Check again for enough space. */
7204 if (&ent->oied_name[encoded_namelen] > buf + OSD_IT_EA_BUFSIZE)
7207 ent->oied_namelen = encoded_namelen;
7209 if (encoded_namelen == namelen)
7210 memcpy(ent->oied_name, name, namelen);
7212 critical_encode(name, namelen, ent->oied_name);
7215 ent->oied_ino = ino;
7216 ent->oied_off = offset;
7217 ent->oied_type = d_type;
7219 it->oie_rd_dirent++;
7220 it->oie_dirent = (void *)ent + round_up(sizeof(*ent) + ent->oied_namelen, 8);
7224 WRAP_FILLDIR_FN(do_, osd_ldiskfs_filldir)
7227 * Calls ->iterate*() to load a directory entry at a time
7228 * and stored it in iterator's in-memory data structure.
7230 * \param di iterator's in memory structure
7232 * \retval 0 on success
7233 * \retval -ve on error
7234 * \retval +1 reach the end of entry
7236 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
7238 struct osd_it_ea *it = (struct osd_it_ea *)di;
7239 struct osd_object *obj = it->oie_obj;
7240 struct htree_lock *hlock = NULL;
7241 struct file *filp = it->oie_file;
7243 struct osd_filldir_cbs buf = {
7244 .ctx.actor = osd_ldiskfs_filldir,
7249 it->oie_dirent = it->oie_buf;
7250 it->oie_rd_dirent = 0;
7253 if (obj->oo_hl_head != NULL) {
7254 hlock = osd_oti_get(env)->oti_hlock;
7255 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
7257 LDISKFS_HLOCK_READDIR);
7259 down_read(&obj->oo_ext_idx_sem);
7263 rc = iterate_dir(filp, &buf.ctx);
7267 if (it->oie_rd_dirent == 0) {
7269 * If it does not get any dirent, it means it has been reached
7270 * to the end of the dir
7272 it->oie_file->f_pos = ldiskfs_get_htree_eof(it->oie_file);
7276 it->oie_dirent = it->oie_buf;
7277 it->oie_it_dirent = 1;
7282 ldiskfs_htree_unlock(hlock);
7284 up_read(&obj->oo_ext_idx_sem);
7291 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7292 * to load a directory entry at a time and stored it in
7293 * iterator's in-memory data structure.
7295 * \param di iterator's in memory structure
7297 * \retval +ve iterator reached to end
7298 * \retval 0 iterator not reached to end
7299 * \retval -ve on error
7301 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
7303 struct osd_it_ea *it = (struct osd_it_ea *)di;
7308 if (it->oie_it_dirent < it->oie_rd_dirent) {
7310 (void *)it->oie_dirent +
7311 round_up(sizeof(struct osd_it_ea_dirent) +
7312 it->oie_dirent->oied_namelen, 8);
7313 it->oie_it_dirent++;
7316 if (it->oie_file->f_pos == ldiskfs_get_htree_eof(it->oie_file))
7319 rc = osd_ldiskfs_it_fill(env, di);
7326 * Returns the key at current position from iterator's in memory structure.
7328 * \param di iterator's in memory structure
7330 * \retval key i.e. struct dt_key on success
7332 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
7333 const struct dt_it *di)
7335 struct osd_it_ea *it = (struct osd_it_ea *)di;
7337 return (struct dt_key *)it->oie_dirent->oied_name;
7341 * Returns key's size at current position from iterator's in memory structure.
7343 * \param di iterator's in memory structure
7345 * \retval key_size i.e. struct dt_key on success
7347 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
7349 struct osd_it_ea *it = (struct osd_it_ea *)di;
7351 return it->oie_dirent->oied_namelen;
7354 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
7355 #undef LDISKFS_DIR_REC_LEN
7356 # if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7357 # define LDISKFS_DIR_REC_LEN(de, dir) LDISKFS_DIR_ENTRY_LEN_((de), (dir))
7359 # define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
7363 #if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7364 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de), NULL)
7366 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de))
7369 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
7371 if (LDISKFS_DIR_REC_LEN_DIR(de) >=
7372 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
7379 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
7380 unsigned int blocksize, bool dotdot)
7383 return osd_dotdot_has_space(de);
7385 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
7386 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
7393 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
7394 handle_t *jh, struct dentry *dentry,
7395 const struct lu_fid *fid, struct buffer_head *bh,
7396 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
7399 struct inode *dir = dentry->d_parent->d_inode;
7400 struct inode *inode = dentry->d_inode;
7401 struct osd_fid_pack *rec;
7402 struct ldiskfs_dentry_param *ldp;
7403 int namelen = dentry->d_name.len;
7405 struct osd_thread_info *info = osd_oti_get(env);
7409 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7412 /* There is enough space to hold the FID-in-dirent. */
7413 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7414 rc = osd_ldiskfs_journal_get_write_access(jh, dir->i_sb, bh,
7419 de->name[namelen] = 0;
7420 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7421 rec->fp_len = sizeof(struct lu_fid) + 1;
7422 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7423 de->file_type |= LDISKFS_DIRENT_LUFID;
7424 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7431 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7435 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7436 osd_get_ldiskfs_dirent_param(ldp, fid);
7437 dentry->d_fsdata = (void *)ldp;
7438 dquot_initialize(dir);
7439 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7441 * It is too bad, we cannot reinsert the name entry back.
7442 * That means we lose it!
7446 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7447 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7448 namelen, dentry->d_name.name, PFID(fid), rc);
7454 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7455 struct osd_it_ea *it, struct lu_fid *fid,
7456 struct osd_inode_id *id, __u32 *attr)
7458 struct osd_thread_info *info = osd_oti_get(env);
7459 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7460 struct osd_device *dev = osd_obj2dev(obj);
7461 struct super_block *sb = osd_sb(dev);
7462 const char *devname = osd_name(dev);
7463 struct osd_it_ea_dirent *ent = it->oie_dirent;
7464 struct inode *dir = obj->oo_inode;
7465 struct htree_lock *hlock = NULL;
7466 struct buffer_head *bh = NULL;
7467 handle_t *jh = NULL;
7468 struct ldiskfs_dir_entry_2 *de;
7469 struct dentry *dentry;
7470 struct inode *inode;
7471 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7474 bool dotdot = false;
7480 if (ent->oied_name[0] == '.') {
7481 if (ent->oied_namelen == 1)
7484 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7488 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7489 inode = osd_iget(info, dev, id);
7490 if (IS_ERR(inode)) {
7491 rc = PTR_ERR(inode);
7492 if (rc == -ENOENT || rc == -ESTALE) {
7494 * Maybe dangling name entry, or
7495 * corrupted directory entry.
7497 *attr |= LUDA_UNKNOWN;
7500 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7501 "check_repair, dir = %lu/%u, name = %.*s, "
7502 "ino = %llu, rc = %d\n",
7503 devname, dir->i_ino, dir->i_generation,
7504 ent->oied_namelen, ent->oied_name,
7511 rc = obj_name2lu_name(obj, ent->oied_name, ent->oied_namelen, &ln);
7515 dentry = osd_child_dentry_by_inode(env, dir, ln.ln_name, ln.ln_namelen);
7516 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7517 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7523 * We need to ensure that the name entry is still valid.
7524 * Because it may be removed or renamed by other already.
7526 * The unlink or rename operation will start journal before PDO lock,
7527 * so to avoid deadlock, here we need to start journal handle before
7528 * related PDO lock also. But because we do not know whether there
7529 * will be something to be repaired before PDO lock, we just start
7530 * journal without conditions.
7532 * We may need to remove the name entry firstly, then insert back.
7533 * One credit is for user quota file update.
7534 * One credit is for group quota file update.
7535 * Two credits are for dirty inode.
7537 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7538 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7540 if (dev->od_dirent_journal != 0) {
7543 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7546 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7547 "check_repair, dir = %lu/%u, credits = %d, "
7548 "name = %.*s, ino = %llu: rc = %d\n",
7549 devname, dir->i_ino, dir->i_generation, credits,
7550 ent->oied_namelen, ent->oied_name,
7553 GOTO(out_inode, rc);
7556 if (obj->oo_hl_head != NULL) {
7557 hlock = osd_oti_get(env)->oti_hlock;
7559 * "0" means exclusive lock for the whole directory.
7560 * We need to prevent others access such name entry
7561 * during the delete + insert. Neither HLOCK_ADD nor
7562 * HLOCK_DEL cannot guarantee the atomicity.
7564 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7566 down_write(&obj->oo_ext_idx_sem);
7569 if (obj->oo_hl_head != NULL) {
7570 hlock = osd_oti_get(env)->oti_hlock;
7571 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7572 LDISKFS_HLOCK_LOOKUP);
7574 down_read(&obj->oo_ext_idx_sem);
7578 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7579 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7580 *attr |= LUDA_IGNORE;
7586 * For dotdot entry, if there is not enough space to hold the
7587 * FID-in-dirent, just keep them there. It only happens when the
7588 * device upgraded from 1.8 or restored from MDT file-level backup.
7589 * For the whole directory, only dotdot entry have no FID-in-dirent
7590 * and needs to get FID from LMA when readdir, it will not affect the
7593 if (dotdot && !osd_dotdot_has_space(de)) {
7594 *attr |= LUDA_UNKNOWN;
7600 if (lu_fid_eq(fid, &lma->lma_self_fid))
7603 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7604 struct lu_fid *tfid = &lma->lma_self_fid;
7606 if (likely(dotdot &&
7607 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7608 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7610 * It must be REMOTE_PARENT_DIR and as the
7611 * 'dotdot' entry of remote directory
7613 *attr |= LUDA_IGNORE;
7615 CDEBUG(D_LFSCK, "%s: expect remote agent "
7616 "parent directory, but got %.*s under "
7617 "dir = %lu/%u with the FID "DFID"\n",
7618 devname, ent->oied_namelen,
7619 ent->oied_name, dir->i_ino,
7620 dir->i_generation, PFID(tfid));
7622 *attr |= LUDA_UNKNOWN;
7629 if (!fid_is_zero(fid)) {
7630 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7632 if (rc == -ENOENT ||
7634 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7636 * linkEA does not recognize the dirent entry,
7637 * it may because the dirent entry corruption
7638 * and points to other's inode.
7640 CDEBUG(D_LFSCK, "%s: the target inode does not "
7641 "recognize the dirent, dir = %lu/%u, "
7642 " name = %.*s, ino = %llu, "
7643 DFID": rc = %d\n", devname, dir->i_ino,
7644 dir->i_generation, ent->oied_namelen,
7645 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7646 *attr |= LUDA_UNKNOWN;
7651 if (rc && rc != -ENODATA) {
7652 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7653 "dir = %lu/%u, name = %.*s, ino = %llu, "
7654 DFID": rc = %d\n", devname, dir->i_ino,
7655 dir->i_generation, ent->oied_namelen,
7656 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7657 *attr |= LUDA_UNKNOWN;
7665 * linkEA recognizes the dirent entry, the FID-in-LMA is
7666 * valid, trusted, in spite of fid_is_sane(fid) or not.
7668 if (*attr & LUDA_VERIFY_DRYRUN) {
7669 *fid = lma->lma_self_fid;
7670 *attr |= LUDA_REPAIR;
7677 dev->od_dirent_journal = 1;
7678 if (hlock != NULL) {
7679 ldiskfs_htree_unlock(hlock);
7682 up_read(&obj->oo_ext_idx_sem);
7688 *fid = lma->lma_self_fid;
7690 /* Update or append the FID-in-dirent. */
7691 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7692 bh, de, hlock, dotdot);
7694 *attr |= LUDA_REPAIR;
7696 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7697 "the dirent, dir = %lu/%u, name = %.*s, "
7698 "ino = %llu, "DFID": rc = %d\n",
7699 devname, dir->i_ino, dir->i_generation,
7700 ent->oied_namelen, ent->oied_name,
7701 ent->oied_ino, PFID(fid), rc);
7703 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7704 if (*attr & LUDA_VERIFY_DRYRUN) {
7705 if (fid_is_sane(fid)) {
7706 *attr |= LUDA_REPAIR;
7707 } else if (dev->od_index == 0) {
7708 lu_igif_build(fid, inode->i_ino,
7709 inode->i_generation);
7710 *attr |= LUDA_UPGRADE;
7718 dev->od_dirent_journal = 1;
7719 if (hlock != NULL) {
7720 ldiskfs_htree_unlock(hlock);
7723 up_read(&obj->oo_ext_idx_sem);
7730 if (unlikely(fid_is_sane(fid))) {
7732 * FID-in-dirent exists, but FID-in-LMA is lost.
7733 * Trust the FID-in-dirent, and add FID-in-LMA.
7735 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7737 *attr |= LUDA_REPAIR;
7739 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7740 "update dirent, dir = %lu/%u, "
7741 "name = %.*s, ino = %llu, "
7743 devname, dir->i_ino, dir->i_generation,
7744 ent->oied_namelen, ent->oied_name,
7745 ent->oied_ino, PFID(fid), rc);
7746 } else if (dev->od_index == 0) {
7747 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7749 * It is probably IGIF object. Only aappend the
7750 * FID-in-dirent. OI scrub will process FID-in-LMA.
7752 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7753 bh, de, hlock, dotdot);
7755 *attr |= LUDA_UPGRADE;
7757 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7758 "after the dirent, dir = %lu/%u, "
7759 "name = %.*s, ino = %llu, "
7761 devname, dir->i_ino, dir->i_generation,
7762 ent->oied_namelen, ent->oied_name,
7763 ent->oied_ino, PFID(fid), rc);
7772 if (hlock != NULL) {
7773 ldiskfs_htree_unlock(hlock);
7775 if (dev->od_dirent_journal != 0)
7776 up_write(&obj->oo_ext_idx_sem);
7778 up_read(&obj->oo_ext_idx_sem);
7782 ldiskfs_journal_stop(jh);
7786 if (rc >= 0 && !dirty)
7787 dev->od_dirent_journal = 0;
7788 if (ln.ln_name != ent->oied_name)
7795 * Returns the value at current position from iterator's in memory structure.
7797 * \param di struct osd_it_ea, iterator's in memory structure
7798 * \param attr attr requested for dirent.
7799 * \param lde lustre dirent
7801 * \retval 0 no error and \param lde has correct lustre dirent.
7802 * \retval -ve on error
7804 static inline int osd_it_ea_rec(const struct lu_env *env,
7805 const struct dt_it *di,
7806 struct dt_rec *dtrec, __u32 attr)
7808 struct osd_it_ea *it = (struct osd_it_ea *)di;
7809 struct osd_object *obj = it->oie_obj;
7810 struct osd_device *dev = osd_obj2dev(obj);
7811 struct osd_thread_info *oti = osd_oti_get(env);
7812 struct osd_inode_id *id = &oti->oti_id;
7813 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7814 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7815 __u32 ino = it->oie_dirent->oied_ino;
7820 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7822 if (attr & LUDA_VERIFY) {
7823 if (unlikely(is_remote_parent_ino(dev, ino))) {
7824 attr |= LUDA_IGNORE;
7826 * If the parent is on remote MDT, and there
7827 * is no FID-in-dirent, then we have to get
7828 * the parent FID from the linkEA.
7830 if (!fid_is_sane(fid) &&
7831 it->oie_dirent->oied_namelen == 2 &&
7832 it->oie_dirent->oied_name[0] == '.' &&
7833 it->oie_dirent->oied_name[1] == '.')
7834 osd_get_pfid_from_linkea(env, obj, fid);
7836 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7840 if (!fid_is_sane(fid))
7841 attr |= LUDA_UNKNOWN;
7843 attr &= ~LU_DIRENT_ATTRS_MASK;
7844 if (!fid_is_sane(fid)) {
7845 bool is_dotdot = false;
7847 if (it->oie_dirent->oied_namelen == 2 &&
7848 it->oie_dirent->oied_name[0] == '.' &&
7849 it->oie_dirent->oied_name[1] == '.')
7852 * If the parent is on remote MDT, and there
7853 * is no FID-in-dirent, then we have to get
7854 * the parent FID from the linkEA.
7856 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7857 rc = osd_get_pfid_from_linkea(env, obj, fid);
7859 if (is_dotdot == false &&
7860 CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7863 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7868 /* Pack the entry anyway, at least the offset is right. */
7869 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7870 it->oie_dirent->oied_name,
7871 it->oie_dirent->oied_namelen,
7872 it->oie_dirent->oied_type, attr);
7874 RETURN(rc > 0 ? 0 : rc);
7878 * Returns the record size size at current position.
7880 * This function will return record(lu_dirent) size in bytes.
7882 * \param[in] env execution environment
7883 * \param[in] di iterator's in memory structure
7884 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7885 * calculate the lu_dirent size.
7887 * \retval record size(in bytes & in memory) of the current lu_dirent
7890 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7893 struct osd_it_ea *it = (struct osd_it_ea *)di;
7895 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7899 * Returns a cookie for current position of the iterator head, so that
7900 * user can use this cookie to load/start the iterator next time.
7902 * \param di iterator's in memory structure
7904 * \retval cookie for current position, on success
7906 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7908 struct osd_it_ea *it = (struct osd_it_ea *)di;
7910 return it->oie_dirent->oied_off;
7914 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7915 * to load a directory entry at a time and stored it i inn,
7916 * in iterator's in-memory data structure.
7918 * \param di struct osd_it_ea, iterator's in memory structure
7920 * \retval +ve on success
7921 * \retval -ve on error
7923 static int osd_it_ea_load(const struct lu_env *env,
7924 const struct dt_it *di, __u64 hash)
7926 struct osd_it_ea *it = (struct osd_it_ea *)di;
7930 it->oie_file->f_pos = hash;
7932 rc = osd_ldiskfs_it_fill(env, di);
7942 int osd_olc_lookup(const struct lu_env *env, struct osd_object *obj,
7943 u64 iversion, struct dt_rec *rec,
7944 const struct lu_name *ln, int *result)
7946 struct osd_thread_info *oti = osd_oti_get(env);
7947 struct osd_lookup_cache *olc = oti->oti_lookup_cache;
7948 struct osd_device *osd = osd_obj2dev(obj);
7949 struct osd_lookup_cache_object *cobj = &oti->oti_cobj;
7952 if (unlikely(olc == NULL))
7955 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
7957 * umount has happened, a new OSD could land to the previous
7958 * address so we can't use it any more, invalidate our cache
7960 memset(olc, 0, sizeof(*olc));
7961 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
7965 memset(cobj, 0, sizeof(*cobj));
7966 cobj->lco_osd = osd;
7967 cobj->lco_ino = obj->oo_inode->i_ino;
7968 cobj->lco_gen = obj->oo_inode->i_generation;
7969 cobj->lco_version = iversion;
7971 for (i = 0; i < OSD_LOOKUP_CACHE_MAX; i++) {
7972 struct osd_lookup_cache_entry *entry;
7974 entry = &olc->olc_entry[i];
7975 /* compare if osd/ino/generation/version match */
7976 if (memcmp(&entry->lce_obj, cobj, sizeof(*cobj)) != 0)
7978 if (entry->lce_namelen != ln->ln_namelen)
7980 if (memcmp(entry->lce_name, ln->ln_name, ln->ln_namelen) != 0)
7983 memcpy(rec, &entry->lce_fid, sizeof(entry->lce_fid));
7984 *result = entry->lce_rc;
7990 void osd_olc_save(const struct lu_env *env, struct osd_object *obj,
7991 struct dt_rec *rec, const struct lu_name *ln,
7992 const int result, u64 iversion)
7994 struct osd_thread_info *oti = osd_oti_get(env);
7995 struct osd_lookup_cache_entry *entry;
7996 struct osd_lookup_cache *olc;
7998 if (unlikely(oti->oti_lookup_cache == NULL)) {
7999 OBD_ALLOC_PTR(oti->oti_lookup_cache);
8000 if (oti->oti_lookup_cache == NULL)
8004 olc = oti->oti_lookup_cache;
8005 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
8006 memset(olc, 0, sizeof(*olc));
8007 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
8010 entry = &olc->olc_entry[olc->olc_cur];
8012 /* invaliate cache slot if needed */
8013 if (entry->lce_obj.lco_osd)
8014 memset(&entry->lce_obj, 0, sizeof(entry->lce_obj));
8016 /* XXX: some kind of LRU */
8017 entry->lce_obj.lco_osd = osd_obj2dev(obj);
8018 entry->lce_obj.lco_ino = obj->oo_inode->i_ino;
8019 entry->lce_obj.lco_gen = obj->oo_inode->i_generation;
8020 entry->lce_obj.lco_version = iversion;
8022 LASSERT(ln->ln_namelen <= LDISKFS_NAME_LEN + 1);
8023 entry->lce_namelen = ln->ln_namelen;
8024 memcpy(entry->lce_name, ln->ln_name, ln->ln_namelen);
8025 memcpy(&entry->lce_fid, rec, sizeof(entry->lce_fid));
8026 entry->lce_rc = result;
8028 if (++olc->olc_cur == OSD_LOOKUP_CACHE_MAX)
8033 * Index lookup function for interoperability mode (b11826).
8035 * \param key, key i.e. file name to be searched
8037 * \retval +ve, on success
8038 * \retval -ve, on error
8040 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
8041 struct dt_rec *rec, const struct dt_key *key)
8043 struct osd_object *obj = osd_dt_obj(dt);
8050 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
8051 LINVRNT(osd_invariant(obj));
8053 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
8058 * grab version before actual lookup, so that we recognize potential
8059 * insert between osd_ea_lookup_rec() and osd_olc_save()
8061 iversion = inode_peek_iversion(obj->oo_inode);
8063 if (osd_olc_lookup(env, obj, iversion, rec, &ln, &result))
8064 GOTO(out, rc = result);
8066 rc = osd_ea_lookup_rec(env, obj, rec, &ln);
8070 osd_olc_save(env, obj, rec, &ln, rc, iversion);
8073 if (ln.ln_name != (char *)key)
8079 * Index and Iterator operations for interoperability
8080 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
8082 static const struct dt_index_operations osd_index_ea_ops = {
8083 .dio_lookup = osd_index_ea_lookup,
8084 .dio_declare_insert = osd_index_declare_ea_insert,
8085 .dio_insert = osd_index_ea_insert,
8086 .dio_declare_delete = osd_index_declare_ea_delete,
8087 .dio_delete = osd_index_ea_delete,
8089 .init = osd_it_ea_init,
8090 .fini = osd_it_ea_fini,
8091 .get = osd_it_ea_get,
8092 .put = osd_it_ea_put,
8093 .next = osd_it_ea_next,
8094 .key = osd_it_ea_key,
8095 .key_size = osd_it_ea_key_size,
8096 .rec = osd_it_ea_rec,
8097 .rec_size = osd_it_ea_rec_size,
8098 .store = osd_it_ea_store,
8099 .load = osd_it_ea_load
8103 static void *osd_key_init(const struct lu_context *ctx,
8104 struct lu_context_key *key)
8106 struct osd_thread_info *info;
8108 OBD_ALLOC_PTR(info);
8110 return ERR_PTR(-ENOMEM);
8112 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8113 if (info->oti_it_ea_buf == NULL)
8116 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
8118 info->oti_hlock = ldiskfs_htree_lock_alloc();
8119 if (info->oti_hlock == NULL)
8125 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8128 return ERR_PTR(-ENOMEM);
8131 static void osd_key_fini(const struct lu_context *ctx,
8132 struct lu_context_key *key, void *data)
8134 struct osd_thread_info *info = data;
8135 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
8136 struct osd_idmap_cache *idc = info->oti_ins_cache;
8138 if (info->oti_dio_pages) {
8140 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
8141 struct page *page = info->oti_dio_pages[i];
8143 LASSERT(PagePrivate2(page));
8144 LASSERT(PageLocked(page));
8145 ClearPagePrivate2(page);
8150 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
8151 PTLRPC_MAX_BRW_PAGES);
8154 if (info->oti_inode != NULL)
8156 if (info->oti_hlock != NULL)
8157 ldiskfs_htree_lock_free(info->oti_hlock);
8158 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8159 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
8160 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
8161 lu_buf_free(&info->oti_big_buf);
8163 LASSERT(info->oti_ins_cache_size > 0);
8164 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
8165 info->oti_ins_cache = NULL;
8166 info->oti_ins_cache_size = 0;
8168 if (info->oti_lookup_cache)
8169 OBD_FREE_PTR(info->oti_lookup_cache);
8173 static void osd_key_exit(const struct lu_context *ctx,
8174 struct lu_context_key *key, void *data)
8176 struct osd_thread_info *info = data;
8177 struct osd_lookup_cache *olc = info->oti_lookup_cache;
8180 memset(olc, 0, sizeof(*olc));
8181 LASSERT(info->oti_r_locks == 0);
8182 LASSERT(info->oti_w_locks == 0);
8183 LASSERT(info->oti_txns == 0);
8184 LASSERTF(info->oti_dio_pages_used == 0, "%d\n",
8185 info->oti_dio_pages_used);
8188 /* type constructor/destructor: osd_type_init, osd_type_fini */
8189 LU_TYPE_INIT_FINI(osd, &osd_key);
8191 struct lu_context_key osd_key = {
8192 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
8193 .lct_init = osd_key_init,
8194 .lct_fini = osd_key_fini,
8195 .lct_exit = osd_key_exit
8199 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
8200 const char *name, struct lu_device *next)
8202 struct osd_device *osd = osd_dev(d);
8204 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
8205 sizeof(osd->od_svname))
8207 return osd_procfs_init(osd, name);
8210 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
8212 struct seq_server_site *ss = osd_seq_site(osd);
8217 if (osd->od_is_ost || osd->od_cl_seq != NULL)
8220 if (unlikely(ss == NULL))
8223 OBD_ALLOC_PTR(osd->od_cl_seq);
8224 if (osd->od_cl_seq == NULL)
8227 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
8228 osd->od_svname, ss->ss_server_seq);
8230 if (ss->ss_node_id == 0) {
8232 * If the OSD on the sequence controller(MDT0), then allocate
8233 * sequence here, otherwise allocate sequence after connected
8234 * to MDT0 (see mdt_register_lwp_callback()).
8236 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
8237 &osd->od_cl_seq->lcs_space, env);
8243 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
8245 if (osd->od_cl_seq == NULL)
8248 seq_client_fini(osd->od_cl_seq);
8249 OBD_FREE_PTR(osd->od_cl_seq);
8250 osd->od_cl_seq = NULL;
8253 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
8257 /* shutdown quota slave instance associated with the device */
8258 if (o->od_quota_slave_md != NULL) {
8259 struct qsd_instance *qsd = o->od_quota_slave_md;
8261 o->od_quota_slave_md = NULL;
8265 if (o->od_quota_slave_dt != NULL) {
8266 struct qsd_instance *qsd = o->od_quota_slave_dt;
8268 o->od_quota_slave_dt = NULL;
8272 osd_fid_fini(env, o);
8273 osd_scrub_cleanup(env, o);
8278 #ifdef HAVE_FLUSH_DELAYED_FPUT
8279 # define cfs_flush_delayed_fput() flush_delayed_fput()
8281 void (*cfs_flush_delayed_fput)(void);
8282 #endif /* HAVE_FLUSH_DELAYED_FPUT */
8284 static void osd_umount(const struct lu_env *env, struct osd_device *o)
8288 atomic_inc(&osd_mount_seq);
8290 if (o->od_mnt != NULL) {
8291 shrink_dcache_sb(osd_sb(o));
8292 osd_sync(env, &o->od_dt_dev);
8293 wait_event(o->od_commit_cb_done,
8294 !atomic_read(&o->od_commit_cb_in_flight));
8300 /* to be sure all delayed fput are finished */
8301 cfs_flush_delayed_fput();
8306 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8307 # ifndef LDISKFS_HAS_INCOMPAT_FEATURE
8308 /* Newer kernels provide the ldiskfs_set_feature_largedir() wrapper already,
8309 * which calls ldiskfs_update_dynamic_rev() to update ancient filesystems.
8310 * All ldiskfs filesystems are already v2, so it is a no-op and unnecessary.
8311 * This avoids maintaining patches to export this otherwise-useless function.
8313 void ldiskfs_update_dynamic_rev(struct super_block *sb)
8320 static int osd_mount(const struct lu_env *env,
8321 struct osd_device *o, struct lustre_cfg *cfg)
8323 const char *name = lustre_cfg_string(cfg, 0);
8324 const char *dev = lustre_cfg_string(cfg, 1);
8326 unsigned long page, s_flags = 0, lmd_flags = 0;
8327 struct page *__page;
8328 struct file_system_type *type;
8329 char *options = NULL;
8331 struct osd_thread_info *info = osd_oti_get(env);
8332 struct lu_fid *fid = &info->oti_fid;
8333 struct inode *inode;
8334 int rc = 0, force_over_1024tb = 0;
8338 if (o->od_mnt != NULL)
8341 if (strlen(dev) >= sizeof(o->od_mntdev))
8343 strcpy(o->od_mntdev, dev);
8345 str = lustre_cfg_buf(cfg, 2);
8346 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
8348 opts = lustre_cfg_string(cfg, 3);
8350 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
8351 CERROR("%s: device %s extents feature is not guaranteed to "
8352 "work on big-endian systems. Use \"bigendian_extents\" "
8353 "mount option to override.\n", name, dev);
8357 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
8358 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
8359 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");
8362 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
8363 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
8364 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");
8367 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8368 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
8369 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");
8373 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
8374 force_over_1024tb = 1;
8376 __page = alloc_page(GFP_KERNEL);
8378 GOTO(out, rc = -ENOMEM);
8379 page = (unsigned long)page_address(__page);
8380 options = (char *)page;
8383 /* strip out the options for back compatiblity */
8384 static const char * const sout[] = {
8391 /* strip out option we processed in osd */
8392 "bigendian_extents",
8396 "force_over_1024tb",
8400 strncat(options, opts, PAGE_SIZE);
8401 for (rc = 0, str = options; sout[rc]; ) {
8402 char *op = strstr(str, sout[rc]);
8409 if (op == options || *(op - 1) == ',') {
8410 str = op + strlen(sout[rc]);
8411 if (*str == ',' || *str == '\0') {
8412 *str == ',' ? str++ : str;
8413 memmove(op, str, strlen(str) + 1);
8416 for (str = op; *str != ',' && *str != '\0'; str++)
8420 strncat(options, "user_xattr,acl", PAGE_SIZE);
8423 /* Glom up mount options */
8424 if (*options != '\0')
8425 strncat(options, ",", PAGE_SIZE);
8426 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
8428 type = get_fs_type("ldiskfs");
8430 CERROR("%s: cannot find ldiskfs module\n", name);
8431 GOTO(out, rc = -ENODEV);
8434 s_flags |= SB_KERNMOUNT;
8435 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
8436 module_put(type->owner);
8438 if (IS_ERR(o->od_mnt)) {
8439 rc = PTR_ERR(o->od_mnt);
8441 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
8445 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
8446 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
8447 force_over_1024tb == 0) {
8448 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",
8450 GOTO(out_mnt, rc = -EINVAL);
8453 if (test_bit(LMD_FLG_DEV_RDONLY, &lmd_flags)) {
8454 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
8457 GOTO(out_mnt, rc = -EOPNOTSUPP);
8460 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
8461 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
8462 GOTO(out_mnt, rc = -EINVAL);
8465 if (ldiskfs_has_feature_fast_commit(o->od_mnt->mnt_sb)) {
8466 CERROR("%s: device %s is mounted with fast_commit that breaks recovery\n",
8468 GOTO(out_mnt, rc = -EOPNOTSUPP);
8471 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8472 #ifdef LDISKFS_MOUNT_DIRDATA
8473 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
8474 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
8475 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
8476 CWARN("%s: device %s was upgraded from Lustre-1.x without "
8477 "enabling the dirdata feature. If you do not want to "
8478 "downgrade to Lustre-1.x again, you can enable it via "
8479 "'tune2fs -O dirdata device'\n", name, dev);
8481 /* enable large_dir on MDTs to avoid REMOTE_PARENT_DIR overflow,
8482 * and on very large OSTs to avoid object directory overflow */
8483 if (unlikely(!ldiskfs_has_feature_largedir(o->od_mnt->mnt_sb) &&
8484 !strstr(name, "MGS"))) {
8485 ldiskfs_set_feature_largedir(o->od_mnt->mnt_sb);
8486 LCONSOLE_INFO("%s: enabled 'large_dir' feature on device %s\n",
8490 inode = osd_sb(o)->s_root->d_inode;
8491 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
8492 if (!o->od_dt_dev.dd_rdonly) {
8493 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
8495 CERROR("%s: failed to set lma on %s root inode\n",
8501 if (test_bit(LMD_FLG_NOSCRUB, &lmd_flags))
8502 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_NEVER;
8504 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
8505 /* do not use pagecache with flash-backed storage */
8506 o->od_writethrough_cache = 0;
8507 o->od_read_cache = 0;
8518 __free_page(__page);
8523 static struct lu_device *osd_device_fini(const struct lu_env *env,
8524 struct lu_device *d)
8526 struct osd_device *o = osd_dev(d);
8530 osd_index_backup(env, o, false);
8531 osd_shutdown(env, o);
8533 if (o->od_oi_table != NULL)
8534 osd_oi_fini(osd_oti_get(env), o);
8535 if (o->od_extent_bytes_percpu)
8536 free_percpu(o->od_extent_bytes_percpu);
8537 osd_obj_map_fini(o);
8543 static int osd_device_init0(const struct lu_env *env,
8544 struct osd_device *o,
8545 struct lustre_cfg *cfg)
8547 struct lu_device *l = osd2lu_dev(o);
8548 struct osd_thread_info *info;
8551 bool restored = false;
8554 /* if the module was re-loaded, env can loose its keys */
8555 rc = lu_env_refill((struct lu_env *)env);
8558 info = osd_oti_get(env);
8561 l->ld_ops = &osd_lu_ops;
8562 o->od_dt_dev.dd_ops = &osd_dt_ops;
8564 spin_lock_init(&o->od_osfs_lock);
8565 mutex_init(&o->od_otable_mutex);
8566 INIT_LIST_HEAD(&o->od_orphan_list);
8567 INIT_LIST_HEAD(&o->od_index_backup_list);
8568 INIT_LIST_HEAD(&o->od_index_restore_list);
8569 spin_lock_init(&o->od_lock);
8570 o->od_index_backup_policy = LIBP_NONE;
8572 init_waitqueue_head(&o->od_commit_cb_done);
8574 o->od_read_cache = 1;
8575 o->od_writethrough_cache = 1;
8576 o->od_enable_projid_xattr = 0;
8577 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
8578 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
8579 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
8580 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_DEFAULT;
8581 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8582 o->od_fallocate_zero_blocks = 0;
8584 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
8585 sizeof(o->od_svname));
8586 if (cplen >= sizeof(o->od_svname)) {
8591 o->od_index_backup_stop = 0;
8592 o->od_index = -1; /* -1 means index is invalid */
8593 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8594 if (rc == LDD_F_SV_TYPE_OST)
8597 o->od_full_scrub_ratio = OFSR_DEFAULT;
8598 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8599 rc = osd_mount(env, o, cfg);
8603 /* Can only check block device after mount */
8604 o->od_nonrotational =
8605 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8607 rc = osd_obj_map_init(env, o);
8611 rc = lu_site_init(&o->od_site, l);
8613 GOTO(out_compat, rc);
8614 o->od_site.ls_bottom_dev = l;
8616 rc = lu_site_init_finish(&o->od_site);
8620 opts = lustre_cfg_string(cfg, 3);
8621 if (opts && strstr(opts, "resetoi"))
8624 INIT_LIST_HEAD(&o->od_ios_list);
8626 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
8628 GOTO(out_brw_stats, rc);
8630 /* setup scrub, including OI files initialization */
8632 rc = osd_scrub_setup(env, o, restored);
8635 GOTO(out_brw_stats, rc);
8637 rc = osd_procfs_init(o, o->od_svname);
8639 CERROR("%s: can't initialize procfs: rc = %d\n",
8641 GOTO(out_scrub, rc);
8644 LASSERT(l->ld_site->ls_linkage.next != NULL);
8645 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8647 /* initialize quota slave instance */
8648 /* currently it's no need to prepare qsd_instance_md for OST */
8649 if (!o->od_is_ost) {
8650 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8651 &o->od_dt_dev, o->od_proc_entry,
8653 if (IS_ERR(o->od_quota_slave_md)) {
8654 rc = PTR_ERR(o->od_quota_slave_md);
8655 o->od_quota_slave_md = NULL;
8656 GOTO(out_procfs, rc);
8660 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8661 o->od_proc_entry, false, true);
8663 if (IS_ERR(o->od_quota_slave_dt)) {
8664 if (o->od_quota_slave_md != NULL) {
8665 qsd_fini(env, o->od_quota_slave_md);
8666 o->od_quota_slave_md = NULL;
8669 rc = PTR_ERR(o->od_quota_slave_dt);
8670 o->od_quota_slave_dt = NULL;
8671 GOTO(out_procfs, rc);
8674 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8675 if (!o->od_extent_bytes_percpu) {
8677 GOTO(out_procfs, rc);
8685 osd_scrub_cleanup(env, o);
8687 lprocfs_fini_brw_stats(&o->od_brw_stats);
8689 lu_site_fini(&o->od_site);
8691 osd_obj_map_fini(o);
8698 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8699 struct lu_device_type *t,
8700 struct lustre_cfg *cfg)
8702 struct osd_device *o;
8707 return ERR_PTR(-ENOMEM);
8709 rc = dt_device_init(&o->od_dt_dev, t);
8712 * Because the ctx might be revived in dt_device_init,
8713 * refill the env here
8715 lu_env_refill((struct lu_env *)env);
8716 rc = osd_device_init0(env, o, cfg);
8718 dt_device_fini(&o->od_dt_dev);
8721 if (unlikely(rc != 0))
8724 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8727 static struct lu_device *osd_device_free(const struct lu_env *env,
8728 struct lu_device *d)
8730 struct osd_device *o = osd_dev(d);
8734 /* XXX: make osd top device in order to release reference */
8735 d->ld_site->ls_top_dev = d;
8736 lu_site_purge(env, d->ld_site, -1);
8737 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8738 D_ERROR, lu_cdebug_printer);
8739 lu_site_fini(&o->od_site);
8740 dt_device_fini(&o->od_dt_dev);
8745 static int osd_process_config(const struct lu_env *env,
8746 struct lu_device *d, struct lustre_cfg *cfg)
8748 struct osd_device *o = osd_dev(d);
8754 switch (cfg->lcfg_command) {
8756 rc = osd_mount(env, o, cfg);
8760 * For the case LCFG_PRE_CLEANUP is not called in advance,
8761 * that may happend if hit failure during mount process.
8763 osd_index_backup(env, o, false);
8764 lu_dev_del_linkage(d->ld_site, d);
8765 rc = osd_shutdown(env, o);
8768 LASSERT(&o->od_dt_dev);
8769 count = class_modify_config(cfg, PARAM_OSD,
8770 &o->od_dt_dev.dd_kobj);
8772 count = class_modify_config(cfg, PARAM_OST,
8773 &o->od_dt_dev.dd_kobj);
8774 rc = count > 0 ? 0 : count;
8776 case LCFG_PRE_CLEANUP:
8778 osd_index_backup(env, o,
8779 o->od_index_backup_policy != LIBP_NONE);
8789 static int osd_recovery_complete(const struct lu_env *env,
8790 struct lu_device *d)
8792 struct osd_device *osd = osd_dev(d);
8797 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8801 * start qsd instance on recovery completion, this notifies the quota
8802 * slave code that we are about to process new requests now
8804 rc = qsd_start(env, osd->od_quota_slave_dt);
8805 if (rc == 0 && osd->od_quota_slave_md != NULL)
8806 rc = qsd_start(env, osd->od_quota_slave_md);
8812 * we use exports to track all osd users
8814 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8815 struct obd_device *obd, struct obd_uuid *cluuid,
8816 struct obd_connect_data *data, void *localdata)
8818 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8819 struct lustre_handle conn;
8824 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8826 rc = class_connect(&conn, obd, cluuid);
8830 *exp = class_conn2export(&conn);
8832 spin_lock(&osd->od_osfs_lock);
8834 spin_unlock(&osd->od_osfs_lock);
8840 * once last export (we don't count self-export) disappeared
8841 * osd can be released
8843 static int osd_obd_disconnect(struct obd_export *exp)
8845 struct obd_device *obd = exp->exp_obd;
8846 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8847 int rc, release = 0;
8851 /* Only disconnect the underlying layers on the final disconnect. */
8852 spin_lock(&osd->od_osfs_lock);
8854 if (osd->od_connects == 0)
8856 spin_unlock(&osd->od_osfs_lock);
8858 rc = class_disconnect(exp); /* bz 9811 */
8860 if (rc == 0 && release)
8861 class_manual_cleanup(obd);
8865 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8866 struct lu_device *dev)
8868 struct osd_device *osd = osd_dev(dev);
8869 struct lr_server_data *lsd =
8870 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8875 if (osd->od_quota_slave_md != NULL) {
8876 /* set up quota slave objects for inode */
8877 result = qsd_prepare(env, osd->od_quota_slave_md);
8882 if (osd->od_quota_slave_dt != NULL) {
8883 /* set up quota slave objects for block */
8884 result = qsd_prepare(env, osd->od_quota_slave_dt);
8890 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8891 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8892 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8893 osd->od_index_in_idif = 1;
8895 osd->od_index_in_idif = 0;
8896 result = osd_register_proc_index_in_idif(osd);
8901 osd->od_index_in_idif = 1;
8905 result = osd_fid_init(env, osd);
8911 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8915 * see include/lu_object.h for the details.
8917 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8918 struct lu_fid *fid, struct lu_object *parent,
8919 const struct lu_name *name)
8921 struct osd_device *osd = osd_dev(d);
8923 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8926 static const struct lu_object_operations osd_lu_obj_ops = {
8927 .loo_object_init = osd_object_init,
8928 .loo_object_delete = osd_object_delete,
8929 .loo_object_release = osd_object_release,
8930 .loo_object_free = osd_object_free,
8931 .loo_object_print = osd_object_print,
8932 .loo_object_invariant = osd_object_invariant
8935 const struct lu_device_operations osd_lu_ops = {
8936 .ldo_object_alloc = osd_object_alloc,
8937 .ldo_process_config = osd_process_config,
8938 .ldo_recovery_complete = osd_recovery_complete,
8939 .ldo_prepare = osd_prepare,
8940 .ldo_fid_alloc = osd_fid_alloc,
8943 static const struct lu_device_type_operations osd_device_type_ops = {
8944 .ldto_init = osd_type_init,
8945 .ldto_fini = osd_type_fini,
8947 .ldto_start = osd_type_start,
8948 .ldto_stop = osd_type_stop,
8950 .ldto_device_alloc = osd_device_alloc,
8951 .ldto_device_free = osd_device_free,
8953 .ldto_device_init = osd_device_init,
8954 .ldto_device_fini = osd_device_fini
8957 static struct lu_device_type osd_device_type = {
8958 .ldt_tags = LU_DEVICE_DT,
8959 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8960 .ldt_ops = &osd_device_type_ops,
8961 .ldt_ctx_tags = LCT_LOCAL,
8964 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8966 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8967 struct super_block *sb = osd_sb(osd);
8969 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8973 * lprocfs legacy support.
8975 static const struct obd_ops osd_obd_device_ops = {
8976 .o_owner = THIS_MODULE,
8977 .o_connect = osd_obd_connect,
8978 .o_disconnect = osd_obd_disconnect,
8979 .o_health_check = osd_health_check,
8982 static ssize_t delayed_unlink_mb_show(struct kobject *kobj,
8983 struct attribute *attr, char *buf)
8985 return snprintf(buf, PAGE_SIZE, "%d\n",
8986 ldiskfs_delayed_unlink_blocks >> 11);
8989 static ssize_t delayed_unlink_mb_store(struct kobject *kobj,
8990 struct attribute *attr,
8991 const char *buffer, size_t count)
8993 u64 delayed_unlink_bytes;
8996 rc = sysfs_memparse(buffer, count, &delayed_unlink_bytes, "MiB");
9000 ldiskfs_delayed_unlink_blocks = delayed_unlink_bytes >> 9;
9004 LUSTRE_RW_ATTR(delayed_unlink_mb);
9007 static ssize_t track_declares_assert_show(struct kobject *kobj,
9008 struct attribute *attr,
9011 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
9014 static ssize_t track_declares_assert_store(struct kobject *kobj,
9015 struct attribute *attr,
9016 const char *buffer, size_t count)
9018 bool track_declares_assert;
9021 rc = kstrtobool(buffer, &track_declares_assert);
9025 ldiskfs_track_declares_assert = track_declares_assert;
9029 LUSTRE_RW_ATTR(track_declares_assert);
9031 static int __init osd_init(void)
9033 struct kobject *kobj;
9036 BUILD_BUG_ON(BH_DXLock >=
9037 sizeof(((struct buffer_head *)0)->b_state) * 8);
9038 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
9039 /* please, try to keep osd_thread_info smaller than a page */
9040 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
9043 rc = libcfs_setup();
9049 rc = lu_kmem_init(ldiskfs_caches);
9053 rc = class_register_type(&osd_obd_device_ops, NULL, true,
9054 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
9056 lu_kmem_fini(ldiskfs_caches);
9060 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9062 rc = sysfs_create_file(kobj,
9063 &lustre_attr_track_declares_assert.attr);
9065 CWARN("%s: track_declares_assert sysfs registration failed: rc = %d\n",
9070 rc = sysfs_create_file(kobj,
9071 &lustre_attr_delayed_unlink_mb.attr);
9073 CWARN("%s: delayed_unlink_mb registration failed: rc = %d\n",
9081 #ifndef HAVE_FLUSH_DELAYED_FPUT
9082 if (unlikely(cfs_flush_delayed_fput == NULL))
9083 cfs_flush_delayed_fput =
9084 cfs_kallsyms_lookup_name("flush_delayed_fput");
9090 static void __exit osd_exit(void)
9092 struct kobject *kobj;
9094 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9096 sysfs_remove_file(kobj,
9097 &lustre_attr_track_declares_assert.attr);
9100 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
9101 lu_kmem_fini(ldiskfs_caches);
9104 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
9105 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
9106 MODULE_VERSION(LUSTRE_VERSION_STRING);
9107 MODULE_LICENSE("GPL");
9109 module_init(osd_init);
9110 module_exit(osd_exit);