X-Git-Url: https://git.camperquake.de/gitweb.cgi?a=blobdiff_plain;f=module%2Fzfs%2Fsa.c;h=581cf4b0daf308e0811f13328062a7925bc539fb;hb=refs%2Fheads%2Frertzinger%2Ffeature-zpool-get--p;hp=d5c985bf13cfcbbc67a7a83e1b34a60998e515e1;hpb=d6320ddb78fa89c4d0fc2af00ae53c7c70992f96;p=zfs.git diff --git a/module/zfs/sa.c b/module/zfs/sa.c index d5c985b..581cf4b 100644 --- a/module/zfs/sa.c +++ b/module/zfs/sa.c @@ -18,8 +18,10 @@ * * CDDL HEADER END */ + /* * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved. + * Copyright (c) 2012 by Delphix. All rights reserved. */ #include @@ -134,7 +136,7 @@ static int sa_modify_attrs(sa_handle_t *hdl, sa_attr_type_t newattr, sa_data_op_t action, sa_data_locator_t *locator, void *datastart, uint16_t buflen, dmu_tx_t *tx); -arc_byteswap_func_t *sa_bswap_table[] = { +arc_byteswap_func_t sa_bswap_table[] = { byteswap_uint64_array, byteswap_uint32_array, byteswap_uint16_array, @@ -201,6 +203,7 @@ sa_attr_type_t sa_dummy_zpl_layout[] = { 0 }; static int sa_legacy_attr_count = 16; static kmem_cache_t *sa_cache = NULL; +static kmem_cache_t *spill_cache = NULL; /*ARGSUSED*/ static int @@ -232,6 +235,8 @@ sa_cache_init(void) sa_cache = kmem_cache_create("sa_cache", sizeof (sa_handle_t), 0, sa_cache_constructor, sa_cache_destructor, NULL, NULL, NULL, 0); + spill_cache = kmem_cache_create("spill_cache", + SPA_MAXBLOCKSIZE, 0, NULL, NULL, NULL, NULL, NULL, 0); } void @@ -239,6 +244,21 @@ sa_cache_fini(void) { if (sa_cache) kmem_cache_destroy(sa_cache); + + if (spill_cache) + kmem_cache_destroy(spill_cache); +} + +void * +sa_spill_alloc(int flags) +{ + return kmem_cache_alloc(spill_cache, flags); +} + +void +sa_spill_free(void *obj) +{ + kmem_cache_free(spill_cache, obj); } static int @@ -398,6 +418,8 @@ sa_attr_op(sa_handle_t *hdl, sa_bulk_attr_t *bulk, int count, if (error) return (error); break; + default: + break; } } return (error); @@ -413,10 +435,10 @@ sa_add_layout_entry(objset_t *os, sa_attr_type_t *attrs, int attr_count, avl_index_t loc; ASSERT(MUTEX_HELD(&sa->sa_lock)); - tb = kmem_zalloc(sizeof (sa_lot_t), KM_SLEEP); + tb = kmem_zalloc(sizeof (sa_lot_t), KM_PUSHPAGE); tb->lot_attr_count = attr_count; tb->lot_attrs = kmem_alloc(sizeof (sa_attr_type_t) * attr_count, - KM_SLEEP); + KM_PUSHPAGE); bcopy(attrs, tb->lot_attrs, sizeof (sa_attr_type_t) * attr_count); tb->lot_num = lot_num; tb->lot_hash = hash; @@ -426,10 +448,9 @@ sa_add_layout_entry(objset_t *os, sa_attr_type_t *attrs, int attr_count, char attr_name[8]; if (sa->sa_layout_attr_obj == 0) { - sa->sa_layout_attr_obj = zap_create(os, - DMU_OT_SA_ATTR_LAYOUTS, DMU_OT_NONE, 0, tx); - VERIFY(zap_add(os, sa->sa_master_obj, SA_LAYOUTS, 8, 1, - &sa->sa_layout_attr_obj, tx) == 0); + sa->sa_layout_attr_obj = zap_create_link(os, + DMU_OT_SA_ATTR_LAYOUTS, + sa->sa_master_obj, SA_LAYOUTS, tx); } (void) snprintf(attr_name, sizeof (attr_name), @@ -551,6 +572,7 @@ sa_find_sizes(sa_os_t *sa, sa_bulk_attr_t *attr_desc, int attr_count, { int var_size = 0; int i; + int j = -1; int full_space; int hdrsize; boolean_t done = B_FALSE; @@ -572,10 +594,12 @@ sa_find_sizes(sa_os_t *sa, sa_bulk_attr_t *attr_desc, int attr_count, sizeof (sa_hdr_phys_t); full_space = (buftype == SA_BONUS) ? DN_MAX_BONUSLEN : db->db_size; + ASSERT(IS_P2ALIGNED(full_space, 8)); for (i = 0; i != attr_count; i++) { boolean_t is_var_sz; + *total = P2ROUNDUP(*total, 8); *total += attr_desc[i].sa_length; if (done) goto next; @@ -588,7 +612,14 @@ sa_find_sizes(sa_os_t *sa, sa_bulk_attr_t *attr_desc, int attr_count, if (is_var_sz && var_size > 1) { if (P2ROUNDUP(hdrsize + sizeof (uint16_t), 8) + *total < full_space) { + /* + * Account for header space used by array of + * optional sizes of variable-length attributes. + * Record the index in case this increase needs + * to be reversed due to spill-over. + */ hdrsize += sizeof (uint16_t); + j = i; } else { done = B_TRUE; *index = i; @@ -605,18 +636,26 @@ sa_find_sizes(sa_os_t *sa, sa_bulk_attr_t *attr_desc, int attr_count, * and spill buffer. */ if (buftype == SA_BONUS && *index == -1 && - P2ROUNDUP(*total + hdrsize, 8) > + (*total + P2ROUNDUP(hdrsize, 8)) > (full_space - sizeof (blkptr_t))) { *index = i; done = B_TRUE; } next: - if (P2ROUNDUP(*total + hdrsize, 8) > full_space && + if ((*total + P2ROUNDUP(hdrsize, 8)) > full_space && buftype == SA_BONUS) *will_spill = B_TRUE; } + /* + * j holds the index of the last variable-sized attribute for + * which hdrsize was increased. Reverse the increase if that + * attribute will be relocated to the spill block. + */ + if (*will_spill && j == *index) + hdrsize -= sizeof (uint16_t); + hdrsize = P2ROUNDUP(hdrsize, 8); return (hdrsize); } @@ -640,7 +679,7 @@ sa_build_layouts(sa_handle_t *hdl, sa_bulk_attr_t *attr_desc, int attr_count, int buf_space; sa_attr_type_t *attrs, *attrs_start; int i, lot_count; - int hdrsize, spillhdrsize; + int hdrsize, spillhdrsize = 0; int used; dmu_object_type_t bonustype; sa_lot_t *lot; @@ -701,18 +740,21 @@ sa_build_layouts(sa_handle_t *hdl, sa_bulk_attr_t *attr_desc, int attr_count, buf_space = hdl->sa_bonus->db_size - hdrsize; attrs_start = attrs = kmem_alloc(sizeof (sa_attr_type_t) * attr_count, - KM_SLEEP); + KM_PUSHPAGE); lot_count = 0; for (i = 0, len_idx = 0, hash = -1ULL; i != attr_count; i++) { uint16_t length; + ASSERT(IS_P2ALIGNED(data_start, 8)); + ASSERT(IS_P2ALIGNED(buf_space, 8)); attrs[i] = attr_desc[i].sa_attr; length = SA_REGISTERED_LEN(sa, attrs[i]); if (length == 0) length = attr_desc[i].sa_length; if (buf_space < length) { /* switch to spill buffer */ + VERIFY(spilling); VERIFY(bonustype == DMU_OT_SA); if (buftype == SA_BONUS && !sa->sa_force_spill) { sa_find_layout(hdl->sa_os, hash, attrs_start, @@ -811,7 +853,7 @@ sa_attr_table_setup(objset_t *os, sa_attr_reg_t *reg_attrs, int count) { sa_os_t *sa = os->os_sa; uint64_t sa_attr_count = 0; - uint64_t sa_reg_count; + uint64_t sa_reg_count = 0; int error = 0; uint64_t attr_value; sa_attr_table_t *tb; @@ -822,7 +864,7 @@ sa_attr_table_setup(objset_t *os, sa_attr_reg_t *reg_attrs, int count) dmu_objset_type_t ostype = dmu_objset_type(os); sa->sa_user_table = - kmem_zalloc(count * sizeof (sa_attr_type_t), KM_SLEEP); + kmem_zalloc(count * sizeof (sa_attr_type_t), KM_PUSHPAGE); sa->sa_user_table_sz = count * sizeof (sa_attr_type_t); if (sa->sa_reg_attr_obj != 0) { @@ -881,7 +923,7 @@ sa_attr_table_setup(objset_t *os, sa_attr_reg_t *reg_attrs, int count) sa->sa_num_attrs = sa_attr_count; tb = sa->sa_attr_table = - kmem_zalloc(sizeof (sa_attr_table_t) * sa_attr_count, KM_SLEEP); + kmem_zalloc(sizeof (sa_attr_table_t) * sa_attr_count, KM_PUSHPAGE); /* * Attribute table is constructed from requested attribute list, @@ -906,7 +948,7 @@ sa_attr_table_setup(objset_t *os, sa_attr_reg_t *reg_attrs, int count) continue; } tb[ATTR_NUM(value)].sa_name = - kmem_zalloc(strlen(za.za_name) +1, KM_SLEEP); + kmem_zalloc(strlen(za.za_name) +1, KM_PUSHPAGE); (void) strlcpy(tb[ATTR_NUM(value)].sa_name, za.za_name, strlen(za.za_name) +1); } @@ -932,7 +974,7 @@ sa_attr_table_setup(objset_t *os, sa_attr_reg_t *reg_attrs, int count) tb[i].sa_registered = B_FALSE; tb[i].sa_name = kmem_zalloc(strlen(sa_legacy_attrs[i].sa_name) +1, - KM_SLEEP); + KM_PUSHPAGE); (void) strlcpy(tb[i].sa_name, sa_legacy_attrs[i].sa_name, strlen(sa_legacy_attrs[i].sa_name) + 1); @@ -950,7 +992,7 @@ sa_attr_table_setup(objset_t *os, sa_attr_reg_t *reg_attrs, int count) tb[attr_id].sa_byteswap = reg_attrs[i].sa_byteswap; tb[attr_id].sa_attr = attr_id; tb[attr_id].sa_name = - kmem_zalloc(strlen(reg_attrs[i].sa_name) + 1, KM_SLEEP); + kmem_zalloc(strlen(reg_attrs[i].sa_name) + 1, KM_PUSHPAGE); (void) strlcpy(tb[attr_id].sa_name, reg_attrs[i].sa_name, strlen(reg_attrs[i].sa_name) + 1); } @@ -987,7 +1029,7 @@ sa_setup(objset_t *os, uint64_t sa_obj, sa_attr_reg_t *reg_attrs, int count, return (0); } - sa = kmem_zalloc(sizeof (sa_os_t), KM_SLEEP); + sa = kmem_zalloc(sizeof (sa_os_t), KM_PUSHPAGE); mutex_init(&sa->sa_lock, NULL, MUTEX_DEFAULT, NULL); sa->sa_master_obj = sa_obj; @@ -1035,7 +1077,7 @@ sa_setup(objset_t *os, uint64_t sa_obj, sa_attr_reg_t *reg_attrs, int count, uint64_t lot_num; lot_attrs = kmem_zalloc(sizeof (sa_attr_type_t) * - za.za_num_integers, KM_SLEEP); + za.za_num_integers, KM_PUSHPAGE); if ((error = (zap_lookup(os, sa->sa_layout_attr_obj, za.za_name, 2, za.za_num_integers, @@ -1103,16 +1145,16 @@ sa_tear_down(objset_t *os) sa_free_attr_table(sa); cookie = NULL; - while (layout = avl_destroy_nodes(&sa->sa_layout_hash_tree, &cookie)) { + while ((layout = avl_destroy_nodes(&sa->sa_layout_hash_tree, &cookie))){ sa_idx_tab_t *tab; - while (tab = list_head(&layout->lot_idx_tab)) { + while ((tab = list_head(&layout->lot_idx_tab))) { ASSERT(refcount_count(&tab->sa_refcount)); sa_idx_tab_rele(os, tab); } } cookie = NULL; - while (layout = avl_destroy_nodes(&sa->sa_layout_num_tree, &cookie)) { + while ((layout = avl_destroy_nodes(&sa->sa_layout_num_tree, &cookie))){ kmem_free(layout->lot_attrs, sizeof (sa_attr_type_t) * layout->lot_attr_count); kmem_free(layout, sizeof (sa_lot_t)); @@ -1204,9 +1246,9 @@ sa_byteswap(sa_handle_t *hdl, sa_buf_type_t buftype) { sa_hdr_phys_t *sa_hdr_phys = SA_GET_HDR(hdl, buftype); dmu_buf_impl_t *db; - sa_os_t *sa = hdl->sa_os->os_sa; int num_lengths = 1; int i; + ASSERTV(sa_os_t *sa = hdl->sa_os->os_sa); ASSERT(MUTEX_HELD(&sa->sa_lock)); if (sa_hdr_phys->sa_magic == SA_MAGIC) @@ -1307,13 +1349,26 @@ sa_idx_tab_rele(objset_t *os, void *arg) static void sa_idx_tab_hold(objset_t *os, sa_idx_tab_t *idx_tab) { - sa_os_t *sa = os->os_sa; + ASSERTV(sa_os_t *sa = os->os_sa); ASSERT(MUTEX_HELD(&sa->sa_lock)); (void) refcount_add(&idx_tab->sa_refcount, NULL); } void +sa_spill_rele(sa_handle_t *hdl) +{ + mutex_enter(&hdl->sa_lock); + if (hdl->sa_spill) { + sa_idx_tab_rele(hdl->sa_os, hdl->sa_spill_tab); + dmu_buf_rele(hdl->sa_spill, NULL); + hdl->sa_spill = NULL; + hdl->sa_spill_tab = NULL; + } + mutex_exit(&hdl->sa_lock); +} + +void sa_handle_destroy(sa_handle_t *hdl) { mutex_enter(&hdl->sa_lock); @@ -1343,10 +1398,10 @@ sa_handle_get_from_db(objset_t *os, dmu_buf_t *db, void *userp, sa_handle_type_t hdl_type, sa_handle_t **handlepp) { int error = 0; - dmu_object_info_t doi; sa_handle_t *handle; - #ifdef ZFS_DEBUG + dmu_object_info_t doi; + dmu_object_info_from_db(db, &doi); ASSERT(doi.doi_bonus_type == DMU_OT_SA || doi.doi_bonus_type == DMU_OT_ZNODE); @@ -1385,7 +1440,7 @@ sa_handle_get(objset_t *objset, uint64_t objid, void *userp, dmu_buf_t *db; int error; - if (error = dmu_bonus_hold(objset, objid, NULL, &db)) + if ((error = dmu_bonus_hold(objset, objid, NULL, &db))) return (error); return (sa_handle_get_from_db(objset, db, userp, hdl_type, @@ -1450,7 +1505,6 @@ sa_lookup_uio(sa_handle_t *hdl, sa_attr_type_t attr, uio_t *uio) } mutex_exit(&hdl->sa_lock); return (error); - } #endif @@ -1508,14 +1562,14 @@ sa_find_idx_tab(objset_t *os, dmu_object_type_t bonustype, void *data) } /* No such luck, create a new entry */ - idx_tab = kmem_zalloc(sizeof (sa_idx_tab_t), KM_SLEEP); + idx_tab = kmem_zalloc(sizeof (sa_idx_tab_t), KM_PUSHPAGE); idx_tab->sa_idx_tab = - kmem_zalloc(sizeof (uint32_t) * sa->sa_num_attrs, KM_SLEEP); + kmem_zalloc(sizeof (uint32_t) * sa->sa_num_attrs, KM_PUSHPAGE); idx_tab->sa_layout = tb; refcount_create(&idx_tab->sa_refcount); if (tb->lot_var_sizes) idx_tab->sa_variable_lengths = kmem_alloc(sizeof (uint16_t) * - tb->lot_var_sizes, KM_SLEEP); + tb->lot_var_sizes, KM_PUSHPAGE); sa_attr_iter(os, hdr, bonustype, sa_build_idx_tab, tb, idx_tab); @@ -1545,16 +1599,15 @@ sa_attr_register_sync(sa_handle_t *hdl, dmu_tx_t *tx) mutex_enter(&sa->sa_lock); - if (!sa->sa_need_attr_registration || sa->sa_master_obj == NULL) { + if (!sa->sa_need_attr_registration || sa->sa_master_obj == 0) { mutex_exit(&sa->sa_lock); return; } - if (sa->sa_reg_attr_obj == NULL) { - sa->sa_reg_attr_obj = zap_create(hdl->sa_os, - DMU_OT_SA_ATTR_REGISTRATION, DMU_OT_NONE, 0, tx); - VERIFY(zap_add(hdl->sa_os, sa->sa_master_obj, - SA_REGISTRY, 8, 1, &sa->sa_reg_attr_obj, tx) == 0); + if (sa->sa_reg_attr_obj == 0) { + sa->sa_reg_attr_obj = zap_create_link(hdl->sa_os, + DMU_OT_SA_ATTR_REGISTRATION, + sa->sa_master_obj, SA_REGISTRY, tx); } for (i = 0; i != sa->sa_num_attrs; i++) { if (sa->sa_attr_table[i].sa_registered) @@ -1617,7 +1670,7 @@ sa_modify_attrs(sa_handle_t *hdl, sa_attr_type_t newattr, sa_bulk_attr_t *attr_desc; void *old_data[2]; int bonus_attr_count = 0; - int bonus_data_size, spill_data_size; + int bonus_data_size = 0; int spill_attr_count = 0; int error; uint16_t length; @@ -1647,8 +1700,8 @@ sa_modify_attrs(sa_handle_t *hdl, sa_attr_type_t newattr, /* Bring spill buffer online if it isn't currently */ if ((error = sa_get_spill(hdl)) == 0) { - spill_data_size = hdl->sa_spill->db_size; - old_data[1] = kmem_alloc(spill_data_size, KM_SLEEP); + ASSERT3U(hdl->sa_spill->db_size, <=, SPA_MAXBLOCKSIZE); + old_data[1] = sa_spill_alloc(KM_SLEEP); bcopy(hdl->sa_spill->db_data, old_data[1], hdl->sa_spill->db_size); spill_attr_count = @@ -1728,7 +1781,7 @@ sa_modify_attrs(sa_handle_t *hdl, sa_attr_type_t newattr, if (old_data[0]) kmem_free(old_data[0], bonus_data_size); if (old_data[1]) - kmem_free(old_data[1], spill_data_size); + sa_spill_free(old_data[1]); kmem_free(attr_desc, sizeof (sa_bulk_attr_t) * attr_count); return (error); @@ -1741,12 +1794,14 @@ sa_bulk_update_impl(sa_handle_t *hdl, sa_bulk_attr_t *bulk, int count, int error; sa_os_t *sa = hdl->sa_os->os_sa; dmu_object_type_t bonustype; - - bonustype = SA_BONUSTYPE_FROM_DB(SA_GET_DB(hdl, SA_BONUS)); + dmu_buf_t *saved_spill; ASSERT(hdl); ASSERT(MUTEX_HELD(&hdl->sa_lock)); + bonustype = SA_BONUSTYPE_FROM_DB(SA_GET_DB(hdl, SA_BONUS)); + saved_spill = hdl->sa_spill; + /* sync out registration table if necessary */ if (sa->sa_need_attr_registration) sa_attr_register_sync(hdl, tx); @@ -1755,6 +1810,24 @@ sa_bulk_update_impl(sa_handle_t *hdl, sa_bulk_attr_t *bulk, int count, if (error == 0 && !IS_SA_BONUSTYPE(bonustype) && sa->sa_update_cb) sa->sa_update_cb(hdl, tx); + /* + * If saved_spill is NULL and current sa_spill is not NULL that + * means we increased the refcount of the spill buffer through + * sa_get_spill() or dmu_spill_hold_by_dnode(). Therefore we + * must release the hold before calling dmu_tx_commit() to avoid + * making a copy of this buffer in dbuf_sync_leaf() due to the + * reference count now being greater than 1. + */ + if (!saved_spill && hdl->sa_spill) { + if (hdl->sa_spill_tab) { + sa_idx_tab_rele(hdl->sa_os, hdl->sa_spill_tab); + hdl->sa_spill_tab = NULL; + } + + dmu_buf_rele((dmu_buf_t *)hdl->sa_spill, NULL); + hdl->sa_spill = NULL; + } + return (error); } @@ -1968,3 +2041,41 @@ sa_handle_unlock(sa_handle_t *hdl) ASSERT(hdl); mutex_exit(&hdl->sa_lock); } + +#ifdef _KERNEL +EXPORT_SYMBOL(sa_handle_get); +EXPORT_SYMBOL(sa_handle_get_from_db); +EXPORT_SYMBOL(sa_handle_destroy); +EXPORT_SYMBOL(sa_buf_hold); +EXPORT_SYMBOL(sa_buf_rele); +EXPORT_SYMBOL(sa_spill_rele); +EXPORT_SYMBOL(sa_lookup); +EXPORT_SYMBOL(sa_update); +EXPORT_SYMBOL(sa_remove); +EXPORT_SYMBOL(sa_bulk_lookup); +EXPORT_SYMBOL(sa_bulk_lookup_locked); +EXPORT_SYMBOL(sa_bulk_update); +EXPORT_SYMBOL(sa_size); +EXPORT_SYMBOL(sa_update_from_cb); +EXPORT_SYMBOL(sa_object_info); +EXPORT_SYMBOL(sa_object_size); +EXPORT_SYMBOL(sa_update_user); +EXPORT_SYMBOL(sa_get_userdata); +EXPORT_SYMBOL(sa_set_userp); +EXPORT_SYMBOL(sa_get_db); +EXPORT_SYMBOL(sa_handle_object); +EXPORT_SYMBOL(sa_register_update_callback); +EXPORT_SYMBOL(sa_setup); +EXPORT_SYMBOL(sa_replace_all_by_template); +EXPORT_SYMBOL(sa_replace_all_by_template_locked); +EXPORT_SYMBOL(sa_enabled); +EXPORT_SYMBOL(sa_cache_init); +EXPORT_SYMBOL(sa_cache_fini); +EXPORT_SYMBOL(sa_spill_alloc); +EXPORT_SYMBOL(sa_spill_free); +EXPORT_SYMBOL(sa_set_sa_object); +EXPORT_SYMBOL(sa_hdrsize); +EXPORT_SYMBOL(sa_handle_lock); +EXPORT_SYMBOL(sa_handle_unlock); +EXPORT_SYMBOL(sa_lookup_uio); +#endif /* _KERNEL */