dbuf.c revision 290754
179110Sjoerg/* 279110Sjoerg * CDDL HEADER START 379110Sjoerg * 479110Sjoerg * The contents of this file are subject to the terms of the 579110Sjoerg * Common Development and Distribution License (the "License"). 679110Sjoerg * You may not use this file except in compliance with the License. 779110Sjoerg * 879110Sjoerg * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 979110Sjoerg * or http://www.opensolaris.org/os/licensing. 1079110Sjoerg * See the License for the specific language governing permissions 1179110Sjoerg * and limitations under the License. 1279110Sjoerg * 1379110Sjoerg * When distributing Covered Code, include this CDDL HEADER in each 1479110Sjoerg * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 1579110Sjoerg * If applicable, add the following below this CDDL HEADER, with the 1679110Sjoerg * fields enclosed by brackets "[]" replaced with your own identifying 1779110Sjoerg * information: Portions Copyright [yyyy] [name of copyright owner] 1879110Sjoerg * 1979110Sjoerg * CDDL HEADER END 2079110Sjoerg */ 2179110Sjoerg/* 2279110Sjoerg * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 2379110Sjoerg * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 2479110Sjoerg * Copyright (c) 2012, 2015 by Delphix. All rights reserved. 2579110Sjoerg * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 2679110Sjoerg * Copyright (c) 2013, Joyent, Inc. All rights reserved. 2779110Sjoerg * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 2879110Sjoerg */ 2987992Sjoerg 3087992Sjoerg#include <sys/zfs_context.h> 3179110Sjoerg#include <sys/dmu.h> 3279110Sjoerg#include <sys/dmu_send.h> 3387992Sjoerg#include <sys/dmu_impl.h> 3479110Sjoerg#include <sys/dbuf.h> 3587992Sjoerg#include <sys/dmu_objset.h> 3679110Sjoerg#include <sys/dsl_dataset.h> 3787992Sjoerg#include <sys/dsl_dir.h> 3879110Sjoerg#include <sys/dmu_tx.h> 3979110Sjoerg#include <sys/spa.h> 4079110Sjoerg#include <sys/zio.h> 4187992Sjoerg#include <sys/dmu_zfetch.h> 4287992Sjoerg#include <sys/sa.h> 4387992Sjoerg#include <sys/sa_impl.h> 4487992Sjoerg#include <sys/zfeature.h> 4587992Sjoerg#include <sys/blkptr.h> 4679110Sjoerg#include <sys/range_tree.h> 4779110Sjoerg 4879110Sjoerg/* 4979110Sjoerg * Number of times that zfs_free_range() took the slow path while doing 5079110Sjoerg * a zfs receive. A nonzero value indicates a potential performance problem. 5179110Sjoerg */ 5279110Sjoerguint64_t zfs_free_range_recv_miss; 5379110Sjoerg 5479110Sjoergstatic void dbuf_destroy(dmu_buf_impl_t *db); 5579110Sjoergstatic boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 5679110Sjoergstatic void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 5779110Sjoerg 5879110Sjoerg/* 5979110Sjoerg * Global data structures and functions for the dbuf cache. 6079110Sjoerg */ 6179110Sjoergstatic kmem_cache_t *dbuf_cache; 62134081Sphkstatic taskq_t *dbu_evict_taskq; 63134081Sphk 64134081Sphk/* ARGSUSED */ 6579110Sjoergstatic int 6679110Sjoergdbuf_cons(void *vdb, void *unused, int kmflag) 6779110Sjoerg{ 6879110Sjoerg dmu_buf_impl_t *db = vdb; 6979110Sjoerg bzero(db, sizeof (dmu_buf_impl_t)); 7079110Sjoerg 7179110Sjoerg mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 7279110Sjoerg cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 7379110Sjoerg refcount_create(&db->db_holds); 7479110Sjoerg 7579110Sjoerg return (0); 7679110Sjoerg} 7779110Sjoerg 7879110Sjoerg/* ARGSUSED */ 7979110Sjoergstatic void 8079110Sjoergdbuf_dest(void *vdb, void *unused) 8179110Sjoerg{ 8279110Sjoerg dmu_buf_impl_t *db = vdb; 8379110Sjoerg mutex_destroy(&db->db_mtx); 8479110Sjoerg cv_destroy(&db->db_changed); 8579110Sjoerg refcount_destroy(&db->db_holds); 8679110Sjoerg} 8779110Sjoerg 8879110Sjoerg/* 8979110Sjoerg * dbuf hash table routines 90134081Sphk */ 91134081Sphkstatic dbuf_hash_table_t dbuf_hash_table; 92126230Sphk 93127522Snyanstatic uint64_t dbuf_hash_count; 94134081Sphk 95274024Snyanstatic uint64_t 96127522Snyandbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 97134081Sphk{ 98127522Snyan uintptr_t osv = (uintptr_t)os; 99134081Sphk uint64_t crc = -1ULL; 100134081Sphk 101134081Sphk ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 102134081Sphk crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 103134081Sphk crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 104134081Sphk crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 105134081Sphk crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 106274024Snyan crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 107139905Sdelphij crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 108127522Snyan 109127522Snyan crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 110127522Snyan 111137459Snyan return (crc); 112134081Sphk} 113134081Sphk 114134081Sphk#define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 115137459Snyan 116137459Snyan#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 117137459Snyan ((dbuf)->db.db_object == (obj) && \ 118139905Sdelphij (dbuf)->db_objset == (os) && \ 119137459Snyan (dbuf)->db_level == (level) && \ 120137459Snyan (dbuf)->db_blkid == (blkid)) 121137459Snyan 122137459Snyandmu_buf_impl_t * 123137459Snyandbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 124137459Snyan{ 125137459Snyan dbuf_hash_table_t *h = &dbuf_hash_table; 126137459Snyan uint64_t hv = DBUF_HASH(os, obj, level, blkid); 127139905Sdelphij uint64_t idx = hv & h->hash_table_mask; 128137459Snyan dmu_buf_impl_t *db; 129127522Snyan 130127522Snyan mutex_enter(DBUF_HASH_MUTEX(h, idx)); 131134081Sphk for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 132137459Snyan if (DBUF_EQUAL(db, os, obj, level, blkid)) { 133134081Sphk mutex_enter(&db->db_mtx); 134137459Snyan if (db->db_state != DB_EVICTING) { 135137459Snyan mutex_exit(DBUF_HASH_MUTEX(h, idx)); 136137459Snyan return (db); 137134081Sphk } 138139905Sdelphij mutex_exit(&db->db_mtx); 139137459Snyan } 140137459Snyan } 141137459Snyan mutex_exit(DBUF_HASH_MUTEX(h, idx)); 142134081Sphk return (NULL); 143134081Sphk} 144134081Sphk 145134081Sphkstatic dmu_buf_impl_t * 146134081Sphkdbuf_find_bonus(objset_t *os, uint64_t object) 147134081Sphk{ 148134081Sphk dnode_t *dn; 149139905Sdelphij dmu_buf_impl_t *db = NULL; 150137459Snyan 15187992Sjoerg if (dnode_hold(os, object, FTAG, &dn) == 0) { 15287992Sjoerg rw_enter(&dn->dn_struct_rwlock, RW_READER); 15387992Sjoerg if (dn->dn_bonus != NULL) { 15487992Sjoerg db = dn->dn_bonus; 155274024Snyan mutex_enter(&db->db_mtx); 156134081Sphk } 157274024Snyan rw_exit(&dn->dn_struct_rwlock); 158139905Sdelphij dnode_rele(dn, FTAG); 15987992Sjoerg } 16087992Sjoerg return (db); 16187992Sjoerg} 16287992Sjoerg 163274024Snyan/* 164134081Sphk * Insert an entry into the hash table. If there is already an element 165274024Snyan * equal to elem in the hash table, then the already existing element 166139905Sdelphij * will be returned and the new element will not be inserted. 16787992Sjoerg * Otherwise returns NULL. 16887992Sjoerg */ 169127522Snyanstatic dmu_buf_impl_t * 17087992Sjoergdbuf_hash_insert(dmu_buf_impl_t *db) 17187992Sjoerg{ 17287992Sjoerg dbuf_hash_table_t *h = &dbuf_hash_table; 17387992Sjoerg objset_t *os = db->db_objset; 17487992Sjoerg uint64_t obj = db->db.db_object; 17587992Sjoerg int level = db->db_level; 17687992Sjoerg uint64_t blkid = db->db_blkid; 17787992Sjoerg uint64_t hv = DBUF_HASH(os, obj, level, blkid); 17887992Sjoerg uint64_t idx = hv & h->hash_table_mask; 17987992Sjoerg dmu_buf_impl_t *dbf; 18087992Sjoerg 18187992Sjoerg mutex_enter(DBUF_HASH_MUTEX(h, idx)); 18287992Sjoerg for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 18387992Sjoerg if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 18487992Sjoerg mutex_enter(&dbf->db_mtx); 18587992Sjoerg if (dbf->db_state != DB_EVICTING) { 18687992Sjoerg mutex_exit(DBUF_HASH_MUTEX(h, idx)); 18787992Sjoerg return (dbf); 18887992Sjoerg } 18987992Sjoerg mutex_exit(&dbf->db_mtx); 19087992Sjoerg } 19187992Sjoerg } 19287992Sjoerg 19387992Sjoerg mutex_enter(&db->db_mtx); 19487992Sjoerg db->db_hash_next = h->hash_table[idx]; 19587992Sjoerg h->hash_table[idx] = db; 19687992Sjoerg mutex_exit(DBUF_HASH_MUTEX(h, idx)); 19787992Sjoerg atomic_inc_64(&dbuf_hash_count); 19887992Sjoerg 19987992Sjoerg return (NULL); 20087992Sjoerg} 20187992Sjoerg 20287992Sjoerg/* 20387992Sjoerg * Remove an entry from the hash table. It must be in the EVICTING state. 20487992Sjoerg */ 20587992Sjoergstatic void 20687992Sjoergdbuf_hash_remove(dmu_buf_impl_t *db) 20787992Sjoerg{ 20887992Sjoerg dbuf_hash_table_t *h = &dbuf_hash_table; 20987992Sjoerg uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 21087992Sjoerg db->db_level, db->db_blkid); 21187992Sjoerg uint64_t idx = hv & h->hash_table_mask; 21287992Sjoerg dmu_buf_impl_t *dbf, **dbp; 21387992Sjoerg 21487992Sjoerg /* 21587992Sjoerg * We musn't hold db_mtx to maintain lock ordering: 21687992Sjoerg * DBUF_HASH_MUTEX > db_mtx. 21787992Sjoerg */ 21887992Sjoerg ASSERT(refcount_is_zero(&db->db_holds)); 21987992Sjoerg ASSERT(db->db_state == DB_EVICTING); 22087992Sjoerg ASSERT(!MUTEX_HELD(&db->db_mtx)); 22187992Sjoerg 22287992Sjoerg mutex_enter(DBUF_HASH_MUTEX(h, idx)); 22387992Sjoerg dbp = &h->hash_table[idx]; 22487992Sjoerg while ((dbf = *dbp) != db) { 22587992Sjoerg dbp = &dbf->db_hash_next; 22687992Sjoerg ASSERT(dbf != NULL); 22787992Sjoerg } 22887992Sjoerg *dbp = db->db_hash_next; 22987992Sjoerg db->db_hash_next = NULL; 23087992Sjoerg mutex_exit(DBUF_HASH_MUTEX(h, idx)); 23187992Sjoerg atomic_dec_64(&dbuf_hash_count); 23287992Sjoerg} 23387992Sjoerg 23487992Sjoergstatic arc_evict_func_t dbuf_do_evict; 23587992Sjoerg 23687992Sjoergtypedef enum { 23787992Sjoerg DBVU_EVICTING, 23887992Sjoerg DBVU_NOT_EVICTING 23987992Sjoerg} dbvu_verify_type_t; 24087992Sjoerg 24187992Sjoergstatic void 24287992Sjoergdbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 24387992Sjoerg{ 24487992Sjoerg#ifdef ZFS_DEBUG 24587992Sjoerg int64_t holds; 24687992Sjoerg 24787992Sjoerg if (db->db_user == NULL) 24887992Sjoerg return; 24987992Sjoerg 25087992Sjoerg /* Only data blocks support the attachment of user data. */ 25187992Sjoerg ASSERT(db->db_level == 0); 25287992Sjoerg 25387992Sjoerg /* Clients must resolve a dbuf before attaching user data. */ 25487992Sjoerg ASSERT(db->db.db_data != NULL); 25587992Sjoerg ASSERT3U(db->db_state, ==, DB_CACHED); 25687992Sjoerg 25787992Sjoerg holds = refcount_count(&db->db_holds); 25887992Sjoerg if (verify_type == DBVU_EVICTING) { 25987992Sjoerg /* 26087992Sjoerg * Immediate eviction occurs when holds == dirtycnt. 26187992Sjoerg * For normal eviction buffers, holds is zero on 26287992Sjoerg * eviction, except when dbuf_fix_old_data() calls 26387992Sjoerg * dbuf_clear_data(). However, the hold count can grow 26487992Sjoerg * during eviction even though db_mtx is held (see 26587992Sjoerg * dmu_bonus_hold() for an example), so we can only 26687992Sjoerg * test the generic invariant that holds >= dirtycnt. 26787992Sjoerg */ 26887992Sjoerg ASSERT3U(holds, >=, db->db_dirtycnt); 26987992Sjoerg } else { 27087992Sjoerg if (db->db_user_immediate_evict == TRUE) 27187992Sjoerg ASSERT3U(holds, >=, db->db_dirtycnt); 27287992Sjoerg else 27387992Sjoerg ASSERT3U(holds, >, 0); 27487992Sjoerg } 27587992Sjoerg#endif 27687992Sjoerg} 27787992Sjoerg 27887992Sjoergstatic void 27987992Sjoergdbuf_evict_user(dmu_buf_impl_t *db) 280127522Snyan{ 28187992Sjoerg dmu_buf_user_t *dbu = db->db_user; 28287992Sjoerg 28387992Sjoerg ASSERT(MUTEX_HELD(&db->db_mtx)); 28487992Sjoerg 28587992Sjoerg if (dbu == NULL) 28687992Sjoerg return; 287134081Sphk 288134081Sphk dbuf_verify_user(db, DBVU_EVICTING); 28987992Sjoerg db->db_user = NULL; 29087992Sjoerg 29187992Sjoerg#ifdef ZFS_DEBUG 292127522Snyan if (dbu->dbu_clear_on_evict_dbufp != NULL) 29387992Sjoerg *dbu->dbu_clear_on_evict_dbufp = NULL; 29487992Sjoerg#endif 29587992Sjoerg 29687992Sjoerg /* 29787992Sjoerg * Invoke the callback from a taskq to avoid lock order reversals 29887992Sjoerg * and limit stack depth. 29987992Sjoerg */ 30087992Sjoerg taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, 30187992Sjoerg &dbu->dbu_tqent); 30287992Sjoerg} 30387992Sjoerg 30487992Sjoergboolean_t 305127522Snyandbuf_is_metadata(dmu_buf_impl_t *db) 30687992Sjoerg{ 30787992Sjoerg if (db->db_level > 0) { 30887992Sjoerg return (B_TRUE); 30987992Sjoerg } else { 31087992Sjoerg boolean_t is_metadata; 31187992Sjoerg 31287992Sjoerg DB_DNODE_ENTER(db); 31387992Sjoerg is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 31487992Sjoerg DB_DNODE_EXIT(db); 31587992Sjoerg 31687992Sjoerg return (is_metadata); 31787992Sjoerg } 31887992Sjoerg} 31987992Sjoerg 32087992Sjoergvoid 32187992Sjoergdbuf_evict(dmu_buf_impl_t *db) 32287992Sjoerg{ 32387992Sjoerg ASSERT(MUTEX_HELD(&db->db_mtx)); 32487992Sjoerg ASSERT(db->db_buf == NULL); 32587992Sjoerg ASSERT(db->db_data_pending == NULL); 32687992Sjoerg 32787992Sjoerg dbuf_clear(db); 32887992Sjoerg dbuf_destroy(db); 32987992Sjoerg} 33087992Sjoerg 33187992Sjoergvoid 33287992Sjoergdbuf_init(void) 33387992Sjoerg{ 33487992Sjoerg uint64_t hsize = 1ULL << 16; 33587992Sjoerg dbuf_hash_table_t *h = &dbuf_hash_table; 33687992Sjoerg int i; 33787992Sjoerg 33887992Sjoerg /* 33987992Sjoerg * The hash table is big enough to fill all of physical memory 34087992Sjoerg * with an average 4K block size. The table will take up 34187992Sjoerg * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 34287992Sjoerg */ 343134081Sphk while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 344134081Sphk hsize <<= 1; 345134081Sphk 346134081Sphkretry: 34787992Sjoerg h->hash_table_mask = hsize - 1; 34887992Sjoerg h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 34987992Sjoerg if (h->hash_table == NULL) { 35087992Sjoerg /* XXX - we should really return an error instead of assert */ 35187992Sjoerg ASSERT(hsize > (1ULL << 10)); 35287992Sjoerg hsize >>= 1; 35387992Sjoerg goto retry; 35487992Sjoerg } 35587992Sjoerg 35687992Sjoerg dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 35787992Sjoerg sizeof (dmu_buf_impl_t), 35887992Sjoerg 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 35987992Sjoerg 36087992Sjoerg for (i = 0; i < DBUF_MUTEXES; i++) 36187992Sjoerg mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 362126230Sphk 36387992Sjoerg /* 36487992Sjoerg * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 36587992Sjoerg * configuration is not required. 36687992Sjoerg */ 36787992Sjoerg dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 36887992Sjoerg} 36987992Sjoerg 37087992Sjoergvoid 37187992Sjoergdbuf_fini(void) 37287992Sjoerg{ 37387992Sjoerg dbuf_hash_table_t *h = &dbuf_hash_table; 37487992Sjoerg int i; 37587992Sjoerg 37687992Sjoerg for (i = 0; i < DBUF_MUTEXES; i++) 37787992Sjoerg mutex_destroy(&h->hash_mutexes[i]); 37887992Sjoerg kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 37987992Sjoerg kmem_cache_destroy(dbuf_cache); 38087992Sjoerg taskq_destroy(dbu_evict_taskq); 38187992Sjoerg} 38287992Sjoerg 38387992Sjoerg/* 38487992Sjoerg * Other stuff. 38587992Sjoerg */ 38687992Sjoerg 38787992Sjoerg#ifdef ZFS_DEBUG 38887992Sjoergstatic void 38987992Sjoergdbuf_verify(dmu_buf_impl_t *db) 39087992Sjoerg{ 39187992Sjoerg dnode_t *dn; 39287992Sjoerg dbuf_dirty_record_t *dr; 393134081Sphk 394134081Sphk ASSERT(MUTEX_HELD(&db->db_mtx)); 39587992Sjoerg 39687992Sjoerg if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 39787992Sjoerg return; 39887992Sjoerg 39987992Sjoerg ASSERT(db->db_objset != NULL); 40087992Sjoerg DB_DNODE_ENTER(db); 40187992Sjoerg dn = DB_DNODE(db); 40287992Sjoerg if (dn == NULL) { 40387992Sjoerg ASSERT(db->db_parent == NULL); 40487992Sjoerg ASSERT(db->db_blkptr == NULL); 40587992Sjoerg } else { 40687992Sjoerg ASSERT3U(db->db.db_object, ==, dn->dn_object); 40787992Sjoerg ASSERT3P(db->db_objset, ==, dn->dn_objset); 40887992Sjoerg ASSERT3U(db->db_level, <, dn->dn_nlevels); 40987992Sjoerg ASSERT(db->db_blkid == DMU_BONUS_BLKID || 41087992Sjoerg db->db_blkid == DMU_SPILL_BLKID || 41187992Sjoerg !avl_is_empty(&dn->dn_dbufs)); 41287992Sjoerg } 41387992Sjoerg if (db->db_blkid == DMU_BONUS_BLKID) { 41487992Sjoerg ASSERT(dn != NULL); 41587992Sjoerg ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 41687992Sjoerg ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 41787992Sjoerg } else if (db->db_blkid == DMU_SPILL_BLKID) { 41887992Sjoerg ASSERT(dn != NULL); 41987992Sjoerg ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 42087992Sjoerg ASSERT0(db->db.db_offset); 42187992Sjoerg } else { 42287992Sjoerg ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 42387992Sjoerg } 42487992Sjoerg 42587992Sjoerg for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 42687992Sjoerg ASSERT(dr->dr_dbuf == db); 42787992Sjoerg 42887992Sjoerg for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 42987992Sjoerg ASSERT(dr->dr_dbuf == db); 43087992Sjoerg 43187992Sjoerg /* 43287992Sjoerg * We can't assert that db_size matches dn_datablksz because it 43387992Sjoerg * can be momentarily different when another thread is doing 43487992Sjoerg * dnode_set_blksz(). 43587992Sjoerg */ 43687992Sjoerg if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 43787992Sjoerg dr = db->db_data_pending; 43887992Sjoerg /* 43987992Sjoerg * It should only be modified in syncing context, so 440126230Sphk * make sure we only have one copy of the data. 441126230Sphk */ 442126230Sphk ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 443134081Sphk } 444134081Sphk 445127522Snyan /* verify db->db_blkptr */ 446127522Snyan if (db->db_blkptr) { 447134081Sphk if (db->db_parent == dn->dn_dbuf) { 44887992Sjoerg /* db is pointed to by the dnode */ 44987992Sjoerg /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 45087992Sjoerg if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 45187992Sjoerg ASSERT(db->db_parent == NULL); 45287992Sjoerg else 45387992Sjoerg ASSERT(db->db_parent != NULL); 45487992Sjoerg if (db->db_blkid != DMU_SPILL_BLKID) 45587992Sjoerg ASSERT3P(db->db_blkptr, ==, 45687992Sjoerg &dn->dn_phys->dn_blkptr[db->db_blkid]); 45787992Sjoerg } else { 45887992Sjoerg /* db is pointed to by an indirect block */ 45987992Sjoerg int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 46087992Sjoerg ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 46187992Sjoerg ASSERT3U(db->db_parent->db.db_object, ==, 46287992Sjoerg db->db.db_object); 46387992Sjoerg /* 46487992Sjoerg * dnode_grow_indblksz() can make this fail if we don't 46587992Sjoerg * have the struct_rwlock. XXX indblksz no longer 46687992Sjoerg * grows. safe to do this now? 46787992Sjoerg */ 46887992Sjoerg if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 46987992Sjoerg ASSERT3P(db->db_blkptr, ==, 47087992Sjoerg ((blkptr_t *)db->db_parent->db.db_data + 47187992Sjoerg db->db_blkid % epb)); 47287992Sjoerg } 47387992Sjoerg } 47487992Sjoerg } 47587992Sjoerg if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 47687992Sjoerg (db->db_buf == NULL || db->db_buf->b_data) && 47787992Sjoerg db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 47887992Sjoerg db->db_state != DB_FILL && !dn->dn_free_txg) { 47987992Sjoerg /* 48087992Sjoerg * If the blkptr isn't set but they have nonzero data, 48187992Sjoerg * it had better be dirty, otherwise we'll lose that 48287992Sjoerg * data when we evict this buffer. 48387992Sjoerg */ 48487992Sjoerg if (db->db_dirtycnt == 0) { 48587992Sjoerg uint64_t *buf = db->db.db_data; 48687992Sjoerg int i; 48787992Sjoerg 48887992Sjoerg for (i = 0; i < db->db.db_size >> 3; i++) { 48987992Sjoerg ASSERT(buf[i] == 0); 49087992Sjoerg } 49187992Sjoerg } 49287992Sjoerg } 49387992Sjoerg DB_DNODE_EXIT(db); 49487992Sjoerg} 49587992Sjoerg#endif 49687992Sjoerg 49787992Sjoergstatic void 49887992Sjoergdbuf_clear_data(dmu_buf_impl_t *db) 49987992Sjoerg{ 50087992Sjoerg ASSERT(MUTEX_HELD(&db->db_mtx)); 50187992Sjoerg dbuf_evict_user(db); 50287992Sjoerg db->db_buf = NULL; 50387992Sjoerg db->db.db_data = NULL; 50487992Sjoerg if (db->db_state != DB_NOFILL) 50587992Sjoerg db->db_state = DB_UNCACHED; 50687992Sjoerg} 50787992Sjoerg 50887992Sjoergstatic void 50987992Sjoergdbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 510{ 511 ASSERT(MUTEX_HELD(&db->db_mtx)); 512 ASSERT(buf != NULL); 513 514 db->db_buf = buf; 515 ASSERT(buf->b_data != NULL); 516 db->db.db_data = buf->b_data; 517 if (!arc_released(buf)) 518 arc_set_callback(buf, dbuf_do_evict, db); 519} 520 521/* 522 * Loan out an arc_buf for read. Return the loaned arc_buf. 523 */ 524arc_buf_t * 525dbuf_loan_arcbuf(dmu_buf_impl_t *db) 526{ 527 arc_buf_t *abuf; 528 529 mutex_enter(&db->db_mtx); 530 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 531 int blksz = db->db.db_size; 532 spa_t *spa = db->db_objset->os_spa; 533 534 mutex_exit(&db->db_mtx); 535 abuf = arc_loan_buf(spa, blksz); 536 bcopy(db->db.db_data, abuf->b_data, blksz); 537 } else { 538 abuf = db->db_buf; 539 arc_loan_inuse_buf(abuf, db); 540 dbuf_clear_data(db); 541 mutex_exit(&db->db_mtx); 542 } 543 return (abuf); 544} 545 546/* 547 * Calculate which level n block references the data at the level 0 offset 548 * provided. 549 */ 550uint64_t 551dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 552{ 553 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 554 /* 555 * The level n blkid is equal to the level 0 blkid divided by 556 * the number of level 0s in a level n block. 557 * 558 * The level 0 blkid is offset >> datablkshift = 559 * offset / 2^datablkshift. 560 * 561 * The number of level 0s in a level n is the number of block 562 * pointers in an indirect block, raised to the power of level. 563 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 564 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 565 * 566 * Thus, the level n blkid is: offset / 567 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 568 * = offset / 2^(datablkshift + level * 569 * (indblkshift - SPA_BLKPTRSHIFT)) 570 * = offset >> (datablkshift + level * 571 * (indblkshift - SPA_BLKPTRSHIFT)) 572 */ 573 return (offset >> (dn->dn_datablkshift + level * 574 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 575 } else { 576 ASSERT3U(offset, <, dn->dn_datablksz); 577 return (0); 578 } 579} 580 581static void 582dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 583{ 584 dmu_buf_impl_t *db = vdb; 585 586 mutex_enter(&db->db_mtx); 587 ASSERT3U(db->db_state, ==, DB_READ); 588 /* 589 * All reads are synchronous, so we must have a hold on the dbuf 590 */ 591 ASSERT(refcount_count(&db->db_holds) > 0); 592 ASSERT(db->db_buf == NULL); 593 ASSERT(db->db.db_data == NULL); 594 if (db->db_level == 0 && db->db_freed_in_flight) { 595 /* we were freed in flight; disregard any error */ 596 arc_release(buf, db); 597 bzero(buf->b_data, db->db.db_size); 598 arc_buf_freeze(buf); 599 db->db_freed_in_flight = FALSE; 600 dbuf_set_data(db, buf); 601 db->db_state = DB_CACHED; 602 } else if (zio == NULL || zio->io_error == 0) { 603 dbuf_set_data(db, buf); 604 db->db_state = DB_CACHED; 605 } else { 606 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 607 ASSERT3P(db->db_buf, ==, NULL); 608 VERIFY(arc_buf_remove_ref(buf, db)); 609 db->db_state = DB_UNCACHED; 610 } 611 cv_broadcast(&db->db_changed); 612 dbuf_rele_and_unlock(db, NULL); 613} 614 615static void 616dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 617{ 618 dnode_t *dn; 619 zbookmark_phys_t zb; 620 arc_flags_t aflags = ARC_FLAG_NOWAIT; 621 622 DB_DNODE_ENTER(db); 623 dn = DB_DNODE(db); 624 ASSERT(!refcount_is_zero(&db->db_holds)); 625 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 626 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 627 ASSERT(MUTEX_HELD(&db->db_mtx)); 628 ASSERT(db->db_state == DB_UNCACHED); 629 ASSERT(db->db_buf == NULL); 630 631 if (db->db_blkid == DMU_BONUS_BLKID) { 632 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 633 634 ASSERT3U(bonuslen, <=, db->db.db_size); 635 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 636 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 637 if (bonuslen < DN_MAX_BONUSLEN) 638 bzero(db->db.db_data, DN_MAX_BONUSLEN); 639 if (bonuslen) 640 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 641 DB_DNODE_EXIT(db); 642 db->db_state = DB_CACHED; 643 mutex_exit(&db->db_mtx); 644 return; 645 } 646 647 /* 648 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 649 * processes the delete record and clears the bp while we are waiting 650 * for the dn_mtx (resulting in a "no" from block_freed). 651 */ 652 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 653 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 654 BP_IS_HOLE(db->db_blkptr)))) { 655 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 656 657 DB_DNODE_EXIT(db); 658 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 659 db->db.db_size, db, type)); 660 bzero(db->db.db_data, db->db.db_size); 661 db->db_state = DB_CACHED; 662 mutex_exit(&db->db_mtx); 663 return; 664 } 665 666 DB_DNODE_EXIT(db); 667 668 db->db_state = DB_READ; 669 mutex_exit(&db->db_mtx); 670 671 if (DBUF_IS_L2CACHEABLE(db)) 672 aflags |= ARC_FLAG_L2CACHE; 673 if (DBUF_IS_L2COMPRESSIBLE(db)) 674 aflags |= ARC_FLAG_L2COMPRESS; 675 676 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 677 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 678 db->db.db_object, db->db_level, db->db_blkid); 679 680 dbuf_add_ref(db, NULL); 681 682 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 683 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 684 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 685 &aflags, &zb); 686} 687 688int 689dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 690{ 691 int err = 0; 692 boolean_t havepzio = (zio != NULL); 693 boolean_t prefetch; 694 dnode_t *dn; 695 696 /* 697 * We don't have to hold the mutex to check db_state because it 698 * can't be freed while we have a hold on the buffer. 699 */ 700 ASSERT(!refcount_is_zero(&db->db_holds)); 701 702 if (db->db_state == DB_NOFILL) 703 return (SET_ERROR(EIO)); 704 705 DB_DNODE_ENTER(db); 706 dn = DB_DNODE(db); 707 if ((flags & DB_RF_HAVESTRUCT) == 0) 708 rw_enter(&dn->dn_struct_rwlock, RW_READER); 709 710 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 711 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 712 DBUF_IS_CACHEABLE(db); 713 714 mutex_enter(&db->db_mtx); 715 if (db->db_state == DB_CACHED) { 716 mutex_exit(&db->db_mtx); 717 if (prefetch) 718 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 719 if ((flags & DB_RF_HAVESTRUCT) == 0) 720 rw_exit(&dn->dn_struct_rwlock); 721 DB_DNODE_EXIT(db); 722 } else if (db->db_state == DB_UNCACHED) { 723 spa_t *spa = dn->dn_objset->os_spa; 724 725 if (zio == NULL) 726 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 727 dbuf_read_impl(db, zio, flags); 728 729 /* dbuf_read_impl has dropped db_mtx for us */ 730 731 if (prefetch) 732 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 733 734 if ((flags & DB_RF_HAVESTRUCT) == 0) 735 rw_exit(&dn->dn_struct_rwlock); 736 DB_DNODE_EXIT(db); 737 738 if (!havepzio) 739 err = zio_wait(zio); 740 } else { 741 /* 742 * Another reader came in while the dbuf was in flight 743 * between UNCACHED and CACHED. Either a writer will finish 744 * writing the buffer (sending the dbuf to CACHED) or the 745 * first reader's request will reach the read_done callback 746 * and send the dbuf to CACHED. Otherwise, a failure 747 * occurred and the dbuf went to UNCACHED. 748 */ 749 mutex_exit(&db->db_mtx); 750 if (prefetch) 751 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 752 if ((flags & DB_RF_HAVESTRUCT) == 0) 753 rw_exit(&dn->dn_struct_rwlock); 754 DB_DNODE_EXIT(db); 755 756 /* Skip the wait per the caller's request. */ 757 mutex_enter(&db->db_mtx); 758 if ((flags & DB_RF_NEVERWAIT) == 0) { 759 while (db->db_state == DB_READ || 760 db->db_state == DB_FILL) { 761 ASSERT(db->db_state == DB_READ || 762 (flags & DB_RF_HAVESTRUCT) == 0); 763 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 764 db, zio_t *, zio); 765 cv_wait(&db->db_changed, &db->db_mtx); 766 } 767 if (db->db_state == DB_UNCACHED) 768 err = SET_ERROR(EIO); 769 } 770 mutex_exit(&db->db_mtx); 771 } 772 773 ASSERT(err || havepzio || db->db_state == DB_CACHED); 774 return (err); 775} 776 777static void 778dbuf_noread(dmu_buf_impl_t *db) 779{ 780 ASSERT(!refcount_is_zero(&db->db_holds)); 781 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 782 mutex_enter(&db->db_mtx); 783 while (db->db_state == DB_READ || db->db_state == DB_FILL) 784 cv_wait(&db->db_changed, &db->db_mtx); 785 if (db->db_state == DB_UNCACHED) { 786 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 787 spa_t *spa = db->db_objset->os_spa; 788 789 ASSERT(db->db_buf == NULL); 790 ASSERT(db->db.db_data == NULL); 791 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 792 db->db_state = DB_FILL; 793 } else if (db->db_state == DB_NOFILL) { 794 dbuf_clear_data(db); 795 } else { 796 ASSERT3U(db->db_state, ==, DB_CACHED); 797 } 798 mutex_exit(&db->db_mtx); 799} 800 801/* 802 * This is our just-in-time copy function. It makes a copy of 803 * buffers, that have been modified in a previous transaction 804 * group, before we modify them in the current active group. 805 * 806 * This function is used in two places: when we are dirtying a 807 * buffer for the first time in a txg, and when we are freeing 808 * a range in a dnode that includes this buffer. 809 * 810 * Note that when we are called from dbuf_free_range() we do 811 * not put a hold on the buffer, we just traverse the active 812 * dbuf list for the dnode. 813 */ 814static void 815dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 816{ 817 dbuf_dirty_record_t *dr = db->db_last_dirty; 818 819 ASSERT(MUTEX_HELD(&db->db_mtx)); 820 ASSERT(db->db.db_data != NULL); 821 ASSERT(db->db_level == 0); 822 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 823 824 if (dr == NULL || 825 (dr->dt.dl.dr_data != 826 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 827 return; 828 829 /* 830 * If the last dirty record for this dbuf has not yet synced 831 * and its referencing the dbuf data, either: 832 * reset the reference to point to a new copy, 833 * or (if there a no active holders) 834 * just null out the current db_data pointer. 835 */ 836 ASSERT(dr->dr_txg >= txg - 2); 837 if (db->db_blkid == DMU_BONUS_BLKID) { 838 /* Note that the data bufs here are zio_bufs */ 839 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 840 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 841 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 842 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 843 int size = db->db.db_size; 844 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 845 spa_t *spa = db->db_objset->os_spa; 846 847 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 848 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 849 } else { 850 dbuf_clear_data(db); 851 } 852} 853 854void 855dbuf_unoverride(dbuf_dirty_record_t *dr) 856{ 857 dmu_buf_impl_t *db = dr->dr_dbuf; 858 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 859 uint64_t txg = dr->dr_txg; 860 861 ASSERT(MUTEX_HELD(&db->db_mtx)); 862 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 863 ASSERT(db->db_level == 0); 864 865 if (db->db_blkid == DMU_BONUS_BLKID || 866 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 867 return; 868 869 ASSERT(db->db_data_pending != dr); 870 871 /* free this block */ 872 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 873 zio_free(db->db_objset->os_spa, txg, bp); 874 875 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 876 dr->dt.dl.dr_nopwrite = B_FALSE; 877 878 /* 879 * Release the already-written buffer, so we leave it in 880 * a consistent dirty state. Note that all callers are 881 * modifying the buffer, so they will immediately do 882 * another (redundant) arc_release(). Therefore, leave 883 * the buf thawed to save the effort of freezing & 884 * immediately re-thawing it. 885 */ 886 arc_release(dr->dt.dl.dr_data, db); 887} 888 889/* 890 * Evict (if its unreferenced) or clear (if its referenced) any level-0 891 * data blocks in the free range, so that any future readers will find 892 * empty blocks. 893 * 894 * This is a no-op if the dataset is in the middle of an incremental 895 * receive; see comment below for details. 896 */ 897void 898dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 899 dmu_tx_t *tx) 900{ 901 dmu_buf_impl_t db_search; 902 dmu_buf_impl_t *db, *db_next; 903 uint64_t txg = tx->tx_txg; 904 avl_index_t where; 905 906 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 907 end_blkid = dn->dn_maxblkid; 908 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 909 910 db_search.db_level = 0; 911 db_search.db_blkid = start_blkid; 912 db_search.db_state = DB_SEARCH; 913 914 mutex_enter(&dn->dn_dbufs_mtx); 915 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 916 /* There can't be any dbufs in this range; no need to search. */ 917#ifdef DEBUG 918 db = avl_find(&dn->dn_dbufs, &db_search, &where); 919 ASSERT3P(db, ==, NULL); 920 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 921 ASSERT(db == NULL || db->db_level > 0); 922#endif 923 mutex_exit(&dn->dn_dbufs_mtx); 924 return; 925 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 926 /* 927 * If we are receiving, we expect there to be no dbufs in 928 * the range to be freed, because receive modifies each 929 * block at most once, and in offset order. If this is 930 * not the case, it can lead to performance problems, 931 * so note that we unexpectedly took the slow path. 932 */ 933 atomic_inc_64(&zfs_free_range_recv_miss); 934 } 935 936 db = avl_find(&dn->dn_dbufs, &db_search, &where); 937 ASSERT3P(db, ==, NULL); 938 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 939 940 for (; db != NULL; db = db_next) { 941 db_next = AVL_NEXT(&dn->dn_dbufs, db); 942 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 943 944 if (db->db_level != 0 || db->db_blkid > end_blkid) { 945 break; 946 } 947 ASSERT3U(db->db_blkid, >=, start_blkid); 948 949 /* found a level 0 buffer in the range */ 950 mutex_enter(&db->db_mtx); 951 if (dbuf_undirty(db, tx)) { 952 /* mutex has been dropped and dbuf destroyed */ 953 continue; 954 } 955 956 if (db->db_state == DB_UNCACHED || 957 db->db_state == DB_NOFILL || 958 db->db_state == DB_EVICTING) { 959 ASSERT(db->db.db_data == NULL); 960 mutex_exit(&db->db_mtx); 961 continue; 962 } 963 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 964 /* will be handled in dbuf_read_done or dbuf_rele */ 965 db->db_freed_in_flight = TRUE; 966 mutex_exit(&db->db_mtx); 967 continue; 968 } 969 if (refcount_count(&db->db_holds) == 0) { 970 ASSERT(db->db_buf); 971 dbuf_clear(db); 972 continue; 973 } 974 /* The dbuf is referenced */ 975 976 if (db->db_last_dirty != NULL) { 977 dbuf_dirty_record_t *dr = db->db_last_dirty; 978 979 if (dr->dr_txg == txg) { 980 /* 981 * This buffer is "in-use", re-adjust the file 982 * size to reflect that this buffer may 983 * contain new data when we sync. 984 */ 985 if (db->db_blkid != DMU_SPILL_BLKID && 986 db->db_blkid > dn->dn_maxblkid) 987 dn->dn_maxblkid = db->db_blkid; 988 dbuf_unoverride(dr); 989 } else { 990 /* 991 * This dbuf is not dirty in the open context. 992 * Either uncache it (if its not referenced in 993 * the open context) or reset its contents to 994 * empty. 995 */ 996 dbuf_fix_old_data(db, txg); 997 } 998 } 999 /* clear the contents if its cached */ 1000 if (db->db_state == DB_CACHED) { 1001 ASSERT(db->db.db_data != NULL); 1002 arc_release(db->db_buf, db); 1003 bzero(db->db.db_data, db->db.db_size); 1004 arc_buf_freeze(db->db_buf); 1005 } 1006 1007 mutex_exit(&db->db_mtx); 1008 } 1009 mutex_exit(&dn->dn_dbufs_mtx); 1010} 1011 1012static int 1013dbuf_block_freeable(dmu_buf_impl_t *db) 1014{ 1015 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 1016 uint64_t birth_txg = 0; 1017 1018 /* 1019 * We don't need any locking to protect db_blkptr: 1020 * If it's syncing, then db_last_dirty will be set 1021 * so we'll ignore db_blkptr. 1022 * 1023 * This logic ensures that only block births for 1024 * filled blocks are considered. 1025 */ 1026 ASSERT(MUTEX_HELD(&db->db_mtx)); 1027 if (db->db_last_dirty && (db->db_blkptr == NULL || 1028 !BP_IS_HOLE(db->db_blkptr))) { 1029 birth_txg = db->db_last_dirty->dr_txg; 1030 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1031 birth_txg = db->db_blkptr->blk_birth; 1032 } 1033 1034 /* 1035 * If this block don't exist or is in a snapshot, it can't be freed. 1036 * Don't pass the bp to dsl_dataset_block_freeable() since we 1037 * are holding the db_mtx lock and might deadlock if we are 1038 * prefetching a dedup-ed block. 1039 */ 1040 if (birth_txg != 0) 1041 return (ds == NULL || 1042 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 1043 else 1044 return (B_FALSE); 1045} 1046 1047void 1048dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1049{ 1050 arc_buf_t *buf, *obuf; 1051 int osize = db->db.db_size; 1052 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1053 dnode_t *dn; 1054 1055 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1056 1057 DB_DNODE_ENTER(db); 1058 dn = DB_DNODE(db); 1059 1060 /* XXX does *this* func really need the lock? */ 1061 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1062 1063 /* 1064 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1065 * is OK, because there can be no other references to the db 1066 * when we are changing its size, so no concurrent DB_FILL can 1067 * be happening. 1068 */ 1069 /* 1070 * XXX we should be doing a dbuf_read, checking the return 1071 * value and returning that up to our callers 1072 */ 1073 dmu_buf_will_dirty(&db->db, tx); 1074 1075 /* create the data buffer for the new block */ 1076 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 1077 1078 /* copy old block data to the new block */ 1079 obuf = db->db_buf; 1080 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1081 /* zero the remainder */ 1082 if (size > osize) 1083 bzero((uint8_t *)buf->b_data + osize, size - osize); 1084 1085 mutex_enter(&db->db_mtx); 1086 dbuf_set_data(db, buf); 1087 VERIFY(arc_buf_remove_ref(obuf, db)); 1088 db->db.db_size = size; 1089 1090 if (db->db_level == 0) { 1091 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1092 db->db_last_dirty->dt.dl.dr_data = buf; 1093 } 1094 mutex_exit(&db->db_mtx); 1095 1096 dnode_willuse_space(dn, size-osize, tx); 1097 DB_DNODE_EXIT(db); 1098} 1099 1100void 1101dbuf_release_bp(dmu_buf_impl_t *db) 1102{ 1103 objset_t *os = db->db_objset; 1104 1105 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1106 ASSERT(arc_released(os->os_phys_buf) || 1107 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1108 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1109 1110 (void) arc_release(db->db_buf, db); 1111} 1112 1113/* 1114 * We already have a dirty record for this TXG, and we are being 1115 * dirtied again. 1116 */ 1117static void 1118dbuf_redirty(dbuf_dirty_record_t *dr) 1119{ 1120 dmu_buf_impl_t *db = dr->dr_dbuf; 1121 1122 ASSERT(MUTEX_HELD(&db->db_mtx)); 1123 1124 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1125 /* 1126 * If this buffer has already been written out, 1127 * we now need to reset its state. 1128 */ 1129 dbuf_unoverride(dr); 1130 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1131 db->db_state != DB_NOFILL) { 1132 /* Already released on initial dirty, so just thaw. */ 1133 ASSERT(arc_released(db->db_buf)); 1134 arc_buf_thaw(db->db_buf); 1135 } 1136 } 1137} 1138 1139dbuf_dirty_record_t * 1140dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1141{ 1142 dnode_t *dn; 1143 objset_t *os; 1144 dbuf_dirty_record_t **drp, *dr; 1145 int drop_struct_lock = FALSE; 1146 boolean_t do_free_accounting = B_FALSE; 1147 int txgoff = tx->tx_txg & TXG_MASK; 1148 1149 ASSERT(tx->tx_txg != 0); 1150 ASSERT(!refcount_is_zero(&db->db_holds)); 1151 DMU_TX_DIRTY_BUF(tx, db); 1152 1153 DB_DNODE_ENTER(db); 1154 dn = DB_DNODE(db); 1155 /* 1156 * Shouldn't dirty a regular buffer in syncing context. Private 1157 * objects may be dirtied in syncing context, but only if they 1158 * were already pre-dirtied in open context. 1159 */ 1160 ASSERT(!dmu_tx_is_syncing(tx) || 1161 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1162 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1163 dn->dn_objset->os_dsl_dataset == NULL); 1164 /* 1165 * We make this assert for private objects as well, but after we 1166 * check if we're already dirty. They are allowed to re-dirty 1167 * in syncing context. 1168 */ 1169 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1170 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1171 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1172 1173 mutex_enter(&db->db_mtx); 1174 /* 1175 * XXX make this true for indirects too? The problem is that 1176 * transactions created with dmu_tx_create_assigned() from 1177 * syncing context don't bother holding ahead. 1178 */ 1179 ASSERT(db->db_level != 0 || 1180 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1181 db->db_state == DB_NOFILL); 1182 1183 mutex_enter(&dn->dn_mtx); 1184 /* 1185 * Don't set dirtyctx to SYNC if we're just modifying this as we 1186 * initialize the objset. 1187 */ 1188 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1189 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1190 dn->dn_dirtyctx = 1191 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1192 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1193 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1194 } 1195 mutex_exit(&dn->dn_mtx); 1196 1197 if (db->db_blkid == DMU_SPILL_BLKID) 1198 dn->dn_have_spill = B_TRUE; 1199 1200 /* 1201 * If this buffer is already dirty, we're done. 1202 */ 1203 drp = &db->db_last_dirty; 1204 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1205 db->db.db_object == DMU_META_DNODE_OBJECT); 1206 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1207 drp = &dr->dr_next; 1208 if (dr && dr->dr_txg == tx->tx_txg) { 1209 DB_DNODE_EXIT(db); 1210 1211 dbuf_redirty(dr); 1212 mutex_exit(&db->db_mtx); 1213 return (dr); 1214 } 1215 1216 /* 1217 * Only valid if not already dirty. 1218 */ 1219 ASSERT(dn->dn_object == 0 || 1220 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1221 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1222 1223 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1224 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1225 dn->dn_phys->dn_nlevels > db->db_level || 1226 dn->dn_next_nlevels[txgoff] > db->db_level || 1227 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1228 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1229 1230 /* 1231 * We should only be dirtying in syncing context if it's the 1232 * mos or we're initializing the os or it's a special object. 1233 * However, we are allowed to dirty in syncing context provided 1234 * we already dirtied it in open context. Hence we must make 1235 * this assertion only if we're not already dirty. 1236 */ 1237 os = dn->dn_objset; 1238 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1239 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1240 ASSERT(db->db.db_size != 0); 1241 1242 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1243 1244 if (db->db_blkid != DMU_BONUS_BLKID) { 1245 /* 1246 * Update the accounting. 1247 * Note: we delay "free accounting" until after we drop 1248 * the db_mtx. This keeps us from grabbing other locks 1249 * (and possibly deadlocking) in bp_get_dsize() while 1250 * also holding the db_mtx. 1251 */ 1252 dnode_willuse_space(dn, db->db.db_size, tx); 1253 do_free_accounting = dbuf_block_freeable(db); 1254 } 1255 1256 /* 1257 * If this buffer is dirty in an old transaction group we need 1258 * to make a copy of it so that the changes we make in this 1259 * transaction group won't leak out when we sync the older txg. 1260 */ 1261 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1262 if (db->db_level == 0) { 1263 void *data_old = db->db_buf; 1264 1265 if (db->db_state != DB_NOFILL) { 1266 if (db->db_blkid == DMU_BONUS_BLKID) { 1267 dbuf_fix_old_data(db, tx->tx_txg); 1268 data_old = db->db.db_data; 1269 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1270 /* 1271 * Release the data buffer from the cache so 1272 * that we can modify it without impacting 1273 * possible other users of this cached data 1274 * block. Note that indirect blocks and 1275 * private objects are not released until the 1276 * syncing state (since they are only modified 1277 * then). 1278 */ 1279 arc_release(db->db_buf, db); 1280 dbuf_fix_old_data(db, tx->tx_txg); 1281 data_old = db->db_buf; 1282 } 1283 ASSERT(data_old != NULL); 1284 } 1285 dr->dt.dl.dr_data = data_old; 1286 } else { 1287 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1288 list_create(&dr->dt.di.dr_children, 1289 sizeof (dbuf_dirty_record_t), 1290 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1291 } 1292 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1293 dr->dr_accounted = db->db.db_size; 1294 dr->dr_dbuf = db; 1295 dr->dr_txg = tx->tx_txg; 1296 dr->dr_next = *drp; 1297 *drp = dr; 1298 1299 /* 1300 * We could have been freed_in_flight between the dbuf_noread 1301 * and dbuf_dirty. We win, as though the dbuf_noread() had 1302 * happened after the free. 1303 */ 1304 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1305 db->db_blkid != DMU_SPILL_BLKID) { 1306 mutex_enter(&dn->dn_mtx); 1307 if (dn->dn_free_ranges[txgoff] != NULL) { 1308 range_tree_clear(dn->dn_free_ranges[txgoff], 1309 db->db_blkid, 1); 1310 } 1311 mutex_exit(&dn->dn_mtx); 1312 db->db_freed_in_flight = FALSE; 1313 } 1314 1315 /* 1316 * This buffer is now part of this txg 1317 */ 1318 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1319 db->db_dirtycnt += 1; 1320 ASSERT3U(db->db_dirtycnt, <=, 3); 1321 1322 mutex_exit(&db->db_mtx); 1323 1324 if (db->db_blkid == DMU_BONUS_BLKID || 1325 db->db_blkid == DMU_SPILL_BLKID) { 1326 mutex_enter(&dn->dn_mtx); 1327 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1328 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1329 mutex_exit(&dn->dn_mtx); 1330 dnode_setdirty(dn, tx); 1331 DB_DNODE_EXIT(db); 1332 return (dr); 1333 } else if (do_free_accounting) { 1334 blkptr_t *bp = db->db_blkptr; 1335 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1336 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1337 /* 1338 * This is only a guess -- if the dbuf is dirty 1339 * in a previous txg, we don't know how much 1340 * space it will use on disk yet. We should 1341 * really have the struct_rwlock to access 1342 * db_blkptr, but since this is just a guess, 1343 * it's OK if we get an odd answer. 1344 */ 1345 ddt_prefetch(os->os_spa, bp); 1346 dnode_willuse_space(dn, -willfree, tx); 1347 } 1348 1349 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1350 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1351 drop_struct_lock = TRUE; 1352 } 1353 1354 if (db->db_level == 0) { 1355 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1356 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1357 } 1358 1359 if (db->db_level+1 < dn->dn_nlevels) { 1360 dmu_buf_impl_t *parent = db->db_parent; 1361 dbuf_dirty_record_t *di; 1362 int parent_held = FALSE; 1363 1364 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1365 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1366 1367 parent = dbuf_hold_level(dn, db->db_level+1, 1368 db->db_blkid >> epbs, FTAG); 1369 ASSERT(parent != NULL); 1370 parent_held = TRUE; 1371 } 1372 if (drop_struct_lock) 1373 rw_exit(&dn->dn_struct_rwlock); 1374 ASSERT3U(db->db_level+1, ==, parent->db_level); 1375 di = dbuf_dirty(parent, tx); 1376 if (parent_held) 1377 dbuf_rele(parent, FTAG); 1378 1379 mutex_enter(&db->db_mtx); 1380 /* 1381 * Since we've dropped the mutex, it's possible that 1382 * dbuf_undirty() might have changed this out from under us. 1383 */ 1384 if (db->db_last_dirty == dr || 1385 dn->dn_object == DMU_META_DNODE_OBJECT) { 1386 mutex_enter(&di->dt.di.dr_mtx); 1387 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1388 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1389 list_insert_tail(&di->dt.di.dr_children, dr); 1390 mutex_exit(&di->dt.di.dr_mtx); 1391 dr->dr_parent = di; 1392 } 1393 mutex_exit(&db->db_mtx); 1394 } else { 1395 ASSERT(db->db_level+1 == dn->dn_nlevels); 1396 ASSERT(db->db_blkid < dn->dn_nblkptr); 1397 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1398 mutex_enter(&dn->dn_mtx); 1399 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1400 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1401 mutex_exit(&dn->dn_mtx); 1402 if (drop_struct_lock) 1403 rw_exit(&dn->dn_struct_rwlock); 1404 } 1405 1406 dnode_setdirty(dn, tx); 1407 DB_DNODE_EXIT(db); 1408 return (dr); 1409} 1410 1411/* 1412 * Undirty a buffer in the transaction group referenced by the given 1413 * transaction. Return whether this evicted the dbuf. 1414 */ 1415static boolean_t 1416dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1417{ 1418 dnode_t *dn; 1419 uint64_t txg = tx->tx_txg; 1420 dbuf_dirty_record_t *dr, **drp; 1421 1422 ASSERT(txg != 0); 1423 1424 /* 1425 * Due to our use of dn_nlevels below, this can only be called 1426 * in open context, unless we are operating on the MOS. 1427 * From syncing context, dn_nlevels may be different from the 1428 * dn_nlevels used when dbuf was dirtied. 1429 */ 1430 ASSERT(db->db_objset == 1431 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1432 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1433 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1434 ASSERT0(db->db_level); 1435 ASSERT(MUTEX_HELD(&db->db_mtx)); 1436 1437 /* 1438 * If this buffer is not dirty, we're done. 1439 */ 1440 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1441 if (dr->dr_txg <= txg) 1442 break; 1443 if (dr == NULL || dr->dr_txg < txg) 1444 return (B_FALSE); 1445 ASSERT(dr->dr_txg == txg); 1446 ASSERT(dr->dr_dbuf == db); 1447 1448 DB_DNODE_ENTER(db); 1449 dn = DB_DNODE(db); 1450 1451 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1452 1453 ASSERT(db->db.db_size != 0); 1454 1455 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1456 dr->dr_accounted, txg); 1457 1458 *drp = dr->dr_next; 1459 1460 /* 1461 * Note that there are three places in dbuf_dirty() 1462 * where this dirty record may be put on a list. 1463 * Make sure to do a list_remove corresponding to 1464 * every one of those list_insert calls. 1465 */ 1466 if (dr->dr_parent) { 1467 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1468 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1469 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1470 } else if (db->db_blkid == DMU_SPILL_BLKID || 1471 db->db_level + 1 == dn->dn_nlevels) { 1472 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1473 mutex_enter(&dn->dn_mtx); 1474 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1475 mutex_exit(&dn->dn_mtx); 1476 } 1477 DB_DNODE_EXIT(db); 1478 1479 if (db->db_state != DB_NOFILL) { 1480 dbuf_unoverride(dr); 1481 1482 ASSERT(db->db_buf != NULL); 1483 ASSERT(dr->dt.dl.dr_data != NULL); 1484 if (dr->dt.dl.dr_data != db->db_buf) 1485 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1486 } 1487 1488 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1489 1490 ASSERT(db->db_dirtycnt > 0); 1491 db->db_dirtycnt -= 1; 1492 1493 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1494 arc_buf_t *buf = db->db_buf; 1495 1496 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1497 dbuf_clear_data(db); 1498 VERIFY(arc_buf_remove_ref(buf, db)); 1499 dbuf_evict(db); 1500 return (B_TRUE); 1501 } 1502 1503 return (B_FALSE); 1504} 1505 1506void 1507dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1508{ 1509 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1510 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1511 1512 ASSERT(tx->tx_txg != 0); 1513 ASSERT(!refcount_is_zero(&db->db_holds)); 1514 1515 /* 1516 * Quick check for dirtyness. For already dirty blocks, this 1517 * reduces runtime of this function by >90%, and overall performance 1518 * by 50% for some workloads (e.g. file deletion with indirect blocks 1519 * cached). 1520 */ 1521 mutex_enter(&db->db_mtx); 1522 dbuf_dirty_record_t *dr; 1523 for (dr = db->db_last_dirty; 1524 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1525 /* 1526 * It's possible that it is already dirty but not cached, 1527 * because there are some calls to dbuf_dirty() that don't 1528 * go through dmu_buf_will_dirty(). 1529 */ 1530 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1531 /* This dbuf is already dirty and cached. */ 1532 dbuf_redirty(dr); 1533 mutex_exit(&db->db_mtx); 1534 return; 1535 } 1536 } 1537 mutex_exit(&db->db_mtx); 1538 1539 DB_DNODE_ENTER(db); 1540 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1541 rf |= DB_RF_HAVESTRUCT; 1542 DB_DNODE_EXIT(db); 1543 (void) dbuf_read(db, NULL, rf); 1544 (void) dbuf_dirty(db, tx); 1545} 1546 1547void 1548dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1549{ 1550 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1551 1552 db->db_state = DB_NOFILL; 1553 1554 dmu_buf_will_fill(db_fake, tx); 1555} 1556 1557void 1558dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1559{ 1560 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1561 1562 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1563 ASSERT(tx->tx_txg != 0); 1564 ASSERT(db->db_level == 0); 1565 ASSERT(!refcount_is_zero(&db->db_holds)); 1566 1567 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1568 dmu_tx_private_ok(tx)); 1569 1570 dbuf_noread(db); 1571 (void) dbuf_dirty(db, tx); 1572} 1573 1574#pragma weak dmu_buf_fill_done = dbuf_fill_done 1575/* ARGSUSED */ 1576void 1577dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1578{ 1579 mutex_enter(&db->db_mtx); 1580 DBUF_VERIFY(db); 1581 1582 if (db->db_state == DB_FILL) { 1583 if (db->db_level == 0 && db->db_freed_in_flight) { 1584 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1585 /* we were freed while filling */ 1586 /* XXX dbuf_undirty? */ 1587 bzero(db->db.db_data, db->db.db_size); 1588 db->db_freed_in_flight = FALSE; 1589 } 1590 db->db_state = DB_CACHED; 1591 cv_broadcast(&db->db_changed); 1592 } 1593 mutex_exit(&db->db_mtx); 1594} 1595 1596void 1597dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1598 bp_embedded_type_t etype, enum zio_compress comp, 1599 int uncompressed_size, int compressed_size, int byteorder, 1600 dmu_tx_t *tx) 1601{ 1602 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1603 struct dirty_leaf *dl; 1604 dmu_object_type_t type; 1605 1606 if (etype == BP_EMBEDDED_TYPE_DATA) { 1607 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1608 SPA_FEATURE_EMBEDDED_DATA)); 1609 } 1610 1611 DB_DNODE_ENTER(db); 1612 type = DB_DNODE(db)->dn_type; 1613 DB_DNODE_EXIT(db); 1614 1615 ASSERT0(db->db_level); 1616 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1617 1618 dmu_buf_will_not_fill(dbuf, tx); 1619 1620 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1621 dl = &db->db_last_dirty->dt.dl; 1622 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1623 data, comp, uncompressed_size, compressed_size); 1624 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1625 BP_SET_TYPE(&dl->dr_overridden_by, type); 1626 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1627 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1628 1629 dl->dr_override_state = DR_OVERRIDDEN; 1630 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1631} 1632 1633/* 1634 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1635 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1636 */ 1637void 1638dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1639{ 1640 ASSERT(!refcount_is_zero(&db->db_holds)); 1641 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1642 ASSERT(db->db_level == 0); 1643 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1644 ASSERT(buf != NULL); 1645 ASSERT(arc_buf_size(buf) == db->db.db_size); 1646 ASSERT(tx->tx_txg != 0); 1647 1648 arc_return_buf(buf, db); 1649 ASSERT(arc_released(buf)); 1650 1651 mutex_enter(&db->db_mtx); 1652 1653 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1654 cv_wait(&db->db_changed, &db->db_mtx); 1655 1656 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1657 1658 if (db->db_state == DB_CACHED && 1659 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1660 mutex_exit(&db->db_mtx); 1661 (void) dbuf_dirty(db, tx); 1662 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1663 VERIFY(arc_buf_remove_ref(buf, db)); 1664 xuio_stat_wbuf_copied(); 1665 return; 1666 } 1667 1668 xuio_stat_wbuf_nocopy(); 1669 if (db->db_state == DB_CACHED) { 1670 dbuf_dirty_record_t *dr = db->db_last_dirty; 1671 1672 ASSERT(db->db_buf != NULL); 1673 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1674 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1675 if (!arc_released(db->db_buf)) { 1676 ASSERT(dr->dt.dl.dr_override_state == 1677 DR_OVERRIDDEN); 1678 arc_release(db->db_buf, db); 1679 } 1680 dr->dt.dl.dr_data = buf; 1681 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1682 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1683 arc_release(db->db_buf, db); 1684 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1685 } 1686 db->db_buf = NULL; 1687 } 1688 ASSERT(db->db_buf == NULL); 1689 dbuf_set_data(db, buf); 1690 db->db_state = DB_FILL; 1691 mutex_exit(&db->db_mtx); 1692 (void) dbuf_dirty(db, tx); 1693 dmu_buf_fill_done(&db->db, tx); 1694} 1695 1696/* 1697 * "Clear" the contents of this dbuf. This will mark the dbuf 1698 * EVICTING and clear *most* of its references. Unfortunately, 1699 * when we are not holding the dn_dbufs_mtx, we can't clear the 1700 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1701 * in this case. For callers from the DMU we will usually see: 1702 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() 1703 * For the arc callback, we will usually see: 1704 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1705 * Sometimes, though, we will get a mix of these two: 1706 * DMU: dbuf_clear()->arc_clear_callback() 1707 * ARC: dbuf_do_evict()->dbuf_destroy() 1708 * 1709 * This routine will dissociate the dbuf from the arc, by calling 1710 * arc_clear_callback(), but will not evict the data from the ARC. 1711 */ 1712void 1713dbuf_clear(dmu_buf_impl_t *db) 1714{ 1715 dnode_t *dn; 1716 dmu_buf_impl_t *parent = db->db_parent; 1717 dmu_buf_impl_t *dndb; 1718 boolean_t dbuf_gone = B_FALSE; 1719 1720 ASSERT(MUTEX_HELD(&db->db_mtx)); 1721 ASSERT(refcount_is_zero(&db->db_holds)); 1722 1723 dbuf_evict_user(db); 1724 1725 if (db->db_state == DB_CACHED) { 1726 ASSERT(db->db.db_data != NULL); 1727 if (db->db_blkid == DMU_BONUS_BLKID) { 1728 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1729 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1730 } 1731 db->db.db_data = NULL; 1732 db->db_state = DB_UNCACHED; 1733 } 1734 1735 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1736 ASSERT(db->db_data_pending == NULL); 1737 1738 db->db_state = DB_EVICTING; 1739 db->db_blkptr = NULL; 1740 1741 DB_DNODE_ENTER(db); 1742 dn = DB_DNODE(db); 1743 dndb = dn->dn_dbuf; 1744 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1745 avl_remove(&dn->dn_dbufs, db); 1746 atomic_dec_32(&dn->dn_dbufs_count); 1747 membar_producer(); 1748 DB_DNODE_EXIT(db); 1749 /* 1750 * Decrementing the dbuf count means that the hold corresponding 1751 * to the removed dbuf is no longer discounted in dnode_move(), 1752 * so the dnode cannot be moved until after we release the hold. 1753 * The membar_producer() ensures visibility of the decremented 1754 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1755 * release any lock. 1756 */ 1757 dnode_rele(dn, db); 1758 db->db_dnode_handle = NULL; 1759 } else { 1760 DB_DNODE_EXIT(db); 1761 } 1762 1763 if (db->db_buf) 1764 dbuf_gone = arc_clear_callback(db->db_buf); 1765 1766 if (!dbuf_gone) 1767 mutex_exit(&db->db_mtx); 1768 1769 /* 1770 * If this dbuf is referenced from an indirect dbuf, 1771 * decrement the ref count on the indirect dbuf. 1772 */ 1773 if (parent && parent != dndb) 1774 dbuf_rele(parent, db); 1775} 1776 1777/* 1778 * Note: While bpp will always be updated if the function returns success, 1779 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 1780 * this happens when the dnode is the meta-dnode, or a userused or groupused 1781 * object. 1782 */ 1783static int 1784dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1785 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1786{ 1787 int nlevels, epbs; 1788 1789 *parentp = NULL; 1790 *bpp = NULL; 1791 1792 ASSERT(blkid != DMU_BONUS_BLKID); 1793 1794 if (blkid == DMU_SPILL_BLKID) { 1795 mutex_enter(&dn->dn_mtx); 1796 if (dn->dn_have_spill && 1797 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1798 *bpp = &dn->dn_phys->dn_spill; 1799 else 1800 *bpp = NULL; 1801 dbuf_add_ref(dn->dn_dbuf, NULL); 1802 *parentp = dn->dn_dbuf; 1803 mutex_exit(&dn->dn_mtx); 1804 return (0); 1805 } 1806 1807 if (dn->dn_phys->dn_nlevels == 0) 1808 nlevels = 1; 1809 else 1810 nlevels = dn->dn_phys->dn_nlevels; 1811 1812 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1813 1814 ASSERT3U(level * epbs, <, 64); 1815 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1816 if (level >= nlevels || 1817 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1818 /* the buffer has no parent yet */ 1819 return (SET_ERROR(ENOENT)); 1820 } else if (level < nlevels-1) { 1821 /* this block is referenced from an indirect block */ 1822 int err = dbuf_hold_impl(dn, level+1, 1823 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 1824 if (err) 1825 return (err); 1826 err = dbuf_read(*parentp, NULL, 1827 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1828 if (err) { 1829 dbuf_rele(*parentp, NULL); 1830 *parentp = NULL; 1831 return (err); 1832 } 1833 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1834 (blkid & ((1ULL << epbs) - 1)); 1835 return (0); 1836 } else { 1837 /* the block is referenced from the dnode */ 1838 ASSERT3U(level, ==, nlevels-1); 1839 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1840 blkid < dn->dn_phys->dn_nblkptr); 1841 if (dn->dn_dbuf) { 1842 dbuf_add_ref(dn->dn_dbuf, NULL); 1843 *parentp = dn->dn_dbuf; 1844 } 1845 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1846 return (0); 1847 } 1848} 1849 1850static dmu_buf_impl_t * 1851dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1852 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1853{ 1854 objset_t *os = dn->dn_objset; 1855 dmu_buf_impl_t *db, *odb; 1856 1857 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1858 ASSERT(dn->dn_type != DMU_OT_NONE); 1859 1860 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1861 1862 db->db_objset = os; 1863 db->db.db_object = dn->dn_object; 1864 db->db_level = level; 1865 db->db_blkid = blkid; 1866 db->db_last_dirty = NULL; 1867 db->db_dirtycnt = 0; 1868 db->db_dnode_handle = dn->dn_handle; 1869 db->db_parent = parent; 1870 db->db_blkptr = blkptr; 1871 1872 db->db_user = NULL; 1873 db->db_user_immediate_evict = FALSE; 1874 db->db_freed_in_flight = FALSE; 1875 db->db_pending_evict = FALSE; 1876 1877 if (blkid == DMU_BONUS_BLKID) { 1878 ASSERT3P(parent, ==, dn->dn_dbuf); 1879 db->db.db_size = DN_MAX_BONUSLEN - 1880 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1881 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1882 db->db.db_offset = DMU_BONUS_BLKID; 1883 db->db_state = DB_UNCACHED; 1884 /* the bonus dbuf is not placed in the hash table */ 1885 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1886 return (db); 1887 } else if (blkid == DMU_SPILL_BLKID) { 1888 db->db.db_size = (blkptr != NULL) ? 1889 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1890 db->db.db_offset = 0; 1891 } else { 1892 int blocksize = 1893 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1894 db->db.db_size = blocksize; 1895 db->db.db_offset = db->db_blkid * blocksize; 1896 } 1897 1898 /* 1899 * Hold the dn_dbufs_mtx while we get the new dbuf 1900 * in the hash table *and* added to the dbufs list. 1901 * This prevents a possible deadlock with someone 1902 * trying to look up this dbuf before its added to the 1903 * dn_dbufs list. 1904 */ 1905 mutex_enter(&dn->dn_dbufs_mtx); 1906 db->db_state = DB_EVICTING; 1907 if ((odb = dbuf_hash_insert(db)) != NULL) { 1908 /* someone else inserted it first */ 1909 kmem_cache_free(dbuf_cache, db); 1910 mutex_exit(&dn->dn_dbufs_mtx); 1911 return (odb); 1912 } 1913 avl_add(&dn->dn_dbufs, db); 1914 if (db->db_level == 0 && db->db_blkid >= 1915 dn->dn_unlisted_l0_blkid) 1916 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1917 db->db_state = DB_UNCACHED; 1918 mutex_exit(&dn->dn_dbufs_mtx); 1919 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1920 1921 if (parent && parent != dn->dn_dbuf) 1922 dbuf_add_ref(parent, db); 1923 1924 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1925 refcount_count(&dn->dn_holds) > 0); 1926 (void) refcount_add(&dn->dn_holds, db); 1927 atomic_inc_32(&dn->dn_dbufs_count); 1928 1929 dprintf_dbuf(db, "db=%p\n", db); 1930 1931 return (db); 1932} 1933 1934static int 1935dbuf_do_evict(void *private) 1936{ 1937 dmu_buf_impl_t *db = private; 1938 1939 if (!MUTEX_HELD(&db->db_mtx)) 1940 mutex_enter(&db->db_mtx); 1941 1942 ASSERT(refcount_is_zero(&db->db_holds)); 1943 1944 if (db->db_state != DB_EVICTING) { 1945 ASSERT(db->db_state == DB_CACHED); 1946 DBUF_VERIFY(db); 1947 db->db_buf = NULL; 1948 dbuf_evict(db); 1949 } else { 1950 mutex_exit(&db->db_mtx); 1951 dbuf_destroy(db); 1952 } 1953 return (0); 1954} 1955 1956static void 1957dbuf_destroy(dmu_buf_impl_t *db) 1958{ 1959 ASSERT(refcount_is_zero(&db->db_holds)); 1960 1961 if (db->db_blkid != DMU_BONUS_BLKID) { 1962 /* 1963 * If this dbuf is still on the dn_dbufs list, 1964 * remove it from that list. 1965 */ 1966 if (db->db_dnode_handle != NULL) { 1967 dnode_t *dn; 1968 1969 DB_DNODE_ENTER(db); 1970 dn = DB_DNODE(db); 1971 mutex_enter(&dn->dn_dbufs_mtx); 1972 avl_remove(&dn->dn_dbufs, db); 1973 atomic_dec_32(&dn->dn_dbufs_count); 1974 mutex_exit(&dn->dn_dbufs_mtx); 1975 DB_DNODE_EXIT(db); 1976 /* 1977 * Decrementing the dbuf count means that the hold 1978 * corresponding to the removed dbuf is no longer 1979 * discounted in dnode_move(), so the dnode cannot be 1980 * moved until after we release the hold. 1981 */ 1982 dnode_rele(dn, db); 1983 db->db_dnode_handle = NULL; 1984 } 1985 dbuf_hash_remove(db); 1986 } 1987 db->db_parent = NULL; 1988 db->db_buf = NULL; 1989 1990 ASSERT(db->db.db_data == NULL); 1991 ASSERT(db->db_hash_next == NULL); 1992 ASSERT(db->db_blkptr == NULL); 1993 ASSERT(db->db_data_pending == NULL); 1994 1995 kmem_cache_free(dbuf_cache, db); 1996 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1997} 1998 1999typedef struct dbuf_prefetch_arg { 2000 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2001 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2002 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2003 int dpa_curlevel; /* The current level that we're reading */ 2004 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2005 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2006 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2007} dbuf_prefetch_arg_t; 2008 2009/* 2010 * Actually issue the prefetch read for the block given. 2011 */ 2012static void 2013dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2014{ 2015 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2016 return; 2017 2018 arc_flags_t aflags = 2019 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2020 2021 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2022 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2023 ASSERT(dpa->dpa_zio != NULL); 2024 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2025 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2026 &aflags, &dpa->dpa_zb); 2027} 2028 2029/* 2030 * Called when an indirect block above our prefetch target is read in. This 2031 * will either read in the next indirect block down the tree or issue the actual 2032 * prefetch if the next block down is our target. 2033 */ 2034static void 2035dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2036{ 2037 dbuf_prefetch_arg_t *dpa = private; 2038 2039 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2040 ASSERT3S(dpa->dpa_curlevel, >, 0); 2041 if (zio != NULL) { 2042 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2043 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2044 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2045 } 2046 2047 dpa->dpa_curlevel--; 2048 2049 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2050 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2051 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2052 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2053 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2054 kmem_free(dpa, sizeof (*dpa)); 2055 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2056 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2057 dbuf_issue_final_prefetch(dpa, bp); 2058 kmem_free(dpa, sizeof (*dpa)); 2059 } else { 2060 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2061 zbookmark_phys_t zb; 2062 2063 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2064 2065 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2066 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2067 2068 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2069 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2070 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2071 &iter_aflags, &zb); 2072 } 2073 (void) arc_buf_remove_ref(abuf, private); 2074} 2075 2076/* 2077 * Issue prefetch reads for the given block on the given level. If the indirect 2078 * blocks above that block are not in memory, we will read them in 2079 * asynchronously. As a result, this call never blocks waiting for a read to 2080 * complete. 2081 */ 2082void 2083dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2084 arc_flags_t aflags) 2085{ 2086 blkptr_t bp; 2087 int epbs, nlevels, curlevel; 2088 uint64_t curblkid; 2089 2090 ASSERT(blkid != DMU_BONUS_BLKID); 2091 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2092 2093 if (blkid > dn->dn_maxblkid) 2094 return; 2095 2096 if (dnode_block_freed(dn, blkid)) 2097 return; 2098 2099 /* 2100 * This dnode hasn't been written to disk yet, so there's nothing to 2101 * prefetch. 2102 */ 2103 nlevels = dn->dn_phys->dn_nlevels; 2104 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2105 return; 2106 2107 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2108 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2109 return; 2110 2111 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2112 level, blkid); 2113 if (db != NULL) { 2114 mutex_exit(&db->db_mtx); 2115 /* 2116 * This dbuf already exists. It is either CACHED, or 2117 * (we assume) about to be read or filled. 2118 */ 2119 return; 2120 } 2121 2122 /* 2123 * Find the closest ancestor (indirect block) of the target block 2124 * that is present in the cache. In this indirect block, we will 2125 * find the bp that is at curlevel, curblkid. 2126 */ 2127 curlevel = level; 2128 curblkid = blkid; 2129 while (curlevel < nlevels - 1) { 2130 int parent_level = curlevel + 1; 2131 uint64_t parent_blkid = curblkid >> epbs; 2132 dmu_buf_impl_t *db; 2133 2134 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2135 FALSE, TRUE, FTAG, &db) == 0) { 2136 blkptr_t *bpp = db->db_buf->b_data; 2137 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2138 dbuf_rele(db, FTAG); 2139 break; 2140 } 2141 2142 curlevel = parent_level; 2143 curblkid = parent_blkid; 2144 } 2145 2146 if (curlevel == nlevels - 1) { 2147 /* No cached indirect blocks found. */ 2148 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2149 bp = dn->dn_phys->dn_blkptr[curblkid]; 2150 } 2151 if (BP_IS_HOLE(&bp)) 2152 return; 2153 2154 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2155 2156 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2157 ZIO_FLAG_CANFAIL); 2158 2159 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2160 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2161 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2162 dn->dn_object, level, blkid); 2163 dpa->dpa_curlevel = curlevel; 2164 dpa->dpa_prio = prio; 2165 dpa->dpa_aflags = aflags; 2166 dpa->dpa_spa = dn->dn_objset->os_spa; 2167 dpa->dpa_epbs = epbs; 2168 dpa->dpa_zio = pio; 2169 2170 /* 2171 * If we have the indirect just above us, no need to do the asynchronous 2172 * prefetch chain; we'll just run the last step ourselves. If we're at 2173 * a higher level, though, we want to issue the prefetches for all the 2174 * indirect blocks asynchronously, so we can go on with whatever we were 2175 * doing. 2176 */ 2177 if (curlevel == level) { 2178 ASSERT3U(curblkid, ==, blkid); 2179 dbuf_issue_final_prefetch(dpa, &bp); 2180 kmem_free(dpa, sizeof (*dpa)); 2181 } else { 2182 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2183 zbookmark_phys_t zb; 2184 2185 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2186 dn->dn_object, curlevel, curblkid); 2187 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2188 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2189 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2190 &iter_aflags, &zb); 2191 } 2192 /* 2193 * We use pio here instead of dpa_zio since it's possible that 2194 * dpa may have already been freed. 2195 */ 2196 zio_nowait(pio); 2197} 2198 2199/* 2200 * Returns with db_holds incremented, and db_mtx not held. 2201 * Note: dn_struct_rwlock must be held. 2202 */ 2203int 2204dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2205 boolean_t fail_sparse, boolean_t fail_uncached, 2206 void *tag, dmu_buf_impl_t **dbp) 2207{ 2208 dmu_buf_impl_t *db, *parent = NULL; 2209 2210 ASSERT(blkid != DMU_BONUS_BLKID); 2211 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2212 ASSERT3U(dn->dn_nlevels, >, level); 2213 2214 *dbp = NULL; 2215top: 2216 /* dbuf_find() returns with db_mtx held */ 2217 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2218 2219 if (db == NULL) { 2220 blkptr_t *bp = NULL; 2221 int err; 2222 2223 if (fail_uncached) 2224 return (SET_ERROR(ENOENT)); 2225 2226 ASSERT3P(parent, ==, NULL); 2227 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2228 if (fail_sparse) { 2229 if (err == 0 && bp && BP_IS_HOLE(bp)) 2230 err = SET_ERROR(ENOENT); 2231 if (err) { 2232 if (parent) 2233 dbuf_rele(parent, NULL); 2234 return (err); 2235 } 2236 } 2237 if (err && err != ENOENT) 2238 return (err); 2239 db = dbuf_create(dn, level, blkid, parent, bp); 2240 } 2241 2242 if (fail_uncached && db->db_state != DB_CACHED) { 2243 mutex_exit(&db->db_mtx); 2244 return (SET_ERROR(ENOENT)); 2245 } 2246 2247 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 2248 arc_buf_add_ref(db->db_buf, db); 2249 if (db->db_buf->b_data == NULL) { 2250 dbuf_clear(db); 2251 if (parent) { 2252 dbuf_rele(parent, NULL); 2253 parent = NULL; 2254 } 2255 goto top; 2256 } 2257 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2258 } 2259 2260 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2261 2262 /* 2263 * If this buffer is currently syncing out, and we are are 2264 * still referencing it from db_data, we need to make a copy 2265 * of it in case we decide we want to dirty it again in this txg. 2266 */ 2267 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2268 dn->dn_object != DMU_META_DNODE_OBJECT && 2269 db->db_state == DB_CACHED && db->db_data_pending) { 2270 dbuf_dirty_record_t *dr = db->db_data_pending; 2271 2272 if (dr->dt.dl.dr_data == db->db_buf) { 2273 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2274 2275 dbuf_set_data(db, 2276 arc_buf_alloc(dn->dn_objset->os_spa, 2277 db->db.db_size, db, type)); 2278 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2279 db->db.db_size); 2280 } 2281 } 2282 2283 (void) refcount_add(&db->db_holds, tag); 2284 DBUF_VERIFY(db); 2285 mutex_exit(&db->db_mtx); 2286 2287 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2288 if (parent) 2289 dbuf_rele(parent, NULL); 2290 2291 ASSERT3P(DB_DNODE(db), ==, dn); 2292 ASSERT3U(db->db_blkid, ==, blkid); 2293 ASSERT3U(db->db_level, ==, level); 2294 *dbp = db; 2295 2296 return (0); 2297} 2298 2299dmu_buf_impl_t * 2300dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2301{ 2302 return (dbuf_hold_level(dn, 0, blkid, tag)); 2303} 2304 2305dmu_buf_impl_t * 2306dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2307{ 2308 dmu_buf_impl_t *db; 2309 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2310 return (err ? NULL : db); 2311} 2312 2313void 2314dbuf_create_bonus(dnode_t *dn) 2315{ 2316 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2317 2318 ASSERT(dn->dn_bonus == NULL); 2319 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2320} 2321 2322int 2323dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2324{ 2325 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2326 dnode_t *dn; 2327 2328 if (db->db_blkid != DMU_SPILL_BLKID) 2329 return (SET_ERROR(ENOTSUP)); 2330 if (blksz == 0) 2331 blksz = SPA_MINBLOCKSIZE; 2332 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2333 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2334 2335 DB_DNODE_ENTER(db); 2336 dn = DB_DNODE(db); 2337 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2338 dbuf_new_size(db, blksz, tx); 2339 rw_exit(&dn->dn_struct_rwlock); 2340 DB_DNODE_EXIT(db); 2341 2342 return (0); 2343} 2344 2345void 2346dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2347{ 2348 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2349} 2350 2351#pragma weak dmu_buf_add_ref = dbuf_add_ref 2352void 2353dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2354{ 2355 int64_t holds = refcount_add(&db->db_holds, tag); 2356 ASSERT(holds > 1); 2357} 2358 2359#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2360boolean_t 2361dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2362 void *tag) 2363{ 2364 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2365 dmu_buf_impl_t *found_db; 2366 boolean_t result = B_FALSE; 2367 2368 if (db->db_blkid == DMU_BONUS_BLKID) 2369 found_db = dbuf_find_bonus(os, obj); 2370 else 2371 found_db = dbuf_find(os, obj, 0, blkid); 2372 2373 if (found_db != NULL) { 2374 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2375 (void) refcount_add(&db->db_holds, tag); 2376 result = B_TRUE; 2377 } 2378 mutex_exit(&db->db_mtx); 2379 } 2380 return (result); 2381} 2382 2383/* 2384 * If you call dbuf_rele() you had better not be referencing the dnode handle 2385 * unless you have some other direct or indirect hold on the dnode. (An indirect 2386 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2387 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2388 * dnode's parent dbuf evicting its dnode handles. 2389 */ 2390void 2391dbuf_rele(dmu_buf_impl_t *db, void *tag) 2392{ 2393 mutex_enter(&db->db_mtx); 2394 dbuf_rele_and_unlock(db, tag); 2395} 2396 2397void 2398dmu_buf_rele(dmu_buf_t *db, void *tag) 2399{ 2400 dbuf_rele((dmu_buf_impl_t *)db, tag); 2401} 2402 2403/* 2404 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2405 * db_dirtycnt and db_holds to be updated atomically. 2406 */ 2407void 2408dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2409{ 2410 int64_t holds; 2411 2412 ASSERT(MUTEX_HELD(&db->db_mtx)); 2413 DBUF_VERIFY(db); 2414 2415 /* 2416 * Remove the reference to the dbuf before removing its hold on the 2417 * dnode so we can guarantee in dnode_move() that a referenced bonus 2418 * buffer has a corresponding dnode hold. 2419 */ 2420 holds = refcount_remove(&db->db_holds, tag); 2421 ASSERT(holds >= 0); 2422 2423 /* 2424 * We can't freeze indirects if there is a possibility that they 2425 * may be modified in the current syncing context. 2426 */ 2427 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2428 arc_buf_freeze(db->db_buf); 2429 2430 if (holds == db->db_dirtycnt && 2431 db->db_level == 0 && db->db_user_immediate_evict) 2432 dbuf_evict_user(db); 2433 2434 if (holds == 0) { 2435 if (db->db_blkid == DMU_BONUS_BLKID) { 2436 dnode_t *dn; 2437 boolean_t evict_dbuf = db->db_pending_evict; 2438 2439 /* 2440 * If the dnode moves here, we cannot cross this 2441 * barrier until the move completes. 2442 */ 2443 DB_DNODE_ENTER(db); 2444 2445 dn = DB_DNODE(db); 2446 atomic_dec_32(&dn->dn_dbufs_count); 2447 2448 /* 2449 * Decrementing the dbuf count means that the bonus 2450 * buffer's dnode hold is no longer discounted in 2451 * dnode_move(). The dnode cannot move until after 2452 * the dnode_rele() below. 2453 */ 2454 DB_DNODE_EXIT(db); 2455 2456 /* 2457 * Do not reference db after its lock is dropped. 2458 * Another thread may evict it. 2459 */ 2460 mutex_exit(&db->db_mtx); 2461 2462 if (evict_dbuf) 2463 dnode_evict_bonus(dn); 2464 2465 dnode_rele(dn, db); 2466 } else if (db->db_buf == NULL) { 2467 /* 2468 * This is a special case: we never associated this 2469 * dbuf with any data allocated from the ARC. 2470 */ 2471 ASSERT(db->db_state == DB_UNCACHED || 2472 db->db_state == DB_NOFILL); 2473 dbuf_evict(db); 2474 } else if (arc_released(db->db_buf)) { 2475 arc_buf_t *buf = db->db_buf; 2476 /* 2477 * This dbuf has anonymous data associated with it. 2478 */ 2479 dbuf_clear_data(db); 2480 VERIFY(arc_buf_remove_ref(buf, db)); 2481 dbuf_evict(db); 2482 } else { 2483 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2484 2485 /* 2486 * A dbuf will be eligible for eviction if either the 2487 * 'primarycache' property is set or a duplicate 2488 * copy of this buffer is already cached in the arc. 2489 * 2490 * In the case of the 'primarycache' a buffer 2491 * is considered for eviction if it matches the 2492 * criteria set in the property. 2493 * 2494 * To decide if our buffer is considered a 2495 * duplicate, we must call into the arc to determine 2496 * if multiple buffers are referencing the same 2497 * block on-disk. If so, then we simply evict 2498 * ourselves. 2499 */ 2500 if (!DBUF_IS_CACHEABLE(db)) { 2501 if (db->db_blkptr != NULL && 2502 !BP_IS_HOLE(db->db_blkptr) && 2503 !BP_IS_EMBEDDED(db->db_blkptr)) { 2504 spa_t *spa = 2505 dmu_objset_spa(db->db_objset); 2506 blkptr_t bp = *db->db_blkptr; 2507 dbuf_clear(db); 2508 arc_freed(spa, &bp); 2509 } else { 2510 dbuf_clear(db); 2511 } 2512 } else if (db->db_pending_evict || 2513 arc_buf_eviction_needed(db->db_buf)) { 2514 dbuf_clear(db); 2515 } else { 2516 mutex_exit(&db->db_mtx); 2517 } 2518 } 2519 } else { 2520 mutex_exit(&db->db_mtx); 2521 } 2522} 2523 2524#pragma weak dmu_buf_refcount = dbuf_refcount 2525uint64_t 2526dbuf_refcount(dmu_buf_impl_t *db) 2527{ 2528 return (refcount_count(&db->db_holds)); 2529} 2530 2531void * 2532dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2533 dmu_buf_user_t *new_user) 2534{ 2535 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2536 2537 mutex_enter(&db->db_mtx); 2538 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2539 if (db->db_user == old_user) 2540 db->db_user = new_user; 2541 else 2542 old_user = db->db_user; 2543 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2544 mutex_exit(&db->db_mtx); 2545 2546 return (old_user); 2547} 2548 2549void * 2550dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2551{ 2552 return (dmu_buf_replace_user(db_fake, NULL, user)); 2553} 2554 2555void * 2556dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2557{ 2558 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2559 2560 db->db_user_immediate_evict = TRUE; 2561 return (dmu_buf_set_user(db_fake, user)); 2562} 2563 2564void * 2565dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2566{ 2567 return (dmu_buf_replace_user(db_fake, user, NULL)); 2568} 2569 2570void * 2571dmu_buf_get_user(dmu_buf_t *db_fake) 2572{ 2573 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2574 2575 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2576 return (db->db_user); 2577} 2578 2579void 2580dmu_buf_user_evict_wait() 2581{ 2582 taskq_wait(dbu_evict_taskq); 2583} 2584 2585boolean_t 2586dmu_buf_freeable(dmu_buf_t *dbuf) 2587{ 2588 boolean_t res = B_FALSE; 2589 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2590 2591 if (db->db_blkptr) 2592 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2593 db->db_blkptr, db->db_blkptr->blk_birth); 2594 2595 return (res); 2596} 2597 2598blkptr_t * 2599dmu_buf_get_blkptr(dmu_buf_t *db) 2600{ 2601 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2602 return (dbi->db_blkptr); 2603} 2604 2605static void 2606dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2607{ 2608 /* ASSERT(dmu_tx_is_syncing(tx) */ 2609 ASSERT(MUTEX_HELD(&db->db_mtx)); 2610 2611 if (db->db_blkptr != NULL) 2612 return; 2613 2614 if (db->db_blkid == DMU_SPILL_BLKID) { 2615 db->db_blkptr = &dn->dn_phys->dn_spill; 2616 BP_ZERO(db->db_blkptr); 2617 return; 2618 } 2619 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2620 /* 2621 * This buffer was allocated at a time when there was 2622 * no available blkptrs from the dnode, or it was 2623 * inappropriate to hook it in (i.e., nlevels mis-match). 2624 */ 2625 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2626 ASSERT(db->db_parent == NULL); 2627 db->db_parent = dn->dn_dbuf; 2628 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2629 DBUF_VERIFY(db); 2630 } else { 2631 dmu_buf_impl_t *parent = db->db_parent; 2632 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2633 2634 ASSERT(dn->dn_phys->dn_nlevels > 1); 2635 if (parent == NULL) { 2636 mutex_exit(&db->db_mtx); 2637 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2638 parent = dbuf_hold_level(dn, db->db_level + 1, 2639 db->db_blkid >> epbs, db); 2640 rw_exit(&dn->dn_struct_rwlock); 2641 mutex_enter(&db->db_mtx); 2642 db->db_parent = parent; 2643 } 2644 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2645 (db->db_blkid & ((1ULL << epbs) - 1)); 2646 DBUF_VERIFY(db); 2647 } 2648} 2649 2650static void 2651dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2652{ 2653 dmu_buf_impl_t *db = dr->dr_dbuf; 2654 dnode_t *dn; 2655 zio_t *zio; 2656 2657 ASSERT(dmu_tx_is_syncing(tx)); 2658 2659 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2660 2661 mutex_enter(&db->db_mtx); 2662 2663 ASSERT(db->db_level > 0); 2664 DBUF_VERIFY(db); 2665 2666 /* Read the block if it hasn't been read yet. */ 2667 if (db->db_buf == NULL) { 2668 mutex_exit(&db->db_mtx); 2669 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2670 mutex_enter(&db->db_mtx); 2671 } 2672 ASSERT3U(db->db_state, ==, DB_CACHED); 2673 ASSERT(db->db_buf != NULL); 2674 2675 DB_DNODE_ENTER(db); 2676 dn = DB_DNODE(db); 2677 /* Indirect block size must match what the dnode thinks it is. */ 2678 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2679 dbuf_check_blkptr(dn, db); 2680 DB_DNODE_EXIT(db); 2681 2682 /* Provide the pending dirty record to child dbufs */ 2683 db->db_data_pending = dr; 2684 2685 mutex_exit(&db->db_mtx); 2686 dbuf_write(dr, db->db_buf, tx); 2687 2688 zio = dr->dr_zio; 2689 mutex_enter(&dr->dt.di.dr_mtx); 2690 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2691 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2692 mutex_exit(&dr->dt.di.dr_mtx); 2693 zio_nowait(zio); 2694} 2695 2696static void 2697dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2698{ 2699 arc_buf_t **datap = &dr->dt.dl.dr_data; 2700 dmu_buf_impl_t *db = dr->dr_dbuf; 2701 dnode_t *dn; 2702 objset_t *os; 2703 uint64_t txg = tx->tx_txg; 2704 2705 ASSERT(dmu_tx_is_syncing(tx)); 2706 2707 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2708 2709 mutex_enter(&db->db_mtx); 2710 /* 2711 * To be synced, we must be dirtied. But we 2712 * might have been freed after the dirty. 2713 */ 2714 if (db->db_state == DB_UNCACHED) { 2715 /* This buffer has been freed since it was dirtied */ 2716 ASSERT(db->db.db_data == NULL); 2717 } else if (db->db_state == DB_FILL) { 2718 /* This buffer was freed and is now being re-filled */ 2719 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2720 } else { 2721 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2722 } 2723 DBUF_VERIFY(db); 2724 2725 DB_DNODE_ENTER(db); 2726 dn = DB_DNODE(db); 2727 2728 if (db->db_blkid == DMU_SPILL_BLKID) { 2729 mutex_enter(&dn->dn_mtx); 2730 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2731 mutex_exit(&dn->dn_mtx); 2732 } 2733 2734 /* 2735 * If this is a bonus buffer, simply copy the bonus data into the 2736 * dnode. It will be written out when the dnode is synced (and it 2737 * will be synced, since it must have been dirty for dbuf_sync to 2738 * be called). 2739 */ 2740 if (db->db_blkid == DMU_BONUS_BLKID) { 2741 dbuf_dirty_record_t **drp; 2742 2743 ASSERT(*datap != NULL); 2744 ASSERT0(db->db_level); 2745 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2746 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2747 DB_DNODE_EXIT(db); 2748 2749 if (*datap != db->db.db_data) { 2750 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2751 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2752 } 2753 db->db_data_pending = NULL; 2754 drp = &db->db_last_dirty; 2755 while (*drp != dr) 2756 drp = &(*drp)->dr_next; 2757 ASSERT(dr->dr_next == NULL); 2758 ASSERT(dr->dr_dbuf == db); 2759 *drp = dr->dr_next; 2760 if (dr->dr_dbuf->db_level != 0) { 2761 list_destroy(&dr->dt.di.dr_children); 2762 mutex_destroy(&dr->dt.di.dr_mtx); 2763 } 2764 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2765 ASSERT(db->db_dirtycnt > 0); 2766 db->db_dirtycnt -= 1; 2767 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2768 return; 2769 } 2770 2771 os = dn->dn_objset; 2772 2773 /* 2774 * This function may have dropped the db_mtx lock allowing a dmu_sync 2775 * operation to sneak in. As a result, we need to ensure that we 2776 * don't check the dr_override_state until we have returned from 2777 * dbuf_check_blkptr. 2778 */ 2779 dbuf_check_blkptr(dn, db); 2780 2781 /* 2782 * If this buffer is in the middle of an immediate write, 2783 * wait for the synchronous IO to complete. 2784 */ 2785 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2786 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2787 cv_wait(&db->db_changed, &db->db_mtx); 2788 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2789 } 2790 2791 if (db->db_state != DB_NOFILL && 2792 dn->dn_object != DMU_META_DNODE_OBJECT && 2793 refcount_count(&db->db_holds) > 1 && 2794 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2795 *datap == db->db_buf) { 2796 /* 2797 * If this buffer is currently "in use" (i.e., there 2798 * are active holds and db_data still references it), 2799 * then make a copy before we start the write so that 2800 * any modifications from the open txg will not leak 2801 * into this write. 2802 * 2803 * NOTE: this copy does not need to be made for 2804 * objects only modified in the syncing context (e.g. 2805 * DNONE_DNODE blocks). 2806 */ 2807 int blksz = arc_buf_size(*datap); 2808 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2809 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2810 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2811 } 2812 db->db_data_pending = dr; 2813 2814 mutex_exit(&db->db_mtx); 2815 2816 dbuf_write(dr, *datap, tx); 2817 2818 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2819 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2820 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2821 DB_DNODE_EXIT(db); 2822 } else { 2823 /* 2824 * Although zio_nowait() does not "wait for an IO", it does 2825 * initiate the IO. If this is an empty write it seems plausible 2826 * that the IO could actually be completed before the nowait 2827 * returns. We need to DB_DNODE_EXIT() first in case 2828 * zio_nowait() invalidates the dbuf. 2829 */ 2830 DB_DNODE_EXIT(db); 2831 zio_nowait(dr->dr_zio); 2832 } 2833} 2834 2835void 2836dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 2837{ 2838 dbuf_dirty_record_t *dr; 2839 2840 while (dr = list_head(list)) { 2841 if (dr->dr_zio != NULL) { 2842 /* 2843 * If we find an already initialized zio then we 2844 * are processing the meta-dnode, and we have finished. 2845 * The dbufs for all dnodes are put back on the list 2846 * during processing, so that we can zio_wait() 2847 * these IOs after initiating all child IOs. 2848 */ 2849 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2850 DMU_META_DNODE_OBJECT); 2851 break; 2852 } 2853 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 2854 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 2855 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 2856 } 2857 list_remove(list, dr); 2858 if (dr->dr_dbuf->db_level > 0) 2859 dbuf_sync_indirect(dr, tx); 2860 else 2861 dbuf_sync_leaf(dr, tx); 2862 } 2863} 2864 2865/* ARGSUSED */ 2866static void 2867dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2868{ 2869 dmu_buf_impl_t *db = vdb; 2870 dnode_t *dn; 2871 blkptr_t *bp = zio->io_bp; 2872 blkptr_t *bp_orig = &zio->io_bp_orig; 2873 spa_t *spa = zio->io_spa; 2874 int64_t delta; 2875 uint64_t fill = 0; 2876 int i; 2877 2878 ASSERT3P(db->db_blkptr, ==, bp); 2879 2880 DB_DNODE_ENTER(db); 2881 dn = DB_DNODE(db); 2882 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2883 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2884 zio->io_prev_space_delta = delta; 2885 2886 if (bp->blk_birth != 0) { 2887 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2888 BP_GET_TYPE(bp) == dn->dn_type) || 2889 (db->db_blkid == DMU_SPILL_BLKID && 2890 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2891 BP_IS_EMBEDDED(bp)); 2892 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2893 } 2894 2895 mutex_enter(&db->db_mtx); 2896 2897#ifdef ZFS_DEBUG 2898 if (db->db_blkid == DMU_SPILL_BLKID) { 2899 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2900 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2901 db->db_blkptr == &dn->dn_phys->dn_spill); 2902 } 2903#endif 2904 2905 if (db->db_level == 0) { 2906 mutex_enter(&dn->dn_mtx); 2907 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2908 db->db_blkid != DMU_SPILL_BLKID) 2909 dn->dn_phys->dn_maxblkid = db->db_blkid; 2910 mutex_exit(&dn->dn_mtx); 2911 2912 if (dn->dn_type == DMU_OT_DNODE) { 2913 dnode_phys_t *dnp = db->db.db_data; 2914 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2915 i--, dnp++) { 2916 if (dnp->dn_type != DMU_OT_NONE) 2917 fill++; 2918 } 2919 } else { 2920 if (BP_IS_HOLE(bp)) { 2921 fill = 0; 2922 } else { 2923 fill = 1; 2924 } 2925 } 2926 } else { 2927 blkptr_t *ibp = db->db.db_data; 2928 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2929 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2930 if (BP_IS_HOLE(ibp)) 2931 continue; 2932 fill += BP_GET_FILL(ibp); 2933 } 2934 } 2935 DB_DNODE_EXIT(db); 2936 2937 if (!BP_IS_EMBEDDED(bp)) 2938 bp->blk_fill = fill; 2939 2940 mutex_exit(&db->db_mtx); 2941} 2942 2943/* 2944 * The SPA will call this callback several times for each zio - once 2945 * for every physical child i/o (zio->io_phys_children times). This 2946 * allows the DMU to monitor the progress of each logical i/o. For example, 2947 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 2948 * block. There may be a long delay before all copies/fragments are completed, 2949 * so this callback allows us to retire dirty space gradually, as the physical 2950 * i/os complete. 2951 */ 2952/* ARGSUSED */ 2953static void 2954dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 2955{ 2956 dmu_buf_impl_t *db = arg; 2957 objset_t *os = db->db_objset; 2958 dsl_pool_t *dp = dmu_objset_pool(os); 2959 dbuf_dirty_record_t *dr; 2960 int delta = 0; 2961 2962 dr = db->db_data_pending; 2963 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 2964 2965 /* 2966 * The callback will be called io_phys_children times. Retire one 2967 * portion of our dirty space each time we are called. Any rounding 2968 * error will be cleaned up by dsl_pool_sync()'s call to 2969 * dsl_pool_undirty_space(). 2970 */ 2971 delta = dr->dr_accounted / zio->io_phys_children; 2972 dsl_pool_undirty_space(dp, delta, zio->io_txg); 2973} 2974 2975/* ARGSUSED */ 2976static void 2977dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2978{ 2979 dmu_buf_impl_t *db = vdb; 2980 blkptr_t *bp_orig = &zio->io_bp_orig; 2981 blkptr_t *bp = db->db_blkptr; 2982 objset_t *os = db->db_objset; 2983 dmu_tx_t *tx = os->os_synctx; 2984 dbuf_dirty_record_t **drp, *dr; 2985 2986 ASSERT0(zio->io_error); 2987 ASSERT(db->db_blkptr == bp); 2988 2989 /* 2990 * For nopwrites and rewrites we ensure that the bp matches our 2991 * original and bypass all the accounting. 2992 */ 2993 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 2994 ASSERT(BP_EQUAL(bp, bp_orig)); 2995 } else { 2996 dsl_dataset_t *ds = os->os_dsl_dataset; 2997 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2998 dsl_dataset_block_born(ds, bp, tx); 2999 } 3000 3001 mutex_enter(&db->db_mtx); 3002 3003 DBUF_VERIFY(db); 3004 3005 drp = &db->db_last_dirty; 3006 while ((dr = *drp) != db->db_data_pending) 3007 drp = &dr->dr_next; 3008 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3009 ASSERT(dr->dr_dbuf == db); 3010 ASSERT(dr->dr_next == NULL); 3011 *drp = dr->dr_next; 3012 3013#ifdef ZFS_DEBUG 3014 if (db->db_blkid == DMU_SPILL_BLKID) { 3015 dnode_t *dn; 3016 3017 DB_DNODE_ENTER(db); 3018 dn = DB_DNODE(db); 3019 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3020 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3021 db->db_blkptr == &dn->dn_phys->dn_spill); 3022 DB_DNODE_EXIT(db); 3023 } 3024#endif 3025 3026 if (db->db_level == 0) { 3027 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3028 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3029 if (db->db_state != DB_NOFILL) { 3030 if (dr->dt.dl.dr_data != db->db_buf) 3031 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 3032 db)); 3033 else if (!arc_released(db->db_buf)) 3034 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3035 } 3036 } else { 3037 dnode_t *dn; 3038 3039 DB_DNODE_ENTER(db); 3040 dn = DB_DNODE(db); 3041 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3042 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3043 if (!BP_IS_HOLE(db->db_blkptr)) { 3044 int epbs = 3045 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3046 ASSERT3U(db->db_blkid, <=, 3047 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3048 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3049 db->db.db_size); 3050 if (!arc_released(db->db_buf)) 3051 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3052 } 3053 DB_DNODE_EXIT(db); 3054 mutex_destroy(&dr->dt.di.dr_mtx); 3055 list_destroy(&dr->dt.di.dr_children); 3056 } 3057 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3058 3059 cv_broadcast(&db->db_changed); 3060 ASSERT(db->db_dirtycnt > 0); 3061 db->db_dirtycnt -= 1; 3062 db->db_data_pending = NULL; 3063 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3064} 3065 3066static void 3067dbuf_write_nofill_ready(zio_t *zio) 3068{ 3069 dbuf_write_ready(zio, NULL, zio->io_private); 3070} 3071 3072static void 3073dbuf_write_nofill_done(zio_t *zio) 3074{ 3075 dbuf_write_done(zio, NULL, zio->io_private); 3076} 3077 3078static void 3079dbuf_write_override_ready(zio_t *zio) 3080{ 3081 dbuf_dirty_record_t *dr = zio->io_private; 3082 dmu_buf_impl_t *db = dr->dr_dbuf; 3083 3084 dbuf_write_ready(zio, NULL, db); 3085} 3086 3087static void 3088dbuf_write_override_done(zio_t *zio) 3089{ 3090 dbuf_dirty_record_t *dr = zio->io_private; 3091 dmu_buf_impl_t *db = dr->dr_dbuf; 3092 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3093 3094 mutex_enter(&db->db_mtx); 3095 if (!BP_EQUAL(zio->io_bp, obp)) { 3096 if (!BP_IS_HOLE(obp)) 3097 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3098 arc_release(dr->dt.dl.dr_data, db); 3099 } 3100 mutex_exit(&db->db_mtx); 3101 3102 dbuf_write_done(zio, NULL, db); 3103} 3104 3105/* Issue I/O to commit a dirty buffer to disk. */ 3106static void 3107dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3108{ 3109 dmu_buf_impl_t *db = dr->dr_dbuf; 3110 dnode_t *dn; 3111 objset_t *os; 3112 dmu_buf_impl_t *parent = db->db_parent; 3113 uint64_t txg = tx->tx_txg; 3114 zbookmark_phys_t zb; 3115 zio_prop_t zp; 3116 zio_t *zio; 3117 int wp_flag = 0; 3118 3119 DB_DNODE_ENTER(db); 3120 dn = DB_DNODE(db); 3121 os = dn->dn_objset; 3122 3123 if (db->db_state != DB_NOFILL) { 3124 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3125 /* 3126 * Private object buffers are released here rather 3127 * than in dbuf_dirty() since they are only modified 3128 * in the syncing context and we don't want the 3129 * overhead of making multiple copies of the data. 3130 */ 3131 if (BP_IS_HOLE(db->db_blkptr)) { 3132 arc_buf_thaw(data); 3133 } else { 3134 dbuf_release_bp(db); 3135 } 3136 } 3137 } 3138 3139 if (parent != dn->dn_dbuf) { 3140 /* Our parent is an indirect block. */ 3141 /* We have a dirty parent that has been scheduled for write. */ 3142 ASSERT(parent && parent->db_data_pending); 3143 /* Our parent's buffer is one level closer to the dnode. */ 3144 ASSERT(db->db_level == parent->db_level-1); 3145 /* 3146 * We're about to modify our parent's db_data by modifying 3147 * our block pointer, so the parent must be released. 3148 */ 3149 ASSERT(arc_released(parent->db_buf)); 3150 zio = parent->db_data_pending->dr_zio; 3151 } else { 3152 /* Our parent is the dnode itself. */ 3153 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3154 db->db_blkid != DMU_SPILL_BLKID) || 3155 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3156 if (db->db_blkid != DMU_SPILL_BLKID) 3157 ASSERT3P(db->db_blkptr, ==, 3158 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3159 zio = dn->dn_zio; 3160 } 3161 3162 ASSERT(db->db_level == 0 || data == db->db_buf); 3163 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3164 ASSERT(zio); 3165 3166 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3167 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3168 db->db.db_object, db->db_level, db->db_blkid); 3169 3170 if (db->db_blkid == DMU_SPILL_BLKID) 3171 wp_flag = WP_SPILL; 3172 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3173 3174 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3175 DB_DNODE_EXIT(db); 3176 3177 if (db->db_level == 0 && 3178 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3179 /* 3180 * The BP for this block has been provided by open context 3181 * (by dmu_sync() or dmu_buf_write_embedded()). 3182 */ 3183 void *contents = (data != NULL) ? data->b_data : NULL; 3184 3185 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3186 db->db_blkptr, contents, db->db.db_size, &zp, 3187 dbuf_write_override_ready, NULL, dbuf_write_override_done, 3188 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3189 mutex_enter(&db->db_mtx); 3190 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3191 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3192 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3193 mutex_exit(&db->db_mtx); 3194 } else if (db->db_state == DB_NOFILL) { 3195 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3196 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3197 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3198 db->db_blkptr, NULL, db->db.db_size, &zp, 3199 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, 3200 ZIO_PRIORITY_ASYNC_WRITE, 3201 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3202 } else { 3203 ASSERT(arc_released(data)); 3204 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3205 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), 3206 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 3207 dbuf_write_physdone, dbuf_write_done, db, 3208 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3209 } 3210} 3211