X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_handler.c;h=7d1b579b2204e9d6588e61c918a7a287d3b7dacf;hp=9c09620b201f400fabc8019663f9042213724a96;hb=40daa59ac41f450b60b42eb2bb0ff42ebd3c998b;hpb=aa64476cef7d5f0050de20839925868013caa0b6 diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c index 9c09620..7d1b579 100644 --- a/lustre/osd-zfs/osd_handler.c +++ b/lustre/osd-zfs/osd_handler.c @@ -23,11 +23,10 @@ * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2016, Intel Corporation. + * Copyright (c) 2012, 2017, 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 @@ -39,7 +38,6 @@ #define DEBUG_SUBSYSTEM S_OSD -#include #include #include #include @@ -47,7 +45,7 @@ #include #include #include -#include +#include #include #include "osd_internal.h" @@ -67,6 +65,8 @@ struct lu_context_key osd_key; +static int osd_txg_sync_delay_us = -1; + /* Slab for OSD object allocation */ struct kmem_cache *osd_object_kmem; @@ -144,8 +144,6 @@ static void osd_trans_commit_cb(void *cb_data, int error) 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) { LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC, @@ -161,12 +159,12 @@ static void osd_trans_commit_cb(void *cb_data, int error) * 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); + qsd_op_end(NULL, osd->od_quota_slave_dt, &oh->ot_quota_trans); + if (osd->od_quota_slave_md != NULL) + qsd_op_end(NULL, osd->od_quota_slave_md, &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); EXIT; @@ -174,8 +172,8 @@ static void osd_trans_commit_cb(void *cb_data, int error) static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb) { - struct osd_thandle *oh = container_of0(th, struct osd_thandle, - ot_super); + 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); @@ -193,26 +191,30 @@ static int osd_trans_cb_add(struct thandle *th, struct dt_txn_commit_cb *dcb) static int osd_trans_start(const struct lu_env *env, struct dt_device *d, struct thandle *th) { - struct osd_thandle *oh; - int rc; + struct osd_device *osd = osd_dt_dev(d); + struct osd_thandle *oh; + int rc; + ENTRY; - oh = container_of0(th, struct osd_thandle, ot_super); + oh = container_of(th, struct osd_thandle, ot_super); LASSERT(oh); LASSERT(oh->ot_tx); rc = dt_txn_hook_start(env, d, th); - if (rc != 0) + if (rc != 0) { + CERROR("%s: dt_txn_hook_start failed: rc = %d\n", + osd->od_svname, rc); RETURN(rc); + } - if (oh->ot_write_commit && OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) + if (OBD_FAIL_CHECK(OBD_FAIL_OSD_TXN_START)) /* Unlike ldiskfs, ZFS checks for available space and returns * -ENOSPC when assigning txg */ - RETURN(-ENOSPC); + RETURN(-EIO); 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" @@ -224,17 +226,15 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d, /* 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); + osd_oti_get(env)->oti_in_trans = 1; lu_device_get(&d->dd_lu_dev); } RETURN(rc); } -static int osd_unlinked_object_free(struct osd_device *osd, uint64_t oid); - -static void osd_unlinked_list_emptify(struct osd_device *osd, +static void osd_unlinked_list_emptify(const struct lu_env *env, + struct osd_device *osd, struct list_head *list, bool free) { struct osd_object *obj; @@ -248,7 +248,7 @@ static void osd_unlinked_list_emptify(struct osd_device *osd, list_del_init(&obj->oo_unlinked_linkage); if (free) - (void)osd_unlinked_object_free(osd, oid); + (void)osd_unlinked_object_free(env, osd, oid); } } @@ -277,25 +277,32 @@ static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt, struct osd_device *osd = osd_dt_dev(th->th_dev); bool sync = (th->th_sync != 0); struct osd_thandle *oh; - struct list_head unlinked; + LIST_HEAD(unlinked); uint64_t txg; int rc; ENTRY; - oh = container_of0(th, struct osd_thandle, ot_super); - INIT_LIST_HEAD(&unlinked); + oh = container_of(th, struct osd_thandle, ot_super); list_splice_init(&oh->ot_unlinked_list, &unlinked); + + osd_oti_get(env)->oti_ins_cache_depth--; /* reset OI cache for safety */ - osd_oti_get(env)->oti_ins_cache_used = 0; + if (osd_oti_get(env)->oti_ins_cache_depth == 0) + osd_oti_get(env)->oti_ins_cache_used = 0; if (oh->ot_assigned == 0) { LASSERT(oh->ot_tx); + CDEBUG(D_OTHER, "%s: transaction is aborted\n", osd->od_svname); + osd_trans_stop_cb(oh, th->th_result); dmu_tx_abort(oh->ot_tx); - osd_object_sa_dirty_rele(oh); - osd_unlinked_list_emptify(osd, &unlinked, false); + osd_object_sa_dirty_rele(env, oh); + osd_unlinked_list_emptify(env, osd, &unlinked, false); /* 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); + qsd_op_end(env, osd->od_quota_slave_dt, &oh->ot_quota_trans); + if (osd->od_quota_slave_md != NULL) + qsd_op_end(env, osd->od_quota_slave_md, + &oh->ot_quota_trans); OBD_FREE_PTR(oh); RETURN(0); } @@ -310,15 +317,20 @@ static int osd_trans_stop(const struct lu_env *env, struct dt_device *dt, LASSERT(oh->ot_tx); txg = oh->ot_tx->tx_txg; - osd_object_sa_dirty_rele(oh); + osd_object_sa_dirty_rele(env, oh); /* XXX: Once dmu_tx_commit() called, oh/th could have been freed * by osd_trans_commit_cb already. */ dmu_tx_commit(oh->ot_tx); + osd_oti_get(env)->oti_in_trans = 0; - osd_unlinked_list_emptify(osd, &unlinked, true); + osd_unlinked_list_emptify(env, osd, &unlinked, true); - if (sync) - txg_wait_synced(dmu_objset_pool(osd->od_os), txg); + if (sync) { + if (osd_txg_sync_delay_us < 0) + txg_wait_synced(dmu_objset_pool(osd->od_os), txg); + else + udelay(osd_txg_sync_delay_us); + } RETURN(rc); } @@ -356,12 +368,13 @@ static struct thandle *osd_trans_create(const struct lu_env *env, INIT_LIST_HEAD(&oh->ot_stop_dcb_list); INIT_LIST_HEAD(&oh->ot_unlinked_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; + + osd_oti_get(env)->oti_ins_cache_depth++; + RETURN(th); } @@ -394,8 +407,8 @@ uint64_t osd_objs_count_estimate(uint64_t usedbytes, uint64_t usedobjs, * gradually disappears as the number of real dnodes grows. It also * avoids the need to check for divide-by-zero computing dn_per_block. */ - CLASSERT(OSD_DNODE_MIN_BLKSHIFT > 0); - CLASSERT(OSD_DNODE_EST_BLKSHIFT > 0); + BUILD_BUG_ON(OSD_DNODE_MIN_BLKSHIFT <= 0); + BUILD_BUG_ON(OSD_DNODE_EST_BLKSHIFT <= 0); est_usedblocks = ((OSD_DNODE_EST_COUNT << OSD_DNODE_EST_BLKSHIFT) + usedbytes) >> est_maxblockshift; @@ -496,7 +509,7 @@ static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs) * Reserve 0.78% of total space, at least 16MB for small filesystems, * for internal files to be created/unlinked when space is tight. */ - CLASSERT(OSD_STATFS_RESERVED_SIZE > 0); + BUILD_BUG_ON(OSD_STATFS_RESERVED_SIZE <= 0); reserved = OSD_STATFS_RESERVED_SIZE >> bshift; if (likely(osfs->os_blocks >= reserved << OSD_STATFS_RESERVED_SHIFT)) reserved = osfs->os_blocks >> OSD_STATFS_RESERVED_SHIFT; @@ -526,7 +539,7 @@ static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs) if (!spa_writeable(dmu_objset_spa(os)) || osd->od_dev_set_rdonly || osd->od_prop_rdonly) - osfs->os_state |= OS_STATE_READONLY; + osfs->os_state |= OS_STATFS_READONLY; return 0; } @@ -535,18 +548,25 @@ static int osd_objset_statfs(struct osd_device *osd, struct obd_statfs *osfs) * Concurrency: shouldn't matter. */ int osd_statfs(const struct lu_env *env, struct dt_device *d, - struct obd_statfs *osfs) + struct obd_statfs *osfs, struct obd_statfs_info *info) { - int rc; + struct osd_device *osd = osd_dt_dev(d); + int rc; ENTRY; - rc = osd_objset_statfs(osd_dt_dev(d), osfs); + rc = osd_objset_statfs(osd, osfs); if (unlikely(rc != 0)) RETURN(rc); osfs->os_bavail -= min_t(u64, OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize, osfs->os_bavail); + + /* ZFS does not support reporting nonrotional status yet, so return + * flag only if user has set nonrotational. + */ + osfs->os_state |= osd->od_nonrotational ? OS_STATFS_NONROT : 0; + RETURN(0); } @@ -561,7 +581,7 @@ static int osd_blk_insert_cost(struct osd_device *osd) /* 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); + BUILD_BUG_ON(DN_MAX_INDBLKSHIFT <= SPA_BLKPTRSHIFT); nr_blkptrshift = DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT; /* max_blockshift / nr_blkptrshift is thus the maximum depth of the @@ -591,7 +611,8 @@ static void osd_conf_get(const struct lu_env *env, param->ddp_mntopts = MNTOPT_USERXATTR; if (osd->od_posix_acl) param->ddp_mntopts |= MNTOPT_ACL; - param->ddp_max_ea_size = DXATTR_MAX_ENTRY_SIZE; + /* Previously DXATTR_MAX_ENTRY_SIZE */ + param->ddp_max_ea_size = OBD_MAX_EA_SIZE; /* for maxbytes, report same value as ZPL */ param->ddp_maxbytes = MAX_LFS_FILESIZE; @@ -612,6 +633,20 @@ static void osd_conf_get(const struct lu_env *env, param->ddp_max_extent_blks = (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)); param->ddp_extent_tax = osd_blk_insert_cost(osd); + + /* Preferred RPC size for efficient disk IO. 1MB shows good + * all-around performance for ZFS, but use blocksize (recordsize) + * by default if larger to avoid read-modify-write. */ + if (osd->od_max_blksz > ONE_MB_BRW_SIZE) + param->ddp_brw_size = osd->od_max_blksz; + else + param->ddp_brw_size = ONE_MB_BRW_SIZE; + +#ifdef HAVE_DMU_OFFSET_NEXT + param->ddp_has_lseek_data_hole = osd->od_sync_on_lseek; +#else + param->ddp_has_lseek_data_hole = false; +#endif } /* @@ -663,17 +698,38 @@ static int osd_ro(const struct lu_env *env, struct dt_device *d) 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, +/* reserve or free quota for some operation */ +static int osd_reserve_or_free_quota(const struct lu_env *env, + struct dt_device *dev, + struct lquota_id_info *qi) +{ + struct osd_device *osd = osd_dt_dev(dev); + struct qsd_instance *qsd = NULL; + int rc; + + ENTRY; + + if (qi->lqi_is_blk) + qsd = osd->od_quota_slave_dt; + else + qsd = osd->od_quota_slave_md; + + rc = qsd_reserve_or_free_quota(env, qsd, qi); + 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_sync = osd_sync, + .dt_commit_async = osd_commit_async, + .dt_ro = osd_ro, + .dt_reserve_or_free_quota = osd_reserve_or_free_quota, }; /* @@ -711,10 +767,11 @@ static void osd_key_fini(const struct lu_context *ctx, if (idc != NULL) { LASSERT(info->oti_ins_cache_size > 0); - OBD_FREE(idc, sizeof(*idc) * info->oti_ins_cache_size); + OBD_FREE_PTR_ARRAY_LARGE(idc, info->oti_ins_cache_size); info->oti_ins_cache = NULL; info->oti_ins_cache_size = 0; } + lu_buf_free(&info->oti_xattr_lbuf); OBD_FREE_PTR(info); } @@ -745,11 +802,21 @@ 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; + if (o->od_quota_slave_md != NULL) { + /* complete all in-flight callbacks */ + osd_sync(env, &o->od_dt_dev); + txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os))); + qsd_fini(env, o->od_quota_slave_md); + o->od_quota_slave_md = NULL; } + if (o->od_quota_slave_dt != NULL) { + /* complete all in-flight callbacks */ + osd_sync(env, &o->od_dt_dev); + txg_wait_callbacks(spa_get_dsl(dmu_objset_spa(o->od_os))); + qsd_fini(env, o->od_quota_slave_dt); + o->od_quota_slave_dt = NULL; + } osd_fid_fini(env, o); RETURN(0); @@ -780,6 +847,14 @@ static void osd_readonly_changed_cb(void *arg, uint64_t newval) osd->od_prop_rdonly = !!newval; } +#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE +static void osd_dnodesize_changed_cb(void *arg, uint64_t newval) +{ + struct osd_device *osd = arg; + + osd->od_dnsize = newval; +} +#endif /* * This function unregisters all registered callbacks. It's harmless to * unregister callbacks that were never registered so it is used to safely @@ -795,6 +870,10 @@ static void osd_objset_unregister_callbacks(struct osd_device *o) osd_recordsize_changed_cb, o); (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY), osd_readonly_changed_cb, o); +#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE), + osd_dnodesize_changed_cb, o); +#endif if (o->arc_prune_cb != NULL) { arc_remove_prune_callback(o->arc_prune_cb); @@ -831,6 +910,13 @@ static int osd_objset_register_callbacks(struct osd_device *o) if (rc) GOTO(err, rc); +#ifdef HAVE_DMU_OBJECT_ALLOC_DNSIZE + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_DNODESIZE), + osd_dnodesize_changed_cb, o); + if (rc) + GOTO(err, rc); +#endif + o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o); err: dsl_pool_config_exit(dp, FTAG); @@ -843,13 +929,14 @@ err: static int osd_objset_open(struct osd_device *o) { uint64_t version = ZPL_VERSION; - uint64_t sa_obj; + uint64_t sa_obj, unlink_obj; int rc; ENTRY; - rc = -dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, + rc = -osd_dmu_objset_own(o->od_mntdev, DMU_OST_ZFS, o->od_dt_dev.dd_rdonly ? B_TRUE : B_FALSE, - o, &o->od_os); + B_TRUE, o, &o->od_os); + if (rc) { CERROR("%s: can't open %s\n", o->od_svname, o->od_mntdev); o->od_os = NULL; @@ -888,7 +975,7 @@ static int osd_objset_open(struct osd_device *o) } rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET, - 8, 1, &o->od_unlinkedid); + 8, 1, &unlink_obj); if (rc) { CERROR("%s: lookup for %s failed: rc = %d\n", o->od_svname, ZFS_UNLINKED_SET, rc); @@ -904,18 +991,26 @@ static int osd_objset_open(struct osd_device *o) GOTO(out, rc = -ENOTSUPP); } + rc = __osd_obj2dnode(o->od_os, unlink_obj, &o->od_unlinked); + if (rc) { + CERROR("%s: can't get dnode for unlinked: rc = %d\n", + o->od_svname, rc); + GOTO(out, rc); + } + out: if (rc != 0 && o->od_os != NULL) { - dmu_objset_disown(o->od_os, o); + osd_dmu_objset_disown(o->od_os, B_TRUE, o); o->od_os = NULL; } RETURN(rc); } -static int -osd_unlinked_object_free(struct osd_device *osd, uint64_t oid) +int osd_unlinked_object_free(const struct lu_env *env, struct osd_device *osd, + uint64_t oid) { + char *key = osd_oti_get(env)->oti_str; int rc; dmu_tx_t *tx; @@ -935,8 +1030,10 @@ osd_unlinked_object_free(struct osd_device *osd, uint64_t oid) } tx = dmu_tx_create(osd->od_os); + dmu_tx_mark_netfree(tx); dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END); - dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL); + osd_tx_hold_zap(tx, osd->od_unlinked->dn_object, osd->od_unlinked, + FALSE, NULL); rc = -dmu_tx_assign(tx, TXG_WAIT); if (rc != 0) { CWARN("%s: Cannot assign tx for %llu: rc = %d\n", @@ -944,7 +1041,9 @@ osd_unlinked_object_free(struct osd_device *osd, uint64_t oid) goto failed; } - rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx); + snprintf(key, sizeof(osd_oti_get(env)->oti_str), "%llx", oid); + rc = osd_zap_remove(osd, osd->od_unlinked->dn_object, + osd->od_unlinked, key, tx); if (rc != 0) { CWARN("%s: Cannot remove %llu from unlinked set: rc = %d\n", osd->od_svname, oid, rc); @@ -974,13 +1073,13 @@ osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd) zap_cursor_t zc; zap_attribute_t *za = &osd_oti_get(env)->oti_za; - zap_cursor_init(&zc, osd->od_os, osd->od_unlinkedid); + zap_cursor_init(&zc, osd->od_os, osd->od_unlinked->dn_object); while (zap_cursor_retrieve(&zc, za) == 0) { /* If cannot free the object, leave it in the unlinked set, * until the OSD is mounted again when obd_unlinked_drain() * will be called. */ - if (osd_unlinked_object_free(osd, za->za_first_integer) != 0) + if (osd_unlinked_object_free(env, osd, za->za_first_integer)) break; zap_cursor_advance(&zc); } @@ -991,12 +1090,15 @@ osd_unlinked_drain(const struct lu_env *env, struct osd_device *osd) static int osd_mount(const struct lu_env *env, struct osd_device *o, struct lustre_cfg *cfg) { - char *mntdev = lustre_cfg_string(cfg, 1); - char *str = lustre_cfg_string(cfg, 2); - char *svname = lustre_cfg_string(cfg, 4); + char *mntdev = lustre_cfg_string(cfg, 1); + char *str = lustre_cfg_string(cfg, 2); + char *svname = lustre_cfg_string(cfg, 4); + time64_t interval = AS_DEFAULT; dnode_t *rootdn; - const char *opts; - int rc; + const char *opts; + bool resetoi = false; + int rc; + ENTRY; if (o->od_os != NULL) @@ -1013,6 +1115,11 @@ static int osd_mount(const struct lu_env *env, if (rc >= sizeof(o->od_svname)) RETURN(-E2BIG); + opts = lustre_cfg_string(cfg, 3); + + o->od_index_backup_stop = 0; + o->od_index = -1; /* -1 means index is invalid */ + rc = server_name2index(o->od_svname, &o->od_index, NULL); str = strstr(str, ":"); if (str) { unsigned long flags; @@ -1026,6 +1133,9 @@ static int osd_mount(const struct lu_env *env, LCONSOLE_WARN("%s: set dev_rdonly on this device\n", svname); } + + if (flags & LMD_FLG_NOSCRUB) + interval = AS_NEVER; } if (server_name_is_ost(o->od_svname)) @@ -1037,23 +1147,33 @@ static int osd_mount(const struct lu_env *env, o->od_xattr_in_sa = B_TRUE; o->od_max_blksz = osd_spa_maxblocksize(o->od_os->os_spa); + o->od_readcache_max_filesize = OSD_MAX_CACHE_SIZE; - rc = osd_objset_register_callbacks(o); + rc = __osd_obj2dnode(o->od_os, o->od_rootid, &rootdn); if (rc) GOTO(err, rc); + o->od_root = rootdn->dn_object; + osd_dnode_rele(rootdn); - rc = __osd_obj2dnode(env, o->od_os, o->od_rootid, &rootdn); + rc = __osd_obj2dnode(o->od_os, DMU_USERUSED_OBJECT, + &o->od_userused_dn); if (rc) GOTO(err, rc); - o->od_root = rootdn->dn_object; - osd_dnode_rele(rootdn); - - /* 1. initialize oi before any file create or file open */ - rc = osd_oi_init(env, o); + rc = __osd_obj2dnode(o->od_os, DMU_GROUPUSED_OBJECT, + &o->od_groupused_dn); if (rc) GOTO(err, rc); +#ifdef ZFS_PROJINHERIT + if (dmu_objset_projectquota_enabled(o->od_os)) { + rc = __osd_obj2dnode(o->od_os, DMU_PROJECTUSED_OBJECT, + &o->od_projectused_dn); + if (rc && rc != -ENOENT) + GOTO(err, rc); + } +#endif + rc = lu_site_init(&o->od_site, osd2lu_dev(o)); if (rc) GOTO(err, rc); @@ -1063,31 +1183,69 @@ static int osd_mount(const struct lu_env *env, if (rc) GOTO(err, rc); + rc = osd_objset_register_callbacks(o); + if (rc) + GOTO(err, rc); + + if (opts && strstr(opts, "resetoi")) + resetoi = true; + + rc = lprocfs_init_brw_stats(&o->od_brw_stats); + if (rc) + GOTO(err, rc); + + o->od_in_init = 1; + rc = osd_scrub_setup(env, o, interval, resetoi); + o->od_in_init = 0; + if (rc) + GOTO(err, rc); + rc = osd_procfs_init(o, o->od_svname); if (rc) GOTO(err, rc); - /* 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; + /* currently it's no need to prepare qsd_instance_md for OST */ + if (!o->od_is_ost) { + o->od_quota_slave_md = qsd_init(env, o->od_svname, + &o->od_dt_dev, o->od_proc_entry, + true, false); + if (IS_ERR(o->od_quota_slave_md)) { + rc = PTR_ERR(o->od_quota_slave_md); + o->od_quota_slave_md = NULL; + GOTO(err, rc); + } + } + + o->od_quota_slave_dt = qsd_init(env, o->od_svname, &o->od_dt_dev, + o->od_proc_entry, false, false); + + if (IS_ERR(o->od_quota_slave_dt)) { + if (o->od_quota_slave_md != NULL) { + qsd_fini(env, o->od_quota_slave_md); + o->od_quota_slave_md = NULL; + } + + rc = PTR_ERR(o->od_quota_slave_dt); + o->od_quota_slave_dt = NULL; GOTO(err, rc); } +#ifdef HAVE_DMU_USEROBJ_ACCOUNTING + if (!osd_dmu_userobj_accounting_available(o)) + CWARN("%s: dnode accounting not enabled: " + "enable feature@userobj_accounting in pool\n", + o->od_mntdev); +#endif + /* 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; osd_unlinked_drain(env, o); -err: - if (rc && o->od_os) { - dmu_objset_disown(o->od_os, o); - o->od_os = NULL; - } + RETURN(0); + +err: RETURN(rc); } @@ -1105,14 +1263,33 @@ static void osd_umount(const struct lu_env *env, struct osd_device *o) CERROR("%s: lost %d pinned dbuf(s)\n", o->od_svname, atomic_read(&o->od_zerocopy_pin)); + if (o->od_unlinked) { + osd_dnode_rele(o->od_unlinked); + o->od_unlinked = NULL; + } + if (o->od_userused_dn) { + osd_dnode_rele(o->od_userused_dn); + o->od_userused_dn = NULL; + } + if (o->od_groupused_dn) { + osd_dnode_rele(o->od_groupused_dn); + o->od_groupused_dn = NULL; + } + +#ifdef ZFS_PROJINHERIT + if (o->od_projectused_dn) { + osd_dnode_rele(o->od_projectused_dn); + o->od_projectused_dn = NULL; + } +#endif + if (o->od_os != NULL) { if (!o->od_dt_dev.dd_rdonly) /* 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); - + osd_dmu_objset_disown(o->od_os, B_TRUE, o); o->od_os = NULL; } @@ -1133,13 +1310,73 @@ static int osd_device_init0(const struct lu_env *env, l->ld_ops = &osd_lu_ops; o->od_dt_dev.dd_ops = &osd_dt_ops; + sema_init(&o->od_otable_sem, 1); + INIT_LIST_HEAD(&o->od_ios_list); + o->od_sync_on_lseek = B_TRUE; + + /* ZFS does not support reporting nonrotional status yet, so this flag + * is only set if explicitly set by the user. + */ + o->od_nonrotational = 0; out: RETURN(rc); } static struct lu_device *osd_device_fini(const struct lu_env *env, - struct lu_device *dev); + struct lu_device *d) +{ + struct osd_device *o = osd_dev(d); + int rc; + + ENTRY; + osd_index_backup(env, o, false); + if (o->od_os) { + osd_objset_unregister_callbacks(o); + if (!o->od_dt_dev.dd_rdonly) { + osd_sync(env, lu2dt_dev(d)); + txg_wait_callbacks( + spa_get_dsl(dmu_objset_spa(o->od_os))); + } + } + + /* now with all the callbacks completed we can cleanup the remainings */ + osd_shutdown(env, o); + osd_scrub_cleanup(env, o); + + 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 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 */ + if (d->ld_site) { + d->ld_site->ls_top_dev = d; + lu_site_purge(env, d->ld_site, -1); + lu_site_print(env, d->ld_site, &d->ld_site->ls_obj_hash.nelems, + D_ERROR, lu_cdebug_printer); + } + if (o->od_site.ls_bottom_dev) + 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_alloc(const struct lu_env *env, struct lu_device_type *type, @@ -1157,77 +1394,32 @@ static struct lu_device *osd_device_alloc(const struct lu_env *env, INIT_LIST_HEAD(&osl->osl_seq_list); rwlock_init(&osl->osl_seq_list_lock); sema_init(&osl->osl_seq_init_sem, 1); + INIT_LIST_HEAD(&dev->od_index_backup_list); + INIT_LIST_HEAD(&dev->od_index_restore_list); + spin_lock_init(&dev->od_lock); + dev->od_index_backup_policy = LIBP_NONE; 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) + if (rc) { osd_device_fini(env, osd2lu_dev(dev)); - } - if (rc) + osd_device_free(env, osd2lu_dev(dev)); + dev = NULL; + } + } else { dt_device_fini(&dev->od_dt_dev); + } } - if (unlikely(rc != 0)) + if (unlikely(rc != 0) && dev) 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); - int rc; - ENTRY; - - - osd_shutdown(env, o); - osd_oi_fini(env, o); - - if (o->od_os) { - osd_objset_unregister_callbacks(o); - if (!o->od_dt_dev.dd_rdonly) { - 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) { @@ -1241,27 +1433,37 @@ static int osd_device_init(const struct lu_env *env, struct lu_device *d, 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; + struct osd_device *o = osd_dev(d); + ssize_t count; + int rc; + ENTRY; switch(cfg->lcfg_command) { case LCFG_SETUP: rc = osd_mount(env, o, cfg); break; case LCFG_CLEANUP: + /* For the case LCFG_PRE_CLEANUP is not called in advance, + * that may happend if hit failure during mount process. */ + osd_index_backup(env, o, false); 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); + count = class_modify_config(cfg, PARAM_OSD, + &o->od_dt_dev.dd_kobj); + if (count < 0) + count = class_modify_config(cfg, PARAM_OST, + &o->od_dt_dev.dd_kobj); + rc = count > 0 ? 0 : count; break; } + case LCFG_PRE_CLEANUP: + osd_scrub_stop(o); + osd_index_backup(env, o, + o->od_index_backup_policy != LIBP_NONE); + rc = 0; + break; default: rc = -ENOTTY; } @@ -1275,12 +1477,14 @@ static int osd_recovery_complete(const struct lu_env *env, struct lu_device *d) int rc = 0; ENTRY; - if (osd->od_quota_slave == NULL) + if (osd->od_quota_slave_md == NULL && osd->od_quota_slave_dt == 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); + rc = qsd_start(env, osd->od_quota_slave_dt); + if (rc == 0 && osd->od_quota_slave_md != NULL) + rc = qsd_start(env, osd->od_quota_slave_md); RETURN(rc); } @@ -1338,8 +1542,8 @@ static int osd_obd_disconnect(struct obd_export *exp) 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; + struct seq_server_site *ss = osd_seq_site(osd); + int rc = 0; ENTRY; if (osd->od_is_ost || osd->od_cl_seq != NULL) @@ -1352,12 +1556,17 @@ static int osd_fid_init(const struct lu_env *env, struct osd_device *osd) 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); + 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; + if (ss->ss_node_id == 0) { + /* + * If the OSD on the sequence controller(MDT0), then allocate + * sequence here, otherwise allocate sequence after connected + * to MDT0 (see mdt_register_lwp_callback()). + */ + rc = seq_server_alloc_meta(osd->od_cl_seq->lcs_srv, + &osd->od_cl_seq->lcs_space, env); } RETURN(rc); @@ -1370,9 +1579,16 @@ static int osd_prepare(const struct lu_env *env, struct lu_device *pdev, int rc = 0; ENTRY; - if (osd->od_quota_slave != NULL) { + if (osd->od_quota_slave_md != NULL) { /* set up quota slave objects */ - rc = qsd_prepare(env, osd->od_quota_slave); + rc = qsd_prepare(env, osd->od_quota_slave_md); + if (rc != 0) + RETURN(rc); + } + + if (osd->od_quota_slave_dt != NULL) { + /* set up quota slave objects */ + rc = qsd_prepare(env, osd->od_quota_slave_dt); if (rc != 0) RETURN(rc); } @@ -1382,11 +1598,28 @@ static int osd_prepare(const struct lu_env *env, struct lu_device *pdev, RETURN(rc); } -struct lu_device_operations osd_lu_ops = { +/** + * Implementation of lu_device_operations::ldo_fid_alloc() for OSD + * + * Allocate FID. + * + * see include/lu_object.h for the details. + */ +static int osd_fid_alloc(const struct lu_env *env, struct lu_device *d, + struct lu_fid *fid, struct lu_object *parent, + const struct lu_name *name) +{ + struct osd_device *osd = osd_dev(d); + + return seq_client_alloc_fid(env, osd->od_cl_seq, fid); +} + +const 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, + .ldo_fid_alloc = osd_fid_alloc, }; static void osd_type_start(struct lu_device_type *t) @@ -1397,15 +1630,7 @@ 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 = { +static const struct lu_device_type_operations osd_device_type_ops = { .ldto_init = osd_type_init, .ldto_fini = osd_type_fini, @@ -1427,11 +1652,10 @@ static struct lu_device_type osd_device_type = { }; -static struct obd_ops osd_obd_device_ops = { +static const 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 }; static int __init osd_init(void) @@ -1446,7 +1670,7 @@ static int __init osd_init(void) if (rc) return rc; - rc = class_register_type(&osd_obd_device_ops, NULL, true, NULL, + rc = class_register_type(&osd_obd_device_ops, NULL, true, LUSTRE_OSD_ZFS_NAME, &osd_device_type); if (rc) lu_kmem_fini(osd_caches); @@ -1459,10 +1683,13 @@ static void __exit osd_exit(void) lu_kmem_fini(osd_caches); } -extern unsigned int osd_oi_count; module_param(osd_oi_count, int, 0444); MODULE_PARM_DESC(osd_oi_count, "Number of Object Index containers to be created, it's only valid for new filesystem."); +module_param(osd_txg_sync_delay_us, int, 0644); +MODULE_PARM_DESC(osd_txg_sync_delay_us, + "When zero or larger delay N usec instead of doing TXG sync"); + MODULE_AUTHOR("OpenSFS, Inc. "); MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")"); MODULE_VERSION(LUSTRE_VERSION_STRING);