4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/osd/osd_handler.c
34 * Top-level entry points into osd module
36 * Author: Nikita Danilov <nikita@clusterfs.com>
37 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
40 #define DEBUG_SUBSYSTEM S_OSD
42 #include <linux/fs_struct.h>
43 #include <linux/kallsyms.h>
44 #include <linux/module.h>
45 #include <linux/user_namespace.h>
46 #include <linux/uidgid.h>
48 /* prerequisite for linux/xattr.h */
49 #include <linux/types.h>
50 /* prerequisite for linux/xattr.h */
52 /* XATTR_{REPLACE,CREATE} */
53 #include <linux/xattr.h>
55 #include <ldiskfs/ldiskfs.h>
56 #include <ldiskfs/xattr.h>
57 #include <ldiskfs/ldiskfs_extents.h>
60 * struct OBD_{ALLOC,FREE}*()
63 #include <obd_support.h>
64 /* struct ptlrpc_thread */
65 #include <lustre_net.h>
66 #include <lustre_fid.h>
68 #include <uapi/linux/lustre/lustre_param.h>
70 #include "osd_internal.h"
71 #include "osd_dynlocks.h"
73 /* llo_* api support */
74 #include <md_object.h>
75 #include <lustre_quota.h>
77 #include <lustre_linkea.h>
79 /* Maximum EA size is limited by LNET_MTU for remote objects */
80 #define OSD_MAX_EA_SIZE 1048364
83 module_param(ldiskfs_pdo, int, 0644);
84 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
86 int ldiskfs_track_declares_assert;
87 module_param(ldiskfs_track_declares_assert, int, 0644);
88 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
90 /* Slab to allocate dynlocks */
91 struct kmem_cache *dynlock_cachep;
93 /* Slab to allocate osd_it_ea */
94 struct kmem_cache *osd_itea_cachep;
96 static struct lu_kmem_descr ldiskfs_caches[] = {
98 .ckd_cache = &dynlock_cachep,
99 .ckd_name = "dynlock_cache",
100 .ckd_size = sizeof(struct dynlock_handle)
103 .ckd_cache = &osd_itea_cachep,
104 .ckd_name = "osd_itea_cache",
105 .ckd_size = sizeof(struct osd_it_ea)
112 static const char dot[] = ".";
113 static const char dotdot[] = "..";
115 static const struct lu_object_operations osd_lu_obj_ops;
116 static const struct dt_object_operations osd_obj_ops;
117 static const struct dt_object_operations osd_obj_otable_it_ops;
118 static const struct dt_index_operations osd_index_iam_ops;
119 static const struct dt_index_operations osd_index_ea_ops;
121 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
122 const struct lu_fid *fid);
123 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
124 struct osd_device *osd);
126 int osd_trans_declare_op2rb[] = {
127 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
128 [OSD_OT_PUNCH] = OSD_OT_MAX,
129 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
130 [OSD_OT_CREATE] = OSD_OT_DESTROY,
131 [OSD_OT_DESTROY] = OSD_OT_CREATE,
132 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
133 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
134 [OSD_OT_WRITE] = OSD_OT_WRITE,
135 [OSD_OT_INSERT] = OSD_OT_DELETE,
136 [OSD_OT_DELETE] = OSD_OT_INSERT,
137 [OSD_OT_QUOTA] = OSD_OT_MAX,
140 static int osd_has_index(const struct osd_object *obj)
142 return obj->oo_dt.do_index_ops != NULL;
145 static int osd_object_invariant(const struct lu_object *l)
147 return osd_invariant(osd_obj(l));
151 * Concurrency: doesn't matter
153 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
155 struct osd_thread_info *oti = osd_oti_get(env);
157 return oti->oti_w_locks > 0 && o->oo_owner == env;
161 * Concurrency: doesn't access mutable data
163 static int osd_root_get(const struct lu_env *env,
164 struct dt_device *dev, struct lu_fid *f)
166 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
171 * the following set of functions are used to maintain per-thread
172 * cache of FID->ino mapping. this mechanism is needed to resolve
173 * FID to inode at dt_insert() which in turn stores ino in the
174 * directory entries to keep ldiskfs compatible with ext[34].
175 * due to locking-originated restrictions we can't lookup ino
176 * using LU cache (deadlock is possible). lookup using OI is quite
177 * expensive. so instead we maintain this cache and methods like
178 * dt_create() fill it. so in the majority of cases dt_insert() is
179 * able to find needed mapping in lockless manner.
181 static struct osd_idmap_cache *
182 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
183 const struct lu_fid *fid)
185 struct osd_thread_info *oti = osd_oti_get(env);
186 struct osd_idmap_cache *idc = oti->oti_ins_cache;
189 for (i = 0; i < oti->oti_ins_cache_used; i++) {
190 if (!lu_fid_eq(&idc[i].oic_fid, fid))
192 if (idc[i].oic_dev != osd)
201 static struct osd_idmap_cache *
202 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
203 const struct lu_fid *fid)
205 struct osd_thread_info *oti = osd_oti_get(env);
206 struct osd_idmap_cache *idc;
209 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
210 i = oti->oti_ins_cache_size * 2;
212 i = OSD_INS_CACHE_SIZE;
213 OBD_ALLOC_PTR_ARRAY(idc, i);
215 return ERR_PTR(-ENOMEM);
216 if (oti->oti_ins_cache != NULL) {
217 memcpy(idc, oti->oti_ins_cache,
218 oti->oti_ins_cache_used * sizeof(*idc));
219 OBD_FREE_PTR_ARRAY(oti->oti_ins_cache,
220 oti->oti_ins_cache_used);
222 oti->oti_ins_cache = idc;
223 oti->oti_ins_cache_size = i;
226 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
229 idc->oic_lid.oii_ino = 0;
230 idc->oic_lid.oii_gen = 0;
237 * lookup mapping for the given fid in the cache, initialize a
238 * new one if not found. the initialization checks whether the
239 * object is local or remote. for local objects, OI is used to
240 * learn ino/generation. the function is used when the caller
241 * has no information about the object, e.g. at dt_insert().
243 static struct osd_idmap_cache *
244 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
245 const struct lu_fid *fid)
247 struct osd_idmap_cache *idc;
250 idc = osd_idc_find(env, osd, fid);
251 LASSERT(!IS_ERR(idc));
255 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
256 osd->od_svname, PFID(fid));
258 /* new mapping is needed */
259 idc = osd_idc_add(env, osd, fid);
261 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
262 osd->od_svname, PFID(fid), PTR_ERR(idc));
267 rc = osd_remote_fid(env, osd, fid);
268 if (unlikely(rc < 0))
272 /* the object is local, lookup in OI */
273 /* XXX: probably cheaper to lookup in LU first? */
274 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
276 if (unlikely(rc < 0)) {
277 CERROR("can't lookup: rc = %d\n", rc);
281 /* the object is remote */
288 static void osd_idc_dump_lma(const struct lu_env *env,
289 struct osd_device *osd,
293 struct osd_thread_info *info = osd_oti_get(env);
294 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
295 const struct lu_fid *fid;
296 struct osd_inode_id lid;
300 inode = osd_ldiskfs_iget(osd_sb(osd), ino);
302 CERROR("%s: can't get inode %lu: rc = %d\n",
303 osd->od_svname, ino, (int)PTR_ERR(inode));
306 if (is_bad_inode(inode)) {
307 CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
310 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
312 CERROR("%s: can't get LMA for %lu: rc = %d\n",
313 osd->od_svname, ino, rc);
316 fid = &loa->loa_lma.lma_self_fid;
317 LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
318 PFID(fid), ino, (unsigned)inode->i_generation);
321 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
323 CERROR("%s: can't lookup "DFID": rc = %d\n",
324 osd->od_svname, PFID(fid), rc);
327 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
328 PFID(fid), lid.oii_ino, lid.oii_gen);
333 static void osd_idc_dump_debug(const struct lu_env *env,
334 struct osd_device *osd,
335 const struct lu_fid *fid,
339 struct osd_inode_id lid;
343 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
345 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
346 osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
347 osd_idc_dump_lma(env, osd, lid.oii_ino, false);
349 CERROR("%s: can't lookup "DFID": rc = %d\n",
350 osd->od_svname, PFID(fid), rc);
353 osd_idc_dump_lma(env, osd, ino1, true);
355 osd_idc_dump_lma(env, osd, ino2, true);
359 * lookup mapping for given FID and fill it from the given object.
360 * the object is lolcal by definition.
362 static int osd_idc_find_and_init(const struct lu_env *env,
363 struct osd_device *osd,
364 struct osd_object *obj)
366 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
367 struct osd_idmap_cache *idc;
369 idc = osd_idc_find(env, osd, fid);
370 LASSERT(!IS_ERR(idc));
372 if (obj->oo_inode == NULL)
374 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
375 if (idc->oic_lid.oii_ino) {
376 osd_idc_dump_debug(env, osd, fid,
377 idc->oic_lid.oii_ino,
378 obj->oo_inode->i_ino);
381 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
382 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
387 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
388 osd->od_svname, PFID(fid));
390 /* new mapping is needed */
391 idc = osd_idc_add(env, osd, fid);
393 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
394 osd->od_svname, PFID(fid), PTR_ERR(idc));
398 if (obj->oo_inode != NULL) {
399 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
400 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
406 * OSD object methods.
410 * Concurrency: no concurrent access is possible that early in object
413 static struct lu_object *osd_object_alloc(const struct lu_env *env,
414 const struct lu_object_header *hdr,
417 struct osd_object *mo;
422 struct lu_object_header *h;
423 struct osd_device *o = osd_dev(d);
425 l = &mo->oo_dt.do_lu;
426 if (unlikely(o->od_in_init)) {
433 lu_object_header_init(h);
434 lu_object_init(l, h, d);
435 lu_object_add_top(h, l);
438 dt_object_init(&mo->oo_dt, NULL, d);
439 mo->oo_header = NULL;
442 mo->oo_dt.do_ops = &osd_obj_ops;
443 l->lo_ops = &osd_lu_obj_ops;
444 init_rwsem(&mo->oo_sem);
445 init_rwsem(&mo->oo_ext_idx_sem);
446 spin_lock_init(&mo->oo_guard);
447 INIT_LIST_HEAD(&mo->oo_xattr_list);
453 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
454 struct dentry *dentry, struct lustre_ost_attrs *loa)
458 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
459 (void *)loa, sizeof(*loa));
461 struct lustre_mdt_attrs *lma = &loa->loa_lma;
463 if (rc < sizeof(*lma))
467 lustre_loa_swab(loa, true);
468 /* Check LMA compatibility */
469 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
471 CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
473 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
474 PFID(&lma->lma_self_fid), inode->i_ino, rc);
476 } else if (rc == 0) {
484 * retrieve object from backend ext fs.
486 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
487 struct osd_inode_id *id)
490 struct inode *inode = NULL;
493 * if we look for an inode withing a running
494 * transaction, then we risk to deadlock
495 * osd_dirent_check_repair() breaks this
497 /* LASSERT(current->journal_info == NULL); */
499 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
501 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
502 id->oii_ino, PTR_ERR(inode));
503 } else if (id->oii_gen != OSD_OII_NOGEN &&
504 inode->i_generation != id->oii_gen) {
505 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
506 "i_generation = %u\n",
507 id->oii_ino, id->oii_gen, inode->i_generation);
509 inode = ERR_PTR(-ESTALE);
510 } else if (inode->i_nlink == 0) {
512 * due to parallel readdir and unlink,
513 * we can have dead inode here.
515 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
517 inode = ERR_PTR(-ESTALE);
518 } else if (is_bad_inode(inode)) {
520 CWARN("%s: bad inode: ino = %u: rc = %d\n",
521 osd_dev2name(dev), id->oii_ino, rc);
524 } else if ((rc = osd_attach_jinode(inode))) {
528 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
529 if (id->oii_gen == OSD_OII_NOGEN)
530 osd_id_gen(id, inode->i_ino, inode->i_generation);
533 * Do not update file c/mtime in ldiskfs.
534 * NB: we don't have any lock to protect this because we don't
535 * have reference on osd_object now, but contention with
536 * another lookup + attr_set can't happen in the tiny window
537 * between if (...) and set S_NOCMTIME.
539 if (!(inode->i_flags & S_NOCMTIME))
540 inode->i_flags |= S_NOCMTIME;
545 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
546 handle_t *handle, struct dentry *child,
547 struct inode *inode, struct htree_lock *hlock)
551 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
552 if (rc == -ENOBUFS || rc == -ENOSPC) {
553 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
554 struct inode *parent = child->d_parent->d_inode;
555 struct lu_fid *fid = NULL;
556 char fidstr[FID_LEN + 1] = "unknown";
558 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
560 fid = &loa->loa_lma.lma_self_fid;
561 } else if (rc2 == -ENODATA) {
562 if (unlikely(is_root_inode(parent))) {
563 fid = &info->oti_fid3;
564 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
565 } else if (!osd->od_is_ost && osd->od_index == 0) {
566 fid = &info->oti_fid3;
567 lu_igif_build(fid, parent->i_ino,
568 parent->i_generation);
573 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
575 /* below message is checked in sanity.sh test_129 */
577 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
578 osd_name(osd), parent->i_ino, fidstr);
580 rc = 0; /* ignore such error now */
581 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
582 osd_name(osd), parent->i_ino, fidstr);
592 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
593 struct osd_inode_id *id, struct lu_fid *fid)
595 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
599 inode = osd_iget(info, dev, id);
603 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
605 *fid = loa->loa_lma.lma_self_fid;
606 } else if (rc == -ENODATA) {
607 if (unlikely(is_root_inode(inode)))
608 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
610 lu_igif_build(fid, inode->i_ino, inode->i_generation);
618 static struct inode *osd_iget_check(struct osd_thread_info *info,
619 struct osd_device *dev,
620 const struct lu_fid *fid,
621 struct osd_inode_id *id,
630 * The cached OI mapping is trustable. If we cannot locate the inode
631 * via the cached OI mapping, then return the failure to the caller
632 * directly without further OI checking.
636 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
639 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
642 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
643 PFID(fid), id->oii_ino, rc);
647 if (is_bad_inode(inode)) {
652 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
653 PFID(fid), id->oii_ino);
657 if (id->oii_gen != OSD_OII_NOGEN &&
658 inode->i_generation != id->oii_gen) {
663 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
664 "oii_gen = %u, i_generation = %u\n", PFID(fid),
665 id->oii_ino, id->oii_gen, inode->i_generation);
669 if (inode->i_nlink == 0) {
674 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
675 PFID(fid), id->oii_ino);
679 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
683 __u32 saved_ino = id->oii_ino;
684 __u32 saved_gen = id->oii_gen;
687 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
689 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
691 * XXX: There are four possible cases:
693 * Backup/restore caused the OI invalid.
695 * Someone unlinked the object but NOT removed
696 * the OI mapping, such as mount target device
697 * as ldiskfs, and modify something directly.
699 * Someone just removed the object between the
700 * former oi_lookup and the iget. It is normal.
701 * 4. Other failure cases.
703 * Generally, when the device is mounted, it will
704 * auto check whether the system is restored from
705 * file-level backup or not. We trust such detect
706 * to distinguish the 1st case from the 2nd case:
707 * if the OI files are consistent but may contain
708 * stale OI mappings because of case 2, if iget()
709 * returns -ENOENT or -ESTALE, then it should be
714 * If the OI mapping was in OI file before the
715 * osd_iget_check(), but now, it is disappear,
716 * then it must be removed by race. That is a
722 * It is the OI scrub updated the OI mapping by race.
723 * The new OI mapping must be valid.
725 if (saved_ino != id->oii_ino ||
726 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
735 if (dev->od_scrub.os_scrub.os_file.sf_flags &
738 * It still can be the case 2, but we cannot
739 * distinguish it from the case 1. So return
740 * -EREMCHG to block current operation until
741 * OI scrub rebuilt the OI mappings.
750 if (inode->i_generation == id->oii_gen)
755 if (id->oii_gen == OSD_OII_NOGEN)
756 osd_id_gen(id, inode->i_ino, inode->i_generation);
759 * Do not update file c/mtime in ldiskfs.
760 * NB: we don't have any lock to protect this because we don't
761 * have reference on osd_object now, but contention with
762 * another lookup + attr_set can't happen in the tiny window
763 * between if (...) and set S_NOCMTIME.
765 if (!(inode->i_flags & S_NOCMTIME))
766 inode->i_flags |= S_NOCMTIME;
783 * \retval +v: new filter_fid does not contain self-fid
784 * \retval 0: filter_fid_18_23, contains self-fid
785 * \retval -v: other failure cases
787 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
788 struct dentry *dentry, struct lu_fid *fid)
790 struct filter_fid *ff = &info->oti_ff;
791 struct ost_id *ostid = &info->oti_ostid;
794 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
795 if (rc == sizeof(struct filter_fid_18_23)) {
796 struct filter_fid_18_23 *ff_old = (void *)ff;
798 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
799 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
801 * XXX: use 0 as the index for compatibility, the caller will
802 * handle index related issues when necessary.
805 ostid_to_fid(fid, ostid, 0);
806 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
808 } else if (rc >= 0) {
815 static int osd_lma_self_repair(struct osd_thread_info *info,
816 struct osd_device *osd, struct inode *inode,
817 const struct lu_fid *fid, __u32 compat)
822 LASSERT(current->journal_info == NULL);
824 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
825 osd_dto_credits_noquota[DTO_XATTR_SET]);
828 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
833 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
835 CWARN("%s: cannot self repair the LMA: rc = %d\n",
837 ldiskfs_journal_stop(jh);
841 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
843 struct osd_thread_info *info = osd_oti_get(env);
844 struct osd_device *osd = osd_obj2dev(obj);
845 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
846 struct lustre_mdt_attrs *lma = &loa->loa_lma;
847 struct inode *inode = obj->oo_inode;
848 struct dentry *dentry = &info->oti_obj_dentry;
849 struct lu_fid *fid = NULL;
850 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
855 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
856 (void *)loa, sizeof(*loa));
857 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
858 fid = &lma->lma_self_fid;
859 rc = osd_get_idif(info, inode, dentry, fid);
860 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
862 * For the given OST-object, if it has neither LMA nor
863 * FID in XATTR_NAME_FID, then the given FID (which is
864 * contained in the @obj, from client RPC for locating
865 * the OST-object) is trusted. We use it to generate
868 osd_lma_self_repair(info, osd, inode, rfid,
879 lustre_lma_swab(lma);
880 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
881 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
882 S_ISREG(inode->i_mode)))) {
883 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
884 "fid = "DFID", ino = %lu\n", osd_name(osd),
885 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
886 PFID(rfid), inode->i_ino);
889 fid = &lma->lma_self_fid;
890 if (lma->lma_compat & LMAC_STRIPE_INFO &&
892 obj->oo_pfid_in_lma = 1;
893 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
895 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
899 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
900 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
901 struct ost_id *oi = &info->oti_ostid;
902 struct lu_fid *fid1 = &info->oti_fid3;
903 __u32 idx = fid_idif_ost_idx(rfid);
906 * For old IDIF, the OST index is not part of the IDIF,
907 * Means that different OSTs may have the same IDIFs.
908 * Under such case, we need to make some compatible
909 * check to make sure to trigger OI scrub properly.
911 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
912 /* Given @rfid is new, LMA is old. */
913 fid_to_ostid(fid, oi);
914 ostid_to_fid(fid1, oi, idx);
915 if (lu_fid_eq(fid1, rfid)) {
916 if (osd->od_index_in_idif)
917 osd_lma_self_repair(info, osd,
931 struct osd_check_lmv_buf {
932 /* please keep it as first member */
933 struct dir_context ctx;
934 struct osd_thread_info *oclb_info;
935 struct osd_device *oclb_dev;
941 * It is called internally by ->iterate*() to filter out the
942 * local slave object's FID of the striped directory.
944 * \retval 1 found the local slave's FID
945 * \retval 0 continue to check next item
946 * \retval -ve for failure
948 #ifdef HAVE_FILLDIR_USE_CTX
949 static int osd_stripe_dir_filldir(struct dir_context *buf,
951 static int osd_stripe_dir_filldir(void *buf,
953 const char *name, int namelen,
954 loff_t offset, __u64 ino, unsigned int d_type)
956 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
957 struct osd_thread_info *oti = oclb->oclb_info;
958 struct lu_fid *fid = &oti->oti_fid3;
959 struct osd_inode_id *id = &oti->oti_id3;
960 struct osd_device *dev = oclb->oclb_dev;
969 sscanf(name + 1, SFID, RFID(fid));
970 if (!fid_is_sane(fid))
973 if (osd_remote_fid(oti->oti_env, dev, fid))
976 osd_id_gen(id, ino, OSD_OII_NOGEN);
977 inode = osd_iget(oti, dev, id);
979 return PTR_ERR(inode);
982 osd_add_oi_cache(oti, dev, id, fid);
983 osd_oii_insert(dev, fid, id, true);
984 oclb->oclb_found = true;
990 * When lookup item under striped directory, we need to locate the master
991 * MDT-object of the striped directory firstly, then the client will send
992 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
993 * and the item's name. If the system is restored from MDT file level backup,
994 * then before the OI scrub completely built the OI files, the OI mappings of
995 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
996 * not a problem for the master MDT-object. Because when locate the master
997 * MDT-object, we will do name based lookup (for the striped directory itself)
998 * firstly, during such process we can setup the correct OI mapping for the
999 * master MDT-object. But it will be trouble for the slave MDT-object. Because
1000 * the client will not trigger name based lookup on the MDT to locate the slave
1001 * MDT-object before locating item under the striped directory, then when
1002 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
1003 * is invalid and does not know what the right OI mapping is, then the MDT has
1004 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
1005 * the OI file, related OI mapping is unknown yet, please try again later. And
1006 * then client will re-try the RPC again and again until related OI mapping has
1007 * been updated. That is quite inefficient.
1009 * To resolve above trouble, we will handle it as the following two cases:
1011 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
1012 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
1013 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
1014 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
1015 * directly. Please check osd_fid_lookup().
1017 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
1018 * Under such case, during lookup the master MDT-object, we will lookup the
1019 * slave MDT-object via readdir against the master MDT-object, because the
1020 * slave MDT-objects information are stored as sub-directories with the name
1021 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
1022 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
1023 * the correct OI mapping for the slave MDT-object.
1025 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
1026 struct inode *inode)
1028 struct lu_buf *buf = &oti->oti_big_buf;
1029 struct dentry *dentry = &oti->oti_obj_dentry;
1031 struct lmv_mds_md_v1 *lmv1;
1032 struct osd_check_lmv_buf oclb = {
1033 .ctx.actor = osd_stripe_dir_filldir,
1036 .oclb_found = false,
1043 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
1045 if (rc == -ERANGE) {
1046 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
1048 lu_buf_realloc(buf, rc);
1049 if (buf->lb_buf == NULL)
1050 GOTO(out, rc = -ENOMEM);
1056 if (unlikely(rc == 0 || rc == -ENODATA))
1062 if (unlikely(buf->lb_buf == NULL)) {
1063 lu_buf_realloc(buf, rc);
1064 if (buf->lb_buf == NULL)
1065 GOTO(out, rc = -ENOMEM);
1071 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1074 filp = osd_quasi_file(oti->oti_env, inode);
1075 rc = osd_security_file_alloc(filp);
1080 oclb.oclb_items = 0;
1081 rc = iterate_dir(filp, &oclb.ctx);
1082 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1083 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1084 inode->i_fop->release(inode, filp);
1089 "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
1090 osd_ino2name(inode), inode->i_ino, inode->i_generation,
1098 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1099 const struct lu_fid *fid,
1100 const struct lu_object_conf *conf)
1102 struct osd_thread_info *info;
1103 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1104 struct osd_device *dev;
1105 struct osd_idmap_cache *oic;
1106 struct osd_inode_id *id;
1107 struct inode *inode = NULL;
1108 struct lustre_scrub *scrub;
1109 struct scrub_file *sf;
1110 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1115 bool remote = false;
1116 bool trusted = true;
1117 bool updated = false;
1118 bool checked = false;
1122 LINVRNT(osd_invariant(obj));
1123 LASSERT(obj->oo_inode == NULL);
1125 if (fid_is_sane(fid) == 0) {
1126 CERROR("%s: invalid FID "DFID"\n", ldev->ld_obd->obd_name,
1132 dev = osd_dev(ldev);
1133 scrub = &dev->od_scrub.os_scrub;
1134 sf = &scrub->os_file;
1135 info = osd_oti_get(env);
1137 oic = &info->oti_cache;
1139 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1143 * For the object is created as locking anchor, or for the object to
1144 * be created on disk. No need to osd_oi_lookup() at here because FID
1145 * shouldn't never be re-used, if it's really a duplicate FID from
1146 * unexpected reason, we should be able to detect it later by calling
1147 * do_create->osd_oi_insert().
1149 if (conf && conf->loc_flags & LOC_F_NEW)
1150 GOTO(out, result = 0);
1152 /* Search order: 1. per-thread cache. */
1153 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1159 if (!list_empty(&scrub->os_inconsistent_items)) {
1160 /* Search order: 2. OI scrub pending list. */
1161 result = osd_oii_lookup(dev, fid, id);
1167 * The OI mapping in the OI file can be updated by the OI scrub
1168 * when we locate the inode via FID. So it may be not trustable.
1172 /* Search order: 3. OI files. */
1173 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1174 if (result == -ENOENT) {
1175 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1176 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1177 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1179 GOTO(out, result = 0);
1184 /* -ESTALE is returned if inode of OST object doesn't exist */
1185 if (result == -ESTALE &&
1186 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1187 GOTO(out, result = 0);
1194 obj->oo_inode = NULL;
1195 /* for later passes through checks, not true on first pass */
1196 if (!IS_ERR_OR_NULL(inode))
1199 inode = osd_iget_check(info, dev, fid, id, trusted);
1200 if (!IS_ERR(inode)) {
1201 obj->oo_inode = inode;
1209 result = PTR_ERR(inode);
1210 if (result == -ENOENT || result == -ESTALE)
1211 GOTO(out, result = 0);
1213 if (result != -EREMCHG)
1218 * We still have chance to get the valid inode: for the
1219 * object which is referenced by remote name entry, the
1220 * object on the local MDT will be linked under the dir
1221 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1223 * We do not know whether the object for the given FID
1224 * is referenced by some remote name entry or not, and
1225 * especially for DNE II, a multiple-linked object may
1226 * have many name entries reside on many MDTs.
1228 * To simplify the operation, OSD will not distinguish
1229 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1230 * only happened for the RPC from other MDT during the
1231 * OI scrub, or for the client side RPC with FID only,
1232 * such as FID to path, or from old connected client.
1235 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1239 flags |= SS_AUTO_PARTIAL;
1240 flags &= ~SS_AUTO_FULL;
1245 if (scrub->os_running) {
1246 if (scrub->os_partial_scan && !scrub->os_in_join)
1249 if (IS_ERR_OR_NULL(inode) || result) {
1250 osd_oii_insert(dev, fid, id, result == -ENOENT);
1251 GOTO(out, result = -EINPROGRESS);
1255 LASSERT(obj->oo_inode == inode);
1257 osd_oii_insert(dev, fid, id, true);
1261 if (dev->od_auto_scrub_interval == AS_NEVER) {
1263 GOTO(out, result = -EREMCHG);
1266 LASSERT(obj->oo_inode == inode);
1268 osd_add_oi_cache(info, dev, id, fid);
1273 rc1 = osd_scrub_start(env, dev, flags);
1274 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
1275 "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
1276 osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
1277 if (rc1 && rc1 != -EALREADY)
1278 GOTO(out, result = -EREMCHG);
1280 if (IS_ERR_OR_NULL(inode) || result) {
1281 osd_oii_insert(dev, fid, id, result == -ENOENT);
1282 GOTO(out, result = -EINPROGRESS);
1286 LASSERT(obj->oo_inode == inode);
1288 osd_oii_insert(dev, fid, id, true);
1293 if (unlikely(obj->oo_header))
1296 result = osd_check_lma(env, obj);
1300 LASSERTF(id->oii_ino == inode->i_ino &&
1301 id->oii_gen == inode->i_generation,
1302 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1303 PFID(fid), id->oii_ino, id->oii_gen,
1304 inode->i_ino, inode->i_generation);
1306 saved_ino = inode->i_ino;
1307 saved_gen = inode->i_generation;
1309 if (unlikely(result == -ENODATA)) {
1311 * If the OI scrub updated the OI mapping by race, it
1312 * must be valid. Trust the inode that has no LMA EA.
1317 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1320 * The OI mapping is still there, the inode is still
1321 * valid. It is just becaues the inode has no LMA EA.
1323 if (saved_ino == id->oii_ino &&
1324 saved_gen == id->oii_gen)
1328 * It is the OI scrub updated the OI mapping by race.
1329 * The new OI mapping must be valid.
1337 * "result == -ENOENT" means that the OI mappinghas been
1338 * removed by race, so the inode belongs to other object.
1340 * Others error can be returned directly.
1342 if (result == -ENOENT) {
1345 obj->oo_inode = NULL;
1350 if (result != -EREMCHG)
1356 * if two OST objects map to the same inode, and inode mode is
1357 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1358 * reserved by precreate, and not written yet, in this case, don't
1359 * set inode for the object whose FID mismatch, so that it can create
1360 * inode and not block precreate.
1362 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1363 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1364 obj->oo_inode = NULL;
1365 GOTO(out, result = 0);
1368 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1370 * "result == -ENOENT" means the cached OI mapping has been removed
1371 * from the OI file by race, above inode belongs to other object.
1373 if (result == -ENOENT) {
1376 obj->oo_inode = NULL;
1377 GOTO(out, result = 0);
1383 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1385 osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
1391 * It is the OI scrub updated the OI mapping by race.
1392 * The new OI mapping must be valid.
1400 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1401 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1403 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1405 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1407 obj->oo_pfid_in_lma = 1;
1408 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1410 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1411 } else if (result != -ENODATA) {
1416 obj->oo_compat_dot_created = 1;
1417 obj->oo_compat_dotdot_created = 1;
1419 if (S_ISDIR(inode->i_mode) &&
1420 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1421 osd_check_lmv(info, dev, inode);
1423 result = osd_attach_jinode(inode);
1428 GOTO(out, result = 0);
1430 LASSERT(!obj->oo_hl_head);
1431 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1433 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1436 if (result || !obj->oo_inode) {
1437 if (!IS_ERR_OR_NULL(inode))
1440 obj->oo_inode = NULL;
1442 fid_zero(&oic->oic_fid);
1445 LINVRNT(osd_invariant(obj));
1450 * Concurrency: shouldn't matter.
1452 static void osd_object_init0(struct osd_object *obj)
1454 LASSERT(obj->oo_inode != NULL);
1455 obj->oo_dt.do_body_ops = &osd_body_ops;
1456 obj->oo_dt.do_lu.lo_header->loh_attr |=
1457 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1461 * Concurrency: no concurrent access is possible that early in object
1464 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1465 const struct lu_object_conf *conf)
1467 struct osd_object *obj = osd_obj(l);
1470 LINVRNT(osd_invariant(obj));
1472 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1473 cfs_fail_val == 2) {
1474 struct osd_thread_info *info = osd_oti_get(env);
1475 struct osd_idmap_cache *oic = &info->oti_cache;
1476 /* invalidate thread cache */
1477 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1479 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1480 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1481 l->lo_header->loh_attr |= LOHA_EXISTS;
1485 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1486 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1487 if (result == 0 && obj->oo_inode != NULL) {
1488 struct osd_thread_info *oti = osd_oti_get(env);
1489 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1491 osd_object_init0(obj);
1492 if (unlikely(obj->oo_header))
1495 result = osd_get_lma(oti, obj->oo_inode,
1496 &oti->oti_obj_dentry, loa);
1499 * Convert LMAI flags to lustre LMA flags
1500 * and cache it to oo_lma_flags
1503 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1504 } else if (result == -ENODATA) {
1508 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
1510 LINVRNT(osd_invariant(obj));
1515 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1516 * The left part is for value, binary mode.
1518 struct osd_xattr_entry {
1519 struct list_head oxe_list;
1523 struct rcu_head oxe_rcu;
1527 static int osd_oxc_get(struct osd_object *obj, const char *name,
1530 struct osd_xattr_entry *tmp;
1531 struct osd_xattr_entry *oxe = NULL;
1532 size_t namelen = strlen(name);
1536 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1537 if (namelen == tmp->oxe_namelen &&
1538 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1545 GOTO(out, rc = -ENOENT);
1547 if (!oxe->oxe_exist)
1548 GOTO(out, rc = -ENODATA);
1551 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1554 if (buf->lb_buf == NULL)
1557 if (buf->lb_len < rc)
1558 GOTO(out, rc = -ERANGE);
1560 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1567 static void osd_oxc_free(struct rcu_head *head)
1569 struct osd_xattr_entry *oxe;
1571 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1572 OBD_FREE(oxe, oxe->oxe_len);
1575 static void osd_oxc_add(struct osd_object *obj, const char *name,
1576 const char *buf, int buflen)
1578 struct osd_xattr_entry *oxe;
1579 struct osd_xattr_entry *old = NULL;
1580 struct osd_xattr_entry *tmp;
1581 size_t namelen = strlen(name);
1582 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1584 OBD_ALLOC(oxe, len);
1588 INIT_LIST_HEAD(&oxe->oxe_list);
1590 oxe->oxe_namelen = namelen;
1591 memcpy(oxe->oxe_buf, name, namelen);
1593 LASSERT(buf != NULL);
1594 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1595 oxe->oxe_exist = true;
1597 oxe->oxe_exist = false;
1600 /* this should be rarely called, just remove old and add new */
1601 spin_lock(&obj->oo_guard);
1602 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1603 if (namelen == tmp->oxe_namelen &&
1604 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1610 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1611 call_rcu(&old->oxe_rcu, osd_oxc_free);
1613 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1615 spin_unlock(&obj->oo_guard);
1618 static void osd_oxc_del(struct osd_object *obj, const char *name)
1620 struct osd_xattr_entry *oxe;
1621 size_t namelen = strlen(name);
1623 spin_lock(&obj->oo_guard);
1624 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1625 if (namelen == oxe->oxe_namelen &&
1626 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1627 list_del_rcu(&oxe->oxe_list);
1628 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1632 spin_unlock(&obj->oo_guard);
1635 static void osd_oxc_fini(struct osd_object *obj)
1637 struct osd_xattr_entry *oxe, *next;
1639 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1640 list_del(&oxe->oxe_list);
1641 OBD_FREE(oxe, oxe->oxe_len);
1646 * Concurrency: no concurrent access is possible that late in object
1649 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1651 struct osd_object *obj = osd_obj(l);
1652 struct lu_object_header *h = obj->oo_header;
1654 LINVRNT(osd_invariant(obj));
1657 dt_object_fini(&obj->oo_dt);
1658 if (obj->oo_hl_head != NULL)
1659 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1660 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1663 lu_object_header_free(h);
1667 * Concurrency: no concurrent access is possible that late in object
1670 static void osd_index_fini(struct osd_object *o)
1672 struct iam_container *bag;
1674 if (o->oo_dir != NULL) {
1675 bag = &o->oo_dir->od_container;
1676 if (o->oo_inode != NULL) {
1677 if (bag->ic_object == o->oo_inode)
1678 iam_container_fini(bag);
1680 OBD_FREE_PTR(o->oo_dir);
1686 OSD_TXN_OI_DELETE_CREDITS = 20,
1687 OSD_TXN_INODE_DELETE_CREDITS = 20
1694 #if OSD_THANDLE_STATS
1696 * Set time when the handle is allocated
1698 static void osd_th_alloced(struct osd_thandle *oth)
1700 oth->oth_alloced = ktime_get();
1704 * Set time when the handle started
1706 static void osd_th_started(struct osd_thandle *oth)
1708 oth->oth_started = ktime_get();
1712 * Check whether the we deal with this handle for too long.
1714 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1715 ktime_t alloced, ktime_t started,
1718 ktime_t now = ktime_get();
1720 LASSERT(dev != NULL);
1722 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1723 ktime_us_delta(started, alloced));
1724 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1725 ktime_us_delta(closed, started));
1726 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1727 ktime_us_delta(now, closed));
1729 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1730 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1731 oth, now, alloced, started, closed);
1732 libcfs_debug_dumpstack(NULL);
1736 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1738 ktime_t __closed = ktime_get(); \
1739 ktime_t __alloced = oth->oth_alloced; \
1740 ktime_t __started = oth->oth_started; \
1743 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1746 #else /* OSD_THANDLE_STATS */
1748 #define osd_th_alloced(h) do {} while(0)
1749 #define osd_th_started(h) do {} while(0)
1750 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1752 #endif /* OSD_THANDLE_STATS */
1755 * Concurrency: doesn't access mutable data.
1757 static int osd_param_is_not_sane(const struct osd_device *dev,
1758 const struct thandle *th)
1760 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1762 return oh->ot_credits > osd_transaction_size(dev);
1766 * Concurrency: shouldn't matter.
1768 static void osd_trans_commit_cb(struct super_block *sb,
1769 struct ldiskfs_journal_cb_entry *jcb, int error)
1771 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1772 struct thandle *th = &oh->ot_super;
1773 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1774 struct osd_device *osd = osd_dev(lud);
1775 struct dt_txn_commit_cb *dcb, *tmp;
1777 LASSERT(oh->ot_handle == NULL);
1780 CERROR("transaction @0x%p commit error: %d\n", th, error);
1782 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1783 /* call per-transaction callbacks if any */
1784 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1786 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1787 "commit callback entry: magic=%x name='%s'\n",
1788 dcb->dcb_magic, dcb->dcb_name);
1789 list_del_init(&dcb->dcb_linkage);
1790 dcb->dcb_func(NULL, th, dcb, error);
1793 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1794 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1795 wake_up(&osd->od_commit_cb_done);
1801 static struct thandle *osd_trans_create(const struct lu_env *env,
1802 struct dt_device *d)
1804 struct osd_thread_info *oti = osd_oti_get(env);
1805 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1806 struct osd_thandle *oh;
1812 CERROR("%s: someone try to start transaction under "
1813 "readonly mode, should be disabled.\n",
1814 osd_name(osd_dt_dev(d)));
1816 RETURN(ERR_PTR(-EROFS));
1819 /* on pending IO in this thread should left from prev. request */
1820 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1822 sb_start_write(osd_sb(osd_dt_dev(d)));
1824 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1826 sb_end_write(osd_sb(osd_dt_dev(d)));
1827 RETURN(ERR_PTR(-ENOMEM));
1830 oh->ot_quota_trans = &oti->oti_quota_trans;
1831 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1836 oh->oh_declared_ext = 0;
1837 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1838 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1839 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1842 memset(oti->oti_declare_ops, 0,
1843 sizeof(oti->oti_declare_ops));
1844 memset(oti->oti_declare_ops_cred, 0,
1845 sizeof(oti->oti_declare_ops_cred));
1846 memset(oti->oti_declare_ops_used, 0,
1847 sizeof(oti->oti_declare_ops_used));
1849 oti->oti_ins_cache_depth++;
1854 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1856 struct osd_thread_info *oti = osd_oti_get(env);
1857 struct osd_thandle *oh;
1859 oh = container_of(th, struct osd_thandle, ot_super);
1860 LASSERT(oh != NULL);
1862 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1863 oti->oti_declare_ops[OSD_OT_CREATE],
1864 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1865 oti->oti_declare_ops_used[OSD_OT_CREATE],
1866 oti->oti_declare_ops[OSD_OT_DESTROY],
1867 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1868 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1869 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1870 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1871 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1872 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1873 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1874 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1875 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1876 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1877 oti->oti_declare_ops[OSD_OT_WRITE],
1878 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1879 oti->oti_declare_ops_used[OSD_OT_WRITE],
1880 oti->oti_declare_ops[OSD_OT_PUNCH],
1881 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1882 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1883 oti->oti_declare_ops[OSD_OT_QUOTA],
1884 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1885 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1886 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1887 oti->oti_declare_ops[OSD_OT_INSERT],
1888 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1889 oti->oti_declare_ops_used[OSD_OT_INSERT],
1890 oti->oti_declare_ops[OSD_OT_DELETE],
1891 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1892 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1893 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1894 oti->oti_declare_ops[OSD_OT_REF_ADD],
1895 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1896 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1897 oti->oti_declare_ops[OSD_OT_REF_DEL],
1898 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1899 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1903 * Concurrency: shouldn't matter.
1905 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1908 struct osd_thread_info *oti = osd_oti_get(env);
1909 struct osd_device *dev = osd_dt_dev(d);
1911 struct osd_thandle *oh;
1916 LASSERT(current->journal_info == NULL);
1918 oh = container_of(th, struct osd_thandle, ot_super);
1919 LASSERT(oh != NULL);
1920 LASSERT(oh->ot_handle == NULL);
1922 rc = dt_txn_hook_start(env, d, th);
1926 if (unlikely(osd_param_is_not_sane(dev, th))) {
1927 static unsigned long last_printed;
1928 static int last_credits;
1931 * don't make noise on a tiny testing systems
1932 * actual credits misuse will be caught anyway
1934 if (last_credits != oh->ot_credits &&
1935 time_after(jiffies, last_printed +
1936 cfs_time_seconds(60)) &&
1937 osd_transaction_size(dev) > 512) {
1938 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1939 oh->ot_credits, osd_transaction_size(dev));
1940 osd_trans_dump_creds(env, th);
1941 libcfs_debug_dumpstack(NULL);
1942 last_credits = oh->ot_credits;
1943 last_printed = jiffies;
1946 * XXX Limit the credits to 'max_transaction_buffers', and
1947 * let the underlying filesystem to catch the error if
1948 * we really need so many credits.
1950 * This should be removed when we can calculate the
1951 * credits precisely.
1953 oh->ot_credits = osd_transaction_size(dev);
1954 } else if (ldiskfs_track_declares_assert != 0) {
1956 * reserve few credits to prevent an assertion in JBD
1957 * our debugging mechanism will be able to detected
1958 * overuse. this can help to debug single-update
1961 oh->ot_credits += 10;
1962 if (unlikely(osd_param_is_not_sane(dev, th)))
1963 oh->ot_credits = osd_transaction_size(dev);
1966 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1967 GOTO(out, rc = -EIO);
1970 * XXX temporary stuff. Some abstraction layer should
1973 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1977 LASSERT(oti->oti_txns == 0);
1979 atomic_inc(&dev->od_commit_cb_in_flight);
1980 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1991 static int osd_seq_exists(const struct lu_env *env,
1992 struct osd_device *osd, u64 seq)
1994 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1995 struct seq_server_site *ss = osd_seq_site(osd);
2000 LASSERT(ss != NULL);
2001 LASSERT(ss->ss_server_fld != NULL);
2003 rc = osd_fld_lookup(env, osd, seq, range);
2006 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2007 osd_name(osd), seq, rc);
2011 RETURN(ss->ss_node_id == range->lsr_index);
2014 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2016 struct dt_txn_commit_cb *dcb;
2017 struct dt_txn_commit_cb *tmp;
2019 /* call per-transaction stop callbacks if any */
2020 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2022 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2023 "commit callback entry: magic=%x name='%s'\n",
2024 dcb->dcb_magic, dcb->dcb_name);
2025 list_del_init(&dcb->dcb_linkage);
2026 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2031 * Concurrency: shouldn't matter.
2033 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2036 struct osd_thread_info *oti = osd_oti_get(env);
2037 struct osd_thandle *oh;
2038 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2039 struct osd_device *osd = osd_dt_dev(th->th_dev);
2040 struct qsd_instance *qsd = osd_def_qsd(osd);
2041 struct lquota_trans *qtrans;
2042 LIST_HEAD(truncates);
2043 int rc = 0, remove_agents = 0;
2047 oh = container_of(th, struct osd_thandle, ot_super);
2049 remove_agents = oh->ot_remove_agents;
2051 qtrans = oh->ot_quota_trans;
2052 oh->ot_quota_trans = NULL;
2054 /* move locks to local list, stop tx, execute truncates */
2055 list_splice(&oh->ot_trunc_locks, &truncates);
2057 if (oh->ot_handle != NULL) {
2060 handle_t *hdl = oh->ot_handle;
2063 * add commit callback
2064 * notice we don't do this in osd_trans_start()
2065 * as underlying transaction can change during truncate
2067 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2070 LASSERT(oti->oti_txns == 1);
2073 rc = dt_txn_hook_stop(env, th);
2075 CERROR("%s: failed in transaction hook: rc = %d\n",
2078 osd_trans_stop_cb(oh, rc);
2079 /* hook functions might modify th_sync */
2080 hdl->h_sync = th->th_sync;
2082 oh->ot_handle = NULL;
2083 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2085 CERROR("%s: failed to stop transaction: rc = %d\n",
2086 osd_name(osd), rc2);
2090 osd_process_truncates(env, &truncates);
2092 osd_trans_stop_cb(oh, th->th_result);
2096 osd_trunc_unlock_all(env, &truncates);
2098 /* inform the quota slave device that the transaction is stopping */
2099 qsd_op_end(env, qsd, qtrans);
2102 * as we want IO to journal and data IO be concurrent, we don't block
2103 * awaiting data IO completion in osd_do_bio(), instead we wait here
2104 * once transaction is submitted to the journal. all reqular requests
2105 * don't do direct IO (except read/write), thus this wait_event becomes
2108 * IMPORTANT: we have to wait till any IO submited by the thread is
2109 * completed otherwise iobuf may be corrupted by different request
2111 wait_event(iobuf->dr_wait,
2112 atomic_read(&iobuf->dr_numreqs) == 0);
2113 osd_fini_iobuf(osd, iobuf);
2115 rc = iobuf->dr_error;
2117 if (unlikely(remove_agents != 0))
2118 osd_process_scheduled_agent_removals(env, osd);
2120 LASSERT(oti->oti_ins_cache_depth > 0);
2121 oti->oti_ins_cache_depth--;
2122 /* reset OI cache for safety */
2123 if (oti->oti_ins_cache_depth == 0)
2124 oti->oti_ins_cache_used = 0;
2126 sb_end_write(osd_sb(osd));
2131 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2133 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2136 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2137 LASSERT(&dcb->dcb_func != NULL);
2138 if (dcb->dcb_flags & DCB_TRANS_STOP)
2139 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2141 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2147 * Called just before object is freed. Releases all resources except for
2148 * object itself (that is released by osd_object_free()).
2150 * Concurrency: no concurrent access is possible that late in object
2153 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2155 struct osd_object *obj = osd_obj(l);
2156 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2157 struct inode *inode = obj->oo_inode;
2162 LINVRNT(osd_invariant(obj));
2165 * If object is unlinked remove fid->ino mapping from object index.
2168 osd_index_fini(obj);
2173 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2174 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2176 uid = i_uid_read(inode);
2177 gid = i_gid_read(inode);
2178 projid = i_projid_read(inode);
2180 obj->oo_inode = NULL;
2183 /* do not rebalance quota if the caller needs to release memory
2184 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2185 * transaction and risk to deadlock - LU-12178 */
2186 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2189 if (!obj->oo_header && qsd) {
2190 struct osd_thread_info *info = osd_oti_get(env);
2191 struct lquota_id_info *qi = &info->oti_qi;
2193 /* Release granted quota to master if necessary */
2194 qi->lqi_id.qid_uid = uid;
2195 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2197 qi->lqi_id.qid_uid = gid;
2198 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2200 qi->lqi_id.qid_uid = projid;
2201 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2206 * Concurrency: ->loo_object_release() is called under site spin-lock.
2208 static void osd_object_release(const struct lu_env *env,
2209 struct lu_object *l)
2211 struct osd_object *o = osd_obj(l);
2214 * nobody should be releasing a non-destroyed object with nlink=0
2215 * the API allows this, but ldiskfs doesn't like and then report
2216 * this inode as deleted
2218 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2219 o->oo_inode->i_nlink == 0));
2223 * Concurrency: shouldn't matter.
2225 static int osd_object_print(const struct lu_env *env, void *cookie,
2226 lu_printer_t p, const struct lu_object *l)
2228 struct osd_object *o = osd_obj(l);
2229 struct iam_descr *d;
2231 if (o->oo_dir != NULL)
2232 d = o->oo_dir->od_container.ic_descr;
2235 return (*p)(env, cookie,
2236 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2238 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2239 o->oo_inode ? o->oo_inode->i_generation : 0,
2240 d ? d->id_ops->id_name : "plain");
2244 * Concurrency: shouldn't matter.
2246 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2247 struct obd_statfs *sfs, struct obd_statfs_info *info)
2249 struct osd_device *osd = osd_dt_dev(d);
2250 struct super_block *sb = osd_sb(osd);
2251 struct kstatfs *ksfs;
2255 if (unlikely(osd->od_mnt == NULL))
2256 return -EINPROGRESS;
2258 /* osd_lproc.c call this without env, allocate ksfs for that case */
2259 if (unlikely(env == NULL)) {
2260 OBD_ALLOC_PTR(ksfs);
2264 ksfs = &osd_oti_get(env)->oti_ksfs;
2267 result = sb->s_op->statfs(sb->s_root, ksfs);
2271 statfs_pack(sfs, ksfs);
2272 if (unlikely(sb->s_flags & SB_RDONLY))
2273 sfs->os_state |= OS_STATFS_READONLY;
2275 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2277 if (ldiskfs_has_feature_extents(sb))
2278 sfs->os_maxbytes = sb->s_maxbytes;
2280 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2283 * Reserve some space so to avoid fragmenting the filesystem too much.
2284 * Fragmentation not only impacts performance, but can also increase
2285 * metadata overhead significantly, causing grant calculation to be
2288 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2290 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2291 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2292 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2293 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2295 sfs->os_blocks -= reserved;
2296 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2297 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2300 if (unlikely(env == NULL))
2306 * Estimate space needed for file creations. We assume the largest filename
2307 * which is 2^64 - 1, hence a filename of 20 chars.
2308 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2310 #ifdef __LDISKFS_DIR_REC_LEN
2311 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2313 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2317 * Concurrency: doesn't access mutable data.
2319 static void osd_conf_get(const struct lu_env *env,
2320 const struct dt_device *dev,
2321 struct dt_device_param *param)
2323 struct osd_device *d = osd_dt_dev(dev);
2324 struct super_block *sb = osd_sb(d);
2325 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2330 * XXX should be taken from not-yet-existing fs abstraction layer.
2332 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2333 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2334 param->ddp_symlink_max = sb->s_blocksize;
2335 param->ddp_mount_type = LDD_MT_LDISKFS;
2336 if (ldiskfs_has_feature_extents(sb))
2337 param->ddp_maxbytes = sb->s_maxbytes;
2339 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2341 * inode are statically allocated, so per-inode space consumption
2342 * is the space consumed by the directory entry
2344 param->ddp_inodespace = PER_OBJ_USAGE;
2346 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2347 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2348 * maximum length to not under-count the actual number of extents
2349 * needed for writing a file if there are sub-optimal block allocations.
2351 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2352 /* worst-case extent insertion metadata overhead */
2353 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2354 param->ddp_mntopts = 0;
2355 if (test_opt(sb, XATTR_USER))
2356 param->ddp_mntopts |= MNTOPT_USERXATTR;
2357 if (test_opt(sb, POSIX_ACL))
2358 param->ddp_mntopts |= MNTOPT_ACL;
2361 * LOD might calculate the max stripe count based on max_ea_size,
2362 * so we need take account in the overhead as well,
2363 * xattr_header + magic + xattr_entry_head
2365 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2366 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2368 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2369 if (ldiskfs_has_feature_ea_inode(sb))
2370 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2374 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2376 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2377 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2380 * Preferred RPC size for efficient disk IO. 4MB shows good
2381 * all-around performance for ldiskfs, but use bigalloc chunk size
2382 * by default if larger.
2384 #if defined(LDISKFS_CLUSTER_SIZE)
2385 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2386 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2389 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2391 param->ddp_t10_cksum_type = 0;
2393 unsigned short interval = blk_integrity_interval(bi);
2394 name = blk_integrity_name(bi);
2402 if (strncmp(name, "T10-DIF-TYPE",
2403 sizeof("T10-DIF-TYPE") - 1) == 0) {
2404 /* also skip "1/3-" at end */
2405 const int type_off = sizeof("T10-DIF-TYPE.");
2406 char type_number = name[type_off - 2];
2408 if (interval != 512 && interval != 4096) {
2409 CERROR("%s: unsupported T10PI sector size %u\n",
2410 d->od_svname, interval);
2411 } else if (type_number != '1' && type_number != '3') {
2412 CERROR("%s: unsupported T10PI type %s\n",
2413 d->od_svname, name);
2414 } else if (strcmp(name + type_off, "CRC") == 0) {
2415 d->od_t10_type = type_number == '1' ?
2416 OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2417 param->ddp_t10_cksum_type = interval == 512 ?
2418 OBD_CKSUM_T10CRC512 :
2420 } else if (strcmp(name + type_off, "IP") == 0) {
2421 d->od_t10_type = type_number == '1' ?
2422 OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2423 param->ddp_t10_cksum_type = interval == 512 ?
2424 OBD_CKSUM_T10IP512 :
2427 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2428 d->od_svname, name);
2432 CERROR("%s: unsupported T10PI type '%s'\n",
2433 d->od_svname, name);
2437 param->ddp_has_lseek_data_hole = true;
2440 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2442 return osd_sb(osd_dt_dev(d));
2446 * Concurrency: shouldn't matter.
2448 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2451 struct super_block *s = osd_sb(osd_dt_dev(d));
2454 down_read(&s->s_umount);
2455 rc = s->s_op->sync_fs(s, 1);
2456 up_read(&s->s_umount);
2458 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2465 * Start commit for OSD device.
2467 * An implementation of dt_commit_async method for OSD device.
2468 * Asychronously starts underlayng fs sync and thereby a transaction
2471 * \param env environment
2472 * \param d dt device
2474 * \see dt_device_operations
2476 static int osd_commit_async(const struct lu_env *env,
2477 struct dt_device *d)
2479 struct super_block *s = osd_sb(osd_dt_dev(d));
2484 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2485 down_read(&s->s_umount);
2486 rc = s->s_op->sync_fs(s, 0);
2487 up_read(&s->s_umount);
2492 static int (*priv_security_file_alloc)(struct file *file);
2494 int osd_security_file_alloc(struct file *file)
2496 if (priv_security_file_alloc)
2497 return priv_security_file_alloc(file);
2502 * Concurrency: shouldn't matter.
2504 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2506 struct super_block *sb = osd_sb(osd_dt_dev(d));
2507 struct block_device *dev = sb->s_bdev;
2508 int rc = -EOPNOTSUPP;
2512 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2513 osd_dt_dev(d)->od_svname, (long)dev, rc);
2519 * Note: we do not count into QUOTA here.
2520 * If we mount with --data_journal we may need more.
2522 const int osd_dto_credits_noquota[DTO_NR] = {
2525 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2526 * SINGLEDATA_TRANS_BLOCKS(8)
2527 * XXX Note: maybe iam need more, since iam have more level than
2530 [DTO_INDEX_INSERT] = 16,
2533 * just modify a single entry, probably merge few within a block
2535 [DTO_INDEX_DELETE] = 1,
2539 [DTO_INDEX_UPDATE] = 16,
2541 * 4(inode, inode bits, groups, GDT)
2542 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2544 [DTO_OBJECT_CREATE] = 4,
2546 * 4(inode, inode bits, groups, GDT)
2547 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2549 [DTO_OBJECT_DELETE] = 4,
2551 * Attr set credits (inode)
2553 [DTO_ATTR_SET_BASE] = 1,
2555 * Xattr set. The same as xattr of EXT3.
2556 * DATA_TRANS_BLOCKS(14)
2557 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2558 * are also counted in. Do not know why?
2560 [DTO_XATTR_SET] = 14,
2562 * credits for inode change during write.
2564 [DTO_WRITE_BASE] = 3,
2566 * credits for single block write.
2568 [DTO_WRITE_BLOCK] = 14,
2570 * Attr set credits for chown.
2571 * This is extra credits for setattr, and it is null without quota
2573 [DTO_ATTR_SET_CHOWN] = 0
2576 static const struct dt_device_operations osd_dt_ops = {
2577 .dt_root_get = osd_root_get,
2578 .dt_statfs = osd_statfs,
2579 .dt_trans_create = osd_trans_create,
2580 .dt_trans_start = osd_trans_start,
2581 .dt_trans_stop = osd_trans_stop,
2582 .dt_trans_cb_add = osd_trans_cb_add,
2583 .dt_conf_get = osd_conf_get,
2584 .dt_mnt_sb_get = osd_mnt_sb_get,
2585 .dt_sync = osd_sync,
2587 .dt_commit_async = osd_commit_async,
2590 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2593 struct osd_object *obj = osd_dt_obj(dt);
2594 struct osd_thread_info *oti = osd_oti_get(env);
2596 LINVRNT(osd_invariant(obj));
2598 LASSERT(obj->oo_owner != env);
2599 down_read_nested(&obj->oo_sem, role);
2601 LASSERT(obj->oo_owner == NULL);
2605 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2608 struct osd_object *obj = osd_dt_obj(dt);
2609 struct osd_thread_info *oti = osd_oti_get(env);
2611 LINVRNT(osd_invariant(obj));
2613 LASSERT(obj->oo_owner != env);
2614 down_write_nested(&obj->oo_sem, role);
2616 LASSERT(obj->oo_owner == NULL);
2617 obj->oo_owner = env;
2621 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2623 struct osd_object *obj = osd_dt_obj(dt);
2624 struct osd_thread_info *oti = osd_oti_get(env);
2626 LINVRNT(osd_invariant(obj));
2628 LASSERT(oti->oti_r_locks > 0);
2630 up_read(&obj->oo_sem);
2633 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2635 struct osd_object *obj = osd_dt_obj(dt);
2636 struct osd_thread_info *oti = osd_oti_get(env);
2638 LINVRNT(osd_invariant(obj));
2640 LASSERT(obj->oo_owner == env);
2641 LASSERT(oti->oti_w_locks > 0);
2643 obj->oo_owner = NULL;
2644 up_write(&obj->oo_sem);
2647 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2649 struct osd_object *obj = osd_dt_obj(dt);
2651 LINVRNT(osd_invariant(obj));
2653 return obj->oo_owner == env;
2656 static void osd_inode_getattr(const struct lu_env *env,
2657 struct inode *inode, struct lu_attr *attr)
2659 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2660 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2661 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2662 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2664 attr->la_atime = inode->i_atime.tv_sec;
2665 attr->la_mtime = inode->i_mtime.tv_sec;
2666 attr->la_ctime = inode->i_ctime.tv_sec;
2667 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2668 attr->la_mode = inode->i_mode;
2669 attr->la_size = i_size_read(inode);
2670 attr->la_blocks = inode->i_blocks;
2671 attr->la_uid = i_uid_read(inode);
2672 attr->la_gid = i_gid_read(inode);
2673 attr->la_projid = i_projid_read(inode);
2674 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2675 attr->la_nlink = inode->i_nlink;
2676 attr->la_rdev = inode->i_rdev;
2677 attr->la_blksize = 1 << inode->i_blkbits;
2678 attr->la_blkbits = inode->i_blkbits;
2680 * Ext4 did not transfer inherit flags from raw inode
2681 * to inode flags, and ext4 internally test raw inode
2682 * @i_flags directly. Instead of patching ext4, we do it here.
2684 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2685 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2688 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2691 struct osd_object *obj = osd_dt_obj(dt);
2692 const struct dt_it_ops *iops;
2698 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2699 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2701 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2702 *count = obj->oo_dirent_count;
2706 /* directory not initialized yet */
2707 if (!dt->do_index_ops) {
2712 iops = &dt->do_index_ops->dio_it;
2713 it = iops->init(env, dt, LUDA_64BITHASH);
2715 RETURN(PTR_ERR(it));
2717 rc = iops->load(env, it, 0);
2719 if (rc == -ENODATA) {
2726 rc = iops->next(env, it);
2728 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2732 if (iops->key_size(env, it) == 0)
2738 obj->oo_dirent_count = *count;
2743 iops->fini(env, it);
2748 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2749 struct lu_attr *attr)
2751 struct osd_object *obj = osd_dt_obj(dt);
2754 if (unlikely(!dt_object_exists(dt)))
2756 if (unlikely(obj->oo_destroyed))
2759 LASSERT(!dt_object_remote(dt));
2760 LINVRNT(osd_invariant(obj));
2762 spin_lock(&obj->oo_guard);
2763 osd_inode_getattr(env, obj->oo_inode, attr);
2764 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2765 attr->la_valid |= LA_FLAGS;
2766 attr->la_flags |= LUSTRE_ORPHAN_FL;
2768 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2769 attr->la_valid |= LA_FLAGS;
2770 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2772 spin_unlock(&obj->oo_guard);
2774 if (S_ISDIR(obj->oo_inode->i_mode) &&
2775 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2776 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2781 static int osd_declare_attr_qid(const struct lu_env *env,
2782 struct osd_object *obj,
2783 struct osd_thandle *oh, long long bspace,
2784 qid_t old_id, qid_t new_id, bool enforce,
2785 unsigned int type, bool ignore_edquot)
2788 struct osd_thread_info *info = osd_oti_get(env);
2789 struct lquota_id_info *qi = &info->oti_qi;
2791 qi->lqi_type = type;
2792 /* inode accounting */
2793 qi->lqi_is_blk = false;
2795 /* one more inode for the new id ... */
2796 qi->lqi_id.qid_uid = new_id;
2798 /* Reserve credits for the new id */
2799 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2800 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2805 /* and one less inode for the current id */
2806 qi->lqi_id.qid_uid = old_id;
2808 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2809 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2814 /* block accounting */
2815 qi->lqi_is_blk = true;
2817 /* more blocks for the new id ... */
2818 qi->lqi_id.qid_uid = new_id;
2819 qi->lqi_space = bspace;
2821 * Credits for the new uid has been reserved, re-use "obj"
2822 * to save credit reservation.
2824 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2825 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2830 /* and finally less blocks for the current uid */
2831 qi->lqi_id.qid_uid = old_id;
2832 qi->lqi_space = -bspace;
2833 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2834 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2840 static int osd_declare_attr_set(const struct lu_env *env,
2841 struct dt_object *dt,
2842 const struct lu_attr *attr,
2843 struct thandle *handle)
2845 struct osd_thandle *oh;
2846 struct osd_object *obj;
2855 LASSERT(dt != NULL);
2856 LASSERT(handle != NULL);
2858 obj = osd_dt_obj(dt);
2859 LASSERT(osd_invariant(obj));
2861 oh = container_of(handle, struct osd_thandle, ot_super);
2862 LASSERT(oh->ot_handle == NULL);
2864 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2865 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2867 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2868 osd_dto_credits_noquota[DTO_XATTR_SET]);
2870 if (attr == NULL || obj->oo_inode == NULL)
2873 bspace = obj->oo_inode->i_blocks << 9;
2874 bspace = toqb(bspace);
2877 * Changing ownership is always preformed by super user, it should not
2878 * fail with EDQUOT unless required explicitly.
2880 * We still need to call the osd_declare_qid() to calculate the journal
2881 * credits for updating quota accounting files and to trigger quota
2882 * space adjustment once the operation is completed.
2884 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2885 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2889 "%s: enforce quota on UID %u, GID %u (quota space is %lld)\n",
2890 osd_ino2name(obj->oo_inode), attr->la_uid,
2891 attr->la_gid, bspace);
2894 uid = i_uid_read(obj->oo_inode);
2895 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2896 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2897 attr->la_uid, enforce, USRQUOTA,
2902 gid = i_gid_read(obj->oo_inode);
2903 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
2904 attr->la_uid, gid, attr->la_gid);
2905 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2906 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
2907 attr->la_gid, enforce, GRPQUOTA,
2913 #ifdef HAVE_PROJECT_QUOTA
2914 if (attr->la_valid & LA_PROJID) {
2915 __u32 projid = i_projid_read(obj->oo_inode);
2917 enforce = (attr->la_valid & LA_PROJID) &&
2918 (attr->la_projid != projid);
2919 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2920 (qid_t)projid, (qid_t)attr->la_projid,
2921 enforce, PRJQUOTA, true);
2929 static int osd_inode_setattr(const struct lu_env *env,
2930 struct inode *inode, const struct lu_attr *attr)
2932 __u64 bits = attr->la_valid;
2934 /* Only allow set size for regular file */
2935 if (!S_ISREG(inode->i_mode))
2936 bits &= ~(LA_SIZE | LA_BLOCKS);
2941 if (bits & LA_ATIME)
2942 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2943 if (bits & LA_CTIME)
2944 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2945 if (bits & LA_MTIME)
2946 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2947 if (bits & LA_SIZE) {
2948 spin_lock(&inode->i_lock);
2949 LDISKFS_I(inode)->i_disksize = attr->la_size;
2950 i_size_write(inode, attr->la_size);
2951 spin_unlock(&inode->i_lock);
2955 * OSD should not change "i_blocks" which is used by quota.
2956 * "i_blocks" should be changed by ldiskfs only.
2959 inode->i_mode = (inode->i_mode & S_IFMT) |
2960 (attr->la_mode & ~S_IFMT);
2962 i_uid_write(inode, attr->la_uid);
2964 i_gid_write(inode, attr->la_gid);
2965 if (bits & LA_PROJID)
2966 i_projid_write(inode, attr->la_projid);
2967 if (bits & LA_NLINK)
2968 set_nlink(inode, attr->la_nlink);
2970 inode->i_rdev = attr->la_rdev;
2972 if (bits & LA_FLAGS) {
2973 /* always keep S_NOCMTIME */
2974 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2976 #if defined(S_ENCRYPTED)
2977 /* Always remove S_ENCRYPTED, because ldiskfs must not be
2978 * aware of encryption status. It is just stored into LMA
2979 * so that it can be forwared to client side.
2981 inode->i_flags &= ~S_ENCRYPTED;
2984 * Ext4 did not transfer inherit flags from
2985 * @inode->i_flags to raw inode i_flags when writing
2986 * flags, we do it explictly here.
2988 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2989 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2991 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2996 #ifdef HAVE_PROJECT_QUOTA
2997 static int osd_transfer_project(struct inode *inode, __u32 projid,
2998 struct thandle *handle)
3000 struct super_block *sb = inode->i_sb;
3001 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3004 struct ldiskfs_iloc iloc;
3005 struct ldiskfs_inode *raw_inode;
3006 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3008 if (!ldiskfs_has_feature_project(sb)) {
3009 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3010 == LDISKFS_DEF_PROJID);
3011 if (projid != LDISKFS_DEF_PROJID)
3017 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3020 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3021 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3024 err = ldiskfs_get_inode_loc(inode, &iloc);
3028 raw_inode = ldiskfs_raw_inode(&iloc);
3029 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3030 struct osd_thandle *oh = container_of(handle,
3034 * try to expand inode size automatically.
3036 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3037 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3045 dquot_initialize(inode);
3046 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3047 if (transfer_to[PRJQUOTA]) {
3048 err = __dquot_transfer(inode, transfer_to);
3049 dqput(transfer_to[PRJQUOTA]);
3058 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3059 struct thandle *handle)
3063 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3064 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3068 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3069 inode->i_ino, i_uid_read(inode), attr->la_uid,
3070 i_gid_read(inode), attr->la_gid);
3072 dquot_initialize(inode);
3074 if (attr->la_valid & LA_UID)
3075 iattr.ia_valid |= ATTR_UID;
3076 if (attr->la_valid & LA_GID)
3077 iattr.ia_valid |= ATTR_GID;
3078 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3079 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3081 rc = dquot_transfer(inode, &iattr);
3083 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3084 osd_ino2name(inode), rc);
3089 /* Handle project id transfer here properly */
3090 if (attr->la_valid & LA_PROJID &&
3091 attr->la_projid != i_projid_read(inode)) {
3092 #ifdef HAVE_PROJECT_QUOTA
3093 rc = osd_transfer_project(inode, attr->la_projid, handle);
3098 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3099 osd_ino2name(inode), rc);
3106 static int osd_attr_set(const struct lu_env *env,
3107 struct dt_object *dt,
3108 const struct lu_attr *attr,
3109 struct thandle *handle)
3111 struct osd_object *obj = osd_dt_obj(dt);
3112 struct inode *inode;
3115 if (!dt_object_exists(dt))
3118 LASSERT(handle != NULL);
3119 LASSERT(!dt_object_remote(dt));
3120 LASSERT(osd_invariant(obj));
3122 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3124 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3125 !osd_obj2dev(obj)->od_is_ost) {
3126 struct osd_thread_info *oti = osd_oti_get(env);
3127 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3128 struct lu_fid *fid1 = &oti->oti_fid;
3129 struct osd_inode_id *id = &oti->oti_id;
3130 struct iam_path_descr *ipd;
3131 struct iam_container *bag;
3132 struct osd_thandle *oh;
3135 fid_cpu_to_be(fid1, fid0);
3136 memset(id, 1, sizeof(*id));
3137 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3138 fid0)->oi_dir.od_container;
3139 ipd = osd_idx_ipd_get(env, bag);
3140 if (unlikely(ipd == NULL))
3143 oh = container_of(handle, struct osd_thandle, ot_super);
3144 rc = iam_update(oh->ot_handle, bag,
3145 (const struct iam_key *)fid1,
3146 (const struct iam_rec *)id, ipd);
3147 osd_ipd_put(env, bag, ipd);
3148 return(rc > 0 ? 0 : rc);
3151 inode = obj->oo_inode;
3153 rc = osd_quota_transfer(inode, attr, handle);
3157 spin_lock(&obj->oo_guard);
3158 rc = osd_inode_setattr(env, inode, attr);
3159 spin_unlock(&obj->oo_guard);
3163 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3165 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3167 if (!(attr->la_valid & LA_FLAGS))
3170 /* Let's check if there are extra flags need to be set into LMA */
3171 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3172 struct osd_thread_info *info = osd_oti_get(env);
3173 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3175 LASSERT(!obj->oo_pfid_in_lma);
3177 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3178 &info->oti_ost_attrs);
3182 lma->lma_incompat |=
3183 lustre_to_lma_flags(attr->la_flags);
3184 lustre_lma_swab(lma);
3186 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3188 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3189 lma, sizeof(*lma), XATTR_REPLACE);
3191 struct osd_device *osd = osd_obj2dev(obj);
3193 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3194 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3195 lma->lma_incompat, rc);
3198 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3200 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3207 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3208 struct osd_object *obj,
3209 const char *name, const int namelen)
3211 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3214 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3215 umode_t mode, struct dt_allocation_hint *hint,
3216 struct thandle *th, struct lu_attr *attr)
3219 struct osd_device *osd = osd_obj2dev(obj);
3220 struct osd_thandle *oth;
3221 struct dt_object *parent = NULL;
3222 struct inode *inode;
3223 struct iattr iattr = {
3224 .ia_valid = ATTR_UID | ATTR_GID |
3225 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3226 .ia_ctime.tv_sec = attr->la_ctime,
3227 .ia_mtime.tv_sec = attr->la_mtime,
3228 .ia_atime.tv_sec = attr->la_atime,
3229 .ia_uid = GLOBAL_ROOT_UID,
3230 .ia_gid = GLOBAL_ROOT_GID,
3232 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3234 if (attr->la_valid & LA_UID)
3235 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3236 if (attr->la_valid & LA_GID)
3237 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3239 LINVRNT(osd_invariant(obj));
3240 LASSERT(obj->oo_inode == NULL);
3241 LASSERT(obj->oo_hl_head == NULL);
3243 if (S_ISDIR(mode) && ldiskfs_pdo) {
3245 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3246 if (obj->oo_hl_head == NULL)
3250 oth = container_of(th, struct osd_thandle, ot_super);
3251 LASSERT(oth->ot_handle->h_transaction != NULL);
3253 if (hint != NULL && hint->dah_parent != NULL &&
3254 !dt_object_remote(hint->dah_parent))
3255 parent = hint->dah_parent;
3257 /* if a time component is not valid set it to UTIME_OMIT */
3258 if (!(attr->la_valid & LA_CTIME))
3259 iattr.ia_ctime = omit;
3260 if (!(attr->la_valid & LA_MTIME))
3261 iattr.ia_mtime = omit;
3262 if (!(attr->la_valid & LA_ATIME))
3263 iattr.ia_atime = omit;
3265 inode = ldiskfs_create_inode(oth->ot_handle,
3266 parent ? osd_dt_obj(parent)->oo_inode :
3267 osd_sb(osd)->s_root->d_inode,
3269 if (!IS_ERR(inode)) {
3270 /* Do not update file c/mtime in ldiskfs. */
3271 inode->i_flags |= S_NOCMTIME;
3274 * For new created object, it must be consistent,
3275 * and it is unnecessary to scrub against it.
3277 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3279 obj->oo_inode = inode;
3282 if (obj->oo_hl_head != NULL) {
3283 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3284 obj->oo_hl_head = NULL;
3286 result = PTR_ERR(inode);
3288 LINVRNT(osd_invariant(obj));
3296 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3297 struct lu_attr *attr,
3298 struct dt_allocation_hint *hint,
3299 struct dt_object_format *dof,
3303 struct osd_thandle *oth;
3304 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3306 LASSERT(S_ISDIR(attr->la_mode));
3308 oth = container_of(th, struct osd_thandle, ot_super);
3309 LASSERT(oth->ot_handle->h_transaction != NULL);
3310 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3311 obj->oo_dirent_count = 0;
3312 result = osd_mkfile(info, obj, mode, hint, th, attr);
3317 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3318 struct lu_attr *attr,
3319 struct dt_allocation_hint *hint,
3320 struct dt_object_format *dof,
3324 struct osd_thandle *oth;
3325 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3327 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3329 LASSERT(S_ISREG(attr->la_mode));
3331 oth = container_of(th, struct osd_thandle, ot_super);
3332 LASSERT(oth->ot_handle->h_transaction != NULL);
3334 result = osd_mkfile(info, obj, mode, hint, th, attr);
3336 LASSERT(obj->oo_inode != NULL);
3337 if (feat->dif_flags & DT_IND_VARKEY)
3338 result = iam_lvar_create(obj->oo_inode,
3339 feat->dif_keysize_max,
3341 feat->dif_recsize_max,
3344 result = iam_lfix_create(obj->oo_inode,
3345 feat->dif_keysize_max,
3347 feat->dif_recsize_max,
3353 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3354 struct lu_attr *attr,
3355 struct dt_allocation_hint *hint,
3356 struct dt_object_format *dof,
3359 LASSERT(S_ISREG(attr->la_mode));
3360 return osd_mkfile(info, obj, (attr->la_mode &
3361 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3365 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3366 struct lu_attr *attr,
3367 struct dt_allocation_hint *hint,
3368 struct dt_object_format *dof,
3371 LASSERT(S_ISLNK(attr->la_mode));
3372 return osd_mkfile(info, obj, (attr->la_mode &
3373 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3377 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3378 struct lu_attr *attr,
3379 struct dt_allocation_hint *hint,
3380 struct dt_object_format *dof,
3383 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3386 LINVRNT(osd_invariant(obj));
3387 LASSERT(obj->oo_inode == NULL);
3388 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3389 S_ISFIFO(mode) || S_ISSOCK(mode));
3391 result = osd_mkfile(info, obj, mode, hint, th, attr);
3393 LASSERT(obj->oo_inode != NULL);
3395 * This inode should be marked dirty for i_rdev. Currently
3396 * that is done in the osd_attr_init().
3398 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3401 LINVRNT(osd_invariant(obj));
3405 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3407 struct dt_allocation_hint *hint,
3408 struct dt_object_format *dof,
3411 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3413 osd_obj_type_f result;
3429 result = osd_mk_index;
3440 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3441 struct dt_object *parent, struct dt_object *child,
3446 ah->dah_parent = parent;
3447 ah->dah_mode = child_mode;
3449 if (parent != NULL && !dt_object_remote(parent)) {
3450 /* will help to find FID->ino at dt_insert("..") */
3451 struct osd_object *pobj = osd_dt_obj(parent);
3453 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3457 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3458 struct lu_attr *attr, struct dt_object_format *dof,
3459 struct thandle *handle)
3461 struct inode *inode = obj->oo_inode;
3462 __u64 valid = attr->la_valid;
3465 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3467 if (dof->dof_type != DFT_NODE)
3468 attr->la_valid &= ~LA_RDEV;
3469 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3470 attr->la_valid &= ~LA_ATIME;
3471 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3472 attr->la_valid &= ~LA_CTIME;
3473 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3474 attr->la_valid &= ~LA_MTIME;
3476 result = osd_quota_transfer(inode, attr, handle);
3480 if (attr->la_valid != 0) {
3481 result = osd_inode_setattr(info->oti_env, inode, attr);
3483 * The osd_inode_setattr() should always succeed here. The
3484 * only error that could be returned is EDQUOT when we are
3485 * trying to change the UID or GID of the inode. However, this
3486 * should not happen since quota enforcement is no longer
3487 * enabled on ldiskfs (lquota takes care of it).
3489 LASSERTF(result == 0, "%d\n", result);
3490 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3493 attr->la_valid = valid;
3497 * Helper function for osd_create()
3499 * \retval 0, on success
3501 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3502 struct lu_attr *attr, struct dt_allocation_hint *hint,
3503 struct dt_object_format *dof, struct thandle *th)
3508 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3510 /* we drop umask so that permissions we pass are not affected */
3511 umask = current->fs->umask;
3512 current->fs->umask = 0;
3514 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3516 if (likely(obj->oo_inode != NULL)) {
3517 LASSERT(obj->oo_inode->i_state & I_NEW);
3520 * Unlock the inode before attr initialization to avoid
3521 * unnecessary dqget operations. LU-6378
3523 unlock_new_inode(obj->oo_inode);
3526 if (likely(result == 0)) {
3527 osd_attr_init(info, obj, attr, dof, th);
3528 osd_object_init0(obj);
3531 /* restore previous umask value */
3532 current->fs->umask = umask;
3534 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3540 * Helper function for osd_create()
3542 * \retval 0, on success
3544 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3545 const struct lu_fid *fid, struct thandle *th)
3547 struct osd_thread_info *info = osd_oti_get(env);
3548 struct osd_inode_id *id = &info->oti_id;
3549 struct osd_device *osd = osd_obj2dev(obj);
3550 struct osd_thandle *oh;
3553 LASSERT(obj->oo_inode != NULL);
3555 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3558 oh = container_of(th, struct osd_thandle, ot_super);
3559 LASSERT(oh->ot_handle);
3560 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3562 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3563 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3564 OI_CHECK_FLD, NULL);
3565 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3566 struct lu_fid next_fid = *fid;
3568 /* insert next object in advance, and map to the same inode */
3570 if (next_fid.f_oid != 0) {
3571 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3572 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3573 OI_CHECK_FLD, NULL);
3574 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3578 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3583 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3584 u64 seq, struct lu_seq_range *range)
3586 struct seq_server_site *ss = osd_seq_site(osd);
3588 if (fid_seq_is_idif(seq)) {
3589 fld_range_set_ost(range);
3590 range->lsr_index = idif_ost_idx(seq);
3594 if (!fid_seq_in_fldb(seq)) {
3595 fld_range_set_mdt(range);
3598 * FIXME: If ss is NULL, it suppose not get lsr_index
3601 range->lsr_index = ss->ss_node_id;
3605 LASSERT(ss != NULL);
3606 fld_range_set_any(range);
3607 /* OSD will only do local fld lookup */
3608 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3611 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3612 struct lu_attr *attr,
3613 struct dt_allocation_hint *hint,
3614 struct dt_object_format *dof,
3615 struct thandle *handle)
3617 struct osd_thandle *oh;
3622 LASSERT(handle != NULL);
3624 oh = container_of(handle, struct osd_thandle, ot_super);
3625 LASSERT(oh->ot_handle == NULL);
3628 * EA object consumes more credits than regular object: osd_mk_index
3629 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3630 * leaf_node, could involves the block, block bitmap, groups, GDT
3631 * change for each block, so add 4 * 2 credits in that case.
3633 osd_trans_declare_op(env, oh, OSD_OT_CREATE,
3634 osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3635 (dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3637 * Reuse idle OI block may cause additional one OI block
3640 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3641 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3642 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3643 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3644 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3646 /* will help to find FID->ino mapping at dt_insert() */
3647 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3655 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3656 attr->la_projid, 1, oh, osd_dt_obj(dt),
3657 NULL, OSD_QID_INODE);
3665 * Called to destroy on-disk representation of the object
3667 * Concurrency: must be locked
3669 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3672 struct osd_object *obj = osd_dt_obj(dt);
3673 struct inode *inode = obj->oo_inode;
3674 struct osd_thandle *oh;
3682 oh = container_of(th, struct osd_thandle, ot_super);
3683 LASSERT(oh->ot_handle == NULL);
3685 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3686 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3688 /* For removing agent entry */
3689 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3690 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3693 * Recycle idle OI leaf may cause additional three OI blocks
3696 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3697 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3698 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3699 /* one less inode */
3700 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3701 i_projid_read(inode), -1, oh, obj, NULL,
3705 /* data to be truncated */
3706 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3707 i_projid_read(inode), 0, oh, obj, NULL,
3713 * will help to find FID->ino when this object is being
3716 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3721 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3724 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3725 struct osd_object *obj = osd_dt_obj(dt);
3726 struct inode *inode = obj->oo_inode;
3727 struct osd_device *osd = osd_obj2dev(obj);
3728 struct osd_thandle *oh;
3733 oh = container_of(th, struct osd_thandle, ot_super);
3734 LASSERT(oh->ot_handle);
3736 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3738 if (unlikely(fid_is_acct(fid)))
3741 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3742 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3744 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3745 osd_name(osd), PFID(fid), result);
3748 if (S_ISDIR(inode->i_mode)) {
3749 if (inode->i_nlink > 2)
3750 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3751 osd_name(osd), PFID(fid), inode->i_ino,
3754 spin_lock(&obj->oo_guard);
3756 spin_unlock(&obj->oo_guard);
3757 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3760 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3762 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3764 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3765 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3766 oh->ot_handle, OI_CHECK_FLD);
3768 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3769 /* XXX: add to ext3 orphan list */
3770 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3772 /* not needed in the cache anymore */
3773 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3774 obj->oo_destroyed = 1;
3780 * Put the fid into lustre_mdt_attrs, and then place the structure
3781 * inode's ea. This fid should not be altered during the life time
3784 * \retval +ve, on success
3785 * \retval -ve, on error
3787 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3789 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3790 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3792 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3793 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3798 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3801 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3804 lustre_loa_init(loa, fid, compat, incompat);
3805 lustre_loa_swab(loa, false);
3808 * For the OST device with 256 bytes inode size by default,
3809 * the PFID EA will be stored together with LMA EA to avoid
3810 * performance trouble. Otherwise the PFID EA can be stored
3811 * independently. LU-8998
3813 if ((compat & LMAC_FID_ON_OST) &&
3814 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3815 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3816 sizeof(*loa), XATTR_CREATE);
3818 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3819 sizeof(*lma), XATTR_CREATE);
3821 * LMA may already exist, but we need to check that all the
3822 * desired compat/incompat flags have been added.
3824 if (unlikely(rc == -EEXIST)) {
3825 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3826 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3830 if (rc < sizeof(*lma))
3833 lustre_loa_swab(loa, true);
3834 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3835 ((compat == 0 && incompat == 0) ||
3836 (!(~lma->lma_compat & compat) &&
3837 !(~lma->lma_incompat & incompat))))
3840 lma->lma_self_fid = *fid;
3841 lma->lma_compat |= compat;
3842 lma->lma_incompat |= incompat;
3843 if (rc == sizeof(*lma)) {
3844 lustre_lma_swab(lma);
3845 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3846 sizeof(*lma), XATTR_REPLACE);
3848 lustre_loa_swab(loa, false);
3849 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3850 sizeof(*loa), XATTR_REPLACE);
3858 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3859 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3860 * To have compatilibility with 1.8 ldiskfs driver we need to have
3861 * magic number at start of fid data.
3862 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3865 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3866 const struct lu_fid *fid)
3868 if (!fid_is_namespace_visible(fid) ||
3869 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3870 param->edp_magic = 0;
3874 param->edp_magic = LDISKFS_LUFID_MAGIC;
3875 param->edp_len = sizeof(struct lu_fid) + 1;
3876 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3880 * Try to read the fid from inode ea into dt_rec.
3882 * \param fid object fid.
3884 * \retval 0 on success
3886 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3887 __u32 ino, struct lu_fid *fid,
3888 struct osd_inode_id *id)
3890 struct osd_thread_info *info = osd_oti_get(env);
3891 struct inode *inode;
3895 osd_id_gen(id, ino, OSD_OII_NOGEN);
3896 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3898 RETURN(PTR_ERR(inode));
3904 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3906 struct inode *parent_dir,
3907 const struct lu_fid *dot_fid,
3908 const struct lu_fid *dot_dot_fid,
3909 struct osd_thandle *oth)
3911 struct ldiskfs_dentry_param *dot_ldp;
3912 struct ldiskfs_dentry_param *dot_dot_ldp;
3913 __u32 saved_nlink = dir->i_nlink;
3916 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
3919 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3920 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3922 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3923 dot_ldp->edp_magic = 0;
3925 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3926 dir, dot_ldp, dot_dot_ldp);
3928 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3929 * the subseqent ref_add() will increase the dir->i_nlink
3930 * as 3. That is incorrect for new created directory.
3932 * It looks like hack, because we want to make the OSD API
3933 * to be order-independent for new created directory object
3934 * between dt_insert(..) and ref_add() operations.
3936 * Here, we only restore the in-RAM dir-inode's nlink attr,
3937 * becuase if the nlink attr is not 2, then there will be
3938 * ref_add() called following the dt_insert(..), such call
3939 * will make both the in-RAM and on-disk dir-inode's nlink
3940 * attr to be set as 2. LU-7447
3942 set_nlink(dir, saved_nlink);
3947 * Create an local agent inode for remote entry
3949 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3950 struct osd_device *osd,
3951 struct osd_object *pobj,
3952 const struct lu_fid *fid,
3956 struct osd_thread_info *info = osd_oti_get(env);
3957 struct inode *local;
3958 struct osd_thandle *oh;
3959 struct iattr iattr = {
3960 .ia_valid = ATTR_UID | ATTR_GID |
3961 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3962 .ia_ctime.tv_nsec = UTIME_OMIT,
3963 .ia_mtime.tv_nsec = UTIME_OMIT,
3964 .ia_atime.tv_nsec = UTIME_OMIT,
3965 .ia_uid = GLOBAL_ROOT_UID,
3966 .ia_gid = GLOBAL_ROOT_GID,
3973 oh = container_of(th, struct osd_thandle, ot_super);
3974 LASSERT(oh->ot_handle->h_transaction != NULL);
3976 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
3978 if (IS_ERR(local)) {
3979 CERROR("%s: create local error %d\n", osd_name(osd),
3980 (int)PTR_ERR(local));
3985 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3986 * correct gid on remote file, not agent here
3988 local->i_gid = current_fsgid();
3989 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
3991 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
3992 * That gives e2fsck something to look at and be happy, and allows
3993 * debugging if we need to determine where this symlink came from.
3995 if (S_ISLNK(type)) {
3996 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
3997 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
3998 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4000 i_size_write(local, rc);
4001 LDISKFS_I(local)->i_disksize = rc;
4003 unlock_new_inode(local);
4005 /* Agent inode should not have project ID */
4006 #ifdef HAVE_PROJECT_QUOTA
4007 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4008 i_projid_read(pobj->oo_inode) != 0) {
4009 rc = osd_transfer_project(local, 0, th);
4011 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4012 osd_ino2name(local), rc);
4013 RETURN(ERR_PTR(rc));
4017 /* Set special LMA flag for local agent inode */
4018 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4020 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4021 osd_name(osd), PFID(fid), rc);
4022 RETURN(ERR_PTR(rc));
4028 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
4029 lu_object_fid(&pobj->oo_dt.do_lu),
4032 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4033 osd_name(osd), PFID(fid), rc);
4034 RETURN(ERR_PTR(rc));
4041 * when direntry is deleted, we have to take care of possible agent inode
4042 * referenced by that. unfortunately we can't do this at that point:
4043 * iget() within a running transaction leads to deadlock and we better do
4044 * not call that every delete declaration to save performance. so we put
4045 * a potention agent inode on a list and process that once the transaction
4046 * is over. Notice it's not any worse in terms of real orphans as regular
4047 * object destroy doesn't put inodes on the on-disk orphan list. this should
4048 * be addressed separately
4050 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4051 struct osd_thandle *oh,
4054 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4055 struct osd_obj_orphan *oor;
4062 oor->oor_env = (struct lu_env *)env;
4063 spin_lock(&osd->od_osfs_lock);
4064 list_add(&oor->oor_list, &osd->od_orphan_list);
4065 spin_unlock(&osd->od_osfs_lock);
4067 oh->ot_remove_agents = 1;
4073 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4074 struct osd_device *osd)
4076 struct osd_thread_info *info = osd_oti_get(env);
4077 struct osd_obj_orphan *oor, *tmp;
4078 struct osd_inode_id id;
4080 struct inode *inode;
4085 spin_lock(&osd->od_osfs_lock);
4086 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4087 if (oor->oor_env == env)
4088 list_move(&oor->oor_list, &list);
4090 spin_unlock(&osd->od_osfs_lock);
4092 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4096 list_del(&oor->oor_list);
4099 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4100 inode = osd_iget_fid(info, osd, &id, &fid);
4104 if (!osd_remote_fid(env, osd, &fid)) {
4109 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4111 mark_inode_dirty(inode);
4112 ldiskfs_journal_stop(jh);
4120 * OSD layer object create function for OST objects (b=11826).
4122 * The FID is inserted into inode xattr here.
4124 * \retval 0, on success
4125 * \retval -ve, on error
4127 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4128 struct lu_attr *attr, struct dt_allocation_hint *hint,
4129 struct dt_object_format *dof, struct thandle *th)
4131 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4132 struct osd_object *obj = osd_dt_obj(dt);
4133 struct osd_thread_info *info = osd_oti_get(env);
4134 int result, on_ost = 0;
4138 if (dt_object_exists(dt))
4141 LINVRNT(osd_invariant(obj));
4142 LASSERT(!dt_object_remote(dt));
4143 LASSERT(osd_is_write_locked(env, obj));
4144 LASSERT(th != NULL);
4146 if (unlikely(fid_is_acct(fid)))
4148 * Quota files can't be created from the kernel any more,
4149 * 'tune2fs -O quota' will take care of creating them
4153 result = __osd_create(info, obj, attr, hint, dof, th);
4155 if (fid_is_idif(fid) &&
4156 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4157 struct lu_fid *tfid = &info->oti_fid;
4158 struct ost_id *oi = &info->oti_ostid;
4160 fid_to_ostid(fid, oi);
4161 ostid_to_fid(tfid, oi, 0);
4163 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4164 LMAC_FID_ON_OST, 0);
4166 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4168 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4169 on_ost ? LMAC_FID_ON_OST : 0,
4172 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4173 obj->oo_dt.do_body_ops = &osd_body_ops;
4176 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4177 struct inode *inode = obj->oo_inode;
4179 result = __osd_oi_insert(env, obj, fid, th);
4180 if (result && inode) {
4181 spin_lock(&obj->oo_guard);
4183 spin_unlock(&obj->oo_guard);
4184 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4185 ldiskfs_set_inode_state(inode,
4186 LDISKFS_STATE_LUSTRE_DESTROY);
4188 obj->oo_inode = NULL;
4193 * a small optimization - dt_insert() isn't usually applied
4194 * to OST objects, so we don't need to cache OI mapping for
4197 if (result == 0 && on_ost == 0) {
4198 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4200 result = osd_idc_find_and_init(env, osd, obj);
4201 LASSERT(result == 0);
4204 LASSERT(ergo(result == 0,
4205 dt_object_exists(dt) && !dt_object_remote(dt)));
4206 LINVRNT(osd_invariant(obj));
4210 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4211 struct thandle *handle)
4213 struct osd_thandle *oh;
4216 /* it's possible that object doesn't exist yet */
4217 LASSERT(handle != NULL);
4219 oh = container_of(handle, struct osd_thandle, ot_super);
4220 LASSERT(oh->ot_handle == NULL);
4222 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4223 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4225 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4232 * Concurrency: @dt is write locked.
4234 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4237 struct osd_object *obj = osd_dt_obj(dt);
4238 struct inode *inode = obj->oo_inode;
4239 struct osd_thandle *oh;
4242 if (!dt_object_exists(dt) || obj->oo_destroyed)
4245 LINVRNT(osd_invariant(obj));
4246 LASSERT(!dt_object_remote(dt));
4247 LASSERT(osd_is_write_locked(env, obj));
4248 LASSERT(th != NULL);
4250 oh = container_of(th, struct osd_thandle, ot_super);
4251 LASSERT(oh->ot_handle != NULL);
4253 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4255 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4256 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4258 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4259 * (65000) subdirectories by storing "1" in i_nlink if the link count
4260 * would otherwise overflow. Directory tranversal tools understand
4261 * that (st_nlink == 1) indicates that the filesystem dose not track
4262 * hard links count on the directory, and will not abort subdirectory
4263 * scanning early once (st_nlink - 2) subdirs have been found.
4265 * This also has to properly handle the case of inodes with nlink == 0
4266 * in case they are being linked into the PENDING directory
4268 spin_lock(&obj->oo_guard);
4269 if (unlikely(inode->i_nlink == 0))
4270 /* inc_nlink from 0 may cause WARN_ON */
4271 set_nlink(inode, 1);
4273 ldiskfs_inc_count(oh->ot_handle, inode);
4274 if (!S_ISDIR(inode->i_mode))
4275 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4277 spin_unlock(&obj->oo_guard);
4279 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4280 LINVRNT(osd_invariant(obj));
4282 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4287 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4288 struct thandle *handle)
4290 struct osd_thandle *oh;
4292 if (!dt_object_exists(dt))
4295 LASSERT(!dt_object_remote(dt));
4296 LASSERT(handle != NULL);
4298 oh = container_of(handle, struct osd_thandle, ot_super);
4299 LASSERT(oh->ot_handle == NULL);
4301 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4302 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4308 * Concurrency: @dt is write locked.
4310 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4313 struct osd_object *obj = osd_dt_obj(dt);
4314 struct inode *inode = obj->oo_inode;
4315 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4316 struct osd_thandle *oh;
4318 if (!dt_object_exists(dt))
4321 LINVRNT(osd_invariant(obj));
4322 LASSERT(!dt_object_remote(dt));
4323 LASSERT(osd_is_write_locked(env, obj));
4324 LASSERT(th != NULL);
4326 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4329 oh = container_of(th, struct osd_thandle, ot_super);
4330 LASSERT(oh->ot_handle != NULL);
4332 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4334 spin_lock(&obj->oo_guard);
4336 * That can be result of upgrade from old Lustre version and
4337 * applied only to local files. Just skip this ref_del call.
4338 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4340 if (inode->i_nlink == 0) {
4341 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4342 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4343 ", maybe an upgraded file? (LU-3915)\n",
4344 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4345 spin_unlock(&obj->oo_guard);
4349 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4350 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4352 ldiskfs_dec_count(oh->ot_handle, inode);
4353 spin_unlock(&obj->oo_guard);
4355 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4356 LINVRNT(osd_invariant(obj));
4358 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4364 * Concurrency: @dt is read locked.
4366 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4367 struct lu_buf *buf, const char *name)
4369 struct osd_object *obj = osd_dt_obj(dt);
4370 struct inode *inode = obj->oo_inode;
4371 struct osd_thread_info *info = osd_oti_get(env);
4372 struct dentry *dentry = &info->oti_obj_dentry;
4373 bool cache_xattr = false;
4378 /* version get is not real XATTR but uses xattr API */
4379 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4380 dt_obj_version_t *ver = buf->lb_buf;
4383 * for version we are just using xattr API but change inode
4386 if (buf->lb_len == 0)
4387 return sizeof(dt_obj_version_t);
4389 if (buf->lb_len < sizeof(dt_obj_version_t))
4392 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4393 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4395 *ver = LDISKFS_I(inode)->i_fs_version;
4397 return sizeof(dt_obj_version_t);
4400 if (!dt_object_exists(dt))
4403 LASSERT(!dt_object_remote(dt));
4404 LASSERT(inode->i_op != NULL);
4405 #ifdef HAVE_IOP_XATTR
4406 LASSERT(inode->i_op->getxattr != NULL);
4409 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4410 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4414 rc = osd_oxc_get(obj, name, buf);
4419 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4420 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4421 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4422 struct filter_fid *ff;
4423 struct ost_layout *ol;
4425 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4429 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4432 if (buf->lb_len == 0 || !buf->lb_buf)
4435 if (buf->lb_len < rc)
4439 ol = &ff->ff_layout;
4440 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4441 PFID_STRIPE_IDX_BITS);
4442 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4443 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4444 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4445 if (lma->lma_compat & LMAC_COMP_INFO) {
4446 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4447 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4448 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4450 ol->ol_comp_start = 0;
4451 ol->ol_comp_end = 0;
4455 rc = __osd_xattr_get(inode, dentry, name,
4456 buf->lb_buf, buf->lb_len);
4460 if (rc == -ENOENT || rc == -ENODATA)
4461 osd_oxc_add(obj, name, NULL, 0);
4462 else if (rc > 0 && buf->lb_buf != NULL)
4463 osd_oxc_add(obj, name, buf->lb_buf, rc);
4469 static int osd_declare_xattr_set(const struct lu_env *env,
4470 struct dt_object *dt,
4471 const struct lu_buf *buf, const char *name,
4472 int fl, struct thandle *handle)
4474 struct osd_thandle *oh;
4476 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4478 LASSERT(handle != NULL);
4480 oh = container_of(handle, struct osd_thandle, ot_super);
4481 LASSERT(oh->ot_handle == NULL);
4483 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4485 * For non-upgrading case, the LMA is set first and
4486 * usually fit inode. But for upgrade case, the LMA
4487 * may be in another separated EA block.
4489 if (dt_object_exists(dt)) {
4490 if (fl == LU_XATTR_REPLACE)
4495 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4497 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4498 /* We may need to delete the old PFID EA. */
4499 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4500 if (fl == LU_XATTR_REPLACE)
4506 * If some name entry resides on remote MDT, then will create
4507 * agent entry under remote parent. On the other hand, if the
4508 * remote entry will be removed, then related agent entry may
4509 * need to be removed from the remote parent. So there may be
4510 * kinds of cases, let's declare enough credits. The credits
4511 * for create agent entry is enough for remove case.
4513 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4514 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4515 if (dt_object_exists(dt))
4516 credits += 1; /* For updating LMA */
4520 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4525 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4527 * learn xattr size from osd_xattr_get if
4528 * attribute has not been read yet
4530 buflen = __osd_xattr_get(
4531 osd_dt_obj(dt)->oo_inode,
4532 &osd_oti_get(env)->oti_obj_dentry,
4537 buflen = buf->lb_len;
4540 if (buflen > sb->s_blocksize) {
4541 credits += osd_calc_bkmap_credits(
4543 (buflen + sb->s_blocksize - 1) >>
4544 sb->s_blocksize_bits);
4548 * xattr set may involve inode quota change, reserve credits for
4549 * dquot_initialize()
4551 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4554 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4559 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4560 const struct lu_buf *buf, int fl,
4561 struct thandle *handle)
4563 struct osd_thread_info *info = osd_oti_get(env);
4564 struct dentry *dentry = &info->oti_obj_dentry;
4565 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4566 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4567 struct inode *inode = obj->oo_inode;
4568 struct filter_fid *ff = buf->lb_buf;
4569 struct ost_layout *ol = &ff->ff_layout;
4570 int flags = XATTR_REPLACE;
4575 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4578 rc = osd_get_lma(info, inode, dentry, loa);
4579 if (rc == -ENODATA) {
4580 /* Usually for upgarding from old device */
4581 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4582 LMAC_FID_ON_OST, 0);
4583 flags = XATTR_CREATE;
4588 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4589 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4592 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4593 /* Separate PFID EA from LMA */
4594 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4595 lustre_lma_swab(lma);
4596 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4597 sizeof(*lma), XATTR_REPLACE);
4599 obj->oo_pfid_in_lma = 0;
4600 rc = LU_XATTR_CREATE;
4606 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4610 * Old client does not send stripe information,
4611 * then store the PFID EA on disk separatedly.
4613 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4614 ol->ol_stripe_size == 0))
4617 /* Remove old PFID EA entry firstly. */
4618 dquot_initialize(inode);
4619 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4620 if (rc == -ENODATA) {
4621 if ((fl & LU_XATTR_REPLACE) && !(fl & LU_XATTR_CREATE))
4628 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4629 if (likely(ol->ol_stripe_size != 0)) {
4630 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4631 PFID_STRIPE_IDX_BITS;
4632 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4633 lma->lma_compat |= LMAC_STRIPE_INFO;
4634 if (ol->ol_comp_id != 0) {
4635 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4636 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4637 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4638 lma->lma_compat |= LMAC_COMP_INFO;
4642 lustre_loa_swab(loa, false);
4644 /* Store the PFID EA inside LMA. */
4645 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4648 obj->oo_pfid_in_lma = 1;
4654 * In DNE environment, the object (in spite of regular file or directory)
4655 * and its name entry may reside on different MDTs. Under such case, we will
4656 * create an agent entry on the MDT where the object resides. The agent entry
4657 * references the object locally, that makes the object to be visible to the
4658 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4659 * such as 'tar' can handle the object properly.
4661 * We handle the agent entry during set linkEA that is the common interface
4662 * for both regular file and directroy, can handle kinds of cases, such as
4663 * create/link/unlink/rename, and so on.
4665 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4667 * XXX: There are two known issues:
4668 * 1. For one object, we will create at most one agent entry even if there
4669 * may be more than one cross-MDTs hard links on the object. So the local
4670 * e2fsck may claim that the object's nlink is larger than the name entries
4671 * that reference such inode. And in further, the e2fsck will fix the nlink
4672 * attribute to match the local references. Then it will cause the object's
4673 * nlink attribute to be inconsistent with the global references. it is bad
4674 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4675 * other hand, the global namespace LFSCK can repair the object's attribute
4676 * according to the linkEA.
4677 * 2. There may be too many hard links on the object as to its linkEA overflow,
4678 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4679 * case happened, then at this point, we do not know whether there are some
4680 * cross-MDTs reference. But there are local references, it guarantees that
4681 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4683 static int osd_xattr_handle_linkea(const struct lu_env *env,
4684 struct osd_device *osd,
4685 struct osd_object *obj,
4686 const struct lu_buf *buf,
4687 struct thandle *handle)
4689 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4690 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4691 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4692 struct lu_name tmpname;
4693 struct osd_thandle *oh;
4695 bool remote = false;
4699 oh = container_of(handle, struct osd_thandle, ot_super);
4700 LASSERT(oh->ot_handle != NULL);
4702 rc = linkea_init_with_rec(&ldata);
4704 linkea_first_entry(&ldata);
4705 while (ldata.ld_lee != NULL && !remote) {
4706 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4708 if (osd_remote_fid(env, osd, tfid) > 0)
4711 linkea_next_entry(&ldata);
4713 } else if (rc == -ENODATA) {
4719 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4720 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4722 CERROR("%s: failed to remove agent entry for "DFID
4723 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4724 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4725 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4727 CERROR("%s: failed to create agent entry for "DFID
4728 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4735 * Concurrency: @dt is write locked.
4737 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4738 const struct lu_buf *buf, const char *name, int fl,
4739 struct thandle *handle)
4741 struct osd_object *obj = osd_dt_obj(dt);
4742 struct osd_device *osd = osd_obj2dev(obj);
4743 struct inode *inode = obj->oo_inode;
4744 struct osd_thread_info *info = osd_oti_get(env);
4754 /* version set is not real XATTR */
4755 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4756 dt_obj_version_t *version = buf->lb_buf;
4759 * for version we are just using xattr API but change inode
4762 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4764 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4765 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4767 LDISKFS_I(inode)->i_fs_version = *version;
4769 * Version is set after all inode operations are finished,
4770 * so we should mark it dirty here
4772 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4777 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4778 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4781 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4784 * For the OST device with 256 bytes inode size by default,
4785 * the PFID EA will be stored together with LMA EA to avoid
4786 * performance trouble. Otherwise the PFID EA can be stored
4787 * independently. LU-8998
4789 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4790 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4791 LASSERT(buf->lb_buf);
4793 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4796 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4797 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4798 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4800 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4804 lma->lma_incompat |= LMAI_STRIPED;
4805 lustre_lma_swab(lma);
4806 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4807 sizeof(*lma), XATTR_REPLACE);
4810 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4811 LASSERT(!osd->od_is_ost);
4813 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4818 if (fl & LU_XATTR_REPLACE)
4819 fs_flags |= XATTR_REPLACE;
4821 if (fl & LU_XATTR_CREATE)
4822 fs_flags |= XATTR_CREATE;
4824 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4825 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4828 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4829 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4830 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4836 * Concurrency: @dt is read locked.
4838 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4839 const struct lu_buf *buf)
4841 struct osd_object *obj = osd_dt_obj(dt);
4842 struct inode *inode = obj->oo_inode;
4843 struct osd_thread_info *info = osd_oti_get(env);
4844 struct dentry *dentry = &info->oti_obj_dentry;
4846 if (!dt_object_exists(dt))
4849 LASSERT(!dt_object_remote(dt));
4850 LASSERT(inode->i_op != NULL);
4851 LASSERT(inode->i_op->listxattr != NULL);
4853 dentry->d_inode = inode;
4854 dentry->d_sb = inode->i_sb;
4855 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4858 static int osd_declare_xattr_del(const struct lu_env *env,
4859 struct dt_object *dt, const char *name,
4860 struct thandle *handle)
4862 struct osd_thandle *oh;
4863 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4865 LASSERT(!dt_object_remote(dt));
4866 LASSERT(handle != NULL);
4868 oh = container_of(handle, struct osd_thandle, ot_super);
4869 LASSERT(oh->ot_handle == NULL);
4871 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4872 osd_dto_credits_noquota[DTO_XATTR_SET]);
4874 * xattr del may involve inode quota change, reserve credits for
4875 * dquot_initialize()
4877 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4883 * Concurrency: @dt is write locked.
4885 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4886 const char *name, struct thandle *handle)
4888 struct osd_object *obj = osd_dt_obj(dt);
4889 struct inode *inode = obj->oo_inode;
4890 struct osd_thread_info *info = osd_oti_get(env);
4891 struct dentry *dentry = &info->oti_obj_dentry;
4894 if (!dt_object_exists(dt))
4897 LASSERT(!dt_object_remote(dt));
4898 LASSERT(inode->i_op != NULL);
4899 LASSERT(handle != NULL);
4900 #ifdef HAVE_IOP_XATTR
4901 LASSERT(inode->i_op->removexattr != NULL);
4904 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4906 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4907 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4909 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4910 &info->oti_ost_attrs);
4912 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4914 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4915 lustre_lma_swab(lma);
4916 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4917 sizeof(*lma), XATTR_REPLACE);
4919 obj->oo_pfid_in_lma = 0;
4922 dquot_initialize(inode);
4923 dentry->d_inode = inode;
4924 dentry->d_sb = inode->i_sb;
4925 rc = ll_vfs_removexattr(dentry, inode, name);
4928 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4931 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4932 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4933 osd_oxc_del(obj, name);
4938 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4939 __u64 start, __u64 end)
4941 struct osd_object *obj = osd_dt_obj(dt);
4942 struct inode *inode = obj->oo_inode;
4943 struct file *file = osd_quasi_file(env, inode);
4948 rc = vfs_fsync_range(file, start, end, 0);
4953 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4958 static bool osd_check_stale(struct dt_object *dt)
4966 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4967 const struct dt_index_features *feat)
4969 struct iam_descr *descr;
4971 if (osd_object_is_root(o))
4972 return feat == &dt_directory_features;
4974 LASSERT(o->oo_dir != NULL);
4976 descr = o->oo_dir->od_container.ic_descr;
4977 if (feat == &dt_directory_features) {
4978 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4983 return feat->dif_keysize_min <= descr->id_key_size &&
4984 descr->id_key_size <= feat->dif_keysize_max &&
4985 feat->dif_recsize_min <= descr->id_rec_size &&
4986 descr->id_rec_size <= feat->dif_recsize_max &&
4987 !(feat->dif_flags & (DT_IND_VARKEY |
4988 DT_IND_VARREC | DT_IND_NONUNQ)) &&
4989 ergo(feat->dif_flags & DT_IND_UPDATE,
4990 1 /* XXX check that object (and fs) is writable */);
4994 static int osd_iam_container_init(const struct lu_env *env,
4995 struct osd_object *obj,
4996 struct osd_directory *dir)
4998 struct iam_container *bag = &dir->od_container;
5001 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5005 result = iam_container_setup(bag);
5007 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5009 iam_container_fini(bag);
5016 * Concurrency: no external locking is necessary.
5018 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5019 const struct dt_index_features *feat)
5023 struct osd_object *obj = osd_dt_obj(dt);
5025 LINVRNT(osd_invariant(obj));
5027 if (osd_object_is_root(obj)) {
5028 dt->do_index_ops = &osd_index_ea_ops;
5030 } else if (feat == &dt_directory_features) {
5031 dt->do_index_ops = &osd_index_ea_ops;
5032 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5037 } else if (unlikely(feat == &dt_otable_features)) {
5038 dt->do_index_ops = &osd_otable_ops;
5040 } else if (unlikely(feat == &dt_acct_features)) {
5041 dt->do_index_ops = &osd_acct_index_ops;
5044 } else if (!osd_has_index(obj)) {
5045 struct osd_directory *dir;
5046 struct osd_device *osd = osd_obj2dev(obj);
5047 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5052 spin_lock(&obj->oo_guard);
5053 if (obj->oo_dir == NULL)
5057 * Concurrent thread allocated container data.
5060 spin_unlock(&obj->oo_guard);
5062 * Now, that we have container data, serialize its
5065 down_write(&obj->oo_ext_idx_sem);
5067 * recheck under lock.
5070 if (osd_has_index(obj)) {
5075 result = osd_iam_container_init(env, obj, obj->oo_dir);
5076 if (result || feat == &dt_lfsck_namespace_features ||
5077 feat == &dt_lfsck_layout_orphan_features ||
5078 feat == &dt_lfsck_layout_dangling_features)
5081 result = osd_index_register(osd, fid,
5082 feat->dif_keysize_max,
5083 feat->dif_recsize_max);
5085 CWARN("%s: failed to register index "
5087 osd_name(osd), PFID(fid), result);
5088 else if (result > 0)
5091 CDEBUG(D_LFSCK, "%s: index object "DFID
5092 " (%d/%d) registered\n",
5093 osd_name(osd), PFID(fid),
5094 (int)feat->dif_keysize_max,
5095 (int)feat->dif_recsize_max);
5098 up_write(&obj->oo_ext_idx_sem);
5106 if (result == 0 && skip_iam == 0) {
5107 if (!osd_iam_index_probe(env, obj, feat))
5110 LINVRNT(osd_invariant(obj));
5115 static int osd_otable_it_attr_get(const struct lu_env *env,
5116 struct dt_object *dt,
5117 struct lu_attr *attr)
5123 static const struct dt_object_operations osd_obj_ops = {
5124 .do_read_lock = osd_read_lock,
5125 .do_write_lock = osd_write_lock,
5126 .do_read_unlock = osd_read_unlock,
5127 .do_write_unlock = osd_write_unlock,
5128 .do_write_locked = osd_write_locked,
5129 .do_attr_get = osd_attr_get,
5130 .do_declare_attr_set = osd_declare_attr_set,
5131 .do_attr_set = osd_attr_set,
5132 .do_ah_init = osd_ah_init,
5133 .do_declare_create = osd_declare_create,
5134 .do_create = osd_create,
5135 .do_declare_destroy = osd_declare_destroy,
5136 .do_destroy = osd_destroy,
5137 .do_index_try = osd_index_try,
5138 .do_declare_ref_add = osd_declare_ref_add,
5139 .do_ref_add = osd_ref_add,
5140 .do_declare_ref_del = osd_declare_ref_del,
5141 .do_ref_del = osd_ref_del,
5142 .do_xattr_get = osd_xattr_get,
5143 .do_declare_xattr_set = osd_declare_xattr_set,
5144 .do_xattr_set = osd_xattr_set,
5145 .do_declare_xattr_del = osd_declare_xattr_del,
5146 .do_xattr_del = osd_xattr_del,
5147 .do_xattr_list = osd_xattr_list,
5148 .do_object_sync = osd_object_sync,
5149 .do_invalidate = osd_invalidate,
5150 .do_check_stale = osd_check_stale,
5153 static const struct dt_object_operations osd_obj_otable_it_ops = {
5154 .do_attr_get = osd_otable_it_attr_get,
5155 .do_index_try = osd_index_try,
5158 static int osd_index_declare_iam_delete(const struct lu_env *env,
5159 struct dt_object *dt,
5160 const struct dt_key *key,
5161 struct thandle *handle)
5163 struct osd_thandle *oh;
5165 oh = container_of(handle, struct osd_thandle, ot_super);
5166 LASSERT(oh->ot_handle == NULL);
5168 /* Recycle may cause additional three blocks to be changed. */
5169 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5170 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5176 * delete a (key, value) pair from index \a dt specified by \a key
5178 * \param dt osd index object
5179 * \param key key for index
5180 * \param rec record reference
5181 * \param handle transaction handler
5184 * \retval -ve failure
5186 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5187 const struct dt_key *key,
5188 struct thandle *handle)
5190 struct osd_thread_info *oti = osd_oti_get(env);
5191 struct osd_object *obj = osd_dt_obj(dt);
5192 struct osd_thandle *oh;
5193 struct iam_path_descr *ipd;
5194 struct iam_container *bag = &obj->oo_dir->od_container;
5199 if (!dt_object_exists(dt))
5202 LINVRNT(osd_invariant(obj));
5203 LASSERT(!dt_object_remote(dt));
5204 LASSERT(bag->ic_object == obj->oo_inode);
5205 LASSERT(handle != NULL);
5207 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5209 ipd = osd_idx_ipd_get(env, bag);
5210 if (unlikely(ipd == NULL))
5213 oh = container_of(handle, struct osd_thandle, ot_super);
5214 LASSERT(oh->ot_handle != NULL);
5215 LASSERT(oh->ot_handle->h_transaction != NULL);
5217 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5218 /* swab quota uid/gid provided by caller */
5219 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5220 key = (const struct dt_key *)&oti->oti_quota_id;
5223 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5224 osd_ipd_put(env, bag, ipd);
5225 LINVRNT(osd_invariant(obj));
5226 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5230 static int osd_index_declare_ea_delete(const struct lu_env *env,
5231 struct dt_object *dt,
5232 const struct dt_key *key,
5233 struct thandle *handle)
5235 struct osd_thandle *oh;
5236 struct inode *inode;
5241 LASSERT(!dt_object_remote(dt));
5242 LASSERT(handle != NULL);
5244 oh = container_of(handle, struct osd_thandle, ot_super);
5245 LASSERT(oh->ot_handle == NULL);
5247 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5248 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5250 inode = osd_dt_obj(dt)->oo_inode;
5254 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5255 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5260 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5263 struct osd_fid_pack *rec;
5266 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5267 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5268 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5269 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5275 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5276 const struct lu_fid *fid)
5278 struct seq_server_site *ss = osd_seq_site(osd);
5282 /* FID seqs not in FLDB, must be local seq */
5283 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5287 * If FLD is not being initialized yet, it only happens during the
5288 * initialization, likely during mgs initialization, and we assume
5289 * this is local FID.
5291 if (ss == NULL || ss->ss_server_fld == NULL)
5294 /* Only check the local FLDB here */
5295 if (osd_seq_exists(env, osd, fid_seq(fid)))
5301 static void osd_take_care_of_agent(const struct lu_env *env,
5302 struct osd_device *osd,
5303 struct osd_thandle *oh,
5304 struct ldiskfs_dir_entry_2 *de)
5306 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5307 struct osd_idmap_cache *idc;
5308 int rc, schedule = 0;
5310 LASSERT(de != NULL);
5312 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5313 if (likely(rc == 0)) {
5314 idc = osd_idc_find_or_init(env, osd, fid);
5315 if (IS_ERR(idc) || idc->oic_remote)
5317 } else if (rc == -ENODATA) {
5319 * can't get FID, postpone to the end of the
5320 * transaction when iget() is safe
5324 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5327 osd_schedule_agent_inode_removal(env, oh,
5328 le32_to_cpu(de->inode));
5332 * Index delete function for interoperability mode (b11826).
5333 * It will remove the directory entry added by osd_index_ea_insert().
5334 * This entry is needed to maintain name->fid mapping.
5336 * \param key, key i.e. file entry to be deleted
5338 * \retval 0, on success
5339 * \retval -ve, on error
5341 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5342 const struct dt_key *key, struct thandle *handle)
5344 struct osd_object *obj = osd_dt_obj(dt);
5345 struct inode *dir = obj->oo_inode;
5346 struct dentry *dentry;
5347 struct osd_thandle *oh;
5348 struct ldiskfs_dir_entry_2 *de = NULL;
5349 struct buffer_head *bh;
5350 struct htree_lock *hlock = NULL;
5351 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5356 if (!dt_object_exists(dt))
5359 LINVRNT(osd_invariant(obj));
5360 LASSERT(!dt_object_remote(dt));
5361 LASSERT(handle != NULL);
5363 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5365 oh = container_of(handle, struct osd_thandle, ot_super);
5366 LASSERT(oh->ot_handle != NULL);
5367 LASSERT(oh->ot_handle->h_transaction != NULL);
5369 dquot_initialize(dir);
5370 dentry = osd_child_dentry_get(env, obj,
5371 (char *)key, strlen((char *)key));
5373 if (obj->oo_hl_head != NULL) {
5374 hlock = osd_oti_get(env)->oti_hlock;
5375 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5376 dir, LDISKFS_HLOCK_DEL);
5378 down_write(&obj->oo_ext_idx_sem);
5381 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5384 * If this is not the ".." entry, it might be a remote DNE
5385 * entry and we need to check if the FID is for a remote
5386 * MDT. If the FID is not in the directory entry (e.g.
5387 * upgraded 1.8 filesystem without dirdata enabled) then
5388 * we need to get the FID from the LMA. For a remote directory
5389 * there HAS to be an LMA, it cannot be an IGIF inode in this
5392 * Delete the entry before the agent inode in order to
5393 * simplify error handling. At worst an error after deleting
5394 * the entry first might leak the agent inode afterward. The
5395 * reverse would need filesystem abort in case of error deleting
5396 * the entry after the agent had been removed, or leave a
5397 * dangling entry pointing at a random inode.
5399 if (strcmp((char *)key, dotdot) != 0)
5400 osd_take_care_of_agent(env, osd, oh, de);
5401 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5407 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5408 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5409 /* NB, dirent count may not be accurate, because it's counted
5412 if (obj->oo_dirent_count)
5413 obj->oo_dirent_count--;
5415 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5418 ldiskfs_htree_unlock(hlock);
5420 up_write(&obj->oo_ext_idx_sem);
5423 LASSERT(osd_invariant(obj));
5424 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5429 * Lookup index for \a key and copy record to \a rec.
5431 * \param dt osd index object
5432 * \param key key for index
5433 * \param rec record reference
5435 * \retval +ve success : exact mach
5436 * \retval 0 return record with key not greater than \a key
5437 * \retval -ve failure
5439 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5440 struct dt_rec *rec, const struct dt_key *key)
5442 struct osd_object *obj = osd_dt_obj(dt);
5443 struct iam_path_descr *ipd;
5444 struct iam_container *bag = &obj->oo_dir->od_container;
5445 struct osd_thread_info *oti = osd_oti_get(env);
5446 struct iam_iterator *it = &oti->oti_idx_it;
5447 struct iam_rec *iam_rec;
5452 if (!dt_object_exists(dt))
5455 LASSERT(osd_invariant(obj));
5456 LASSERT(!dt_object_remote(dt));
5457 LASSERT(bag->ic_object == obj->oo_inode);
5459 ipd = osd_idx_ipd_get(env, bag);
5463 /* got ipd now we can start iterator. */
5464 iam_it_init(it, bag, 0, ipd);
5466 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5467 /* swab quota uid/gid provided by caller */
5468 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5469 key = (const struct dt_key *)&oti->oti_quota_id;
5472 rc = iam_it_get(it, (struct iam_key *)key);
5474 if (S_ISDIR(obj->oo_inode->i_mode))
5475 iam_rec = (struct iam_rec *)oti->oti_ldp;
5477 iam_rec = (struct iam_rec *)rec;
5479 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5481 if (S_ISDIR(obj->oo_inode->i_mode))
5482 osd_fid_unpack((struct lu_fid *)rec,
5483 (struct osd_fid_pack *)iam_rec);
5484 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5485 osd_quota_unpack(obj, rec);
5490 osd_ipd_put(env, bag, ipd);
5492 LINVRNT(osd_invariant(obj));
5497 static int osd_index_declare_iam_insert(const struct lu_env *env,
5498 struct dt_object *dt,
5499 const struct dt_rec *rec,
5500 const struct dt_key *key,
5501 struct thandle *handle)
5503 struct osd_thandle *oh;
5505 LASSERT(handle != NULL);
5507 oh = container_of(handle, struct osd_thandle, ot_super);
5508 LASSERT(oh->ot_handle == NULL);
5510 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5511 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5517 * Inserts (key, value) pair in \a dt index object.
5519 * \param dt osd index object
5520 * \param key key for index
5521 * \param rec record reference
5522 * \param th transaction handler
5525 * \retval -ve failure
5527 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5528 const struct dt_rec *rec,
5529 const struct dt_key *key, struct thandle *th)
5531 struct osd_object *obj = osd_dt_obj(dt);
5532 struct iam_path_descr *ipd;
5533 struct osd_thandle *oh;
5534 struct iam_container *bag;
5535 struct osd_thread_info *oti = osd_oti_get(env);
5536 struct iam_rec *iam_rec;
5541 if (!dt_object_exists(dt))
5544 LINVRNT(osd_invariant(obj));
5545 LASSERT(!dt_object_remote(dt));
5547 bag = &obj->oo_dir->od_container;
5548 LASSERT(bag->ic_object == obj->oo_inode);
5549 LASSERT(th != NULL);
5551 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5553 ipd = osd_idx_ipd_get(env, bag);
5554 if (unlikely(ipd == NULL))
5557 oh = container_of(th, struct osd_thandle, ot_super);
5558 LASSERT(oh->ot_handle != NULL);
5559 LASSERT(oh->ot_handle->h_transaction != NULL);
5560 if (S_ISDIR(obj->oo_inode->i_mode)) {
5561 iam_rec = (struct iam_rec *)oti->oti_ldp;
5562 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5564 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5565 /* pack quota uid/gid */
5566 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5567 key = (const struct dt_key *)&oti->oti_quota_id;
5568 /* pack quota record */
5569 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5570 iam_rec = (struct iam_rec *)rec;
5572 iam_rec = (struct iam_rec *)rec;
5575 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5577 osd_ipd_put(env, bag, ipd);
5578 LINVRNT(osd_invariant(obj));
5579 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5584 * Calls ldiskfs_add_entry() to add directory entry
5585 * into the directory. This is required for
5586 * interoperability mode (b11826)
5588 * \retval 0, on success
5589 * \retval -ve, on error
5591 static int __osd_ea_add_rec(struct osd_thread_info *info,
5592 struct osd_object *pobj, struct inode *cinode,
5593 const char *name, const struct lu_fid *fid,
5594 struct htree_lock *hlock, struct thandle *th)
5596 struct ldiskfs_dentry_param *ldp;
5597 struct dentry *child;
5598 struct osd_thandle *oth;
5601 oth = container_of(th, struct osd_thandle, ot_super);
5602 LASSERT(oth->ot_handle != NULL);
5603 LASSERT(oth->ot_handle->h_transaction != NULL);
5604 LASSERT(pobj->oo_inode);
5606 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5607 if (unlikely(osd_object_is_root(pobj)))
5610 osd_get_ldiskfs_dirent_param(ldp, fid);
5611 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5612 child->d_fsdata = (void *)ldp;
5613 dquot_initialize(pobj->oo_inode);
5614 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5615 child, cinode, hlock);
5616 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5617 struct ldiskfs_dir_entry_2 *de;
5618 struct buffer_head *bh;
5621 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5624 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5627 if (S_ISDIR(cinode->i_mode))
5628 de->file_type = LDISKFS_DIRENT_LUFID |
5629 LDISKFS_FT_REG_FILE;
5631 de->file_type = LDISKFS_DIRENT_LUFID |
5633 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5644 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5645 * into the directory.Also sets flags into osd object to
5646 * indicate dot and dotdot are created. This is required for
5647 * interoperability mode (b11826)
5649 * \param dir directory for dot and dotdot fixup.
5650 * \param obj child object for linking
5652 * \retval 0, on success
5653 * \retval -ve, on error
5655 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5656 struct osd_object *dir,
5657 struct inode *parent_dir, const char *name,
5658 const struct lu_fid *dot_fid,
5659 const struct lu_fid *dot_dot_fid,
5662 struct inode *inode = dir->oo_inode;
5663 struct osd_thandle *oth;
5666 oth = container_of(th, struct osd_thandle, ot_super);
5667 LASSERT(oth->ot_handle->h_transaction != NULL);
5668 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5670 if (strcmp(name, dot) == 0) {
5671 if (dir->oo_compat_dot_created) {
5674 LASSERT(inode->i_ino == parent_dir->i_ino);
5675 dir->oo_compat_dot_created = 1;
5678 } else if (strcmp(name, dotdot) == 0) {
5679 if (!dir->oo_compat_dot_created)
5681 /* in case of rename, dotdot is already created */
5682 if (dir->oo_compat_dotdot_created) {
5683 return __osd_ea_add_rec(info, dir, parent_dir, name,
5684 dot_dot_fid, NULL, th);
5687 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5688 struct lu_fid tfid = *dot_dot_fid;
5691 result = osd_add_dot_dotdot_internal(info,
5692 dir->oo_inode, parent_dir, dot_fid,
5695 result = osd_add_dot_dotdot_internal(info,
5696 dir->oo_inode, parent_dir, dot_fid,
5701 dir->oo_compat_dotdot_created = 1;
5709 * It will call the appropriate osd_add* function and return the
5710 * value, return by respective functions.
5712 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5713 struct inode *cinode, const char *name,
5714 const struct lu_fid *fid, struct thandle *th)
5716 struct osd_thread_info *info = osd_oti_get(env);
5717 struct htree_lock *hlock;
5720 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5722 if (name[0] == '.' && (name[1] == '\0' ||
5723 (name[1] == '.' && name[2] == '\0'))) {
5724 if (hlock != NULL) {
5725 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5728 down_write(&pobj->oo_ext_idx_sem);
5731 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5732 lu_object_fid(&pobj->oo_dt.do_lu),
5735 if (hlock != NULL) {
5736 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5737 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5739 down_write(&pobj->oo_ext_idx_sem);
5742 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5743 struct lu_fid *tfid = &info->oti_fid;
5747 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5750 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5754 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
5755 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
5756 pobj->oo_dirent_count++;
5759 ldiskfs_htree_unlock(hlock);
5761 up_write(&pobj->oo_ext_idx_sem);
5767 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5768 const struct lu_fid *fid, struct osd_inode_id *id)
5770 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5771 struct inode *inode = NULL;
5778 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5781 if (scrub->os_running && scrub->os_pos_current > id->oii_ino)
5784 if (dev->od_auto_scrub_interval == AS_NEVER ||
5785 ktime_get_real_seconds() <
5786 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5790 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5791 if (rc == -ENOENT) {
5792 __u32 gen = id->oii_gen;
5798 inode = osd_iget(oti, dev, id);
5799 /* The inode has been removed (by race maybe). */
5800 if (IS_ERR(inode)) {
5801 rc = PTR_ERR(inode);
5803 RETURN(rc == -ESTALE ? -ENOENT : rc);
5806 /* The OI mapping is lost. */
5807 if (gen != OSD_OII_NOGEN)
5811 * The inode may has been reused by others, we do not know,
5812 * leave it to be handled by subsequent osd_fid_lookup().
5815 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5822 if (scrub->os_running) {
5823 if (inode == NULL) {
5824 inode = osd_iget(oti, dev, id);
5825 /* The inode has been removed (by race maybe). */
5826 if (IS_ERR(inode)) {
5827 rc = PTR_ERR(inode);
5829 RETURN(rc == -ESTALE ? -ENOENT : rc);
5833 rc = osd_oii_insert(dev, fid, id, insert);
5835 * There is race condition between osd_oi_lookup and OI scrub.
5836 * The OI scrub finished just after osd_oi_lookup() failure.
5837 * Under such case, it is unnecessary to trigger OI scrub again,
5838 * but try to call osd_oi_lookup() again.
5840 if (unlikely(rc == -EAGAIN))
5843 if (!S_ISDIR(inode->i_mode))
5846 rc = osd_check_lmv(oti, dev, inode);
5851 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5852 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5853 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5854 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
5855 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
5856 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
5857 if (rc == 0 || rc == -EALREADY)
5869 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5870 struct osd_device *dev,
5871 struct lu_fid *fid, __u32 ino)
5873 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5874 struct osd_idmap_cache *oic = &oti->oti_cache;
5875 struct inode *inode;
5878 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5879 inode = osd_iget(oti, dev, &oic->oic_lid);
5880 if (IS_ERR(inode)) {
5881 fid_zero(&oic->oic_fid);
5882 return PTR_ERR(inode);
5885 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5888 fid_zero(&oic->oic_fid);
5890 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5894 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5895 struct osd_inode_id *id, const struct lu_fid *fid)
5897 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5898 id->oii_ino, id->oii_gen, info);
5899 info->oti_cache.oic_lid = *id;
5900 info->oti_cache.oic_fid = *fid;
5901 info->oti_cache.oic_dev = osd;
5905 * Get parent FID from the linkEA.
5907 * For a directory which parent resides on remote MDT, to satisfy the
5908 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5909 * the other hand, to make the lookup(..) on the directory can return
5910 * the real parent FID, we append the real parent FID after its ".."
5911 * name entry in the /REMOTE_PARENT_DIR.
5913 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5914 * backup. So after the restore, we cannot get the right parent FID from
5915 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5916 * we have stored the real parent FID in the directory object's linkEA,
5917 * we can parse the linkEA for the real parent FID.
5919 * \param[in] env pointer to the thread context
5920 * \param[in] obj pointer to the object to be handled
5921 * \param[out]fid pointer to the buffer to hold the parent FID
5923 * \retval 0 for getting the real parent FID successfully
5924 * \retval negative error number on failure
5926 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5927 struct osd_object *obj,
5930 struct osd_thread_info *oti = osd_oti_get(env);
5931 struct lu_buf *buf = &oti->oti_big_buf;
5932 struct dentry *dentry = &oti->oti_obj_dentry;
5933 struct inode *inode = obj->oo_inode;
5934 struct linkea_data ldata = { NULL };
5940 if (!S_ISDIR(inode->i_mode))
5944 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5945 buf->lb_buf, buf->lb_len);
5946 if (rc == -ERANGE) {
5947 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5950 lu_buf_realloc(buf, rc);
5951 if (buf->lb_buf == NULL)
5958 if (unlikely(rc == 0))
5964 if (unlikely(buf->lb_buf == NULL)) {
5965 lu_buf_realloc(buf, rc);
5966 if (buf->lb_buf == NULL)
5973 rc = linkea_init_with_rec(&ldata);
5975 linkea_first_entry(&ldata);
5976 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5982 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5983 struct inode *inode,
5984 const struct lu_fid *pfid,
5985 const char *name, const int namelen)
5987 struct osd_thread_info *oti = osd_oti_get(env);
5988 struct lu_buf *buf = &oti->oti_big_buf;
5989 struct dentry *dentry = &oti->oti_obj_dentry;
5990 struct linkea_data ldata = { NULL };
5991 struct lu_name cname = { .ln_name = name,
5992 .ln_namelen = namelen };
5998 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5999 buf->lb_buf, buf->lb_len);
6001 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6006 if (unlikely(rc == 0))
6009 if (buf->lb_len < rc) {
6010 lu_buf_realloc(buf, rc);
6011 if (buf->lb_buf == NULL)
6018 rc = linkea_init_with_rec(&ldata);
6020 rc = linkea_links_find(&ldata, &cname, pfid);
6026 * Calls ->lookup() to find dentry. From dentry get inode and
6027 * read inode's ea to get fid. This is required for interoperability
6030 * \retval 0, on success
6031 * \retval -ve, on error
6033 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6034 struct dt_rec *rec, const struct dt_key *key)
6036 struct inode *dir = obj->oo_inode;
6037 struct dentry *dentry;
6038 struct ldiskfs_dir_entry_2 *de;
6039 struct buffer_head *bh;
6040 struct lu_fid *fid = (struct lu_fid *)rec;
6041 struct htree_lock *hlock = NULL;
6047 LASSERT(dir->i_op != NULL);
6048 LASSERT(dir->i_op->lookup != NULL);
6050 dentry = osd_child_dentry_get(env, obj,
6051 (char *)key, strlen((char *)key));
6053 if (obj->oo_hl_head != NULL) {
6054 hlock = osd_oti_get(env)->oti_hlock;
6055 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6056 dir, LDISKFS_HLOCK_LOOKUP);
6058 down_read(&obj->oo_ext_idx_sem);
6061 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6063 struct osd_thread_info *oti = osd_oti_get(env);
6064 struct osd_inode_id *id = &oti->oti_id;
6065 struct osd_device *dev = osd_obj2dev(obj);
6067 ino = le32_to_cpu(de->inode);
6068 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6070 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6074 rc = osd_get_fid_from_dentry(de, rec);
6076 /* done with de, release bh */
6079 if (unlikely(is_remote_parent_ino(dev, ino))) {
6080 const char *name = (const char *)key;
6083 * If the parent is on remote MDT, and there
6084 * is no FID-in-dirent, then we have to get
6085 * the parent FID from the linkEA.
6087 if (likely(strlen(name) == 2 &&
6088 name[0] == '.' && name[1] == '.'))
6089 rc = osd_get_pfid_from_linkea(env, obj,
6092 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6095 osd_id_gen(id, ino, OSD_OII_NOGEN);
6098 if (rc != 0 || osd_remote_fid(env, dev, fid))
6101 rc = osd_consistency_check(oti, dev, fid, id);
6102 if (rc != -ENOENT) {
6103 /* Other error should not affect lookup result. */
6106 /* Normal file mapping should be added into OI cache
6107 * after FID in LMA check, but for local files like
6108 * hsm_actions, their FIDs are not stored in OI files,
6109 * see osd_initial_OI_scrub(), and here is the only
6110 * place to load mapping into OI cache.
6112 if (!fid_is_namespace_visible(fid))
6113 osd_add_oi_cache(osd_oti_get(env),
6114 osd_obj2dev(obj), id, fid);
6124 ldiskfs_htree_unlock(hlock);
6126 up_read(&obj->oo_ext_idx_sem);
6130 static int osd_index_declare_ea_insert(const struct lu_env *env,
6131 struct dt_object *dt,
6132 const struct dt_rec *rec,
6133 const struct dt_key *key,
6134 struct thandle *handle)
6136 struct osd_thandle *oh;
6137 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6138 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6139 const struct lu_fid *fid = rec1->rec_fid;
6140 int credits, rc = 0;
6141 struct osd_idmap_cache *idc;
6145 LASSERT(!dt_object_remote(dt));
6146 LASSERT(handle != NULL);
6147 LASSERT(fid != NULL);
6148 LASSERT(rec1->rec_type != 0);
6150 oh = container_of(handle, struct osd_thandle, ot_super);
6151 LASSERT(oh->ot_handle == NULL);
6153 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6156 * we can't call iget() while a transactions is running
6157 * (this can lead to a deadlock), but we need to know
6158 * inum and object type. so we find this information at
6159 * declaration and cache in per-thread info
6161 idc = osd_idc_find_or_init(env, osd, fid);
6163 RETURN(PTR_ERR(idc));
6164 if (idc->oic_remote) {
6166 * a reference to remote inode is represented by an
6167 * agent inode which we have to create
6169 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6170 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6173 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6175 if (osd_dt_obj(dt)->oo_inode != NULL) {
6176 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6179 * We ignore block quota on meta pool (MDTs), so needn't
6180 * calculate how many blocks will be consumed by this index
6183 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6185 i_projid_read(inode), 0,
6186 oh, osd_dt_obj(dt), NULL,
6191 #ifdef HAVE_PROJECT_QUOTA
6193 * Reserve credits for local agent inode to transfer
6194 * to 0, quota enforcement is ignored in this case.
6196 if (idc->oic_remote &&
6197 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6198 i_projid_read(inode) != 0)
6199 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6200 0, i_projid_read(inode),
6201 0, false, PRJQUOTA, true);
6209 * Index add function for interoperability mode (b11826).
6210 * It will add the directory entry.This entry is needed to
6211 * maintain name->fid mapping.
6213 * \param key it is key i.e. file entry to be inserted
6214 * \param rec it is value of given key i.e. fid
6216 * \retval 0, on success
6217 * \retval -ve, on error
6219 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6220 const struct dt_rec *rec,
6221 const struct dt_key *key, struct thandle *th)
6223 struct osd_object *obj = osd_dt_obj(dt);
6224 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6225 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6226 const struct lu_fid *fid = rec1->rec_fid;
6227 const char *name = (const char *)key;
6228 struct osd_thread_info *oti = osd_oti_get(env);
6229 struct inode *child_inode = NULL;
6230 struct osd_idmap_cache *idc;
6235 if (!dt_object_exists(dt))
6238 LASSERT(osd_invariant(obj));
6239 LASSERT(!dt_object_remote(dt));
6240 LASSERT(th != NULL);
6242 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6244 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6246 idc = osd_idc_find(env, osd, fid);
6247 if (unlikely(idc == NULL)) {
6248 idc = osd_idc_find_or_init(env, osd, fid);
6251 * this dt_insert() wasn't declared properly, so
6252 * FID is missing in OI cache. we better do not
6253 * lookup FID in FLDB/OI and don't risk to deadlock,
6254 * but in some special cases (lfsck testing, etc)
6255 * it's much simpler than fixing a caller.
6257 * normally this error should be placed after the first
6258 * find, but migrate may attach source stripes to
6259 * target, which doesn't create stripes.
6261 CERROR("%s: "DFID" wasn't declared for insert\n",
6262 osd_name(osd), PFID(fid));
6264 RETURN(PTR_ERR(idc));
6268 if (idc->oic_remote) {
6269 /* Insert remote entry */
6270 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6272 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6274 child_inode = osd_create_local_agent_inode(env, osd,
6275 obj, fid, rec1->rec_type & S_IFMT, th);
6276 if (IS_ERR(child_inode))
6277 RETURN(PTR_ERR(child_inode));
6280 /* Insert local entry */
6281 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6282 /* for a reason OI cache wasn't filled properly */
6283 CERROR("%s: OIC for "DFID" isn't filled\n",
6284 osd_name(osd), PFID(fid));
6287 child_inode = oti->oti_inode;
6288 if (unlikely(child_inode == NULL)) {
6289 struct ldiskfs_inode_info *lii;
6294 child_inode = oti->oti_inode = &lii->vfs_inode;
6296 child_inode->i_sb = osd_sb(osd);
6297 child_inode->i_ino = idc->oic_lid.oii_ino;
6298 child_inode->i_mode = rec1->rec_type & S_IFMT;
6301 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6303 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6304 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6306 if (child_inode && child_inode != oti->oti_inode)
6308 LASSERT(osd_invariant(obj));
6309 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6315 * Initialize osd Iterator for given osd index object.
6317 * \param dt osd index object
6320 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6321 struct dt_object *dt,
6324 struct osd_it_iam *it;
6325 struct osd_object *obj = osd_dt_obj(dt);
6326 struct lu_object *lo = &dt->do_lu;
6327 struct iam_path_descr *ipd;
6328 struct iam_container *bag = &obj->oo_dir->od_container;
6330 if (!dt_object_exists(dt))
6331 return ERR_PTR(-ENOENT);
6335 return ERR_PTR(-ENOMEM);
6337 ipd = osd_it_ipd_get(env, bag);
6338 if (likely(ipd != NULL)) {
6342 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6343 return (struct dt_it *)it;
6346 return ERR_PTR(-ENOMEM);
6351 * free given Iterator.
6353 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6355 struct osd_it_iam *it = (struct osd_it_iam *)di;
6356 struct osd_object *obj = it->oi_obj;
6358 iam_it_fini(&it->oi_it);
6359 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6360 osd_object_put(env, obj);
6365 * Move Iterator to record specified by \a key
6367 * \param di osd iterator
6368 * \param key key for index
6370 * \retval +ve di points to record with least key not larger than key
6371 * \retval 0 di points to exact matched key
6372 * \retval -ve failure
6375 static int osd_it_iam_get(const struct lu_env *env,
6376 struct dt_it *di, const struct dt_key *key)
6378 struct osd_thread_info *oti = osd_oti_get(env);
6379 struct osd_it_iam *it = (struct osd_it_iam *)di;
6381 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6382 /* swab quota uid/gid */
6383 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6384 key = (struct dt_key *)&oti->oti_quota_id;
6387 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6393 * \param di osd iterator
6395 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6397 struct osd_it_iam *it = (struct osd_it_iam *)di;
6399 iam_it_put(&it->oi_it);
6403 * Move iterator by one record
6405 * \param di osd iterator
6407 * \retval +1 end of container reached
6409 * \retval -ve failure
6412 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6414 struct osd_it_iam *it = (struct osd_it_iam *)di;
6416 return iam_it_next(&it->oi_it);
6420 * Return pointer to the key under iterator.
6423 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6424 const struct dt_it *di)
6426 struct osd_thread_info *oti = osd_oti_get(env);
6427 struct osd_it_iam *it = (struct osd_it_iam *)di;
6428 struct osd_object *obj = it->oi_obj;
6431 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6433 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6434 /* swab quota uid/gid */
6435 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6436 key = (struct dt_key *)&oti->oti_quota_id;
6443 * Return size of key under iterator (in bytes)
6446 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6448 struct osd_it_iam *it = (struct osd_it_iam *)di;
6450 return iam_it_key_size(&it->oi_it);
6454 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6456 /* check if file type is required */
6457 if (ent->lde_attrs & LUDA_TYPE) {
6458 struct luda_type *lt;
6459 int align = sizeof(*lt) - 1;
6461 len = (len + align) & ~align;
6462 lt = (struct luda_type *)(ent->lde_name + len);
6463 lt->lt_type = cpu_to_le16(DTTOIF(type));
6466 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6470 * build lu direct from backend fs dirent.
6474 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6475 char *name, __u16 namelen, __u16 type, __u32 attr)
6477 ent->lde_attrs = attr | LUDA_FID;
6478 fid_cpu_to_le(&ent->lde_fid, fid);
6480 ent->lde_hash = cpu_to_le64(offset);
6481 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6483 strncpy(ent->lde_name, name, namelen);
6484 ent->lde_name[namelen] = '\0';
6485 ent->lde_namelen = cpu_to_le16(namelen);
6487 /* append lustre attributes */
6488 osd_it_append_attrs(ent, namelen, type);
6492 * Return pointer to the record under iterator.
6494 static int osd_it_iam_rec(const struct lu_env *env,
6495 const struct dt_it *di,
6496 struct dt_rec *dtrec, __u32 attr)
6498 struct osd_it_iam *it = (struct osd_it_iam *)di;
6499 struct osd_thread_info *info = osd_oti_get(env);
6503 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6504 const struct osd_fid_pack *rec;
6505 struct lu_fid *fid = &info->oti_fid;
6506 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6512 name = (char *)iam_it_key_get(&it->oi_it);
6514 RETURN(PTR_ERR(name));
6516 namelen = iam_it_key_size(&it->oi_it);
6518 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6520 RETURN(PTR_ERR(rec));
6522 rc = osd_fid_unpack(fid, rec);
6526 hash = iam_it_store(&it->oi_it);
6528 /* IAM does not store object type in IAM index (dir) */
6529 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6531 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6532 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6533 (struct iam_rec *)dtrec);
6534 osd_quota_unpack(it->oi_obj, dtrec);
6536 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6537 (struct iam_rec *)dtrec);
6544 * Returns cookie for current Iterator position.
6546 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6548 struct osd_it_iam *it = (struct osd_it_iam *)di;
6550 return iam_it_store(&it->oi_it);
6554 * Restore iterator from cookie.
6556 * \param di osd iterator
6557 * \param hash Iterator location cookie
6559 * \retval +ve di points to record with least key not larger than key.
6560 * \retval 0 di points to exact matched key
6561 * \retval -ve failure
6564 static int osd_it_iam_load(const struct lu_env *env,
6565 const struct dt_it *di, __u64 hash)
6567 struct osd_it_iam *it = (struct osd_it_iam *)di;
6569 return iam_it_load(&it->oi_it, hash);
6572 static const struct dt_index_operations osd_index_iam_ops = {
6573 .dio_lookup = osd_index_iam_lookup,
6574 .dio_declare_insert = osd_index_declare_iam_insert,
6575 .dio_insert = osd_index_iam_insert,
6576 .dio_declare_delete = osd_index_declare_iam_delete,
6577 .dio_delete = osd_index_iam_delete,
6579 .init = osd_it_iam_init,
6580 .fini = osd_it_iam_fini,
6581 .get = osd_it_iam_get,
6582 .put = osd_it_iam_put,
6583 .next = osd_it_iam_next,
6584 .key = osd_it_iam_key,
6585 .key_size = osd_it_iam_key_size,
6586 .rec = osd_it_iam_rec,
6587 .store = osd_it_iam_store,
6588 .load = osd_it_iam_load
6594 * Creates or initializes iterator context.
6596 * \retval struct osd_it_ea, iterator structure on success
6599 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6600 struct dt_object *dt,
6603 struct osd_object *obj = osd_dt_obj(dt);
6604 struct osd_thread_info *info = osd_oti_get(env);
6605 struct osd_it_ea *oie;
6607 struct lu_object *lo = &dt->do_lu;
6608 struct dentry *obj_dentry;
6612 if (!dt_object_exists(dt) || obj->oo_destroyed)
6613 RETURN(ERR_PTR(-ENOENT));
6615 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6617 RETURN(ERR_PTR(-ENOMEM));
6618 obj_dentry = &oie->oie_dentry;
6620 obj_dentry->d_inode = obj->oo_inode;
6621 obj_dentry->d_sb = osd_sb(osd_obj2dev(obj));
6622 obj_dentry->d_name.hash = 0;
6624 oie->oie_rd_dirent = 0;
6625 oie->oie_it_dirent = 0;
6626 oie->oie_dirent = NULL;
6627 if (unlikely(!info->oti_it_ea_buf_used)) {
6628 oie->oie_buf = info->oti_it_ea_buf;
6629 info->oti_it_ea_buf_used = 1;
6631 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6632 if (oie->oie_buf == NULL)
6633 RETURN(ERR_PTR(-ENOMEM));
6637 file = &oie->oie_file;
6639 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6640 if (attr & LUDA_64BITHASH)
6641 file->f_mode = FMODE_64BITHASH;
6643 file->f_mode = FMODE_32BITHASH;
6644 file->f_path.dentry = obj_dentry;
6645 file->f_mapping = obj->oo_inode->i_mapping;
6646 file->f_op = obj->oo_inode->i_fop;
6647 file->f_inode = obj->oo_inode;
6650 RETURN((struct dt_it *)oie);
6654 * Destroy or finishes iterator context.
6656 * \param di iterator structure to be destroyed
6658 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6660 struct osd_thread_info *info = osd_oti_get(env);
6661 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6662 struct osd_object *obj = oie->oie_obj;
6663 struct inode *inode = obj->oo_inode;
6666 oie->oie_file.f_op->release(inode, &oie->oie_file);
6667 osd_object_put(env, obj);
6668 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6669 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6671 info->oti_it_ea_buf_used = 0;
6672 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6677 * It position the iterator at given key, so that next lookup continues from
6678 * that key Or it is similar to dio_it->load() but based on a key,
6679 * rather than file position.
6681 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6684 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6686 static int osd_it_ea_get(const struct lu_env *env,
6687 struct dt_it *di, const struct dt_key *key)
6689 struct osd_it_ea *it = (struct osd_it_ea *)di;
6692 LASSERT(((const char *)key)[0] == '\0');
6693 it->oie_file.f_pos = 0;
6694 it->oie_rd_dirent = 0;
6695 it->oie_it_dirent = 0;
6696 it->oie_dirent = NULL;
6704 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6708 struct osd_filldir_cbs {
6709 struct dir_context ctx;
6710 struct osd_it_ea *it;
6713 * It is called internally by ->iterate*(). It fills the
6714 * iterator's in-memory data structure with required
6715 * information i.e. name, namelen, rec_size etc.
6717 * \param buf in which information to be filled in.
6718 * \param name name of the file in given dir
6720 * \retval 0 on success
6721 * \retval 1 on buffer full
6723 #ifdef HAVE_FILLDIR_USE_CTX
6724 static int osd_ldiskfs_filldir(struct dir_context *buf,
6726 static int osd_ldiskfs_filldir(void *buf,
6728 const char *name, int namelen,
6729 loff_t offset, __u64 ino, unsigned int d_type)
6731 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6732 struct osd_object *obj = it->oie_obj;
6733 struct osd_it_ea_dirent *ent = it->oie_dirent;
6734 struct lu_fid *fid = &ent->oied_fid;
6735 struct osd_fid_pack *rec;
6739 /* this should never happen */
6740 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6741 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6745 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6749 /* "." is just the object itself. */
6750 if (namelen == 1 && name[0] == '.') {
6751 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6752 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6753 rec = (struct osd_fid_pack *)(name + namelen + 1);
6754 if (osd_fid_unpack(fid, rec) != 0)
6759 d_type &= ~LDISKFS_DIRENT_LUFID;
6761 /* NOT export local root. */
6762 if (unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6763 ino = obj->oo_inode->i_ino;
6764 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6767 ent->oied_ino = ino;
6768 ent->oied_off = offset;
6769 ent->oied_namelen = namelen;
6770 ent->oied_type = d_type;
6772 memcpy(ent->oied_name, name, namelen);
6774 it->oie_rd_dirent++;
6775 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6780 * Calls ->iterate*() to load a directory entry at a time
6781 * and stored it in iterator's in-memory data structure.
6783 * \param di iterator's in memory structure
6785 * \retval 0 on success
6786 * \retval -ve on error
6787 * \retval +1 reach the end of entry
6789 static int osd_ldiskfs_it_fill(const struct lu_env *env,
6790 const struct dt_it *di)
6792 struct osd_it_ea *it = (struct osd_it_ea *)di;
6793 struct osd_object *obj = it->oie_obj;
6794 struct inode *inode = obj->oo_inode;
6795 struct htree_lock *hlock = NULL;
6796 struct file *filp = &it->oie_file;
6798 struct osd_filldir_cbs buf = {
6799 .ctx.actor = osd_ldiskfs_filldir,
6804 it->oie_dirent = it->oie_buf;
6805 it->oie_rd_dirent = 0;
6807 if (obj->oo_hl_head != NULL) {
6808 hlock = osd_oti_get(env)->oti_hlock;
6809 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6810 inode, LDISKFS_HLOCK_READDIR);
6812 down_read(&obj->oo_ext_idx_sem);
6815 filp->f_cred = current_cred();
6816 rc = osd_security_file_alloc(filp);
6820 filp->f_flags |= O_NOATIME;
6821 filp->f_mode |= FMODE_NONOTIFY;
6822 rc = iterate_dir(filp, &buf.ctx);
6826 if (it->oie_rd_dirent == 0) {
6828 * If it does not get any dirent, it means it has been reached
6829 * to the end of the dir
6831 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6835 it->oie_dirent = it->oie_buf;
6836 it->oie_it_dirent = 1;
6840 ldiskfs_htree_unlock(hlock);
6842 up_read(&obj->oo_ext_idx_sem);
6848 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
6849 * to load a directory entry at a time and stored it in
6850 * iterator's in-memory data structure.
6852 * \param di iterator's in memory structure
6854 * \retval +ve iterator reached to end
6855 * \retval 0 iterator not reached to end
6856 * \retval -ve on error
6858 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6860 struct osd_it_ea *it = (struct osd_it_ea *)di;
6865 if (it->oie_it_dirent < it->oie_rd_dirent) {
6867 (void *)it->oie_dirent +
6868 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6869 it->oie_dirent->oied_namelen);
6870 it->oie_it_dirent++;
6873 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6876 rc = osd_ldiskfs_it_fill(env, di);
6883 * Returns the key at current position from iterator's in memory structure.
6885 * \param di iterator's in memory structure
6887 * \retval key i.e. struct dt_key on success
6889 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6890 const struct dt_it *di)
6892 struct osd_it_ea *it = (struct osd_it_ea *)di;
6894 return (struct dt_key *)it->oie_dirent->oied_name;
6898 * Returns key's size at current position from iterator's in memory structure.
6900 * \param di iterator's in memory structure
6902 * \retval key_size i.e. struct dt_key on success
6904 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6906 struct osd_it_ea *it = (struct osd_it_ea *)di;
6908 return it->oie_dirent->oied_namelen;
6911 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
6912 #undef LDISKFS_DIR_REC_LEN
6913 #define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
6916 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6918 if (LDISKFS_DIR_REC_LEN(de) >=
6919 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6926 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6927 unsigned int blocksize, bool dotdot)
6930 return osd_dotdot_has_space(de);
6932 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6933 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6940 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6941 handle_t *jh, struct dentry *dentry,
6942 const struct lu_fid *fid, struct buffer_head *bh,
6943 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6946 struct inode *dir = dentry->d_parent->d_inode;
6947 struct inode *inode = dentry->d_inode;
6948 struct osd_fid_pack *rec;
6949 struct ldiskfs_dentry_param *ldp;
6950 int namelen = dentry->d_name.len;
6952 struct osd_thread_info *info = osd_oti_get(env);
6956 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
6959 /* There is enough space to hold the FID-in-dirent. */
6960 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
6961 rc = ldiskfs_journal_get_write_access(jh, bh);
6965 de->name[namelen] = 0;
6966 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
6967 rec->fp_len = sizeof(struct lu_fid) + 1;
6968 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
6969 de->file_type |= LDISKFS_DIRENT_LUFID;
6970 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
6977 rc = ldiskfs_delete_entry(jh, dir, de, bh);
6981 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
6982 osd_get_ldiskfs_dirent_param(ldp, fid);
6983 dentry->d_fsdata = (void *)ldp;
6984 dquot_initialize(dir);
6985 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
6987 * It is too bad, we cannot reinsert the name entry back.
6988 * That means we lose it!
6992 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
6993 osd_ino2name(inode), dir->i_ino, dir->i_generation,
6994 namelen, dentry->d_name.name, PFID(fid), rc);
7000 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7001 struct osd_it_ea *it, struct lu_fid *fid,
7002 struct osd_inode_id *id, __u32 *attr)
7004 struct osd_thread_info *info = osd_oti_get(env);
7005 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7006 struct osd_device *dev = osd_obj2dev(obj);
7007 struct super_block *sb = osd_sb(dev);
7008 const char *devname = osd_name(dev);
7009 struct osd_it_ea_dirent *ent = it->oie_dirent;
7010 struct inode *dir = obj->oo_inode;
7011 struct htree_lock *hlock = NULL;
7012 struct buffer_head *bh = NULL;
7013 handle_t *jh = NULL;
7014 struct ldiskfs_dir_entry_2 *de;
7015 struct dentry *dentry;
7016 struct inode *inode;
7017 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7020 bool dotdot = false;
7025 if (ent->oied_name[0] == '.') {
7026 if (ent->oied_namelen == 1)
7029 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7033 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7034 inode = osd_iget(info, dev, id);
7035 if (IS_ERR(inode)) {
7036 rc = PTR_ERR(inode);
7037 if (rc == -ENOENT || rc == -ESTALE) {
7039 * Maybe dangling name entry, or
7040 * corrupted directory entry.
7042 *attr |= LUDA_UNKNOWN;
7045 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7046 "check_repair, dir = %lu/%u, name = %.*s, "
7047 "ino = %llu, rc = %d\n",
7048 devname, dir->i_ino, dir->i_generation,
7049 ent->oied_namelen, ent->oied_name,
7056 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
7058 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7059 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7065 * We need to ensure that the name entry is still valid.
7066 * Because it may be removed or renamed by other already.
7068 * The unlink or rename operation will start journal before PDO lock,
7069 * so to avoid deadlock, here we need to start journal handle before
7070 * related PDO lock also. But because we do not know whether there
7071 * will be something to be repaired before PDO lock, we just start
7072 * journal without conditions.
7074 * We may need to remove the name entry firstly, then insert back.
7075 * One credit is for user quota file update.
7076 * One credit is for group quota file update.
7077 * Two credits are for dirty inode.
7079 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7080 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7082 if (dev->od_dirent_journal != 0) {
7085 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7088 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7089 "check_repair, dir = %lu/%u, credits = %d, "
7090 "name = %.*s, ino = %llu: rc = %d\n",
7091 devname, dir->i_ino, dir->i_generation, credits,
7092 ent->oied_namelen, ent->oied_name,
7095 GOTO(out_inode, rc);
7098 if (obj->oo_hl_head != NULL) {
7099 hlock = osd_oti_get(env)->oti_hlock;
7101 * "0" means exclusive lock for the whole directory.
7102 * We need to prevent others access such name entry
7103 * during the delete + insert. Neither HLOCK_ADD nor
7104 * HLOCK_DEL cannot guarantee the atomicity.
7106 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7108 down_write(&obj->oo_ext_idx_sem);
7111 if (obj->oo_hl_head != NULL) {
7112 hlock = osd_oti_get(env)->oti_hlock;
7113 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7114 LDISKFS_HLOCK_LOOKUP);
7116 down_read(&obj->oo_ext_idx_sem);
7120 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7121 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7122 *attr |= LUDA_IGNORE;
7128 * For dotdot entry, if there is not enough space to hold the
7129 * FID-in-dirent, just keep them there. It only happens when the
7130 * device upgraded from 1.8 or restored from MDT file-level backup.
7131 * For the whole directory, only dotdot entry have no FID-in-dirent
7132 * and needs to get FID from LMA when readdir, it will not affect the
7135 if (dotdot && !osd_dotdot_has_space(de)) {
7136 *attr |= LUDA_UNKNOWN;
7142 if (lu_fid_eq(fid, &lma->lma_self_fid))
7145 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7146 struct lu_fid *tfid = &lma->lma_self_fid;
7148 if (likely(dotdot &&
7149 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7150 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7152 * It must be REMOTE_PARENT_DIR and as the
7153 * 'dotdot' entry of remote directory
7155 *attr |= LUDA_IGNORE;
7157 CDEBUG(D_LFSCK, "%s: expect remote agent "
7158 "parent directory, but got %.*s under "
7159 "dir = %lu/%u with the FID "DFID"\n",
7160 devname, ent->oied_namelen,
7161 ent->oied_name, dir->i_ino,
7162 dir->i_generation, PFID(tfid));
7164 *attr |= LUDA_UNKNOWN;
7171 if (!fid_is_zero(fid)) {
7172 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7174 if (rc == -ENOENT ||
7176 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7178 * linkEA does not recognize the dirent entry,
7179 * it may because the dirent entry corruption
7180 * and points to other's inode.
7182 CDEBUG(D_LFSCK, "%s: the target inode does not "
7183 "recognize the dirent, dir = %lu/%u, "
7184 " name = %.*s, ino = %llu, "
7185 DFID": rc = %d\n", devname, dir->i_ino,
7186 dir->i_generation, ent->oied_namelen,
7187 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7188 *attr |= LUDA_UNKNOWN;
7193 if (rc && rc != -ENODATA) {
7194 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7195 "dir = %lu/%u, name = %.*s, ino = %llu, "
7196 DFID": rc = %d\n", devname, dir->i_ino,
7197 dir->i_generation, ent->oied_namelen,
7198 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7199 *attr |= LUDA_UNKNOWN;
7207 * linkEA recognizes the dirent entry, the FID-in-LMA is
7208 * valid, trusted, in spite of fid_is_sane(fid) or not.
7210 if (*attr & LUDA_VERIFY_DRYRUN) {
7211 *fid = lma->lma_self_fid;
7212 *attr |= LUDA_REPAIR;
7219 dev->od_dirent_journal = 1;
7220 if (hlock != NULL) {
7221 ldiskfs_htree_unlock(hlock);
7224 up_read(&obj->oo_ext_idx_sem);
7230 *fid = lma->lma_self_fid;
7232 /* Update or append the FID-in-dirent. */
7233 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7234 bh, de, hlock, dotdot);
7236 *attr |= LUDA_REPAIR;
7238 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7239 "the dirent, dir = %lu/%u, name = %.*s, "
7240 "ino = %llu, "DFID": rc = %d\n",
7241 devname, dir->i_ino, dir->i_generation,
7242 ent->oied_namelen, ent->oied_name,
7243 ent->oied_ino, PFID(fid), rc);
7245 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7246 if (*attr & LUDA_VERIFY_DRYRUN) {
7247 if (fid_is_sane(fid)) {
7248 *attr |= LUDA_REPAIR;
7249 } else if (dev->od_index == 0) {
7250 lu_igif_build(fid, inode->i_ino,
7251 inode->i_generation);
7252 *attr |= LUDA_UPGRADE;
7260 dev->od_dirent_journal = 1;
7261 if (hlock != NULL) {
7262 ldiskfs_htree_unlock(hlock);
7265 up_read(&obj->oo_ext_idx_sem);
7272 if (unlikely(fid_is_sane(fid))) {
7274 * FID-in-dirent exists, but FID-in-LMA is lost.
7275 * Trust the FID-in-dirent, and add FID-in-LMA.
7277 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7279 *attr |= LUDA_REPAIR;
7281 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7282 "update dirent, dir = %lu/%u, "
7283 "name = %.*s, ino = %llu, "
7285 devname, dir->i_ino, dir->i_generation,
7286 ent->oied_namelen, ent->oied_name,
7287 ent->oied_ino, PFID(fid), rc);
7288 } else if (dev->od_index == 0) {
7289 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7291 * It is probably IGIF object. Only aappend the
7292 * FID-in-dirent. OI scrub will process FID-in-LMA.
7294 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7295 bh, de, hlock, dotdot);
7297 *attr |= LUDA_UPGRADE;
7299 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7300 "after the dirent, dir = %lu/%u, "
7301 "name = %.*s, ino = %llu, "
7303 devname, dir->i_ino, dir->i_generation,
7304 ent->oied_namelen, ent->oied_name,
7305 ent->oied_ino, PFID(fid), rc);
7314 if (hlock != NULL) {
7315 ldiskfs_htree_unlock(hlock);
7317 if (dev->od_dirent_journal != 0)
7318 up_write(&obj->oo_ext_idx_sem);
7320 up_read(&obj->oo_ext_idx_sem);
7324 ldiskfs_journal_stop(jh);
7328 if (rc >= 0 && !dirty)
7329 dev->od_dirent_journal = 0;
7335 * Returns the value at current position from iterator's in memory structure.
7337 * \param di struct osd_it_ea, iterator's in memory structure
7338 * \param attr attr requested for dirent.
7339 * \param lde lustre dirent
7341 * \retval 0 no error and \param lde has correct lustre dirent.
7342 * \retval -ve on error
7344 static inline int osd_it_ea_rec(const struct lu_env *env,
7345 const struct dt_it *di,
7346 struct dt_rec *dtrec, __u32 attr)
7348 struct osd_it_ea *it = (struct osd_it_ea *)di;
7349 struct osd_object *obj = it->oie_obj;
7350 struct osd_device *dev = osd_obj2dev(obj);
7351 struct osd_thread_info *oti = osd_oti_get(env);
7352 struct osd_inode_id *id = &oti->oti_id;
7353 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7354 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7355 __u32 ino = it->oie_dirent->oied_ino;
7360 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7362 if (attr & LUDA_VERIFY) {
7363 if (unlikely(is_remote_parent_ino(dev, ino))) {
7364 attr |= LUDA_IGNORE;
7366 * If the parent is on remote MDT, and there
7367 * is no FID-in-dirent, then we have to get
7368 * the parent FID from the linkEA.
7370 if (!fid_is_sane(fid) &&
7371 it->oie_dirent->oied_namelen == 2 &&
7372 it->oie_dirent->oied_name[0] == '.' &&
7373 it->oie_dirent->oied_name[1] == '.')
7374 osd_get_pfid_from_linkea(env, obj, fid);
7376 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7380 if (!fid_is_sane(fid))
7381 attr |= LUDA_UNKNOWN;
7383 attr &= ~LU_DIRENT_ATTRS_MASK;
7384 if (!fid_is_sane(fid)) {
7385 bool is_dotdot = false;
7387 if (it->oie_dirent->oied_namelen == 2 &&
7388 it->oie_dirent->oied_name[0] == '.' &&
7389 it->oie_dirent->oied_name[1] == '.')
7392 * If the parent is on remote MDT, and there
7393 * is no FID-in-dirent, then we have to get
7394 * the parent FID from the linkEA.
7396 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7397 rc = osd_get_pfid_from_linkea(env, obj, fid);
7399 if (is_dotdot == false &&
7400 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7403 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7408 /* Pack the entry anyway, at least the offset is right. */
7409 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7410 it->oie_dirent->oied_name,
7411 it->oie_dirent->oied_namelen,
7412 it->oie_dirent->oied_type, attr);
7414 RETURN(rc > 0 ? 0 : rc);
7418 * Returns the record size size at current position.
7420 * This function will return record(lu_dirent) size in bytes.
7422 * \param[in] env execution environment
7423 * \param[in] di iterator's in memory structure
7424 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7425 * calculate the lu_dirent size.
7427 * \retval record size(in bytes & in memory) of the current lu_dirent
7430 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7433 struct osd_it_ea *it = (struct osd_it_ea *)di;
7435 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7439 * Returns a cookie for current position of the iterator head, so that
7440 * user can use this cookie to load/start the iterator next time.
7442 * \param di iterator's in memory structure
7444 * \retval cookie for current position, on success
7446 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7448 struct osd_it_ea *it = (struct osd_it_ea *)di;
7450 return it->oie_dirent->oied_off;
7454 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7455 * to load a directory entry at a time and stored it i inn,
7456 * in iterator's in-memory data structure.
7458 * \param di struct osd_it_ea, iterator's in memory structure
7460 * \retval +ve on success
7461 * \retval -ve on error
7463 static int osd_it_ea_load(const struct lu_env *env,
7464 const struct dt_it *di, __u64 hash)
7466 struct osd_it_ea *it = (struct osd_it_ea *)di;
7470 it->oie_file.f_pos = hash;
7472 rc = osd_ldiskfs_it_fill(env, di);
7483 * Index lookup function for interoperability mode (b11826).
7485 * \param key, key i.e. file name to be searched
7487 * \retval +ve, on success
7488 * \retval -ve, on error
7490 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7491 struct dt_rec *rec, const struct dt_key *key)
7493 struct osd_object *obj = osd_dt_obj(dt);
7498 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7499 LINVRNT(osd_invariant(obj));
7501 rc = osd_ea_lookup_rec(env, obj, rec, key);
7508 * Index and Iterator operations for interoperability
7509 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7511 static const struct dt_index_operations osd_index_ea_ops = {
7512 .dio_lookup = osd_index_ea_lookup,
7513 .dio_declare_insert = osd_index_declare_ea_insert,
7514 .dio_insert = osd_index_ea_insert,
7515 .dio_declare_delete = osd_index_declare_ea_delete,
7516 .dio_delete = osd_index_ea_delete,
7518 .init = osd_it_ea_init,
7519 .fini = osd_it_ea_fini,
7520 .get = osd_it_ea_get,
7521 .put = osd_it_ea_put,
7522 .next = osd_it_ea_next,
7523 .key = osd_it_ea_key,
7524 .key_size = osd_it_ea_key_size,
7525 .rec = osd_it_ea_rec,
7526 .rec_size = osd_it_ea_rec_size,
7527 .store = osd_it_ea_store,
7528 .load = osd_it_ea_load
7532 static void *osd_key_init(const struct lu_context *ctx,
7533 struct lu_context_key *key)
7535 struct osd_thread_info *info;
7537 OBD_ALLOC_PTR(info);
7539 return ERR_PTR(-ENOMEM);
7541 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7542 if (info->oti_it_ea_buf == NULL)
7545 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7547 info->oti_hlock = ldiskfs_htree_lock_alloc();
7548 if (info->oti_hlock == NULL)
7554 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7557 return ERR_PTR(-ENOMEM);
7560 static void osd_key_fini(const struct lu_context *ctx,
7561 struct lu_context_key *key, void *data)
7563 struct osd_thread_info *info = data;
7564 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7565 struct osd_idmap_cache *idc = info->oti_ins_cache;
7567 if (info->oti_dio_pages) {
7569 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7570 struct page *page = info->oti_dio_pages[i];
7572 LASSERT(PagePrivate2(page));
7573 LASSERT(PageLocked(page));
7574 ClearPagePrivate2(page);
7579 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7580 PTLRPC_MAX_BRW_PAGES);
7583 if (info->oti_inode != NULL)
7585 if (info->oti_hlock != NULL)
7586 ldiskfs_htree_lock_free(info->oti_hlock);
7587 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7588 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7589 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7590 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7591 lu_buf_free(&info->oti_big_buf);
7593 LASSERT(info->oti_ins_cache_size > 0);
7594 OBD_FREE_PTR_ARRAY(idc, info->oti_ins_cache_size);
7595 info->oti_ins_cache = NULL;
7596 info->oti_ins_cache_size = 0;
7601 static void osd_key_exit(const struct lu_context *ctx,
7602 struct lu_context_key *key, void *data)
7604 struct osd_thread_info *info = data;
7606 LASSERT(info->oti_r_locks == 0);
7607 LASSERT(info->oti_w_locks == 0);
7608 LASSERT(info->oti_txns == 0);
7611 /* type constructor/destructor: osd_type_init, osd_type_fini */
7612 LU_TYPE_INIT_FINI(osd, &osd_key);
7614 struct lu_context_key osd_key = {
7615 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7616 .lct_init = osd_key_init,
7617 .lct_fini = osd_key_fini,
7618 .lct_exit = osd_key_exit
7622 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7623 const char *name, struct lu_device *next)
7625 struct osd_device *osd = osd_dev(d);
7627 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7628 sizeof(osd->od_svname))
7630 return osd_procfs_init(osd, name);
7633 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7635 struct seq_server_site *ss = osd_seq_site(osd);
7640 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7643 if (unlikely(ss == NULL))
7646 OBD_ALLOC_PTR(osd->od_cl_seq);
7647 if (osd->od_cl_seq == NULL)
7650 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7651 osd->od_svname, ss->ss_server_seq);
7653 if (ss->ss_node_id == 0) {
7655 * If the OSD on the sequence controller(MDT0), then allocate
7656 * sequence here, otherwise allocate sequence after connected
7657 * to MDT0 (see mdt_register_lwp_callback()).
7659 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7660 &osd->od_cl_seq->lcs_space, env);
7666 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7668 if (osd->od_cl_seq == NULL)
7671 seq_client_fini(osd->od_cl_seq);
7672 OBD_FREE_PTR(osd->od_cl_seq);
7673 osd->od_cl_seq = NULL;
7676 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7680 /* shutdown quota slave instance associated with the device */
7681 if (o->od_quota_slave_md != NULL) {
7682 struct qsd_instance *qsd = o->od_quota_slave_md;
7684 o->od_quota_slave_md = NULL;
7688 if (o->od_quota_slave_dt != NULL) {
7689 struct qsd_instance *qsd = o->od_quota_slave_dt;
7691 o->od_quota_slave_dt = NULL;
7695 osd_fid_fini(env, o);
7696 osd_scrub_cleanup(env, o);
7701 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7705 if (o->od_mnt != NULL) {
7706 shrink_dcache_sb(osd_sb(o));
7707 osd_sync(env, &o->od_dt_dev);
7708 wait_event(o->od_commit_cb_done,
7709 !atomic_read(&o->od_commit_cb_in_flight));
7718 static int osd_mount(const struct lu_env *env,
7719 struct osd_device *o, struct lustre_cfg *cfg)
7721 const char *name = lustre_cfg_string(cfg, 0);
7722 const char *dev = lustre_cfg_string(cfg, 1);
7724 unsigned long page, s_flags = 0, lmd_flags = 0;
7725 struct page *__page;
7726 struct file_system_type *type;
7727 char *options = NULL;
7729 struct osd_thread_info *info = osd_oti_get(env);
7730 struct lu_fid *fid = &info->oti_fid;
7731 struct inode *inode;
7732 int rc = 0, force_over_1024tb = 0;
7736 if (o->od_mnt != NULL)
7739 if (strlen(dev) >= sizeof(o->od_mntdev))
7741 strcpy(o->od_mntdev, dev);
7743 str = lustre_cfg_buf(cfg, 2);
7744 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
7746 opts = lustre_cfg_string(cfg, 3);
7748 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7749 CERROR("%s: device %s extents feature is not guaranteed to "
7750 "work on big-endian systems. Use \"bigendian_extents\" "
7751 "mount option to override.\n", name, dev);
7755 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7756 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7757 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");
7760 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7761 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7762 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");
7765 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7766 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
7767 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");
7771 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
7772 force_over_1024tb = 1;
7774 __page = alloc_page(GFP_KERNEL);
7776 GOTO(out, rc = -ENOMEM);
7777 page = (unsigned long)page_address(__page);
7778 options = (char *)page;
7781 /* strip out the options for back compatiblity */
7782 static const char * const sout[] = {
7789 /* strip out option we processed in osd */
7790 "bigendian_extents",
7794 "force_over_1024tb",
7798 strncat(options, opts, PAGE_SIZE);
7799 for (rc = 0, str = options; sout[rc]; ) {
7800 char *op = strstr(str, sout[rc]);
7807 if (op == options || *(op - 1) == ',') {
7808 str = op + strlen(sout[rc]);
7809 if (*str == ',' || *str == '\0') {
7810 *str == ',' ? str++ : str;
7811 memmove(op, str, strlen(str) + 1);
7814 for (str = op; *str != ',' && *str != '\0'; str++)
7818 strncat(options, "user_xattr,acl", PAGE_SIZE);
7821 /* Glom up mount options */
7822 if (*options != '\0')
7823 strncat(options, ",", PAGE_SIZE);
7824 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7826 type = get_fs_type("ldiskfs");
7828 CERROR("%s: cannot find ldiskfs module\n", name);
7829 GOTO(out, rc = -ENODEV);
7832 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7833 module_put(type->owner);
7835 if (IS_ERR(o->od_mnt)) {
7836 rc = PTR_ERR(o->od_mnt);
7838 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7842 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7843 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
7844 force_over_1024tb == 0) {
7845 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",
7847 GOTO(out_mnt, rc = -EINVAL);
7850 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7851 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
7854 GOTO(out_mnt, rc = -EOPNOTSUPP);
7857 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7858 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7859 GOTO(out_mnt, rc = -EINVAL);
7862 #ifdef LDISKFS_MOUNT_DIRDATA
7863 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7864 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7865 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7866 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7867 "enabling the dirdata feature. If you do not want to "
7868 "downgrade to Lustre-1.x again, you can enable it via "
7869 "'tune2fs -O dirdata device'\n", name, dev);
7871 inode = osd_sb(o)->s_root->d_inode;
7872 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7873 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7875 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7879 if (lmd_flags & LMD_FLG_NOSCRUB)
7880 o->od_auto_scrub_interval = AS_NEVER;
7882 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7883 /* do not use pagecache with flash-backed storage */
7884 o->od_writethrough_cache = 0;
7885 o->od_read_cache = 0;
7896 __free_page(__page);
7901 static struct lu_device *osd_device_fini(const struct lu_env *env,
7902 struct lu_device *d)
7904 struct osd_device *o = osd_dev(d);
7908 osd_index_backup(env, o, false);
7909 osd_shutdown(env, o);
7911 if (o->od_oi_table != NULL)
7912 osd_oi_fini(osd_oti_get(env), o);
7913 if (o->od_extent_bytes_percpu)
7914 free_percpu(o->od_extent_bytes_percpu);
7915 osd_obj_map_fini(o);
7921 static int osd_device_init0(const struct lu_env *env,
7922 struct osd_device *o,
7923 struct lustre_cfg *cfg)
7925 struct lu_device *l = osd2lu_dev(o);
7926 struct osd_thread_info *info;
7929 bool restored = false;
7932 /* if the module was re-loaded, env can loose its keys */
7933 rc = lu_env_refill((struct lu_env *)env);
7936 info = osd_oti_get(env);
7939 l->ld_ops = &osd_lu_ops;
7940 o->od_dt_dev.dd_ops = &osd_dt_ops;
7942 spin_lock_init(&o->od_osfs_lock);
7943 mutex_init(&o->od_otable_mutex);
7944 INIT_LIST_HEAD(&o->od_orphan_list);
7945 INIT_LIST_HEAD(&o->od_index_backup_list);
7946 INIT_LIST_HEAD(&o->od_index_restore_list);
7947 spin_lock_init(&o->od_lock);
7948 o->od_index_backup_policy = LIBP_NONE;
7950 init_waitqueue_head(&o->od_commit_cb_done);
7952 o->od_read_cache = 1;
7953 o->od_writethrough_cache = 1;
7954 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7955 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
7956 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
7957 o->od_auto_scrub_interval = AS_DEFAULT;
7958 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
7959 o->od_fallocate_zero_blocks = 0;
7961 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
7962 sizeof(o->od_svname));
7963 if (cplen >= sizeof(o->od_svname)) {
7968 o->od_index_backup_stop = 0;
7969 o->od_index = -1; /* -1 means index is invalid */
7970 rc = server_name2index(o->od_svname, &o->od_index, NULL);
7971 if (rc == LDD_F_SV_TYPE_OST)
7974 o->od_full_scrub_ratio = OFSR_DEFAULT;
7975 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
7976 rc = osd_mount(env, o, cfg);
7980 /* Can only check block device after mount */
7981 o->od_nonrotational =
7982 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
7984 rc = osd_obj_map_init(env, o);
7988 rc = lu_site_init(&o->od_site, l);
7990 GOTO(out_compat, rc);
7991 o->od_site.ls_bottom_dev = l;
7993 rc = lu_site_init_finish(&o->od_site);
7997 opts = lustre_cfg_string(cfg, 3);
7998 if (opts && strstr(opts, "resetoi"))
8001 INIT_LIST_HEAD(&o->od_ios_list);
8002 /* setup scrub, including OI files initialization */
8004 rc = osd_scrub_setup(env, o, restored);
8009 rc = osd_procfs_init(o, o->od_svname);
8011 CERROR("%s: can't initialize procfs: rc = %d\n",
8013 GOTO(out_scrub, rc);
8016 LASSERT(l->ld_site->ls_linkage.next != NULL);
8017 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8019 /* initialize quota slave instance */
8020 /* currently it's no need to prepare qsd_instance_md for OST */
8021 if (!o->od_is_ost) {
8022 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8024 o->od_proc_entry, true);
8025 if (IS_ERR(o->od_quota_slave_md)) {
8026 rc = PTR_ERR(o->od_quota_slave_md);
8027 o->od_quota_slave_md = NULL;
8028 GOTO(out_procfs, rc);
8032 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8033 o->od_proc_entry, false);
8035 if (IS_ERR(o->od_quota_slave_dt)) {
8036 if (o->od_quota_slave_md != NULL) {
8037 qsd_fini(env, o->od_quota_slave_md);
8038 o->od_quota_slave_md = NULL;
8041 rc = PTR_ERR(o->od_quota_slave_dt);
8042 o->od_quota_slave_dt = NULL;
8043 GOTO(out_procfs, rc);
8046 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8047 if (!o->od_extent_bytes_percpu) {
8049 GOTO(out_procfs, rc);
8057 osd_scrub_cleanup(env, o);
8059 lu_site_fini(&o->od_site);
8061 osd_obj_map_fini(o);
8068 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8069 struct lu_device_type *t,
8070 struct lustre_cfg *cfg)
8072 struct osd_device *o;
8077 return ERR_PTR(-ENOMEM);
8079 rc = dt_device_init(&o->od_dt_dev, t);
8082 * Because the ctx might be revived in dt_device_init,
8083 * refill the env here
8085 lu_env_refill((struct lu_env *)env);
8086 rc = osd_device_init0(env, o, cfg);
8088 dt_device_fini(&o->od_dt_dev);
8091 if (unlikely(rc != 0))
8094 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8097 static struct lu_device *osd_device_free(const struct lu_env *env,
8098 struct lu_device *d)
8100 struct osd_device *o = osd_dev(d);
8104 /* XXX: make osd top device in order to release reference */
8105 d->ld_site->ls_top_dev = d;
8106 lu_site_purge(env, d->ld_site, -1);
8107 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8108 D_ERROR, lu_cdebug_printer);
8109 lu_site_fini(&o->od_site);
8110 dt_device_fini(&o->od_dt_dev);
8115 static int osd_process_config(const struct lu_env *env,
8116 struct lu_device *d, struct lustre_cfg *cfg)
8118 struct osd_device *o = osd_dev(d);
8124 switch (cfg->lcfg_command) {
8126 rc = osd_mount(env, o, cfg);
8130 * For the case LCFG_PRE_CLEANUP is not called in advance,
8131 * that may happend if hit failure during mount process.
8133 osd_index_backup(env, o, false);
8134 lu_dev_del_linkage(d->ld_site, d);
8135 rc = osd_shutdown(env, o);
8138 LASSERT(&o->od_dt_dev);
8139 count = class_modify_config(cfg, PARAM_OSD,
8140 &o->od_dt_dev.dd_kobj);
8142 count = class_modify_config(cfg, PARAM_OST,
8143 &o->od_dt_dev.dd_kobj);
8144 rc = count > 0 ? 0 : count;
8146 case LCFG_PRE_CLEANUP:
8148 osd_index_backup(env, o,
8149 o->od_index_backup_policy != LIBP_NONE);
8159 static int osd_recovery_complete(const struct lu_env *env,
8160 struct lu_device *d)
8162 struct osd_device *osd = osd_dev(d);
8167 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8171 * start qsd instance on recovery completion, this notifies the quota
8172 * slave code that we are about to process new requests now
8174 rc = qsd_start(env, osd->od_quota_slave_dt);
8175 if (rc == 0 && osd->od_quota_slave_md != NULL)
8176 rc = qsd_start(env, osd->od_quota_slave_md);
8182 * we use exports to track all osd users
8184 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8185 struct obd_device *obd, struct obd_uuid *cluuid,
8186 struct obd_connect_data *data, void *localdata)
8188 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8189 struct lustre_handle conn;
8194 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8196 rc = class_connect(&conn, obd, cluuid);
8200 *exp = class_conn2export(&conn);
8202 spin_lock(&osd->od_osfs_lock);
8204 spin_unlock(&osd->od_osfs_lock);
8210 * once last export (we don't count self-export) disappeared
8211 * osd can be released
8213 static int osd_obd_disconnect(struct obd_export *exp)
8215 struct obd_device *obd = exp->exp_obd;
8216 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8217 int rc, release = 0;
8221 /* Only disconnect the underlying layers on the final disconnect. */
8222 spin_lock(&osd->od_osfs_lock);
8224 if (osd->od_connects == 0)
8226 spin_unlock(&osd->od_osfs_lock);
8228 rc = class_disconnect(exp); /* bz 9811 */
8230 if (rc == 0 && release)
8231 class_manual_cleanup(obd);
8235 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8236 struct lu_device *dev)
8238 struct osd_device *osd = osd_dev(dev);
8239 struct lr_server_data *lsd =
8240 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8245 if (osd->od_quota_slave_md != NULL) {
8246 /* set up quota slave objects for inode */
8247 result = qsd_prepare(env, osd->od_quota_slave_md);
8252 if (osd->od_quota_slave_dt != NULL) {
8253 /* set up quota slave objects for block */
8254 result = qsd_prepare(env, osd->od_quota_slave_dt);
8260 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8261 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8262 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8263 osd->od_index_in_idif = 1;
8265 osd->od_index_in_idif = 0;
8266 result = osd_register_proc_index_in_idif(osd);
8271 osd->od_index_in_idif = 1;
8275 result = osd_fid_init(env, osd);
8281 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8285 * see include/lu_object.h for the details.
8287 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8288 struct lu_fid *fid, struct lu_object *parent,
8289 const struct lu_name *name)
8291 struct osd_device *osd = osd_dev(d);
8293 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8296 static const struct lu_object_operations osd_lu_obj_ops = {
8297 .loo_object_init = osd_object_init,
8298 .loo_object_delete = osd_object_delete,
8299 .loo_object_release = osd_object_release,
8300 .loo_object_free = osd_object_free,
8301 .loo_object_print = osd_object_print,
8302 .loo_object_invariant = osd_object_invariant
8305 const struct lu_device_operations osd_lu_ops = {
8306 .ldo_object_alloc = osd_object_alloc,
8307 .ldo_process_config = osd_process_config,
8308 .ldo_recovery_complete = osd_recovery_complete,
8309 .ldo_prepare = osd_prepare,
8310 .ldo_fid_alloc = osd_fid_alloc,
8313 static const struct lu_device_type_operations osd_device_type_ops = {
8314 .ldto_init = osd_type_init,
8315 .ldto_fini = osd_type_fini,
8317 .ldto_start = osd_type_start,
8318 .ldto_stop = osd_type_stop,
8320 .ldto_device_alloc = osd_device_alloc,
8321 .ldto_device_free = osd_device_free,
8323 .ldto_device_init = osd_device_init,
8324 .ldto_device_fini = osd_device_fini
8327 static struct lu_device_type osd_device_type = {
8328 .ldt_tags = LU_DEVICE_DT,
8329 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8330 .ldt_ops = &osd_device_type_ops,
8331 .ldt_ctx_tags = LCT_LOCAL,
8334 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8336 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8337 struct super_block *sb = osd_sb(osd);
8339 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8343 * lprocfs legacy support.
8345 static const struct obd_ops osd_obd_device_ops = {
8346 .o_owner = THIS_MODULE,
8347 .o_connect = osd_obd_connect,
8348 .o_disconnect = osd_obd_disconnect,
8349 .o_health_check = osd_health_check,
8352 static ssize_t track_declares_assert_show(struct kobject *kobj,
8353 struct attribute *attr,
8356 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8359 static ssize_t track_declares_assert_store(struct kobject *kobj,
8360 struct attribute *attr,
8361 const char *buffer, size_t count)
8363 bool track_declares_assert;
8366 rc = kstrtobool(buffer, &track_declares_assert);
8370 ldiskfs_track_declares_assert = track_declares_assert;
8374 LUSTRE_RW_ATTR(track_declares_assert);
8376 static int __init osd_init(void)
8378 struct kobject *kobj;
8381 BUILD_BUG_ON(BH_DXLock >=
8382 sizeof(((struct buffer_head *)0)->b_state) * 8);
8383 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8384 /* please, try to keep osd_thread_info smaller than a page */
8385 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8390 rc = lu_kmem_init(ldiskfs_caches);
8394 #ifdef CONFIG_KALLSYMS
8395 priv_security_file_alloc =
8396 (void *)kallsyms_lookup_name("security_file_alloc");
8399 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8400 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8402 lu_kmem_fini(ldiskfs_caches);
8406 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8408 rc = sysfs_create_file(kobj,
8409 &lustre_attr_track_declares_assert.attr);
8412 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8419 static void __exit osd_exit(void)
8421 struct kobject *kobj;
8423 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8425 sysfs_remove_file(kobj,
8426 &lustre_attr_track_declares_assert.attr);
8429 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8430 lu_kmem_fini(ldiskfs_caches);
8433 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8434 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8435 MODULE_VERSION(LUSTRE_VERSION_STRING);
8436 MODULE_LICENSE("GPL");
8438 module_init(osd_init);
8439 module_exit(osd_exit);