X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_handler.c;h=dfbd45f6b49a5e69247254099bd1490c97950429;hb=26995a3588e72d65f609b12772d24a879c9deb7f;hp=4efe29e221e5f35e9304e4ba6fe6a28e394b1093;hpb=79b4ae9139cba69b2a81db25276cb2c0ae12a31c;p=fs%2Flustre-release.git diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c index 4efe29e..dfbd45f 100644 --- a/lustre/osd-zfs/osd_handler.c +++ b/lustre/osd-zfs/osd_handler.c @@ -28,7 +28,7 @@ * Use is subject to license terms. */ /* - * Copyright (c) 2011, 2012 Whamcloud, Inc. + * Copyright (c) 2012, 2013, Intel Corporation. * Use is subject to license terms. * */ @@ -44,9 +44,6 @@ * Author: Johann Lombardi */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_OSD #include @@ -57,6 +54,8 @@ #include #include #include +#include +#include #include "osd_internal.h" @@ -75,10 +74,8 @@ struct lu_context_key osd_key; -static char *root_tag = "osd_mount, rootdb"; - /* Slab for OSD object allocation */ -cfs_mem_cache_t *osd_object_kmem; +struct kmem_cache *osd_object_kmem; static struct lu_kmem_descr osd_caches[] = { { @@ -131,6 +128,7 @@ 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; @@ -148,14 +146,22 @@ static void osd_trans_commit_cb(void *cb_data, int error) dt_txn_hook_commit(th); /* call per-transaction callbacks if any */ - cfs_list_for_each_entry_safe(dcb, tmp, &oh->ot_dcb_list, dcb_linkage) + 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); - OBD_FREE_PTR(oh); + thandle_put(&oh->ot_super); EXIT; } @@ -165,7 +171,7 @@ 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); - cfs_list_add(&dcb->dcb_linkage, &oh->ot_dcb_list); + list_add(&dcb->dcb_linkage, &oh->ot_dcb_list); return 0; } @@ -200,11 +206,10 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d, 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_dt_dev.dd_lu_dev.ld_obd->obd_name); + "grant_ratio is too low.\n", osd->od_svname); else CERROR("%s: can't assign tx: rc = %d\n", - osd->od_dt_dev.dd_lu_dev.ld_obd->obd_name, rc); + osd->od_svname, rc); } else { /* add commit callback */ dmu_tx_callback_register(oh->ot_tx, osd_trans_commit_cb, oh); @@ -220,7 +225,8 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d, /* * Concurrency: shouldn't matter. */ -static int osd_trans_stop(const struct lu_env *env, struct thandle *th) +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; @@ -234,10 +240,23 @@ static int osd_trans_stop(const struct lu_env *env, struct thandle *th) LASSERT(oh->ot_tx); dmu_tx_abort(oh->ot_tx); osd_object_sa_dirty_rele(oh); - OBD_FREE_PTR(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", @@ -250,7 +269,7 @@ static int osd_trans_stop(const struct lu_env *env, struct thandle *th) dmu_tx_commit(oh->ot_tx); if (th->th_sync) - txg_wait_synced(dmu_objset_pool(osd->od_objset.os), txg); + txg_wait_synced(dmu_objset_pool(osd->od_os), txg); RETURN(rc); } @@ -264,7 +283,7 @@ static struct thandle *osd_trans_create(const struct lu_env *env, dmu_tx_t *tx; ENTRY; - tx = dmu_tx_create(osd->od_objset.os); + tx = dmu_tx_create(osd->od_os); if (tx == NULL) RETURN(ERR_PTR(-ENOMEM)); @@ -276,16 +295,162 @@ static struct thandle *osd_trans_create(const struct lu_env *env, } oh->ot_tx = tx; - CFS_INIT_LIST_HEAD(&oh->ot_dcb_list); - CFS_INIT_LIST_HEAD(&oh->ot_sa_list); - cfs_sema_init(&oh->ot_sa_lock, 1); + 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. */ @@ -296,15 +461,36 @@ int osd_statfs(const struct lu_env *env, struct dt_device *d, int rc; ENTRY; - rc = udmu_objset_statfs(&osd->od_objset, osfs); - if (rc) + rc = osd_objset_statfs(osd->od_os, osfs); + if (unlikely(rc != 0)) RETURN(rc); + osfs->os_bavail -= min_t(obd_size, 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<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_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 | MNTOPT_ACL; - param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE; + 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; + 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 @@ -339,7 +529,7 @@ static void osd_conf_get(const struct lu_env *env, * 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 = udmu_blk_insert_cost(); + param->ddp_grant_frag = osd_blk_insert_cost(); } /* @@ -349,16 +539,17 @@ static int osd_sync(const struct lu_env *env, struct dt_device *d) { struct osd_device *osd = osd_dt_dev(d); CDEBUG(D_HA, "syncing OSD %s\n", LUSTRE_OSD_ZFS_NAME); - txg_wait_synced(dmu_objset_pool(osd->od_objset.os), 0ULL); + txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL); 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_objset.os)->dp_tx; + 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; @@ -380,7 +571,7 @@ static int osd_ro(const struct lu_env *env, struct dt_device *d) 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_objset.os)); + spa_freeze(dmu_objset_spa(osd->od_os)); RETURN(0); } @@ -403,23 +594,6 @@ static int osd_init_capa_ctxt(const struct lu_env *env, struct dt_device *d, RETURN(0); } -static char *osd_label_get(const struct lu_env *env, const struct dt_device *d) -{ - struct osd_device *dev = osd_dt_dev(d); - int rc; - ENTRY; - - rc = -udmu_userprop_get_str(&dev->od_objset, DMU_OSD_SVNAME, - dev->od_svname, sizeof(dev->od_svname)); - if (rc != 0) { - if (rc == -EOVERFLOW) - CWARN("%s: buffer too small\n", dev->od_svname); - RETURN(NULL); - } - - RETURN(&dev->od_svname[0]); -} - static struct dt_device_operations osd_dt_ops = { .dt_root_get = osd_root_get, .dt_statfs = osd_statfs, @@ -486,73 +660,183 @@ struct lu_context_key osd_key = { 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; + } + 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) { - char *dev = lustre_cfg_string(cfg, 0); - struct lustre_mount_info *lmi; - struct lustre_sb_info *lsi; - dmu_buf_t *rootdb; - char *label; - int rc; + 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_objset.os != NULL) + if (o->od_os != NULL) RETURN(0); - lmi = server_get_mount(dev); - if (lmi == NULL) { - CERROR("Unknown mount point: '%s'\n", dev); - RETURN(-ENODEV); - } + if (mntdev == NULL || svname == NULL) + RETURN(-EINVAL); - lsi = s2lsi(lmi->lmi_sb); - dev = lsi->lsi_lmd->lmd_dev; + rc = strlcpy(o->od_mntdev, mntdev, sizeof(o->od_mntdev)); + if (rc >= sizeof(o->od_mntdev)) + RETURN(-E2BIG); - if (strlen(dev) >= sizeof(o->od_mntdev)) + rc = strlcpy(o->od_svname, svname, sizeof(o->od_svname)); + if (rc >= sizeof(o->od_svname)) RETURN(-E2BIG); - strcpy(o->od_mntdev, dev); + if (server_name_is_ost(o->od_svname)) + o->od_is_ost = 1; - rc = -udmu_objset_open(o->od_mntdev + 1, &o->od_objset); + rc = osd_objset_open(o); if (rc) { - CERROR("can't open objset %s: %d\n", o->od_mntdev, rc); + CERROR("%s: can't open objset %s: rc = %d\n", o->od_svname, + o->od_mntdev, rc); RETURN(rc); } - rc = __osd_obj2dbuf(env, o->od_objset.os, o->od_objset.root, - &rootdb, root_tag); + 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("udmu_obj2dbuf() failed with error %d\n", rc); - udmu_objset_close(&o->od_objset); + 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, root_tag); + 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); - label = osd_label_get(env, &o->od_dt_dev); - if (label == NULL) - GOTO(err, rc = -ENODEV); + 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, label); + 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); } @@ -561,18 +845,25 @@ static void osd_umount(const struct lu_env *env, struct osd_device *o) { ENTRY; - if (cfs_atomic_read(&o->od_zerocopy_alloc)) + if (atomic_read(&o->od_zerocopy_alloc)) CERROR("%s: lost %d allocated page(s)\n", o->od_svname, - cfs_atomic_read(&o->od_zerocopy_alloc)); - if (cfs_atomic_read(&o->od_zerocopy_loan)) + atomic_read(&o->od_zerocopy_alloc)); + if (atomic_read(&o->od_zerocopy_loan)) CERROR("%s: lost %d loaned abuf(s)\n", o->od_svname, - cfs_atomic_read(&o->od_zerocopy_loan)); - if (cfs_atomic_read(&o->od_zerocopy_pin)) + atomic_read(&o->od_zerocopy_loan)); + if (atomic_read(&o->od_zerocopy_pin)) CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname, - cfs_atomic_read(&o->od_zerocopy_pin)); + atomic_read(&o->od_zerocopy_pin)); - if (o->od_objset.os != NULL) - udmu_objset_close(&o->od_objset); + 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; } @@ -600,28 +891,36 @@ 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 *t, + struct lu_device_type *type, struct lustre_cfg *cfg) { - struct osd_device *o; - int rc; + struct osd_device *dev; + int rc; - OBD_ALLOC_PTR(o); - if (o == NULL) + OBD_ALLOC_PTR(dev); + if (dev == NULL) return ERR_PTR(-ENOMEM); - rc = dt_device_init(&o->od_dt_dev, t); + rc = dt_device_init(&dev->od_dt_dev, type); if (rc == 0) { - rc = osd_device_init0(env, o, cfg); + 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(&o->od_dt_dev); + dt_device_fini(&dev->od_dt_dev); } if (unlikely(rc != 0)) - OBD_FREE_PTR(o); + OBD_FREE_PTR(dev); - return rc == 0 ? osd2lu_dev(o) : ERR_PTR(rc); + return rc == 0 ? osd2lu_dev(dev) : ERR_PTR(rc); } static struct lu_device *osd_device_free(const struct lu_env *env, @@ -632,9 +931,13 @@ static struct lu_device *osd_device_free(const struct lu_env *env, cleanup_capa_hash(o->od_capa_hash); /* XXX: make osd top device in order to release reference */ - /*d->ld_site->ls_top_dev = d; + d->ld_site->ls_top_dev = d; lu_site_purge(env, d->ld_site, -1); - lu_site_fini(&o->od_site);*/ + 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); @@ -644,19 +947,27 @@ static struct lu_device *osd_device_free(const struct lu_env *env, static struct lu_device *osd_device_fini(const struct lu_env *env, struct lu_device *d) { - struct osd_device *o = osd_dev(d); - struct lustre_mount_info *lmi; - int rc; + 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_objset.os) { - arc_remove_prune_callback(o->arc_prune_cb); - o->arc_prune_cb = NULL; + 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_objset.os))); + txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os))); } rc = osd_procfs_fini(o); @@ -665,13 +976,9 @@ static struct lu_device *osd_device_fini(const struct lu_env *env, RETURN(ERR_PTR(rc)); } - if (o->od_objset.os) + if (o->od_os) osd_umount(env, o); - lmi = server_get_mount_2(o->od_svname); - LASSERT(lmi); - server_put_mount(lmi->lmi_name, lmi->lmi_mnt); - RETURN(NULL); } @@ -689,33 +996,112 @@ 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 err; + int rc; ENTRY; switch(cfg->lcfg_command) { case LCFG_SETUP: - err = osd_mount(env, o, cfg); + rc = osd_mount(env, o, cfg); break; case LCFG_CLEANUP: - err = osd_shutdown(env, o); + 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: - err = -ENOTTY; + rc = -ENOTTY; } - RETURN(err); + 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_prepare(const struct lu_env *env, struct lu_device *pdev, struct lu_device *dev) { - return 0; + 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); + + RETURN(rc); } struct lu_device_operations osd_lu_ops = { @@ -733,6 +1119,14 @@ 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, @@ -757,6 +1151,9 @@ static struct lu_device_type osd_device_type = { 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) @@ -771,8 +1168,7 @@ int __init osd_init(void) if (rc) return rc; - rc = class_register_type(&osd_obd_device_ops, NULL, - lprocfs_osd_module_vars, + 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);