4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/osd/osd_handler.c
33 * Top-level entry points into osd module
35 * Author: Nikita Danilov <nikita@clusterfs.com>
36 * Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
39 #define DEBUG_SUBSYSTEM S_OSD
41 #include <linux/fs_struct.h>
42 #include <linux/module.h>
43 #include <linux/user_namespace.h>
44 #include <linux/uidgid.h>
46 /* prerequisite for linux/xattr.h */
47 #include <linux/types.h>
48 /* prerequisite for linux/xattr.h */
50 /* XATTR_{REPLACE,CREATE} */
51 #include <linux/xattr.h>
53 #include <ldiskfs/ldiskfs.h>
54 #include <ldiskfs/xattr.h>
55 #include <ldiskfs/ldiskfs_extents.h>
58 * struct OBD_{ALLOC,FREE}*()
61 #include <obd_support.h>
62 /* struct ptlrpc_thread */
63 #include <lustre_net.h>
64 #include <lustre_fid.h>
66 #include <uapi/linux/lustre/lustre_param.h>
67 #include <uapi/linux/lustre/lustre_disk.h>
69 #include "osd_internal.h"
70 #include "osd_dynlocks.h"
72 /* llo_* api support */
73 #include <md_object.h>
74 #include <lustre_quota.h>
76 #include <lustre_linkea.h>
78 /* encoding routines */
79 #include <lustre_crypto.h>
81 /* Maximum EA size is limited by LNET_MTU for remote objects */
82 #define OSD_MAX_EA_SIZE 1048364
85 module_param(ldiskfs_pdo, int, 0644);
86 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
88 int ldiskfs_track_declares_assert;
89 module_param(ldiskfs_track_declares_assert, int, 0644);
90 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
92 /* Slab to allocate dynlocks */
93 struct kmem_cache *dynlock_cachep;
95 /* Slab to allocate osd_it_ea */
96 struct kmem_cache *osd_itea_cachep;
98 static struct lu_kmem_descr ldiskfs_caches[] = {
100 .ckd_cache = &dynlock_cachep,
101 .ckd_name = "dynlock_cache",
102 .ckd_size = sizeof(struct dynlock_handle)
105 .ckd_cache = &osd_itea_cachep,
106 .ckd_name = "osd_itea_cache",
107 .ckd_size = sizeof(struct osd_it_ea)
114 static const char dot[] = ".";
115 static const char dotdot[] = "..";
117 static const struct lu_object_operations osd_lu_obj_ops;
118 static const struct dt_object_operations osd_obj_ops;
119 static const struct dt_object_operations osd_obj_otable_it_ops;
120 static const struct dt_index_operations osd_index_iam_ops;
121 static const struct dt_index_operations osd_index_ea_ops;
123 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
124 const struct lu_fid *fid);
125 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
126 struct osd_device *osd);
127 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
128 const struct lu_buf *buf, const char *name, int fl,
129 struct thandle *handle);
131 int osd_trans_declare_op2rb[] = {
132 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
133 [OSD_OT_PUNCH] = OSD_OT_MAX,
134 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
135 [OSD_OT_CREATE] = OSD_OT_DESTROY,
136 [OSD_OT_DESTROY] = OSD_OT_CREATE,
137 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
138 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
139 [OSD_OT_WRITE] = OSD_OT_WRITE,
140 [OSD_OT_INSERT] = OSD_OT_DELETE,
141 [OSD_OT_DELETE] = OSD_OT_INSERT,
142 [OSD_OT_QUOTA] = OSD_OT_MAX,
145 static int osd_has_index(const struct osd_object *obj)
147 return obj->oo_dt.do_index_ops != NULL;
150 static int osd_object_invariant(const struct lu_object *l)
152 return osd_invariant(osd_obj(l));
156 * Concurrency: doesn't matter
158 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
160 struct osd_thread_info *oti = osd_oti_get(env);
162 return oti->oti_w_locks > 0 && o->oo_owner == env;
166 * Concurrency: doesn't access mutable data
168 static int osd_root_get(const struct lu_env *env,
169 struct dt_device *dev, struct lu_fid *f)
171 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
176 * the following set of functions are used to maintain per-thread
177 * cache of FID->ino mapping. this mechanism is needed to resolve
178 * FID to inode at dt_insert() which in turn stores ino in the
179 * directory entries to keep ldiskfs compatible with ext[34].
180 * due to locking-originated restrictions we can't lookup ino
181 * using LU cache (deadlock is possible). lookup using OI is quite
182 * expensive. so instead we maintain this cache and methods like
183 * dt_create() fill it. so in the majority of cases dt_insert() is
184 * able to find needed mapping in lockless manner.
186 static struct osd_idmap_cache *
187 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
188 const struct lu_fid *fid)
190 struct osd_thread_info *oti = osd_oti_get(env);
191 struct osd_idmap_cache *idc = oti->oti_ins_cache;
194 for (i = 0; i < oti->oti_ins_cache_used; i++) {
195 if (!lu_fid_eq(&idc[i].oic_fid, fid))
197 if (idc[i].oic_dev != osd)
206 static struct osd_idmap_cache *
207 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
208 const struct lu_fid *fid)
210 struct osd_thread_info *oti = osd_oti_get(env);
211 struct osd_idmap_cache *idc;
214 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
215 i = oti->oti_ins_cache_size * 2;
217 i = OSD_INS_CACHE_SIZE;
218 OBD_ALLOC_PTR_ARRAY_LARGE(idc, i);
220 return ERR_PTR(-ENOMEM);
221 if (oti->oti_ins_cache != NULL) {
222 memcpy(idc, oti->oti_ins_cache,
223 oti->oti_ins_cache_used * sizeof(*idc));
224 OBD_FREE_PTR_ARRAY_LARGE(oti->oti_ins_cache,
225 oti->oti_ins_cache_used);
227 oti->oti_ins_cache = idc;
228 oti->oti_ins_cache_size = i;
231 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
234 idc->oic_lid.oii_ino = 0;
235 idc->oic_lid.oii_gen = 0;
242 * lookup mapping for the given fid in the cache, initialize a
243 * new one if not found. the initialization checks whether the
244 * object is local or remote. for local objects, OI is used to
245 * learn ino/generation. the function is used when the caller
246 * has no information about the object, e.g. at dt_insert().
248 static struct osd_idmap_cache *
249 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
250 const struct lu_fid *fid)
252 struct osd_idmap_cache *idc;
255 idc = osd_idc_find(env, osd, fid);
256 LASSERT(!IS_ERR(idc));
260 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
261 osd->od_svname, PFID(fid));
263 /* new mapping is needed */
264 idc = osd_idc_add(env, osd, fid);
266 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
267 osd->od_svname, PFID(fid), PTR_ERR(idc));
272 rc = osd_remote_fid(env, osd, fid);
273 if (unlikely(rc < 0))
277 /* the object is local, lookup in OI */
278 /* XXX: probably cheaper to lookup in LU first? */
279 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
281 if (unlikely(rc < 0)) {
282 CERROR("can't lookup: rc = %d\n", rc);
286 /* the object is remote */
293 static void osd_idc_dump_lma(const struct lu_env *env,
294 struct osd_device *osd,
298 struct osd_thread_info *info = osd_oti_get(env);
299 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
300 const struct lu_fid *fid;
301 struct osd_inode_id lid;
305 inode = osd_ldiskfs_iget(osd_sb(osd), ino);
307 CERROR("%s: can't get inode %lu: rc = %d\n",
308 osd->od_svname, ino, (int)PTR_ERR(inode));
311 if (is_bad_inode(inode)) {
312 CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
315 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
317 CERROR("%s: can't get LMA for %lu: rc = %d\n",
318 osd->od_svname, ino, rc);
321 fid = &loa->loa_lma.lma_self_fid;
322 LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
323 PFID(fid), ino, (unsigned)inode->i_generation);
326 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
328 CERROR("%s: can't lookup "DFID": rc = %d\n",
329 osd->od_svname, PFID(fid), rc);
332 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
333 PFID(fid), lid.oii_ino, lid.oii_gen);
338 static void osd_idc_dump_debug(const struct lu_env *env,
339 struct osd_device *osd,
340 const struct lu_fid *fid,
344 struct osd_inode_id lid;
348 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
350 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
351 osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
352 osd_idc_dump_lma(env, osd, lid.oii_ino, false);
354 CERROR("%s: can't lookup "DFID": rc = %d\n",
355 osd->od_svname, PFID(fid), rc);
358 osd_idc_dump_lma(env, osd, ino1, true);
360 osd_idc_dump_lma(env, osd, ino2, true);
364 * lookup mapping for given FID and fill it from the given object.
365 * the object is lolcal by definition.
367 static int osd_idc_find_and_init(const struct lu_env *env,
368 struct osd_device *osd,
369 struct osd_object *obj)
371 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
372 struct osd_idmap_cache *idc;
374 idc = osd_idc_find(env, osd, fid);
375 LASSERT(!IS_ERR(idc));
377 if (obj->oo_inode == NULL)
379 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
380 if (idc->oic_lid.oii_ino) {
381 osd_idc_dump_debug(env, osd, fid,
382 idc->oic_lid.oii_ino,
383 obj->oo_inode->i_ino);
386 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
387 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
392 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
393 osd->od_svname, PFID(fid));
395 /* new mapping is needed */
396 idc = osd_idc_add(env, osd, fid);
398 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
399 osd->od_svname, PFID(fid), PTR_ERR(idc));
403 if (obj->oo_inode != NULL) {
404 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
405 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
411 * OSD object methods.
415 * Concurrency: no concurrent access is possible that early in object
418 static struct lu_object *osd_object_alloc(const struct lu_env *env,
419 const struct lu_object_header *hdr,
422 struct osd_object *mo;
427 struct lu_object_header *h;
428 struct osd_device *o = osd_dev(d);
430 l = &mo->oo_dt.do_lu;
431 if (unlikely(o->od_in_init)) {
438 lu_object_header_init(h);
439 lu_object_init(l, h, d);
440 lu_object_add_top(h, l);
443 dt_object_init(&mo->oo_dt, NULL, d);
444 mo->oo_header = NULL;
447 mo->oo_dt.do_ops = &osd_obj_ops;
448 l->lo_ops = &osd_lu_obj_ops;
449 init_rwsem(&mo->oo_sem);
450 init_rwsem(&mo->oo_ext_idx_sem);
451 spin_lock_init(&mo->oo_guard);
452 INIT_LIST_HEAD(&mo->oo_xattr_list);
458 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
459 struct dentry *dentry, struct lustre_ost_attrs *loa)
463 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
464 (void *)loa, sizeof(*loa));
466 struct lustre_mdt_attrs *lma = &loa->loa_lma;
468 if (rc < sizeof(*lma))
472 lustre_loa_swab(loa, true);
473 /* Check LMA compatibility */
474 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
476 CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
478 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
479 PFID(&lma->lma_self_fid), inode->i_ino, rc);
481 } else if (rc == 0) {
489 * retrieve object from backend ext fs.
491 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
492 struct osd_inode_id *id)
495 struct inode *inode = NULL;
498 * if we look for an inode withing a running
499 * transaction, then we risk to deadlock
500 * osd_dirent_check_repair() breaks this
502 /* LASSERT(current->journal_info == NULL); */
504 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
506 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
507 id->oii_ino, PTR_ERR(inode));
508 } else if (id->oii_gen != OSD_OII_NOGEN &&
509 inode->i_generation != id->oii_gen) {
510 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
511 "i_generation = %u\n",
512 id->oii_ino, id->oii_gen, inode->i_generation);
514 inode = ERR_PTR(-ESTALE);
515 } else if (inode->i_nlink == 0) {
517 * due to parallel readdir and unlink,
518 * we can have dead inode here.
520 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
522 inode = ERR_PTR(-ESTALE);
523 } else if (is_bad_inode(inode)) {
524 CWARN("%s: bad inode: ino = %u: rc = %d\n",
525 osd_dev2name(dev), id->oii_ino, -ENOENT);
527 inode = ERR_PTR(-ENOENT);
528 } else if (osd_is_ea_inode(inode)) {
530 * EA inode is internal ldiskfs object, should don't visible
533 CDEBUG(D_INODE, "EA inode: ino = %u\n", id->oii_ino);
535 inode = ERR_PTR(-ENOENT);
536 } else if ((rc = osd_attach_jinode(inode))) {
540 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
541 if (id->oii_gen == OSD_OII_NOGEN)
542 osd_id_gen(id, inode->i_ino, inode->i_generation);
545 * Do not update file c/mtime in ldiskfs.
546 * NB: we don't have any lock to protect this because we don't
547 * have reference on osd_object now, but contention with
548 * another lookup + attr_set can't happen in the tiny window
549 * between if (...) and set S_NOCMTIME.
551 if (!(inode->i_flags & S_NOCMTIME))
552 inode->i_flags |= S_NOCMTIME;
557 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
558 handle_t *handle, struct dentry *child,
559 struct inode *inode, struct htree_lock *hlock)
563 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
564 if (rc == -ENOBUFS || rc == -ENOSPC) {
565 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
566 struct inode *parent = child->d_parent->d_inode;
567 struct lu_fid *fid = NULL;
568 char fidstr[FID_LEN + 1] = "unknown";
570 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
572 fid = &loa->loa_lma.lma_self_fid;
573 } else if (rc2 == -ENODATA) {
574 if (unlikely(is_root_inode(parent))) {
575 fid = &info->oti_fid3;
576 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
577 } else if (!osd->od_is_ost && osd->od_index == 0) {
578 fid = &info->oti_fid3;
579 lu_igif_build(fid, parent->i_ino,
580 parent->i_generation);
585 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
587 /* below message is checked in sanity.sh test_129 */
589 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
590 osd_name(osd), parent->i_ino, fidstr);
592 rc = 0; /* ignore such error now */
593 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
594 osd_name(osd), parent->i_ino, fidstr);
604 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
605 struct osd_inode_id *id, struct lu_fid *fid)
607 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
611 inode = osd_iget(info, dev, id);
615 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
617 *fid = loa->loa_lma.lma_self_fid;
618 } else if (rc == -ENODATA) {
619 if (unlikely(is_root_inode(inode)))
620 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
622 lu_igif_build(fid, inode->i_ino, inode->i_generation);
630 static struct inode *osd_iget_check(struct osd_thread_info *info,
631 struct osd_device *dev,
632 const struct lu_fid *fid,
633 struct osd_inode_id *id,
642 * The cached OI mapping is trustable. If we cannot locate the inode
643 * via the cached OI mapping, then return the failure to the caller
644 * directly without further OI checking.
648 inode = osd_iget(info, dev, id);
651 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
654 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
655 PFID(fid), id->oii_ino, rc);
661 __u32 saved_ino = id->oii_ino;
662 __u32 saved_gen = id->oii_gen;
665 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
667 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
669 * XXX: There are four possible cases:
671 * Backup/restore caused the OI invalid.
673 * Someone unlinked the object but NOT removed
674 * the OI mapping, such as mount target device
675 * as ldiskfs, and modify something directly.
677 * Someone just removed the object between the
678 * former oi_lookup and the iget. It is normal.
679 * 4. Other failure cases.
681 * Generally, when the device is mounted, it will
682 * auto check whether the system is restored from
683 * file-level backup or not. We trust such detect
684 * to distinguish the 1st case from the 2nd case:
685 * if the OI files are consistent but may contain
686 * stale OI mappings because of case 2, if iget()
687 * returns -ENOENT or -ESTALE, then it should be
692 * If the OI mapping was in OI file before the
693 * osd_iget_check(), but now, it is disappear,
694 * then it must be removed by race. That is a
700 * It is the OI scrub updated the OI mapping by race.
701 * The new OI mapping must be valid.
703 if (saved_ino != id->oii_ino ||
704 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
713 if (dev->od_scrub.os_scrub.os_file.sf_flags &
716 * It still can be the case 2, but we cannot
717 * distinguish it from the case 1. So return
718 * -EREMCHG to block current operation until
719 * OI scrub rebuilt the OI mappings.
728 if (inode->i_generation == id->oii_gen)
748 * \retval +v: new filter_fid does not contain self-fid
749 * \retval 0: filter_fid_18_23, contains self-fid
750 * \retval -v: other failure cases
752 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
753 struct dentry *dentry, struct lu_fid *fid)
755 struct filter_fid *ff = &info->oti_ff;
756 struct ost_id *ostid = &info->oti_ostid;
759 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
760 if (rc == sizeof(struct filter_fid_18_23)) {
761 struct filter_fid_18_23 *ff_old = (void *)ff;
763 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
764 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
766 * XXX: use 0 as the index for compatibility, the caller will
767 * handle index related issues when necessary.
770 ostid_to_fid(fid, ostid, 0);
771 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
773 } else if (rc >= 0) {
780 static int osd_lma_self_repair(struct osd_thread_info *info,
781 struct osd_device *osd, struct inode *inode,
782 const struct lu_fid *fid, __u32 compat)
787 LASSERT(current->journal_info == NULL);
789 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
790 osd_dto_credits_noquota[DTO_XATTR_SET]);
793 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
798 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
800 CWARN("%s: cannot self repair the LMA: rc = %d\n",
802 ldiskfs_journal_stop(jh);
806 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
808 struct osd_thread_info *info = osd_oti_get(env);
809 struct osd_device *osd = osd_obj2dev(obj);
810 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
811 struct lustre_mdt_attrs *lma = &loa->loa_lma;
812 struct inode *inode = obj->oo_inode;
813 struct dentry *dentry = &info->oti_obj_dentry;
814 struct lu_fid *fid = NULL;
815 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
820 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
821 (void *)loa, sizeof(*loa));
822 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
823 fid = &lma->lma_self_fid;
824 rc = osd_get_idif(info, inode, dentry, fid);
825 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
827 * For the given OST-object, if it has neither LMA nor
828 * FID in XATTR_NAME_FID, then the given FID (which is
829 * contained in the @obj, from client RPC for locating
830 * the OST-object) is trusted. We use it to generate
833 osd_lma_self_repair(info, osd, inode, rfid,
844 lustre_lma_swab(lma);
845 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
846 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
847 S_ISREG(inode->i_mode)))) {
848 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
849 "fid = "DFID", ino = %lu\n", osd_name(osd),
850 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
851 PFID(rfid), inode->i_ino);
854 fid = &lma->lma_self_fid;
855 if (lma->lma_compat & LMAC_STRIPE_INFO &&
857 obj->oo_pfid_in_lma = 1;
858 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
860 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
864 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
865 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
866 struct ost_id *oi = &info->oti_ostid;
867 struct lu_fid *fid1 = &info->oti_fid3;
868 __u32 idx = fid_idif_ost_idx(rfid);
871 * For old IDIF, the OST index is not part of the IDIF,
872 * Means that different OSTs may have the same IDIFs.
873 * Under such case, we need to make some compatible
874 * check to make sure to trigger OI scrub properly.
876 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
877 /* Given @rfid is new, LMA is old. */
878 fid_to_ostid(fid, oi);
879 ostid_to_fid(fid1, oi, idx);
880 if (lu_fid_eq(fid1, rfid)) {
881 if (osd->od_index_in_idif)
882 osd_lma_self_repair(info, osd,
896 struct osd_check_lmv_buf {
897 /* please keep it as first member */
898 struct dir_context ctx;
899 struct osd_thread_info *oclb_info;
900 struct osd_device *oclb_dev;
906 * It is called internally by ->iterate*() to filter out the
907 * local slave object's FID of the striped directory.
909 * \retval 1 found the local slave's FID
910 * \retval 0 continue to check next item
911 * \retval -ve for failure
913 #ifdef HAVE_FILLDIR_USE_CTX
914 static FILLDIR_TYPE do_osd_stripe_dir_filldir(struct dir_context *buf,
916 static int osd_stripe_dir_filldir(void *buf,
918 const char *name, int namelen,
919 loff_t offset, __u64 ino, unsigned int d_type)
921 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
922 struct osd_thread_info *oti = oclb->oclb_info;
923 struct lu_fid *fid = &oti->oti_fid3;
924 struct osd_inode_id *id = &oti->oti_id3;
925 struct osd_device *dev = oclb->oclb_dev;
934 sscanf(name + 1, SFID, RFID(fid));
935 if (!fid_is_sane(fid))
938 if (osd_remote_fid(oti->oti_env, dev, fid))
941 osd_id_gen(id, ino, OSD_OII_NOGEN);
942 inode = osd_iget(oti, dev, id);
944 return PTR_ERR(inode);
947 osd_add_oi_cache(oti, dev, id, fid);
948 osd_scrub_oi_insert(dev, fid, id, true);
949 oclb->oclb_found = true;
954 WRAP_FILLDIR_FN(do_, osd_stripe_dir_filldir)
957 * When lookup item under striped directory, we need to locate the master
958 * MDT-object of the striped directory firstly, then the client will send
959 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
960 * and the item's name. If the system is restored from MDT file level backup,
961 * then before the OI scrub completely built the OI files, the OI mappings of
962 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
963 * not a problem for the master MDT-object. Because when locate the master
964 * MDT-object, we will do name based lookup (for the striped directory itself)
965 * firstly, during such process we can setup the correct OI mapping for the
966 * master MDT-object. But it will be trouble for the slave MDT-object. Because
967 * the client will not trigger name based lookup on the MDT to locate the slave
968 * MDT-object before locating item under the striped directory, then when
969 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
970 * is invalid and does not know what the right OI mapping is, then the MDT has
971 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
972 * the OI file, related OI mapping is unknown yet, please try again later. And
973 * then client will re-try the RPC again and again until related OI mapping has
974 * been updated. That is quite inefficient.
976 * To resolve above trouble, we will handle it as the following two cases:
978 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
979 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
980 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
981 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
982 * directly. Please check osd_fid_lookup().
984 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
985 * Under such case, during lookup the master MDT-object, we will lookup the
986 * slave MDT-object via readdir against the master MDT-object, because the
987 * slave MDT-objects information are stored as sub-directories with the name
988 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
989 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
990 * the correct OI mapping for the slave MDT-object.
992 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
995 struct lu_buf *buf = &oti->oti_big_buf;
997 struct lmv_mds_md_v1 *lmv1;
998 struct osd_check_lmv_buf oclb = {
999 .ctx.actor = osd_stripe_dir_filldir,
1002 .oclb_found = false,
1007 /* We should use the VFS layer to create a real dentry. */
1008 oti->oti_obj_dentry.d_inode = inode;
1009 oti->oti_obj_dentry.d_sb = inode->i_sb;
1011 filp = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
1016 filp->f_mode |= FMODE_64BITHASH;
1020 rc = __osd_xattr_get(inode, filp->f_path.dentry, XATTR_NAME_LMV,
1021 buf->lb_buf, buf->lb_len);
1022 if (rc == -ERANGE) {
1023 rc = __osd_xattr_get(inode, filp->f_path.dentry,
1024 XATTR_NAME_LMV, NULL, 0);
1026 lu_buf_realloc(buf, rc);
1027 if (buf->lb_buf == NULL)
1028 GOTO(out, rc = -ENOMEM);
1034 if (unlikely(rc == 0 || rc == -ENODATA))
1040 if (unlikely(buf->lb_buf == NULL)) {
1041 lu_buf_realloc(buf, rc);
1042 if (buf->lb_buf == NULL)
1043 GOTO(out, rc = -ENOMEM);
1049 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1053 oclb.oclb_items = 0;
1054 rc = iterate_dir(filp, &oclb.ctx);
1055 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1056 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1061 "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
1062 osd_ino2name(inode), inode->i_ino, inode->i_generation,
1071 * Is object in scrub inconsistent/stale list.
1073 * \a scrub has two lists, os_inconsistent_items contains mappings to fix, while
1074 * os_stale_items contains mappings failed to fix.
1076 static bool fid_in_scrub_list(struct lustre_scrub *scrub,
1077 const struct list_head *list,
1078 const struct lu_fid *fid)
1080 struct osd_inconsistent_item *oii;
1082 if (list_empty(list))
1085 spin_lock(&scrub->os_lock);
1086 list_for_each_entry(oii, list, oii_list) {
1087 if (lu_fid_eq(fid, &oii->oii_cache.oic_fid)) {
1088 spin_unlock(&scrub->os_lock);
1092 spin_unlock(&scrub->os_lock);
1097 static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
1098 const struct lu_fid *fid,
1099 const struct lu_object_conf *conf)
1101 struct osd_thread_info *info;
1102 struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
1103 struct osd_device *dev;
1104 struct osd_idmap_cache *oic;
1105 struct osd_inode_id *id;
1106 struct inode *inode = NULL;
1107 struct lustre_scrub *scrub;
1108 struct scrub_file *sf;
1109 __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
1114 bool remote = false;
1115 bool trusted = true;
1116 bool updated = false;
1117 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)
1152 /* Search order: 1. per-thread cache. */
1153 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1158 /* Search order: 2. OI scrub pending list. */
1160 memset(id, 0, sizeof(struct osd_inode_id));
1161 if (fid_in_scrub_list(scrub, &scrub->os_inconsistent_items, fid) &&
1163 RETURN(-EINPROGRESS);
1165 stale = fid_in_scrub_list(scrub, &scrub->os_stale_items, fid);
1166 if (stale && OBD_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1170 * The OI mapping in the OI file can be updated by the OI scrub
1171 * when we locate the inode via FID. So it may be not trustable.
1175 /* Search order: 3. OI files. */
1176 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1177 if (result == -ENOENT) {
1178 if (!fid_is_norm(fid) ||
1179 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1180 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1182 GOTO(out, result = 0);
1187 /* -ESTALE is returned if inode of OST object doesn't exist */
1188 if (result == -ESTALE &&
1189 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1190 GOTO(out, result = 0);
1197 obj->oo_inode = NULL;
1198 /* for later passes through checks, not true on first pass */
1199 if (!IS_ERR_OR_NULL(inode))
1202 inode = osd_iget_check(info, dev, fid, id, trusted);
1203 if (!IS_ERR(inode)) {
1204 obj->oo_inode = inode;
1212 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_SCRUB_STALE))
1215 result = PTR_ERR(inode);
1216 if (result == -ENOENT || result == -ESTALE)
1217 GOTO(out, result = 0);
1219 if (result != -EREMCHG)
1223 /* don't trigger repeatedly for stale mapping */
1225 GOTO(out, result = -ESTALE);
1228 * We still have chance to get the valid inode: for the
1229 * object which is referenced by remote name entry, the
1230 * object on the local MDT will be linked under the dir
1231 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1233 * We do not know whether the object for the given FID
1234 * is referenced by some remote name entry or not, and
1235 * especially for DNE II, a multiple-linked object may
1236 * have many name entries reside on many MDTs.
1238 * To simplify the operation, OSD will not distinguish
1239 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1240 * only happened for the RPC from other MDT during the
1241 * OI scrub, or for the client side RPC with FID only,
1242 * such as FID to path, or from old connected client.
1245 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1249 flags |= SS_AUTO_PARTIAL;
1250 flags &= ~SS_AUTO_FULL;
1255 if (scrub->os_running) {
1256 if (scrub->os_partial_scan && !scrub->os_in_join)
1259 if (IS_ERR_OR_NULL(inode) || result) {
1260 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1261 GOTO(out, result = -EINPROGRESS);
1265 LASSERT(obj->oo_inode == inode);
1267 osd_scrub_oi_insert(dev, fid, id, true);
1271 if (dev->od_scrub.os_scrub.os_auto_scrub_interval == AS_NEVER) {
1273 GOTO(out, result = -EREMCHG);
1276 LASSERT(obj->oo_inode == inode);
1278 osd_add_oi_cache(info, dev, id, fid);
1283 if (IS_ERR_OR_NULL(inode) || result)
1284 osd_scrub_oi_insert(dev, fid, id, result == -ENOENT);
1286 rc1 = osd_scrub_start(env, dev, flags);
1287 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
1288 "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
1289 osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
1290 if (rc1 && rc1 != -EALREADY)
1291 GOTO(out, result = -EREMCHG);
1293 if (IS_ERR_OR_NULL(inode) || result)
1294 GOTO(out, result = -EINPROGRESS);
1297 LASSERT(obj->oo_inode == inode);
1302 if (unlikely(obj->oo_header))
1305 result = osd_check_lma(env, obj);
1309 LASSERTF(id->oii_ino == inode->i_ino &&
1310 id->oii_gen == inode->i_generation,
1311 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1312 PFID(fid), id->oii_ino, id->oii_gen,
1313 inode->i_ino, inode->i_generation);
1315 saved_ino = inode->i_ino;
1316 saved_gen = inode->i_generation;
1318 if (unlikely(result == -ENODATA)) {
1320 * If the OI scrub updated the OI mapping by race, it
1321 * must be valid. Trust the inode that has no LMA EA.
1326 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1329 * The OI mapping is still there, the inode is still
1330 * valid. It is just becaues the inode has no LMA EA.
1332 if (saved_ino == id->oii_ino &&
1333 saved_gen == id->oii_gen)
1337 * It is the OI scrub updated the OI mapping by race.
1338 * The new OI mapping must be valid.
1346 * "result == -ENOENT" means that the OI mappinghas been
1347 * removed by race, so the inode belongs to other object.
1349 * Others error can be returned directly.
1351 if (result == -ENOENT) {
1354 obj->oo_inode = NULL;
1359 if (result != -EREMCHG)
1365 * if two OST objects map to the same inode, and inode mode is
1366 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1367 * reserved by precreate, and not written yet, in this case, don't
1368 * set inode for the object whose FID mismatch, so that it can create
1369 * inode and not block precreate.
1371 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1372 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1373 obj->oo_inode = NULL;
1374 GOTO(out, result = 0);
1377 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1379 * "result == -ENOENT" means the cached OI mapping has been removed
1380 * from the OI file by race, above inode belongs to other object.
1382 if (result == -ENOENT) {
1385 obj->oo_inode = NULL;
1386 GOTO(out, result = 0);
1392 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1394 osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
1400 * It is the OI scrub updated the OI mapping by race.
1401 * The new OI mapping must be valid.
1409 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1410 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1412 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1414 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1416 obj->oo_pfid_in_lma = 1;
1417 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1419 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1420 } else if (result != -ENODATA) {
1425 obj->oo_compat_dot_created = 1;
1426 obj->oo_compat_dotdot_created = 1;
1428 if (S_ISDIR(inode->i_mode) &&
1429 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1430 osd_check_lmv(info, dev, inode);
1432 result = osd_attach_jinode(inode);
1437 GOTO(out, result = 0);
1439 LASSERT(!obj->oo_hl_head);
1440 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1442 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1445 if (!result && stale)
1446 osd_scrub_oi_resurrect(scrub, fid);
1448 if (result || !obj->oo_inode) {
1449 if (!IS_ERR_OR_NULL(inode))
1452 obj->oo_inode = NULL;
1454 fid_zero(&oic->oic_fid);
1457 LINVRNT(osd_invariant(obj));
1462 * Concurrency: shouldn't matter.
1464 static void osd_object_init0(struct osd_object *obj)
1466 LASSERT(obj->oo_inode != NULL);
1467 obj->oo_dt.do_body_ops = &osd_body_ops;
1468 obj->oo_dt.do_lu.lo_header->loh_attr |=
1469 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1473 * Concurrency: no concurrent access is possible that early in object
1476 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1477 const struct lu_object_conf *conf)
1479 struct osd_object *obj = osd_obj(l);
1482 LINVRNT(osd_invariant(obj));
1484 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1485 cfs_fail_val == 2) {
1486 struct osd_thread_info *info = osd_oti_get(env);
1487 struct osd_idmap_cache *oic = &info->oti_cache;
1488 /* invalidate thread cache */
1489 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1491 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1492 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1493 l->lo_header->loh_attr |= LOHA_EXISTS;
1497 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1498 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1499 if (result == 0 && obj->oo_inode != NULL) {
1500 struct osd_thread_info *oti = osd_oti_get(env);
1501 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1503 osd_object_init0(obj);
1504 if (unlikely(obj->oo_header))
1507 result = osd_get_lma(oti, obj->oo_inode,
1508 &oti->oti_obj_dentry, loa);
1511 * Convert LMAI flags to lustre LMA flags
1512 * and cache it to oo_lma_flags
1515 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1516 } else if (result == -ENODATA) {
1520 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
1522 LINVRNT(osd_invariant(obj));
1527 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1528 * The left part is for value, binary mode.
1530 struct osd_xattr_entry {
1531 struct list_head oxe_list;
1535 struct rcu_head oxe_rcu;
1539 static int osd_oxc_get(struct osd_object *obj, const char *name,
1542 struct osd_xattr_entry *tmp;
1543 struct osd_xattr_entry *oxe = NULL;
1544 size_t namelen = strlen(name);
1548 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1549 if (namelen == tmp->oxe_namelen &&
1550 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1557 GOTO(out, rc = -ENOENT);
1559 if (!oxe->oxe_exist)
1560 GOTO(out, rc = -ENODATA);
1563 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1566 if (buf->lb_buf == NULL)
1569 if (buf->lb_len < rc)
1570 GOTO(out, rc = -ERANGE);
1572 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1579 static void osd_oxc_free(struct rcu_head *head)
1581 struct osd_xattr_entry *oxe;
1583 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1584 OBD_FREE(oxe, oxe->oxe_len);
1587 static void osd_oxc_add(struct osd_object *obj, const char *name,
1588 const char *buf, int buflen)
1590 struct osd_xattr_entry *oxe;
1591 struct osd_xattr_entry *old = NULL;
1592 struct osd_xattr_entry *tmp;
1593 size_t namelen = strlen(name);
1594 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1596 OBD_ALLOC(oxe, len);
1600 INIT_LIST_HEAD(&oxe->oxe_list);
1602 oxe->oxe_namelen = namelen;
1603 memcpy(oxe->oxe_buf, name, namelen);
1605 LASSERT(buf != NULL);
1606 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1607 oxe->oxe_exist = true;
1609 oxe->oxe_exist = false;
1612 /* this should be rarely called, just remove old and add new */
1613 spin_lock(&obj->oo_guard);
1614 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1615 if (namelen == tmp->oxe_namelen &&
1616 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1622 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1623 call_rcu(&old->oxe_rcu, osd_oxc_free);
1625 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1627 spin_unlock(&obj->oo_guard);
1630 static void osd_oxc_del(struct osd_object *obj, const char *name)
1632 struct osd_xattr_entry *oxe;
1633 size_t namelen = strlen(name);
1635 spin_lock(&obj->oo_guard);
1636 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1637 if (namelen == oxe->oxe_namelen &&
1638 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1639 list_del_rcu(&oxe->oxe_list);
1640 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1644 spin_unlock(&obj->oo_guard);
1647 static void osd_oxc_fini(struct osd_object *obj)
1649 struct osd_xattr_entry *oxe, *next;
1651 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1652 list_del(&oxe->oxe_list);
1653 OBD_FREE(oxe, oxe->oxe_len);
1658 * Concurrency: no concurrent access is possible that late in object
1661 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1663 struct osd_object *obj = osd_obj(l);
1664 struct lu_object_header *h = obj->oo_header;
1666 LINVRNT(osd_invariant(obj));
1669 dt_object_fini(&obj->oo_dt);
1670 if (obj->oo_hl_head != NULL)
1671 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1672 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1675 lu_object_header_free(h);
1679 * Concurrency: no concurrent access is possible that late in object
1682 static void osd_index_fini(struct osd_object *o)
1684 struct iam_container *bag;
1686 if (o->oo_dir != NULL) {
1687 bag = &o->oo_dir->od_container;
1688 if (o->oo_inode != NULL) {
1689 if (bag->ic_object == o->oo_inode)
1690 iam_container_fini(bag);
1692 OBD_FREE_PTR(o->oo_dir);
1698 OSD_TXN_OI_DELETE_CREDITS = 20,
1699 OSD_TXN_INODE_DELETE_CREDITS = 20
1706 #if OSD_THANDLE_STATS
1708 * Set time when the handle is allocated
1710 static void osd_th_alloced(struct osd_thandle *oth)
1712 oth->oth_alloced = ktime_get();
1716 * Set time when the handle started
1718 static void osd_th_started(struct osd_thandle *oth)
1720 oth->oth_started = ktime_get();
1724 * Check whether the we deal with this handle for too long.
1726 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1727 ktime_t alloced, ktime_t started,
1730 ktime_t now = ktime_get();
1732 LASSERT(dev != NULL);
1734 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1735 ktime_us_delta(started, alloced));
1736 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1737 ktime_us_delta(closed, started));
1738 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1739 ktime_us_delta(now, closed));
1741 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1742 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1743 oth, now, alloced, started, closed);
1744 libcfs_debug_dumpstack(NULL);
1748 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1750 ktime_t __closed = ktime_get(); \
1751 ktime_t __alloced = oth->oth_alloced; \
1752 ktime_t __started = oth->oth_started; \
1755 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1758 #else /* OSD_THANDLE_STATS */
1760 #define osd_th_alloced(h) do {} while(0)
1761 #define osd_th_started(h) do {} while(0)
1762 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1764 #endif /* OSD_THANDLE_STATS */
1767 * in some cases (like overstriped files) the same operations on the same
1768 * objects are declared many times and this may lead to huge number of
1769 * credits which can be a problem and/or cause performance degradation.
1770 * this function is to remember what declarations have been made within
1771 * a given thandle and then skip duplications.
1772 * limit it's scope so that regular small transactions don't need all
1773 * this overhead with allocations, lists.
1774 * also, limit scope to the specific objects like llogs, etc.
1776 static inline bool osd_check_special_fid(const struct lu_fid *f)
1778 if (fid_seq_is_llog(f->f_seq))
1780 if (f->f_seq == FID_SEQ_LOCAL_FILE &&
1781 f->f_oid == MDD_LOV_OBJ_OID)
1786 bool osd_tx_was_declared(const struct lu_env *env, struct osd_thandle *oth,
1787 struct dt_object *dt, enum dt_txn_op op, loff_t pos)
1789 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
1790 struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1791 struct osd_thread_info *oti = osd_oti_get(env);
1792 struct osd_obj_declare *old;
1797 /* small transactions don't need this overhead */
1798 if (oti->oti_declare_ops[DTO_OBJECT_CREATE] < 10 &&
1799 oti->oti_declare_ops[DTO_WRITE_BASE] < 10)
1802 if (osd_check_special_fid(fid) == 0)
1805 list_for_each_entry(old, &oth->ot_declare_list, old_list) {
1806 if (old->old_op == op && old->old_pos == pos &&
1807 lu_fid_eq(&old->old_fid, fid))
1811 if (unlikely(old == NULL))
1813 old->old_fid = *lu_object_fid(&dt->do_lu);
1816 list_add(&old->old_list, &oth->ot_declare_list);
1820 void osd_tx_declaration_free(struct osd_thandle *oth)
1822 struct osd_obj_declare *old, *tmp;
1824 list_for_each_entry_safe(old, tmp, &oth->ot_declare_list, old_list) {
1825 list_del_init(&old->old_list);
1831 * Concurrency: doesn't access mutable data.
1833 static int osd_param_is_not_sane(const struct osd_device *dev,
1834 const struct thandle *th)
1836 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1838 return oh->ot_credits > osd_transaction_size(dev);
1842 * Concurrency: shouldn't matter.
1844 static void osd_trans_commit_cb(struct super_block *sb,
1845 struct ldiskfs_journal_cb_entry *jcb, int error)
1847 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1848 struct thandle *th = &oh->ot_super;
1849 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1850 struct osd_device *osd = osd_dev(lud);
1851 struct dt_txn_commit_cb *dcb, *tmp;
1853 LASSERT(oh->ot_handle == NULL);
1856 CERROR("transaction @0x%p commit error: %d\n", th, error);
1858 /* call per-transaction callbacks if any */
1859 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1861 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1862 "commit callback entry: magic=%x name='%s'\n",
1863 dcb->dcb_magic, dcb->dcb_name);
1864 list_del_init(&dcb->dcb_linkage);
1865 dcb->dcb_func(NULL, th, dcb, error);
1868 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1869 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1870 wake_up(&osd->od_commit_cb_done);
1876 static struct thandle *osd_trans_create(const struct lu_env *env,
1877 struct dt_device *d)
1879 struct osd_thread_info *oti = osd_oti_get(env);
1880 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1881 struct osd_thandle *oh;
1887 CERROR("%s: someone try to start transaction under "
1888 "readonly mode, should be disabled.\n",
1889 osd_name(osd_dt_dev(d)));
1891 RETURN(ERR_PTR(-EROFS));
1894 /* on pending IO in this thread should left from prev. request */
1895 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1897 sb_start_write(osd_sb(osd_dt_dev(d)));
1899 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1901 sb_end_write(osd_sb(osd_dt_dev(d)));
1902 RETURN(ERR_PTR(-ENOMEM));
1905 oh->ot_quota_trans = &oti->oti_quota_trans;
1906 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1911 oh->oh_declared_ext = 0;
1912 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1913 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1914 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1915 INIT_LIST_HEAD(&oh->ot_declare_list);
1918 memset(oti->oti_declare_ops, 0,
1919 sizeof(oti->oti_declare_ops));
1920 memset(oti->oti_declare_ops_cred, 0,
1921 sizeof(oti->oti_declare_ops_cred));
1922 memset(oti->oti_declare_ops_used, 0,
1923 sizeof(oti->oti_declare_ops_used));
1925 oti->oti_ins_cache_depth++;
1930 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1932 struct osd_thread_info *oti = osd_oti_get(env);
1933 struct osd_thandle *oh;
1935 oh = container_of(th, struct osd_thandle, ot_super);
1936 LASSERT(oh != NULL);
1938 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1939 oti->oti_declare_ops[OSD_OT_CREATE],
1940 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1941 oti->oti_declare_ops_used[OSD_OT_CREATE],
1942 oti->oti_declare_ops[OSD_OT_DESTROY],
1943 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1944 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1945 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1946 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1947 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1948 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1949 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1950 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1951 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1952 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1953 oti->oti_declare_ops[OSD_OT_WRITE],
1954 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1955 oti->oti_declare_ops_used[OSD_OT_WRITE],
1956 oti->oti_declare_ops[OSD_OT_PUNCH],
1957 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1958 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1959 oti->oti_declare_ops[OSD_OT_QUOTA],
1960 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1961 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1962 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1963 oti->oti_declare_ops[OSD_OT_INSERT],
1964 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1965 oti->oti_declare_ops_used[OSD_OT_INSERT],
1966 oti->oti_declare_ops[OSD_OT_DELETE],
1967 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1968 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1969 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1970 oti->oti_declare_ops[OSD_OT_REF_ADD],
1971 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1972 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1973 oti->oti_declare_ops[OSD_OT_REF_DEL],
1974 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1975 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1979 * Concurrency: shouldn't matter.
1981 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1984 struct osd_thread_info *oti = osd_oti_get(env);
1985 struct osd_device *dev = osd_dt_dev(d);
1987 struct osd_thandle *oh;
1992 LASSERT(current->journal_info == NULL);
1994 oh = container_of(th, struct osd_thandle, ot_super);
1995 LASSERT(oh != NULL);
1996 LASSERT(oh->ot_handle == NULL);
1997 if (unlikely(ldiskfs_track_declares_assert != 0)) {
1998 LASSERT(oti->oti_r_locks == 0);
1999 LASSERT(oti->oti_w_locks == 0);
2002 rc = dt_txn_hook_start(env, d, th);
2006 if (unlikely(osd_param_is_not_sane(dev, th))) {
2007 static unsigned long last_printed;
2008 static int last_credits;
2010 lprocfs_counter_add(dev->od_stats,
2011 LPROC_OSD_TOO_MANY_CREDITS, 1);
2014 * don't make noise on a tiny testing systems
2015 * actual credits misuse will be caught anyway
2017 if (last_credits != oh->ot_credits &&
2018 time_after(jiffies, last_printed +
2019 cfs_time_seconds(60)) &&
2020 osd_transaction_size(dev) > 512) {
2021 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
2022 oh->ot_credits, osd_transaction_size(dev));
2023 osd_trans_dump_creds(env, th);
2024 libcfs_debug_dumpstack(NULL);
2025 last_credits = oh->ot_credits;
2026 last_printed = jiffies;
2029 * XXX Limit the credits to 'max_transaction_buffers', and
2030 * let the underlying filesystem to catch the error if
2031 * we really need so many credits.
2033 * This should be removed when we can calculate the
2034 * credits precisely.
2036 oh->ot_credits = osd_transaction_size(dev);
2037 } else if (ldiskfs_track_declares_assert != 0) {
2039 * reserve few credits to prevent an assertion in JBD
2040 * our debugging mechanism will be able to detected
2041 * overuse. this can help to debug single-update
2044 oh->ot_credits += 10;
2045 if (unlikely(osd_param_is_not_sane(dev, th)))
2046 oh->ot_credits = osd_transaction_size(dev);
2049 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
2050 GOTO(out, rc = -EIO);
2053 * we ignore quota checks for system-owned files, but still
2054 * need to count blocks for uid/gid/projid
2056 osd_trans_declare_op(env, oh, OSD_OT_QUOTA, 3);
2059 * XXX temporary stuff. Some abstraction layer should
2062 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
2066 LASSERT(oti->oti_txns == 0);
2068 atomic_inc(&dev->od_commit_cb_in_flight);
2069 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
2080 static int osd_seq_exists(const struct lu_env *env,
2081 struct osd_device *osd, u64 seq)
2083 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
2084 struct seq_server_site *ss = osd_seq_site(osd);
2089 LASSERT(ss != NULL);
2090 LASSERT(ss->ss_server_fld != NULL);
2092 rc = osd_fld_lookup(env, osd, seq, range);
2095 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2096 osd_name(osd), seq, rc);
2100 RETURN(ss->ss_node_id == range->lsr_index);
2103 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2105 struct dt_txn_commit_cb *dcb;
2106 struct dt_txn_commit_cb *tmp;
2108 /* call per-transaction stop callbacks if any */
2109 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2111 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2112 "commit callback entry: magic=%x name='%s'\n",
2113 dcb->dcb_magic, dcb->dcb_name);
2114 list_del_init(&dcb->dcb_linkage);
2115 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2120 * Concurrency: shouldn't matter.
2122 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2125 struct osd_thread_info *oti = osd_oti_get(env);
2126 struct osd_thandle *oh;
2127 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2128 struct osd_device *osd = osd_dt_dev(th->th_dev);
2129 struct qsd_instance *qsd = osd_def_qsd(osd);
2130 struct lquota_trans *qtrans;
2131 LIST_HEAD(truncates);
2132 int rc = 0, remove_agents = 0;
2136 oh = container_of(th, struct osd_thandle, ot_super);
2138 remove_agents = oh->ot_remove_agents;
2140 qtrans = oh->ot_quota_trans;
2141 oh->ot_quota_trans = NULL;
2143 osd_tx_declaration_free(oh);
2145 /* move locks to local list, stop tx, execute truncates */
2146 list_splice(&oh->ot_trunc_locks, &truncates);
2148 if (oh->ot_handle != NULL) {
2151 handle_t *hdl = oh->ot_handle;
2154 * add commit callback
2155 * notice we don't do this in osd_trans_start()
2156 * as underlying transaction can change during truncate
2158 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2161 LASSERT(oti->oti_txns == 1);
2164 rc = dt_txn_hook_stop(env, th);
2166 CERROR("%s: failed in transaction hook: rc = %d\n",
2169 osd_trans_stop_cb(oh, rc);
2170 /* hook functions might modify th_sync */
2171 hdl->h_sync = th->th_sync;
2173 oh->ot_handle = NULL;
2174 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2176 CERROR("%s: failed to stop transaction: rc = %d\n",
2177 osd_name(osd), rc2);
2181 /* We preserve the origin behavior of ignoring any
2182 * failures with the underlying punch / truncate
2183 * operation. We do record for debugging if an error
2184 * does occur in the lctl dk logs.
2186 rc2 = osd_process_truncates(env, &truncates);
2188 CERROR("%s: failed truncate process: rc = %d\n",
2189 osd_name(osd), rc2);
2191 osd_trans_stop_cb(oh, th->th_result);
2195 osd_trunc_unlock_all(env, &truncates);
2197 /* inform the quota slave device that the transaction is stopping */
2198 qsd_op_end(env, qsd, qtrans);
2201 * as we want IO to journal and data IO be concurrent, we don't block
2202 * awaiting data IO completion in osd_do_bio(), instead we wait here
2203 * once transaction is submitted to the journal. all reqular requests
2204 * don't do direct IO (except read/write), thus this wait_event becomes
2207 * IMPORTANT: we have to wait till any IO submited by the thread is
2208 * completed otherwise iobuf may be corrupted by different request
2210 wait_event(iobuf->dr_wait,
2211 atomic_read(&iobuf->dr_numreqs) == 0);
2212 osd_fini_iobuf(osd, iobuf);
2214 rc = iobuf->dr_error;
2216 if (unlikely(remove_agents != 0))
2217 osd_process_scheduled_agent_removals(env, osd);
2219 LASSERT(oti->oti_ins_cache_depth > 0);
2220 oti->oti_ins_cache_depth--;
2221 /* reset OI cache for safety */
2222 if (oti->oti_ins_cache_depth == 0)
2223 oti->oti_ins_cache_used = 0;
2225 sb_end_write(osd_sb(osd));
2230 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2232 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2235 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2236 LASSERT(&dcb->dcb_func != NULL);
2237 if (dcb->dcb_flags & DCB_TRANS_STOP)
2238 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2240 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2246 * Called just before object is freed. Releases all resources except for
2247 * object itself (that is released by osd_object_free()).
2249 * Concurrency: no concurrent access is possible that late in object
2252 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2254 struct osd_object *obj = osd_obj(l);
2255 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2256 struct inode *inode = obj->oo_inode;
2261 LINVRNT(osd_invariant(obj));
2264 * If object is unlinked remove fid->ino mapping from object index.
2267 osd_index_fini(obj);
2272 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2273 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2275 uid = i_uid_read(inode);
2276 gid = i_gid_read(inode);
2277 projid = i_projid_read(inode);
2279 obj->oo_inode = NULL;
2282 /* do not rebalance quota if the caller needs to release memory
2283 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2284 * transaction and risk to deadlock - LU-12178 */
2285 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2288 if (!obj->oo_header && qsd) {
2289 struct osd_thread_info *info = osd_oti_get(env);
2290 struct lquota_id_info *qi = &info->oti_qi;
2292 /* Release granted quota to master if necessary */
2293 qi->lqi_id.qid_uid = uid;
2294 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2296 qi->lqi_id.qid_uid = gid;
2297 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2299 qi->lqi_id.qid_uid = projid;
2300 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2305 * Concurrency: ->loo_object_release() is called under site spin-lock.
2307 static void osd_object_release(const struct lu_env *env,
2308 struct lu_object *l)
2310 struct osd_object *o = osd_obj(l);
2313 * nobody should be releasing a non-destroyed object with nlink=0
2314 * the API allows this, but ldiskfs doesn't like and then report
2315 * this inode as deleted
2317 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2318 o->oo_inode->i_nlink == 0));
2322 * Concurrency: shouldn't matter.
2324 static int osd_object_print(const struct lu_env *env, void *cookie,
2325 lu_printer_t p, const struct lu_object *l)
2327 struct osd_object *o = osd_obj(l);
2328 struct iam_descr *d;
2330 if (o->oo_dir != NULL)
2331 d = o->oo_dir->od_container.ic_descr;
2334 return (*p)(env, cookie,
2335 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2337 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2338 o->oo_inode ? o->oo_inode->i_generation : 0,
2339 d ? d->id_ops->id_name : "plain");
2343 * Concurrency: shouldn't matter.
2345 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2346 struct obd_statfs *sfs, struct obd_statfs_info *info)
2348 struct osd_device *osd = osd_dt_dev(d);
2349 struct super_block *sb = osd_sb(osd);
2350 struct kstatfs *ksfs;
2354 if (unlikely(osd->od_mnt == NULL))
2355 return -EINPROGRESS;
2357 /* osd_lproc.c call this without env, allocate ksfs for that case */
2358 if (unlikely(env == NULL)) {
2359 OBD_ALLOC_PTR(ksfs);
2363 ksfs = &osd_oti_get(env)->oti_ksfs;
2366 result = sb->s_op->statfs(sb->s_root, ksfs);
2370 statfs_pack(sfs, ksfs);
2371 if (unlikely(sb->s_flags & SB_RDONLY))
2372 sfs->os_state |= OS_STATFS_READONLY;
2374 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2376 if (ldiskfs_has_feature_extents(sb))
2377 sfs->os_maxbytes = sb->s_maxbytes;
2379 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2382 * Reserve some space so to avoid fragmenting the filesystem too much.
2383 * Fragmentation not only impacts performance, but can also increase
2384 * metadata overhead significantly, causing grant calculation to be
2387 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2389 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2390 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2391 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2392 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2394 sfs->os_blocks -= reserved;
2395 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2396 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2399 if (unlikely(env == NULL))
2405 * Estimate space needed for file creations. We assume the largest filename
2406 * which is 2^64 - 1, hence a filename of 20 chars.
2407 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2409 #ifdef __LDISKFS_DIR_REC_LEN
2410 # define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2411 #elif defined LDISKFS_DIR_REC_LEN_WITH_DIR
2412 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20, NULL)
2414 # define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2418 * Concurrency: doesn't access mutable data.
2420 static void osd_conf_get(const struct lu_env *env,
2421 const struct dt_device *dev,
2422 struct dt_device_param *param)
2424 struct osd_device *d = osd_dt_dev(dev);
2425 struct super_block *sb = osd_sb(d);
2426 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2431 * XXX should be taken from not-yet-existing fs abstraction layer.
2433 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2434 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2435 param->ddp_symlink_max = sb->s_blocksize;
2436 param->ddp_mount_type = LDD_MT_LDISKFS;
2437 if (ldiskfs_has_feature_extents(sb))
2438 param->ddp_maxbytes = sb->s_maxbytes;
2440 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2442 * inode are statically allocated, so per-inode space consumption
2443 * is the space consumed by the directory entry
2445 param->ddp_inodespace = PER_OBJ_USAGE;
2447 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2448 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2449 * maximum length to not under-count the actual number of extents
2450 * needed for writing a file if there are sub-optimal block allocations.
2452 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2453 /* worst-case extent insertion metadata overhead */
2454 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2455 param->ddp_mntopts = 0;
2456 if (test_opt(sb, XATTR_USER))
2457 param->ddp_mntopts |= MNTOPT_USERXATTR;
2458 if (test_opt(sb, POSIX_ACL))
2459 param->ddp_mntopts |= MNTOPT_ACL;
2462 * LOD might calculate the max stripe count based on max_ea_size,
2463 * so we need take account in the overhead as well,
2464 * xattr_header + magic + xattr_entry_head
2466 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2467 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2469 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2470 if (ldiskfs_has_feature_ea_inode(sb))
2471 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2475 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2477 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2478 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2481 * Preferred RPC size for efficient disk IO. 4MB shows good
2482 * all-around performance for ldiskfs, but use bigalloc chunk size
2483 * by default if larger.
2485 #if defined(LDISKFS_CLUSTER_SIZE)
2486 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2487 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2490 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2492 param->ddp_t10_cksum_type = 0;
2494 unsigned short interval = blk_integrity_interval(bi);
2495 name = blk_integrity_name(bi);
2505 if (strncmp(name, "T10-DIF-TYPE",
2506 sizeof("T10-DIF-TYPE") - 1) == 0) {
2507 /* also skip "1/2/3-" at end */
2508 const int type_off = sizeof("T10-DIF-TYPE.");
2509 char type_number = name[type_off - 2];
2511 if (interval != 512 && interval != 4096) {
2512 CERROR("%s: unsupported T10PI sector size %u\n",
2513 d->od_svname, interval);
2516 switch (type_number) {
2518 d->od_t10_type = OSD_T10_TYPE1;
2521 d->od_t10_type = OSD_T10_TYPE2;
2524 d->od_t10_type = OSD_T10_TYPE3;
2527 CERROR("%s: unsupported T10PI type %s\n",
2528 d->od_svname, name);
2531 if (strcmp(name + type_off, "CRC") == 0) {
2532 d->od_t10_type |= OSD_T10_TYPE_CRC;
2533 param->ddp_t10_cksum_type = interval == 512 ?
2534 OBD_CKSUM_T10CRC512 :
2536 } else if (strcmp(name + type_off, "IP") == 0) {
2537 d->od_t10_type |= OSD_T10_TYPE_IP;
2538 param->ddp_t10_cksum_type = interval == 512 ?
2539 OBD_CKSUM_T10IP512 :
2542 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2543 d->od_svname, name);
2548 CERROR("%s: unsupported T10PI type '%s'\n",
2549 d->od_svname, name);
2554 param->ddp_has_lseek_data_hole = true;
2557 static struct vfsmount *osd_mnt_get(const struct dt_device *d)
2559 return osd_dt_dev(d)->od_mnt;
2563 * Concurrency: shouldn't matter.
2565 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2568 struct super_block *s = osd_sb(osd_dt_dev(d));
2571 down_read(&s->s_umount);
2572 rc = s->s_op->sync_fs(s, 1);
2573 up_read(&s->s_umount);
2575 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2582 * Start commit for OSD device.
2584 * An implementation of dt_commit_async method for OSD device.
2585 * Asychronously starts underlayng fs sync and thereby a transaction
2588 * \param env environment
2589 * \param d dt device
2591 * \see dt_device_operations
2593 static int osd_commit_async(const struct lu_env *env,
2594 struct dt_device *d)
2596 struct super_block *s = osd_sb(osd_dt_dev(d));
2601 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2602 down_read(&s->s_umount);
2603 rc = s->s_op->sync_fs(s, 0);
2604 up_read(&s->s_umount);
2610 * Concurrency: shouldn't matter.
2612 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2614 struct super_block *sb = osd_sb(osd_dt_dev(d));
2615 struct block_device *dev = sb->s_bdev;
2616 int rc = -EOPNOTSUPP;
2620 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2621 osd_dt_dev(d)->od_svname, (long)dev, rc);
2627 * Note: we do not count into QUOTA here.
2628 * If we mount with --data_journal we may need more.
2630 const int osd_dto_credits_noquota[DTO_NR] = {
2633 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2634 * SINGLEDATA_TRANS_BLOCKS(8)
2635 * XXX Note: maybe iam need more, since iam have more level than
2638 [DTO_INDEX_INSERT] = 16,
2641 * just modify a single entry, probably merge few within a block
2643 [DTO_INDEX_DELETE] = 1,
2647 [DTO_INDEX_UPDATE] = 16,
2649 * 4(inode, inode bits, groups, GDT)
2650 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2652 [DTO_OBJECT_CREATE] = 4,
2654 * 4(inode, inode bits, groups, GDT)
2655 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2657 [DTO_OBJECT_DELETE] = 4,
2659 * Attr set credits (inode)
2661 [DTO_ATTR_SET_BASE] = 1,
2663 * Xattr set. The same as xattr of EXT3.
2664 * DATA_TRANS_BLOCKS(14)
2665 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2666 * are also counted in. Do not know why?
2668 [DTO_XATTR_SET] = 14,
2670 * credits for inode change during write.
2672 [DTO_WRITE_BASE] = 3,
2674 * credits for single block write.
2676 [DTO_WRITE_BLOCK] = 14,
2678 * Attr set credits for chown.
2679 * This is extra credits for setattr, and it is null without quota
2681 [DTO_ATTR_SET_CHOWN] = 0
2684 /* reserve or free quota for some operation */
2685 static int osd_reserve_or_free_quota(const struct lu_env *env,
2686 struct dt_device *dev,
2687 struct lquota_id_info *qi)
2689 struct osd_device *osd = osd_dt_dev(dev);
2690 struct qsd_instance *qsd = NULL;
2696 qsd = osd->od_quota_slave_dt;
2698 qsd = osd->od_quota_slave_md;
2700 rc = qsd_reserve_or_free_quota(env, qsd, qi);
2704 static const struct dt_device_operations osd_dt_ops = {
2705 .dt_root_get = osd_root_get,
2706 .dt_statfs = osd_statfs,
2707 .dt_trans_create = osd_trans_create,
2708 .dt_trans_start = osd_trans_start,
2709 .dt_trans_stop = osd_trans_stop,
2710 .dt_trans_cb_add = osd_trans_cb_add,
2711 .dt_conf_get = osd_conf_get,
2712 .dt_mnt_get = osd_mnt_get,
2713 .dt_sync = osd_sync,
2715 .dt_commit_async = osd_commit_async,
2716 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
2719 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2722 struct osd_object *obj = osd_dt_obj(dt);
2723 struct osd_thread_info *oti = osd_oti_get(env);
2725 LINVRNT(osd_invariant(obj));
2727 LASSERT(obj->oo_owner != env);
2728 down_read_nested(&obj->oo_sem, role);
2730 LASSERT(obj->oo_owner == NULL);
2734 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2737 struct osd_object *obj = osd_dt_obj(dt);
2738 struct osd_thread_info *oti = osd_oti_get(env);
2740 LINVRNT(osd_invariant(obj));
2742 LASSERT(obj->oo_owner != env);
2743 down_write_nested(&obj->oo_sem, role);
2745 LASSERT(obj->oo_owner == NULL);
2746 obj->oo_owner = env;
2750 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2752 struct osd_object *obj = osd_dt_obj(dt);
2753 struct osd_thread_info *oti = osd_oti_get(env);
2755 LINVRNT(osd_invariant(obj));
2757 LASSERT(oti->oti_r_locks > 0);
2759 up_read(&obj->oo_sem);
2762 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2764 struct osd_object *obj = osd_dt_obj(dt);
2765 struct osd_thread_info *oti = osd_oti_get(env);
2767 LINVRNT(osd_invariant(obj));
2769 LASSERT(obj->oo_owner == env);
2770 LASSERT(oti->oti_w_locks > 0);
2772 obj->oo_owner = NULL;
2773 up_write(&obj->oo_sem);
2776 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2778 struct osd_object *obj = osd_dt_obj(dt);
2780 LINVRNT(osd_invariant(obj));
2782 return obj->oo_owner == env;
2785 static void osd_inode_getattr(const struct lu_env *env,
2786 struct inode *inode, struct lu_attr *attr)
2788 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2789 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2790 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2791 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2793 attr->la_atime = inode->i_atime.tv_sec;
2794 attr->la_mtime = inode->i_mtime.tv_sec;
2795 attr->la_ctime = inode->i_ctime.tv_sec;
2796 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2797 attr->la_mode = inode->i_mode;
2798 attr->la_size = i_size_read(inode);
2799 attr->la_blocks = inode->i_blocks;
2800 attr->la_uid = i_uid_read(inode);
2801 attr->la_gid = i_gid_read(inode);
2802 attr->la_projid = i_projid_read(inode);
2803 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2804 attr->la_nlink = inode->i_nlink;
2805 attr->la_rdev = inode->i_rdev;
2806 attr->la_blksize = 1 << inode->i_blkbits;
2807 attr->la_blkbits = inode->i_blkbits;
2809 * Ext4 did not transfer inherit flags from raw inode
2810 * to inode flags, and ext4 internally test raw inode
2811 * @i_flags directly. Instead of patching ext4, we do it here.
2813 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2814 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2817 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2820 struct osd_object *obj = osd_dt_obj(dt);
2821 const struct dt_it_ops *iops;
2827 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2828 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2830 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2831 *count = obj->oo_dirent_count;
2835 /* directory not initialized yet */
2836 if (!dt->do_index_ops) {
2841 iops = &dt->do_index_ops->dio_it;
2842 it = iops->init(env, dt, LUDA_64BITHASH);
2844 RETURN(PTR_ERR(it));
2846 rc = iops->load(env, it, 0);
2848 if (rc == -ENODATA) {
2855 rc = iops->next(env, it);
2857 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2861 if (iops->key_size(env, it) == 0)
2867 obj->oo_dirent_count = *count;
2872 iops->fini(env, it);
2877 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2878 struct lu_attr *attr)
2880 struct osd_object *obj = osd_dt_obj(dt);
2883 if (unlikely(!dt_object_exists(dt)))
2885 if (unlikely(obj->oo_destroyed))
2888 LASSERT(!dt_object_remote(dt));
2889 LINVRNT(osd_invariant(obj));
2891 spin_lock(&obj->oo_guard);
2892 osd_inode_getattr(env, obj->oo_inode, attr);
2893 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2894 attr->la_valid |= LA_FLAGS;
2895 attr->la_flags |= LUSTRE_ORPHAN_FL;
2897 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2898 attr->la_valid |= LA_FLAGS;
2899 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2901 spin_unlock(&obj->oo_guard);
2903 if (S_ISDIR(obj->oo_inode->i_mode) &&
2904 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2905 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2910 static int osd_declare_attr_qid(const struct lu_env *env,
2911 struct osd_object *obj,
2912 struct osd_thandle *oh, long long bspace,
2913 qid_t old_id, qid_t new_id, bool enforce,
2917 struct osd_thread_info *info = osd_oti_get(env);
2918 struct lquota_id_info *qi = &info->oti_qi;
2920 qi->lqi_type = type;
2921 /* inode accounting */
2922 qi->lqi_is_blk = false;
2924 /* one more inode for the new id ... */
2925 qi->lqi_id.qid_uid = new_id;
2927 /* Reserve credits for the new id */
2928 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2929 if (rc == -EDQUOT || rc == -EINPROGRESS)
2934 /* and one less inode for the current id */
2935 qi->lqi_id.qid_uid = old_id;
2937 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2938 if (rc == -EDQUOT || rc == -EINPROGRESS)
2943 /* block accounting */
2944 qi->lqi_is_blk = true;
2946 /* more blocks for the new id ... */
2947 qi->lqi_id.qid_uid = new_id;
2948 qi->lqi_space = bspace;
2950 * Credits for the new uid has been reserved, re-use "obj"
2951 * to save credit reservation.
2953 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2954 if (rc == -EDQUOT || rc == -EINPROGRESS)
2959 /* and finally less blocks for the current uid */
2960 qi->lqi_id.qid_uid = old_id;
2961 qi->lqi_space = -bspace;
2962 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2963 if (rc == -EDQUOT || rc == -EINPROGRESS)
2969 static int osd_declare_attr_set(const struct lu_env *env,
2970 struct dt_object *dt,
2971 const struct lu_attr *attr,
2972 struct thandle *handle)
2974 struct osd_thandle *oh;
2975 struct osd_object *obj;
2984 LASSERT(dt != NULL);
2985 LASSERT(handle != NULL);
2987 obj = osd_dt_obj(dt);
2988 LASSERT(osd_invariant(obj));
2990 oh = container_of(handle, struct osd_thandle, ot_super);
2991 LASSERT(oh->ot_handle == NULL);
2993 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2994 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2996 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2997 osd_dto_credits_noquota[DTO_XATTR_SET]);
2999 if (attr == NULL || obj->oo_inode == NULL)
3002 bspace = obj->oo_inode->i_blocks << 9;
3003 bspace = toqb(bspace);
3006 * Changing ownership is always preformed by super user, it should not
3007 * fail with EDQUOT unless required explicitly.
3009 * We still need to call the osd_declare_qid() to calculate the journal
3010 * credits for updating quota accounting files and to trigger quota
3011 * space adjustment once the operation is completed.
3013 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
3015 uid = i_uid_read(obj->oo_inode);
3016 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
3017 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
3018 attr->la_uid, enforce, USRQUOTA);
3022 gid = i_gid_read(obj->oo_inode);
3023 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
3024 attr->la_uid, gid, attr->la_gid);
3025 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
3026 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
3027 attr->la_gid, enforce, GRPQUOTA);
3032 #ifdef HAVE_PROJECT_QUOTA
3033 if (attr->la_valid & LA_PROJID) {
3034 __u32 projid = i_projid_read(obj->oo_inode);
3036 enforce = (attr->la_valid & LA_PROJID) &&
3037 (attr->la_projid != projid);
3038 rc = osd_declare_attr_qid(env, obj, oh, bspace,
3039 (qid_t)projid, (qid_t)attr->la_projid,
3045 /* punch must be aware we are dealing with an encrypted file */
3046 if (attr->la_valid & LA_FLAGS && attr->la_flags & LUSTRE_ENCRYPT_FL)
3047 obj->oo_lma_flags |= LUSTRE_ENCRYPT_FL;
3052 static int osd_inode_setattr(const struct lu_env *env,
3053 struct inode *inode, const struct lu_attr *attr)
3055 __u64 bits = attr->la_valid;
3057 /* Only allow set size for regular file */
3058 if (!S_ISREG(inode->i_mode))
3059 bits &= ~(LA_SIZE | LA_BLOCKS);
3064 if (bits & LA_ATIME)
3065 inode->i_atime = osd_inode_time(inode, attr->la_atime);
3066 if (bits & LA_CTIME)
3067 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
3068 if (bits & LA_MTIME)
3069 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
3070 if (bits & LA_SIZE) {
3071 spin_lock(&inode->i_lock);
3072 LDISKFS_I(inode)->i_disksize = attr->la_size;
3073 i_size_write(inode, attr->la_size);
3074 spin_unlock(&inode->i_lock);
3078 * OSD should not change "i_blocks" which is used by quota.
3079 * "i_blocks" should be changed by ldiskfs only.
3082 inode->i_mode = (inode->i_mode & S_IFMT) |
3083 (attr->la_mode & ~S_IFMT);
3085 i_uid_write(inode, attr->la_uid);
3087 i_gid_write(inode, attr->la_gid);
3088 if (bits & LA_PROJID)
3089 i_projid_write(inode, attr->la_projid);
3090 if (bits & LA_NLINK)
3091 set_nlink(inode, attr->la_nlink);
3093 inode->i_rdev = attr->la_rdev;
3095 if (bits & LA_FLAGS) {
3096 /* always keep S_NOCMTIME */
3097 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
3099 #if defined(S_ENCRYPTED)
3100 /* Always remove S_ENCRYPTED, because ldiskfs must not be
3101 * aware of encryption status. It is just stored into LMA
3102 * so that it can be forwared to client side.
3104 inode->i_flags &= ~S_ENCRYPTED;
3107 * Ext4 did not transfer inherit flags from
3108 * @inode->i_flags to raw inode i_flags when writing
3109 * flags, we do it explictly here.
3111 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
3112 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
3114 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
3119 #ifdef HAVE_PROJECT_QUOTA
3120 static int osd_transfer_project(struct inode *inode, __u32 projid,
3121 struct thandle *handle)
3123 struct super_block *sb = inode->i_sb;
3124 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3127 struct ldiskfs_iloc iloc;
3128 struct ldiskfs_inode *raw_inode;
3129 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3131 if (!ldiskfs_has_feature_project(sb)) {
3132 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3133 == LDISKFS_DEF_PROJID);
3134 if (projid != LDISKFS_DEF_PROJID)
3140 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3143 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3144 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3147 err = ldiskfs_get_inode_loc(inode, &iloc);
3151 raw_inode = ldiskfs_raw_inode(&iloc);
3152 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3153 struct osd_thandle *oh = container_of(handle,
3157 * try to expand inode size automatically.
3159 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3160 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3168 dquot_initialize(inode);
3169 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3170 if (transfer_to[PRJQUOTA]) {
3171 lock_dquot_transfer(inode);
3172 err = __dquot_transfer(inode, transfer_to);
3173 unlock_dquot_transfer(inode);
3174 dqput(transfer_to[PRJQUOTA]);
3183 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3184 struct thandle *handle)
3188 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3189 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3193 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3194 inode->i_ino, i_uid_read(inode), attr->la_uid,
3195 i_gid_read(inode), attr->la_gid);
3197 dquot_initialize(inode);
3199 if (attr->la_valid & LA_UID)
3200 iattr.ia_valid |= ATTR_UID;
3201 if (attr->la_valid & LA_GID)
3202 iattr.ia_valid |= ATTR_GID;
3203 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3204 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3206 lock_dquot_transfer(inode);
3207 rc = osd_dquot_transfer(&init_user_ns, inode, &iattr);
3208 unlock_dquot_transfer(inode);
3210 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3211 osd_ino2name(inode), rc);
3216 /* Handle project id transfer here properly */
3217 if (attr->la_valid & LA_PROJID &&
3218 attr->la_projid != i_projid_read(inode)) {
3219 if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
3221 #ifdef HAVE_PROJECT_QUOTA
3222 rc = osd_transfer_project(inode, attr->la_projid, handle);
3227 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3228 osd_ino2name(inode), rc);
3235 static int osd_attr_set(const struct lu_env *env,
3236 struct dt_object *dt,
3237 const struct lu_attr *attr,
3238 struct thandle *handle)
3240 struct osd_thread_info *info = osd_oti_get(env);
3241 struct osd_object *obj = osd_dt_obj(dt);
3242 struct osd_device *osd = osd_obj2dev(obj);
3243 struct inode *inode;
3246 if (!dt_object_exists(dt))
3249 LASSERT(handle != NULL);
3250 LASSERT(!dt_object_remote(dt));
3251 LASSERT(osd_invariant(obj));
3253 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3255 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3256 !osd_obj2dev(obj)->od_is_ost) {
3257 struct osd_thread_info *oti = osd_oti_get(env);
3258 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3259 struct lu_fid *fid1 = &oti->oti_fid;
3260 struct osd_inode_id *id = &oti->oti_id;
3261 struct iam_path_descr *ipd;
3262 struct iam_container *bag;
3263 struct osd_thandle *oh;
3266 fid_cpu_to_be(fid1, fid0);
3267 memset(id, 1, sizeof(*id));
3268 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3269 fid0)->oi_dir.od_container;
3270 ipd = osd_idx_ipd_get(env, bag);
3271 if (unlikely(ipd == NULL))
3274 oh = container_of(handle, struct osd_thandle, ot_super);
3275 rc = iam_update(oh->ot_handle, bag,
3276 (const struct iam_key *)fid1,
3277 (const struct iam_rec *)id, ipd);
3278 osd_ipd_put(env, bag, ipd);
3279 return(rc > 0 ? 0 : rc);
3282 inode = obj->oo_inode;
3284 rc = osd_quota_transfer(inode, attr, handle);
3288 spin_lock(&obj->oo_guard);
3289 rc = osd_inode_setattr(env, inode, attr);
3290 spin_unlock(&obj->oo_guard);
3294 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3296 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3298 if (!(attr->la_valid & LA_FLAGS))
3301 /* If setting LUSTRE_ENCRYPT_FL on an OST object, also set a dummy
3302 * enc ctx xattr, with 2 benefits:
3303 * - setting the LL_XATTR_NAME_ENCRYPTION_CONTEXT xattr internally sets
3304 * the LDISKFS_ENCRYPT_FL flag on the on-disk inode;
3305 * - it makes e2fsprogs happy to see an enc ctx for an inode that has
3306 * the LDISKFS_ENCRYPT_FL flag
3307 * We do not need the actual encryption context on OST objects, it is
3308 * only stored on MDT inodes, at file creation time.
3310 if (!(LDISKFS_I(obj->oo_inode)->i_flags & LDISKFS_ENCRYPT_FL) &&
3311 attr->la_flags & LUSTRE_ENCRYPT_FL && osd->od_is_ost &&
3312 !OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) {
3315 /* use a dummy enc ctx, fine with e2fsprogs */
3316 buf.lb_buf = "\xFF";
3318 rc = osd_xattr_set(env, dt, &buf,
3319 LL_XATTR_NAME_ENCRYPTION_CONTEXT,
3322 CWARN("%s: set "DFID" enc ctx failed: rc = %d\n",
3323 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3327 /* Let's check if there are extra flags need to be set into LMA */
3328 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3329 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3331 LASSERT(!obj->oo_pfid_in_lma);
3333 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3334 &info->oti_ost_attrs);
3338 if ((lma->lma_incompat & lustre_to_lma_flags(attr->la_flags)) ==
3339 lustre_to_lma_flags(attr->la_flags))
3340 /* if lma incompat already has the flags,
3341 * save a useless call to xattr_set
3345 lma->lma_incompat |=
3346 lustre_to_lma_flags(attr->la_flags);
3347 lustre_lma_swab(lma);
3349 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3351 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3352 lma, sizeof(*lma), XATTR_REPLACE);
3354 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3355 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3356 lma->lma_incompat, rc);
3359 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3360 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3367 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3368 struct osd_object *obj,
3369 const char *name, const int namelen)
3371 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3374 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3375 umode_t mode, struct dt_allocation_hint *hint,
3376 struct thandle *th, struct lu_attr *attr)
3379 struct osd_device *osd = osd_obj2dev(obj);
3380 struct osd_thandle *oth;
3381 struct dt_object *parent = NULL;
3382 struct inode *inode;
3383 struct iattr iattr = {
3384 .ia_valid = ATTR_UID | ATTR_GID |
3385 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3386 .ia_ctime.tv_sec = attr->la_ctime,
3387 .ia_mtime.tv_sec = attr->la_mtime,
3388 .ia_atime.tv_sec = attr->la_atime,
3389 .ia_uid = GLOBAL_ROOT_UID,
3390 .ia_gid = GLOBAL_ROOT_GID,
3392 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3394 if (attr->la_valid & LA_UID)
3395 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3396 if (attr->la_valid & LA_GID)
3397 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3399 LINVRNT(osd_invariant(obj));
3400 LASSERT(obj->oo_inode == NULL);
3401 LASSERT(obj->oo_hl_head == NULL);
3403 if (S_ISDIR(mode) && ldiskfs_pdo) {
3405 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3406 if (obj->oo_hl_head == NULL)
3410 oth = container_of(th, struct osd_thandle, ot_super);
3411 LASSERT(oth->ot_handle->h_transaction != NULL);
3413 if (hint != NULL && hint->dah_parent != NULL &&
3414 !dt_object_remote(hint->dah_parent))
3415 parent = hint->dah_parent;
3417 /* if a time component is not valid set it to UTIME_OMIT */
3418 if (!(attr->la_valid & LA_CTIME))
3419 iattr.ia_ctime = omit;
3420 if (!(attr->la_valid & LA_MTIME))
3421 iattr.ia_mtime = omit;
3422 if (!(attr->la_valid & LA_ATIME))
3423 iattr.ia_atime = omit;
3425 inode = ldiskfs_create_inode(oth->ot_handle,
3426 parent ? osd_dt_obj(parent)->oo_inode :
3427 osd_sb(osd)->s_root->d_inode,
3429 if (!IS_ERR(inode)) {
3430 /* Do not update file c/mtime in ldiskfs. */
3431 inode->i_flags |= S_NOCMTIME;
3434 * For new created object, it must be consistent,
3435 * and it is unnecessary to scrub against it.
3437 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3439 obj->oo_inode = inode;
3442 if (obj->oo_hl_head != NULL) {
3443 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3444 obj->oo_hl_head = NULL;
3446 result = PTR_ERR(inode);
3448 LINVRNT(osd_invariant(obj));
3456 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3457 struct lu_attr *attr,
3458 struct dt_allocation_hint *hint,
3459 struct dt_object_format *dof,
3463 struct osd_thandle *oth;
3464 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3466 LASSERT(S_ISDIR(attr->la_mode));
3468 oth = container_of(th, struct osd_thandle, ot_super);
3469 LASSERT(oth->ot_handle->h_transaction != NULL);
3470 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3471 obj->oo_dirent_count = 0;
3472 result = osd_mkfile(info, obj, mode, hint, th, attr);
3477 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3478 struct lu_attr *attr,
3479 struct dt_allocation_hint *hint,
3480 struct dt_object_format *dof,
3484 struct osd_thandle *oth;
3485 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3487 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3489 LASSERT(S_ISREG(attr->la_mode));
3491 oth = container_of(th, struct osd_thandle, ot_super);
3492 LASSERT(oth->ot_handle->h_transaction != NULL);
3494 result = osd_mkfile(info, obj, mode, hint, th, attr);
3496 LASSERT(obj->oo_inode != NULL);
3497 if (feat->dif_flags & DT_IND_VARKEY)
3498 result = iam_lvar_create(obj->oo_inode,
3499 feat->dif_keysize_max,
3501 feat->dif_recsize_max,
3504 result = iam_lfix_create(obj->oo_inode,
3505 feat->dif_keysize_max,
3507 feat->dif_recsize_max,
3513 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3514 struct lu_attr *attr,
3515 struct dt_allocation_hint *hint,
3516 struct dt_object_format *dof,
3519 LASSERT(S_ISREG(attr->la_mode));
3520 return osd_mkfile(info, obj, (attr->la_mode &
3521 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3525 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3526 struct lu_attr *attr,
3527 struct dt_allocation_hint *hint,
3528 struct dt_object_format *dof,
3531 LASSERT(S_ISLNK(attr->la_mode));
3532 return osd_mkfile(info, obj, (attr->la_mode &
3533 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3537 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3538 struct lu_attr *attr,
3539 struct dt_allocation_hint *hint,
3540 struct dt_object_format *dof,
3543 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3546 LINVRNT(osd_invariant(obj));
3547 LASSERT(obj->oo_inode == NULL);
3548 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3549 S_ISFIFO(mode) || S_ISSOCK(mode));
3551 result = osd_mkfile(info, obj, mode, hint, th, attr);
3553 LASSERT(obj->oo_inode != NULL);
3555 * This inode should be marked dirty for i_rdev. Currently
3556 * that is done in the osd_attr_init().
3558 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3561 LINVRNT(osd_invariant(obj));
3565 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3567 struct dt_allocation_hint *hint,
3568 struct dt_object_format *dof,
3571 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3573 osd_obj_type_f result;
3589 result = osd_mk_index;
3599 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3600 struct dt_object *parent, struct dt_object *child,
3605 ah->dah_parent = parent;
3607 if (parent != NULL && !dt_object_remote(parent)) {
3608 /* will help to find FID->ino at dt_insert("..") */
3609 struct osd_object *pobj = osd_dt_obj(parent);
3611 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3615 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3616 struct lu_attr *attr, struct dt_object_format *dof,
3617 struct thandle *handle)
3619 struct inode *inode = obj->oo_inode;
3620 __u64 valid = attr->la_valid;
3623 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3625 if (dof->dof_type != DFT_NODE)
3626 attr->la_valid &= ~LA_RDEV;
3627 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3628 attr->la_valid &= ~LA_ATIME;
3629 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3630 attr->la_valid &= ~LA_CTIME;
3631 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3632 attr->la_valid &= ~LA_MTIME;
3634 result = osd_quota_transfer(inode, attr, handle);
3638 if (attr->la_valid != 0) {
3639 result = osd_inode_setattr(info->oti_env, inode, attr);
3641 * The osd_inode_setattr() should always succeed here. The
3642 * only error that could be returned is EDQUOT when we are
3643 * trying to change the UID or GID of the inode. However, this
3644 * should not happen since quota enforcement is no longer
3645 * enabled on ldiskfs (lquota takes care of it).
3647 LASSERTF(result == 0, "%d\n", result);
3648 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3651 attr->la_valid = valid;
3655 * Helper function for osd_create()
3657 * \retval 0, on success
3659 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3660 struct lu_attr *attr, struct dt_allocation_hint *hint,
3661 struct dt_object_format *dof, struct thandle *th)
3666 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3668 /* we drop umask so that permissions we pass are not affected */
3669 umask = current->fs->umask;
3670 current->fs->umask = 0;
3672 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3674 if (likely(obj->oo_inode != NULL)) {
3675 LASSERT(obj->oo_inode->i_state & I_NEW);
3678 * Unlock the inode before attr initialization to avoid
3679 * unnecessary dqget operations. LU-6378
3681 unlock_new_inode(obj->oo_inode);
3684 if (likely(result == 0)) {
3685 osd_attr_init(info, obj, attr, dof, th);
3686 osd_object_init0(obj);
3689 /* restore previous umask value */
3690 current->fs->umask = umask;
3692 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3698 * Helper function for osd_create()
3700 * \retval 0, on success
3702 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3703 const struct lu_fid *fid, struct thandle *th)
3705 struct osd_thread_info *info = osd_oti_get(env);
3706 struct osd_inode_id *id = &info->oti_id;
3707 struct osd_device *osd = osd_obj2dev(obj);
3708 struct osd_thandle *oh;
3711 LASSERT(obj->oo_inode != NULL);
3713 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3716 oh = container_of(th, struct osd_thandle, ot_super);
3717 LASSERT(oh->ot_handle);
3718 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3720 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3721 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3722 OI_CHECK_FLD, NULL);
3723 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3724 struct lu_fid next_fid = *fid;
3726 /* insert next object in advance, and map to the same inode */
3728 if (next_fid.f_oid != 0) {
3729 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3730 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3731 OI_CHECK_FLD, NULL);
3732 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3736 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3741 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3742 u64 seq, struct lu_seq_range *range)
3744 struct seq_server_site *ss = osd_seq_site(osd);
3746 if (fid_seq_is_idif(seq)) {
3747 fld_range_set_ost(range);
3748 range->lsr_index = idif_ost_idx(seq);
3752 if (!fid_seq_in_fldb(seq)) {
3753 fld_range_set_mdt(range);
3756 * FIXME: If ss is NULL, it suppose not get lsr_index
3759 range->lsr_index = ss->ss_node_id;
3763 LASSERT(ss != NULL);
3764 fld_range_set_any(range);
3765 /* OSD will only do local fld lookup */
3766 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3769 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3770 struct lu_attr *attr,
3771 struct dt_allocation_hint *hint,
3772 struct dt_object_format *dof,
3773 struct thandle *handle)
3775 struct osd_thandle *oh;
3776 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3782 LASSERT(handle != NULL);
3784 oh = container_of(handle, struct osd_thandle, ot_super);
3785 LASSERT(oh->ot_handle == NULL);
3787 if (osd_tx_was_declared(env, oh, dt, DTO_OBJECT_CREATE, 0))
3791 * EA object consumes more credits than regular object: osd_mk_index
3792 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3793 * leaf_node, could involves the block, block bitmap, groups, GDT
3794 * change for each block, so add 4 * 2 credits in that case.
3796 * The default ACL initialization may consume an additional 16 blocks
3798 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3799 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3802 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3803 * credits for possible default ACL creation in new inode
3805 if (hint && hint->dah_acl_len)
3806 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3807 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3808 sb->s_blocksize_bits);
3810 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3813 * Reuse idle OI block may cause additional one OI block
3816 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3817 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3818 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3819 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3820 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3822 /* will help to find FID->ino mapping at dt_insert() */
3823 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3831 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3832 attr->la_projid, 1, oh, osd_dt_obj(dt),
3833 NULL, OSD_QID_INODE);
3841 * Called to destroy on-disk representation of the object
3843 * Concurrency: must be locked
3845 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3848 struct osd_object *obj = osd_dt_obj(dt);
3849 struct inode *inode = obj->oo_inode;
3850 struct osd_thandle *oh;
3858 oh = container_of(th, struct osd_thandle, ot_super);
3859 LASSERT(oh->ot_handle == NULL);
3861 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3862 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3864 /* For removing agent entry */
3865 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3866 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3869 * Recycle idle OI leaf may cause additional three OI blocks
3872 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3873 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3874 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3875 /* one less inode */
3876 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3877 i_projid_read(inode), -1, oh, obj, NULL,
3881 /* data to be truncated */
3882 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3883 i_projid_read(inode), 0, oh, obj, NULL,
3889 * will help to find FID->ino when this object is being
3892 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3897 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3900 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3901 struct osd_object *obj = osd_dt_obj(dt);
3902 struct inode *inode = obj->oo_inode;
3903 struct osd_device *osd = osd_obj2dev(obj);
3904 struct osd_thandle *oh;
3909 oh = container_of(th, struct osd_thandle, ot_super);
3910 LASSERT(oh->ot_handle);
3912 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3914 if (unlikely(fid_is_acct(fid)))
3917 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3918 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3920 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3921 osd_name(osd), PFID(fid), result);
3924 if (S_ISDIR(inode->i_mode)) {
3925 if (inode->i_nlink > 2)
3926 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3927 osd_name(osd), PFID(fid), inode->i_ino,
3930 spin_lock(&obj->oo_guard);
3932 spin_unlock(&obj->oo_guard);
3933 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3936 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3938 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3940 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3941 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3942 oh->ot_handle, OI_CHECK_FLD);
3944 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3945 /* XXX: add to ext3 orphan list */
3946 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3948 /* not needed in the cache anymore */
3949 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3950 obj->oo_destroyed = 1;
3956 * Put the fid into lustre_mdt_attrs, and then place the structure
3957 * inode's ea. This fid should not be altered during the life time
3960 * \retval +ve, on success
3961 * \retval -ve, on error
3963 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3965 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3966 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3968 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3969 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3974 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3977 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3980 lustre_loa_init(loa, fid, compat, incompat);
3981 lustre_loa_swab(loa, false);
3984 * For the OST device with 256 bytes inode size by default,
3985 * the PFID EA will be stored together with LMA EA to avoid
3986 * performance trouble. Otherwise the PFID EA can be stored
3987 * independently. LU-8998
3989 if ((compat & LMAC_FID_ON_OST) &&
3990 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3991 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3992 sizeof(*loa), XATTR_CREATE);
3994 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3995 sizeof(*lma), XATTR_CREATE);
3997 * LMA may already exist, but we need to check that all the
3998 * desired compat/incompat flags have been added.
4000 if (unlikely(rc == -EEXIST)) {
4001 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
4002 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
4006 if (rc < sizeof(*lma))
4009 lustre_loa_swab(loa, true);
4010 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
4011 ((compat == 0 && incompat == 0) ||
4012 (!(~lma->lma_compat & compat) &&
4013 !(~lma->lma_incompat & incompat))))
4016 lma->lma_self_fid = *fid;
4017 lma->lma_compat |= compat;
4018 lma->lma_incompat |= incompat;
4019 if (rc == sizeof(*lma)) {
4020 lustre_lma_swab(lma);
4021 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4022 sizeof(*lma), XATTR_REPLACE);
4024 lustre_loa_swab(loa, false);
4025 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
4026 sizeof(*loa), XATTR_REPLACE);
4034 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
4035 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
4036 * To have compatilibility with 1.8 ldiskfs driver we need to have
4037 * magic number at start of fid data.
4038 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
4041 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
4042 const struct lu_fid *fid)
4044 if (!fid_is_namespace_visible(fid) ||
4045 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
4046 param->edp_magic = 0;
4050 param->edp_magic = LDISKFS_LUFID_MAGIC;
4051 param->edp_len = sizeof(struct lu_fid) + 1;
4052 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
4056 * Try to read the fid from inode ea into dt_rec.
4058 * \param fid object fid.
4060 * \retval 0 on success
4062 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
4063 __u32 ino, struct lu_fid *fid,
4064 struct osd_inode_id *id)
4066 struct osd_thread_info *info = osd_oti_get(env);
4067 struct inode *inode;
4071 osd_id_gen(id, ino, OSD_OII_NOGEN);
4072 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
4074 RETURN(PTR_ERR(inode));
4080 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
4082 struct inode *parent_dir,
4083 const struct lu_fid *dot_fid,
4084 const struct lu_fid *dot_dot_fid,
4085 struct osd_thandle *oth)
4087 struct ldiskfs_dentry_param *dot_ldp;
4088 struct ldiskfs_dentry_param *dot_dot_ldp;
4089 __u32 saved_nlink = dir->i_nlink;
4092 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
4095 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
4096 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
4098 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
4099 dot_ldp->edp_magic = 0;
4101 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
4102 dir, dot_ldp, dot_dot_ldp);
4104 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
4105 * the subseqent ref_add() will increase the dir->i_nlink
4106 * as 3. That is incorrect for new created directory.
4108 * It looks like hack, because we want to make the OSD API
4109 * to be order-independent for new created directory object
4110 * between dt_insert(..) and ref_add() operations.
4112 * Here, we only restore the in-RAM dir-inode's nlink attr,
4113 * becuase if the nlink attr is not 2, then there will be
4114 * ref_add() called following the dt_insert(..), such call
4115 * will make both the in-RAM and on-disk dir-inode's nlink
4116 * attr to be set as 2. LU-7447
4118 set_nlink(dir, saved_nlink);
4123 * Create an local agent inode for remote entry
4125 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
4126 struct osd_device *osd,
4127 struct osd_object *pobj,
4128 const struct lu_fid *fid,
4132 struct osd_thread_info *info = osd_oti_get(env);
4133 struct inode *local;
4134 struct osd_thandle *oh;
4135 struct iattr iattr = {
4136 .ia_valid = ATTR_UID | ATTR_GID |
4137 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
4138 .ia_ctime.tv_nsec = UTIME_OMIT,
4139 .ia_mtime.tv_nsec = UTIME_OMIT,
4140 .ia_atime.tv_nsec = UTIME_OMIT,
4141 .ia_uid = GLOBAL_ROOT_UID,
4142 .ia_gid = GLOBAL_ROOT_GID,
4149 oh = container_of(th, struct osd_thandle, ot_super);
4150 LASSERT(oh->ot_handle->h_transaction != NULL);
4152 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
4154 if (IS_ERR(local)) {
4155 CERROR("%s: create local error %d\n", osd_name(osd),
4156 (int)PTR_ERR(local));
4161 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
4162 * correct gid on remote file, not agent here
4164 local->i_gid = current_fsgid();
4165 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4167 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4168 * That gives e2fsck something to look at and be happy, and allows
4169 * debugging if we need to determine where this symlink came from.
4171 if (S_ISLNK(type)) {
4172 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4173 ldiskfs_clear_inode_flag(local, LDISKFS_INODE_EXTENTS);
4174 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4175 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4177 i_size_write(local, rc);
4178 LDISKFS_I(local)->i_disksize = rc;
4180 unlock_new_inode(local);
4182 /* Agent inode should not have project ID */
4183 #ifdef HAVE_PROJECT_QUOTA
4184 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4185 i_projid_read(pobj->oo_inode) != 0) {
4186 rc = osd_transfer_project(local, 0, th);
4188 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4189 osd_ino2name(local), rc);
4190 RETURN(ERR_PTR(rc));
4194 /* Set special LMA flag for local agent inode */
4195 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4197 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4198 osd_name(osd), PFID(fid), rc);
4199 RETURN(ERR_PTR(rc));
4205 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode, fid,
4206 lu_object_fid(&pobj->oo_dt.do_lu),
4209 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4210 osd_name(osd), PFID(fid), rc);
4211 RETURN(ERR_PTR(rc));
4218 * when direntry is deleted, we have to take care of possible agent inode
4219 * referenced by that. unfortunately we can't do this at that point:
4220 * iget() within a running transaction leads to deadlock and we better do
4221 * not call that every delete declaration to save performance. so we put
4222 * a potention agent inode on a list and process that once the transaction
4223 * is over. Notice it's not any worse in terms of real orphans as regular
4224 * object destroy doesn't put inodes on the on-disk orphan list. this should
4225 * be addressed separately
4227 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4228 struct osd_thandle *oh,
4231 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4232 struct osd_obj_orphan *oor;
4239 oor->oor_env = (struct lu_env *)env;
4240 spin_lock(&osd->od_osfs_lock);
4241 list_add(&oor->oor_list, &osd->od_orphan_list);
4242 spin_unlock(&osd->od_osfs_lock);
4244 oh->ot_remove_agents = 1;
4250 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4251 struct osd_device *osd)
4253 struct osd_thread_info *info = osd_oti_get(env);
4254 struct osd_obj_orphan *oor, *tmp;
4255 struct osd_inode_id id;
4257 struct inode *inode;
4262 spin_lock(&osd->od_osfs_lock);
4263 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4264 if (oor->oor_env == env)
4265 list_move(&oor->oor_list, &list);
4267 spin_unlock(&osd->od_osfs_lock);
4269 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4273 list_del(&oor->oor_list);
4276 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4277 inode = osd_iget_fid(info, osd, &id, &fid);
4281 if (!osd_remote_fid(env, osd, &fid)) {
4286 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4288 mark_inode_dirty(inode);
4289 ldiskfs_journal_stop(jh);
4297 * OSD layer object create function for OST objects (b=11826).
4299 * The FID is inserted into inode xattr here.
4301 * \retval 0, on success
4302 * \retval -ve, on error
4304 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4305 struct lu_attr *attr, struct dt_allocation_hint *hint,
4306 struct dt_object_format *dof, struct thandle *th)
4308 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4309 struct osd_object *obj = osd_dt_obj(dt);
4310 struct osd_thread_info *info = osd_oti_get(env);
4311 int result, on_ost = 0;
4315 if (dt_object_exists(dt))
4318 LINVRNT(osd_invariant(obj));
4319 LASSERT(!dt_object_remote(dt));
4320 LASSERT(osd_is_write_locked(env, obj));
4321 LASSERT(th != NULL);
4323 if (unlikely(fid_is_acct(fid)))
4325 * Quota files can't be created from the kernel any more,
4326 * 'tune2fs -O quota' will take care of creating them
4330 result = __osd_create(info, obj, attr, hint, dof, th);
4332 if (fid_is_idif(fid) &&
4333 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4334 struct lu_fid *tfid = &info->oti_fid;
4335 struct ost_id *oi = &info->oti_ostid;
4337 fid_to_ostid(fid, oi);
4338 ostid_to_fid(tfid, oi, 0);
4340 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4341 LMAC_FID_ON_OST, 0);
4343 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4345 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4346 on_ost ? LMAC_FID_ON_OST : 0,
4349 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4350 obj->oo_dt.do_body_ops = &osd_body_ops;
4353 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4354 struct inode *inode = obj->oo_inode;
4356 result = __osd_oi_insert(env, obj, fid, th);
4357 if (result && inode) {
4358 spin_lock(&obj->oo_guard);
4360 spin_unlock(&obj->oo_guard);
4361 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4362 ldiskfs_set_inode_state(inode,
4363 LDISKFS_STATE_LUSTRE_DESTROY);
4365 obj->oo_inode = NULL;
4370 * a small optimization - dt_insert() isn't usually applied
4371 * to OST objects, so we don't need to cache OI mapping for
4374 if (result == 0 && on_ost == 0) {
4375 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4377 result = osd_idc_find_and_init(env, osd, obj);
4378 LASSERT(result == 0);
4381 LASSERT(ergo(result == 0,
4382 dt_object_exists(dt) && !dt_object_remote(dt)));
4383 LINVRNT(osd_invariant(obj));
4387 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4388 struct thandle *handle)
4390 struct osd_thandle *oh;
4393 /* it's possible that object doesn't exist yet */
4394 LASSERT(handle != NULL);
4396 oh = container_of(handle, struct osd_thandle, ot_super);
4397 LASSERT(oh->ot_handle == NULL);
4399 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4400 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4402 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4409 * Concurrency: @dt is write locked.
4411 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4414 struct osd_object *obj = osd_dt_obj(dt);
4415 struct inode *inode = obj->oo_inode;
4416 struct osd_thandle *oh;
4419 if (!dt_object_exists(dt) || obj->oo_destroyed)
4422 LINVRNT(osd_invariant(obj));
4423 LASSERT(!dt_object_remote(dt));
4424 LASSERT(osd_is_write_locked(env, obj));
4425 LASSERT(th != NULL);
4427 oh = container_of(th, struct osd_thandle, ot_super);
4428 LASSERT(oh->ot_handle != NULL);
4430 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4432 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4433 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4435 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4436 * (65000) subdirectories by storing "1" in i_nlink if the link count
4437 * would otherwise overflow. Directory tranversal tools understand
4438 * that (st_nlink == 1) indicates that the filesystem dose not track
4439 * hard links count on the directory, and will not abort subdirectory
4440 * scanning early once (st_nlink - 2) subdirs have been found.
4442 * This also has to properly handle the case of inodes with nlink == 0
4443 * in case they are being linked into the PENDING directory
4445 spin_lock(&obj->oo_guard);
4446 if (unlikely(inode->i_nlink == 0))
4447 /* inc_nlink from 0 may cause WARN_ON */
4448 set_nlink(inode, 1);
4450 osd_ldiskfs_inc_count(oh->ot_handle, inode);
4451 if (!S_ISDIR(inode->i_mode))
4452 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4454 spin_unlock(&obj->oo_guard);
4456 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4457 LINVRNT(osd_invariant(obj));
4459 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4464 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4465 struct thandle *handle)
4467 struct osd_thandle *oh;
4469 if (!dt_object_exists(dt))
4472 LASSERT(!dt_object_remote(dt));
4473 LASSERT(handle != NULL);
4475 oh = container_of(handle, struct osd_thandle, ot_super);
4476 LASSERT(oh->ot_handle == NULL);
4478 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4479 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4485 * Concurrency: @dt is write locked.
4487 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4490 struct osd_object *obj = osd_dt_obj(dt);
4491 struct inode *inode = obj->oo_inode;
4492 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4493 struct osd_thandle *oh;
4495 if (!dt_object_exists(dt))
4498 LINVRNT(osd_invariant(obj));
4499 LASSERT(!dt_object_remote(dt));
4500 LASSERT(osd_is_write_locked(env, obj));
4501 LASSERT(th != NULL);
4503 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4506 oh = container_of(th, struct osd_thandle, ot_super);
4507 LASSERT(oh->ot_handle != NULL);
4509 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4511 spin_lock(&obj->oo_guard);
4513 * That can be result of upgrade from old Lustre version and
4514 * applied only to local files. Just skip this ref_del call.
4515 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4517 if (inode->i_nlink == 0) {
4518 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4519 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4520 ", maybe an upgraded file? (LU-3915)\n",
4521 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4522 spin_unlock(&obj->oo_guard);
4526 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4527 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4529 osd_ldiskfs_dec_count(oh->ot_handle, inode);
4530 spin_unlock(&obj->oo_guard);
4532 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4533 LINVRNT(osd_invariant(obj));
4535 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4541 * Concurrency: @dt is read locked.
4543 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4544 struct lu_buf *buf, const char *name)
4546 struct osd_object *obj = osd_dt_obj(dt);
4547 struct inode *inode = obj->oo_inode;
4548 struct osd_thread_info *info = osd_oti_get(env);
4549 struct dentry *dentry = &info->oti_obj_dentry;
4550 bool cache_xattr = false;
4555 /* version get is not real XATTR but uses xattr API */
4556 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4557 dt_obj_version_t *ver = buf->lb_buf;
4560 * for version we are just using xattr API but change inode
4563 if (buf->lb_len == 0)
4564 return sizeof(dt_obj_version_t);
4566 if (buf->lb_len < sizeof(dt_obj_version_t))
4569 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4570 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4572 *ver = LDISKFS_I(inode)->i_fs_version;
4574 return sizeof(dt_obj_version_t);
4577 if (!dt_object_exists(dt))
4580 LASSERT(!dt_object_remote(dt));
4581 LASSERT(inode->i_op != NULL);
4582 #ifdef HAVE_IOP_XATTR
4583 LASSERT(inode->i_op->getxattr != NULL);
4586 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4587 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4591 rc = osd_oxc_get(obj, name, buf);
4596 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4597 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4598 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4599 struct filter_fid *ff;
4600 struct ost_layout *ol;
4602 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4606 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4609 if (buf->lb_len == 0 || !buf->lb_buf)
4612 if (buf->lb_len < rc)
4616 ol = &ff->ff_layout;
4617 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4618 PFID_STRIPE_IDX_BITS);
4619 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4620 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4621 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4622 if (lma->lma_compat & LMAC_COMP_INFO) {
4623 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4624 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4625 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4627 ol->ol_comp_start = 0;
4628 ol->ol_comp_end = 0;
4632 /* Get enc context xattr directly from ldiskfs instead of going
4633 * through the VFS, as there is no xattr handler for
4636 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0)
4637 rc = ldiskfs_xattr_get(inode,
4638 LDISKFS_XATTR_INDEX_ENCRYPTION,
4639 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
4640 buf->lb_buf, buf->lb_len);
4642 rc = __osd_xattr_get(inode, dentry, name,
4643 buf->lb_buf, buf->lb_len);
4647 if (rc == -ENOENT || rc == -ENODATA)
4648 osd_oxc_add(obj, name, NULL, 0);
4649 else if (rc > 0 && buf->lb_buf != NULL)
4650 osd_oxc_add(obj, name, buf->lb_buf, rc);
4656 static int osd_declare_xattr_set(const struct lu_env *env,
4657 struct dt_object *dt,
4658 const struct lu_buf *buf, const char *name,
4659 int fl, struct thandle *handle)
4661 struct osd_thandle *oh;
4663 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4665 LASSERT(handle != NULL);
4667 oh = container_of(handle, struct osd_thandle, ot_super);
4668 LASSERT(oh->ot_handle == NULL);
4670 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4672 * For non-upgrading case, the LMA is set first and
4673 * usually fit inode. But for upgrade case, the LMA
4674 * may be in another separated EA block.
4676 if (dt_object_exists(dt)) {
4677 if (fl == LU_XATTR_REPLACE)
4682 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4684 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4685 /* We may need to delete the old PFID EA. */
4686 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4687 if (fl == LU_XATTR_REPLACE)
4693 * If some name entry resides on remote MDT, then will create
4694 * agent entry under remote parent. On the other hand, if the
4695 * remote entry will be removed, then related agent entry may
4696 * need to be removed from the remote parent. So there may be
4697 * kinds of cases, let's declare enough credits. The credits
4698 * for create agent entry is enough for remove case.
4700 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4701 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4702 if (dt_object_exists(dt))
4703 credits += 1; /* For updating LMA */
4707 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4712 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4714 * learn xattr size from osd_xattr_get if
4715 * attribute has not been read yet
4717 buflen = __osd_xattr_get(
4718 osd_dt_obj(dt)->oo_inode,
4719 &osd_oti_get(env)->oti_obj_dentry,
4724 buflen = buf->lb_len;
4727 if (buflen > sb->s_blocksize) {
4728 credits += osd_calc_bkmap_credits(
4730 (buflen + sb->s_blocksize - 1) >>
4731 sb->s_blocksize_bits);
4735 * xattr set may involve inode quota change, reserve credits for
4736 * dquot_initialize()
4738 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4741 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4746 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4747 const struct lu_buf *buf, int fl,
4748 struct thandle *handle)
4750 struct osd_thread_info *info = osd_oti_get(env);
4751 struct dentry *dentry = &info->oti_obj_dentry;
4752 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4753 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4754 struct inode *inode = obj->oo_inode;
4755 struct filter_fid *ff = buf->lb_buf;
4756 struct ost_layout *ol = &ff->ff_layout;
4757 int flags = XATTR_REPLACE;
4762 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4765 rc = osd_get_lma(info, inode, dentry, loa);
4766 if (rc == -ENODATA) {
4767 /* Usually for upgarding from old device */
4768 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4769 LMAC_FID_ON_OST, 0);
4770 flags = XATTR_CREATE;
4775 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4776 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4779 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4780 /* Separate PFID EA from LMA */
4781 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4782 lustre_lma_swab(lma);
4783 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4784 sizeof(*lma), XATTR_REPLACE);
4786 obj->oo_pfid_in_lma = 0;
4787 rc = LU_XATTR_CREATE;
4793 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4797 * Old client does not send stripe information,
4798 * then store the PFID EA on disk separatedly.
4800 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4801 ol->ol_stripe_size == 0))
4804 /* Remove old PFID EA entry firstly. */
4805 dquot_initialize(inode);
4806 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4807 if (rc == -ENODATA) {
4808 /* XATTR_NAME_FID is already absent */
4815 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4816 if (likely(ol->ol_stripe_size != 0)) {
4817 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4818 PFID_STRIPE_IDX_BITS;
4819 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4820 lma->lma_compat |= LMAC_STRIPE_INFO;
4821 if (ol->ol_comp_id != 0) {
4822 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4823 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4824 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4825 lma->lma_compat |= LMAC_COMP_INFO;
4829 lustre_loa_swab(loa, false);
4831 /* Store the PFID EA inside LMA. */
4832 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4835 obj->oo_pfid_in_lma = 1;
4841 * In DNE environment, the object (in spite of regular file or directory)
4842 * and its name entry may reside on different MDTs. Under such case, we will
4843 * create an agent entry on the MDT where the object resides. The agent entry
4844 * references the object locally, that makes the object to be visible to the
4845 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4846 * such as 'tar' can handle the object properly.
4848 * We handle the agent entry during set linkEA that is the common interface
4849 * for both regular file and directroy, can handle kinds of cases, such as
4850 * create/link/unlink/rename, and so on.
4852 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4854 * XXX: There are two known issues:
4855 * 1. For one object, we will create at most one agent entry even if there
4856 * may be more than one cross-MDTs hard links on the object. So the local
4857 * e2fsck may claim that the object's nlink is larger than the name entries
4858 * that reference such inode. And in further, the e2fsck will fix the nlink
4859 * attribute to match the local references. Then it will cause the object's
4860 * nlink attribute to be inconsistent with the global references. it is bad
4861 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4862 * other hand, the global namespace LFSCK can repair the object's attribute
4863 * according to the linkEA.
4864 * 2. There may be too many hard links on the object as to its linkEA overflow,
4865 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4866 * case happened, then at this point, we do not know whether there are some
4867 * cross-MDTs reference. But there are local references, it guarantees that
4868 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4870 static int osd_xattr_handle_linkea(const struct lu_env *env,
4871 struct osd_device *osd,
4872 struct osd_object *obj,
4873 const struct lu_buf *buf,
4874 struct thandle *handle)
4876 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4877 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4878 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4879 struct lu_name tmpname;
4880 struct osd_thandle *oh;
4882 bool remote = false;
4886 oh = container_of(handle, struct osd_thandle, ot_super);
4887 LASSERT(oh->ot_handle != NULL);
4889 rc = linkea_init_with_rec(&ldata);
4891 linkea_first_entry(&ldata);
4892 while (ldata.ld_lee != NULL && !remote) {
4893 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4895 if (osd_remote_fid(env, osd, tfid) > 0)
4898 linkea_next_entry(&ldata);
4900 } else if (rc == -ENODATA) {
4906 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4907 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4909 CERROR("%s: failed to remove agent entry for "DFID
4910 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4911 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4912 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4914 CERROR("%s: failed to create agent entry for "DFID
4915 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4922 * Concurrency: @dt is write locked.
4924 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4925 const struct lu_buf *buf, const char *name, int fl,
4926 struct thandle *handle)
4928 struct osd_object *obj = osd_dt_obj(dt);
4929 struct osd_device *osd = osd_obj2dev(obj);
4930 struct inode *inode = obj->oo_inode;
4931 struct osd_thread_info *info = osd_oti_get(env);
4941 /* version set is not real XATTR */
4942 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4943 dt_obj_version_t *version = buf->lb_buf;
4946 * for version we are just using xattr API but change inode
4949 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4951 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4952 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4954 LDISKFS_I(inode)->i_fs_version = *version;
4956 * Version is set after all inode operations are finished,
4957 * so we should mark it dirty here
4959 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4964 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4965 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4968 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4971 * For the OST device with 256 bytes inode size by default,
4972 * the PFID EA will be stored together with LMA EA to avoid
4973 * performance trouble. Otherwise the PFID EA can be stored
4974 * independently. LU-8998
4976 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4977 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4978 LASSERT(buf->lb_buf);
4980 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4983 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4984 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4985 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4987 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4991 lma->lma_incompat |= LMAI_STRIPED;
4992 lustre_lma_swab(lma);
4993 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4994 sizeof(*lma), XATTR_REPLACE);
4997 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4998 LASSERT(!osd->od_is_ost);
5000 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
5005 if (fl & LU_XATTR_REPLACE)
5006 fs_flags |= XATTR_REPLACE;
5008 if (fl & LU_XATTR_CREATE)
5009 fs_flags |= XATTR_CREATE;
5011 if (strcmp(name, LL_XATTR_NAME_ENCRYPTION_CONTEXT) == 0) {
5012 /* Set enc context xattr directly in ldiskfs instead of going
5013 * through the VFS, as there is no xattr handler for
5016 struct osd_thandle *oth = container_of(handle,
5020 if (!oth->ot_handle)
5021 /* this should be already part of a transaction */
5024 rc = ldiskfs_xattr_set_handle(oth->ot_handle, inode,
5025 LDISKFS_XATTR_INDEX_ENCRYPTION,
5026 LDISKFS_XATTR_NAME_ENCRYPTION_CONTEXT,
5027 buf->lb_buf, len, fs_flags);
5029 rc = __osd_xattr_set(info, inode, name,
5030 buf->lb_buf, len, fs_flags);
5032 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5035 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5036 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5037 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
5043 * Concurrency: @dt is read locked.
5045 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
5046 const struct lu_buf *buf)
5048 struct osd_object *obj = osd_dt_obj(dt);
5049 struct osd_device *dev = osd_obj2dev(obj);
5050 struct inode *inode = obj->oo_inode;
5051 struct osd_thread_info *info = osd_oti_get(env);
5052 struct dentry *dentry = &info->oti_obj_dentry;
5055 if (!dt_object_exists(dt))
5058 LASSERT(!dt_object_remote(dt));
5059 LASSERT(inode->i_op != NULL);
5060 LASSERT(inode->i_op->listxattr != NULL);
5062 dentry->d_inode = inode;
5063 dentry->d_sb = inode->i_sb;
5064 rc = inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
5066 if (rc < 0 || buf->lb_buf == NULL)
5069 /* Hide virtual project ID xattr from list if disabled */
5070 if (!dev->od_enable_projid_xattr) {
5071 char *end = (char *)buf->lb_buf + rc;
5072 char *p = buf->lb_buf;
5075 char *next = p + strlen(p) + 1;
5077 if (strcmp(p, XATTR_NAME_PROJID) == 0) {
5079 memmove(p, next, end - next);
5091 static int osd_declare_xattr_del(const struct lu_env *env,
5092 struct dt_object *dt, const char *name,
5093 struct thandle *handle)
5095 struct osd_thandle *oh;
5096 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
5098 LASSERT(!dt_object_remote(dt));
5099 LASSERT(handle != NULL);
5101 oh = container_of(handle, struct osd_thandle, ot_super);
5102 LASSERT(oh->ot_handle == NULL);
5104 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
5105 osd_dto_credits_noquota[DTO_XATTR_SET]);
5107 * xattr del may involve inode quota change, reserve credits for
5108 * dquot_initialize()
5110 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
5116 * Concurrency: @dt is write locked.
5118 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
5119 const char *name, struct thandle *handle)
5121 struct osd_object *obj = osd_dt_obj(dt);
5122 struct inode *inode = obj->oo_inode;
5123 struct osd_thread_info *info = osd_oti_get(env);
5124 struct dentry *dentry = &info->oti_obj_dentry;
5127 if (!dt_object_exists(dt))
5130 LASSERT(!dt_object_remote(dt));
5131 LASSERT(inode->i_op != NULL);
5132 LASSERT(handle != NULL);
5133 #ifdef HAVE_IOP_XATTR
5134 LASSERT(inode->i_op->removexattr != NULL);
5137 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
5139 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
5140 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
5142 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
5143 &info->oti_ost_attrs);
5145 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
5147 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
5148 lustre_lma_swab(lma);
5149 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
5150 sizeof(*lma), XATTR_REPLACE);
5152 obj->oo_pfid_in_lma = 0;
5155 dquot_initialize(inode);
5156 dentry->d_inode = inode;
5157 dentry->d_sb = inode->i_sb;
5158 rc = ll_vfs_removexattr(dentry, inode, name);
5161 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
5164 (strcmp(name, XATTR_NAME_LOV) == 0 ||
5165 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
5166 osd_oxc_del(obj, name);
5171 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
5172 __u64 start, __u64 end)
5174 struct osd_object *obj = osd_dt_obj(dt);
5175 struct osd_device *dev = osd_obj2dev(obj);
5176 struct inode *inode = obj->oo_inode;
5181 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
5184 RETURN(PTR_ERR(file));
5186 file->f_mode |= FMODE_64BITHASH;
5187 rc = vfs_fsync_range(file, start, end, 0);
5194 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
5199 static bool osd_check_stale(struct dt_object *dt)
5207 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
5208 const struct dt_index_features *feat)
5210 struct iam_descr *descr;
5212 if (osd_object_is_root(o))
5213 return feat == &dt_directory_features;
5215 LASSERT(o->oo_dir != NULL);
5217 descr = o->oo_dir->od_container.ic_descr;
5218 if (feat == &dt_directory_features) {
5219 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
5224 return feat->dif_keysize_min <= descr->id_key_size &&
5225 descr->id_key_size <= feat->dif_keysize_max &&
5226 feat->dif_recsize_min <= descr->id_rec_size &&
5227 descr->id_rec_size <= feat->dif_recsize_max &&
5228 !(feat->dif_flags & (DT_IND_VARKEY |
5229 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5230 ergo(feat->dif_flags & DT_IND_UPDATE,
5231 1 /* XXX check that object (and fs) is writable */);
5235 static int osd_iam_container_init(const struct lu_env *env,
5236 struct osd_object *obj,
5237 struct osd_directory *dir)
5239 struct iam_container *bag = &dir->od_container;
5242 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5246 result = iam_container_setup(bag);
5248 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5250 iam_container_fini(bag);
5257 * Concurrency: no external locking is necessary.
5259 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5260 const struct dt_index_features *feat)
5264 struct osd_object *obj = osd_dt_obj(dt);
5266 LINVRNT(osd_invariant(obj));
5268 if (osd_object_is_root(obj)) {
5269 dt->do_index_ops = &osd_index_ea_ops;
5271 } else if (feat == &dt_directory_features) {
5272 dt->do_index_ops = &osd_index_ea_ops;
5273 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5278 } else if (unlikely(feat == &dt_otable_features)) {
5279 dt->do_index_ops = &osd_otable_ops;
5281 } else if (unlikely(feat == &dt_acct_features)) {
5282 dt->do_index_ops = &osd_acct_index_ops;
5285 } else if (!osd_has_index(obj)) {
5286 struct osd_directory *dir;
5287 struct osd_device *osd = osd_obj2dev(obj);
5288 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5293 spin_lock(&obj->oo_guard);
5294 if (obj->oo_dir == NULL)
5298 * Concurrent thread allocated container data.
5301 spin_unlock(&obj->oo_guard);
5303 * Now, that we have container data, serialize its
5306 down_write(&obj->oo_ext_idx_sem);
5308 * recheck under lock.
5311 if (osd_has_index(obj)) {
5316 result = osd_iam_container_init(env, obj, obj->oo_dir);
5317 if (result || feat == &dt_lfsck_namespace_features ||
5318 feat == &dt_lfsck_layout_orphan_features ||
5319 feat == &dt_lfsck_layout_dangling_features)
5322 result = osd_index_register(osd, fid,
5323 feat->dif_keysize_max,
5324 feat->dif_recsize_max);
5326 CWARN("%s: failed to register index "
5328 osd_name(osd), PFID(fid), result);
5329 else if (result > 0)
5332 CDEBUG(D_LFSCK, "%s: index object "DFID
5333 " (%d/%d) registered\n",
5334 osd_name(osd), PFID(fid),
5335 (int)feat->dif_keysize_max,
5336 (int)feat->dif_recsize_max);
5339 up_write(&obj->oo_ext_idx_sem);
5347 if (result == 0 && skip_iam == 0) {
5348 if (!osd_iam_index_probe(env, obj, feat))
5351 LINVRNT(osd_invariant(obj));
5356 static int osd_otable_it_attr_get(const struct lu_env *env,
5357 struct dt_object *dt,
5358 struct lu_attr *attr)
5364 static const struct dt_object_operations osd_obj_ops = {
5365 .do_read_lock = osd_read_lock,
5366 .do_write_lock = osd_write_lock,
5367 .do_read_unlock = osd_read_unlock,
5368 .do_write_unlock = osd_write_unlock,
5369 .do_write_locked = osd_write_locked,
5370 .do_attr_get = osd_attr_get,
5371 .do_declare_attr_set = osd_declare_attr_set,
5372 .do_attr_set = osd_attr_set,
5373 .do_ah_init = osd_ah_init,
5374 .do_declare_create = osd_declare_create,
5375 .do_create = osd_create,
5376 .do_declare_destroy = osd_declare_destroy,
5377 .do_destroy = osd_destroy,
5378 .do_index_try = osd_index_try,
5379 .do_declare_ref_add = osd_declare_ref_add,
5380 .do_ref_add = osd_ref_add,
5381 .do_declare_ref_del = osd_declare_ref_del,
5382 .do_ref_del = osd_ref_del,
5383 .do_xattr_get = osd_xattr_get,
5384 .do_declare_xattr_set = osd_declare_xattr_set,
5385 .do_xattr_set = osd_xattr_set,
5386 .do_declare_xattr_del = osd_declare_xattr_del,
5387 .do_xattr_del = osd_xattr_del,
5388 .do_xattr_list = osd_xattr_list,
5389 .do_object_sync = osd_object_sync,
5390 .do_invalidate = osd_invalidate,
5391 .do_check_stale = osd_check_stale,
5394 static const struct dt_object_operations osd_obj_otable_it_ops = {
5395 .do_attr_get = osd_otable_it_attr_get,
5396 .do_index_try = osd_index_try,
5399 static int osd_index_declare_iam_delete(const struct lu_env *env,
5400 struct dt_object *dt,
5401 const struct dt_key *key,
5402 struct thandle *handle)
5404 struct osd_thandle *oh;
5406 oh = container_of(handle, struct osd_thandle, ot_super);
5407 LASSERT(oh->ot_handle == NULL);
5409 /* Recycle may cause additional three blocks to be changed. */
5410 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5411 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5417 * delete a (key, value) pair from index \a dt specified by \a key
5419 * \param dt osd index object
5420 * \param key key for index
5421 * \param rec record reference
5422 * \param handle transaction handler
5425 * \retval -ve failure
5427 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5428 const struct dt_key *key,
5429 struct thandle *handle)
5431 struct osd_thread_info *oti = osd_oti_get(env);
5432 struct osd_object *obj = osd_dt_obj(dt);
5433 struct osd_thandle *oh;
5434 struct iam_path_descr *ipd;
5435 struct iam_container *bag = &obj->oo_dir->od_container;
5440 if (!dt_object_exists(dt))
5443 LINVRNT(osd_invariant(obj));
5444 LASSERT(!dt_object_remote(dt));
5445 LASSERT(bag->ic_object == obj->oo_inode);
5446 LASSERT(handle != NULL);
5448 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5450 ipd = osd_idx_ipd_get(env, bag);
5451 if (unlikely(ipd == NULL))
5454 oh = container_of(handle, struct osd_thandle, ot_super);
5455 LASSERT(oh->ot_handle != NULL);
5456 LASSERT(oh->ot_handle->h_transaction != NULL);
5458 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5459 /* swab quota uid/gid provided by caller */
5460 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5461 key = (const struct dt_key *)&oti->oti_quota_id;
5464 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5465 osd_ipd_put(env, bag, ipd);
5466 LINVRNT(osd_invariant(obj));
5467 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5471 static int osd_index_declare_ea_delete(const struct lu_env *env,
5472 struct dt_object *dt,
5473 const struct dt_key *key,
5474 struct thandle *handle)
5476 struct osd_thandle *oh;
5477 struct inode *inode;
5482 LASSERT(!dt_object_remote(dt));
5483 LASSERT(handle != NULL);
5485 oh = container_of(handle, struct osd_thandle, ot_super);
5486 LASSERT(oh->ot_handle == NULL);
5488 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5489 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5491 inode = osd_dt_obj(dt)->oo_inode;
5495 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5496 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5501 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5504 struct osd_fid_pack *rec;
5507 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5508 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5509 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5510 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5516 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5517 const struct lu_fid *fid)
5519 struct seq_server_site *ss = osd_seq_site(osd);
5523 /* FID seqs not in FLDB, must be local seq */
5524 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5528 * If FLD is not being initialized yet, it only happens during the
5529 * initialization, likely during mgs initialization, and we assume
5530 * this is local FID.
5532 if (ss == NULL || ss->ss_server_fld == NULL)
5535 /* Only check the local FLDB here */
5536 if (osd_seq_exists(env, osd, fid_seq(fid)))
5542 static void osd_take_care_of_agent(const struct lu_env *env,
5543 struct osd_device *osd,
5544 struct osd_thandle *oh,
5545 struct ldiskfs_dir_entry_2 *de)
5547 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5548 struct osd_idmap_cache *idc;
5549 int rc, schedule = 0;
5551 LASSERT(de != NULL);
5553 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5554 if (likely(rc == 0)) {
5555 idc = osd_idc_find_or_init(env, osd, fid);
5556 if (IS_ERR(idc) || idc->oic_remote)
5558 } else if (rc == -ENODATA) {
5560 * can't get FID, postpone to the end of the
5561 * transaction when iget() is safe
5565 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5568 osd_schedule_agent_inode_removal(env, oh,
5569 le32_to_cpu(de->inode));
5573 * Utility function to get real name from object name
5575 * \param[in] obj pointer to the object to be handled
5576 * \param[in] name object name
5577 * \param[in] len object name len
5578 * \param[out]ln pointer to the struct lu_name to hold the real name
5580 * If file is not encrypted, real name is just the object name.
5581 * If file is encrypted, object name needs to be decoded. In
5582 * this case a new buffer is allocated, and ln->ln_name needs to be freed by
5585 * \retval 0, on success
5586 * \retval -ve, on error
5588 static int obj_name2lu_name(struct osd_object *obj, const char *name,
5589 int len, struct lu_name *ln)
5591 if (!(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
5593 ln->ln_namelen = len;
5595 char *buf = kmalloc(len, GFP_NOFS);
5600 len = critical_decode(name, len, buf);
5602 ln->ln_namelen = len;
5609 * Index delete function for interoperability mode (b11826).
5610 * It will remove the directory entry added by osd_index_ea_insert().
5611 * This entry is needed to maintain name->fid mapping.
5613 * \param key, key i.e. file entry to be deleted
5615 * \retval 0, on success
5616 * \retval -ve, on error
5618 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5619 const struct dt_key *key, struct thandle *handle)
5621 struct osd_object *obj = osd_dt_obj(dt);
5622 struct inode *dir = obj->oo_inode;
5623 struct dentry *dentry;
5624 struct osd_thandle *oh;
5625 struct ldiskfs_dir_entry_2 *de = NULL;
5626 struct buffer_head *bh;
5627 struct htree_lock *hlock = NULL;
5628 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5634 if (!dt_object_exists(dt))
5637 LINVRNT(osd_invariant(obj));
5638 LASSERT(!dt_object_remote(dt));
5639 LASSERT(handle != NULL);
5641 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
5645 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5647 oh = container_of(handle, struct osd_thandle, ot_super);
5648 LASSERT(oh->ot_handle != NULL);
5649 LASSERT(oh->ot_handle->h_transaction != NULL);
5651 dquot_initialize(dir);
5652 dentry = osd_child_dentry_get(env, obj, ln.ln_name, ln.ln_namelen);
5654 if (obj->oo_hl_head != NULL) {
5655 hlock = osd_oti_get(env)->oti_hlock;
5656 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5657 dir, LDISKFS_HLOCK_DEL);
5659 down_write(&obj->oo_ext_idx_sem);
5662 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5665 * If this is not the ".." entry, it might be a remote DNE
5666 * entry and we need to check if the FID is for a remote
5667 * MDT. If the FID is not in the directory entry (e.g.
5668 * upgraded 1.8 filesystem without dirdata enabled) then
5669 * we need to get the FID from the LMA. For a remote directory
5670 * there HAS to be an LMA, it cannot be an IGIF inode in this
5673 * Delete the entry before the agent inode in order to
5674 * simplify error handling. At worst an error after deleting
5675 * the entry first might leak the agent inode afterward. The
5676 * reverse would need filesystem abort in case of error deleting
5677 * the entry after the agent had been removed, or leave a
5678 * dangling entry pointing at a random inode.
5680 if (strcmp((char *)key, dotdot) != 0)
5681 osd_take_care_of_agent(env, osd, oh, de);
5682 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5688 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5689 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5690 /* NB, dirent count may not be accurate, because it's counted
5693 if (obj->oo_dirent_count)
5694 obj->oo_dirent_count--;
5696 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5699 ldiskfs_htree_unlock(hlock);
5701 up_write(&obj->oo_ext_idx_sem);
5704 LASSERT(osd_invariant(obj));
5705 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5706 if (ln.ln_name != (char *)key)
5712 * Lookup index for \a key and copy record to \a rec.
5714 * \param dt osd index object
5715 * \param key key for index
5716 * \param rec record reference
5718 * \retval +ve success : exact mach
5719 * \retval 0 return record with key not greater than \a key
5720 * \retval -ve failure
5722 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5723 struct dt_rec *rec, const struct dt_key *key)
5725 struct osd_object *obj = osd_dt_obj(dt);
5726 struct iam_path_descr *ipd;
5727 struct iam_container *bag = &obj->oo_dir->od_container;
5728 struct osd_thread_info *oti = osd_oti_get(env);
5729 struct iam_iterator *it = &oti->oti_idx_it;
5730 struct iam_rec *iam_rec;
5735 if (!dt_object_exists(dt))
5738 LASSERT(osd_invariant(obj));
5739 LASSERT(!dt_object_remote(dt));
5740 LASSERT(bag->ic_object == obj->oo_inode);
5742 ipd = osd_idx_ipd_get(env, bag);
5746 /* got ipd now we can start iterator. */
5747 iam_it_init(it, bag, 0, ipd);
5749 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5750 /* swab quota uid/gid provided by caller */
5751 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5752 key = (const struct dt_key *)&oti->oti_quota_id;
5755 rc = iam_it_get(it, (struct iam_key *)key);
5757 if (S_ISDIR(obj->oo_inode->i_mode))
5758 iam_rec = (struct iam_rec *)oti->oti_ldp;
5760 iam_rec = (struct iam_rec *)rec;
5762 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5764 if (S_ISDIR(obj->oo_inode->i_mode))
5765 osd_fid_unpack((struct lu_fid *)rec,
5766 (struct osd_fid_pack *)iam_rec);
5767 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5768 osd_quota_unpack(obj, rec);
5773 osd_ipd_put(env, bag, ipd);
5775 LINVRNT(osd_invariant(obj));
5780 static int osd_index_declare_iam_insert(const struct lu_env *env,
5781 struct dt_object *dt,
5782 const struct dt_rec *rec,
5783 const struct dt_key *key,
5784 struct thandle *handle)
5786 struct osd_thandle *oh;
5788 LASSERT(handle != NULL);
5790 oh = container_of(handle, struct osd_thandle, ot_super);
5791 LASSERT(oh->ot_handle == NULL);
5793 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5794 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5800 * Inserts (key, value) pair in \a dt index object.
5802 * \param dt osd index object
5803 * \param key key for index
5804 * \param rec record reference
5805 * \param th transaction handler
5808 * \retval -ve failure
5810 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5811 const struct dt_rec *rec,
5812 const struct dt_key *key, struct thandle *th)
5814 struct osd_object *obj = osd_dt_obj(dt);
5815 struct iam_path_descr *ipd;
5816 struct osd_thandle *oh;
5817 struct iam_container *bag;
5818 struct osd_thread_info *oti = osd_oti_get(env);
5819 struct iam_rec *iam_rec;
5824 if (!dt_object_exists(dt))
5827 LINVRNT(osd_invariant(obj));
5828 LASSERT(!dt_object_remote(dt));
5830 bag = &obj->oo_dir->od_container;
5831 LASSERT(bag->ic_object == obj->oo_inode);
5832 LASSERT(th != NULL);
5834 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5836 ipd = osd_idx_ipd_get(env, bag);
5837 if (unlikely(ipd == NULL))
5840 oh = container_of(th, struct osd_thandle, ot_super);
5841 LASSERT(oh->ot_handle != NULL);
5842 LASSERT(oh->ot_handle->h_transaction != NULL);
5843 if (S_ISDIR(obj->oo_inode->i_mode)) {
5844 iam_rec = (struct iam_rec *)oti->oti_ldp;
5845 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5847 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5848 /* pack quota uid/gid */
5849 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5850 key = (const struct dt_key *)&oti->oti_quota_id;
5851 /* pack quota record */
5852 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5853 iam_rec = (struct iam_rec *)rec;
5855 iam_rec = (struct iam_rec *)rec;
5858 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5860 osd_ipd_put(env, bag, ipd);
5861 LINVRNT(osd_invariant(obj));
5862 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5867 * Calls ldiskfs_add_entry() to add directory entry
5868 * into the directory. This is required for
5869 * interoperability mode (b11826)
5871 * \retval 0, on success
5872 * \retval -ve, on error
5874 static int __osd_ea_add_rec(struct osd_thread_info *info,
5875 struct osd_object *pobj, struct inode *cinode,
5876 const char *name, const struct lu_fid *fid,
5877 struct htree_lock *hlock, struct thandle *th)
5879 struct ldiskfs_dentry_param *ldp;
5880 struct dentry *child;
5881 struct osd_thandle *oth;
5885 oth = container_of(th, struct osd_thandle, ot_super);
5886 LASSERT(oth->ot_handle != NULL);
5887 LASSERT(oth->ot_handle->h_transaction != NULL);
5888 LASSERT(pobj->oo_inode);
5890 rc = obj_name2lu_name(pobj, name, strlen(name), &ln);
5894 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5895 if (unlikely(osd_object_is_root(pobj)))
5898 osd_get_ldiskfs_dirent_param(ldp, fid);
5899 child = osd_child_dentry_get(info->oti_env, pobj,
5900 ln.ln_name, ln.ln_namelen);
5901 child->d_fsdata = (void *)ldp;
5902 dquot_initialize(pobj->oo_inode);
5903 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5904 child, cinode, hlock);
5905 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5906 struct ldiskfs_dir_entry_2 *de;
5907 struct buffer_head *bh;
5910 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5913 rc1 = osd_ldiskfs_journal_get_write_access(
5914 oth->ot_handle, pobj->oo_inode->i_sb, bh,
5917 if (S_ISDIR(cinode->i_mode))
5918 de->file_type = LDISKFS_DIRENT_LUFID |
5919 LDISKFS_FT_REG_FILE;
5921 de->file_type = LDISKFS_DIRENT_LUFID |
5923 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5930 if (ln.ln_name != name)
5936 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5937 * into the directory.Also sets flags into osd object to
5938 * indicate dot and dotdot are created. This is required for
5939 * interoperability mode (b11826)
5941 * \param dir directory for dot and dotdot fixup.
5942 * \param obj child object for linking
5944 * \retval 0, on success
5945 * \retval -ve, on error
5947 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5948 struct osd_object *dir,
5949 struct inode *parent_dir, const char *name,
5950 const struct lu_fid *dot_fid,
5951 const struct lu_fid *dot_dot_fid,
5954 struct inode *inode = dir->oo_inode;
5955 struct osd_thandle *oth;
5958 oth = container_of(th, struct osd_thandle, ot_super);
5959 LASSERT(oth->ot_handle->h_transaction != NULL);
5960 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5962 if (strcmp(name, dot) == 0) {
5963 if (dir->oo_compat_dot_created) {
5966 LASSERT(inode->i_ino == parent_dir->i_ino);
5967 dir->oo_compat_dot_created = 1;
5970 } else if (strcmp(name, dotdot) == 0) {
5971 if (!dir->oo_compat_dot_created)
5973 /* in case of rename, dotdot is already created */
5974 if (dir->oo_compat_dotdot_created) {
5975 return __osd_ea_add_rec(info, dir, parent_dir, name,
5976 dot_dot_fid, NULL, th);
5979 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5980 struct lu_fid tfid = *dot_dot_fid;
5983 result = osd_add_dot_dotdot_internal(info,
5984 dir->oo_inode, parent_dir, dot_fid,
5987 result = osd_add_dot_dotdot_internal(info,
5988 dir->oo_inode, parent_dir, dot_fid,
5993 dir->oo_compat_dotdot_created = 1;
6001 * It will call the appropriate osd_add* function and return the
6002 * value, return by respective functions.
6004 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
6005 struct inode *cinode, const char *name,
6006 const struct lu_fid *fid, struct thandle *th)
6008 struct osd_thread_info *info = osd_oti_get(env);
6009 struct htree_lock *hlock;
6012 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
6014 if (name[0] == '.' && (name[1] == '\0' ||
6015 (name[1] == '.' && name[2] == '\0'))) {
6016 if (hlock != NULL) {
6017 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6020 down_write(&pobj->oo_ext_idx_sem);
6023 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
6024 lu_object_fid(&pobj->oo_dt.do_lu),
6027 if (hlock != NULL) {
6028 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
6029 pobj->oo_inode, LDISKFS_HLOCK_ADD);
6031 down_write(&pobj->oo_ext_idx_sem);
6034 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
6035 struct lu_fid *tfid = &info->oti_fid;
6039 rc = __osd_ea_add_rec(info, pobj, cinode, name,
6042 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
6046 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
6047 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
6048 pobj->oo_dirent_count++;
6051 ldiskfs_htree_unlock(hlock);
6053 up_write(&pobj->oo_ext_idx_sem);
6059 osd_ldiskfs_consistency_check(struct osd_thread_info *oti,
6060 struct osd_device *dev,
6061 const struct lu_fid *fid,
6062 struct osd_inode_id *id)
6064 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
6065 struct inode *inode = NULL;
6071 if (!scrub_needs_check(scrub, fid, id->oii_ino))
6074 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
6075 if (rc == -ENOENT) {
6076 __u32 gen = id->oii_gen;
6082 inode = osd_iget(oti, dev, id);
6083 /* The inode has been removed (by race maybe). */
6084 if (IS_ERR(inode)) {
6085 rc = PTR_ERR(inode);
6087 RETURN(rc == -ESTALE ? -ENOENT : rc);
6090 /* The OI mapping is lost. */
6091 if (gen != OSD_OII_NOGEN)
6095 * The inode may has been reused by others, we do not know,
6096 * leave it to be handled by subsequent osd_fid_lookup().
6099 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
6106 if (scrub->os_running) {
6107 if (inode == NULL) {
6108 inode = osd_iget(oti, dev, id);
6109 /* The inode has been removed (by race maybe). */
6110 if (IS_ERR(inode)) {
6111 rc = PTR_ERR(inode);
6113 RETURN(rc == -ESTALE ? -ENOENT : rc);
6117 rc = osd_scrub_oi_insert(dev, fid, id, insert);
6119 * There is race condition between osd_oi_lookup and OI scrub.
6120 * The OI scrub finished just after osd_oi_lookup() failure.
6121 * Under such case, it is unnecessary to trigger OI scrub again,
6122 * but try to call osd_oi_lookup() again.
6124 if (unlikely(rc == -EAGAIN))
6127 if (!S_ISDIR(inode->i_mode))
6130 rc = osd_check_lmv(oti, dev, inode);
6135 if (dev->od_scrub.os_scrub.os_auto_scrub_interval != AS_NEVER &&
6137 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
6138 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
6139 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
6140 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
6141 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
6142 if (rc == 0 || rc == -EALREADY)
6154 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
6155 struct osd_device *dev,
6156 struct lu_fid *fid, __u32 ino)
6158 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
6159 struct osd_idmap_cache *oic = &oti->oti_cache;
6160 struct inode *inode;
6163 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
6164 inode = osd_iget(oti, dev, &oic->oic_lid);
6165 if (IS_ERR(inode)) {
6166 fid_zero(&oic->oic_fid);
6167 return PTR_ERR(inode);
6170 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
6173 fid_zero(&oic->oic_fid);
6175 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
6179 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
6180 struct osd_inode_id *id, const struct lu_fid *fid)
6182 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
6183 id->oii_ino, id->oii_gen, info);
6184 info->oti_cache.oic_lid = *id;
6185 info->oti_cache.oic_fid = *fid;
6186 info->oti_cache.oic_dev = osd;
6190 * Get parent FID from the linkEA.
6192 * For a directory which parent resides on remote MDT, to satisfy the
6193 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
6194 * the other hand, to make the lookup(..) on the directory can return
6195 * the real parent FID, we append the real parent FID after its ".."
6196 * name entry in the /REMOTE_PARENT_DIR.
6198 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
6199 * backup. So after the restore, we cannot get the right parent FID from
6200 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
6201 * we have stored the real parent FID in the directory object's linkEA,
6202 * we can parse the linkEA for the real parent FID.
6204 * \param[in] env pointer to the thread context
6205 * \param[in] obj pointer to the object to be handled
6206 * \param[out]fid pointer to the buffer to hold the parent FID
6208 * \retval 0 for getting the real parent FID successfully
6209 * \retval negative error number on failure
6211 static int osd_get_pfid_from_linkea(const struct lu_env *env,
6212 struct osd_object *obj,
6215 struct osd_thread_info *oti = osd_oti_get(env);
6216 struct lu_buf *buf = &oti->oti_big_buf;
6217 struct dentry *dentry = &oti->oti_obj_dentry;
6218 struct inode *inode = obj->oo_inode;
6219 struct linkea_data ldata = { NULL };
6225 if (!S_ISDIR(inode->i_mode))
6229 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6230 buf->lb_buf, buf->lb_len);
6231 if (rc == -ERANGE) {
6232 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6235 lu_buf_realloc(buf, rc);
6236 if (buf->lb_buf == NULL)
6243 if (unlikely(rc == 0))
6249 if (unlikely(buf->lb_buf == NULL)) {
6250 lu_buf_realloc(buf, rc);
6251 if (buf->lb_buf == NULL)
6258 rc = linkea_init_with_rec(&ldata);
6260 linkea_first_entry(&ldata);
6261 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
6267 static int osd_verify_ent_by_linkea(const struct lu_env *env,
6268 struct inode *inode,
6269 const struct lu_fid *pfid,
6270 const char *name, const int namelen)
6272 struct osd_thread_info *oti = osd_oti_get(env);
6273 struct lu_buf *buf = &oti->oti_big_buf;
6274 struct dentry *dentry = &oti->oti_obj_dentry;
6275 struct linkea_data ldata = { NULL };
6276 struct lu_name cname = { .ln_name = name,
6277 .ln_namelen = namelen };
6283 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6284 buf->lb_buf, buf->lb_len);
6286 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6291 if (unlikely(rc == 0))
6294 if (buf->lb_len < rc) {
6295 lu_buf_realloc(buf, rc);
6296 if (buf->lb_buf == NULL)
6303 rc = linkea_init_with_rec(&ldata);
6305 rc = linkea_links_find(&ldata, &cname, pfid);
6311 * Calls ->lookup() to find dentry. From dentry get inode and
6312 * read inode's ea to get fid. This is required for interoperability
6315 * \retval 0, on success
6316 * \retval -ve, on error
6318 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6319 struct dt_rec *rec, const struct dt_key *key)
6321 struct inode *dir = obj->oo_inode;
6322 struct dentry *dentry;
6323 struct ldiskfs_dir_entry_2 *de;
6324 struct buffer_head *bh;
6325 struct lu_fid *fid = (struct lu_fid *)rec;
6326 struct htree_lock *hlock = NULL;
6333 LASSERT(dir->i_op != NULL);
6334 LASSERT(dir->i_op->lookup != NULL);
6336 rc = obj_name2lu_name(obj, (char *)key, strlen((char *)key), &ln);
6340 dentry = osd_child_dentry_get(env, obj, ln.ln_name, ln.ln_namelen);
6342 if (obj->oo_hl_head != NULL) {
6343 hlock = osd_oti_get(env)->oti_hlock;
6344 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6345 dir, LDISKFS_HLOCK_LOOKUP);
6347 down_read(&obj->oo_ext_idx_sem);
6350 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6352 struct osd_thread_info *oti = osd_oti_get(env);
6353 struct osd_inode_id *id = &oti->oti_id;
6354 struct osd_device *dev = osd_obj2dev(obj);
6356 ino = le32_to_cpu(de->inode);
6357 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6359 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6363 rc = osd_get_fid_from_dentry(de, rec);
6365 /* done with de, release bh */
6368 if (unlikely(is_remote_parent_ino(dev, ino))) {
6369 const char *name = (const char *)key;
6372 * If the parent is on remote MDT, and there
6373 * is no FID-in-dirent, then we have to get
6374 * the parent FID from the linkEA.
6376 if (likely(strlen(name) == 2 &&
6377 name[0] == '.' && name[1] == '.'))
6378 rc = osd_get_pfid_from_linkea(env, obj,
6381 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6384 osd_id_gen(id, ino, OSD_OII_NOGEN);
6387 if (rc != 0 || osd_remote_fid(env, dev, fid))
6390 rc = osd_ldiskfs_consistency_check(oti, dev, fid, id);
6391 if (rc != -ENOENT) {
6392 /* Other error should not affect lookup result. */
6395 /* Normal file mapping should be added into OI cache
6396 * after FID in LMA check, but for local files like
6397 * hsm_actions, their FIDs are not stored in OI files,
6398 * see osd_initial_OI_scrub(), and here is the only
6399 * place to load mapping into OI cache.
6401 if (!fid_is_namespace_visible(fid))
6402 osd_add_oi_cache(osd_oti_get(env),
6403 osd_obj2dev(obj), id, fid);
6405 CDEBUG(D_INODE, DFID"/"DNAME" => "DFID"\n",
6406 PFID(lu_object_fid(&obj->oo_dt.do_lu)), PNAME(&ln),
6416 ldiskfs_htree_unlock(hlock);
6418 up_read(&obj->oo_ext_idx_sem);
6419 if (ln.ln_name != (char *)key)
6424 static int osd_index_declare_ea_insert(const struct lu_env *env,
6425 struct dt_object *dt,
6426 const struct dt_rec *rec,
6427 const struct dt_key *key,
6428 struct thandle *handle)
6430 struct osd_thandle *oh;
6431 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6432 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6433 const struct lu_fid *fid = rec1->rec_fid;
6434 int credits, rc = 0;
6435 struct osd_idmap_cache *idc;
6439 LASSERT(!dt_object_remote(dt));
6440 LASSERT(handle != NULL);
6441 LASSERT(fid != NULL);
6442 LASSERT(rec1->rec_type != 0);
6444 oh = container_of(handle, struct osd_thandle, ot_super);
6445 LASSERT(oh->ot_handle == NULL);
6447 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6450 * we can't call iget() while a transactions is running
6451 * (this can lead to a deadlock), but we need to know
6452 * inum and object type. so we find this information at
6453 * declaration and cache in per-thread info
6455 idc = osd_idc_find_or_init(env, osd, fid);
6457 RETURN(PTR_ERR(idc));
6458 if (idc->oic_remote) {
6460 * a reference to remote inode is represented by an
6461 * agent inode which we have to create
6463 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6464 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6467 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6469 if (osd_dt_obj(dt)->oo_inode != NULL) {
6470 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6473 * We ignore block quota on meta pool (MDTs), so needn't
6474 * calculate how many blocks will be consumed by this index
6477 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6479 i_projid_read(inode), 0,
6480 oh, osd_dt_obj(dt), NULL,
6485 #ifdef HAVE_PROJECT_QUOTA
6487 * Reserve credits for local agent inode to transfer
6488 * to 0, quota enforcement is ignored in this case.
6490 if (idc->oic_remote &&
6491 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6492 i_projid_read(inode) != 0)
6493 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6494 0, i_projid_read(inode),
6495 0, false, PRJQUOTA);
6503 * Index add function for interoperability mode (b11826).
6504 * It will add the directory entry.This entry is needed to
6505 * maintain name->fid mapping.
6507 * \param key it is key i.e. file entry to be inserted
6508 * \param rec it is value of given key i.e. fid
6510 * \retval 0, on success
6511 * \retval -ve, on error
6513 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6514 const struct dt_rec *rec,
6515 const struct dt_key *key, struct thandle *th)
6517 struct osd_object *obj = osd_dt_obj(dt);
6518 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6519 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6520 const struct lu_fid *fid = rec1->rec_fid;
6521 const char *name = (const char *)key;
6522 struct osd_thread_info *oti = osd_oti_get(env);
6523 struct inode *child_inode = NULL;
6524 struct osd_idmap_cache *idc;
6529 if (!dt_object_exists(dt))
6532 LASSERT(osd_invariant(obj));
6533 LASSERT(!dt_object_remote(dt));
6534 LASSERT(th != NULL);
6536 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6538 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6540 idc = osd_idc_find(env, osd, fid);
6541 if (unlikely(idc == NULL)) {
6542 idc = osd_idc_find_or_init(env, osd, fid);
6545 * this dt_insert() wasn't declared properly, so
6546 * FID is missing in OI cache. we better do not
6547 * lookup FID in FLDB/OI and don't risk to deadlock,
6548 * but in some special cases (lfsck testing, etc)
6549 * it's much simpler than fixing a caller.
6551 * normally this error should be placed after the first
6552 * find, but migrate may attach source stripes to
6553 * target, which doesn't create stripes.
6555 CERROR("%s: "DFID" wasn't declared for insert\n",
6556 osd_name(osd), PFID(fid));
6558 RETURN(PTR_ERR(idc));
6562 if (idc->oic_remote) {
6563 /* Insert remote entry */
6564 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6566 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6568 child_inode = osd_create_local_agent_inode(env, osd,
6569 obj, fid, rec1->rec_type & S_IFMT, th);
6570 if (IS_ERR(child_inode))
6571 RETURN(PTR_ERR(child_inode));
6574 /* Insert local entry */
6575 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6576 /* for a reason OI cache wasn't filled properly */
6577 CERROR("%s: OIC for "DFID" isn't filled\n",
6578 osd_name(osd), PFID(fid));
6581 child_inode = oti->oti_inode;
6582 if (unlikely(child_inode == NULL)) {
6583 struct ldiskfs_inode_info *lii;
6588 child_inode = oti->oti_inode = &lii->vfs_inode;
6590 child_inode->i_sb = osd_sb(osd);
6591 child_inode->i_ino = idc->oic_lid.oii_ino;
6592 child_inode->i_mode = rec1->rec_type & S_IFMT;
6595 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6597 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6598 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6600 if (child_inode && child_inode != oti->oti_inode)
6602 LASSERT(osd_invariant(obj));
6603 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6609 * Initialize osd Iterator for given osd index object.
6611 * \param dt osd index object
6614 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6615 struct dt_object *dt,
6618 struct osd_it_iam *it;
6619 struct osd_object *obj = osd_dt_obj(dt);
6620 struct lu_object *lo = &dt->do_lu;
6621 struct iam_path_descr *ipd;
6622 struct iam_container *bag = &obj->oo_dir->od_container;
6624 if (!dt_object_exists(dt))
6625 return ERR_PTR(-ENOENT);
6629 return ERR_PTR(-ENOMEM);
6631 ipd = osd_it_ipd_get(env, bag);
6632 if (likely(ipd != NULL)) {
6636 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6637 return (struct dt_it *)it;
6640 return ERR_PTR(-ENOMEM);
6645 * free given Iterator.
6647 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6649 struct osd_it_iam *it = (struct osd_it_iam *)di;
6650 struct osd_object *obj = it->oi_obj;
6652 iam_it_fini(&it->oi_it);
6653 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6654 osd_object_put(env, obj);
6659 * Move Iterator to record specified by \a key
6661 * \param di osd iterator
6662 * \param key key for index
6664 * \retval +ve di points to record with least key not larger than key
6665 * \retval 0 di points to exact matched key
6666 * \retval -ve failure
6669 static int osd_it_iam_get(const struct lu_env *env,
6670 struct dt_it *di, const struct dt_key *key)
6672 struct osd_thread_info *oti = osd_oti_get(env);
6673 struct osd_it_iam *it = (struct osd_it_iam *)di;
6675 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6676 /* swab quota uid/gid */
6677 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6678 key = (struct dt_key *)&oti->oti_quota_id;
6681 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6687 * \param di osd iterator
6689 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6691 struct osd_it_iam *it = (struct osd_it_iam *)di;
6693 iam_it_put(&it->oi_it);
6697 * Move iterator by one record
6699 * \param di osd iterator
6701 * \retval +1 end of container reached
6703 * \retval -ve failure
6706 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6708 struct osd_it_iam *it = (struct osd_it_iam *)di;
6710 return iam_it_next(&it->oi_it);
6714 * Return pointer to the key under iterator.
6717 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6718 const struct dt_it *di)
6720 struct osd_thread_info *oti = osd_oti_get(env);
6721 struct osd_it_iam *it = (struct osd_it_iam *)di;
6722 struct osd_object *obj = it->oi_obj;
6725 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6727 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6728 /* swab quota uid/gid */
6729 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6730 key = (struct dt_key *)&oti->oti_quota_id;
6737 * Return size of key under iterator (in bytes)
6740 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6742 struct osd_it_iam *it = (struct osd_it_iam *)di;
6744 return iam_it_key_size(&it->oi_it);
6748 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6750 /* check if file type is required */
6751 if (ent->lde_attrs & LUDA_TYPE) {
6752 struct luda_type *lt;
6753 int align = sizeof(*lt) - 1;
6755 len = (len + align) & ~align;
6756 lt = (struct luda_type *)(ent->lde_name + len);
6757 lt->lt_type = cpu_to_le16(DTTOIF(type));
6760 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6764 * build lu direct from backend fs dirent.
6768 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6769 char *name, __u16 namelen, __u16 type, __u32 attr)
6771 ent->lde_attrs = attr | LUDA_FID;
6772 fid_cpu_to_le(&ent->lde_fid, fid);
6774 ent->lde_hash = cpu_to_le64(offset);
6775 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6777 strncpy(ent->lde_name, name, namelen);
6778 ent->lde_name[namelen] = '\0';
6779 ent->lde_namelen = cpu_to_le16(namelen);
6781 /* append lustre attributes */
6782 osd_it_append_attrs(ent, namelen, type);
6786 * Return pointer to the record under iterator.
6788 static int osd_it_iam_rec(const struct lu_env *env,
6789 const struct dt_it *di,
6790 struct dt_rec *dtrec, __u32 attr)
6792 struct osd_it_iam *it = (struct osd_it_iam *)di;
6793 struct osd_thread_info *info = osd_oti_get(env);
6797 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6798 const struct osd_fid_pack *rec;
6799 struct lu_fid *fid = &info->oti_fid;
6800 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6806 name = (char *)iam_it_key_get(&it->oi_it);
6808 RETURN(PTR_ERR(name));
6810 namelen = iam_it_key_size(&it->oi_it);
6812 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6814 RETURN(PTR_ERR(rec));
6816 rc = osd_fid_unpack(fid, rec);
6820 hash = iam_it_store(&it->oi_it);
6822 /* IAM does not store object type in IAM index (dir) */
6823 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6825 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6826 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6827 (struct iam_rec *)dtrec);
6828 osd_quota_unpack(it->oi_obj, dtrec);
6830 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6831 (struct iam_rec *)dtrec);
6838 * Returns cookie for current Iterator position.
6840 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6842 struct osd_it_iam *it = (struct osd_it_iam *)di;
6844 return iam_it_store(&it->oi_it);
6848 * Restore iterator from cookie.
6850 * \param di osd iterator
6851 * \param hash Iterator location cookie
6853 * \retval +ve di points to record with least key not larger than key.
6854 * \retval 0 di points to exact matched key
6855 * \retval -ve failure
6858 static int osd_it_iam_load(const struct lu_env *env,
6859 const struct dt_it *di, __u64 hash)
6861 struct osd_it_iam *it = (struct osd_it_iam *)di;
6863 return iam_it_load(&it->oi_it, hash);
6866 static const struct dt_index_operations osd_index_iam_ops = {
6867 .dio_lookup = osd_index_iam_lookup,
6868 .dio_declare_insert = osd_index_declare_iam_insert,
6869 .dio_insert = osd_index_iam_insert,
6870 .dio_declare_delete = osd_index_declare_iam_delete,
6871 .dio_delete = osd_index_iam_delete,
6873 .init = osd_it_iam_init,
6874 .fini = osd_it_iam_fini,
6875 .get = osd_it_iam_get,
6876 .put = osd_it_iam_put,
6877 .next = osd_it_iam_next,
6878 .key = osd_it_iam_key,
6879 .key_size = osd_it_iam_key_size,
6880 .rec = osd_it_iam_rec,
6881 .store = osd_it_iam_store,
6882 .load = osd_it_iam_load
6886 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
6887 struct osd_device *dev,
6888 struct inode *inode, u32 attr)
6890 struct osd_thread_info *info = osd_oti_get(env);
6891 struct osd_it_ea *oie;
6895 file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
6898 RETURN(ERR_CAST(file));
6900 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6901 if (attr & LUDA_64BITHASH)
6902 file->f_mode |= FMODE_64BITHASH;
6904 file->f_mode |= FMODE_32BITHASH;
6907 OBD_SLAB_ALLOC_PTR(oie, osd_itea_cachep);
6911 oie->oie_rd_dirent = 0;
6912 oie->oie_it_dirent = 0;
6913 oie->oie_dirent = NULL;
6914 if (unlikely(!info->oti_it_ea_buf_used)) {
6915 oie->oie_buf = info->oti_it_ea_buf;
6916 info->oti_it_ea_buf_used = 1;
6918 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6922 oie->oie_obj = NULL;
6923 oie->oie_file = file;
6928 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6932 return ERR_PTR(-ENOMEM);
6936 * Creates or initializes iterator context.
6938 * \retval struct osd_it_ea, iterator structure on success
6941 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6942 struct dt_object *dt,
6945 struct osd_object *obj = osd_dt_obj(dt);
6946 struct osd_device *dev = osd_obj2dev(obj);
6947 struct lu_object *lo = &dt->do_lu;
6948 struct osd_it_ea *oie;
6952 if (!dt_object_exists(dt) || obj->oo_destroyed)
6953 RETURN(ERR_PTR(-ENOENT));
6955 oie = osd_it_dir_init(env, dev, obj->oo_inode, attr);
6957 RETURN(ERR_CAST(oie));
6961 RETURN((struct dt_it *)oie);
6964 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
6965 struct inode *inode)
6967 struct osd_thread_info *info = osd_oti_get(env);
6970 fput(oie->oie_file);
6971 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6972 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6974 info->oti_it_ea_buf_used = 0;
6975 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6980 * Destroy or finishes iterator context.
6982 * \param di iterator structure to be destroyed
6984 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6986 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6987 struct osd_object *obj = oie->oie_obj;
6988 struct inode *inode = obj->oo_inode;
6991 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
6992 osd_object_put(env, obj);
6997 * It position the iterator at given key, so that next lookup continues from
6998 * that key Or it is similar to dio_it->load() but based on a key,
6999 * rather than file position.
7001 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
7004 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
7006 static int osd_it_ea_get(const struct lu_env *env,
7007 struct dt_it *di, const struct dt_key *key)
7009 struct osd_it_ea *it = (struct osd_it_ea *)di;
7012 LASSERT(((const char *)key)[0] == '\0');
7013 it->oie_file->f_pos = 0;
7014 it->oie_rd_dirent = 0;
7015 it->oie_it_dirent = 0;
7016 it->oie_dirent = NULL;
7024 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
7028 struct osd_filldir_cbs {
7029 struct dir_context ctx;
7030 struct osd_it_ea *it;
7033 * It is called internally by ->iterate*(). It fills the
7034 * iterator's in-memory data structure with required
7035 * information i.e. name, namelen, rec_size etc.
7037 * \param buf in which information to be filled in.
7038 * \param name name of the file in given dir
7040 * \retval 0 on success
7041 * \retval 1 on buffer full
7043 #ifdef HAVE_FILLDIR_USE_CTX
7044 static FILLDIR_TYPE do_osd_ldiskfs_filldir(struct dir_context *ctx,
7046 static int osd_ldiskfs_filldir(void *ctx,
7048 const char *name, int namelen,
7049 loff_t offset, __u64 ino, unsigned int d_type)
7051 struct osd_it_ea *it = ((struct osd_filldir_cbs *)ctx)->it;
7052 struct osd_object *obj = it->oie_obj;
7053 struct osd_it_ea_dirent *ent = it->oie_dirent;
7054 struct lu_fid *fid = &ent->oied_fid;
7055 char *buf = it->oie_buf;
7056 struct osd_fid_pack *rec;
7059 /* this should never happen */
7060 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
7061 CERROR("ldiskfs return invalid namelen %d\n", namelen);
7065 /* Check for enough space. Note oied_name is not NUL terminated. */
7066 if (&ent->oied_name[namelen] > buf + OSD_IT_EA_BUFSIZE)
7069 /* "." is just the object itself. */
7070 if (namelen == 1 && name[0] == '.') {
7072 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7073 } else if (d_type & LDISKFS_DIRENT_LUFID) {
7074 rec = (struct osd_fid_pack *)(name + namelen + 1);
7075 if (osd_fid_unpack(fid, rec) != 0)
7080 d_type &= ~LDISKFS_DIRENT_LUFID;
7082 /* NOT export local root. */
7084 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
7085 ino = obj->oo_inode->i_ino;
7086 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
7089 if (obj == NULL || !(obj->oo_lma_flags & LUSTRE_ENCRYPT_FL)) {
7090 ent->oied_namelen = namelen;
7091 memcpy(ent->oied_name, name, namelen);
7093 int encoded_namelen = critical_chars(name, namelen);
7095 /* Check again for enough space. */
7096 if (&ent->oied_name[encoded_namelen] > buf + OSD_IT_EA_BUFSIZE)
7099 ent->oied_namelen = encoded_namelen;
7101 if (encoded_namelen == namelen)
7102 memcpy(ent->oied_name, name, namelen);
7104 critical_encode(name, namelen, ent->oied_name);
7107 ent->oied_ino = ino;
7108 ent->oied_off = offset;
7109 ent->oied_type = d_type;
7111 it->oie_rd_dirent++;
7112 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + ent->oied_namelen);
7116 WRAP_FILLDIR_FN(do_, osd_ldiskfs_filldir)
7119 * Calls ->iterate*() to load a directory entry at a time
7120 * and stored it in iterator's in-memory data structure.
7122 * \param di iterator's in memory structure
7124 * \retval 0 on success
7125 * \retval -ve on error
7126 * \retval +1 reach the end of entry
7128 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
7130 struct osd_it_ea *it = (struct osd_it_ea *)di;
7131 struct osd_object *obj = it->oie_obj;
7132 struct htree_lock *hlock = NULL;
7133 struct file *filp = it->oie_file;
7135 struct osd_filldir_cbs buf = {
7136 .ctx.actor = osd_ldiskfs_filldir,
7141 it->oie_dirent = it->oie_buf;
7142 it->oie_rd_dirent = 0;
7145 if (obj->oo_hl_head != NULL) {
7146 hlock = osd_oti_get(env)->oti_hlock;
7147 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
7149 LDISKFS_HLOCK_READDIR);
7151 down_read(&obj->oo_ext_idx_sem);
7155 rc = iterate_dir(filp, &buf.ctx);
7159 if (it->oie_rd_dirent == 0) {
7161 * If it does not get any dirent, it means it has been reached
7162 * to the end of the dir
7164 it->oie_file->f_pos = ldiskfs_get_htree_eof(it->oie_file);
7168 it->oie_dirent = it->oie_buf;
7169 it->oie_it_dirent = 1;
7174 ldiskfs_htree_unlock(hlock);
7176 up_read(&obj->oo_ext_idx_sem);
7183 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7184 * to load a directory entry at a time and stored it in
7185 * iterator's in-memory data structure.
7187 * \param di iterator's in memory structure
7189 * \retval +ve iterator reached to end
7190 * \retval 0 iterator not reached to end
7191 * \retval -ve on error
7193 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
7195 struct osd_it_ea *it = (struct osd_it_ea *)di;
7200 if (it->oie_it_dirent < it->oie_rd_dirent) {
7202 (void *)it->oie_dirent +
7203 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
7204 it->oie_dirent->oied_namelen);
7205 it->oie_it_dirent++;
7208 if (it->oie_file->f_pos == ldiskfs_get_htree_eof(it->oie_file))
7211 rc = osd_ldiskfs_it_fill(env, di);
7218 * Returns the key at current position from iterator's in memory structure.
7220 * \param di iterator's in memory structure
7222 * \retval key i.e. struct dt_key on success
7224 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
7225 const struct dt_it *di)
7227 struct osd_it_ea *it = (struct osd_it_ea *)di;
7229 return (struct dt_key *)it->oie_dirent->oied_name;
7233 * Returns key's size at current position from iterator's in memory structure.
7235 * \param di iterator's in memory structure
7237 * \retval key_size i.e. struct dt_key on success
7239 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
7241 struct osd_it_ea *it = (struct osd_it_ea *)di;
7243 return it->oie_dirent->oied_namelen;
7246 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
7247 #undef LDISKFS_DIR_REC_LEN
7248 # if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7249 # define LDISKFS_DIR_REC_LEN(de, dir) LDISKFS_DIR_ENTRY_LEN_((de), (dir))
7251 # define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
7255 #if defined LDISKFS_DIR_REC_LEN_WITH_DIR
7256 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de), NULL)
7258 # define LDISKFS_DIR_REC_LEN_DIR(de) LDISKFS_DIR_REC_LEN((de))
7261 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
7263 if (LDISKFS_DIR_REC_LEN_DIR(de) >=
7264 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
7271 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
7272 unsigned int blocksize, bool dotdot)
7275 return osd_dotdot_has_space(de);
7277 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
7278 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
7285 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
7286 handle_t *jh, struct dentry *dentry,
7287 const struct lu_fid *fid, struct buffer_head *bh,
7288 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
7291 struct inode *dir = dentry->d_parent->d_inode;
7292 struct inode *inode = dentry->d_inode;
7293 struct osd_fid_pack *rec;
7294 struct ldiskfs_dentry_param *ldp;
7295 int namelen = dentry->d_name.len;
7297 struct osd_thread_info *info = osd_oti_get(env);
7301 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7304 /* There is enough space to hold the FID-in-dirent. */
7305 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7306 rc = osd_ldiskfs_journal_get_write_access(jh, dir->i_sb, bh,
7311 de->name[namelen] = 0;
7312 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7313 rec->fp_len = sizeof(struct lu_fid) + 1;
7314 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7315 de->file_type |= LDISKFS_DIRENT_LUFID;
7316 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7323 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7327 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7328 osd_get_ldiskfs_dirent_param(ldp, fid);
7329 dentry->d_fsdata = (void *)ldp;
7330 dquot_initialize(dir);
7331 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7333 * It is too bad, we cannot reinsert the name entry back.
7334 * That means we lose it!
7338 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7339 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7340 namelen, dentry->d_name.name, PFID(fid), rc);
7346 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7347 struct osd_it_ea *it, struct lu_fid *fid,
7348 struct osd_inode_id *id, __u32 *attr)
7350 struct osd_thread_info *info = osd_oti_get(env);
7351 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7352 struct osd_device *dev = osd_obj2dev(obj);
7353 struct super_block *sb = osd_sb(dev);
7354 const char *devname = osd_name(dev);
7355 struct osd_it_ea_dirent *ent = it->oie_dirent;
7356 struct inode *dir = obj->oo_inode;
7357 struct htree_lock *hlock = NULL;
7358 struct buffer_head *bh = NULL;
7359 handle_t *jh = NULL;
7360 struct ldiskfs_dir_entry_2 *de;
7361 struct dentry *dentry;
7362 struct inode *inode;
7363 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7366 bool dotdot = false;
7372 if (ent->oied_name[0] == '.') {
7373 if (ent->oied_namelen == 1)
7376 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7380 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7381 inode = osd_iget(info, dev, id);
7382 if (IS_ERR(inode)) {
7383 rc = PTR_ERR(inode);
7384 if (rc == -ENOENT || rc == -ESTALE) {
7386 * Maybe dangling name entry, or
7387 * corrupted directory entry.
7389 *attr |= LUDA_UNKNOWN;
7392 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7393 "check_repair, dir = %lu/%u, name = %.*s, "
7394 "ino = %llu, rc = %d\n",
7395 devname, dir->i_ino, dir->i_generation,
7396 ent->oied_namelen, ent->oied_name,
7403 rc = obj_name2lu_name(obj, ent->oied_name, ent->oied_namelen, &ln);
7407 dentry = osd_child_dentry_by_inode(env, dir, ln.ln_name, ln.ln_namelen);
7408 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7409 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7415 * We need to ensure that the name entry is still valid.
7416 * Because it may be removed or renamed by other already.
7418 * The unlink or rename operation will start journal before PDO lock,
7419 * so to avoid deadlock, here we need to start journal handle before
7420 * related PDO lock also. But because we do not know whether there
7421 * will be something to be repaired before PDO lock, we just start
7422 * journal without conditions.
7424 * We may need to remove the name entry firstly, then insert back.
7425 * One credit is for user quota file update.
7426 * One credit is for group quota file update.
7427 * Two credits are for dirty inode.
7429 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7430 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7432 if (dev->od_dirent_journal != 0) {
7435 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7438 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7439 "check_repair, dir = %lu/%u, credits = %d, "
7440 "name = %.*s, ino = %llu: rc = %d\n",
7441 devname, dir->i_ino, dir->i_generation, credits,
7442 ent->oied_namelen, ent->oied_name,
7445 GOTO(out_inode, rc);
7448 if (obj->oo_hl_head != NULL) {
7449 hlock = osd_oti_get(env)->oti_hlock;
7451 * "0" means exclusive lock for the whole directory.
7452 * We need to prevent others access such name entry
7453 * during the delete + insert. Neither HLOCK_ADD nor
7454 * HLOCK_DEL cannot guarantee the atomicity.
7456 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7458 down_write(&obj->oo_ext_idx_sem);
7461 if (obj->oo_hl_head != NULL) {
7462 hlock = osd_oti_get(env)->oti_hlock;
7463 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7464 LDISKFS_HLOCK_LOOKUP);
7466 down_read(&obj->oo_ext_idx_sem);
7470 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7471 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7472 *attr |= LUDA_IGNORE;
7478 * For dotdot entry, if there is not enough space to hold the
7479 * FID-in-dirent, just keep them there. It only happens when the
7480 * device upgraded from 1.8 or restored from MDT file-level backup.
7481 * For the whole directory, only dotdot entry have no FID-in-dirent
7482 * and needs to get FID from LMA when readdir, it will not affect the
7485 if (dotdot && !osd_dotdot_has_space(de)) {
7486 *attr |= LUDA_UNKNOWN;
7492 if (lu_fid_eq(fid, &lma->lma_self_fid))
7495 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7496 struct lu_fid *tfid = &lma->lma_self_fid;
7498 if (likely(dotdot &&
7499 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7500 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7502 * It must be REMOTE_PARENT_DIR and as the
7503 * 'dotdot' entry of remote directory
7505 *attr |= LUDA_IGNORE;
7507 CDEBUG(D_LFSCK, "%s: expect remote agent "
7508 "parent directory, but got %.*s under "
7509 "dir = %lu/%u with the FID "DFID"\n",
7510 devname, ent->oied_namelen,
7511 ent->oied_name, dir->i_ino,
7512 dir->i_generation, PFID(tfid));
7514 *attr |= LUDA_UNKNOWN;
7521 if (!fid_is_zero(fid)) {
7522 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7524 if (rc == -ENOENT ||
7526 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7528 * linkEA does not recognize the dirent entry,
7529 * it may because the dirent entry corruption
7530 * and points to other's inode.
7532 CDEBUG(D_LFSCK, "%s: the target inode does not "
7533 "recognize the dirent, dir = %lu/%u, "
7534 " name = %.*s, ino = %llu, "
7535 DFID": rc = %d\n", devname, dir->i_ino,
7536 dir->i_generation, ent->oied_namelen,
7537 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7538 *attr |= LUDA_UNKNOWN;
7543 if (rc && rc != -ENODATA) {
7544 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7545 "dir = %lu/%u, name = %.*s, ino = %llu, "
7546 DFID": rc = %d\n", devname, dir->i_ino,
7547 dir->i_generation, ent->oied_namelen,
7548 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7549 *attr |= LUDA_UNKNOWN;
7557 * linkEA recognizes the dirent entry, the FID-in-LMA is
7558 * valid, trusted, in spite of fid_is_sane(fid) or not.
7560 if (*attr & LUDA_VERIFY_DRYRUN) {
7561 *fid = lma->lma_self_fid;
7562 *attr |= LUDA_REPAIR;
7569 dev->od_dirent_journal = 1;
7570 if (hlock != NULL) {
7571 ldiskfs_htree_unlock(hlock);
7574 up_read(&obj->oo_ext_idx_sem);
7580 *fid = lma->lma_self_fid;
7582 /* Update or append the FID-in-dirent. */
7583 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7584 bh, de, hlock, dotdot);
7586 *attr |= LUDA_REPAIR;
7588 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7589 "the dirent, dir = %lu/%u, name = %.*s, "
7590 "ino = %llu, "DFID": rc = %d\n",
7591 devname, dir->i_ino, dir->i_generation,
7592 ent->oied_namelen, ent->oied_name,
7593 ent->oied_ino, PFID(fid), rc);
7595 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7596 if (*attr & LUDA_VERIFY_DRYRUN) {
7597 if (fid_is_sane(fid)) {
7598 *attr |= LUDA_REPAIR;
7599 } else if (dev->od_index == 0) {
7600 lu_igif_build(fid, inode->i_ino,
7601 inode->i_generation);
7602 *attr |= LUDA_UPGRADE;
7610 dev->od_dirent_journal = 1;
7611 if (hlock != NULL) {
7612 ldiskfs_htree_unlock(hlock);
7615 up_read(&obj->oo_ext_idx_sem);
7622 if (unlikely(fid_is_sane(fid))) {
7624 * FID-in-dirent exists, but FID-in-LMA is lost.
7625 * Trust the FID-in-dirent, and add FID-in-LMA.
7627 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7629 *attr |= LUDA_REPAIR;
7631 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7632 "update dirent, dir = %lu/%u, "
7633 "name = %.*s, ino = %llu, "
7635 devname, dir->i_ino, dir->i_generation,
7636 ent->oied_namelen, ent->oied_name,
7637 ent->oied_ino, PFID(fid), rc);
7638 } else if (dev->od_index == 0) {
7639 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7641 * It is probably IGIF object. Only aappend the
7642 * FID-in-dirent. OI scrub will process FID-in-LMA.
7644 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7645 bh, de, hlock, dotdot);
7647 *attr |= LUDA_UPGRADE;
7649 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7650 "after the dirent, dir = %lu/%u, "
7651 "name = %.*s, ino = %llu, "
7653 devname, dir->i_ino, dir->i_generation,
7654 ent->oied_namelen, ent->oied_name,
7655 ent->oied_ino, PFID(fid), rc);
7664 if (hlock != NULL) {
7665 ldiskfs_htree_unlock(hlock);
7667 if (dev->od_dirent_journal != 0)
7668 up_write(&obj->oo_ext_idx_sem);
7670 up_read(&obj->oo_ext_idx_sem);
7674 ldiskfs_journal_stop(jh);
7678 if (rc >= 0 && !dirty)
7679 dev->od_dirent_journal = 0;
7680 if (ln.ln_name != ent->oied_name)
7687 * Returns the value at current position from iterator's in memory structure.
7689 * \param di struct osd_it_ea, iterator's in memory structure
7690 * \param attr attr requested for dirent.
7691 * \param lde lustre dirent
7693 * \retval 0 no error and \param lde has correct lustre dirent.
7694 * \retval -ve on error
7696 static inline int osd_it_ea_rec(const struct lu_env *env,
7697 const struct dt_it *di,
7698 struct dt_rec *dtrec, __u32 attr)
7700 struct osd_it_ea *it = (struct osd_it_ea *)di;
7701 struct osd_object *obj = it->oie_obj;
7702 struct osd_device *dev = osd_obj2dev(obj);
7703 struct osd_thread_info *oti = osd_oti_get(env);
7704 struct osd_inode_id *id = &oti->oti_id;
7705 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7706 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7707 __u32 ino = it->oie_dirent->oied_ino;
7712 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7714 if (attr & LUDA_VERIFY) {
7715 if (unlikely(is_remote_parent_ino(dev, ino))) {
7716 attr |= LUDA_IGNORE;
7718 * If the parent is on remote MDT, and there
7719 * is no FID-in-dirent, then we have to get
7720 * the parent FID from the linkEA.
7722 if (!fid_is_sane(fid) &&
7723 it->oie_dirent->oied_namelen == 2 &&
7724 it->oie_dirent->oied_name[0] == '.' &&
7725 it->oie_dirent->oied_name[1] == '.')
7726 osd_get_pfid_from_linkea(env, obj, fid);
7728 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7732 if (!fid_is_sane(fid))
7733 attr |= LUDA_UNKNOWN;
7735 attr &= ~LU_DIRENT_ATTRS_MASK;
7736 if (!fid_is_sane(fid)) {
7737 bool is_dotdot = false;
7739 if (it->oie_dirent->oied_namelen == 2 &&
7740 it->oie_dirent->oied_name[0] == '.' &&
7741 it->oie_dirent->oied_name[1] == '.')
7744 * If the parent is on remote MDT, and there
7745 * is no FID-in-dirent, then we have to get
7746 * the parent FID from the linkEA.
7748 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7749 rc = osd_get_pfid_from_linkea(env, obj, fid);
7751 if (is_dotdot == false &&
7752 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7755 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7760 /* Pack the entry anyway, at least the offset is right. */
7761 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7762 it->oie_dirent->oied_name,
7763 it->oie_dirent->oied_namelen,
7764 it->oie_dirent->oied_type, attr);
7766 RETURN(rc > 0 ? 0 : rc);
7770 * Returns the record size size at current position.
7772 * This function will return record(lu_dirent) size in bytes.
7774 * \param[in] env execution environment
7775 * \param[in] di iterator's in memory structure
7776 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7777 * calculate the lu_dirent size.
7779 * \retval record size(in bytes & in memory) of the current lu_dirent
7782 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7785 struct osd_it_ea *it = (struct osd_it_ea *)di;
7787 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7791 * Returns a cookie for current position of the iterator head, so that
7792 * user can use this cookie to load/start the iterator next time.
7794 * \param di iterator's in memory structure
7796 * \retval cookie for current position, on success
7798 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7800 struct osd_it_ea *it = (struct osd_it_ea *)di;
7802 return it->oie_dirent->oied_off;
7806 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7807 * to load a directory entry at a time and stored it i inn,
7808 * in iterator's in-memory data structure.
7810 * \param di struct osd_it_ea, iterator's in memory structure
7812 * \retval +ve on success
7813 * \retval -ve on error
7815 static int osd_it_ea_load(const struct lu_env *env,
7816 const struct dt_it *di, __u64 hash)
7818 struct osd_it_ea *it = (struct osd_it_ea *)di;
7822 it->oie_file->f_pos = hash;
7824 rc = osd_ldiskfs_it_fill(env, di);
7835 * Index lookup function for interoperability mode (b11826).
7837 * \param key, key i.e. file name to be searched
7839 * \retval +ve, on success
7840 * \retval -ve, on error
7842 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7843 struct dt_rec *rec, const struct dt_key *key)
7845 struct osd_object *obj = osd_dt_obj(dt);
7850 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7851 LINVRNT(osd_invariant(obj));
7853 rc = osd_ea_lookup_rec(env, obj, rec, key);
7860 * Index and Iterator operations for interoperability
7861 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7863 static const struct dt_index_operations osd_index_ea_ops = {
7864 .dio_lookup = osd_index_ea_lookup,
7865 .dio_declare_insert = osd_index_declare_ea_insert,
7866 .dio_insert = osd_index_ea_insert,
7867 .dio_declare_delete = osd_index_declare_ea_delete,
7868 .dio_delete = osd_index_ea_delete,
7870 .init = osd_it_ea_init,
7871 .fini = osd_it_ea_fini,
7872 .get = osd_it_ea_get,
7873 .put = osd_it_ea_put,
7874 .next = osd_it_ea_next,
7875 .key = osd_it_ea_key,
7876 .key_size = osd_it_ea_key_size,
7877 .rec = osd_it_ea_rec,
7878 .rec_size = osd_it_ea_rec_size,
7879 .store = osd_it_ea_store,
7880 .load = osd_it_ea_load
7884 static void *osd_key_init(const struct lu_context *ctx,
7885 struct lu_context_key *key)
7887 struct osd_thread_info *info;
7889 OBD_ALLOC_PTR(info);
7891 return ERR_PTR(-ENOMEM);
7893 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7894 if (info->oti_it_ea_buf == NULL)
7897 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7899 info->oti_hlock = ldiskfs_htree_lock_alloc();
7900 if (info->oti_hlock == NULL)
7906 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7909 return ERR_PTR(-ENOMEM);
7912 static void osd_key_fini(const struct lu_context *ctx,
7913 struct lu_context_key *key, void *data)
7915 struct osd_thread_info *info = data;
7916 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7917 struct osd_idmap_cache *idc = info->oti_ins_cache;
7919 if (info->oti_dio_pages) {
7921 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7922 struct page *page = info->oti_dio_pages[i];
7924 LASSERT(PagePrivate2(page));
7925 LASSERT(PageLocked(page));
7926 ClearPagePrivate2(page);
7931 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7932 PTLRPC_MAX_BRW_PAGES);
7935 if (info->oti_inode != NULL)
7937 if (info->oti_hlock != NULL)
7938 ldiskfs_htree_lock_free(info->oti_hlock);
7939 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7940 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7941 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7942 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7943 lu_buf_free(&info->oti_big_buf);
7945 LASSERT(info->oti_ins_cache_size > 0);
7946 OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size);
7947 info->oti_ins_cache = NULL;
7948 info->oti_ins_cache_size = 0;
7953 static void osd_key_exit(const struct lu_context *ctx,
7954 struct lu_context_key *key, void *data)
7956 struct osd_thread_info *info = data;
7958 LASSERT(info->oti_r_locks == 0);
7959 LASSERT(info->oti_w_locks == 0);
7960 LASSERT(info->oti_txns == 0);
7961 LASSERTF(info->oti_dio_pages_used == 0, "%d\n",
7962 info->oti_dio_pages_used);
7965 /* type constructor/destructor: osd_type_init, osd_type_fini */
7966 LU_TYPE_INIT_FINI(osd, &osd_key);
7968 struct lu_context_key osd_key = {
7969 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7970 .lct_init = osd_key_init,
7971 .lct_fini = osd_key_fini,
7972 .lct_exit = osd_key_exit
7976 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7977 const char *name, struct lu_device *next)
7979 struct osd_device *osd = osd_dev(d);
7981 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7982 sizeof(osd->od_svname))
7984 return osd_procfs_init(osd, name);
7987 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7989 struct seq_server_site *ss = osd_seq_site(osd);
7994 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7997 if (unlikely(ss == NULL))
8000 OBD_ALLOC_PTR(osd->od_cl_seq);
8001 if (osd->od_cl_seq == NULL)
8004 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
8005 osd->od_svname, ss->ss_server_seq);
8007 if (ss->ss_node_id == 0) {
8009 * If the OSD on the sequence controller(MDT0), then allocate
8010 * sequence here, otherwise allocate sequence after connected
8011 * to MDT0 (see mdt_register_lwp_callback()).
8013 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
8014 &osd->od_cl_seq->lcs_space, env);
8020 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
8022 if (osd->od_cl_seq == NULL)
8025 seq_client_fini(osd->od_cl_seq);
8026 OBD_FREE_PTR(osd->od_cl_seq);
8027 osd->od_cl_seq = NULL;
8030 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
8034 /* shutdown quota slave instance associated with the device */
8035 if (o->od_quota_slave_md != NULL) {
8036 struct qsd_instance *qsd = o->od_quota_slave_md;
8038 o->od_quota_slave_md = NULL;
8042 if (o->od_quota_slave_dt != NULL) {
8043 struct qsd_instance *qsd = o->od_quota_slave_dt;
8045 o->od_quota_slave_dt = NULL;
8049 osd_fid_fini(env, o);
8050 osd_scrub_cleanup(env, o);
8055 static void osd_umount(const struct lu_env *env, struct osd_device *o)
8059 if (o->od_mnt != NULL) {
8060 shrink_dcache_sb(osd_sb(o));
8061 osd_sync(env, &o->od_dt_dev);
8062 wait_event(o->od_commit_cb_done,
8063 !atomic_read(&o->od_commit_cb_in_flight));
8072 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8073 # ifndef LDISKFS_HAS_INCOMPAT_FEATURE
8074 /* Newer kernels provide the ldiskfs_set_feature_largedir() wrapper already,
8075 * which calls ldiskfs_update_dynamic_rev() to update ancient filesystems.
8076 * All ldiskfs filesystems are already v2, so it is a no-op and unnecessary.
8077 * This avoids maintaining patches to export this otherwise-useless function.
8079 void ldiskfs_update_dynamic_rev(struct super_block *sb)
8086 static int osd_mount(const struct lu_env *env,
8087 struct osd_device *o, struct lustre_cfg *cfg)
8089 const char *name = lustre_cfg_string(cfg, 0);
8090 const char *dev = lustre_cfg_string(cfg, 1);
8092 unsigned long page, s_flags = 0, lmd_flags = 0;
8093 struct page *__page;
8094 struct file_system_type *type;
8095 char *options = NULL;
8097 struct osd_thread_info *info = osd_oti_get(env);
8098 struct lu_fid *fid = &info->oti_fid;
8099 struct inode *inode;
8100 int rc = 0, force_over_1024tb = 0;
8104 if (o->od_mnt != NULL)
8107 if (strlen(dev) >= sizeof(o->od_mntdev))
8109 strcpy(o->od_mntdev, dev);
8111 str = lustre_cfg_buf(cfg, 2);
8112 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
8114 opts = lustre_cfg_string(cfg, 3);
8116 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
8117 CERROR("%s: device %s extents feature is not guaranteed to "
8118 "work on big-endian systems. Use \"bigendian_extents\" "
8119 "mount option to override.\n", name, dev);
8123 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
8124 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
8125 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");
8128 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
8129 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
8130 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");
8133 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8134 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
8135 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");
8139 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
8140 force_over_1024tb = 1;
8142 __page = alloc_page(GFP_KERNEL);
8144 GOTO(out, rc = -ENOMEM);
8145 page = (unsigned long)page_address(__page);
8146 options = (char *)page;
8149 /* strip out the options for back compatiblity */
8150 static const char * const sout[] = {
8157 /* strip out option we processed in osd */
8158 "bigendian_extents",
8162 "force_over_1024tb",
8166 strncat(options, opts, PAGE_SIZE);
8167 for (rc = 0, str = options; sout[rc]; ) {
8168 char *op = strstr(str, sout[rc]);
8175 if (op == options || *(op - 1) == ',') {
8176 str = op + strlen(sout[rc]);
8177 if (*str == ',' || *str == '\0') {
8178 *str == ',' ? str++ : str;
8179 memmove(op, str, strlen(str) + 1);
8182 for (str = op; *str != ',' && *str != '\0'; str++)
8186 strncat(options, "user_xattr,acl", PAGE_SIZE);
8189 /* Glom up mount options */
8190 if (*options != '\0')
8191 strncat(options, ",", PAGE_SIZE);
8192 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
8194 type = get_fs_type("ldiskfs");
8196 CERROR("%s: cannot find ldiskfs module\n", name);
8197 GOTO(out, rc = -ENODEV);
8200 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
8201 module_put(type->owner);
8203 if (IS_ERR(o->od_mnt)) {
8204 rc = PTR_ERR(o->od_mnt);
8206 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
8210 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
8211 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
8212 force_over_1024tb == 0) {
8213 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",
8215 GOTO(out_mnt, rc = -EINVAL);
8218 if (test_bit(LMD_FLG_DEV_RDONLY, &lmd_flags)) {
8219 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
8222 GOTO(out_mnt, rc = -EOPNOTSUPP);
8225 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
8226 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
8227 GOTO(out_mnt, rc = -EINVAL);
8230 if (ldiskfs_has_feature_fast_commit(o->od_mnt->mnt_sb)) {
8231 CERROR("%s: device %s is mounted with fast_commit that breaks recovery\n",
8233 GOTO(out_mnt, rc = -EOPNOTSUPP);
8236 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
8237 #ifdef LDISKFS_MOUNT_DIRDATA
8238 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
8239 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
8240 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
8241 CWARN("%s: device %s was upgraded from Lustre-1.x without "
8242 "enabling the dirdata feature. If you do not want to "
8243 "downgrade to Lustre-1.x again, you can enable it via "
8244 "'tune2fs -O dirdata device'\n", name, dev);
8246 /* enable large_dir on MDTs to avoid REMOTE_PARENT_DIR overflow,
8247 * and on very large OSTs to avoid object directory overflow */
8248 if (unlikely(!ldiskfs_has_feature_largedir(o->od_mnt->mnt_sb) &&
8249 !strstr(name, "MGS"))) {
8250 ldiskfs_set_feature_largedir(o->od_mnt->mnt_sb);
8251 LCONSOLE_INFO("%s: enabled 'large_dir' feature on device %s\n",
8255 inode = osd_sb(o)->s_root->d_inode;
8256 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
8257 if (!o->od_dt_dev.dd_rdonly) {
8258 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
8260 CERROR("%s: failed to set lma on %s root inode\n",
8266 if (test_bit(LMD_FLG_NOSCRUB, &lmd_flags))
8267 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_NEVER;
8269 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
8270 /* do not use pagecache with flash-backed storage */
8271 o->od_writethrough_cache = 0;
8272 o->od_read_cache = 0;
8283 __free_page(__page);
8288 static struct lu_device *osd_device_fini(const struct lu_env *env,
8289 struct lu_device *d)
8291 struct osd_device *o = osd_dev(d);
8295 osd_index_backup(env, o, false);
8296 osd_shutdown(env, o);
8298 if (o->od_oi_table != NULL)
8299 osd_oi_fini(osd_oti_get(env), o);
8300 if (o->od_extent_bytes_percpu)
8301 free_percpu(o->od_extent_bytes_percpu);
8302 osd_obj_map_fini(o);
8308 static int osd_device_init0(const struct lu_env *env,
8309 struct osd_device *o,
8310 struct lustre_cfg *cfg)
8312 struct lu_device *l = osd2lu_dev(o);
8313 struct osd_thread_info *info;
8316 bool restored = false;
8319 /* if the module was re-loaded, env can loose its keys */
8320 rc = lu_env_refill((struct lu_env *)env);
8323 info = osd_oti_get(env);
8326 l->ld_ops = &osd_lu_ops;
8327 o->od_dt_dev.dd_ops = &osd_dt_ops;
8329 spin_lock_init(&o->od_osfs_lock);
8330 mutex_init(&o->od_otable_mutex);
8331 INIT_LIST_HEAD(&o->od_orphan_list);
8332 INIT_LIST_HEAD(&o->od_index_backup_list);
8333 INIT_LIST_HEAD(&o->od_index_restore_list);
8334 spin_lock_init(&o->od_lock);
8335 o->od_index_backup_policy = LIBP_NONE;
8337 init_waitqueue_head(&o->od_commit_cb_done);
8339 o->od_read_cache = 1;
8340 o->od_writethrough_cache = 1;
8341 o->od_enable_projid_xattr = 0;
8342 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
8343 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
8344 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
8345 o->od_scrub.os_scrub.os_auto_scrub_interval = AS_DEFAULT;
8346 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8347 o->od_fallocate_zero_blocks = 0;
8349 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
8350 sizeof(o->od_svname));
8351 if (cplen >= sizeof(o->od_svname)) {
8356 o->od_index_backup_stop = 0;
8357 o->od_index = -1; /* -1 means index is invalid */
8358 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8359 if (rc == LDD_F_SV_TYPE_OST)
8362 o->od_full_scrub_ratio = OFSR_DEFAULT;
8363 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8364 rc = osd_mount(env, o, cfg);
8368 /* Can only check block device after mount */
8369 o->od_nonrotational =
8370 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8372 rc = osd_obj_map_init(env, o);
8376 rc = lu_site_init(&o->od_site, l);
8378 GOTO(out_compat, rc);
8379 o->od_site.ls_bottom_dev = l;
8381 rc = lu_site_init_finish(&o->od_site);
8385 opts = lustre_cfg_string(cfg, 3);
8386 if (opts && strstr(opts, "resetoi"))
8389 INIT_LIST_HEAD(&o->od_ios_list);
8391 rc = lprocfs_init_brw_stats(&o->od_brw_stats);
8393 GOTO(out_brw_stats, rc);
8395 /* setup scrub, including OI files initialization */
8397 rc = osd_scrub_setup(env, o, restored);
8400 GOTO(out_brw_stats, rc);
8402 rc = osd_procfs_init(o, o->od_svname);
8404 CERROR("%s: can't initialize procfs: rc = %d\n",
8406 GOTO(out_scrub, rc);
8409 LASSERT(l->ld_site->ls_linkage.next != NULL);
8410 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8412 /* initialize quota slave instance */
8413 /* currently it's no need to prepare qsd_instance_md for OST */
8414 if (!o->od_is_ost) {
8415 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8416 &o->od_dt_dev, o->od_proc_entry,
8418 if (IS_ERR(o->od_quota_slave_md)) {
8419 rc = PTR_ERR(o->od_quota_slave_md);
8420 o->od_quota_slave_md = NULL;
8421 GOTO(out_procfs, rc);
8425 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8426 o->od_proc_entry, false, true);
8428 if (IS_ERR(o->od_quota_slave_dt)) {
8429 if (o->od_quota_slave_md != NULL) {
8430 qsd_fini(env, o->od_quota_slave_md);
8431 o->od_quota_slave_md = NULL;
8434 rc = PTR_ERR(o->od_quota_slave_dt);
8435 o->od_quota_slave_dt = NULL;
8436 GOTO(out_procfs, rc);
8439 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8440 if (!o->od_extent_bytes_percpu) {
8442 GOTO(out_procfs, rc);
8450 osd_scrub_cleanup(env, o);
8452 lprocfs_fini_brw_stats(&o->od_brw_stats);
8454 lu_site_fini(&o->od_site);
8456 osd_obj_map_fini(o);
8463 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8464 struct lu_device_type *t,
8465 struct lustre_cfg *cfg)
8467 struct osd_device *o;
8472 return ERR_PTR(-ENOMEM);
8474 rc = dt_device_init(&o->od_dt_dev, t);
8477 * Because the ctx might be revived in dt_device_init,
8478 * refill the env here
8480 lu_env_refill((struct lu_env *)env);
8481 rc = osd_device_init0(env, o, cfg);
8483 dt_device_fini(&o->od_dt_dev);
8486 if (unlikely(rc != 0))
8489 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8492 static struct lu_device *osd_device_free(const struct lu_env *env,
8493 struct lu_device *d)
8495 struct osd_device *o = osd_dev(d);
8499 /* XXX: make osd top device in order to release reference */
8500 d->ld_site->ls_top_dev = d;
8501 lu_site_purge(env, d->ld_site, -1);
8502 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8503 D_ERROR, lu_cdebug_printer);
8504 lu_site_fini(&o->od_site);
8505 dt_device_fini(&o->od_dt_dev);
8510 static int osd_process_config(const struct lu_env *env,
8511 struct lu_device *d, struct lustre_cfg *cfg)
8513 struct osd_device *o = osd_dev(d);
8519 switch (cfg->lcfg_command) {
8521 rc = osd_mount(env, o, cfg);
8525 * For the case LCFG_PRE_CLEANUP is not called in advance,
8526 * that may happend if hit failure during mount process.
8528 osd_index_backup(env, o, false);
8529 lu_dev_del_linkage(d->ld_site, d);
8530 rc = osd_shutdown(env, o);
8533 LASSERT(&o->od_dt_dev);
8534 count = class_modify_config(cfg, PARAM_OSD,
8535 &o->od_dt_dev.dd_kobj);
8537 count = class_modify_config(cfg, PARAM_OST,
8538 &o->od_dt_dev.dd_kobj);
8539 rc = count > 0 ? 0 : count;
8541 case LCFG_PRE_CLEANUP:
8543 osd_index_backup(env, o,
8544 o->od_index_backup_policy != LIBP_NONE);
8554 static int osd_recovery_complete(const struct lu_env *env,
8555 struct lu_device *d)
8557 struct osd_device *osd = osd_dev(d);
8562 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8566 * start qsd instance on recovery completion, this notifies the quota
8567 * slave code that we are about to process new requests now
8569 rc = qsd_start(env, osd->od_quota_slave_dt);
8570 if (rc == 0 && osd->od_quota_slave_md != NULL)
8571 rc = qsd_start(env, osd->od_quota_slave_md);
8577 * we use exports to track all osd users
8579 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8580 struct obd_device *obd, struct obd_uuid *cluuid,
8581 struct obd_connect_data *data, void *localdata)
8583 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8584 struct lustre_handle conn;
8589 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8591 rc = class_connect(&conn, obd, cluuid);
8595 *exp = class_conn2export(&conn);
8597 spin_lock(&osd->od_osfs_lock);
8599 spin_unlock(&osd->od_osfs_lock);
8605 * once last export (we don't count self-export) disappeared
8606 * osd can be released
8608 static int osd_obd_disconnect(struct obd_export *exp)
8610 struct obd_device *obd = exp->exp_obd;
8611 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8612 int rc, release = 0;
8616 /* Only disconnect the underlying layers on the final disconnect. */
8617 spin_lock(&osd->od_osfs_lock);
8619 if (osd->od_connects == 0)
8621 spin_unlock(&osd->od_osfs_lock);
8623 rc = class_disconnect(exp); /* bz 9811 */
8625 if (rc == 0 && release)
8626 class_manual_cleanup(obd);
8630 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8631 struct lu_device *dev)
8633 struct osd_device *osd = osd_dev(dev);
8634 struct lr_server_data *lsd =
8635 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8640 if (osd->od_quota_slave_md != NULL) {
8641 /* set up quota slave objects for inode */
8642 result = qsd_prepare(env, osd->od_quota_slave_md);
8647 if (osd->od_quota_slave_dt != NULL) {
8648 /* set up quota slave objects for block */
8649 result = qsd_prepare(env, osd->od_quota_slave_dt);
8655 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8656 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8657 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8658 osd->od_index_in_idif = 1;
8660 osd->od_index_in_idif = 0;
8661 result = osd_register_proc_index_in_idif(osd);
8666 osd->od_index_in_idif = 1;
8670 result = osd_fid_init(env, osd);
8676 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8680 * see include/lu_object.h for the details.
8682 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8683 struct lu_fid *fid, struct lu_object *parent,
8684 const struct lu_name *name)
8686 struct osd_device *osd = osd_dev(d);
8688 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8691 static const struct lu_object_operations osd_lu_obj_ops = {
8692 .loo_object_init = osd_object_init,
8693 .loo_object_delete = osd_object_delete,
8694 .loo_object_release = osd_object_release,
8695 .loo_object_free = osd_object_free,
8696 .loo_object_print = osd_object_print,
8697 .loo_object_invariant = osd_object_invariant
8700 const struct lu_device_operations osd_lu_ops = {
8701 .ldo_object_alloc = osd_object_alloc,
8702 .ldo_process_config = osd_process_config,
8703 .ldo_recovery_complete = osd_recovery_complete,
8704 .ldo_prepare = osd_prepare,
8705 .ldo_fid_alloc = osd_fid_alloc,
8708 static const struct lu_device_type_operations osd_device_type_ops = {
8709 .ldto_init = osd_type_init,
8710 .ldto_fini = osd_type_fini,
8712 .ldto_start = osd_type_start,
8713 .ldto_stop = osd_type_stop,
8715 .ldto_device_alloc = osd_device_alloc,
8716 .ldto_device_free = osd_device_free,
8718 .ldto_device_init = osd_device_init,
8719 .ldto_device_fini = osd_device_fini
8722 static struct lu_device_type osd_device_type = {
8723 .ldt_tags = LU_DEVICE_DT,
8724 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8725 .ldt_ops = &osd_device_type_ops,
8726 .ldt_ctx_tags = LCT_LOCAL,
8729 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8731 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8732 struct super_block *sb = osd_sb(osd);
8734 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8738 * lprocfs legacy support.
8740 static const struct obd_ops osd_obd_device_ops = {
8741 .o_owner = THIS_MODULE,
8742 .o_connect = osd_obd_connect,
8743 .o_disconnect = osd_obd_disconnect,
8744 .o_health_check = osd_health_check,
8747 static ssize_t track_declares_assert_show(struct kobject *kobj,
8748 struct attribute *attr,
8751 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8754 static ssize_t track_declares_assert_store(struct kobject *kobj,
8755 struct attribute *attr,
8756 const char *buffer, size_t count)
8758 bool track_declares_assert;
8761 rc = kstrtobool(buffer, &track_declares_assert);
8765 ldiskfs_track_declares_assert = track_declares_assert;
8769 LUSTRE_RW_ATTR(track_declares_assert);
8771 static int __init osd_init(void)
8773 struct kobject *kobj;
8776 BUILD_BUG_ON(BH_DXLock >=
8777 sizeof(((struct buffer_head *)0)->b_state) * 8);
8778 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8779 /* please, try to keep osd_thread_info smaller than a page */
8780 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8785 rc = lu_kmem_init(ldiskfs_caches);
8789 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8790 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8792 lu_kmem_fini(ldiskfs_caches);
8796 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8798 rc = sysfs_create_file(kobj,
8799 &lustre_attr_track_declares_assert.attr);
8802 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8810 static void __exit osd_exit(void)
8812 struct kobject *kobj;
8814 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8816 sysfs_remove_file(kobj,
8817 &lustre_attr_track_declares_assert.attr);
8820 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8821 lu_kmem_fini(ldiskfs_caches);
8824 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8825 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8826 MODULE_VERSION(LUSTRE_VERSION_STRING);
8827 MODULE_LICENSE("GPL");
8829 module_init(osd_init);
8830 module_exit(osd_exit);