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/kallsyms.h>
43 #include <linux/module.h>
44 #include <linux/user_namespace.h>
45 #include <linux/uidgid.h>
47 /* prerequisite for linux/xattr.h */
48 #include <linux/types.h>
49 /* prerequisite for linux/xattr.h */
51 /* XATTR_{REPLACE,CREATE} */
52 #include <linux/xattr.h>
54 #include <ldiskfs/ldiskfs.h>
55 #include <ldiskfs/xattr.h>
56 #include <ldiskfs/ldiskfs_extents.h>
59 * struct OBD_{ALLOC,FREE}*()
62 #include <obd_support.h>
63 /* struct ptlrpc_thread */
64 #include <lustre_net.h>
65 #include <lustre_fid.h>
67 #include <uapi/linux/lustre/lustre_param.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 /* Maximum EA size is limited by LNET_MTU for remote objects */
79 #define OSD_MAX_EA_SIZE 1048364
82 module_param(ldiskfs_pdo, int, 0644);
83 MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");
85 int ldiskfs_track_declares_assert;
86 module_param(ldiskfs_track_declares_assert, int, 0644);
87 MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");
89 /* Slab to allocate dynlocks */
90 struct kmem_cache *dynlock_cachep;
92 /* Slab to allocate osd_it_ea */
93 struct kmem_cache *osd_itea_cachep;
95 static struct lu_kmem_descr ldiskfs_caches[] = {
97 .ckd_cache = &dynlock_cachep,
98 .ckd_name = "dynlock_cache",
99 .ckd_size = sizeof(struct dynlock_handle)
102 .ckd_cache = &osd_itea_cachep,
103 .ckd_name = "osd_itea_cache",
104 .ckd_size = sizeof(struct osd_it_ea)
111 static const char dot[] = ".";
112 static const char dotdot[] = "..";
114 static const struct lu_object_operations osd_lu_obj_ops;
115 static const struct dt_object_operations osd_obj_ops;
116 static const struct dt_object_operations osd_obj_otable_it_ops;
117 static const struct dt_index_operations osd_index_iam_ops;
118 static const struct dt_index_operations osd_index_ea_ops;
120 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
121 const struct lu_fid *fid);
122 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
123 struct osd_device *osd);
125 int osd_trans_declare_op2rb[] = {
126 [OSD_OT_ATTR_SET] = OSD_OT_ATTR_SET,
127 [OSD_OT_PUNCH] = OSD_OT_MAX,
128 [OSD_OT_XATTR_SET] = OSD_OT_XATTR_SET,
129 [OSD_OT_CREATE] = OSD_OT_DESTROY,
130 [OSD_OT_DESTROY] = OSD_OT_CREATE,
131 [OSD_OT_REF_ADD] = OSD_OT_REF_DEL,
132 [OSD_OT_REF_DEL] = OSD_OT_REF_ADD,
133 [OSD_OT_WRITE] = OSD_OT_WRITE,
134 [OSD_OT_INSERT] = OSD_OT_DELETE,
135 [OSD_OT_DELETE] = OSD_OT_INSERT,
136 [OSD_OT_QUOTA] = OSD_OT_MAX,
139 static int osd_has_index(const struct osd_object *obj)
141 return obj->oo_dt.do_index_ops != NULL;
144 static int osd_object_invariant(const struct lu_object *l)
146 return osd_invariant(osd_obj(l));
150 * Concurrency: doesn't matter
152 static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
154 struct osd_thread_info *oti = osd_oti_get(env);
156 return oti->oti_w_locks > 0 && o->oo_owner == env;
160 * Concurrency: doesn't access mutable data
162 static int osd_root_get(const struct lu_env *env,
163 struct dt_device *dev, struct lu_fid *f)
165 lu_local_obj_fid(f, OSD_FS_ROOT_OID);
170 * the following set of functions are used to maintain per-thread
171 * cache of FID->ino mapping. this mechanism is needed to resolve
172 * FID to inode at dt_insert() which in turn stores ino in the
173 * directory entries to keep ldiskfs compatible with ext[34].
174 * due to locking-originated restrictions we can't lookup ino
175 * using LU cache (deadlock is possible). lookup using OI is quite
176 * expensive. so instead we maintain this cache and methods like
177 * dt_create() fill it. so in the majority of cases dt_insert() is
178 * able to find needed mapping in lockless manner.
180 static struct osd_idmap_cache *
181 osd_idc_find(const struct lu_env *env, struct osd_device *osd,
182 const struct lu_fid *fid)
184 struct osd_thread_info *oti = osd_oti_get(env);
185 struct osd_idmap_cache *idc = oti->oti_ins_cache;
188 for (i = 0; i < oti->oti_ins_cache_used; i++) {
189 if (!lu_fid_eq(&idc[i].oic_fid, fid))
191 if (idc[i].oic_dev != osd)
200 static struct osd_idmap_cache *
201 osd_idc_add(const struct lu_env *env, struct osd_device *osd,
202 const struct lu_fid *fid)
204 struct osd_thread_info *oti = osd_oti_get(env);
205 struct osd_idmap_cache *idc;
208 if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
209 i = oti->oti_ins_cache_size * 2;
211 i = OSD_INS_CACHE_SIZE;
212 OBD_ALLOC_PTR_ARRAY(idc, i);
214 return ERR_PTR(-ENOMEM);
215 if (oti->oti_ins_cache != NULL) {
216 memcpy(idc, oti->oti_ins_cache,
217 oti->oti_ins_cache_used * sizeof(*idc));
218 OBD_FREE_PTR_ARRAY(oti->oti_ins_cache,
219 oti->oti_ins_cache_used);
221 oti->oti_ins_cache = idc;
222 oti->oti_ins_cache_size = i;
225 idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
228 idc->oic_lid.oii_ino = 0;
229 idc->oic_lid.oii_gen = 0;
236 * lookup mapping for the given fid in the cache, initialize a
237 * new one if not found. the initialization checks whether the
238 * object is local or remote. for local objects, OI is used to
239 * learn ino/generation. the function is used when the caller
240 * has no information about the object, e.g. at dt_insert().
242 static struct osd_idmap_cache *
243 osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
244 const struct lu_fid *fid)
246 struct osd_idmap_cache *idc;
249 idc = osd_idc_find(env, osd, fid);
250 LASSERT(!IS_ERR(idc));
254 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
255 osd->od_svname, PFID(fid));
257 /* new mapping is needed */
258 idc = osd_idc_add(env, osd, fid);
260 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
261 osd->od_svname, PFID(fid), PTR_ERR(idc));
266 rc = osd_remote_fid(env, osd, fid);
267 if (unlikely(rc < 0))
271 /* the object is local, lookup in OI */
272 /* XXX: probably cheaper to lookup in LU first? */
273 rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
275 if (unlikely(rc < 0)) {
276 CERROR("can't lookup: rc = %d\n", rc);
280 /* the object is remote */
287 static void osd_idc_dump_lma(const struct lu_env *env,
288 struct osd_device *osd,
292 struct osd_thread_info *info = osd_oti_get(env);
293 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
294 const struct lu_fid *fid;
295 struct osd_inode_id lid;
299 inode = osd_ldiskfs_iget(osd_sb(osd), ino);
301 CERROR("%s: can't get inode %lu: rc = %d\n",
302 osd->od_svname, ino, (int)PTR_ERR(inode));
305 if (is_bad_inode(inode)) {
306 CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
309 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
311 CERROR("%s: can't get LMA for %lu: rc = %d\n",
312 osd->od_svname, ino, rc);
315 fid = &loa->loa_lma.lma_self_fid;
316 LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
317 PFID(fid), ino, (unsigned)inode->i_generation);
320 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
322 CERROR("%s: can't lookup "DFID": rc = %d\n",
323 osd->od_svname, PFID(fid), rc);
326 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
327 PFID(fid), lid.oii_ino, lid.oii_gen);
332 static void osd_idc_dump_debug(const struct lu_env *env,
333 struct osd_device *osd,
334 const struct lu_fid *fid,
338 struct osd_inode_id lid;
342 rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
344 LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
345 osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
346 osd_idc_dump_lma(env, osd, lid.oii_ino, false);
348 CERROR("%s: can't lookup "DFID": rc = %d\n",
349 osd->od_svname, PFID(fid), rc);
352 osd_idc_dump_lma(env, osd, ino1, true);
354 osd_idc_dump_lma(env, osd, ino2, true);
358 * lookup mapping for given FID and fill it from the given object.
359 * the object is lolcal by definition.
361 static int osd_idc_find_and_init(const struct lu_env *env,
362 struct osd_device *osd,
363 struct osd_object *obj)
365 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
366 struct osd_idmap_cache *idc;
368 idc = osd_idc_find(env, osd, fid);
369 LASSERT(!IS_ERR(idc));
371 if (obj->oo_inode == NULL)
373 if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
374 if (idc->oic_lid.oii_ino) {
375 osd_idc_dump_debug(env, osd, fid,
376 idc->oic_lid.oii_ino,
377 obj->oo_inode->i_ino);
380 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
381 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
386 CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
387 osd->od_svname, PFID(fid));
389 /* new mapping is needed */
390 idc = osd_idc_add(env, osd, fid);
392 CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
393 osd->od_svname, PFID(fid), PTR_ERR(idc));
397 if (obj->oo_inode != NULL) {
398 idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
399 idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
405 * OSD object methods.
409 * Concurrency: no concurrent access is possible that early in object
412 static struct lu_object *osd_object_alloc(const struct lu_env *env,
413 const struct lu_object_header *hdr,
416 struct osd_object *mo;
421 struct lu_object_header *h;
422 struct osd_device *o = osd_dev(d);
424 l = &mo->oo_dt.do_lu;
425 if (unlikely(o->od_in_init)) {
432 lu_object_header_init(h);
433 lu_object_init(l, h, d);
434 lu_object_add_top(h, l);
437 dt_object_init(&mo->oo_dt, NULL, d);
438 mo->oo_header = NULL;
441 mo->oo_dt.do_ops = &osd_obj_ops;
442 l->lo_ops = &osd_lu_obj_ops;
443 init_rwsem(&mo->oo_sem);
444 init_rwsem(&mo->oo_ext_idx_sem);
445 spin_lock_init(&mo->oo_guard);
446 INIT_LIST_HEAD(&mo->oo_xattr_list);
452 int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
453 struct dentry *dentry, struct lustre_ost_attrs *loa)
457 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
458 (void *)loa, sizeof(*loa));
460 struct lustre_mdt_attrs *lma = &loa->loa_lma;
462 if (rc < sizeof(*lma))
466 lustre_loa_swab(loa, true);
467 /* Check LMA compatibility */
468 if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
470 CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
472 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
473 PFID(&lma->lma_self_fid), inode->i_ino, rc);
475 } else if (rc == 0) {
483 * retrieve object from backend ext fs.
485 struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
486 struct osd_inode_id *id)
489 struct inode *inode = NULL;
492 * if we look for an inode withing a running
493 * transaction, then we risk to deadlock
494 * osd_dirent_check_repair() breaks this
496 /* LASSERT(current->journal_info == NULL); */
498 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
500 CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
501 id->oii_ino, PTR_ERR(inode));
502 } else if (id->oii_gen != OSD_OII_NOGEN &&
503 inode->i_generation != id->oii_gen) {
504 CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
505 "i_generation = %u\n",
506 id->oii_ino, id->oii_gen, inode->i_generation);
508 inode = ERR_PTR(-ESTALE);
509 } else if (inode->i_nlink == 0) {
511 * due to parallel readdir and unlink,
512 * we can have dead inode here.
514 CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
516 inode = ERR_PTR(-ESTALE);
517 } else if (is_bad_inode(inode)) {
519 CWARN("%s: bad inode: ino = %u: rc = %d\n",
520 osd_dev2name(dev), id->oii_ino, rc);
523 } else if ((rc = osd_attach_jinode(inode))) {
527 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
528 if (id->oii_gen == OSD_OII_NOGEN)
529 osd_id_gen(id, inode->i_ino, inode->i_generation);
532 * Do not update file c/mtime in ldiskfs.
533 * NB: we don't have any lock to protect this because we don't
534 * have reference on osd_object now, but contention with
535 * another lookup + attr_set can't happen in the tiny window
536 * between if (...) and set S_NOCMTIME.
538 if (!(inode->i_flags & S_NOCMTIME))
539 inode->i_flags |= S_NOCMTIME;
544 int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
545 handle_t *handle, struct dentry *child,
546 struct inode *inode, struct htree_lock *hlock)
550 rc = __ldiskfs_add_entry(handle, child, inode, hlock);
551 if (rc == -ENOBUFS || rc == -ENOSPC) {
552 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
553 struct inode *parent = child->d_parent->d_inode;
554 struct lu_fid *fid = NULL;
555 char fidstr[FID_LEN + 1] = "unknown";
557 rc2 = osd_get_lma(info, parent, child->d_parent, loa);
559 fid = &loa->loa_lma.lma_self_fid;
560 } else if (rc2 == -ENODATA) {
561 if (unlikely(is_root_inode(parent))) {
562 fid = &info->oti_fid3;
563 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
564 } else if (!osd->od_is_ost && osd->od_index == 0) {
565 fid = &info->oti_fid3;
566 lu_igif_build(fid, parent->i_ino,
567 parent->i_generation);
572 snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));
574 /* below message is checked in sanity.sh test_129 */
576 CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
577 osd_name(osd), parent->i_ino, fidstr);
579 rc = 0; /* ignore such error now */
580 CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
581 osd_name(osd), parent->i_ino, fidstr);
591 osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
592 struct osd_inode_id *id, struct lu_fid *fid)
594 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
598 inode = osd_iget(info, dev, id);
602 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
604 *fid = loa->loa_lma.lma_self_fid;
605 } else if (rc == -ENODATA) {
606 if (unlikely(is_root_inode(inode)))
607 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
609 lu_igif_build(fid, inode->i_ino, inode->i_generation);
617 static struct inode *osd_iget_check(struct osd_thread_info *info,
618 struct osd_device *dev,
619 const struct lu_fid *fid,
620 struct osd_inode_id *id,
629 * The cached OI mapping is trustable. If we cannot locate the inode
630 * via the cached OI mapping, then return the failure to the caller
631 * directly without further OI checking.
635 inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
638 if (!trusted && (rc == -ENOENT || rc == -ESTALE))
641 CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
642 PFID(fid), id->oii_ino, rc);
646 if (is_bad_inode(inode)) {
651 CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
652 PFID(fid), id->oii_ino);
656 if (id->oii_gen != OSD_OII_NOGEN &&
657 inode->i_generation != id->oii_gen) {
662 CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
663 "oii_gen = %u, i_generation = %u\n", PFID(fid),
664 id->oii_ino, id->oii_gen, inode->i_generation);
668 if (inode->i_nlink == 0) {
673 CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
674 PFID(fid), id->oii_ino);
678 ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
682 __u32 saved_ino = id->oii_ino;
683 __u32 saved_gen = id->oii_gen;
686 LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);
688 rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
690 * XXX: There are four possible cases:
692 * Backup/restore caused the OI invalid.
694 * Someone unlinked the object but NOT removed
695 * the OI mapping, such as mount target device
696 * as ldiskfs, and modify something directly.
698 * Someone just removed the object between the
699 * former oi_lookup and the iget. It is normal.
700 * 4. Other failure cases.
702 * Generally, when the device is mounted, it will
703 * auto check whether the system is restored from
704 * file-level backup or not. We trust such detect
705 * to distinguish the 1st case from the 2nd case:
706 * if the OI files are consistent but may contain
707 * stale OI mappings because of case 2, if iget()
708 * returns -ENOENT or -ESTALE, then it should be
713 * If the OI mapping was in OI file before the
714 * osd_iget_check(), but now, it is disappear,
715 * then it must be removed by race. That is a
721 * It is the OI scrub updated the OI mapping by race.
722 * The new OI mapping must be valid.
724 if (saved_ino != id->oii_ino ||
725 (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
734 if (dev->od_scrub.os_scrub.os_file.sf_flags &
737 * It still can be the case 2, but we cannot
738 * distinguish it from the case 1. So return
739 * -EREMCHG to block current operation until
740 * OI scrub rebuilt the OI mappings.
749 if (inode->i_generation == id->oii_gen)
754 if (id->oii_gen == OSD_OII_NOGEN)
755 osd_id_gen(id, inode->i_ino, inode->i_generation);
758 * Do not update file c/mtime in ldiskfs.
759 * NB: we don't have any lock to protect this because we don't
760 * have reference on osd_object now, but contention with
761 * another lookup + attr_set can't happen in the tiny window
762 * between if (...) and set S_NOCMTIME.
764 if (!(inode->i_flags & S_NOCMTIME))
765 inode->i_flags |= S_NOCMTIME;
782 * \retval +v: new filter_fid does not contain self-fid
783 * \retval 0: filter_fid_18_23, contains self-fid
784 * \retval -v: other failure cases
786 int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
787 struct dentry *dentry, struct lu_fid *fid)
789 struct filter_fid *ff = &info->oti_ff;
790 struct ost_id *ostid = &info->oti_ostid;
793 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
794 if (rc == sizeof(struct filter_fid_18_23)) {
795 struct filter_fid_18_23 *ff_old = (void *)ff;
797 ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
798 rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
800 * XXX: use 0 as the index for compatibility, the caller will
801 * handle index related issues when necessary.
804 ostid_to_fid(fid, ostid, 0);
805 } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
807 } else if (rc >= 0) {
814 static int osd_lma_self_repair(struct osd_thread_info *info,
815 struct osd_device *osd, struct inode *inode,
816 const struct lu_fid *fid, __u32 compat)
821 LASSERT(current->journal_info == NULL);
823 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
824 osd_dto_credits_noquota[DTO_XATTR_SET]);
827 CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
832 rc = osd_ea_fid_set(info, inode, fid, compat, 0);
834 CWARN("%s: cannot self repair the LMA: rc = %d\n",
836 ldiskfs_journal_stop(jh);
840 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
842 struct osd_thread_info *info = osd_oti_get(env);
843 struct osd_device *osd = osd_obj2dev(obj);
844 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
845 struct lustre_mdt_attrs *lma = &loa->loa_lma;
846 struct inode *inode = obj->oo_inode;
847 struct dentry *dentry = &info->oti_obj_dentry;
848 struct lu_fid *fid = NULL;
849 const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
854 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
855 (void *)loa, sizeof(*loa));
856 if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
857 fid = &lma->lma_self_fid;
858 rc = osd_get_idif(info, inode, dentry, fid);
859 if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
861 * For the given OST-object, if it has neither LMA nor
862 * FID in XATTR_NAME_FID, then the given FID (which is
863 * contained in the @obj, from client RPC for locating
864 * the OST-object) is trusted. We use it to generate
867 osd_lma_self_repair(info, osd, inode, rfid,
878 lustre_lma_swab(lma);
879 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
880 (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
881 S_ISREG(inode->i_mode)))) {
882 CWARN("%s: unsupported incompat LMA feature(s) %#x for "
883 "fid = "DFID", ino = %lu\n", osd_name(osd),
884 lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
885 PFID(rfid), inode->i_ino);
888 fid = &lma->lma_self_fid;
889 if (lma->lma_compat & LMAC_STRIPE_INFO &&
891 obj->oo_pfid_in_lma = 1;
892 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
894 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
898 if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
899 if (fid_is_idif(rfid) && fid_is_idif(fid)) {
900 struct ost_id *oi = &info->oti_ostid;
901 struct lu_fid *fid1 = &info->oti_fid3;
902 __u32 idx = fid_idif_ost_idx(rfid);
905 * For old IDIF, the OST index is not part of the IDIF,
906 * Means that different OSTs may have the same IDIFs.
907 * Under such case, we need to make some compatible
908 * check to make sure to trigger OI scrub properly.
910 if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
911 /* Given @rfid is new, LMA is old. */
912 fid_to_ostid(fid, oi);
913 ostid_to_fid(fid1, oi, idx);
914 if (lu_fid_eq(fid1, rfid)) {
915 if (osd->od_index_in_idif)
916 osd_lma_self_repair(info, osd,
930 struct osd_check_lmv_buf {
931 /* please keep it as first member */
932 struct dir_context ctx;
933 struct osd_thread_info *oclb_info;
934 struct osd_device *oclb_dev;
940 * It is called internally by ->iterate*() to filter out the
941 * local slave object's FID of the striped directory.
943 * \retval 1 found the local slave's FID
944 * \retval 0 continue to check next item
945 * \retval -ve for failure
947 #ifdef HAVE_FILLDIR_USE_CTX
948 static int osd_stripe_dir_filldir(struct dir_context *buf,
950 static int osd_stripe_dir_filldir(void *buf,
952 const char *name, int namelen,
953 loff_t offset, __u64 ino, unsigned int d_type)
955 struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
956 struct osd_thread_info *oti = oclb->oclb_info;
957 struct lu_fid *fid = &oti->oti_fid3;
958 struct osd_inode_id *id = &oti->oti_id3;
959 struct osd_device *dev = oclb->oclb_dev;
968 sscanf(name + 1, SFID, RFID(fid));
969 if (!fid_is_sane(fid))
972 if (osd_remote_fid(oti->oti_env, dev, fid))
975 osd_id_gen(id, ino, OSD_OII_NOGEN);
976 inode = osd_iget(oti, dev, id);
978 return PTR_ERR(inode);
981 osd_add_oi_cache(oti, dev, id, fid);
982 osd_oii_insert(dev, fid, id, true);
983 oclb->oclb_found = true;
989 * When lookup item under striped directory, we need to locate the master
990 * MDT-object of the striped directory firstly, then the client will send
991 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
992 * and the item's name. If the system is restored from MDT file level backup,
993 * then before the OI scrub completely built the OI files, the OI mappings of
994 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
995 * not a problem for the master MDT-object. Because when locate the master
996 * MDT-object, we will do name based lookup (for the striped directory itself)
997 * firstly, during such process we can setup the correct OI mapping for the
998 * master MDT-object. But it will be trouble for the slave MDT-object. Because
999 * the client will not trigger name based lookup on the MDT to locate the slave
1000 * MDT-object before locating item under the striped directory, then when
1001 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
1002 * is invalid and does not know what the right OI mapping is, then the MDT has
1003 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
1004 * the OI file, related OI mapping is unknown yet, please try again later. And
1005 * then client will re-try the RPC again and again until related OI mapping has
1006 * been updated. That is quite inefficient.
1008 * To resolve above trouble, we will handle it as the following two cases:
1010 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
1011 * It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
1012 * is linked under /REMOTE_PARENT_DIR with the name of its FID string.
1013 * We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
1014 * directly. Please check osd_fid_lookup().
1016 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
1017 * Under such case, during lookup the master MDT-object, we will lookup the
1018 * slave MDT-object via readdir against the master MDT-object, because the
1019 * slave MDT-objects information are stored as sub-directories with the name
1020 * "${FID}:${index}". Then when find the local slave MDT-object, its OI
1021 * mapping will be recorded. Then subsequent osd_fid_lookup() will know
1022 * the correct OI mapping for the slave MDT-object.
1024 static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
1025 struct inode *inode)
1027 struct lu_buf *buf = &oti->oti_big_buf;
1028 struct dentry *dentry = &oti->oti_obj_dentry;
1030 struct lmv_mds_md_v1 *lmv1;
1031 struct osd_check_lmv_buf oclb = {
1032 .ctx.actor = osd_stripe_dir_filldir,
1035 .oclb_found = false,
1042 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
1044 if (rc == -ERANGE) {
1045 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
1047 lu_buf_realloc(buf, rc);
1048 if (buf->lb_buf == NULL)
1049 GOTO(out, rc = -ENOMEM);
1055 if (unlikely(rc == 0 || rc == -ENODATA))
1061 if (unlikely(buf->lb_buf == NULL)) {
1062 lu_buf_realloc(buf, rc);
1063 if (buf->lb_buf == NULL)
1064 GOTO(out, rc = -ENOMEM);
1070 if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
1073 filp = osd_quasi_file(oti->oti_env, inode);
1074 rc = osd_security_file_alloc(filp);
1079 oclb.oclb_items = 0;
1080 rc = iterate_dir(filp, &oclb.ctx);
1081 } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
1082 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
1083 inode->i_fop->release(inode, filp);
1088 "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
1089 osd_ino2name(inode), inode->i_ino, inode->i_generation,
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;
1121 LINVRNT(osd_invariant(obj));
1122 LASSERT(obj->oo_inode == NULL);
1124 if (fid_is_sane(fid) == 0) {
1125 CERROR("%s: invalid FID "DFID"\n", ldev->ld_obd->obd_name,
1131 dev = osd_dev(ldev);
1132 scrub = &dev->od_scrub.os_scrub;
1133 sf = &scrub->os_file;
1134 info = osd_oti_get(env);
1136 oic = &info->oti_cache;
1138 if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
1142 * For the object is created as locking anchor, or for the object to
1143 * be created on disk. No need to osd_oi_lookup() at here because FID
1144 * shouldn't never be re-used, if it's really a duplicate FID from
1145 * unexpected reason, we should be able to detect it later by calling
1146 * do_create->osd_oi_insert().
1148 if (conf && conf->loc_flags & LOC_F_NEW)
1149 GOTO(out, result = 0);
1151 /* Search order: 1. per-thread cache. */
1152 if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
1158 if (!list_empty(&scrub->os_inconsistent_items)) {
1159 /* Search order: 2. OI scrub pending list. */
1160 result = osd_oii_lookup(dev, fid, id);
1166 * The OI mapping in the OI file can be updated by the OI scrub
1167 * when we locate the inode via FID. So it may be not trustable.
1171 /* Search order: 3. OI files. */
1172 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1173 if (result == -ENOENT) {
1174 if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
1175 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
1176 !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
1178 GOTO(out, result = 0);
1183 /* -ESTALE is returned if inode of OST object doesn't exist */
1184 if (result == -ESTALE &&
1185 fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
1186 GOTO(out, result = 0);
1193 obj->oo_inode = NULL;
1194 /* for later passes through checks, not true on first pass */
1195 if (!IS_ERR_OR_NULL(inode))
1198 inode = osd_iget_check(info, dev, fid, id, trusted);
1199 if (!IS_ERR(inode)) {
1200 obj->oo_inode = inode;
1208 result = PTR_ERR(inode);
1209 if (result == -ENOENT || result == -ESTALE)
1210 GOTO(out, result = 0);
1212 if (result != -EREMCHG)
1217 * We still have chance to get the valid inode: for the
1218 * object which is referenced by remote name entry, the
1219 * object on the local MDT will be linked under the dir
1220 * of "/REMOTE_PARENT_DIR" with its FID string as name.
1222 * We do not know whether the object for the given FID
1223 * is referenced by some remote name entry or not, and
1224 * especially for DNE II, a multiple-linked object may
1225 * have many name entries reside on many MDTs.
1227 * To simplify the operation, OSD will not distinguish
1228 * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
1229 * only happened for the RPC from other MDT during the
1230 * OI scrub, or for the client side RPC with FID only,
1231 * such as FID to path, or from old connected client.
1234 rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
1238 flags |= SS_AUTO_PARTIAL;
1239 flags &= ~SS_AUTO_FULL;
1244 if (scrub->os_running) {
1245 if (scrub->os_partial_scan && !scrub->os_in_join)
1248 if (IS_ERR_OR_NULL(inode) || result) {
1249 osd_oii_insert(dev, fid, id, result == -ENOENT);
1250 GOTO(out, result = -EINPROGRESS);
1254 LASSERT(obj->oo_inode == inode);
1256 osd_oii_insert(dev, fid, id, true);
1260 if (dev->od_auto_scrub_interval == AS_NEVER) {
1262 GOTO(out, result = -EREMCHG);
1265 LASSERT(obj->oo_inode == inode);
1267 osd_add_oi_cache(info, dev, id, fid);
1272 rc1 = osd_scrub_start(env, dev, flags);
1273 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
1274 "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
1275 osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
1276 if (rc1 && rc1 != -EALREADY)
1277 GOTO(out, result = -EREMCHG);
1279 if (IS_ERR_OR_NULL(inode) || result) {
1280 osd_oii_insert(dev, fid, id, result == -ENOENT);
1281 GOTO(out, result = -EINPROGRESS);
1285 LASSERT(obj->oo_inode == inode);
1287 osd_oii_insert(dev, fid, id, true);
1292 if (unlikely(obj->oo_header))
1295 result = osd_check_lma(env, obj);
1299 LASSERTF(id->oii_ino == inode->i_ino &&
1300 id->oii_gen == inode->i_generation,
1301 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
1302 PFID(fid), id->oii_ino, id->oii_gen,
1303 inode->i_ino, inode->i_generation);
1305 saved_ino = inode->i_ino;
1306 saved_gen = inode->i_generation;
1308 if (unlikely(result == -ENODATA)) {
1310 * If the OI scrub updated the OI mapping by race, it
1311 * must be valid. Trust the inode that has no LMA EA.
1316 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1319 * The OI mapping is still there, the inode is still
1320 * valid. It is just becaues the inode has no LMA EA.
1322 if (saved_ino == id->oii_ino &&
1323 saved_gen == id->oii_gen)
1327 * It is the OI scrub updated the OI mapping by race.
1328 * The new OI mapping must be valid.
1336 * "result == -ENOENT" means that the OI mappinghas been
1337 * removed by race, so the inode belongs to other object.
1339 * Others error can be returned directly.
1341 if (result == -ENOENT) {
1344 obj->oo_inode = NULL;
1349 if (result != -EREMCHG)
1355 * if two OST objects map to the same inode, and inode mode is
1356 * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
1357 * reserved by precreate, and not written yet, in this case, don't
1358 * set inode for the object whose FID mismatch, so that it can create
1359 * inode and not block precreate.
1361 if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
1362 inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
1363 obj->oo_inode = NULL;
1364 GOTO(out, result = 0);
1367 result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
1369 * "result == -ENOENT" means the cached OI mapping has been removed
1370 * from the OI file by race, above inode belongs to other object.
1372 if (result == -ENOENT) {
1375 obj->oo_inode = NULL;
1376 GOTO(out, result = 0);
1382 if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
1384 osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
1390 * It is the OI scrub updated the OI mapping by race.
1391 * The new OI mapping must be valid.
1399 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
1400 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
1402 result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
1404 if (lma->lma_compat & LMAC_STRIPE_INFO &&
1406 obj->oo_pfid_in_lma = 1;
1407 if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
1409 lu_object_set_agent_entry(&obj->oo_dt.do_lu);
1410 } else if (result != -ENODATA) {
1415 obj->oo_compat_dot_created = 1;
1416 obj->oo_compat_dotdot_created = 1;
1418 if (S_ISDIR(inode->i_mode) &&
1419 (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
1420 osd_check_lmv(info, dev, inode);
1422 result = osd_attach_jinode(inode);
1427 GOTO(out, result = 0);
1429 LASSERT(!obj->oo_hl_head);
1430 obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
1432 GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));
1435 if (result || !obj->oo_inode) {
1436 if (!IS_ERR_OR_NULL(inode))
1439 obj->oo_inode = NULL;
1441 fid_zero(&oic->oic_fid);
1444 LINVRNT(osd_invariant(obj));
1449 * Concurrency: shouldn't matter.
1451 static void osd_object_init0(struct osd_object *obj)
1453 LASSERT(obj->oo_inode != NULL);
1454 obj->oo_dt.do_body_ops = &osd_body_ops;
1455 obj->oo_dt.do_lu.lo_header->loh_attr |=
1456 (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
1460 * Concurrency: no concurrent access is possible that early in object
1463 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
1464 const struct lu_object_conf *conf)
1466 struct osd_object *obj = osd_obj(l);
1469 LINVRNT(osd_invariant(obj));
1471 if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
1472 cfs_fail_val == 2) {
1473 struct osd_thread_info *info = osd_oti_get(env);
1474 struct osd_idmap_cache *oic = &info->oti_cache;
1475 /* invalidate thread cache */
1476 memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
1478 if (fid_is_otable_it(&l->lo_header->loh_fid)) {
1479 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
1480 l->lo_header->loh_attr |= LOHA_EXISTS;
1484 result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
1485 obj->oo_dt.do_body_ops = &osd_body_ops_new;
1486 if (result == 0 && obj->oo_inode != NULL) {
1487 struct osd_thread_info *oti = osd_oti_get(env);
1488 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
1490 osd_object_init0(obj);
1491 if (unlikely(obj->oo_header))
1494 result = osd_get_lma(oti, obj->oo_inode,
1495 &oti->oti_obj_dentry, loa);
1498 * Convert LMAI flags to lustre LMA flags
1499 * and cache it to oo_lma_flags
1502 lma_to_lustre_flags(loa->loa_lma.lma_incompat);
1503 } else if (result == -ENODATA) {
1507 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
1509 LINVRNT(osd_invariant(obj));
1514 * The first part of oxe_buf is xattr name, and is '\0' terminated.
1515 * The left part is for value, binary mode.
1517 struct osd_xattr_entry {
1518 struct list_head oxe_list;
1522 struct rcu_head oxe_rcu;
1526 static int osd_oxc_get(struct osd_object *obj, const char *name,
1529 struct osd_xattr_entry *tmp;
1530 struct osd_xattr_entry *oxe = NULL;
1531 size_t namelen = strlen(name);
1535 list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
1536 if (namelen == tmp->oxe_namelen &&
1537 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1544 GOTO(out, rc = -ENOENT);
1546 if (!oxe->oxe_exist)
1547 GOTO(out, rc = -ENODATA);
1550 rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
1553 if (buf->lb_buf == NULL)
1556 if (buf->lb_len < rc)
1557 GOTO(out, rc = -ERANGE);
1559 memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
1566 static void osd_oxc_free(struct rcu_head *head)
1568 struct osd_xattr_entry *oxe;
1570 oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
1571 OBD_FREE(oxe, oxe->oxe_len);
1574 static void osd_oxc_add(struct osd_object *obj, const char *name,
1575 const char *buf, int buflen)
1577 struct osd_xattr_entry *oxe;
1578 struct osd_xattr_entry *old = NULL;
1579 struct osd_xattr_entry *tmp;
1580 size_t namelen = strlen(name);
1581 size_t len = sizeof(*oxe) + namelen + 1 + buflen;
1583 OBD_ALLOC(oxe, len);
1587 INIT_LIST_HEAD(&oxe->oxe_list);
1589 oxe->oxe_namelen = namelen;
1590 memcpy(oxe->oxe_buf, name, namelen);
1592 LASSERT(buf != NULL);
1593 memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
1594 oxe->oxe_exist = true;
1596 oxe->oxe_exist = false;
1599 /* this should be rarely called, just remove old and add new */
1600 spin_lock(&obj->oo_guard);
1601 list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
1602 if (namelen == tmp->oxe_namelen &&
1603 strncmp(name, tmp->oxe_buf, namelen) == 0) {
1609 list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
1610 call_rcu(&old->oxe_rcu, osd_oxc_free);
1612 list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
1614 spin_unlock(&obj->oo_guard);
1617 static void osd_oxc_del(struct osd_object *obj, const char *name)
1619 struct osd_xattr_entry *oxe;
1620 size_t namelen = strlen(name);
1622 spin_lock(&obj->oo_guard);
1623 list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
1624 if (namelen == oxe->oxe_namelen &&
1625 strncmp(name, oxe->oxe_buf, namelen) == 0) {
1626 list_del_rcu(&oxe->oxe_list);
1627 call_rcu(&oxe->oxe_rcu, osd_oxc_free);
1631 spin_unlock(&obj->oo_guard);
1634 static void osd_oxc_fini(struct osd_object *obj)
1636 struct osd_xattr_entry *oxe, *next;
1638 list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
1639 list_del(&oxe->oxe_list);
1640 OBD_FREE(oxe, oxe->oxe_len);
1645 * Concurrency: no concurrent access is possible that late in object
1648 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
1650 struct osd_object *obj = osd_obj(l);
1651 struct lu_object_header *h = obj->oo_header;
1653 LINVRNT(osd_invariant(obj));
1656 dt_object_fini(&obj->oo_dt);
1657 if (obj->oo_hl_head != NULL)
1658 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
1659 /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
1662 lu_object_header_free(h);
1666 * Concurrency: no concurrent access is possible that late in object
1669 static void osd_index_fini(struct osd_object *o)
1671 struct iam_container *bag;
1673 if (o->oo_dir != NULL) {
1674 bag = &o->oo_dir->od_container;
1675 if (o->oo_inode != NULL) {
1676 if (bag->ic_object == o->oo_inode)
1677 iam_container_fini(bag);
1679 OBD_FREE_PTR(o->oo_dir);
1685 OSD_TXN_OI_DELETE_CREDITS = 20,
1686 OSD_TXN_INODE_DELETE_CREDITS = 20
1693 #if OSD_THANDLE_STATS
1695 * Set time when the handle is allocated
1697 static void osd_th_alloced(struct osd_thandle *oth)
1699 oth->oth_alloced = ktime_get();
1703 * Set time when the handle started
1705 static void osd_th_started(struct osd_thandle *oth)
1707 oth->oth_started = ktime_get();
1711 * Check whether the we deal with this handle for too long.
1713 static void __osd_th_check_slow(void *oth, struct osd_device *dev,
1714 ktime_t alloced, ktime_t started,
1717 ktime_t now = ktime_get();
1719 LASSERT(dev != NULL);
1721 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
1722 ktime_us_delta(started, alloced));
1723 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
1724 ktime_us_delta(closed, started));
1725 lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
1726 ktime_us_delta(now, closed));
1728 if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
1729 CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
1730 oth, now, alloced, started, closed);
1731 libcfs_debug_dumpstack(NULL);
1735 #define OSD_CHECK_SLOW_TH(oth, dev, expr) \
1737 ktime_t __closed = ktime_get(); \
1738 ktime_t __alloced = oth->oth_alloced; \
1739 ktime_t __started = oth->oth_started; \
1742 __osd_th_check_slow(oth, dev, __alloced, __started, __closed); \
1745 #else /* OSD_THANDLE_STATS */
1747 #define osd_th_alloced(h) do {} while(0)
1748 #define osd_th_started(h) do {} while(0)
1749 #define OSD_CHECK_SLOW_TH(oth, dev, expr) expr
1751 #endif /* OSD_THANDLE_STATS */
1754 * Concurrency: doesn't access mutable data.
1756 static int osd_param_is_not_sane(const struct osd_device *dev,
1757 const struct thandle *th)
1759 struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);
1761 return oh->ot_credits > osd_transaction_size(dev);
1765 * Concurrency: shouldn't matter.
1767 static void osd_trans_commit_cb(struct super_block *sb,
1768 struct ldiskfs_journal_cb_entry *jcb, int error)
1770 struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
1771 struct thandle *th = &oh->ot_super;
1772 struct lu_device *lud = &th->th_dev->dd_lu_dev;
1773 struct osd_device *osd = osd_dev(lud);
1774 struct dt_txn_commit_cb *dcb, *tmp;
1776 LASSERT(oh->ot_handle == NULL);
1779 CERROR("transaction @0x%p commit error: %d\n", th, error);
1781 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
1782 /* call per-transaction callbacks if any */
1783 list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
1785 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
1786 "commit callback entry: magic=%x name='%s'\n",
1787 dcb->dcb_magic, dcb->dcb_name);
1788 list_del_init(&dcb->dcb_linkage);
1789 dcb->dcb_func(NULL, th, dcb, error);
1792 lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
1793 if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
1794 wake_up(&osd->od_commit_cb_done);
1800 static struct thandle *osd_trans_create(const struct lu_env *env,
1801 struct dt_device *d)
1803 struct osd_thread_info *oti = osd_oti_get(env);
1804 struct osd_iobuf *iobuf = &oti->oti_iobuf;
1805 struct osd_thandle *oh;
1811 CERROR("%s: someone try to start transaction under "
1812 "readonly mode, should be disabled.\n",
1813 osd_name(osd_dt_dev(d)));
1815 RETURN(ERR_PTR(-EROFS));
1818 /* on pending IO in this thread should left from prev. request */
1819 LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);
1821 sb_start_write(osd_sb(osd_dt_dev(d)));
1823 OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
1825 sb_end_write(osd_sb(osd_dt_dev(d)));
1826 RETURN(ERR_PTR(-ENOMEM));
1829 oh->ot_quota_trans = &oti->oti_quota_trans;
1830 memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
1835 oh->oh_declared_ext = 0;
1836 INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
1837 INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
1838 INIT_LIST_HEAD(&oh->ot_trunc_locks);
1841 memset(oti->oti_declare_ops, 0,
1842 sizeof(oti->oti_declare_ops));
1843 memset(oti->oti_declare_ops_cred, 0,
1844 sizeof(oti->oti_declare_ops_cred));
1845 memset(oti->oti_declare_ops_used, 0,
1846 sizeof(oti->oti_declare_ops_used));
1848 oti->oti_ins_cache_depth++;
1853 void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
1855 struct osd_thread_info *oti = osd_oti_get(env);
1856 struct osd_thandle *oh;
1858 oh = container_of(th, struct osd_thandle, ot_super);
1859 LASSERT(oh != NULL);
1861 CWARN(" create: %u/%u/%u, destroy: %u/%u/%u\n",
1862 oti->oti_declare_ops[OSD_OT_CREATE],
1863 oti->oti_declare_ops_cred[OSD_OT_CREATE],
1864 oti->oti_declare_ops_used[OSD_OT_CREATE],
1865 oti->oti_declare_ops[OSD_OT_DESTROY],
1866 oti->oti_declare_ops_cred[OSD_OT_DESTROY],
1867 oti->oti_declare_ops_used[OSD_OT_DESTROY]);
1868 CWARN(" attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
1869 oti->oti_declare_ops[OSD_OT_ATTR_SET],
1870 oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
1871 oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
1872 oti->oti_declare_ops[OSD_OT_XATTR_SET],
1873 oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
1874 oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
1875 CWARN(" write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
1876 oti->oti_declare_ops[OSD_OT_WRITE],
1877 oti->oti_declare_ops_cred[OSD_OT_WRITE],
1878 oti->oti_declare_ops_used[OSD_OT_WRITE],
1879 oti->oti_declare_ops[OSD_OT_PUNCH],
1880 oti->oti_declare_ops_cred[OSD_OT_PUNCH],
1881 oti->oti_declare_ops_used[OSD_OT_PUNCH],
1882 oti->oti_declare_ops[OSD_OT_QUOTA],
1883 oti->oti_declare_ops_cred[OSD_OT_QUOTA],
1884 oti->oti_declare_ops_used[OSD_OT_QUOTA]);
1885 CWARN(" insert: %u/%u/%u, delete: %u/%u/%u\n",
1886 oti->oti_declare_ops[OSD_OT_INSERT],
1887 oti->oti_declare_ops_cred[OSD_OT_INSERT],
1888 oti->oti_declare_ops_used[OSD_OT_INSERT],
1889 oti->oti_declare_ops[OSD_OT_DELETE],
1890 oti->oti_declare_ops_cred[OSD_OT_DELETE],
1891 oti->oti_declare_ops_used[OSD_OT_DELETE]);
1892 CWARN(" ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
1893 oti->oti_declare_ops[OSD_OT_REF_ADD],
1894 oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
1895 oti->oti_declare_ops_used[OSD_OT_REF_ADD],
1896 oti->oti_declare_ops[OSD_OT_REF_DEL],
1897 oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
1898 oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
1902 * Concurrency: shouldn't matter.
1904 static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
1907 struct osd_thread_info *oti = osd_oti_get(env);
1908 struct osd_device *dev = osd_dt_dev(d);
1910 struct osd_thandle *oh;
1915 LASSERT(current->journal_info == NULL);
1917 oh = container_of(th, struct osd_thandle, ot_super);
1918 LASSERT(oh != NULL);
1919 LASSERT(oh->ot_handle == NULL);
1921 rc = dt_txn_hook_start(env, d, th);
1925 if (unlikely(osd_param_is_not_sane(dev, th))) {
1926 static unsigned long last_printed;
1927 static int last_credits;
1930 * don't make noise on a tiny testing systems
1931 * actual credits misuse will be caught anyway
1933 if (last_credits != oh->ot_credits &&
1934 time_after(jiffies, last_printed +
1935 cfs_time_seconds(60)) &&
1936 osd_transaction_size(dev) > 512) {
1937 CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
1938 oh->ot_credits, osd_transaction_size(dev));
1939 osd_trans_dump_creds(env, th);
1940 libcfs_debug_dumpstack(NULL);
1941 last_credits = oh->ot_credits;
1942 last_printed = jiffies;
1945 * XXX Limit the credits to 'max_transaction_buffers', and
1946 * let the underlying filesystem to catch the error if
1947 * we really need so many credits.
1949 * This should be removed when we can calculate the
1950 * credits precisely.
1952 oh->ot_credits = osd_transaction_size(dev);
1953 } else if (ldiskfs_track_declares_assert != 0) {
1955 * reserve few credits to prevent an assertion in JBD
1956 * our debugging mechanism will be able to detected
1957 * overuse. this can help to debug single-update
1960 oh->ot_credits += 10;
1961 if (unlikely(osd_param_is_not_sane(dev, th)))
1962 oh->ot_credits = osd_transaction_size(dev);
1965 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
1966 GOTO(out, rc = -EIO);
1969 * XXX temporary stuff. Some abstraction layer should
1972 jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
1976 LASSERT(oti->oti_txns == 0);
1978 atomic_inc(&dev->od_commit_cb_in_flight);
1979 lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
1990 static int osd_seq_exists(const struct lu_env *env,
1991 struct osd_device *osd, u64 seq)
1993 struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
1994 struct seq_server_site *ss = osd_seq_site(osd);
1999 LASSERT(ss != NULL);
2000 LASSERT(ss->ss_server_fld != NULL);
2002 rc = osd_fld_lookup(env, osd, seq, range);
2005 CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
2006 osd_name(osd), seq, rc);
2010 RETURN(ss->ss_node_id == range->lsr_index);
2013 static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
2015 struct dt_txn_commit_cb *dcb;
2016 struct dt_txn_commit_cb *tmp;
2018 /* call per-transaction stop callbacks if any */
2019 list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
2021 LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
2022 "commit callback entry: magic=%x name='%s'\n",
2023 dcb->dcb_magic, dcb->dcb_name);
2024 list_del_init(&dcb->dcb_linkage);
2025 dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
2030 * Concurrency: shouldn't matter.
2032 static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
2035 struct osd_thread_info *oti = osd_oti_get(env);
2036 struct osd_thandle *oh;
2037 struct osd_iobuf *iobuf = &oti->oti_iobuf;
2038 struct osd_device *osd = osd_dt_dev(th->th_dev);
2039 struct qsd_instance *qsd = osd_def_qsd(osd);
2040 struct lquota_trans *qtrans;
2041 LIST_HEAD(truncates);
2042 int rc = 0, remove_agents = 0;
2046 oh = container_of(th, struct osd_thandle, ot_super);
2048 remove_agents = oh->ot_remove_agents;
2050 qtrans = oh->ot_quota_trans;
2051 oh->ot_quota_trans = NULL;
2053 /* move locks to local list, stop tx, execute truncates */
2054 list_splice(&oh->ot_trunc_locks, &truncates);
2056 if (oh->ot_handle != NULL) {
2059 handle_t *hdl = oh->ot_handle;
2062 * add commit callback
2063 * notice we don't do this in osd_trans_start()
2064 * as underlying transaction can change during truncate
2066 ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
2069 LASSERT(oti->oti_txns == 1);
2072 rc = dt_txn_hook_stop(env, th);
2074 CERROR("%s: failed in transaction hook: rc = %d\n",
2077 osd_trans_stop_cb(oh, rc);
2078 /* hook functions might modify th_sync */
2079 hdl->h_sync = th->th_sync;
2081 oh->ot_handle = NULL;
2082 OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
2084 CERROR("%s: failed to stop transaction: rc = %d\n",
2085 osd_name(osd), rc2);
2089 osd_process_truncates(env, &truncates);
2091 osd_trans_stop_cb(oh, th->th_result);
2095 osd_trunc_unlock_all(env, &truncates);
2097 /* inform the quota slave device that the transaction is stopping */
2098 qsd_op_end(env, qsd, qtrans);
2101 * as we want IO to journal and data IO be concurrent, we don't block
2102 * awaiting data IO completion in osd_do_bio(), instead we wait here
2103 * once transaction is submitted to the journal. all reqular requests
2104 * don't do direct IO (except read/write), thus this wait_event becomes
2107 * IMPORTANT: we have to wait till any IO submited by the thread is
2108 * completed otherwise iobuf may be corrupted by different request
2110 wait_event(iobuf->dr_wait,
2111 atomic_read(&iobuf->dr_numreqs) == 0);
2112 osd_fini_iobuf(osd, iobuf);
2114 rc = iobuf->dr_error;
2116 if (unlikely(remove_agents != 0))
2117 osd_process_scheduled_agent_removals(env, osd);
2119 LASSERT(oti->oti_ins_cache_depth > 0);
2120 oti->oti_ins_cache_depth--;
2121 /* reset OI cache for safety */
2122 if (oti->oti_ins_cache_depth == 0)
2123 oti->oti_ins_cache_used = 0;
2125 sb_end_write(osd_sb(osd));
2130 static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
2132 struct osd_thandle *oh = container_of(th, struct osd_thandle,
2135 LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
2136 LASSERT(&dcb->dcb_func != NULL);
2137 if (dcb->dcb_flags & DCB_TRANS_STOP)
2138 list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
2140 list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);
2146 * Called just before object is freed. Releases all resources except for
2147 * object itself (that is released by osd_object_free()).
2149 * Concurrency: no concurrent access is possible that late in object
2152 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
2154 struct osd_object *obj = osd_obj(l);
2155 struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
2156 struct inode *inode = obj->oo_inode;
2161 LINVRNT(osd_invariant(obj));
2164 * If object is unlinked remove fid->ino mapping from object index.
2167 osd_index_fini(obj);
2172 if (osd_has_index(obj) && obj->oo_dt.do_index_ops == &osd_index_iam_ops)
2173 ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);
2175 uid = i_uid_read(inode);
2176 gid = i_gid_read(inode);
2177 projid = i_projid_read(inode);
2179 obj->oo_inode = NULL;
2182 /* do not rebalance quota if the caller needs to release memory
2183 * otherwise qsd_refresh_usage() may went into a new ldiskfs
2184 * transaction and risk to deadlock - LU-12178 */
2185 if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
2188 if (!obj->oo_header && qsd) {
2189 struct osd_thread_info *info = osd_oti_get(env);
2190 struct lquota_id_info *qi = &info->oti_qi;
2192 /* Release granted quota to master if necessary */
2193 qi->lqi_id.qid_uid = uid;
2194 qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);
2196 qi->lqi_id.qid_uid = gid;
2197 qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);
2199 qi->lqi_id.qid_uid = projid;
2200 qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
2205 * Concurrency: ->loo_object_release() is called under site spin-lock.
2207 static void osd_object_release(const struct lu_env *env,
2208 struct lu_object *l)
2210 struct osd_object *o = osd_obj(l);
2213 * nobody should be releasing a non-destroyed object with nlink=0
2214 * the API allows this, but ldiskfs doesn't like and then report
2215 * this inode as deleted
2217 LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
2218 o->oo_inode->i_nlink == 0));
2222 * Concurrency: shouldn't matter.
2224 static int osd_object_print(const struct lu_env *env, void *cookie,
2225 lu_printer_t p, const struct lu_object *l)
2227 struct osd_object *o = osd_obj(l);
2228 struct iam_descr *d;
2230 if (o->oo_dir != NULL)
2231 d = o->oo_dir->od_container.ic_descr;
2234 return (*p)(env, cookie,
2235 LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
2237 o->oo_inode ? o->oo_inode->i_ino : 0UL,
2238 o->oo_inode ? o->oo_inode->i_generation : 0,
2239 d ? d->id_ops->id_name : "plain");
2243 * Concurrency: shouldn't matter.
2245 int osd_statfs(const struct lu_env *env, struct dt_device *d,
2246 struct obd_statfs *sfs, struct obd_statfs_info *info)
2248 struct osd_device *osd = osd_dt_dev(d);
2249 struct super_block *sb = osd_sb(osd);
2250 struct kstatfs *ksfs;
2254 if (unlikely(osd->od_mnt == NULL))
2255 return -EINPROGRESS;
2257 /* osd_lproc.c call this without env, allocate ksfs for that case */
2258 if (unlikely(env == NULL)) {
2259 OBD_ALLOC_PTR(ksfs);
2263 ksfs = &osd_oti_get(env)->oti_ksfs;
2266 result = sb->s_op->statfs(sb->s_root, ksfs);
2270 statfs_pack(sfs, ksfs);
2271 if (unlikely(sb->s_flags & SB_RDONLY))
2272 sfs->os_state |= OS_STATFS_READONLY;
2274 sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;
2276 if (ldiskfs_has_feature_extents(sb))
2277 sfs->os_maxbytes = sb->s_maxbytes;
2279 sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2282 * Reserve some space so to avoid fragmenting the filesystem too much.
2283 * Fragmentation not only impacts performance, but can also increase
2284 * metadata overhead significantly, causing grant calculation to be
2287 * Reserve 0.78% of total space, at least 8MB for small filesystems.
2289 BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
2290 reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
2291 if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
2292 reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
2294 sfs->os_blocks -= reserved;
2295 sfs->os_bfree -= min(reserved, sfs->os_bfree);
2296 sfs->os_bavail -= min(reserved, sfs->os_bavail);
2299 if (unlikely(env == NULL))
2305 * Estimate space needed for file creations. We assume the largest filename
2306 * which is 2^64 - 1, hence a filename of 20 chars.
2307 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
2309 #ifdef __LDISKFS_DIR_REC_LEN
2310 #define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
2312 #define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
2316 * Concurrency: doesn't access mutable data.
2318 static void osd_conf_get(const struct lu_env *env,
2319 const struct dt_device *dev,
2320 struct dt_device_param *param)
2322 struct osd_device *d = osd_dt_dev(dev);
2323 struct super_block *sb = osd_sb(d);
2324 struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
2329 * XXX should be taken from not-yet-existing fs abstraction layer.
2331 param->ddp_max_name_len = LDISKFS_NAME_LEN;
2332 param->ddp_max_nlink = LDISKFS_LINK_MAX;
2333 param->ddp_symlink_max = sb->s_blocksize;
2334 param->ddp_mount_type = LDD_MT_LDISKFS;
2335 if (ldiskfs_has_feature_extents(sb))
2336 param->ddp_maxbytes = sb->s_maxbytes;
2338 param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
2340 * inode are statically allocated, so per-inode space consumption
2341 * is the space consumed by the directory entry
2343 param->ddp_inodespace = PER_OBJ_USAGE;
2345 * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
2346 * is 128MB) which is unlikely to be hit in real life. Report a smaller
2347 * maximum length to not under-count the actual number of extents
2348 * needed for writing a file if there are sub-optimal block allocations.
2350 param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
2351 /* worst-case extent insertion metadata overhead */
2352 param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
2353 param->ddp_mntopts = 0;
2354 if (test_opt(sb, XATTR_USER))
2355 param->ddp_mntopts |= MNTOPT_USERXATTR;
2356 if (test_opt(sb, POSIX_ACL))
2357 param->ddp_mntopts |= MNTOPT_ACL;
2360 * LOD might calculate the max stripe count based on max_ea_size,
2361 * so we need take account in the overhead as well,
2362 * xattr_header + magic + xattr_entry_head
2364 ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
2365 LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);
2367 #if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
2368 if (ldiskfs_has_feature_ea_inode(sb))
2369 param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
2373 param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;
2375 if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
2376 param->ddp_max_ea_size = OBD_MAX_EA_SIZE;
2379 * Preferred RPC size for efficient disk IO. 4MB shows good
2380 * all-around performance for ldiskfs, but use bigalloc chunk size
2381 * by default if larger.
2383 #if defined(LDISKFS_CLUSTER_SIZE)
2384 if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
2385 param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
2388 param->ddp_brw_size = DT_DEF_BRW_SIZE;
2390 param->ddp_t10_cksum_type = 0;
2392 unsigned short interval = blk_integrity_interval(bi);
2393 name = blk_integrity_name(bi);
2401 if (strncmp(name, "T10-DIF-TYPE",
2402 sizeof("T10-DIF-TYPE") - 1) == 0) {
2403 /* also skip "1/3-" at end */
2404 const int type_off = sizeof("T10-DIF-TYPE.");
2405 char type_number = name[type_off - 2];
2407 if (interval != 512 && interval != 4096) {
2408 CERROR("%s: unsupported T10PI sector size %u\n",
2409 d->od_svname, interval);
2410 } else if (type_number != '1' && type_number != '3') {
2411 CERROR("%s: unsupported T10PI type %s\n",
2412 d->od_svname, name);
2413 } else if (strcmp(name + type_off, "CRC") == 0) {
2414 d->od_t10_type = type_number == '1' ?
2415 OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
2416 param->ddp_t10_cksum_type = interval == 512 ?
2417 OBD_CKSUM_T10CRC512 :
2419 } else if (strcmp(name + type_off, "IP") == 0) {
2420 d->od_t10_type = type_number == '1' ?
2421 OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
2422 param->ddp_t10_cksum_type = interval == 512 ?
2423 OBD_CKSUM_T10IP512 :
2426 CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
2427 d->od_svname, name);
2431 CERROR("%s: unsupported T10PI type '%s'\n",
2432 d->od_svname, name);
2436 param->ddp_has_lseek_data_hole = true;
2439 static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
2441 return osd_sb(osd_dt_dev(d));
2445 * Concurrency: shouldn't matter.
2447 static int osd_sync(const struct lu_env *env, struct dt_device *d)
2450 struct super_block *s = osd_sb(osd_dt_dev(d));
2453 down_read(&s->s_umount);
2454 rc = s->s_op->sync_fs(s, 1);
2455 up_read(&s->s_umount);
2457 CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
2464 * Start commit for OSD device.
2466 * An implementation of dt_commit_async method for OSD device.
2467 * Asychronously starts underlayng fs sync and thereby a transaction
2470 * \param env environment
2471 * \param d dt device
2473 * \see dt_device_operations
2475 static int osd_commit_async(const struct lu_env *env,
2476 struct dt_device *d)
2478 struct super_block *s = osd_sb(osd_dt_dev(d));
2483 CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
2484 down_read(&s->s_umount);
2485 rc = s->s_op->sync_fs(s, 0);
2486 up_read(&s->s_umount);
2491 static int (*priv_security_file_alloc)(struct file *file);
2493 int osd_security_file_alloc(struct file *file)
2495 if (priv_security_file_alloc)
2496 return priv_security_file_alloc(file);
2501 * Concurrency: shouldn't matter.
2503 static int osd_ro(const struct lu_env *env, struct dt_device *d)
2505 struct super_block *sb = osd_sb(osd_dt_dev(d));
2506 struct block_device *dev = sb->s_bdev;
2507 int rc = -EOPNOTSUPP;
2511 CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
2512 osd_dt_dev(d)->od_svname, (long)dev, rc);
2518 * Note: we do not count into QUOTA here.
2519 * If we mount with --data_journal we may need more.
2521 const int osd_dto_credits_noquota[DTO_NR] = {
2524 * INDEX_EXTRA_TRANS_BLOCKS(8) +
2525 * SINGLEDATA_TRANS_BLOCKS(8)
2526 * XXX Note: maybe iam need more, since iam have more level than
2529 [DTO_INDEX_INSERT] = 16,
2532 * just modify a single entry, probably merge few within a block
2534 [DTO_INDEX_DELETE] = 1,
2538 [DTO_INDEX_UPDATE] = 16,
2540 * 4(inode, inode bits, groups, GDT)
2541 * notice: OI updates are counted separately with DTO_INDEX_INSERT
2543 [DTO_OBJECT_CREATE] = 4,
2545 * 4(inode, inode bits, groups, GDT)
2546 * notice: OI updates are counted separately with DTO_INDEX_DELETE
2548 [DTO_OBJECT_DELETE] = 4,
2550 * Attr set credits (inode)
2552 [DTO_ATTR_SET_BASE] = 1,
2554 * Xattr set. The same as xattr of EXT3.
2555 * DATA_TRANS_BLOCKS(14)
2556 * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
2557 * are also counted in. Do not know why?
2559 [DTO_XATTR_SET] = 14,
2561 * credits for inode change during write.
2563 [DTO_WRITE_BASE] = 3,
2565 * credits for single block write.
2567 [DTO_WRITE_BLOCK] = 14,
2569 * Attr set credits for chown.
2570 * This is extra credits for setattr, and it is null without quota
2572 [DTO_ATTR_SET_CHOWN] = 0
2575 static const struct dt_device_operations osd_dt_ops = {
2576 .dt_root_get = osd_root_get,
2577 .dt_statfs = osd_statfs,
2578 .dt_trans_create = osd_trans_create,
2579 .dt_trans_start = osd_trans_start,
2580 .dt_trans_stop = osd_trans_stop,
2581 .dt_trans_cb_add = osd_trans_cb_add,
2582 .dt_conf_get = osd_conf_get,
2583 .dt_mnt_sb_get = osd_mnt_sb_get,
2584 .dt_sync = osd_sync,
2586 .dt_commit_async = osd_commit_async,
2589 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2592 struct osd_object *obj = osd_dt_obj(dt);
2593 struct osd_thread_info *oti = osd_oti_get(env);
2595 LINVRNT(osd_invariant(obj));
2597 LASSERT(obj->oo_owner != env);
2598 down_read_nested(&obj->oo_sem, role);
2600 LASSERT(obj->oo_owner == NULL);
2604 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2607 struct osd_object *obj = osd_dt_obj(dt);
2608 struct osd_thread_info *oti = osd_oti_get(env);
2610 LINVRNT(osd_invariant(obj));
2612 LASSERT(obj->oo_owner != env);
2613 down_write_nested(&obj->oo_sem, role);
2615 LASSERT(obj->oo_owner == NULL);
2616 obj->oo_owner = env;
2620 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2622 struct osd_object *obj = osd_dt_obj(dt);
2623 struct osd_thread_info *oti = osd_oti_get(env);
2625 LINVRNT(osd_invariant(obj));
2627 LASSERT(oti->oti_r_locks > 0);
2629 up_read(&obj->oo_sem);
2632 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2634 struct osd_object *obj = osd_dt_obj(dt);
2635 struct osd_thread_info *oti = osd_oti_get(env);
2637 LINVRNT(osd_invariant(obj));
2639 LASSERT(obj->oo_owner == env);
2640 LASSERT(oti->oti_w_locks > 0);
2642 obj->oo_owner = NULL;
2643 up_write(&obj->oo_sem);
2646 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2648 struct osd_object *obj = osd_dt_obj(dt);
2650 LINVRNT(osd_invariant(obj));
2652 return obj->oo_owner == env;
2655 static void osd_inode_getattr(const struct lu_env *env,
2656 struct inode *inode, struct lu_attr *attr)
2658 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2659 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2660 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2661 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2663 attr->la_atime = inode->i_atime.tv_sec;
2664 attr->la_mtime = inode->i_mtime.tv_sec;
2665 attr->la_ctime = inode->i_ctime.tv_sec;
2666 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2667 attr->la_mode = inode->i_mode;
2668 attr->la_size = i_size_read(inode);
2669 attr->la_blocks = inode->i_blocks;
2670 attr->la_uid = i_uid_read(inode);
2671 attr->la_gid = i_gid_read(inode);
2672 attr->la_projid = i_projid_read(inode);
2673 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2674 attr->la_nlink = inode->i_nlink;
2675 attr->la_rdev = inode->i_rdev;
2676 attr->la_blksize = 1 << inode->i_blkbits;
2677 attr->la_blkbits = inode->i_blkbits;
2679 * Ext4 did not transfer inherit flags from raw inode
2680 * to inode flags, and ext4 internally test raw inode
2681 * @i_flags directly. Instead of patching ext4, we do it here.
2683 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2684 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2687 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2690 struct osd_object *obj = osd_dt_obj(dt);
2691 const struct dt_it_ops *iops;
2697 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2698 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2700 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2701 *count = obj->oo_dirent_count;
2705 /* directory not initialized yet */
2706 if (!dt->do_index_ops) {
2711 iops = &dt->do_index_ops->dio_it;
2712 it = iops->init(env, dt, LUDA_64BITHASH);
2714 RETURN(PTR_ERR(it));
2716 rc = iops->load(env, it, 0);
2718 if (rc == -ENODATA) {
2725 rc = iops->next(env, it);
2727 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2731 if (iops->key_size(env, it) == 0)
2737 obj->oo_dirent_count = *count;
2742 iops->fini(env, it);
2747 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2748 struct lu_attr *attr)
2750 struct osd_object *obj = osd_dt_obj(dt);
2753 if (unlikely(!dt_object_exists(dt)))
2755 if (unlikely(obj->oo_destroyed))
2758 LASSERT(!dt_object_remote(dt));
2759 LINVRNT(osd_invariant(obj));
2761 spin_lock(&obj->oo_guard);
2762 osd_inode_getattr(env, obj->oo_inode, attr);
2763 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2764 attr->la_valid |= LA_FLAGS;
2765 attr->la_flags |= LUSTRE_ORPHAN_FL;
2767 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2768 attr->la_valid |= LA_FLAGS;
2769 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2771 spin_unlock(&obj->oo_guard);
2773 if (S_ISDIR(obj->oo_inode->i_mode) &&
2774 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2775 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2780 static int osd_declare_attr_qid(const struct lu_env *env,
2781 struct osd_object *obj,
2782 struct osd_thandle *oh, long long bspace,
2783 qid_t old_id, qid_t new_id, bool enforce,
2784 unsigned int type, bool ignore_edquot)
2787 struct osd_thread_info *info = osd_oti_get(env);
2788 struct lquota_id_info *qi = &info->oti_qi;
2790 qi->lqi_type = type;
2791 /* inode accounting */
2792 qi->lqi_is_blk = false;
2794 /* one more inode for the new id ... */
2795 qi->lqi_id.qid_uid = new_id;
2797 /* Reserve credits for the new id */
2798 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2799 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2804 /* and one less inode for the current id */
2805 qi->lqi_id.qid_uid = old_id;
2807 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2808 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2813 /* block accounting */
2814 qi->lqi_is_blk = true;
2816 /* more blocks for the new id ... */
2817 qi->lqi_id.qid_uid = new_id;
2818 qi->lqi_space = bspace;
2820 * Credits for the new uid has been reserved, re-use "obj"
2821 * to save credit reservation.
2823 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2824 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2829 /* and finally less blocks for the current uid */
2830 qi->lqi_id.qid_uid = old_id;
2831 qi->lqi_space = -bspace;
2832 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2833 if (ignore_edquot && (rc == -EDQUOT || rc == -EINPROGRESS))
2839 static int osd_declare_attr_set(const struct lu_env *env,
2840 struct dt_object *dt,
2841 const struct lu_attr *attr,
2842 struct thandle *handle)
2844 struct osd_thandle *oh;
2845 struct osd_object *obj;
2854 LASSERT(dt != NULL);
2855 LASSERT(handle != NULL);
2857 obj = osd_dt_obj(dt);
2858 LASSERT(osd_invariant(obj));
2860 oh = container_of(handle, struct osd_thandle, ot_super);
2861 LASSERT(oh->ot_handle == NULL);
2863 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2864 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2866 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2867 osd_dto_credits_noquota[DTO_XATTR_SET]);
2869 if (attr == NULL || obj->oo_inode == NULL)
2872 bspace = obj->oo_inode->i_blocks << 9;
2873 bspace = toqb(bspace);
2876 * Changing ownership is always preformed by super user, it should not
2877 * fail with EDQUOT unless required explicitly.
2879 * We still need to call the osd_declare_qid() to calculate the journal
2880 * credits for updating quota accounting files and to trigger quota
2881 * space adjustment once the operation is completed.
2883 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2884 bool ignore_edquot = !(attr->la_flags & LUSTRE_SET_SYNC_FL);
2888 "%s: enforce quota on UID %u, GID %u (quota space is %lld)\n",
2889 osd_ino2name(obj->oo_inode), attr->la_uid,
2890 attr->la_gid, bspace);
2893 uid = i_uid_read(obj->oo_inode);
2894 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2895 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2896 attr->la_uid, enforce, USRQUOTA,
2901 gid = i_gid_read(obj->oo_inode);
2902 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
2903 attr->la_uid, gid, attr->la_gid);
2904 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2905 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
2906 attr->la_gid, enforce, GRPQUOTA,
2912 #ifdef HAVE_PROJECT_QUOTA
2913 if (attr->la_valid & LA_PROJID) {
2914 __u32 projid = i_projid_read(obj->oo_inode);
2916 enforce = (attr->la_valid & LA_PROJID) &&
2917 (attr->la_projid != projid);
2918 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2919 (qid_t)projid, (qid_t)attr->la_projid,
2920 enforce, PRJQUOTA, true);
2928 static int osd_inode_setattr(const struct lu_env *env,
2929 struct inode *inode, const struct lu_attr *attr)
2931 __u64 bits = attr->la_valid;
2933 /* Only allow set size for regular file */
2934 if (!S_ISREG(inode->i_mode))
2935 bits &= ~(LA_SIZE | LA_BLOCKS);
2940 if (bits & LA_ATIME)
2941 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2942 if (bits & LA_CTIME)
2943 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2944 if (bits & LA_MTIME)
2945 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2946 if (bits & LA_SIZE) {
2947 spin_lock(&inode->i_lock);
2948 LDISKFS_I(inode)->i_disksize = attr->la_size;
2949 i_size_write(inode, attr->la_size);
2950 spin_unlock(&inode->i_lock);
2954 * OSD should not change "i_blocks" which is used by quota.
2955 * "i_blocks" should be changed by ldiskfs only.
2958 inode->i_mode = (inode->i_mode & S_IFMT) |
2959 (attr->la_mode & ~S_IFMT);
2961 i_uid_write(inode, attr->la_uid);
2963 i_gid_write(inode, attr->la_gid);
2964 if (bits & LA_PROJID)
2965 i_projid_write(inode, attr->la_projid);
2966 if (bits & LA_NLINK)
2967 set_nlink(inode, attr->la_nlink);
2969 inode->i_rdev = attr->la_rdev;
2971 if (bits & LA_FLAGS) {
2972 /* always keep S_NOCMTIME */
2973 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2975 #if defined(S_ENCRYPTED)
2976 /* Always remove S_ENCRYPTED, because ldiskfs must not be
2977 * aware of encryption status. It is just stored into LMA
2978 * so that it can be forwared to client side.
2980 inode->i_flags &= ~S_ENCRYPTED;
2983 * Ext4 did not transfer inherit flags from
2984 * @inode->i_flags to raw inode i_flags when writing
2985 * flags, we do it explictly here.
2987 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
2988 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
2990 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
2995 #ifdef HAVE_PROJECT_QUOTA
2996 static int osd_transfer_project(struct inode *inode, __u32 projid,
2997 struct thandle *handle)
2999 struct super_block *sb = inode->i_sb;
3000 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3003 struct ldiskfs_iloc iloc;
3004 struct ldiskfs_inode *raw_inode;
3005 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3007 if (!ldiskfs_has_feature_project(sb)) {
3008 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3009 == LDISKFS_DEF_PROJID);
3010 if (projid != LDISKFS_DEF_PROJID)
3016 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3019 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3020 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3023 err = ldiskfs_get_inode_loc(inode, &iloc);
3027 raw_inode = ldiskfs_raw_inode(&iloc);
3028 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3029 struct osd_thandle *oh = container_of(handle,
3033 * try to expand inode size automatically.
3035 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3036 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3044 dquot_initialize(inode);
3045 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3046 if (transfer_to[PRJQUOTA]) {
3047 err = __dquot_transfer(inode, transfer_to);
3048 dqput(transfer_to[PRJQUOTA]);
3057 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3058 struct thandle *handle)
3062 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3063 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3067 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3068 inode->i_ino, i_uid_read(inode), attr->la_uid,
3069 i_gid_read(inode), attr->la_gid);
3071 dquot_initialize(inode);
3073 if (attr->la_valid & LA_UID)
3074 iattr.ia_valid |= ATTR_UID;
3075 if (attr->la_valid & LA_GID)
3076 iattr.ia_valid |= ATTR_GID;
3077 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3078 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3080 rc = dquot_transfer(inode, &iattr);
3082 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3083 osd_ino2name(inode), rc);
3088 /* Handle project id transfer here properly */
3089 if (attr->la_valid & LA_PROJID &&
3090 attr->la_projid != i_projid_read(inode)) {
3091 #ifdef HAVE_PROJECT_QUOTA
3092 rc = osd_transfer_project(inode, attr->la_projid, handle);
3097 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3098 osd_ino2name(inode), rc);
3105 static int osd_attr_set(const struct lu_env *env,
3106 struct dt_object *dt,
3107 const struct lu_attr *attr,
3108 struct thandle *handle)
3110 struct osd_object *obj = osd_dt_obj(dt);
3111 struct inode *inode;
3114 if (!dt_object_exists(dt))
3117 LASSERT(handle != NULL);
3118 LASSERT(!dt_object_remote(dt));
3119 LASSERT(osd_invariant(obj));
3121 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3123 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3124 !osd_obj2dev(obj)->od_is_ost) {
3125 struct osd_thread_info *oti = osd_oti_get(env);
3126 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3127 struct lu_fid *fid1 = &oti->oti_fid;
3128 struct osd_inode_id *id = &oti->oti_id;
3129 struct iam_path_descr *ipd;
3130 struct iam_container *bag;
3131 struct osd_thandle *oh;
3134 fid_cpu_to_be(fid1, fid0);
3135 memset(id, 1, sizeof(*id));
3136 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3137 fid0)->oi_dir.od_container;
3138 ipd = osd_idx_ipd_get(env, bag);
3139 if (unlikely(ipd == NULL))
3142 oh = container_of(handle, struct osd_thandle, ot_super);
3143 rc = iam_update(oh->ot_handle, bag,
3144 (const struct iam_key *)fid1,
3145 (const struct iam_rec *)id, ipd);
3146 osd_ipd_put(env, bag, ipd);
3147 return(rc > 0 ? 0 : rc);
3150 inode = obj->oo_inode;
3152 rc = osd_quota_transfer(inode, attr, handle);
3156 spin_lock(&obj->oo_guard);
3157 rc = osd_inode_setattr(env, inode, attr);
3158 spin_unlock(&obj->oo_guard);
3162 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3164 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3166 if (!(attr->la_valid & LA_FLAGS))
3169 /* Let's check if there are extra flags need to be set into LMA */
3170 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3171 struct osd_thread_info *info = osd_oti_get(env);
3172 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3174 LASSERT(!obj->oo_pfid_in_lma);
3176 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3177 &info->oti_ost_attrs);
3181 lma->lma_incompat |=
3182 lustre_to_lma_flags(attr->la_flags);
3183 lustre_lma_swab(lma);
3185 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3187 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3188 lma, sizeof(*lma), XATTR_REPLACE);
3190 struct osd_device *osd = osd_obj2dev(obj);
3192 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3193 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3194 lma->lma_incompat, rc);
3197 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3199 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3206 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3207 struct osd_object *obj,
3208 const char *name, const int namelen)
3210 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3213 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3214 umode_t mode, struct dt_allocation_hint *hint,
3215 struct thandle *th, struct lu_attr *attr)
3218 struct osd_device *osd = osd_obj2dev(obj);
3219 struct osd_thandle *oth;
3220 struct dt_object *parent = NULL;
3221 struct inode *inode;
3222 struct iattr iattr = {
3223 .ia_valid = ATTR_UID | ATTR_GID |
3224 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3225 .ia_ctime.tv_sec = attr->la_ctime,
3226 .ia_mtime.tv_sec = attr->la_mtime,
3227 .ia_atime.tv_sec = attr->la_atime,
3228 .ia_uid = GLOBAL_ROOT_UID,
3229 .ia_gid = GLOBAL_ROOT_GID,
3231 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3233 if (attr->la_valid & LA_UID)
3234 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3235 if (attr->la_valid & LA_GID)
3236 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3238 LINVRNT(osd_invariant(obj));
3239 LASSERT(obj->oo_inode == NULL);
3240 LASSERT(obj->oo_hl_head == NULL);
3242 if (S_ISDIR(mode) && ldiskfs_pdo) {
3244 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3245 if (obj->oo_hl_head == NULL)
3249 oth = container_of(th, struct osd_thandle, ot_super);
3250 LASSERT(oth->ot_handle->h_transaction != NULL);
3252 if (hint != NULL && hint->dah_parent != NULL &&
3253 !dt_object_remote(hint->dah_parent))
3254 parent = hint->dah_parent;
3256 /* if a time component is not valid set it to UTIME_OMIT */
3257 if (!(attr->la_valid & LA_CTIME))
3258 iattr.ia_ctime = omit;
3259 if (!(attr->la_valid & LA_MTIME))
3260 iattr.ia_mtime = omit;
3261 if (!(attr->la_valid & LA_ATIME))
3262 iattr.ia_atime = omit;
3264 inode = ldiskfs_create_inode(oth->ot_handle,
3265 parent ? osd_dt_obj(parent)->oo_inode :
3266 osd_sb(osd)->s_root->d_inode,
3268 if (!IS_ERR(inode)) {
3269 /* Do not update file c/mtime in ldiskfs. */
3270 inode->i_flags |= S_NOCMTIME;
3273 * For new created object, it must be consistent,
3274 * and it is unnecessary to scrub against it.
3276 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3278 obj->oo_inode = inode;
3281 if (obj->oo_hl_head != NULL) {
3282 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3283 obj->oo_hl_head = NULL;
3285 result = PTR_ERR(inode);
3287 LINVRNT(osd_invariant(obj));
3295 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3296 struct lu_attr *attr,
3297 struct dt_allocation_hint *hint,
3298 struct dt_object_format *dof,
3302 struct osd_thandle *oth;
3303 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3305 LASSERT(S_ISDIR(attr->la_mode));
3307 oth = container_of(th, struct osd_thandle, ot_super);
3308 LASSERT(oth->ot_handle->h_transaction != NULL);
3309 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3310 obj->oo_dirent_count = 0;
3311 result = osd_mkfile(info, obj, mode, hint, th, attr);
3316 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3317 struct lu_attr *attr,
3318 struct dt_allocation_hint *hint,
3319 struct dt_object_format *dof,
3323 struct osd_thandle *oth;
3324 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3326 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3328 LASSERT(S_ISREG(attr->la_mode));
3330 oth = container_of(th, struct osd_thandle, ot_super);
3331 LASSERT(oth->ot_handle->h_transaction != NULL);
3333 result = osd_mkfile(info, obj, mode, hint, th, attr);
3335 LASSERT(obj->oo_inode != NULL);
3336 if (feat->dif_flags & DT_IND_VARKEY)
3337 result = iam_lvar_create(obj->oo_inode,
3338 feat->dif_keysize_max,
3340 feat->dif_recsize_max,
3343 result = iam_lfix_create(obj->oo_inode,
3344 feat->dif_keysize_max,
3346 feat->dif_recsize_max,
3352 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3353 struct lu_attr *attr,
3354 struct dt_allocation_hint *hint,
3355 struct dt_object_format *dof,
3358 LASSERT(S_ISREG(attr->la_mode));
3359 return osd_mkfile(info, obj, (attr->la_mode &
3360 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3364 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3365 struct lu_attr *attr,
3366 struct dt_allocation_hint *hint,
3367 struct dt_object_format *dof,
3370 LASSERT(S_ISLNK(attr->la_mode));
3371 return osd_mkfile(info, obj, (attr->la_mode &
3372 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3376 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3377 struct lu_attr *attr,
3378 struct dt_allocation_hint *hint,
3379 struct dt_object_format *dof,
3382 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3385 LINVRNT(osd_invariant(obj));
3386 LASSERT(obj->oo_inode == NULL);
3387 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3388 S_ISFIFO(mode) || S_ISSOCK(mode));
3390 result = osd_mkfile(info, obj, mode, hint, th, attr);
3392 LASSERT(obj->oo_inode != NULL);
3394 * This inode should be marked dirty for i_rdev. Currently
3395 * that is done in the osd_attr_init().
3397 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3400 LINVRNT(osd_invariant(obj));
3404 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3406 struct dt_allocation_hint *hint,
3407 struct dt_object_format *dof,
3410 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3412 osd_obj_type_f result;
3428 result = osd_mk_index;
3438 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3439 struct dt_object *parent, struct dt_object *child,
3444 ah->dah_parent = parent;
3445 ah->dah_mode = child_mode;
3447 if (parent != NULL && !dt_object_remote(parent)) {
3448 /* will help to find FID->ino at dt_insert("..") */
3449 struct osd_object *pobj = osd_dt_obj(parent);
3451 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3455 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3456 struct lu_attr *attr, struct dt_object_format *dof,
3457 struct thandle *handle)
3459 struct inode *inode = obj->oo_inode;
3460 __u64 valid = attr->la_valid;
3463 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3465 if (dof->dof_type != DFT_NODE)
3466 attr->la_valid &= ~LA_RDEV;
3467 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3468 attr->la_valid &= ~LA_ATIME;
3469 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3470 attr->la_valid &= ~LA_CTIME;
3471 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3472 attr->la_valid &= ~LA_MTIME;
3474 result = osd_quota_transfer(inode, attr, handle);
3478 if (attr->la_valid != 0) {
3479 result = osd_inode_setattr(info->oti_env, inode, attr);
3481 * The osd_inode_setattr() should always succeed here. The
3482 * only error that could be returned is EDQUOT when we are
3483 * trying to change the UID or GID of the inode. However, this
3484 * should not happen since quota enforcement is no longer
3485 * enabled on ldiskfs (lquota takes care of it).
3487 LASSERTF(result == 0, "%d\n", result);
3488 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3491 attr->la_valid = valid;
3495 * Helper function for osd_create()
3497 * \retval 0, on success
3499 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3500 struct lu_attr *attr, struct dt_allocation_hint *hint,
3501 struct dt_object_format *dof, struct thandle *th)
3506 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3508 /* we drop umask so that permissions we pass are not affected */
3509 umask = current->fs->umask;
3510 current->fs->umask = 0;
3512 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3514 if (likely(obj->oo_inode != NULL)) {
3515 LASSERT(obj->oo_inode->i_state & I_NEW);
3518 * Unlock the inode before attr initialization to avoid
3519 * unnecessary dqget operations. LU-6378
3521 unlock_new_inode(obj->oo_inode);
3524 if (likely(result == 0)) {
3525 osd_attr_init(info, obj, attr, dof, th);
3526 osd_object_init0(obj);
3529 /* restore previous umask value */
3530 current->fs->umask = umask;
3532 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3538 * Helper function for osd_create()
3540 * \retval 0, on success
3542 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3543 const struct lu_fid *fid, struct thandle *th)
3545 struct osd_thread_info *info = osd_oti_get(env);
3546 struct osd_inode_id *id = &info->oti_id;
3547 struct osd_device *osd = osd_obj2dev(obj);
3548 struct osd_thandle *oh;
3551 LASSERT(obj->oo_inode != NULL);
3553 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3556 oh = container_of(th, struct osd_thandle, ot_super);
3557 LASSERT(oh->ot_handle);
3558 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3560 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3561 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3562 OI_CHECK_FLD, NULL);
3563 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3564 struct lu_fid next_fid = *fid;
3566 /* insert next object in advance, and map to the same inode */
3568 if (next_fid.f_oid != 0) {
3569 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3570 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3571 OI_CHECK_FLD, NULL);
3572 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3576 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3581 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3582 u64 seq, struct lu_seq_range *range)
3584 struct seq_server_site *ss = osd_seq_site(osd);
3586 if (fid_seq_is_idif(seq)) {
3587 fld_range_set_ost(range);
3588 range->lsr_index = idif_ost_idx(seq);
3592 if (!fid_seq_in_fldb(seq)) {
3593 fld_range_set_mdt(range);
3596 * FIXME: If ss is NULL, it suppose not get lsr_index
3599 range->lsr_index = ss->ss_node_id;
3603 LASSERT(ss != NULL);
3604 fld_range_set_any(range);
3605 /* OSD will only do local fld lookup */
3606 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3609 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3610 struct lu_attr *attr,
3611 struct dt_allocation_hint *hint,
3612 struct dt_object_format *dof,
3613 struct thandle *handle)
3615 struct osd_thandle *oh;
3616 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3622 LASSERT(handle != NULL);
3624 oh = container_of(handle, struct osd_thandle, ot_super);
3625 LASSERT(oh->ot_handle == NULL);
3628 * EA object consumes more credits than regular object: osd_mk_index
3629 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3630 * leaf_node, could involves the block, block bitmap, groups, GDT
3631 * change for each block, so add 4 * 2 credits in that case.
3633 * The default ACL initialization may consume an additional 16 blocks
3635 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3636 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3639 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3640 * credits for possible default ACL creation in new inode
3642 if (hint && hint->dah_acl_len)
3643 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3644 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3645 sb->s_blocksize_bits);
3647 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3650 * Reuse idle OI block may cause additional one OI block
3653 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3654 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3655 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3656 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3657 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3659 /* will help to find FID->ino mapping at dt_insert() */
3660 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3668 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3669 attr->la_projid, 1, oh, osd_dt_obj(dt),
3670 NULL, OSD_QID_INODE);
3678 * Called to destroy on-disk representation of the object
3680 * Concurrency: must be locked
3682 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3685 struct osd_object *obj = osd_dt_obj(dt);
3686 struct inode *inode = obj->oo_inode;
3687 struct osd_thandle *oh;
3695 oh = container_of(th, struct osd_thandle, ot_super);
3696 LASSERT(oh->ot_handle == NULL);
3698 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3699 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3701 /* For removing agent entry */
3702 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3703 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3706 * Recycle idle OI leaf may cause additional three OI blocks
3709 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3710 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3711 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3712 /* one less inode */
3713 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3714 i_projid_read(inode), -1, oh, obj, NULL,
3718 /* data to be truncated */
3719 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3720 i_projid_read(inode), 0, oh, obj, NULL,
3726 * will help to find FID->ino when this object is being
3729 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3734 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3737 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3738 struct osd_object *obj = osd_dt_obj(dt);
3739 struct inode *inode = obj->oo_inode;
3740 struct osd_device *osd = osd_obj2dev(obj);
3741 struct osd_thandle *oh;
3746 oh = container_of(th, struct osd_thandle, ot_super);
3747 LASSERT(oh->ot_handle);
3749 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3751 if (unlikely(fid_is_acct(fid)))
3754 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3755 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3757 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3758 osd_name(osd), PFID(fid), result);
3761 if (S_ISDIR(inode->i_mode)) {
3762 if (inode->i_nlink > 2)
3763 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3764 osd_name(osd), PFID(fid), inode->i_ino,
3767 spin_lock(&obj->oo_guard);
3769 spin_unlock(&obj->oo_guard);
3770 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3773 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3775 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3777 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3778 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3779 oh->ot_handle, OI_CHECK_FLD);
3781 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3782 /* XXX: add to ext3 orphan list */
3783 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3785 /* not needed in the cache anymore */
3786 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3787 obj->oo_destroyed = 1;
3793 * Put the fid into lustre_mdt_attrs, and then place the structure
3794 * inode's ea. This fid should not be altered during the life time
3797 * \retval +ve, on success
3798 * \retval -ve, on error
3800 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3802 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3803 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3805 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3806 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3811 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3814 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3817 lustre_loa_init(loa, fid, compat, incompat);
3818 lustre_loa_swab(loa, false);
3821 * For the OST device with 256 bytes inode size by default,
3822 * the PFID EA will be stored together with LMA EA to avoid
3823 * performance trouble. Otherwise the PFID EA can be stored
3824 * independently. LU-8998
3826 if ((compat & LMAC_FID_ON_OST) &&
3827 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3828 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3829 sizeof(*loa), XATTR_CREATE);
3831 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3832 sizeof(*lma), XATTR_CREATE);
3834 * LMA may already exist, but we need to check that all the
3835 * desired compat/incompat flags have been added.
3837 if (unlikely(rc == -EEXIST)) {
3838 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3839 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3843 if (rc < sizeof(*lma))
3846 lustre_loa_swab(loa, true);
3847 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3848 ((compat == 0 && incompat == 0) ||
3849 (!(~lma->lma_compat & compat) &&
3850 !(~lma->lma_incompat & incompat))))
3853 lma->lma_self_fid = *fid;
3854 lma->lma_compat |= compat;
3855 lma->lma_incompat |= incompat;
3856 if (rc == sizeof(*lma)) {
3857 lustre_lma_swab(lma);
3858 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3859 sizeof(*lma), XATTR_REPLACE);
3861 lustre_loa_swab(loa, false);
3862 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3863 sizeof(*loa), XATTR_REPLACE);
3871 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3872 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3873 * To have compatilibility with 1.8 ldiskfs driver we need to have
3874 * magic number at start of fid data.
3875 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3878 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3879 const struct lu_fid *fid)
3881 if (!fid_is_namespace_visible(fid) ||
3882 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3883 param->edp_magic = 0;
3887 param->edp_magic = LDISKFS_LUFID_MAGIC;
3888 param->edp_len = sizeof(struct lu_fid) + 1;
3889 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3893 * Try to read the fid from inode ea into dt_rec.
3895 * \param fid object fid.
3897 * \retval 0 on success
3899 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3900 __u32 ino, struct lu_fid *fid,
3901 struct osd_inode_id *id)
3903 struct osd_thread_info *info = osd_oti_get(env);
3904 struct inode *inode;
3908 osd_id_gen(id, ino, OSD_OII_NOGEN);
3909 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3911 RETURN(PTR_ERR(inode));
3917 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3919 struct inode *parent_dir,
3920 const struct lu_fid *dot_fid,
3921 const struct lu_fid *dot_dot_fid,
3922 struct osd_thandle *oth)
3924 struct ldiskfs_dentry_param *dot_ldp;
3925 struct ldiskfs_dentry_param *dot_dot_ldp;
3926 __u32 saved_nlink = dir->i_nlink;
3929 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
3932 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3933 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3935 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3936 dot_ldp->edp_magic = 0;
3938 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3939 dir, dot_ldp, dot_dot_ldp);
3941 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3942 * the subseqent ref_add() will increase the dir->i_nlink
3943 * as 3. That is incorrect for new created directory.
3945 * It looks like hack, because we want to make the OSD API
3946 * to be order-independent for new created directory object
3947 * between dt_insert(..) and ref_add() operations.
3949 * Here, we only restore the in-RAM dir-inode's nlink attr,
3950 * becuase if the nlink attr is not 2, then there will be
3951 * ref_add() called following the dt_insert(..), such call
3952 * will make both the in-RAM and on-disk dir-inode's nlink
3953 * attr to be set as 2. LU-7447
3955 set_nlink(dir, saved_nlink);
3960 * Create an local agent inode for remote entry
3962 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3963 struct osd_device *osd,
3964 struct osd_object *pobj,
3965 const struct lu_fid *fid,
3969 struct osd_thread_info *info = osd_oti_get(env);
3970 struct inode *local;
3971 struct osd_thandle *oh;
3972 struct iattr iattr = {
3973 .ia_valid = ATTR_UID | ATTR_GID |
3974 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3975 .ia_ctime.tv_nsec = UTIME_OMIT,
3976 .ia_mtime.tv_nsec = UTIME_OMIT,
3977 .ia_atime.tv_nsec = UTIME_OMIT,
3978 .ia_uid = GLOBAL_ROOT_UID,
3979 .ia_gid = GLOBAL_ROOT_GID,
3986 oh = container_of(th, struct osd_thandle, ot_super);
3987 LASSERT(oh->ot_handle->h_transaction != NULL);
3989 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
3991 if (IS_ERR(local)) {
3992 CERROR("%s: create local error %d\n", osd_name(osd),
3993 (int)PTR_ERR(local));
3998 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
3999 * correct gid on remote file, not agent here
4001 local->i_gid = current_fsgid();
4002 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4004 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4005 * That gives e2fsck something to look at and be happy, and allows
4006 * debugging if we need to determine where this symlink came from.
4008 if (S_ISLNK(type)) {
4009 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4010 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4011 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4013 i_size_write(local, rc);
4014 LDISKFS_I(local)->i_disksize = rc;
4016 unlock_new_inode(local);
4018 /* Agent inode should not have project ID */
4019 #ifdef HAVE_PROJECT_QUOTA
4020 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4021 i_projid_read(pobj->oo_inode) != 0) {
4022 rc = osd_transfer_project(local, 0, th);
4024 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4025 osd_ino2name(local), rc);
4026 RETURN(ERR_PTR(rc));
4030 /* Set special LMA flag for local agent inode */
4031 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4033 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4034 osd_name(osd), PFID(fid), rc);
4035 RETURN(ERR_PTR(rc));
4041 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
4042 lu_object_fid(&pobj->oo_dt.do_lu),
4045 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4046 osd_name(osd), PFID(fid), rc);
4047 RETURN(ERR_PTR(rc));
4054 * when direntry is deleted, we have to take care of possible agent inode
4055 * referenced by that. unfortunately we can't do this at that point:
4056 * iget() within a running transaction leads to deadlock and we better do
4057 * not call that every delete declaration to save performance. so we put
4058 * a potention agent inode on a list and process that once the transaction
4059 * is over. Notice it's not any worse in terms of real orphans as regular
4060 * object destroy doesn't put inodes on the on-disk orphan list. this should
4061 * be addressed separately
4063 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4064 struct osd_thandle *oh,
4067 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4068 struct osd_obj_orphan *oor;
4075 oor->oor_env = (struct lu_env *)env;
4076 spin_lock(&osd->od_osfs_lock);
4077 list_add(&oor->oor_list, &osd->od_orphan_list);
4078 spin_unlock(&osd->od_osfs_lock);
4080 oh->ot_remove_agents = 1;
4086 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4087 struct osd_device *osd)
4089 struct osd_thread_info *info = osd_oti_get(env);
4090 struct osd_obj_orphan *oor, *tmp;
4091 struct osd_inode_id id;
4093 struct inode *inode;
4098 spin_lock(&osd->od_osfs_lock);
4099 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4100 if (oor->oor_env == env)
4101 list_move(&oor->oor_list, &list);
4103 spin_unlock(&osd->od_osfs_lock);
4105 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4109 list_del(&oor->oor_list);
4112 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4113 inode = osd_iget_fid(info, osd, &id, &fid);
4117 if (!osd_remote_fid(env, osd, &fid)) {
4122 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4124 mark_inode_dirty(inode);
4125 ldiskfs_journal_stop(jh);
4133 * OSD layer object create function for OST objects (b=11826).
4135 * The FID is inserted into inode xattr here.
4137 * \retval 0, on success
4138 * \retval -ve, on error
4140 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4141 struct lu_attr *attr, struct dt_allocation_hint *hint,
4142 struct dt_object_format *dof, struct thandle *th)
4144 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4145 struct osd_object *obj = osd_dt_obj(dt);
4146 struct osd_thread_info *info = osd_oti_get(env);
4147 int result, on_ost = 0;
4151 if (dt_object_exists(dt))
4154 LINVRNT(osd_invariant(obj));
4155 LASSERT(!dt_object_remote(dt));
4156 LASSERT(osd_is_write_locked(env, obj));
4157 LASSERT(th != NULL);
4159 if (unlikely(fid_is_acct(fid)))
4161 * Quota files can't be created from the kernel any more,
4162 * 'tune2fs -O quota' will take care of creating them
4166 result = __osd_create(info, obj, attr, hint, dof, th);
4168 if (fid_is_idif(fid) &&
4169 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4170 struct lu_fid *tfid = &info->oti_fid;
4171 struct ost_id *oi = &info->oti_ostid;
4173 fid_to_ostid(fid, oi);
4174 ostid_to_fid(tfid, oi, 0);
4176 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4177 LMAC_FID_ON_OST, 0);
4179 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4181 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4182 on_ost ? LMAC_FID_ON_OST : 0,
4185 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4186 obj->oo_dt.do_body_ops = &osd_body_ops;
4189 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4190 struct inode *inode = obj->oo_inode;
4192 result = __osd_oi_insert(env, obj, fid, th);
4193 if (result && inode) {
4194 spin_lock(&obj->oo_guard);
4196 spin_unlock(&obj->oo_guard);
4197 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4198 ldiskfs_set_inode_state(inode,
4199 LDISKFS_STATE_LUSTRE_DESTROY);
4201 obj->oo_inode = NULL;
4206 * a small optimization - dt_insert() isn't usually applied
4207 * to OST objects, so we don't need to cache OI mapping for
4210 if (result == 0 && on_ost == 0) {
4211 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4213 result = osd_idc_find_and_init(env, osd, obj);
4214 LASSERT(result == 0);
4217 LASSERT(ergo(result == 0,
4218 dt_object_exists(dt) && !dt_object_remote(dt)));
4219 LINVRNT(osd_invariant(obj));
4223 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4224 struct thandle *handle)
4226 struct osd_thandle *oh;
4229 /* it's possible that object doesn't exist yet */
4230 LASSERT(handle != NULL);
4232 oh = container_of(handle, struct osd_thandle, ot_super);
4233 LASSERT(oh->ot_handle == NULL);
4235 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4236 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4238 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4245 * Concurrency: @dt is write locked.
4247 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4250 struct osd_object *obj = osd_dt_obj(dt);
4251 struct inode *inode = obj->oo_inode;
4252 struct osd_thandle *oh;
4255 if (!dt_object_exists(dt) || obj->oo_destroyed)
4258 LINVRNT(osd_invariant(obj));
4259 LASSERT(!dt_object_remote(dt));
4260 LASSERT(osd_is_write_locked(env, obj));
4261 LASSERT(th != NULL);
4263 oh = container_of(th, struct osd_thandle, ot_super);
4264 LASSERT(oh->ot_handle != NULL);
4266 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4268 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4269 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4271 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4272 * (65000) subdirectories by storing "1" in i_nlink if the link count
4273 * would otherwise overflow. Directory tranversal tools understand
4274 * that (st_nlink == 1) indicates that the filesystem dose not track
4275 * hard links count on the directory, and will not abort subdirectory
4276 * scanning early once (st_nlink - 2) subdirs have been found.
4278 * This also has to properly handle the case of inodes with nlink == 0
4279 * in case they are being linked into the PENDING directory
4281 spin_lock(&obj->oo_guard);
4282 if (unlikely(inode->i_nlink == 0))
4283 /* inc_nlink from 0 may cause WARN_ON */
4284 set_nlink(inode, 1);
4286 ldiskfs_inc_count(oh->ot_handle, inode);
4287 if (!S_ISDIR(inode->i_mode))
4288 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4290 spin_unlock(&obj->oo_guard);
4292 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4293 LINVRNT(osd_invariant(obj));
4295 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4300 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4301 struct thandle *handle)
4303 struct osd_thandle *oh;
4305 if (!dt_object_exists(dt))
4308 LASSERT(!dt_object_remote(dt));
4309 LASSERT(handle != NULL);
4311 oh = container_of(handle, struct osd_thandle, ot_super);
4312 LASSERT(oh->ot_handle == NULL);
4314 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4315 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4321 * Concurrency: @dt is write locked.
4323 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4326 struct osd_object *obj = osd_dt_obj(dt);
4327 struct inode *inode = obj->oo_inode;
4328 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4329 struct osd_thandle *oh;
4331 if (!dt_object_exists(dt))
4334 LINVRNT(osd_invariant(obj));
4335 LASSERT(!dt_object_remote(dt));
4336 LASSERT(osd_is_write_locked(env, obj));
4337 LASSERT(th != NULL);
4339 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4342 oh = container_of(th, struct osd_thandle, ot_super);
4343 LASSERT(oh->ot_handle != NULL);
4345 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4347 spin_lock(&obj->oo_guard);
4349 * That can be result of upgrade from old Lustre version and
4350 * applied only to local files. Just skip this ref_del call.
4351 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4353 if (inode->i_nlink == 0) {
4354 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4355 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4356 ", maybe an upgraded file? (LU-3915)\n",
4357 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4358 spin_unlock(&obj->oo_guard);
4362 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4363 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4365 ldiskfs_dec_count(oh->ot_handle, inode);
4366 spin_unlock(&obj->oo_guard);
4368 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4369 LINVRNT(osd_invariant(obj));
4371 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4377 * Concurrency: @dt is read locked.
4379 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4380 struct lu_buf *buf, const char *name)
4382 struct osd_object *obj = osd_dt_obj(dt);
4383 struct inode *inode = obj->oo_inode;
4384 struct osd_thread_info *info = osd_oti_get(env);
4385 struct dentry *dentry = &info->oti_obj_dentry;
4386 bool cache_xattr = false;
4391 /* version get is not real XATTR but uses xattr API */
4392 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4393 dt_obj_version_t *ver = buf->lb_buf;
4396 * for version we are just using xattr API but change inode
4399 if (buf->lb_len == 0)
4400 return sizeof(dt_obj_version_t);
4402 if (buf->lb_len < sizeof(dt_obj_version_t))
4405 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4406 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4408 *ver = LDISKFS_I(inode)->i_fs_version;
4410 return sizeof(dt_obj_version_t);
4413 if (!dt_object_exists(dt))
4416 LASSERT(!dt_object_remote(dt));
4417 LASSERT(inode->i_op != NULL);
4418 #ifdef HAVE_IOP_XATTR
4419 LASSERT(inode->i_op->getxattr != NULL);
4422 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4423 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4427 rc = osd_oxc_get(obj, name, buf);
4432 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4433 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4434 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4435 struct filter_fid *ff;
4436 struct ost_layout *ol;
4438 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4442 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4445 if (buf->lb_len == 0 || !buf->lb_buf)
4448 if (buf->lb_len < rc)
4452 ol = &ff->ff_layout;
4453 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4454 PFID_STRIPE_IDX_BITS);
4455 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4456 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4457 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4458 if (lma->lma_compat & LMAC_COMP_INFO) {
4459 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4460 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4461 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4463 ol->ol_comp_start = 0;
4464 ol->ol_comp_end = 0;
4468 rc = __osd_xattr_get(inode, dentry, name,
4469 buf->lb_buf, buf->lb_len);
4473 if (rc == -ENOENT || rc == -ENODATA)
4474 osd_oxc_add(obj, name, NULL, 0);
4475 else if (rc > 0 && buf->lb_buf != NULL)
4476 osd_oxc_add(obj, name, buf->lb_buf, rc);
4482 static int osd_declare_xattr_set(const struct lu_env *env,
4483 struct dt_object *dt,
4484 const struct lu_buf *buf, const char *name,
4485 int fl, struct thandle *handle)
4487 struct osd_thandle *oh;
4489 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4491 LASSERT(handle != NULL);
4493 oh = container_of(handle, struct osd_thandle, ot_super);
4494 LASSERT(oh->ot_handle == NULL);
4496 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4498 * For non-upgrading case, the LMA is set first and
4499 * usually fit inode. But for upgrade case, the LMA
4500 * may be in another separated EA block.
4502 if (dt_object_exists(dt)) {
4503 if (fl == LU_XATTR_REPLACE)
4508 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4510 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4511 /* We may need to delete the old PFID EA. */
4512 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4513 if (fl == LU_XATTR_REPLACE)
4519 * If some name entry resides on remote MDT, then will create
4520 * agent entry under remote parent. On the other hand, if the
4521 * remote entry will be removed, then related agent entry may
4522 * need to be removed from the remote parent. So there may be
4523 * kinds of cases, let's declare enough credits. The credits
4524 * for create agent entry is enough for remove case.
4526 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4527 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4528 if (dt_object_exists(dt))
4529 credits += 1; /* For updating LMA */
4533 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4538 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4540 * learn xattr size from osd_xattr_get if
4541 * attribute has not been read yet
4543 buflen = __osd_xattr_get(
4544 osd_dt_obj(dt)->oo_inode,
4545 &osd_oti_get(env)->oti_obj_dentry,
4550 buflen = buf->lb_len;
4553 if (buflen > sb->s_blocksize) {
4554 credits += osd_calc_bkmap_credits(
4556 (buflen + sb->s_blocksize - 1) >>
4557 sb->s_blocksize_bits);
4561 * xattr set may involve inode quota change, reserve credits for
4562 * dquot_initialize()
4564 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4567 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4572 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4573 const struct lu_buf *buf, int fl,
4574 struct thandle *handle)
4576 struct osd_thread_info *info = osd_oti_get(env);
4577 struct dentry *dentry = &info->oti_obj_dentry;
4578 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4579 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4580 struct inode *inode = obj->oo_inode;
4581 struct filter_fid *ff = buf->lb_buf;
4582 struct ost_layout *ol = &ff->ff_layout;
4583 int flags = XATTR_REPLACE;
4588 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4591 rc = osd_get_lma(info, inode, dentry, loa);
4592 if (rc == -ENODATA) {
4593 /* Usually for upgarding from old device */
4594 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4595 LMAC_FID_ON_OST, 0);
4596 flags = XATTR_CREATE;
4601 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4602 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4605 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4606 /* Separate PFID EA from LMA */
4607 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4608 lustre_lma_swab(lma);
4609 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4610 sizeof(*lma), XATTR_REPLACE);
4612 obj->oo_pfid_in_lma = 0;
4613 rc = LU_XATTR_CREATE;
4619 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4623 * Old client does not send stripe information,
4624 * then store the PFID EA on disk separatedly.
4626 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4627 ol->ol_stripe_size == 0))
4630 /* Remove old PFID EA entry firstly. */
4631 dquot_initialize(inode);
4632 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4633 if (rc == -ENODATA) {
4634 /* XATTR_NAME_FID is already absent */
4641 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4642 if (likely(ol->ol_stripe_size != 0)) {
4643 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4644 PFID_STRIPE_IDX_BITS;
4645 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4646 lma->lma_compat |= LMAC_STRIPE_INFO;
4647 if (ol->ol_comp_id != 0) {
4648 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4649 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4650 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4651 lma->lma_compat |= LMAC_COMP_INFO;
4655 lustre_loa_swab(loa, false);
4657 /* Store the PFID EA inside LMA. */
4658 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4661 obj->oo_pfid_in_lma = 1;
4667 * In DNE environment, the object (in spite of regular file or directory)
4668 * and its name entry may reside on different MDTs. Under such case, we will
4669 * create an agent entry on the MDT where the object resides. The agent entry
4670 * references the object locally, that makes the object to be visible to the
4671 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4672 * such as 'tar' can handle the object properly.
4674 * We handle the agent entry during set linkEA that is the common interface
4675 * for both regular file and directroy, can handle kinds of cases, such as
4676 * create/link/unlink/rename, and so on.
4678 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4680 * XXX: There are two known issues:
4681 * 1. For one object, we will create at most one agent entry even if there
4682 * may be more than one cross-MDTs hard links on the object. So the local
4683 * e2fsck may claim that the object's nlink is larger than the name entries
4684 * that reference such inode. And in further, the e2fsck will fix the nlink
4685 * attribute to match the local references. Then it will cause the object's
4686 * nlink attribute to be inconsistent with the global references. it is bad
4687 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4688 * other hand, the global namespace LFSCK can repair the object's attribute
4689 * according to the linkEA.
4690 * 2. There may be too many hard links on the object as to its linkEA overflow,
4691 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4692 * case happened, then at this point, we do not know whether there are some
4693 * cross-MDTs reference. But there are local references, it guarantees that
4694 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4696 static int osd_xattr_handle_linkea(const struct lu_env *env,
4697 struct osd_device *osd,
4698 struct osd_object *obj,
4699 const struct lu_buf *buf,
4700 struct thandle *handle)
4702 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4703 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4704 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4705 struct lu_name tmpname;
4706 struct osd_thandle *oh;
4708 bool remote = false;
4712 oh = container_of(handle, struct osd_thandle, ot_super);
4713 LASSERT(oh->ot_handle != NULL);
4715 rc = linkea_init_with_rec(&ldata);
4717 linkea_first_entry(&ldata);
4718 while (ldata.ld_lee != NULL && !remote) {
4719 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4721 if (osd_remote_fid(env, osd, tfid) > 0)
4724 linkea_next_entry(&ldata);
4726 } else if (rc == -ENODATA) {
4732 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4733 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4735 CERROR("%s: failed to remove agent entry for "DFID
4736 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4737 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4738 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4740 CERROR("%s: failed to create agent entry for "DFID
4741 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4748 * Concurrency: @dt is write locked.
4750 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4751 const struct lu_buf *buf, const char *name, int fl,
4752 struct thandle *handle)
4754 struct osd_object *obj = osd_dt_obj(dt);
4755 struct osd_device *osd = osd_obj2dev(obj);
4756 struct inode *inode = obj->oo_inode;
4757 struct osd_thread_info *info = osd_oti_get(env);
4767 /* version set is not real XATTR */
4768 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4769 dt_obj_version_t *version = buf->lb_buf;
4772 * for version we are just using xattr API but change inode
4775 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4777 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4778 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4780 LDISKFS_I(inode)->i_fs_version = *version;
4782 * Version is set after all inode operations are finished,
4783 * so we should mark it dirty here
4785 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4790 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4791 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4794 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4797 * For the OST device with 256 bytes inode size by default,
4798 * the PFID EA will be stored together with LMA EA to avoid
4799 * performance trouble. Otherwise the PFID EA can be stored
4800 * independently. LU-8998
4802 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4803 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4804 LASSERT(buf->lb_buf);
4806 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4809 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4810 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4811 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4813 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4817 lma->lma_incompat |= LMAI_STRIPED;
4818 lustre_lma_swab(lma);
4819 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4820 sizeof(*lma), XATTR_REPLACE);
4823 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4824 LASSERT(!osd->od_is_ost);
4826 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4831 if (fl & LU_XATTR_REPLACE)
4832 fs_flags |= XATTR_REPLACE;
4834 if (fl & LU_XATTR_CREATE)
4835 fs_flags |= XATTR_CREATE;
4837 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4838 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4841 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4842 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4843 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4849 * Concurrency: @dt is read locked.
4851 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4852 const struct lu_buf *buf)
4854 struct osd_object *obj = osd_dt_obj(dt);
4855 struct inode *inode = obj->oo_inode;
4856 struct osd_thread_info *info = osd_oti_get(env);
4857 struct dentry *dentry = &info->oti_obj_dentry;
4859 if (!dt_object_exists(dt))
4862 LASSERT(!dt_object_remote(dt));
4863 LASSERT(inode->i_op != NULL);
4864 LASSERT(inode->i_op->listxattr != NULL);
4866 dentry->d_inode = inode;
4867 dentry->d_sb = inode->i_sb;
4868 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4871 static int osd_declare_xattr_del(const struct lu_env *env,
4872 struct dt_object *dt, const char *name,
4873 struct thandle *handle)
4875 struct osd_thandle *oh;
4876 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4878 LASSERT(!dt_object_remote(dt));
4879 LASSERT(handle != NULL);
4881 oh = container_of(handle, struct osd_thandle, ot_super);
4882 LASSERT(oh->ot_handle == NULL);
4884 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4885 osd_dto_credits_noquota[DTO_XATTR_SET]);
4887 * xattr del may involve inode quota change, reserve credits for
4888 * dquot_initialize()
4890 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4896 * Concurrency: @dt is write locked.
4898 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4899 const char *name, struct thandle *handle)
4901 struct osd_object *obj = osd_dt_obj(dt);
4902 struct inode *inode = obj->oo_inode;
4903 struct osd_thread_info *info = osd_oti_get(env);
4904 struct dentry *dentry = &info->oti_obj_dentry;
4907 if (!dt_object_exists(dt))
4910 LASSERT(!dt_object_remote(dt));
4911 LASSERT(inode->i_op != NULL);
4912 LASSERT(handle != NULL);
4913 #ifdef HAVE_IOP_XATTR
4914 LASSERT(inode->i_op->removexattr != NULL);
4917 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4919 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4920 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4922 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4923 &info->oti_ost_attrs);
4925 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4927 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4928 lustre_lma_swab(lma);
4929 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4930 sizeof(*lma), XATTR_REPLACE);
4932 obj->oo_pfid_in_lma = 0;
4935 dquot_initialize(inode);
4936 dentry->d_inode = inode;
4937 dentry->d_sb = inode->i_sb;
4938 rc = ll_vfs_removexattr(dentry, inode, name);
4941 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4944 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4945 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4946 osd_oxc_del(obj, name);
4951 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4952 __u64 start, __u64 end)
4954 struct osd_object *obj = osd_dt_obj(dt);
4955 struct inode *inode = obj->oo_inode;
4956 struct file *file = osd_quasi_file(env, inode);
4961 rc = vfs_fsync_range(file, start, end, 0);
4966 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4971 static bool osd_check_stale(struct dt_object *dt)
4979 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4980 const struct dt_index_features *feat)
4982 struct iam_descr *descr;
4984 if (osd_object_is_root(o))
4985 return feat == &dt_directory_features;
4987 LASSERT(o->oo_dir != NULL);
4989 descr = o->oo_dir->od_container.ic_descr;
4990 if (feat == &dt_directory_features) {
4991 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
4996 return feat->dif_keysize_min <= descr->id_key_size &&
4997 descr->id_key_size <= feat->dif_keysize_max &&
4998 feat->dif_recsize_min <= descr->id_rec_size &&
4999 descr->id_rec_size <= feat->dif_recsize_max &&
5000 !(feat->dif_flags & (DT_IND_VARKEY |
5001 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5002 ergo(feat->dif_flags & DT_IND_UPDATE,
5003 1 /* XXX check that object (and fs) is writable */);
5007 static int osd_iam_container_init(const struct lu_env *env,
5008 struct osd_object *obj,
5009 struct osd_directory *dir)
5011 struct iam_container *bag = &dir->od_container;
5014 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5018 result = iam_container_setup(bag);
5020 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5022 iam_container_fini(bag);
5029 * Concurrency: no external locking is necessary.
5031 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5032 const struct dt_index_features *feat)
5036 struct osd_object *obj = osd_dt_obj(dt);
5038 LINVRNT(osd_invariant(obj));
5040 if (osd_object_is_root(obj)) {
5041 dt->do_index_ops = &osd_index_ea_ops;
5043 } else if (feat == &dt_directory_features) {
5044 dt->do_index_ops = &osd_index_ea_ops;
5045 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5050 } else if (unlikely(feat == &dt_otable_features)) {
5051 dt->do_index_ops = &osd_otable_ops;
5053 } else if (unlikely(feat == &dt_acct_features)) {
5054 dt->do_index_ops = &osd_acct_index_ops;
5057 } else if (!osd_has_index(obj)) {
5058 struct osd_directory *dir;
5059 struct osd_device *osd = osd_obj2dev(obj);
5060 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5065 spin_lock(&obj->oo_guard);
5066 if (obj->oo_dir == NULL)
5070 * Concurrent thread allocated container data.
5073 spin_unlock(&obj->oo_guard);
5075 * Now, that we have container data, serialize its
5078 down_write(&obj->oo_ext_idx_sem);
5080 * recheck under lock.
5083 if (osd_has_index(obj)) {
5088 result = osd_iam_container_init(env, obj, obj->oo_dir);
5089 if (result || feat == &dt_lfsck_namespace_features ||
5090 feat == &dt_lfsck_layout_orphan_features ||
5091 feat == &dt_lfsck_layout_dangling_features)
5094 result = osd_index_register(osd, fid,
5095 feat->dif_keysize_max,
5096 feat->dif_recsize_max);
5098 CWARN("%s: failed to register index "
5100 osd_name(osd), PFID(fid), result);
5101 else if (result > 0)
5104 CDEBUG(D_LFSCK, "%s: index object "DFID
5105 " (%d/%d) registered\n",
5106 osd_name(osd), PFID(fid),
5107 (int)feat->dif_keysize_max,
5108 (int)feat->dif_recsize_max);
5111 up_write(&obj->oo_ext_idx_sem);
5119 if (result == 0 && skip_iam == 0) {
5120 if (!osd_iam_index_probe(env, obj, feat))
5123 LINVRNT(osd_invariant(obj));
5128 static int osd_otable_it_attr_get(const struct lu_env *env,
5129 struct dt_object *dt,
5130 struct lu_attr *attr)
5136 static const struct dt_object_operations osd_obj_ops = {
5137 .do_read_lock = osd_read_lock,
5138 .do_write_lock = osd_write_lock,
5139 .do_read_unlock = osd_read_unlock,
5140 .do_write_unlock = osd_write_unlock,
5141 .do_write_locked = osd_write_locked,
5142 .do_attr_get = osd_attr_get,
5143 .do_declare_attr_set = osd_declare_attr_set,
5144 .do_attr_set = osd_attr_set,
5145 .do_ah_init = osd_ah_init,
5146 .do_declare_create = osd_declare_create,
5147 .do_create = osd_create,
5148 .do_declare_destroy = osd_declare_destroy,
5149 .do_destroy = osd_destroy,
5150 .do_index_try = osd_index_try,
5151 .do_declare_ref_add = osd_declare_ref_add,
5152 .do_ref_add = osd_ref_add,
5153 .do_declare_ref_del = osd_declare_ref_del,
5154 .do_ref_del = osd_ref_del,
5155 .do_xattr_get = osd_xattr_get,
5156 .do_declare_xattr_set = osd_declare_xattr_set,
5157 .do_xattr_set = osd_xattr_set,
5158 .do_declare_xattr_del = osd_declare_xattr_del,
5159 .do_xattr_del = osd_xattr_del,
5160 .do_xattr_list = osd_xattr_list,
5161 .do_object_sync = osd_object_sync,
5162 .do_invalidate = osd_invalidate,
5163 .do_check_stale = osd_check_stale,
5166 static const struct dt_object_operations osd_obj_otable_it_ops = {
5167 .do_attr_get = osd_otable_it_attr_get,
5168 .do_index_try = osd_index_try,
5171 static int osd_index_declare_iam_delete(const struct lu_env *env,
5172 struct dt_object *dt,
5173 const struct dt_key *key,
5174 struct thandle *handle)
5176 struct osd_thandle *oh;
5178 oh = container_of(handle, struct osd_thandle, ot_super);
5179 LASSERT(oh->ot_handle == NULL);
5181 /* Recycle may cause additional three blocks to be changed. */
5182 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5183 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5189 * delete a (key, value) pair from index \a dt specified by \a key
5191 * \param dt osd index object
5192 * \param key key for index
5193 * \param rec record reference
5194 * \param handle transaction handler
5197 * \retval -ve failure
5199 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5200 const struct dt_key *key,
5201 struct thandle *handle)
5203 struct osd_thread_info *oti = osd_oti_get(env);
5204 struct osd_object *obj = osd_dt_obj(dt);
5205 struct osd_thandle *oh;
5206 struct iam_path_descr *ipd;
5207 struct iam_container *bag = &obj->oo_dir->od_container;
5212 if (!dt_object_exists(dt))
5215 LINVRNT(osd_invariant(obj));
5216 LASSERT(!dt_object_remote(dt));
5217 LASSERT(bag->ic_object == obj->oo_inode);
5218 LASSERT(handle != NULL);
5220 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5222 ipd = osd_idx_ipd_get(env, bag);
5223 if (unlikely(ipd == NULL))
5226 oh = container_of(handle, struct osd_thandle, ot_super);
5227 LASSERT(oh->ot_handle != NULL);
5228 LASSERT(oh->ot_handle->h_transaction != NULL);
5230 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5231 /* swab quota uid/gid provided by caller */
5232 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5233 key = (const struct dt_key *)&oti->oti_quota_id;
5236 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5237 osd_ipd_put(env, bag, ipd);
5238 LINVRNT(osd_invariant(obj));
5239 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5243 static int osd_index_declare_ea_delete(const struct lu_env *env,
5244 struct dt_object *dt,
5245 const struct dt_key *key,
5246 struct thandle *handle)
5248 struct osd_thandle *oh;
5249 struct inode *inode;
5254 LASSERT(!dt_object_remote(dt));
5255 LASSERT(handle != NULL);
5257 oh = container_of(handle, struct osd_thandle, ot_super);
5258 LASSERT(oh->ot_handle == NULL);
5260 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5261 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5263 inode = osd_dt_obj(dt)->oo_inode;
5267 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5268 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5273 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5276 struct osd_fid_pack *rec;
5279 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5280 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5281 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5282 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5288 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5289 const struct lu_fid *fid)
5291 struct seq_server_site *ss = osd_seq_site(osd);
5295 /* FID seqs not in FLDB, must be local seq */
5296 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5300 * If FLD is not being initialized yet, it only happens during the
5301 * initialization, likely during mgs initialization, and we assume
5302 * this is local FID.
5304 if (ss == NULL || ss->ss_server_fld == NULL)
5307 /* Only check the local FLDB here */
5308 if (osd_seq_exists(env, osd, fid_seq(fid)))
5314 static void osd_take_care_of_agent(const struct lu_env *env,
5315 struct osd_device *osd,
5316 struct osd_thandle *oh,
5317 struct ldiskfs_dir_entry_2 *de)
5319 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5320 struct osd_idmap_cache *idc;
5321 int rc, schedule = 0;
5323 LASSERT(de != NULL);
5325 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5326 if (likely(rc == 0)) {
5327 idc = osd_idc_find_or_init(env, osd, fid);
5328 if (IS_ERR(idc) || idc->oic_remote)
5330 } else if (rc == -ENODATA) {
5332 * can't get FID, postpone to the end of the
5333 * transaction when iget() is safe
5337 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5340 osd_schedule_agent_inode_removal(env, oh,
5341 le32_to_cpu(de->inode));
5345 * Index delete function for interoperability mode (b11826).
5346 * It will remove the directory entry added by osd_index_ea_insert().
5347 * This entry is needed to maintain name->fid mapping.
5349 * \param key, key i.e. file entry to be deleted
5351 * \retval 0, on success
5352 * \retval -ve, on error
5354 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5355 const struct dt_key *key, struct thandle *handle)
5357 struct osd_object *obj = osd_dt_obj(dt);
5358 struct inode *dir = obj->oo_inode;
5359 struct dentry *dentry;
5360 struct osd_thandle *oh;
5361 struct ldiskfs_dir_entry_2 *de = NULL;
5362 struct buffer_head *bh;
5363 struct htree_lock *hlock = NULL;
5364 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5369 if (!dt_object_exists(dt))
5372 LINVRNT(osd_invariant(obj));
5373 LASSERT(!dt_object_remote(dt));
5374 LASSERT(handle != NULL);
5376 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5378 oh = container_of(handle, struct osd_thandle, ot_super);
5379 LASSERT(oh->ot_handle != NULL);
5380 LASSERT(oh->ot_handle->h_transaction != NULL);
5382 dquot_initialize(dir);
5383 dentry = osd_child_dentry_get(env, obj,
5384 (char *)key, strlen((char *)key));
5386 if (obj->oo_hl_head != NULL) {
5387 hlock = osd_oti_get(env)->oti_hlock;
5388 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5389 dir, LDISKFS_HLOCK_DEL);
5391 down_write(&obj->oo_ext_idx_sem);
5394 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5397 * If this is not the ".." entry, it might be a remote DNE
5398 * entry and we need to check if the FID is for a remote
5399 * MDT. If the FID is not in the directory entry (e.g.
5400 * upgraded 1.8 filesystem without dirdata enabled) then
5401 * we need to get the FID from the LMA. For a remote directory
5402 * there HAS to be an LMA, it cannot be an IGIF inode in this
5405 * Delete the entry before the agent inode in order to
5406 * simplify error handling. At worst an error after deleting
5407 * the entry first might leak the agent inode afterward. The
5408 * reverse would need filesystem abort in case of error deleting
5409 * the entry after the agent had been removed, or leave a
5410 * dangling entry pointing at a random inode.
5412 if (strcmp((char *)key, dotdot) != 0)
5413 osd_take_care_of_agent(env, osd, oh, de);
5414 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5420 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5421 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5422 /* NB, dirent count may not be accurate, because it's counted
5425 if (obj->oo_dirent_count)
5426 obj->oo_dirent_count--;
5428 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5431 ldiskfs_htree_unlock(hlock);
5433 up_write(&obj->oo_ext_idx_sem);
5436 LASSERT(osd_invariant(obj));
5437 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5442 * Lookup index for \a key and copy record to \a rec.
5444 * \param dt osd index object
5445 * \param key key for index
5446 * \param rec record reference
5448 * \retval +ve success : exact mach
5449 * \retval 0 return record with key not greater than \a key
5450 * \retval -ve failure
5452 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5453 struct dt_rec *rec, const struct dt_key *key)
5455 struct osd_object *obj = osd_dt_obj(dt);
5456 struct iam_path_descr *ipd;
5457 struct iam_container *bag = &obj->oo_dir->od_container;
5458 struct osd_thread_info *oti = osd_oti_get(env);
5459 struct iam_iterator *it = &oti->oti_idx_it;
5460 struct iam_rec *iam_rec;
5465 if (!dt_object_exists(dt))
5468 LASSERT(osd_invariant(obj));
5469 LASSERT(!dt_object_remote(dt));
5470 LASSERT(bag->ic_object == obj->oo_inode);
5472 ipd = osd_idx_ipd_get(env, bag);
5476 /* got ipd now we can start iterator. */
5477 iam_it_init(it, bag, 0, ipd);
5479 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5480 /* swab quota uid/gid provided by caller */
5481 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5482 key = (const struct dt_key *)&oti->oti_quota_id;
5485 rc = iam_it_get(it, (struct iam_key *)key);
5487 if (S_ISDIR(obj->oo_inode->i_mode))
5488 iam_rec = (struct iam_rec *)oti->oti_ldp;
5490 iam_rec = (struct iam_rec *)rec;
5492 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5494 if (S_ISDIR(obj->oo_inode->i_mode))
5495 osd_fid_unpack((struct lu_fid *)rec,
5496 (struct osd_fid_pack *)iam_rec);
5497 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5498 osd_quota_unpack(obj, rec);
5503 osd_ipd_put(env, bag, ipd);
5505 LINVRNT(osd_invariant(obj));
5510 static int osd_index_declare_iam_insert(const struct lu_env *env,
5511 struct dt_object *dt,
5512 const struct dt_rec *rec,
5513 const struct dt_key *key,
5514 struct thandle *handle)
5516 struct osd_thandle *oh;
5518 LASSERT(handle != NULL);
5520 oh = container_of(handle, struct osd_thandle, ot_super);
5521 LASSERT(oh->ot_handle == NULL);
5523 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5524 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5530 * Inserts (key, value) pair in \a dt index object.
5532 * \param dt osd index object
5533 * \param key key for index
5534 * \param rec record reference
5535 * \param th transaction handler
5538 * \retval -ve failure
5540 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5541 const struct dt_rec *rec,
5542 const struct dt_key *key, struct thandle *th)
5544 struct osd_object *obj = osd_dt_obj(dt);
5545 struct iam_path_descr *ipd;
5546 struct osd_thandle *oh;
5547 struct iam_container *bag;
5548 struct osd_thread_info *oti = osd_oti_get(env);
5549 struct iam_rec *iam_rec;
5554 if (!dt_object_exists(dt))
5557 LINVRNT(osd_invariant(obj));
5558 LASSERT(!dt_object_remote(dt));
5560 bag = &obj->oo_dir->od_container;
5561 LASSERT(bag->ic_object == obj->oo_inode);
5562 LASSERT(th != NULL);
5564 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5566 ipd = osd_idx_ipd_get(env, bag);
5567 if (unlikely(ipd == NULL))
5570 oh = container_of(th, struct osd_thandle, ot_super);
5571 LASSERT(oh->ot_handle != NULL);
5572 LASSERT(oh->ot_handle->h_transaction != NULL);
5573 if (S_ISDIR(obj->oo_inode->i_mode)) {
5574 iam_rec = (struct iam_rec *)oti->oti_ldp;
5575 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5577 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5578 /* pack quota uid/gid */
5579 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5580 key = (const struct dt_key *)&oti->oti_quota_id;
5581 /* pack quota record */
5582 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5583 iam_rec = (struct iam_rec *)rec;
5585 iam_rec = (struct iam_rec *)rec;
5588 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5590 osd_ipd_put(env, bag, ipd);
5591 LINVRNT(osd_invariant(obj));
5592 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5597 * Calls ldiskfs_add_entry() to add directory entry
5598 * into the directory. This is required for
5599 * interoperability mode (b11826)
5601 * \retval 0, on success
5602 * \retval -ve, on error
5604 static int __osd_ea_add_rec(struct osd_thread_info *info,
5605 struct osd_object *pobj, struct inode *cinode,
5606 const char *name, const struct lu_fid *fid,
5607 struct htree_lock *hlock, struct thandle *th)
5609 struct ldiskfs_dentry_param *ldp;
5610 struct dentry *child;
5611 struct osd_thandle *oth;
5614 oth = container_of(th, struct osd_thandle, ot_super);
5615 LASSERT(oth->ot_handle != NULL);
5616 LASSERT(oth->ot_handle->h_transaction != NULL);
5617 LASSERT(pobj->oo_inode);
5619 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5620 if (unlikely(osd_object_is_root(pobj)))
5623 osd_get_ldiskfs_dirent_param(ldp, fid);
5624 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5625 child->d_fsdata = (void *)ldp;
5626 dquot_initialize(pobj->oo_inode);
5627 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5628 child, cinode, hlock);
5629 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5630 struct ldiskfs_dir_entry_2 *de;
5631 struct buffer_head *bh;
5634 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5637 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5640 if (S_ISDIR(cinode->i_mode))
5641 de->file_type = LDISKFS_DIRENT_LUFID |
5642 LDISKFS_FT_REG_FILE;
5644 de->file_type = LDISKFS_DIRENT_LUFID |
5646 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5657 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5658 * into the directory.Also sets flags into osd object to
5659 * indicate dot and dotdot are created. This is required for
5660 * interoperability mode (b11826)
5662 * \param dir directory for dot and dotdot fixup.
5663 * \param obj child object for linking
5665 * \retval 0, on success
5666 * \retval -ve, on error
5668 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5669 struct osd_object *dir,
5670 struct inode *parent_dir, const char *name,
5671 const struct lu_fid *dot_fid,
5672 const struct lu_fid *dot_dot_fid,
5675 struct inode *inode = dir->oo_inode;
5676 struct osd_thandle *oth;
5679 oth = container_of(th, struct osd_thandle, ot_super);
5680 LASSERT(oth->ot_handle->h_transaction != NULL);
5681 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5683 if (strcmp(name, dot) == 0) {
5684 if (dir->oo_compat_dot_created) {
5687 LASSERT(inode->i_ino == parent_dir->i_ino);
5688 dir->oo_compat_dot_created = 1;
5691 } else if (strcmp(name, dotdot) == 0) {
5692 if (!dir->oo_compat_dot_created)
5694 /* in case of rename, dotdot is already created */
5695 if (dir->oo_compat_dotdot_created) {
5696 return __osd_ea_add_rec(info, dir, parent_dir, name,
5697 dot_dot_fid, NULL, th);
5700 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5701 struct lu_fid tfid = *dot_dot_fid;
5704 result = osd_add_dot_dotdot_internal(info,
5705 dir->oo_inode, parent_dir, dot_fid,
5708 result = osd_add_dot_dotdot_internal(info,
5709 dir->oo_inode, parent_dir, dot_fid,
5714 dir->oo_compat_dotdot_created = 1;
5722 * It will call the appropriate osd_add* function and return the
5723 * value, return by respective functions.
5725 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5726 struct inode *cinode, const char *name,
5727 const struct lu_fid *fid, struct thandle *th)
5729 struct osd_thread_info *info = osd_oti_get(env);
5730 struct htree_lock *hlock;
5733 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5735 if (name[0] == '.' && (name[1] == '\0' ||
5736 (name[1] == '.' && name[2] == '\0'))) {
5737 if (hlock != NULL) {
5738 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5741 down_write(&pobj->oo_ext_idx_sem);
5744 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5745 lu_object_fid(&pobj->oo_dt.do_lu),
5748 if (hlock != NULL) {
5749 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5750 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5752 down_write(&pobj->oo_ext_idx_sem);
5755 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5756 struct lu_fid *tfid = &info->oti_fid;
5760 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5763 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5767 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
5768 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
5769 pobj->oo_dirent_count++;
5772 ldiskfs_htree_unlock(hlock);
5774 up_write(&pobj->oo_ext_idx_sem);
5780 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5781 const struct lu_fid *fid, struct osd_inode_id *id)
5783 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5784 struct inode *inode = NULL;
5791 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5794 if (scrub->os_running && scrub->os_pos_current > id->oii_ino)
5797 if (dev->od_auto_scrub_interval == AS_NEVER ||
5798 ktime_get_real_seconds() <
5799 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5803 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5804 if (rc == -ENOENT) {
5805 __u32 gen = id->oii_gen;
5811 inode = osd_iget(oti, dev, id);
5812 /* The inode has been removed (by race maybe). */
5813 if (IS_ERR(inode)) {
5814 rc = PTR_ERR(inode);
5816 RETURN(rc == -ESTALE ? -ENOENT : rc);
5819 /* The OI mapping is lost. */
5820 if (gen != OSD_OII_NOGEN)
5824 * The inode may has been reused by others, we do not know,
5825 * leave it to be handled by subsequent osd_fid_lookup().
5828 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5835 if (scrub->os_running) {
5836 if (inode == NULL) {
5837 inode = osd_iget(oti, dev, id);
5838 /* The inode has been removed (by race maybe). */
5839 if (IS_ERR(inode)) {
5840 rc = PTR_ERR(inode);
5842 RETURN(rc == -ESTALE ? -ENOENT : rc);
5846 rc = osd_oii_insert(dev, fid, id, insert);
5848 * There is race condition between osd_oi_lookup and OI scrub.
5849 * The OI scrub finished just after osd_oi_lookup() failure.
5850 * Under such case, it is unnecessary to trigger OI scrub again,
5851 * but try to call osd_oi_lookup() again.
5853 if (unlikely(rc == -EAGAIN))
5856 if (!S_ISDIR(inode->i_mode))
5859 rc = osd_check_lmv(oti, dev, inode);
5864 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5865 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5866 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5867 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
5868 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
5869 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
5870 if (rc == 0 || rc == -EALREADY)
5882 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5883 struct osd_device *dev,
5884 struct lu_fid *fid, __u32 ino)
5886 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5887 struct osd_idmap_cache *oic = &oti->oti_cache;
5888 struct inode *inode;
5891 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5892 inode = osd_iget(oti, dev, &oic->oic_lid);
5893 if (IS_ERR(inode)) {
5894 fid_zero(&oic->oic_fid);
5895 return PTR_ERR(inode);
5898 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5901 fid_zero(&oic->oic_fid);
5903 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5907 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5908 struct osd_inode_id *id, const struct lu_fid *fid)
5910 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5911 id->oii_ino, id->oii_gen, info);
5912 info->oti_cache.oic_lid = *id;
5913 info->oti_cache.oic_fid = *fid;
5914 info->oti_cache.oic_dev = osd;
5918 * Get parent FID from the linkEA.
5920 * For a directory which parent resides on remote MDT, to satisfy the
5921 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5922 * the other hand, to make the lookup(..) on the directory can return
5923 * the real parent FID, we append the real parent FID after its ".."
5924 * name entry in the /REMOTE_PARENT_DIR.
5926 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5927 * backup. So after the restore, we cannot get the right parent FID from
5928 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5929 * we have stored the real parent FID in the directory object's linkEA,
5930 * we can parse the linkEA for the real parent FID.
5932 * \param[in] env pointer to the thread context
5933 * \param[in] obj pointer to the object to be handled
5934 * \param[out]fid pointer to the buffer to hold the parent FID
5936 * \retval 0 for getting the real parent FID successfully
5937 * \retval negative error number on failure
5939 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5940 struct osd_object *obj,
5943 struct osd_thread_info *oti = osd_oti_get(env);
5944 struct lu_buf *buf = &oti->oti_big_buf;
5945 struct dentry *dentry = &oti->oti_obj_dentry;
5946 struct inode *inode = obj->oo_inode;
5947 struct linkea_data ldata = { NULL };
5953 if (!S_ISDIR(inode->i_mode))
5957 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5958 buf->lb_buf, buf->lb_len);
5959 if (rc == -ERANGE) {
5960 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5963 lu_buf_realloc(buf, rc);
5964 if (buf->lb_buf == NULL)
5971 if (unlikely(rc == 0))
5977 if (unlikely(buf->lb_buf == NULL)) {
5978 lu_buf_realloc(buf, rc);
5979 if (buf->lb_buf == NULL)
5986 rc = linkea_init_with_rec(&ldata);
5988 linkea_first_entry(&ldata);
5989 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
5995 static int osd_verify_ent_by_linkea(const struct lu_env *env,
5996 struct inode *inode,
5997 const struct lu_fid *pfid,
5998 const char *name, const int namelen)
6000 struct osd_thread_info *oti = osd_oti_get(env);
6001 struct lu_buf *buf = &oti->oti_big_buf;
6002 struct dentry *dentry = &oti->oti_obj_dentry;
6003 struct linkea_data ldata = { NULL };
6004 struct lu_name cname = { .ln_name = name,
6005 .ln_namelen = namelen };
6011 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6012 buf->lb_buf, buf->lb_len);
6014 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6019 if (unlikely(rc == 0))
6022 if (buf->lb_len < rc) {
6023 lu_buf_realloc(buf, rc);
6024 if (buf->lb_buf == NULL)
6031 rc = linkea_init_with_rec(&ldata);
6033 rc = linkea_links_find(&ldata, &cname, pfid);
6039 * Calls ->lookup() to find dentry. From dentry get inode and
6040 * read inode's ea to get fid. This is required for interoperability
6043 * \retval 0, on success
6044 * \retval -ve, on error
6046 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6047 struct dt_rec *rec, const struct dt_key *key)
6049 struct inode *dir = obj->oo_inode;
6050 struct dentry *dentry;
6051 struct ldiskfs_dir_entry_2 *de;
6052 struct buffer_head *bh;
6053 struct lu_fid *fid = (struct lu_fid *)rec;
6054 struct htree_lock *hlock = NULL;
6060 LASSERT(dir->i_op != NULL);
6061 LASSERT(dir->i_op->lookup != NULL);
6063 dentry = osd_child_dentry_get(env, obj,
6064 (char *)key, strlen((char *)key));
6066 if (obj->oo_hl_head != NULL) {
6067 hlock = osd_oti_get(env)->oti_hlock;
6068 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6069 dir, LDISKFS_HLOCK_LOOKUP);
6071 down_read(&obj->oo_ext_idx_sem);
6074 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6076 struct osd_thread_info *oti = osd_oti_get(env);
6077 struct osd_inode_id *id = &oti->oti_id;
6078 struct osd_device *dev = osd_obj2dev(obj);
6080 ino = le32_to_cpu(de->inode);
6081 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6083 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6087 rc = osd_get_fid_from_dentry(de, rec);
6089 /* done with de, release bh */
6092 if (unlikely(is_remote_parent_ino(dev, ino))) {
6093 const char *name = (const char *)key;
6096 * If the parent is on remote MDT, and there
6097 * is no FID-in-dirent, then we have to get
6098 * the parent FID from the linkEA.
6100 if (likely(strlen(name) == 2 &&
6101 name[0] == '.' && name[1] == '.'))
6102 rc = osd_get_pfid_from_linkea(env, obj,
6105 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6108 osd_id_gen(id, ino, OSD_OII_NOGEN);
6111 if (rc != 0 || osd_remote_fid(env, dev, fid))
6114 rc = osd_consistency_check(oti, dev, fid, id);
6115 if (rc != -ENOENT) {
6116 /* Other error should not affect lookup result. */
6119 /* Normal file mapping should be added into OI cache
6120 * after FID in LMA check, but for local files like
6121 * hsm_actions, their FIDs are not stored in OI files,
6122 * see osd_initial_OI_scrub(), and here is the only
6123 * place to load mapping into OI cache.
6125 if (!fid_is_namespace_visible(fid))
6126 osd_add_oi_cache(osd_oti_get(env),
6127 osd_obj2dev(obj), id, fid);
6137 ldiskfs_htree_unlock(hlock);
6139 up_read(&obj->oo_ext_idx_sem);
6143 static int osd_index_declare_ea_insert(const struct lu_env *env,
6144 struct dt_object *dt,
6145 const struct dt_rec *rec,
6146 const struct dt_key *key,
6147 struct thandle *handle)
6149 struct osd_thandle *oh;
6150 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6151 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6152 const struct lu_fid *fid = rec1->rec_fid;
6153 int credits, rc = 0;
6154 struct osd_idmap_cache *idc;
6158 LASSERT(!dt_object_remote(dt));
6159 LASSERT(handle != NULL);
6160 LASSERT(fid != NULL);
6161 LASSERT(rec1->rec_type != 0);
6163 oh = container_of(handle, struct osd_thandle, ot_super);
6164 LASSERT(oh->ot_handle == NULL);
6166 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6169 * we can't call iget() while a transactions is running
6170 * (this can lead to a deadlock), but we need to know
6171 * inum and object type. so we find this information at
6172 * declaration and cache in per-thread info
6174 idc = osd_idc_find_or_init(env, osd, fid);
6176 RETURN(PTR_ERR(idc));
6177 if (idc->oic_remote) {
6179 * a reference to remote inode is represented by an
6180 * agent inode which we have to create
6182 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6183 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6186 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6188 if (osd_dt_obj(dt)->oo_inode != NULL) {
6189 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6192 * We ignore block quota on meta pool (MDTs), so needn't
6193 * calculate how many blocks will be consumed by this index
6196 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6198 i_projid_read(inode), 0,
6199 oh, osd_dt_obj(dt), NULL,
6204 #ifdef HAVE_PROJECT_QUOTA
6206 * Reserve credits for local agent inode to transfer
6207 * to 0, quota enforcement is ignored in this case.
6209 if (idc->oic_remote &&
6210 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6211 i_projid_read(inode) != 0)
6212 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6213 0, i_projid_read(inode),
6214 0, false, PRJQUOTA, true);
6222 * Index add function for interoperability mode (b11826).
6223 * It will add the directory entry.This entry is needed to
6224 * maintain name->fid mapping.
6226 * \param key it is key i.e. file entry to be inserted
6227 * \param rec it is value of given key i.e. fid
6229 * \retval 0, on success
6230 * \retval -ve, on error
6232 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6233 const struct dt_rec *rec,
6234 const struct dt_key *key, struct thandle *th)
6236 struct osd_object *obj = osd_dt_obj(dt);
6237 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6238 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6239 const struct lu_fid *fid = rec1->rec_fid;
6240 const char *name = (const char *)key;
6241 struct osd_thread_info *oti = osd_oti_get(env);
6242 struct inode *child_inode = NULL;
6243 struct osd_idmap_cache *idc;
6248 if (!dt_object_exists(dt))
6251 LASSERT(osd_invariant(obj));
6252 LASSERT(!dt_object_remote(dt));
6253 LASSERT(th != NULL);
6255 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6257 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6259 idc = osd_idc_find(env, osd, fid);
6260 if (unlikely(idc == NULL)) {
6261 idc = osd_idc_find_or_init(env, osd, fid);
6264 * this dt_insert() wasn't declared properly, so
6265 * FID is missing in OI cache. we better do not
6266 * lookup FID in FLDB/OI and don't risk to deadlock,
6267 * but in some special cases (lfsck testing, etc)
6268 * it's much simpler than fixing a caller.
6270 * normally this error should be placed after the first
6271 * find, but migrate may attach source stripes to
6272 * target, which doesn't create stripes.
6274 CERROR("%s: "DFID" wasn't declared for insert\n",
6275 osd_name(osd), PFID(fid));
6277 RETURN(PTR_ERR(idc));
6281 if (idc->oic_remote) {
6282 /* Insert remote entry */
6283 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6285 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6287 child_inode = osd_create_local_agent_inode(env, osd,
6288 obj, fid, rec1->rec_type & S_IFMT, th);
6289 if (IS_ERR(child_inode))
6290 RETURN(PTR_ERR(child_inode));
6293 /* Insert local entry */
6294 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6295 /* for a reason OI cache wasn't filled properly */
6296 CERROR("%s: OIC for "DFID" isn't filled\n",
6297 osd_name(osd), PFID(fid));
6300 child_inode = oti->oti_inode;
6301 if (unlikely(child_inode == NULL)) {
6302 struct ldiskfs_inode_info *lii;
6307 child_inode = oti->oti_inode = &lii->vfs_inode;
6309 child_inode->i_sb = osd_sb(osd);
6310 child_inode->i_ino = idc->oic_lid.oii_ino;
6311 child_inode->i_mode = rec1->rec_type & S_IFMT;
6314 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6316 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6317 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6319 if (child_inode && child_inode != oti->oti_inode)
6321 LASSERT(osd_invariant(obj));
6322 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6328 * Initialize osd Iterator for given osd index object.
6330 * \param dt osd index object
6333 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6334 struct dt_object *dt,
6337 struct osd_it_iam *it;
6338 struct osd_object *obj = osd_dt_obj(dt);
6339 struct lu_object *lo = &dt->do_lu;
6340 struct iam_path_descr *ipd;
6341 struct iam_container *bag = &obj->oo_dir->od_container;
6343 if (!dt_object_exists(dt))
6344 return ERR_PTR(-ENOENT);
6348 return ERR_PTR(-ENOMEM);
6350 ipd = osd_it_ipd_get(env, bag);
6351 if (likely(ipd != NULL)) {
6355 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6356 return (struct dt_it *)it;
6359 return ERR_PTR(-ENOMEM);
6364 * free given Iterator.
6366 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6368 struct osd_it_iam *it = (struct osd_it_iam *)di;
6369 struct osd_object *obj = it->oi_obj;
6371 iam_it_fini(&it->oi_it);
6372 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6373 osd_object_put(env, obj);
6378 * Move Iterator to record specified by \a key
6380 * \param di osd iterator
6381 * \param key key for index
6383 * \retval +ve di points to record with least key not larger than key
6384 * \retval 0 di points to exact matched key
6385 * \retval -ve failure
6388 static int osd_it_iam_get(const struct lu_env *env,
6389 struct dt_it *di, const struct dt_key *key)
6391 struct osd_thread_info *oti = osd_oti_get(env);
6392 struct osd_it_iam *it = (struct osd_it_iam *)di;
6394 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6395 /* swab quota uid/gid */
6396 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6397 key = (struct dt_key *)&oti->oti_quota_id;
6400 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6406 * \param di osd iterator
6408 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6410 struct osd_it_iam *it = (struct osd_it_iam *)di;
6412 iam_it_put(&it->oi_it);
6416 * Move iterator by one record
6418 * \param di osd iterator
6420 * \retval +1 end of container reached
6422 * \retval -ve failure
6425 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6427 struct osd_it_iam *it = (struct osd_it_iam *)di;
6429 return iam_it_next(&it->oi_it);
6433 * Return pointer to the key under iterator.
6436 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6437 const struct dt_it *di)
6439 struct osd_thread_info *oti = osd_oti_get(env);
6440 struct osd_it_iam *it = (struct osd_it_iam *)di;
6441 struct osd_object *obj = it->oi_obj;
6444 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6446 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6447 /* swab quota uid/gid */
6448 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6449 key = (struct dt_key *)&oti->oti_quota_id;
6456 * Return size of key under iterator (in bytes)
6459 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6461 struct osd_it_iam *it = (struct osd_it_iam *)di;
6463 return iam_it_key_size(&it->oi_it);
6467 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6469 /* check if file type is required */
6470 if (ent->lde_attrs & LUDA_TYPE) {
6471 struct luda_type *lt;
6472 int align = sizeof(*lt) - 1;
6474 len = (len + align) & ~align;
6475 lt = (struct luda_type *)(ent->lde_name + len);
6476 lt->lt_type = cpu_to_le16(DTTOIF(type));
6479 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6483 * build lu direct from backend fs dirent.
6487 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6488 char *name, __u16 namelen, __u16 type, __u32 attr)
6490 ent->lde_attrs = attr | LUDA_FID;
6491 fid_cpu_to_le(&ent->lde_fid, fid);
6493 ent->lde_hash = cpu_to_le64(offset);
6494 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6496 strncpy(ent->lde_name, name, namelen);
6497 ent->lde_name[namelen] = '\0';
6498 ent->lde_namelen = cpu_to_le16(namelen);
6500 /* append lustre attributes */
6501 osd_it_append_attrs(ent, namelen, type);
6505 * Return pointer to the record under iterator.
6507 static int osd_it_iam_rec(const struct lu_env *env,
6508 const struct dt_it *di,
6509 struct dt_rec *dtrec, __u32 attr)
6511 struct osd_it_iam *it = (struct osd_it_iam *)di;
6512 struct osd_thread_info *info = osd_oti_get(env);
6516 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6517 const struct osd_fid_pack *rec;
6518 struct lu_fid *fid = &info->oti_fid;
6519 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6525 name = (char *)iam_it_key_get(&it->oi_it);
6527 RETURN(PTR_ERR(name));
6529 namelen = iam_it_key_size(&it->oi_it);
6531 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6533 RETURN(PTR_ERR(rec));
6535 rc = osd_fid_unpack(fid, rec);
6539 hash = iam_it_store(&it->oi_it);
6541 /* IAM does not store object type in IAM index (dir) */
6542 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6544 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6545 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6546 (struct iam_rec *)dtrec);
6547 osd_quota_unpack(it->oi_obj, dtrec);
6549 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6550 (struct iam_rec *)dtrec);
6557 * Returns cookie for current Iterator position.
6559 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6561 struct osd_it_iam *it = (struct osd_it_iam *)di;
6563 return iam_it_store(&it->oi_it);
6567 * Restore iterator from cookie.
6569 * \param di osd iterator
6570 * \param hash Iterator location cookie
6572 * \retval +ve di points to record with least key not larger than key.
6573 * \retval 0 di points to exact matched key
6574 * \retval -ve failure
6577 static int osd_it_iam_load(const struct lu_env *env,
6578 const struct dt_it *di, __u64 hash)
6580 struct osd_it_iam *it = (struct osd_it_iam *)di;
6582 return iam_it_load(&it->oi_it, hash);
6585 static const struct dt_index_operations osd_index_iam_ops = {
6586 .dio_lookup = osd_index_iam_lookup,
6587 .dio_declare_insert = osd_index_declare_iam_insert,
6588 .dio_insert = osd_index_iam_insert,
6589 .dio_declare_delete = osd_index_declare_iam_delete,
6590 .dio_delete = osd_index_iam_delete,
6592 .init = osd_it_iam_init,
6593 .fini = osd_it_iam_fini,
6594 .get = osd_it_iam_get,
6595 .put = osd_it_iam_put,
6596 .next = osd_it_iam_next,
6597 .key = osd_it_iam_key,
6598 .key_size = osd_it_iam_key_size,
6599 .rec = osd_it_iam_rec,
6600 .store = osd_it_iam_store,
6601 .load = osd_it_iam_load
6605 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
6606 struct inode *inode, __u32 attr)
6608 struct osd_thread_info *info = osd_oti_get(env);
6609 struct osd_it_ea *oie;
6611 struct dentry *obj_dentry;
6615 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6617 RETURN(ERR_PTR(-ENOMEM));
6618 obj_dentry = &oie->oie_dentry;
6620 obj_dentry->d_inode = inode;
6621 obj_dentry->d_sb = inode->i_sb;
6622 obj_dentry->d_name.hash = 0;
6624 oie->oie_rd_dirent = 0;
6625 oie->oie_it_dirent = 0;
6626 oie->oie_dirent = NULL;
6627 if (unlikely(!info->oti_it_ea_buf_used)) {
6628 oie->oie_buf = info->oti_it_ea_buf;
6629 info->oti_it_ea_buf_used = 1;
6631 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6632 if (oie->oie_buf == NULL)
6633 RETURN(ERR_PTR(-ENOMEM));
6635 oie->oie_obj = NULL;
6637 file = &oie->oie_file;
6639 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6640 if (attr & LUDA_64BITHASH)
6641 file->f_mode = FMODE_64BITHASH;
6643 file->f_mode = FMODE_32BITHASH;
6644 file->f_path.dentry = obj_dentry;
6645 file->f_mapping = inode->i_mapping;
6646 file->f_op = inode->i_fop;
6647 file->f_inode = inode;
6653 * Creates or initializes iterator context.
6655 * \retval struct osd_it_ea, iterator structure on success
6658 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6659 struct dt_object *dt,
6662 struct osd_object *obj = osd_dt_obj(dt);
6663 struct lu_object *lo = &dt->do_lu;
6664 struct osd_it_ea *oie;
6668 if (!dt_object_exists(dt) || obj->oo_destroyed)
6669 RETURN(ERR_PTR(-ENOENT));
6671 oie = osd_it_dir_init(env, obj->oo_inode, attr);
6673 RETURN((struct dt_it *)oie);
6677 RETURN((struct dt_it *)oie);
6680 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
6681 struct inode *inode)
6683 struct osd_thread_info *info = osd_oti_get(env);
6686 oie->oie_file.f_op->release(inode, &oie->oie_file);
6687 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6688 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6690 info->oti_it_ea_buf_used = 0;
6691 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6696 * Destroy or finishes iterator context.
6698 * \param di iterator structure to be destroyed
6700 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6702 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6703 struct osd_object *obj = oie->oie_obj;
6704 struct inode *inode = obj->oo_inode;
6707 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
6708 osd_object_put(env, obj);
6713 * It position the iterator at given key, so that next lookup continues from
6714 * that key Or it is similar to dio_it->load() but based on a key,
6715 * rather than file position.
6717 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6720 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6722 static int osd_it_ea_get(const struct lu_env *env,
6723 struct dt_it *di, const struct dt_key *key)
6725 struct osd_it_ea *it = (struct osd_it_ea *)di;
6728 LASSERT(((const char *)key)[0] == '\0');
6729 it->oie_file.f_pos = 0;
6730 it->oie_rd_dirent = 0;
6731 it->oie_it_dirent = 0;
6732 it->oie_dirent = NULL;
6740 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6744 struct osd_filldir_cbs {
6745 struct dir_context ctx;
6746 struct osd_it_ea *it;
6749 * It is called internally by ->iterate*(). It fills the
6750 * iterator's in-memory data structure with required
6751 * information i.e. name, namelen, rec_size etc.
6753 * \param buf in which information to be filled in.
6754 * \param name name of the file in given dir
6756 * \retval 0 on success
6757 * \retval 1 on buffer full
6759 #ifdef HAVE_FILLDIR_USE_CTX
6760 static int osd_ldiskfs_filldir(struct dir_context *buf,
6762 static int osd_ldiskfs_filldir(void *buf,
6764 const char *name, int namelen,
6765 loff_t offset, __u64 ino, unsigned int d_type)
6767 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6768 struct osd_object *obj = it->oie_obj;
6769 struct osd_it_ea_dirent *ent = it->oie_dirent;
6770 struct lu_fid *fid = &ent->oied_fid;
6771 struct osd_fid_pack *rec;
6775 /* this should never happen */
6776 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6777 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6781 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6785 /* "." is just the object itself. */
6786 if (namelen == 1 && name[0] == '.') {
6788 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6789 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6790 rec = (struct osd_fid_pack *)(name + namelen + 1);
6791 if (osd_fid_unpack(fid, rec) != 0)
6796 d_type &= ~LDISKFS_DIRENT_LUFID;
6798 /* NOT export local root. */
6800 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6801 ino = obj->oo_inode->i_ino;
6802 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6805 ent->oied_ino = ino;
6806 ent->oied_off = offset;
6807 ent->oied_namelen = namelen;
6808 ent->oied_type = d_type;
6810 memcpy(ent->oied_name, name, namelen);
6812 it->oie_rd_dirent++;
6813 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6818 * Calls ->iterate*() to load a directory entry at a time
6819 * and stored it in iterator's in-memory data structure.
6821 * \param di iterator's in memory structure
6823 * \retval 0 on success
6824 * \retval -ve on error
6825 * \retval +1 reach the end of entry
6827 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
6829 struct osd_it_ea *it = (struct osd_it_ea *)di;
6830 struct osd_object *obj = it->oie_obj;
6831 struct htree_lock *hlock = NULL;
6832 struct file *filp = &it->oie_file;
6834 struct osd_filldir_cbs buf = {
6835 .ctx.actor = osd_ldiskfs_filldir,
6840 it->oie_dirent = it->oie_buf;
6841 it->oie_rd_dirent = 0;
6844 if (obj->oo_hl_head != NULL) {
6845 hlock = osd_oti_get(env)->oti_hlock;
6846 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6848 LDISKFS_HLOCK_READDIR);
6850 down_read(&obj->oo_ext_idx_sem);
6854 filp->f_cred = current_cred();
6855 rc = osd_security_file_alloc(filp);
6859 filp->f_flags |= O_NOATIME;
6860 filp->f_mode |= FMODE_NONOTIFY;
6861 rc = iterate_dir(filp, &buf.ctx);
6865 if (it->oie_rd_dirent == 0) {
6867 * If it does not get any dirent, it means it has been reached
6868 * to the end of the dir
6870 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6874 it->oie_dirent = it->oie_buf;
6875 it->oie_it_dirent = 1;
6880 ldiskfs_htree_unlock(hlock);
6882 up_read(&obj->oo_ext_idx_sem);
6889 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
6890 * to load a directory entry at a time and stored it in
6891 * iterator's in-memory data structure.
6893 * \param di iterator's in memory structure
6895 * \retval +ve iterator reached to end
6896 * \retval 0 iterator not reached to end
6897 * \retval -ve on error
6899 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6901 struct osd_it_ea *it = (struct osd_it_ea *)di;
6906 if (it->oie_it_dirent < it->oie_rd_dirent) {
6908 (void *)it->oie_dirent +
6909 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6910 it->oie_dirent->oied_namelen);
6911 it->oie_it_dirent++;
6914 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6917 rc = osd_ldiskfs_it_fill(env, di);
6924 * Returns the key at current position from iterator's in memory structure.
6926 * \param di iterator's in memory structure
6928 * \retval key i.e. struct dt_key on success
6930 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6931 const struct dt_it *di)
6933 struct osd_it_ea *it = (struct osd_it_ea *)di;
6935 return (struct dt_key *)it->oie_dirent->oied_name;
6939 * Returns key's size at current position from iterator's in memory structure.
6941 * \param di iterator's in memory structure
6943 * \retval key_size i.e. struct dt_key on success
6945 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6947 struct osd_it_ea *it = (struct osd_it_ea *)di;
6949 return it->oie_dirent->oied_namelen;
6952 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
6953 #undef LDISKFS_DIR_REC_LEN
6954 #define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
6957 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6959 if (LDISKFS_DIR_REC_LEN(de) >=
6960 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6967 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6968 unsigned int blocksize, bool dotdot)
6971 return osd_dotdot_has_space(de);
6973 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6974 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6981 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6982 handle_t *jh, struct dentry *dentry,
6983 const struct lu_fid *fid, struct buffer_head *bh,
6984 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
6987 struct inode *dir = dentry->d_parent->d_inode;
6988 struct inode *inode = dentry->d_inode;
6989 struct osd_fid_pack *rec;
6990 struct ldiskfs_dentry_param *ldp;
6991 int namelen = dentry->d_name.len;
6993 struct osd_thread_info *info = osd_oti_get(env);
6997 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7000 /* There is enough space to hold the FID-in-dirent. */
7001 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7002 rc = ldiskfs_journal_get_write_access(jh, bh);
7006 de->name[namelen] = 0;
7007 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7008 rec->fp_len = sizeof(struct lu_fid) + 1;
7009 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7010 de->file_type |= LDISKFS_DIRENT_LUFID;
7011 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7018 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7022 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7023 osd_get_ldiskfs_dirent_param(ldp, fid);
7024 dentry->d_fsdata = (void *)ldp;
7025 dquot_initialize(dir);
7026 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7028 * It is too bad, we cannot reinsert the name entry back.
7029 * That means we lose it!
7033 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7034 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7035 namelen, dentry->d_name.name, PFID(fid), rc);
7041 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7042 struct osd_it_ea *it, struct lu_fid *fid,
7043 struct osd_inode_id *id, __u32 *attr)
7045 struct osd_thread_info *info = osd_oti_get(env);
7046 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7047 struct osd_device *dev = osd_obj2dev(obj);
7048 struct super_block *sb = osd_sb(dev);
7049 const char *devname = osd_name(dev);
7050 struct osd_it_ea_dirent *ent = it->oie_dirent;
7051 struct inode *dir = obj->oo_inode;
7052 struct htree_lock *hlock = NULL;
7053 struct buffer_head *bh = NULL;
7054 handle_t *jh = NULL;
7055 struct ldiskfs_dir_entry_2 *de;
7056 struct dentry *dentry;
7057 struct inode *inode;
7058 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7061 bool dotdot = false;
7066 if (ent->oied_name[0] == '.') {
7067 if (ent->oied_namelen == 1)
7070 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7074 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7075 inode = osd_iget(info, dev, id);
7076 if (IS_ERR(inode)) {
7077 rc = PTR_ERR(inode);
7078 if (rc == -ENOENT || rc == -ESTALE) {
7080 * Maybe dangling name entry, or
7081 * corrupted directory entry.
7083 *attr |= LUDA_UNKNOWN;
7086 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7087 "check_repair, dir = %lu/%u, name = %.*s, "
7088 "ino = %llu, rc = %d\n",
7089 devname, dir->i_ino, dir->i_generation,
7090 ent->oied_namelen, ent->oied_name,
7097 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
7099 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7100 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7106 * We need to ensure that the name entry is still valid.
7107 * Because it may be removed or renamed by other already.
7109 * The unlink or rename operation will start journal before PDO lock,
7110 * so to avoid deadlock, here we need to start journal handle before
7111 * related PDO lock also. But because we do not know whether there
7112 * will be something to be repaired before PDO lock, we just start
7113 * journal without conditions.
7115 * We may need to remove the name entry firstly, then insert back.
7116 * One credit is for user quota file update.
7117 * One credit is for group quota file update.
7118 * Two credits are for dirty inode.
7120 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7121 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7123 if (dev->od_dirent_journal != 0) {
7126 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7129 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7130 "check_repair, dir = %lu/%u, credits = %d, "
7131 "name = %.*s, ino = %llu: rc = %d\n",
7132 devname, dir->i_ino, dir->i_generation, credits,
7133 ent->oied_namelen, ent->oied_name,
7136 GOTO(out_inode, rc);
7139 if (obj->oo_hl_head != NULL) {
7140 hlock = osd_oti_get(env)->oti_hlock;
7142 * "0" means exclusive lock for the whole directory.
7143 * We need to prevent others access such name entry
7144 * during the delete + insert. Neither HLOCK_ADD nor
7145 * HLOCK_DEL cannot guarantee the atomicity.
7147 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7149 down_write(&obj->oo_ext_idx_sem);
7152 if (obj->oo_hl_head != NULL) {
7153 hlock = osd_oti_get(env)->oti_hlock;
7154 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7155 LDISKFS_HLOCK_LOOKUP);
7157 down_read(&obj->oo_ext_idx_sem);
7161 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7162 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7163 *attr |= LUDA_IGNORE;
7169 * For dotdot entry, if there is not enough space to hold the
7170 * FID-in-dirent, just keep them there. It only happens when the
7171 * device upgraded from 1.8 or restored from MDT file-level backup.
7172 * For the whole directory, only dotdot entry have no FID-in-dirent
7173 * and needs to get FID from LMA when readdir, it will not affect the
7176 if (dotdot && !osd_dotdot_has_space(de)) {
7177 *attr |= LUDA_UNKNOWN;
7183 if (lu_fid_eq(fid, &lma->lma_self_fid))
7186 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7187 struct lu_fid *tfid = &lma->lma_self_fid;
7189 if (likely(dotdot &&
7190 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7191 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7193 * It must be REMOTE_PARENT_DIR and as the
7194 * 'dotdot' entry of remote directory
7196 *attr |= LUDA_IGNORE;
7198 CDEBUG(D_LFSCK, "%s: expect remote agent "
7199 "parent directory, but got %.*s under "
7200 "dir = %lu/%u with the FID "DFID"\n",
7201 devname, ent->oied_namelen,
7202 ent->oied_name, dir->i_ino,
7203 dir->i_generation, PFID(tfid));
7205 *attr |= LUDA_UNKNOWN;
7212 if (!fid_is_zero(fid)) {
7213 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7215 if (rc == -ENOENT ||
7217 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7219 * linkEA does not recognize the dirent entry,
7220 * it may because the dirent entry corruption
7221 * and points to other's inode.
7223 CDEBUG(D_LFSCK, "%s: the target inode does not "
7224 "recognize the dirent, dir = %lu/%u, "
7225 " name = %.*s, ino = %llu, "
7226 DFID": rc = %d\n", devname, dir->i_ino,
7227 dir->i_generation, ent->oied_namelen,
7228 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7229 *attr |= LUDA_UNKNOWN;
7234 if (rc && rc != -ENODATA) {
7235 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7236 "dir = %lu/%u, name = %.*s, ino = %llu, "
7237 DFID": rc = %d\n", devname, dir->i_ino,
7238 dir->i_generation, ent->oied_namelen,
7239 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7240 *attr |= LUDA_UNKNOWN;
7248 * linkEA recognizes the dirent entry, the FID-in-LMA is
7249 * valid, trusted, in spite of fid_is_sane(fid) or not.
7251 if (*attr & LUDA_VERIFY_DRYRUN) {
7252 *fid = lma->lma_self_fid;
7253 *attr |= LUDA_REPAIR;
7260 dev->od_dirent_journal = 1;
7261 if (hlock != NULL) {
7262 ldiskfs_htree_unlock(hlock);
7265 up_read(&obj->oo_ext_idx_sem);
7271 *fid = lma->lma_self_fid;
7273 /* Update or append the FID-in-dirent. */
7274 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7275 bh, de, hlock, dotdot);
7277 *attr |= LUDA_REPAIR;
7279 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7280 "the dirent, dir = %lu/%u, name = %.*s, "
7281 "ino = %llu, "DFID": rc = %d\n",
7282 devname, dir->i_ino, dir->i_generation,
7283 ent->oied_namelen, ent->oied_name,
7284 ent->oied_ino, PFID(fid), rc);
7286 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7287 if (*attr & LUDA_VERIFY_DRYRUN) {
7288 if (fid_is_sane(fid)) {
7289 *attr |= LUDA_REPAIR;
7290 } else if (dev->od_index == 0) {
7291 lu_igif_build(fid, inode->i_ino,
7292 inode->i_generation);
7293 *attr |= LUDA_UPGRADE;
7301 dev->od_dirent_journal = 1;
7302 if (hlock != NULL) {
7303 ldiskfs_htree_unlock(hlock);
7306 up_read(&obj->oo_ext_idx_sem);
7313 if (unlikely(fid_is_sane(fid))) {
7315 * FID-in-dirent exists, but FID-in-LMA is lost.
7316 * Trust the FID-in-dirent, and add FID-in-LMA.
7318 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7320 *attr |= LUDA_REPAIR;
7322 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7323 "update dirent, dir = %lu/%u, "
7324 "name = %.*s, ino = %llu, "
7326 devname, dir->i_ino, dir->i_generation,
7327 ent->oied_namelen, ent->oied_name,
7328 ent->oied_ino, PFID(fid), rc);
7329 } else if (dev->od_index == 0) {
7330 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7332 * It is probably IGIF object. Only aappend the
7333 * FID-in-dirent. OI scrub will process FID-in-LMA.
7335 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7336 bh, de, hlock, dotdot);
7338 *attr |= LUDA_UPGRADE;
7340 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7341 "after the dirent, dir = %lu/%u, "
7342 "name = %.*s, ino = %llu, "
7344 devname, dir->i_ino, dir->i_generation,
7345 ent->oied_namelen, ent->oied_name,
7346 ent->oied_ino, PFID(fid), rc);
7355 if (hlock != NULL) {
7356 ldiskfs_htree_unlock(hlock);
7358 if (dev->od_dirent_journal != 0)
7359 up_write(&obj->oo_ext_idx_sem);
7361 up_read(&obj->oo_ext_idx_sem);
7365 ldiskfs_journal_stop(jh);
7369 if (rc >= 0 && !dirty)
7370 dev->od_dirent_journal = 0;
7376 * Returns the value at current position from iterator's in memory structure.
7378 * \param di struct osd_it_ea, iterator's in memory structure
7379 * \param attr attr requested for dirent.
7380 * \param lde lustre dirent
7382 * \retval 0 no error and \param lde has correct lustre dirent.
7383 * \retval -ve on error
7385 static inline int osd_it_ea_rec(const struct lu_env *env,
7386 const struct dt_it *di,
7387 struct dt_rec *dtrec, __u32 attr)
7389 struct osd_it_ea *it = (struct osd_it_ea *)di;
7390 struct osd_object *obj = it->oie_obj;
7391 struct osd_device *dev = osd_obj2dev(obj);
7392 struct osd_thread_info *oti = osd_oti_get(env);
7393 struct osd_inode_id *id = &oti->oti_id;
7394 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7395 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7396 __u32 ino = it->oie_dirent->oied_ino;
7401 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7403 if (attr & LUDA_VERIFY) {
7404 if (unlikely(is_remote_parent_ino(dev, ino))) {
7405 attr |= LUDA_IGNORE;
7407 * If the parent is on remote MDT, and there
7408 * is no FID-in-dirent, then we have to get
7409 * the parent FID from the linkEA.
7411 if (!fid_is_sane(fid) &&
7412 it->oie_dirent->oied_namelen == 2 &&
7413 it->oie_dirent->oied_name[0] == '.' &&
7414 it->oie_dirent->oied_name[1] == '.')
7415 osd_get_pfid_from_linkea(env, obj, fid);
7417 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7421 if (!fid_is_sane(fid))
7422 attr |= LUDA_UNKNOWN;
7424 attr &= ~LU_DIRENT_ATTRS_MASK;
7425 if (!fid_is_sane(fid)) {
7426 bool is_dotdot = false;
7428 if (it->oie_dirent->oied_namelen == 2 &&
7429 it->oie_dirent->oied_name[0] == '.' &&
7430 it->oie_dirent->oied_name[1] == '.')
7433 * If the parent is on remote MDT, and there
7434 * is no FID-in-dirent, then we have to get
7435 * the parent FID from the linkEA.
7437 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7438 rc = osd_get_pfid_from_linkea(env, obj, fid);
7440 if (is_dotdot == false &&
7441 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7444 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7449 /* Pack the entry anyway, at least the offset is right. */
7450 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7451 it->oie_dirent->oied_name,
7452 it->oie_dirent->oied_namelen,
7453 it->oie_dirent->oied_type, attr);
7455 RETURN(rc > 0 ? 0 : rc);
7459 * Returns the record size size at current position.
7461 * This function will return record(lu_dirent) size in bytes.
7463 * \param[in] env execution environment
7464 * \param[in] di iterator's in memory structure
7465 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7466 * calculate the lu_dirent size.
7468 * \retval record size(in bytes & in memory) of the current lu_dirent
7471 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7474 struct osd_it_ea *it = (struct osd_it_ea *)di;
7476 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7480 * Returns a cookie for current position of the iterator head, so that
7481 * user can use this cookie to load/start the iterator next time.
7483 * \param di iterator's in memory structure
7485 * \retval cookie for current position, on success
7487 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7489 struct osd_it_ea *it = (struct osd_it_ea *)di;
7491 return it->oie_dirent->oied_off;
7495 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7496 * to load a directory entry at a time and stored it i inn,
7497 * in iterator's in-memory data structure.
7499 * \param di struct osd_it_ea, iterator's in memory structure
7501 * \retval +ve on success
7502 * \retval -ve on error
7504 static int osd_it_ea_load(const struct lu_env *env,
7505 const struct dt_it *di, __u64 hash)
7507 struct osd_it_ea *it = (struct osd_it_ea *)di;
7511 it->oie_file.f_pos = hash;
7513 rc = osd_ldiskfs_it_fill(env, di);
7524 * Index lookup function for interoperability mode (b11826).
7526 * \param key, key i.e. file name to be searched
7528 * \retval +ve, on success
7529 * \retval -ve, on error
7531 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7532 struct dt_rec *rec, const struct dt_key *key)
7534 struct osd_object *obj = osd_dt_obj(dt);
7539 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7540 LINVRNT(osd_invariant(obj));
7542 rc = osd_ea_lookup_rec(env, obj, rec, key);
7549 * Index and Iterator operations for interoperability
7550 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7552 static const struct dt_index_operations osd_index_ea_ops = {
7553 .dio_lookup = osd_index_ea_lookup,
7554 .dio_declare_insert = osd_index_declare_ea_insert,
7555 .dio_insert = osd_index_ea_insert,
7556 .dio_declare_delete = osd_index_declare_ea_delete,
7557 .dio_delete = osd_index_ea_delete,
7559 .init = osd_it_ea_init,
7560 .fini = osd_it_ea_fini,
7561 .get = osd_it_ea_get,
7562 .put = osd_it_ea_put,
7563 .next = osd_it_ea_next,
7564 .key = osd_it_ea_key,
7565 .key_size = osd_it_ea_key_size,
7566 .rec = osd_it_ea_rec,
7567 .rec_size = osd_it_ea_rec_size,
7568 .store = osd_it_ea_store,
7569 .load = osd_it_ea_load
7573 static void *osd_key_init(const struct lu_context *ctx,
7574 struct lu_context_key *key)
7576 struct osd_thread_info *info;
7578 OBD_ALLOC_PTR(info);
7580 return ERR_PTR(-ENOMEM);
7582 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7583 if (info->oti_it_ea_buf == NULL)
7586 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7588 info->oti_hlock = ldiskfs_htree_lock_alloc();
7589 if (info->oti_hlock == NULL)
7595 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7598 return ERR_PTR(-ENOMEM);
7601 static void osd_key_fini(const struct lu_context *ctx,
7602 struct lu_context_key *key, void *data)
7604 struct osd_thread_info *info = data;
7605 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7606 struct osd_idmap_cache *idc = info->oti_ins_cache;
7608 if (info->oti_dio_pages) {
7610 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7611 struct page *page = info->oti_dio_pages[i];
7613 LASSERT(PagePrivate2(page));
7614 LASSERT(PageLocked(page));
7615 ClearPagePrivate2(page);
7620 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7621 PTLRPC_MAX_BRW_PAGES);
7624 if (info->oti_inode != NULL)
7626 if (info->oti_hlock != NULL)
7627 ldiskfs_htree_lock_free(info->oti_hlock);
7628 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7629 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7630 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7631 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7632 lu_buf_free(&info->oti_big_buf);
7634 LASSERT(info->oti_ins_cache_size > 0);
7635 OBD_FREE_PTR_ARRAY(idc, info->oti_ins_cache_size);
7636 info->oti_ins_cache = NULL;
7637 info->oti_ins_cache_size = 0;
7642 static void osd_key_exit(const struct lu_context *ctx,
7643 struct lu_context_key *key, void *data)
7645 struct osd_thread_info *info = data;
7647 LASSERT(info->oti_r_locks == 0);
7648 LASSERT(info->oti_w_locks == 0);
7649 LASSERT(info->oti_txns == 0);
7652 /* type constructor/destructor: osd_type_init, osd_type_fini */
7653 LU_TYPE_INIT_FINI(osd, &osd_key);
7655 struct lu_context_key osd_key = {
7656 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7657 .lct_init = osd_key_init,
7658 .lct_fini = osd_key_fini,
7659 .lct_exit = osd_key_exit
7663 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7664 const char *name, struct lu_device *next)
7666 struct osd_device *osd = osd_dev(d);
7668 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7669 sizeof(osd->od_svname))
7671 return osd_procfs_init(osd, name);
7674 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7676 struct seq_server_site *ss = osd_seq_site(osd);
7681 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7684 if (unlikely(ss == NULL))
7687 OBD_ALLOC_PTR(osd->od_cl_seq);
7688 if (osd->od_cl_seq == NULL)
7691 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7692 osd->od_svname, ss->ss_server_seq);
7694 if (ss->ss_node_id == 0) {
7696 * If the OSD on the sequence controller(MDT0), then allocate
7697 * sequence here, otherwise allocate sequence after connected
7698 * to MDT0 (see mdt_register_lwp_callback()).
7700 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7701 &osd->od_cl_seq->lcs_space, env);
7707 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7709 if (osd->od_cl_seq == NULL)
7712 seq_client_fini(osd->od_cl_seq);
7713 OBD_FREE_PTR(osd->od_cl_seq);
7714 osd->od_cl_seq = NULL;
7717 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7721 /* shutdown quota slave instance associated with the device */
7722 if (o->od_quota_slave_md != NULL) {
7723 struct qsd_instance *qsd = o->od_quota_slave_md;
7725 o->od_quota_slave_md = NULL;
7729 if (o->od_quota_slave_dt != NULL) {
7730 struct qsd_instance *qsd = o->od_quota_slave_dt;
7732 o->od_quota_slave_dt = NULL;
7736 osd_fid_fini(env, o);
7737 osd_scrub_cleanup(env, o);
7742 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7746 if (o->od_mnt != NULL) {
7747 shrink_dcache_sb(osd_sb(o));
7748 osd_sync(env, &o->od_dt_dev);
7749 wait_event(o->od_commit_cb_done,
7750 !atomic_read(&o->od_commit_cb_in_flight));
7759 static int osd_mount(const struct lu_env *env,
7760 struct osd_device *o, struct lustre_cfg *cfg)
7762 const char *name = lustre_cfg_string(cfg, 0);
7763 const char *dev = lustre_cfg_string(cfg, 1);
7765 unsigned long page, s_flags = 0, lmd_flags = 0;
7766 struct page *__page;
7767 struct file_system_type *type;
7768 char *options = NULL;
7770 struct osd_thread_info *info = osd_oti_get(env);
7771 struct lu_fid *fid = &info->oti_fid;
7772 struct inode *inode;
7773 int rc = 0, force_over_1024tb = 0;
7777 if (o->od_mnt != NULL)
7780 if (strlen(dev) >= sizeof(o->od_mntdev))
7782 strcpy(o->od_mntdev, dev);
7784 str = lustre_cfg_buf(cfg, 2);
7785 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
7787 opts = lustre_cfg_string(cfg, 3);
7789 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7790 CERROR("%s: device %s extents feature is not guaranteed to "
7791 "work on big-endian systems. Use \"bigendian_extents\" "
7792 "mount option to override.\n", name, dev);
7796 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7797 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7798 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");
7801 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7802 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7803 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");
7806 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7807 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
7808 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");
7812 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
7813 force_over_1024tb = 1;
7815 __page = alloc_page(GFP_KERNEL);
7817 GOTO(out, rc = -ENOMEM);
7818 page = (unsigned long)page_address(__page);
7819 options = (char *)page;
7822 /* strip out the options for back compatiblity */
7823 static const char * const sout[] = {
7830 /* strip out option we processed in osd */
7831 "bigendian_extents",
7835 "force_over_1024tb",
7839 strncat(options, opts, PAGE_SIZE);
7840 for (rc = 0, str = options; sout[rc]; ) {
7841 char *op = strstr(str, sout[rc]);
7848 if (op == options || *(op - 1) == ',') {
7849 str = op + strlen(sout[rc]);
7850 if (*str == ',' || *str == '\0') {
7851 *str == ',' ? str++ : str;
7852 memmove(op, str, strlen(str) + 1);
7855 for (str = op; *str != ',' && *str != '\0'; str++)
7859 strncat(options, "user_xattr,acl", PAGE_SIZE);
7862 /* Glom up mount options */
7863 if (*options != '\0')
7864 strncat(options, ",", PAGE_SIZE);
7865 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7867 type = get_fs_type("ldiskfs");
7869 CERROR("%s: cannot find ldiskfs module\n", name);
7870 GOTO(out, rc = -ENODEV);
7873 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7874 module_put(type->owner);
7876 if (IS_ERR(o->od_mnt)) {
7877 rc = PTR_ERR(o->od_mnt);
7879 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7883 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7884 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
7885 force_over_1024tb == 0) {
7886 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",
7888 GOTO(out_mnt, rc = -EINVAL);
7891 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7892 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
7895 GOTO(out_mnt, rc = -EOPNOTSUPP);
7898 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7899 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7900 GOTO(out_mnt, rc = -EINVAL);
7903 #ifdef LDISKFS_MOUNT_DIRDATA
7904 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7905 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7906 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7907 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7908 "enabling the dirdata feature. If you do not want to "
7909 "downgrade to Lustre-1.x again, you can enable it via "
7910 "'tune2fs -O dirdata device'\n", name, dev);
7912 inode = osd_sb(o)->s_root->d_inode;
7913 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7914 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7916 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7920 if (lmd_flags & LMD_FLG_NOSCRUB)
7921 o->od_auto_scrub_interval = AS_NEVER;
7923 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7924 /* do not use pagecache with flash-backed storage */
7925 o->od_writethrough_cache = 0;
7926 o->od_read_cache = 0;
7937 __free_page(__page);
7942 static struct lu_device *osd_device_fini(const struct lu_env *env,
7943 struct lu_device *d)
7945 struct osd_device *o = osd_dev(d);
7949 osd_index_backup(env, o, false);
7950 osd_shutdown(env, o);
7952 if (o->od_oi_table != NULL)
7953 osd_oi_fini(osd_oti_get(env), o);
7954 if (o->od_extent_bytes_percpu)
7955 free_percpu(o->od_extent_bytes_percpu);
7956 osd_obj_map_fini(o);
7962 static int osd_device_init0(const struct lu_env *env,
7963 struct osd_device *o,
7964 struct lustre_cfg *cfg)
7966 struct lu_device *l = osd2lu_dev(o);
7967 struct osd_thread_info *info;
7970 bool restored = false;
7973 /* if the module was re-loaded, env can loose its keys */
7974 rc = lu_env_refill((struct lu_env *)env);
7977 info = osd_oti_get(env);
7980 l->ld_ops = &osd_lu_ops;
7981 o->od_dt_dev.dd_ops = &osd_dt_ops;
7983 spin_lock_init(&o->od_osfs_lock);
7984 mutex_init(&o->od_otable_mutex);
7985 INIT_LIST_HEAD(&o->od_orphan_list);
7986 INIT_LIST_HEAD(&o->od_index_backup_list);
7987 INIT_LIST_HEAD(&o->od_index_restore_list);
7988 spin_lock_init(&o->od_lock);
7989 o->od_index_backup_policy = LIBP_NONE;
7991 init_waitqueue_head(&o->od_commit_cb_done);
7993 o->od_read_cache = 1;
7994 o->od_writethrough_cache = 1;
7995 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
7996 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
7997 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
7998 o->od_auto_scrub_interval = AS_DEFAULT;
7999 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8000 o->od_fallocate_zero_blocks = 0;
8002 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
8003 sizeof(o->od_svname));
8004 if (cplen >= sizeof(o->od_svname)) {
8009 o->od_index_backup_stop = 0;
8010 o->od_index = -1; /* -1 means index is invalid */
8011 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8012 if (rc == LDD_F_SV_TYPE_OST)
8015 o->od_full_scrub_ratio = OFSR_DEFAULT;
8016 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8017 rc = osd_mount(env, o, cfg);
8021 /* Can only check block device after mount */
8022 o->od_nonrotational =
8023 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8025 rc = osd_obj_map_init(env, o);
8029 rc = lu_site_init(&o->od_site, l);
8031 GOTO(out_compat, rc);
8032 o->od_site.ls_bottom_dev = l;
8034 rc = lu_site_init_finish(&o->od_site);
8038 opts = lustre_cfg_string(cfg, 3);
8039 if (opts && strstr(opts, "resetoi"))
8042 INIT_LIST_HEAD(&o->od_ios_list);
8043 /* setup scrub, including OI files initialization */
8045 rc = osd_scrub_setup(env, o, restored);
8050 rc = osd_procfs_init(o, o->od_svname);
8052 CERROR("%s: can't initialize procfs: rc = %d\n",
8054 GOTO(out_scrub, rc);
8057 LASSERT(l->ld_site->ls_linkage.next != NULL);
8058 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8060 /* initialize quota slave instance */
8061 /* currently it's no need to prepare qsd_instance_md for OST */
8062 if (!o->od_is_ost) {
8063 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8065 o->od_proc_entry, true);
8066 if (IS_ERR(o->od_quota_slave_md)) {
8067 rc = PTR_ERR(o->od_quota_slave_md);
8068 o->od_quota_slave_md = NULL;
8069 GOTO(out_procfs, rc);
8073 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8074 o->od_proc_entry, false);
8076 if (IS_ERR(o->od_quota_slave_dt)) {
8077 if (o->od_quota_slave_md != NULL) {
8078 qsd_fini(env, o->od_quota_slave_md);
8079 o->od_quota_slave_md = NULL;
8082 rc = PTR_ERR(o->od_quota_slave_dt);
8083 o->od_quota_slave_dt = NULL;
8084 GOTO(out_procfs, rc);
8087 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8088 if (!o->od_extent_bytes_percpu) {
8090 GOTO(out_procfs, rc);
8098 osd_scrub_cleanup(env, o);
8100 lu_site_fini(&o->od_site);
8102 osd_obj_map_fini(o);
8109 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8110 struct lu_device_type *t,
8111 struct lustre_cfg *cfg)
8113 struct osd_device *o;
8118 return ERR_PTR(-ENOMEM);
8120 rc = dt_device_init(&o->od_dt_dev, t);
8123 * Because the ctx might be revived in dt_device_init,
8124 * refill the env here
8126 lu_env_refill((struct lu_env *)env);
8127 rc = osd_device_init0(env, o, cfg);
8129 dt_device_fini(&o->od_dt_dev);
8132 if (unlikely(rc != 0))
8135 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8138 static struct lu_device *osd_device_free(const struct lu_env *env,
8139 struct lu_device *d)
8141 struct osd_device *o = osd_dev(d);
8145 /* XXX: make osd top device in order to release reference */
8146 d->ld_site->ls_top_dev = d;
8147 lu_site_purge(env, d->ld_site, -1);
8148 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8149 D_ERROR, lu_cdebug_printer);
8150 lu_site_fini(&o->od_site);
8151 dt_device_fini(&o->od_dt_dev);
8156 static int osd_process_config(const struct lu_env *env,
8157 struct lu_device *d, struct lustre_cfg *cfg)
8159 struct osd_device *o = osd_dev(d);
8165 switch (cfg->lcfg_command) {
8167 rc = osd_mount(env, o, cfg);
8171 * For the case LCFG_PRE_CLEANUP is not called in advance,
8172 * that may happend if hit failure during mount process.
8174 osd_index_backup(env, o, false);
8175 lu_dev_del_linkage(d->ld_site, d);
8176 rc = osd_shutdown(env, o);
8179 LASSERT(&o->od_dt_dev);
8180 count = class_modify_config(cfg, PARAM_OSD,
8181 &o->od_dt_dev.dd_kobj);
8183 count = class_modify_config(cfg, PARAM_OST,
8184 &o->od_dt_dev.dd_kobj);
8185 rc = count > 0 ? 0 : count;
8187 case LCFG_PRE_CLEANUP:
8189 osd_index_backup(env, o,
8190 o->od_index_backup_policy != LIBP_NONE);
8200 static int osd_recovery_complete(const struct lu_env *env,
8201 struct lu_device *d)
8203 struct osd_device *osd = osd_dev(d);
8208 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8212 * start qsd instance on recovery completion, this notifies the quota
8213 * slave code that we are about to process new requests now
8215 rc = qsd_start(env, osd->od_quota_slave_dt);
8216 if (rc == 0 && osd->od_quota_slave_md != NULL)
8217 rc = qsd_start(env, osd->od_quota_slave_md);
8223 * we use exports to track all osd users
8225 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8226 struct obd_device *obd, struct obd_uuid *cluuid,
8227 struct obd_connect_data *data, void *localdata)
8229 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8230 struct lustre_handle conn;
8235 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8237 rc = class_connect(&conn, obd, cluuid);
8241 *exp = class_conn2export(&conn);
8243 spin_lock(&osd->od_osfs_lock);
8245 spin_unlock(&osd->od_osfs_lock);
8251 * once last export (we don't count self-export) disappeared
8252 * osd can be released
8254 static int osd_obd_disconnect(struct obd_export *exp)
8256 struct obd_device *obd = exp->exp_obd;
8257 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8258 int rc, release = 0;
8262 /* Only disconnect the underlying layers on the final disconnect. */
8263 spin_lock(&osd->od_osfs_lock);
8265 if (osd->od_connects == 0)
8267 spin_unlock(&osd->od_osfs_lock);
8269 rc = class_disconnect(exp); /* bz 9811 */
8271 if (rc == 0 && release)
8272 class_manual_cleanup(obd);
8276 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8277 struct lu_device *dev)
8279 struct osd_device *osd = osd_dev(dev);
8280 struct lr_server_data *lsd =
8281 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8286 if (osd->od_quota_slave_md != NULL) {
8287 /* set up quota slave objects for inode */
8288 result = qsd_prepare(env, osd->od_quota_slave_md);
8293 if (osd->od_quota_slave_dt != NULL) {
8294 /* set up quota slave objects for block */
8295 result = qsd_prepare(env, osd->od_quota_slave_dt);
8301 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8302 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8303 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8304 osd->od_index_in_idif = 1;
8306 osd->od_index_in_idif = 0;
8307 result = osd_register_proc_index_in_idif(osd);
8312 osd->od_index_in_idif = 1;
8316 result = osd_fid_init(env, osd);
8322 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8326 * see include/lu_object.h for the details.
8328 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8329 struct lu_fid *fid, struct lu_object *parent,
8330 const struct lu_name *name)
8332 struct osd_device *osd = osd_dev(d);
8334 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8337 static const struct lu_object_operations osd_lu_obj_ops = {
8338 .loo_object_init = osd_object_init,
8339 .loo_object_delete = osd_object_delete,
8340 .loo_object_release = osd_object_release,
8341 .loo_object_free = osd_object_free,
8342 .loo_object_print = osd_object_print,
8343 .loo_object_invariant = osd_object_invariant
8346 const struct lu_device_operations osd_lu_ops = {
8347 .ldo_object_alloc = osd_object_alloc,
8348 .ldo_process_config = osd_process_config,
8349 .ldo_recovery_complete = osd_recovery_complete,
8350 .ldo_prepare = osd_prepare,
8351 .ldo_fid_alloc = osd_fid_alloc,
8354 static const struct lu_device_type_operations osd_device_type_ops = {
8355 .ldto_init = osd_type_init,
8356 .ldto_fini = osd_type_fini,
8358 .ldto_start = osd_type_start,
8359 .ldto_stop = osd_type_stop,
8361 .ldto_device_alloc = osd_device_alloc,
8362 .ldto_device_free = osd_device_free,
8364 .ldto_device_init = osd_device_init,
8365 .ldto_device_fini = osd_device_fini
8368 static struct lu_device_type osd_device_type = {
8369 .ldt_tags = LU_DEVICE_DT,
8370 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8371 .ldt_ops = &osd_device_type_ops,
8372 .ldt_ctx_tags = LCT_LOCAL,
8375 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8377 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8378 struct super_block *sb = osd_sb(osd);
8380 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8384 * lprocfs legacy support.
8386 static const struct obd_ops osd_obd_device_ops = {
8387 .o_owner = THIS_MODULE,
8388 .o_connect = osd_obd_connect,
8389 .o_disconnect = osd_obd_disconnect,
8390 .o_health_check = osd_health_check,
8393 static ssize_t track_declares_assert_show(struct kobject *kobj,
8394 struct attribute *attr,
8397 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8400 static ssize_t track_declares_assert_store(struct kobject *kobj,
8401 struct attribute *attr,
8402 const char *buffer, size_t count)
8404 bool track_declares_assert;
8407 rc = kstrtobool(buffer, &track_declares_assert);
8411 ldiskfs_track_declares_assert = track_declares_assert;
8415 LUSTRE_RW_ATTR(track_declares_assert);
8417 static int __init osd_init(void)
8419 struct kobject *kobj;
8422 BUILD_BUG_ON(BH_DXLock >=
8423 sizeof(((struct buffer_head *)0)->b_state) * 8);
8424 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8425 /* please, try to keep osd_thread_info smaller than a page */
8426 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8431 rc = lu_kmem_init(ldiskfs_caches);
8435 #ifdef CONFIG_KALLSYMS
8436 priv_security_file_alloc =
8437 (void *)cfs_kallsyms_lookup_name("security_file_alloc");
8440 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8441 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8443 lu_kmem_fini(ldiskfs_caches);
8447 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8449 rc = sysfs_create_file(kobj,
8450 &lustre_attr_track_declares_assert.attr);
8453 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8460 static void __exit osd_exit(void)
8462 struct kobject *kobj;
8464 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8466 sysfs_remove_file(kobj,
8467 &lustre_attr_track_declares_assert.attr);
8470 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8471 lu_kmem_fini(ldiskfs_caches);
8474 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8475 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8476 MODULE_VERSION(LUSTRE_VERSION_STRING);
8477 MODULE_LICENSE("GPL");
8479 module_init(osd_init);
8480 module_exit(osd_exit);