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 /* reserve or free quota for some operation */
2576 static int osd_reserve_or_free_quota(const struct lu_env *env,
2577 struct dt_device *dev,
2578 enum quota_type type, __u64 uid,
2579 __u64 gid, __s64 count, bool is_md)
2582 struct osd_device *osd = osd_dt_dev(dev);
2583 struct osd_thread_info *info = osd_oti_get(env);
2584 struct lquota_id_info *qi = &info->oti_qi;
2585 struct qsd_instance *qsd = NULL;
2590 qsd = osd->od_quota_slave_md;
2592 qsd = osd->od_quota_slave_dt;
2594 rc = quota_reserve_or_free(env, qsd, qi, type, uid, gid, count, is_md);
2598 static const struct dt_device_operations osd_dt_ops = {
2599 .dt_root_get = osd_root_get,
2600 .dt_statfs = osd_statfs,
2601 .dt_trans_create = osd_trans_create,
2602 .dt_trans_start = osd_trans_start,
2603 .dt_trans_stop = osd_trans_stop,
2604 .dt_trans_cb_add = osd_trans_cb_add,
2605 .dt_conf_get = osd_conf_get,
2606 .dt_mnt_sb_get = osd_mnt_sb_get,
2607 .dt_sync = osd_sync,
2609 .dt_commit_async = osd_commit_async,
2610 .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
2613 static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
2616 struct osd_object *obj = osd_dt_obj(dt);
2617 struct osd_thread_info *oti = osd_oti_get(env);
2619 LINVRNT(osd_invariant(obj));
2621 LASSERT(obj->oo_owner != env);
2622 down_read_nested(&obj->oo_sem, role);
2624 LASSERT(obj->oo_owner == NULL);
2628 static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
2631 struct osd_object *obj = osd_dt_obj(dt);
2632 struct osd_thread_info *oti = osd_oti_get(env);
2634 LINVRNT(osd_invariant(obj));
2636 LASSERT(obj->oo_owner != env);
2637 down_write_nested(&obj->oo_sem, role);
2639 LASSERT(obj->oo_owner == NULL);
2640 obj->oo_owner = env;
2644 static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
2646 struct osd_object *obj = osd_dt_obj(dt);
2647 struct osd_thread_info *oti = osd_oti_get(env);
2649 LINVRNT(osd_invariant(obj));
2651 LASSERT(oti->oti_r_locks > 0);
2653 up_read(&obj->oo_sem);
2656 static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
2658 struct osd_object *obj = osd_dt_obj(dt);
2659 struct osd_thread_info *oti = osd_oti_get(env);
2661 LINVRNT(osd_invariant(obj));
2663 LASSERT(obj->oo_owner == env);
2664 LASSERT(oti->oti_w_locks > 0);
2666 obj->oo_owner = NULL;
2667 up_write(&obj->oo_sem);
2670 static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
2672 struct osd_object *obj = osd_dt_obj(dt);
2674 LINVRNT(osd_invariant(obj));
2676 return obj->oo_owner == env;
2679 static void osd_inode_getattr(const struct lu_env *env,
2680 struct inode *inode, struct lu_attr *attr)
2682 attr->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
2683 LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
2684 LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
2685 LA_BLKSIZE | LA_TYPE | LA_BTIME;
2687 attr->la_atime = inode->i_atime.tv_sec;
2688 attr->la_mtime = inode->i_mtime.tv_sec;
2689 attr->la_ctime = inode->i_ctime.tv_sec;
2690 attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
2691 attr->la_mode = inode->i_mode;
2692 attr->la_size = i_size_read(inode);
2693 attr->la_blocks = inode->i_blocks;
2694 attr->la_uid = i_uid_read(inode);
2695 attr->la_gid = i_gid_read(inode);
2696 attr->la_projid = i_projid_read(inode);
2697 attr->la_flags = ll_inode_to_ext_flags(inode->i_flags);
2698 attr->la_nlink = inode->i_nlink;
2699 attr->la_rdev = inode->i_rdev;
2700 attr->la_blksize = 1 << inode->i_blkbits;
2701 attr->la_blkbits = inode->i_blkbits;
2703 * Ext4 did not transfer inherit flags from raw inode
2704 * to inode flags, and ext4 internally test raw inode
2705 * @i_flags directly. Instead of patching ext4, we do it here.
2707 if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
2708 attr->la_flags |= LUSTRE_PROJINHERIT_FL;
2711 static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
2714 struct osd_object *obj = osd_dt_obj(dt);
2715 const struct dt_it_ops *iops;
2721 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
2722 LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));
2724 if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
2725 *count = obj->oo_dirent_count;
2729 /* directory not initialized yet */
2730 if (!dt->do_index_ops) {
2735 iops = &dt->do_index_ops->dio_it;
2736 it = iops->init(env, dt, LUDA_64BITHASH);
2738 RETURN(PTR_ERR(it));
2740 rc = iops->load(env, it, 0);
2742 if (rc == -ENODATA) {
2749 rc = iops->next(env, it);
2751 for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
2755 if (iops->key_size(env, it) == 0)
2761 obj->oo_dirent_count = *count;
2766 iops->fini(env, it);
2771 static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
2772 struct lu_attr *attr)
2774 struct osd_object *obj = osd_dt_obj(dt);
2777 if (unlikely(!dt_object_exists(dt)))
2779 if (unlikely(obj->oo_destroyed))
2782 LASSERT(!dt_object_remote(dt));
2783 LINVRNT(osd_invariant(obj));
2785 spin_lock(&obj->oo_guard);
2786 osd_inode_getattr(env, obj->oo_inode, attr);
2787 if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
2788 attr->la_valid |= LA_FLAGS;
2789 attr->la_flags |= LUSTRE_ORPHAN_FL;
2791 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
2792 attr->la_valid |= LA_FLAGS;
2793 attr->la_flags |= LUSTRE_ENCRYPT_FL;
2795 spin_unlock(&obj->oo_guard);
2797 if (S_ISDIR(obj->oo_inode->i_mode) &&
2798 fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
2799 rc = osd_dirent_count(env, dt, &attr->la_dirent_count);
2804 static int osd_declare_attr_qid(const struct lu_env *env,
2805 struct osd_object *obj,
2806 struct osd_thandle *oh, long long bspace,
2807 qid_t old_id, qid_t new_id, bool enforce,
2811 struct osd_thread_info *info = osd_oti_get(env);
2812 struct lquota_id_info *qi = &info->oti_qi;
2814 qi->lqi_type = type;
2815 /* inode accounting */
2816 qi->lqi_is_blk = false;
2818 /* one more inode for the new id ... */
2819 qi->lqi_id.qid_uid = new_id;
2821 /* Reserve credits for the new id */
2822 rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
2823 if (rc == -EDQUOT || rc == -EINPROGRESS)
2828 /* and one less inode for the current id */
2829 qi->lqi_id.qid_uid = old_id;
2831 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2832 if (rc == -EDQUOT || rc == -EINPROGRESS)
2837 /* block accounting */
2838 qi->lqi_is_blk = true;
2840 /* more blocks for the new id ... */
2841 qi->lqi_id.qid_uid = new_id;
2842 qi->lqi_space = bspace;
2844 * Credits for the new uid has been reserved, re-use "obj"
2845 * to save credit reservation.
2847 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2848 if (rc == -EDQUOT || rc == -EINPROGRESS)
2853 /* and finally less blocks for the current uid */
2854 qi->lqi_id.qid_uid = old_id;
2855 qi->lqi_space = -bspace;
2856 rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
2857 if (rc == -EDQUOT || rc == -EINPROGRESS)
2863 static int osd_declare_attr_set(const struct lu_env *env,
2864 struct dt_object *dt,
2865 const struct lu_attr *attr,
2866 struct thandle *handle)
2868 struct osd_thandle *oh;
2869 struct osd_object *obj;
2878 LASSERT(dt != NULL);
2879 LASSERT(handle != NULL);
2881 obj = osd_dt_obj(dt);
2882 LASSERT(osd_invariant(obj));
2884 oh = container_of(handle, struct osd_thandle, ot_super);
2885 LASSERT(oh->ot_handle == NULL);
2887 osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
2888 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
2890 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
2891 osd_dto_credits_noquota[DTO_XATTR_SET]);
2893 if (attr == NULL || obj->oo_inode == NULL)
2896 bspace = obj->oo_inode->i_blocks << 9;
2897 bspace = toqb(bspace);
2900 * Changing ownership is always preformed by super user, it should not
2901 * fail with EDQUOT unless required explicitly.
2903 * We still need to call the osd_declare_qid() to calculate the journal
2904 * credits for updating quota accounting files and to trigger quota
2905 * space adjustment once the operation is completed.
2907 if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
2909 uid = i_uid_read(obj->oo_inode);
2910 enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
2911 rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
2912 attr->la_uid, enforce, USRQUOTA);
2916 gid = i_gid_read(obj->oo_inode);
2917 CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
2918 attr->la_uid, gid, attr->la_gid);
2919 enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
2920 rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
2921 attr->la_gid, enforce, GRPQUOTA);
2926 #ifdef HAVE_PROJECT_QUOTA
2927 if (attr->la_valid & LA_PROJID) {
2928 __u32 projid = i_projid_read(obj->oo_inode);
2930 enforce = (attr->la_valid & LA_PROJID) &&
2931 (attr->la_projid != projid);
2932 rc = osd_declare_attr_qid(env, obj, oh, bspace,
2933 (qid_t)projid, (qid_t)attr->la_projid,
2942 static int osd_inode_setattr(const struct lu_env *env,
2943 struct inode *inode, const struct lu_attr *attr)
2945 __u64 bits = attr->la_valid;
2947 /* Only allow set size for regular file */
2948 if (!S_ISREG(inode->i_mode))
2949 bits &= ~(LA_SIZE | LA_BLOCKS);
2954 if (bits & LA_ATIME)
2955 inode->i_atime = osd_inode_time(inode, attr->la_atime);
2956 if (bits & LA_CTIME)
2957 inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
2958 if (bits & LA_MTIME)
2959 inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
2960 if (bits & LA_SIZE) {
2961 spin_lock(&inode->i_lock);
2962 LDISKFS_I(inode)->i_disksize = attr->la_size;
2963 i_size_write(inode, attr->la_size);
2964 spin_unlock(&inode->i_lock);
2968 * OSD should not change "i_blocks" which is used by quota.
2969 * "i_blocks" should be changed by ldiskfs only.
2972 inode->i_mode = (inode->i_mode & S_IFMT) |
2973 (attr->la_mode & ~S_IFMT);
2975 i_uid_write(inode, attr->la_uid);
2977 i_gid_write(inode, attr->la_gid);
2978 if (bits & LA_PROJID)
2979 i_projid_write(inode, attr->la_projid);
2980 if (bits & LA_NLINK)
2981 set_nlink(inode, attr->la_nlink);
2983 inode->i_rdev = attr->la_rdev;
2985 if (bits & LA_FLAGS) {
2986 /* always keep S_NOCMTIME */
2987 inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
2989 #if defined(S_ENCRYPTED)
2990 /* Always remove S_ENCRYPTED, because ldiskfs must not be
2991 * aware of encryption status. It is just stored into LMA
2992 * so that it can be forwared to client side.
2994 inode->i_flags &= ~S_ENCRYPTED;
2997 * Ext4 did not transfer inherit flags from
2998 * @inode->i_flags to raw inode i_flags when writing
2999 * flags, we do it explictly here.
3001 if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
3002 LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
3004 LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
3009 #ifdef HAVE_PROJECT_QUOTA
3010 static int osd_transfer_project(struct inode *inode, __u32 projid,
3011 struct thandle *handle)
3013 struct super_block *sb = inode->i_sb;
3014 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
3017 struct ldiskfs_iloc iloc;
3018 struct ldiskfs_inode *raw_inode;
3019 struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };
3021 if (!ldiskfs_has_feature_project(sb)) {
3022 LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
3023 == LDISKFS_DEF_PROJID);
3024 if (projid != LDISKFS_DEF_PROJID)
3030 if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
3033 kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
3034 if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
3037 err = ldiskfs_get_inode_loc(inode, &iloc);
3041 raw_inode = ldiskfs_raw_inode(&iloc);
3042 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3043 struct osd_thandle *oh = container_of(handle,
3047 * try to expand inode size automatically.
3049 ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
3050 if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
3058 dquot_initialize(inode);
3059 transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
3060 if (transfer_to[PRJQUOTA]) {
3061 err = __dquot_transfer(inode, transfer_to);
3062 dqput(transfer_to[PRJQUOTA]);
3071 static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
3072 struct thandle *handle)
3076 if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
3077 (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
3081 "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
3082 inode->i_ino, i_uid_read(inode), attr->la_uid,
3083 i_gid_read(inode), attr->la_gid);
3085 dquot_initialize(inode);
3087 if (attr->la_valid & LA_UID)
3088 iattr.ia_valid |= ATTR_UID;
3089 if (attr->la_valid & LA_GID)
3090 iattr.ia_valid |= ATTR_GID;
3091 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3092 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3094 rc = dquot_transfer(inode, &iattr);
3096 CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3097 osd_ino2name(inode), rc);
3102 /* Handle project id transfer here properly */
3103 if (attr->la_valid & LA_PROJID &&
3104 attr->la_projid != i_projid_read(inode)) {
3105 if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
3107 #ifdef HAVE_PROJECT_QUOTA
3108 rc = osd_transfer_project(inode, attr->la_projid, handle);
3113 CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
3114 osd_ino2name(inode), rc);
3121 static int osd_attr_set(const struct lu_env *env,
3122 struct dt_object *dt,
3123 const struct lu_attr *attr,
3124 struct thandle *handle)
3126 struct osd_object *obj = osd_dt_obj(dt);
3127 struct inode *inode;
3130 if (!dt_object_exists(dt))
3133 LASSERT(handle != NULL);
3134 LASSERT(!dt_object_remote(dt));
3135 LASSERT(osd_invariant(obj));
3137 osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);
3139 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
3140 !osd_obj2dev(obj)->od_is_ost) {
3141 struct osd_thread_info *oti = osd_oti_get(env);
3142 const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
3143 struct lu_fid *fid1 = &oti->oti_fid;
3144 struct osd_inode_id *id = &oti->oti_id;
3145 struct iam_path_descr *ipd;
3146 struct iam_container *bag;
3147 struct osd_thandle *oh;
3150 fid_cpu_to_be(fid1, fid0);
3151 memset(id, 1, sizeof(*id));
3152 bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
3153 fid0)->oi_dir.od_container;
3154 ipd = osd_idx_ipd_get(env, bag);
3155 if (unlikely(ipd == NULL))
3158 oh = container_of(handle, struct osd_thandle, ot_super);
3159 rc = iam_update(oh->ot_handle, bag,
3160 (const struct iam_key *)fid1,
3161 (const struct iam_rec *)id, ipd);
3162 osd_ipd_put(env, bag, ipd);
3163 return(rc > 0 ? 0 : rc);
3166 inode = obj->oo_inode;
3168 rc = osd_quota_transfer(inode, attr, handle);
3172 spin_lock(&obj->oo_guard);
3173 rc = osd_inode_setattr(env, inode, attr);
3174 spin_unlock(&obj->oo_guard);
3178 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3180 osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);
3182 if (!(attr->la_valid & LA_FLAGS))
3185 /* Let's check if there are extra flags need to be set into LMA */
3186 if (attr->la_flags & LUSTRE_LMA_FL_MASKS) {
3187 struct osd_thread_info *info = osd_oti_get(env);
3188 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
3190 LASSERT(!obj->oo_pfid_in_lma);
3192 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
3193 &info->oti_ost_attrs);
3197 lma->lma_incompat |=
3198 lustre_to_lma_flags(attr->la_flags);
3199 lustre_lma_swab(lma);
3201 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
3203 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA,
3204 lma, sizeof(*lma), XATTR_REPLACE);
3206 struct osd_device *osd = osd_obj2dev(obj);
3208 CWARN("%s: set "DFID" lma flags %u failed: rc = %d\n",
3209 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)),
3210 lma->lma_incompat, rc);
3213 attr->la_flags & LUSTRE_LMA_FL_MASKS;
3215 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
3222 static struct dentry *osd_child_dentry_get(const struct lu_env *env,
3223 struct osd_object *obj,
3224 const char *name, const int namelen)
3226 return osd_child_dentry_by_inode(env, obj->oo_inode, name, namelen);
3229 static int osd_mkfile(struct osd_thread_info *info, struct osd_object *obj,
3230 umode_t mode, struct dt_allocation_hint *hint,
3231 struct thandle *th, struct lu_attr *attr)
3234 struct osd_device *osd = osd_obj2dev(obj);
3235 struct osd_thandle *oth;
3236 struct dt_object *parent = NULL;
3237 struct inode *inode;
3238 struct iattr iattr = {
3239 .ia_valid = ATTR_UID | ATTR_GID |
3240 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3241 .ia_ctime.tv_sec = attr->la_ctime,
3242 .ia_mtime.tv_sec = attr->la_mtime,
3243 .ia_atime.tv_sec = attr->la_atime,
3244 .ia_uid = GLOBAL_ROOT_UID,
3245 .ia_gid = GLOBAL_ROOT_GID,
3247 const struct osd_timespec omit = { .tv_nsec = UTIME_OMIT };
3249 if (attr->la_valid & LA_UID)
3250 iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
3251 if (attr->la_valid & LA_GID)
3252 iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);
3254 LINVRNT(osd_invariant(obj));
3255 LASSERT(obj->oo_inode == NULL);
3256 LASSERT(obj->oo_hl_head == NULL);
3258 if (S_ISDIR(mode) && ldiskfs_pdo) {
3260 ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);
3261 if (obj->oo_hl_head == NULL)
3265 oth = container_of(th, struct osd_thandle, ot_super);
3266 LASSERT(oth->ot_handle->h_transaction != NULL);
3268 if (hint != NULL && hint->dah_parent != NULL &&
3269 !dt_object_remote(hint->dah_parent))
3270 parent = hint->dah_parent;
3272 /* if a time component is not valid set it to UTIME_OMIT */
3273 if (!(attr->la_valid & LA_CTIME))
3274 iattr.ia_ctime = omit;
3275 if (!(attr->la_valid & LA_MTIME))
3276 iattr.ia_mtime = omit;
3277 if (!(attr->la_valid & LA_ATIME))
3278 iattr.ia_atime = omit;
3280 inode = ldiskfs_create_inode(oth->ot_handle,
3281 parent ? osd_dt_obj(parent)->oo_inode :
3282 osd_sb(osd)->s_root->d_inode,
3284 if (!IS_ERR(inode)) {
3285 /* Do not update file c/mtime in ldiskfs. */
3286 inode->i_flags |= S_NOCMTIME;
3289 * For new created object, it must be consistent,
3290 * and it is unnecessary to scrub against it.
3292 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_NOSCRUB);
3294 obj->oo_inode = inode;
3297 if (obj->oo_hl_head != NULL) {
3298 ldiskfs_htree_lock_head_free(obj->oo_hl_head);
3299 obj->oo_hl_head = NULL;
3301 result = PTR_ERR(inode);
3303 LINVRNT(osd_invariant(obj));
3311 static int osd_mkdir(struct osd_thread_info *info, struct osd_object *obj,
3312 struct lu_attr *attr,
3313 struct dt_allocation_hint *hint,
3314 struct dt_object_format *dof,
3318 struct osd_thandle *oth;
3319 __u32 mode = (attr->la_mode & (S_IFMT | S_IRWXUGO | S_ISVTX | S_ISGID));
3321 LASSERT(S_ISDIR(attr->la_mode));
3323 oth = container_of(th, struct osd_thandle, ot_super);
3324 LASSERT(oth->ot_handle->h_transaction != NULL);
3325 if (fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)))
3326 obj->oo_dirent_count = 0;
3327 result = osd_mkfile(info, obj, mode, hint, th, attr);
3332 static int osd_mk_index(struct osd_thread_info *info, struct osd_object *obj,
3333 struct lu_attr *attr,
3334 struct dt_allocation_hint *hint,
3335 struct dt_object_format *dof,
3339 struct osd_thandle *oth;
3340 const struct dt_index_features *feat = dof->u.dof_idx.di_feat;
3342 __u32 mode = (attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX));
3344 LASSERT(S_ISREG(attr->la_mode));
3346 oth = container_of(th, struct osd_thandle, ot_super);
3347 LASSERT(oth->ot_handle->h_transaction != NULL);
3349 result = osd_mkfile(info, obj, mode, hint, th, attr);
3351 LASSERT(obj->oo_inode != NULL);
3352 if (feat->dif_flags & DT_IND_VARKEY)
3353 result = iam_lvar_create(obj->oo_inode,
3354 feat->dif_keysize_max,
3356 feat->dif_recsize_max,
3359 result = iam_lfix_create(obj->oo_inode,
3360 feat->dif_keysize_max,
3362 feat->dif_recsize_max,
3368 static int osd_mkreg(struct osd_thread_info *info, struct osd_object *obj,
3369 struct lu_attr *attr,
3370 struct dt_allocation_hint *hint,
3371 struct dt_object_format *dof,
3374 LASSERT(S_ISREG(attr->la_mode));
3375 return osd_mkfile(info, obj, (attr->la_mode &
3376 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3380 static int osd_mksym(struct osd_thread_info *info, struct osd_object *obj,
3381 struct lu_attr *attr,
3382 struct dt_allocation_hint *hint,
3383 struct dt_object_format *dof,
3386 LASSERT(S_ISLNK(attr->la_mode));
3387 return osd_mkfile(info, obj, (attr->la_mode &
3388 (S_IFMT | S_IALLUGO | S_ISVTX)), hint, th,
3392 static int osd_mknod(struct osd_thread_info *info, struct osd_object *obj,
3393 struct lu_attr *attr,
3394 struct dt_allocation_hint *hint,
3395 struct dt_object_format *dof,
3398 umode_t mode = attr->la_mode & (S_IFMT | S_IALLUGO | S_ISVTX);
3401 LINVRNT(osd_invariant(obj));
3402 LASSERT(obj->oo_inode == NULL);
3403 LASSERT(S_ISCHR(mode) || S_ISBLK(mode) ||
3404 S_ISFIFO(mode) || S_ISSOCK(mode));
3406 result = osd_mkfile(info, obj, mode, hint, th, attr);
3408 LASSERT(obj->oo_inode != NULL);
3410 * This inode should be marked dirty for i_rdev. Currently
3411 * that is done in the osd_attr_init().
3413 init_special_inode(obj->oo_inode, obj->oo_inode->i_mode,
3416 LINVRNT(osd_invariant(obj));
3420 typedef int (*osd_obj_type_f)(struct osd_thread_info *, struct osd_object *,
3422 struct dt_allocation_hint *hint,
3423 struct dt_object_format *dof,
3426 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
3428 osd_obj_type_f result;
3444 result = osd_mk_index;
3454 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
3455 struct dt_object *parent, struct dt_object *child,
3460 ah->dah_parent = parent;
3461 ah->dah_mode = child_mode;
3463 if (parent != NULL && !dt_object_remote(parent)) {
3464 /* will help to find FID->ino at dt_insert("..") */
3465 struct osd_object *pobj = osd_dt_obj(parent);
3467 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
3471 static void osd_attr_init(struct osd_thread_info *info, struct osd_object *obj,
3472 struct lu_attr *attr, struct dt_object_format *dof,
3473 struct thandle *handle)
3475 struct inode *inode = obj->oo_inode;
3476 __u64 valid = attr->la_valid;
3479 attr->la_valid &= ~(LA_TYPE | LA_MODE);
3481 if (dof->dof_type != DFT_NODE)
3482 attr->la_valid &= ~LA_RDEV;
3483 if ((valid & LA_ATIME) && (attr->la_atime == inode->i_atime.tv_sec))
3484 attr->la_valid &= ~LA_ATIME;
3485 if ((valid & LA_CTIME) && (attr->la_ctime == inode->i_ctime.tv_sec))
3486 attr->la_valid &= ~LA_CTIME;
3487 if ((valid & LA_MTIME) && (attr->la_mtime == inode->i_mtime.tv_sec))
3488 attr->la_valid &= ~LA_MTIME;
3490 result = osd_quota_transfer(inode, attr, handle);
3494 if (attr->la_valid != 0) {
3495 result = osd_inode_setattr(info->oti_env, inode, attr);
3497 * The osd_inode_setattr() should always succeed here. The
3498 * only error that could be returned is EDQUOT when we are
3499 * trying to change the UID or GID of the inode. However, this
3500 * should not happen since quota enforcement is no longer
3501 * enabled on ldiskfs (lquota takes care of it).
3503 LASSERTF(result == 0, "%d\n", result);
3504 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3507 attr->la_valid = valid;
3511 * Helper function for osd_create()
3513 * \retval 0, on success
3515 static int __osd_create(struct osd_thread_info *info, struct osd_object *obj,
3516 struct lu_attr *attr, struct dt_allocation_hint *hint,
3517 struct dt_object_format *dof, struct thandle *th)
3522 osd_trans_exec_op(info->oti_env, th, OSD_OT_CREATE);
3524 /* we drop umask so that permissions we pass are not affected */
3525 umask = current->fs->umask;
3526 current->fs->umask = 0;
3528 result = osd_create_type_f(dof->dof_type)(info, obj, attr, hint, dof,
3530 if (likely(obj->oo_inode != NULL)) {
3531 LASSERT(obj->oo_inode->i_state & I_NEW);
3534 * Unlock the inode before attr initialization to avoid
3535 * unnecessary dqget operations. LU-6378
3537 unlock_new_inode(obj->oo_inode);
3540 if (likely(result == 0)) {
3541 osd_attr_init(info, obj, attr, dof, th);
3542 osd_object_init0(obj);
3545 /* restore previous umask value */
3546 current->fs->umask = umask;
3548 osd_trans_exec_check(info->oti_env, th, OSD_OT_CREATE);
3554 * Helper function for osd_create()
3556 * \retval 0, on success
3558 static int __osd_oi_insert(const struct lu_env *env, struct osd_object *obj,
3559 const struct lu_fid *fid, struct thandle *th)
3561 struct osd_thread_info *info = osd_oti_get(env);
3562 struct osd_inode_id *id = &info->oti_id;
3563 struct osd_device *osd = osd_obj2dev(obj);
3564 struct osd_thandle *oh;
3567 LASSERT(obj->oo_inode != NULL);
3569 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_OI_ENOSPC))
3572 oh = container_of(th, struct osd_thandle, ot_super);
3573 LASSERT(oh->ot_handle);
3574 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3576 osd_id_gen(id, obj->oo_inode->i_ino, obj->oo_inode->i_generation);
3577 rc = osd_oi_insert(info, osd, fid, id, oh->ot_handle,
3578 OI_CHECK_FLD, NULL);
3579 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP) && osd->od_is_ost) {
3580 struct lu_fid next_fid = *fid;
3582 /* insert next object in advance, and map to the same inode */
3584 if (next_fid.f_oid != 0) {
3585 osd_trans_exec_op(env, th, OSD_OT_INSERT);
3586 osd_oi_insert(info, osd, &next_fid, id, oh->ot_handle,
3587 OI_CHECK_FLD, NULL);
3588 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3592 osd_trans_exec_check(env, th, OSD_OT_INSERT);
3597 int osd_fld_lookup(const struct lu_env *env, struct osd_device *osd,
3598 u64 seq, struct lu_seq_range *range)
3600 struct seq_server_site *ss = osd_seq_site(osd);
3602 if (fid_seq_is_idif(seq)) {
3603 fld_range_set_ost(range);
3604 range->lsr_index = idif_ost_idx(seq);
3608 if (!fid_seq_in_fldb(seq)) {
3609 fld_range_set_mdt(range);
3612 * FIXME: If ss is NULL, it suppose not get lsr_index
3615 range->lsr_index = ss->ss_node_id;
3619 LASSERT(ss != NULL);
3620 fld_range_set_any(range);
3621 /* OSD will only do local fld lookup */
3622 return fld_local_lookup(env, ss->ss_server_fld, seq, range);
3625 static int osd_declare_create(const struct lu_env *env, struct dt_object *dt,
3626 struct lu_attr *attr,
3627 struct dt_allocation_hint *hint,
3628 struct dt_object_format *dof,
3629 struct thandle *handle)
3631 struct osd_thandle *oh;
3632 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
3638 LASSERT(handle != NULL);
3640 oh = container_of(handle, struct osd_thandle, ot_super);
3641 LASSERT(oh->ot_handle == NULL);
3644 * EA object consumes more credits than regular object: osd_mk_index
3645 * vs. osd_mkreg: osd_mk_index will create 2 blocks for root_node and
3646 * leaf_node, could involves the block, block bitmap, groups, GDT
3647 * change for each block, so add 4 * 2 credits in that case.
3649 * The default ACL initialization may consume an additional 16 blocks
3651 credits = osd_dto_credits_noquota[DTO_OBJECT_CREATE] +
3652 ((dof->dof_type == DFT_INDEX) ? 4 * 2 : 0);
3655 * While ldiskfs_new_inode() calls ldiskfs_init_acl() we have to add
3656 * credits for possible default ACL creation in new inode
3658 if (hint && hint->dah_acl_len)
3659 credits += osd_calc_bkmap_credits(sb, NULL, 0, -1,
3660 (hint->dah_acl_len + sb->s_blocksize - 1) >>
3661 sb->s_blocksize_bits);
3663 osd_trans_declare_op(env, oh, OSD_OT_CREATE, credits);
3666 * Reuse idle OI block may cause additional one OI block
3669 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3670 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3671 if (CFS_FAIL_CHECK(OBD_FAIL_OSD_DUPLICATE_MAP))
3672 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
3673 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1);
3675 /* will help to find FID->ino mapping at dt_insert() */
3676 rc = osd_idc_find_and_init(env, osd_obj2dev(osd_dt_obj(dt)),
3684 rc = osd_declare_inode_qid(env, attr->la_uid, attr->la_gid,
3685 attr->la_projid, 1, oh, osd_dt_obj(dt),
3686 NULL, OSD_QID_INODE);
3694 * Called to destroy on-disk representation of the object
3696 * Concurrency: must be locked
3698 static int osd_declare_destroy(const struct lu_env *env, struct dt_object *dt,
3701 struct osd_object *obj = osd_dt_obj(dt);
3702 struct inode *inode = obj->oo_inode;
3703 struct osd_thandle *oh;
3711 oh = container_of(th, struct osd_thandle, ot_super);
3712 LASSERT(oh->ot_handle == NULL);
3714 osd_trans_declare_op(env, oh, OSD_OT_DESTROY,
3715 osd_dto_credits_noquota[DTO_OBJECT_DELETE]);
3717 /* For removing agent entry */
3718 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu))
3719 oh->ot_credits += osd_dto_credits_noquota[DTO_INDEX_DELETE];
3722 * Recycle idle OI leaf may cause additional three OI blocks
3725 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3726 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
3727 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
3728 /* one less inode */
3729 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3730 i_projid_read(inode), -1, oh, obj, NULL,
3734 /* data to be truncated */
3735 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
3736 i_projid_read(inode), 0, oh, obj, NULL,
3742 * will help to find FID->ino when this object is being
3745 rc = osd_idc_find_and_init(env, osd_obj2dev(obj), obj);
3750 static int osd_destroy(const struct lu_env *env, struct dt_object *dt,
3753 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
3754 struct osd_object *obj = osd_dt_obj(dt);
3755 struct inode *inode = obj->oo_inode;
3756 struct osd_device *osd = osd_obj2dev(obj);
3757 struct osd_thandle *oh;
3762 oh = container_of(th, struct osd_thandle, ot_super);
3763 LASSERT(oh->ot_handle);
3765 LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
3767 if (unlikely(fid_is_acct(fid)))
3770 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu)) {
3771 result = osd_delete_from_remote_parent(env, osd, obj, oh, true);
3773 CERROR("%s: remove agent entry "DFID": rc = %d\n",
3774 osd_name(osd), PFID(fid), result);
3777 if (S_ISDIR(inode->i_mode)) {
3778 if (inode->i_nlink > 2)
3779 CERROR("%s: directory "DFID" ino %lu link count is %u at unlink. run e2fsck to repair\n",
3780 osd_name(osd), PFID(fid), inode->i_ino,
3783 spin_lock(&obj->oo_guard);
3785 spin_unlock(&obj->oo_guard);
3786 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
3789 osd_trans_exec_op(env, th, OSD_OT_DESTROY);
3791 ldiskfs_set_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
3793 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_LOST_MDTOBJ2))
3794 result = osd_oi_delete(osd_oti_get(env), osd, fid,
3795 oh->ot_handle, OI_CHECK_FLD);
3797 osd_trans_exec_check(env, th, OSD_OT_DESTROY);
3798 /* XXX: add to ext3 orphan list */
3799 /* rc = ext3_orphan_add(handle_t *handle, struct inode *inode) */
3801 /* not needed in the cache anymore */
3802 set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
3803 obj->oo_destroyed = 1;
3809 * Put the fid into lustre_mdt_attrs, and then place the structure
3810 * inode's ea. This fid should not be altered during the life time
3813 * \retval +ve, on success
3814 * \retval -ve, on error
3816 * FIXME: It is good to have/use ldiskfs_xattr_set_handle() here
3818 int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
3819 const struct lu_fid *fid, __u32 compat, __u32 incompat)
3821 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
3822 struct lustre_mdt_attrs *lma = &loa->loa_lma;
3827 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INLMA))
3830 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_OST_EA_FID_SET))
3833 lustre_loa_init(loa, fid, compat, incompat);
3834 lustre_loa_swab(loa, false);
3837 * For the OST device with 256 bytes inode size by default,
3838 * the PFID EA will be stored together with LMA EA to avoid
3839 * performance trouble. Otherwise the PFID EA can be stored
3840 * independently. LU-8998
3842 if ((compat & LMAC_FID_ON_OST) &&
3843 LDISKFS_INODE_SIZE(inode->i_sb) <= 256)
3844 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3845 sizeof(*loa), XATTR_CREATE);
3847 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3848 sizeof(*lma), XATTR_CREATE);
3850 * LMA may already exist, but we need to check that all the
3851 * desired compat/incompat flags have been added.
3853 if (unlikely(rc == -EEXIST)) {
3854 rc = __osd_xattr_get(inode, &info->oti_obj_dentry,
3855 XATTR_NAME_LMA, (void *)loa, sizeof(*loa));
3859 if (rc < sizeof(*lma))
3862 lustre_loa_swab(loa, true);
3863 if (lu_fid_eq(fid, &lma->lma_self_fid) &&
3864 ((compat == 0 && incompat == 0) ||
3865 (!(~lma->lma_compat & compat) &&
3866 !(~lma->lma_incompat & incompat))))
3869 lma->lma_self_fid = *fid;
3870 lma->lma_compat |= compat;
3871 lma->lma_incompat |= incompat;
3872 if (rc == sizeof(*lma)) {
3873 lustre_lma_swab(lma);
3874 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
3875 sizeof(*lma), XATTR_REPLACE);
3877 lustre_loa_swab(loa, false);
3878 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa,
3879 sizeof(*loa), XATTR_REPLACE);
3887 * ldiskfs supports fid in dirent, it is passed in dentry->d_fsdata.
3888 * lustre 1.8 also uses d_fsdata for passing other info to ldiskfs.
3889 * To have compatilibility with 1.8 ldiskfs driver we need to have
3890 * magic number at start of fid data.
3891 * \ldiskfs_dentry_param is used only to pass fid from osd to ldiskfs.
3894 static void osd_get_ldiskfs_dirent_param(struct ldiskfs_dentry_param *param,
3895 const struct lu_fid *fid)
3897 if (!fid_is_namespace_visible(fid) ||
3898 OBD_FAIL_CHECK(OBD_FAIL_FID_IGIF)) {
3899 param->edp_magic = 0;
3903 param->edp_magic = LDISKFS_LUFID_MAGIC;
3904 param->edp_len = sizeof(struct lu_fid) + 1;
3905 fid_cpu_to_be((struct lu_fid *)param->edp_data, (struct lu_fid *)fid);
3909 * Try to read the fid from inode ea into dt_rec.
3911 * \param fid object fid.
3913 * \retval 0 on success
3915 static int osd_ea_fid_get(const struct lu_env *env, struct osd_object *obj,
3916 __u32 ino, struct lu_fid *fid,
3917 struct osd_inode_id *id)
3919 struct osd_thread_info *info = osd_oti_get(env);
3920 struct inode *inode;
3924 osd_id_gen(id, ino, OSD_OII_NOGEN);
3925 inode = osd_iget_fid(info, osd_obj2dev(obj), id, fid);
3927 RETURN(PTR_ERR(inode));
3933 static int osd_add_dot_dotdot_internal(struct osd_thread_info *info,
3935 struct inode *parent_dir,
3936 const struct lu_fid *dot_fid,
3937 const struct lu_fid *dot_dot_fid,
3938 struct osd_thandle *oth)
3940 struct ldiskfs_dentry_param *dot_ldp;
3941 struct ldiskfs_dentry_param *dot_dot_ldp;
3942 __u32 saved_nlink = dir->i_nlink;
3945 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_DOTDOT_ENOSPC))
3948 dot_dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp2;
3949 osd_get_ldiskfs_dirent_param(dot_dot_ldp, dot_dot_fid);
3951 dot_ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
3952 dot_ldp->edp_magic = 0;
3954 rc = ldiskfs_add_dot_dotdot(oth->ot_handle, parent_dir,
3955 dir, dot_ldp, dot_dot_ldp);
3957 * The ldiskfs_add_dot_dotdot() may dir->i_nlink as 2, then
3958 * the subseqent ref_add() will increase the dir->i_nlink
3959 * as 3. That is incorrect for new created directory.
3961 * It looks like hack, because we want to make the OSD API
3962 * to be order-independent for new created directory object
3963 * between dt_insert(..) and ref_add() operations.
3965 * Here, we only restore the in-RAM dir-inode's nlink attr,
3966 * becuase if the nlink attr is not 2, then there will be
3967 * ref_add() called following the dt_insert(..), such call
3968 * will make both the in-RAM and on-disk dir-inode's nlink
3969 * attr to be set as 2. LU-7447
3971 set_nlink(dir, saved_nlink);
3976 * Create an local agent inode for remote entry
3978 static struct inode *osd_create_local_agent_inode(const struct lu_env *env,
3979 struct osd_device *osd,
3980 struct osd_object *pobj,
3981 const struct lu_fid *fid,
3985 struct osd_thread_info *info = osd_oti_get(env);
3986 struct inode *local;
3987 struct osd_thandle *oh;
3988 struct iattr iattr = {
3989 .ia_valid = ATTR_UID | ATTR_GID |
3990 ATTR_CTIME | ATTR_MTIME | ATTR_ATIME,
3991 .ia_ctime.tv_nsec = UTIME_OMIT,
3992 .ia_mtime.tv_nsec = UTIME_OMIT,
3993 .ia_atime.tv_nsec = UTIME_OMIT,
3994 .ia_uid = GLOBAL_ROOT_UID,
3995 .ia_gid = GLOBAL_ROOT_GID,
4002 oh = container_of(th, struct osd_thandle, ot_super);
4003 LASSERT(oh->ot_handle->h_transaction != NULL);
4005 local = ldiskfs_create_inode(oh->ot_handle, pobj->oo_inode,
4007 if (IS_ERR(local)) {
4008 CERROR("%s: create local error %d\n", osd_name(osd),
4009 (int)PTR_ERR(local));
4014 * restore i_gid in case S_ISGID is set, we will inherit S_ISGID and set
4015 * correct gid on remote file, not agent here
4017 local->i_gid = current_fsgid();
4018 ldiskfs_set_inode_state(local, LDISKFS_STATE_LUSTRE_NOSCRUB);
4020 /* e2fsck doesn't like empty symlinks. Store remote FID as symlink.
4021 * That gives e2fsck something to look at and be happy, and allows
4022 * debugging if we need to determine where this symlink came from.
4024 if (S_ISLNK(type)) {
4025 BUILD_BUG_ON(LDISKFS_N_BLOCKS * 4 < FID_LEN + 1);
4026 rc = scnprintf((char *)LDISKFS_I(local)->i_data,
4027 LDISKFS_N_BLOCKS * 4, DFID, PFID(fid));
4029 i_size_write(local, rc);
4030 LDISKFS_I(local)->i_disksize = rc;
4032 unlock_new_inode(local);
4034 /* Agent inode should not have project ID */
4035 #ifdef HAVE_PROJECT_QUOTA
4036 if (LDISKFS_I(pobj->oo_inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
4037 i_projid_read(pobj->oo_inode) != 0) {
4038 rc = osd_transfer_project(local, 0, th);
4040 CERROR("%s: quota transfer failed:. Is project quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
4041 osd_ino2name(local), rc);
4042 RETURN(ERR_PTR(rc));
4046 /* Set special LMA flag for local agent inode */
4047 rc = osd_ea_fid_set(info, local, fid, 0, LMAI_AGENT);
4049 CERROR("%s: set LMA for "DFID" remote inode failed: rc = %d\n",
4050 osd_name(osd), PFID(fid), rc);
4051 RETURN(ERR_PTR(rc));
4057 rc = osd_add_dot_dotdot_internal(info, local, pobj->oo_inode,
4058 lu_object_fid(&pobj->oo_dt.do_lu),
4061 CERROR("%s: "DFID" add dot dotdot error: rc = %d\n",
4062 osd_name(osd), PFID(fid), rc);
4063 RETURN(ERR_PTR(rc));
4070 * when direntry is deleted, we have to take care of possible agent inode
4071 * referenced by that. unfortunately we can't do this at that point:
4072 * iget() within a running transaction leads to deadlock and we better do
4073 * not call that every delete declaration to save performance. so we put
4074 * a potention agent inode on a list and process that once the transaction
4075 * is over. Notice it's not any worse in terms of real orphans as regular
4076 * object destroy doesn't put inodes on the on-disk orphan list. this should
4077 * be addressed separately
4079 static int osd_schedule_agent_inode_removal(const struct lu_env *env,
4080 struct osd_thandle *oh,
4083 struct osd_device *osd = osd_dt_dev(oh->ot_super.th_dev);
4084 struct osd_obj_orphan *oor;
4091 oor->oor_env = (struct lu_env *)env;
4092 spin_lock(&osd->od_osfs_lock);
4093 list_add(&oor->oor_list, &osd->od_orphan_list);
4094 spin_unlock(&osd->od_osfs_lock);
4096 oh->ot_remove_agents = 1;
4102 static int osd_process_scheduled_agent_removals(const struct lu_env *env,
4103 struct osd_device *osd)
4105 struct osd_thread_info *info = osd_oti_get(env);
4106 struct osd_obj_orphan *oor, *tmp;
4107 struct osd_inode_id id;
4109 struct inode *inode;
4114 spin_lock(&osd->od_osfs_lock);
4115 list_for_each_entry_safe(oor, tmp, &osd->od_orphan_list, oor_list) {
4116 if (oor->oor_env == env)
4117 list_move(&oor->oor_list, &list);
4119 spin_unlock(&osd->od_osfs_lock);
4121 list_for_each_entry_safe(oor, tmp, &list, oor_list) {
4125 list_del(&oor->oor_list);
4128 osd_id_gen(&id, ino, OSD_OII_NOGEN);
4129 inode = osd_iget_fid(info, osd, &id, &fid);
4133 if (!osd_remote_fid(env, osd, &fid)) {
4138 jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC, 1);
4140 mark_inode_dirty(inode);
4141 ldiskfs_journal_stop(jh);
4149 * OSD layer object create function for OST objects (b=11826).
4151 * The FID is inserted into inode xattr here.
4153 * \retval 0, on success
4154 * \retval -ve, on error
4156 static int osd_create(const struct lu_env *env, struct dt_object *dt,
4157 struct lu_attr *attr, struct dt_allocation_hint *hint,
4158 struct dt_object_format *dof, struct thandle *th)
4160 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
4161 struct osd_object *obj = osd_dt_obj(dt);
4162 struct osd_thread_info *info = osd_oti_get(env);
4163 int result, on_ost = 0;
4167 if (dt_object_exists(dt))
4170 LINVRNT(osd_invariant(obj));
4171 LASSERT(!dt_object_remote(dt));
4172 LASSERT(osd_is_write_locked(env, obj));
4173 LASSERT(th != NULL);
4175 if (unlikely(fid_is_acct(fid)))
4177 * Quota files can't be created from the kernel any more,
4178 * 'tune2fs -O quota' will take care of creating them
4182 result = __osd_create(info, obj, attr, hint, dof, th);
4184 if (fid_is_idif(fid) &&
4185 !osd_dev(dt->do_lu.lo_dev)->od_index_in_idif) {
4186 struct lu_fid *tfid = &info->oti_fid;
4187 struct ost_id *oi = &info->oti_ostid;
4189 fid_to_ostid(fid, oi);
4190 ostid_to_fid(tfid, oi, 0);
4192 result = osd_ea_fid_set(info, obj->oo_inode, tfid,
4193 LMAC_FID_ON_OST, 0);
4195 on_ost = fid_is_on_ost(info, osd_obj2dev(obj),
4197 result = osd_ea_fid_set(info, obj->oo_inode, fid,
4198 on_ost ? LMAC_FID_ON_OST : 0,
4201 if (obj->oo_dt.do_body_ops == &osd_body_ops_new)
4202 obj->oo_dt.do_body_ops = &osd_body_ops;
4205 if (!result && !CFS_FAIL_CHECK(OBD_FAIL_OSD_NO_OI_ENTRY)) {
4206 struct inode *inode = obj->oo_inode;
4208 result = __osd_oi_insert(env, obj, fid, th);
4209 if (result && inode) {
4210 spin_lock(&obj->oo_guard);
4212 spin_unlock(&obj->oo_guard);
4213 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4214 ldiskfs_set_inode_state(inode,
4215 LDISKFS_STATE_LUSTRE_DESTROY);
4217 obj->oo_inode = NULL;
4222 * a small optimization - dt_insert() isn't usually applied
4223 * to OST objects, so we don't need to cache OI mapping for
4226 if (result == 0 && on_ost == 0) {
4227 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4229 result = osd_idc_find_and_init(env, osd, obj);
4230 LASSERT(result == 0);
4233 LASSERT(ergo(result == 0,
4234 dt_object_exists(dt) && !dt_object_remote(dt)));
4235 LINVRNT(osd_invariant(obj));
4239 static int osd_declare_ref_add(const struct lu_env *env, struct dt_object *dt,
4240 struct thandle *handle)
4242 struct osd_thandle *oh;
4245 /* it's possible that object doesn't exist yet */
4246 LASSERT(handle != NULL);
4248 oh = container_of(handle, struct osd_thandle, ot_super);
4249 LASSERT(oh->ot_handle == NULL);
4251 osd_trans_declare_op(env, oh, OSD_OT_REF_ADD,
4252 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4254 rc = osd_idc_find_and_init(env, osd_dev(dt->do_lu.lo_dev),
4261 * Concurrency: @dt is write locked.
4263 static int osd_ref_add(const struct lu_env *env, struct dt_object *dt,
4266 struct osd_object *obj = osd_dt_obj(dt);
4267 struct inode *inode = obj->oo_inode;
4268 struct osd_thandle *oh;
4271 if (!dt_object_exists(dt) || obj->oo_destroyed)
4274 LINVRNT(osd_invariant(obj));
4275 LASSERT(!dt_object_remote(dt));
4276 LASSERT(osd_is_write_locked(env, obj));
4277 LASSERT(th != NULL);
4279 oh = container_of(th, struct osd_thandle, ot_super);
4280 LASSERT(oh->ot_handle != NULL);
4282 osd_trans_exec_op(env, th, OSD_OT_REF_ADD);
4284 CDEBUG(D_INODE, DFID" increase nlink %d\n",
4285 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4287 * The DIR_NLINK feature allows directories to exceed LDISKFS_LINK_MAX
4288 * (65000) subdirectories by storing "1" in i_nlink if the link count
4289 * would otherwise overflow. Directory tranversal tools understand
4290 * that (st_nlink == 1) indicates that the filesystem dose not track
4291 * hard links count on the directory, and will not abort subdirectory
4292 * scanning early once (st_nlink - 2) subdirs have been found.
4294 * This also has to properly handle the case of inodes with nlink == 0
4295 * in case they are being linked into the PENDING directory
4297 spin_lock(&obj->oo_guard);
4298 if (unlikely(inode->i_nlink == 0))
4299 /* inc_nlink from 0 may cause WARN_ON */
4300 set_nlink(inode, 1);
4302 ldiskfs_inc_count(oh->ot_handle, inode);
4303 if (!S_ISDIR(inode->i_mode))
4304 LASSERT(inode->i_nlink <= LDISKFS_LINK_MAX);
4306 spin_unlock(&obj->oo_guard);
4308 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4309 LINVRNT(osd_invariant(obj));
4311 osd_trans_exec_check(env, th, OSD_OT_REF_ADD);
4316 static int osd_declare_ref_del(const struct lu_env *env, struct dt_object *dt,
4317 struct thandle *handle)
4319 struct osd_thandle *oh;
4321 if (!dt_object_exists(dt))
4324 LASSERT(!dt_object_remote(dt));
4325 LASSERT(handle != NULL);
4327 oh = container_of(handle, struct osd_thandle, ot_super);
4328 LASSERT(oh->ot_handle == NULL);
4330 osd_trans_declare_op(env, oh, OSD_OT_REF_DEL,
4331 osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);
4337 * Concurrency: @dt is write locked.
4339 static int osd_ref_del(const struct lu_env *env, struct dt_object *dt,
4342 struct osd_object *obj = osd_dt_obj(dt);
4343 struct inode *inode = obj->oo_inode;
4344 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
4345 struct osd_thandle *oh;
4347 if (!dt_object_exists(dt))
4350 LINVRNT(osd_invariant(obj));
4351 LASSERT(!dt_object_remote(dt));
4352 LASSERT(osd_is_write_locked(env, obj));
4353 LASSERT(th != NULL);
4355 if (OBD_FAIL_CHECK(OBD_FAIL_OSD_REF_DEL))
4358 oh = container_of(th, struct osd_thandle, ot_super);
4359 LASSERT(oh->ot_handle != NULL);
4361 osd_trans_exec_op(env, th, OSD_OT_REF_DEL);
4363 spin_lock(&obj->oo_guard);
4365 * That can be result of upgrade from old Lustre version and
4366 * applied only to local files. Just skip this ref_del call.
4367 * ext4_unlink() only treats this as a warning, don't LASSERT here.
4369 if (inode->i_nlink == 0) {
4370 CDEBUG_LIMIT(fid_is_norm(lu_object_fid(&dt->do_lu)) ?
4371 D_ERROR : D_INODE, "%s: nlink == 0 on "DFID
4372 ", maybe an upgraded file? (LU-3915)\n",
4373 osd_name(osd), PFID(lu_object_fid(&dt->do_lu)));
4374 spin_unlock(&obj->oo_guard);
4378 CDEBUG(D_INODE, DFID" decrease nlink %d\n",
4379 PFID(lu_object_fid(&dt->do_lu)), inode->i_nlink);
4381 ldiskfs_dec_count(oh->ot_handle, inode);
4382 spin_unlock(&obj->oo_guard);
4384 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4385 LINVRNT(osd_invariant(obj));
4387 osd_trans_exec_check(env, th, OSD_OT_REF_DEL);
4393 * Concurrency: @dt is read locked.
4395 static int osd_xattr_get(const struct lu_env *env, struct dt_object *dt,
4396 struct lu_buf *buf, const char *name)
4398 struct osd_object *obj = osd_dt_obj(dt);
4399 struct inode *inode = obj->oo_inode;
4400 struct osd_thread_info *info = osd_oti_get(env);
4401 struct dentry *dentry = &info->oti_obj_dentry;
4402 bool cache_xattr = false;
4407 /* version get is not real XATTR but uses xattr API */
4408 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4409 dt_obj_version_t *ver = buf->lb_buf;
4412 * for version we are just using xattr API but change inode
4415 if (buf->lb_len == 0)
4416 return sizeof(dt_obj_version_t);
4418 if (buf->lb_len < sizeof(dt_obj_version_t))
4421 CDEBUG(D_INODE, "Get version %#llx for inode %lu\n",
4422 LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4424 *ver = LDISKFS_I(inode)->i_fs_version;
4426 return sizeof(dt_obj_version_t);
4429 if (!dt_object_exists(dt))
4432 LASSERT(!dt_object_remote(dt));
4433 LASSERT(inode->i_op != NULL);
4434 #ifdef HAVE_IOP_XATTR
4435 LASSERT(inode->i_op->getxattr != NULL);
4438 if (strcmp(name, XATTR_NAME_LOV) == 0 ||
4439 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0)
4443 rc = osd_oxc_get(obj, name, buf);
4448 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4449 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4450 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4451 struct filter_fid *ff;
4452 struct ost_layout *ol;
4454 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4458 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4461 if (buf->lb_len == 0 || !buf->lb_buf)
4464 if (buf->lb_len < rc)
4468 ol = &ff->ff_layout;
4469 ol->ol_stripe_count = cpu_to_le32(loa->loa_parent_fid.f_ver >>
4470 PFID_STRIPE_IDX_BITS);
4471 ol->ol_stripe_size = cpu_to_le32(loa->loa_stripe_size);
4472 loa->loa_parent_fid.f_ver &= PFID_STRIPE_COUNT_MASK;
4473 fid_cpu_to_le(&ff->ff_parent, &loa->loa_parent_fid);
4474 if (lma->lma_compat & LMAC_COMP_INFO) {
4475 ol->ol_comp_start = cpu_to_le64(loa->loa_comp_start);
4476 ol->ol_comp_end = cpu_to_le64(loa->loa_comp_end);
4477 ol->ol_comp_id = cpu_to_le32(loa->loa_comp_id);
4479 ol->ol_comp_start = 0;
4480 ol->ol_comp_end = 0;
4484 rc = __osd_xattr_get(inode, dentry, name,
4485 buf->lb_buf, buf->lb_len);
4489 if (rc == -ENOENT || rc == -ENODATA)
4490 osd_oxc_add(obj, name, NULL, 0);
4491 else if (rc > 0 && buf->lb_buf != NULL)
4492 osd_oxc_add(obj, name, buf->lb_buf, rc);
4498 static int osd_declare_xattr_set(const struct lu_env *env,
4499 struct dt_object *dt,
4500 const struct lu_buf *buf, const char *name,
4501 int fl, struct thandle *handle)
4503 struct osd_thandle *oh;
4505 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4507 LASSERT(handle != NULL);
4509 oh = container_of(handle, struct osd_thandle, ot_super);
4510 LASSERT(oh->ot_handle == NULL);
4512 if (strcmp(name, XATTR_NAME_LMA) == 0) {
4514 * For non-upgrading case, the LMA is set first and
4515 * usually fit inode. But for upgrade case, the LMA
4516 * may be in another separated EA block.
4518 if (dt_object_exists(dt)) {
4519 if (fl == LU_XATTR_REPLACE)
4524 } else if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4526 } else if (strcmp(name, XATTR_NAME_FID) == 0) {
4527 /* We may need to delete the old PFID EA. */
4528 credits = LDISKFS_MAXQUOTAS_DEL_BLOCKS(sb);
4529 if (fl == LU_XATTR_REPLACE)
4535 * If some name entry resides on remote MDT, then will create
4536 * agent entry under remote parent. On the other hand, if the
4537 * remote entry will be removed, then related agent entry may
4538 * need to be removed from the remote parent. So there may be
4539 * kinds of cases, let's declare enough credits. The credits
4540 * for create agent entry is enough for remove case.
4542 if (strcmp(name, XATTR_NAME_LINK) == 0) {
4543 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
4544 if (dt_object_exists(dt))
4545 credits += 1; /* For updating LMA */
4549 credits += osd_dto_credits_noquota[DTO_XATTR_SET];
4554 if (buf->lb_buf == NULL && dt_object_exists(dt)) {
4556 * learn xattr size from osd_xattr_get if
4557 * attribute has not been read yet
4559 buflen = __osd_xattr_get(
4560 osd_dt_obj(dt)->oo_inode,
4561 &osd_oti_get(env)->oti_obj_dentry,
4566 buflen = buf->lb_len;
4569 if (buflen > sb->s_blocksize) {
4570 credits += osd_calc_bkmap_credits(
4572 (buflen + sb->s_blocksize - 1) >>
4573 sb->s_blocksize_bits);
4577 * xattr set may involve inode quota change, reserve credits for
4578 * dquot_initialize()
4580 credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4583 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET, credits);
4588 static int osd_xattr_set_pfid(const struct lu_env *env, struct osd_object *obj,
4589 const struct lu_buf *buf, int fl,
4590 struct thandle *handle)
4592 struct osd_thread_info *info = osd_oti_get(env);
4593 struct dentry *dentry = &info->oti_obj_dentry;
4594 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4595 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4596 struct inode *inode = obj->oo_inode;
4597 struct filter_fid *ff = buf->lb_buf;
4598 struct ost_layout *ol = &ff->ff_layout;
4599 int flags = XATTR_REPLACE;
4604 if (buf->lb_len != sizeof(*ff) && buf->lb_len != sizeof(struct lu_fid))
4607 rc = osd_get_lma(info, inode, dentry, loa);
4608 if (rc == -ENODATA) {
4609 /* Usually for upgarding from old device */
4610 lustre_loa_init(loa, lu_object_fid(&obj->oo_dt.do_lu),
4611 LMAC_FID_ON_OST, 0);
4612 flags = XATTR_CREATE;
4617 if (!rc && lma->lma_compat & LMAC_STRIPE_INFO) {
4618 if ((fl & LU_XATTR_CREATE) && !(fl & LU_XATTR_REPLACE))
4621 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256) {
4622 /* Separate PFID EA from LMA */
4623 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4624 lustre_lma_swab(lma);
4625 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4626 sizeof(*lma), XATTR_REPLACE);
4628 obj->oo_pfid_in_lma = 0;
4629 rc = LU_XATTR_CREATE;
4635 if (LDISKFS_INODE_SIZE(inode->i_sb) > 256)
4639 * Old client does not send stripe information,
4640 * then store the PFID EA on disk separatedly.
4642 if (unlikely(buf->lb_len == sizeof(struct lu_fid) ||
4643 ol->ol_stripe_size == 0))
4646 /* Remove old PFID EA entry firstly. */
4647 dquot_initialize(inode);
4648 rc = ll_vfs_removexattr(dentry, inode, XATTR_NAME_FID);
4649 if (rc == -ENODATA) {
4650 /* XATTR_NAME_FID is already absent */
4657 fid_le_to_cpu(&loa->loa_parent_fid, &ff->ff_parent);
4658 if (likely(ol->ol_stripe_size != 0)) {
4659 loa->loa_parent_fid.f_ver |= le32_to_cpu(ol->ol_stripe_count) <<
4660 PFID_STRIPE_IDX_BITS;
4661 loa->loa_stripe_size = le32_to_cpu(ol->ol_stripe_size);
4662 lma->lma_compat |= LMAC_STRIPE_INFO;
4663 if (ol->ol_comp_id != 0) {
4664 loa->loa_comp_id = le32_to_cpu(ol->ol_comp_id);
4665 loa->loa_comp_start = le64_to_cpu(ol->ol_comp_start);
4666 loa->loa_comp_end = le64_to_cpu(ol->ol_comp_end);
4667 lma->lma_compat |= LMAC_COMP_INFO;
4671 lustre_loa_swab(loa, false);
4673 /* Store the PFID EA inside LMA. */
4674 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, loa, sizeof(*loa),
4677 obj->oo_pfid_in_lma = 1;
4683 * In DNE environment, the object (in spite of regular file or directory)
4684 * and its name entry may reside on different MDTs. Under such case, we will
4685 * create an agent entry on the MDT where the object resides. The agent entry
4686 * references the object locally, that makes the object to be visible to the
4687 * userspace when mounted as 'ldiskfs' directly. Then the userspace tools,
4688 * such as 'tar' can handle the object properly.
4690 * We handle the agent entry during set linkEA that is the common interface
4691 * for both regular file and directroy, can handle kinds of cases, such as
4692 * create/link/unlink/rename, and so on.
4694 * NOTE: we can NOT do that when ea_{insert,delete} that is only for directory.
4696 * XXX: There are two known issues:
4697 * 1. For one object, we will create at most one agent entry even if there
4698 * may be more than one cross-MDTs hard links on the object. So the local
4699 * e2fsck may claim that the object's nlink is larger than the name entries
4700 * that reference such inode. And in further, the e2fsck will fix the nlink
4701 * attribute to match the local references. Then it will cause the object's
4702 * nlink attribute to be inconsistent with the global references. it is bad
4703 * but not fatal. The ref_del() can handle the zero-referenced case. On the
4704 * other hand, the global namespace LFSCK can repair the object's attribute
4705 * according to the linkEA.
4706 * 2. There may be too many hard links on the object as to its linkEA overflow,
4707 * then the linkEA entry for cross-MDTs reference may be discarded. If such
4708 * case happened, then at this point, we do not know whether there are some
4709 * cross-MDTs reference. But there are local references, it guarantees that
4710 * object is visible to userspace when mounted as 'ldiskfs'. That is enough.
4712 static int osd_xattr_handle_linkea(const struct lu_env *env,
4713 struct osd_device *osd,
4714 struct osd_object *obj,
4715 const struct lu_buf *buf,
4716 struct thandle *handle)
4718 const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
4719 struct lu_fid *tfid = &osd_oti_get(env)->oti_fid3;
4720 struct linkea_data ldata = { .ld_buf = (struct lu_buf *)buf };
4721 struct lu_name tmpname;
4722 struct osd_thandle *oh;
4724 bool remote = false;
4728 oh = container_of(handle, struct osd_thandle, ot_super);
4729 LASSERT(oh->ot_handle != NULL);
4731 rc = linkea_init_with_rec(&ldata);
4733 linkea_first_entry(&ldata);
4734 while (ldata.ld_lee != NULL && !remote) {
4735 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen,
4737 if (osd_remote_fid(env, osd, tfid) > 0)
4740 linkea_next_entry(&ldata);
4742 } else if (rc == -ENODATA) {
4748 if (lu_object_has_agent_entry(&obj->oo_dt.do_lu) && !remote) {
4749 rc = osd_delete_from_remote_parent(env, osd, obj, oh, false);
4751 CERROR("%s: failed to remove agent entry for "DFID
4752 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4753 } else if (!lu_object_has_agent_entry(&obj->oo_dt.do_lu) && remote) {
4754 rc = osd_add_to_remote_parent(env, osd, obj, oh);
4756 CERROR("%s: failed to create agent entry for "DFID
4757 ": rc = %d\n", osd_name(osd), PFID(fid), rc);
4764 * Concurrency: @dt is write locked.
4766 static int osd_xattr_set(const struct lu_env *env, struct dt_object *dt,
4767 const struct lu_buf *buf, const char *name, int fl,
4768 struct thandle *handle)
4770 struct osd_object *obj = osd_dt_obj(dt);
4771 struct osd_device *osd = osd_obj2dev(obj);
4772 struct inode *inode = obj->oo_inode;
4773 struct osd_thread_info *info = osd_oti_get(env);
4783 /* version set is not real XATTR */
4784 if (strcmp(name, XATTR_NAME_VERSION) == 0) {
4785 dt_obj_version_t *version = buf->lb_buf;
4788 * for version we are just using xattr API but change inode
4791 LASSERT(buf->lb_len == sizeof(dt_obj_version_t));
4793 CDEBUG(D_INODE, "Set version %#llx (old %#llx) for inode %lu\n",
4794 *version, LDISKFS_I(inode)->i_fs_version, inode->i_ino);
4796 LDISKFS_I(inode)->i_fs_version = *version;
4798 * Version is set after all inode operations are finished,
4799 * so we should mark it dirty here
4801 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
4806 CDEBUG(D_INODE, DFID" set xattr '%s' with size %zu\n",
4807 PFID(lu_object_fid(&dt->do_lu)), name, buf->lb_len);
4810 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4813 * For the OST device with 256 bytes inode size by default,
4814 * the PFID EA will be stored together with LMA EA to avoid
4815 * performance trouble. Otherwise the PFID EA can be stored
4816 * independently. LU-8998
4818 if (strcmp(name, XATTR_NAME_FID) == 0 && osd->od_is_ost &&
4819 (LDISKFS_INODE_SIZE(inode->i_sb) <= 256 || obj->oo_pfid_in_lma)) {
4820 LASSERT(buf->lb_buf);
4822 fl = osd_xattr_set_pfid(env, obj, buf, fl, handle);
4825 } else if (strcmp(name, XATTR_NAME_LMV) == 0) {
4826 struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
4827 struct lustre_mdt_attrs *lma = &loa->loa_lma;
4829 rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
4833 lma->lma_incompat |= LMAI_STRIPED;
4834 lustre_lma_swab(lma);
4835 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4836 sizeof(*lma), XATTR_REPLACE);
4839 } else if (strcmp(name, XATTR_NAME_LINK) == 0) {
4840 LASSERT(!osd->od_is_ost);
4842 rc = osd_xattr_handle_linkea(env, osd, obj, buf, handle);
4847 if (fl & LU_XATTR_REPLACE)
4848 fs_flags |= XATTR_REPLACE;
4850 if (fl & LU_XATTR_CREATE)
4851 fs_flags |= XATTR_CREATE;
4853 rc = __osd_xattr_set(info, inode, name, buf->lb_buf, len, fs_flags);
4854 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4857 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4858 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4859 osd_oxc_add(obj, name, buf->lb_buf, buf->lb_len);
4865 * Concurrency: @dt is read locked.
4867 static int osd_xattr_list(const struct lu_env *env, struct dt_object *dt,
4868 const struct lu_buf *buf)
4870 struct osd_object *obj = osd_dt_obj(dt);
4871 struct inode *inode = obj->oo_inode;
4872 struct osd_thread_info *info = osd_oti_get(env);
4873 struct dentry *dentry = &info->oti_obj_dentry;
4875 if (!dt_object_exists(dt))
4878 LASSERT(!dt_object_remote(dt));
4879 LASSERT(inode->i_op != NULL);
4880 LASSERT(inode->i_op->listxattr != NULL);
4882 dentry->d_inode = inode;
4883 dentry->d_sb = inode->i_sb;
4884 return inode->i_op->listxattr(dentry, buf->lb_buf, buf->lb_len);
4887 static int osd_declare_xattr_del(const struct lu_env *env,
4888 struct dt_object *dt, const char *name,
4889 struct thandle *handle)
4891 struct osd_thandle *oh;
4892 struct super_block *sb = osd_sb(osd_dev(dt->do_lu.lo_dev));
4894 LASSERT(!dt_object_remote(dt));
4895 LASSERT(handle != NULL);
4897 oh = container_of(handle, struct osd_thandle, ot_super);
4898 LASSERT(oh->ot_handle == NULL);
4900 osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
4901 osd_dto_credits_noquota[DTO_XATTR_SET]);
4903 * xattr del may involve inode quota change, reserve credits for
4904 * dquot_initialize()
4906 oh->ot_credits += LDISKFS_MAXQUOTAS_INIT_BLOCKS(sb);
4912 * Concurrency: @dt is write locked.
4914 static int osd_xattr_del(const struct lu_env *env, struct dt_object *dt,
4915 const char *name, struct thandle *handle)
4917 struct osd_object *obj = osd_dt_obj(dt);
4918 struct inode *inode = obj->oo_inode;
4919 struct osd_thread_info *info = osd_oti_get(env);
4920 struct dentry *dentry = &info->oti_obj_dentry;
4923 if (!dt_object_exists(dt))
4926 LASSERT(!dt_object_remote(dt));
4927 LASSERT(inode->i_op != NULL);
4928 LASSERT(handle != NULL);
4929 #ifdef HAVE_IOP_XATTR
4930 LASSERT(inode->i_op->removexattr != NULL);
4933 osd_trans_exec_op(env, handle, OSD_OT_XATTR_SET);
4935 if (strcmp(name, XATTR_NAME_FID) == 0 && obj->oo_pfid_in_lma) {
4936 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
4938 rc = osd_get_lma(info, inode, &info->oti_obj_dentry,
4939 &info->oti_ost_attrs);
4941 LASSERT(lma->lma_compat & LMAC_STRIPE_INFO);
4943 lma->lma_compat &= ~(LMAC_STRIPE_INFO | LMAC_COMP_INFO);
4944 lustre_lma_swab(lma);
4945 rc = __osd_xattr_set(info, inode, XATTR_NAME_LMA, lma,
4946 sizeof(*lma), XATTR_REPLACE);
4948 obj->oo_pfid_in_lma = 0;
4951 dquot_initialize(inode);
4952 dentry->d_inode = inode;
4953 dentry->d_sb = inode->i_sb;
4954 rc = ll_vfs_removexattr(dentry, inode, name);
4957 osd_trans_exec_check(env, handle, OSD_OT_XATTR_SET);
4960 (strcmp(name, XATTR_NAME_LOV) == 0 ||
4961 strcmp(name, XATTR_NAME_DEFAULT_LMV) == 0))
4962 osd_oxc_del(obj, name);
4967 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
4968 __u64 start, __u64 end)
4970 struct osd_object *obj = osd_dt_obj(dt);
4971 struct inode *inode = obj->oo_inode;
4972 struct file *file = osd_quasi_file(env, inode);
4977 rc = vfs_fsync_range(file, start, end, 0);
4982 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
4987 static bool osd_check_stale(struct dt_object *dt)
4995 static int osd_iam_index_probe(const struct lu_env *env, struct osd_object *o,
4996 const struct dt_index_features *feat)
4998 struct iam_descr *descr;
5000 if (osd_object_is_root(o))
5001 return feat == &dt_directory_features;
5003 LASSERT(o->oo_dir != NULL);
5005 descr = o->oo_dir->od_container.ic_descr;
5006 if (feat == &dt_directory_features) {
5007 if (descr->id_rec_size == sizeof(struct osd_fid_pack))
5012 return feat->dif_keysize_min <= descr->id_key_size &&
5013 descr->id_key_size <= feat->dif_keysize_max &&
5014 feat->dif_recsize_min <= descr->id_rec_size &&
5015 descr->id_rec_size <= feat->dif_recsize_max &&
5016 !(feat->dif_flags & (DT_IND_VARKEY |
5017 DT_IND_VARREC | DT_IND_NONUNQ)) &&
5018 ergo(feat->dif_flags & DT_IND_UPDATE,
5019 1 /* XXX check that object (and fs) is writable */);
5023 static int osd_iam_container_init(const struct lu_env *env,
5024 struct osd_object *obj,
5025 struct osd_directory *dir)
5027 struct iam_container *bag = &dir->od_container;
5030 result = iam_container_init(bag, &dir->od_descr, obj->oo_inode);
5034 result = iam_container_setup(bag);
5036 obj->oo_dt.do_index_ops = &osd_index_iam_ops;
5038 iam_container_fini(bag);
5045 * Concurrency: no external locking is necessary.
5047 static int osd_index_try(const struct lu_env *env, struct dt_object *dt,
5048 const struct dt_index_features *feat)
5052 struct osd_object *obj = osd_dt_obj(dt);
5054 LINVRNT(osd_invariant(obj));
5056 if (osd_object_is_root(obj)) {
5057 dt->do_index_ops = &osd_index_ea_ops;
5059 } else if (feat == &dt_directory_features) {
5060 dt->do_index_ops = &osd_index_ea_ops;
5061 if (obj->oo_inode == NULL || S_ISDIR(obj->oo_inode->i_mode))
5066 } else if (unlikely(feat == &dt_otable_features)) {
5067 dt->do_index_ops = &osd_otable_ops;
5069 } else if (unlikely(feat == &dt_acct_features)) {
5070 dt->do_index_ops = &osd_acct_index_ops;
5073 } else if (!osd_has_index(obj)) {
5074 struct osd_directory *dir;
5075 struct osd_device *osd = osd_obj2dev(obj);
5076 const struct lu_fid *fid = lu_object_fid(&dt->do_lu);
5081 spin_lock(&obj->oo_guard);
5082 if (obj->oo_dir == NULL)
5086 * Concurrent thread allocated container data.
5089 spin_unlock(&obj->oo_guard);
5091 * Now, that we have container data, serialize its
5094 down_write(&obj->oo_ext_idx_sem);
5096 * recheck under lock.
5099 if (osd_has_index(obj)) {
5104 result = osd_iam_container_init(env, obj, obj->oo_dir);
5105 if (result || feat == &dt_lfsck_namespace_features ||
5106 feat == &dt_lfsck_layout_orphan_features ||
5107 feat == &dt_lfsck_layout_dangling_features)
5110 result = osd_index_register(osd, fid,
5111 feat->dif_keysize_max,
5112 feat->dif_recsize_max);
5114 CWARN("%s: failed to register index "
5116 osd_name(osd), PFID(fid), result);
5117 else if (result > 0)
5120 CDEBUG(D_LFSCK, "%s: index object "DFID
5121 " (%d/%d) registered\n",
5122 osd_name(osd), PFID(fid),
5123 (int)feat->dif_keysize_max,
5124 (int)feat->dif_recsize_max);
5127 up_write(&obj->oo_ext_idx_sem);
5135 if (result == 0 && skip_iam == 0) {
5136 if (!osd_iam_index_probe(env, obj, feat))
5139 LINVRNT(osd_invariant(obj));
5144 static int osd_otable_it_attr_get(const struct lu_env *env,
5145 struct dt_object *dt,
5146 struct lu_attr *attr)
5152 static const struct dt_object_operations osd_obj_ops = {
5153 .do_read_lock = osd_read_lock,
5154 .do_write_lock = osd_write_lock,
5155 .do_read_unlock = osd_read_unlock,
5156 .do_write_unlock = osd_write_unlock,
5157 .do_write_locked = osd_write_locked,
5158 .do_attr_get = osd_attr_get,
5159 .do_declare_attr_set = osd_declare_attr_set,
5160 .do_attr_set = osd_attr_set,
5161 .do_ah_init = osd_ah_init,
5162 .do_declare_create = osd_declare_create,
5163 .do_create = osd_create,
5164 .do_declare_destroy = osd_declare_destroy,
5165 .do_destroy = osd_destroy,
5166 .do_index_try = osd_index_try,
5167 .do_declare_ref_add = osd_declare_ref_add,
5168 .do_ref_add = osd_ref_add,
5169 .do_declare_ref_del = osd_declare_ref_del,
5170 .do_ref_del = osd_ref_del,
5171 .do_xattr_get = osd_xattr_get,
5172 .do_declare_xattr_set = osd_declare_xattr_set,
5173 .do_xattr_set = osd_xattr_set,
5174 .do_declare_xattr_del = osd_declare_xattr_del,
5175 .do_xattr_del = osd_xattr_del,
5176 .do_xattr_list = osd_xattr_list,
5177 .do_object_sync = osd_object_sync,
5178 .do_invalidate = osd_invalidate,
5179 .do_check_stale = osd_check_stale,
5182 static const struct dt_object_operations osd_obj_otable_it_ops = {
5183 .do_attr_get = osd_otable_it_attr_get,
5184 .do_index_try = osd_index_try,
5187 static int osd_index_declare_iam_delete(const struct lu_env *env,
5188 struct dt_object *dt,
5189 const struct dt_key *key,
5190 struct thandle *handle)
5192 struct osd_thandle *oh;
5194 oh = container_of(handle, struct osd_thandle, ot_super);
5195 LASSERT(oh->ot_handle == NULL);
5197 /* Recycle may cause additional three blocks to be changed. */
5198 osd_trans_declare_op(env, oh, OSD_OT_DELETE,
5199 osd_dto_credits_noquota[DTO_INDEX_DELETE] + 3);
5205 * delete a (key, value) pair from index \a dt specified by \a key
5207 * \param dt osd index object
5208 * \param key key for index
5209 * \param rec record reference
5210 * \param handle transaction handler
5213 * \retval -ve failure
5215 static int osd_index_iam_delete(const struct lu_env *env, struct dt_object *dt,
5216 const struct dt_key *key,
5217 struct thandle *handle)
5219 struct osd_thread_info *oti = osd_oti_get(env);
5220 struct osd_object *obj = osd_dt_obj(dt);
5221 struct osd_thandle *oh;
5222 struct iam_path_descr *ipd;
5223 struct iam_container *bag = &obj->oo_dir->od_container;
5228 if (!dt_object_exists(dt))
5231 LINVRNT(osd_invariant(obj));
5232 LASSERT(!dt_object_remote(dt));
5233 LASSERT(bag->ic_object == obj->oo_inode);
5234 LASSERT(handle != NULL);
5236 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5238 ipd = osd_idx_ipd_get(env, bag);
5239 if (unlikely(ipd == NULL))
5242 oh = container_of(handle, struct osd_thandle, ot_super);
5243 LASSERT(oh->ot_handle != NULL);
5244 LASSERT(oh->ot_handle->h_transaction != NULL);
5246 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5247 /* swab quota uid/gid provided by caller */
5248 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5249 key = (const struct dt_key *)&oti->oti_quota_id;
5252 rc = iam_delete(oh->ot_handle, bag, (const struct iam_key *)key, ipd);
5253 osd_ipd_put(env, bag, ipd);
5254 LINVRNT(osd_invariant(obj));
5255 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5259 static int osd_index_declare_ea_delete(const struct lu_env *env,
5260 struct dt_object *dt,
5261 const struct dt_key *key,
5262 struct thandle *handle)
5264 struct osd_thandle *oh;
5265 struct inode *inode;
5270 LASSERT(!dt_object_remote(dt));
5271 LASSERT(handle != NULL);
5273 oh = container_of(handle, struct osd_thandle, ot_super);
5274 LASSERT(oh->ot_handle == NULL);
5276 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE];
5277 osd_trans_declare_op(env, oh, OSD_OT_DELETE, credits);
5279 inode = osd_dt_obj(dt)->oo_inode;
5283 rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
5284 i_projid_read(inode), 0, oh, osd_dt_obj(dt),
5289 static inline int osd_get_fid_from_dentry(struct ldiskfs_dir_entry_2 *de,
5292 struct osd_fid_pack *rec;
5295 if (de->file_type & LDISKFS_DIRENT_LUFID) {
5296 rec = (struct osd_fid_pack *)(de->name + de->name_len + 1);
5297 rc = osd_fid_unpack((struct lu_fid *)fid, rec);
5298 if (rc == 0 && unlikely(!fid_is_sane((struct lu_fid *)fid)))
5304 static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
5305 const struct lu_fid *fid)
5307 struct seq_server_site *ss = osd_seq_site(osd);
5311 /* FID seqs not in FLDB, must be local seq */
5312 if (unlikely(!fid_seq_in_fldb(fid_seq(fid))))
5316 * If FLD is not being initialized yet, it only happens during the
5317 * initialization, likely during mgs initialization, and we assume
5318 * this is local FID.
5320 if (ss == NULL || ss->ss_server_fld == NULL)
5323 /* Only check the local FLDB here */
5324 if (osd_seq_exists(env, osd, fid_seq(fid)))
5330 static void osd_take_care_of_agent(const struct lu_env *env,
5331 struct osd_device *osd,
5332 struct osd_thandle *oh,
5333 struct ldiskfs_dir_entry_2 *de)
5335 struct lu_fid *fid = &osd_oti_get(env)->oti_fid;
5336 struct osd_idmap_cache *idc;
5337 int rc, schedule = 0;
5339 LASSERT(de != NULL);
5341 rc = osd_get_fid_from_dentry(de, (struct dt_rec *)fid);
5342 if (likely(rc == 0)) {
5343 idc = osd_idc_find_or_init(env, osd, fid);
5344 if (IS_ERR(idc) || idc->oic_remote)
5346 } else if (rc == -ENODATA) {
5348 * can't get FID, postpone to the end of the
5349 * transaction when iget() is safe
5353 CERROR("%s: can't get FID: rc = %d\n", osd_name(osd), rc);
5356 osd_schedule_agent_inode_removal(env, oh,
5357 le32_to_cpu(de->inode));
5361 * Index delete function for interoperability mode (b11826).
5362 * It will remove the directory entry added by osd_index_ea_insert().
5363 * This entry is needed to maintain name->fid mapping.
5365 * \param key, key i.e. file entry to be deleted
5367 * \retval 0, on success
5368 * \retval -ve, on error
5370 static int osd_index_ea_delete(const struct lu_env *env, struct dt_object *dt,
5371 const struct dt_key *key, struct thandle *handle)
5373 struct osd_object *obj = osd_dt_obj(dt);
5374 struct inode *dir = obj->oo_inode;
5375 struct dentry *dentry;
5376 struct osd_thandle *oh;
5377 struct ldiskfs_dir_entry_2 *de = NULL;
5378 struct buffer_head *bh;
5379 struct htree_lock *hlock = NULL;
5380 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
5385 if (!dt_object_exists(dt))
5388 LINVRNT(osd_invariant(obj));
5389 LASSERT(!dt_object_remote(dt));
5390 LASSERT(handle != NULL);
5392 osd_trans_exec_op(env, handle, OSD_OT_DELETE);
5394 oh = container_of(handle, struct osd_thandle, ot_super);
5395 LASSERT(oh->ot_handle != NULL);
5396 LASSERT(oh->ot_handle->h_transaction != NULL);
5398 dquot_initialize(dir);
5399 dentry = osd_child_dentry_get(env, obj,
5400 (char *)key, strlen((char *)key));
5402 if (obj->oo_hl_head != NULL) {
5403 hlock = osd_oti_get(env)->oti_hlock;
5404 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
5405 dir, LDISKFS_HLOCK_DEL);
5407 down_write(&obj->oo_ext_idx_sem);
5410 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
5413 * If this is not the ".." entry, it might be a remote DNE
5414 * entry and we need to check if the FID is for a remote
5415 * MDT. If the FID is not in the directory entry (e.g.
5416 * upgraded 1.8 filesystem without dirdata enabled) then
5417 * we need to get the FID from the LMA. For a remote directory
5418 * there HAS to be an LMA, it cannot be an IGIF inode in this
5421 * Delete the entry before the agent inode in order to
5422 * simplify error handling. At worst an error after deleting
5423 * the entry first might leak the agent inode afterward. The
5424 * reverse would need filesystem abort in case of error deleting
5425 * the entry after the agent had been removed, or leave a
5426 * dangling entry pointing at a random inode.
5428 if (strcmp((char *)key, dotdot) != 0)
5429 osd_take_care_of_agent(env, osd, oh, de);
5430 rc = ldiskfs_delete_entry(oh->ot_handle, dir, de, bh);
5436 if (!rc && fid_is_namespace_visible(lu_object_fid(&dt->do_lu)) &&
5437 obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
5438 /* NB, dirent count may not be accurate, because it's counted
5441 if (obj->oo_dirent_count)
5442 obj->oo_dirent_count--;
5444 obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;
5447 ldiskfs_htree_unlock(hlock);
5449 up_write(&obj->oo_ext_idx_sem);
5452 LASSERT(osd_invariant(obj));
5453 osd_trans_exec_check(env, handle, OSD_OT_DELETE);
5458 * Lookup index for \a key and copy record to \a rec.
5460 * \param dt osd index object
5461 * \param key key for index
5462 * \param rec record reference
5464 * \retval +ve success : exact mach
5465 * \retval 0 return record with key not greater than \a key
5466 * \retval -ve failure
5468 static int osd_index_iam_lookup(const struct lu_env *env, struct dt_object *dt,
5469 struct dt_rec *rec, const struct dt_key *key)
5471 struct osd_object *obj = osd_dt_obj(dt);
5472 struct iam_path_descr *ipd;
5473 struct iam_container *bag = &obj->oo_dir->od_container;
5474 struct osd_thread_info *oti = osd_oti_get(env);
5475 struct iam_iterator *it = &oti->oti_idx_it;
5476 struct iam_rec *iam_rec;
5481 if (!dt_object_exists(dt))
5484 LASSERT(osd_invariant(obj));
5485 LASSERT(!dt_object_remote(dt));
5486 LASSERT(bag->ic_object == obj->oo_inode);
5488 ipd = osd_idx_ipd_get(env, bag);
5492 /* got ipd now we can start iterator. */
5493 iam_it_init(it, bag, 0, ipd);
5495 if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5496 /* swab quota uid/gid provided by caller */
5497 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5498 key = (const struct dt_key *)&oti->oti_quota_id;
5501 rc = iam_it_get(it, (struct iam_key *)key);
5503 if (S_ISDIR(obj->oo_inode->i_mode))
5504 iam_rec = (struct iam_rec *)oti->oti_ldp;
5506 iam_rec = (struct iam_rec *)rec;
5508 iam_reccpy(&it->ii_path.ip_leaf, (struct iam_rec *)iam_rec);
5510 if (S_ISDIR(obj->oo_inode->i_mode))
5511 osd_fid_unpack((struct lu_fid *)rec,
5512 (struct osd_fid_pack *)iam_rec);
5513 else if (fid_is_quota(lu_object_fid(&dt->do_lu)))
5514 osd_quota_unpack(obj, rec);
5519 osd_ipd_put(env, bag, ipd);
5521 LINVRNT(osd_invariant(obj));
5526 static int osd_index_declare_iam_insert(const struct lu_env *env,
5527 struct dt_object *dt,
5528 const struct dt_rec *rec,
5529 const struct dt_key *key,
5530 struct thandle *handle)
5532 struct osd_thandle *oh;
5534 LASSERT(handle != NULL);
5536 oh = container_of(handle, struct osd_thandle, ot_super);
5537 LASSERT(oh->ot_handle == NULL);
5539 osd_trans_declare_op(env, oh, OSD_OT_INSERT,
5540 osd_dto_credits_noquota[DTO_INDEX_INSERT]);
5546 * Inserts (key, value) pair in \a dt index object.
5548 * \param dt osd index object
5549 * \param key key for index
5550 * \param rec record reference
5551 * \param th transaction handler
5554 * \retval -ve failure
5556 static int osd_index_iam_insert(const struct lu_env *env, struct dt_object *dt,
5557 const struct dt_rec *rec,
5558 const struct dt_key *key, struct thandle *th)
5560 struct osd_object *obj = osd_dt_obj(dt);
5561 struct iam_path_descr *ipd;
5562 struct osd_thandle *oh;
5563 struct iam_container *bag;
5564 struct osd_thread_info *oti = osd_oti_get(env);
5565 struct iam_rec *iam_rec;
5570 if (!dt_object_exists(dt))
5573 LINVRNT(osd_invariant(obj));
5574 LASSERT(!dt_object_remote(dt));
5576 bag = &obj->oo_dir->od_container;
5577 LASSERT(bag->ic_object == obj->oo_inode);
5578 LASSERT(th != NULL);
5580 osd_trans_exec_op(env, th, OSD_OT_INSERT);
5582 ipd = osd_idx_ipd_get(env, bag);
5583 if (unlikely(ipd == NULL))
5586 oh = container_of(th, struct osd_thandle, ot_super);
5587 LASSERT(oh->ot_handle != NULL);
5588 LASSERT(oh->ot_handle->h_transaction != NULL);
5589 if (S_ISDIR(obj->oo_inode->i_mode)) {
5590 iam_rec = (struct iam_rec *)oti->oti_ldp;
5591 osd_fid_pack((struct osd_fid_pack *)iam_rec, rec,
5593 } else if (fid_is_quota(lu_object_fid(&dt->do_lu))) {
5594 /* pack quota uid/gid */
5595 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
5596 key = (const struct dt_key *)&oti->oti_quota_id;
5597 /* pack quota record */
5598 rec = osd_quota_pack(obj, rec, &oti->oti_quota_rec);
5599 iam_rec = (struct iam_rec *)rec;
5601 iam_rec = (struct iam_rec *)rec;
5604 rc = iam_insert(oh->ot_handle, bag, (const struct iam_key *)key,
5606 osd_ipd_put(env, bag, ipd);
5607 LINVRNT(osd_invariant(obj));
5608 osd_trans_exec_check(env, th, OSD_OT_INSERT);
5613 * Calls ldiskfs_add_entry() to add directory entry
5614 * into the directory. This is required for
5615 * interoperability mode (b11826)
5617 * \retval 0, on success
5618 * \retval -ve, on error
5620 static int __osd_ea_add_rec(struct osd_thread_info *info,
5621 struct osd_object *pobj, struct inode *cinode,
5622 const char *name, const struct lu_fid *fid,
5623 struct htree_lock *hlock, struct thandle *th)
5625 struct ldiskfs_dentry_param *ldp;
5626 struct dentry *child;
5627 struct osd_thandle *oth;
5630 oth = container_of(th, struct osd_thandle, ot_super);
5631 LASSERT(oth->ot_handle != NULL);
5632 LASSERT(oth->ot_handle->h_transaction != NULL);
5633 LASSERT(pobj->oo_inode);
5635 ldp = (struct ldiskfs_dentry_param *)info->oti_ldp;
5636 if (unlikely(osd_object_is_root(pobj)))
5639 osd_get_ldiskfs_dirent_param(ldp, fid);
5640 child = osd_child_dentry_get(info->oti_env, pobj, name, strlen(name));
5641 child->d_fsdata = (void *)ldp;
5642 dquot_initialize(pobj->oo_inode);
5643 rc = osd_ldiskfs_add_entry(info, osd_obj2dev(pobj), oth->ot_handle,
5644 child, cinode, hlock);
5645 if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_TYPE)) {
5646 struct ldiskfs_dir_entry_2 *de;
5647 struct buffer_head *bh;
5650 bh = osd_ldiskfs_find_entry(pobj->oo_inode, &child->d_name, &de,
5653 rc1 = ldiskfs_journal_get_write_access(oth->ot_handle,
5656 if (S_ISDIR(cinode->i_mode))
5657 de->file_type = LDISKFS_DIRENT_LUFID |
5658 LDISKFS_FT_REG_FILE;
5660 de->file_type = LDISKFS_DIRENT_LUFID |
5662 ldiskfs_handle_dirty_metadata(oth->ot_handle,
5673 * Calls ldiskfs_add_dot_dotdot() to add dot and dotdot entries
5674 * into the directory.Also sets flags into osd object to
5675 * indicate dot and dotdot are created. This is required for
5676 * interoperability mode (b11826)
5678 * \param dir directory for dot and dotdot fixup.
5679 * \param obj child object for linking
5681 * \retval 0, on success
5682 * \retval -ve, on error
5684 static int osd_add_dot_dotdot(struct osd_thread_info *info,
5685 struct osd_object *dir,
5686 struct inode *parent_dir, const char *name,
5687 const struct lu_fid *dot_fid,
5688 const struct lu_fid *dot_dot_fid,
5691 struct inode *inode = dir->oo_inode;
5692 struct osd_thandle *oth;
5695 oth = container_of(th, struct osd_thandle, ot_super);
5696 LASSERT(oth->ot_handle->h_transaction != NULL);
5697 LASSERT(S_ISDIR(dir->oo_inode->i_mode));
5699 if (strcmp(name, dot) == 0) {
5700 if (dir->oo_compat_dot_created) {
5703 LASSERT(inode->i_ino == parent_dir->i_ino);
5704 dir->oo_compat_dot_created = 1;
5707 } else if (strcmp(name, dotdot) == 0) {
5708 if (!dir->oo_compat_dot_created)
5710 /* in case of rename, dotdot is already created */
5711 if (dir->oo_compat_dotdot_created) {
5712 return __osd_ea_add_rec(info, dir, parent_dir, name,
5713 dot_dot_fid, NULL, th);
5716 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_PARENT)) {
5717 struct lu_fid tfid = *dot_dot_fid;
5720 result = osd_add_dot_dotdot_internal(info,
5721 dir->oo_inode, parent_dir, dot_fid,
5724 result = osd_add_dot_dotdot_internal(info,
5725 dir->oo_inode, parent_dir, dot_fid,
5730 dir->oo_compat_dotdot_created = 1;
5738 * It will call the appropriate osd_add* function and return the
5739 * value, return by respective functions.
5741 static int osd_ea_add_rec(const struct lu_env *env, struct osd_object *pobj,
5742 struct inode *cinode, const char *name,
5743 const struct lu_fid *fid, struct thandle *th)
5745 struct osd_thread_info *info = osd_oti_get(env);
5746 struct htree_lock *hlock;
5749 hlock = pobj->oo_hl_head != NULL ? info->oti_hlock : NULL;
5751 if (name[0] == '.' && (name[1] == '\0' ||
5752 (name[1] == '.' && name[2] == '\0'))) {
5753 if (hlock != NULL) {
5754 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5757 down_write(&pobj->oo_ext_idx_sem);
5760 rc = osd_add_dot_dotdot(info, pobj, cinode, name,
5761 lu_object_fid(&pobj->oo_dt.do_lu),
5764 if (hlock != NULL) {
5765 ldiskfs_htree_lock(hlock, pobj->oo_hl_head,
5766 pobj->oo_inode, LDISKFS_HLOCK_ADD);
5768 down_write(&pobj->oo_ext_idx_sem);
5771 if (OBD_FAIL_CHECK(OBD_FAIL_FID_INDIR)) {
5772 struct lu_fid *tfid = &info->oti_fid;
5776 rc = __osd_ea_add_rec(info, pobj, cinode, name,
5779 rc = __osd_ea_add_rec(info, pobj, cinode, name, fid,
5783 if (!rc && fid_is_namespace_visible(lu_object_fid(&pobj->oo_dt.do_lu))
5784 && pobj->oo_dirent_count != LU_DIRENT_COUNT_UNSET)
5785 pobj->oo_dirent_count++;
5788 ldiskfs_htree_unlock(hlock);
5790 up_write(&pobj->oo_ext_idx_sem);
5796 osd_consistency_check(struct osd_thread_info *oti, struct osd_device *dev,
5797 const struct lu_fid *fid, struct osd_inode_id *id)
5799 struct lustre_scrub *scrub = &dev->od_scrub.os_scrub;
5800 struct inode *inode = NULL;
5807 if (!fid_is_norm(fid) && !fid_is_igif(fid))
5810 if (scrub->os_running && scrub->os_pos_current > id->oii_ino)
5813 if (dev->od_auto_scrub_interval == AS_NEVER ||
5814 ktime_get_real_seconds() <
5815 scrub->os_file.sf_time_last_complete + dev->od_auto_scrub_interval)
5819 rc = osd_oi_lookup(oti, dev, fid, &oti->oti_id, 0);
5820 if (rc == -ENOENT) {
5821 __u32 gen = id->oii_gen;
5827 inode = osd_iget(oti, dev, id);
5828 /* The inode has been removed (by race maybe). */
5829 if (IS_ERR(inode)) {
5830 rc = PTR_ERR(inode);
5832 RETURN(rc == -ESTALE ? -ENOENT : rc);
5835 /* The OI mapping is lost. */
5836 if (gen != OSD_OII_NOGEN)
5840 * The inode may has been reused by others, we do not know,
5841 * leave it to be handled by subsequent osd_fid_lookup().
5844 } else if (rc || osd_id_eq(id, &oti->oti_id)) {
5851 if (scrub->os_running) {
5852 if (inode == NULL) {
5853 inode = osd_iget(oti, dev, id);
5854 /* The inode has been removed (by race maybe). */
5855 if (IS_ERR(inode)) {
5856 rc = PTR_ERR(inode);
5858 RETURN(rc == -ESTALE ? -ENOENT : rc);
5862 rc = osd_oii_insert(dev, fid, id, insert);
5864 * There is race condition between osd_oi_lookup and OI scrub.
5865 * The OI scrub finished just after osd_oi_lookup() failure.
5866 * Under such case, it is unnecessary to trigger OI scrub again,
5867 * but try to call osd_oi_lookup() again.
5869 if (unlikely(rc == -EAGAIN))
5872 if (!S_ISDIR(inode->i_mode))
5875 rc = osd_check_lmv(oti, dev, inode);
5880 if (dev->od_auto_scrub_interval != AS_NEVER && ++once == 1) {
5881 rc = osd_scrub_start(oti->oti_env, dev, SS_AUTO_PARTIAL |
5882 SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT);
5883 CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
5884 "%s: trigger partial OI scrub for RPC inconsistency, checking FID "DFID"/%u: rc = %d\n",
5885 osd_dev2name(dev), PFID(fid), id->oii_ino, rc);
5886 if (rc == 0 || rc == -EALREADY)
5898 static int osd_fail_fid_lookup(struct osd_thread_info *oti,
5899 struct osd_device *dev,
5900 struct lu_fid *fid, __u32 ino)
5902 struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;
5903 struct osd_idmap_cache *oic = &oti->oti_cache;
5904 struct inode *inode;
5907 osd_id_gen(&oic->oic_lid, ino, OSD_OII_NOGEN);
5908 inode = osd_iget(oti, dev, &oic->oic_lid);
5909 if (IS_ERR(inode)) {
5910 fid_zero(&oic->oic_fid);
5911 return PTR_ERR(inode);
5914 rc = osd_get_lma(oti, inode, &oti->oti_obj_dentry, loa);
5917 fid_zero(&oic->oic_fid);
5919 *fid = oic->oic_fid = loa->loa_lma.lma_self_fid;
5923 void osd_add_oi_cache(struct osd_thread_info *info, struct osd_device *osd,
5924 struct osd_inode_id *id, const struct lu_fid *fid)
5926 CDEBUG(D_INODE, "add "DFID" %u:%u to info %p\n", PFID(fid),
5927 id->oii_ino, id->oii_gen, info);
5928 info->oti_cache.oic_lid = *id;
5929 info->oti_cache.oic_fid = *fid;
5930 info->oti_cache.oic_dev = osd;
5934 * Get parent FID from the linkEA.
5936 * For a directory which parent resides on remote MDT, to satisfy the
5937 * local e2fsck, we insert it into the /REMOTE_PARENT_DIR locally. On
5938 * the other hand, to make the lookup(..) on the directory can return
5939 * the real parent FID, we append the real parent FID after its ".."
5940 * name entry in the /REMOTE_PARENT_DIR.
5942 * Unfortunately, such PFID-in-dirent cannot be preserved via file-level
5943 * backup. So after the restore, we cannot get the right parent FID from
5944 * its ".." name entry in the /REMOTE_PARENT_DIR. Under such case, since
5945 * we have stored the real parent FID in the directory object's linkEA,
5946 * we can parse the linkEA for the real parent FID.
5948 * \param[in] env pointer to the thread context
5949 * \param[in] obj pointer to the object to be handled
5950 * \param[out]fid pointer to the buffer to hold the parent FID
5952 * \retval 0 for getting the real parent FID successfully
5953 * \retval negative error number on failure
5955 static int osd_get_pfid_from_linkea(const struct lu_env *env,
5956 struct osd_object *obj,
5959 struct osd_thread_info *oti = osd_oti_get(env);
5960 struct lu_buf *buf = &oti->oti_big_buf;
5961 struct dentry *dentry = &oti->oti_obj_dentry;
5962 struct inode *inode = obj->oo_inode;
5963 struct linkea_data ldata = { NULL };
5969 if (!S_ISDIR(inode->i_mode))
5973 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5974 buf->lb_buf, buf->lb_len);
5975 if (rc == -ERANGE) {
5976 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
5979 lu_buf_realloc(buf, rc);
5980 if (buf->lb_buf == NULL)
5987 if (unlikely(rc == 0))
5993 if (unlikely(buf->lb_buf == NULL)) {
5994 lu_buf_realloc(buf, rc);
5995 if (buf->lb_buf == NULL)
6002 rc = linkea_init_with_rec(&ldata);
6004 linkea_first_entry(&ldata);
6005 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, NULL, fid);
6011 static int osd_verify_ent_by_linkea(const struct lu_env *env,
6012 struct inode *inode,
6013 const struct lu_fid *pfid,
6014 const char *name, const int namelen)
6016 struct osd_thread_info *oti = osd_oti_get(env);
6017 struct lu_buf *buf = &oti->oti_big_buf;
6018 struct dentry *dentry = &oti->oti_obj_dentry;
6019 struct linkea_data ldata = { NULL };
6020 struct lu_name cname = { .ln_name = name,
6021 .ln_namelen = namelen };
6027 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK,
6028 buf->lb_buf, buf->lb_len);
6030 rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LINK, NULL, 0);
6035 if (unlikely(rc == 0))
6038 if (buf->lb_len < rc) {
6039 lu_buf_realloc(buf, rc);
6040 if (buf->lb_buf == NULL)
6047 rc = linkea_init_with_rec(&ldata);
6049 rc = linkea_links_find(&ldata, &cname, pfid);
6055 * Calls ->lookup() to find dentry. From dentry get inode and
6056 * read inode's ea to get fid. This is required for interoperability
6059 * \retval 0, on success
6060 * \retval -ve, on error
6062 static int osd_ea_lookup_rec(const struct lu_env *env, struct osd_object *obj,
6063 struct dt_rec *rec, const struct dt_key *key)
6065 struct inode *dir = obj->oo_inode;
6066 struct dentry *dentry;
6067 struct ldiskfs_dir_entry_2 *de;
6068 struct buffer_head *bh;
6069 struct lu_fid *fid = (struct lu_fid *)rec;
6070 struct htree_lock *hlock = NULL;
6076 LASSERT(dir->i_op != NULL);
6077 LASSERT(dir->i_op->lookup != NULL);
6079 dentry = osd_child_dentry_get(env, obj,
6080 (char *)key, strlen((char *)key));
6082 if (obj->oo_hl_head != NULL) {
6083 hlock = osd_oti_get(env)->oti_hlock;
6084 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6085 dir, LDISKFS_HLOCK_LOOKUP);
6087 down_read(&obj->oo_ext_idx_sem);
6090 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
6092 struct osd_thread_info *oti = osd_oti_get(env);
6093 struct osd_inode_id *id = &oti->oti_id;
6094 struct osd_device *dev = osd_obj2dev(obj);
6096 ino = le32_to_cpu(de->inode);
6097 if (OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP)) {
6099 rc = osd_fail_fid_lookup(oti, dev, fid, ino);
6103 rc = osd_get_fid_from_dentry(de, rec);
6105 /* done with de, release bh */
6108 if (unlikely(is_remote_parent_ino(dev, ino))) {
6109 const char *name = (const char *)key;
6112 * If the parent is on remote MDT, and there
6113 * is no FID-in-dirent, then we have to get
6114 * the parent FID from the linkEA.
6116 if (likely(strlen(name) == 2 &&
6117 name[0] == '.' && name[1] == '.'))
6118 rc = osd_get_pfid_from_linkea(env, obj,
6121 rc = osd_ea_fid_get(env, obj, ino, fid, id);
6124 osd_id_gen(id, ino, OSD_OII_NOGEN);
6127 if (rc != 0 || osd_remote_fid(env, dev, fid))
6130 rc = osd_consistency_check(oti, dev, fid, id);
6131 if (rc != -ENOENT) {
6132 /* Other error should not affect lookup result. */
6135 /* Normal file mapping should be added into OI cache
6136 * after FID in LMA check, but for local files like
6137 * hsm_actions, their FIDs are not stored in OI files,
6138 * see osd_initial_OI_scrub(), and here is the only
6139 * place to load mapping into OI cache.
6141 if (!fid_is_namespace_visible(fid))
6142 osd_add_oi_cache(osd_oti_get(env),
6143 osd_obj2dev(obj), id, fid);
6153 ldiskfs_htree_unlock(hlock);
6155 up_read(&obj->oo_ext_idx_sem);
6159 static int osd_index_declare_ea_insert(const struct lu_env *env,
6160 struct dt_object *dt,
6161 const struct dt_rec *rec,
6162 const struct dt_key *key,
6163 struct thandle *handle)
6165 struct osd_thandle *oh;
6166 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6167 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6168 const struct lu_fid *fid = rec1->rec_fid;
6169 int credits, rc = 0;
6170 struct osd_idmap_cache *idc;
6174 LASSERT(!dt_object_remote(dt));
6175 LASSERT(handle != NULL);
6176 LASSERT(fid != NULL);
6177 LASSERT(rec1->rec_type != 0);
6179 oh = container_of(handle, struct osd_thandle, ot_super);
6180 LASSERT(oh->ot_handle == NULL);
6182 credits = osd_dto_credits_noquota[DTO_INDEX_INSERT];
6185 * we can't call iget() while a transactions is running
6186 * (this can lead to a deadlock), but we need to know
6187 * inum and object type. so we find this information at
6188 * declaration and cache in per-thread info
6190 idc = osd_idc_find_or_init(env, osd, fid);
6192 RETURN(PTR_ERR(idc));
6193 if (idc->oic_remote) {
6195 * a reference to remote inode is represented by an
6196 * agent inode which we have to create
6198 credits += osd_dto_credits_noquota[DTO_OBJECT_CREATE];
6199 credits += osd_dto_credits_noquota[DTO_INDEX_INSERT];
6202 osd_trans_declare_op(env, oh, OSD_OT_INSERT, credits);
6204 if (osd_dt_obj(dt)->oo_inode != NULL) {
6205 struct inode *inode = osd_dt_obj(dt)->oo_inode;
6208 * We ignore block quota on meta pool (MDTs), so needn't
6209 * calculate how many blocks will be consumed by this index
6212 rc = osd_declare_inode_qid(env, i_uid_read(inode),
6214 i_projid_read(inode), 0,
6215 oh, osd_dt_obj(dt), NULL,
6220 #ifdef HAVE_PROJECT_QUOTA
6222 * Reserve credits for local agent inode to transfer
6223 * to 0, quota enforcement is ignored in this case.
6225 if (idc->oic_remote &&
6226 LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL &&
6227 i_projid_read(inode) != 0)
6228 rc = osd_declare_attr_qid(env, osd_dt_obj(dt), oh,
6229 0, i_projid_read(inode),
6230 0, false, PRJQUOTA);
6238 * Index add function for interoperability mode (b11826).
6239 * It will add the directory entry.This entry is needed to
6240 * maintain name->fid mapping.
6242 * \param key it is key i.e. file entry to be inserted
6243 * \param rec it is value of given key i.e. fid
6245 * \retval 0, on success
6246 * \retval -ve, on error
6248 static int osd_index_ea_insert(const struct lu_env *env, struct dt_object *dt,
6249 const struct dt_rec *rec,
6250 const struct dt_key *key, struct thandle *th)
6252 struct osd_object *obj = osd_dt_obj(dt);
6253 struct osd_device *osd = osd_dev(dt->do_lu.lo_dev);
6254 struct dt_insert_rec *rec1 = (struct dt_insert_rec *)rec;
6255 const struct lu_fid *fid = rec1->rec_fid;
6256 const char *name = (const char *)key;
6257 struct osd_thread_info *oti = osd_oti_get(env);
6258 struct inode *child_inode = NULL;
6259 struct osd_idmap_cache *idc;
6264 if (!dt_object_exists(dt))
6267 LASSERT(osd_invariant(obj));
6268 LASSERT(!dt_object_remote(dt));
6269 LASSERT(th != NULL);
6271 osd_trans_exec_op(env, th, OSD_OT_INSERT);
6273 LASSERTF(fid_is_sane(fid), "fid"DFID" is insane!\n", PFID(fid));
6275 idc = osd_idc_find(env, osd, fid);
6276 if (unlikely(idc == NULL)) {
6277 idc = osd_idc_find_or_init(env, osd, fid);
6280 * this dt_insert() wasn't declared properly, so
6281 * FID is missing in OI cache. we better do not
6282 * lookup FID in FLDB/OI and don't risk to deadlock,
6283 * but in some special cases (lfsck testing, etc)
6284 * it's much simpler than fixing a caller.
6286 * normally this error should be placed after the first
6287 * find, but migrate may attach source stripes to
6288 * target, which doesn't create stripes.
6290 CERROR("%s: "DFID" wasn't declared for insert\n",
6291 osd_name(osd), PFID(fid));
6293 RETURN(PTR_ERR(idc));
6297 if (idc->oic_remote) {
6298 /* Insert remote entry */
6299 if (strcmp(name, dotdot) == 0 && strlen(name) == 2) {
6301 igrab(osd->od_mdt_map->omm_remote_parent->d_inode);
6303 child_inode = osd_create_local_agent_inode(env, osd,
6304 obj, fid, rec1->rec_type & S_IFMT, th);
6305 if (IS_ERR(child_inode))
6306 RETURN(PTR_ERR(child_inode));
6309 /* Insert local entry */
6310 if (unlikely(idc->oic_lid.oii_ino == 0)) {
6311 /* for a reason OI cache wasn't filled properly */
6312 CERROR("%s: OIC for "DFID" isn't filled\n",
6313 osd_name(osd), PFID(fid));
6316 child_inode = oti->oti_inode;
6317 if (unlikely(child_inode == NULL)) {
6318 struct ldiskfs_inode_info *lii;
6323 child_inode = oti->oti_inode = &lii->vfs_inode;
6325 child_inode->i_sb = osd_sb(osd);
6326 child_inode->i_ino = idc->oic_lid.oii_ino;
6327 child_inode->i_mode = rec1->rec_type & S_IFMT;
6330 rc = osd_ea_add_rec(env, obj, child_inode, name, fid, th);
6332 CDEBUG(D_INODE, "parent %lu insert %s:%lu rc = %d\n",
6333 obj->oo_inode->i_ino, name, child_inode->i_ino, rc);
6335 if (child_inode && child_inode != oti->oti_inode)
6337 LASSERT(osd_invariant(obj));
6338 osd_trans_exec_check(env, th, OSD_OT_INSERT);
6344 * Initialize osd Iterator for given osd index object.
6346 * \param dt osd index object
6349 static struct dt_it *osd_it_iam_init(const struct lu_env *env,
6350 struct dt_object *dt,
6353 struct osd_it_iam *it;
6354 struct osd_object *obj = osd_dt_obj(dt);
6355 struct lu_object *lo = &dt->do_lu;
6356 struct iam_path_descr *ipd;
6357 struct iam_container *bag = &obj->oo_dir->od_container;
6359 if (!dt_object_exists(dt))
6360 return ERR_PTR(-ENOENT);
6364 return ERR_PTR(-ENOMEM);
6366 ipd = osd_it_ipd_get(env, bag);
6367 if (likely(ipd != NULL)) {
6371 iam_it_init(&it->oi_it, bag, IAM_IT_MOVE, ipd);
6372 return (struct dt_it *)it;
6375 return ERR_PTR(-ENOMEM);
6380 * free given Iterator.
6382 static void osd_it_iam_fini(const struct lu_env *env, struct dt_it *di)
6384 struct osd_it_iam *it = (struct osd_it_iam *)di;
6385 struct osd_object *obj = it->oi_obj;
6387 iam_it_fini(&it->oi_it);
6388 osd_ipd_put(env, &obj->oo_dir->od_container, it->oi_ipd);
6389 osd_object_put(env, obj);
6394 * Move Iterator to record specified by \a key
6396 * \param di osd iterator
6397 * \param key key for index
6399 * \retval +ve di points to record with least key not larger than key
6400 * \retval 0 di points to exact matched key
6401 * \retval -ve failure
6404 static int osd_it_iam_get(const struct lu_env *env,
6405 struct dt_it *di, const struct dt_key *key)
6407 struct osd_thread_info *oti = osd_oti_get(env);
6408 struct osd_it_iam *it = (struct osd_it_iam *)di;
6410 if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6411 /* swab quota uid/gid */
6412 oti->oti_quota_id = cpu_to_le64(*((__u64 *)key));
6413 key = (struct dt_key *)&oti->oti_quota_id;
6416 return iam_it_get(&it->oi_it, (const struct iam_key *)key);
6422 * \param di osd iterator
6424 static void osd_it_iam_put(const struct lu_env *env, struct dt_it *di)
6426 struct osd_it_iam *it = (struct osd_it_iam *)di;
6428 iam_it_put(&it->oi_it);
6432 * Move iterator by one record
6434 * \param di osd iterator
6436 * \retval +1 end of container reached
6438 * \retval -ve failure
6441 static int osd_it_iam_next(const struct lu_env *env, struct dt_it *di)
6443 struct osd_it_iam *it = (struct osd_it_iam *)di;
6445 return iam_it_next(&it->oi_it);
6449 * Return pointer to the key under iterator.
6452 static struct dt_key *osd_it_iam_key(const struct lu_env *env,
6453 const struct dt_it *di)
6455 struct osd_thread_info *oti = osd_oti_get(env);
6456 struct osd_it_iam *it = (struct osd_it_iam *)di;
6457 struct osd_object *obj = it->oi_obj;
6460 key = (struct dt_key *)iam_it_key_get(&it->oi_it);
6462 if (!IS_ERR(key) && fid_is_quota(lu_object_fid(&obj->oo_dt.do_lu))) {
6463 /* swab quota uid/gid */
6464 oti->oti_quota_id = le64_to_cpu(*((__u64 *)key));
6465 key = (struct dt_key *)&oti->oti_quota_id;
6472 * Return size of key under iterator (in bytes)
6475 static int osd_it_iam_key_size(const struct lu_env *env, const struct dt_it *di)
6477 struct osd_it_iam *it = (struct osd_it_iam *)di;
6479 return iam_it_key_size(&it->oi_it);
6483 osd_it_append_attrs(struct lu_dirent *ent, int len, __u16 type)
6485 /* check if file type is required */
6486 if (ent->lde_attrs & LUDA_TYPE) {
6487 struct luda_type *lt;
6488 int align = sizeof(*lt) - 1;
6490 len = (len + align) & ~align;
6491 lt = (struct luda_type *)(ent->lde_name + len);
6492 lt->lt_type = cpu_to_le16(DTTOIF(type));
6495 ent->lde_attrs = cpu_to_le32(ent->lde_attrs);
6499 * build lu direct from backend fs dirent.
6503 osd_it_pack_dirent(struct lu_dirent *ent, struct lu_fid *fid, __u64 offset,
6504 char *name, __u16 namelen, __u16 type, __u32 attr)
6506 ent->lde_attrs = attr | LUDA_FID;
6507 fid_cpu_to_le(&ent->lde_fid, fid);
6509 ent->lde_hash = cpu_to_le64(offset);
6510 ent->lde_reclen = cpu_to_le16(lu_dirent_calc_size(namelen, attr));
6512 strncpy(ent->lde_name, name, namelen);
6513 ent->lde_name[namelen] = '\0';
6514 ent->lde_namelen = cpu_to_le16(namelen);
6516 /* append lustre attributes */
6517 osd_it_append_attrs(ent, namelen, type);
6521 * Return pointer to the record under iterator.
6523 static int osd_it_iam_rec(const struct lu_env *env,
6524 const struct dt_it *di,
6525 struct dt_rec *dtrec, __u32 attr)
6527 struct osd_it_iam *it = (struct osd_it_iam *)di;
6528 struct osd_thread_info *info = osd_oti_get(env);
6532 if (S_ISDIR(it->oi_obj->oo_inode->i_mode)) {
6533 const struct osd_fid_pack *rec;
6534 struct lu_fid *fid = &info->oti_fid;
6535 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
6541 name = (char *)iam_it_key_get(&it->oi_it);
6543 RETURN(PTR_ERR(name));
6545 namelen = iam_it_key_size(&it->oi_it);
6547 rec = (const struct osd_fid_pack *)iam_it_rec_get(&it->oi_it);
6549 RETURN(PTR_ERR(rec));
6551 rc = osd_fid_unpack(fid, rec);
6555 hash = iam_it_store(&it->oi_it);
6557 /* IAM does not store object type in IAM index (dir) */
6558 osd_it_pack_dirent(lde, fid, hash, name, namelen,
6560 } else if (fid_is_quota(lu_object_fid(&it->oi_obj->oo_dt.do_lu))) {
6561 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6562 (struct iam_rec *)dtrec);
6563 osd_quota_unpack(it->oi_obj, dtrec);
6565 iam_reccpy(&it->oi_it.ii_path.ip_leaf,
6566 (struct iam_rec *)dtrec);
6573 * Returns cookie for current Iterator position.
6575 static __u64 osd_it_iam_store(const struct lu_env *env, const struct dt_it *di)
6577 struct osd_it_iam *it = (struct osd_it_iam *)di;
6579 return iam_it_store(&it->oi_it);
6583 * Restore iterator from cookie.
6585 * \param di osd iterator
6586 * \param hash Iterator location cookie
6588 * \retval +ve di points to record with least key not larger than key.
6589 * \retval 0 di points to exact matched key
6590 * \retval -ve failure
6593 static int osd_it_iam_load(const struct lu_env *env,
6594 const struct dt_it *di, __u64 hash)
6596 struct osd_it_iam *it = (struct osd_it_iam *)di;
6598 return iam_it_load(&it->oi_it, hash);
6601 static const struct dt_index_operations osd_index_iam_ops = {
6602 .dio_lookup = osd_index_iam_lookup,
6603 .dio_declare_insert = osd_index_declare_iam_insert,
6604 .dio_insert = osd_index_iam_insert,
6605 .dio_declare_delete = osd_index_declare_iam_delete,
6606 .dio_delete = osd_index_iam_delete,
6608 .init = osd_it_iam_init,
6609 .fini = osd_it_iam_fini,
6610 .get = osd_it_iam_get,
6611 .put = osd_it_iam_put,
6612 .next = osd_it_iam_next,
6613 .key = osd_it_iam_key,
6614 .key_size = osd_it_iam_key_size,
6615 .rec = osd_it_iam_rec,
6616 .store = osd_it_iam_store,
6617 .load = osd_it_iam_load
6621 struct osd_it_ea *osd_it_dir_init(const struct lu_env *env,
6622 struct inode *inode, __u32 attr)
6624 struct osd_thread_info *info = osd_oti_get(env);
6625 struct osd_it_ea *oie;
6627 struct dentry *obj_dentry;
6631 OBD_SLAB_ALLOC_PTR_GFP(oie, osd_itea_cachep, GFP_NOFS);
6633 RETURN(ERR_PTR(-ENOMEM));
6634 obj_dentry = &oie->oie_dentry;
6636 obj_dentry->d_inode = inode;
6637 obj_dentry->d_sb = inode->i_sb;
6638 obj_dentry->d_name.hash = 0;
6640 oie->oie_rd_dirent = 0;
6641 oie->oie_it_dirent = 0;
6642 oie->oie_dirent = NULL;
6643 if (unlikely(!info->oti_it_ea_buf_used)) {
6644 oie->oie_buf = info->oti_it_ea_buf;
6645 info->oti_it_ea_buf_used = 1;
6647 OBD_ALLOC(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6648 if (oie->oie_buf == NULL)
6649 RETURN(ERR_PTR(-ENOMEM));
6651 oie->oie_obj = NULL;
6653 file = &oie->oie_file;
6655 /* Only FMODE_64BITHASH or FMODE_32BITHASH should be set, NOT both. */
6656 if (attr & LUDA_64BITHASH)
6657 file->f_mode = FMODE_64BITHASH;
6659 file->f_mode = FMODE_32BITHASH;
6660 file->f_path.dentry = obj_dentry;
6661 file->f_mapping = inode->i_mapping;
6662 file->f_op = inode->i_fop;
6663 file->f_inode = inode;
6669 * Creates or initializes iterator context.
6671 * \retval struct osd_it_ea, iterator structure on success
6674 static struct dt_it *osd_it_ea_init(const struct lu_env *env,
6675 struct dt_object *dt,
6678 struct osd_object *obj = osd_dt_obj(dt);
6679 struct lu_object *lo = &dt->do_lu;
6680 struct osd_it_ea *oie;
6684 if (!dt_object_exists(dt) || obj->oo_destroyed)
6685 RETURN(ERR_PTR(-ENOENT));
6687 oie = osd_it_dir_init(env, obj->oo_inode, attr);
6689 RETURN((struct dt_it *)oie);
6693 RETURN((struct dt_it *)oie);
6696 void osd_it_dir_fini(const struct lu_env *env, struct osd_it_ea *oie,
6697 struct inode *inode)
6699 struct osd_thread_info *info = osd_oti_get(env);
6702 oie->oie_file.f_op->release(inode, &oie->oie_file);
6703 if (unlikely(oie->oie_buf != info->oti_it_ea_buf))
6704 OBD_FREE(oie->oie_buf, OSD_IT_EA_BUFSIZE);
6706 info->oti_it_ea_buf_used = 0;
6707 OBD_SLAB_FREE_PTR(oie, osd_itea_cachep);
6712 * Destroy or finishes iterator context.
6714 * \param di iterator structure to be destroyed
6716 static void osd_it_ea_fini(const struct lu_env *env, struct dt_it *di)
6718 struct osd_it_ea *oie = (struct osd_it_ea *)di;
6719 struct osd_object *obj = oie->oie_obj;
6720 struct inode *inode = obj->oo_inode;
6723 osd_it_dir_fini(env, (struct osd_it_ea *)di, inode);
6724 osd_object_put(env, obj);
6729 * It position the iterator at given key, so that next lookup continues from
6730 * that key Or it is similar to dio_it->load() but based on a key,
6731 * rather than file position.
6733 * As a special convention, osd_it_ea_get(env, di, "") has to rewind iterator
6736 * TODO: Presently return +1 considering it is only used by mdd_dir_is_empty().
6738 static int osd_it_ea_get(const struct lu_env *env,
6739 struct dt_it *di, const struct dt_key *key)
6741 struct osd_it_ea *it = (struct osd_it_ea *)di;
6744 LASSERT(((const char *)key)[0] == '\0');
6745 it->oie_file.f_pos = 0;
6746 it->oie_rd_dirent = 0;
6747 it->oie_it_dirent = 0;
6748 it->oie_dirent = NULL;
6756 static void osd_it_ea_put(const struct lu_env *env, struct dt_it *di)
6760 struct osd_filldir_cbs {
6761 struct dir_context ctx;
6762 struct osd_it_ea *it;
6765 * It is called internally by ->iterate*(). It fills the
6766 * iterator's in-memory data structure with required
6767 * information i.e. name, namelen, rec_size etc.
6769 * \param buf in which information to be filled in.
6770 * \param name name of the file in given dir
6772 * \retval 0 on success
6773 * \retval 1 on buffer full
6775 #ifdef HAVE_FILLDIR_USE_CTX
6776 static int osd_ldiskfs_filldir(struct dir_context *buf,
6778 static int osd_ldiskfs_filldir(void *buf,
6780 const char *name, int namelen,
6781 loff_t offset, __u64 ino, unsigned int d_type)
6783 struct osd_it_ea *it = ((struct osd_filldir_cbs *)buf)->it;
6784 struct osd_object *obj = it->oie_obj;
6785 struct osd_it_ea_dirent *ent = it->oie_dirent;
6786 struct lu_fid *fid = &ent->oied_fid;
6787 struct osd_fid_pack *rec;
6791 /* this should never happen */
6792 if (unlikely(namelen == 0 || namelen > LDISKFS_NAME_LEN)) {
6793 CERROR("ldiskfs return invalid namelen %d\n", namelen);
6797 if ((void *)ent - it->oie_buf + sizeof(*ent) + namelen >
6801 /* "." is just the object itself. */
6802 if (namelen == 1 && name[0] == '.') {
6804 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6805 } else if (d_type & LDISKFS_DIRENT_LUFID) {
6806 rec = (struct osd_fid_pack *)(name + namelen + 1);
6807 if (osd_fid_unpack(fid, rec) != 0)
6812 d_type &= ~LDISKFS_DIRENT_LUFID;
6814 /* NOT export local root. */
6816 unlikely(osd_sb(osd_obj2dev(obj))->s_root->d_inode->i_ino == ino)) {
6817 ino = obj->oo_inode->i_ino;
6818 *fid = obj->oo_dt.do_lu.lo_header->loh_fid;
6821 ent->oied_ino = ino;
6822 ent->oied_off = offset;
6823 ent->oied_namelen = namelen;
6824 ent->oied_type = d_type;
6826 memcpy(ent->oied_name, name, namelen);
6828 it->oie_rd_dirent++;
6829 it->oie_dirent = (void *)ent + cfs_size_round(sizeof(*ent) + namelen);
6834 * Calls ->iterate*() to load a directory entry at a time
6835 * and stored it in iterator's in-memory data structure.
6837 * \param di iterator's in memory structure
6839 * \retval 0 on success
6840 * \retval -ve on error
6841 * \retval +1 reach the end of entry
6843 int osd_ldiskfs_it_fill(const struct lu_env *env, const struct dt_it *di)
6845 struct osd_it_ea *it = (struct osd_it_ea *)di;
6846 struct osd_object *obj = it->oie_obj;
6847 struct htree_lock *hlock = NULL;
6848 struct file *filp = &it->oie_file;
6850 struct osd_filldir_cbs buf = {
6851 .ctx.actor = osd_ldiskfs_filldir,
6856 it->oie_dirent = it->oie_buf;
6857 it->oie_rd_dirent = 0;
6860 if (obj->oo_hl_head != NULL) {
6861 hlock = osd_oti_get(env)->oti_hlock;
6862 ldiskfs_htree_lock(hlock, obj->oo_hl_head,
6864 LDISKFS_HLOCK_READDIR);
6866 down_read(&obj->oo_ext_idx_sem);
6870 filp->f_cred = current_cred();
6871 rc = osd_security_file_alloc(filp);
6875 filp->f_flags |= O_NOATIME;
6876 filp->f_mode |= FMODE_NONOTIFY;
6877 rc = iterate_dir(filp, &buf.ctx);
6881 if (it->oie_rd_dirent == 0) {
6883 * If it does not get any dirent, it means it has been reached
6884 * to the end of the dir
6886 it->oie_file.f_pos = ldiskfs_get_htree_eof(&it->oie_file);
6890 it->oie_dirent = it->oie_buf;
6891 it->oie_it_dirent = 1;
6896 ldiskfs_htree_unlock(hlock);
6898 up_read(&obj->oo_ext_idx_sem);
6905 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
6906 * to load a directory entry at a time and stored it in
6907 * iterator's in-memory data structure.
6909 * \param di iterator's in memory structure
6911 * \retval +ve iterator reached to end
6912 * \retval 0 iterator not reached to end
6913 * \retval -ve on error
6915 static int osd_it_ea_next(const struct lu_env *env, struct dt_it *di)
6917 struct osd_it_ea *it = (struct osd_it_ea *)di;
6922 if (it->oie_it_dirent < it->oie_rd_dirent) {
6924 (void *)it->oie_dirent +
6925 cfs_size_round(sizeof(struct osd_it_ea_dirent) +
6926 it->oie_dirent->oied_namelen);
6927 it->oie_it_dirent++;
6930 if (it->oie_file.f_pos == ldiskfs_get_htree_eof(&it->oie_file))
6933 rc = osd_ldiskfs_it_fill(env, di);
6940 * Returns the key at current position from iterator's in memory structure.
6942 * \param di iterator's in memory structure
6944 * \retval key i.e. struct dt_key on success
6946 static struct dt_key *osd_it_ea_key(const struct lu_env *env,
6947 const struct dt_it *di)
6949 struct osd_it_ea *it = (struct osd_it_ea *)di;
6951 return (struct dt_key *)it->oie_dirent->oied_name;
6955 * Returns key's size at current position from iterator's in memory structure.
6957 * \param di iterator's in memory structure
6959 * \retval key_size i.e. struct dt_key on success
6961 static int osd_it_ea_key_size(const struct lu_env *env, const struct dt_it *di)
6963 struct osd_it_ea *it = (struct osd_it_ea *)di;
6965 return it->oie_dirent->oied_namelen;
6968 #if defined LDISKFS_DIR_ENTRY_LEN && defined LDISKFS_DIR_ENTRY_LEN_
6969 #undef LDISKFS_DIR_REC_LEN
6970 #define LDISKFS_DIR_REC_LEN(de) LDISKFS_DIR_ENTRY_LEN_((de))
6973 static inline bool osd_dotdot_has_space(struct ldiskfs_dir_entry_2 *de)
6975 if (LDISKFS_DIR_REC_LEN(de) >=
6976 __LDISKFS_DIR_REC_LEN(2 + 1 + sizeof(struct osd_fid_pack)))
6983 osd_dirent_has_space(struct ldiskfs_dir_entry_2 *de, __u16 namelen,
6984 unsigned int blocksize, bool dotdot)
6987 return osd_dotdot_has_space(de);
6989 if (ldiskfs_rec_len_from_disk(de->rec_len, blocksize) >=
6990 __LDISKFS_DIR_REC_LEN(namelen + 1 + sizeof(struct osd_fid_pack)))
6997 osd_dirent_reinsert(const struct lu_env *env, struct osd_device *dev,
6998 handle_t *jh, struct dentry *dentry,
6999 const struct lu_fid *fid, struct buffer_head *bh,
7000 struct ldiskfs_dir_entry_2 *de, struct htree_lock *hlock,
7003 struct inode *dir = dentry->d_parent->d_inode;
7004 struct inode *inode = dentry->d_inode;
7005 struct osd_fid_pack *rec;
7006 struct ldiskfs_dentry_param *ldp;
7007 int namelen = dentry->d_name.len;
7009 struct osd_thread_info *info = osd_oti_get(env);
7013 if (!ldiskfs_has_feature_dirdata(inode->i_sb))
7016 /* There is enough space to hold the FID-in-dirent. */
7017 if (osd_dirent_has_space(de, namelen, dir->i_sb->s_blocksize, dotdot)) {
7018 rc = ldiskfs_journal_get_write_access(jh, bh);
7022 de->name[namelen] = 0;
7023 rec = (struct osd_fid_pack *)(de->name + namelen + 1);
7024 rec->fp_len = sizeof(struct lu_fid) + 1;
7025 fid_cpu_to_be((struct lu_fid *)rec->fp_area, fid);
7026 de->file_type |= LDISKFS_DIRENT_LUFID;
7027 rc = ldiskfs_handle_dirty_metadata(jh, NULL, bh);
7034 rc = ldiskfs_delete_entry(jh, dir, de, bh);
7038 ldp = (struct ldiskfs_dentry_param *)osd_oti_get(env)->oti_ldp;
7039 osd_get_ldiskfs_dirent_param(ldp, fid);
7040 dentry->d_fsdata = (void *)ldp;
7041 dquot_initialize(dir);
7042 rc = osd_ldiskfs_add_entry(info, dev, jh, dentry, inode, hlock);
7044 * It is too bad, we cannot reinsert the name entry back.
7045 * That means we lose it!
7049 "%s: fail to reinsert the dirent, dir = %lu/%u, name = %.*s, "DFID": rc = %d\n",
7050 osd_ino2name(inode), dir->i_ino, dir->i_generation,
7051 namelen, dentry->d_name.name, PFID(fid), rc);
7057 osd_dirent_check_repair(const struct lu_env *env, struct osd_object *obj,
7058 struct osd_it_ea *it, struct lu_fid *fid,
7059 struct osd_inode_id *id, __u32 *attr)
7061 struct osd_thread_info *info = osd_oti_get(env);
7062 struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;
7063 struct osd_device *dev = osd_obj2dev(obj);
7064 struct super_block *sb = osd_sb(dev);
7065 const char *devname = osd_name(dev);
7066 struct osd_it_ea_dirent *ent = it->oie_dirent;
7067 struct inode *dir = obj->oo_inode;
7068 struct htree_lock *hlock = NULL;
7069 struct buffer_head *bh = NULL;
7070 handle_t *jh = NULL;
7071 struct ldiskfs_dir_entry_2 *de;
7072 struct dentry *dentry;
7073 struct inode *inode;
7074 const struct lu_fid *pfid = lu_object_fid(&obj->oo_dt.do_lu);
7077 bool dotdot = false;
7082 if (ent->oied_name[0] == '.') {
7083 if (ent->oied_namelen == 1)
7086 if (ent->oied_namelen == 2 && ent->oied_name[1] == '.')
7090 osd_id_gen(id, ent->oied_ino, OSD_OII_NOGEN);
7091 inode = osd_iget(info, dev, id);
7092 if (IS_ERR(inode)) {
7093 rc = PTR_ERR(inode);
7094 if (rc == -ENOENT || rc == -ESTALE) {
7096 * Maybe dangling name entry, or
7097 * corrupted directory entry.
7099 *attr |= LUDA_UNKNOWN;
7102 CDEBUG(D_LFSCK, "%s: fail to iget() for dirent "
7103 "check_repair, dir = %lu/%u, name = %.*s, "
7104 "ino = %llu, rc = %d\n",
7105 devname, dir->i_ino, dir->i_generation,
7106 ent->oied_namelen, ent->oied_name,
7113 dentry = osd_child_dentry_by_inode(env, dir, ent->oied_name,
7115 rc = osd_get_lma(info, inode, dentry, &info->oti_ost_attrs);
7116 if (rc == -ENODATA || !fid_is_sane(&lma->lma_self_fid))
7122 * We need to ensure that the name entry is still valid.
7123 * Because it may be removed or renamed by other already.
7125 * The unlink or rename operation will start journal before PDO lock,
7126 * so to avoid deadlock, here we need to start journal handle before
7127 * related PDO lock also. But because we do not know whether there
7128 * will be something to be repaired before PDO lock, we just start
7129 * journal without conditions.
7131 * We may need to remove the name entry firstly, then insert back.
7132 * One credit is for user quota file update.
7133 * One credit is for group quota file update.
7134 * Two credits are for dirty inode.
7136 credits = osd_dto_credits_noquota[DTO_INDEX_DELETE] +
7137 osd_dto_credits_noquota[DTO_INDEX_INSERT] + 1 + 1 + 2;
7139 if (dev->od_dirent_journal != 0) {
7142 jh = osd_journal_start_sb(sb, LDISKFS_HT_MISC, credits);
7145 CDEBUG(D_LFSCK, "%s: fail to start trans for dirent "
7146 "check_repair, dir = %lu/%u, credits = %d, "
7147 "name = %.*s, ino = %llu: rc = %d\n",
7148 devname, dir->i_ino, dir->i_generation, credits,
7149 ent->oied_namelen, ent->oied_name,
7152 GOTO(out_inode, rc);
7155 if (obj->oo_hl_head != NULL) {
7156 hlock = osd_oti_get(env)->oti_hlock;
7158 * "0" means exclusive lock for the whole directory.
7159 * We need to prevent others access such name entry
7160 * during the delete + insert. Neither HLOCK_ADD nor
7161 * HLOCK_DEL cannot guarantee the atomicity.
7163 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir, 0);
7165 down_write(&obj->oo_ext_idx_sem);
7168 if (obj->oo_hl_head != NULL) {
7169 hlock = osd_oti_get(env)->oti_hlock;
7170 ldiskfs_htree_lock(hlock, obj->oo_hl_head, dir,
7171 LDISKFS_HLOCK_LOOKUP);
7173 down_read(&obj->oo_ext_idx_sem);
7177 bh = osd_ldiskfs_find_entry(dir, &dentry->d_name, &de, NULL, hlock);
7178 if (IS_ERR(bh) || le32_to_cpu(de->inode) != inode->i_ino) {
7179 *attr |= LUDA_IGNORE;
7185 * For dotdot entry, if there is not enough space to hold the
7186 * FID-in-dirent, just keep them there. It only happens when the
7187 * device upgraded from 1.8 or restored from MDT file-level backup.
7188 * For the whole directory, only dotdot entry have no FID-in-dirent
7189 * and needs to get FID from LMA when readdir, it will not affect the
7192 if (dotdot && !osd_dotdot_has_space(de)) {
7193 *attr |= LUDA_UNKNOWN;
7199 if (lu_fid_eq(fid, &lma->lma_self_fid))
7202 if (unlikely(lma->lma_compat & LMAC_NOT_IN_OI)) {
7203 struct lu_fid *tfid = &lma->lma_self_fid;
7205 if (likely(dotdot &&
7206 fid_seq(tfid) == FID_SEQ_LOCAL_FILE &&
7207 fid_oid(tfid) == REMOTE_PARENT_DIR_OID)) {
7209 * It must be REMOTE_PARENT_DIR and as the
7210 * 'dotdot' entry of remote directory
7212 *attr |= LUDA_IGNORE;
7214 CDEBUG(D_LFSCK, "%s: expect remote agent "
7215 "parent directory, but got %.*s under "
7216 "dir = %lu/%u with the FID "DFID"\n",
7217 devname, ent->oied_namelen,
7218 ent->oied_name, dir->i_ino,
7219 dir->i_generation, PFID(tfid));
7221 *attr |= LUDA_UNKNOWN;
7228 if (!fid_is_zero(fid)) {
7229 rc = osd_verify_ent_by_linkea(env, inode, pfid, ent->oied_name,
7231 if (rc == -ENOENT ||
7233 !(dev->od_scrub.os_scrub.os_file.sf_flags & SF_UPGRADE))) {
7235 * linkEA does not recognize the dirent entry,
7236 * it may because the dirent entry corruption
7237 * and points to other's inode.
7239 CDEBUG(D_LFSCK, "%s: the target inode does not "
7240 "recognize the dirent, dir = %lu/%u, "
7241 " name = %.*s, ino = %llu, "
7242 DFID": rc = %d\n", devname, dir->i_ino,
7243 dir->i_generation, ent->oied_namelen,
7244 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7245 *attr |= LUDA_UNKNOWN;
7250 if (rc && rc != -ENODATA) {
7251 CDEBUG(D_LFSCK, "%s: fail to verify FID in the dirent, "
7252 "dir = %lu/%u, name = %.*s, ino = %llu, "
7253 DFID": rc = %d\n", devname, dir->i_ino,
7254 dir->i_generation, ent->oied_namelen,
7255 ent->oied_name, ent->oied_ino, PFID(fid), rc);
7256 *attr |= LUDA_UNKNOWN;
7264 * linkEA recognizes the dirent entry, the FID-in-LMA is
7265 * valid, trusted, in spite of fid_is_sane(fid) or not.
7267 if (*attr & LUDA_VERIFY_DRYRUN) {
7268 *fid = lma->lma_self_fid;
7269 *attr |= LUDA_REPAIR;
7276 dev->od_dirent_journal = 1;
7277 if (hlock != NULL) {
7278 ldiskfs_htree_unlock(hlock);
7281 up_read(&obj->oo_ext_idx_sem);
7287 *fid = lma->lma_self_fid;
7289 /* Update or append the FID-in-dirent. */
7290 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7291 bh, de, hlock, dotdot);
7293 *attr |= LUDA_REPAIR;
7295 CDEBUG(D_LFSCK, "%s: fail to re-insert FID after "
7296 "the dirent, dir = %lu/%u, name = %.*s, "
7297 "ino = %llu, "DFID": rc = %d\n",
7298 devname, dir->i_ino, dir->i_generation,
7299 ent->oied_namelen, ent->oied_name,
7300 ent->oied_ino, PFID(fid), rc);
7302 /* lma is NULL, trust the FID-in-dirent if it is valid. */
7303 if (*attr & LUDA_VERIFY_DRYRUN) {
7304 if (fid_is_sane(fid)) {
7305 *attr |= LUDA_REPAIR;
7306 } else if (dev->od_index == 0) {
7307 lu_igif_build(fid, inode->i_ino,
7308 inode->i_generation);
7309 *attr |= LUDA_UPGRADE;
7317 dev->od_dirent_journal = 1;
7318 if (hlock != NULL) {
7319 ldiskfs_htree_unlock(hlock);
7322 up_read(&obj->oo_ext_idx_sem);
7329 if (unlikely(fid_is_sane(fid))) {
7331 * FID-in-dirent exists, but FID-in-LMA is lost.
7332 * Trust the FID-in-dirent, and add FID-in-LMA.
7334 rc = osd_ea_fid_set(info, inode, fid, 0, 0);
7336 *attr |= LUDA_REPAIR;
7338 CDEBUG(D_LFSCK, "%s: fail to set LMA for "
7339 "update dirent, dir = %lu/%u, "
7340 "name = %.*s, ino = %llu, "
7342 devname, dir->i_ino, dir->i_generation,
7343 ent->oied_namelen, ent->oied_name,
7344 ent->oied_ino, PFID(fid), rc);
7345 } else if (dev->od_index == 0) {
7346 lu_igif_build(fid, inode->i_ino, inode->i_generation);
7348 * It is probably IGIF object. Only aappend the
7349 * FID-in-dirent. OI scrub will process FID-in-LMA.
7351 rc = osd_dirent_reinsert(env, dev, jh, dentry, fid,
7352 bh, de, hlock, dotdot);
7354 *attr |= LUDA_UPGRADE;
7356 CDEBUG(D_LFSCK, "%s: fail to append IGIF "
7357 "after the dirent, dir = %lu/%u, "
7358 "name = %.*s, ino = %llu, "
7360 devname, dir->i_ino, dir->i_generation,
7361 ent->oied_namelen, ent->oied_name,
7362 ent->oied_ino, PFID(fid), rc);
7371 if (hlock != NULL) {
7372 ldiskfs_htree_unlock(hlock);
7374 if (dev->od_dirent_journal != 0)
7375 up_write(&obj->oo_ext_idx_sem);
7377 up_read(&obj->oo_ext_idx_sem);
7381 ldiskfs_journal_stop(jh);
7385 if (rc >= 0 && !dirty)
7386 dev->od_dirent_journal = 0;
7392 * Returns the value at current position from iterator's in memory structure.
7394 * \param di struct osd_it_ea, iterator's in memory structure
7395 * \param attr attr requested for dirent.
7396 * \param lde lustre dirent
7398 * \retval 0 no error and \param lde has correct lustre dirent.
7399 * \retval -ve on error
7401 static inline int osd_it_ea_rec(const struct lu_env *env,
7402 const struct dt_it *di,
7403 struct dt_rec *dtrec, __u32 attr)
7405 struct osd_it_ea *it = (struct osd_it_ea *)di;
7406 struct osd_object *obj = it->oie_obj;
7407 struct osd_device *dev = osd_obj2dev(obj);
7408 struct osd_thread_info *oti = osd_oti_get(env);
7409 struct osd_inode_id *id = &oti->oti_id;
7410 struct lu_fid *fid = &it->oie_dirent->oied_fid;
7411 struct lu_dirent *lde = (struct lu_dirent *)dtrec;
7412 __u32 ino = it->oie_dirent->oied_ino;
7417 LASSERT(!is_remote_parent_ino(dev, obj->oo_inode->i_ino));
7419 if (attr & LUDA_VERIFY) {
7420 if (unlikely(is_remote_parent_ino(dev, ino))) {
7421 attr |= LUDA_IGNORE;
7423 * If the parent is on remote MDT, and there
7424 * is no FID-in-dirent, then we have to get
7425 * the parent FID from the linkEA.
7427 if (!fid_is_sane(fid) &&
7428 it->oie_dirent->oied_namelen == 2 &&
7429 it->oie_dirent->oied_name[0] == '.' &&
7430 it->oie_dirent->oied_name[1] == '.')
7431 osd_get_pfid_from_linkea(env, obj, fid);
7433 rc = osd_dirent_check_repair(env, obj, it, fid, id,
7437 if (!fid_is_sane(fid))
7438 attr |= LUDA_UNKNOWN;
7440 attr &= ~LU_DIRENT_ATTRS_MASK;
7441 if (!fid_is_sane(fid)) {
7442 bool is_dotdot = false;
7444 if (it->oie_dirent->oied_namelen == 2 &&
7445 it->oie_dirent->oied_name[0] == '.' &&
7446 it->oie_dirent->oied_name[1] == '.')
7449 * If the parent is on remote MDT, and there
7450 * is no FID-in-dirent, then we have to get
7451 * the parent FID from the linkEA.
7453 if (is_remote_parent_ino(dev, ino) && is_dotdot) {
7454 rc = osd_get_pfid_from_linkea(env, obj, fid);
7456 if (is_dotdot == false &&
7457 OBD_FAIL_CHECK(OBD_FAIL_FID_LOOKUP))
7460 rc = osd_ea_fid_get(env, obj, ino, fid, id);
7465 /* Pack the entry anyway, at least the offset is right. */
7466 osd_it_pack_dirent(lde, fid, it->oie_dirent->oied_off,
7467 it->oie_dirent->oied_name,
7468 it->oie_dirent->oied_namelen,
7469 it->oie_dirent->oied_type, attr);
7471 RETURN(rc > 0 ? 0 : rc);
7475 * Returns the record size size at current position.
7477 * This function will return record(lu_dirent) size in bytes.
7479 * \param[in] env execution environment
7480 * \param[in] di iterator's in memory structure
7481 * \param[in] attr attribute of the entry, only requires LUDA_TYPE to
7482 * calculate the lu_dirent size.
7484 * \retval record size(in bytes & in memory) of the current lu_dirent
7487 static int osd_it_ea_rec_size(const struct lu_env *env, const struct dt_it *di,
7490 struct osd_it_ea *it = (struct osd_it_ea *)di;
7492 return lu_dirent_calc_size(it->oie_dirent->oied_namelen, attr);
7496 * Returns a cookie for current position of the iterator head, so that
7497 * user can use this cookie to load/start the iterator next time.
7499 * \param di iterator's in memory structure
7501 * \retval cookie for current position, on success
7503 static __u64 osd_it_ea_store(const struct lu_env *env, const struct dt_it *di)
7505 struct osd_it_ea *it = (struct osd_it_ea *)di;
7507 return it->oie_dirent->oied_off;
7511 * It calls osd_ldiskfs_it_fill() which will use ->iterate*()
7512 * to load a directory entry at a time and stored it i inn,
7513 * in iterator's in-memory data structure.
7515 * \param di struct osd_it_ea, iterator's in memory structure
7517 * \retval +ve on success
7518 * \retval -ve on error
7520 static int osd_it_ea_load(const struct lu_env *env,
7521 const struct dt_it *di, __u64 hash)
7523 struct osd_it_ea *it = (struct osd_it_ea *)di;
7527 it->oie_file.f_pos = hash;
7529 rc = osd_ldiskfs_it_fill(env, di);
7540 * Index lookup function for interoperability mode (b11826).
7542 * \param key, key i.e. file name to be searched
7544 * \retval +ve, on success
7545 * \retval -ve, on error
7547 static int osd_index_ea_lookup(const struct lu_env *env, struct dt_object *dt,
7548 struct dt_rec *rec, const struct dt_key *key)
7550 struct osd_object *obj = osd_dt_obj(dt);
7555 LASSERT(S_ISDIR(obj->oo_inode->i_mode));
7556 LINVRNT(osd_invariant(obj));
7558 rc = osd_ea_lookup_rec(env, obj, rec, key);
7565 * Index and Iterator operations for interoperability
7566 * mode (i.e. to run 2.0 mds on 1.8 disk) (b11826)
7568 static const struct dt_index_operations osd_index_ea_ops = {
7569 .dio_lookup = osd_index_ea_lookup,
7570 .dio_declare_insert = osd_index_declare_ea_insert,
7571 .dio_insert = osd_index_ea_insert,
7572 .dio_declare_delete = osd_index_declare_ea_delete,
7573 .dio_delete = osd_index_ea_delete,
7575 .init = osd_it_ea_init,
7576 .fini = osd_it_ea_fini,
7577 .get = osd_it_ea_get,
7578 .put = osd_it_ea_put,
7579 .next = osd_it_ea_next,
7580 .key = osd_it_ea_key,
7581 .key_size = osd_it_ea_key_size,
7582 .rec = osd_it_ea_rec,
7583 .rec_size = osd_it_ea_rec_size,
7584 .store = osd_it_ea_store,
7585 .load = osd_it_ea_load
7589 static void *osd_key_init(const struct lu_context *ctx,
7590 struct lu_context_key *key)
7592 struct osd_thread_info *info;
7594 OBD_ALLOC_PTR(info);
7596 return ERR_PTR(-ENOMEM);
7598 OBD_ALLOC(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7599 if (info->oti_it_ea_buf == NULL)
7602 info->oti_env = container_of(ctx, struct lu_env, le_ctx);
7604 info->oti_hlock = ldiskfs_htree_lock_alloc();
7605 if (info->oti_hlock == NULL)
7611 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7614 return ERR_PTR(-ENOMEM);
7617 static void osd_key_fini(const struct lu_context *ctx,
7618 struct lu_context_key *key, void *data)
7620 struct osd_thread_info *info = data;
7621 struct ldiskfs_inode_info *lli = LDISKFS_I(info->oti_inode);
7622 struct osd_idmap_cache *idc = info->oti_ins_cache;
7624 if (info->oti_dio_pages) {
7626 for (i = 0; i < PTLRPC_MAX_BRW_PAGES; i++) {
7627 struct page *page = info->oti_dio_pages[i];
7629 LASSERT(PagePrivate2(page));
7630 LASSERT(PageLocked(page));
7631 ClearPagePrivate2(page);
7636 OBD_FREE_PTR_ARRAY_LARGE(info->oti_dio_pages,
7637 PTLRPC_MAX_BRW_PAGES);
7640 if (info->oti_inode != NULL)
7642 if (info->oti_hlock != NULL)
7643 ldiskfs_htree_lock_free(info->oti_hlock);
7644 OBD_FREE(info->oti_it_ea_buf, OSD_IT_EA_BUFSIZE);
7645 lu_buf_free(&info->oti_iobuf.dr_pg_buf);
7646 lu_buf_free(&info->oti_iobuf.dr_bl_buf);
7647 lu_buf_free(&info->oti_iobuf.dr_lnb_buf);
7648 lu_buf_free(&info->oti_big_buf);
7650 LASSERT(info->oti_ins_cache_size > 0);
7651 OBD_FREE_PTR_ARRAY(idc, info->oti_ins_cache_size);
7652 info->oti_ins_cache = NULL;
7653 info->oti_ins_cache_size = 0;
7658 static void osd_key_exit(const struct lu_context *ctx,
7659 struct lu_context_key *key, void *data)
7661 struct osd_thread_info *info = data;
7663 LASSERT(info->oti_r_locks == 0);
7664 LASSERT(info->oti_w_locks == 0);
7665 LASSERT(info->oti_txns == 0);
7668 /* type constructor/destructor: osd_type_init, osd_type_fini */
7669 LU_TYPE_INIT_FINI(osd, &osd_key);
7671 struct lu_context_key osd_key = {
7672 .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
7673 .lct_init = osd_key_init,
7674 .lct_fini = osd_key_fini,
7675 .lct_exit = osd_key_exit
7679 static int osd_device_init(const struct lu_env *env, struct lu_device *d,
7680 const char *name, struct lu_device *next)
7682 struct osd_device *osd = osd_dev(d);
7684 if (strlcpy(osd->od_svname, name, sizeof(osd->od_svname)) >=
7685 sizeof(osd->od_svname))
7687 return osd_procfs_init(osd, name);
7690 static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
7692 struct seq_server_site *ss = osd_seq_site(osd);
7697 if (osd->od_is_ost || osd->od_cl_seq != NULL)
7700 if (unlikely(ss == NULL))
7703 OBD_ALLOC_PTR(osd->od_cl_seq);
7704 if (osd->od_cl_seq == NULL)
7707 seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
7708 osd->od_svname, ss->ss_server_seq);
7710 if (ss->ss_node_id == 0) {
7712 * If the OSD on the sequence controller(MDT0), then allocate
7713 * sequence here, otherwise allocate sequence after connected
7714 * to MDT0 (see mdt_register_lwp_callback()).
7716 rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv,
7717 &osd->od_cl_seq->lcs_space, env);
7723 static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
7725 if (osd->od_cl_seq == NULL)
7728 seq_client_fini(osd->od_cl_seq);
7729 OBD_FREE_PTR(osd->od_cl_seq);
7730 osd->od_cl_seq = NULL;
7733 static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
7737 /* shutdown quota slave instance associated with the device */
7738 if (o->od_quota_slave_md != NULL) {
7739 struct qsd_instance *qsd = o->od_quota_slave_md;
7741 o->od_quota_slave_md = NULL;
7745 if (o->od_quota_slave_dt != NULL) {
7746 struct qsd_instance *qsd = o->od_quota_slave_dt;
7748 o->od_quota_slave_dt = NULL;
7752 osd_fid_fini(env, o);
7753 osd_scrub_cleanup(env, o);
7758 static void osd_umount(const struct lu_env *env, struct osd_device *o)
7762 if (o->od_mnt != NULL) {
7763 shrink_dcache_sb(osd_sb(o));
7764 osd_sync(env, &o->od_dt_dev);
7765 wait_event(o->od_commit_cb_done,
7766 !atomic_read(&o->od_commit_cb_in_flight));
7775 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7776 # ifndef LDISKFS_HAS_INCOMPAT_FEATURE
7777 /* Newer kernels provide the ldiskfs_set_feature_largedir() wrapper already,
7778 * which calls ldiskfs_update_dynamic_rev() to update ancient filesystems.
7779 * All ldiskfs filesystems are already v2, so it is a no-op and unnecessary.
7780 * This avoids maintaining patches to export this otherwise-useless function.
7782 void ldiskfs_update_dynamic_rev(struct super_block *sb)
7789 static int osd_mount(const struct lu_env *env,
7790 struct osd_device *o, struct lustre_cfg *cfg)
7792 const char *name = lustre_cfg_string(cfg, 0);
7793 const char *dev = lustre_cfg_string(cfg, 1);
7795 unsigned long page, s_flags = 0, lmd_flags = 0;
7796 struct page *__page;
7797 struct file_system_type *type;
7798 char *options = NULL;
7800 struct osd_thread_info *info = osd_oti_get(env);
7801 struct lu_fid *fid = &info->oti_fid;
7802 struct inode *inode;
7803 int rc = 0, force_over_1024tb = 0;
7807 if (o->od_mnt != NULL)
7810 if (strlen(dev) >= sizeof(o->od_mntdev))
7812 strcpy(o->od_mntdev, dev);
7814 str = lustre_cfg_buf(cfg, 2);
7815 sscanf(str, "%lu:%lu", &s_flags, &lmd_flags);
7817 opts = lustre_cfg_string(cfg, 3);
7819 if (opts == NULL || strstr(opts, "bigendian_extents") == NULL) {
7820 CERROR("%s: device %s extents feature is not guaranteed to "
7821 "work on big-endian systems. Use \"bigendian_extents\" "
7822 "mount option to override.\n", name, dev);
7826 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
7827 if (opts != NULL && strstr(opts, "force_over_128tb") != NULL) {
7828 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");
7831 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 1, 53, 0)
7832 if (opts != NULL && strstr(opts, "force_over_256tb") != NULL) {
7833 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");
7836 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7837 if (opts != NULL && strstr(opts, "force_over_512tb") != NULL) {
7838 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");
7842 if (opts != NULL && strstr(opts, "force_over_1024tb") != NULL)
7843 force_over_1024tb = 1;
7845 __page = alloc_page(GFP_KERNEL);
7847 GOTO(out, rc = -ENOMEM);
7848 page = (unsigned long)page_address(__page);
7849 options = (char *)page;
7852 /* strip out the options for back compatiblity */
7853 static const char * const sout[] = {
7860 /* strip out option we processed in osd */
7861 "bigendian_extents",
7865 "force_over_1024tb",
7869 strncat(options, opts, PAGE_SIZE);
7870 for (rc = 0, str = options; sout[rc]; ) {
7871 char *op = strstr(str, sout[rc]);
7878 if (op == options || *(op - 1) == ',') {
7879 str = op + strlen(sout[rc]);
7880 if (*str == ',' || *str == '\0') {
7881 *str == ',' ? str++ : str;
7882 memmove(op, str, strlen(str) + 1);
7885 for (str = op; *str != ',' && *str != '\0'; str++)
7889 strncat(options, "user_xattr,acl", PAGE_SIZE);
7892 /* Glom up mount options */
7893 if (*options != '\0')
7894 strncat(options, ",", PAGE_SIZE);
7895 strncat(options, "no_mbcache,nodelalloc", PAGE_SIZE);
7897 type = get_fs_type("ldiskfs");
7899 CERROR("%s: cannot find ldiskfs module\n", name);
7900 GOTO(out, rc = -ENODEV);
7903 o->od_mnt = vfs_kern_mount(type, s_flags, dev, options);
7904 module_put(type->owner);
7906 if (IS_ERR(o->od_mnt)) {
7907 rc = PTR_ERR(o->od_mnt);
7909 CERROR("%s: can't mount %s: %d\n", name, dev, rc);
7913 if (ldiskfs_blocks_count(LDISKFS_SB(osd_sb(o))->s_es) <<
7914 osd_sb(o)->s_blocksize_bits > 1024ULL << 40 &&
7915 force_over_1024tb == 0) {
7916 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",
7918 GOTO(out_mnt, rc = -EINVAL);
7921 if (lmd_flags & LMD_FLG_DEV_RDONLY) {
7922 LCONSOLE_WARN("%s: not support dev_rdonly on this device\n",
7925 GOTO(out_mnt, rc = -EOPNOTSUPP);
7928 if (!ldiskfs_has_feature_journal(o->od_mnt->mnt_sb)) {
7929 CERROR("%s: device %s is mounted w/o journal\n", name, dev);
7930 GOTO(out_mnt, rc = -EINVAL);
7933 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 53, 0)
7934 #ifdef LDISKFS_MOUNT_DIRDATA
7935 if (ldiskfs_has_feature_dirdata(o->od_mnt->mnt_sb))
7936 LDISKFS_SB(osd_sb(o))->s_mount_opt |= LDISKFS_MOUNT_DIRDATA;
7937 else if (strstr(name, "MDT")) /* don't complain for MGT or OSTs */
7938 CWARN("%s: device %s was upgraded from Lustre-1.x without "
7939 "enabling the dirdata feature. If you do not want to "
7940 "downgrade to Lustre-1.x again, you can enable it via "
7941 "'tune2fs -O dirdata device'\n", name, dev);
7943 /* enable large_dir on MDTs to avoid REMOTE_PARENT_DIR overflow,
7944 * and on very large OSTs to avoid object directory overflow */
7945 if (unlikely(!ldiskfs_has_feature_largedir(o->od_mnt->mnt_sb) &&
7946 !strstr(name, "MGS"))) {
7947 ldiskfs_set_feature_largedir(o->od_mnt->mnt_sb);
7948 LCONSOLE_INFO("%s: enable 'large_dir' feature on device '%s'\n",
7952 inode = osd_sb(o)->s_root->d_inode;
7953 lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
7954 rc = osd_ea_fid_set(info, inode, fid, LMAC_NOT_IN_OI, 0);
7956 CERROR("%s: failed to set lma on %s root inode\n", name, dev);
7960 if (lmd_flags & LMD_FLG_NOSCRUB)
7961 o->od_auto_scrub_interval = AS_NEVER;
7963 if (blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev))) {
7964 /* do not use pagecache with flash-backed storage */
7965 o->od_writethrough_cache = 0;
7966 o->od_read_cache = 0;
7977 __free_page(__page);
7982 static struct lu_device *osd_device_fini(const struct lu_env *env,
7983 struct lu_device *d)
7985 struct osd_device *o = osd_dev(d);
7989 osd_index_backup(env, o, false);
7990 osd_shutdown(env, o);
7992 if (o->od_oi_table != NULL)
7993 osd_oi_fini(osd_oti_get(env), o);
7994 if (o->od_extent_bytes_percpu)
7995 free_percpu(o->od_extent_bytes_percpu);
7996 osd_obj_map_fini(o);
8002 static int osd_device_init0(const struct lu_env *env,
8003 struct osd_device *o,
8004 struct lustre_cfg *cfg)
8006 struct lu_device *l = osd2lu_dev(o);
8007 struct osd_thread_info *info;
8010 bool restored = false;
8013 /* if the module was re-loaded, env can loose its keys */
8014 rc = lu_env_refill((struct lu_env *)env);
8017 info = osd_oti_get(env);
8020 l->ld_ops = &osd_lu_ops;
8021 o->od_dt_dev.dd_ops = &osd_dt_ops;
8023 spin_lock_init(&o->od_osfs_lock);
8024 mutex_init(&o->od_otable_mutex);
8025 INIT_LIST_HEAD(&o->od_orphan_list);
8026 INIT_LIST_HEAD(&o->od_index_backup_list);
8027 INIT_LIST_HEAD(&o->od_index_restore_list);
8028 spin_lock_init(&o->od_lock);
8029 o->od_index_backup_policy = LIBP_NONE;
8031 init_waitqueue_head(&o->od_commit_cb_done);
8033 o->od_read_cache = 1;
8034 o->od_writethrough_cache = 1;
8035 o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE;
8036 o->od_readcache_max_iosize = OSD_READCACHE_MAX_IO_MB << 20;
8037 o->od_writethrough_max_iosize = OSD_WRITECACHE_MAX_IO_MB << 20;
8038 o->od_auto_scrub_interval = AS_DEFAULT;
8039 /* default fallocate to unwritten extents: LU-14326/LU-14333 */
8040 o->od_fallocate_zero_blocks = 0;
8042 cplen = strlcpy(o->od_svname, lustre_cfg_string(cfg, 4),
8043 sizeof(o->od_svname));
8044 if (cplen >= sizeof(o->od_svname)) {
8049 o->od_index_backup_stop = 0;
8050 o->od_index = -1; /* -1 means index is invalid */
8051 rc = server_name2index(o->od_svname, &o->od_index, NULL);
8052 if (rc == LDD_F_SV_TYPE_OST)
8055 o->od_full_scrub_ratio = OFSR_DEFAULT;
8056 o->od_full_scrub_threshold_rate = FULL_SCRUB_THRESHOLD_RATE_DEFAULT;
8057 rc = osd_mount(env, o, cfg);
8061 /* Can only check block device after mount */
8062 o->od_nonrotational =
8063 blk_queue_nonrot(bdev_get_queue(osd_sb(o)->s_bdev));
8065 rc = osd_obj_map_init(env, o);
8069 rc = lu_site_init(&o->od_site, l);
8071 GOTO(out_compat, rc);
8072 o->od_site.ls_bottom_dev = l;
8074 rc = lu_site_init_finish(&o->od_site);
8078 opts = lustre_cfg_string(cfg, 3);
8079 if (opts && strstr(opts, "resetoi"))
8082 INIT_LIST_HEAD(&o->od_ios_list);
8083 /* setup scrub, including OI files initialization */
8085 rc = osd_scrub_setup(env, o, restored);
8090 rc = osd_procfs_init(o, o->od_svname);
8092 CERROR("%s: can't initialize procfs: rc = %d\n",
8094 GOTO(out_scrub, rc);
8097 LASSERT(l->ld_site->ls_linkage.next != NULL);
8098 LASSERT(l->ld_site->ls_linkage.prev != NULL);
8100 /* initialize quota slave instance */
8101 /* currently it's no need to prepare qsd_instance_md for OST */
8102 if (!o->od_is_ost) {
8103 o->od_quota_slave_md = qsd_init(env, o->od_svname,
8105 o->od_proc_entry, true);
8106 if (IS_ERR(o->od_quota_slave_md)) {
8107 rc = PTR_ERR(o->od_quota_slave_md);
8108 o->od_quota_slave_md = NULL;
8109 GOTO(out_procfs, rc);
8113 o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev,
8114 o->od_proc_entry, false);
8116 if (IS_ERR(o->od_quota_slave_dt)) {
8117 if (o->od_quota_slave_md != NULL) {
8118 qsd_fini(env, o->od_quota_slave_md);
8119 o->od_quota_slave_md = NULL;
8122 rc = PTR_ERR(o->od_quota_slave_dt);
8123 o->od_quota_slave_dt = NULL;
8124 GOTO(out_procfs, rc);
8127 o->od_extent_bytes_percpu = alloc_percpu(unsigned int);
8128 if (!o->od_extent_bytes_percpu) {
8130 GOTO(out_procfs, rc);
8138 osd_scrub_cleanup(env, o);
8140 lu_site_fini(&o->od_site);
8142 osd_obj_map_fini(o);
8149 static struct lu_device *osd_device_alloc(const struct lu_env *env,
8150 struct lu_device_type *t,
8151 struct lustre_cfg *cfg)
8153 struct osd_device *o;
8158 return ERR_PTR(-ENOMEM);
8160 rc = dt_device_init(&o->od_dt_dev, t);
8163 * Because the ctx might be revived in dt_device_init,
8164 * refill the env here
8166 lu_env_refill((struct lu_env *)env);
8167 rc = osd_device_init0(env, o, cfg);
8169 dt_device_fini(&o->od_dt_dev);
8172 if (unlikely(rc != 0))
8175 return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc);
8178 static struct lu_device *osd_device_free(const struct lu_env *env,
8179 struct lu_device *d)
8181 struct osd_device *o = osd_dev(d);
8185 /* XXX: make osd top device in order to release reference */
8186 d->ld_site->ls_top_dev = d;
8187 lu_site_purge(env, d->ld_site, -1);
8188 lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems,
8189 D_ERROR, lu_cdebug_printer);
8190 lu_site_fini(&o->od_site);
8191 dt_device_fini(&o->od_dt_dev);
8196 static int osd_process_config(const struct lu_env *env,
8197 struct lu_device *d, struct lustre_cfg *cfg)
8199 struct osd_device *o = osd_dev(d);
8205 switch (cfg->lcfg_command) {
8207 rc = osd_mount(env, o, cfg);
8211 * For the case LCFG_PRE_CLEANUP is not called in advance,
8212 * that may happend if hit failure during mount process.
8214 osd_index_backup(env, o, false);
8215 lu_dev_del_linkage(d->ld_site, d);
8216 rc = osd_shutdown(env, o);
8219 LASSERT(&o->od_dt_dev);
8220 count = class_modify_config(cfg, PARAM_OSD,
8221 &o->od_dt_dev.dd_kobj);
8223 count = class_modify_config(cfg, PARAM_OST,
8224 &o->od_dt_dev.dd_kobj);
8225 rc = count > 0 ? 0 : count;
8227 case LCFG_PRE_CLEANUP:
8229 osd_index_backup(env, o,
8230 o->od_index_backup_policy != LIBP_NONE);
8240 static int osd_recovery_complete(const struct lu_env *env,
8241 struct lu_device *d)
8243 struct osd_device *osd = osd_dev(d);
8248 if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == NULL)
8252 * start qsd instance on recovery completion, this notifies the quota
8253 * slave code that we are about to process new requests now
8255 rc = qsd_start(env, osd->od_quota_slave_dt);
8256 if (rc == 0 && osd->od_quota_slave_md != NULL)
8257 rc = qsd_start(env, osd->od_quota_slave_md);
8263 * we use exports to track all osd users
8265 static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
8266 struct obd_device *obd, struct obd_uuid *cluuid,
8267 struct obd_connect_data *data, void *localdata)
8269 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8270 struct lustre_handle conn;
8275 CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);
8277 rc = class_connect(&conn, obd, cluuid);
8281 *exp = class_conn2export(&conn);
8283 spin_lock(&osd->od_osfs_lock);
8285 spin_unlock(&osd->od_osfs_lock);
8291 * once last export (we don't count self-export) disappeared
8292 * osd can be released
8294 static int osd_obd_disconnect(struct obd_export *exp)
8296 struct obd_device *obd = exp->exp_obd;
8297 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8298 int rc, release = 0;
8302 /* Only disconnect the underlying layers on the final disconnect. */
8303 spin_lock(&osd->od_osfs_lock);
8305 if (osd->od_connects == 0)
8307 spin_unlock(&osd->od_osfs_lock);
8309 rc = class_disconnect(exp); /* bz 9811 */
8311 if (rc == 0 && release)
8312 class_manual_cleanup(obd);
8316 static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
8317 struct lu_device *dev)
8319 struct osd_device *osd = osd_dev(dev);
8320 struct lr_server_data *lsd =
8321 &osd->od_dt_dev.dd_lu_dev.ld_site->ls_tgt->lut_lsd;
8326 if (osd->od_quota_slave_md != NULL) {
8327 /* set up quota slave objects for inode */
8328 result = qsd_prepare(env, osd->od_quota_slave_md);
8333 if (osd->od_quota_slave_dt != NULL) {
8334 /* set up quota slave objects for block */
8335 result = qsd_prepare(env, osd->od_quota_slave_dt);
8341 if (lsd->lsd_feature_incompat & OBD_COMPAT_OST) {
8342 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
8343 if (lsd->lsd_feature_rocompat & OBD_ROCOMPAT_IDX_IN_IDIF) {
8344 osd->od_index_in_idif = 1;
8346 osd->od_index_in_idif = 0;
8347 result = osd_register_proc_index_in_idif(osd);
8352 osd->od_index_in_idif = 1;
8356 result = osd_fid_init(env, osd);
8362 * Implementation of lu_device_operations::ldo_fid_alloc() for OSD
8366 * see include/lu_object.h for the details.
8368 static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d,
8369 struct lu_fid *fid, struct lu_object *parent,
8370 const struct lu_name *name)
8372 struct osd_device *osd = osd_dev(d);
8374 return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
8377 static const struct lu_object_operations osd_lu_obj_ops = {
8378 .loo_object_init = osd_object_init,
8379 .loo_object_delete = osd_object_delete,
8380 .loo_object_release = osd_object_release,
8381 .loo_object_free = osd_object_free,
8382 .loo_object_print = osd_object_print,
8383 .loo_object_invariant = osd_object_invariant
8386 const struct lu_device_operations osd_lu_ops = {
8387 .ldo_object_alloc = osd_object_alloc,
8388 .ldo_process_config = osd_process_config,
8389 .ldo_recovery_complete = osd_recovery_complete,
8390 .ldo_prepare = osd_prepare,
8391 .ldo_fid_alloc = osd_fid_alloc,
8394 static const struct lu_device_type_operations osd_device_type_ops = {
8395 .ldto_init = osd_type_init,
8396 .ldto_fini = osd_type_fini,
8398 .ldto_start = osd_type_start,
8399 .ldto_stop = osd_type_stop,
8401 .ldto_device_alloc = osd_device_alloc,
8402 .ldto_device_free = osd_device_free,
8404 .ldto_device_init = osd_device_init,
8405 .ldto_device_fini = osd_device_fini
8408 static struct lu_device_type osd_device_type = {
8409 .ldt_tags = LU_DEVICE_DT,
8410 .ldt_name = LUSTRE_OSD_LDISKFS_NAME,
8411 .ldt_ops = &osd_device_type_ops,
8412 .ldt_ctx_tags = LCT_LOCAL,
8415 static int osd_health_check(const struct lu_env *env, struct obd_device *obd)
8417 struct osd_device *osd = osd_dev(obd->obd_lu_dev);
8418 struct super_block *sb = osd_sb(osd);
8420 return (osd->od_mnt == NULL || sb->s_flags & SB_RDONLY);
8424 * lprocfs legacy support.
8426 static const struct obd_ops osd_obd_device_ops = {
8427 .o_owner = THIS_MODULE,
8428 .o_connect = osd_obd_connect,
8429 .o_disconnect = osd_obd_disconnect,
8430 .o_health_check = osd_health_check,
8433 static ssize_t track_declares_assert_show(struct kobject *kobj,
8434 struct attribute *attr,
8437 return sprintf(buf, "%d\n", ldiskfs_track_declares_assert);
8440 static ssize_t track_declares_assert_store(struct kobject *kobj,
8441 struct attribute *attr,
8442 const char *buffer, size_t count)
8444 bool track_declares_assert;
8447 rc = kstrtobool(buffer, &track_declares_assert);
8451 ldiskfs_track_declares_assert = track_declares_assert;
8455 LUSTRE_RW_ATTR(track_declares_assert);
8457 static int __init osd_init(void)
8459 struct kobject *kobj;
8462 BUILD_BUG_ON(BH_DXLock >=
8463 sizeof(((struct buffer_head *)0)->b_state) * 8);
8464 #if !defined(CONFIG_DEBUG_MUTEXES) && !defined(CONFIG_DEBUG_SPINLOCK)
8465 /* please, try to keep osd_thread_info smaller than a page */
8466 BUILD_BUG_ON(sizeof(struct osd_thread_info) > PAGE_SIZE);
8471 rc = lu_kmem_init(ldiskfs_caches);
8475 #ifdef CONFIG_KALLSYMS
8476 priv_security_file_alloc =
8477 (void *)cfs_kallsyms_lookup_name("security_file_alloc");
8480 rc = class_register_type(&osd_obd_device_ops, NULL, true,
8481 LUSTRE_OSD_LDISKFS_NAME, &osd_device_type);
8483 lu_kmem_fini(ldiskfs_caches);
8487 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8489 rc = sysfs_create_file(kobj,
8490 &lustre_attr_track_declares_assert.attr);
8493 CWARN("osd-ldiskfs: track_declares_assert failed to register with sysfs\n");
8500 static void __exit osd_exit(void)
8502 struct kobject *kobj;
8504 kobj = kset_find_obj(lustre_kset, LUSTRE_OSD_LDISKFS_NAME);
8506 sysfs_remove_file(kobj,
8507 &lustre_attr_track_declares_assert.attr);
8510 class_unregister_type(LUSTRE_OSD_LDISKFS_NAME);
8511 lu_kmem_fini(ldiskfs_caches);
8514 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
8515 MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_LDISKFS_NAME")");
8516 MODULE_VERSION(LUSTRE_VERSION_STRING);
8517 MODULE_LICENSE("GPL");
8519 module_init(osd_init);
8520 module_exit(osd_exit);