[fs/lustre-release.git] / lustre / osd-zfs / 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.sun.com/software/products/lustre/docs/GPLv2.pdf
 *
 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
 * CA 95054 USA or visit www.sun.com if you need additional information or
 * have any questions.
 *
 * GPL HEADER END
 */
/*
 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
 * Use is subject to license terms.
 *
 * Copyright (c) 2012, 2014, Intel Corporation.
 */
/*
 * This file is part of Lustre, http://www.lustre.org/
 * Lustre is a trademark of Sun Microsystems, Inc.
 *
 * lustre/osd-zfs/osd_handler.c
 * Top-level entry points into osd module
 *
 * Author: Alex Zhuravlev <bzzz@whamcloud.com>
 * Author: Mike Pershin <tappro@whamcloud.com>
 * Author: Johann Lombardi <johann@whamcloud.com>
 */

#define DEBUG_SUBSYSTEM S_OSD

#include <lustre_ver.h>
#include <libcfs/libcfs.h>
#include <obd_support.h>
#include <lustre_net.h>
#include <obd.h>
#include <obd_class.h>
#include <lustre_disk.h>
#include <lustre_fid.h>
#include <lustre_param.h>
#include <md_object.h>

#include "osd_internal.h"

#include <sys/dnode.h>
#include <sys/dbuf.h>
#include <sys/spa.h>
#include <sys/stat.h>
#include <sys/zap.h>
#include <sys/spa_impl.h>
#include <sys/zfs_znode.h>
#include <sys/dmu_tx.h>
#include <sys/dmu_objset.h>
#include <sys/dsl_prop.h>
#include <sys/sa_impl.h>
#include <sys/txg.h>

struct lu_context_key   osd_key;

/* Slab for OSD object allocation */
struct kmem_cache *osd_object_kmem;

/* Slab to allocate osd_zap_it */
struct kmem_cache *osd_zapit_cachep;

static struct lu_kmem_descr osd_caches[] = {
        {
                .ckd_cache = &osd_object_kmem,
                .ckd_name  = "zfs_osd_obj",
                .ckd_size  = sizeof(struct osd_object)
        },
        {
                .ckd_cache = &osd_zapit_cachep,
                .ckd_name  = "osd_zapit_cache",
                .ckd_size  = sizeof(struct osd_zap_it)
        },
        {
                .ckd_cache = NULL
        }
};

static void arc_prune_func(int64_t bytes, void *private)
{
        struct osd_device *od = private;
        struct lu_site    *site = &od->od_site;
        struct lu_env      env;
        int rc;

        rc = lu_env_init(&env, LCT_SHRINKER);
        if (rc) {
                CERROR("%s: can't initialize shrinker env: rc = %d\n",
                       od->od_svname, rc);
                return;
        }

        lu_site_purge(&env, site, (bytes >> 10));

        lu_env_fini(&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;
}

/*
 * OSD object methods.
 */

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

        ENTRY;

        if (error) {
                if (error == ECANCELED)
                        CWARN("%s: transaction @0x%p was aborted\n",
                              osd_dt_dev(th->th_dev)->od_svname, th);
                else
                        CERROR("%s: transaction @0x%p commit error: rc = %d\n",
                                osd_dt_dev(th->th_dev)->od_svname, th, error);
        }

        dt_txn_hook_commit(th);

        /* call per-transaction callbacks if any */
        list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage)
                dcb->dcb_func(NULL, th, dcb, error);

        /* Unlike ldiskfs, zfs updates space accounting at commit time.
         * As a consequence, op_end is called only now to inform the quota slave
         * component that reserved quota space is now accounted in usage and
         * should be released. Quota space won't be adjusted at this point since
         * we can't provide a suitable environment. It will be performed
         * asynchronously by a lquota thread. */
        qsd_op_end(NULL, osd->od_quota_slave, &oh->ot_quota_trans);

        lu_device_put(lud);
        th->th_dev = NULL;
        lu_context_exit(&th->th_ctx);
        lu_context_fini(&th->th_ctx);
        thandle_put(&oh->ot_super);

        EXIT;
}

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

        oh = container_of0(th, struct osd_thandle, ot_super);
        list_add(&dcb->dcb_linkage, &oh->ot_dcb_list);

        return 0;
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_trans_start(const struct lu_env *env, struct dt_device *d,
                           struct thandle *th)
{
        struct osd_thandle      *oh;
        int                     rc;
        ENTRY;

        oh = container_of0(th, struct osd_thandle, ot_super);
        LASSERT(oh);
        LASSERT(oh->ot_tx);

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

        if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC))
                /* Unlike ldiskfs, ZFS checks for available space and returns
                 * -ENOSPC when assigning txg */
                RETURN(-ENOSPC);

        rc = -dmu_tx_assign(oh->ot_tx, TXG_WAIT);
        if (unlikely(rc != 0)) {
                struct osd_device *osd = osd_dt_dev(d);
                /* dmu will call commit callback with error code during abort */
                if (!lu_device_is_md(&d->dd_lu_dev) && rc == -ENOSPC)
                        CERROR("%s: failed to start transaction due to ENOSPC. "
                               "Metadata overhead is underestimated or "
                               "grant_ratio is too low.\n", osd->od_svname);
                else
                        CERROR("%s: can't assign tx: rc = %d\n",
                               osd->od_svname, rc);
        } else {
                /* add commit callback */
                dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh);
                oh->ot_assigned = 1;
                lu_context_init(&th->th_ctx, th->th_tags);
                lu_context_enter(&th->th_ctx);
                lu_device_get(&d->dd_lu_dev);
        }

        RETURN(rc);
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt,
                          struct thandle *th)
{
        struct osd_device       *osd = osd_dt_dev(th->th_dev);
        struct osd_thandle      *oh;
        uint64_t                 txg;
        int                      rc;
        ENTRY;

        oh = container_of0(th, struct osd_thandle, ot_super);

        if (oh->ot_assigned == 0) {
                LASSERT(oh->ot_tx);
                dmu_tx_abort(oh->ot_tx);
                osd_object_sa_dirty_rele(oh);
                /* there won't be any commit, release reserved quota space now,
                 * if any */
                qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);
                thandle_put(&oh->ot_super);
                RETURN(0);
        }

        /* When doing our own inode accounting, the ZAPs storing per-uid/gid
         * usage are updated at operation execution time, so we should call
         * qsd_op_end() straight away. Otherwise (for blk accounting maintained
         * by ZFS and when #inode is estimated from #blks) accounting is updated
         * at commit time and the call to qsd_op_end() must be delayed */
        if (oh->ot_quota_trans.lqt_id_cnt > 0 &&
                        !oh->ot_quota_trans.lqt_ids[0].lqi_is_blk &&
                        !osd->od_quota_iused_est)
                qsd_op_end(env, osd->od_quota_slave, &oh->ot_quota_trans);

        rc = dt_txn_hook_stop(env, th);
        if (rc != 0)
                CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n",
                       osd->od_svname, rc);

        LASSERT(oh->ot_tx);
        txg = oh->ot_tx->tx_txg;

        osd_object_sa_dirty_rele(oh);
        dmu_tx_commit(oh->ot_tx);

        if (th->th_sync)
                txg_wait_synced(dmu_objset_pool(osd->od_os), txg);

        RETURN(rc);
}

static struct thandle *osd_trans_create(const struct lu_env *env,
                                        struct dt_device *dt)
{
        struct osd_device       *osd = osd_dt_dev(dt);
        struct osd_thandle      *oh;
        struct thandle          *th;
        dmu_tx_t                *tx;
        ENTRY;

        tx = dmu_tx_create(osd->od_os);
        if (tx == NULL)
                RETURN(ERR_PTR(-ENOMEM));

        /* alloc callback data */
        OBD_ALLOC_PTR(oh);
        if (oh == NULL) {
                dmu_tx_abort(tx);
                RETURN(ERR_PTR(-ENOMEM));
        }

        oh->ot_tx = tx;
        INIT_LIST_HEAD(&oh->ot_dcb_list);
        INIT_LIST_HEAD(&oh->ot_sa_list);
        sema_init(&oh->ot_sa_lock, 1);
        memset(&oh->ot_quota_trans, 0, sizeof(oh->ot_quota_trans));
        th = &oh->ot_super;
        th->th_dev = dt;
        th->th_result = 0;
        th->th_tags = LCT_TX_HANDLE;
        atomic_set(&th->th_refc, 1);
        th->th_alloc_size = sizeof(*oh);
        RETURN(th);
}

/* Estimate the number of objects from a number of blocks */
uint64_t osd_objs_count_estimate(uint64_t refdbytes, uint64_t usedobjs,
                                 uint64_t nrblocks)
{
        uint64_t est_objs, est_refdblocks, est_usedobjs;

        /* Compute an nrblocks estimate based on the actual number of
         * dnodes that could fit in the space.  Since we don't know the
         * overhead associated with each dnode (xattrs, SAs, VDEV overhead,
         * etc) just using DNODE_SHIFT isn't going to give a good estimate.
         * Instead, compute an estimate based on the average space usage per
         * dnode, with an upper and lower cap.
         *
         * In case there aren't many dnodes or blocks used yet, add a small
         * correction factor using OSD_DNODE_EST_SHIFT.  This correction
         * factor gradually disappears as the number of real dnodes grows.
         * This also avoids the need to check for divide-by-zero later.
         */
        CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0);
        CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0);

        est_refdblocks = (refdbytes >> SPA_MAXBLOCKSHIFT) +
                         (OSD_DNODE_EST_COUNT >> OSD_DNODE_EST_BLKSHIFT);
        est_usedobjs   = usedobjs + OSD_DNODE_EST_COUNT;

        /* Average space/dnode more than maximum dnode size, use max dnode
         * size to estimate free dnodes from adjusted free blocks count.
         * OSTs typically use more than one block dnode so this case applies. */
        if (est_usedobjs <= est_refdblocks * 2) {
                est_objs = nrblocks;

        /* Average space/dnode smaller than min dnode size (probably due to
         * metadnode compression), use min dnode size to estimate the number of
         * objects.
         * An MDT typically uses below 512 bytes/dnode so this case applies. */
        } else if (est_usedobjs >= (est_refdblocks << OSD_DNODE_MIN_BLKSHIFT)) {
                est_objs = nrblocks << OSD_DNODE_MIN_BLKSHIFT;

                /* Between the extremes, we try to use the average size of
                 * existing dnodes to compute the number of dnodes that fit
                 * into nrblocks:
                 *
                 * est_objs = nrblocks * (est_usedobjs / est_refblocks);
                 *
                 * but this may overflow 64 bits or become 0 if not handled well
                 *
                 * We know nrblocks is below (64 - 17 = 47) bits from
                 * SPA_MAXBLKSHIFT, and est_usedobjs is under 48 bits due to
                 * DN_MAX_OBJECT_SHIFT, which means that multiplying them may
                 * get as large as 2 ^ 95.
                 *
                 * We also know (est_usedobjs / est_refdblocks) is between 2 and
                 * 256, due to above checks, we can safely compute this first.
                 * We care more about accuracy on the MDT (many dnodes/block)
                 * which is good because this is where truncation errors are
                 * smallest.  This adds 8 bits to nrblocks so we can use 7 bits
                 * to compute a fixed-point fraction and nrblocks can still fit
                 * in 64 bits. */
        } else {
                unsigned dnodes_per_block = (est_usedobjs << 7)/est_refdblocks;

                est_objs = (nrblocks * dnodes_per_block) >> 7;
        }
        return est_objs;
}

static int osd_objset_statfs(struct objset *os, struct obd_statfs *osfs)
{
        uint64_t refdbytes, availbytes, usedobjs, availobjs;
        uint64_t est_availobjs;
        uint64_t reserved;

        dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs,
                         &availobjs);

        /*
         * ZFS allows multiple block sizes.  For statfs, Linux makes no
         * proper distinction between bsize and frsize.  For calculations
         * of free and used blocks incorrectly uses bsize instead of frsize,
         * but bsize is also used as the optimal blocksize.  We return the
         * largest possible block size as IO size for the optimum performance
         * and scale the free and used blocks count appropriately.
         */
        osfs->os_bsize = 1ULL << SPA_MAXBLOCKSHIFT;

        osfs->os_blocks = (refdbytes + availbytes) >> SPA_MAXBLOCKSHIFT;
        osfs->os_bfree = availbytes >> SPA_MAXBLOCKSHIFT;
        osfs->os_bavail = osfs->os_bfree; /* no extra root reservation */

        /* Take replication (i.e. number of copies) into account */
        osfs->os_bavail /= os->os_copies;

        /*
         * Reserve some space so we don't run into ENOSPC due to grants not
         * accounting for metadata overhead in ZFS, and to avoid fragmentation.
         * Rather than report this via os_bavail (which makes users unhappy if
         * they can't fill the filesystem 100%), reduce os_blocks as well.
         *
         * Reserve 0.78% of total space, at least 4MB for small filesystems,
         * for internal files to be created/unlinked when space is tight.
         */
        CLASSERT(OSD_STATFS_RESERVED_BLKS > 0);
        if (likely(osfs->os_blocks >=
                        OSD_STATFS_RESERVED_BLKS << OSD_STATFS_RESERVED_SHIFT))
                reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT;
        else
                reserved = OSD_STATFS_RESERVED_BLKS;

        osfs->os_blocks -= reserved;
        osfs->os_bfree  -= MIN(reserved, osfs->os_bfree);
        osfs->os_bavail -= MIN(reserved, osfs->os_bavail);

        /*
         * The availobjs value returned from dmu_objset_space() is largely
         * useless, since it reports the number of objects that might
         * theoretically still fit into the dataset, independent of minor
         * issues like how much space is actually available in the pool.
         * Compute a better estimate in udmu_objs_count_estimate().
         */
        est_availobjs = osd_objs_count_estimate(refdbytes, usedobjs,
                                                osfs->os_bfree);

        osfs->os_ffree = min(availobjs, est_availobjs);
        osfs->os_files = osfs->os_ffree + usedobjs;

        /* ZFS XXX: fill in backing dataset FSID/UUID
           memcpy(osfs->os_fsid, .... );*/

        /* We're a zfs filesystem. */
        osfs->os_type = UBERBLOCK_MAGIC;

        /* ZFS XXX: fill in appropriate OS_STATE_{DEGRADED,READONLY} flags
           osfs->os_state = vf_to_stf(vfsp->vfs_flag);
           if (sb->s_flags & MS_RDONLY)
           osfs->os_state = OS_STATE_READONLY;
         */

        osfs->os_namelen = MAXNAMELEN;
        osfs->os_maxbytes = OBD_OBJECT_EOF;

        return 0;
}

/*
 * Concurrency: shouldn't matter.
 */
int osd_statfs(const struct lu_env *env, struct dt_device *d,
               struct obd_statfs *osfs)
{
        struct osd_device *osd = osd_dt_dev(d);
        int                rc;
        ENTRY;

        rc = osd_objset_statfs(osd->od_os, osfs);
        if (unlikely(rc != 0))
                RETURN(rc);

        osfs->os_bavail -= min_t(u64,
                                 OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize,
                                 osfs->os_bavail);
        RETURN(0);
}

static int osd_blk_insert_cost(void)
{
        int max_blockshift, nr_blkptrshift;

        /* max_blockshift is the log2 of the number of blocks needed to reach
         * the maximum filesize (that's to say 2^64) */
        max_blockshift = DN_MAX_OFFSET_SHIFT - SPA_MAXBLOCKSHIFT;

        /* nr_blkptrshift is the log2 of the number of block pointers that can
         * be stored in an indirect block */
        CLASSERT(DN_MAX_INDBLKSHIFT > SPA_BLKPTRSHIFT);
        nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT;

        /* max_blockshift / nr_blkptrshift is thus the maximum depth of the
         * tree. We add +1 for rounding purpose.
         * The tree depth times the indirect block size gives us the maximum
         * cost of inserting a block in the tree */
        return (max_blockshift / nr_blkptrshift + 1) * (1<<DN_MAX_INDBLKSHIFT);
}

/*
 * 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 *osd = osd_dt_dev(dev);

        /*
         * XXX should be taken from not-yet-existing fs abstraction layer.
         */
        param->ddp_max_name_len = MAXNAMELEN;
        param->ddp_max_nlink    = 1 << 31; /* it's 8byte on a disk */
        param->ddp_block_shift  = 12; /* XXX */
        param->ddp_mount_type   = LDD_MT_ZFS;

        param->ddp_mntopts      = MNTOPT_USERXATTR;
        if (osd->od_posix_acl)
                param->ddp_mntopts |= MNTOPT_ACL;
        param->ddp_max_ea_size  = DXATTR_MAX_ENTRY_SIZE;

        /* for maxbytes, report same value as ZPL */
        param->ddp_maxbytes     = MAX_LFS_FILESIZE;

        /* Default reserved fraction of the available space that should be kept
         * for error margin. Unfortunately, there are many factors that can
         * impact the overhead with zfs, so let's be very cautious for now and
         * reserve 20% of the available space which is not given out as grant.
         * This tunable can be changed on a live system via procfs if needed. */
        param->ddp_grant_reserved = 20;

        /* inodes are dynamically allocated, so we report the per-inode space
         * consumption to upper layers. This static value is not really accurate
         * and we should use the same logic as in udmu_objset_statfs() to
         * estimate the real size consumed by an object */
        param->ddp_inodespace = OSD_DNODE_EST_COUNT;
        /* per-fragment overhead to be used by the client code */
        param->ddp_grant_frag = osd_blk_insert_cost();
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_sync(const struct lu_env *env, struct dt_device *d)
{
        struct osd_device  *osd = osd_dt_dev(d);
        CDEBUG(D_CACHE, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME);
        txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
        CDEBUG(D_CACHE, "synced OSD %s\n", LUSTRE_OSD_ZFS_NAME);
        return 0;
}

static int osd_commit_async(const struct lu_env *env, struct dt_device *dev)
{
        struct osd_device *osd = osd_dt_dev(dev);
        tx_state_t        *tx = &dmu_objset_pool(osd->od_os)->dp_tx;
        uint64_t           txg;

        mutex_enter(&tx->tx_sync_lock);
        txg = tx->tx_open_txg + 1;
        if (tx->tx_quiesce_txg_waiting < txg) {
                tx->tx_quiesce_txg_waiting = txg;
                cv_broadcast(&tx->tx_quiesce_more_cv);
        }
        mutex_exit(&tx->tx_sync_lock);

        return 0;
}

/*
 * Concurrency: shouldn't matter.
 */
static int osd_ro(const struct lu_env *env, struct dt_device *d)
{
        struct osd_device  *osd = osd_dt_dev(d);
        ENTRY;

        CERROR("%s: *** setting device %s read-only ***\n",
               osd->od_svname, LUSTRE_OSD_ZFS_NAME);
        osd->od_rdonly = 1;
        spa_freeze(dmu_objset_spa(osd->od_os));

        RETURN(0);
}

static 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_sync                = osd_sync,
        .dt_commit_async        = osd_commit_async,
        .dt_ro                  = osd_ro,
};

/*
 * DMU OSD device type methods
 */
static int osd_type_init(struct lu_device_type *t)
{
        LU_CONTEXT_KEY_INIT(&osd_key);
        return lu_context_key_register(&osd_key);
}

static void osd_type_fini(struct lu_device_type *t)
{
        lu_context_key_degister(&osd_key);
}

static void *osd_key_init(const struct lu_context *ctx,
                          struct lu_context_key *key)
{
        struct osd_thread_info *info;

        OBD_ALLOC_PTR(info);
        if (info != NULL)
                info->oti_env = container_of(ctx, struct lu_env, le_ctx);
        else
                info = ERR_PTR(-ENOMEM);
        return info;
}

static void osd_key_fini(const struct lu_context *ctx,
                         struct lu_context_key *key, void *data)
{
        struct osd_thread_info *info = data;

        OBD_FREE_PTR(info);
}

static void osd_key_exit(const struct lu_context *ctx,
                         struct lu_context_key *key, void *data)
{
        struct osd_thread_info *info = data;

        memset(info, 0, sizeof(*info));
}

struct lu_context_key osd_key = {
        .lct_tags = LCT_DT_THREAD | LCT_MD_THREAD | LCT_MG_THREAD | LCT_LOCAL,
        .lct_init = osd_key_init,
        .lct_fini = osd_key_fini,
        .lct_exit = osd_key_exit
};

static void osd_fid_fini(const struct lu_env *env, struct osd_device *osd)
{
        if (osd->od_cl_seq == NULL)
                return;

        seq_client_fini(osd->od_cl_seq);
        OBD_FREE_PTR(osd->od_cl_seq);
        osd->od_cl_seq = NULL;
}

static int osd_shutdown(const struct lu_env *env, struct osd_device *o)
{
        ENTRY;

        /* shutdown quota slave instance associated with the device */
        if (o->od_quota_slave != NULL) {
                qsd_fini(env, o->od_quota_slave);
                o->od_quota_slave = NULL;
        }

        osd_fid_fini(env, o);

        RETURN(0);
}

static void osd_xattr_changed_cb(void *arg, uint64_t newval)
{
        struct osd_device *osd = arg;

        osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA);
}

static int osd_objset_open(struct osd_device *o)
{
        uint64_t        version = ZPL_VERSION;
        uint64_t        sa_obj;
        int             rc;
        ENTRY;

        rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, B_FALSE, o, &o->od_os);
        if (rc) {
                o->od_os = NULL;
                goto out;
        }

        /* Check ZFS version */
        rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
                         ZPL_VERSION_STR, 8, 1, &version);
        if (rc) {
                CERROR("%s: Error looking up ZPL VERSION\n", o->od_mntdev);
                /*
                 * We can't return ENOENT because that would mean the objset
                 * didn't exist.
                 */
                GOTO(out, rc = -EIO);
        }

        rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ,
                         ZFS_SA_ATTRS, 8, 1, &sa_obj);
        if (rc)
                GOTO(out, rc);

        rc = -sa_setup(o->od_os, sa_obj, zfs_attr_table,
                       ZPL_END, &o->z_attr_table);
        if (rc)
                GOTO(out, rc);

        rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_ROOT_OBJ,
                         8, 1, &o->od_rootid);
        if (rc) {
                CERROR("%s: lookup for root failed: rc = %d\n",
                        o->od_svname, rc);
                GOTO(out, rc);
        }

        /* Check that user/group usage tracking is supported */
        if (!dmu_objset_userused_enabled(o->od_os) ||
            DMU_USERUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED ||
            DMU_GROUPUSED_DNODE(o->od_os)->dn_type != DMU_OT_USERGROUP_USED) {
                CERROR("%s: Space accounting not supported by this target, "
                        "aborting\n", o->od_svname);
                GOTO(out, -ENOTSUPP);
        }

out:
        if (rc != 0 && o->od_os != NULL)
                dmu_objset_disown(o->od_os, o);

        RETURN(rc);
}

static int osd_mount(const struct lu_env *env,
                     struct osd_device *o, struct lustre_cfg *cfg)
{
        struct dsl_dataset      *ds;
        char                    *mntdev = lustre_cfg_string(cfg, 1);
        char                    *svname = lustre_cfg_string(cfg, 4);
        dmu_buf_t               *rootdb;
        dsl_pool_t              *dp;
        const char              *opts;
        int                      rc;
        ENTRY;

        if (o->od_os != NULL)
                RETURN(0);

        if (mntdev == NULL || svname == NULL)
                RETURN(-EINVAL);

        rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev));
        if (rc >= sizeof(o->od_mntdev))
                RETURN(-E2BIG);

        rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname));
        if (rc >= sizeof(o->od_svname))
                RETURN(-E2BIG);

        if (server_name_is_ost(o->od_svname))
                o->od_is_ost = 1;

        rc = osd_objset_open(o);
        if (rc) {
                CERROR("%s: can't open objset %s: rc = %d\n", o->od_svname,
                        o->od_mntdev, rc);
                RETURN(rc);
        }

        ds = dmu_objset_ds(o->od_os);
        dp = dmu_objset_pool(o->od_os);
        LASSERT(ds);
        LASSERT(dp);
        dsl_pool_config_enter(dp, FTAG);
        rc = dsl_prop_register(ds, "xattr", osd_xattr_changed_cb, o);
        dsl_pool_config_exit(dp, FTAG);
        if (rc)
                CWARN("%s: can't register xattr callback, ignore: rc=%d\n",
                      o->od_svname, rc);

        rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb);
        if (rc) {
                CERROR("%s: obj2dbuf() failed: rc = %d\n", o->od_svname, rc);
                dmu_objset_disown(o->od_os, o);
                o->od_os = NULL;
                RETURN(rc);
        }

        o->od_root = rootdb->db_object;
        sa_buf_rele(rootdb, osd_obj_tag);

        /* 1. initialize oi before any file create or file open */
        rc = osd_oi_init(env, o);
        if (rc)
                GOTO(err, rc);

        rc = lu_site_init(&o->od_site, osd2lu_dev(o));
        if (rc)
                GOTO(err, rc);
        o->od_site.ls_bottom_dev = osd2lu_dev(o);

        rc = lu_site_init_finish(&o->od_site);
        if (rc)
                GOTO(err, rc);

        rc = osd_convert_root_to_new_seq(env, o);
        if (rc)
                GOTO(err, rc);

        /* Use our own ZAP for inode accounting by default, this can be changed
         * via procfs to estimate the inode usage from the block usage */
        o->od_quota_iused_est = 0;

        rc = osd_procfs_init(o, o->od_svname);
        if (rc)
                GOTO(err, rc);

        o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o);

        /* initialize quota slave instance */
        o->od_quota_slave = qsd_init(env, o->od_svname, &o->od_dt_dev,
                                     o->od_proc_entry);
        if (IS_ERR(o->od_quota_slave)) {
                rc = PTR_ERR(o->od_quota_slave);
                o->od_quota_slave = NULL;
                GOTO(err, rc);
        }

        /* parse mount option "noacl", and enable ACL by default */
        opts = lustre_cfg_string(cfg, 3);
        if (opts == NULL || strstr(opts, "noacl") == NULL)
                o->od_posix_acl = 1;

err:
        RETURN(rc);
}

static void osd_umount(const struct lu_env *env, struct osd_device *o)
{
        ENTRY;

        if (atomic_read(&o->od_zerocopy_alloc))
                CERROR("%s: lost %d allocated page(s)\n", o->od_svname,
                       atomic_read(&o->od_zerocopy_alloc));
        if (atomic_read(&o->od_zerocopy_loan))
                CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname,
                       atomic_read(&o->od_zerocopy_loan));
        if (atomic_read(&o->od_zerocopy_pin))
                CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname,
                       atomic_read(&o->od_zerocopy_pin));

        if (o->od_os != NULL) {
                /* force a txg sync to get all commit callbacks */
                txg_wait_synced(dmu_objset_pool(o->od_os), 0ULL);

                /* close the object set */
                dmu_objset_disown(o->od_os, o);

                o->od_os = NULL;
        }

        EXIT;
}

static int osd_device_init0(const struct lu_env *env,
                            struct osd_device *o,
                            struct lustre_cfg *cfg)
{
        struct lu_device        *l = osd2lu_dev(o);
        int                      rc;

        /* if the module was re-loaded, env can loose its keys */
        rc = lu_env_refill((struct lu_env *) env);
        if (rc)
                GOTO(out, rc);

        l->ld_ops = &osd_lu_ops;
        o->od_dt_dev.dd_ops = &osd_dt_ops;

out:
        RETURN(rc);
}

static struct lu_device *osd_device_fini(const struct lu_env *env,
                                         struct lu_device *dev);

static struct lu_device *osd_device_alloc(const struct lu_env *env,
                                          struct lu_device_type *type,
                                          struct lustre_cfg *cfg)
{
        struct osd_device *dev;
        int                rc;

        OBD_ALLOC_PTR(dev);
        if (dev == NULL)
                return ERR_PTR(-ENOMEM);

        rc = dt_device_init(&dev->od_dt_dev, type);
        if (rc == 0) {
                rc = osd_device_init0(env, dev, cfg);
                if (rc == 0) {
                        rc = osd_mount(env, dev, cfg);
                        if (rc)
                                osd_device_fini(env, osd2lu_dev(dev));
                }
                if (rc)
                        dt_device_fini(&dev->od_dt_dev);
        }

        if (unlikely(rc != 0))
                OBD_FREE_PTR(dev);

        return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc);
}

static struct lu_device *osd_device_free(const struct lu_env *env,
                                         struct lu_device *d)
{
        struct osd_device *o = osd_dev(d);
        ENTRY;

        /* XXX: make osd top device in order to release reference */
        d->ld_site->ls_top_dev = d;
        lu_site_purge(env, d->ld_site, -1);
        if (!cfs_hash_is_empty(d->ld_site->ls_obj_hash)) {
                LIBCFS_DEBUG_MSG_DATA_DECL(msgdata, D_ERROR, NULL);
                lu_site_print(env, d->ld_site, &msgdata, lu_cdebug_printer);
        }
        lu_site_fini(&o->od_site);
        dt_device_fini(&o->od_dt_dev);
        OBD_FREE_PTR(o);

        RETURN (NULL);
}

static struct lu_device *osd_device_fini(const struct lu_env *env,
                                         struct lu_device *d)
{
        struct osd_device *o = osd_dev(d);
        struct dsl_dataset *ds;
        int                rc;
        ENTRY;


        osd_shutdown(env, o);
        osd_oi_fini(env, o);

        if (o->od_os) {
                ds = dmu_objset_ds(o->od_os);
                rc = dsl_prop_unregister(ds, "xattr", osd_xattr_changed_cb, o);
                if (rc)
                        CERROR("%s: dsl_prop_unregister xattr error %d\n",
                                o->od_svname, rc);
                if (o->arc_prune_cb != NULL) {
                        arc_remove_prune_callback(o->arc_prune_cb);
                        o->arc_prune_cb = NULL;
                }
                osd_sync(env, lu2dt_dev(d));
                txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os)));
        }

        rc = osd_procfs_fini(o);
        if (rc) {
                CERROR("proc fini error %d\n", rc);
                RETURN(ERR_PTR(rc));
        }

        if (o->od_os)
                osd_umount(env, o);

        RETURN(NULL);
}

static int osd_device_init(const struct lu_env *env, struct lu_device *d,
                           const char *name, struct lu_device *next)
{
        return 0;
}

/*
 * To be removed, setup is performed by osd_device_{init,alloc} and
 * cleanup is performed by osd_device_{fini,free).
 */
static int osd_process_config(const struct lu_env *env,
                              struct lu_device *d, struct lustre_cfg *cfg)
{
        struct osd_device       *o = osd_dev(d);
        int                     rc;
        ENTRY;

        switch(cfg->lcfg_command) {
        case LCFG_SETUP:
                rc = osd_mount(env, o, cfg);
                break;
        case LCFG_CLEANUP:
                rc = osd_shutdown(env, o);
                break;
        case LCFG_PARAM: {
                LASSERT(&o->od_dt_dev);
                rc = class_process_proc_param(PARAM_OSD, lprocfs_osd_obd_vars,
                                              cfg, &o->od_dt_dev);
                if (rc > 0 || rc == -ENOSYS)
                        rc = class_process_proc_param(PARAM_OST,
                                                      lprocfs_osd_obd_vars,
                                                      cfg, &o->od_dt_dev);
                break;
        }
        default:
                rc = -ENOTTY;
        }

        RETURN(rc);
}

static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d)
{
        struct osd_device       *osd = osd_dev(d);
        int                      rc = 0;
        ENTRY;

        if (osd->od_quota_slave == NULL)
                RETURN(0);

        /* start qsd instance on recovery completion, this notifies the quota
         * slave code that we are about to process new requests now */
        rc = qsd_start(env, osd->od_quota_slave);
        RETURN(rc);
}

/*
 * we use exports to track all osd users
 */
static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp,
                           struct obd_device *obd, struct obd_uuid *cluuid,
                           struct obd_connect_data *data, void *localdata)
{
        struct osd_device    *osd = osd_dev(obd->obd_lu_dev);
        struct lustre_handle  conn;
        int                   rc;
        ENTRY;

        CDEBUG(D_CONFIG, "connect #%d\n", osd->od_connects);

        rc = class_connect(&conn, obd, cluuid);
        if (rc)
                RETURN(rc);

        *exp = class_conn2export(&conn);

        spin_lock(&obd->obd_dev_lock);
        osd->od_connects++;
        spin_unlock(&obd->obd_dev_lock);

        RETURN(0);
}

/*
 * once last export (we don't count self-export) disappeared
 * osd can be released
 */
static int osd_obd_disconnect(struct obd_export *exp)
{
        struct obd_device *obd = exp->exp_obd;
        struct osd_device *osd = osd_dev(obd->obd_lu_dev);
        int                rc, release = 0;
        ENTRY;

        /* Only disconnect the underlying layers on the final disconnect. */
        spin_lock(&obd->obd_dev_lock);
        osd->od_connects--;
        if (osd->od_connects == 0)
                release = 1;
        spin_unlock(&obd->obd_dev_lock);

        rc = class_disconnect(exp); /* bz 9811 */

        if (rc == 0 && release)
                class_manual_cleanup(obd);
        RETURN(rc);
}

static int osd_fid_init(const struct lu_env *env, struct osd_device *osd)
{
        struct seq_server_site  *ss = osd_seq_site(osd);
        int                     rc;
        ENTRY;

        if (osd->od_is_ost || osd->od_cl_seq != NULL)
                RETURN(0);

        if (unlikely(ss == NULL))
                RETURN(-ENODEV);

        OBD_ALLOC_PTR(osd->od_cl_seq);
        if (osd->od_cl_seq == NULL)
                RETURN(-ENOMEM);

        rc = seq_client_init(osd->od_cl_seq, NULL, LUSTRE_SEQ_METADATA,
                             osd->od_svname, ss->ss_server_seq);

        if (rc != 0) {
                OBD_FREE_PTR(osd->od_cl_seq);
                osd->od_cl_seq = NULL;
        }

        RETURN(rc);
}

static int osd_prepare(const struct lu_env *env, struct lu_device *pdev,
                       struct lu_device *dev)
{
        struct osd_device       *osd = osd_dev(dev);
        int                      rc = 0;
        ENTRY;

        if (osd->od_quota_slave != NULL) {
                /* set up quota slave objects */
                rc = qsd_prepare(env, osd->od_quota_slave);
                if (rc != 0)
                        RETURN(rc);
        }

        rc = osd_fid_init(env, osd);

        RETURN(rc);
}

struct lu_device_operations osd_lu_ops = {
        .ldo_object_alloc       = osd_object_alloc,
        .ldo_process_config     = osd_process_config,
        .ldo_recovery_complete  = osd_recovery_complete,
        .ldo_prepare            = osd_prepare,
};

static void osd_type_start(struct lu_device_type *t)
{
}

static void osd_type_stop(struct lu_device_type *t)
{
}

int osd_fid_alloc(const struct lu_env *env, struct obd_export *exp,
                  struct lu_fid *fid, struct md_op_data *op_data)
{
        struct osd_device *osd = osd_dev(exp->exp_obd->obd_lu_dev);

        return seq_client_alloc_fid(env, osd->od_cl_seq, fid);
}

static struct lu_device_type_operations osd_device_type_ops = {
        .ldto_init              = osd_type_init,
        .ldto_fini              = osd_type_fini,

        .ldto_start             = osd_type_start,
        .ldto_stop              = osd_type_stop,

        .ldto_device_alloc      = osd_device_alloc,
        .ldto_device_free       = osd_device_free,

        .ldto_device_init       = osd_device_init,
        .ldto_device_fini       = osd_device_fini
};

static struct lu_device_type osd_device_type = {
        .ldt_tags     = LU_DEVICE_DT,
        .ldt_name     = LUSTRE_OSD_ZFS_NAME,
        .ldt_ops      = &osd_device_type_ops,
        .ldt_ctx_tags = LCT_LOCAL
};


static struct obd_ops osd_obd_device_ops = {
        .o_owner       = THIS_MODULE,
        .o_connect      = osd_obd_connect,
        .o_disconnect   = osd_obd_disconnect,
        .o_fid_alloc    = osd_fid_alloc
};

int __init osd_init(void)
{
        int rc;

        rc = osd_options_init();
        if (rc)
                return rc;

        rc = lu_kmem_init(osd_caches);
        if (rc)
                return rc;

        rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL,
                                 LUSTRE_OSD_ZFS_NAME, &osd_device_type);
        if (rc)
                lu_kmem_fini(osd_caches);
        return rc;
}

void __exit osd_exit(void)
{
        class_unregister_type(LUSTRE_OSD_ZFS_NAME);
        lu_kmem_fini(osd_caches);
}

extern unsigned int osd_oi_count;
CFS_MODULE_PARM(osd_oi_count, "i", int, 0444,
                "Number of Object Index containers to be created, "
                "it's only valid for new filesystem.");

MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")");
MODULE_LICENSE("GPL");

cfs_module(osd, LUSTRE_VERSION_STRING, osd_init, osd_exit);