4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/osd/osd_handler.c
33 * Top-level entry points into osd module
35 * Author: Nikita Danilov <nikita@clusterfs.com>
36 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
39 #define DEBUG_SUBSYSTEM S_OSD
41 #include <linux/fs_struct.h>
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #include <linux/uidgid.h>
45 #ifdef HAVE_INODE_IVERSION
46 #include <linux/iversion.h>
48 #define inode_peek_iversion(__inode) ((__inode)->i_version)
51 /* prerequisite for linux/xattr.h */
52 #include <linux/types.h>
53 /* prerequisite for linux/xattr.h */
55 /* XATTR_{REPLACE,CREATE} */
56 #include <linux/xattr.h>
58 #include <ldiskfs/ldiskfs.h>
59 #include <ldiskfs/xattr.h>
60 #include <ldiskfs/ldiskfs_extents.h>
63 * struct OBD_{ALLOC,FREE}*()
65 #include <obd_support.h>
66 #include <libcfs/libcfs.h>
67 /* struct ptlrpc_thread */
68 #include <lustre_net.h>
69 #include <lustre_fid.h>
71 #include <uapi/linux/lustre/lustre_param.h>
72 #include <uapi/linux/lustre/lustre_disk.h>
74 #include "osd_internal.h"
75 #include "osd_dynlocks.h"
77 /* llo_* api support */
78 #include <md_object.h>
79 #include <lustre_quota.h>
81 #include <lustre_linkea.h>
83 /* encoding routines */
84 #include <lustre_crypto.h>
86 /* Maximum EA size is limited by LNET_MTU for remote objects */
87 #define OSD_MAX_EA_SIZE 1048364
90 module_param(ldiskfs_pdo, int, 0644);
91 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
93 int ldiskfs_track_declares_assert;
94 module_param(ldiskfs_track_declares_assert, int, 0644);
95 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
97 /* 1 GiB in 512-byte sectors */
98 int ldiskfs_delayed_unlink_blocks = (1 << (30 - 9));
100 /* Slab to allocate dynlocks */
101 struct kmem_cache *dynlock_cachep;
103 /* Slab to allocate osd_it_ea */
104 struct kmem_cache *osd_itea_cachep;
106 static struct lu_kmem_descr ldiskfs_caches[] = {
108 .ckd_cache = &biop_cachep,
109 .ckd_name = "biop_cache",
110 .ckd_size = sizeof(struct osd_bio_private)
113 .ckd_cache = &dynlock_cachep,
114 .ckd_name = "dynlock_cache",
115 .ckd_size = sizeof(struct dynlock_handle)
118 .ckd_cache = &osd_itea_cachep,
119 .ckd_name = "osd_itea_cache",
120 .ckd_size = sizeof(struct osd_it_ea)
127 static atomic_t osd_mount_seq;
129 static const char dot[] = ".";
130 static const char dotdot[] = "..";
132 static const struct lu_object_operations osd_lu_obj_ops;
133 static const struct dt_object_operations osd_obj_ops;
134 static const struct dt_object_operations osd_obj_otable_it_ops;
135 static const struct dt_index_operations osd_index_iam_ops;
136 static const struct dt_index_operations osd_index_ea_ops;
138 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
139 const struct lu_fid *fid);
140 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
141 struct osd_device *osd);
142 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
143 const struct lu_buf *buf, const char *name, int fl,
144 struct thandle *handle);
146 int osd_trans_declare_op2rb[] = {
147 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
148 [OSD_OT_PUNCH] = OSD_OT_MAX,
149 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
150 [OSD_OT_CREATE] = OSD_OT_DESTROY,
151 [OSD_OT_DESTROY] = OSD_OT_CREATE,
152 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
153 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
154 [OSD_OT_WRITE] = OSD_OT_WRITE,
155 [OSD_OT_INSERT] = OSD_OT_DELETE,
156 [OSD_OT_DELETE] = OSD_OT_INSERT,
157 [OSD_OT_QUOTA] = OSD_OT_MAX,
160 static int osd_has_index(const struct osd_object *obj)
162 return obj->oo_dt.do_index_ops != NULL;
165 static int osd_object_invariant(const struct lu_object *l)
167 return osd_invariant(osd_obj(l));
171 * Concurrency: doesn't matter
173 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
175 struct osd_thread_info *oti = osd_oti_get(env);
177 return oti->oti_w_locks > 0 && o->oo_owner == env;
181 * Concurrency: doesn't access mutable data
183 static int osd_root_get(const struct lu_env *env,
184 struct dt_device *dev, struct lu_fid *f)
186 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
191 * the following set of functions are used to maintain per-thread
192 * cache of FID->ino mapping. this mechanism is needed to resolve
193 * FID to inode at dt_insert() which in turn stores ino in the
194 * directory entries to keep ldiskfs compatible with ext[34].
195 * due to locking-originated restrictions we can't lookup ino
196 * using LU cache (deadlock is possible). lookup using OI is quite
197 * expensive. so instead we maintain this cache and methods like
198 * dt_create() fill it. so in the majority of cases dt_insert() is
199 * able to find needed mapping in lockless manner.
201 static struct osd_idmap_cache *
202 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
203 const struct lu_fid *fid)
205 struct osd_thread_info *oti = osd_oti_get(env);
206 struct osd_idmap_cache *idc = oti->oti_ins_cache;
209 for (i = 0; i < oti->oti_ins_cache_used; i++) {
210 if (!lu_fid_eq(&idc[i].oic_fid, fid))
212 if (idc[i].oic_dev != osd)
221 static struct osd_idmap_cache *
222 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
223 const struct lu_fid *fid)
225 struct osd_thread_info *oti = osd_oti_get(env);
226 struct osd_idmap_cache *idc;
229 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
230 i = oti->oti_ins_cache_size * 2;
232 i = OSD_INS_CACHE_SIZE;
233 OBD_ALLOC_PTR_ARRAY_LARGE(idc, i);
235 return ERR_PTR(-ENOMEM);
236 if (oti->oti_ins_cache != NULL) {
237 memcpy(idc, oti->oti_ins_cache,
238 oti->oti_ins_cache_used * sizeof(*idc));
239 OBD_FREE_PTR_ARRAY_LARGE(oti->oti_ins_cache,
240 oti->oti_ins_cache_used);
242 oti->oti_ins_cache = idc;
243 oti->oti_ins_cache_size = i;
246 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
249 idc->oic_lid.oii_ino = 0;
250 idc->oic_lid.oii_gen = 0;
257 * lookup mapping for the given fid in the cache, initialize a
258 * new one if not found. the initialization checks whether the
259 * object is local or remote. for local objects, OI is used to
260 * learn ino/generation. the function is used when the caller
261 * has no information about the object, e.g. at dt_insert().
263 static struct osd_idmap_cache *
264 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
265 const struct lu_fid *fid)
267 struct osd_idmap_cache *idc;
270 idc = osd_idc_find(env, osd, fid);
271 LASSERT(!IS_ERR(idc));
275 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
276 osd->od_svname, PFID(fid));
278 /* new mapping is needed */
279 idc = osd_idc_add(env, osd, fid);
281 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
282 osd->od_svname, PFID(fid), PTR_ERR(idc));
287 rc = osd_remote_fid(env, osd, fid);
288 if (unlikely(rc < 0))
292 /* the object is local, lookup in OI */
293 /* XXX: probably cheaper to lookup in LU first? */
294 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
296 if (unlikely(rc < 0)) {
297 CERROR("can't lookup: rc = %d\n", rc);
301 /* the object is remote */
308 static void osd_idc_dump_lma(const struct lu_env *env,
309 struct osd_device *osd,
313 struct osd_thread_info *info = osd_oti_get(env);
314 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
315 const struct lu_fid *fid;
316 struct osd_inode_id lid;
320 inode = osd_ldiskfs_iget(osd_sb(osd), ino);
322 CERROR("%s: can't get inode %lu: rc = %d\n",
323 osd->od_svname, ino, (int)PTR_ERR(inode));
326 if (is_bad_inode(inode)) {
327 CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
330 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
332 CERROR("%s: can't get LMA for %lu: rc = %d\n",
333 osd->od_svname, ino, rc);
336 fid = &loa->loa_lma.lma_self_fid;
337 LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
338 PFID(fid), ino, (unsigned)inode->i_generation);
341 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
343 CERROR("%s: can't lookup "DFID": rc = %d\n",
344 osd->od_svname, PFID(fid), rc);
347 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
348 PFID(fid), lid.oii_ino, lid.oii_gen);
353 static void osd_idc_dump_debug(const struct lu_env *env,
354 struct osd_device *osd,
355 const struct lu_fid *fid,
359 struct osd_inode_id lid;
363 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
365 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
366 osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
367 osd_idc_dump_lma(env, osd, lid.oii_ino, false);
369 CERROR("%s: can't lookup "DFID": rc = %d\n",
370 osd->od_svname, PFID(fid), rc);
373 osd_idc_dump_lma(env, osd, ino1, true);
375 osd_idc_dump_lma(env, osd, ino2, true);
379 * lookup mapping for given FID and fill it from the given object.
380 * the object is lolcal by definition.
382 static int osd_idc_find_and_init(const struct lu_env *env,
383 struct osd_device *osd,
384 struct osd_object *obj)
386 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
387 struct osd_idmap_cache *idc;
389 idc = osd_idc_find(env, osd, fid);
390 LASSERT(!IS_ERR(idc));
392 if (obj->oo_inode == NULL)
394 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
395 if (idc->oic_lid.oii_ino) {
396 osd_idc_dump_debug(env, osd, fid,
397 idc->oic_lid.oii_ino,
398 obj->oo_inode->i_ino);
401 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
402 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
407 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
408 osd->od_svname, PFID(fid));
410 /* new mapping is needed */
411 idc = osd_idc_add(env, osd, fid);
413 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
414 osd->od_svname, PFID(fid), PTR_ERR(idc));
418 if (obj->oo_inode != NULL) {
419 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
420 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
426 * OSD object methods.
430 * Concurrency: no concurrent access is possible that early in object
433 static struct lu_object *osd_object_alloc(const struct lu_env *env,
434 const struct lu_object_header *hdr,
437 struct osd_object *mo;
442 struct lu_object_header *h;
443 struct osd_device *o = osd_dev(d);
445 l = &mo->oo_dt.do_lu;
446 if (unlikely(o->od_in_init)) {
453 lu_object_header_init(h);
454 lu_object_init(l, h, d);
455 lu_object_add_top(h, l);
458 dt_object_init(&mo->oo_dt, NULL, d);
459 mo->oo_header = NULL;
462 mo->oo_dt.do_ops = &osd_obj_ops;
463 l->lo_ops = &osd_lu_obj_ops;
464 init_rwsem(&mo->oo_sem);
465 init_rwsem(&mo->oo_ext_idx_sem);
466 spin_lock_init(&mo->oo_guard);
467 INIT_LIST_HEAD(&mo->oo_xattr_list);
473 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
474 struct dentry *dentry, struct lustre_ost_attrs *loa)
478 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
479 (void *)loa, sizeof(*loa));
481 struct lustre_mdt_attrs *lma = &loa->loa_lma;
483 if (rc < sizeof(*lma))
487 lustre_loa_swab(loa, true);
488 /* Check LMA compatibility */
489 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
491 CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
493 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
494 PFID(&lma->lma_self_fid), inode->i_ino, rc);
496 } else if (rc == 0) {
504 * retrieve object from backend ext fs.
506 static struct inode *osd_iget2(struct osd_thread_info *info,
507 struct osd_device *dev, struct osd_inode_id *id,
508 bool special, int *err)
510 struct inode *inode = NULL;
514 * if we look for an inode withing a running
515 * transaction, then we risk to deadlock
516 * osd_dirent_check_repair() breaks this
518 /* LASSERT(current->journal_info == NULL); */
520 inode = osd_ldiskfs_iget_special(osd_sb(dev), id->oii_ino, special);
522 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
523 id->oii_ino, PTR_ERR(inode));
524 } else if (id->oii_gen != OSD_OII_NOGEN &&
525 inode->i_generation != id->oii_gen) {
526 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
527 "i_generation = %u\n",
528 id->oii_ino, id->oii_gen, inode->i_generation);
530 } else if (inode->i_nlink == 0) {
532 * due to parallel readdir and unlink,
533 * we can have dead inode here.
535 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
537 } else if (is_bad_inode(inode)) {
538 CWARN("%s: bad inode: ino = %u: rc = %d\n",
539 osd_dev2name(dev), id->oii_ino, -ENOENT);
541 } else if (osd_is_ea_inode(inode)) {
543 * EA inode is internal ldiskfs object, should don't visible
546 CDEBUG(D_INODE, "EA inode: ino = %u\n", id->oii_ino);
548 } else if ((rc = osd_attach_jinode(inode))) {
549 CDEBUG(D_INODE, "jbd: ino = %u rc = %d\n", id->oii_ino, rc);
551 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
552 if (id->oii_gen == OSD_OII_NOGEN)
553 osd_id_gen(id, inode->i_ino, inode->i_generation);
556 * Do not update file c/mtime in ldiskfs.
557 * NB: we don't have any lock to protect this because we don't
558 * have reference on osd_object now, but contention with
559 * another lookup + attr_set can't happen in the tiny window
560 * between if (...) and set S_NOCMTIME.
562 if (!(inode->i_flags & S_NOCMTIME))
563 inode->i_flags |= S_NOCMTIME;
571 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
572 struct osd_inode_id *id)
577 inode = osd_iget2(info, dev, id, 0, &rc);
587 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
588 handle_t *handle, struct dentry *child,
589 struct inode *inode, struct htree_lock *hlock)
593 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
594 if (rc == -ENOBUFS || rc == -ENOSPC) {
595 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
596 struct inode *parent = child->d_parent->d_inode;
597 struct lu_fid *fid = NULL;
598 char fidstr[FID_LEN + 1] = "unknown";
600 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
602 fid = &loa->loa_lma.lma_self_fid;
603 } else if (rc2 == -ENODATA) {
604 if (unlikely(is_root_inode(parent))) {
605 fid = &info->oti_fid3;
606 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
607 } else if (!osd->od_is_ost && osd->od_index == 0) {
608 fid = &info->oti_fid3;
609 lu_igif_build(fid, parent->i_ino,
610 parent->i_generation);
615 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
617 /* below message is checked in sanity.sh test_129 */
619 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
620 osd_name(osd), parent->i_ino, fidstr);
622 rc = 0; /* ignore such error now */
623 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
624 osd_name(osd), parent->i_ino, fidstr);
634 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
635 struct osd_inode_id *id, struct lu_fid *fid)
637 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
641 inode = osd_iget(info, dev, id);
645 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
647 *fid = loa->loa_lma.lma_self_fid;
648 } else if (rc == -ENODATA) {
649 if (unlikely(is_root_inode(inode)))
650 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
652 lu_igif_build(fid, inode->i_ino, inode->i_generation);
660 static struct inode *osd_iget_check(struct osd_thread_info *info,
661 struct osd_device *dev,
662 const struct lu_fid *fid,
663 struct osd_inode_id *id,
667 bool special = false;
673 * The cached OI mapping is trustable. If we cannot locate the inode
674 * via the cached OI mapping, then return the failure to the caller
675 * directly without further OI checking.
678 if (unlikely(fid_is_acct(fid)))
681 inode = osd_iget2(info, dev, id, special, &rc);
683 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
686 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
687 PFID(fid), id->oii_ino, rc);
693 __u32 saved_ino = id->oii_ino;
694 __u32 saved_gen = id->oii_gen;
697 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
699 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
701 * XXX: There are four possible cases:
703 * Backup/restore caused the OI invalid.
705 * Someone unlinked the object but NOT removed
706 * the OI mapping, such as mount target device
707 * as ldiskfs, and modify something directly.
709 * Someone just removed the object between the
710 * former oi_lookup and the iget. It is normal.
711 * 4. Other failure cases.
713 * Generally, when the device is mounted, it will
714 * auto check whether the system is restored from
715 * file-level backup or not. We trust such detect
716 * to distinguish the 1st case from the 2nd case:
717 * if the OI files are consistent but may contain
718 * stale OI mappings because of case 2, if iget()
719 * returns -ENOENT or -ESTALE, then it should be
724 * If the OI mapping was in OI file before the
725 * osd_iget_check(), but now, it is disappear,
726 * then it must be removed by race. That is a
732 * It is the OI scrub updated the OI mapping by race.
733 * The new OI mapping must be valid.
735 if (saved_ino != id->oii_ino ||
736 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
745 if (dev->od_scrub.os_scrub.os_file.sf_flags &
748 * It still can be the case 2, but we cannot
749 * distinguish it from the case 1. So return
750 * -EREMCHG to block current operation until
751 * OI scrub rebuilt the OI mappings.
760 if (inode->i_generation == id->oii_gen)
780 * \retval +v: new filter_fid does not contain self-fid
781 * \retval 0: filter_fid_18_23, contains self-fid
782 * \retval -v: other failure cases
784 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
785 struct dentry *dentry, struct lu_fid *fid)
787 struct filter_fid *ff = &info->oti_ff;
788 struct ost_id *ostid = &info->oti_ostid;
791 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
792 if (rc == sizeof(struct filter_fid_18_23)) {
793 struct filter_fid_18_23 *ff_old = (void *)ff;
795 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
796 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
798 * XXX: use 0 as the index for compatibility, the caller will
799 * handle index related issues when necessary.
802 ostid_to_fid(fid, ostid, 0);
803 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
805 } else if (rc >= 0) {
812 static int osd_lma_self_repair(struct osd_thread_info *info,
813 struct osd_device *osd, struct inode *inode,
814 const struct lu_fid *fid, __u32 compat)
819 LASSERT(current->journal_info == NULL);
821 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
822 osd_dto_credits_noquota[DTO_XATTR_SET]);
825 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
830 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
832 CWARN("%s: cannot self repair the LMA: rc = %d\n",
834 ldiskfs_journal_stop(jh);
838 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
840 struct osd_thread_info *info = osd_oti_get(env);
841 struct osd_device *osd = osd_obj2dev(obj);
842 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
843 struct lustre_mdt_attrs *lma = &loa->loa_lma;
844 struct inode *inode = obj->oo_inode;
845 struct dentry *dentry = &info->oti_obj_dentry;
846 struct lu_fid *fid = NULL;
847 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
852 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
853 (void *)loa, sizeof(*loa));
854 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
855 fid = &lma->lma_self_fid;
856 rc = osd_get_idif(info, inode, dentry, fid);
857 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
859 * For the given OST-object, if it has neither LMA nor
860 * FID in XATTR_NAME_FID, then the given FID (which is
861 * contained in the @obj, from client RPC for locating
862 * the OST-object) is trusted. We use it to generate
865 osd_lma_self_repair(info, osd, inode, rfid,
876 lustre_lma_swab(lma);
877 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
878 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
879 S_ISREG(inode->i_mode)))) {
880 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
881 "fid = "DFID", ino = %lu\n", osd_name(osd),
882 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
883 PFID(rfid), inode->i_ino);
886 fid = &lma->lma_self_fid;
887 if (lma->lma_compat & LMAC_STRIPE_INFO &&
889 obj->oo_pfid_in_lma = 1;
890 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
892 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
896 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
897 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
898 struct ost_id *oi = &info->oti_ostid;
899 struct lu_fid *fid1 = &info->oti_fid3;
900 __u32 idx = fid_idif_ost_idx(rfid);
903 * For old IDIF, the OST index is not part of the IDIF,
904 * Means that different OSTs may have the same IDIFs.
905 * Under such case, we need to make some compatible
906 * check to make sure to trigger OI scrub properly.
908 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
909 /* Given @rfid is new, LMA is old. */
910 fid_to_ostid(fid, oi);
911 ostid_to_fid(fid1, oi, idx);
912 if (lu_fid_eq(fid1, rfid)) {
913 if (osd->od_index_in_idif)
914 osd_lma_self_repair(info, osd,
928 struct osd_check_lmv_buf {
929 /* please keep it as first member */
930 struct dir_context ctx;
931 struct osd_thread_info *oclb_info;
932 struct osd_device *oclb_dev;
938 * It is called internally by ->iterate*() to filter out the
939 * local slave object's FID of the striped directory.
941 * \retval 1 found the local slave's FID
942 * \retval 0 continue to check next item
943 * \retval -ve for failure
945 #ifdef HAVE_FILLDIR_USE_CTX
946 static FILLDIR_TYPE do_osd_stripe_dir_filldir(struct dir_context *buf,
948 static int osd_stripe_dir_filldir(void *buf,
950 const char *name, int namelen,
951 loff_t offset, __u64 ino, unsigned int d_type)
953 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
954 struct osd_thread_info *oti = oclb->oclb_info;
955 struct lu_fid *fid = &oti->oti_fid3;
956 struct osd_inode_id *id = &oti->oti_id3;
957 struct osd_inode_id id2;
958 struct osd_device *dev = oclb->oclb_dev;
967 sscanf(name + 1, SFID, RFID(fid));
968 if (!fid_is_sane(fid))
971 if (osd_remote_fid(oti->oti_env, dev, fid))
974 osd_id_gen(id, ino, OSD_OII_NOGEN);
975 inode = osd_iget(oti, dev, id);
977 return PTR_ERR(inode);
980 osd_add_oi_cache(oti, dev, id, fid);
981 /* Check shard by scrub only if it has a problem with OI */
982 if (osd_oi_lookup(oti, dev, fid, &id2, 0) || !osd_id_eq(id, &id2))
983 osd_scrub_oi_insert(dev, fid, id, true);
984 oclb->oclb_found = true;
989 WRAP_FILLDIR_FN(do_, osd_stripe_dir_filldir)
992 * When lookup item under striped directory, we need to locate the master
993 * MDT-object of the striped directory firstly, then the client will send
994 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
995 * and the item's name. If the system is restored from MDT file level backup,
996 * then before the OI scrub completely built the OI files, the OI mappings of
997 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
998 * not a problem for the master MDT-object. Because when locate the master
999 * MDT-object, we will do name based lookup (for the striped directory itself)
1000 * firstly, during such process we can setup the correct OI mapping for the
1001 * master MDT-object. But it will be trouble for the slave MDT-object. Because
1002 * the client will not trigger name based lookup on the MDT to locate the slave
1003 * MDT-object before locating item under the striped directory, then when
1004 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
1005 * is invalid and does not know what the right OI mapping is, then the MDT has
1006 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
1007 * the OI file, related OI mapping is unknown yet, please try again later. And
1008 * then client will re-try the RPC again and again until related OI mapping has
1009 * been updated. That is quite inefficient.
1011 * To resolve above trouble, we will handle it as the following two cases:
1013 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
1014 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
1015 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
1016 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
1017 * directly. Please check osd_fid_lookup().
1019 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
1020 * Under such case, during lookup the master MDT-object, we will lookup the
1021 * slave MDT-object via readdir against the master MDT-object, because the
1022 * slave MDT-objects information are stored as sub-directories with the name
1023 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
1024 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
1025 * the correct OI mapping for the slave MDT-object.
1027 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
1028 struct inode *inode)
1030 struct lu_buf *buf = &oti->oti_big_buf;
1032 struct lmv_mds_md_v1 *lmv1;
1033 struct osd_check_lmv_buf oclb = {
1034 .ctx.actor = osd_stripe_dir_filldir,
1037 .oclb_found = false,
1042 /* We should use the VFS layer to create a real dentry. */
1043 oti->oti_obj_dentry.d_inode = inode;
1044 oti->oti_obj_dentry.d_sb = inode->i_sb;
1046 filp = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
1051 filp->f_mode |= FMODE_64BITHASH;
1055 rc = __osd_xattr_get(inode, filp->f_path.dentry, XATTR_NAME_LMV,
1056 buf->lb_buf, buf->lb_len);
1057 if (rc == -ERANGE) {
1058 rc = __osd_xattr_get(inode, filp->f_path.dentry,
1059 XATTR_NAME_LMV, NULL, 0);
1061 lu_buf_realloc(buf, rc);
1062 if (buf->lb_buf == NULL)
1063 GOTO(out, rc = -ENOMEM);
1069 if (unlikely(rc == 0 || rc == -ENODATA))
1075 if (unlikely(buf->lb_buf == NULL)) {
1076 lu_buf_realloc(buf, rc);
1077 if (buf->lb_buf == NULL)
1078 GOTO(out, rc = -ENOMEM);
1084 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1088 oclb.oclb_items = 0;
1089 rc = iterate_dir(filp, &oclb.ctx);
1090 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1091 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1096 "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
1097 osd_ino2name(inode), inode->i_ino, inode->i_generation,
1106 * Is object in scrub inconsistent/stale list.
1108 * \a scrub has two lists, os_inconsistent_items contains mappings to fix, while
1109 * os_stale_items contains mappings failed to fix.
1111 static bool fid_in_scrub_list(struct lustre_scrub *scrub,
1112 const struct list_head *list,
1113 const struct lu_fid *fid)
1115 struct osd_inconsistent_item *oii;
1117 if (list_empty(list))
1120 spin_lock(&scrub->os_lock);
1121 list_for_each_entry(oii, list, oii_list) {
1122 if (lu_fid_eq(fid, &oii->oii_cache.oic_fid)) {
1123 spin_unlock(&scrub->os_lock);
1127 spin_unlock(&scrub->os_lock);
1132 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1133 const struct lu_fid *fid,
1134 const struct lu_object_conf *conf)
1136 struct osd_thread_info *info;
1137 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1138 struct osd_device *dev;
1139 struct osd_idmap_cache *oic;
1140 struct osd_inode_id *id;
1141 struct inode *inode = NULL;
1142 struct lustre_scrub *scrub;
1143 struct scrub_file *sf;
1144 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1149 bool remote = false;
1150 bool trusted = true;
1151 bool updated = false;
1152 bool checked = false;
1157 LINVRNT(osd_invariant(obj));
1158 LASSERT(obj->oo_inode == NULL);
1160 if (fid_is_sane(fid) == 0) {
1161 CERROR("%s: invalid FID "DFID"\n", ldev->ld_obd->obd_name,
1167 dev = osd_dev(ldev);
1168 scrub = &dev->od_scrub.os_scrub;
1169 sf = &scrub->os_file;
1170 info = osd_oti_get(env);
1172 oic = &info->oti_cache;
1174 if (CFS_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1178 * For the object is created as locking anchor, or for the object to
1179 * be created on disk. No need to osd_oi_lookup() at here because FID
1180 * shouldn't never be re-used, if it's really a duplicate FID from
1181 * unexpected reason, we should be able to detect it later by calling
1182 * do_create->osd_oi_insert().
1184 if (conf && conf->loc_flags & LOC_F_NEW)
1187 /* Search order: 1. per-thread cache. */
1188 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1193 /* Search order: 2. OI scrub pending list. */
1195 memset(id, 0, sizeof(struct osd_inode_id));
1196 if (fid_in_scrub_list(scrub, &scrub->os_inconsistent_items, fid) &&
1198 RETURN(-EINPROGRESS);
1200 stale = fid_in_scrub_list(scrub, &scrub->os_stale_items, fid);
1201 if (stale && CFS_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1205 * The OI mapping in the OI file can be updated by the OI scrub
1206 * when we locate the inode via FID. So it may be not trustable.
1210 /* Search order: 3. OI files. */
1211 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1212 if (result == -ENOENT) {
1213 if (!fid_is_norm(fid) ||
1214 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1215 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1217 GOTO(out, result = 0);
1222 /* -ESTALE is returned if inode of OST object doesn't exist */
1223 if (result == -ESTALE &&
1224 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1225 GOTO(out, result = 0);
1232 obj->oo_inode = NULL;
1233 /* for later passes through checks, not true on first pass */
1234 if (!IS_ERR_OR_NULL(inode))
1237 inode = osd_iget_check(info, dev, fid, id, trusted);
1238 if (!IS_ERR(inode)) {
1239 obj->oo_inode = inode;
1247 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1250 result = PTR_ERR(inode);
1251 if (result == -ENOENT || result == -ESTALE)
1252 GOTO(out, result = 0);
1254 if (result != -EREMCHG)
1258 /* don't trigger repeatedly for stale mapping */
1260 GOTO(out, result = -ESTALE);
1263 * We still have chance to get the valid inode: for the
1264 * object which is referenced by remote name entry, the
1265 * object on the local MDT will be linked under the dir
1266 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1268 * We do not know whether the object for the given FID
1269 * is referenced by some remote name entry or not, and
1270 * especially for DNE II, a multiple-linked object may
1271 * have many name entries reside on many MDTs.
1273 * To simplify the operation, OSD will not distinguish
1274 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1275 * only happened for the RPC from other MDT during the
1276 * OI scrub, or for the client side RPC with FID only,
1277 * such as FID to path, or from old connected client.
1280 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1284 flags |= SS_AUTO_PARTIAL;
1285 flags &= ~SS_AUTO_FULL;
1290 if (scrub->os_running) {
1291 if (scrub->os_partial_scan && !scrub->os_in_join)
1294 if (IS_ERR_OR_NULL(inode) || result) {
1295 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1296 GOTO(out, result = -EINPROGRESS);
1300 LASSERT(obj->oo_inode == inode);
1302 osd_scrub_oi_insert(dev, fid, id, true);
1306 if (dev->od_scrub.os_scrub.os_auto_scrub_interval == AS_NEVER) {
1308 GOTO(out, result = -EREMCHG);
1311 LASSERT(obj->oo_inode == inode);
1313 osd_add_oi_cache(info, dev, id, fid);
1318 if (IS_ERR_OR_NULL(inode) || result)
1319 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1321 rc1 = osd_scrub_start(env, dev, flags);
1322 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
1323 "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
1324 osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
1325 if (rc1 && rc1 != -EALREADY)
1326 GOTO(out, result = -EREMCHG);
1328 if (IS_ERR_OR_NULL(inode) || result)
1329 GOTO(out, result = -EINPROGRESS);
1332 LASSERT(obj->oo_inode == inode);
1337 if (unlikely(obj->oo_header))
1340 result = osd_check_lma(env, obj);
1344 LASSERTF(id->oii_ino == inode->i_ino &&
1345 id->oii_gen == inode->i_generation,
1346 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1347 PFID(fid), id->oii_ino, id->oii_gen,
1348 inode->i_ino, inode->i_generation);
1350 saved_ino = inode->i_ino;
1351 saved_gen = inode->i_generation;
1353 if (unlikely(result == -ENODATA)) {
1355 * If the OI scrub updated the OI mapping by race, it
1356 * must be valid. Trust the inode that has no LMA EA.
1361 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1364 * The OI mapping is still there, the inode is still
1365 * valid. It is just becaues the inode has no LMA EA.
1367 if (saved_ino == id->oii_ino &&
1368 saved_gen == id->oii_gen)
1372 * It is the OI scrub updated the OI mapping by race.
1373 * The new OI mapping must be valid.
1381 * "result == -ENOENT" means that the OI mappinghas been
1382 * removed by race, so the inode belongs to other object.
1384 * Others error can be returned directly.
1386 if (result == -ENOENT) {
1387 obj->oo_inode = NULL;
1392 if (result != -EREMCHG)
1398 * if two OST objects map to the same inode, and inode mode is
1399 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1400 * reserved by precreate, and not written yet, in this case, don't
1401 * set inode for the object whose FID mismatch, so that it can create
1402 * inode and not block precreate.
1404 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1405 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1406 obj->oo_inode = NULL;
1407 GOTO(out, result = 0);
1410 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1412 * "result == -ENOENT" means the cached OI mapping has been removed
1413 * from the OI file by race, above inode belongs to other object.
1415 if (result == -ENOENT) {
1416 obj->oo_inode = NULL;
1417 GOTO(out, result = 0);
1423 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1425 osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
1431 * It is the OI scrub updated the OI mapping by race.
1432 * The new OI mapping must be valid.
1440 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1441 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1443 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1445 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1447 obj->oo_pfid_in_lma = 1;
1448 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1450 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1451 } else if (result != -ENODATA) {
1456 obj->oo_compat_dot_created = 1;
1457 obj->oo_compat_dotdot_created = 1;
1459 if (S_ISDIR(inode->i_mode) &&
1460 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1461 osd_check_lmv(info, dev, inode);
1463 result = osd_attach_jinode(inode);
1468 GOTO(out, result = 0);
1470 LASSERT(!obj->oo_hl_head);
1471 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1473 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1476 if (!result && stale)
1477 osd_scrub_oi_resurrect(scrub, fid);
1479 if (result || !obj->oo_inode) {
1480 if (!IS_ERR_OR_NULL(inode))
1483 obj->oo_inode = NULL;
1485 fid_zero(&oic->oic_fid);
1488 LINVRNT(osd_invariant(obj));
1493 * Concurrency: shouldn't matter.
1495 static void osd_object_init0(struct osd_object *obj)
1497 LASSERT(obj->oo_inode != NULL);
1498 obj->oo_dt.do_body_ops = &osd_body_ops;
1499 obj->oo_dt.do_lu.lo_header->loh_attr |=
1500 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1504 * Concurrency: no concurrent access is possible that early in object
1507 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1508 const struct lu_object_conf *conf)
1510 struct osd_object *obj = osd_obj(l);
1513 LINVRNT(osd_invariant(obj));
1515 if (CFS_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1516 cfs_fail_val == 2) {
1517 struct osd_thread_info *info = osd_oti_get(env);
1518 struct osd_idmap_cache *oic = &info->oti_cache;
1519 /* invalidate thread cache */
1520 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1522 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1523 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1524 l->lo_header->loh_attr |= LOHA_EXISTS;
1528 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1529 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1530 if (result == 0 && obj->oo_inode != NULL) {
1531 struct osd_thread_info *oti = osd_oti_get(env);
1532 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1534 osd_object_init0(obj);
1535 if (unlikely(obj->oo_header))
1538 result = osd_get_lma(oti, obj->oo_inode,
1539 &oti->oti_obj_dentry, loa);
1542 * Convert LMAI flags to lustre LMA flags
1543 * and cache it to oo_lma_flags
1546 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1547 } else if (result == -ENODATA) {
1551 atomic_set(&obj->oo_dirent_count, LU_DIRENT_COUNT_UNSET);
1553 LINVRNT(osd_invariant(obj));
1558 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1559 * The left part is for value, binary mode.
1561 struct osd_xattr_entry {
1562 struct list_head oxe_list;
1566 struct rcu_head oxe_rcu;
1570 static int osd_oxc_get(struct osd_object *obj, const char *name,
1573 struct osd_xattr_entry *tmp;
1574 struct osd_xattr_entry *oxe = NULL;
1575 size_t namelen = strlen(name);
1579 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1580 if (namelen == tmp->oxe_namelen &&
1581 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1588 GOTO(out, rc = -ENOENT);
1590 if (!oxe->oxe_exist)
1591 GOTO(out, rc = -ENODATA);
1594 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1597 if (buf->lb_buf == NULL)
1600 if (buf->lb_len < rc)
1601 GOTO(out, rc = -ERANGE);
1603 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1610 static void osd_oxc_free(struct rcu_head *head)
1612 struct osd_xattr_entry *oxe;
1614 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1615 OBD_FREE(oxe, oxe->oxe_len);
1618 static void osd_oxc_add(struct osd_object *obj, const char *name,
1619 const char *buf, int buflen)
1621 struct osd_xattr_entry *oxe;
1622 struct osd_xattr_entry *old = NULL;
1623 struct osd_xattr_entry *tmp;
1624 size_t namelen = strlen(name);
1625 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1627 OBD_ALLOC(oxe, len);
1631 INIT_LIST_HEAD(&oxe->oxe_list);
1633 oxe->oxe_namelen = namelen;
1634 memcpy(oxe->oxe_buf, name, namelen);
1636 LASSERT(buf != NULL);
1637 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1638 oxe->oxe_exist = true;
1640 oxe->oxe_exist = false;
1643 /* this should be rarely called, just remove old and add new */
1644 spin_lock(&obj->oo_guard);
1645 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1646 if (namelen == tmp->oxe_namelen &&
1647 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1653 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1654 call_rcu(&old->oxe_rcu, osd_oxc_free);
1656 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1658 spin_unlock(&obj->oo_guard);
1661 static void osd_oxc_del(struct osd_object *obj, const char *name)
1663 struct osd_xattr_entry *oxe;
1664 size_t namelen = strlen(name);
1666 spin_lock(&obj->oo_guard);
1667 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1668 if (namelen == oxe->oxe_namelen &&
1669 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1670 list_del_rcu(&oxe->oxe_list);
1671 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1675 spin_unlock(&obj->oo_guard);
1678 static void osd_oxc_fini(struct osd_object *obj)
1680 struct osd_xattr_entry *oxe, *next;
1682 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1683 list_del(&oxe->oxe_list);
1684 OBD_FREE(oxe, oxe->oxe_len);
1689 * Concurrency: no concurrent access is possible that late in object
1692 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1694 struct osd_object *obj = osd_obj(l);
1695 struct lu_object_header *h = obj->oo_header;
1697 LINVRNT(osd_invariant(obj));
1700 dt_object_fini(&obj->oo_dt);
1701 if (obj->oo_hl_head != NULL)
1702 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1703 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1706 lu_object_header_free(h);
1710 * Concurrency: no concurrent access is possible that late in object
1713 static void osd_index_fini(struct osd_object *o)
1715 struct iam_container *bag;
1717 if (o->oo_dir != NULL) {
1718 bag = &o->oo_dir->od_container;
1719 if (o->oo_inode != NULL) {
1720 if (bag->ic_object == o->oo_inode)
1721 iam_container_fini(bag);
1723 OBD_FREE_PTR(o->oo_dir);
1729 OSD_TXN_OI_DELETE_CREDITS = 20,
1730 OSD_TXN_INODE_DELETE_CREDITS = 20
1737 #if OSD_THANDLE_STATS
1739 * Set time when the handle is allocated
1741 static void osd_th_alloced(struct osd_thandle *oth)
1743 oth->oth_alloced = ktime_get();
1747 * Set time when the handle started
1749 static void osd_th_started(struct osd_thandle *oth)
1751 oth->oth_started = ktime_get();
1755 * Check whether the we deal with this handle for too long.
1757 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1758 ktime_t alloced, ktime_t started,
1761 ktime_t now = ktime_get();
1763 LASSERT(dev != NULL);
1765 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1766 ktime_us_delta(started, alloced));
1767 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1768 ktime_us_delta(closed, started));
1769 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1770 ktime_us_delta(now, closed));
1772 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1773 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1774 oth, now, alloced, started, closed);
1779 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1781 ktime_t __closed = ktime_get(); \
1782 ktime_t __alloced = oth->oth_alloced; \
1783 ktime_t __started = oth->oth_started; \
1786 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1789 #else /* OSD_THANDLE_STATS */
1791 #define osd_th_alloced(h) do {} while(0)
1792 #define osd_th_started(h) do {} while(0)
1793 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1795 #endif /* OSD_THANDLE_STATS */
1798 * in some cases (like overstriped files) the same operations on the same
1799 * objects are declared many times and this may lead to huge number of
1800 * credits which can be a problem and/or cause performance degradation.
1801 * this function is to remember what declarations have been made within
1802 * a given thandle and then skip duplications.
1803 * limit it's scope so that regular small transactions don't need all
1804 * this overhead with allocations, lists.
1805 * also, limit scope to the specific objects like llogs, etc.
1807 static inline bool osd_check_special_fid(const struct lu_fid *f)
1809 if (fid_seq_is_llog(f->f_seq))
1811 if (f->f_seq == FID_SEQ_LOCAL_FILE &&
1812 f->f_oid == MDD_LOV_OBJ_OID)
1817 bool osd_tx_was_declared(const struct lu_env *env, struct osd_thandle *oth,
1818 struct dt_object *dt, enum dt_txn_op op, loff_t pos)
1820 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1821 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1822 struct osd_thread_info *oti = osd_oti_get(env);
1823 struct osd_obj_declare *old;
1828 /* small transactions don't need this overhead */
1829 if (oti->oti_declare_ops[DTO_OBJECT_CREATE] < 10 &&
1830 oti->oti_declare_ops[DTO_WRITE_BASE] < 10)
1833 if (osd_check_special_fid(fid) == 0)
1836 list_for_each_entry(old, &oth->ot_declare_list, old_list) {
1837 if (old->old_op == op && old->old_pos == pos &&
1838 lu_fid_eq(&old->old_fid, fid))
1842 if (unlikely(old == NULL))
1844 old->old_fid = *lu_object_fid(&dt->do_lu);
1847 list_add(&old->old_list, &oth->ot_declare_list);
1851 void osd_tx_declaration_free(struct osd_thandle *oth)
1853 struct osd_obj_declare *old, *tmp;
1855 list_for_each_entry_safe(old, tmp, &oth->ot_declare_list, old_list) {
1856 list_del_init(&old->old_list);
1862 * Concurrency: doesn't access mutable data.
1864 static int osd_param_is_not_sane(const struct osd_device *dev,
1865 const struct thandle *th)
1867 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1869 return oh->ot_credits > osd_transaction_size(dev);
1873 * Concurrency: shouldn't matter.
1875 static void osd_trans_commit_cb(struct super_block *sb,
1876 struct ldiskfs_journal_cb_entry *jcb, int error)
1878 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1879 struct thandle *th = &oh->ot_super;
1880 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1881 struct osd_device *osd = osd_dev(lud);
1882 struct dt_txn_commit_cb *dcb, *tmp;
1884 LASSERT(oh->ot_handle == NULL);
1887 CERROR("transaction @0x%p commit error: %d\n", th, error);
1889 /* call per-transaction callbacks if any */
1890 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1892 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1893 "commit callback entry: magic=%x name='%s'\n",
1894 dcb->dcb_magic, dcb->dcb_name);
1895 list_del_init(&dcb->dcb_linkage);
1896 dcb->dcb_func(NULL, th, dcb, error);
1899 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1900 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1901 wake_up(&osd->od_commit_cb_done);
1907 static struct thandle *osd_trans_create(const struct lu_env *env,
1908 struct dt_device *d)
1910 struct osd_thread_info *oti = osd_oti_get(env);
1911 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1912 struct osd_thandle *oh;
1918 CERROR("%s: someone try to start transaction under "
1919 "readonly mode, should be disabled.\n",
1920 osd_name(osd_dt_dev(d)));
1922 RETURN(ERR_PTR(-EROFS));
1925 /* on pending IO in this thread should left from prev. request */
1926 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1928 sb_start_write(osd_sb(osd_dt_dev(d)));
1930 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1932 sb_end_write(osd_sb(osd_dt_dev(d)));
1933 RETURN(ERR_PTR(-ENOMEM));
1936 oh->ot_quota_trans = &oti->oti_quota_trans;
1937 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1942 oh->oh_declared_ext = 0;
1943 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1944 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1945 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1946 INIT_LIST_HEAD(&oh->ot_declare_list);
1949 memset(oti->oti_declare_ops, 0,
1950 sizeof(oti->oti_declare_ops));
1951 memset(oti->oti_declare_ops_cred, 0,
1952 sizeof(oti->oti_declare_ops_cred));
1953 memset(oti->oti_declare_ops_used, 0,
1954 sizeof(oti->oti_declare_ops_used));
1956 oti->oti_ins_cache_depth++;
1961 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1963 struct osd_thread_info *oti = osd_oti_get(env);
1964 struct osd_thandle *oh;
1966 oh = container_of(th, struct osd_thandle, ot_super);
1967 LASSERT(oh != NULL);
1969 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1970 oti->oti_declare_ops[OSD_OT_CREATE],
1971 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1972 oti->oti_declare_ops_used[OSD_OT_CREATE],
1973 oti->oti_declare_ops[OSD_OT_DESTROY],
1974 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1975 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1976 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1977 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1978 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1979 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1980 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1981 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1982 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1983 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1984 oti->oti_declare_ops[OSD_OT_WRITE],
1985 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1986 oti->oti_declare_ops_used[OSD_OT_WRITE],
1987 oti->oti_declare_ops[OSD_OT_PUNCH],
1988 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1989 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1990 oti->oti_declare_ops[OSD_OT_QUOTA],
1991 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1992 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1993 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1994 oti->oti_declare_ops[OSD_OT_INSERT],
1995 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1996 oti->oti_declare_ops_used[OSD_OT_INSERT],
1997 oti->oti_declare_ops[OSD_OT_DELETE],
1998 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1999 oti->oti_declare_ops_used[OSD_OT_DELETE]);
2000 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
2001 oti->oti_declare_ops[OSD_OT_REF_ADD],
2002 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
2003 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
2004 oti->oti_declare_ops[OSD_OT_REF_DEL],
2005 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
2006 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
2009 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
2010 void osd_ldiskfs_credits_for_revoke(struct osd_device *osd,
2011 struct osd_thandle *oh,
2012 int *credits, int *revoke)
2014 int blocks = LDISKFS_MAX_EXTENT_DEPTH * oh->oh_declared_ext;
2015 *revoke += ldiskfs_trans_default_revoke_credits(osd_sb(osd)) + blocks;
2018 void osd_ldiskfs_credits_for_revoke(struct osd_device *osd,
2019 struct osd_thandle *oh,
2020 int *credits, int *revoke)
2022 struct journal_s *journal = LDISKFS_SB(osd_sb(osd))->s_journal;
2023 int blocks, jbsize, records_per_block;
2025 blocks = LDISKFS_MAX_EXTENT_DEPTH * oh->oh_declared_ext;
2026 jbsize = journal->j_blocksize;
2027 jbsize -= sizeof(struct jbd2_journal_block_tail) +
2028 sizeof(jbd2_journal_revoke_header_t);
2029 records_per_block = jbsize / 8;
2030 *credits += (blocks + records_per_block - 1) / records_per_block;
2035 * Concurrency: shouldn't matter.
2037 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
2040 struct osd_thread_info *oti = osd_oti_get(env);
2041 struct osd_device *dev = osd_dt_dev(d);
2043 struct osd_thandle *oh;
2048 LASSERT(current->journal_info == NULL);
2050 oh = container_of(th, struct osd_thandle, ot_super);
2051 LASSERT(oh != NULL);
2052 LASSERT(oh->ot_handle == NULL);
2053 if (unlikely(ldiskfs_track_declares_assert != 0)) {
2054 LASSERT(oti->oti_r_locks == 0);
2055 LASSERT(oti->oti_w_locks == 0);
2058 rc = dt_txn_hook_start(env, d, th);
2062 if (unlikely(osd_param_is_not_sane(dev, th))) {
2063 static unsigned long last_printed;
2064 static int last_credits;
2066 lprocfs_counter_add(dev->od_stats,
2067 LPROC_OSD_TOO_MANY_CREDITS, 1);
2070 * don't make noise on a tiny testing systems
2071 * actual credits misuse will be caught anyway
2073 if (last_credits != oh->ot_credits &&
2074 time_after(jiffies, last_printed +
2075 cfs_time_seconds(60)) &&
2076 osd_transaction_size(dev) > 512) {
2077 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
2078 oh->ot_credits, osd_transaction_size(dev));
2079 osd_trans_dump_creds(env, th);
2081 last_credits = oh->ot_credits;
2082 last_printed = jiffies;
2085 * XXX Limit the credits to 'max_transaction_buffers', and
2086 * let the underlying filesystem to catch the error if
2087 * we really need so many credits.
2089 * This should be removed when we can calculate the
2090 * credits precisely.
2092 oh->ot_credits = osd_transaction_size(dev);
2093 } else if (ldiskfs_track_declares_assert != 0) {
2095 * reserve few credits to prevent an assertion in JBD
2096 * our debugging mechanism will be able to detected
2097 * overuse. this can help to debug single-update
2100 oh->ot_credits += 10;
2101 if (unlikely(osd_param_is_not_sane(dev, th)))
2102 oh->ot_credits = osd_transaction_size(dev);
2105 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
2106 GOTO(out, rc = -EIO);
2109 * we ignore quota checks for system-owned files, but still
2110 * need to count blocks for uid/gid/projid
2112 osd_trans_declare_op(env, oh, OSD_OT_QUOTA, 3);
2115 * XXX temporary stuff. Some abstraction layer should
2118 osd_ldiskfs_credits_for_revoke(dev, oh, &oh->ot_credits, &revoke);
2119 jh = osd_journal_start_with_revoke(osd_sb(dev), LDISKFS_HT_MISC,
2120 oh->ot_credits, revoke);
2124 LASSERT(oti->oti_txns == 0);
2126 atomic_inc(&dev->od_commit_cb_in_flight);
2127 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
2138 static int osd_seq_exists(const struct lu_env *env,
2139 struct osd_device *osd, u64 seq)
2141 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
2142 struct seq_server_site *ss = osd_seq_site(osd);
2147 LASSERT(ss != NULL);
2148 LASSERT(ss->ss_server_fld != NULL);
2150 rc = osd_fld_lookup(env, osd, seq, range);
2153 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2154 osd_name(osd), seq, rc);
2158 RETURN(ss->ss_node_id == range->lsr_index);
2161 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2163 struct dt_txn_commit_cb *dcb;
2164 struct dt_txn_commit_cb *tmp;
2166 /* call per-transaction stop callbacks if any */
2167 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2169 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2170 "commit callback entry: magic=%x name='%s'\n",
2171 dcb->dcb_magic, dcb->dcb_name);
2172 list_del_init(&dcb->dcb_linkage);
2173 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2178 * Concurrency: shouldn't matter.
2180 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2183 struct osd_thread_info *oti = osd_oti_get(env);
2184 struct osd_thandle *oh;
2185 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2186 struct osd_device *osd = osd_dt_dev(th->th_dev);
2187 struct qsd_instance *qsd = osd_def_qsd(osd);
2188 struct lquota_trans *qtrans;
2189 LIST_HEAD(truncates);
2190 int rc = 0, remove_agents = 0;
2194 oh = container_of(th, struct osd_thandle, ot_super);
2196 remove_agents = oh->ot_remove_agents;
2198 qtrans = oh->ot_quota_trans;
2199 oh->ot_quota_trans = NULL;
2201 osd_tx_declaration_free(oh);
2203 /* move locks to local list, stop tx, execute truncates */
2204 list_splice(&oh->ot_trunc_locks, &truncates);
2206 if (oh->ot_handle != NULL) {
2209 handle_t *hdl = oh->ot_handle;
2212 * add commit callback
2213 * notice we don't do this in osd_trans_start()
2214 * as underlying transaction can change during truncate
2216 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2219 LASSERT(oti->oti_txns == 1);
2222 rc = dt_txn_hook_stop(env, th);
2224 CERROR("%s: failed in transaction hook: rc = %d\n",
2227 osd_trans_stop_cb(oh, rc);
2228 /* hook functions might modify th_sync */
2229 hdl->h_sync = th->th_sync;
2231 oh->ot_handle = NULL;
2232 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2234 CERROR("%s: failed to stop transaction: rc = %d\n",
2235 osd_name(osd), rc2);
2239 /* We preserve the origin behavior of ignoring any
2240 * failures with the underlying punch / truncate
2241 * operation. We do record for debugging if an error
2242 * does occur in the lctl dk logs.
2244 rc2 = osd_process_truncates(env, &truncates);
2246 CERROR("%s: failed truncate process: rc = %d\n",
2247 osd_name(osd), rc2);
2249 osd_trans_stop_cb(oh, th->th_result);
2253 osd_trunc_unlock_all(env, &truncates);
2255 /* inform the quota slave device that the transaction is stopping */
2256 qsd_op_end(env, qsd, qtrans);
2259 * as we want IO to journal and data IO be concurrent, we don't block
2260 * awaiting data IO completion in osd_do_bio(), instead we wait here
2261 * once transaction is submitted to the journal. all reqular requests
2262 * don't do direct IO (except read/write), thus this wait_event becomes
2265 * IMPORTANT: we have to wait till any IO submited by the thread is
2266 * completed otherwise iobuf may be corrupted by different request
2268 wait_event(iobuf->dr_wait,
2269 atomic_read(&iobuf->dr_numreqs) == 0);
2272 rc = iobuf->dr_error;
2274 osd_fini_iobuf(osd, iobuf);
2276 if (unlikely(remove_agents != 0))
2277 osd_process_scheduled_agent_removals(env, osd);
2279 LASSERT(oti->oti_ins_cache_depth > 0);
2280 oti->oti_ins_cache_depth--;
2281 /* reset OI cache for safety */
2282 if (oti->oti_ins_cache_depth == 0)
2283 oti->oti_ins_cache_used = 0;
2285 sb_end_write(osd_sb(osd));
2290 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2292 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2295 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2296 LASSERT(&dcb->dcb_func != NULL);
2297 if (dcb->dcb_flags & DCB_TRANS_STOP)
2298 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2300 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2305 struct osd_delayed_iput_work {
2306 struct work_struct diw_work;
2307 struct inode *diw_inode;
2310 static void osd_delayed_iput_fn(struct work_struct *work)
2312 struct osd_delayed_iput_work *diwork;
2313 struct inode *inode;
2315 diwork = container_of(work, struct osd_delayed_iput_work, diw_work);
2316 inode = diwork->diw_inode;
2317 CDEBUG(D_INODE, "%s: delayed iput (ino=%lu)\n",
2318 inode->i_sb->s_id, inode->i_ino);
2320 OBD_FREE_PTR(diwork);
2323 noinline void osd_delayed_iput(struct inode *inode,
2324 struct osd_delayed_iput_work *diwork)
2329 INIT_WORK(&diwork->diw_work, osd_delayed_iput_fn);
2330 diwork->diw_inode = inode;
2331 queue_work(LDISKFS_SB(inode->i_sb)->s_misc_wq,
2337 * Called just before object is freed. Releases all resources except for
2338 * object itself (that is released by osd_object_free()).
2340 * Concurrency: no concurrent access is possible that late in object
2343 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2345 struct osd_object *obj = osd_obj(l);
2346 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2347 struct inode *inode = obj->oo_inode;
2348 struct osd_delayed_iput_work *diwork = NULL;
2353 LINVRNT(osd_invariant(obj));
2356 * If object is unlinked remove fid->ino mapping from object index.
2359 osd_index_fini(obj);
2364 if (inode->i_blocks > ldiskfs_delayed_unlink_blocks)
2365 OBD_ALLOC(diwork, sizeof(*diwork));
2367 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2368 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2370 uid = i_uid_read(inode);
2371 gid = i_gid_read(inode);
2372 projid = i_projid_read(inode);
2374 obj->oo_inode = NULL;
2375 osd_delayed_iput(inode, diwork);
2377 /* do not rebalance quota if the caller needs to release memory
2378 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2379 * transaction and risk to deadlock - LU-12178 */
2380 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2383 if (!obj->oo_header && qsd) {
2384 struct osd_thread_info *info = osd_oti_get(env);
2385 struct lquota_id_info *qi = &info->oti_qi;
2387 /* Release granted quota to master if necessary */
2388 qi->lqi_id.qid_uid = uid;
2389 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2391 qi->lqi_id.qid_uid = gid;
2392 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2394 qi->lqi_id.qid_uid = projid;
2395 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2400 * Concurrency: ->loo_object_release() is called under site spin-lock.
2402 static void osd_object_release(const struct lu_env *env,
2403 struct lu_object *l)
2405 struct osd_object *o = osd_obj(l);
2408 * nobody should be releasing a non-destroyed object with nlink=0
2409 * the API allows this, but ldiskfs doesn't like and then report
2410 * this inode as deleted
2412 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2413 o->oo_inode->i_nlink == 0));
2417 * Concurrency: shouldn't matter.
2419 static int osd_object_print(const struct lu_env *env, void *cookie,
2420 lu_printer_t p, const struct lu_object *l)
2422 struct osd_object *o = osd_obj(l);
2423 struct iam_descr *d;
2425 if (o->oo_dir != NULL)
2426 d = o->oo_dir->od_container.ic_descr;
2429 return (*p)(env, cookie,
2430 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2432 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2433 o->oo_inode ? o->oo_inode->i_generation : 0,
2434 d ? d->id_ops->id_name : "plain");
2438 * Concurrency: shouldn't matter.
2440 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2441 struct obd_statfs *sfs, struct obd_statfs_info *info)
2443 struct osd_device *osd = osd_dt_dev(d);
2444 struct super_block *sb = osd_sb(osd);
2445 struct kstatfs *ksfs;
2449 if (unlikely(osd->od_mnt == NULL))
2450 return -EINPROGRESS;
2452 /* osd_lproc.c call this without env, allocate ksfs for that case */
2453 if (unlikely(env == NULL)) {
2454 OBD_ALLOC_PTR(ksfs);
2458 ksfs = &osd_oti_get(env)->oti_ksfs;
2461 result = sb->s_op->statfs(sb->s_root, ksfs);
2465 statfs_pack(sfs, ksfs);
2466 if (unlikely(sb->s_flags & SB_RDONLY))
2467 sfs->os_state |= OS_STATFS_READONLY;
2469 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2471 if (ldiskfs_has_feature_extents(sb))
2472 sfs->os_maxbytes = sb->s_maxbytes;
2474 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2477 * Reserve some space so to avoid fragmenting the filesystem too much.
2478 * Fragmentation not only impacts performance, but can also increase
2479 * metadata overhead significantly, causing grant calculation to be
2482 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2484 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2485 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2486 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2487 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2489 sfs->os_blocks -= reserved;
2490 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2491 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2494 if (unlikely(env == NULL))
2500 * Estimate space needed for file creations. We assume the largest filename
2501 * which is 2^64 - 1, hence a filename of 20 chars.
2502 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2504 #ifdef __LDISKFS_DIR_REC_LEN
2505 # define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2506 #elif defined LDISKFS_DIR_REC_LEN_WITH_DIR
2507 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20, NULL)
2509 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2513 * Concurrency: doesn't access mutable data.
2515 static void osd_conf_get(const struct lu_env *env,
2516 const struct dt_device *dev,
2517 struct dt_device_param *param)
2519 struct osd_device *d = osd_dt_dev(dev);
2520 struct super_block *sb = osd_sb(d);
2521 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2526 * XXX should be taken from not-yet-existing fs abstraction layer.
2528 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2529 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2530 param->ddp_symlink_max = sb->s_blocksize;
2531 param->ddp_mount_type = LDD_MT_LDISKFS;
2532 if (ldiskfs_has_feature_extents(sb))
2533 param->ddp_maxbytes = sb->s_maxbytes;
2535 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2537 * inode are statically allocated, so per-inode space consumption
2538 * is the space consumed by the directory entry
2540 param->ddp_inodespace = PER_OBJ_USAGE;
2542 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2543 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2544 * maximum length to not under-count the actual number of extents
2545 * needed for writing a file if there are sub-optimal block allocations.
2547 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2548 /* worst-case extent insertion metadata overhead */
2549 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2550 param->ddp_mntopts = 0;
2551 if (test_opt(sb, XATTR_USER))
2552 param->ddp_mntopts |= MNTOPT_USERXATTR;
2553 if (test_opt(sb, POSIX_ACL))
2554 param->ddp_mntopts |= MNTOPT_ACL;
2557 * LOD might calculate the max stripe count based on max_ea_size,
2558 * so we need take account in the overhead as well,
2559 * xattr_header + magic + xattr_entry_head
2561 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2562 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2564 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2565 if (ldiskfs_has_feature_ea_inode(sb))
2566 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2570 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2572 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2573 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2576 * Preferred RPC size for efficient disk IO. 4MB shows good
2577 * all-around performance for ldiskfs, but use bigalloc chunk size
2578 * by default if larger.
2580 #if defined(LDISKFS_CLUSTER_SIZE)
2581 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2582 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2585 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2587 param->ddp_t10_cksum_type = 0;
2589 unsigned short interval = blk_integrity_interval(bi);
2590 name = blk_integrity_name(bi);
2600 if (strncmp(name, "T10-DIF-TYPE",
2601 sizeof("T10-DIF-TYPE") - 1) == 0) {
2602 /* also skip "1/2/3-" at end */
2603 const int type_off = sizeof("T10-DIF-TYPE.");
2604 char type_number = name[type_off - 2];
2606 if (interval != 512 && interval != 4096) {
2607 CERROR("%s: unsupported T10PI sector size %u\n",
2608 d->od_svname, interval);
2611 switch (type_number) {
2613 d->od_t10_type = OSD_T10_TYPE1;
2616 d->od_t10_type = OSD_T10_TYPE2;
2619 d->od_t10_type = OSD_T10_TYPE3;
2622 CERROR("%s: unsupported T10PI type %s\n",
2623 d->od_svname, name);
2626 if (strcmp(name + type_off, "CRC") == 0) {
2627 d->od_t10_type |= OSD_T10_TYPE_CRC;
2628 param->ddp_t10_cksum_type = interval == 512 ?
2629 OBD_CKSUM_T10CRC512 :
2631 } else if (strcmp(name + type_off, "IP") == 0) {
2632 d->od_t10_type |= OSD_T10_TYPE_IP;
2633 param->ddp_t10_cksum_type = interval == 512 ?
2634 OBD_CKSUM_T10IP512 :
2637 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2638 d->od_svname, name);
2643 CERROR("%s: unsupported T10PI type '%s'\n",
2644 d->od_svname, name);
2649 param->ddp_has_lseek_data_hole = true;
2652 static struct vfsmount *osd_mnt_get(const struct dt_device *d)
2654 return osd_dt_dev(d)->od_mnt;
2658 * Concurrency: shouldn't matter.
2660 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2663 struct super_block *s = osd_sb(osd_dt_dev(d));
2666 down_read(&s->s_umount);
2667 rc = s->s_op->sync_fs(s, 1);
2668 up_read(&s->s_umount);
2670 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2677 * Start commit for OSD device.
2679 * An implementation of dt_commit_async method for OSD device.
2680 * Asychronously starts underlayng fs sync and thereby a transaction
2683 * \param env environment
2684 * \param d dt device
2686 * \see dt_device_operations
2688 static int osd_commit_async(const struct lu_env *env,
2689 struct dt_device *d)
2691 struct super_block *s = osd_sb(osd_dt_dev(d));
2696 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2697 down_read(&s->s_umount);
2698 rc = s->s_op->sync_fs(s, 0);
2699 up_read(&s->s_umount);
2705 * Concurrency: shouldn't matter.
2707 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2709 struct super_block *sb = osd_sb(osd_dt_dev(d));
2710 struct block_device *dev = sb->s_bdev;
2711 int rc = -EOPNOTSUPP;
2715 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2716 osd_dt_dev(d)->od_svname, (long)dev, rc);
2722 * Note: we do not count into QUOTA here.
2723 * If we mount with --data_journal we may need more.
2725 const int osd_dto_credits_noquota[DTO_NR] = {
2728 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2729 * SINGLEDATA_TRANS_BLOCKS(8)
2730 * XXX Note: maybe iam need more, since iam have more level than
2733 [DTO_INDEX_INSERT] = 16,
2736 * just modify a single entry, probably merge few within a block
2738 [DTO_INDEX_DELETE] = 1,
2742 [DTO_INDEX_UPDATE] = 16,
2744 * 4(inode, inode bits, groups, GDT)
2745 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2747 [DTO_OBJECT_CREATE] = 4,
2749 * 4(inode, inode bits, groups, GDT)
2750 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2752 [DTO_OBJECT_DELETE] = 4,
2754 * Attr set credits (inode)
2756 [DTO_ATTR_SET_BASE] = 1,
2758 * Xattr set. The same as xattr of EXT3.
2759 * DATA_TRANS_BLOCKS(14)
2760 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2761 * are also counted in. Do not know why?
2763 [DTO_XATTR_SET] = 14,
2765 * credits for inode change during write.
2767 [DTO_WRITE_BASE] = 3,
2769 * credits for single block write.
2771 [DTO_WRITE_BLOCK] = 14,
2773 * Attr set credits for chown.
2774 * This is extra credits for setattr, and it is null without quota
2776 [DTO_ATTR_SET_CHOWN] = 0
2779 /* reserve or free quota for some operation */
2780 static int osd_reserve_or_free_quota(const struct lu_env *env,
2781 struct dt_device *dev,
2782 struct lquota_id_info *qi)
2784 struct osd_device *osd = osd_dt_dev(dev);
2785 struct qsd_instance *qsd = NULL;
2791 qsd = osd->od_quota_slave_dt;
2793 qsd = osd->od_quota_slave_md;
2795 rc = qsd_reserve_or_free_quota(env, qsd, qi);
2799 static const struct dt_device_operations osd_dt_ops = {
2800 .dt_root_get = osd_root_get,
2801 .dt_statfs = osd_statfs,
2802 .dt_trans_create = osd_trans_create,
2803 .dt_trans_start = osd_trans_start,
2804 .dt_trans_stop = osd_trans_stop,
2805 .dt_trans_cb_add = osd_trans_cb_add,
2806 .dt_conf_get = osd_conf_get,
2807 .dt_mnt_get = osd_mnt_get,
2808 .dt_sync = osd_sync,
2810 .dt_commit_async = osd_commit_async,
2811 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
2814 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2817 struct osd_object *obj = osd_dt_obj(dt);
2818 struct osd_thread_info *oti = osd_oti_get(env);
2820 LINVRNT(osd_invariant(obj));
2822 LASSERT(obj->oo_owner != env);
2823 down_read_nested(&obj->oo_sem, role);
2825 LASSERT(obj->oo_owner == NULL);
2829 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2832 struct osd_object *obj = osd_dt_obj(dt);
2833 struct osd_thread_info *oti = osd_oti_get(env);
2835 LINVRNT(osd_invariant(obj));
2837 LASSERT(obj->oo_owner != env);
2838 down_write_nested(&obj->oo_sem, role);
2840 LASSERT(obj->oo_owner == NULL);
2841 obj->oo_owner = env;
2845 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2847 struct osd_object *obj = osd_dt_obj(dt);
2848 struct osd_thread_info *oti = osd_oti_get(env);
2850 LINVRNT(osd_invariant(obj));
2852 LASSERT(oti->oti_r_locks > 0);
2854 up_read(&obj->oo_sem);
2857 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2859 struct osd_object *obj = osd_dt_obj(dt);
2860 struct osd_thread_info *oti = osd_oti_get(env);
2862 LINVRNT(osd_invariant(obj));
2864 LASSERT(obj->oo_owner == env);
2865 LASSERT(oti->oti_w_locks > 0);
2867 obj->oo_owner = NULL;
2868 up_write(&obj->oo_sem);
2871 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2873 struct osd_object *obj = osd_dt_obj(dt);
2875 LINVRNT(osd_invariant(obj));
2877 return obj->oo_owner == env;
2880 static void osd_inode_getattr(const struct lu_env *env,
2881 struct inode *inode, struct lu_attr *attr)
2883 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2884 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2885 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2886 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2888 attr->la_atime = inode_get_atime_sec(inode);
2889 attr->la_mtime = inode_get_mtime_sec(inode);
2890 attr->la_ctime = inode_get_ctime_sec(inode);
2891 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2892 attr->la_mode = inode->i_mode;
2893 attr->la_size = i_size_read(inode);
2894 attr->la_blocks = inode->i_blocks;
2895 attr->la_uid = i_uid_read(inode);
2896 attr->la_gid = i_gid_read(inode);
2897 attr->la_projid = i_projid_read(inode);
2898 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2899 attr->la_nlink = inode->i_nlink;
2900 attr->la_rdev = inode->i_rdev;
2901 attr->la_blksize = 1 << inode->i_blkbits;
2902 attr->la_blkbits = inode->i_blkbits;
2904 * Ext4 did not transfer inherit flags from raw inode
2905 * to inode flags, and ext4 internally test raw inode
2906 * @i_flags directly. Instead of patching ext4, we do it here.
2908 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2909 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2912 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2915 struct osd_object *obj = osd_dt_obj(dt);
2916 const struct dt_it_ops *iops;
2922 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2923 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2925 /* directory not initialized yet */
2926 if (!dt->do_index_ops) {
2931 spin_lock(&obj->oo_guard);
2932 *count = atomic_read(&obj->oo_dirent_count);
2933 if (*count == LU_DIRENT_COUNT_UNSET)
2934 atomic_set(&obj->oo_dirent_count, 0);
2935 spin_unlock(&obj->oo_guard);
2936 if (*count != LU_DIRENT_COUNT_UNSET)
2940 iops = &dt->do_index_ops->dio_it;
2941 it = iops->init(env, dt, LUDA_64BITHASH);
2943 GOTO(out, rc = PTR_ERR(it));
2945 rc = iops->load(env, it, 0);
2952 rc = iops->next(env, it);
2954 for (; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2958 if (iops->key_size(env, it) == 0)
2963 if (rc == 1 || rc == -ESTALE)
2967 iops->fini(env, it);
2969 /* If counting dirents failed, use the current count (if any).
2971 * At worst this means the directory will not be split until the
2972 * count can be completed successfully (remount or oo_dirent_count
2973 * incremented by adding new entries). This avoids re-walking
2974 * the whole directory on each access and hitting the same error.
2976 if (rc && *count == 0)
2977 *count = LU_DIRENT_COUNT_UNSET;
2978 atomic_set(&obj->oo_dirent_count, *count);
2982 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2983 struct lu_attr *attr)
2985 struct osd_object *obj = osd_dt_obj(dt);
2988 if (unlikely(!dt_object_exists(dt)))
2990 if (unlikely(obj->oo_destroyed))
2993 LASSERT(!dt_object_remote(dt));
2994 LINVRNT(osd_invariant(obj));
2996 spin_lock(&obj->oo_guard);
2997 osd_inode_getattr(env, obj->oo_inode, attr);
2998 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2999 attr->la_valid |= LA_FLAGS;
3000 attr->la_flags |= LUSTRE_ORPHAN_FL;
3002 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
3003 attr->la_valid |= LA_FLAGS;
3004 attr->la_flags |= LUSTRE_ENCRYPT_FL;
3006 spin_unlock(&obj->oo_guard);
3008 if (S_ISDIR(obj->oo_inode->i_mode) &&
3009 (attr->la_valid & LA_DIRENT_CNT) &&
3010 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
3011 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
3013 attr->la_valid &= ~LA_DIRENT_CNT;
3018 static int osd_declare_attr_qid(const struct lu_env *env,
3019 struct osd_object *obj,
3020 struct osd_thandle *oh, long long bspace,
3021 qid_t old_id, qid_t new_id, bool enforce,
3025 struct osd_thread_info *info = osd_oti_get(env);
3026 struct lquota_id_info *qi = &info->oti_qi;
3028 qi->lqi_type = type;
3029 /* inode accounting */
3030 qi->lqi_is_blk = false;
3032 /* one more inode for the new id ... */
3033 qi->lqi_id.qid_uid = new_id;
3035 /* Reserve credits for the new id */
3036 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
3037 if (rc == -EDQUOT || rc == -EINPROGRESS)
3042 /* and one less inode for the current id */
3043 qi->lqi_id.qid_uid = old_id;
3045 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3046 if (rc == -EDQUOT || rc == -EINPROGRESS)
3051 /* block accounting */
3052 qi->lqi_is_blk = true;
3054 /* more blocks for the new id ... */
3055 qi->lqi_id.qid_uid = new_id;
3056 qi->lqi_space = bspace;
3058 * Credits for the new uid has been reserved, re-use "obj"
3059 * to save credit reservation.
3061 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3062 if (rc == -EDQUOT || rc == -EINPROGRESS)
3067 /* and finally less blocks for the current uid */
3068 qi->lqi_id.qid_uid = old_id;
3069 qi->lqi_space = -bspace;
3070 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
3071 if (rc == -EDQUOT || rc == -EINPROGRESS)
3077 static int osd_declare_attr_set(const struct lu_env *env,
3078 struct dt_object *dt,
3079 const struct lu_attr *attr,
3080 struct thandle *handle)
3082 struct osd_thandle *oh;
3083 struct osd_object *obj;
3092 LASSERT(dt != NULL);
3093 LASSERT(handle != NULL);
3095 obj = osd_dt_obj(dt);
3096 LASSERT(osd_invariant(obj));
3098 oh = container_of(handle, struct osd_thandle, ot_super);
3099 LASSERT(oh->ot_handle == NULL);
3101 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
3102 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
3104 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
3105 osd_dto_credits_noquota[DTO_XATTR_SET]);
3107 if (attr == NULL || obj->oo_inode == NULL)
3110 bspace = obj->oo_inode->i_blocks << 9;
3111 bspace = toqb(bspace);
3114 * Changing ownership is always preformed by super user, it should not
3115 * fail with EDQUOT unless required explicitly.
3117 * We still need to call the osd_declare_qid() to calculate the journal
3118 * credits for updating quota accounting files and to trigger quota
3119 * space adjustment once the operation is completed.
3121 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
3123 uid = i_uid_read(obj->oo_inode);
3124 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
3125 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
3126 attr->la_uid, enforce, USRQUOTA);
3130 gid = i_gid_read(obj->oo_inode);
3131 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
3132 attr->la_uid, gid, attr->la_gid);
3133 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
3134 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
3135 attr->la_gid, enforce, GRPQUOTA);
3140 #ifdef HAVE_PROJECT_QUOTA
3141 if (attr->la_valid & LA_PROJID) {
3142 __u32 projid = i_projid_read(obj->oo_inode);
3144 enforce = (attr->la_valid & LA_PROJID) &&
3145 (attr->la_projid != projid);
3146 rc = osd_declare_attr_qid(env, obj, oh, bspace,
3147 (qid_t)projid, (qid_t)attr->la_projid,
3153 /* punch must be aware we are dealing with an encrypted file */
3154 if (attr->la_valid & LA_FLAGS && attr->la_flags & LUSTRE_ENCRYPT_FL)
3155 obj->oo_lma_flags |= LUSTRE_ENCRYPT_FL;
3160 static int osd_inode_setattr(const struct lu_env *env,
3161 struct inode *inode, const struct lu_attr *attr)
3163 __u64 bits = attr->la_valid;
3165 /* Only allow set size for regular file */
3166 if (!S_ISREG(inode->i_mode))
3167 bits &= ~(LA_SIZE | LA_BLOCKS);
3172 if (bits & LA_ATIME)
3173 inode_set_atime_to_ts(inode,
3174 osd_inode_time(inode, attr->la_atime));
3175 if (bits & LA_CTIME)
3176 inode_set_ctime_to_ts(inode,
3177 osd_inode_time(inode, attr->la_ctime));
3178 if (bits & LA_MTIME)
3179 inode_set_mtime_to_ts(inode,
3180 osd_inode_time(inode, attr->la_mtime));
3181 if (bits & LA_SIZE) {
3182 spin_lock(&inode->i_lock);
3183 LDISKFS_I(inode)->i_disksize = attr->la_size;
3184 i_size_write(inode, attr->la_size);
3185 spin_unlock(&inode->i_lock);
3189 * OSD should not change "i_blocks" which is used by quota.
3190 * "i_blocks" should be changed by ldiskfs only.
3193 inode->i_mode = (inode->i_mode & S_IFMT) |
3194 (attr->la_mode & ~S_IFMT);
3196 i_uid_write(inode, attr->la_uid);
3198 i_gid_write(inode, attr->la_gid);
3199 if (bits & LA_PROJID)
3200 i_projid_write(inode, attr->la_projid);
3201 if (bits & LA_NLINK)
3202 set_nlink(inode, attr->la_nlink);
3204 inode->i_rdev = attr->la_rdev;
3206 if (bits & LA_FLAGS) {
3207 /* always keep S_NOCMTIME */
3208 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
3210 #if defined(S_ENCRYPTED)
3211 /* Always remove S_ENCRYPTED, because ldiskfs must not be
3212 * aware of encryption status. It is just stored into LMA
3213 * so that it can be forwared to client side.
3215 inode->i_flags &= ~S_ENCRYPTED;
3218 * Ext4 did not transfer inherit flags from
3219 * @inode->i_flags to raw inode i_flags when writing
3220 * flags, we do it explictly here.
3222 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
3223 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
3225 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
3230 #ifdef HAVE_PROJECT_QUOTA
3231 static int osd_transfer_project(struct inode *inode, __u32 projid,
3232 struct thandle *handle)
3234 struct super_block *sb = inode->i_sb;
3235 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3238 struct ldiskfs_iloc iloc;
3239 struct ldiskfs_inode *raw_inode;
3240 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3242 if (!ldiskfs_has_feature_project(sb)) {
3243 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3244 == LDISKFS_DEF_PROJID);
3245 if (projid != LDISKFS_DEF_PROJID)
3251 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3254 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3255 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3258 err = ldiskfs_get_inode_loc(inode, &iloc);
3262 raw_inode = ldiskfs_raw_inode(&iloc);
3263 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3264 struct osd_thandle *oh = container_of(handle,
3268 * try to expand inode size automatically.
3270 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3271 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3279 dquot_initialize(inode);
3280 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3281 if (transfer_to[PRJQUOTA]) {
3282 lock_dquot_transfer(inode);
3283 err = __dquot_transfer(inode, transfer_to);
3284 unlock_dquot_transfer(inode);
3285 dqput(transfer_to[PRJQUOTA]);
3294 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3295 struct thandle *handle)
3299 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3300 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3304 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3305 inode->i_ino, i_uid_read(inode), attr->la_uid,
3306 i_gid_read(inode), attr->la_gid);
3308 dquot_initialize(inode);
3310 if (attr->la_valid & LA_UID)
3311 iattr.ia_valid |= ATTR_UID;
3312 if (attr->la_valid & LA_GID)
3313 iattr.ia_valid |= ATTR_GID;
3314 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3315 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3317 lock_dquot_transfer(inode);
3318 rc = osd_dquot_transfer(&nop_mnt_idmap, inode, &iattr);
3319 unlock_dquot_transfer(inode);
3321 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3322 osd_ino2name(inode), rc);
3327 /* Handle project id transfer here properly */
3328 if (attr->la_valid & LA_PROJID &&
3329 attr->la_projid != i_projid_read(inode)) {
3330 if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
3332 #ifdef HAVE_PROJECT_QUOTA
3333 rc = osd_transfer_project(inode, attr->la_projid, handle);
3338 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3339 osd_ino2name(inode), rc);
3346 static int osd_attr_set(const struct lu_env *env,
3347 struct dt_object *dt,
3348 const struct lu_attr *attr,
3349 struct thandle *handle)
3351 struct osd_thread_info *info = osd_oti_get(env);
3352 struct osd_object *obj = osd_dt_obj(dt);
3353 struct osd_device *osd = osd_obj2dev(obj);
3354 struct inode *inode;
3357 if (!dt_object_exists(dt))
3360 LASSERT(handle != NULL);
3361 LASSERT(!dt_object_remote(dt));
3362 LASSERT(osd_invariant(obj));
3364 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3366 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3367 !osd_obj2dev(obj)->od_is_ost) {
3368 struct osd_thread_info *oti = osd_oti_get(env);
3369 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3370 struct lu_fid *fid1 = &oti->oti_fid;
3371 struct osd_inode_id *id = &oti->oti_id;
3372 struct iam_path_descr *ipd;
3373 struct iam_container *bag;
3374 struct osd_thandle *oh;
3377 fid_cpu_to_be(fid1, fid0);
3378 memset(id, 1, sizeof(*id));
3379 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3380 fid0)->oi_dir.od_container;
3381 ipd = osd_idx_ipd_get(env, bag);
3382 if (unlikely(ipd == NULL))
3385 oh = container_of(handle, struct osd_thandle, ot_super);
3386 rc = iam_update(oh->ot_handle, bag,
3387 (const struct iam_key *)fid1,
3388 (const struct iam_rec *)id, ipd);
3389 osd_ipd_put(env, bag, ipd);
3390 return(rc > 0 ? 0 : rc);
3393 inode = obj->oo_inode;
3395 rc = osd_quota_transfer(inode, attr, handle);
3399 spin_lock(&obj->oo_guard);
3400 rc = osd_inode_setattr(env, inode, attr);
3401 spin_unlock(&obj->oo_guard);
3405 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3407 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3409 if (!(attr->la_valid & LA_FLAGS))
3412 /* If setting LUSTRE_ENCRYPT_FL on an OST object, also set a dummy
3413 * enc ctx xattr, with 2 benefits:
3414 * - setting the LL_XATTR_NAME_ENCRYPTION_CONTEXT xattr internally sets
3415 * the LDISKFS_ENCRYPT_FL flag on the on-disk inode;
3416 * - it makes e2fsprogs happy to see an enc ctx for an inode that has
3417 * the LDISKFS_ENCRYPT_FL flag
3418 * We do not need the actual encryption context on OST objects, it is
3419 * only stored on MDT inodes, at file creation time.
3421 if (!(LDISKFS_I(obj->oo_inode)->i_flags & LDISKFS_ENCRYPT_FL) &&
3422 attr->la_flags & LUSTRE_ENCRYPT_FL && osd->od_is_ost &&
3423 !CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) {
3426 /* use a dummy enc ctx, fine with e2fsprogs */
3427 buf.lb_buf = "\xFF";
3429 rc = osd_xattr_set(env, dt, &buf,
3430 LL_XATTR_NAME_ENCRYPTION_CONTEXT,
3433 CWARN("%s: set "DFID" enc ctx failed: rc = %d\n",
3434 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3438 /* Let's check if there are extra flags need to be set into LMA */
3439 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3440 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3442 LASSERT(!obj->oo_pfid_in_lma);
3444 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3445 &info->oti_ost_attrs);
3449 if ((lma->lma_incompat & lustre_to_lma_flags(attr->la_flags)) ==
3450 lustre_to_lma_flags(attr->la_flags))
3451 /* if lma incompat already has the flags,
3452 * save a useless call to xattr_set
3456 lma->lma_incompat |=
3457 lustre_to_lma_flags(attr->la_flags);
3458 lustre_lma_swab(lma);
3460 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3462 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3463 lma, sizeof(*lma), XATTR_REPLACE);
3465 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3466 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3467 lma->lma_incompat, rc);
3470 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3471 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3478 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3479 struct osd_object *obj,
3480 const char *name, const int namelen)
3482 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3485 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3486 umode_t mode, struct dt_allocation_hint *hint,
3487 struct thandle *th, struct lu_attr *attr)
3490 struct osd_device *osd = osd_obj2dev(obj);
3491 struct osd_thandle *oth;
3492 struct dt_object *parent = NULL;
3493 struct inode *inode;
3494 struct iattr iattr = {
3495 .ia_valid = ATTR_UID | ATTR_GID |
3496 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3497 .ia_ctime.tv_sec = attr->la_ctime,
3498 .ia_mtime.tv_sec = attr->la_mtime,
3499 .ia_atime.tv_sec = attr->la_atime,
3500 .ia_uid = GLOBAL_ROOT_UID,
3501 .ia_gid = GLOBAL_ROOT_GID,
3503 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3505 if (attr->la_valid & LA_UID)
3506 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3507 if (attr->la_valid & LA_GID)
3508 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3510 LINVRNT(osd_invariant(obj));
3511 LASSERT(obj->oo_inode == NULL);
3512 LASSERT(obj->oo_hl_head == NULL);
3514 if (S_ISDIR(mode) && ldiskfs_pdo) {
3516 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3517 if (obj->oo_hl_head == NULL)
3521 oth = container_of(th, struct osd_thandle, ot_super);
3522 LASSERT(oth->ot_handle->h_transaction != NULL);
3524 if (hint != NULL && hint->dah_parent != NULL &&
3525 !dt_object_remote(hint->dah_parent))
3526 parent = hint->dah_parent;
3528 /* if a time component is not valid set it to UTIME_OMIT */
3529 if (!(attr->la_valid & LA_CTIME))
3530 iattr.ia_ctime = omit;
3531 if (!(attr->la_valid & LA_MTIME))
3532 iattr.ia_mtime = omit;
3533 if (!(attr->la_valid & LA_ATIME))
3534 iattr.ia_atime = omit;
3536 inode = ldiskfs_create_inode(oth->ot_handle,
3537 parent ? osd_dt_obj(parent)->oo_inode :
3538 osd_sb(osd)->s_root->d_inode,
3540 if (!IS_ERR(inode)) {
3541 /* Do not update file c/mtime in ldiskfs. */
3542 inode->i_flags |= S_NOCMTIME;
3545 * For new created object, it must be consistent,
3546 * and it is unnecessary to scrub against it.
3548 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3550 obj->oo_inode = inode;
3553 if (obj->oo_hl_head != NULL) {
3554 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3555 obj->oo_hl_head = NULL;
3557 result = PTR_ERR(inode);
3559 LINVRNT(osd_invariant(obj));
3567 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3568 struct lu_attr *attr,
3569 struct dt_allocation_hint *hint,
3570 struct dt_object_format *dof,
3574 struct osd_thandle *oth;
3575 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3577 LASSERT(S_ISDIR(attr->la_mode));
3579 oth = container_of(th, struct osd_thandle, ot_super);
3580 LASSERT(oth->ot_handle->h_transaction != NULL);
3581 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3582 atomic_set(&obj->oo_dirent_count, 0);
3583 result = osd_mkfile(info, obj, mode, hint, th, attr);
3588 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3589 struct lu_attr *attr,
3590 struct dt_allocation_hint *hint,
3591 struct dt_object_format *dof,
3595 struct osd_thandle *oth;
3596 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3598 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3600 LASSERT(S_ISREG(attr->la_mode));
3602 oth = container_of(th, struct osd_thandle, ot_super);
3603 LASSERT(oth->ot_handle->h_transaction != NULL);
3605 result = osd_mkfile(info, obj, mode, hint, th, attr);
3607 LASSERT(obj->oo_inode != NULL);
3608 if (feat->dif_flags & DT_IND_VARKEY)
3609 result = iam_lvar_create(obj->oo_inode,
3610 feat->dif_keysize_max,
3612 feat->dif_recsize_max,
3615 result = iam_lfix_create(obj->oo_inode,
3616 feat->dif_keysize_max,
3618 feat->dif_recsize_max,
3624 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3625 struct lu_attr *attr,
3626 struct dt_allocation_hint *hint,
3627 struct dt_object_format *dof,
3630 LASSERT(S_ISREG(attr->la_mode));
3631 return osd_mkfile(info, obj, (attr->la_mode &
3632 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3636 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3637 struct lu_attr *attr,
3638 struct dt_allocation_hint *hint,
3639 struct dt_object_format *dof,
3642 LASSERT(S_ISLNK(attr->la_mode));
3643 return osd_mkfile(info, obj, (attr->la_mode &
3644 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3648 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3649 struct lu_attr *attr,
3650 struct dt_allocation_hint *hint,
3651 struct dt_object_format *dof,
3654 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3657 LINVRNT(osd_invariant(obj));
3658 LASSERT(obj->oo_inode == NULL);
3659 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3660 S_ISFIFO(mode) || S_ISSOCK(mode));
3662 result = osd_mkfile(info, obj, mode, hint, th, attr);
3664 LASSERT(obj->oo_inode != NULL);
3666 * This inode should be marked dirty for i_rdev. Currently
3667 * that is done in the osd_attr_init().
3669 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3672 LINVRNT(osd_invariant(obj));
3676 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3678 struct dt_allocation_hint *hint,
3679 struct dt_object_format *dof,
3682 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3684 osd_obj_type_f result;
3700 result = osd_mk_index;
3710 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3711 struct dt_object *parent, struct dt_object *child,
3716 ah->dah_parent = parent;
3718 if (parent != NULL && !dt_object_remote(parent)) {
3719 /* will help to find FID->ino at dt_insert("..") */
3720 struct osd_object *pobj = osd_dt_obj(parent);
3722 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3726 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3727 struct lu_attr *attr, struct dt_object_format *dof,
3728 struct thandle *handle)
3730 struct inode *inode = obj->oo_inode;
3731 __u64 valid = attr->la_valid;
3734 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3736 if (dof->dof_type != DFT_NODE)
3737 attr->la_valid &= ~LA_RDEV;
3738 if ((valid & LA_ATIME) &&
3739 (attr->la_atime == inode_get_atime_sec(inode)))
3740 attr->la_valid &= ~LA_ATIME;
3741 if ((valid & LA_CTIME) &&
3742 (attr->la_ctime == inode_get_ctime_sec(inode)))
3743 attr->la_valid &= ~LA_CTIME;
3744 if ((valid & LA_MTIME) &&
3745 (attr->la_mtime == inode_get_mtime_sec(inode)))
3746 attr->la_valid &= ~LA_MTIME;
3748 result = osd_quota_transfer(inode, attr, handle);
3752 if (attr->la_valid != 0) {
3753 result = osd_inode_setattr(info->oti_env, inode, attr);
3755 * The osd_inode_setattr() should always succeed here. The
3756 * only error that could be returned is EDQUOT when we are
3757 * trying to change the UID or GID of the inode. However, this
3758 * should not happen since quota enforcement is no longer
3759 * enabled on ldiskfs (lquota takes care of it).
3761 LASSERTF(result == 0, "%d\n", result);
3762 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3765 attr->la_valid = valid;
3769 * Helper function for osd_create()
3771 * \retval 0, on success
3773 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3774 struct lu_attr *attr, struct dt_allocation_hint *hint,
3775 struct dt_object_format *dof, struct thandle *th)
3780 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3782 /* we drop umask so that permissions we pass are not affected */
3783 umask = current->fs->umask;
3784 current->fs->umask = 0;
3786 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3788 if (likely(obj->oo_inode != NULL)) {
3789 LASSERT(obj->oo_inode->i_state & I_NEW);
3792 * Unlock the inode before attr initialization to avoid
3793 * unnecessary dqget operations. LU-6378
3795 unlock_new_inode(obj->oo_inode);
3798 if (likely(result == 0)) {
3799 osd_attr_init(info, obj, attr, dof, th);
3800 osd_object_init0(obj);
3803 /* restore previous umask value */
3804 current->fs->umask = umask;
3806 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3812 * Helper function for osd_create()
3814 * \retval 0, on success
3816 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3817 const struct lu_fid *fid, struct thandle *th)
3819 struct osd_thread_info *info = osd_oti_get(env);
3820 struct osd_inode_id *id = &info->oti_id;
3821 struct osd_device *osd = osd_obj2dev(obj);
3822 struct osd_thandle *oh;
3825 LASSERT(obj->oo_inode != NULL);
3827 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3830 oh = container_of(th, struct osd_thandle, ot_super);
3831 LASSERT(oh->ot_handle);
3832 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3834 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3835 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_FID_REUSE) && osd->od_is_ost &&
3837 struct lu_fid tfid = *fid;
3840 osd_oi_insert(info, osd, &tfid, id, oh->ot_handle,
3841 OI_CHECK_FLD, NULL);
3842 /* clear NOSCRUB flag so that it can be scrubbed immediately */
3843 ldiskfs_clear_inode_state(obj->oo_inode,
3844 LDISKFS_STATE_LUSTRE_NOSCRUB);
3847 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3848 OI_CHECK_FLD, NULL);
3850 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3851 struct lu_fid next_fid = *fid;
3853 /* insert next object in advance, and map to the same inode */
3855 if (next_fid.f_oid != 0) {
3856 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3857 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3858 OI_CHECK_FLD, NULL);
3859 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3863 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3868 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3869 u64 seq, struct lu_seq_range *range)
3871 struct seq_server_site *ss = osd_seq_site(osd);
3873 if (fid_seq_is_idif(seq)) {
3874 fld_range_set_ost(range);
3875 range->lsr_index = idif_ost_idx(seq);
3879 if (!fid_seq_in_fldb(seq)) {
3880 fld_range_set_mdt(range);
3883 * FIXME: If ss is NULL, it suppose not get lsr_index
3886 range->lsr_index = ss->ss_node_id;
3890 LASSERT(ss != NULL);
3891 fld_range_set_any(range);
3892 /* OSD will only do local fld lookup */
3893 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3896 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3897 struct lu_attr *attr,
3898 struct dt_allocation_hint *hint,
3899 struct dt_object_format *dof,
3900 struct thandle *handle)
3902 struct osd_thandle *oh;
3903 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3909 LASSERT(handle != NULL);
3911 oh = container_of(handle, struct osd_thandle, ot_super);
3912 LASSERT(oh->ot_handle == NULL);
3914 if (osd_tx_was_declared(env, oh, dt, DTO_OBJECT_CREATE, 0))
3918 * EA object consumes more credits than regular object: osd_mk_index
3919 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3920 * leaf_node, could involves the block, block bitmap, groups, GDT
3921 * change for each block, so add 4 * 2 credits in that case.
3923 * The default ACL initialization may consume an additional 16 blocks
3925 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3926 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3929 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3930 * credits for possible default ACL creation in new inode
3932 if (hint && hint->dah_acl_len)
3933 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3934 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3935 sb->s_blocksize_bits);
3937 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3940 * Reuse idle OI block may cause additional one OI block
3943 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3944 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3945 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3946 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3947 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3949 /* will help to find FID->ino mapping at dt_insert() */
3950 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3958 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3959 attr->la_projid, 1, oh, osd_dt_obj(dt),
3960 NULL, OSD_QID_INODE);
3968 * Called to destroy on-disk representation of the object
3970 * Concurrency: must be locked
3972 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3975 struct osd_object *obj = osd_dt_obj(dt);
3976 struct inode *inode = obj->oo_inode;
3977 struct osd_thandle *oh;
3985 oh = container_of(th, struct osd_thandle, ot_super);
3986 LASSERT(oh->ot_handle == NULL);
3988 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3989 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3991 /* For removing agent entry */
3992 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3993 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3996 * Recycle idle OI leaf may cause additional three OI blocks
3999 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
4000 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
4001 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
4002 /* one less inode */
4003 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4004 i_projid_read(inode), -1, oh, obj, NULL,
4008 /* data to be truncated */
4009 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
4010 i_projid_read(inode), 0, oh, obj, NULL,
4016 * will help to find FID->ino when this object is being
4019 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
4024 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
4027 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4028 struct osd_object *obj = osd_dt_obj(dt);
4029 struct inode *inode = obj->oo_inode;
4030 struct osd_device *osd = osd_obj2dev(obj);
4031 struct osd_thandle *oh;
4036 oh = container_of(th, struct osd_thandle, ot_super);
4037 LASSERT(oh->ot_handle);
4039 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
4041 if (unlikely(fid_is_acct(fid)))
4044 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
4045 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
4047 CERROR("%s: remove agent entry "DFID": rc = %d\n",
4048 osd_name(osd), PFID(fid), result);
4051 if (S_ISDIR(inode->i_mode)) {
4052 if (inode->i_nlink > 2)
4053 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
4054 osd_name(osd), PFID(fid), inode->i_ino,
4057 spin_lock(&obj->oo_guard);
4059 spin_unlock(&obj->oo_guard);
4060 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4063 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
4065 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
4067 if (!CFS_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
4068 result = osd_oi_delete(osd_oti_get(env), osd, fid,
4069 oh->ot_handle, OI_CHECK_FLD);
4071 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
4072 /* XXX: add to ext3 orphan list */
4073 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
4075 /* not needed in the cache anymore */
4076 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
4077 obj->oo_destroyed = 1;
4083 * Put the fid into lustre_mdt_attrs, and then place the structure
4084 * inode's ea. This fid should not be altered during the life time
4087 * \retval +ve, on success
4088 * \retval -ve, on error
4090 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
4092 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
4093 const struct lu_fid *fid, __u32 compat, __u32 incompat)
4095 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4096 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4101 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INLMA))
4104 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
4107 lustre_loa_init(loa, fid, compat, incompat);
4108 lustre_loa_swab(loa, false);
4111 * For the OST device with 256 bytes inode size by default,
4112 * the PFID EA will be stored together with LMA EA to avoid
4113 * performance trouble. Otherwise the PFID EA can be stored
4114 * independently. LU-8998
4116 if ((compat & LMAC_FID_ON_OST) &&
4117 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
4118 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4119 sizeof(*loa), XATTR_CREATE);
4121 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4122 sizeof(*lma), XATTR_CREATE);
4124 * LMA may already exist, but we need to check that all the
4125 * desired compat/incompat flags have been added.
4127 if (unlikely(rc == -EEXIST)) {
4128 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
4129 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
4133 if (rc < sizeof(*lma))
4136 lustre_loa_swab(loa, true);
4137 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
4138 ((compat == 0 && incompat == 0) ||
4139 (!(~lma->lma_compat & compat) &&
4140 !(~lma->lma_incompat & incompat))))
4143 lma->lma_self_fid = *fid;
4144 lma->lma_compat |= compat;
4145 lma->lma_incompat |= incompat;
4146 if (rc == sizeof(*lma)) {
4147 lustre_lma_swab(lma);
4148 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4149 sizeof(*lma), XATTR_REPLACE);
4151 lustre_loa_swab(loa, false);
4152 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4153 sizeof(*loa), XATTR_REPLACE);
4161 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
4162 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
4163 * To have compatilibility with 1.8 ldiskfs driver we need to have
4164 * magic number at start of fid data.
4165 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
4168 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
4169 const struct lu_fid *fid)
4171 if (!fid_is_namespace_visible(fid) ||
4172 CFS_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
4173 param->edp_magic = 0;
4177 param->edp_magic = LDISKFS_LUFID_MAGIC;
4178 param->edp_len = sizeof(struct lu_fid) + 1;
4179 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
4183 * Try to read the fid from inode ea into dt_rec.
4185 * \param fid object fid.
4187 * \retval 0 on success
4189 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
4190 __u32 ino, struct lu_fid *fid,
4191 struct osd_inode_id *id)
4193 struct osd_thread_info *info = osd_oti_get(env);
4194 struct inode *inode;
4198 osd_id_gen(id, ino, OSD_OII_NOGEN);
4199 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
4201 RETURN(PTR_ERR(inode));
4207 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
4209 struct inode *parent_dir,
4210 const struct lu_fid *dot_fid,
4211 const struct lu_fid *dot_dot_fid,
4212 struct osd_thandle *oth)
4214 struct ldiskfs_dentry_param *dot_ldp;
4215 struct ldiskfs_dentry_param *dot_dot_ldp;
4216 __u32 saved_nlink = dir->i_nlink;
4219 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
4222 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
4223 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
4225 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4226 dot_ldp->edp_magic = 0;
4228 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
4229 dir, dot_ldp, dot_dot_ldp);
4231 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
4232 * the subseqent ref_add() will increase the dir->i_nlink
4233 * as 3. That is incorrect for new created directory.
4235 * It looks like hack, because we want to make the OSD API
4236 * to be order-independent for new created directory object
4237 * between dt_insert(..) and ref_add() operations.
4239 * Here, we only restore the in-RAM dir-inode's nlink attr,
4240 * becuase if the nlink attr is not 2, then there will be
4241 * ref_add() called following the dt_insert(..), such call
4242 * will make both the in-RAM and on-disk dir-inode's nlink
4243 * attr to be set as 2. LU-7447
4245 set_nlink(dir, saved_nlink);
4250 * Create an local agent inode for remote entry
4252 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
4253 struct osd_device *osd,
4254 struct osd_object *pobj,
4255 const struct lu_fid *fid,
4259 struct osd_thread_info *info = osd_oti_get(env);
4260 struct inode *local;
4261 struct osd_thandle *oh;
4262 struct iattr iattr = {
4263 .ia_valid = ATTR_UID | ATTR_GID |
4264 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
4265 .ia_ctime.tv_nsec = UTIME_OMIT,
4266 .ia_mtime.tv_nsec = UTIME_OMIT,
4267 .ia_atime.tv_nsec = UTIME_OMIT,
4268 .ia_uid = GLOBAL_ROOT_UID,
4269 .ia_gid = GLOBAL_ROOT_GID,
4276 oh = container_of(th, struct osd_thandle, ot_super);
4277 LASSERT(oh->ot_handle->h_transaction != NULL);
4279 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
4281 if (IS_ERR(local)) {
4282 CERROR("%s: create local error %d\n", osd_name(osd),
4283 (int)PTR_ERR(local));
4288 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
4289 * correct gid on remote file, not agent here
4291 local->i_gid = current_fsgid();
4292 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4294 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4295 * That gives e2fsck something to look at and be happy, and allows
4296 * debugging if we need to determine where this symlink came from.
4298 if (S_ISLNK(type)) {
4299 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4300 ldiskfs_clear_inode_flag(local, LDISKFS_INODE_EXTENTS);
4301 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4302 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4304 i_size_write(local, rc);
4305 LDISKFS_I(local)->i_disksize = rc;
4307 unlock_new_inode(local);
4309 /* Agent inode should not have project ID */
4310 #ifdef HAVE_PROJECT_QUOTA
4311 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4312 i_projid_read(pobj->oo_inode) != 0) {
4313 rc = osd_transfer_project(local, 0, th);
4315 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4316 osd_ino2name(local), rc);
4317 RETURN(ERR_PTR(rc));
4321 /* Set special LMA flag for local agent inode */
4322 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4324 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4325 osd_name(osd), PFID(fid), rc);
4326 RETURN(ERR_PTR(rc));
4332 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode, fid,
4333 lu_object_fid(&pobj->oo_dt.do_lu),
4336 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4337 osd_name(osd), PFID(fid), rc);
4338 RETURN(ERR_PTR(rc));
4345 * when direntry is deleted, we have to take care of possible agent inode
4346 * referenced by that. unfortunately we can't do this at that point:
4347 * iget() within a running transaction leads to deadlock and we better do
4348 * not call that every delete declaration to save performance. so we put
4349 * a potention agent inode on a list and process that once the transaction
4350 * is over. Notice it's not any worse in terms of real orphans as regular
4351 * object destroy doesn't put inodes on the on-disk orphan list. this should
4352 * be addressed separately
4354 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4355 struct osd_thandle *oh,
4358 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4359 struct osd_obj_orphan *oor;
4366 oor->oor_env = (struct lu_env *)env;
4367 spin_lock(&osd->od_osfs_lock);
4368 list_add(&oor->oor_list, &osd->od_orphan_list);
4369 spin_unlock(&osd->od_osfs_lock);
4371 oh->ot_remove_agents = 1;
4377 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4378 struct osd_device *osd)
4380 struct osd_thread_info *info = osd_oti_get(env);
4381 struct osd_obj_orphan *oor, *tmp;
4382 struct osd_inode_id id;
4384 struct inode *inode;
4389 spin_lock(&osd->od_osfs_lock);
4390 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4391 if (oor->oor_env == env)
4392 list_move(&oor->oor_list, &list);
4394 spin_unlock(&osd->od_osfs_lock);
4396 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4400 list_del(&oor->oor_list);
4403 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4404 inode = osd_iget_fid(info, osd, &id, &fid);
4408 if (!osd_remote_fid(env, osd, &fid)) {
4413 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4415 mark_inode_dirty(inode);
4416 ldiskfs_journal_stop(jh);
4424 * OSD layer object create function for OST objects (b=11826).
4426 * The FID is inserted into inode xattr here.
4428 * \retval 0, on success
4429 * \retval -ve, on error
4431 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4432 struct lu_attr *attr, struct dt_allocation_hint *hint,
4433 struct dt_object_format *dof, struct thandle *th)
4435 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4436 struct osd_object *obj = osd_dt_obj(dt);
4437 struct osd_thread_info *info = osd_oti_get(env);
4438 int result, on_ost = 0;
4442 if (dt_object_exists(dt))
4445 LINVRNT(osd_invariant(obj));
4446 LASSERT(!dt_object_remote(dt));
4447 LASSERT(osd_is_write_locked(env, obj));
4448 LASSERT(th != NULL);
4450 if (unlikely(fid_is_acct(fid)))
4452 * Quota files can't be created from the kernel any more,
4453 * 'tune2fs -O quota' will take care of creating them
4457 result = __osd_create(info, obj, attr, hint, dof, th);
4459 if (fid_is_idif(fid) &&
4460 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4461 struct lu_fid *tfid = &info->oti_fid;
4462 struct ost_id *oi = &info->oti_ostid;
4464 fid_to_ostid(fid, oi);
4465 ostid_to_fid(tfid, oi, 0);
4467 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4468 LMAC_FID_ON_OST, 0);
4470 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4472 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4473 on_ost ? LMAC_FID_ON_OST : 0,
4476 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4477 obj->oo_dt.do_body_ops = &osd_body_ops;
4480 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4481 struct inode *inode = obj->oo_inode;
4483 result = __osd_oi_insert(env, obj, fid, th);
4484 if (result && inode) {
4485 spin_lock(&obj->oo_guard);
4487 spin_unlock(&obj->oo_guard);
4488 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4489 ldiskfs_set_inode_state(inode,
4490 LDISKFS_STATE_LUSTRE_DESTROY);
4492 obj->oo_inode = NULL;
4497 * a small optimization - dt_insert() isn't usually applied
4498 * to OST objects, so we don't need to cache OI mapping for
4501 if (result == 0 && on_ost == 0) {
4502 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4504 result = osd_idc_find_and_init(env, osd, obj);
4505 LASSERT(result == 0);
4508 LASSERT(ergo(result == 0,
4509 dt_object_exists(dt) && !dt_object_remote(dt)));
4510 LINVRNT(osd_invariant(obj));
4514 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4515 struct thandle *handle)
4517 struct osd_thandle *oh;
4520 /* it's possible that object doesn't exist yet */
4521 LASSERT(handle != NULL);
4523 oh = container_of(handle, struct osd_thandle, ot_super);
4524 LASSERT(oh->ot_handle == NULL);
4526 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4527 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4529 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4536 * Concurrency: @dt is write locked.
4538 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4541 struct osd_object *obj = osd_dt_obj(dt);
4542 struct inode *inode = obj->oo_inode;
4543 struct osd_thandle *oh;
4546 if (!dt_object_exists(dt) || obj->oo_destroyed)
4549 LINVRNT(osd_invariant(obj));
4550 LASSERT(!dt_object_remote(dt));
4551 LASSERT(osd_is_write_locked(env, obj));
4552 LASSERT(th != NULL);
4554 oh = container_of(th, struct osd_thandle, ot_super);
4555 LASSERT(oh->ot_handle != NULL);
4557 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4559 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4560 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4562 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4563 * (65000) subdirectories by storing "1" in i_nlink if the link count
4564 * would otherwise overflow. Directory tranversal tools understand
4565 * that (st_nlink == 1) indicates that the filesystem dose not track
4566 * hard links count on the directory, and will not abort subdirectory
4567 * scanning early once (st_nlink - 2) subdirs have been found.
4569 * This also has to properly handle the case of inodes with nlink == 0
4570 * in case they are being linked into the PENDING directory
4572 spin_lock(&obj->oo_guard);
4573 if (unlikely(inode->i_nlink == 0))
4574 /* inc_nlink from 0 may cause WARN_ON */
4575 set_nlink(inode, 1);
4577 osd_ldiskfs_inc_count(oh->ot_handle, inode);
4578 if (!S_ISDIR(inode->i_mode))
4579 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4581 spin_unlock(&obj->oo_guard);
4583 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4584 LINVRNT(osd_invariant(obj));
4586 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4591 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4592 struct thandle *handle)
4594 struct osd_thandle *oh;
4596 if (!dt_object_exists(dt))
4599 LASSERT(!dt_object_remote(dt));
4600 LASSERT(handle != NULL);
4602 oh = container_of(handle, struct osd_thandle, ot_super);
4603 LASSERT(oh->ot_handle == NULL);
4605 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4606 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4612 * Concurrency: @dt is write locked.
4614 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4617 struct osd_object *obj = osd_dt_obj(dt);
4618 struct inode *inode = obj->oo_inode;
4619 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4620 struct osd_thandle *oh;
4622 if (!dt_object_exists(dt))
4625 LINVRNT(osd_invariant(obj));
4626 LASSERT(!dt_object_remote(dt));
4627 LASSERT(osd_is_write_locked(env, obj));
4628 LASSERT(th != NULL);
4630 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4633 oh = container_of(th, struct osd_thandle, ot_super);
4634 LASSERT(oh->ot_handle != NULL);
4636 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4638 spin_lock(&obj->oo_guard);
4640 * That can be result of upgrade from old Lustre version and
4641 * applied only to local files. Just skip this ref_del call.
4642 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4644 if (inode->i_nlink == 0) {
4645 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4646 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4647 ", maybe an upgraded file? (LU-3915)\n",
4648 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4649 spin_unlock(&obj->oo_guard);
4653 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4654 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4656 osd_ldiskfs_dec_count(oh->ot_handle, inode);
4657 spin_unlock(&obj->oo_guard);
4659 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4660 LINVRNT(osd_invariant(obj));
4662 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4668 * Concurrency: @dt is read locked.
4670 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4671 struct lu_buf *buf, const char *name)
4673 struct osd_object *obj = osd_dt_obj(dt);
4674 struct inode *inode = obj->oo_inode;
4675 struct osd_thread_info *info = osd_oti_get(env);
4676 struct dentry *dentry = &info->oti_obj_dentry;
4677 bool cache_xattr = false;
4682 /* version get is not real XATTR but uses xattr API */
4683 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4684 dt_obj_version_t *ver = buf->lb_buf;
4687 * for version we are just using xattr API but change inode
4690 if (buf->lb_len == 0)
4691 return sizeof(dt_obj_version_t);
4693 if (buf->lb_len < sizeof(dt_obj_version_t))
4696 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4697 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4699 *ver = LDISKFS_I(inode)->i_fs_version;
4701 return sizeof(dt_obj_version_t);
4704 if (!dt_object_exists(dt))
4707 LASSERT(!dt_object_remote(dt));
4708 LASSERT(inode->i_op != NULL);
4709 #ifdef HAVE_IOP_XATTR
4710 LASSERT(inode->i_op->getxattr != NULL);
4713 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4714 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4718 rc = osd_oxc_get(obj, name, buf);
4723 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4724 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4725 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4726 struct filter_fid *ff;
4727 struct ost_layout *ol;
4729 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4733 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4736 if (buf->lb_len == 0 || !buf->lb_buf)
4739 if (buf->lb_len < rc)
4743 ol = &ff->ff_layout;
4744 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4745 PFID_STRIPE_IDX_BITS);
4746 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4747 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4748 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4749 if (lma->lma_compat & LMAC_COMP_INFO) {
4750 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4751 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4752 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4754 ol->ol_comp_start = 0;
4755 ol->ol_comp_end = 0;
4759 /* Get enc context xattr directly from ldiskfs instead of going
4760 * through the VFS, as there is no xattr handler for
4763 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0)
4764 rc = ldiskfs_xattr_get(inode,
4765 LDISKFS_XATTR_INDEX_ENCRYPTION,
4766 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
4767 buf->lb_buf, buf->lb_len);
4769 rc = __osd_xattr_get(inode, dentry, name,
4770 buf->lb_buf, buf->lb_len);
4774 if (rc == -ENOENT || rc == -ENODATA)
4775 osd_oxc_add(obj, name, NULL, 0);
4776 else if (rc > 0 && buf->lb_buf != NULL)
4777 osd_oxc_add(obj, name, buf->lb_buf, rc);
4783 static int osd_declare_xattr_set(const struct lu_env *env,
4784 struct dt_object *dt,
4785 const struct lu_buf *buf, const char *name,
4786 int fl, struct thandle *handle)
4788 struct osd_thandle *oh;
4790 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4792 LASSERT(handle != NULL);
4794 oh = container_of(handle, struct osd_thandle, ot_super);
4795 LASSERT(oh->ot_handle == NULL);
4797 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4799 * For non-upgrading case, the LMA is set first and
4800 * usually fit inode. But for upgrade case, the LMA
4801 * may be in another separated EA block.
4803 if (dt_object_exists(dt)) {
4804 if (fl == LU_XATTR_REPLACE)
4809 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4811 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4812 /* We may need to delete the old PFID EA. */
4813 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4814 if (fl == LU_XATTR_REPLACE)
4820 * If some name entry resides on remote MDT, then will create
4821 * agent entry under remote parent. On the other hand, if the
4822 * remote entry will be removed, then related agent entry may
4823 * need to be removed from the remote parent. So there may be
4824 * kinds of cases, let's declare enough credits. The credits
4825 * for create agent entry is enough for remove case.
4827 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4828 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4829 if (dt_object_exists(dt))
4830 credits += 1; /* For updating LMA */
4834 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4839 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4841 * learn xattr size from osd_xattr_get if
4842 * attribute has not been read yet
4844 buflen = __osd_xattr_get(
4845 osd_dt_obj(dt)->oo_inode,
4846 &osd_oti_get(env)->oti_obj_dentry,
4851 buflen = buf->lb_len;
4854 if (buflen > sb->s_blocksize) {
4855 credits += osd_calc_bkmap_credits(
4857 (buflen + sb->s_blocksize - 1) >>
4858 sb->s_blocksize_bits);
4862 * xattr set may involve inode quota change, reserve credits for
4863 * dquot_initialize()
4865 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4868 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4873 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4874 const struct lu_buf *buf, int fl,
4875 struct thandle *handle)
4877 struct osd_thread_info *info = osd_oti_get(env);
4878 struct dentry *dentry = &info->oti_obj_dentry;
4879 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4880 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4881 struct inode *inode = obj->oo_inode;
4882 struct filter_fid *ff = buf->lb_buf;
4883 struct ost_layout *ol = &ff->ff_layout;
4884 int flags = XATTR_REPLACE;
4889 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4892 rc = osd_get_lma(info, inode, dentry, loa);
4893 if (rc == -ENODATA) {
4894 /* Usually for upgarding from old device */
4895 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4896 LMAC_FID_ON_OST, 0);
4897 flags = XATTR_CREATE;
4902 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4903 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4906 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4907 /* Separate PFID EA from LMA */
4908 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4909 lustre_lma_swab(lma);
4910 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4911 sizeof(*lma), XATTR_REPLACE);
4913 obj->oo_pfid_in_lma = 0;
4914 rc = LU_XATTR_CREATE;
4920 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4924 * Old client does not send stripe information,
4925 * then store the PFID EA on disk separatedly.
4927 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4928 ol->ol_stripe_size == 0))
4931 /* Remove old PFID EA entry firstly. */
4932 dquot_initialize(inode);
4933 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4934 if (rc == -ENODATA) {
4935 /* XATTR_NAME_FID is already absent */
4942 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4943 if (likely(ol->ol_stripe_size != 0)) {
4944 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4945 PFID_STRIPE_IDX_BITS;
4946 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4947 lma->lma_compat |= LMAC_STRIPE_INFO;
4948 if (ol->ol_comp_id != 0) {
4949 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4950 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4951 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4952 lma->lma_compat |= LMAC_COMP_INFO;
4956 lustre_loa_swab(loa, false);
4958 /* Store the PFID EA inside LMA. */
4959 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4962 obj->oo_pfid_in_lma = 1;
4968 * In DNE environment, the object (in spite of regular file or directory)
4969 * and its name entry may reside on different MDTs. Under such case, we will
4970 * create an agent entry on the MDT where the object resides. The agent entry
4971 * references the object locally, that makes the object to be visible to the
4972 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4973 * such as 'tar' can handle the object properly.
4975 * We handle the agent entry during set linkEA that is the common interface
4976 * for both regular file and directroy, can handle kinds of cases, such as
4977 * create/link/unlink/rename, and so on.
4979 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4981 * XXX: There are two known issues:
4982 * 1. For one object, we will create at most one agent entry even if there
4983 * may be more than one cross-MDTs hard links on the object. So the local
4984 * e2fsck may claim that the object's nlink is larger than the name entries
4985 * that reference such inode. And in further, the e2fsck will fix the nlink
4986 * attribute to match the local references. Then it will cause the object's
4987 * nlink attribute to be inconsistent with the global references. it is bad
4988 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4989 * other hand, the global namespace LFSCK can repair the object's attribute
4990 * according to the linkEA.
4991 * 2. There may be too many hard links on the object as to its linkEA overflow,
4992 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4993 * case happened, then at this point, we do not know whether there are some
4994 * cross-MDTs reference. But there are local references, it guarantees that
4995 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4997 static int osd_xattr_handle_linkea(const struct lu_env *env,
4998 struct osd_device *osd,
4999 struct osd_object *obj,
5000 const struct lu_buf *buf,
5001 struct thandle *handle)
5003 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
5004 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
5005 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
5006 struct lu_name tmpname;
5007 struct osd_thandle *oh;
5009 bool remote = false;
5013 oh = container_of(handle, struct osd_thandle, ot_super);
5014 LASSERT(oh->ot_handle != NULL);
5016 rc = linkea_init_with_rec(&ldata);
5018 linkea_first_entry(&ldata);
5019 while (ldata.ld_lee != NULL && !remote) {
5020 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
5022 if (osd_remote_fid(env, osd, tfid) > 0)
5025 linkea_next_entry(&ldata);
5027 } else if (rc == -ENODATA) {
5033 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
5034 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
5036 CERROR("%s: failed to remove agent entry for "DFID
5037 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5038 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
5039 rc = osd_add_to_remote_parent(env, osd, obj, oh);
5041 CERROR("%s: failed to create agent entry for "DFID
5042 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
5049 * Concurrency: @dt is write locked.
5051 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
5052 const struct lu_buf *buf, const char *name, int fl,
5053 struct thandle *handle)
5055 struct osd_object *obj = osd_dt_obj(dt);
5056 struct osd_device *osd = osd_obj2dev(obj);
5057 struct inode *inode = obj->oo_inode;
5058 struct osd_thread_info *info = osd_oti_get(env);
5068 /* version set is not real XATTR */
5069 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
5070 dt_obj_version_t *version = buf->lb_buf;
5073 * for version we are just using xattr API but change inode
5076 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
5079 DFID" set version %#llx (old %#llx) for inode %lu\n",
5080 PFID(lu_object_fid(&dt->do_lu)), *version,
5081 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
5083 LDISKFS_I(inode)->i_fs_version = *version;
5085 * Version is set after all inode operations are finished,
5086 * so we should mark it dirty here
5088 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
5093 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
5094 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
5097 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5100 * For the OST device with 256 bytes inode size by default,
5101 * the PFID EA will be stored together with LMA EA to avoid
5102 * performance trouble. Otherwise the PFID EA can be stored
5103 * independently. LU-8998
5105 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
5106 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
5107 LASSERT(buf->lb_buf);
5109 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
5112 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
5113 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
5114 struct lustre_mdt_attrs *lma = &loa->loa_lma;
5116 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
5120 lma->lma_incompat |= LMAI_STRIPED;
5121 lustre_lma_swab(lma);
5122 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5123 sizeof(*lma), XATTR_REPLACE);
5126 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
5127 LASSERT(!osd->od_is_ost);
5129 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
5134 if (fl & LU_XATTR_REPLACE)
5135 fs_flags |= XATTR_REPLACE;
5137 if (fl & LU_XATTR_CREATE)
5138 fs_flags |= XATTR_CREATE;
5140 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0) {
5141 /* Set enc context xattr directly in ldiskfs instead of going
5142 * through the VFS, as there is no xattr handler for
5145 struct osd_thandle *oth = container_of(handle,
5149 if (!oth->ot_handle)
5150 /* this should be already part of a transaction */
5153 rc = ldiskfs_xattr_set_handle(oth->ot_handle, inode,
5154 LDISKFS_XATTR_INDEX_ENCRYPTION,
5155 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
5156 buf->lb_buf, len, fs_flags);
5158 rc = __osd_xattr_set(info, inode, name,
5159 buf->lb_buf, len, fs_flags);
5161 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5164 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5165 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5166 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
5172 * Concurrency: @dt is read locked.
5174 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
5175 const struct lu_buf *buf)
5177 struct osd_object *obj = osd_dt_obj(dt);
5178 struct osd_device *dev = osd_obj2dev(obj);
5179 struct inode *inode = obj->oo_inode;
5180 struct osd_thread_info *info = osd_oti_get(env);
5181 struct dentry *dentry = &info->oti_obj_dentry;
5184 if (!dt_object_exists(dt))
5187 LASSERT(!dt_object_remote(dt));
5188 LASSERT(inode->i_op != NULL);
5189 LASSERT(inode->i_op->listxattr != NULL);
5191 dentry->d_inode = inode;
5192 dentry->d_sb = inode->i_sb;
5193 rc = inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
5195 if (rc < 0 || buf->lb_buf == NULL)
5198 /* Hide virtual project ID xattr from list if disabled */
5199 if (!dev->od_enable_projid_xattr) {
5200 char *end = (char *)buf->lb_buf + rc;
5201 char *p = buf->lb_buf;
5204 char *next = p + strlen(p) + 1;
5206 if (strcmp(p, XATTR_NAME_PROJID) == 0) {
5208 memmove(p, next, end - next);
5220 static int osd_declare_xattr_del(const struct lu_env *env,
5221 struct dt_object *dt, const char *name,
5222 struct thandle *handle)
5224 struct osd_thandle *oh;
5225 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
5227 LASSERT(!dt_object_remote(dt));
5228 LASSERT(handle != NULL);
5230 oh = container_of(handle, struct osd_thandle, ot_super);
5231 LASSERT(oh->ot_handle == NULL);
5233 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
5234 osd_dto_credits_noquota[DTO_XATTR_SET]);
5236 * xattr del may involve inode quota change, reserve credits for
5237 * dquot_initialize()
5239 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
5245 * Concurrency: @dt is write locked.
5247 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
5248 const char *name, struct thandle *handle)
5250 struct osd_object *obj = osd_dt_obj(dt);
5251 struct inode *inode = obj->oo_inode;
5252 struct osd_thread_info *info = osd_oti_get(env);
5253 struct dentry *dentry = &info->oti_obj_dentry;
5256 if (!dt_object_exists(dt))
5259 LASSERT(!dt_object_remote(dt));
5260 LASSERT(inode->i_op != NULL);
5261 LASSERT(handle != NULL);
5262 #ifdef HAVE_IOP_XATTR
5263 LASSERT(inode->i_op->removexattr != NULL);
5266 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5268 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
5269 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
5271 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
5272 &info->oti_ost_attrs);
5274 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
5276 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
5277 lustre_lma_swab(lma);
5278 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5279 sizeof(*lma), XATTR_REPLACE);
5281 obj->oo_pfid_in_lma = 0;
5284 dquot_initialize(inode);
5285 dentry->d_inode = inode;
5286 dentry->d_sb = inode->i_sb;
5287 rc = ll_vfs_removexattr(dentry, inode, name);
5290 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5293 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5294 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5295 osd_oxc_del(obj, name);
5300 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
5301 __u64 start, __u64 end)
5303 struct osd_object *obj = osd_dt_obj(dt);
5304 struct osd_device *dev = osd_obj2dev(obj);
5305 struct inode *inode = obj->oo_inode;
5310 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
5313 RETURN(PTR_ERR(file));
5315 file->f_mode |= FMODE_64BITHASH;
5316 rc = vfs_fsync_range(file, start, end, 0);
5323 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
5328 static bool osd_check_stale(struct dt_object *dt)
5336 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
5337 const struct dt_index_features *feat)
5339 struct iam_descr *descr;
5341 if (osd_object_is_root(o))
5342 return feat == &dt_directory_features;
5344 LASSERT(o->oo_dir != NULL);
5346 descr = o->oo_dir->od_container.ic_descr;
5347 if (feat == &dt_directory_features) {
5348 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
5353 return feat->dif_keysize_min <= descr->id_key_size &&
5354 descr->id_key_size <= feat->dif_keysize_max &&
5355 feat->dif_recsize_min <= descr->id_rec_size &&
5356 descr->id_rec_size <= feat->dif_recsize_max &&
5357 !(feat->dif_flags & (DT_IND_VARKEY |
5358 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5359 ergo(feat->dif_flags & DT_IND_UPDATE,
5360 1 /* XXX check that object (and fs) is writable */);
5364 static int osd_iam_container_init(const struct lu_env *env,
5365 struct osd_object *obj,
5366 struct osd_directory *dir)
5368 struct iam_container *bag = &dir->od_container;
5371 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5375 result = iam_container_setup(bag);
5377 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5379 iam_container_fini(bag);
5386 * Concurrency: no external locking is necessary.
5388 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5389 const struct dt_index_features *feat)
5393 struct osd_object *obj = osd_dt_obj(dt);
5395 LINVRNT(osd_invariant(obj));
5397 if (osd_object_is_root(obj)) {
5398 dt->do_index_ops = &osd_index_ea_ops;
5400 } else if (feat == &dt_directory_features) {
5401 dt->do_index_ops = &osd_index_ea_ops;
5402 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5407 } else if (unlikely(feat == &dt_otable_features)) {
5408 dt->do_index_ops = &osd_otable_ops;
5410 } else if (unlikely(feat == &dt_acct_features)) {
5411 dt->do_index_ops = &osd_acct_index_ops;
5414 } else if (!osd_has_index(obj)) {
5415 struct osd_directory *dir;
5416 struct osd_device *osd = osd_obj2dev(obj);
5417 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5422 spin_lock(&obj->oo_guard);
5423 if (obj->oo_dir == NULL)
5427 * Concurrent thread allocated container data.
5430 spin_unlock(&obj->oo_guard);
5432 * Now, that we have container data, serialize its
5435 down_write(&obj->oo_ext_idx_sem);
5437 * recheck under lock.
5440 if (osd_has_index(obj)) {
5445 result = osd_iam_container_init(env, obj, obj->oo_dir);
5446 if (result || feat == &dt_lfsck_namespace_features ||
5447 feat == &dt_lfsck_layout_orphan_features ||
5448 feat == &dt_lfsck_layout_dangling_features)
5451 result = osd_index_register(osd, fid,
5452 feat->dif_keysize_max,
5453 feat->dif_recsize_max);
5455 CWARN("%s: failed to register index "
5457 osd_name(osd), PFID(fid), result);
5458 else if (result > 0)
5461 CDEBUG(D_LFSCK, "%s: index object "DFID
5462 " (%d/%d) registered\n",
5463 osd_name(osd), PFID(fid),
5464 (int)feat->dif_keysize_max,
5465 (int)feat->dif_recsize_max);
5468 up_write(&obj->oo_ext_idx_sem);
5476 if (result == 0 && skip_iam == 0) {
5477 if (!osd_iam_index_probe(env, obj, feat))
5480 LINVRNT(osd_invariant(obj));
5485 static int osd_otable_it_attr_get(const struct lu_env *env,
5486 struct dt_object *dt,
5487 struct lu_attr *attr)
5493 static const struct dt_object_operations osd_obj_ops = {
5494 .do_read_lock = osd_read_lock,
5495 .do_write_lock = osd_write_lock,
5496 .do_read_unlock = osd_read_unlock,
5497 .do_write_unlock = osd_write_unlock,
5498 .do_write_locked = osd_write_locked,
5499 .do_attr_get = osd_attr_get,
5500 .do_declare_attr_set = osd_declare_attr_set,
5501 .do_attr_set = osd_attr_set,
5502 .do_ah_init = osd_ah_init,
5503 .do_declare_create = osd_declare_create,
5504 .do_create = osd_create,
5505 .do_declare_destroy = osd_declare_destroy,
5506 .do_destroy = osd_destroy,
5507 .do_index_try = osd_index_try,
5508 .do_declare_ref_add = osd_declare_ref_add,
5509 .do_ref_add = osd_ref_add,
5510 .do_declare_ref_del = osd_declare_ref_del,
5511 .do_ref_del = osd_ref_del,
5512 .do_xattr_get = osd_xattr_get,
5513 .do_declare_xattr_set = osd_declare_xattr_set,
5514 .do_xattr_set = osd_xattr_set,
5515 .do_declare_xattr_del = osd_declare_xattr_del,
5516 .do_xattr_del = osd_xattr_del,
5517 .do_xattr_list = osd_xattr_list,
5518 .do_object_sync = osd_object_sync,
5519 .do_invalidate = osd_invalidate,
5520 .do_check_stale = osd_check_stale,
5523 static const struct dt_object_operations osd_obj_otable_it_ops = {
5524 .do_attr_get = osd_otable_it_attr_get,
5525 .do_index_try = osd_index_try,
5528 static int osd_index_declare_iam_delete(const struct lu_env *env,
5529 struct dt_object *dt,
5530 const struct dt_key *key,
5531 struct thandle *handle)
5533 struct osd_thandle *oh;
5535 oh = container_of(handle, struct osd_thandle, ot_super);
5536 LASSERT(oh->ot_handle == NULL);
5538 /* Recycle may cause additional three blocks to be changed. */
5539 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5540 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5546 * delete a (key, value) pair from index \a dt specified by \a key
5548 * \param dt osd index object
5549 * \param key key for index
5550 * \param rec record reference
5551 * \param handle transaction handler
5554 * \retval -ve failure
5556 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5557 const struct dt_key *key,
5558 struct thandle *handle)
5560 struct osd_thread_info *oti = osd_oti_get(env);
5561 struct osd_object *obj = osd_dt_obj(dt);
5562 struct osd_thandle *oh;
5563 struct iam_path_descr *ipd;
5564 struct iam_container *bag = &obj->oo_dir->od_container;
5569 if (!dt_object_exists(dt))
5572 LINVRNT(osd_invariant(obj));
5573 LASSERT(!dt_object_remote(dt));
5574 LASSERT(bag->ic_object == obj->oo_inode);
5575 LASSERT(handle != NULL);
5577 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5579 ipd = osd_idx_ipd_get(env, bag);
5580 if (unlikely(ipd == NULL))
5583 oh = container_of(handle, struct osd_thandle, ot_super);
5584 LASSERT(oh->ot_handle != NULL);
5585 LASSERT(oh->ot_handle->h_transaction != NULL);
5587 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5588 /* swab quota uid/gid provided by caller */
5589 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5590 key = (const struct dt_key *)&oti->oti_quota_id;
5593 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5594 osd_ipd_put(env, bag, ipd);
5595 LINVRNT(osd_invariant(obj));
5596 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5600 static int osd_index_declare_ea_delete(const struct lu_env *env,
5601 struct dt_object *dt,
5602 const struct dt_key *key,
5603 struct thandle *handle)
5605 struct osd_thandle *oh;
5606 struct inode *inode;
5611 LASSERT(!dt_object_remote(dt));
5612 LASSERT(handle != NULL);
5614 oh = container_of(handle, struct osd_thandle, ot_super);
5615 LASSERT(oh->ot_handle == NULL);
5617 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5618 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5620 inode = osd_dt_obj(dt)->oo_inode;
5624 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5625 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5630 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5633 struct osd_fid_pack *rec;
5636 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5637 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5638 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5639 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5645 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5646 const struct lu_fid *fid)
5648 struct seq_server_site *ss = osd_seq_site(osd);
5652 /* FID seqs not in FLDB, must be local seq */
5653 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5657 * If FLD is not being initialized yet, it only happens during the
5658 * initialization, likely during mgs initialization, and we assume
5659 * this is local FID.
5661 if (ss == NULL || ss->ss_server_fld == NULL)
5664 /* Only check the local FLDB here */
5665 if (osd_seq_exists(env, osd, fid_seq(fid)))
5671 static void osd_take_care_of_agent(const struct lu_env *env,
5672 struct osd_device *osd,
5673 struct osd_thandle *oh,
5674 struct ldiskfs_dir_entry_2 *de)
5676 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5677 struct osd_idmap_cache *idc;
5678 int rc, schedule = 0;
5680 LASSERT(de != NULL);
5682 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5683 if (likely(rc == 0)) {
5684 idc = osd_idc_find_or_init(env, osd, fid);
5685 if (IS_ERR(idc) || idc->oic_remote)
5687 } else if (rc == -ENODATA) {
5689 * can't get FID, postpone to the end of the
5690 * transaction when iget() is safe
5694 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5697 osd_schedule_agent_inode_removal(env, oh,
5698 le32_to_cpu(de->inode));
5702 * Utility function to get real name from object name
5704 * \param[in] obj pointer to the object to be handled
5705 * \param[in] name object name
5706 * \param[in] len object name len
5707 * \param[out]ln pointer to the struct lu_name to hold the real name
5709 * If file is not encrypted, real name is just the object name.
5710 * If file is encrypted, object name needs to be decoded. In
5711 * this case a new buffer is allocated, and ln->ln_name needs to be freed by
5714 * \retval 0, on success
5715 * \retval -ve, on error
5717 static int obj_name2lu_name(struct osd_object *obj, const char *name,
5718 int len, struct lu_name *ln)
5720 if (!(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
5722 ln->ln_namelen = len;
5724 char *buf = kmalloc(len, GFP_NOFS);
5729 len = critical_decode(name, len, buf);
5731 ln->ln_namelen = len;
5738 * Index delete function for interoperability mode (b11826).
5739 * It will remove the directory entry added by osd_index_ea_insert().
5740 * This entry is needed to maintain name->fid mapping.
5742 * \param key, key i.e. file entry to be deleted
5744 * \retval 0, on success
5745 * \retval -ve, on error
5747 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5748 const struct dt_key *key, struct thandle *handle)
5750 struct osd_object *obj = osd_dt_obj(dt);
5751 struct inode *dir = obj->oo_inode;
5752 struct dentry *dentry;
5753 struct osd_thandle *oh;
5754 struct ldiskfs_dir_entry_2 *de = NULL;
5755 struct buffer_head *bh;
5756 struct htree_lock *hlock = NULL;
5757 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5763 if (!dt_object_exists(dt))
5766 LINVRNT(osd_invariant(obj));
5767 LASSERT(!dt_object_remote(dt));
5768 LASSERT(handle != NULL);
5770 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
5774 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5776 oh = container_of(handle, struct osd_thandle, ot_super);
5777 LASSERT(oh->ot_handle != NULL);
5778 LASSERT(oh->ot_handle->h_transaction != NULL);
5780 dquot_initialize(dir);
5781 dentry = osd_child_dentry_get(env, obj, ln.ln_name, ln.ln_namelen);
5783 if (obj->oo_hl_head != NULL) {
5784 hlock = osd_oti_get(env)->oti_hlock;
5785 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5786 dir, LDISKFS_HLOCK_DEL);
5788 down_write(&obj->oo_ext_idx_sem);
5791 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5794 * If this is not the ".." entry, it might be a remote DNE
5795 * entry and we need to check if the FID is for a remote
5796 * MDT. If the FID is not in the directory entry (e.g.
5797 * upgraded 1.8 filesystem without dirdata enabled) then
5798 * we need to get the FID from the LMA. For a remote directory
5799 * there HAS to be an LMA, it cannot be an IGIF inode in this
5802 * Delete the entry before the agent inode in order to
5803 * simplify error handling. At worst an error after deleting
5804 * the entry first might leak the agent inode afterward. The
5805 * reverse would need filesystem abort in case of error deleting
5806 * the entry after the agent had been removed, or leave a
5807 * dangling entry pointing at a random inode.
5809 if (strcmp((char *)key, dotdot) != 0)
5810 osd_take_care_of_agent(env, osd, oh, de);
5811 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5817 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
5818 atomic_dec_if_positive(&obj->oo_dirent_count);
5820 ldiskfs_htree_unlock(hlock);
5822 up_write(&obj->oo_ext_idx_sem);
5825 LASSERT(osd_invariant(obj));
5826 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5827 if (ln.ln_name != (char *)key)
5833 * Lookup index for \a key and copy record to \a rec.
5835 * \param dt osd index object
5836 * \param key key for index
5837 * \param rec record reference
5839 * \retval +ve success : exact mach
5840 * \retval 0 return record with key not greater than \a key
5841 * \retval -ve failure
5843 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5844 struct dt_rec *rec, const struct dt_key *key)
5846 struct osd_object *obj = osd_dt_obj(dt);
5847 struct iam_path_descr *ipd;
5848 struct iam_container *bag = &obj->oo_dir->od_container;
5849 struct osd_thread_info *oti = osd_oti_get(env);
5850 struct iam_iterator *it = &oti->oti_idx_it;
5851 struct iam_rec *iam_rec;
5856 if (!dt_object_exists(dt))
5859 LASSERT(osd_invariant(obj));
5860 LASSERT(!dt_object_remote(dt));
5861 LASSERT(bag->ic_object == obj->oo_inode);
5863 ipd = osd_idx_ipd_get(env, bag);
5867 /* got ipd now we can start iterator. */
5868 iam_it_init(it, bag, 0, ipd);
5870 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5871 /* swab quota uid/gid provided by caller */
5872 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5873 key = (const struct dt_key *)&oti->oti_quota_id;
5876 rc = iam_it_get(it, (struct iam_key *)key);
5878 if (S_ISDIR(obj->oo_inode->i_mode))
5879 iam_rec = (struct iam_rec *)oti->oti_ldp;
5881 iam_rec = (struct iam_rec *)rec;
5883 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5885 if (S_ISDIR(obj->oo_inode->i_mode))
5886 osd_fid_unpack((struct lu_fid *)rec,
5887 (struct osd_fid_pack *)iam_rec);
5888 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5889 osd_quota_unpack(obj, rec);
5894 osd_ipd_put(env, bag, ipd);
5896 LINVRNT(osd_invariant(obj));
5901 static int osd_index_declare_iam_insert(const struct lu_env *env,
5902 struct dt_object *dt,
5903 const struct dt_rec *rec,
5904 const struct dt_key *key,
5905 struct thandle *handle)
5907 struct osd_thandle *oh;
5909 LASSERT(handle != NULL);
5911 oh = container_of(handle, struct osd_thandle, ot_super);
5912 LASSERT(oh->ot_handle == NULL);
5914 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5915 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5921 * Inserts (key, value) pair in \a dt index object.
5923 * \param dt osd index object
5924 * \param key key for index
5925 * \param rec record reference
5926 * \param th transaction handler
5929 * \retval -ve failure
5931 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5932 const struct dt_rec *rec,
5933 const struct dt_key *key, struct thandle *th)
5935 struct osd_object *obj = osd_dt_obj(dt);
5936 struct iam_path_descr *ipd;
5937 struct osd_thandle *oh;
5938 struct iam_container *bag;
5939 struct osd_thread_info *oti = osd_oti_get(env);
5940 struct iam_rec *iam_rec;
5945 if (!dt_object_exists(dt))
5948 LINVRNT(osd_invariant(obj));
5949 LASSERT(!dt_object_remote(dt));
5951 bag = &obj->oo_dir->od_container;
5952 LASSERT(bag->ic_object == obj->oo_inode);
5953 LASSERT(th != NULL);
5955 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5957 ipd = osd_idx_ipd_get(env, bag);
5958 if (unlikely(ipd == NULL))
5961 oh = container_of(th, struct osd_thandle, ot_super);
5962 LASSERT(oh->ot_handle != NULL);
5963 LASSERT(oh->ot_handle->h_transaction != NULL);
5964 if (S_ISDIR(obj->oo_inode->i_mode)) {
5965 iam_rec = (struct iam_rec *)oti->oti_ldp;
5966 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5968 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5969 /* pack quota uid/gid */
5970 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5971 key = (const struct dt_key *)&oti->oti_quota_id;
5972 /* pack quota record */
5973 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5974 iam_rec = (struct iam_rec *)rec;
5976 iam_rec = (struct iam_rec *)rec;
5979 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5981 osd_ipd_put(env, bag, ipd);
5982 LINVRNT(osd_invariant(obj));
5983 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5988 * Calls ldiskfs_add_entry() to add directory entry
5989 * into the directory. This is required for
5990 * interoperability mode (b11826)
5992 * \retval 0, on success
5993 * \retval -ve, on error
5995 static int __osd_ea_add_rec(struct osd_thread_info *info,
5996 struct osd_object *pobj, struct inode *cinode,
5997 const char *name, const struct lu_fid *fid,
5998 struct htree_lock *hlock, struct thandle *th)
6000 struct ldiskfs_dentry_param *ldp;
6001 struct dentry *child;
6002 struct osd_thandle *oth;
6006 oth = container_of(th, struct osd_thandle, ot_super);
6007 LASSERT(oth->ot_handle != NULL);
6008 LASSERT(oth->ot_handle->h_transaction != NULL);
6009 LASSERT(pobj->oo_inode);
6011 rc = obj_name2lu_name(pobj, name, strlen(name), &ln);
6015 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
6016 if (unlikely(osd_object_is_root(pobj)))
6019 osd_get_ldiskfs_dirent_param(ldp, fid);
6020 child = osd_child_dentry_get(info->oti_env, pobj,
6021 ln.ln_name, ln.ln_namelen);
6022 child->d_fsdata = (void *)ldp;
6023 dquot_initialize(pobj->oo_inode);
6024 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
6025 child, cinode, hlock);
6026 if (rc == 0 && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
6027 struct ldiskfs_dir_entry_2 *de;
6028 struct buffer_head *bh;
6031 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
6034 rc1 = osd_ldiskfs_journal_get_write_access(
6035 oth->ot_handle, pobj->oo_inode->i_sb, bh,
6038 if (S_ISDIR(cinode->i_mode))
6039 de->file_type = LDISKFS_DIRENT_LUFID |
6040 LDISKFS_FT_REG_FILE;
6042 de->file_type = LDISKFS_DIRENT_LUFID |
6044 ldiskfs_handle_dirty_metadata(oth->ot_handle,
6051 if (ln.ln_name != name)
6057 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
6058 * into the directory.Also sets flags into osd object to
6059 * indicate dot and dotdot are created. This is required for
6060 * interoperability mode (b11826)
6062 * \param dir directory for dot and dotdot fixup.
6063 * \param obj child object for linking
6065 * \retval 0, on success
6066 * \retval -ve, on error
6068 static int osd_add_dot_dotdot(struct osd_thread_info *info,
6069 struct osd_object *dir,
6070 struct inode *parent_dir, const char *name,
6071 const struct lu_fid *dot_fid,
6072 const struct lu_fid *dot_dot_fid,
6075 struct inode *inode = dir->oo_inode;
6076 struct osd_thandle *oth;
6079 oth = container_of(th, struct osd_thandle, ot_super);
6080 LASSERT(oth->ot_handle->h_transaction != NULL);
6081 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
6083 if (strcmp(name, dot) == 0) {
6084 if (dir->oo_compat_dot_created) {
6087 LASSERT(inode->i_ino == parent_dir->i_ino);
6088 dir->oo_compat_dot_created = 1;
6091 } else if (strcmp(name, dotdot) == 0) {
6092 if (!dir->oo_compat_dot_created)
6094 /* in case of rename, dotdot is already created */
6095 if (dir->oo_compat_dotdot_created) {
6096 return __osd_ea_add_rec(info, dir, parent_dir, name,
6097 dot_dot_fid, NULL, th);
6100 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
6101 struct lu_fid tfid = *dot_dot_fid;
6104 result = osd_add_dot_dotdot_internal(info,
6105 dir->oo_inode, parent_dir, dot_fid,
6108 result = osd_add_dot_dotdot_internal(info,
6109 dir->oo_inode, parent_dir, dot_fid,
6114 dir->oo_compat_dotdot_created = 1;
6122 * It will call the appropriate osd_add* function and return the
6123 * value, return by respective functions.
6125 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
6126 struct inode *cinode, const char *name,
6127 const struct lu_fid *fid, struct thandle *th)
6129 struct osd_thread_info *info = osd_oti_get(env);
6130 struct htree_lock *hlock;
6133 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
6135 if (name[0] == '.' && (name[1] == '\0' ||
6136 (name[1] == '.' && name[2] == '\0'))) {
6137 if (hlock != NULL) {
6138 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6141 down_write(&pobj->oo_ext_idx_sem);
6144 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
6145 lu_object_fid(&pobj->oo_dt.do_lu),
6148 if (hlock != NULL) {
6149 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6150 pobj->oo_inode, LDISKFS_HLOCK_ADD);
6152 down_write(&pobj->oo_ext_idx_sem);
6155 if (CFS_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
6156 struct lu_fid *tfid = &info->oti_fid;
6160 rc = __osd_ea_add_rec(info, pobj, cinode, name,
6163 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
6167 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))){
6168 int dirent_count = atomic_read(&pobj->oo_dirent_count);
6170 /* avoid extremely unlikely 2B-entry directory overflow case */
6171 if (dirent_count != LU_DIRENT_COUNT_UNSET &&
6172 likely(dirent_count < INT_MAX - NR_CPUS))
6173 atomic_inc(&pobj->oo_dirent_count);
6177 ldiskfs_htree_unlock(hlock);
6179 up_write(&pobj->oo_ext_idx_sem);
6185 osd_ldiskfs_consistency_check(struct osd_thread_info *oti,
6186 struct osd_device *dev,
6187 const struct lu_fid *fid,
6188 struct osd_inode_id *id)
6190 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
6191 struct inode *inode = NULL;
6197 if (!scrub_needs_check(scrub, fid, id->oii_ino))
6200 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
6201 if (rc == -ENOENT) {
6202 __u32 gen = id->oii_gen;
6208 inode = osd_iget(oti, dev, id);
6209 /* The inode has been removed (by race maybe). */
6210 if (IS_ERR(inode)) {
6211 rc = PTR_ERR(inode);
6213 RETURN(rc == -ESTALE ? -ENOENT : rc);
6216 /* The OI mapping is lost. */
6217 if (gen != OSD_OII_NOGEN)
6221 * The inode may has been reused by others, we do not know,
6222 * leave it to be handled by subsequent osd_fid_lookup().
6225 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
6232 if (scrub->os_running) {
6233 if (inode == NULL) {
6234 inode = osd_iget(oti, dev, id);
6235 /* The inode has been removed (by race maybe). */
6236 if (IS_ERR(inode)) {
6237 rc = PTR_ERR(inode);
6239 RETURN(rc == -ESTALE ? -ENOENT : rc);
6243 rc = osd_scrub_oi_insert(dev, fid, id, insert);
6245 * There is race condition between osd_oi_lookup and OI scrub.
6246 * The OI scrub finished just after osd_oi_lookup() failure.
6247 * Under such case, it is unnecessary to trigger OI scrub again,
6248 * but try to call osd_oi_lookup() again.
6250 if (unlikely(rc == -EAGAIN))
6253 if (!S_ISDIR(inode->i_mode))
6256 rc = osd_check_lmv(oti, dev, inode);
6261 if (dev->od_scrub.os_scrub.os_auto_scrub_interval != AS_NEVER &&
6263 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
6264 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
6265 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
6266 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
6267 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
6268 if (rc == 0 || rc == -EALREADY)
6280 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
6281 struct osd_device *dev,
6282 struct lu_fid *fid, __u32 ino)
6284 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
6285 struct osd_idmap_cache *oic = &oti->oti_cache;
6286 struct inode *inode;
6289 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
6290 inode = osd_iget(oti, dev, &oic->oic_lid);
6291 if (IS_ERR(inode)) {
6292 fid_zero(&oic->oic_fid);
6293 return PTR_ERR(inode);
6296 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
6299 fid_zero(&oic->oic_fid);
6301 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
6305 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
6306 struct osd_inode_id *id, const struct lu_fid *fid)
6308 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
6309 id->oii_ino, id->oii_gen, info);
6310 info->oti_cache.oic_lid = *id;
6311 info->oti_cache.oic_fid = *fid;
6312 info->oti_cache.oic_dev = osd;
6316 * Get parent FID from the linkEA.
6318 * For a directory which parent resides on remote MDT, to satisfy the
6319 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
6320 * the other hand, to make the lookup(..) on the directory can return
6321 * the real parent FID, we append the real parent FID after its ".."
6322 * name entry in the /REMOTE_PARENT_DIR.
6324 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
6325 * backup. So after the restore, we cannot get the right parent FID from
6326 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
6327 * we have stored the real parent FID in the directory object's linkEA,
6328 * we can parse the linkEA for the real parent FID.
6330 * \param[in] env pointer to the thread context
6331 * \param[in] obj pointer to the object to be handled
6332 * \param[out]fid pointer to the buffer to hold the parent FID
6334 * \retval 0 for getting the real parent FID successfully
6335 * \retval negative error number on failure
6337 static int osd_get_pfid_from_linkea(const struct lu_env *env,
6338 struct osd_object *obj,
6341 struct osd_thread_info *oti = osd_oti_get(env);
6342 struct lu_buf *buf = &oti->oti_big_buf;
6343 struct dentry *dentry = &oti->oti_obj_dentry;
6344 struct inode *inode = obj->oo_inode;
6345 struct linkea_data ldata = { NULL };
6351 if (!S_ISDIR(inode->i_mode))
6355 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6356 buf->lb_buf, buf->lb_len);
6357 if (rc == -ERANGE) {
6358 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6361 lu_buf_realloc(buf, rc);
6362 if (buf->lb_buf == NULL)
6369 if (unlikely(rc == 0))
6375 if (unlikely(buf->lb_buf == NULL)) {
6376 lu_buf_realloc(buf, rc);
6377 if (buf->lb_buf == NULL)
6384 rc = linkea_init_with_rec(&ldata);
6386 linkea_first_entry(&ldata);
6387 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
6393 static int osd_verify_ent_by_linkea(const struct lu_env *env,
6394 struct inode *inode,
6395 const struct lu_fid *pfid,
6396 const char *name, const int namelen)
6398 struct osd_thread_info *oti = osd_oti_get(env);
6399 struct lu_buf *buf = &oti->oti_big_buf;
6400 struct dentry *dentry = &oti->oti_obj_dentry;
6401 struct linkea_data ldata = { NULL };
6402 struct lu_name cname = { .ln_name = name,
6403 .ln_namelen = namelen };
6409 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6410 buf->lb_buf, buf->lb_len);
6412 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6417 if (unlikely(rc == 0))
6420 if (buf->lb_len < rc) {
6421 lu_buf_realloc(buf, rc);
6422 if (buf->lb_buf == NULL)
6429 rc = linkea_init_with_rec(&ldata);
6431 rc = linkea_links_find(&ldata, &cname, pfid);
6437 * Calls ->lookup() to find dentry. From dentry get inode and
6438 * read inode's ea to get fid. This is required for interoperability
6441 * \retval 0, on success
6442 * \retval -ve, on error
6444 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6445 struct dt_rec *rec, const struct lu_name *ln)
6447 struct inode *dir = obj->oo_inode;
6448 struct dentry *dentry;
6449 struct ldiskfs_dir_entry_2 *de;
6450 struct buffer_head *bh;
6451 struct lu_fid *fid = (struct lu_fid *)rec;
6452 struct htree_lock *hlock = NULL;
6458 LASSERT(dir->i_op != NULL);
6459 LASSERT(dir->i_op->lookup != NULL);
6461 dentry = osd_child_dentry_get(env, obj, ln->ln_name, ln->ln_namelen);
6463 if (obj->oo_hl_head != NULL) {
6464 hlock = osd_oti_get(env)->oti_hlock;
6465 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6466 dir, LDISKFS_HLOCK_LOOKUP);
6468 down_read(&obj->oo_ext_idx_sem);
6471 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6473 struct osd_thread_info *oti = osd_oti_get(env);
6474 struct osd_inode_id *id = &oti->oti_id;
6475 struct osd_device *dev = osd_obj2dev(obj);
6477 ino = le32_to_cpu(de->inode);
6478 if (CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6480 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6484 rc = osd_get_fid_from_dentry(de, rec);
6486 /* done with de, release bh */
6489 if (unlikely(is_remote_parent_ino(dev, ino))) {
6491 * If the parent is on remote MDT, and there
6492 * is no FID-in-dirent, then we have to get
6493 * the parent FID from the linkEA.
6495 if (likely(ln->ln_namelen == 2 &&
6496 ln->ln_name[0] == '.' && ln->ln_name[1] == '.'))
6497 rc = osd_get_pfid_from_linkea(env, obj,
6500 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6503 osd_id_gen(id, ino, OSD_OII_NOGEN);
6506 if (rc != 0 || osd_remote_fid(env, dev, fid))
6509 rc = osd_ldiskfs_consistency_check(oti, dev, fid, id);
6510 if (rc != -ENOENT) {
6511 /* Other error should not affect lookup result. */
6514 /* Normal file mapping should be added into OI cache
6515 * after FID in LMA check, but for local files like
6516 * hsm_actions, their FIDs are not stored in OI files,
6517 * see osd_initial_OI_scrub(), and here is the only
6518 * place to load mapping into OI cache.
6520 if (!fid_is_namespace_visible(fid))
6521 osd_add_oi_cache(osd_oti_get(env),
6522 osd_obj2dev(obj), id, fid);
6524 CDEBUG(D_INODE, DFID"/"DNAME" => "DFID"\n",
6525 PFID(lu_object_fid(&obj->oo_dt.do_lu)), PNAME(ln),
6535 ldiskfs_htree_unlock(hlock);
6537 up_read(&obj->oo_ext_idx_sem);
6541 static int osd_index_declare_ea_insert(const struct lu_env *env,
6542 struct dt_object *dt,
6543 const struct dt_rec *rec,
6544 const struct dt_key *key,
6545 struct thandle *handle)
6547 struct osd_thandle *oh;
6548 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6549 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6550 const struct lu_fid *fid = rec1->rec_fid;
6551 int credits, rc = 0;
6552 struct osd_idmap_cache *idc;
6556 LASSERT(!dt_object_remote(dt));
6557 LASSERT(handle != NULL);
6558 LASSERT(fid != NULL);
6559 LASSERT(rec1->rec_type != 0);
6561 oh = container_of(handle, struct osd_thandle, ot_super);
6562 LASSERT(oh->ot_handle == NULL);
6564 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6567 * we can't call iget() while a transactions is running
6568 * (this can lead to a deadlock), but we need to know
6569 * inum and object type. so we find this information at
6570 * declaration and cache in per-thread info
6572 idc = osd_idc_find_or_init(env, osd, fid);
6574 RETURN(PTR_ERR(idc));
6575 if (idc->oic_remote) {
6577 * a reference to remote inode is represented by an
6578 * agent inode which we have to create
6580 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6581 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6584 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6586 if (osd_dt_obj(dt)->oo_inode != NULL) {
6587 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6590 * We ignore block quota on meta pool (MDTs), so needn't
6591 * calculate how many blocks will be consumed by this index
6594 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6596 i_projid_read(inode), 0,
6597 oh, osd_dt_obj(dt), NULL,
6602 #ifdef HAVE_PROJECT_QUOTA
6604 * Reserve credits for local agent inode to transfer
6605 * to 0, quota enforcement is ignored in this case.
6607 if (idc->oic_remote &&
6608 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6609 i_projid_read(inode) != 0)
6610 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6611 0, i_projid_read(inode),
6612 0, false, PRJQUOTA);
6620 * Index add function for interoperability mode (b11826).
6621 * It will add the directory entry.This entry is needed to
6622 * maintain name->fid mapping.
6624 * \param key it is key i.e. file entry to be inserted
6625 * \param rec it is value of given key i.e. fid
6627 * \retval 0, on success
6628 * \retval -ve, on error
6630 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6631 const struct dt_rec *rec,
6632 const struct dt_key *key, struct thandle *th)
6634 struct osd_object *obj = osd_dt_obj(dt);
6635 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6636 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6637 const struct lu_fid *fid = rec1->rec_fid;
6638 const char *name = (const char *)key;
6639 struct osd_thread_info *oti = osd_oti_get(env);
6640 struct inode *child_inode = NULL;
6641 struct osd_idmap_cache *idc;
6646 if (!dt_object_exists(dt))
6649 LASSERT(osd_invariant(obj));
6650 LASSERT(!dt_object_remote(dt));
6651 LASSERT(th != NULL);
6653 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6655 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6657 idc = osd_idc_find(env, osd, fid);
6658 if (unlikely(idc == NULL)) {
6659 idc = osd_idc_find_or_init(env, osd, fid);
6662 * this dt_insert() wasn't declared properly, so
6663 * FID is missing in OI cache. we better do not
6664 * lookup FID in FLDB/OI and don't risk to deadlock,
6665 * but in some special cases (lfsck testing, etc)
6666 * it's much simpler than fixing a caller.
6668 * normally this error should be placed after the first
6669 * find, but migrate may attach source stripes to
6670 * target, which doesn't create stripes.
6672 CERROR("%s: "DFID" wasn't declared for insert\n",
6673 osd_name(osd), PFID(fid));
6675 RETURN(PTR_ERR(idc));
6679 if (idc->oic_remote) {
6680 /* Insert remote entry */
6681 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6683 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6685 child_inode = osd_create_local_agent_inode(env, osd,
6686 obj, fid, rec1->rec_type & S_IFMT, th);
6687 if (IS_ERR(child_inode))
6688 RETURN(PTR_ERR(child_inode));
6691 /* Insert local entry */
6692 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6693 /* for a reason OI cache wasn't filled properly */
6694 CERROR("%s: OIC for "DFID" isn't filled\n",
6695 osd_name(osd), PFID(fid));
6698 child_inode = oti->oti_inode;
6699 if (unlikely(child_inode == NULL)) {
6700 struct ldiskfs_inode_info *lii;
6705 child_inode = oti->oti_inode = &lii->vfs_inode;
6707 child_inode->i_sb = osd_sb(osd);
6708 child_inode->i_ino = idc->oic_lid.oii_ino;
6709 child_inode->i_mode = rec1->rec_type & S_IFMT;
6712 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6714 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6715 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6717 if (child_inode && child_inode != oti->oti_inode)
6719 LASSERT(osd_invariant(obj));
6720 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6726 * Initialize osd Iterator for given osd index object.
6728 * \param dt osd index object
6731 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6732 struct dt_object *dt,
6735 struct osd_it_iam *it;
6736 struct osd_object *obj = osd_dt_obj(dt);
6737 struct lu_object *lo = &dt->do_lu;
6738 struct iam_path_descr *ipd;
6739 struct iam_container *bag = &obj->oo_dir->od_container;
6741 if (!dt_object_exists(dt))
6742 return ERR_PTR(-ENOENT);
6746 return ERR_PTR(-ENOMEM);
6748 ipd = osd_it_ipd_get(env, bag);
6749 if (likely(ipd != NULL)) {
6753 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6754 return (struct dt_it *)it;
6757 return ERR_PTR(-ENOMEM);
6762 * free given Iterator.
6764 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6766 struct osd_it_iam *it = (struct osd_it_iam *)di;
6767 struct osd_object *obj = it->oi_obj;
6769 iam_it_fini(&it->oi_it);
6770 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6771 osd_object_put(env, obj);
6776 * Move Iterator to record specified by \a key
6778 * \param di osd iterator
6779 * \param key key for index
6781 * \retval +ve di points to record with least key not larger than key
6782 * \retval 0 di points to exact matched key
6783 * \retval -ve failure
6786 static int osd_it_iam_get(const struct lu_env *env,
6787 struct dt_it *di, const struct dt_key *key)
6789 struct osd_thread_info *oti = osd_oti_get(env);
6790 struct osd_it_iam *it = (struct osd_it_iam *)di;
6792 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6793 /* swab quota uid/gid */
6794 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6795 key = (struct dt_key *)&oti->oti_quota_id;
6798 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6804 * \param di osd iterator
6806 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6808 struct osd_it_iam *it = (struct osd_it_iam *)di;
6810 iam_it_put(&it->oi_it);
6814 * Move iterator by one record
6816 * \param di osd iterator
6818 * \retval +1 end of container reached
6820 * \retval -ve failure
6823 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6825 struct osd_it_iam *it = (struct osd_it_iam *)di;
6827 return iam_it_next(&it->oi_it);
6831 * Return pointer to the key under iterator.
6834 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6835 const struct dt_it *di)
6837 struct osd_thread_info *oti = osd_oti_get(env);
6838 struct osd_it_iam *it = (struct osd_it_iam *)di;
6839 struct osd_object *obj = it->oi_obj;
6842 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6844 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6845 /* swab quota uid/gid */
6846 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6847 key = (struct dt_key *)&oti->oti_quota_id;
6854 * Return size of key under iterator (in bytes)
6857 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6859 struct osd_it_iam *it = (struct osd_it_iam *)di;
6861 return iam_it_key_size(&it->oi_it);
6865 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6867 /* check if file type is required */
6868 if (ent->lde_attrs & LUDA_TYPE) {
6869 struct luda_type *lt;
6870 int align = sizeof(*lt) - 1;
6872 len = (len + align) & ~align;
6873 lt = (struct luda_type *)(ent->lde_name + len);
6874 lt->lt_type = cpu_to_le16(DTTOIF(type));
6877 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6881 * build lu direct from backend fs dirent.
6885 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6886 char *name, __u16 namelen, __u16 type, __u32 attr)
6888 ent->lde_attrs = attr | LUDA_FID;
6889 fid_cpu_to_le(&ent->lde_fid, fid);
6891 ent->lde_hash = cpu_to_le64(offset);
6892 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6894 strncpy(ent->lde_name, name, namelen);
6895 ent->lde_name[namelen] = '\0';
6896 ent->lde_namelen = cpu_to_le16(namelen);
6898 /* append lustre attributes */
6899 osd_it_append_attrs(ent, namelen, type);
6903 * Return pointer to the record under iterator.
6905 static int osd_it_iam_rec(const struct lu_env *env,
6906 const struct dt_it *di,
6907 struct dt_rec *dtrec, __u32 attr)
6909 struct osd_it_iam *it = (struct osd_it_iam *)di;
6910 struct osd_thread_info *info = osd_oti_get(env);
6914 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6915 const struct osd_fid_pack *rec;
6916 struct lu_fid *fid = &info->oti_fid;
6917 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6923 name = (char *)iam_it_key_get(&it->oi_it);
6925 RETURN(PTR_ERR(name));
6927 namelen = iam_it_key_size(&it->oi_it);
6929 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6931 RETURN(PTR_ERR(rec));
6933 rc = osd_fid_unpack(fid, rec);
6937 hash = iam_it_store(&it->oi_it);
6939 /* IAM does not store object type in IAM index (dir) */
6940 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6942 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6943 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6944 (struct iam_rec *)dtrec);
6945 osd_quota_unpack(it->oi_obj, dtrec);
6947 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6948 (struct iam_rec *)dtrec);
6955 * Returns cookie for current Iterator position.
6957 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6959 struct osd_it_iam *it = (struct osd_it_iam *)di;
6961 return iam_it_store(&it->oi_it);
6965 * Restore iterator from cookie.
6967 * \param di osd iterator
6968 * \param hash Iterator location cookie
6970 * \retval +ve di points to record with least key not larger than key.
6971 * \retval 0 di points to exact matched key
6972 * \retval -ve failure
6975 static int osd_it_iam_load(const struct lu_env *env,
6976 const struct dt_it *di, __u64 hash)
6978 struct osd_it_iam *it = (struct osd_it_iam *)di;
6980 return iam_it_load(&it->oi_it, hash);
6983 static const struct dt_index_operations osd_index_iam_ops = {
6984 .dio_lookup = osd_index_iam_lookup,
6985 .dio_declare_insert = osd_index_declare_iam_insert,
6986 .dio_insert = osd_index_iam_insert,
6987 .dio_declare_delete = osd_index_declare_iam_delete,
6988 .dio_delete = osd_index_iam_delete,
6990 .init = osd_it_iam_init,
6991 .fini = osd_it_iam_fini,
6992 .get = osd_it_iam_get,
6993 .put = osd_it_iam_put,
6994 .next = osd_it_iam_next,
6995 .key = osd_it_iam_key,
6996 .key_size = osd_it_iam_key_size,
6997 .rec = osd_it_iam_rec,
6998 .store = osd_it_iam_store,
6999 .load = osd_it_iam_load
7003 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
7004 struct osd_device *dev,
7005 struct inode *inode, u32 attr)
7007 struct osd_thread_info *info = osd_oti_get(env);
7008 struct osd_it_ea *oie;
7012 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
7015 RETURN(ERR_CAST(file));
7017 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
7018 if (attr & LUDA_64BITHASH)
7019 file->f_mode |= FMODE_64BITHASH;
7021 file->f_mode |= FMODE_32BITHASH;
7024 OBD_SLAB_ALLOC_PTR(oie, osd_itea_cachep);
7028 oie->oie_rd_dirent = 0;
7029 oie->oie_it_dirent = 0;
7030 oie->oie_dirent = NULL;
7031 if (unlikely(!info->oti_it_ea_buf_used)) {
7032 oie->oie_buf = info->oti_it_ea_buf;
7033 info->oti_it_ea_buf_used = 1;
7035 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7039 oie->oie_obj = NULL;
7040 oie->oie_file = file;
7045 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7049 return ERR_PTR(-ENOMEM);
7053 * Creates or initializes iterator context.
7055 * \retval struct osd_it_ea, iterator structure on success
7058 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
7059 struct dt_object *dt,
7062 struct osd_object *obj = osd_dt_obj(dt);
7063 struct osd_device *dev = osd_obj2dev(obj);
7064 struct lu_object *lo = &dt->do_lu;
7065 struct osd_it_ea *oie;
7069 if (!dt_object_exists(dt) || obj->oo_destroyed)
7070 RETURN(ERR_PTR(-ENOENT));
7072 oie = osd_it_dir_init(env, dev, obj->oo_inode, attr);
7074 RETURN(ERR_CAST(oie));
7078 RETURN((struct dt_it *)oie);
7081 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
7082 struct inode *inode)
7084 struct osd_thread_info *info = osd_oti_get(env);
7087 fput(oie->oie_file);
7088 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
7089 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
7091 info->oti_it_ea_buf_used = 0;
7092 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
7097 * Destroy or finishes iterator context.
7099 * \param di iterator structure to be destroyed
7101 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
7103 struct osd_it_ea *oie = (struct osd_it_ea *)di;
7104 struct osd_object *obj = oie->oie_obj;
7105 struct inode *inode = obj->oo_inode;
7108 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
7109 osd_object_put(env, obj);
7114 * It position the iterator at given key, so that next lookup continues from
7115 * that key Or it is similar to dio_it->load() but based on a key,
7116 * rather than file position.
7118 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
7121 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
7123 static int osd_it_ea_get(const struct lu_env *env,
7124 struct dt_it *di, const struct dt_key *key)
7126 struct osd_it_ea *it = (struct osd_it_ea *)di;
7129 LASSERT(((const char *)key)[0] == '\0');
7130 it->oie_file->f_pos = 0;
7131 it->oie_rd_dirent = 0;
7132 it->oie_it_dirent = 0;
7133 it->oie_dirent = NULL;
7141 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
7145 struct osd_filldir_cbs {
7146 struct dir_context ctx;
7147 struct osd_it_ea *it;
7150 * It is called internally by ->iterate*(). It fills the
7151 * iterator's in-memory data structure with required
7152 * information i.e. name, namelen, rec_size etc.
7154 * \param buf in which information to be filled in.
7155 * \param name name of the file in given dir
7157 * \retval 0 on success
7158 * \retval 1 on buffer full
7160 #ifdef HAVE_FILLDIR_USE_CTX
7161 static FILLDIR_TYPE do_osd_ldiskfs_filldir(struct dir_context *ctx,
7163 static int osd_ldiskfs_filldir(void *ctx,
7165 const char *name, int namelen,
7166 loff_t offset, __u64 ino, unsigned int d_type)
7168 struct osd_it_ea *it = ((struct osd_filldir_cbs *)ctx)->it;
7169 struct osd_object *obj = it->oie_obj;
7170 struct osd_it_ea_dirent *ent = it->oie_dirent;
7171 struct lu_fid *fid = &ent->oied_fid;
7172 char *buf = it->oie_buf;
7173 struct osd_fid_pack *rec;
7176 /* this should never happen */
7177 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
7178 CERROR("ldiskfs return invalid namelen %d\n", namelen);
7182 /* Check for enough space. Note oied_name is not NUL terminated. */
7183 if (&ent->oied_name[namelen] > buf + OSD_IT_EA_BUFSIZE)
7186 /* "." is just the object itself. */
7187 if (namelen == 1 && name[0] == '.') {
7189 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7190 } else if (d_type & LDISKFS_DIRENT_LUFID) {
7191 rec = (struct osd_fid_pack *)(name + namelen + 1);
7192 if (osd_fid_unpack(fid, rec) != 0)
7197 d_type &= ~LDISKFS_DIRENT_LUFID;
7199 /* NOT export local root. */
7201 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
7202 ino = obj->oo_inode->i_ino;
7203 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7206 if (obj == NULL || !(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
7207 ent->oied_namelen = namelen;
7208 memcpy(ent->oied_name, name, namelen);
7210 int encoded_namelen = critical_chars(name, namelen);
7212 /* Check again for enough space. */
7213 if (&ent->oied_name[encoded_namelen] > buf + OSD_IT_EA_BUFSIZE)
7216 ent->oied_namelen = encoded_namelen;
7218 if (encoded_namelen == namelen)
7219 memcpy(ent->oied_name, name, namelen);
7221 critical_encode(name, namelen, ent->oied_name);
7224 ent->oied_ino = ino;
7225 ent->oied_off = offset;
7226 ent->oied_type = d_type;
7228 it->oie_rd_dirent++;
7229 it->oie_dirent = (void *)ent + round_up(sizeof(*ent) + ent->oied_namelen, 8);
7233 WRAP_FILLDIR_FN(do_, osd_ldiskfs_filldir)
7236 * Calls ->iterate*() to load a directory entry at a time
7237 * and stored it in iterator's in-memory data structure.
7239 * \param di iterator's in memory structure
7241 * \retval 0 on success
7242 * \retval -ve on error
7243 * \retval +1 reach the end of entry
7245 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
7247 struct osd_it_ea *it = (struct osd_it_ea *)di;
7248 struct osd_object *obj = it->oie_obj;
7249 struct htree_lock *hlock = NULL;
7250 struct file *filp = it->oie_file;
7252 struct osd_filldir_cbs buf = {
7253 .ctx.actor = osd_ldiskfs_filldir,
7258 it->oie_dirent = it->oie_buf;
7259 it->oie_rd_dirent = 0;
7262 if (obj->oo_hl_head != NULL) {
7263 hlock = osd_oti_get(env)->oti_hlock;
7264 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
7266 LDISKFS_HLOCK_READDIR);
7268 down_read(&obj->oo_ext_idx_sem);
7272 rc = iterate_dir(filp, &buf.ctx);
7276 if (it->oie_rd_dirent == 0) {
7278 * If it does not get any dirent, it means it has been reached
7279 * to the end of the dir
7281 it->oie_file->f_pos = ldiskfs_get_htree_eof(it->oie_file);
7285 it->oie_dirent = it->oie_buf;
7286 it->oie_it_dirent = 1;
7291 ldiskfs_htree_unlock(hlock);
7293 up_read(&obj->oo_ext_idx_sem);
7300 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7301 * to load a directory entry at a time and stored it in
7302 * iterator's in-memory data structure.
7304 * \param di iterator's in memory structure
7306 * \retval +ve iterator reached to end
7307 * \retval 0 iterator not reached to end
7308 * \retval -ve on error
7310 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
7312 struct osd_it_ea *it = (struct osd_it_ea *)di;
7317 if (it->oie_it_dirent < it->oie_rd_dirent) {
7319 (void *)it->oie_dirent +
7320 round_up(sizeof(struct osd_it_ea_dirent) +
7321 it->oie_dirent->oied_namelen, 8);
7322 it->oie_it_dirent++;
7325 if (it->oie_file->f_pos == ldiskfs_get_htree_eof(it->oie_file))
7328 rc = osd_ldiskfs_it_fill(env, di);
7335 * Returns the key at current position from iterator's in memory structure.
7337 * \param di iterator's in memory structure
7339 * \retval key i.e. struct dt_key on success
7341 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
7342 const struct dt_it *di)
7344 struct osd_it_ea *it = (struct osd_it_ea *)di;
7346 return (struct dt_key *)it->oie_dirent->oied_name;
7350 * Returns key's size at current position from iterator's in memory structure.
7352 * \param di iterator's in memory structure
7354 * \retval key_size i.e. struct dt_key on success
7356 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
7358 struct osd_it_ea *it = (struct osd_it_ea *)di;
7360 return it->oie_dirent->oied_namelen;
7363 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
7364 #undef LDISKFS_DIR_REC_LEN
7365 # if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7366 # define LDISKFS_DIR_REC_LEN(de, dir) LDISKFS_DIR_ENTRY_LEN_((de), (dir))
7368 # define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
7372 #if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7373 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de), NULL)
7375 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de))
7378 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
7380 if (LDISKFS_DIR_REC_LEN_DIR(de) >=
7381 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
7388 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
7389 unsigned int blocksize, bool dotdot)
7392 return osd_dotdot_has_space(de);
7394 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
7395 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
7402 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
7403 handle_t *jh, struct dentry *dentry,
7404 const struct lu_fid *fid, struct buffer_head *bh,
7405 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
7408 struct inode *dir = dentry->d_parent->d_inode;
7409 struct inode *inode = dentry->d_inode;
7410 struct osd_fid_pack *rec;
7411 struct ldiskfs_dentry_param *ldp;
7412 int namelen = dentry->d_name.len;
7414 struct osd_thread_info *info = osd_oti_get(env);
7418 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7421 /* There is enough space to hold the FID-in-dirent. */
7422 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7423 rc = osd_ldiskfs_journal_get_write_access(jh, dir->i_sb, bh,
7428 de->name[namelen] = 0;
7429 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7430 rec->fp_len = sizeof(struct lu_fid) + 1;
7431 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7432 de->file_type |= LDISKFS_DIRENT_LUFID;
7433 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7440 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7444 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7445 osd_get_ldiskfs_dirent_param(ldp, fid);
7446 dentry->d_fsdata = (void *)ldp;
7447 dquot_initialize(dir);
7448 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7450 * It is too bad, we cannot reinsert the name entry back.
7451 * That means we lose it!
7455 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7456 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7457 namelen, dentry->d_name.name, PFID(fid), rc);
7463 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7464 struct osd_it_ea *it, struct lu_fid *fid,
7465 struct osd_inode_id *id, __u32 *attr)
7467 struct osd_thread_info *info = osd_oti_get(env);
7468 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7469 struct osd_device *dev = osd_obj2dev(obj);
7470 struct super_block *sb = osd_sb(dev);
7471 const char *devname = osd_name(dev);
7472 struct osd_it_ea_dirent *ent = it->oie_dirent;
7473 struct inode *dir = obj->oo_inode;
7474 struct htree_lock *hlock = NULL;
7475 struct buffer_head *bh = NULL;
7476 handle_t *jh = NULL;
7477 struct ldiskfs_dir_entry_2 *de;
7478 struct dentry *dentry;
7479 struct inode *inode;
7480 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7483 bool dotdot = false;
7489 if (ent->oied_name[0] == '.') {
7490 if (ent->oied_namelen == 1)
7493 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7497 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7498 inode = osd_iget(info, dev, id);
7499 if (IS_ERR(inode)) {
7500 rc = PTR_ERR(inode);
7501 if (rc == -ENOENT || rc == -ESTALE) {
7503 * Maybe dangling name entry, or
7504 * corrupted directory entry.
7506 *attr |= LUDA_UNKNOWN;
7509 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7510 "check_repair, dir = %lu/%u, name = %.*s, "
7511 "ino = %llu, rc = %d\n",
7512 devname, dir->i_ino, dir->i_generation,
7513 ent->oied_namelen, ent->oied_name,
7520 rc = obj_name2lu_name(obj, ent->oied_name, ent->oied_namelen, &ln);
7524 dentry = osd_child_dentry_by_inode(env, dir, ln.ln_name, ln.ln_namelen);
7525 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7526 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7532 * We need to ensure that the name entry is still valid.
7533 * Because it may be removed or renamed by other already.
7535 * The unlink or rename operation will start journal before PDO lock,
7536 * so to avoid deadlock, here we need to start journal handle before
7537 * related PDO lock also. But because we do not know whether there
7538 * will be something to be repaired before PDO lock, we just start
7539 * journal without conditions.
7541 * We may need to remove the name entry firstly, then insert back.
7542 * One credit is for user quota file update.
7543 * One credit is for group quota file update.
7544 * Two credits are for dirty inode.
7546 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7547 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7549 if (dev->od_dirent_journal != 0) {
7552 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7555 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7556 "check_repair, dir = %lu/%u, credits = %d, "
7557 "name = %.*s, ino = %llu: rc = %d\n",
7558 devname, dir->i_ino, dir->i_generation, credits,
7559 ent->oied_namelen, ent->oied_name,
7562 GOTO(out_inode, rc);
7565 if (obj->oo_hl_head != NULL) {
7566 hlock = osd_oti_get(env)->oti_hlock;
7568 * "0" means exclusive lock for the whole directory.
7569 * We need to prevent others access such name entry
7570 * during the delete + insert. Neither HLOCK_ADD nor
7571 * HLOCK_DEL cannot guarantee the atomicity.
7573 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7575 down_write(&obj->oo_ext_idx_sem);
7578 if (obj->oo_hl_head != NULL) {
7579 hlock = osd_oti_get(env)->oti_hlock;
7580 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7581 LDISKFS_HLOCK_LOOKUP);
7583 down_read(&obj->oo_ext_idx_sem);
7587 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7588 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7589 *attr |= LUDA_IGNORE;
7595 * For dotdot entry, if there is not enough space to hold the
7596 * FID-in-dirent, just keep them there. It only happens when the
7597 * device upgraded from 1.8 or restored from MDT file-level backup.
7598 * For the whole directory, only dotdot entry have no FID-in-dirent
7599 * and needs to get FID from LMA when readdir, it will not affect the
7602 if (dotdot && !osd_dotdot_has_space(de)) {
7603 *attr |= LUDA_UNKNOWN;
7609 if (lu_fid_eq(fid, &lma->lma_self_fid))
7612 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7613 struct lu_fid *tfid = &lma->lma_self_fid;
7615 if (likely(dotdot &&
7616 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7617 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7619 * It must be REMOTE_PARENT_DIR and as the
7620 * 'dotdot' entry of remote directory
7622 *attr |= LUDA_IGNORE;
7624 CDEBUG(D_LFSCK, "%s: expect remote agent "
7625 "parent directory, but got %.*s under "
7626 "dir = %lu/%u with the FID "DFID"\n",
7627 devname, ent->oied_namelen,
7628 ent->oied_name, dir->i_ino,
7629 dir->i_generation, PFID(tfid));
7631 *attr |= LUDA_UNKNOWN;
7638 if (!fid_is_zero(fid)) {
7639 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7641 if (rc == -ENOENT ||
7643 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7645 * linkEA does not recognize the dirent entry,
7646 * it may because the dirent entry corruption
7647 * and points to other's inode.
7649 CDEBUG(D_LFSCK, "%s: the target inode does not "
7650 "recognize the dirent, dir = %lu/%u, "
7651 " name = %.*s, ino = %llu, "
7652 DFID": rc = %d\n", devname, dir->i_ino,
7653 dir->i_generation, ent->oied_namelen,
7654 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7655 *attr |= LUDA_UNKNOWN;
7660 if (rc && rc != -ENODATA) {
7661 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7662 "dir = %lu/%u, name = %.*s, ino = %llu, "
7663 DFID": rc = %d\n", devname, dir->i_ino,
7664 dir->i_generation, ent->oied_namelen,
7665 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7666 *attr |= LUDA_UNKNOWN;
7674 * linkEA recognizes the dirent entry, the FID-in-LMA is
7675 * valid, trusted, in spite of fid_is_sane(fid) or not.
7677 if (*attr & LUDA_VERIFY_DRYRUN) {
7678 *fid = lma->lma_self_fid;
7679 *attr |= LUDA_REPAIR;
7686 dev->od_dirent_journal = 1;
7687 if (hlock != NULL) {
7688 ldiskfs_htree_unlock(hlock);
7691 up_read(&obj->oo_ext_idx_sem);
7697 *fid = lma->lma_self_fid;
7699 /* Update or append the FID-in-dirent. */
7700 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7701 bh, de, hlock, dotdot);
7703 *attr |= LUDA_REPAIR;
7705 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7706 "the dirent, dir = %lu/%u, name = %.*s, "
7707 "ino = %llu, "DFID": rc = %d\n",
7708 devname, dir->i_ino, dir->i_generation,
7709 ent->oied_namelen, ent->oied_name,
7710 ent->oied_ino, PFID(fid), rc);
7712 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7713 if (*attr & LUDA_VERIFY_DRYRUN) {
7714 if (fid_is_sane(fid)) {
7715 *attr |= LUDA_REPAIR;
7716 } else if (dev->od_index == 0) {
7717 lu_igif_build(fid, inode->i_ino,
7718 inode->i_generation);
7719 *attr |= LUDA_UPGRADE;
7727 dev->od_dirent_journal = 1;
7728 if (hlock != NULL) {
7729 ldiskfs_htree_unlock(hlock);
7732 up_read(&obj->oo_ext_idx_sem);
7739 if (unlikely(fid_is_sane(fid))) {
7741 * FID-in-dirent exists, but FID-in-LMA is lost.
7742 * Trust the FID-in-dirent, and add FID-in-LMA.
7744 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7746 *attr |= LUDA_REPAIR;
7748 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7749 "update dirent, dir = %lu/%u, "
7750 "name = %.*s, ino = %llu, "
7752 devname, dir->i_ino, dir->i_generation,
7753 ent->oied_namelen, ent->oied_name,
7754 ent->oied_ino, PFID(fid), rc);
7755 } else if (dev->od_index == 0) {
7756 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7758 * It is probably IGIF object. Only aappend the
7759 * FID-in-dirent. OI scrub will process FID-in-LMA.
7761 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7762 bh, de, hlock, dotdot);
7764 *attr |= LUDA_UPGRADE;
7766 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7767 "after the dirent, dir = %lu/%u, "
7768 "name = %.*s, ino = %llu, "
7770 devname, dir->i_ino, dir->i_generation,
7771 ent->oied_namelen, ent->oied_name,
7772 ent->oied_ino, PFID(fid), rc);
7781 if (hlock != NULL) {
7782 ldiskfs_htree_unlock(hlock);
7784 if (dev->od_dirent_journal != 0)
7785 up_write(&obj->oo_ext_idx_sem);
7787 up_read(&obj->oo_ext_idx_sem);
7791 ldiskfs_journal_stop(jh);
7795 if (rc >= 0 && !dirty)
7796 dev->od_dirent_journal = 0;
7797 if (ln.ln_name != ent->oied_name)
7804 * Returns the value at current position from iterator's in memory structure.
7806 * \param di struct osd_it_ea, iterator's in memory structure
7807 * \param attr attr requested for dirent.
7808 * \param lde lustre dirent
7810 * \retval 0 no error and \param lde has correct lustre dirent.
7811 * \retval -ve on error
7813 static inline int osd_it_ea_rec(const struct lu_env *env,
7814 const struct dt_it *di,
7815 struct dt_rec *dtrec, __u32 attr)
7817 struct osd_it_ea *it = (struct osd_it_ea *)di;
7818 struct osd_object *obj = it->oie_obj;
7819 struct osd_device *dev = osd_obj2dev(obj);
7820 struct osd_thread_info *oti = osd_oti_get(env);
7821 struct osd_inode_id *id = &oti->oti_id;
7822 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7823 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7824 __u32 ino = it->oie_dirent->oied_ino;
7829 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7831 if (attr & LUDA_VERIFY) {
7832 if (unlikely(is_remote_parent_ino(dev, ino))) {
7833 attr |= LUDA_IGNORE;
7835 * If the parent is on remote MDT, and there
7836 * is no FID-in-dirent, then we have to get
7837 * the parent FID from the linkEA.
7839 if (!fid_is_sane(fid) &&
7840 it->oie_dirent->oied_namelen == 2 &&
7841 it->oie_dirent->oied_name[0] == '.' &&
7842 it->oie_dirent->oied_name[1] == '.')
7843 osd_get_pfid_from_linkea(env, obj, fid);
7845 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7849 if (!fid_is_sane(fid))
7850 attr |= LUDA_UNKNOWN;
7852 attr &= ~LU_DIRENT_ATTRS_MASK;
7853 if (!fid_is_sane(fid)) {
7854 bool is_dotdot = false;
7856 if (it->oie_dirent->oied_namelen == 2 &&
7857 it->oie_dirent->oied_name[0] == '.' &&
7858 it->oie_dirent->oied_name[1] == '.')
7861 * If the parent is on remote MDT, and there
7862 * is no FID-in-dirent, then we have to get
7863 * the parent FID from the linkEA.
7865 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7866 rc = osd_get_pfid_from_linkea(env, obj, fid);
7868 if (is_dotdot == false &&
7869 CFS_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7872 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7877 /* Pack the entry anyway, at least the offset is right. */
7878 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7879 it->oie_dirent->oied_name,
7880 it->oie_dirent->oied_namelen,
7881 it->oie_dirent->oied_type, attr);
7883 RETURN(rc > 0 ? 0 : rc);
7887 * Returns the record size size at current position.
7889 * This function will return record(lu_dirent) size in bytes.
7891 * \param[in] env execution environment
7892 * \param[in] di iterator's in memory structure
7893 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7894 * calculate the lu_dirent size.
7896 * \retval record size(in bytes & in memory) of the current lu_dirent
7899 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7902 struct osd_it_ea *it = (struct osd_it_ea *)di;
7904 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7908 * Returns a cookie for current position of the iterator head, so that
7909 * user can use this cookie to load/start the iterator next time.
7911 * \param di iterator's in memory structure
7913 * \retval cookie for current position, on success
7915 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7917 struct osd_it_ea *it = (struct osd_it_ea *)di;
7919 return it->oie_dirent->oied_off;
7923 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7924 * to load a directory entry at a time and stored it i inn,
7925 * in iterator's in-memory data structure.
7927 * \param di struct osd_it_ea, iterator's in memory structure
7929 * \retval +ve on success
7930 * \retval -ve on error
7932 static int osd_it_ea_load(const struct lu_env *env,
7933 const struct dt_it *di, __u64 hash)
7935 struct osd_it_ea *it = (struct osd_it_ea *)di;
7939 it->oie_file->f_pos = hash;
7941 rc = osd_ldiskfs_it_fill(env, di);
7951 int osd_olc_lookup(const struct lu_env *env, struct osd_object *obj,
7952 u64 iversion, struct dt_rec *rec,
7953 const struct lu_name *ln, int *result)
7955 struct osd_thread_info *oti = osd_oti_get(env);
7956 struct osd_lookup_cache *olc = oti->oti_lookup_cache;
7957 struct osd_device *osd = osd_obj2dev(obj);
7958 struct osd_lookup_cache_object *cobj = &oti->oti_cobj;
7961 if (unlikely(olc == NULL))
7964 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
7966 * umount has happened, a new OSD could land to the previous
7967 * address so we can't use it any more, invalidate our cache
7969 memset(olc, 0, sizeof(*olc));
7970 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
7974 memset(cobj, 0, sizeof(*cobj));
7975 cobj->lco_osd = osd;
7976 cobj->lco_ino = obj->oo_inode->i_ino;
7977 cobj->lco_gen = obj->oo_inode->i_generation;
7978 cobj->lco_version = iversion;
7980 for (i = 0; i < OSD_LOOKUP_CACHE_MAX; i++) {
7981 struct osd_lookup_cache_entry *entry;
7983 entry = &olc->olc_entry[i];
7984 /* compare if osd/ino/generation/version match */
7985 if (memcmp(&entry->lce_obj, cobj, sizeof(*cobj)) != 0)
7987 if (entry->lce_namelen != ln->ln_namelen)
7989 if (memcmp(entry->lce_name, ln->ln_name, ln->ln_namelen) != 0)
7992 memcpy(rec, &entry->lce_fid, sizeof(entry->lce_fid));
7993 *result = entry->lce_rc;
7999 void osd_olc_save(const struct lu_env *env, struct osd_object *obj,
8000 struct dt_rec *rec, const struct lu_name *ln,
8001 const int result, u64 iversion)
8003 struct osd_thread_info *oti = osd_oti_get(env);
8004 struct osd_lookup_cache_entry *entry;
8005 struct osd_lookup_cache *olc;
8007 if (unlikely(oti->oti_lookup_cache == NULL)) {
8008 OBD_ALLOC_PTR(oti->oti_lookup_cache);
8009 if (oti->oti_lookup_cache == NULL)
8013 olc = oti->oti_lookup_cache;
8014 if (unlikely(atomic_read(&osd_mount_seq) != olc->olc_mount_seq)) {
8015 memset(olc, 0, sizeof(*olc));
8016 olc->olc_mount_seq = atomic_read(&osd_mount_seq);
8019 entry = &olc->olc_entry[olc->olc_cur];
8021 /* invaliate cache slot if needed */
8022 if (entry->lce_obj.lco_osd)
8023 memset(&entry->lce_obj, 0, sizeof(entry->lce_obj));
8025 /* XXX: some kind of LRU */
8026 entry->lce_obj.lco_osd = osd_obj2dev(obj);
8027 entry->lce_obj.lco_ino = obj->oo_inode->i_ino;
8028 entry->lce_obj.lco_gen = obj->oo_inode->i_generation;
8029 entry->lce_obj.lco_version = iversion;
8031 LASSERT(ln->ln_namelen <= LDISKFS_NAME_LEN + 1);
8032 entry->lce_namelen = ln->ln_namelen;
8033 memcpy(entry->lce_name, ln->ln_name, ln->ln_namelen);
8034 memcpy(&entry->lce_fid, rec, sizeof(entry->lce_fid));
8035 entry->lce_rc = result;
8037 if (++olc->olc_cur == OSD_LOOKUP_CACHE_MAX)
8042 * Index lookup function for interoperability mode (b11826).
8044 * \param key, key i.e. file name to be searched
8046 * \retval +ve, on success
8047 * \retval -ve, on error
8049 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
8050 struct dt_rec *rec, const struct dt_key *key)
8052 struct osd_object *obj = osd_dt_obj(dt);
8059 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
8060 LINVRNT(osd_invariant(obj));
8062 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
8067 * grab version before actual lookup, so that we recognize potential
8068 * insert between osd_ea_lookup_rec() and osd_olc_save()
8070 iversion = inode_peek_iversion(obj->oo_inode);
8072 if (osd_olc_lookup(env, obj, iversion, rec, &ln, &result))
8073 GOTO(out, rc = result);
8075 rc = osd_ea_lookup_rec(env, obj, rec, &ln);
8079 osd_olc_save(env, obj, rec, &ln, rc, iversion);
8082 if (ln.ln_name != (char *)key)
8088 * Index and Iterator operations for interoperability
8089 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
8091 static const struct dt_index_operations osd_index_ea_ops = {
8092 .dio_lookup = osd_index_ea_lookup,
8093 .dio_declare_insert = osd_index_declare_ea_insert,
8094 .dio_insert = osd_index_ea_insert,
8095 .dio_declare_delete = osd_index_declare_ea_delete,
8096 .dio_delete = osd_index_ea_delete,
8098 .init = osd_it_ea_init,
8099 .fini = osd_it_ea_fini,
8100 .get = osd_it_ea_get,
8101 .put = osd_it_ea_put,
8102 .next = osd_it_ea_next,
8103 .key = osd_it_ea_key,
8104 .key_size = osd_it_ea_key_size,
8105 .rec = osd_it_ea_rec,
8106 .rec_size = osd_it_ea_rec_size,
8107 .store = osd_it_ea_store,
8108 .load = osd_it_ea_load
8112 static void *osd_key_init(const struct lu_context *ctx,
8113 struct lu_context_key *key)
8115 struct osd_thread_info *info;
8117 OBD_ALLOC_PTR(info);
8119 return ERR_PTR(-ENOMEM);
8121 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8122 if (info->oti_it_ea_buf == NULL)
8125 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
8127 info->oti_hlock = ldiskfs_htree_lock_alloc();
8128 if (info->oti_hlock == NULL)
8134 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8137 return ERR_PTR(-ENOMEM);
8140 static void osd_key_fini(const struct lu_context *ctx,
8141 struct lu_context_key *key, void *data)
8143 struct osd_thread_info *info = data;
8144 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
8145 struct osd_idmap_cache *idc = info->oti_ins_cache;
8147 if (info->oti_dio_pages) {
8149 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
8150 struct page *page = info->oti_dio_pages[i];
8152 LASSERT(PagePrivate2(page));
8153 LASSERT(PageLocked(page));
8154 ClearPagePrivate2(page);
8159 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
8160 PTLRPC_MAX_BRW_PAGES);
8163 if (info->oti_inode != NULL)
8165 if (info->oti_hlock != NULL)
8166 ldiskfs_htree_lock_free(info->oti_hlock);
8167 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
8168 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
8169 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
8170 lu_buf_free(&info->oti_big_buf);
8172 LASSERT(info->oti_ins_cache_size > 0);
8173 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
8174 info->oti_ins_cache = NULL;
8175 info->oti_ins_cache_size = 0;
8177 if (info->oti_lookup_cache)
8178 OBD_FREE_PTR(info->oti_lookup_cache);
8182 static void osd_key_exit(const struct lu_context *ctx,
8183 struct lu_context_key *key, void *data)
8185 struct osd_thread_info *info = data;
8186 struct osd_lookup_cache *olc = info->oti_lookup_cache;
8189 memset(olc, 0, sizeof(*olc));
8190 LASSERT(info->oti_r_locks == 0);
8191 LASSERT(info->oti_w_locks == 0);
8192 LASSERT(info->oti_txns == 0);
8193 LASSERTF(info->oti_dio_pages_used == 0, "%d\n",
8194 info->oti_dio_pages_used);
8197 /* type constructor/destructor: osd_type_init, osd_type_fini */
8198 LU_TYPE_INIT_FINI(osd, &osd_key);
8200 struct lu_context_key osd_key = {
8201 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
8202 .lct_init = osd_key_init,
8203 .lct_fini = osd_key_fini,
8204 .lct_exit = osd_key_exit
8208 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
8209 const char *name, struct lu_device *next)
8211 struct osd_device *osd = osd_dev(d);
8212 int cplen = strscpy(osd->od_svname, name, sizeof(osd->od_svname));
8216 return osd_procfs_init(osd, name);
8219 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
8221 struct seq_server_site *ss = osd_seq_site(osd);
8226 if (osd->od_is_ost || osd->od_cl_seq != NULL)
8229 if (unlikely(ss == NULL))
8232 OBD_ALLOC_PTR(osd->od_cl_seq);
8233 if (osd->od_cl_seq == NULL)
8236 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
8237 osd->od_svname, ss->ss_server_seq);
8239 if (ss->ss_node_id == 0) {
8241 * If the OSD on the sequence controller(MDT0), then allocate
8242 * sequence here, otherwise allocate sequence after connected
8243 * to MDT0 (see mdt_register_lwp_callback()).
8245 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
8246 &osd->od_cl_seq->lcs_space, env);
8252 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
8254 if (osd->od_cl_seq == NULL)
8257 seq_client_fini(osd->od_cl_seq);
8258 OBD_FREE_PTR(osd->od_cl_seq);
8259 osd->od_cl_seq = NULL;
8262 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
8266 /* shutdown quota slave instance associated with the device */
8267 if (o->od_quota_slave_md != NULL) {
8268 struct qsd_instance *qsd = o->od_quota_slave_md;
8270 o->od_quota_slave_md = NULL;
8274 if (o->od_quota_slave_dt != NULL) {
8275 struct qsd_instance *qsd = o->od_quota_slave_dt;
8277 o->od_quota_slave_dt = NULL;
8281 osd_fid_fini(env, o);
8282 osd_scrub_cleanup(env, o);
8287 #ifdef HAVE_FLUSH_DELAYED_FPUT
8288 # define cfs_flush_delayed_fput() flush_delayed_fput()
8290 void (*cfs_flush_delayed_fput)(void);
8291 #endif /* HAVE_FLUSH_DELAYED_FPUT */
8293 static void osd_umount(const struct lu_env *env, struct osd_device *o)
8297 atomic_inc(&osd_mount_seq);
8299 if (o->od_mnt != NULL) {
8300 shrink_dcache_sb(osd_sb(o));
8301 osd_sync(env, &o->od_dt_dev);
8302 wait_event(o->od_commit_cb_done,
8303 !atomic_read(&o->od_commit_cb_in_flight));
8309 /* to be sure all delayed fput are finished */
8310 cfs_flush_delayed_fput();
8315 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8316 # ifndef LDISKFS_HAS_INCOMPAT_FEATURE
8317 /* Newer kernels provide the ldiskfs_set_feature_largedir() wrapper already,
8318 * which calls ldiskfs_update_dynamic_rev() to update ancient filesystems.
8319 * All ldiskfs filesystems are already v2, so it is a no-op and unnecessary.
8320 * This avoids maintaining patches to export this otherwise-useless function.
8322 void ldiskfs_update_dynamic_rev(struct super_block *sb)
8329 static int osd_mount(const struct lu_env *env,
8330 struct osd_device *o, struct lustre_cfg *cfg)
8332 const char *name = lustre_cfg_string(cfg, 0);
8333 const char *dev = lustre_cfg_string(cfg, 1);
8335 unsigned long page, s_flags = 0, lmd_flags = 0;
8336 struct page *__page;
8337 struct file_system_type *type;
8338 char *options = NULL;
8340 struct osd_thread_info *info = osd_oti_get(env);
8341 struct lu_fid *fid = &info->oti_fid;
8342 struct inode *inode;
8343 int rc = 0, force_over_1024tb = 0;
8347 if (o->od_mnt != NULL)
8350 if (strlen(dev) >= sizeof(o->od_mntdev))
8352 strcpy(o->od_mntdev, dev);
8354 str = lustre_cfg_buf(cfg, 2);
8355 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
8357 opts = lustre_cfg_string(cfg, 3);
8359 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
8360 CERROR("%s: device %s extents feature is not guaranteed to "
8361 "work on big-endian systems. Use \"bigendian_extents\" "
8362 "mount option to override.\n", name, dev);
8366 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
8367 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
8368 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");
8371 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
8372 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
8373 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");
8376 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8377 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
8378 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");
8382 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
8383 force_over_1024tb = 1;
8385 __page = alloc_page(GFP_KERNEL);
8387 GOTO(out, rc = -ENOMEM);
8388 page = (unsigned long)page_address(__page);
8389 options = (char *)page;
8392 /* strip out the options for back compatiblity */
8393 static const char * const sout[] = {
8400 /* strip out option we processed in osd */
8401 "bigendian_extents",
8405 "force_over_1024tb",
8410 strncat(options, opts, PAGE_SIZE);
8411 for (rc = 0, str = options; sout[rc]; ) {
8412 char *op = strstr(str, sout[rc]);
8419 if (op == options || *(op - 1) == ',') {
8420 str = op + strlen(sout[rc]);
8421 if (*str == ',' || *str == '\0') {
8422 *str == ',' ? str++ : str;
8423 memmove(op, str, strlen(str) + 1);
8426 for (str = op; *str != ',' && *str != '\0'; str++)
8430 strncat(options, "user_xattr,acl", PAGE_SIZE);
8433 /* Glom up mount options */
8434 if (*options != '\0')
8435 strncat(options, ",", PAGE_SIZE);
8436 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
8438 type = get_fs_type("ldiskfs");
8440 CERROR("%s: cannot find ldiskfs module\n", name);
8441 GOTO(out, rc = -ENODEV);
8444 s_flags |= SB_KERNMOUNT;
8445 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
8446 module_put(type->owner);
8448 if (IS_ERR(o->od_mnt)) {
8449 rc = PTR_ERR(o->od_mnt);
8451 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
8455 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
8456 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
8457 force_over_1024tb == 0) {
8458 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",
8460 GOTO(out_mnt, rc = -EINVAL);
8463 if (test_bit(LMD_FLG_DEV_RDONLY, &lmd_flags)) {
8464 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
8467 GOTO(out_mnt, rc = -EOPNOTSUPP);
8470 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
8471 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
8472 GOTO(out_mnt, rc = -EINVAL);
8475 if (ldiskfs_has_feature_fast_commit(o->od_mnt->mnt_sb)) {
8476 CERROR("%s: device %s is mounted with fast_commit that breaks recovery\n",
8478 GOTO(out_mnt, rc = -EOPNOTSUPP);
8481 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8482 #ifdef LDISKFS_MOUNT_DIRDATA
8483 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
8484 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
8485 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
8486 CWARN("%s: device %s was upgraded from Lustre-1.x without "
8487 "enabling the dirdata feature. If you do not want to "
8488 "downgrade to Lustre-1.x again, you can enable it via "
8489 "'tune2fs -O dirdata device'\n", name, dev);
8491 /* enable large_dir on MDTs to avoid REMOTE_PARENT_DIR overflow,
8492 * and on very large OSTs to avoid object directory overflow */
8493 if (unlikely(!ldiskfs_has_feature_largedir(o->od_mnt->mnt_sb) &&
8494 !strstr(name, "MGS"))) {
8495 ldiskfs_set_feature_largedir(o->od_mnt->mnt_sb);
8496 LCONSOLE_INFO("%s: enabled 'large_dir' feature on device %s\n",
8500 inode = osd_sb(o)->s_root->d_inode;
8501 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
8502 if (!o->od_dt_dev.dd_rdonly) {
8503 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
8505 CERROR("%s: failed to set lma on %s root inode\n",
8511 if (test_bit(LMD_FLG_NOSCRUB, &lmd_flags))
8512 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_NEVER;
8514 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
8515 /* do not use pagecache with flash-backed storage */
8516 o->od_writethrough_cache = 0;
8517 o->od_read_cache = 0;
8528 __free_page(__page);
8533 static struct lu_device *osd_device_fini(const struct lu_env *env,
8534 struct lu_device *d)
8536 struct osd_device *o = osd_dev(d);
8540 osd_index_backup(env, o, false);
8541 osd_shutdown(env, o);
8543 if (o->od_oi_table != NULL)
8544 osd_oi_fini(osd_oti_get(env), o);
8545 if (o->od_extent_bytes_percpu)
8546 free_percpu(o->od_extent_bytes_percpu);
8547 osd_obj_map_fini(o);
8553 static int osd_device_init0(const struct lu_env *env,
8554 struct osd_device *o,
8555 struct lustre_cfg *cfg)
8557 struct lu_device *l = osd2lu_dev(o);
8558 struct osd_thread_info *info;
8561 bool restored = false;
8564 /* if the module was re-loaded, env can loose its keys */
8565 rc = lu_env_refill((struct lu_env *)env);
8568 info = osd_oti_get(env);
8571 l->ld_ops = &osd_lu_ops;
8572 o->od_dt_dev.dd_ops = &osd_dt_ops;
8574 spin_lock_init(&o->od_osfs_lock);
8575 mutex_init(&o->od_otable_mutex);
8576 INIT_LIST_HEAD(&o->od_orphan_list);
8577 INIT_LIST_HEAD(&o->od_index_backup_list);
8578 INIT_LIST_HEAD(&o->od_index_restore_list);
8579 spin_lock_init(&o->od_lock);
8580 o->od_index_backup_policy = LIBP_NONE;
8582 init_waitqueue_head(&o->od_commit_cb_done);
8584 o->od_read_cache = 1;
8585 o->od_writethrough_cache = 1;
8586 o->od_enable_projid_xattr = 0;
8587 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
8588 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
8589 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
8590 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_DEFAULT;
8591 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8592 o->od_fallocate_zero_blocks = 0;
8594 cplen = strscpy(o->od_svname, lustre_cfg_string(cfg, 4),
8595 sizeof(o->od_svname));
8597 GOTO(out, rc = cplen);
8599 o->od_index_backup_stop = 0;
8600 o->od_index = -1; /* -1 means index is invalid */
8601 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8602 if (rc == LDD_F_SV_TYPE_OST)
8605 o->od_full_scrub_ratio = OFSR_DEFAULT;
8606 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8607 rc = osd_mount(env, o, cfg);
8611 /* Can only check block device after mount */
8612 o->od_nonrotational =
8613 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8615 rc = osd_obj_map_init(env, o);
8619 rc = lu_site_init(&o->od_site, l);
8621 GOTO(out_compat, rc);
8622 o->od_site.ls_bottom_dev = l;
8624 rc = lu_site_init_finish(&o->od_site);
8628 opts = lustre_cfg_string(cfg, 3);
8629 if (opts && strstr(opts, "resetoi"))
8632 INIT_LIST_HEAD(&o->od_ios_list);
8634 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
8636 GOTO(out_brw_stats, rc);
8638 /* setup scrub, including OI files initialization */
8640 rc = osd_scrub_setup(env, o, restored);
8643 GOTO(out_brw_stats, rc);
8645 rc = osd_procfs_init(o, o->od_svname);
8647 CERROR("%s: can't initialize procfs: rc = %d\n",
8649 GOTO(out_scrub, rc);
8652 LASSERT(l->ld_site->ls_linkage.next != NULL);
8653 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8655 /* initialize quota slave instance */
8656 /* currently it's no need to prepare qsd_instance_md for OST */
8657 if (!o->od_is_ost) {
8658 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8659 &o->od_dt_dev, o->od_proc_entry,
8661 if (IS_ERR(o->od_quota_slave_md)) {
8662 rc = PTR_ERR(o->od_quota_slave_md);
8663 o->od_quota_slave_md = NULL;
8664 GOTO(out_procfs, rc);
8668 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8669 o->od_proc_entry, false, true);
8671 if (IS_ERR(o->od_quota_slave_dt)) {
8672 if (o->od_quota_slave_md != NULL) {
8673 qsd_fini(env, o->od_quota_slave_md);
8674 o->od_quota_slave_md = NULL;
8677 rc = PTR_ERR(o->od_quota_slave_dt);
8678 o->od_quota_slave_dt = NULL;
8679 GOTO(out_procfs, rc);
8682 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8683 if (!o->od_extent_bytes_percpu) {
8685 GOTO(out_procfs, rc);
8693 osd_scrub_cleanup(env, o);
8695 lprocfs_fini_brw_stats(&o->od_brw_stats);
8697 lu_site_fini(&o->od_site);
8699 osd_obj_map_fini(o);
8706 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8707 struct lu_device_type *t,
8708 struct lustre_cfg *cfg)
8710 struct osd_device *o;
8715 return ERR_PTR(-ENOMEM);
8717 rc = dt_device_init(&o->od_dt_dev, t);
8720 * Because the ctx might be revived in dt_device_init,
8721 * refill the env here
8723 lu_env_refill((struct lu_env *)env);
8724 rc = osd_device_init0(env, o, cfg);
8726 dt_device_fini(&o->od_dt_dev);
8729 if (unlikely(rc != 0))
8732 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8735 static struct lu_device *osd_device_free(const struct lu_env *env,
8736 struct lu_device *d)
8738 struct osd_device *o = osd_dev(d);
8742 /* XXX: make osd top device in order to release reference */
8743 d->ld_site->ls_top_dev = d;
8744 lu_site_purge(env, d->ld_site, -1);
8745 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8746 D_ERROR, lu_cdebug_printer);
8747 lu_site_fini(&o->od_site);
8748 dt_device_fini(&o->od_dt_dev);
8753 static int osd_process_config(const struct lu_env *env,
8754 struct lu_device *d, struct lustre_cfg *cfg)
8756 struct osd_device *o = osd_dev(d);
8762 switch (cfg->lcfg_command) {
8764 rc = osd_mount(env, o, cfg);
8768 * For the case LCFG_PRE_CLEANUP is not called in advance,
8769 * that may happend if hit failure during mount process.
8771 osd_index_backup(env, o, false);
8772 lu_dev_del_linkage(d->ld_site, d);
8773 rc = osd_shutdown(env, o);
8776 LASSERT(&o->od_dt_dev);
8777 count = class_modify_config(cfg, PARAM_OSD,
8778 &o->od_dt_dev.dd_kobj);
8780 count = class_modify_config(cfg, PARAM_OST,
8781 &o->od_dt_dev.dd_kobj);
8782 rc = count > 0 ? 0 : count;
8784 case LCFG_PRE_CLEANUP:
8786 osd_index_backup(env, o,
8787 o->od_index_backup_policy != LIBP_NONE);
8797 static int osd_recovery_complete(const struct lu_env *env,
8798 struct lu_device *d)
8800 struct osd_device *osd = osd_dev(d);
8805 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8809 * start qsd instance on recovery completion, this notifies the quota
8810 * slave code that we are about to process new requests now
8812 rc = qsd_start(env, osd->od_quota_slave_dt);
8813 if (rc == 0 && osd->od_quota_slave_md != NULL)
8814 rc = qsd_start(env, osd->od_quota_slave_md);
8820 * we use exports to track all osd users
8822 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8823 struct obd_device *obd, struct obd_uuid *cluuid,
8824 struct obd_connect_data *data, void *localdata)
8826 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8827 struct lustre_handle conn;
8832 CDEBUG(D_CONFIG, "connect #%d\n", atomic_read(&osd->od_connects));
8834 rc = class_connect(&conn, obd, cluuid);
8838 *exp = class_conn2export(&conn);
8840 atomic_inc(&osd->od_connects);
8846 * once last export (we don't count self-export) disappeared
8847 * osd can be released
8849 static int osd_obd_disconnect(struct obd_export *exp)
8851 struct obd_device *obd = exp->exp_obd;
8852 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8853 int rc, release = 0;
8857 /* Only disconnect the underlying layers on the final disconnect. */
8858 release = atomic_dec_and_test(&osd->od_connects);
8860 rc = class_disconnect(exp); /* bz 9811 */
8862 if (rc == 0 && release)
8863 class_manual_cleanup(obd);
8867 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8868 struct lu_device *dev)
8870 struct osd_device *osd = osd_dev(dev);
8871 struct lr_server_data *lsd =
8872 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8877 if (osd->od_quota_slave_md != NULL) {
8878 /* set up quota slave objects for inode */
8879 result = qsd_prepare(env, osd->od_quota_slave_md);
8884 if (osd->od_quota_slave_dt != NULL) {
8885 /* set up quota slave objects for block */
8886 result = qsd_prepare(env, osd->od_quota_slave_dt);
8892 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8893 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8894 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8895 osd->od_index_in_idif = 1;
8897 osd->od_index_in_idif = 0;
8898 result = osd_register_proc_index_in_idif(osd);
8903 osd->od_index_in_idif = 1;
8907 result = osd_fid_init(env, osd);
8913 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8917 * see include/lu_object.h for the details.
8919 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8920 struct lu_fid *fid, struct lu_object *parent,
8921 const struct lu_name *name)
8923 struct osd_device *osd = osd_dev(d);
8925 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8928 static const struct lu_object_operations osd_lu_obj_ops = {
8929 .loo_object_init = osd_object_init,
8930 .loo_object_delete = osd_object_delete,
8931 .loo_object_release = osd_object_release,
8932 .loo_object_free = osd_object_free,
8933 .loo_object_print = osd_object_print,
8934 .loo_object_invariant = osd_object_invariant
8937 const struct lu_device_operations osd_lu_ops = {
8938 .ldo_object_alloc = osd_object_alloc,
8939 .ldo_process_config = osd_process_config,
8940 .ldo_recovery_complete = osd_recovery_complete,
8941 .ldo_prepare = osd_prepare,
8942 .ldo_fid_alloc = osd_fid_alloc,
8945 static const struct lu_device_type_operations osd_device_type_ops = {
8946 .ldto_init = osd_type_init,
8947 .ldto_fini = osd_type_fini,
8949 .ldto_start = osd_type_start,
8950 .ldto_stop = osd_type_stop,
8952 .ldto_device_alloc = osd_device_alloc,
8953 .ldto_device_free = osd_device_free,
8955 .ldto_device_init = osd_device_init,
8956 .ldto_device_fini = osd_device_fini
8959 static struct lu_device_type osd_device_type = {
8960 .ldt_tags = LU_DEVICE_DT,
8961 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8962 .ldt_ops = &osd_device_type_ops,
8963 .ldt_ctx_tags = LCT_LOCAL,
8966 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8968 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8969 struct super_block *sb = osd_sb(osd);
8971 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8975 * lprocfs legacy support.
8977 static const struct obd_ops osd_obd_device_ops = {
8978 .o_owner = THIS_MODULE,
8979 .o_connect = osd_obd_connect,
8980 .o_disconnect = osd_obd_disconnect,
8981 .o_health_check = osd_health_check,
8984 static ssize_t delayed_unlink_mb_show(struct kobject *kobj,
8985 struct attribute *attr, char *buf)
8987 return snprintf(buf, PAGE_SIZE, "%d\n",
8988 ldiskfs_delayed_unlink_blocks >> 11);
8991 static ssize_t delayed_unlink_mb_store(struct kobject *kobj,
8992 struct attribute *attr,
8993 const char *buffer, size_t count)
8995 u64 delayed_unlink_bytes;
8998 rc = sysfs_memparse(buffer, count, &delayed_unlink_bytes, "MiB");
9002 ldiskfs_delayed_unlink_blocks = delayed_unlink_bytes >> 9;
9006 LUSTRE_RW_ATTR(delayed_unlink_mb);
9009 static ssize_t track_declares_assert_show(struct kobject *kobj,
9010 struct attribute *attr,
9013 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
9016 static ssize_t track_declares_assert_store(struct kobject *kobj,
9017 struct attribute *attr,
9018 const char *buffer, size_t count)
9020 bool track_declares_assert;
9023 rc = kstrtobool(buffer, &track_declares_assert);
9027 ldiskfs_track_declares_assert = track_declares_assert;
9031 LUSTRE_RW_ATTR(track_declares_assert);
9033 static int __init osd_init(void)
9035 struct kobject *kobj;
9038 BUILD_BUG_ON(BH_DXLock >=
9039 sizeof(((struct buffer_head *)0)->b_state) * 8);
9040 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
9041 /* please, try to keep osd_thread_info smaller than a page */
9042 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
9045 rc = libcfs_setup();
9051 rc = lu_kmem_init(ldiskfs_caches);
9055 rc = class_register_type(&osd_obd_device_ops, NULL, true,
9056 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
9058 lu_kmem_fini(ldiskfs_caches);
9062 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9064 rc = sysfs_create_file(kobj,
9065 &lustre_attr_track_declares_assert.attr);
9067 CWARN("%s: track_declares_assert sysfs registration failed: rc = %d\n",
9072 rc = sysfs_create_file(kobj,
9073 &lustre_attr_delayed_unlink_mb.attr);
9075 CWARN("%s: delayed_unlink_mb registration failed: rc = %d\n",
9083 #ifndef HAVE_FLUSH_DELAYED_FPUT
9084 if (unlikely(cfs_flush_delayed_fput == NULL))
9085 cfs_flush_delayed_fput =
9086 cfs_kallsyms_lookup_name("flush_delayed_fput");
9092 static void __exit osd_exit(void)
9094 struct kobject *kobj;
9096 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
9098 sysfs_remove_file(kobj,
9099 &lustre_attr_track_declares_assert.attr);
9102 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
9103 lu_kmem_fini(ldiskfs_caches);
9106 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
9107 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
9108 MODULE_VERSION(LUSTRE_VERSION_STRING);
9109 MODULE_LICENSE("GPL");
9111 module_init(osd_init);
9112 module_exit(osd_exit);