X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_handler.c;h=bb420236a3275f14cfd796c8fece335037b63351;hp=3c2abe52e3d75a9f014c2bda772a792df237da67;hb=ba7c99a5b18535305d3ae3b1687116c1fff5b4a8;hpb=fd3feec5956d92561789f9db519b7cdb0e00d5ac diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c index 3c2abe5..bb42023 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" @@ -78,7 +77,7 @@ 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 +130,7 @@ 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 +148,22 @@ 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 +173,7 @@ 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 +208,10 @@ 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 +227,8 @@ 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 +242,23 @@ 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", @@ -276,13 +297,16 @@ 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); } @@ -297,7 +321,7 @@ int osd_statfs(const struct lu_env *env, struct dt_device *d, ENTRY; rc = udmu_objset_statfs(&osd->od_objset, osfs); - if (rc) + if (unlikely(rc)) RETURN(rc); osfs->os_bavail -= min_t(obd_size, OSD_GRANT_FOR_LOCAL_OIDS / osfs->os_bsize, @@ -312,19 +336,23 @@ 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_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 @@ -359,6 +387,7 @@ static int osd_commit_async(const struct lu_env *env, struct dt_device *dev) tx_state_t *tx = &dmu_objset_pool(osd->od_objset.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; @@ -480,33 +509,41 @@ static int osd_shutdown(const struct lu_env *env, struct osd_device *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_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; - 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) 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); - strcpy(o->od_svname, lsi->lsi_ldd->ldd_svname); + if (server_name_is_ost(o->od_svname)) + o->od_is_ost = 1; rc = -udmu_objset_open(o->od_mntdev, &o->od_objset); if (rc) { @@ -514,6 +551,17 @@ static int osd_mount(const struct lu_env *env, RETURN(rc); } + ds = dmu_objset_ds(o->od_objset.os); + dp = dmu_objset_pool(o->od_objset.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) + CERROR("%s: cat not register xattr callback, ignore: %d\n", + o->od_svname, rc); + rc = __osd_obj2dbuf(env, o->od_objset.os, o->od_objset.root, &rootdb, root_tag); if (rc) { @@ -530,6 +578,19 @@ static int osd_mount(const struct lu_env *env, 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; @@ -540,6 +601,20 @@ static int osd_mount(const struct lu_env *env, 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); } @@ -548,15 +623,15 @@ 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); @@ -587,28 +662,36 @@ 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, @@ -619,9 +702,13 @@ 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); @@ -631,17 +718,25 @@ 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; + ds = dmu_objset_ds(o->od_objset.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))); } @@ -655,10 +750,6 @@ static struct lu_device *osd_device_fini(const struct lu_env *env, if (o->od_objset.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); } @@ -676,29 +767,101 @@ 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_seq_param(PARAM_OSD, + lprocfs_osd_obd_vars, cfg, + &o->od_dt_dev); + if (rc > 0 || rc == -ENOSYS) + rc = class_process_proc_seq_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(&osd->od_objset.lock); + osd->od_connects++; + spin_unlock(&osd->od_objset.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(&osd->od_objset.lock); + osd->od_connects--; + if (osd->od_connects == 0) + release = 1; + spin_unlock(&osd->od_objset.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) { @@ -706,13 +869,9 @@ static int osd_prepare(const struct lu_env *env, struct lu_device *pdev, int rc = 0; ENTRY; - /* initialize quota slave instance */ - osd->od_quota_slave = qsd_init(env, osd->od_svname, &osd->od_dt_dev, - osd->od_proc_entry); - if (IS_ERR(osd->od_quota_slave)) { - rc = PTR_ERR(osd->od_quota_slave); - osd->od_quota_slave = NULL; - } + if (osd->od_quota_slave != NULL) + /* set up quota slave objects */ + rc = qsd_prepare(env, osd->od_quota_slave); RETURN(rc); } @@ -732,6 +891,14 @@ 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, @@ -756,6 +923,9 @@ 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) @@ -770,8 +940,10 @@ 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, +#ifndef HAVE_ONLY_PROCFS_SEQ + NULL, +#endif LUSTRE_OSD_ZFS_NAME, &osd_device_type); if (rc) lu_kmem_fini(osd_caches);