LU-14927 scrub: create shared scrub_needs_check() function.

[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_handler.c

/*
 * GPL HEADER START
 *
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License version 2 only,
 * as published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License version 2 for more details (a copy is included
 * in the LICENSE file that accompanied this code).
 *
 * You should have received a copy of the GNU General Public License
 * version 2 along with this program; If not, see
 * http://www.gnu.org/licenses/gpl-2.0.html
 *
 * GPL HEADER END
 */
/*
 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
 * Use is subject to license terms.
 *
 * Copyright (c) 2011, 2017, Intel Corporation.
 */
/*
 * This file is part of Lustre, http://www.lustre.org/
 *
 * lustre/osd/osd_handler.c
 *
 * Top-level entry points into osd module
 *
 * Author: Nikita Danilov <nikita@clusterfs.com>
 *         Pravin Shelar <pravin.shelar@sun.com> : Added fid in dirent
 */

#define DEBUG_SUBSYSTEM S_OSD

#include <linux/fs_struct.h>
#include <linux/kallsyms.h>
#include <linux/module.h>
#include <linux/user_namespace.h>
#include <linux/uidgid.h>

/* prerequisite for linux/xattr.h */
#include <linux/types.h>
/* prerequisite for linux/xattr.h */
#include <linux/fs.h>
/* XATTR_{REPLACE,CREATE} */
#include <linux/xattr.h>

#include <ldiskfs/ldiskfs.h>
#include <ldiskfs/xattr.h>
#include <ldiskfs/ldiskfs_extents.h>
#undef ENTRY
/*
 * struct OBD_{ALLOC,FREE}*()
 * OBD_FAIL_CHECK
 */
#include <obd_support.h>
/* struct ptlrpc_thread */
#include <lustre_net.h>
#include <lustre_fid.h>
/* process_config */
#include <uapi/linux/lustre/lustre_param.h>

#include "osd_internal.h"
#include "osd_dynlocks.h"

/* llo_* api support */
#include <md_object.h>
#include <lustre_quota.h>

#include <lustre_linkea.h>

/* encoding routines */
#include <lustre_crypto.h>

/* Maximum EA size is limited by LNET_MTU for remote objects */
#define OSD_MAX_EA_SIZE 1048364

int ldiskfs_pdo = 1;
module_param(ldiskfs_pdo, int, 0644);
MODULE_PARM_DESC(ldiskfs_pdo, "ldiskfs with parallel directory operations");

int ldiskfs_track_declares_assert;
module_param(ldiskfs_track_declares_assert, int, 0644);
MODULE_PARM_DESC(ldiskfs_track_declares_assert, "LBUG during tracking of declares");

/* Slab to allocate dynlocks */
struct kmem_cache *dynlock_cachep;

/* Slab to allocate osd_it_ea */
struct kmem_cache *osd_itea_cachep;

static struct lu_kmem_descr ldiskfs_caches[] = {
        {
                .ckd_cache = &dynlock_cachep,
                .ckd_name  = "dynlock_cache",
                .ckd_size  = sizeof(struct dynlock_handle)
        },
        {
                .ckd_cache = &osd_itea_cachep,
                .ckd_name  = "osd_itea_cache",
                .ckd_size  = sizeof(struct osd_it_ea)
        },
        {
                .ckd_cache = NULL
        }
};

static const char dot[] = ".";
static const char dotdot[] = "..";

static const struct lu_object_operations      osd_lu_obj_ops;
static const struct dt_object_operations      osd_obj_ops;
static const struct dt_object_operations      osd_obj_otable_it_ops;
static const struct dt_index_operations       osd_index_iam_ops;
static const struct dt_index_operations       osd_index_ea_ops;

static int osd_remote_fid(const struct lu_env *env, struct osd_device *osd,
                          const struct lu_fid *fid);
static int osd_process_scheduled_agent_removals(const struct lu_env *env,
                                                struct osd_device *osd);

int osd_trans_declare_op2rb[] = {
        [OSD_OT_ATTR_SET]       = OSD_OT_ATTR_SET,
        [OSD_OT_PUNCH]          = OSD_OT_MAX,
        [OSD_OT_XATTR_SET]      = OSD_OT_XATTR_SET,
        [OSD_OT_CREATE]         = OSD_OT_DESTROY,
        [OSD_OT_DESTROY]        = OSD_OT_CREATE,
        [OSD_OT_REF_ADD]        = OSD_OT_REF_DEL,
        [OSD_OT_REF_DEL]        = OSD_OT_REF_ADD,
        [OSD_OT_WRITE]          = OSD_OT_WRITE,
        [OSD_OT_INSERT]         = OSD_OT_DELETE,
        [OSD_OT_DELETE]         = OSD_OT_INSERT,
        [OSD_OT_QUOTA]          = OSD_OT_MAX,
};

static int osd_has_index(const struct osd_object *obj)
{
        return obj->oo_dt.do_index_ops != NULL;
}

static int osd_object_invariant(const struct lu_object *l)
{
        return osd_invariant(osd_obj(l));
}

/*
 * Concurrency: doesn't matter
 */
static int osd_is_write_locked(const struct lu_env *env, struct osd_object *o)
{
        struct osd_thread_info *oti = osd_oti_get(env);

        return oti->oti_w_locks > 0 && o->oo_owner == env;
}

/*
 * Concurrency: doesn't access mutable data
 */
static int osd_root_get(const struct lu_env *env,
                        struct dt_device *dev, struct lu_fid *f)
{
        lu_local_obj_fid(f, OSD_FS_ROOT_OID);
        return 0;
}

/*
 * the following set of functions are used to maintain per-thread
 * cache of FID->ino mapping. this mechanism is needed to resolve
 * FID to inode at dt_insert() which in turn stores ino in the
 * directory entries to keep ldiskfs compatible with ext[34].
 * due to locking-originated restrictions we can't lookup ino
 * using LU cache (deadlock is possible). lookup using OI is quite
 * expensive. so instead we maintain this cache and methods like
 * dt_create() fill it. so in the majority of cases dt_insert() is
 * able to find needed mapping in lockless manner.
 */
static struct osd_idmap_cache *
osd_idc_find(const struct lu_env *env, struct osd_device *osd,
             const struct lu_fid *fid)
{
        struct osd_thread_info *oti = osd_oti_get(env);
        struct osd_idmap_cache *idc = oti->oti_ins_cache;
        int i;

        for (i = 0; i < oti->oti_ins_cache_used; i++) {
                if (!lu_fid_eq(&idc[i].oic_fid, fid))
                        continue;
                if (idc[i].oic_dev != osd)
                        continue;

                return idc + i;
        }

        return NULL;
}

static struct osd_idmap_cache *
osd_idc_add(const struct lu_env *env, struct osd_device *osd,
            const struct lu_fid *fid)
{
        struct osd_thread_info *oti   = osd_oti_get(env);
        struct osd_idmap_cache *idc;
        int i;

        if (unlikely(oti->oti_ins_cache_used >= oti->oti_ins_cache_size)) {
                i = oti->oti_ins_cache_size * 2;
                if (i == 0)
                        i = OSD_INS_CACHE_SIZE;
                OBD_ALLOC_PTR_ARRAY(idc, i);
                if (idc == NULL)
                        return ERR_PTR(-ENOMEM);
                if (oti->oti_ins_cache != NULL) {
                        memcpy(idc, oti->oti_ins_cache,
                               oti->oti_ins_cache_used * sizeof(*idc));
                        OBD_FREE_PTR_ARRAY(oti->oti_ins_cache,
                                           oti->oti_ins_cache_used);
                }
                oti->oti_ins_cache = idc;
                oti->oti_ins_cache_size = i;
        }

        idc = oti->oti_ins_cache + oti->oti_ins_cache_used++;
        idc->oic_fid = *fid;
        idc->oic_dev = osd;
        idc->oic_lid.oii_ino = 0;
        idc->oic_lid.oii_gen = 0;
        idc->oic_remote = 0;

        return idc;
}

/*
 * lookup mapping for the given fid in the cache, initialize a
 * new one if not found. the initialization checks whether the
 * object is local or remote. for local objects, OI is used to
 * learn ino/generation. the function is used when the caller
 * has no information about the object, e.g. at dt_insert().
 */
static struct osd_idmap_cache *
osd_idc_find_or_init(const struct lu_env *env, struct osd_device *osd,
                     const struct lu_fid *fid)
{
        struct osd_idmap_cache *idc;
        int rc;

        idc = osd_idc_find(env, osd, fid);
        LASSERT(!IS_ERR(idc));
        if (idc != NULL)
                return idc;

        CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
               osd->od_svname, PFID(fid));

        /* new mapping is needed */
        idc = osd_idc_add(env, osd, fid);
        if (IS_ERR(idc)) {
                CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
                       osd->od_svname, PFID(fid), PTR_ERR(idc));
                return idc;
        }

        /* initialize it */
        rc = osd_remote_fid(env, osd, fid);
        if (unlikely(rc < 0))
                return ERR_PTR(rc);

        if (rc == 0) {
                /* the object is local, lookup in OI */
                /* XXX: probably cheaper to lookup in LU first? */
                rc = osd_oi_lookup(osd_oti_get(env), osd, fid,
                                   &idc->oic_lid, 0);
                if (unlikely(rc < 0)) {
                        CERROR("can't lookup: rc = %d\n", rc);
                        return ERR_PTR(rc);
                }
        } else {
                /* the object is remote */
                idc->oic_remote = 1;
        }

        return idc;
}

static void osd_idc_dump_lma(const struct lu_env *env,
                                struct osd_device *osd,
                                unsigned long ino,
                                bool check_in_oi)
{
        struct osd_thread_info *info = osd_oti_get(env);
        struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
        const struct lu_fid *fid;
        struct osd_inode_id lid;
        struct inode *inode;
        int rc;

        inode = osd_ldiskfs_iget(osd_sb(osd), ino);
        if (IS_ERR(inode)) {
                CERROR("%s: can't get inode %lu: rc = %d\n",
                       osd->od_svname, ino, (int)PTR_ERR(inode));
                return;
        }
        if (is_bad_inode(inode)) {
                CERROR("%s: bad inode %lu\n", osd->od_svname, ino);
                goto put;
        }
        rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
        if (rc) {
                CERROR("%s: can't get LMA for %lu: rc = %d\n",
                       osd->od_svname, ino, rc);
                goto put;
        }
        fid = &loa->loa_lma.lma_self_fid;
        LCONSOLE(D_INFO, "%s: "DFID" in inode %lu/%u\n", osd->od_svname,
                      PFID(fid), ino, (unsigned)inode->i_generation);
        if (!check_in_oi)
                goto put;
        rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
        if (rc) {
                CERROR("%s: can't lookup "DFID": rc = %d\n",
                       osd->od_svname, PFID(fid), rc);
                goto put;
        }
        LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n", osd->od_svname,
                      PFID(fid), lid.oii_ino, lid.oii_gen);
put:
        iput(inode);
}

static void osd_idc_dump_debug(const struct lu_env *env,
                                struct osd_device *osd,
                                const struct lu_fid *fid,
                                unsigned long ino1,
                                unsigned long ino2)
{
        struct osd_inode_id lid;

        int rc;

        rc = osd_oi_lookup(osd_oti_get(env), osd, fid, &lid, 0);
        if (!rc) {
                LCONSOLE(D_INFO, "%s: "DFID" maps to %u/%u\n",
                        osd->od_svname, PFID(fid), lid.oii_ino, lid.oii_gen);
                osd_idc_dump_lma(env, osd, lid.oii_ino, false);
        } else {
                CERROR("%s: can't lookup "DFID": rc = %d\n",
                       osd->od_svname, PFID(fid), rc);
        }
        if (ino1)
                osd_idc_dump_lma(env, osd, ino1, true);
        if (ino2)
                osd_idc_dump_lma(env, osd, ino2, true);
}

/*
 * lookup mapping for given FID and fill it from the given object.
 * the object is lolcal by definition.
 */
static int osd_idc_find_and_init(const struct lu_env *env,
                                 struct osd_device *osd,
                                 struct osd_object *obj)
{
        const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
        struct osd_idmap_cache *idc;

        idc = osd_idc_find(env, osd, fid);
        LASSERT(!IS_ERR(idc));
        if (idc != NULL) {
                if (obj->oo_inode == NULL)
                        return 0;
                if (idc->oic_lid.oii_ino != obj->oo_inode->i_ino) {
                        if (idc->oic_lid.oii_ino) {
                                osd_idc_dump_debug(env, osd, fid,
                                                   idc->oic_lid.oii_ino,
                                                   obj->oo_inode->i_ino);
                                return -EINVAL;
                        }
                        idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
                        idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
                }
                return 0;
        }

        CDEBUG(D_INODE, "%s: FID "DFID" not in the id map cache\n",
               osd->od_svname, PFID(fid));

        /* new mapping is needed */
        idc = osd_idc_add(env, osd, fid);
        if (IS_ERR(idc)) {
                CERROR("%s: FID "DFID" add id map cache failed: %ld\n",
                       osd->od_svname, PFID(fid), PTR_ERR(idc));
                return PTR_ERR(idc);
        }

        if (obj->oo_inode != NULL) {
                idc->oic_lid.oii_ino = obj->oo_inode->i_ino;
                idc->oic_lid.oii_gen = obj->oo_inode->i_generation;
        }
        return 0;
}

/*
 * OSD object methods.
 */

/*
 * Concurrency: no concurrent access is possible that early in object
 * life-cycle.
 */
static struct lu_object *osd_object_alloc(const struct lu_env *env,
                                          const struct lu_object_header *hdr,
                                          struct lu_device *d)
{
        struct osd_object *mo;

        OBD_ALLOC_PTR(mo);
        if (mo != NULL) {
                struct lu_object *l;
                struct lu_object_header *h;
                struct osd_device *o = osd_dev(d);

                l = &mo->oo_dt.do_lu;
                if (unlikely(o->od_in_init)) {
                        OBD_ALLOC_PTR(h);
                        if (!h) {
                                OBD_FREE_PTR(mo);
                                return NULL;
                        }

                        lu_object_header_init(h);
                        lu_object_init(l, h, d);
                        lu_object_add_top(h, l);
                        mo->oo_header = h;
                } else {
                        dt_object_init(&mo->oo_dt, NULL, d);
                        mo->oo_header = NULL;
                }

                mo->oo_dt.do_ops = &osd_obj_ops;
                l->lo_ops = &osd_lu_obj_ops;
                init_rwsem(&mo->oo_sem);
                init_rwsem(&mo->oo_ext_idx_sem);
                spin_lock_init(&mo->oo_guard);
                INIT_LIST_HEAD(&mo->oo_xattr_list);
                return l;
        }
        return NULL;
}

int osd_get_lma(struct osd_thread_info *info, struct inode *inode,
                struct dentry *dentry, struct lustre_ost_attrs *loa)
{
        int rc;

        rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
                             (void *)loa, sizeof(*loa));
        if (rc > 0) {
                struct lustre_mdt_attrs *lma = &loa->loa_lma;

                if (rc < sizeof(*lma))
                        return -EINVAL;

                rc = 0;
                lustre_loa_swab(loa, true);
                /* Check LMA compatibility */
                if (lma->lma_incompat & ~LMA_INCOMPAT_SUPP) {
                        rc = -EOPNOTSUPP;
                        CWARN("%s: unsupported incompat LMA feature(s) %#x for fid = "DFID", ino = %lu: rc = %d\n",
                              osd_ino2name(inode),
                              lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
                              PFID(&lma->lma_self_fid), inode->i_ino, rc);
                }
        } else if (rc == 0) {
                rc = -ENODATA;
        }

        return rc;
}

/*
 * retrieve object from backend ext fs.
 **/
struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
                       struct osd_inode_id *id)
{
        int rc;
        struct inode *inode = NULL;

        /*
         * if we look for an inode withing a running
         * transaction, then we risk to deadlock
         * osd_dirent_check_repair() breaks this
         */
         /* LASSERT(current->journal_info == NULL); */

        inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
        if (IS_ERR(inode)) {
                CDEBUG(D_INODE, "no inode: ino = %u, rc = %ld\n",
                       id->oii_ino, PTR_ERR(inode));
        } else if (id->oii_gen != OSD_OII_NOGEN &&
                   inode->i_generation != id->oii_gen) {
                CDEBUG(D_INODE, "unmatched inode: ino = %u, oii_gen = %u, "
                       "i_generation = %u\n",
                       id->oii_ino, id->oii_gen, inode->i_generation);
                iput(inode);
                inode = ERR_PTR(-ESTALE);
        } else if (inode->i_nlink == 0) {
                /*
                 * due to parallel readdir and unlink,
                 * we can have dead inode here.
                 */
                CDEBUG(D_INODE, "stale inode: ino = %u\n", id->oii_ino);
                iput(inode);
                inode = ERR_PTR(-ESTALE);
        } else if (is_bad_inode(inode)) {
                rc = -ENOENT;
                CWARN("%s: bad inode: ino = %u: rc = %d\n",
                      osd_dev2name(dev), id->oii_ino, rc);
                iput(inode);
                inode = ERR_PTR(rc);
        } else if ((rc = osd_attach_jinode(inode))) {
                iput(inode);
                inode = ERR_PTR(rc);
        } else {
                ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);
                if (id->oii_gen == OSD_OII_NOGEN)
                        osd_id_gen(id, inode->i_ino, inode->i_generation);

                /*
                 * Do not update file c/mtime in ldiskfs.
                 * NB: we don't have any lock to protect this because we don't
                 * have reference on osd_object now, but contention with
                 * another lookup + attr_set can't happen in the tiny window
                 * between if (...) and set S_NOCMTIME.
                 */
                if (!(inode->i_flags & S_NOCMTIME))
                        inode->i_flags |= S_NOCMTIME;
        }
        return inode;
}

int osd_ldiskfs_add_entry(struct osd_thread_info *info, struct osd_device *osd,
                          handle_t *handle, struct dentry *child,
                          struct inode *inode, struct htree_lock *hlock)
{
        int rc, rc2;

        rc = __ldiskfs_add_entry(handle, child, inode, hlock);
        if (rc == -ENOBUFS || rc == -ENOSPC) {
                struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
                struct inode *parent = child->d_parent->d_inode;
                struct lu_fid *fid = NULL;
                char fidstr[FID_LEN + 1] = "unknown";

                rc2 = osd_get_lma(info, parent, child->d_parent, loa);
                if (!rc2) {
                        fid = &loa->loa_lma.lma_self_fid;
                } else if (rc2 == -ENODATA) {
                        if (unlikely(is_root_inode(parent))) {
                                fid = &info->oti_fid3;
                                lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
                        } else if (!osd->od_is_ost && osd->od_index == 0) {
                                fid = &info->oti_fid3;
                                lu_igif_build(fid, parent->i_ino,
                                              parent->i_generation);
                        }
                }

                if (fid != NULL)
                        snprintf(fidstr, sizeof(fidstr), DFID, PFID(fid));

                /* below message is checked in sanity.sh test_129 */
                if (rc == -ENOSPC) {
                        CWARN("%s: directory (inode: %lu, FID: %s) has reached max size limit\n",
                              osd_name(osd), parent->i_ino, fidstr);
                } else {
                        rc = 0; /* ignore such error now */
                        CWARN("%s: directory (inode: %lu, FID: %s) is approaching max size limit\n",
                              osd_name(osd), parent->i_ino, fidstr);
                }

        }

        return rc;
}


struct inode *
osd_iget_fid(struct osd_thread_info *info, struct osd_device *dev,
             struct osd_inode_id *id, struct lu_fid *fid)
{
        struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
        struct inode *inode;
        int rc;

        inode = osd_iget(info, dev, id);
        if (IS_ERR(inode))
                return inode;

        rc = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
        if (!rc) {
                *fid = loa->loa_lma.lma_self_fid;
        } else if (rc == -ENODATA) {
                if (unlikely(is_root_inode(inode)))
                        lu_local_obj_fid(fid, OSD_FS_ROOT_OID);
                else
                        lu_igif_build(fid, inode->i_ino, inode->i_generation);
        } else {
                iput(inode);
                inode = ERR_PTR(rc);
        }
        return inode;
}

static struct inode *osd_iget_check(struct osd_thread_info *info,
                                    struct osd_device *dev,
                                    const struct lu_fid *fid,
                                    struct osd_inode_id *id,
                                    bool trusted)
{
        struct inode *inode;
        int rc = 0;

        ENTRY;

        /*
         * The cached OI mapping is trustable. If we cannot locate the inode
         * via the cached OI mapping, then return the failure to the caller
         * directly without further OI checking.
         */

again:
        inode = osd_ldiskfs_iget(osd_sb(dev), id->oii_ino);
        if (IS_ERR(inode)) {
                rc = PTR_ERR(inode);
                if (!trusted && (rc == -ENOENT || rc == -ESTALE))
                        goto check_oi;

                CDEBUG(D_INODE, "no inode for FID: "DFID", ino = %u, rc = %d\n",
                       PFID(fid), id->oii_ino, rc);
                GOTO(put, rc);
        }

        if (is_bad_inode(inode)) {
                rc = -ENOENT;
                if (!trusted)
                        goto check_oi;

                CDEBUG(D_INODE, "bad inode for FID: "DFID", ino = %u\n",
                       PFID(fid), id->oii_ino);
                GOTO(put, rc);
        }

        if (id->oii_gen != OSD_OII_NOGEN &&
            inode->i_generation != id->oii_gen) {
                rc = -ESTALE;
                if (!trusted)
                        goto check_oi;

                CDEBUG(D_INODE, "unmatched inode for FID: "DFID", ino = %u, "
                       "oii_gen = %u, i_generation = %u\n", PFID(fid),
                       id->oii_ino, id->oii_gen, inode->i_generation);
                GOTO(put, rc);
        }

        if (inode->i_nlink == 0) {
                rc = -ENOENT;
                if (!trusted)
                        goto check_oi;

                CDEBUG(D_INODE, "stale inode for FID: "DFID", ino = %u\n",
                       PFID(fid), id->oii_ino);
                GOTO(put, rc);
        }

        ldiskfs_clear_inode_state(inode, LDISKFS_STATE_LUSTRE_DESTROY);

check_oi:
        if (rc != 0) {
                __u32 saved_ino = id->oii_ino;
                __u32 saved_gen = id->oii_gen;

                LASSERT(!trusted);
                LASSERTF(rc == -ESTALE || rc == -ENOENT, "rc = %d\n", rc);

                rc = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
                /*
                 * XXX: There are four possible cases:
                 *      1. rc = 0.
                 *         Backup/restore caused the OI invalid.
                 *      2. rc = 0.
                 *         Someone unlinked the object but NOT removed
                 *         the OI mapping, such as mount target device
                 *         as ldiskfs, and modify something directly.
                 *      3. rc = -ENOENT.
                 *         Someone just removed the object between the
                 *         former oi_lookup and the iget. It is normal.
                 *      4. Other failure cases.
                 *
                 *      Generally, when the device is mounted, it will
                 *      auto check whether the system is restored from
                 *      file-level backup or not. We trust such detect
                 *      to distinguish the 1st case from the 2nd case:
                 *      if the OI files are consistent but may contain
                 *      stale OI mappings because of case 2, if iget()
                 *      returns -ENOENT or -ESTALE, then it should be
                 *      the case 2.
                 */
                if (rc != 0)
                        /*
                         * If the OI mapping was in OI file before the
                         * osd_iget_check(), but now, it is disappear,
                         * then it must be removed by race. That is a
                         * normal race case.
                         */
                        GOTO(put, rc);

                /*
                 * It is the OI scrub updated the OI mapping by race.
                 * The new OI mapping must be valid.
                 */
                if (saved_ino != id->oii_ino ||
                    (saved_gen != id->oii_gen && saved_gen != OSD_OII_NOGEN)) {
                        if (!IS_ERR(inode))
                                iput(inode);

                        trusted = true;
                        goto again;
                }

                if (IS_ERR(inode)) {
                        if (dev->od_scrub.os_scrub.os_file.sf_flags &
                            SF_INCONSISTENT)
                                /*
                                 * It still can be the case 2, but we cannot
                                 * distinguish it from the case 1. So return
                                 * -EREMCHG to block current operation until
                                 *  OI scrub rebuilt the OI mappings.
                                 */
                                rc = -EREMCHG;
                        else
                                rc = -ENOENT;

                        GOTO(put, rc);
                }

                if (inode->i_generation == id->oii_gen)
                        rc = -ENOENT;
                else
                        rc = -EREMCHG;
        } else {
                if (id->oii_gen == OSD_OII_NOGEN)
                        osd_id_gen(id, inode->i_ino, inode->i_generation);

                /*
                 * Do not update file c/mtime in ldiskfs.
                 * NB: we don't have any lock to protect this because we don't
                 * have reference on osd_object now, but contention with
                 * another lookup + attr_set can't happen in the tiny window
                 * between if (...) and set S_NOCMTIME.
                 */
                if (!(inode->i_flags & S_NOCMTIME))
                        inode->i_flags |= S_NOCMTIME;
        }

        GOTO(put, rc);

put:
        if (rc != 0) {
                if (!IS_ERR(inode))
                        iput(inode);

                inode = ERR_PTR(rc);
        }

        return inode;
}

/**
 * \retval +v: new filter_fid does not contain self-fid
 * \retval 0:  filter_fid_18_23, contains self-fid
 * \retval -v: other failure cases
 */
int osd_get_idif(struct osd_thread_info *info, struct inode *inode,
                 struct dentry *dentry, struct lu_fid *fid)
{
        struct filter_fid *ff = &info->oti_ff;
        struct ost_id *ostid = &info->oti_ostid;
        int rc;

        rc = __osd_xattr_get(inode, dentry, XATTR_NAME_FID, ff, sizeof(*ff));
        if (rc == sizeof(struct filter_fid_18_23)) {
                struct filter_fid_18_23 *ff_old = (void *)ff;

                ostid_set_seq(ostid, le64_to_cpu(ff_old->ff_seq));
                rc = ostid_set_id(ostid, le64_to_cpu(ff_old->ff_objid));
                /*
                 * XXX: use 0 as the index for compatibility, the caller will
                 * handle index related issues when necessary.
                 */
                if (!rc)
                        ostid_to_fid(fid, ostid, 0);
        } else if (rc >= (int)sizeof(struct filter_fid_24_29)) {
                rc = 1;
        } else if (rc >= 0) {
                rc = -EINVAL;
        }

        return rc;
}

static int osd_lma_self_repair(struct osd_thread_info *info,
                               struct osd_device *osd, struct inode *inode,
                               const struct lu_fid *fid, __u32 compat)
{
        handle_t *jh;
        int rc;

        LASSERT(current->journal_info == NULL);

        jh = osd_journal_start_sb(osd_sb(osd), LDISKFS_HT_MISC,
                                  osd_dto_credits_noquota[DTO_XATTR_SET]);
        if (IS_ERR(jh)) {
                rc = PTR_ERR(jh);
                CWARN("%s: cannot start journal for lma_self_repair: rc = %d\n",
                      osd_name(osd), rc);
                return rc;
        }

        rc = osd_ea_fid_set(info, inode, fid, compat, 0);
        if (rc != 0)
                CWARN("%s: cannot self repair the LMA: rc = %d\n",
                      osd_name(osd), rc);
        ldiskfs_journal_stop(jh);
        return rc;
}

static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
{
        struct osd_thread_info *info = osd_oti_get(env);
        struct osd_device *osd = osd_obj2dev(obj);
        struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
        struct lustre_mdt_attrs *lma = &loa->loa_lma;
        struct inode *inode = obj->oo_inode;
        struct dentry *dentry = &info->oti_obj_dentry;
        struct lu_fid *fid = NULL;
        const struct lu_fid *rfid = lu_object_fid(&obj->oo_dt.do_lu);
        int rc;

        ENTRY;

        rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMA,
                             (void *)loa, sizeof(*loa));
        if (rc == -ENODATA && !fid_is_igif(rfid) && osd->od_check_ff) {
                fid = &lma->lma_self_fid;
                rc = osd_get_idif(info, inode, dentry, fid);
                if (rc > 0 || (rc == -ENODATA && osd->od_index_in_idif)) {
                        /*
                         * For the given OST-object, if it has neither LMA nor
                         * FID in XATTR_NAME_FID, then the given FID (which is
                         * contained in the @obj, from client RPC for locating
                         * the OST-object) is trusted. We use it to generate
                         * the LMA.
                         */
                        osd_lma_self_repair(info, osd, inode, rfid,
                                            LMAC_FID_ON_OST);
                        RETURN(0);
                }
        }

        if (rc < 0)
                RETURN(rc);

        if (rc > 0) {
                rc = 0;
                lustre_lma_swab(lma);
                if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
                             (CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT) &&
                              S_ISREG(inode->i_mode)))) {
                        CWARN("%s: unsupported incompat LMA feature(s) %#x for "
                              "fid = "DFID", ino = %lu\n", osd_name(osd),
                              lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
                              PFID(rfid), inode->i_ino);
                        rc = -EOPNOTSUPP;
                } else {
                        fid = &lma->lma_self_fid;
                        if (lma->lma_compat & LMAC_STRIPE_INFO &&
                            osd->od_is_ost)
                                obj->oo_pfid_in_lma = 1;
                        if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
                            !osd->od_is_ost)
                                lu_object_set_agent_entry(&obj->oo_dt.do_lu);
                }
        }

        if (fid != NULL && unlikely(!lu_fid_eq(rfid, fid))) {
                if (fid_is_idif(rfid) && fid_is_idif(fid)) {
                        struct ost_id   *oi   = &info->oti_ostid;
                        struct lu_fid   *fid1 = &info->oti_fid3;
                        __u32            idx  = fid_idif_ost_idx(rfid);

                        /*
                         * For old IDIF, the OST index is not part of the IDIF,
                         * Means that different OSTs may have the same IDIFs.
                         * Under such case, we need to make some compatible
                         * check to make sure to trigger OI scrub properly.
                         */
                        if (idx != 0 && fid_idif_ost_idx(fid) == 0) {
                                /* Given @rfid is new, LMA is old. */
                                fid_to_ostid(fid, oi);
                                ostid_to_fid(fid1, oi, idx);
                                if (lu_fid_eq(fid1, rfid)) {
                                        if (osd->od_index_in_idif)
                                                osd_lma_self_repair(info, osd,
                                                        inode, rfid,
                                                        LMAC_FID_ON_OST);
                                        RETURN(0);
                                }
                        }
                }

                rc = -EREMCHG;
        }

        RETURN(rc);
}

struct osd_check_lmv_buf {
        /* please keep it as first member */
        struct dir_context ctx;
        struct osd_thread_info *oclb_info;
        struct osd_device *oclb_dev;
        int oclb_items;
        bool oclb_found;
};

/**
 * It is called internally by ->iterate*() to filter out the
 * local slave object's FID of the striped directory.
 *
 * \retval      1 found the local slave's FID
 * \retval      0 continue to check next item
 * \retval      -ve for failure
 */
#ifdef HAVE_FILLDIR_USE_CTX
static int osd_stripe_dir_filldir(struct dir_context *buf,
#else
static int osd_stripe_dir_filldir(void *buf,
#endif
                                  const char *name, int namelen,
                                  loff_t offset, __u64 ino, unsigned int d_type)
{
        struct osd_check_lmv_buf *oclb = (struct osd_check_lmv_buf *)buf;
        struct osd_thread_info *oti = oclb->oclb_info;
        struct lu_fid *fid = &oti->oti_fid3;
        struct osd_inode_id *id = &oti->oti_id3;
        struct osd_device *dev = oclb->oclb_dev;
        struct inode *inode;

        oclb->oclb_items++;

        if (name[0] == '.')
                return 0;

        fid_zero(fid);
        sscanf(name + 1, SFID, RFID(fid));
        if (!fid_is_sane(fid))
                return 0;

        if (osd_remote_fid(oti->oti_env, dev, fid))
                return 0;

        osd_id_gen(id, ino, OSD_OII_NOGEN);
        inode = osd_iget(oti, dev, id);
        if (IS_ERR(inode))
                return PTR_ERR(inode);

        iput(inode);
        osd_add_oi_cache(oti, dev, id, fid);
        osd_oii_insert(dev, fid, id, true);
        oclb->oclb_found = true;

        return 1;
}

/*
 * When lookup item under striped directory, we need to locate the master
 * MDT-object of the striped directory firstly, then the client will send
 * lookup (getattr_by_name) RPC to the MDT with some slave MDT-object's FID
 * and the item's name. If the system is restored from MDT file level backup,
 * then before the OI scrub completely built the OI files, the OI mappings of
 * the master MDT-object and slave MDT-object may be invalid. Usually, it is
 * not a problem for the master MDT-object. Because when locate the master
 * MDT-object, we will do name based lookup (for the striped directory itself)
 * firstly, during such process we can setup the correct OI mapping for the
 * master MDT-object. But it will be trouble for the slave MDT-object. Because
 * the client will not trigger name based lookup on the MDT to locate the slave
 * MDT-object before locating item under the striped directory, then when
 * osd_fid_lookup(), it will find that the OI mapping for the slave MDT-object
 * is invalid and does not know what the right OI mapping is, then the MDT has
 * to return -EINPROGRESS to the client to notify that the OI scrub is rebuiding
 * the OI file, related OI mapping is unknown yet, please try again later. And
 * then client will re-try the RPC again and again until related OI mapping has
 * been updated. That is quite inefficient.
 *
 * To resolve above trouble, we will handle it as the following two cases:
 *
 * 1) The slave MDT-object and the master MDT-object are on different MDTs.
 *    It is relative easy. Be as one of remote MDT-objects, the slave MDT-object
 *    is linked under /REMOTE_PARENT_DIR with the name of its FID string.
 *    We can locate the slave MDT-object via lookup the /REMOTE_PARENT_DIR
 *    directly. Please check osd_fid_lookup().
 *
 * 2) The slave MDT-object and the master MDT-object reside on the same MDT.
 *    Under such case, during lookup the master MDT-object, we will lookup the
 *    slave MDT-object via readdir against the master MDT-object, because the
 *    slave MDT-objects information are stored as sub-directories with the name
 *    "${FID}:${index}". Then when find the local slave MDT-object, its OI
 *    mapping will be recorded. Then subsequent osd_fid_lookup() will know
 *    the correct OI mapping for the slave MDT-object.
 */
static int osd_check_lmv(struct osd_thread_info *oti, struct osd_device *dev,
                         struct inode *inode)
{
        struct lu_buf *buf = &oti->oti_big_buf;
        struct dentry *dentry = &oti->oti_obj_dentry;
        struct file *filp;
        struct lmv_mds_md_v1 *lmv1;
        struct osd_check_lmv_buf oclb = {
                .ctx.actor = osd_stripe_dir_filldir,
                .oclb_info = oti,
                .oclb_dev = dev,
                .oclb_found = false,
        };
        int rc = 0;

        ENTRY;

again:
        rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, buf->lb_buf,
                             buf->lb_len);
        if (rc == -ERANGE) {
                rc = __osd_xattr_get(inode, dentry, XATTR_NAME_LMV, NULL, 0);
                if (rc > 0) {
                        lu_buf_realloc(buf, rc);
                        if (buf->lb_buf == NULL)
                                GOTO(out, rc = -ENOMEM);

                        goto again;
                }
        }

        if (unlikely(rc == 0 || rc == -ENODATA))
                GOTO(out, rc = 0);

        if (rc < 0)
                GOTO(out, rc);

        if (unlikely(buf->lb_buf == NULL)) {
                lu_buf_realloc(buf, rc);
                if (buf->lb_buf == NULL)
                        GOTO(out, rc = -ENOMEM);

                goto again;
        }

        lmv1 = buf->lb_buf;
        if (le32_to_cpu(lmv1->lmv_magic) != LMV_MAGIC_V1)
                GOTO(out, rc = 0);

        filp = osd_quasi_file(oti->oti_env, inode);
        rc = osd_security_file_alloc(filp);
        if (rc)
                goto out;

        do {
                oclb.oclb_items = 0;
                rc = iterate_dir(filp, &oclb.ctx);
        } while (rc >= 0 && oclb.oclb_items > 0 && !oclb.oclb_found &&
                 filp->f_pos != LDISKFS_HTREE_EOF_64BIT);
        inode->i_fop->release(inode, filp);

out:
        if (rc < 0)
                CDEBUG(D_LFSCK,
                       "%s: cannot check LMV, ino = %lu/%u: rc = %d\n",
                       osd_ino2name(inode), inode->i_ino, inode->i_generation,
                       rc);
        else
                rc = 0;

        RETURN(rc);
}

static int osd_fid_lookup(const struct lu_env *env, struct osd_object *obj,
                          const struct lu_fid *fid,
                          const struct lu_object_conf *conf)
{
        struct osd_thread_info *info;
        struct lu_device *ldev = obj->oo_dt.do_lu.lo_dev;
        struct osd_device *dev;
        struct osd_idmap_cache *oic;
        struct osd_inode_id *id;
        struct inode *inode = NULL;
        struct lustre_scrub *scrub;
        struct scrub_file *sf;
        __u32 flags = SS_CLEAR_DRYRUN | SS_CLEAR_FAILOUT | SS_AUTO_FULL;
        __u32 saved_ino;
        __u32 saved_gen;
        int result = 0;
        int rc1 = 0;
        bool remote = false;
        bool trusted = true;
        bool updated = false;
        bool checked = false;

        ENTRY;

        LINVRNT(osd_invariant(obj));
        LASSERT(obj->oo_inode == NULL);

        if (fid_is_sane(fid) == 0) {
                CERROR("%s: invalid FID "DFID"\n", ldev->ld_obd->obd_name,
                       PFID(fid));
                dump_stack();
                RETURN(-EINVAL);
        }

        dev = osd_dev(ldev);
        scrub = &dev->od_scrub.os_scrub;
        sf = &scrub->os_file;
        info = osd_oti_get(env);
        LASSERT(info);
        oic = &info->oti_cache;

        if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT))
                RETURN(-ENOENT);

        /*
         * For the object is created as locking anchor, or for the object to
         * be created on disk. No need to osd_oi_lookup() at here because FID
         * shouldn't never be re-used, if it's really a duplicate FID from
         * unexpected reason, we should be able to detect it later by calling
         * do_create->osd_oi_insert().
         */
        if (conf && conf->loc_flags & LOC_F_NEW)
                GOTO(out, result = 0);

        /* Search order: 1. per-thread cache. */
        if (lu_fid_eq(fid, &oic->oic_fid) && likely(oic->oic_dev == dev)) {
                id = &oic->oic_lid;
                goto iget;
        }

        id = &info->oti_id;
        if (!list_empty(&scrub->os_inconsistent_items)) {
                /* Search order: 2. OI scrub pending list. */
                result = osd_oii_lookup(dev, fid, id);
                if (!result)
                        goto iget;
        }

        /*
         * The OI mapping in the OI file can be updated by the OI scrub
         * when we locate the inode via FID. So it may be not trustable.
         */
        trusted = false;

        /* Search order: 3. OI files. */
        result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
        if (result == -ENOENT) {
                if (!(fid_is_norm(fid) || fid_is_igif(fid)) ||
                    fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) ||
                    !ldiskfs_test_bit(osd_oi_fid2idx(dev, fid),
                                      sf->sf_oi_bitmap))
                        GOTO(out, result = 0);

                goto trigger;
        }

        /* -ESTALE is returned if inode of OST object doesn't exist */
        if (result == -ESTALE &&
            fid_is_on_ost(info, dev, fid, OI_CHECK_FLD)) {
                GOTO(out, result = 0);
        }

        if (result)
                GOTO(out, result);

iget:
        obj->oo_inode = NULL;
        /* for later passes through checks, not true on first pass */
        if (!IS_ERR_OR_NULL(inode))
                iput(inode);

        inode = osd_iget_check(info, dev, fid, id, trusted);
        if (!IS_ERR(inode)) {
                obj->oo_inode = inode;
                result = 0;
                if (remote)
                        goto trigger;

                goto check_lma;
        }

        result = PTR_ERR(inode);
        if (result == -ENOENT || result == -ESTALE)
                GOTO(out, result = 0);

        if (result != -EREMCHG)
                GOTO(out, result);

trigger:
        /*
         * We still have chance to get the valid inode: for the
         * object which is referenced by remote name entry, the
         * object on the local MDT will be linked under the dir
         * of "/REMOTE_PARENT_DIR" with its FID string as name.
         *
         * We do not know whether the object for the given FID
         * is referenced by some remote name entry or not, and
         * especially for DNE II, a multiple-linked object may
         * have many name entries reside on many MDTs.
         *
         * To simplify the operation, OSD will not distinguish
         * more, just lookup "/REMOTE_PARENT_DIR". Usually, it
         * only happened for the RPC from other MDT during the
         * OI scrub, or for the client side RPC with FID only,
         * such as FID to path, or from old connected client.
         */
        if (!remote) {
                rc1 = osd_lookup_in_remote_parent(info, dev, fid, id);
                if (!rc1) {
                        remote = true;
                        trusted = true;
                        flags |= SS_AUTO_PARTIAL;
                        flags &= ~SS_AUTO_FULL;
                        goto iget;
                }
        }

        if (scrub->os_running) {
                if (scrub->os_partial_scan && !scrub->os_in_join)
                        goto join;

                if (IS_ERR_OR_NULL(inode) || result) {
                        osd_oii_insert(dev, fid, id, result == -ENOENT);
                        GOTO(out, result = -EINPROGRESS);
                }

                LASSERT(remote);
                LASSERT(obj->oo_inode == inode);

                osd_oii_insert(dev, fid, id, true);
                goto found;
        }

        if (dev->od_scrub.os_scrub.os_auto_scrub_interval == AS_NEVER) {
                if (!remote)
                        GOTO(out, result = -EREMCHG);

                LASSERT(!result);
                LASSERT(obj->oo_inode == inode);

                osd_add_oi_cache(info, dev, id, fid);
                goto found;
        }

join:
        rc1 = osd_scrub_start(env, dev, flags);
        CDEBUG_LIMIT(D_LFSCK | D_CONSOLE | D_WARNING,
                     "%s: trigger OI scrub by RPC for "DFID"/%u with flags %#x: rc = %d\n",
                     osd_name(dev), PFID(fid), id->oii_ino, flags, rc1);
        if (rc1 && rc1 != -EALREADY)
                GOTO(out, result = -EREMCHG);

        if (IS_ERR_OR_NULL(inode) || result) {
                osd_oii_insert(dev, fid, id, result == -ENOENT);
                GOTO(out, result = -EINPROGRESS);
        }

        LASSERT(remote);
        LASSERT(obj->oo_inode == inode);

        osd_oii_insert(dev, fid, id, true);
        goto found;

check_lma:
        checked = true;
        if (unlikely(obj->oo_header))
                goto found;

        result = osd_check_lma(env, obj);
        if (!result)
                goto found;

        LASSERTF(id->oii_ino == inode->i_ino &&
                 id->oii_gen == inode->i_generation,
                 "locate wrong inode for FID: "DFID", %u/%u => %ld/%u\n",
                 PFID(fid), id->oii_ino, id->oii_gen,
                 inode->i_ino, inode->i_generation);

        saved_ino = inode->i_ino;
        saved_gen = inode->i_generation;

        if (unlikely(result == -ENODATA)) {
                /*
                 * If the OI scrub updated the OI mapping by race, it
                 * must be valid. Trust the inode that has no LMA EA.
                 */
                if (updated)
                        goto found;

                result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
                if (!result) {
                        /*
                         * The OI mapping is still there, the inode is still
                         * valid. It is just becaues the inode has no LMA EA.
                         */
                        if (saved_ino == id->oii_ino &&
                            saved_gen == id->oii_gen)
                                goto found;

                        /*
                         * It is the OI scrub updated the OI mapping by race.
                         * The new OI mapping must be valid.
                         */
                        trusted = true;
                        updated = true;
                        goto iget;
                }

                /*
                 * "result == -ENOENT" means that the OI mappinghas been
                 * removed by race, so the inode belongs to other object.
                 *
                 * Others error can be returned  directly.
                 */
                if (result == -ENOENT) {
                        LASSERT(trusted);

                        obj->oo_inode = NULL;
                        result = 0;
                }
        }

        if (result != -EREMCHG)
                GOTO(out, result);

        LASSERT(!updated);

        /*
         * if two OST objects map to the same inode, and inode mode is
         * (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666), which means it's
         * reserved by precreate, and not written yet, in this case, don't
         * set inode for the object whose FID mismatch, so that it can create
         * inode and not block precreate.
         */
        if (fid_is_on_ost(info, dev, fid, OI_CHECK_FLD) &&
            inode->i_mode == (S_IFREG | S_ISUID | S_ISGID | S_ISVTX | 0666)) {
                obj->oo_inode = NULL;
                GOTO(out, result = 0);
        }

        result = osd_oi_lookup(info, dev, fid, id, OI_CHECK_FLD);
        /*
         * "result == -ENOENT" means the cached OI mapping has been removed
         * from the OI file by race, above inode belongs to other object.
         */
        if (result == -ENOENT) {
                LASSERT(trusted);

                obj->oo_inode = NULL;
                GOTO(out, result = 0);
        }

        if (result)
                GOTO(out, result);

        if (saved_ino == id->oii_ino && saved_gen == id->oii_gen) {
                result = -EREMCHG;
                osd_scrub_refresh_mapping(info, dev, fid, id, DTO_INDEX_DELETE,
                                          true, 0, NULL);
                goto trigger;
        }

        /*
         * It is the OI scrub updated the OI mapping by race.
         * The new OI mapping must be valid.
         */
        trusted = true;
        updated = true;
        goto iget;

found:
        if (!checked) {
                struct lustre_ost_attrs *loa = &info->oti_ost_attrs;
                struct lustre_mdt_attrs *lma = &info->oti_ost_attrs.loa_lma;

                result = osd_get_lma(info, inode, &info->oti_obj_dentry, loa);
                if (!result) {
                        if (lma->lma_compat & LMAC_STRIPE_INFO &&
                            dev->od_is_ost)
                                obj->oo_pfid_in_lma = 1;
                        if (unlikely(lma->lma_incompat & LMAI_REMOTE_PARENT) &&
                            !dev->od_is_ost)
                                lu_object_set_agent_entry(&obj->oo_dt.do_lu);
                } else if (result != -ENODATA) {
                        GOTO(out, result);
                }
        }

        obj->oo_compat_dot_created = 1;
        obj->oo_compat_dotdot_created = 1;

        if (S_ISDIR(inode->i_mode) &&
            (flags & SS_AUTO_PARTIAL || sf->sf_status == SS_SCANNING))
                osd_check_lmv(info, dev, inode);

        result = osd_attach_jinode(inode);
        if (result)
                GOTO(out, result);

        if (!ldiskfs_pdo)
                GOTO(out, result = 0);

        LASSERT(!obj->oo_hl_head);
        obj->oo_hl_head = ldiskfs_htree_lock_head_alloc(HTREE_HBITS_DEF);

        GOTO(out, result = (!obj->oo_hl_head ? -ENOMEM : 0));

out:
        if (result || !obj->oo_inode) {
                if (!IS_ERR_OR_NULL(inode))
                        iput(inode);

                obj->oo_inode = NULL;
                if (trusted)
                        fid_zero(&oic->oic_fid);
        }

        LINVRNT(osd_invariant(obj));
        return result;
}

/*
 * Concurrency: shouldn't matter.
 */
static void osd_object_init0(struct osd_object *obj)
{
        LASSERT(obj->oo_inode != NULL);
        obj->oo_dt.do_body_ops = &osd_body_ops;
        obj->oo_dt.do_lu.lo_header->loh_attr |=
                (LOHA_EXISTS | (obj->oo_inode->i_mode & S_IFMT));
}

/*
 * Concurrency: no concurrent access is possible that early in object
 * life-cycle.
 */
static int osd_object_init(const struct lu_env *env, struct lu_object *l,
                           const struct lu_object_conf *conf)
{
        struct osd_object *obj = osd_obj(l);
        int result;

        LINVRNT(osd_invariant(obj));

        if (OBD_FAIL_PRECHECK(OBD_FAIL_MDS_LLOG_UMOUNT_RACE) &&
            cfs_fail_val == 2) {
                struct osd_thread_info *info = osd_oti_get(env);
                struct osd_idmap_cache *oic = &info->oti_cache;
                /* invalidate thread cache */
                memset(&oic->oic_fid, 0, sizeof(oic->oic_fid));
        }
        if (fid_is_otable_it(&l->lo_header->loh_fid)) {
                obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
                l->lo_header->loh_attr |= LOHA_EXISTS;
                return 0;
        }

        result = osd_fid_lookup(env, obj, lu_object_fid(l), conf);
        obj->oo_dt.do_body_ops = &osd_body_ops_new;
        if (result == 0 && obj->oo_inode != NULL) {
                struct osd_thread_info *oti = osd_oti_get(env);
                struct lustre_ost_attrs *loa = &oti->oti_ost_attrs;

                osd_object_init0(obj);
                if (unlikely(obj->oo_header))
                        return 0;

                result = osd_get_lma(oti, obj->oo_inode,
                                     &oti->oti_obj_dentry, loa);
                if (!result) {
                        /*
                         * Convert LMAI flags to lustre LMA flags
                         * and cache it to oo_lma_flags
                         */
                        obj->oo_lma_flags =
                                lma_to_lustre_flags(loa->loa_lma.lma_incompat);
                } else if (result == -ENODATA) {
                        result = 0;
                }
        }
        obj->oo_dirent_count = LU_DIRENT_COUNT_UNSET;

        LINVRNT(osd_invariant(obj));
        return result;
}

/*
 * The first part of oxe_buf is xattr name, and is '\0' terminated.
 * The left part is for value, binary mode.
 */
struct osd_xattr_entry {
        struct list_head        oxe_list;
        size_t                  oxe_len;
        size_t                  oxe_namelen;
        bool                    oxe_exist;
        struct rcu_head         oxe_rcu;
        char                    oxe_buf[0];
};

static int osd_oxc_get(struct osd_object *obj, const char *name,
                       struct lu_buf *buf)
{
        struct osd_xattr_entry *tmp;
        struct osd_xattr_entry *oxe = NULL;
        size_t namelen = strlen(name);
        int rc;

        rcu_read_lock();
        list_for_each_entry_rcu(tmp, &obj->oo_xattr_list, oxe_list) {
                if (namelen == tmp->oxe_namelen &&
                    strncmp(name, tmp->oxe_buf, namelen) == 0) {
                        oxe = tmp;
                        break;
                }
        }

        if (oxe == NULL)
                GOTO(out, rc = -ENOENT);

        if (!oxe->oxe_exist)
                GOTO(out, rc = -ENODATA);

        /* vallen */
        rc = oxe->oxe_len - sizeof(*oxe) - oxe->oxe_namelen - 1;
        LASSERT(rc > 0);

        if (buf->lb_buf == NULL)
                GOTO(out, rc);

        if (buf->lb_len < rc)
                GOTO(out, rc = -ERANGE);

        memcpy(buf->lb_buf, &oxe->oxe_buf[namelen + 1], rc);
out:
        rcu_read_unlock();

        return rc;
}

static void osd_oxc_free(struct rcu_head *head)
{
        struct osd_xattr_entry *oxe;

        oxe = container_of(head, struct osd_xattr_entry, oxe_rcu);
        OBD_FREE(oxe, oxe->oxe_len);
}

static void osd_oxc_add(struct osd_object *obj, const char *name,
                        const char *buf, int buflen)
{
        struct osd_xattr_entry *oxe;
        struct osd_xattr_entry *old = NULL;
        struct osd_xattr_entry *tmp;
        size_t namelen = strlen(name);
        size_t len = sizeof(*oxe) + namelen + 1 + buflen;

        OBD_ALLOC(oxe, len);
        if (oxe == NULL)
                return;

        INIT_LIST_HEAD(&oxe->oxe_list);
        oxe->oxe_len = len;
        oxe->oxe_namelen = namelen;
        memcpy(oxe->oxe_buf, name, namelen);
        if (buflen > 0) {
                LASSERT(buf != NULL);
                memcpy(oxe->oxe_buf + namelen + 1, buf, buflen);
                oxe->oxe_exist = true;
        } else {
                oxe->oxe_exist = false;
        }

        /* this should be rarely called, just remove old and add new */
        spin_lock(&obj->oo_guard);
        list_for_each_entry(tmp, &obj->oo_xattr_list, oxe_list) {
                if (namelen == tmp->oxe_namelen &&
                    strncmp(name, tmp->oxe_buf, namelen) == 0) {
                        old = tmp;
                        break;
                }
        }
        if (old != NULL) {
                list_replace_rcu(&old->oxe_list, &oxe->oxe_list);
                call_rcu(&old->oxe_rcu, osd_oxc_free);
        } else {
                list_add_tail_rcu(&oxe->oxe_list, &obj->oo_xattr_list);
        }
        spin_unlock(&obj->oo_guard);
}

static void osd_oxc_del(struct osd_object *obj, const char *name)
{
        struct osd_xattr_entry *oxe;
        size_t namelen = strlen(name);

        spin_lock(&obj->oo_guard);
        list_for_each_entry(oxe, &obj->oo_xattr_list, oxe_list) {
                if (namelen == oxe->oxe_namelen &&
                    strncmp(name, oxe->oxe_buf, namelen) == 0) {
                        list_del_rcu(&oxe->oxe_list);
                        call_rcu(&oxe->oxe_rcu, osd_oxc_free);
                        break;
                }
        }
        spin_unlock(&obj->oo_guard);
}

static void osd_oxc_fini(struct osd_object *obj)
{
        struct osd_xattr_entry *oxe, *next;

        list_for_each_entry_safe(oxe, next, &obj->oo_xattr_list, oxe_list) {
                list_del(&oxe->oxe_list);
                OBD_FREE(oxe, oxe->oxe_len);
        }
}

/*
 * Concurrency: no concurrent access is possible that late in object
 * life-cycle.
 */
static void osd_object_free(const struct lu_env *env, struct lu_object *l)
{
        struct osd_object *obj = osd_obj(l);
        struct lu_object_header *h = obj->oo_header;

        LINVRNT(osd_invariant(obj));

        osd_oxc_fini(obj);
        dt_object_fini(&obj->oo_dt);
        if (obj->oo_hl_head != NULL)
                ldiskfs_htree_lock_head_free(obj->oo_hl_head);
        /* obj doesn't contain an lu_object_header, so we don't need call_rcu */
        OBD_FREE_PTR(obj);
        if (unlikely(h))
                lu_object_header_free(h);
}

/*
 * Concurrency: no concurrent access is possible that late in object
 * life-cycle.
 */
static void osd_index_fini(struct osd_object *o)
{
        struct iam_container *bag;

        if (o->oo_dir != NULL) {
                bag = &o->oo_dir->od_container;
                if (o->oo_inode != NULL) {
                        if (bag->ic_object == o->oo_inode)
                                iam_container_fini(bag);
                }
                OBD_FREE_PTR(o->oo_dir);
                o->oo_dir = NULL;
        }
}

enum {
        OSD_TXN_OI_DELETE_CREDITS    = 20,
        OSD_TXN_INODE_DELETE_CREDITS = 20
};

/*
 * Journal
 */

#if OSD_THANDLE_STATS
/**
 * Set time when the handle is allocated
 */
static void osd_th_alloced(struct osd_thandle *oth)
{
        oth->oth_alloced = ktime_get();
}

/**
 * Set time when the handle started
 */
static void osd_th_started(struct osd_thandle *oth)
{
        oth->oth_started = ktime_get();
}

/**
 * Check whether the we deal with this handle for too long.
 */
static void __osd_th_check_slow(void *oth, struct osd_device *dev,
                                ktime_t alloced, ktime_t started,
                                ktime_t closed)
{
        ktime_t now = ktime_get();

        LASSERT(dev != NULL);

        lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_STARTING,
                            ktime_us_delta(started, alloced));
        lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_OPEN,
                            ktime_us_delta(closed, started));
        lprocfs_counter_add(dev->od_stats, LPROC_OSD_THANDLE_CLOSING,
                            ktime_us_delta(now, closed));

        if (ktime_before(ktime_add_ns(alloced, 30 * NSEC_PER_SEC), now)) {
                CWARN("transaction handle %p was open for too long: now %lld, alloced %lld, started %lld, closed %lld\n",
                                oth, now, alloced, started, closed);
                libcfs_debug_dumpstack(NULL);
        }
}

#define OSD_CHECK_SLOW_TH(oth, dev, expr)                               \
{                                                                       \
        ktime_t __closed = ktime_get();                                 \
        ktime_t __alloced = oth->oth_alloced;                           \
        ktime_t __started = oth->oth_started;                           \
                                                                        \
        expr;                                                           \
        __osd_th_check_slow(oth, dev, __alloced, __started, __closed);  \
}

#else /* OSD_THANDLE_STATS */

#define osd_th_alloced(h)                  do {} while(0)
#define osd_th_started(h)                  do {} while(0)
#define OSD_CHECK_SLOW_TH(oth, dev, expr)  expr

#endif /* OSD_THANDLE_STATS */

/*
 * Concurrency: doesn't access mutable data.
 */
static int osd_param_is_not_sane(const struct osd_device *dev,
                                 const struct thandle *th)
{
        struct osd_thandle *oh = container_of(th, typeof(*oh), ot_super);

        return oh->ot_credits > osd_transaction_size(dev);
}

/*
 * Concurrency: shouldn't matter.
 */
static void osd_trans_commit_cb(struct super_block *sb,
                                struct ldiskfs_journal_cb_entry *jcb, int error)
{
        struct osd_thandle *oh = container_of(jcb, struct osd_thandle, ot_jcb);
        struct thandle *th = &oh->ot_super;
        struct lu_device *lud = &th->th_dev->dd_lu_dev;
        struct osd_device *osd = osd_dev(lud);
        struct dt_txn_commit_cb *dcb, *tmp;

        LASSERT(oh->ot_handle == NULL);

        if (error)
                CERROR("transaction @0x%p commit error: %d\n", th, error);

        OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DELAY_TRANS, 40);
        /* call per-transaction callbacks if any */
        list_for_each_entry_safe(dcb, tmp, &oh->ot_commit_dcb_list,
                                 dcb_linkage) {
                LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
                         "commit callback entry: magic=%x name='%s'\n",
                         dcb->dcb_magic, dcb->dcb_name);
                list_del_init(&dcb->dcb_linkage);
                dcb->dcb_func(NULL, th, dcb, error);
        }

        lu_ref_del_at(&lud->ld_reference, &oh->ot_dev_link, "osd-tx", th);
        if (atomic_dec_and_test(&osd->od_commit_cb_in_flight))
                wake_up(&osd->od_commit_cb_done);
        th->th_dev = NULL;

        OBD_FREE_PTR(oh);
}

static struct thandle *osd_trans_create(const struct lu_env *env,
                                        struct dt_device *d)
{
        struct osd_thread_info *oti = osd_oti_get(env);
        struct osd_iobuf *iobuf = &oti->oti_iobuf;
        struct osd_thandle *oh;
        struct thandle *th;

        ENTRY;

        if (d->dd_rdonly) {
                CERROR("%s: someone try to start transaction under "
                       "readonly mode, should be disabled.\n",
                       osd_name(osd_dt_dev(d)));
                dump_stack();
                RETURN(ERR_PTR(-EROFS));
        }

        /* on pending IO in this thread should left from prev. request */
        LASSERT(atomic_read(&iobuf->dr_numreqs) == 0);

        sb_start_write(osd_sb(osd_dt_dev(d)));

        OBD_ALLOC_GFP(oh, sizeof(*oh), GFP_NOFS);
        if (!oh) {
                sb_end_write(osd_sb(osd_dt_dev(d)));
                RETURN(ERR_PTR(-ENOMEM));
        }

        oh->ot_quota_trans = &oti->oti_quota_trans;
        memset(oh->ot_quota_trans, 0, sizeof(*oh->ot_quota_trans));
        th = &oh->ot_super;
        th->th_dev = d;
        th->th_result = 0;
        oh->ot_credits = 0;
        oh->oh_declared_ext = 0;
        INIT_LIST_HEAD(&oh->ot_commit_dcb_list);
        INIT_LIST_HEAD(&oh->ot_stop_dcb_list);
        INIT_LIST_HEAD(&oh->ot_trunc_locks);
        osd_th_alloced(oh);

        memset(oti->oti_declare_ops, 0,
               sizeof(oti->oti_declare_ops));
        memset(oti->oti_declare_ops_cred, 0,
               sizeof(oti->oti_declare_ops_cred));
        memset(oti->oti_declare_ops_used, 0,
               sizeof(oti->oti_declare_ops_used));

        oti->oti_ins_cache_depth++;

        RETURN(th);
}

void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th)
{
        struct osd_thread_info *oti = osd_oti_get(env);
        struct osd_thandle *oh;

        oh = container_of(th, struct osd_thandle, ot_super);
        LASSERT(oh != NULL);

        CWARN("  create: %u/%u/%u, destroy: %u/%u/%u\n",
              oti->oti_declare_ops[OSD_OT_CREATE],
              oti->oti_declare_ops_cred[OSD_OT_CREATE],
              oti->oti_declare_ops_used[OSD_OT_CREATE],
              oti->oti_declare_ops[OSD_OT_DESTROY],
              oti->oti_declare_ops_cred[OSD_OT_DESTROY],
              oti->oti_declare_ops_used[OSD_OT_DESTROY]);
        CWARN("  attr_set: %u/%u/%u, xattr_set: %u/%u/%u\n",
              oti->oti_declare_ops[OSD_OT_ATTR_SET],
              oti->oti_declare_ops_cred[OSD_OT_ATTR_SET],
              oti->oti_declare_ops_used[OSD_OT_ATTR_SET],
              oti->oti_declare_ops[OSD_OT_XATTR_SET],
              oti->oti_declare_ops_cred[OSD_OT_XATTR_SET],
              oti->oti_declare_ops_used[OSD_OT_XATTR_SET]);
        CWARN("  write: %u/%u/%u, punch: %u/%u/%u, quota %u/%u/%u\n",
              oti->oti_declare_ops[OSD_OT_WRITE],
              oti->oti_declare_ops_cred[OSD_OT_WRITE],
              oti->oti_declare_ops_used[OSD_OT_WRITE],
              oti->oti_declare_ops[OSD_OT_PUNCH],
              oti->oti_declare_ops_cred[OSD_OT_PUNCH],
              oti->oti_declare_ops_used[OSD_OT_PUNCH],
              oti->oti_declare_ops[OSD_OT_QUOTA],
              oti->oti_declare_ops_cred[OSD_OT_QUOTA],
              oti->oti_declare_ops_used[OSD_OT_QUOTA]);
        CWARN("  insert: %u/%u/%u, delete: %u/%u/%u\n",
              oti->oti_declare_ops[OSD_OT_INSERT],
              oti->oti_declare_ops_cred[OSD_OT_INSERT],
              oti->oti_declare_ops_used[OSD_OT_INSERT],
              oti->oti_declare_ops[OSD_OT_DELETE],
              oti->oti_declare_ops_cred[OSD_OT_DELETE],
              oti->oti_declare_ops_used[OSD_OT_DELETE]);
        CWARN("  ref_add: %u/%u/%u, ref_del: %u/%u/%u\n",
              oti->oti_declare_ops[OSD_OT_REF_ADD],
              oti->oti_declare_ops_cred[OSD_OT_REF_ADD],
              oti->oti_declare_ops_used[OSD_OT_REF_ADD],
              oti->oti_declare_ops[OSD_OT_REF_DEL],
              oti->oti_declare_ops_cred[OSD_OT_REF_DEL],
              oti->oti_declare_ops_used[OSD_OT_REF_DEL]);
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
                           struct thandle *th)
{
        struct osd_thread_info *oti = osd_oti_get(env);
        struct osd_device *dev = osd_dt_dev(d);
        handle_t *jh;
        struct osd_thandle *oh;
        int rc;

        ENTRY;

        LASSERT(current->journal_info == NULL);

        oh = container_of(th, struct osd_thandle, ot_super);
        LASSERT(oh != NULL);
        LASSERT(oh->ot_handle == NULL);

        rc = dt_txn_hook_start(env, d, th);
        if (rc != 0)
                GOTO(out, rc);

        if (unlikely(osd_param_is_not_sane(dev, th))) {
                static unsigned long last_printed;
                static int last_credits;

                /*
                 * don't make noise on a tiny testing systems
                 * actual credits misuse will be caught anyway
                 */
                if (last_credits != oh->ot_credits &&
                    time_after(jiffies, last_printed +
                               cfs_time_seconds(60)) &&
                    osd_transaction_size(dev) > 512) {
                        CWARN("%s: credits %u > trans_max %u\n", osd_name(dev),
                              oh->ot_credits, osd_transaction_size(dev));
                        osd_trans_dump_creds(env, th);
                        libcfs_debug_dumpstack(NULL);
                        last_credits = oh->ot_credits;
                        last_printed = jiffies;
                }
                /*
                 * XXX Limit the credits to 'max_transaction_buffers', and
                 *     let the underlying filesystem to catch the error if
                 *     we really need so many credits.
                 *
                 *     This should be removed when we can calculate the
                 *     credits precisely.
                 */
                oh->ot_credits = osd_transaction_size(dev);
        } else if (ldiskfs_track_declares_assert != 0) {
                /*
                 * reserve few credits to prevent an assertion in JBD
                 * our debugging mechanism will be able to detected
                 * overuse. this can help to debug single-update
                 * transactions
                 */
                oh->ot_credits += 10;
                if (unlikely(osd_param_is_not_sane(dev, th)))
                        oh->ot_credits = osd_transaction_size(dev);
        }

        if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START))
                GOTO(out, rc = -EIO);

        /*
         * XXX temporary stuff. Some abstraction layer should
         * be used.
         */
        jh = osd_journal_start_sb(osd_sb(dev), LDISKFS_HT_MISC, oh->ot_credits);
        osd_th_started(oh);
        if (!IS_ERR(jh)) {
                oh->ot_handle = jh;
                LASSERT(oti->oti_txns == 0);

                atomic_inc(&dev->od_commit_cb_in_flight);
                lu_ref_add_at(&d->dd_lu_dev.ld_reference, &oh->ot_dev_link,
                              "osd-tx", th);
                oti->oti_txns++;
                rc = 0;
        } else {
                rc = PTR_ERR(jh);
        }
out:
        RETURN(rc);
}

static int osd_seq_exists(const struct lu_env *env,
                          struct osd_device *osd, u64 seq)
{
        struct lu_seq_range *range = &osd_oti_get(env)->oti_seq_range;
        struct seq_server_site *ss = osd_seq_site(osd);
        int rc;

        ENTRY;

        LASSERT(ss != NULL);
        LASSERT(ss->ss_server_fld != NULL);

        rc = osd_fld_lookup(env, osd, seq, range);
        if (rc != 0) {
                if (rc != -ENOENT)
                        CERROR("%s: can't lookup FLD sequence %#llx: rc = %d\n",
                               osd_name(osd), seq, rc);
                RETURN(0);
        }

        RETURN(ss->ss_node_id == range->lsr_index);
}

static void osd_trans_stop_cb(struct osd_thandle *oth, int result)
{
        struct dt_txn_commit_cb *dcb;
        struct dt_txn_commit_cb *tmp;

        /* call per-transaction stop callbacks if any */
        list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list,
                                 dcb_linkage) {
                LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC,
                         "commit callback entry: magic=%x name='%s'\n",
                         dcb->dcb_magic, dcb->dcb_name);
                list_del_init(&dcb->dcb_linkage);
                dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
        }
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
                          struct thandle *th)
{
        struct osd_thread_info *oti = osd_oti_get(env);
        struct osd_thandle *oh;
        struct osd_iobuf *iobuf = &oti->oti_iobuf;
        struct osd_device *osd = osd_dt_dev(th->th_dev);
        struct qsd_instance *qsd = osd_def_qsd(osd);
        struct lquota_trans *qtrans;
        LIST_HEAD(truncates);
        int rc = 0, remove_agents = 0;

        ENTRY;

        oh = container_of(th, struct osd_thandle, ot_super);

        remove_agents = oh->ot_remove_agents;

        qtrans = oh->ot_quota_trans;
        oh->ot_quota_trans = NULL;

        /* move locks to local list, stop tx, execute truncates */
        list_splice(&oh->ot_trunc_locks, &truncates);

        if (oh->ot_handle != NULL) {
                int rc2;

                handle_t *hdl = oh->ot_handle;

                /*
                 * add commit callback
                 * notice we don't do this in osd_trans_start()
                 * as underlying transaction can change during truncate
                 */
                ldiskfs_journal_callback_add(hdl, osd_trans_commit_cb,
                                             &oh->ot_jcb);

                LASSERT(oti->oti_txns == 1);
                oti->oti_txns--;

                rc = dt_txn_hook_stop(env, th);
                if (rc != 0)
                        CERROR("%s: failed in transaction hook: rc = %d\n",
                               osd_name(osd), rc);

                osd_trans_stop_cb(oh, rc);
                /* hook functions might modify th_sync */
                hdl->h_sync = th->th_sync;

                oh->ot_handle = NULL;
                OSD_CHECK_SLOW_TH(oh, osd, rc2 = ldiskfs_journal_stop(hdl));
                if (rc2 != 0)
                        CERROR("%s: failed to stop transaction: rc = %d\n",
                               osd_name(osd), rc2);
                if (!rc)
                        rc = rc2;

                osd_process_truncates(env, &truncates);
        } else {
                osd_trans_stop_cb(oh, th->th_result);
                OBD_FREE_PTR(oh);
        }

        osd_trunc_unlock_all(env, &truncates);

        /* inform the quota slave device that the transaction is stopping */
        qsd_op_end(env, qsd, qtrans);

        /*
         * as we want IO to journal and data IO be concurrent, we don't block
         * awaiting data IO completion in osd_do_bio(), instead we wait here
         * once transaction is submitted to the journal. all reqular requests
         * don't do direct IO (except read/write), thus this wait_event becomes
         * no-op for them.
         *
         * IMPORTANT: we have to wait till any IO submited by the thread is
         * completed otherwise iobuf may be corrupted by different request
         */
        wait_event(iobuf->dr_wait,
                       atomic_read(&iobuf->dr_numreqs) == 0);
        osd_fini_iobuf(osd, iobuf);
        if (!rc)
                rc = iobuf->dr_error;

        if (unlikely(remove_agents != 0))
                osd_process_scheduled_agent_removals(env, osd);

        LASSERT(oti->oti_ins_cache_depth > 0);
        oti->oti_ins_cache_depth--;
        /* reset OI cache for safety */
        if (oti->oti_ins_cache_depth == 0)
                oti->oti_ins_cache_used = 0;

        sb_end_write(osd_sb(osd));

        RETURN(rc);
}

static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb)
{
        struct osd_thandle *oh = container_of(th, struct osd_thandle,
                                              ot_super);

        LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC);
        LASSERT(&dcb->dcb_func != NULL);
        if (dcb->dcb_flags & DCB_TRANS_STOP)
                list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list);
        else
                list_add(&dcb->dcb_linkage, &oh->ot_commit_dcb_list);

        return 0;
}

/*
 * Called just before object is freed. Releases all resources except for
 * object itself (that is released by osd_object_free()).
 *
 * Concurrency: no concurrent access is possible that late in object
 * life-cycle.
 */
static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
{
        struct osd_object *obj = osd_obj(l);
        struct qsd_instance *qsd = osd_def_qsd(osd_obj2dev(obj));
        struct inode *inode = obj->oo_inode;
        __u64 projid;
        qid_t uid;
        qid_t gid;

        LINVRNT(osd_invariant(obj));

        /*
         * If object is unlinked remove fid->ino mapping from object index.
         */

        osd_index_fini(obj);

        if (!inode)
                return;

        if (osd_has_index(obj) &&  obj->oo_dt.do_index_ops == &osd_index_iam_ops)
                ldiskfs_set_inode_flag(inode, LDISKFS_INODE_JOURNAL_DATA);

        uid = i_uid_read(inode);
        gid = i_gid_read(inode);
        projid = i_projid_read(inode);

        obj->oo_inode = NULL;
        iput(inode);

        /* do not rebalance quota if the caller needs to release memory
         * otherwise qsd_refresh_usage() may went into a new ldiskfs
         * transaction and risk to deadlock - LU-12178 */
        if (current->flags & (PF_MEMALLOC | PF_KSWAPD))
                return;

        if (!obj->oo_header && qsd) {
                struct osd_thread_info *info = osd_oti_get(env);
                struct lquota_id_info *qi = &info->oti_qi;

                /* Release granted quota to master if necessary */
                qi->lqi_id.qid_uid = uid;
                qsd_op_adjust(env, qsd, &qi->lqi_id, USRQUOTA);

                qi->lqi_id.qid_uid = gid;
                qsd_op_adjust(env, qsd, &qi->lqi_id, GRPQUOTA);

                qi->lqi_id.qid_uid = projid;
                qsd_op_adjust(env, qsd, &qi->lqi_id, PRJQUOTA);
        }
}

/*
 * Concurrency: ->loo_object_release() is called under site spin-lock.
 */
static void osd_object_release(const struct lu_env *env,
                               struct lu_object *l)
{
        struct osd_object *o = osd_obj(l);

        /*
         * nobody should be releasing a non-destroyed object with nlink=0
         * the API allows this, but ldiskfs doesn't like and then report
         * this inode as deleted
         */
        LASSERT(!(o->oo_destroyed == 0 && o->oo_inode &&
                  o->oo_inode->i_nlink == 0));
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_object_print(const struct lu_env *env, void *cookie,
                            lu_printer_t p, const struct lu_object *l)
{
        struct osd_object *o = osd_obj(l);
        struct iam_descr *d;

        if (o->oo_dir != NULL)
                d = o->oo_dir->od_container.ic_descr;
        else
                d = NULL;
        return (*p)(env, cookie,
                    LUSTRE_OSD_LDISKFS_NAME"-object@%p(i:%p:%lu/%u)[%s]",
                    o, o->oo_inode,
                    o->oo_inode ? o->oo_inode->i_ino : 0UL,
                    o->oo_inode ? o->oo_inode->i_generation : 0,
                    d ? d->id_ops->id_name : "plain");
}

/*
 * Concurrency: shouldn't matter.
 */
int osd_statfs(const struct lu_env *env, struct dt_device *d,
                struct obd_statfs *sfs, struct obd_statfs_info *info)
{
        struct osd_device *osd = osd_dt_dev(d);
        struct super_block *sb = osd_sb(osd);
        struct kstatfs *ksfs;
        __u64 reserved;
        int result = 0;

        if (unlikely(osd->od_mnt == NULL))
                return -EINPROGRESS;

        /* osd_lproc.c call this without env, allocate ksfs for that case */
        if (unlikely(env == NULL)) {
                OBD_ALLOC_PTR(ksfs);
                if (ksfs == NULL)
                        return -ENOMEM;
        } else {
                ksfs = &osd_oti_get(env)->oti_ksfs;
        }

        result = sb->s_op->statfs(sb->s_root, ksfs);
        if (result)
                goto out;

        statfs_pack(sfs, ksfs);
        if (unlikely(sb->s_flags & SB_RDONLY))
                sfs->os_state |= OS_STATFS_READONLY;

        sfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0;

        if (ldiskfs_has_feature_extents(sb))
                sfs->os_maxbytes = sb->s_maxbytes;
        else
                sfs->os_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;

        /*
         * Reserve some space so to avoid fragmenting the filesystem too much.
         * Fragmentation not only impacts performance, but can also increase
         * metadata overhead significantly, causing grant calculation to be
         * wrong.
         *
         * Reserve 0.78% of total space, at least 8MB for small filesystems.
         */
        BUILD_BUG_ON(OSD_STATFS_RESERVED <= LDISKFS_MAX_BLOCK_SIZE);
        reserved = OSD_STATFS_RESERVED >> sb->s_blocksize_bits;
        if (likely(sfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT))
                reserved = sfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;

        sfs->os_blocks -= reserved;
        sfs->os_bfree  -= min(reserved, sfs->os_bfree);
        sfs->os_bavail -= min(reserved, sfs->os_bavail);

out:
        if (unlikely(env == NULL))
                OBD_FREE_PTR(ksfs);
        return result;
}

/**
 * Estimate space needed for file creations. We assume the largest filename
 * which is 2^64 - 1, hence a filename of 20 chars.
 * This is 28 bytes per object which is 28MB for 1M objects ... no so bad.
 */
#ifdef __LDISKFS_DIR_REC_LEN
#define PER_OBJ_USAGE __LDISKFS_DIR_REC_LEN(20)
#else
#define PER_OBJ_USAGE LDISKFS_DIR_REC_LEN(20)
#endif

/*
 * Concurrency: doesn't access mutable data.
 */
static void osd_conf_get(const struct lu_env *env,
                         const struct dt_device *dev,
                         struct dt_device_param *param)
{
        struct osd_device *d = osd_dt_dev(dev);
        struct super_block *sb = osd_sb(d);
        struct blk_integrity *bi = bdev_get_integrity(sb->s_bdev);
        const char *name;
        int ea_overhead;

        /*
         * XXX should be taken from not-yet-existing fs abstraction layer.
         */
        param->ddp_max_name_len = LDISKFS_NAME_LEN;
        param->ddp_max_nlink    = LDISKFS_LINK_MAX;
        param->ddp_symlink_max  = sb->s_blocksize;
        param->ddp_mount_type   = LDD_MT_LDISKFS;
        if (ldiskfs_has_feature_extents(sb))
                param->ddp_maxbytes = sb->s_maxbytes;
        else
                param->ddp_maxbytes = LDISKFS_SB(sb)->s_bitmap_maxbytes;
        /*
         * inode are statically allocated, so per-inode space consumption
         * is the space consumed by the directory entry
         */
        param->ddp_inodespace     = PER_OBJ_USAGE;
        /*
         * EXT_INIT_MAX_LEN is the theoretical maximum extent size (32k blocks
         * is 128MB) which is unlikely to be hit in real life. Report a smaller
         * maximum length to not under-count the actual number of extents
         * needed for writing a file if there are sub-optimal block allocations.
         */
        param->ddp_max_extent_blks = EXT_INIT_MAX_LEN >> 1;
        /* worst-case extent insertion metadata overhead */
        param->ddp_extent_tax = 6 * LDISKFS_BLOCK_SIZE(sb);
        param->ddp_mntopts = 0;
        if (test_opt(sb, XATTR_USER))
                param->ddp_mntopts |= MNTOPT_USERXATTR;
        if (test_opt(sb, POSIX_ACL))
                param->ddp_mntopts |= MNTOPT_ACL;

        /*
         * LOD might calculate the max stripe count based on max_ea_size,
         * so we need take account in the overhead as well,
         * xattr_header + magic + xattr_entry_head
         */
        ea_overhead = sizeof(struct ldiskfs_xattr_header) + sizeof(__u32) +
                      LDISKFS_XATTR_LEN(XATTR_NAME_MAX_LEN);

#if defined(LDISKFS_FEATURE_INCOMPAT_EA_INODE)
        if (ldiskfs_has_feature_ea_inode(sb))
                param->ddp_max_ea_size = LDISKFS_XATTR_MAX_LARGE_EA_SIZE -
                                                                ea_overhead;
        else
#endif
                param->ddp_max_ea_size = sb->s_blocksize - ea_overhead;

        if (param->ddp_max_ea_size > OBD_MAX_EA_SIZE)
                param->ddp_max_ea_size = OBD_MAX_EA_SIZE;

        /*
         * Preferred RPC size for efficient disk IO.  4MB shows good
         * all-around performance for ldiskfs, but use bigalloc chunk size
         * by default if larger.
         */
#if defined(LDISKFS_CLUSTER_SIZE)
        if (LDISKFS_CLUSTER_SIZE(sb) > DT_DEF_BRW_SIZE)
                param->ddp_brw_size = LDISKFS_CLUSTER_SIZE(sb);
        else
#endif
                param->ddp_brw_size = DT_DEF_BRW_SIZE;

        param->ddp_t10_cksum_type = 0;
        if (bi) {
                unsigned short interval = blk_integrity_interval(bi);
                name = blk_integrity_name(bi);
                /*
                 * Expected values:
                 * T10-DIF-TYPE1-CRC
                 * T10-DIF-TYPE3-CRC
                 * T10-DIF-TYPE1-IP
                 * T10-DIF-TYPE3-IP
                 */
                if (strncmp(name, "T10-DIF-TYPE",
                            sizeof("T10-DIF-TYPE") - 1) == 0) {
                        /* also skip "1/3-" at end */
                        const int type_off = sizeof("T10-DIF-TYPE.");
                        char type_number = name[type_off - 2];

                        if (interval != 512 && interval != 4096) {
                                CERROR("%s: unsupported T10PI sector size %u\n",
                                       d->od_svname, interval);
                        } else if (type_number != '1' && type_number != '3') {
                                CERROR("%s: unsupported T10PI type %s\n",
                                       d->od_svname, name);
                        } else if (strcmp(name + type_off, "CRC") == 0) {
                                d->od_t10_type = type_number == '1' ?
                                        OSD_T10_TYPE1_CRC : OSD_T10_TYPE3_CRC;
                                param->ddp_t10_cksum_type = interval == 512 ?
                                        OBD_CKSUM_T10CRC512 :
                                        OBD_CKSUM_T10CRC4K;
                        } else if (strcmp(name + type_off, "IP") == 0) {
                                d->od_t10_type = type_number == '1' ?
                                        OSD_T10_TYPE1_IP : OSD_T10_TYPE3_IP;
                                param->ddp_t10_cksum_type = interval == 512 ?
                                        OBD_CKSUM_T10IP512 :
                                        OBD_CKSUM_T10IP4K;
                        } else {
                                CERROR("%s: unsupported checksum type of T10PI type '%s'\n",
                                       d->od_svname, name);
                        }

                } else {
                        CERROR("%s: unsupported T10PI type '%s'\n",
                               d->od_svname, name);
                }
        }

        param->ddp_has_lseek_data_hole = true;
}

static struct super_block *osd_mnt_sb_get(const struct dt_device *d)
{
        return osd_sb(osd_dt_dev(d));
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_sync(const struct lu_env *env, struct dt_device *d)
{
        int rc;
        struct super_block *s = osd_sb(osd_dt_dev(d));
        ENTRY;

        down_read(&s->s_umount);
        rc = s->s_op->sync_fs(s, 1);
        up_read(&s->s_umount);

        CDEBUG(D_CACHE, "%s: synced OSD: rc = %d\n", osd_dt_dev(d)->od_svname,
               rc);

        return rc;
}

/**
 * Start commit for OSD device.
 *
 * An implementation of dt_commit_async method for OSD device.
 * Asychronously starts underlayng fs sync and thereby a transaction
 * commit.
 *
 * \param env environment
 * \param d dt device
 *
 * \see dt_device_operations
 */
static int osd_commit_async(const struct lu_env *env,
                            struct dt_device *d)
{
        struct super_block *s = osd_sb(osd_dt_dev(d));
        int rc;

        ENTRY;

        CDEBUG(D_HA, "%s: async commit OSD\n", osd_dt_dev(d)->od_svname);
        down_read(&s->s_umount);
        rc = s->s_op->sync_fs(s, 0);
        up_read(&s->s_umount);

        RETURN(rc);
}

static int (*priv_security_file_alloc)(struct file *file);

int osd_security_file_alloc(struct file *file)
{
        if (priv_security_file_alloc)
                return priv_security_file_alloc(file);
        return 0;
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_ro(const struct lu_env *env, struct dt_device *d)
{
        struct super_block *sb = osd_sb(osd_dt_dev(d));
        struct block_device *dev = sb->s_bdev;
        int rc = -EOPNOTSUPP;

        ENTRY;

        CERROR("%s: %lx CANNOT BE SET READONLY: rc = %d\n",
               osd_dt_dev(d)->od_svname, (long)dev, rc);

        RETURN(rc);
}

/**
 * Note: we do not count into QUOTA here.
 * If we mount with --data_journal we may need more.
 */
const int osd_dto_credits_noquota[DTO_NR] = {
        /**
         * Insert.
         * INDEX_EXTRA_TRANS_BLOCKS(8) +
         * SINGLEDATA_TRANS_BLOCKS(8)
         * XXX Note: maybe iam need more, since iam have more level than
         *           EXT3 htree.
         */
        [DTO_INDEX_INSERT]  = 16,
        /**
         * Delete
         * just modify a single entry, probably merge few within a block
         */
        [DTO_INDEX_DELETE]  = 1,
        /**
         * Used for OI scrub
         */
        [DTO_INDEX_UPDATE]  = 16,
        /**
         * 4(inode, inode bits, groups, GDT)
         *   notice: OI updates are counted separately with DTO_INDEX_INSERT
         */
        [DTO_OBJECT_CREATE] = 4,
        /**
         * 4(inode, inode bits, groups, GDT)
         *   notice: OI updates are counted separately with DTO_INDEX_DELETE
         */
        [DTO_OBJECT_DELETE] = 4,
        /**
         * Attr set credits (inode)
         */
        [DTO_ATTR_SET_BASE] = 1,
        /**
         * Xattr set. The same as xattr of EXT3.
         * DATA_TRANS_BLOCKS(14)
         * XXX Note: in original MDS implmentation INDEX_EXTRA_TRANS_BLOCKS
         * are also counted in. Do not know why?
         */
        [DTO_XATTR_SET]     = 14,
        /**
         * credits for inode change during write.
         */
        [DTO_WRITE_BASE]    = 3,
        /**
         * credits for single block write.
         */
        [DTO_WRITE_BLOCK]   = 14,
        /**
         * Attr set credits for chown.
         * This is extra credits for setattr, and it is null without quota
         */
        [DTO_ATTR_SET_CHOWN] = 0
};

/* reserve or free quota for some operation */
static int osd_reserve_or_free_quota(const struct lu_env *env,
                                     struct dt_device *dev,
                                     enum quota_type type, __u64 uid,
                                     __u64 gid, __s64 count, bool is_md)
{
        int rc;
        struct osd_device       *osd = osd_dt_dev(dev);
        struct osd_thread_info  *info = osd_oti_get(env);
        struct lquota_id_info   *qi = &info->oti_qi;
        struct qsd_instance     *qsd = NULL;

        ENTRY;

        if (is_md)
                qsd = osd->od_quota_slave_md;
        else
                qsd = osd->od_quota_slave_dt;

        rc = quota_reserve_or_free(env, qsd, qi, type, uid, gid, count, is_md);
        RETURN(rc);
}

static const struct dt_device_operations osd_dt_ops = {
        .dt_root_get              = osd_root_get,
        .dt_statfs                = osd_statfs,
        .dt_trans_create          = osd_trans_create,
        .dt_trans_start           = osd_trans_start,
        .dt_trans_stop            = osd_trans_stop,
        .dt_trans_cb_add          = osd_trans_cb_add,
        .dt_conf_get              = osd_conf_get,
        .dt_mnt_sb_get            = osd_mnt_sb_get,
        .dt_sync                  = osd_sync,
        .dt_ro                    = osd_ro,
        .dt_commit_async          = osd_commit_async,
        .dt_reserve_or_free_quota = osd_reserve_or_free_quota,
};

static void osd_read_lock(const struct lu_env *env, struct dt_object *dt,
                          unsigned int role)
{
        struct osd_object *obj = osd_dt_obj(dt);
        struct osd_thread_info *oti = osd_oti_get(env);

        LINVRNT(osd_invariant(obj));

        LASSERT(obj->oo_owner != env);
        down_read_nested(&obj->oo_sem, role);

        LASSERT(obj->oo_owner == NULL);
        oti->oti_r_locks++;
}

static void osd_write_lock(const struct lu_env *env, struct dt_object *dt,
                           unsigned int role)
{
        struct osd_object *obj = osd_dt_obj(dt);
        struct osd_thread_info *oti = osd_oti_get(env);

        LINVRNT(osd_invariant(obj));

        LASSERT(obj->oo_owner != env);
        down_write_nested(&obj->oo_sem, role);

        LASSERT(obj->oo_owner == NULL);
        obj->oo_owner = env;
        oti->oti_w_locks++;
}

static void osd_read_unlock(const struct lu_env *env, struct dt_object *dt)
{
        struct osd_object *obj = osd_dt_obj(dt);
        struct osd_thread_info *oti = osd_oti_get(env);

        LINVRNT(osd_invariant(obj));

        LASSERT(oti->oti_r_locks > 0);
        oti->oti_r_locks--;
        up_read(&obj->oo_sem);
}

static void osd_write_unlock(const struct lu_env *env, struct dt_object *dt)
{
        struct osd_object *obj = osd_dt_obj(dt);
        struct osd_thread_info *oti = osd_oti_get(env);

        LINVRNT(osd_invariant(obj));

        LASSERT(obj->oo_owner == env);
        LASSERT(oti->oti_w_locks > 0);
        oti->oti_w_locks--;
        obj->oo_owner = NULL;
        up_write(&obj->oo_sem);
}

static int osd_write_locked(const struct lu_env *env, struct dt_object *dt)
{
        struct osd_object *obj = osd_dt_obj(dt);

        LINVRNT(osd_invariant(obj));

        return obj->oo_owner == env;
}

static void osd_inode_getattr(const struct lu_env *env,
                              struct inode *inode, struct lu_attr *attr)
{
        attr->la_valid  |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
                           LA_SIZE | LA_BLOCKS | LA_UID | LA_GID |
                           LA_PROJID | LA_FLAGS | LA_NLINK | LA_RDEV |
                           LA_BLKSIZE | LA_TYPE | LA_BTIME;

        attr->la_atime = inode->i_atime.tv_sec;
        attr->la_mtime = inode->i_mtime.tv_sec;
        attr->la_ctime = inode->i_ctime.tv_sec;
        attr->la_btime = LDISKFS_I(inode)->i_crtime.tv_sec;
        attr->la_mode    = inode->i_mode;
        attr->la_size    = i_size_read(inode);
        attr->la_blocks  = inode->i_blocks;
        attr->la_uid     = i_uid_read(inode);
        attr->la_gid     = i_gid_read(inode);
        attr->la_projid  = i_projid_read(inode);
        attr->la_flags   = ll_inode_to_ext_flags(inode->i_flags);
        attr->la_nlink   = inode->i_nlink;
        attr->la_rdev    = inode->i_rdev;
        attr->la_blksize = 1 << inode->i_blkbits;
        attr->la_blkbits = inode->i_blkbits;
        /*
         * Ext4 did not transfer inherit flags from raw inode
         * to inode flags, and ext4 internally test raw inode
         * @i_flags directly. Instead of patching ext4, we do it here.
         */
        if (LDISKFS_I(inode)->i_flags & LUSTRE_PROJINHERIT_FL)
                attr->la_flags |= LUSTRE_PROJINHERIT_FL;
}

static int osd_dirent_count(const struct lu_env *env, struct dt_object *dt,
                            u64 *count)
{
        struct osd_object *obj = osd_dt_obj(dt);
        const struct dt_it_ops *iops;
        struct dt_it *it;
        int rc;

        ENTRY;

        LASSERT(S_ISDIR(obj->oo_inode->i_mode));
        LASSERT(fid_is_namespace_visible(lu_object_fid(&obj->oo_dt.do_lu)));

        if (obj->oo_dirent_count != LU_DIRENT_COUNT_UNSET) {
                *count = obj->oo_dirent_count;
                RETURN(0);
        }

        /* directory not initialized yet */
        if (!dt->do_index_ops) {
                *count = 0;
                RETURN(0);
        }

        iops = &dt->do_index_ops->dio_it;
        it = iops->init(env, dt, LUDA_64BITHASH);
        if (IS_ERR(it))
                RETURN(PTR_ERR(it));

        rc = iops->load(env, it, 0);
        if (rc < 0) {
                if (rc == -ENODATA) {
                        rc = 0;
                        *count = 0;
                }
                GOTO(out, rc);
        }
        if (rc > 0)
                rc = iops->next(env, it);

        for (*count = 0; rc == 0 || rc == -ESTALE; rc = iops->next(env, it)) {
                if (rc == -ESTALE)
                        continue;

                if (iops->key_size(env, it) == 0)
                        continue;

                (*count)++;
        }
        if (rc == 1) {
                obj->oo_dirent_count = *count;
                rc = 0;
        }
out:
        iops->put(env, it);
        iops->fini(env, it);

        RETURN(rc);
}

static int osd_attr_get(const struct lu_env *env, struct dt_object *dt,
                        struct lu_attr *attr)
{
        struct osd_object *obj = osd_dt_obj(dt);
        int rc = 0;

        if (unlikely(!dt_object_exists(dt)))
                return -ENOENT;
        if (unlikely(obj->oo_destroyed))
                return -ENOENT;

        LASSERT(!dt_object_remote(dt));
        LINVRNT(osd_invariant(obj));

        spin_lock(&obj->oo_guard);
        osd_inode_getattr(env, obj->oo_inode, attr);
        if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL) {
                attr->la_valid |= LA_FLAGS;
                attr->la_flags |= LUSTRE_ORPHAN_FL;
        }
        if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL) {
                attr->la_valid |= LA_FLAGS;
                attr->la_flags |= LUSTRE_ENCRYPT_FL;
        }
        spin_unlock(&obj->oo_guard);

        if (S_ISDIR(obj->oo_inode->i_mode) &&
            fid_is_namespace_visible(lu_object_fid(&dt->do_lu)))
                rc = osd_dirent_count(env, dt, &attr->la_dirent_count);

        return rc;
}

static int osd_declare_attr_qid(const struct lu_env *env,
                                struct osd_object *obj,
                                struct osd_thandle *oh, long long bspace,
                                qid_t old_id, qid_t new_id, bool enforce,
                                unsigned int type)
{
        int rc;
        struct osd_thread_info *info = osd_oti_get(env);
        struct lquota_id_info  *qi = &info->oti_qi;

        qi->lqi_type = type;
        /* inode accounting */
        qi->lqi_is_blk = false;

        /* one more inode for the new id ... */
        qi->lqi_id.qid_uid = new_id;
        qi->lqi_space      = 1;
        /* Reserve credits for the new id */
        rc = osd_declare_qid(env, oh, qi, NULL, enforce, NULL);
        if (rc == -EDQUOT || rc == -EINPROGRESS)
                rc = 0;
        if (rc)
                RETURN(rc);

        /* and one less inode for the current id */
        qi->lqi_id.qid_uid = old_id;
        qi->lqi_space = -1;
        rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
        if (rc == -EDQUOT || rc == -EINPROGRESS)
                rc = 0;
        if (rc)
                RETURN(rc);

        /* block accounting */
        qi->lqi_is_blk = true;

        /* more blocks for the new id ... */
        qi->lqi_id.qid_uid = new_id;
        qi->lqi_space      = bspace;
        /*
         * Credits for the new uid has been reserved, re-use "obj"
         * to save credit reservation.
         */
        rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
        if (rc == -EDQUOT || rc == -EINPROGRESS)
                rc = 0;
        if (rc)
                RETURN(rc);

        /* and finally less blocks for the current uid */
        qi->lqi_id.qid_uid = old_id;
        qi->lqi_space      = -bspace;
        rc = osd_declare_qid(env, oh, qi, obj, enforce, NULL);
        if (rc == -EDQUOT || rc == -EINPROGRESS)
                rc = 0;

        RETURN(rc);
}

static int osd_declare_attr_set(const struct lu_env *env,
                                struct dt_object *dt,
                                const struct lu_attr *attr,
                                struct thandle *handle)
{
        struct osd_thandle *oh;
        struct osd_object *obj;
        qid_t uid;
        qid_t gid;
        long long bspace;
        int rc = 0;
        bool enforce;

        ENTRY;

        LASSERT(dt != NULL);
        LASSERT(handle != NULL);

        obj = osd_dt_obj(dt);
        LASSERT(osd_invariant(obj));

        oh = container_of(handle, struct osd_thandle, ot_super);
        LASSERT(oh->ot_handle == NULL);

        osd_trans_declare_op(env, oh, OSD_OT_ATTR_SET,
                             osd_dto_credits_noquota[DTO_ATTR_SET_BASE]);

        osd_trans_declare_op(env, oh, OSD_OT_XATTR_SET,
                             osd_dto_credits_noquota[DTO_XATTR_SET]);

        if (attr == NULL || obj->oo_inode == NULL)
                RETURN(rc);

        bspace   = obj->oo_inode->i_blocks << 9;
        bspace   = toqb(bspace);

        /*
         * Changing ownership is always preformed by super user, it should not
         * fail with EDQUOT unless required explicitly.
         *
         * We still need to call the osd_declare_qid() to calculate the journal
         * credits for updating quota accounting files and to trigger quota
         * space adjustment once the operation is completed.
         */
        if (attr->la_valid & LA_UID || attr->la_valid & LA_GID) {
                /* USERQUOTA */
                uid = i_uid_read(obj->oo_inode);
                enforce = (attr->la_valid & LA_UID) && (attr->la_uid != uid);
                rc = osd_declare_attr_qid(env, obj, oh, bspace, uid,
                                          attr->la_uid, enforce, USRQUOTA);
                if (rc)
                        RETURN(rc);

                gid = i_gid_read(obj->oo_inode);
                CDEBUG(D_QUOTA, "declare uid %d -> %d gid %d -> %d\n", uid,
                       attr->la_uid, gid, attr->la_gid);
                enforce = (attr->la_valid & LA_GID) && (attr->la_gid != gid);
                rc = osd_declare_attr_qid(env, obj, oh, bspace, gid,
                                          attr->la_gid, enforce, GRPQUOTA);
                if (rc)
                        RETURN(rc);

        }
#ifdef HAVE_PROJECT_QUOTA
        if (attr->la_valid & LA_PROJID) {
                __u32 projid = i_projid_read(obj->oo_inode);

                enforce = (attr->la_valid & LA_PROJID) &&
                                        (attr->la_projid != projid);
                rc = osd_declare_attr_qid(env, obj, oh, bspace,
                                          (qid_t)projid, (qid_t)attr->la_projid,
                                          enforce, PRJQUOTA);
                if (rc)
                        RETURN(rc);
        }
#endif
        RETURN(rc);
}

static int osd_inode_setattr(const struct lu_env *env,
                             struct inode *inode, const struct lu_attr *attr)
{
        __u64 bits = attr->la_valid;

        /* Only allow set size for regular file */
        if (!S_ISREG(inode->i_mode))
                bits &= ~(LA_SIZE | LA_BLOCKS);

        if (bits == 0)
                return 0;

        if (bits & LA_ATIME)
                inode->i_atime = osd_inode_time(inode, attr->la_atime);
        if (bits & LA_CTIME)
                inode->i_ctime = osd_inode_time(inode, attr->la_ctime);
        if (bits & LA_MTIME)
                inode->i_mtime = osd_inode_time(inode, attr->la_mtime);
        if (bits & LA_SIZE) {
                spin_lock(&inode->i_lock);
                LDISKFS_I(inode)->i_disksize = attr->la_size;
                i_size_write(inode, attr->la_size);
                spin_unlock(&inode->i_lock);
        }

        /*
         * OSD should not change "i_blocks" which is used by quota.
         * "i_blocks" should be changed by ldiskfs only.
         */
        if (bits & LA_MODE)
                inode->i_mode = (inode->i_mode & S_IFMT) |
                                (attr->la_mode & ~S_IFMT);
        if (bits & LA_UID)
                i_uid_write(inode, attr->la_uid);
        if (bits & LA_GID)
                i_gid_write(inode, attr->la_gid);
        if (bits & LA_PROJID)
                i_projid_write(inode, attr->la_projid);
        if (bits & LA_NLINK)
                set_nlink(inode, attr->la_nlink);
        if (bits & LA_RDEV)
                inode->i_rdev = attr->la_rdev;

        if (bits & LA_FLAGS) {
                /* always keep S_NOCMTIME */
                inode->i_flags = ll_ext_to_inode_flags(attr->la_flags) |
                                 S_NOCMTIME;
#if defined(S_ENCRYPTED)
                /* Always remove S_ENCRYPTED, because ldiskfs must not be
                 * aware of encryption status. It is just stored into LMA
                 * so that it can be forwared to client side.
                 */
                inode->i_flags &= ~S_ENCRYPTED;
#endif
                /*
                 * Ext4 did not transfer inherit flags from
                 * @inode->i_flags to raw inode i_flags when writing
                 * flags, we do it explictly here.
                 */
                if (attr->la_flags & LUSTRE_PROJINHERIT_FL)
                        LDISKFS_I(inode)->i_flags |= LUSTRE_PROJINHERIT_FL;
                else
                        LDISKFS_I(inode)->i_flags &= ~LUSTRE_PROJINHERIT_FL;
        }
        return 0;
}

#ifdef HAVE_PROJECT_QUOTA
static int osd_transfer_project(struct inode *inode, __u32 projid,
                                struct thandle *handle)
{
        struct super_block *sb = inode->i_sb;
        struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
        int err;
        kprojid_t kprojid;
        struct ldiskfs_iloc iloc;
        struct ldiskfs_inode *raw_inode;
        struct dquot *transfer_to[LDISKFS_MAXQUOTAS] = { };

        if (!ldiskfs_has_feature_project(sb)) {
                LASSERT(__kprojid_val(LDISKFS_I(inode)->i_projid)
                        == LDISKFS_DEF_PROJID);
                if (projid != LDISKFS_DEF_PROJID)
                        return -EOPNOTSUPP;
                else
                        return 0;
        }

        if (LDISKFS_INODE_SIZE(sb) <= LDISKFS_GOOD_OLD_INODE_SIZE)
                return -EOPNOTSUPP;

        kprojid = make_kprojid(&init_user_ns, (projid_t)projid);
        if (projid_eq(kprojid, LDISKFS_I(inode)->i_projid))
                return 0;

        err = ldiskfs_get_inode_loc(inode, &iloc);
        if (err)
                return err;

        raw_inode = ldiskfs_raw_inode(&iloc);
        if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
                struct osd_thandle *oh = container_of(handle,
                                                      struct osd_thandle,
                                                      ot_super);
                /**
                 * try to expand inode size automatically.
                 */
                ldiskfs_mark_inode_dirty(oh->ot_handle, inode);
                if (!LDISKFS_FITS_IN_INODE(raw_inode, ei, i_projid)) {
                        err = -EOVERFLOW;
                        brelse(iloc.bh);
                        return err;
                }
        }
        brelse(iloc.bh);

        dquot_initialize(inode);
        transfer_to[PRJQUOTA] = dqget(sb, make_kqid_projid(kprojid));
        if (transfer_to[PRJQUOTA]) {
                err = __dquot_transfer(inode, transfer_to);
                dqput(transfer_to[PRJQUOTA]);
                if (err)
                        return err;
        }

        return err;
}
#endif

static int osd_quota_transfer(struct inode *inode, const struct lu_attr *attr,
                              struct thandle *handle)
{
        int rc;

        if ((attr->la_valid & LA_UID && attr->la_uid != i_uid_read(inode)) ||
            (attr->la_valid & LA_GID && attr->la_gid != i_gid_read(inode))) {
                struct iattr iattr;

                CDEBUG(D_QUOTA,
                       "executing dquot_transfer inode %ld uid %d -> %d gid %d -> %d\n",
                       inode->i_ino, i_uid_read(inode), attr->la_uid,
                       i_gid_read(inode), attr->la_gid);

                dquot_initialize(inode);
                iattr.ia_valid = 0;
                if (attr->la_valid & LA_UID)
                        iattr.ia_valid |= ATTR_UID;
                if (attr->la_valid & LA_GID)
                        iattr.ia_valid |= ATTR_GID;
                iattr.ia_uid = make_kuid(&init_user_ns, attr->la_uid);
                iattr.ia_gid = make_kgid(&init_user_ns, attr->la_gid);

                rc = dquot_transfer(inode, &iattr);
                if (rc) {
                        CERROR("%s: quota transfer failed. Is quota enforcement enabled on the ldiskfs filesystem? rc = %d\n",
                               osd_ino2name(inode), rc);
                        return rc;
                }
        }

        /* Handle project id transfer here properly */
        if (attr->la_valid & LA_PROJID &&
            attr->la_projid != i_projid_read(inode)) {
                if (!projid_valid(make_kprojid(&init_user_ns, attr->la_projid)))
                        return -EINVAL;
#ifdef HAVE_PROJECT_QUOTA
                rc = osd_transfer_project(inode, attr->la_projid, handle);
#else
                rc = -ENOTSUPP;
#endif
                if (rc) {
                        CERROR("%s: quota transfer failed. Is project enforcement enabled on the ldiskfs filesystem? rc = %d\n",
                               osd_ino2name(inode), rc);
                        return rc;
                }
        }
        return 0;
}

static int osd_attr_set(const struct lu_env *env,
                        struct dt_object *dt,
                        const struct lu_attr *attr,
                        struct thandle *handle)
{
        struct osd_object *obj = osd_dt_obj(dt);
        struct inode *inode;
        int rc;

        if (!dt_object_exists(dt))
                return -ENOENT;

        LASSERT(handle != NULL);
        LASSERT(!dt_object_remote(dt));
        LASSERT(osd_invariant(obj));

        osd_trans_exec_op(env, handle, OSD_OT_ATTR_SET);

        if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FID_MAPPING) &&
            !osd_obj2dev(obj)->od_is_ost) {
                struct osd_thread_info *oti = osd_oti_get(env);
                const struct lu_fid *fid0 = lu_object_fid(&dt->do_lu);
                struct lu_fid *fid1 = &oti->oti_fid;
                struct osd_inode_id *id = &oti->oti_id;
                struct iam_path_descr *ipd;
                struct iam_container *bag;
                struct osd_thandle *oh;
                int rc;

                fid_cpu_to_be(fid1, fid0);
                memset(id, 1, sizeof(*id));
                bag = &osd_fid2oi(osd_dev(dt->do_lu.lo_dev),
                                  fid0)->oi_dir.od_container;
                ipd = osd_idx_ipd_get(env, bag);
                if (unlikely(ipd == NULL))
                        RETURN(-ENOMEM);

                oh = container_of(handle, struct osd_thandle, ot_super);
                rc = iam_update(oh->ot_handle, bag,
                                (const struct iam_key *)fid1,
                                (const struct iam_rec *)id, ipd);
                osd_ipd_put(env, bag, ipd);
                return(rc > 0 ? 0 : rc);
        }

        inode = obj->oo_inode;

        rc = osd_quota_transfer(inode, attr, handle);
        if (rc)
                return rc;

        spin_lock(&obj->oo_guard);
        rc = osd_inode_setattr(env, inode, attr);
        spin_unlock(&obj->oo_guard);
        if (rc != 0)
                GOTO(out, rc);

        osd_dirty_inode(inode, I_DIRTY_DATASYNC);

        osd_trans_exec_check(env, handle, OSD_OT_ATTR_SET);