X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fosd-zfs%2Fosd_handler.c;h=334310d98da5b46fe41e8d758173556051005d43;hb=60270c6488b01db756eb216548f83f2826972854;hp=77f479933af282d68b714319445d2b5986a97abf;hpb=3270bfc2370884933628f95122da00d430db6072;p=fs%2Flustre-release.git diff --git a/lustre/osd-zfs/osd_handler.c b/lustre/osd-zfs/osd_handler.c index 77f4799..334310d 100644 --- a/lustre/osd-zfs/osd_handler.c +++ b/lustre/osd-zfs/osd_handler.c @@ -26,11 +26,8 @@ /* * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. - */ -/* - * Copyright (c) 2012, 2013, Intel Corporation. - * Use is subject to license terms. * + * Copyright (c) 2012, 2015, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -54,6 +51,7 @@ #include #include #include +#include #include #include "osd_internal.h" @@ -73,11 +71,12 @@ struct lu_context_key osd_key; -static char *root_tag = "osd_mount, rootdb"; - /* 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, @@ -85,6 +84,11 @@ static struct lu_kmem_descr osd_caches[] = { .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 } }; @@ -147,7 +151,7 @@ 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. @@ -169,10 +173,15 @@ 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; + struct osd_thandle *oh = container_of0(th, struct osd_thandle, + ot_super); - oh = container_of0(th, struct osd_thandle, ot_super); - cfs_list_add(&dcb->dcb_linkage, &oh->ot_dcb_list); + LASSERT(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC); + LASSERT(&dcb->dcb_func != NULL); + if (dcb->dcb_flags & DCB_TRANS_STOP) + list_add(&dcb->dcb_linkage, &oh->ot_stop_dcb_list); + else + list_add(&dcb->dcb_linkage, &oh->ot_dcb_list); return 0; } @@ -205,9 +214,8 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d, 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); + CERROR("%s: failed to start transaction due to ENOSPC" + "\n", osd->od_svname); else CERROR("%s: can't assign tx: rc = %d\n", osd->od_svname, rc); @@ -223,23 +231,65 @@ static int osd_trans_start(const struct lu_env *env, struct dt_device *d, 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, + struct list_head *list, bool free) +{ + struct osd_object *obj; + uint64_t oid; + + while (!list_empty(list)) { + obj = list_entry(list->next, + struct osd_object, oo_unlinked_linkage); + LASSERT(obj->oo_db != NULL); + oid = obj->oo_db->db_object; + + list_del_init(&obj->oo_unlinked_linkage); + if (free) + (void)osd_unlinked_object_free(osd, oid); + } +} + +static void osd_trans_stop_cb(struct osd_thandle *oth, int result) +{ + struct dt_txn_commit_cb *dcb; + struct dt_txn_commit_cb *tmp; + + /* call per-transaction stop callbacks if any */ + list_for_each_entry_safe(dcb, tmp, &oth->ot_stop_dcb_list, + dcb_linkage) { + LASSERTF(dcb->dcb_magic == TRANS_COMMIT_CB_MAGIC, + "commit callback entry: magic=%x name='%s'\n", + dcb->dcb_magic, dcb->dcb_name); + list_del_init(&dcb->dcb_linkage); + dcb->dcb_func(NULL, &oth->ot_super, dcb, result); + } +} + /* * Concurrency: shouldn't matter. */ -static int osd_trans_stop(const struct lu_env *env, struct 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); + bool sync = (th->th_sync != 0); struct osd_thandle *oh; + struct list_head unlinked; uint64_t txg; int rc; ENTRY; oh = container_of0(th, struct osd_thandle, ot_super); + INIT_LIST_HEAD(&unlinked); + list_splice_init(&oh->ot_unlinked_list, &unlinked); if (oh->ot_assigned == 0) { LASSERT(oh->ot_tx); dmu_tx_abort(oh->ot_tx); osd_object_sa_dirty_rele(oh); + osd_unlinked_list_emptify(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); @@ -262,14 +312,20 @@ static int osd_trans_stop(const struct lu_env *env, struct thandle *th) CDEBUG(D_OTHER, "%s: transaction hook failed: rc = %d\n", osd->od_svname, rc); + osd_trans_stop_cb(oh, rc); + LASSERT(oh->ot_tx); txg = oh->ot_tx->tx_txg; osd_object_sa_dirty_rele(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); - if (th->th_sync) - txg_wait_synced(dmu_objset_pool(osd->od_objset.os), txg); + osd_unlinked_list_emptify(osd, &unlinked, true); + + if (sync) + txg_wait_synced(dmu_objset_pool(osd->od_os), txg); RETURN(rc); } @@ -283,7 +339,7 @@ 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)); @@ -295,8 +351,10 @@ 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); + INIT_LIST_HEAD(&oh->ot_dcb_list); + 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; @@ -306,25 +364,189 @@ static struct thandle *osd_trans_create(const struct lu_env *env, 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_maxblockshift) +{ + 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 >> est_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 osd_device *osd, struct obd_statfs *osfs) +{ + struct objset *os = osd->od_os; + uint64_t refdbytes, availbytes, usedobjs, availobjs; + uint64_t est_availobjs; + uint64_t reserved; + uint64_t bshift; + + dmu_objset_space(os, &refdbytes, &availbytes, &usedobjs, &availobjs); + + memset(osfs, 0, sizeof(*osfs)); + + /* We're a zfs filesystem. */ + osfs->os_type = UBERBLOCK_MAGIC; + + /* + * 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 = osd->od_max_blksz; + bshift = fls64(osfs->os_bsize) - 1; + + osfs->os_blocks = (refdbytes + availbytes) >> bshift; + osfs->os_bfree = availbytes >> bshift; + 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 16MB for small filesystems, + * for internal files to be created/unlinked when space is tight. + */ + CLASSERT(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; + + 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, bshift); + + 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, .... );*/ + + osfs->os_namelen = MAXNAMELEN; + osfs->os_maxbytes = OBD_OBJECT_EOF; + + if (!spa_writeable(dmu_objset_spa(os)) || + osd->od_dev_set_rdonly || osd->od_prop_rdonly) + osfs->os_state |= OS_STATE_READONLY; + + 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 = udmu_objset_statfs(&osd->od_objset, osfs); - if (unlikely(rc)) + rc = osd_objset_statfs(osd_dt_dev(d), osfs); + if (unlikely(rc != 0)) RETURN(rc); - osfs->os_bavail -= min_t(obd_size, + + 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(struct osd_device *osd) +{ + int max_blockshift, nr_blkptrshift, bshift; + + /* max_blockshift is the log2 of the number of blocks needed to reach + * the maximum filesize (that's to say 2^64) */ + bshift = osd_spa_maxblockshift(dmu_objset_spa(osd->od_os)); + max_blockshift = DN_MAX_OFFSET_SHIFT - bshift; + + /* 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_symlink_max = PATH_MAX; + 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; - - /* 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; + param->ddp_maxbytes = MAX_LFS_FILESIZE; /* 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 = udmu_blk_insert_cost(); + /* Although ZFS isn't an extent-based filesystem, the metadata overhead + * (i.e. 7 levels of indirect blocks, see osd_blk_insert_cost()) should + * not be accounted for every single new block insertion. + * Instead, the maximum extent size is set to the number of blocks that + * can fit into a single contiguous indirect block. There would be some + * cases where this crosses indirect blocks, but it also won't have 7 + * new levels of indirect blocks in that case either, so it will still + * have enough reserved space for the extra indirect block */ + param->ddp_max_extent_blks = + (1 << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT)); + param->ddp_extent_tax = osd_blk_insert_cost(osd); } /* @@ -368,15 +596,16 @@ static void osd_conf_get(const struct lu_env *env, 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); + 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_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); @@ -400,26 +629,8 @@ 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)); - - RETURN(0); -} - -/* - * Concurrency: serialization provided by callers. - */ -static int osd_init_capa_ctxt(const struct lu_env *env, struct dt_device *d, - int mode, unsigned long timeout, __u32 alg, - struct lustre_capa_key *keys) -{ - struct osd_device *dev = osd_dt_dev(d); - ENTRY; - - dev->od_fl_capa = mode; - dev->od_capa_timeout = timeout; - dev->od_capa_alg = alg; - dev->od_capa_keys = keys; + osd->od_dev_set_rdonly = 1; + spa_freeze(dmu_objset_spa(osd->od_os)); RETURN(0); } @@ -435,7 +646,6 @@ static struct dt_device_operations osd_dt_ops = { .dt_sync = osd_sync, .dt_commit_async = osd_commit_async, .dt_ro = osd_ro, - .dt_init_capa_ctxt = osd_init_capa_ctxt, }; /* @@ -488,6 +698,16 @@ struct lu_context_key osd_key = { .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; @@ -498,6 +718,8 @@ static int osd_shutdown(const struct lu_env *env, struct osd_device *o) o->od_quota_slave = NULL; } + osd_fid_fini(env, o); + RETURN(0); } @@ -508,56 +730,264 @@ static void osd_xattr_changed_cb(void *arg, uint64_t newval) osd->od_xattr_in_sa = (newval == ZFS_XATTR_SA); } +static void osd_recordsize_changed_cb(void *arg, uint64_t newval) +{ + struct osd_device *osd = arg; + + LASSERT(newval <= osd_spa_maxblocksize(dmu_objset_spa(osd->od_os))); + LASSERT(newval >= SPA_MINBLOCKSIZE); + LASSERT(ISP2(newval)); + + osd->od_max_blksz = newval; +} + +static void osd_readonly_changed_cb(void *arg, uint64_t newval) +{ + struct osd_device *osd = arg; + + osd->od_prop_rdonly = !!newval; +} + +/* + * This function unregisters all registered callbacks. It's harmless to + * unregister callbacks that were never registered so it is used to safely + * unwind a partially completed call to osd_objset_register_callbacks(). + */ +static void osd_objset_unregister_callbacks(struct osd_device *o) +{ + struct dsl_dataset *ds = dmu_objset_ds(o->od_os); + + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_XATTR), + osd_xattr_changed_cb, o); + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE), + osd_recordsize_changed_cb, o); + (void) dsl_prop_unregister(ds, zfs_prop_to_name(ZFS_PROP_READONLY), + osd_readonly_changed_cb, o); + + if (o->arc_prune_cb != NULL) { + arc_remove_prune_callback(o->arc_prune_cb); + o->arc_prune_cb = NULL; + } +} + +/* + * Register the required callbacks to be notified when zfs properties + * are modified using the 'zfs(8)' command line utility. + */ +static int osd_objset_register_callbacks(struct osd_device *o) +{ + struct dsl_dataset *ds = dmu_objset_ds(o->od_os); + dsl_pool_t *dp = dmu_objset_pool(o->od_os); + int rc; + + LASSERT(ds); + LASSERT(dp); + + dsl_pool_config_enter(dp, FTAG); + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_XATTR), + osd_xattr_changed_cb, o); + if (rc) + GOTO(err, rc); + + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_RECORDSIZE), + osd_recordsize_changed_cb, o); + if (rc) + GOTO(err, rc); + + rc = -dsl_prop_register(ds, zfs_prop_to_name(ZFS_PROP_READONLY), + osd_readonly_changed_cb, o); + if (rc) + GOTO(err, rc); + + o->arc_prune_cb = arc_add_prune_callback(arc_prune_func, o); +err: + dsl_pool_config_exit(dp, FTAG); + if (rc) + osd_objset_unregister_callbacks(o); + + RETURN(rc); +} + +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); + } + + rc = -zap_lookup(o->od_os, MASTER_NODE_OBJ, ZFS_UNLINKED_SET, + 8, 1, &o->od_unlinkedid); + if (rc) { + CERROR("%s: lookup for %s failed: rc = %d\n", + o->od_svname, ZFS_UNLINKED_SET, 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); + o->od_os = NULL; + } + + RETURN(rc); +} + +static int +osd_unlinked_object_free(struct osd_device *osd, uint64_t oid) +{ + int rc; + dmu_tx_t *tx; + + rc = -dmu_free_long_range(osd->od_os, oid, 0, DMU_OBJECT_END); + if (rc != 0) { + CWARN("%s: Cannot truncate "LPU64": rc = %d\n", + osd->od_svname, oid, rc); + return rc; + } + + tx = dmu_tx_create(osd->od_os); + dmu_tx_hold_free(tx, oid, 0, DMU_OBJECT_END); + dmu_tx_hold_zap(tx, osd->od_unlinkedid, FALSE, NULL); + rc = -dmu_tx_assign(tx, TXG_WAIT); + if (rc != 0) { + CWARN("%s: Cannot assign tx for "LPU64": rc = %d\n", + osd->od_svname, oid, rc); + goto failed; + } + + rc = -zap_remove_int(osd->od_os, osd->od_unlinkedid, oid, tx); + if (rc != 0) { + CWARN("%s: Cannot remove "LPU64" from unlinked set: rc = %d\n", + osd->od_svname, oid, rc); + goto failed; + } + + rc = -dmu_object_free(osd->od_os, oid, tx); + if (rc != 0) { + CWARN("%s: Cannot free "LPU64": rc = %d\n", + osd->od_svname, oid, rc); + goto failed; + } + dmu_tx_commit(tx); + + return 0; + +failed: + LASSERT(rc != 0); + dmu_tx_abort(tx); + + return rc; +} + +static void +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); + + 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) + break; + zap_cursor_advance(&zc); + } + + zap_cursor_fini(&zc); +} + static int osd_mount(const struct lu_env *env, struct osd_device *o, struct lustre_cfg *cfg) { - struct dsl_dataset *ds; - char *dev = lustre_cfg_string(cfg, 1); + 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); - if (strlen(dev) >= sizeof(o->od_mntdev)) + 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); - strcpy(o->od_mntdev, dev); - strncpy(o->od_svname, lustre_cfg_string(cfg, 4), - sizeof(o->od_svname) - 1); + 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 = -udmu_objset_open(o->od_mntdev, &o->od_objset); - if (rc) { - CERROR("can't open objset %s: %d\n", o->od_mntdev, rc); - RETURN(rc); - } + rc = osd_objset_open(o); + if (rc) + GOTO(err, 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); + o->od_xattr_in_sa = B_TRUE; + o->od_max_blksz = SPA_OLD_MAXBLOCKSIZE; + + rc = osd_objset_register_callbacks(o); if (rc) - CERROR("%s: cat not register xattr callback, ignore: %d\n", - o->od_svname, rc); + GOTO(err, rc); - rc = __osd_obj2dbuf(env, o->od_objset.os, o->od_objset.root, - &rootdb, root_tag); - if (rc) { - CERROR("udmu_obj2dbuf() failed with error %d\n", rc); - udmu_objset_close(&o->od_objset); - RETURN(rc); - } + rc = __osd_obj2dbuf(env, o->od_os, o->od_rootid, &rootdb); + if (rc) + GOTO(err, 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); @@ -573,10 +1003,6 @@ static int osd_mount(const struct lu_env *env, 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; @@ -585,8 +1011,6 @@ static int osd_mount(const struct lu_env *env, 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); @@ -595,7 +1019,20 @@ static int osd_mount(const struct lu_env *env, 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; + + osd_unlinked_drain(env, o); err: + if (rc) { + if (o->od_os) + dmu_objset_disown(o->od_os, o); + o->od_os = NULL; + } + RETURN(rc); } @@ -603,18 +1040,25 @@ 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; } @@ -634,10 +1078,6 @@ 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; - o->od_capa_hash = init_capa_hash(); - if (o->od_capa_hash == NULL) - GOTO(out, rc = -ENOMEM); - out: RETURN(rc); } @@ -649,13 +1089,19 @@ 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; + struct osd_device *dev; + struct osd_seq_list *osl; + int rc; OBD_ALLOC_PTR(dev); if (dev == NULL) return ERR_PTR(-ENOMEM); + osl = &dev->od_seq_list; + INIT_LIST_HEAD(&osl->osl_seq_list); + rwlock_init(&osl->osl_seq_list_lock); + sema_init(&osl->osl_seq_init_sem, 1); + rc = dt_device_init(&dev->od_dt_dev, type); if (rc == 0) { rc = osd_device_init0(env, dev, cfg); @@ -680,7 +1126,6 @@ static struct lu_device *osd_device_free(const struct lu_env *env, struct osd_device *o = osd_dev(d); ENTRY; - cleanup_capa_hash(o->od_capa_hash); /* 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); @@ -699,7 +1144,6 @@ 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; @@ -707,16 +1151,10 @@ static struct lu_device *osd_device_fini(const struct lu_env *env, osd_shutdown(env, o); osd_oi_fini(env, o); - if (o->od_objset.os) { - 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); - arc_remove_prune_callback(o->arc_prune_cb); - o->arc_prune_cb = NULL; + if (o->od_os) { + osd_objset_unregister_callbacks(o); 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); @@ -725,7 +1163,7 @@ 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); RETURN(NULL); @@ -745,21 +1183,31 @@ 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) @@ -797,9 +1245,9 @@ static int osd_obd_connect(const struct lu_env *env, struct obd_export **exp, *exp = class_conn2export(&conn); - spin_lock(&osd->od_objset.lock); + spin_lock(&obd->obd_dev_lock); osd->od_connects++; - spin_unlock(&osd->od_objset.lock); + spin_unlock(&obd->obd_dev_lock); RETURN(0); } @@ -816,11 +1264,11 @@ static int osd_obd_disconnect(struct obd_export *exp) ENTRY; /* Only disconnect the underlying layers on the final disconnect. */ - spin_lock(&osd->od_objset.lock); + spin_lock(&obd->obd_dev_lock); osd->od_connects--; if (osd->od_connects == 0) release = 1; - spin_unlock(&osd->od_objset.lock); + spin_unlock(&obd->obd_dev_lock); rc = class_disconnect(exp); /* bz 9811 */ @@ -829,6 +1277,33 @@ static int osd_obd_disconnect(struct obd_export *exp) 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) { @@ -836,9 +1311,14 @@ 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 != 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); } @@ -858,6 +1338,14 @@ 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, @@ -883,10 +1371,11 @@ 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_disconnect = osd_obd_disconnect, + .o_fid_alloc = osd_fid_alloc }; -int __init osd_init(void) +static int __init osd_init(void) { int rc; @@ -898,29 +1387,27 @@ int __init osd_init(void) if (rc) return rc; - rc = class_register_type(&osd_obd_device_ops, NULL, NULL, -#ifndef HAVE_ONLY_PROCFS_SEQ - lprocfs_osd_module_vars, -#endif - LUSTRE_OSD_ZFS_NAME, &osd_device_type); + 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) +static 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_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_AUTHOR("Sun Microsystems, Inc. "); +MODULE_AUTHOR("OpenSFS, Inc. "); MODULE_DESCRIPTION("Lustre Object Storage Device ("LUSTRE_OSD_ZFS_NAME")"); +MODULE_VERSION(LUSTRE_VERSION_STRING); MODULE_LICENSE("GPL"); -cfs_module(osd, LUSTRE_VERSION_STRING, osd_init, osd_exit); +module_init(osd_init); +module_exit(osd_exit);