1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5f65e61c0Sahrens * Common Development and Distribution License (the "License"). 6f65e61c0Sahrens * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 2206e0070dSMark Shellenbaum * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 233f2366c2SGordon Ross * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 2494c2d0ebSMatthew Ahrens * Copyright (c) 2012, 2017 by Delphix. All rights reserved. 25aad02571SSaso Kiselkov * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26810e43b2SBill Pijewski * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27bc9014e6SJustin Gibbs * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 28c3d26abcSMatthew Ahrens * Copyright (c) 2014 Integros [integros.com] 29fa9e4066Sahrens */ 30fa9e4066Sahrens 31fa9e4066Sahrens #include <sys/zfs_context.h> 32fa9e4066Sahrens #include <sys/dmu.h> 332f3d8780SMatthew Ahrens #include <sys/dmu_send.h> 34fa9e4066Sahrens #include <sys/dmu_impl.h> 35fa9e4066Sahrens #include <sys/dbuf.h> 36fa9e4066Sahrens #include <sys/dmu_objset.h> 37fa9e4066Sahrens #include <sys/dsl_dataset.h> 38fa9e4066Sahrens #include <sys/dsl_dir.h> 39fa9e4066Sahrens #include <sys/dmu_tx.h> 40fa9e4066Sahrens #include <sys/spa.h> 41fa9e4066Sahrens #include <sys/zio.h> 42fa9e4066Sahrens #include <sys/dmu_zfetch.h> 430a586ceaSMark Shellenbaum #include <sys/sa.h> 440a586ceaSMark Shellenbaum #include <sys/sa_impl.h> 455d7b4d43SMatthew Ahrens #include <sys/zfeature.h> 465d7b4d43SMatthew Ahrens #include <sys/blkptr.h> 47bf16b11eSMatthew Ahrens #include <sys/range_tree.h> 48dcbf3bd6SGeorge Wilson #include <sys/callb.h> 49770499e1SDan Kimmel #include <sys/abd.h> 50dcbf3bd6SGeorge Wilson 51dcbf3bd6SGeorge Wilson uint_t zfs_dbuf_evict_key; 52fa9e4066Sahrens 533b2aab18SMatthew Ahrens static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 54088f3894Sahrens static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 55fa9e4066Sahrens 56bc9014e6SJustin Gibbs #ifndef __lint 57bc9014e6SJustin Gibbs extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, 5840510e8eSJosef 'Jeff' Sipek dmu_buf_evict_func_t *evict_func_sync, 5940510e8eSJosef 'Jeff' Sipek dmu_buf_evict_func_t *evict_func_async, 6040510e8eSJosef 'Jeff' Sipek dmu_buf_t **clear_on_evict_dbufp); 61bc9014e6SJustin Gibbs #endif /* ! __lint */ 62bc9014e6SJustin Gibbs 63fa9e4066Sahrens /* 64fa9e4066Sahrens * Global data structures and functions for the dbuf cache. 65fa9e4066Sahrens */ 66dcbf3bd6SGeorge Wilson static kmem_cache_t *dbuf_kmem_cache; 67bc9014e6SJustin Gibbs static taskq_t *dbu_evict_taskq; 68fa9e4066Sahrens 69dcbf3bd6SGeorge Wilson static kthread_t *dbuf_cache_evict_thread; 70dcbf3bd6SGeorge Wilson static kmutex_t dbuf_evict_lock; 71dcbf3bd6SGeorge Wilson static kcondvar_t dbuf_evict_cv; 72dcbf3bd6SGeorge Wilson static boolean_t dbuf_evict_thread_exit; 73dcbf3bd6SGeorge Wilson 74dcbf3bd6SGeorge Wilson /* 75dcbf3bd6SGeorge Wilson * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that 76dcbf3bd6SGeorge Wilson * are not currently held but have been recently released. These dbufs 77dcbf3bd6SGeorge Wilson * are not eligible for arc eviction until they are aged out of the cache. 78dcbf3bd6SGeorge Wilson * Dbufs are added to the dbuf cache once the last hold is released. If a 79dcbf3bd6SGeorge Wilson * dbuf is later accessed and still exists in the dbuf cache, then it will 80dcbf3bd6SGeorge Wilson * be removed from the cache and later re-added to the head of the cache. 81dcbf3bd6SGeorge Wilson * Dbufs that are aged out of the cache will be immediately destroyed and 82dcbf3bd6SGeorge Wilson * become eligible for arc eviction. 83dcbf3bd6SGeorge Wilson */ 8494c2d0ebSMatthew Ahrens static multilist_t *dbuf_cache; 85dcbf3bd6SGeorge Wilson static refcount_t dbuf_cache_size; 86dcbf3bd6SGeorge Wilson uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024; 87dcbf3bd6SGeorge Wilson 88dcbf3bd6SGeorge Wilson /* Cap the size of the dbuf cache to log2 fraction of arc size. */ 89dcbf3bd6SGeorge Wilson int dbuf_cache_max_shift = 5; 90dcbf3bd6SGeorge Wilson 91dcbf3bd6SGeorge Wilson /* 92dcbf3bd6SGeorge Wilson * The dbuf cache uses a three-stage eviction policy: 93dcbf3bd6SGeorge Wilson * - A low water marker designates when the dbuf eviction thread 94dcbf3bd6SGeorge Wilson * should stop evicting from the dbuf cache. 95dcbf3bd6SGeorge Wilson * - When we reach the maximum size (aka mid water mark), we 96dcbf3bd6SGeorge Wilson * signal the eviction thread to run. 97dcbf3bd6SGeorge Wilson * - The high water mark indicates when the eviction thread 98dcbf3bd6SGeorge Wilson * is unable to keep up with the incoming load and eviction must 99dcbf3bd6SGeorge Wilson * happen in the context of the calling thread. 100dcbf3bd6SGeorge Wilson * 101dcbf3bd6SGeorge Wilson * The dbuf cache: 102dcbf3bd6SGeorge Wilson * (max size) 103dcbf3bd6SGeorge Wilson * low water mid water hi water 104dcbf3bd6SGeorge Wilson * +----------------------------------------+----------+----------+ 105dcbf3bd6SGeorge Wilson * | | | | 106dcbf3bd6SGeorge Wilson * | | | | 107dcbf3bd6SGeorge Wilson * | | | | 108dcbf3bd6SGeorge Wilson * | | | | 109dcbf3bd6SGeorge Wilson * +----------------------------------------+----------+----------+ 110dcbf3bd6SGeorge Wilson * stop signal evict 111dcbf3bd6SGeorge Wilson * evicting eviction directly 112dcbf3bd6SGeorge Wilson * thread 113dcbf3bd6SGeorge Wilson * 114dcbf3bd6SGeorge Wilson * The high and low water marks indicate the operating range for the eviction 115dcbf3bd6SGeorge Wilson * thread. The low water mark is, by default, 90% of the total size of the 116dcbf3bd6SGeorge Wilson * cache and the high water mark is at 110% (both of these percentages can be 117dcbf3bd6SGeorge Wilson * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, 118dcbf3bd6SGeorge Wilson * respectively). The eviction thread will try to ensure that the cache remains 119dcbf3bd6SGeorge Wilson * within this range by waking up every second and checking if the cache is 120dcbf3bd6SGeorge Wilson * above the low water mark. The thread can also be woken up by callers adding 121dcbf3bd6SGeorge Wilson * elements into the cache if the cache is larger than the mid water (i.e max 122dcbf3bd6SGeorge Wilson * cache size). Once the eviction thread is woken up and eviction is required, 123dcbf3bd6SGeorge Wilson * it will continue evicting buffers until it's able to reduce the cache size 124dcbf3bd6SGeorge Wilson * to the low water mark. If the cache size continues to grow and hits the high 125dcbf3bd6SGeorge Wilson * water mark, then callers adding elments to the cache will begin to evict 126dcbf3bd6SGeorge Wilson * directly from the cache until the cache is no longer above the high water 127dcbf3bd6SGeorge Wilson * mark. 128dcbf3bd6SGeorge Wilson */ 129dcbf3bd6SGeorge Wilson 130dcbf3bd6SGeorge Wilson /* 131dcbf3bd6SGeorge Wilson * The percentage above and below the maximum cache size. 132dcbf3bd6SGeorge Wilson */ 133dcbf3bd6SGeorge Wilson uint_t dbuf_cache_hiwater_pct = 10; 134dcbf3bd6SGeorge Wilson uint_t dbuf_cache_lowater_pct = 10; 135dcbf3bd6SGeorge Wilson 136fa9e4066Sahrens /* ARGSUSED */ 137fa9e4066Sahrens static int 138fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag) 139fa9e4066Sahrens { 140fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 141fa9e4066Sahrens bzero(db, sizeof (dmu_buf_impl_t)); 142fa9e4066Sahrens 143fa9e4066Sahrens mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 144fa9e4066Sahrens cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 145dcbf3bd6SGeorge Wilson multilist_link_init(&db->db_cache_link); 146fa9e4066Sahrens refcount_create(&db->db_holds); 1470f6d88adSAlex Reece 148fa9e4066Sahrens return (0); 149fa9e4066Sahrens } 150fa9e4066Sahrens 151fa9e4066Sahrens /* ARGSUSED */ 152fa9e4066Sahrens static void 153fa9e4066Sahrens dbuf_dest(void *vdb, void *unused) 154fa9e4066Sahrens { 155fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 156fa9e4066Sahrens mutex_destroy(&db->db_mtx); 157fa9e4066Sahrens cv_destroy(&db->db_changed); 158dcbf3bd6SGeorge Wilson ASSERT(!multilist_link_active(&db->db_cache_link)); 159fa9e4066Sahrens refcount_destroy(&db->db_holds); 160fa9e4066Sahrens } 161fa9e4066Sahrens 162fa9e4066Sahrens /* 163fa9e4066Sahrens * dbuf hash table routines 164fa9e4066Sahrens */ 165fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table; 166fa9e4066Sahrens 167fa9e4066Sahrens static uint64_t dbuf_hash_count; 168fa9e4066Sahrens 169fa9e4066Sahrens static uint64_t 170fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 171fa9e4066Sahrens { 172fa9e4066Sahrens uintptr_t osv = (uintptr_t)os; 173fa9e4066Sahrens uint64_t crc = -1ULL; 174fa9e4066Sahrens 175fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 176fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 177fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 178fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 179fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 180fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 181fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 182fa9e4066Sahrens 183fa9e4066Sahrens crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 184fa9e4066Sahrens 185fa9e4066Sahrens return (crc); 186fa9e4066Sahrens } 187fa9e4066Sahrens 188fa9e4066Sahrens #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 189fa9e4066Sahrens ((dbuf)->db.db_object == (obj) && \ 190fa9e4066Sahrens (dbuf)->db_objset == (os) && \ 191fa9e4066Sahrens (dbuf)->db_level == (level) && \ 192fa9e4066Sahrens (dbuf)->db_blkid == (blkid)) 193fa9e4066Sahrens 194fa9e4066Sahrens dmu_buf_impl_t * 195e57a022bSJustin T. Gibbs dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 196fa9e4066Sahrens { 197fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 198dcbf3bd6SGeorge Wilson uint64_t hv = dbuf_hash(os, obj, level, blkid); 199fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 200fa9e4066Sahrens dmu_buf_impl_t *db; 201fa9e4066Sahrens 202fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 203fa9e4066Sahrens for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 204fa9e4066Sahrens if (DBUF_EQUAL(db, os, obj, level, blkid)) { 205fa9e4066Sahrens mutex_enter(&db->db_mtx); 206ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 207fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 208fa9e4066Sahrens return (db); 209fa9e4066Sahrens } 210fa9e4066Sahrens mutex_exit(&db->db_mtx); 211fa9e4066Sahrens } 212fa9e4066Sahrens } 213fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 214fa9e4066Sahrens return (NULL); 215fa9e4066Sahrens } 216fa9e4066Sahrens 217e57a022bSJustin T. Gibbs static dmu_buf_impl_t * 218e57a022bSJustin T. Gibbs dbuf_find_bonus(objset_t *os, uint64_t object) 219e57a022bSJustin T. Gibbs { 220e57a022bSJustin T. Gibbs dnode_t *dn; 221e57a022bSJustin T. Gibbs dmu_buf_impl_t *db = NULL; 222e57a022bSJustin T. Gibbs 223e57a022bSJustin T. Gibbs if (dnode_hold(os, object, FTAG, &dn) == 0) { 224e57a022bSJustin T. Gibbs rw_enter(&dn->dn_struct_rwlock, RW_READER); 225e57a022bSJustin T. Gibbs if (dn->dn_bonus != NULL) { 226e57a022bSJustin T. Gibbs db = dn->dn_bonus; 227e57a022bSJustin T. Gibbs mutex_enter(&db->db_mtx); 228e57a022bSJustin T. Gibbs } 229e57a022bSJustin T. Gibbs rw_exit(&dn->dn_struct_rwlock); 230e57a022bSJustin T. Gibbs dnode_rele(dn, FTAG); 231e57a022bSJustin T. Gibbs } 232e57a022bSJustin T. Gibbs return (db); 233e57a022bSJustin T. Gibbs } 234e57a022bSJustin T. Gibbs 235fa9e4066Sahrens /* 236fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 237fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 238fa9e4066Sahrens * will be returned and the new element will not be inserted. 239fa9e4066Sahrens * Otherwise returns NULL. 240fa9e4066Sahrens */ 241fa9e4066Sahrens static dmu_buf_impl_t * 242fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db) 243fa9e4066Sahrens { 244fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 245503ad85cSMatthew Ahrens objset_t *os = db->db_objset; 246fa9e4066Sahrens uint64_t obj = db->db.db_object; 247fa9e4066Sahrens int level = db->db_level; 248fa9e4066Sahrens uint64_t blkid = db->db_blkid; 249dcbf3bd6SGeorge Wilson uint64_t hv = dbuf_hash(os, obj, level, blkid); 250fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 251fa9e4066Sahrens dmu_buf_impl_t *dbf; 252fa9e4066Sahrens 253fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 254fa9e4066Sahrens for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 255fa9e4066Sahrens if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 256fa9e4066Sahrens mutex_enter(&dbf->db_mtx); 257ea8dc4b6Seschrock if (dbf->db_state != DB_EVICTING) { 258fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 259fa9e4066Sahrens return (dbf); 260fa9e4066Sahrens } 261fa9e4066Sahrens mutex_exit(&dbf->db_mtx); 262fa9e4066Sahrens } 263fa9e4066Sahrens } 264fa9e4066Sahrens 265fa9e4066Sahrens mutex_enter(&db->db_mtx); 266fa9e4066Sahrens db->db_hash_next = h->hash_table[idx]; 267fa9e4066Sahrens h->hash_table[idx] = db; 268fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 2691a5e258fSJosef 'Jeff' Sipek atomic_inc_64(&dbuf_hash_count); 270fa9e4066Sahrens 271fa9e4066Sahrens return (NULL); 272fa9e4066Sahrens } 273fa9e4066Sahrens 274fa9e4066Sahrens /* 275bbfa8ea8SMatthew Ahrens * Remove an entry from the hash table. It must be in the EVICTING state. 276fa9e4066Sahrens */ 277fa9e4066Sahrens static void 278fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db) 279fa9e4066Sahrens { 280fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 281dcbf3bd6SGeorge Wilson uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object, 282fa9e4066Sahrens db->db_level, db->db_blkid); 283fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 284fa9e4066Sahrens dmu_buf_impl_t *dbf, **dbp; 285fa9e4066Sahrens 286fa9e4066Sahrens /* 287bbfa8ea8SMatthew Ahrens * We musn't hold db_mtx to maintain lock ordering: 288fa9e4066Sahrens * DBUF_HASH_MUTEX > db_mtx. 289fa9e4066Sahrens */ 290fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 291ea8dc4b6Seschrock ASSERT(db->db_state == DB_EVICTING); 292fa9e4066Sahrens ASSERT(!MUTEX_HELD(&db->db_mtx)); 293fa9e4066Sahrens 294fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 295fa9e4066Sahrens dbp = &h->hash_table[idx]; 296fa9e4066Sahrens while ((dbf = *dbp) != db) { 297fa9e4066Sahrens dbp = &dbf->db_hash_next; 298fa9e4066Sahrens ASSERT(dbf != NULL); 299fa9e4066Sahrens } 300fa9e4066Sahrens *dbp = db->db_hash_next; 301fa9e4066Sahrens db->db_hash_next = NULL; 302fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 3031a5e258fSJosef 'Jeff' Sipek atomic_dec_64(&dbuf_hash_count); 304fa9e4066Sahrens } 305fa9e4066Sahrens 306bc9014e6SJustin Gibbs typedef enum { 307bc9014e6SJustin Gibbs DBVU_EVICTING, 308bc9014e6SJustin Gibbs DBVU_NOT_EVICTING 309bc9014e6SJustin Gibbs } dbvu_verify_type_t; 310bc9014e6SJustin Gibbs 311bc9014e6SJustin Gibbs static void 312bc9014e6SJustin Gibbs dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 313bc9014e6SJustin Gibbs { 314bc9014e6SJustin Gibbs #ifdef ZFS_DEBUG 315bc9014e6SJustin Gibbs int64_t holds; 316bc9014e6SJustin Gibbs 317bc9014e6SJustin Gibbs if (db->db_user == NULL) 318bc9014e6SJustin Gibbs return; 319bc9014e6SJustin Gibbs 320bc9014e6SJustin Gibbs /* Only data blocks support the attachment of user data. */ 321bc9014e6SJustin Gibbs ASSERT(db->db_level == 0); 322bc9014e6SJustin Gibbs 323bc9014e6SJustin Gibbs /* Clients must resolve a dbuf before attaching user data. */ 324bc9014e6SJustin Gibbs ASSERT(db->db.db_data != NULL); 325bc9014e6SJustin Gibbs ASSERT3U(db->db_state, ==, DB_CACHED); 326bc9014e6SJustin Gibbs 327bc9014e6SJustin Gibbs holds = refcount_count(&db->db_holds); 328bc9014e6SJustin Gibbs if (verify_type == DBVU_EVICTING) { 329bc9014e6SJustin Gibbs /* 330bc9014e6SJustin Gibbs * Immediate eviction occurs when holds == dirtycnt. 331bc9014e6SJustin Gibbs * For normal eviction buffers, holds is zero on 332bc9014e6SJustin Gibbs * eviction, except when dbuf_fix_old_data() calls 333bc9014e6SJustin Gibbs * dbuf_clear_data(). However, the hold count can grow 334bc9014e6SJustin Gibbs * during eviction even though db_mtx is held (see 335bc9014e6SJustin Gibbs * dmu_bonus_hold() for an example), so we can only 336bc9014e6SJustin Gibbs * test the generic invariant that holds >= dirtycnt. 337bc9014e6SJustin Gibbs */ 338bc9014e6SJustin Gibbs ASSERT3U(holds, >=, db->db_dirtycnt); 339bc9014e6SJustin Gibbs } else { 340d2058105SJustin T. Gibbs if (db->db_user_immediate_evict == TRUE) 341bc9014e6SJustin Gibbs ASSERT3U(holds, >=, db->db_dirtycnt); 342bc9014e6SJustin Gibbs else 343bc9014e6SJustin Gibbs ASSERT3U(holds, >, 0); 344bc9014e6SJustin Gibbs } 345bc9014e6SJustin Gibbs #endif 346bc9014e6SJustin Gibbs } 347bc9014e6SJustin Gibbs 348fa9e4066Sahrens static void 349fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db) 350fa9e4066Sahrens { 351bc9014e6SJustin Gibbs dmu_buf_user_t *dbu = db->db_user; 352bc9014e6SJustin Gibbs 353fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 354fa9e4066Sahrens 355bc9014e6SJustin Gibbs if (dbu == NULL) 356fa9e4066Sahrens return; 357fa9e4066Sahrens 358bc9014e6SJustin Gibbs dbuf_verify_user(db, DBVU_EVICTING); 359bc9014e6SJustin Gibbs db->db_user = NULL; 360bc9014e6SJustin Gibbs 361bc9014e6SJustin Gibbs #ifdef ZFS_DEBUG 362bc9014e6SJustin Gibbs if (dbu->dbu_clear_on_evict_dbufp != NULL) 363bc9014e6SJustin Gibbs *dbu->dbu_clear_on_evict_dbufp = NULL; 364bc9014e6SJustin Gibbs #endif 365bc9014e6SJustin Gibbs 366bc9014e6SJustin Gibbs /* 36740510e8eSJosef 'Jeff' Sipek * There are two eviction callbacks - one that we call synchronously 36840510e8eSJosef 'Jeff' Sipek * and one that we invoke via a taskq. The async one is useful for 36940510e8eSJosef 'Jeff' Sipek * avoiding lock order reversals and limiting stack depth. 37040510e8eSJosef 'Jeff' Sipek * 37140510e8eSJosef 'Jeff' Sipek * Note that if we have a sync callback but no async callback, 37240510e8eSJosef 'Jeff' Sipek * it's likely that the sync callback will free the structure 37340510e8eSJosef 'Jeff' Sipek * containing the dbu. In that case we need to take care to not 37440510e8eSJosef 'Jeff' Sipek * dereference dbu after calling the sync evict func. 375bc9014e6SJustin Gibbs */ 37640510e8eSJosef 'Jeff' Sipek boolean_t has_async = (dbu->dbu_evict_func_async != NULL); 37740510e8eSJosef 'Jeff' Sipek 37840510e8eSJosef 'Jeff' Sipek if (dbu->dbu_evict_func_sync != NULL) 37940510e8eSJosef 'Jeff' Sipek dbu->dbu_evict_func_sync(dbu); 38040510e8eSJosef 'Jeff' Sipek 38140510e8eSJosef 'Jeff' Sipek if (has_async) { 38240510e8eSJosef 'Jeff' Sipek taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, 38340510e8eSJosef 'Jeff' Sipek dbu, 0, &dbu->dbu_tqent); 38440510e8eSJosef 'Jeff' Sipek } 385fa9e4066Sahrens } 386fa9e4066Sahrens 387744947dcSTom Erickson boolean_t 388744947dcSTom Erickson dbuf_is_metadata(dmu_buf_impl_t *db) 389744947dcSTom Erickson { 390744947dcSTom Erickson if (db->db_level > 0) { 391744947dcSTom Erickson return (B_TRUE); 392744947dcSTom Erickson } else { 393744947dcSTom Erickson boolean_t is_metadata; 394744947dcSTom Erickson 395744947dcSTom Erickson DB_DNODE_ENTER(db); 396ad135b5dSChristopher Siden is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 397744947dcSTom Erickson DB_DNODE_EXIT(db); 398744947dcSTom Erickson 399744947dcSTom Erickson return (is_metadata); 400744947dcSTom Erickson } 401744947dcSTom Erickson } 402744947dcSTom Erickson 403dcbf3bd6SGeorge Wilson /* 404dcbf3bd6SGeorge Wilson * This function *must* return indices evenly distributed between all 405dcbf3bd6SGeorge Wilson * sublists of the multilist. This is needed due to how the dbuf eviction 406dcbf3bd6SGeorge Wilson * code is laid out; dbuf_evict_thread() assumes dbufs are evenly 407dcbf3bd6SGeorge Wilson * distributed between all sublists and uses this assumption when 408dcbf3bd6SGeorge Wilson * deciding which sublist to evict from and how much to evict from it. 409dcbf3bd6SGeorge Wilson */ 410dcbf3bd6SGeorge Wilson unsigned int 411dcbf3bd6SGeorge Wilson dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) 412ea8dc4b6Seschrock { 413dcbf3bd6SGeorge Wilson dmu_buf_impl_t *db = obj; 414ea8dc4b6Seschrock 415dcbf3bd6SGeorge Wilson /* 416dcbf3bd6SGeorge Wilson * The assumption here, is the hash value for a given 417dcbf3bd6SGeorge Wilson * dmu_buf_impl_t will remain constant throughout it's lifetime 418dcbf3bd6SGeorge Wilson * (i.e. it's objset, object, level and blkid fields don't change). 419dcbf3bd6SGeorge Wilson * Thus, we don't need to store the dbuf's sublist index 420dcbf3bd6SGeorge Wilson * on insertion, as this index can be recalculated on removal. 421dcbf3bd6SGeorge Wilson * 422dcbf3bd6SGeorge Wilson * Also, the low order bits of the hash value are thought to be 423dcbf3bd6SGeorge Wilson * distributed evenly. Otherwise, in the case that the multilist 424dcbf3bd6SGeorge Wilson * has a power of two number of sublists, each sublists' usage 425dcbf3bd6SGeorge Wilson * would not be evenly distributed. 426dcbf3bd6SGeorge Wilson */ 427dcbf3bd6SGeorge Wilson return (dbuf_hash(db->db_objset, db->db.db_object, 428dcbf3bd6SGeorge Wilson db->db_level, db->db_blkid) % 429dcbf3bd6SGeorge Wilson multilist_get_num_sublists(ml)); 430dcbf3bd6SGeorge Wilson } 431dcbf3bd6SGeorge Wilson 432dcbf3bd6SGeorge Wilson static inline boolean_t 433dcbf3bd6SGeorge Wilson dbuf_cache_above_hiwater(void) 434dcbf3bd6SGeorge Wilson { 435dcbf3bd6SGeorge Wilson uint64_t dbuf_cache_hiwater_bytes = 436dcbf3bd6SGeorge Wilson (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100; 437dcbf3bd6SGeorge Wilson 438dcbf3bd6SGeorge Wilson return (refcount_count(&dbuf_cache_size) > 439dcbf3bd6SGeorge Wilson dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes); 440dcbf3bd6SGeorge Wilson } 441dcbf3bd6SGeorge Wilson 442dcbf3bd6SGeorge Wilson static inline boolean_t 443dcbf3bd6SGeorge Wilson dbuf_cache_above_lowater(void) 444dcbf3bd6SGeorge Wilson { 445dcbf3bd6SGeorge Wilson uint64_t dbuf_cache_lowater_bytes = 446dcbf3bd6SGeorge Wilson (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100; 447dcbf3bd6SGeorge Wilson 448dcbf3bd6SGeorge Wilson return (refcount_count(&dbuf_cache_size) > 449dcbf3bd6SGeorge Wilson dbuf_cache_max_bytes - dbuf_cache_lowater_bytes); 450dcbf3bd6SGeorge Wilson } 451dcbf3bd6SGeorge Wilson 452dcbf3bd6SGeorge Wilson /* 453dcbf3bd6SGeorge Wilson * Evict the oldest eligible dbuf from the dbuf cache. 454dcbf3bd6SGeorge Wilson */ 455dcbf3bd6SGeorge Wilson static void 456dcbf3bd6SGeorge Wilson dbuf_evict_one(void) 457dcbf3bd6SGeorge Wilson { 45894c2d0ebSMatthew Ahrens int idx = multilist_get_random_index(dbuf_cache); 45994c2d0ebSMatthew Ahrens multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx); 460dcbf3bd6SGeorge Wilson 461dcbf3bd6SGeorge Wilson ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); 462dcbf3bd6SGeorge Wilson 463dcbf3bd6SGeorge Wilson /* 464dcbf3bd6SGeorge Wilson * Set the thread's tsd to indicate that it's processing evictions. 465dcbf3bd6SGeorge Wilson * Once a thread stops evicting from the dbuf cache it will 466dcbf3bd6SGeorge Wilson * reset its tsd to NULL. 467dcbf3bd6SGeorge Wilson */ 468dcbf3bd6SGeorge Wilson ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL); 469dcbf3bd6SGeorge Wilson (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE); 470dcbf3bd6SGeorge Wilson 471dcbf3bd6SGeorge Wilson dmu_buf_impl_t *db = multilist_sublist_tail(mls); 472dcbf3bd6SGeorge Wilson while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { 473dcbf3bd6SGeorge Wilson db = multilist_sublist_prev(mls, db); 474dcbf3bd6SGeorge Wilson } 475dcbf3bd6SGeorge Wilson 476dcbf3bd6SGeorge Wilson DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, 477dcbf3bd6SGeorge Wilson multilist_sublist_t *, mls); 478dcbf3bd6SGeorge Wilson 479dcbf3bd6SGeorge Wilson if (db != NULL) { 480dcbf3bd6SGeorge Wilson multilist_sublist_remove(mls, db); 481dcbf3bd6SGeorge Wilson multilist_sublist_unlock(mls); 482dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dbuf_cache_size, 483dcbf3bd6SGeorge Wilson db->db.db_size, db); 484ea8dc4b6Seschrock dbuf_destroy(db); 485dcbf3bd6SGeorge Wilson } else { 486dcbf3bd6SGeorge Wilson multilist_sublist_unlock(mls); 487dcbf3bd6SGeorge Wilson } 488dcbf3bd6SGeorge Wilson (void) tsd_set(zfs_dbuf_evict_key, NULL); 489dcbf3bd6SGeorge Wilson } 490dcbf3bd6SGeorge Wilson 491dcbf3bd6SGeorge Wilson /* 492dcbf3bd6SGeorge Wilson * The dbuf evict thread is responsible for aging out dbufs from the 493dcbf3bd6SGeorge Wilson * cache. Once the cache has reached it's maximum size, dbufs are removed 494dcbf3bd6SGeorge Wilson * and destroyed. The eviction thread will continue running until the size 495dcbf3bd6SGeorge Wilson * of the dbuf cache is at or below the maximum size. Once the dbuf is aged 496dcbf3bd6SGeorge Wilson * out of the cache it is destroyed and becomes eligible for arc eviction. 497dcbf3bd6SGeorge Wilson */ 498dcbf3bd6SGeorge Wilson static void 499dcbf3bd6SGeorge Wilson dbuf_evict_thread(void) 500dcbf3bd6SGeorge Wilson { 501dcbf3bd6SGeorge Wilson callb_cpr_t cpr; 502dcbf3bd6SGeorge Wilson 503dcbf3bd6SGeorge Wilson CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); 504dcbf3bd6SGeorge Wilson 505dcbf3bd6SGeorge Wilson mutex_enter(&dbuf_evict_lock); 506dcbf3bd6SGeorge Wilson while (!dbuf_evict_thread_exit) { 507dcbf3bd6SGeorge Wilson while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 508dcbf3bd6SGeorge Wilson CALLB_CPR_SAFE_BEGIN(&cpr); 509dcbf3bd6SGeorge Wilson (void) cv_timedwait_hires(&dbuf_evict_cv, 510dcbf3bd6SGeorge Wilson &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 511dcbf3bd6SGeorge Wilson CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); 512dcbf3bd6SGeorge Wilson } 513dcbf3bd6SGeorge Wilson mutex_exit(&dbuf_evict_lock); 514dcbf3bd6SGeorge Wilson 515dcbf3bd6SGeorge Wilson /* 516dcbf3bd6SGeorge Wilson * Keep evicting as long as we're above the low water mark 517dcbf3bd6SGeorge Wilson * for the cache. We do this without holding the locks to 518dcbf3bd6SGeorge Wilson * minimize lock contention. 519dcbf3bd6SGeorge Wilson */ 520dcbf3bd6SGeorge Wilson while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 521dcbf3bd6SGeorge Wilson dbuf_evict_one(); 522dcbf3bd6SGeorge Wilson } 523dcbf3bd6SGeorge Wilson 524dcbf3bd6SGeorge Wilson mutex_enter(&dbuf_evict_lock); 525dcbf3bd6SGeorge Wilson } 526dcbf3bd6SGeorge Wilson 527dcbf3bd6SGeorge Wilson dbuf_evict_thread_exit = B_FALSE; 528dcbf3bd6SGeorge Wilson cv_broadcast(&dbuf_evict_cv); 529dcbf3bd6SGeorge Wilson CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ 530dcbf3bd6SGeorge Wilson thread_exit(); 531dcbf3bd6SGeorge Wilson } 532dcbf3bd6SGeorge Wilson 533dcbf3bd6SGeorge Wilson /* 534dcbf3bd6SGeorge Wilson * Wake up the dbuf eviction thread if the dbuf cache is at its max size. 535dcbf3bd6SGeorge Wilson * If the dbuf cache is at its high water mark, then evict a dbuf from the 536dcbf3bd6SGeorge Wilson * dbuf cache using the callers context. 537dcbf3bd6SGeorge Wilson */ 538dcbf3bd6SGeorge Wilson static void 539dcbf3bd6SGeorge Wilson dbuf_evict_notify(void) 540dcbf3bd6SGeorge Wilson { 541dcbf3bd6SGeorge Wilson 542dcbf3bd6SGeorge Wilson /* 543dcbf3bd6SGeorge Wilson * We use thread specific data to track when a thread has 544dcbf3bd6SGeorge Wilson * started processing evictions. This allows us to avoid deeply 545dcbf3bd6SGeorge Wilson * nested stacks that would have a call flow similar to this: 546dcbf3bd6SGeorge Wilson * 547dcbf3bd6SGeorge Wilson * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() 548dcbf3bd6SGeorge Wilson * ^ | 549dcbf3bd6SGeorge Wilson * | | 550dcbf3bd6SGeorge Wilson * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ 551dcbf3bd6SGeorge Wilson * 552dcbf3bd6SGeorge Wilson * The dbuf_eviction_thread will always have its tsd set until 553dcbf3bd6SGeorge Wilson * that thread exits. All other threads will only set their tsd 554dcbf3bd6SGeorge Wilson * if they are participating in the eviction process. This only 555dcbf3bd6SGeorge Wilson * happens if the eviction thread is unable to process evictions 556dcbf3bd6SGeorge Wilson * fast enough. To keep the dbuf cache size in check, other threads 557dcbf3bd6SGeorge Wilson * can evict from the dbuf cache directly. Those threads will set 558dcbf3bd6SGeorge Wilson * their tsd values so that we ensure that they only evict one dbuf 559dcbf3bd6SGeorge Wilson * from the dbuf cache. 560dcbf3bd6SGeorge Wilson */ 561dcbf3bd6SGeorge Wilson if (tsd_get(zfs_dbuf_evict_key) != NULL) 562dcbf3bd6SGeorge Wilson return; 563dcbf3bd6SGeorge Wilson 564dcbf3bd6SGeorge Wilson if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) { 565dcbf3bd6SGeorge Wilson boolean_t evict_now = B_FALSE; 566dcbf3bd6SGeorge Wilson 567dcbf3bd6SGeorge Wilson mutex_enter(&dbuf_evict_lock); 568dcbf3bd6SGeorge Wilson if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) { 569dcbf3bd6SGeorge Wilson evict_now = dbuf_cache_above_hiwater(); 570dcbf3bd6SGeorge Wilson cv_signal(&dbuf_evict_cv); 571dcbf3bd6SGeorge Wilson } 572dcbf3bd6SGeorge Wilson mutex_exit(&dbuf_evict_lock); 573dcbf3bd6SGeorge Wilson 574dcbf3bd6SGeorge Wilson if (evict_now) { 575dcbf3bd6SGeorge Wilson dbuf_evict_one(); 576dcbf3bd6SGeorge Wilson } 577dcbf3bd6SGeorge Wilson } 578ea8dc4b6Seschrock } 579ea8dc4b6Seschrock 580ea8dc4b6Seschrock void 581fa9e4066Sahrens dbuf_init(void) 582fa9e4066Sahrens { 583ea8dc4b6Seschrock uint64_t hsize = 1ULL << 16; 584fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 585fa9e4066Sahrens int i; 586fa9e4066Sahrens 587fa9e4066Sahrens /* 588fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 589ea8dc4b6Seschrock * with an average 4K block size. The table will take up 590ea8dc4b6Seschrock * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 591fa9e4066Sahrens */ 592ea8dc4b6Seschrock while (hsize * 4096 < physmem * PAGESIZE) 593fa9e4066Sahrens hsize <<= 1; 594fa9e4066Sahrens 595ea8dc4b6Seschrock retry: 596fa9e4066Sahrens h->hash_table_mask = hsize - 1; 597ea8dc4b6Seschrock h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 598ea8dc4b6Seschrock if (h->hash_table == NULL) { 599ea8dc4b6Seschrock /* XXX - we should really return an error instead of assert */ 600ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 10)); 601ea8dc4b6Seschrock hsize >>= 1; 602ea8dc4b6Seschrock goto retry; 603ea8dc4b6Seschrock } 604fa9e4066Sahrens 605dcbf3bd6SGeorge Wilson dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", 606fa9e4066Sahrens sizeof (dmu_buf_impl_t), 607fa9e4066Sahrens 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 608fa9e4066Sahrens 609fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 610fa9e4066Sahrens mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 611bc9014e6SJustin Gibbs 612bc9014e6SJustin Gibbs /* 613dcbf3bd6SGeorge Wilson * Setup the parameters for the dbuf cache. We cap the size of the 614dcbf3bd6SGeorge Wilson * dbuf cache to 1/32nd (default) of the size of the ARC. 615dcbf3bd6SGeorge Wilson */ 616dcbf3bd6SGeorge Wilson dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes, 617dcbf3bd6SGeorge Wilson arc_max_bytes() >> dbuf_cache_max_shift); 618dcbf3bd6SGeorge Wilson 619dcbf3bd6SGeorge Wilson /* 620bc9014e6SJustin Gibbs * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 621bc9014e6SJustin Gibbs * configuration is not required. 622bc9014e6SJustin Gibbs */ 623bc9014e6SJustin Gibbs dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 624dcbf3bd6SGeorge Wilson 62594c2d0ebSMatthew Ahrens dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t), 626dcbf3bd6SGeorge Wilson offsetof(dmu_buf_impl_t, db_cache_link), 627dcbf3bd6SGeorge Wilson dbuf_cache_multilist_index_func); 628dcbf3bd6SGeorge Wilson refcount_create(&dbuf_cache_size); 629dcbf3bd6SGeorge Wilson 630dcbf3bd6SGeorge Wilson tsd_create(&zfs_dbuf_evict_key, NULL); 631dcbf3bd6SGeorge Wilson dbuf_evict_thread_exit = B_FALSE; 632dcbf3bd6SGeorge Wilson mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); 633dcbf3bd6SGeorge Wilson cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); 634dcbf3bd6SGeorge Wilson dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, 635dcbf3bd6SGeorge Wilson NULL, 0, &p0, TS_RUN, minclsyspri); 636fa9e4066Sahrens } 637fa9e4066Sahrens 638fa9e4066Sahrens void 639fa9e4066Sahrens dbuf_fini(void) 640fa9e4066Sahrens { 641fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 642fa9e4066Sahrens int i; 643fa9e4066Sahrens 644fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 645fa9e4066Sahrens mutex_destroy(&h->hash_mutexes[i]); 646fa9e4066Sahrens kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 647dcbf3bd6SGeorge Wilson kmem_cache_destroy(dbuf_kmem_cache); 648bc9014e6SJustin Gibbs taskq_destroy(dbu_evict_taskq); 649dcbf3bd6SGeorge Wilson 650dcbf3bd6SGeorge Wilson mutex_enter(&dbuf_evict_lock); 651dcbf3bd6SGeorge Wilson dbuf_evict_thread_exit = B_TRUE; 652dcbf3bd6SGeorge Wilson while (dbuf_evict_thread_exit) { 653dcbf3bd6SGeorge Wilson cv_signal(&dbuf_evict_cv); 654dcbf3bd6SGeorge Wilson cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); 655dcbf3bd6SGeorge Wilson } 656dcbf3bd6SGeorge Wilson mutex_exit(&dbuf_evict_lock); 657dcbf3bd6SGeorge Wilson tsd_destroy(&zfs_dbuf_evict_key); 658dcbf3bd6SGeorge Wilson 659dcbf3bd6SGeorge Wilson mutex_destroy(&dbuf_evict_lock); 660dcbf3bd6SGeorge Wilson cv_destroy(&dbuf_evict_cv); 661dcbf3bd6SGeorge Wilson 662dcbf3bd6SGeorge Wilson refcount_destroy(&dbuf_cache_size); 66394c2d0ebSMatthew Ahrens multilist_destroy(dbuf_cache); 664fa9e4066Sahrens } 665fa9e4066Sahrens 666fa9e4066Sahrens /* 667fa9e4066Sahrens * Other stuff. 668fa9e4066Sahrens */ 669fa9e4066Sahrens 6709c9dc39aSek110237 #ifdef ZFS_DEBUG 671fa9e4066Sahrens static void 672fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db) 673fa9e4066Sahrens { 674744947dcSTom Erickson dnode_t *dn; 675b24ab676SJeff Bonwick dbuf_dirty_record_t *dr; 676fa9e4066Sahrens 677fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 678fa9e4066Sahrens 679fa9e4066Sahrens if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 680fa9e4066Sahrens return; 681fa9e4066Sahrens 682fa9e4066Sahrens ASSERT(db->db_objset != NULL); 683744947dcSTom Erickson DB_DNODE_ENTER(db); 684744947dcSTom Erickson dn = DB_DNODE(db); 685fa9e4066Sahrens if (dn == NULL) { 686fa9e4066Sahrens ASSERT(db->db_parent == NULL); 687fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 688fa9e4066Sahrens } else { 689fa9e4066Sahrens ASSERT3U(db->db.db_object, ==, dn->dn_object); 690fa9e4066Sahrens ASSERT3P(db->db_objset, ==, dn->dn_objset); 691fa9e4066Sahrens ASSERT3U(db->db_level, <, dn->dn_nlevels); 692744947dcSTom Erickson ASSERT(db->db_blkid == DMU_BONUS_BLKID || 693744947dcSTom Erickson db->db_blkid == DMU_SPILL_BLKID || 6940f6d88adSAlex Reece !avl_is_empty(&dn->dn_dbufs)); 695fa9e4066Sahrens } 6960a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 697fa9e4066Sahrens ASSERT(dn != NULL); 6981934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 6990a586ceaSMark Shellenbaum ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 7000a586ceaSMark Shellenbaum } else if (db->db_blkid == DMU_SPILL_BLKID) { 7010a586ceaSMark Shellenbaum ASSERT(dn != NULL); 7020a586ceaSMark Shellenbaum ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 703fb09f5aaSMadhav Suresh ASSERT0(db->db.db_offset); 704fa9e4066Sahrens } else { 705fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 706fa9e4066Sahrens } 707fa9e4066Sahrens 708b24ab676SJeff Bonwick for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 709b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 710b24ab676SJeff Bonwick 711b24ab676SJeff Bonwick for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 712b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 713b24ab676SJeff Bonwick 71488b7b0f2SMatthew Ahrens /* 71588b7b0f2SMatthew Ahrens * We can't assert that db_size matches dn_datablksz because it 71688b7b0f2SMatthew Ahrens * can be momentarily different when another thread is doing 71788b7b0f2SMatthew Ahrens * dnode_set_blksz(). 71888b7b0f2SMatthew Ahrens */ 71988b7b0f2SMatthew Ahrens if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 720b24ab676SJeff Bonwick dr = db->db_data_pending; 721fa9e4066Sahrens /* 72288b7b0f2SMatthew Ahrens * It should only be modified in syncing context, so 72388b7b0f2SMatthew Ahrens * make sure we only have one copy of the data. 724fa9e4066Sahrens */ 725c717a561Smaybee ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 726fa9e4066Sahrens } 727fa9e4066Sahrens 728fa9e4066Sahrens /* verify db->db_blkptr */ 729fa9e4066Sahrens if (db->db_blkptr) { 730fa9e4066Sahrens if (db->db_parent == dn->dn_dbuf) { 731fa9e4066Sahrens /* db is pointed to by the dnode */ 732fa9e4066Sahrens /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 73314843421SMatthew Ahrens if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 734fa9e4066Sahrens ASSERT(db->db_parent == NULL); 735fa9e4066Sahrens else 736fa9e4066Sahrens ASSERT(db->db_parent != NULL); 7370a586ceaSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID) 738fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 739fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 740fa9e4066Sahrens } else { 741fa9e4066Sahrens /* db is pointed to by an indirect block */ 742fa9e4066Sahrens int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 743fa9e4066Sahrens ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 744fa9e4066Sahrens ASSERT3U(db->db_parent->db.db_object, ==, 745fa9e4066Sahrens db->db.db_object); 746fa9e4066Sahrens /* 747fa9e4066Sahrens * dnode_grow_indblksz() can make this fail if we don't 748fa9e4066Sahrens * have the struct_rwlock. XXX indblksz no longer 749fa9e4066Sahrens * grows. safe to do this now? 750fa9e4066Sahrens */ 751744947dcSTom Erickson if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 752fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 753fa9e4066Sahrens ((blkptr_t *)db->db_parent->db.db_data + 754fa9e4066Sahrens db->db_blkid % epb)); 755fa9e4066Sahrens } 756fa9e4066Sahrens } 757fa9e4066Sahrens } 758fa9e4066Sahrens if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 7593f9d6ad7SLin Ling (db->db_buf == NULL || db->db_buf->b_data) && 7600a586ceaSMark Shellenbaum db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 761fa9e4066Sahrens db->db_state != DB_FILL && !dn->dn_free_txg) { 762fa9e4066Sahrens /* 763fa9e4066Sahrens * If the blkptr isn't set but they have nonzero data, 764fa9e4066Sahrens * it had better be dirty, otherwise we'll lose that 765fa9e4066Sahrens * data when we evict this buffer. 7668df0bcf0SPaul Dagnelie * 7678df0bcf0SPaul Dagnelie * There is an exception to this rule for indirect blocks; in 7688df0bcf0SPaul Dagnelie * this case, if the indirect block is a hole, we fill in a few 7698df0bcf0SPaul Dagnelie * fields on each of the child blocks (importantly, birth time) 7708df0bcf0SPaul Dagnelie * to prevent hole birth times from being lost when you 7718df0bcf0SPaul Dagnelie * partially fill in a hole. 772fa9e4066Sahrens */ 773fa9e4066Sahrens if (db->db_dirtycnt == 0) { 7748df0bcf0SPaul Dagnelie if (db->db_level == 0) { 775fa9e4066Sahrens uint64_t *buf = db->db.db_data; 776fa9e4066Sahrens int i; 777fa9e4066Sahrens 778fa9e4066Sahrens for (i = 0; i < db->db.db_size >> 3; i++) { 779fa9e4066Sahrens ASSERT(buf[i] == 0); 780fa9e4066Sahrens } 7818df0bcf0SPaul Dagnelie } else { 7828df0bcf0SPaul Dagnelie blkptr_t *bps = db->db.db_data; 7838df0bcf0SPaul Dagnelie ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, 7848df0bcf0SPaul Dagnelie db->db.db_size); 7858df0bcf0SPaul Dagnelie /* 7868df0bcf0SPaul Dagnelie * We want to verify that all the blkptrs in the 7878df0bcf0SPaul Dagnelie * indirect block are holes, but we may have 7888df0bcf0SPaul Dagnelie * automatically set up a few fields for them. 7898df0bcf0SPaul Dagnelie * We iterate through each blkptr and verify 7908df0bcf0SPaul Dagnelie * they only have those fields set. 7918df0bcf0SPaul Dagnelie */ 7928df0bcf0SPaul Dagnelie for (int i = 0; 7938df0bcf0SPaul Dagnelie i < db->db.db_size / sizeof (blkptr_t); 7948df0bcf0SPaul Dagnelie i++) { 7958df0bcf0SPaul Dagnelie blkptr_t *bp = &bps[i]; 7968df0bcf0SPaul Dagnelie ASSERT(ZIO_CHECKSUM_IS_ZERO( 7978df0bcf0SPaul Dagnelie &bp->blk_cksum)); 7988df0bcf0SPaul Dagnelie ASSERT( 7998df0bcf0SPaul Dagnelie DVA_IS_EMPTY(&bp->blk_dva[0]) && 8008df0bcf0SPaul Dagnelie DVA_IS_EMPTY(&bp->blk_dva[1]) && 8018df0bcf0SPaul Dagnelie DVA_IS_EMPTY(&bp->blk_dva[2])); 8028df0bcf0SPaul Dagnelie ASSERT0(bp->blk_fill); 8038df0bcf0SPaul Dagnelie ASSERT0(bp->blk_pad[0]); 8048df0bcf0SPaul Dagnelie ASSERT0(bp->blk_pad[1]); 8058df0bcf0SPaul Dagnelie ASSERT(!BP_IS_EMBEDDED(bp)); 8068df0bcf0SPaul Dagnelie ASSERT(BP_IS_HOLE(bp)); 8078df0bcf0SPaul Dagnelie ASSERT0(bp->blk_phys_birth); 8088df0bcf0SPaul Dagnelie } 8098df0bcf0SPaul Dagnelie } 810fa9e4066Sahrens } 811fa9e4066Sahrens } 812744947dcSTom Erickson DB_DNODE_EXIT(db); 813fa9e4066Sahrens } 8149c9dc39aSek110237 #endif 815fa9e4066Sahrens 816fa9e4066Sahrens static void 817bc9014e6SJustin Gibbs dbuf_clear_data(dmu_buf_impl_t *db) 818fa9e4066Sahrens { 819fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 820ea8dc4b6Seschrock dbuf_evict_user(db); 821dcbf3bd6SGeorge Wilson ASSERT3P(db->db_buf, ==, NULL); 822ea8dc4b6Seschrock db->db.db_data = NULL; 82382c9918fSTim Haley if (db->db_state != DB_NOFILL) 824ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 825ea8dc4b6Seschrock } 826bc9014e6SJustin Gibbs 827bc9014e6SJustin Gibbs static void 828bc9014e6SJustin Gibbs dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 829bc9014e6SJustin Gibbs { 830bc9014e6SJustin Gibbs ASSERT(MUTEX_HELD(&db->db_mtx)); 831bc9014e6SJustin Gibbs ASSERT(buf != NULL); 832bc9014e6SJustin Gibbs 833bc9014e6SJustin Gibbs db->db_buf = buf; 834bc9014e6SJustin Gibbs ASSERT(buf->b_data != NULL); 835bc9014e6SJustin Gibbs db->db.db_data = buf->b_data; 836fa9e4066Sahrens } 837fa9e4066Sahrens 838c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* 839c242f9a0Schunli zhang - Sun Microsystems - Irvine United States * Loan out an arc_buf for read. Return the loaned arc_buf. 840c242f9a0Schunli zhang - Sun Microsystems - Irvine United States */ 841c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t * 842c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dbuf_loan_arcbuf(dmu_buf_impl_t *db) 843c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 844c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t *abuf; 845c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 846dcbf3bd6SGeorge Wilson ASSERT(db->db_blkid != DMU_BONUS_BLKID); 847c242f9a0Schunli zhang - Sun Microsystems - Irvine United States mutex_enter(&db->db_mtx); 848c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 849c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int blksz = db->db.db_size; 85043466aaeSMax Grossman spa_t *spa = db->db_objset->os_spa; 851744947dcSTom Erickson 852c242f9a0Schunli zhang - Sun Microsystems - Irvine United States mutex_exit(&db->db_mtx); 8535602294fSDan Kimmel abuf = arc_loan_buf(spa, B_FALSE, blksz); 854c242f9a0Schunli zhang - Sun Microsystems - Irvine United States bcopy(db->db.db_data, abuf->b_data, blksz); 855c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } else { 856c242f9a0Schunli zhang - Sun Microsystems - Irvine United States abuf = db->db_buf; 857c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_loan_inuse_buf(abuf, db); 858dcbf3bd6SGeorge Wilson db->db_buf = NULL; 859bc9014e6SJustin Gibbs dbuf_clear_data(db); 860c242f9a0Schunli zhang - Sun Microsystems - Irvine United States mutex_exit(&db->db_mtx); 861c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 862c242f9a0Schunli zhang - Sun Microsystems - Irvine United States return (abuf); 863c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 864c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 865a2cdcdd2SPaul Dagnelie /* 866a2cdcdd2SPaul Dagnelie * Calculate which level n block references the data at the level 0 offset 867a2cdcdd2SPaul Dagnelie * provided. 868a2cdcdd2SPaul Dagnelie */ 869fa9e4066Sahrens uint64_t 870a2cdcdd2SPaul Dagnelie dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 871fa9e4066Sahrens { 872a2cdcdd2SPaul Dagnelie if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 873a2cdcdd2SPaul Dagnelie /* 874a2cdcdd2SPaul Dagnelie * The level n blkid is equal to the level 0 blkid divided by 875a2cdcdd2SPaul Dagnelie * the number of level 0s in a level n block. 876a2cdcdd2SPaul Dagnelie * 877a2cdcdd2SPaul Dagnelie * The level 0 blkid is offset >> datablkshift = 878a2cdcdd2SPaul Dagnelie * offset / 2^datablkshift. 879a2cdcdd2SPaul Dagnelie * 880a2cdcdd2SPaul Dagnelie * The number of level 0s in a level n is the number of block 881a2cdcdd2SPaul Dagnelie * pointers in an indirect block, raised to the power of level. 882a2cdcdd2SPaul Dagnelie * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 883a2cdcdd2SPaul Dagnelie * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 884a2cdcdd2SPaul Dagnelie * 885a2cdcdd2SPaul Dagnelie * Thus, the level n blkid is: offset / 886a2cdcdd2SPaul Dagnelie * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 887a2cdcdd2SPaul Dagnelie * = offset / 2^(datablkshift + level * 888a2cdcdd2SPaul Dagnelie * (indblkshift - SPA_BLKPTRSHIFT)) 889a2cdcdd2SPaul Dagnelie * = offset >> (datablkshift + level * 890a2cdcdd2SPaul Dagnelie * (indblkshift - SPA_BLKPTRSHIFT)) 891a2cdcdd2SPaul Dagnelie */ 892a2cdcdd2SPaul Dagnelie return (offset >> (dn->dn_datablkshift + level * 893a2cdcdd2SPaul Dagnelie (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 894fa9e4066Sahrens } else { 895fa9e4066Sahrens ASSERT3U(offset, <, dn->dn_datablksz); 896fa9e4066Sahrens return (0); 897fa9e4066Sahrens } 898fa9e4066Sahrens } 899fa9e4066Sahrens 900fa9e4066Sahrens static void 901fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 902fa9e4066Sahrens { 903fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 904fa9e4066Sahrens 905fa9e4066Sahrens mutex_enter(&db->db_mtx); 906fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_READ); 907fa9e4066Sahrens /* 908fa9e4066Sahrens * All reads are synchronous, so we must have a hold on the dbuf 909fa9e4066Sahrens */ 910fa9e4066Sahrens ASSERT(refcount_count(&db->db_holds) > 0); 911ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 912fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 913c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 914fa9e4066Sahrens /* we were freed in flight; disregard any error */ 915fa9e4066Sahrens arc_release(buf, db); 916fa9e4066Sahrens bzero(buf->b_data, db->db.db_size); 9176b4acc8bSahrens arc_buf_freeze(buf); 918c717a561Smaybee db->db_freed_in_flight = FALSE; 919fa9e4066Sahrens dbuf_set_data(db, buf); 920fa9e4066Sahrens db->db_state = DB_CACHED; 921fa9e4066Sahrens } else if (zio == NULL || zio->io_error == 0) { 922fa9e4066Sahrens dbuf_set_data(db, buf); 923fa9e4066Sahrens db->db_state = DB_CACHED; 924fa9e4066Sahrens } else { 9250a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 926fa9e4066Sahrens ASSERT3P(db->db_buf, ==, NULL); 927dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, db); 928ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 929fa9e4066Sahrens } 930fa9e4066Sahrens cv_broadcast(&db->db_changed); 9313f9d6ad7SLin Ling dbuf_rele_and_unlock(db, NULL); 932fa9e4066Sahrens } 933fa9e4066Sahrens 934ea8dc4b6Seschrock static void 935cf6106c8SMatthew Ahrens dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 936fa9e4066Sahrens { 937744947dcSTom Erickson dnode_t *dn; 9387802d7bfSMatthew Ahrens zbookmark_phys_t zb; 9397adb730bSGeorge Wilson arc_flags_t aflags = ARC_FLAG_NOWAIT; 940fa9e4066Sahrens 941744947dcSTom Erickson DB_DNODE_ENTER(db); 942744947dcSTom Erickson dn = DB_DNODE(db); 943fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 944fa9e4066Sahrens /* We need the struct_rwlock to prevent db_blkptr from changing. */ 945088f3894Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 946ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 947ea8dc4b6Seschrock ASSERT(db->db_state == DB_UNCACHED); 948ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 949fa9e4066Sahrens 9500a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 951cf04dda1SMark Maybee int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 9521934e92fSmaybee 9531934e92fSmaybee ASSERT3U(bonuslen, <=, db->db.db_size); 954ea8dc4b6Seschrock db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 9555a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 9561934e92fSmaybee if (bonuslen < DN_MAX_BONUSLEN) 957ea8dc4b6Seschrock bzero(db->db.db_data, DN_MAX_BONUSLEN); 958cf04dda1SMark Maybee if (bonuslen) 959cf04dda1SMark Maybee bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 960744947dcSTom Erickson DB_DNODE_EXIT(db); 961fa9e4066Sahrens db->db_state = DB_CACHED; 962fa9e4066Sahrens mutex_exit(&db->db_mtx); 963fa9e4066Sahrens return; 964fa9e4066Sahrens } 965fa9e4066Sahrens 9661c8564a7SMark Maybee /* 9671c8564a7SMark Maybee * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 9681c8564a7SMark Maybee * processes the delete record and clears the bp while we are waiting 9691c8564a7SMark Maybee * for the dn_mtx (resulting in a "no" from block_freed). 9701c8564a7SMark Maybee */ 971088f3894Sahrens if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 9721c8564a7SMark Maybee (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 9731c8564a7SMark Maybee BP_IS_HOLE(db->db_blkptr)))) { 974ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 975ad23a2dbSjohansen 9765602294fSDan Kimmel dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, 9775602294fSDan Kimmel db->db.db_size)); 978fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 9798df0bcf0SPaul Dagnelie 9808df0bcf0SPaul Dagnelie if (db->db_blkptr != NULL && db->db_level > 0 && 9818df0bcf0SPaul Dagnelie BP_IS_HOLE(db->db_blkptr) && 9828df0bcf0SPaul Dagnelie db->db_blkptr->blk_birth != 0) { 9838df0bcf0SPaul Dagnelie blkptr_t *bps = db->db.db_data; 9848df0bcf0SPaul Dagnelie for (int i = 0; i < ((1 << 9858df0bcf0SPaul Dagnelie DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); 9868df0bcf0SPaul Dagnelie i++) { 9878df0bcf0SPaul Dagnelie blkptr_t *bp = &bps[i]; 9888df0bcf0SPaul Dagnelie ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 9898df0bcf0SPaul Dagnelie 1 << dn->dn_indblkshift); 9908df0bcf0SPaul Dagnelie BP_SET_LSIZE(bp, 9918df0bcf0SPaul Dagnelie BP_GET_LEVEL(db->db_blkptr) == 1 ? 9928df0bcf0SPaul Dagnelie dn->dn_datablksz : 9938df0bcf0SPaul Dagnelie BP_GET_LSIZE(db->db_blkptr)); 9948df0bcf0SPaul Dagnelie BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); 9958df0bcf0SPaul Dagnelie BP_SET_LEVEL(bp, 9968df0bcf0SPaul Dagnelie BP_GET_LEVEL(db->db_blkptr) - 1); 9978df0bcf0SPaul Dagnelie BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); 9988df0bcf0SPaul Dagnelie } 9998df0bcf0SPaul Dagnelie } 10008df0bcf0SPaul Dagnelie DB_DNODE_EXIT(db); 1001fa9e4066Sahrens db->db_state = DB_CACHED; 1002fa9e4066Sahrens mutex_exit(&db->db_mtx); 1003fa9e4066Sahrens return; 1004fa9e4066Sahrens } 1005fa9e4066Sahrens 1006744947dcSTom Erickson DB_DNODE_EXIT(db); 1007744947dcSTom Erickson 1008fa9e4066Sahrens db->db_state = DB_READ; 1009fa9e4066Sahrens mutex_exit(&db->db_mtx); 1010fa9e4066Sahrens 10113baa08fcSek110237 if (DBUF_IS_L2CACHEABLE(db)) 10127adb730bSGeorge Wilson aflags |= ARC_FLAG_L2CACHE; 10133baa08fcSek110237 1014b24ab676SJeff Bonwick SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 1015b24ab676SJeff Bonwick db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 1016b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 1017ea8dc4b6Seschrock 1018ea8dc4b6Seschrock dbuf_add_ref(db, NULL); 1019088f3894Sahrens 102043466aaeSMax Grossman (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 1021fa9e4066Sahrens dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 1022cf6106c8SMatthew Ahrens (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 102313506d1eSmaybee &aflags, &zb); 1024fa9e4066Sahrens } 1025fa9e4066Sahrens 10265602294fSDan Kimmel /* 10275602294fSDan Kimmel * This is our just-in-time copy function. It makes a copy of buffers that 10285602294fSDan Kimmel * have been modified in a previous transaction group before we access them in 10295602294fSDan Kimmel * the current active group. 10305602294fSDan Kimmel * 10315602294fSDan Kimmel * This function is used in three places: when we are dirtying a buffer for the 10325602294fSDan Kimmel * first time in a txg, when we are freeing a range in a dnode that includes 10335602294fSDan Kimmel * this buffer, and when we are accessing a buffer which was received compressed 10345602294fSDan Kimmel * and later referenced in a WRITE_BYREF record. 10355602294fSDan Kimmel * 10365602294fSDan Kimmel * Note that when we are called from dbuf_free_range() we do not put a hold on 10375602294fSDan Kimmel * the buffer, we just traverse the active dbuf list for the dnode. 10385602294fSDan Kimmel */ 10395602294fSDan Kimmel static void 10405602294fSDan Kimmel dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 10415602294fSDan Kimmel { 10425602294fSDan Kimmel dbuf_dirty_record_t *dr = db->db_last_dirty; 10435602294fSDan Kimmel 10445602294fSDan Kimmel ASSERT(MUTEX_HELD(&db->db_mtx)); 10455602294fSDan Kimmel ASSERT(db->db.db_data != NULL); 10465602294fSDan Kimmel ASSERT(db->db_level == 0); 10475602294fSDan Kimmel ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 10485602294fSDan Kimmel 10495602294fSDan Kimmel if (dr == NULL || 10505602294fSDan Kimmel (dr->dt.dl.dr_data != 10515602294fSDan Kimmel ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 10525602294fSDan Kimmel return; 10535602294fSDan Kimmel 10545602294fSDan Kimmel /* 10555602294fSDan Kimmel * If the last dirty record for this dbuf has not yet synced 10565602294fSDan Kimmel * and its referencing the dbuf data, either: 10575602294fSDan Kimmel * reset the reference to point to a new copy, 10585602294fSDan Kimmel * or (if there a no active holders) 10595602294fSDan Kimmel * just null out the current db_data pointer. 10605602294fSDan Kimmel */ 10615602294fSDan Kimmel ASSERT(dr->dr_txg >= txg - 2); 10625602294fSDan Kimmel if (db->db_blkid == DMU_BONUS_BLKID) { 10635602294fSDan Kimmel /* Note that the data bufs here are zio_bufs */ 10645602294fSDan Kimmel dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 10655602294fSDan Kimmel arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 10665602294fSDan Kimmel bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 10675602294fSDan Kimmel } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 10685602294fSDan Kimmel int size = arc_buf_size(db->db_buf); 10695602294fSDan Kimmel arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 10705602294fSDan Kimmel spa_t *spa = db->db_objset->os_spa; 10715602294fSDan Kimmel enum zio_compress compress_type = 10725602294fSDan Kimmel arc_get_compression(db->db_buf); 10735602294fSDan Kimmel 10745602294fSDan Kimmel if (compress_type == ZIO_COMPRESS_OFF) { 10755602294fSDan Kimmel dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); 10765602294fSDan Kimmel } else { 10775602294fSDan Kimmel ASSERT3U(type, ==, ARC_BUFC_DATA); 10785602294fSDan Kimmel dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, 10795602294fSDan Kimmel size, arc_buf_lsize(db->db_buf), compress_type); 10805602294fSDan Kimmel } 10815602294fSDan Kimmel bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 10825602294fSDan Kimmel } else { 10835602294fSDan Kimmel db->db_buf = NULL; 10845602294fSDan Kimmel dbuf_clear_data(db); 10855602294fSDan Kimmel } 10865602294fSDan Kimmel } 10875602294fSDan Kimmel 1088ea8dc4b6Seschrock int 1089ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 1090fa9e4066Sahrens { 1091ea8dc4b6Seschrock int err = 0; 109243466aaeSMax Grossman boolean_t prefetch; 1093744947dcSTom Erickson dnode_t *dn; 1094fa9e4066Sahrens 1095fa9e4066Sahrens /* 1096fa9e4066Sahrens * We don't have to hold the mutex to check db_state because it 1097fa9e4066Sahrens * can't be freed while we have a hold on the buffer. 1098fa9e4066Sahrens */ 1099fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1100fa9e4066Sahrens 110182c9918fSTim Haley if (db->db_state == DB_NOFILL) 1102be6fd75aSMatthew Ahrens return (SET_ERROR(EIO)); 110382c9918fSTim Haley 1104744947dcSTom Erickson DB_DNODE_ENTER(db); 1105744947dcSTom Erickson dn = DB_DNODE(db); 1106fa9e4066Sahrens if ((flags & DB_RF_HAVESTRUCT) == 0) 1107744947dcSTom Erickson rw_enter(&dn->dn_struct_rwlock, RW_READER); 1108fa9e4066Sahrens 11090a586ceaSMark Shellenbaum prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1110744947dcSTom Erickson (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 11113baa08fcSek110237 DBUF_IS_CACHEABLE(db); 111213506d1eSmaybee 1113fa9e4066Sahrens mutex_enter(&db->db_mtx); 1114ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 11155602294fSDan Kimmel /* 11165602294fSDan Kimmel * If the arc buf is compressed, we need to decompress it to 11175602294fSDan Kimmel * read the data. This could happen during the "zfs receive" of 11185602294fSDan Kimmel * a stream which is compressed and deduplicated. 11195602294fSDan Kimmel */ 11205602294fSDan Kimmel if (db->db_buf != NULL && 11215602294fSDan Kimmel arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) { 11225602294fSDan Kimmel dbuf_fix_old_data(db, 11235602294fSDan Kimmel spa_syncing_txg(dmu_objset_spa(db->db_objset))); 11245602294fSDan Kimmel err = arc_decompress(db->db_buf); 11255602294fSDan Kimmel dbuf_set_data(db, db->db_buf); 11265602294fSDan Kimmel } 1127ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 112813506d1eSmaybee if (prefetch) 1129cb92f413SAlexander Motin dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1130ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 1131744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 1132744947dcSTom Erickson DB_DNODE_EXIT(db); 1133ea8dc4b6Seschrock } else if (db->db_state == DB_UNCACHED) { 1134744947dcSTom Erickson spa_t *spa = dn->dn_objset->os_spa; 1135def4fac5SMatthew Ahrens boolean_t need_wait = B_FALSE; 1136744947dcSTom Erickson 1137def4fac5SMatthew Ahrens if (zio == NULL && 1138def4fac5SMatthew Ahrens db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1139744947dcSTom Erickson zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 1140def4fac5SMatthew Ahrens need_wait = B_TRUE; 1141def4fac5SMatthew Ahrens } 1142cf6106c8SMatthew Ahrens dbuf_read_impl(db, zio, flags); 114313506d1eSmaybee 1144ea8dc4b6Seschrock /* dbuf_read_impl has dropped db_mtx for us */ 1145ea8dc4b6Seschrock 114613506d1eSmaybee if (prefetch) 1147cb92f413SAlexander Motin dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1148ea8dc4b6Seschrock 1149ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 1150744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 1151744947dcSTom Erickson DB_DNODE_EXIT(db); 1152ea8dc4b6Seschrock 1153def4fac5SMatthew Ahrens if (need_wait) 1154ea8dc4b6Seschrock err = zio_wait(zio); 1155ea8dc4b6Seschrock } else { 11563e30c24aSWill Andrews /* 11573e30c24aSWill Andrews * Another reader came in while the dbuf was in flight 11583e30c24aSWill Andrews * between UNCACHED and CACHED. Either a writer will finish 11593e30c24aSWill Andrews * writing the buffer (sending the dbuf to CACHED) or the 11603e30c24aSWill Andrews * first reader's request will reach the read_done callback 11613e30c24aSWill Andrews * and send the dbuf to CACHED. Otherwise, a failure 11623e30c24aSWill Andrews * occurred and the dbuf went to UNCACHED. 11633e30c24aSWill Andrews */ 116413506d1eSmaybee mutex_exit(&db->db_mtx); 116513506d1eSmaybee if (prefetch) 1166cb92f413SAlexander Motin dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1167ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 1168744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 1169744947dcSTom Erickson DB_DNODE_EXIT(db); 117013506d1eSmaybee 11713e30c24aSWill Andrews /* Skip the wait per the caller's request. */ 117213506d1eSmaybee mutex_enter(&db->db_mtx); 1173ea8dc4b6Seschrock if ((flags & DB_RF_NEVERWAIT) == 0) { 1174ea8dc4b6Seschrock while (db->db_state == DB_READ || 1175ea8dc4b6Seschrock db->db_state == DB_FILL) { 1176fa9e4066Sahrens ASSERT(db->db_state == DB_READ || 1177fa9e4066Sahrens (flags & DB_RF_HAVESTRUCT) == 0); 1178f6164ad6SAdam H. Leventhal DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 1179f6164ad6SAdam H. Leventhal db, zio_t *, zio); 1180fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 1181fa9e4066Sahrens } 1182ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED) 1183be6fd75aSMatthew Ahrens err = SET_ERROR(EIO); 1184ea8dc4b6Seschrock } 1185fa9e4066Sahrens mutex_exit(&db->db_mtx); 1186fa9e4066Sahrens } 1187fa9e4066Sahrens 1188ea8dc4b6Seschrock return (err); 1189fa9e4066Sahrens } 1190fa9e4066Sahrens 1191fa9e4066Sahrens static void 1192fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db) 1193fa9e4066Sahrens { 1194fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 11950a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1196fa9e4066Sahrens mutex_enter(&db->db_mtx); 1197fa9e4066Sahrens while (db->db_state == DB_READ || db->db_state == DB_FILL) 1198fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 1199fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 1200ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 120143466aaeSMax Grossman spa_t *spa = db->db_objset->os_spa; 1202ad23a2dbSjohansen 1203ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 1204fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 12055602294fSDan Kimmel dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); 1206fa9e4066Sahrens db->db_state = DB_FILL; 120782c9918fSTim Haley } else if (db->db_state == DB_NOFILL) { 1208bc9014e6SJustin Gibbs dbuf_clear_data(db); 1209fa9e4066Sahrens } else { 1210fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 1211fa9e4066Sahrens } 1212fa9e4066Sahrens mutex_exit(&db->db_mtx); 1213fa9e4066Sahrens } 1214fa9e4066Sahrens 1215fa9e4066Sahrens void 1216c717a561Smaybee dbuf_unoverride(dbuf_dirty_record_t *dr) 1217fa9e4066Sahrens { 1218c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 1219b24ab676SJeff Bonwick blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 1220c717a561Smaybee uint64_t txg = dr->dr_txg; 1221c5c6ffa0Smaybee 1222c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 122340713f2bSAlan Somers /* 122440713f2bSAlan Somers * This assert is valid because dmu_sync() expects to be called by 122540713f2bSAlan Somers * a zilog's get_data while holding a range lock. This call only 122640713f2bSAlan Somers * comes from dbuf_dirty() callers who must also hold a range lock. 122740713f2bSAlan Somers */ 1228c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 1229c717a561Smaybee ASSERT(db->db_level == 0); 1230c717a561Smaybee 12310a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID || 1232c717a561Smaybee dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 1233c717a561Smaybee return; 1234c717a561Smaybee 1235b24ab676SJeff Bonwick ASSERT(db->db_data_pending != dr); 1236b24ab676SJeff Bonwick 1237fa9e4066Sahrens /* free this block */ 123843466aaeSMax Grossman if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 123943466aaeSMax Grossman zio_free(db->db_objset->os_spa, txg, bp); 1240b24ab676SJeff Bonwick 1241c717a561Smaybee dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 124280901aeaSGeorge Wilson dr->dt.dl.dr_nopwrite = B_FALSE; 124380901aeaSGeorge Wilson 12446b4acc8bSahrens /* 12456b4acc8bSahrens * Release the already-written buffer, so we leave it in 12466b4acc8bSahrens * a consistent dirty state. Note that all callers are 12476b4acc8bSahrens * modifying the buffer, so they will immediately do 12486b4acc8bSahrens * another (redundant) arc_release(). Therefore, leave 12496b4acc8bSahrens * the buf thawed to save the effort of freezing & 12506b4acc8bSahrens * immediately re-thawing it. 12516b4acc8bSahrens */ 1252c717a561Smaybee arc_release(dr->dt.dl.dr_data, db); 1253fa9e4066Sahrens } 1254fa9e4066Sahrens 1255cdb0ab79Smaybee /* 1256cdb0ab79Smaybee * Evict (if its unreferenced) or clear (if its referenced) any level-0 1257cdb0ab79Smaybee * data blocks in the free range, so that any future readers will find 125843466aaeSMax Grossman * empty blocks. 1259cdb0ab79Smaybee */ 1260fa9e4066Sahrens void 12610f6d88adSAlex Reece dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 12620f6d88adSAlex Reece dmu_tx_t *tx) 1263fa9e4066Sahrens { 1264bc9014e6SJustin Gibbs dmu_buf_impl_t db_search; 1265bc9014e6SJustin Gibbs dmu_buf_impl_t *db, *db_next; 1266fa9e4066Sahrens uint64_t txg = tx->tx_txg; 12670f6d88adSAlex Reece avl_index_t where; 1268fa9e4066Sahrens 1269653af1b8SStephen Blinick if (end_blkid > dn->dn_maxblkid && 1270653af1b8SStephen Blinick !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) 12710f6d88adSAlex Reece end_blkid = dn->dn_maxblkid; 12720f6d88adSAlex Reece dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 12730f6d88adSAlex Reece 12740f6d88adSAlex Reece db_search.db_level = 0; 12750f6d88adSAlex Reece db_search.db_blkid = start_blkid; 127686bb58aeSAlex Reece db_search.db_state = DB_SEARCH; 12772f3d8780SMatthew Ahrens 1278713d6c20SMatthew Ahrens mutex_enter(&dn->dn_dbufs_mtx); 12790f6d88adSAlex Reece db = avl_find(&dn->dn_dbufs, &db_search, &where); 12800f6d88adSAlex Reece ASSERT3P(db, ==, NULL); 12812f3d8780SMatthew Ahrens 12820f6d88adSAlex Reece db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 12830f6d88adSAlex Reece 12840f6d88adSAlex Reece for (; db != NULL; db = db_next) { 12850f6d88adSAlex Reece db_next = AVL_NEXT(&dn->dn_dbufs, db); 12860a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1287cdb0ab79Smaybee 12880f6d88adSAlex Reece if (db->db_level != 0 || db->db_blkid > end_blkid) { 12890f6d88adSAlex Reece break; 12900f6d88adSAlex Reece } 12910f6d88adSAlex Reece ASSERT3U(db->db_blkid, >=, start_blkid); 1292fa9e4066Sahrens 1293fa9e4066Sahrens /* found a level 0 buffer in the range */ 1294fa9e4066Sahrens mutex_enter(&db->db_mtx); 12953b2aab18SMatthew Ahrens if (dbuf_undirty(db, tx)) { 12963b2aab18SMatthew Ahrens /* mutex has been dropped and dbuf destroyed */ 12973b2aab18SMatthew Ahrens continue; 12983b2aab18SMatthew Ahrens } 12993b2aab18SMatthew Ahrens 1300ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED || 130182c9918fSTim Haley db->db_state == DB_NOFILL || 1302ea8dc4b6Seschrock db->db_state == DB_EVICTING) { 1303fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1304fa9e4066Sahrens mutex_exit(&db->db_mtx); 1305fa9e4066Sahrens continue; 1306fa9e4066Sahrens } 1307c543ec06Sahrens if (db->db_state == DB_READ || db->db_state == DB_FILL) { 1308c543ec06Sahrens /* will be handled in dbuf_read_done or dbuf_rele */ 1309c717a561Smaybee db->db_freed_in_flight = TRUE; 1310fa9e4066Sahrens mutex_exit(&db->db_mtx); 1311fa9e4066Sahrens continue; 1312fa9e4066Sahrens } 1313ea8dc4b6Seschrock if (refcount_count(&db->db_holds) == 0) { 1314ea8dc4b6Seschrock ASSERT(db->db_buf); 1315dcbf3bd6SGeorge Wilson dbuf_destroy(db); 1316ea8dc4b6Seschrock continue; 1317ea8dc4b6Seschrock } 1318c717a561Smaybee /* The dbuf is referenced */ 1319fa9e4066Sahrens 1320c717a561Smaybee if (db->db_last_dirty != NULL) { 1321c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 1322c717a561Smaybee 1323c717a561Smaybee if (dr->dr_txg == txg) { 1324ea8dc4b6Seschrock /* 1325c717a561Smaybee * This buffer is "in-use", re-adjust the file 1326c717a561Smaybee * size to reflect that this buffer may 1327c717a561Smaybee * contain new data when we sync. 1328ea8dc4b6Seschrock */ 132906e0070dSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID && 133006e0070dSMark Shellenbaum db->db_blkid > dn->dn_maxblkid) 133144eda4d7Smaybee dn->dn_maxblkid = db->db_blkid; 1332c717a561Smaybee dbuf_unoverride(dr); 1333c717a561Smaybee } else { 1334c717a561Smaybee /* 1335c717a561Smaybee * This dbuf is not dirty in the open context. 1336c717a561Smaybee * Either uncache it (if its not referenced in 1337c717a561Smaybee * the open context) or reset its contents to 1338c717a561Smaybee * empty. 1339c717a561Smaybee */ 1340c717a561Smaybee dbuf_fix_old_data(db, txg); 134144eda4d7Smaybee } 1342c717a561Smaybee } 1343c717a561Smaybee /* clear the contents if its cached */ 1344ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 1345ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 1346fa9e4066Sahrens arc_release(db->db_buf, db); 1347fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 13486b4acc8bSahrens arc_buf_freeze(db->db_buf); 1349fa9e4066Sahrens } 1350ea8dc4b6Seschrock 1351fa9e4066Sahrens mutex_exit(&db->db_mtx); 1352fa9e4066Sahrens } 1353fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1354fa9e4066Sahrens } 1355fa9e4066Sahrens 1356fa9e4066Sahrens void 1357fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1358fa9e4066Sahrens { 1359fa9e4066Sahrens arc_buf_t *buf, *obuf; 1360fa9e4066Sahrens int osize = db->db.db_size; 1361ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1362744947dcSTom Erickson dnode_t *dn; 1363fa9e4066Sahrens 13640a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1365ea8dc4b6Seschrock 1366744947dcSTom Erickson DB_DNODE_ENTER(db); 1367744947dcSTom Erickson dn = DB_DNODE(db); 1368744947dcSTom Erickson 1369fa9e4066Sahrens /* XXX does *this* func really need the lock? */ 1370744947dcSTom Erickson ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1371fa9e4066Sahrens 1372fa9e4066Sahrens /* 137343466aaeSMax Grossman * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1374fa9e4066Sahrens * is OK, because there can be no other references to the db 1375fa9e4066Sahrens * when we are changing its size, so no concurrent DB_FILL can 1376fa9e4066Sahrens * be happening. 1377fa9e4066Sahrens */ 1378ea8dc4b6Seschrock /* 1379ea8dc4b6Seschrock * XXX we should be doing a dbuf_read, checking the return 1380ea8dc4b6Seschrock * value and returning that up to our callers 1381ea8dc4b6Seschrock */ 138243466aaeSMax Grossman dmu_buf_will_dirty(&db->db, tx); 1383fa9e4066Sahrens 1384fa9e4066Sahrens /* create the data buffer for the new block */ 13855602294fSDan Kimmel buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); 1386fa9e4066Sahrens 1387fa9e4066Sahrens /* copy old block data to the new block */ 1388fa9e4066Sahrens obuf = db->db_buf; 1389f65e61c0Sahrens bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1390fa9e4066Sahrens /* zero the remainder */ 1391f65e61c0Sahrens if (size > osize) 1392fa9e4066Sahrens bzero((uint8_t *)buf->b_data + osize, size - osize); 1393fa9e4066Sahrens 1394fa9e4066Sahrens mutex_enter(&db->db_mtx); 1395fa9e4066Sahrens dbuf_set_data(db, buf); 1396dcbf3bd6SGeorge Wilson arc_buf_destroy(obuf, db); 1397fa9e4066Sahrens db->db.db_size = size; 1398fa9e4066Sahrens 1399c717a561Smaybee if (db->db_level == 0) { 1400c717a561Smaybee ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1401c717a561Smaybee db->db_last_dirty->dt.dl.dr_data = buf; 1402c717a561Smaybee } 1403fa9e4066Sahrens mutex_exit(&db->db_mtx); 1404fa9e4066Sahrens 140561e255ceSMatthew Ahrens dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); 1406744947dcSTom Erickson DB_DNODE_EXIT(db); 1407fa9e4066Sahrens } 1408fa9e4066Sahrens 14093f9d6ad7SLin Ling void 14103f9d6ad7SLin Ling dbuf_release_bp(dmu_buf_impl_t *db) 14113f9d6ad7SLin Ling { 141243466aaeSMax Grossman objset_t *os = db->db_objset; 14133f9d6ad7SLin Ling 14143f9d6ad7SLin Ling ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 14153f9d6ad7SLin Ling ASSERT(arc_released(os->os_phys_buf) || 14163f9d6ad7SLin Ling list_link_active(&os->os_dsl_dataset->ds_synced_link)); 14173f9d6ad7SLin Ling ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 14183f9d6ad7SLin Ling 14191b912ec7SGeorge Wilson (void) arc_release(db->db_buf, db); 14203f9d6ad7SLin Ling } 14213f9d6ad7SLin Ling 14220f2e7d03SMatthew Ahrens /* 14230f2e7d03SMatthew Ahrens * We already have a dirty record for this TXG, and we are being 14240f2e7d03SMatthew Ahrens * dirtied again. 14250f2e7d03SMatthew Ahrens */ 14260f2e7d03SMatthew Ahrens static void 14270f2e7d03SMatthew Ahrens dbuf_redirty(dbuf_dirty_record_t *dr) 14280f2e7d03SMatthew Ahrens { 14290f2e7d03SMatthew Ahrens dmu_buf_impl_t *db = dr->dr_dbuf; 14300f2e7d03SMatthew Ahrens 14310f2e7d03SMatthew Ahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 14320f2e7d03SMatthew Ahrens 14330f2e7d03SMatthew Ahrens if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 14340f2e7d03SMatthew Ahrens /* 14350f2e7d03SMatthew Ahrens * If this buffer has already been written out, 14360f2e7d03SMatthew Ahrens * we now need to reset its state. 14370f2e7d03SMatthew Ahrens */ 14380f2e7d03SMatthew Ahrens dbuf_unoverride(dr); 14390f2e7d03SMatthew Ahrens if (db->db.db_object != DMU_META_DNODE_OBJECT && 14400f2e7d03SMatthew Ahrens db->db_state != DB_NOFILL) { 14410f2e7d03SMatthew Ahrens /* Already released on initial dirty, so just thaw. */ 14420f2e7d03SMatthew Ahrens ASSERT(arc_released(db->db_buf)); 14430f2e7d03SMatthew Ahrens arc_buf_thaw(db->db_buf); 14440f2e7d03SMatthew Ahrens } 14450f2e7d03SMatthew Ahrens } 14460f2e7d03SMatthew Ahrens } 14470f2e7d03SMatthew Ahrens 1448c717a561Smaybee dbuf_dirty_record_t * 1449fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1450fa9e4066Sahrens { 1451744947dcSTom Erickson dnode_t *dn; 1452744947dcSTom Erickson objset_t *os; 1453c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 1454fa9e4066Sahrens int drop_struct_lock = FALSE; 1455fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 1456fa9e4066Sahrens 1457fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1458fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 14599c9dc39aSek110237 DMU_TX_DIRTY_BUF(tx, db); 1460fa9e4066Sahrens 1461744947dcSTom Erickson DB_DNODE_ENTER(db); 1462744947dcSTom Erickson dn = DB_DNODE(db); 1463fa9e4066Sahrens /* 1464fa9e4066Sahrens * Shouldn't dirty a regular buffer in syncing context. Private 1465fa9e4066Sahrens * objects may be dirtied in syncing context, but only if they 1466fa9e4066Sahrens * were already pre-dirtied in open context. 1467fa9e4066Sahrens */ 1468c166b69dSPaul Dagnelie #ifdef DEBUG 1469c166b69dSPaul Dagnelie if (dn->dn_objset->os_dsl_dataset != NULL) { 1470c166b69dSPaul Dagnelie rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1471c166b69dSPaul Dagnelie RW_READER, FTAG); 1472c166b69dSPaul Dagnelie } 1473c717a561Smaybee ASSERT(!dmu_tx_is_syncing(tx) || 1474c717a561Smaybee BP_IS_HOLE(dn->dn_objset->os_rootbp) || 147514843421SMatthew Ahrens DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 147614843421SMatthew Ahrens dn->dn_objset->os_dsl_dataset == NULL); 1477c166b69dSPaul Dagnelie if (dn->dn_objset->os_dsl_dataset != NULL) 1478c166b69dSPaul Dagnelie rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); 1479c166b69dSPaul Dagnelie #endif 1480fa9e4066Sahrens /* 1481fa9e4066Sahrens * We make this assert for private objects as well, but after we 1482fa9e4066Sahrens * check if we're already dirty. They are allowed to re-dirty 1483fa9e4066Sahrens * in syncing context. 1484fa9e4066Sahrens */ 1485ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1486c717a561Smaybee dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1487fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1488fa9e4066Sahrens 1489fa9e4066Sahrens mutex_enter(&db->db_mtx); 1490fa9e4066Sahrens /* 1491c717a561Smaybee * XXX make this true for indirects too? The problem is that 1492c717a561Smaybee * transactions created with dmu_tx_create_assigned() from 1493c717a561Smaybee * syncing context don't bother holding ahead. 1494fa9e4066Sahrens */ 1495c717a561Smaybee ASSERT(db->db_level != 0 || 149682c9918fSTim Haley db->db_state == DB_CACHED || db->db_state == DB_FILL || 149782c9918fSTim Haley db->db_state == DB_NOFILL); 1498fa9e4066Sahrens 1499fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1500fa9e4066Sahrens /* 1501fa9e4066Sahrens * Don't set dirtyctx to SYNC if we're just modifying this as we 1502fa9e4066Sahrens * initialize the objset. 1503fa9e4066Sahrens */ 1504c166b69dSPaul Dagnelie if (dn->dn_dirtyctx == DN_UNDIRTIED) { 1505c166b69dSPaul Dagnelie if (dn->dn_objset->os_dsl_dataset != NULL) { 1506c166b69dSPaul Dagnelie rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1507c166b69dSPaul Dagnelie RW_READER, FTAG); 1508c166b69dSPaul Dagnelie } 1509c166b69dSPaul Dagnelie if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1510c166b69dSPaul Dagnelie dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? 1511c166b69dSPaul Dagnelie DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1512fa9e4066Sahrens ASSERT(dn->dn_dirtyctx_firstset == NULL); 1513fa9e4066Sahrens dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1514fa9e4066Sahrens } 1515c166b69dSPaul Dagnelie if (dn->dn_objset->os_dsl_dataset != NULL) { 1516c166b69dSPaul Dagnelie rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1517c166b69dSPaul Dagnelie FTAG); 1518c166b69dSPaul Dagnelie } 1519c166b69dSPaul Dagnelie } 1520fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1521fa9e4066Sahrens 15220a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) 15230a586ceaSMark Shellenbaum dn->dn_have_spill = B_TRUE; 15240a586ceaSMark Shellenbaum 1525fa9e4066Sahrens /* 1526fa9e4066Sahrens * If this buffer is already dirty, we're done. 1527fa9e4066Sahrens */ 1528c717a561Smaybee drp = &db->db_last_dirty; 1529c717a561Smaybee ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1530c717a561Smaybee db->db.db_object == DMU_META_DNODE_OBJECT); 15317e2186e3Sbonwick while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 15327e2186e3Sbonwick drp = &dr->dr_next; 15337e2186e3Sbonwick if (dr && dr->dr_txg == tx->tx_txg) { 1534744947dcSTom Erickson DB_DNODE_EXIT(db); 1535744947dcSTom Erickson 15360f2e7d03SMatthew Ahrens dbuf_redirty(dr); 1537fa9e4066Sahrens mutex_exit(&db->db_mtx); 15387e2186e3Sbonwick return (dr); 1539fa9e4066Sahrens } 1540fa9e4066Sahrens 1541fa9e4066Sahrens /* 1542fa9e4066Sahrens * Only valid if not already dirty. 1543fa9e4066Sahrens */ 154414843421SMatthew Ahrens ASSERT(dn->dn_object == 0 || 154514843421SMatthew Ahrens dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1546fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1547fa9e4066Sahrens 1548fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, db->db_level); 1549fa9e4066Sahrens ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1550fa9e4066Sahrens dn->dn_phys->dn_nlevels > db->db_level || 1551fa9e4066Sahrens dn->dn_next_nlevels[txgoff] > db->db_level || 1552fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1553fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1554fa9e4066Sahrens 1555fa9e4066Sahrens /* 1556fa9e4066Sahrens * We should only be dirtying in syncing context if it's the 155714843421SMatthew Ahrens * mos or we're initializing the os or it's a special object. 155814843421SMatthew Ahrens * However, we are allowed to dirty in syncing context provided 155914843421SMatthew Ahrens * we already dirtied it in open context. Hence we must make 156014843421SMatthew Ahrens * this assertion only if we're not already dirty. 1561fa9e4066Sahrens */ 1562744947dcSTom Erickson os = dn->dn_objset; 15633991b535SGeorge Wilson VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); 1564c166b69dSPaul Dagnelie #ifdef DEBUG 1565c166b69dSPaul Dagnelie if (dn->dn_objset->os_dsl_dataset != NULL) 1566c166b69dSPaul Dagnelie rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); 156714843421SMatthew Ahrens ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 156814843421SMatthew Ahrens os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1569c166b69dSPaul Dagnelie if (dn->dn_objset->os_dsl_dataset != NULL) 1570c166b69dSPaul Dagnelie rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); 1571c166b69dSPaul Dagnelie #endif 1572fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1573fa9e4066Sahrens 1574fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1575fa9e4066Sahrens 15760a586ceaSMark Shellenbaum if (db->db_blkid != DMU_BONUS_BLKID) { 157761e255ceSMatthew Ahrens dmu_objset_willuse_space(os, db->db.db_size, tx); 15781934e92fSmaybee } 15791934e92fSmaybee 1580ea8dc4b6Seschrock /* 1581ea8dc4b6Seschrock * If this buffer is dirty in an old transaction group we need 1582ea8dc4b6Seschrock * to make a copy of it so that the changes we make in this 1583ea8dc4b6Seschrock * transaction group won't leak out when we sync the older txg. 1584ea8dc4b6Seschrock */ 1585c717a561Smaybee dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1586c717a561Smaybee if (db->db_level == 0) { 1587c717a561Smaybee void *data_old = db->db_buf; 1588c717a561Smaybee 158982c9918fSTim Haley if (db->db_state != DB_NOFILL) { 15900a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 1591c717a561Smaybee dbuf_fix_old_data(db, tx->tx_txg); 1592c717a561Smaybee data_old = db->db.db_data; 1593c717a561Smaybee } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1594fa9e4066Sahrens /* 159582c9918fSTim Haley * Release the data buffer from the cache so 159682c9918fSTim Haley * that we can modify it without impacting 159782c9918fSTim Haley * possible other users of this cached data 159882c9918fSTim Haley * block. Note that indirect blocks and 159982c9918fSTim Haley * private objects are not released until the 160082c9918fSTim Haley * syncing state (since they are only modified 160182c9918fSTim Haley * then). 1602fa9e4066Sahrens */ 1603fa9e4066Sahrens arc_release(db->db_buf, db); 1604fa9e4066Sahrens dbuf_fix_old_data(db, tx->tx_txg); 1605c717a561Smaybee data_old = db->db_buf; 1606fa9e4066Sahrens } 1607c717a561Smaybee ASSERT(data_old != NULL); 160882c9918fSTim Haley } 1609c717a561Smaybee dr->dt.dl.dr_data = data_old; 1610c717a561Smaybee } else { 1611c717a561Smaybee mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1612c717a561Smaybee list_create(&dr->dt.di.dr_children, 1613c717a561Smaybee sizeof (dbuf_dirty_record_t), 1614c717a561Smaybee offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1615fa9e4066Sahrens } 161669962b56SMatthew Ahrens if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 161769962b56SMatthew Ahrens dr->dr_accounted = db->db.db_size; 1618c717a561Smaybee dr->dr_dbuf = db; 1619c717a561Smaybee dr->dr_txg = tx->tx_txg; 1620c717a561Smaybee dr->dr_next = *drp; 1621c717a561Smaybee *drp = dr; 1622fa9e4066Sahrens 1623fa9e4066Sahrens /* 1624fa9e4066Sahrens * We could have been freed_in_flight between the dbuf_noread 1625fa9e4066Sahrens * and dbuf_dirty. We win, as though the dbuf_noread() had 1626fa9e4066Sahrens * happened after the free. 1627fa9e4066Sahrens */ 16280a586ceaSMark Shellenbaum if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 16290a586ceaSMark Shellenbaum db->db_blkid != DMU_SPILL_BLKID) { 1630c717a561Smaybee mutex_enter(&dn->dn_mtx); 1631bf16b11eSMatthew Ahrens if (dn->dn_free_ranges[txgoff] != NULL) { 1632bf16b11eSMatthew Ahrens range_tree_clear(dn->dn_free_ranges[txgoff], 1633bf16b11eSMatthew Ahrens db->db_blkid, 1); 1634bf16b11eSMatthew Ahrens } 1635fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1636c717a561Smaybee db->db_freed_in_flight = FALSE; 1637c717a561Smaybee } 1638fa9e4066Sahrens 1639fa9e4066Sahrens /* 1640fa9e4066Sahrens * This buffer is now part of this txg 1641fa9e4066Sahrens */ 1642fa9e4066Sahrens dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1643fa9e4066Sahrens db->db_dirtycnt += 1; 1644fa9e4066Sahrens ASSERT3U(db->db_dirtycnt, <=, 3); 1645fa9e4066Sahrens 1646fa9e4066Sahrens mutex_exit(&db->db_mtx); 1647fa9e4066Sahrens 16480a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID || 16490a586ceaSMark Shellenbaum db->db_blkid == DMU_SPILL_BLKID) { 1650c717a561Smaybee mutex_enter(&dn->dn_mtx); 1651c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1652c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1653c717a561Smaybee mutex_exit(&dn->dn_mtx); 1654fa9e4066Sahrens dnode_setdirty(dn, tx); 1655744947dcSTom Erickson DB_DNODE_EXIT(db); 1656c717a561Smaybee return (dr); 165792654925SMatthew Ahrens } 165892654925SMatthew Ahrens 165992654925SMatthew Ahrens /* 166092654925SMatthew Ahrens * The dn_struct_rwlock prevents db_blkptr from changing 166192654925SMatthew Ahrens * due to a write from syncing context completing 166292654925SMatthew Ahrens * while we are running, so we want to acquire it before 166392654925SMatthew Ahrens * looking at db_blkptr. 166492654925SMatthew Ahrens */ 166592654925SMatthew Ahrens if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 166692654925SMatthew Ahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 166792654925SMatthew Ahrens drop_struct_lock = TRUE; 166892654925SMatthew Ahrens } 166992654925SMatthew Ahrens 1670d3469faaSMark Maybee /* 167161e255ceSMatthew Ahrens * If we are overwriting a dedup BP, then unless it is snapshotted, 167261e255ceSMatthew Ahrens * when we get to syncing context we will need to decrement its 167361e255ceSMatthew Ahrens * refcount in the DDT. Prefetch the relevant DDT block so that 167461e255ceSMatthew Ahrens * syncing context won't have to wait for the i/o. 1675d3469faaSMark Maybee */ 167661e255ceSMatthew Ahrens ddt_prefetch(os->os_spa, db->db_blkptr); 1677fa9e4066Sahrens 16788346f03fSJonathan W Adams if (db->db_level == 0) { 16798346f03fSJonathan W Adams dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 16808346f03fSJonathan W Adams ASSERT(dn->dn_maxblkid >= db->db_blkid); 16818346f03fSJonathan W Adams } 16828346f03fSJonathan W Adams 168344eda4d7Smaybee if (db->db_level+1 < dn->dn_nlevels) { 1684c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 1685c717a561Smaybee dbuf_dirty_record_t *di; 1686c717a561Smaybee int parent_held = FALSE; 1687c717a561Smaybee 1688c717a561Smaybee if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1689fa9e4066Sahrens int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1690c717a561Smaybee 1691fa9e4066Sahrens parent = dbuf_hold_level(dn, db->db_level+1, 1692fa9e4066Sahrens db->db_blkid >> epbs, FTAG); 169301025c89SJohn Harres ASSERT(parent != NULL); 1694c717a561Smaybee parent_held = TRUE; 1695c717a561Smaybee } 1696fa9e4066Sahrens if (drop_struct_lock) 1697fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1698c717a561Smaybee ASSERT3U(db->db_level+1, ==, parent->db_level); 1699c717a561Smaybee di = dbuf_dirty(parent, tx); 1700c717a561Smaybee if (parent_held) 1701ea8dc4b6Seschrock dbuf_rele(parent, FTAG); 1702c717a561Smaybee 1703c717a561Smaybee mutex_enter(&db->db_mtx); 170469962b56SMatthew Ahrens /* 170569962b56SMatthew Ahrens * Since we've dropped the mutex, it's possible that 170669962b56SMatthew Ahrens * dbuf_undirty() might have changed this out from under us. 170769962b56SMatthew Ahrens */ 1708c717a561Smaybee if (db->db_last_dirty == dr || 1709c717a561Smaybee dn->dn_object == DMU_META_DNODE_OBJECT) { 1710c717a561Smaybee mutex_enter(&di->dt.di.dr_mtx); 1711c717a561Smaybee ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1712c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1713c717a561Smaybee list_insert_tail(&di->dt.di.dr_children, dr); 1714c717a561Smaybee mutex_exit(&di->dt.di.dr_mtx); 1715c717a561Smaybee dr->dr_parent = di; 1716c717a561Smaybee } 1717c717a561Smaybee mutex_exit(&db->db_mtx); 1718fa9e4066Sahrens } else { 1719c717a561Smaybee ASSERT(db->db_level+1 == dn->dn_nlevels); 1720c717a561Smaybee ASSERT(db->db_blkid < dn->dn_nblkptr); 1721744947dcSTom Erickson ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1722c717a561Smaybee mutex_enter(&dn->dn_mtx); 1723c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1724c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1725c717a561Smaybee mutex_exit(&dn->dn_mtx); 1726fa9e4066Sahrens if (drop_struct_lock) 1727fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1728fa9e4066Sahrens } 1729fa9e4066Sahrens 1730fa9e4066Sahrens dnode_setdirty(dn, tx); 1731744947dcSTom Erickson DB_DNODE_EXIT(db); 1732c717a561Smaybee return (dr); 1733fa9e4066Sahrens } 1734fa9e4066Sahrens 17353b2aab18SMatthew Ahrens /* 17363e30c24aSWill Andrews * Undirty a buffer in the transaction group referenced by the given 17373e30c24aSWill Andrews * transaction. Return whether this evicted the dbuf. 17383b2aab18SMatthew Ahrens */ 17393b2aab18SMatthew Ahrens static boolean_t 1740fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1741fa9e4066Sahrens { 1742744947dcSTom Erickson dnode_t *dn; 1743c717a561Smaybee uint64_t txg = tx->tx_txg; 174417f17c2dSbonwick dbuf_dirty_record_t *dr, **drp; 1745fa9e4066Sahrens 1746c717a561Smaybee ASSERT(txg != 0); 174746e1baa6SMatthew Ahrens 174846e1baa6SMatthew Ahrens /* 174946e1baa6SMatthew Ahrens * Due to our use of dn_nlevels below, this can only be called 175046e1baa6SMatthew Ahrens * in open context, unless we are operating on the MOS. 175146e1baa6SMatthew Ahrens * From syncing context, dn_nlevels may be different from the 175246e1baa6SMatthew Ahrens * dn_nlevels used when dbuf was dirtied. 175346e1baa6SMatthew Ahrens */ 175446e1baa6SMatthew Ahrens ASSERT(db->db_objset == 175546e1baa6SMatthew Ahrens dmu_objset_pool(db->db_objset)->dp_meta_objset || 175646e1baa6SMatthew Ahrens txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 17570a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 17583b2aab18SMatthew Ahrens ASSERT0(db->db_level); 17593b2aab18SMatthew Ahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 1760fa9e4066Sahrens 1761fa9e4066Sahrens /* 1762fa9e4066Sahrens * If this buffer is not dirty, we're done. 1763fa9e4066Sahrens */ 176417f17c2dSbonwick for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1765c717a561Smaybee if (dr->dr_txg <= txg) 1766c717a561Smaybee break; 17673b2aab18SMatthew Ahrens if (dr == NULL || dr->dr_txg < txg) 17683b2aab18SMatthew Ahrens return (B_FALSE); 1769c717a561Smaybee ASSERT(dr->dr_txg == txg); 1770b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 1771fa9e4066Sahrens 1772744947dcSTom Erickson DB_DNODE_ENTER(db); 1773744947dcSTom Erickson dn = DB_DNODE(db); 1774744947dcSTom Erickson 1775fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1776fa9e4066Sahrens 1777fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1778fa9e4066Sahrens 177946e1baa6SMatthew Ahrens dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 178046e1baa6SMatthew Ahrens dr->dr_accounted, txg); 1781fa9e4066Sahrens 178217f17c2dSbonwick *drp = dr->dr_next; 1783c717a561Smaybee 17843f2366c2SGordon Ross /* 17853f2366c2SGordon Ross * Note that there are three places in dbuf_dirty() 17863f2366c2SGordon Ross * where this dirty record may be put on a list. 17873f2366c2SGordon Ross * Make sure to do a list_remove corresponding to 17883f2366c2SGordon Ross * every one of those list_insert calls. 17893f2366c2SGordon Ross */ 1790c717a561Smaybee if (dr->dr_parent) { 1791c717a561Smaybee mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1792c717a561Smaybee list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1793c717a561Smaybee mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 17943f2366c2SGordon Ross } else if (db->db_blkid == DMU_SPILL_BLKID || 17953f2366c2SGordon Ross db->db_level + 1 == dn->dn_nlevels) { 1796cdb0ab79Smaybee ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1797fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1798c717a561Smaybee list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1799fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1800c717a561Smaybee } 1801744947dcSTom Erickson DB_DNODE_EXIT(db); 1802c717a561Smaybee 180382c9918fSTim Haley if (db->db_state != DB_NOFILL) { 1804c717a561Smaybee dbuf_unoverride(dr); 1805c717a561Smaybee 1806c717a561Smaybee ASSERT(db->db_buf != NULL); 1807c717a561Smaybee ASSERT(dr->dt.dl.dr_data != NULL); 1808c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 1809dcbf3bd6SGeorge Wilson arc_buf_destroy(dr->dt.dl.dr_data, db); 1810c717a561Smaybee } 1811d2b3cbbdSJorgen Lundman 1812c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1813fa9e4066Sahrens 1814fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1815fa9e4066Sahrens db->db_dirtycnt -= 1; 1816fa9e4066Sahrens 1817c717a561Smaybee if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1818dcbf3bd6SGeorge Wilson ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); 1819dcbf3bd6SGeorge Wilson dbuf_destroy(db); 18203b2aab18SMatthew Ahrens return (B_TRUE); 1821fa9e4066Sahrens } 1822fa9e4066Sahrens 18233b2aab18SMatthew Ahrens return (B_FALSE); 1824fa9e4066Sahrens } 1825fa9e4066Sahrens 1826fa9e4066Sahrens void 182743466aaeSMax Grossman dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1828fa9e4066Sahrens { 182943466aaeSMax Grossman dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 18301ab7f2deSmaybee int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1831fa9e4066Sahrens 1832fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1833fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1834fa9e4066Sahrens 18350f2e7d03SMatthew Ahrens /* 18360f2e7d03SMatthew Ahrens * Quick check for dirtyness. For already dirty blocks, this 18370f2e7d03SMatthew Ahrens * reduces runtime of this function by >90%, and overall performance 18380f2e7d03SMatthew Ahrens * by 50% for some workloads (e.g. file deletion with indirect blocks 18390f2e7d03SMatthew Ahrens * cached). 18400f2e7d03SMatthew Ahrens */ 18410f2e7d03SMatthew Ahrens mutex_enter(&db->db_mtx); 18420f2e7d03SMatthew Ahrens dbuf_dirty_record_t *dr; 18430f2e7d03SMatthew Ahrens for (dr = db->db_last_dirty; 18440f2e7d03SMatthew Ahrens dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 18450f2e7d03SMatthew Ahrens /* 18460f2e7d03SMatthew Ahrens * It's possible that it is already dirty but not cached, 18470f2e7d03SMatthew Ahrens * because there are some calls to dbuf_dirty() that don't 18480f2e7d03SMatthew Ahrens * go through dmu_buf_will_dirty(). 18490f2e7d03SMatthew Ahrens */ 18500f2e7d03SMatthew Ahrens if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 18510f2e7d03SMatthew Ahrens /* This dbuf is already dirty and cached. */ 18520f2e7d03SMatthew Ahrens dbuf_redirty(dr); 18530f2e7d03SMatthew Ahrens mutex_exit(&db->db_mtx); 18540f2e7d03SMatthew Ahrens return; 18550f2e7d03SMatthew Ahrens } 18560f2e7d03SMatthew Ahrens } 18570f2e7d03SMatthew Ahrens mutex_exit(&db->db_mtx); 18580f2e7d03SMatthew Ahrens 1859744947dcSTom Erickson DB_DNODE_ENTER(db); 1860744947dcSTom Erickson if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1861fa9e4066Sahrens rf |= DB_RF_HAVESTRUCT; 1862744947dcSTom Erickson DB_DNODE_EXIT(db); 1863ea8dc4b6Seschrock (void) dbuf_read(db, NULL, rf); 1864c717a561Smaybee (void) dbuf_dirty(db, tx); 1865fa9e4066Sahrens } 1866fa9e4066Sahrens 1867fa9e4066Sahrens void 186882c9918fSTim Haley dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 186982c9918fSTim Haley { 187082c9918fSTim Haley dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 187182c9918fSTim Haley 187282c9918fSTim Haley db->db_state = DB_NOFILL; 187382c9918fSTim Haley 187482c9918fSTim Haley dmu_buf_will_fill(db_fake, tx); 187582c9918fSTim Haley } 187682c9918fSTim Haley 187782c9918fSTim Haley void 1878ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1879fa9e4066Sahrens { 1880ea8dc4b6Seschrock dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1881ea8dc4b6Seschrock 18820a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1883fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1884fa9e4066Sahrens ASSERT(db->db_level == 0); 1885fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1886fa9e4066Sahrens 1887ea8dc4b6Seschrock ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1888fa9e4066Sahrens dmu_tx_private_ok(tx)); 1889fa9e4066Sahrens 1890fa9e4066Sahrens dbuf_noread(db); 1891c717a561Smaybee (void) dbuf_dirty(db, tx); 1892fa9e4066Sahrens } 1893fa9e4066Sahrens 1894fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done 1895fa9e4066Sahrens /* ARGSUSED */ 1896fa9e4066Sahrens void 1897fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1898fa9e4066Sahrens { 1899fa9e4066Sahrens mutex_enter(&db->db_mtx); 19009c9dc39aSek110237 DBUF_VERIFY(db); 1901fa9e4066Sahrens 1902fa9e4066Sahrens if (db->db_state == DB_FILL) { 1903c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 19040a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1905fa9e4066Sahrens /* we were freed while filling */ 1906fa9e4066Sahrens /* XXX dbuf_undirty? */ 1907fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 1908c717a561Smaybee db->db_freed_in_flight = FALSE; 1909fa9e4066Sahrens } 1910fa9e4066Sahrens db->db_state = DB_CACHED; 1911fa9e4066Sahrens cv_broadcast(&db->db_changed); 1912fa9e4066Sahrens } 1913fa9e4066Sahrens mutex_exit(&db->db_mtx); 1914fa9e4066Sahrens } 1915fa9e4066Sahrens 19165d7b4d43SMatthew Ahrens void 19175d7b4d43SMatthew Ahrens dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 19185d7b4d43SMatthew Ahrens bp_embedded_type_t etype, enum zio_compress comp, 19195d7b4d43SMatthew Ahrens int uncompressed_size, int compressed_size, int byteorder, 19205d7b4d43SMatthew Ahrens dmu_tx_t *tx) 19215d7b4d43SMatthew Ahrens { 19225d7b4d43SMatthew Ahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 19235d7b4d43SMatthew Ahrens struct dirty_leaf *dl; 19245d7b4d43SMatthew Ahrens dmu_object_type_t type; 19255d7b4d43SMatthew Ahrens 1926ca0cc391SMatthew Ahrens if (etype == BP_EMBEDDED_TYPE_DATA) { 1927ca0cc391SMatthew Ahrens ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1928ca0cc391SMatthew Ahrens SPA_FEATURE_EMBEDDED_DATA)); 1929ca0cc391SMatthew Ahrens } 1930ca0cc391SMatthew Ahrens 19315d7b4d43SMatthew Ahrens DB_DNODE_ENTER(db); 19325d7b4d43SMatthew Ahrens type = DB_DNODE(db)->dn_type; 19335d7b4d43SMatthew Ahrens DB_DNODE_EXIT(db); 19345d7b4d43SMatthew Ahrens 19355d7b4d43SMatthew Ahrens ASSERT0(db->db_level); 19365d7b4d43SMatthew Ahrens ASSERT(db->db_blkid != DMU_BONUS_BLKID); 19375d7b4d43SMatthew Ahrens 19385d7b4d43SMatthew Ahrens dmu_buf_will_not_fill(dbuf, tx); 19395d7b4d43SMatthew Ahrens 19405d7b4d43SMatthew Ahrens ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 19415d7b4d43SMatthew Ahrens dl = &db->db_last_dirty->dt.dl; 19425d7b4d43SMatthew Ahrens encode_embedded_bp_compressed(&dl->dr_overridden_by, 19435d7b4d43SMatthew Ahrens data, comp, uncompressed_size, compressed_size); 19445d7b4d43SMatthew Ahrens BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 19455d7b4d43SMatthew Ahrens BP_SET_TYPE(&dl->dr_overridden_by, type); 19465d7b4d43SMatthew Ahrens BP_SET_LEVEL(&dl->dr_overridden_by, 0); 19475d7b4d43SMatthew Ahrens BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 19485d7b4d43SMatthew Ahrens 19495d7b4d43SMatthew Ahrens dl->dr_override_state = DR_OVERRIDDEN; 19505d7b4d43SMatthew Ahrens dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 19515d7b4d43SMatthew Ahrens } 19525d7b4d43SMatthew Ahrens 1953ea8dc4b6Seschrock /* 19542fdbea25SAleksandr Guzovskiy * Directly assign a provided arc buf to a given dbuf if it's not referenced 19552fdbea25SAleksandr Guzovskiy * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 19562fdbea25SAleksandr Guzovskiy */ 19572fdbea25SAleksandr Guzovskiy void 19582fdbea25SAleksandr Guzovskiy dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 19592fdbea25SAleksandr Guzovskiy { 19602fdbea25SAleksandr Guzovskiy ASSERT(!refcount_is_zero(&db->db_holds)); 19610a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 19622fdbea25SAleksandr Guzovskiy ASSERT(db->db_level == 0); 19635602294fSDan Kimmel ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); 19642fdbea25SAleksandr Guzovskiy ASSERT(buf != NULL); 19655602294fSDan Kimmel ASSERT(arc_buf_lsize(buf) == db->db.db_size); 19662fdbea25SAleksandr Guzovskiy ASSERT(tx->tx_txg != 0); 19672fdbea25SAleksandr Guzovskiy 19682fdbea25SAleksandr Guzovskiy arc_return_buf(buf, db); 19692fdbea25SAleksandr Guzovskiy ASSERT(arc_released(buf)); 19702fdbea25SAleksandr Guzovskiy 19712fdbea25SAleksandr Guzovskiy mutex_enter(&db->db_mtx); 19722fdbea25SAleksandr Guzovskiy 19732fdbea25SAleksandr Guzovskiy while (db->db_state == DB_READ || db->db_state == DB_FILL) 19742fdbea25SAleksandr Guzovskiy cv_wait(&db->db_changed, &db->db_mtx); 19752fdbea25SAleksandr Guzovskiy 19762fdbea25SAleksandr Guzovskiy ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 19772fdbea25SAleksandr Guzovskiy 19782fdbea25SAleksandr Guzovskiy if (db->db_state == DB_CACHED && 19792fdbea25SAleksandr Guzovskiy refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 19802fdbea25SAleksandr Guzovskiy mutex_exit(&db->db_mtx); 19812fdbea25SAleksandr Guzovskiy (void) dbuf_dirty(db, tx); 19822fdbea25SAleksandr Guzovskiy bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1983dcbf3bd6SGeorge Wilson arc_buf_destroy(buf, db); 1984c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_wbuf_copied(); 19852fdbea25SAleksandr Guzovskiy return; 19862fdbea25SAleksandr Guzovskiy } 19872fdbea25SAleksandr Guzovskiy 1988c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_wbuf_nocopy(); 19892fdbea25SAleksandr Guzovskiy if (db->db_state == DB_CACHED) { 19902fdbea25SAleksandr Guzovskiy dbuf_dirty_record_t *dr = db->db_last_dirty; 19912fdbea25SAleksandr Guzovskiy 19922fdbea25SAleksandr Guzovskiy ASSERT(db->db_buf != NULL); 19932fdbea25SAleksandr Guzovskiy if (dr != NULL && dr->dr_txg == tx->tx_txg) { 19942fdbea25SAleksandr Guzovskiy ASSERT(dr->dt.dl.dr_data == db->db_buf); 19952fdbea25SAleksandr Guzovskiy if (!arc_released(db->db_buf)) { 19962fdbea25SAleksandr Guzovskiy ASSERT(dr->dt.dl.dr_override_state == 19972fdbea25SAleksandr Guzovskiy DR_OVERRIDDEN); 19982fdbea25SAleksandr Guzovskiy arc_release(db->db_buf, db); 19992fdbea25SAleksandr Guzovskiy } 20002fdbea25SAleksandr Guzovskiy dr->dt.dl.dr_data = buf; 2001dcbf3bd6SGeorge Wilson arc_buf_destroy(db->db_buf, db); 20022fdbea25SAleksandr Guzovskiy } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 20032fdbea25SAleksandr Guzovskiy arc_release(db->db_buf, db); 2004dcbf3bd6SGeorge Wilson arc_buf_destroy(db->db_buf, db); 20052fdbea25SAleksandr Guzovskiy } 20062fdbea25SAleksandr Guzovskiy db->db_buf = NULL; 20072fdbea25SAleksandr Guzovskiy } 20082fdbea25SAleksandr Guzovskiy ASSERT(db->db_buf == NULL); 20092fdbea25SAleksandr Guzovskiy dbuf_set_data(db, buf); 20102fdbea25SAleksandr Guzovskiy db->db_state = DB_FILL; 20112fdbea25SAleksandr Guzovskiy mutex_exit(&db->db_mtx); 20122fdbea25SAleksandr Guzovskiy (void) dbuf_dirty(db, tx); 201343466aaeSMax Grossman dmu_buf_fill_done(&db->db, tx); 20142fdbea25SAleksandr Guzovskiy } 20152fdbea25SAleksandr Guzovskiy 2016ea8dc4b6Seschrock void 2017dcbf3bd6SGeorge Wilson dbuf_destroy(dmu_buf_impl_t *db) 2018fa9e4066Sahrens { 2019744947dcSTom Erickson dnode_t *dn; 2020ea8dc4b6Seschrock dmu_buf_impl_t *parent = db->db_parent; 2021744947dcSTom Erickson dmu_buf_impl_t *dndb; 2022fa9e4066Sahrens 2023fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 2024fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 2025fa9e4066Sahrens 2026dcbf3bd6SGeorge Wilson if (db->db_buf != NULL) { 2027dcbf3bd6SGeorge Wilson arc_buf_destroy(db->db_buf, db); 2028dcbf3bd6SGeorge Wilson db->db_buf = NULL; 2029dcbf3bd6SGeorge Wilson } 2030ea8dc4b6Seschrock 20310a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 2032dcbf3bd6SGeorge Wilson ASSERT(db->db.db_data != NULL); 2033ea8dc4b6Seschrock zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 20345a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2035fa9e4066Sahrens db->db_state = DB_UNCACHED; 2036fa9e4066Sahrens } 2037fa9e4066Sahrens 2038dcbf3bd6SGeorge Wilson dbuf_clear_data(db); 2039dcbf3bd6SGeorge Wilson 2040dcbf3bd6SGeorge Wilson if (multilist_link_active(&db->db_cache_link)) { 204194c2d0ebSMatthew Ahrens multilist_remove(dbuf_cache, db); 2042dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dbuf_cache_size, 2043dcbf3bd6SGeorge Wilson db->db.db_size, db); 2044dcbf3bd6SGeorge Wilson } 2045dcbf3bd6SGeorge Wilson 204682c9918fSTim Haley ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 2047fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 2048fa9e4066Sahrens 2049ea8dc4b6Seschrock db->db_state = DB_EVICTING; 2050ea8dc4b6Seschrock db->db_blkptr = NULL; 2051ea8dc4b6Seschrock 2052dcbf3bd6SGeorge Wilson /* 2053dcbf3bd6SGeorge Wilson * Now that db_state is DB_EVICTING, nobody else can find this via 2054dcbf3bd6SGeorge Wilson * the hash table. We can now drop db_mtx, which allows us to 2055dcbf3bd6SGeorge Wilson * acquire the dn_dbufs_mtx. 2056dcbf3bd6SGeorge Wilson */ 2057dcbf3bd6SGeorge Wilson mutex_exit(&db->db_mtx); 2058dcbf3bd6SGeorge Wilson 2059744947dcSTom Erickson DB_DNODE_ENTER(db); 2060744947dcSTom Erickson dn = DB_DNODE(db); 2061744947dcSTom Erickson dndb = dn->dn_dbuf; 2062dcbf3bd6SGeorge Wilson if (db->db_blkid != DMU_BONUS_BLKID) { 2063dcbf3bd6SGeorge Wilson boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); 2064dcbf3bd6SGeorge Wilson if (needlock) 2065dcbf3bd6SGeorge Wilson mutex_enter(&dn->dn_dbufs_mtx); 20660f6d88adSAlex Reece avl_remove(&dn->dn_dbufs, db); 2067640c1670SJosef 'Jeff' Sipek atomic_dec_32(&dn->dn_dbufs_count); 2068744947dcSTom Erickson membar_producer(); 2069744947dcSTom Erickson DB_DNODE_EXIT(db); 2070dcbf3bd6SGeorge Wilson if (needlock) 2071dcbf3bd6SGeorge Wilson mutex_exit(&dn->dn_dbufs_mtx); 2072744947dcSTom Erickson /* 2073744947dcSTom Erickson * Decrementing the dbuf count means that the hold corresponding 2074744947dcSTom Erickson * to the removed dbuf is no longer discounted in dnode_move(), 2075744947dcSTom Erickson * so the dnode cannot be moved until after we release the hold. 2076744947dcSTom Erickson * The membar_producer() ensures visibility of the decremented 2077744947dcSTom Erickson * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 2078744947dcSTom Erickson * release any lock. 2079744947dcSTom Erickson */ 2080ea8dc4b6Seschrock dnode_rele(dn, db); 2081744947dcSTom Erickson db->db_dnode_handle = NULL; 2082dcbf3bd6SGeorge Wilson 2083dcbf3bd6SGeorge Wilson dbuf_hash_remove(db); 2084744947dcSTom Erickson } else { 2085744947dcSTom Erickson DB_DNODE_EXIT(db); 2086ea8dc4b6Seschrock } 2087ea8dc4b6Seschrock 2088dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&db->db_holds)); 2089ea8dc4b6Seschrock 2090dcbf3bd6SGeorge Wilson db->db_parent = NULL; 2091dcbf3bd6SGeorge Wilson 2092dcbf3bd6SGeorge Wilson ASSERT(db->db_buf == NULL); 2093dcbf3bd6SGeorge Wilson ASSERT(db->db.db_data == NULL); 2094dcbf3bd6SGeorge Wilson ASSERT(db->db_hash_next == NULL); 2095dcbf3bd6SGeorge Wilson ASSERT(db->db_blkptr == NULL); 2096dcbf3bd6SGeorge Wilson ASSERT(db->db_data_pending == NULL); 2097dcbf3bd6SGeorge Wilson ASSERT(!multilist_link_active(&db->db_cache_link)); 2098dcbf3bd6SGeorge Wilson 2099dcbf3bd6SGeorge Wilson kmem_cache_free(dbuf_kmem_cache, db); 2100dcbf3bd6SGeorge Wilson arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2101fa9e4066Sahrens 2102fa9e4066Sahrens /* 2103744947dcSTom Erickson * If this dbuf is referenced from an indirect dbuf, 2104fa9e4066Sahrens * decrement the ref count on the indirect dbuf. 2105fa9e4066Sahrens */ 2106c543ec06Sahrens if (parent && parent != dndb) 2107ea8dc4b6Seschrock dbuf_rele(parent, db); 2108fa9e4066Sahrens } 2109fa9e4066Sahrens 2110a2cdcdd2SPaul Dagnelie /* 2111a2cdcdd2SPaul Dagnelie * Note: While bpp will always be updated if the function returns success, 2112a2cdcdd2SPaul Dagnelie * parentp will not be updated if the dnode does not have dn_dbuf filled in; 2113a2cdcdd2SPaul Dagnelie * this happens when the dnode is the meta-dnode, or a userused or groupused 2114a2cdcdd2SPaul Dagnelie * object. 2115a2cdcdd2SPaul Dagnelie */ 2116fa9e4066Sahrens static int 2117fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 2118fa9e4066Sahrens dmu_buf_impl_t **parentp, blkptr_t **bpp) 2119fa9e4066Sahrens { 21200b69c2f0Sahrens *parentp = NULL; 21210b69c2f0Sahrens *bpp = NULL; 21220b69c2f0Sahrens 21230a586ceaSMark Shellenbaum ASSERT(blkid != DMU_BONUS_BLKID); 21240a586ceaSMark Shellenbaum 21250a586ceaSMark Shellenbaum if (blkid == DMU_SPILL_BLKID) { 21260a586ceaSMark Shellenbaum mutex_enter(&dn->dn_mtx); 212706e0070dSMark Shellenbaum if (dn->dn_have_spill && 212806e0070dSMark Shellenbaum (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 21290a586ceaSMark Shellenbaum *bpp = &dn->dn_phys->dn_spill; 21300a586ceaSMark Shellenbaum else 21310a586ceaSMark Shellenbaum *bpp = NULL; 21320a586ceaSMark Shellenbaum dbuf_add_ref(dn->dn_dbuf, NULL); 21330a586ceaSMark Shellenbaum *parentp = dn->dn_dbuf; 21340a586ceaSMark Shellenbaum mutex_exit(&dn->dn_mtx); 21350a586ceaSMark Shellenbaum return (0); 21360a586ceaSMark Shellenbaum } 2137ea8dc4b6Seschrock 21387de35a3eSPaul Dagnelie int nlevels = 21397de35a3eSPaul Dagnelie (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; 21407de35a3eSPaul Dagnelie int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2141fa9e4066Sahrens 2142fa9e4066Sahrens ASSERT3U(level * epbs, <, 64); 2143fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 21447de35a3eSPaul Dagnelie /* 21457de35a3eSPaul Dagnelie * This assertion shouldn't trip as long as the max indirect block size 21467de35a3eSPaul Dagnelie * is less than 1M. The reason for this is that up to that point, 21477de35a3eSPaul Dagnelie * the number of levels required to address an entire object with blocks 21487de35a3eSPaul Dagnelie * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In 21497de35a3eSPaul Dagnelie * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 21507de35a3eSPaul Dagnelie * (i.e. we can address the entire object), objects will all use at most 21517de35a3eSPaul Dagnelie * N-1 levels and the assertion won't overflow. However, once epbs is 21527de35a3eSPaul Dagnelie * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be 21537de35a3eSPaul Dagnelie * enough to address an entire object, so objects will have 5 levels, 21547de35a3eSPaul Dagnelie * but then this assertion will overflow. 21557de35a3eSPaul Dagnelie * 21567de35a3eSPaul Dagnelie * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we 21577de35a3eSPaul Dagnelie * need to redo this logic to handle overflows. 21587de35a3eSPaul Dagnelie */ 21597de35a3eSPaul Dagnelie ASSERT(level >= nlevels || 21607de35a3eSPaul Dagnelie ((nlevels - level - 1) * epbs) + 21617de35a3eSPaul Dagnelie highbit64(dn->dn_phys->dn_nblkptr) <= 64); 2162ea8dc4b6Seschrock if (level >= nlevels || 21637de35a3eSPaul Dagnelie blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << 21647de35a3eSPaul Dagnelie ((nlevels - level - 1) * epbs)) || 21657de35a3eSPaul Dagnelie (fail_sparse && 21667de35a3eSPaul Dagnelie blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 2167fa9e4066Sahrens /* the buffer has no parent yet */ 2168be6fd75aSMatthew Ahrens return (SET_ERROR(ENOENT)); 2169fa9e4066Sahrens } else if (level < nlevels-1) { 2170fa9e4066Sahrens /* this block is referenced from an indirect block */ 2171fa9e4066Sahrens int err = dbuf_hold_impl(dn, level+1, 2172a2cdcdd2SPaul Dagnelie blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 2173fa9e4066Sahrens if (err) 2174fa9e4066Sahrens return (err); 2175ea8dc4b6Seschrock err = dbuf_read(*parentp, NULL, 2176ea8dc4b6Seschrock (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 2177c543ec06Sahrens if (err) { 2178c543ec06Sahrens dbuf_rele(*parentp, NULL); 2179c543ec06Sahrens *parentp = NULL; 2180c543ec06Sahrens return (err); 2181c543ec06Sahrens } 2182fa9e4066Sahrens *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 2183fa9e4066Sahrens (blkid & ((1ULL << epbs) - 1)); 21847de35a3eSPaul Dagnelie if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) 21857de35a3eSPaul Dagnelie ASSERT(BP_IS_HOLE(*bpp)); 2186c543ec06Sahrens return (0); 2187fa9e4066Sahrens } else { 2188fa9e4066Sahrens /* the block is referenced from the dnode */ 2189fa9e4066Sahrens ASSERT3U(level, ==, nlevels-1); 2190fa9e4066Sahrens ASSERT(dn->dn_phys->dn_nblkptr == 0 || 2191fa9e4066Sahrens blkid < dn->dn_phys->dn_nblkptr); 2192c543ec06Sahrens if (dn->dn_dbuf) { 2193c543ec06Sahrens dbuf_add_ref(dn->dn_dbuf, NULL); 2194fa9e4066Sahrens *parentp = dn->dn_dbuf; 2195c543ec06Sahrens } 2196fa9e4066Sahrens *bpp = &dn->dn_phys->dn_blkptr[blkid]; 2197fa9e4066Sahrens return (0); 2198fa9e4066Sahrens } 2199fa9e4066Sahrens } 2200fa9e4066Sahrens 2201fa9e4066Sahrens static dmu_buf_impl_t * 2202fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 2203fa9e4066Sahrens dmu_buf_impl_t *parent, blkptr_t *blkptr) 2204fa9e4066Sahrens { 2205503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 2206fa9e4066Sahrens dmu_buf_impl_t *db, *odb; 2207fa9e4066Sahrens 2208fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2209fa9e4066Sahrens ASSERT(dn->dn_type != DMU_OT_NONE); 2210fa9e4066Sahrens 2211dcbf3bd6SGeorge Wilson db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); 2212fa9e4066Sahrens 2213fa9e4066Sahrens db->db_objset = os; 2214fa9e4066Sahrens db->db.db_object = dn->dn_object; 2215fa9e4066Sahrens db->db_level = level; 2216fa9e4066Sahrens db->db_blkid = blkid; 2217c717a561Smaybee db->db_last_dirty = NULL; 2218ea8dc4b6Seschrock db->db_dirtycnt = 0; 2219744947dcSTom Erickson db->db_dnode_handle = dn->dn_handle; 2220ea8dc4b6Seschrock db->db_parent = parent; 2221ea8dc4b6Seschrock db->db_blkptr = blkptr; 2222fa9e4066Sahrens 2223bc9014e6SJustin Gibbs db->db_user = NULL; 2224d2058105SJustin T. Gibbs db->db_user_immediate_evict = FALSE; 2225d2058105SJustin T. Gibbs db->db_freed_in_flight = FALSE; 2226d2058105SJustin T. Gibbs db->db_pending_evict = FALSE; 2227ea8dc4b6Seschrock 22280a586ceaSMark Shellenbaum if (blkid == DMU_BONUS_BLKID) { 2229ea8dc4b6Seschrock ASSERT3P(parent, ==, dn->dn_dbuf); 22301934e92fSmaybee db->db.db_size = DN_MAX_BONUSLEN - 22311934e92fSmaybee (dn->dn_nblkptr-1) * sizeof (blkptr_t); 22321934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 22330a586ceaSMark Shellenbaum db->db.db_offset = DMU_BONUS_BLKID; 2234ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 2235ea8dc4b6Seschrock /* the bonus dbuf is not placed in the hash table */ 22365a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2237ea8dc4b6Seschrock return (db); 22380a586ceaSMark Shellenbaum } else if (blkid == DMU_SPILL_BLKID) { 22390a586ceaSMark Shellenbaum db->db.db_size = (blkptr != NULL) ? 22400a586ceaSMark Shellenbaum BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 22410a586ceaSMark Shellenbaum db->db.db_offset = 0; 2242fa9e4066Sahrens } else { 2243fa9e4066Sahrens int blocksize = 2244fa9e4066Sahrens db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 2245fa9e4066Sahrens db->db.db_size = blocksize; 2246fa9e4066Sahrens db->db.db_offset = db->db_blkid * blocksize; 2247fa9e4066Sahrens } 2248fa9e4066Sahrens 2249fa9e4066Sahrens /* 2250fa9e4066Sahrens * Hold the dn_dbufs_mtx while we get the new dbuf 2251fa9e4066Sahrens * in the hash table *and* added to the dbufs list. 2252fa9e4066Sahrens * This prevents a possible deadlock with someone 2253fa9e4066Sahrens * trying to look up this dbuf before its added to the 2254fa9e4066Sahrens * dn_dbufs list. 2255fa9e4066Sahrens */ 2256fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 2257ea8dc4b6Seschrock db->db_state = DB_EVICTING; 2258fa9e4066Sahrens if ((odb = dbuf_hash_insert(db)) != NULL) { 2259fa9e4066Sahrens /* someone else inserted it first */ 2260dcbf3bd6SGeorge Wilson kmem_cache_free(dbuf_kmem_cache, db); 2261fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 2262fa9e4066Sahrens return (odb); 2263fa9e4066Sahrens } 22640f6d88adSAlex Reece avl_add(&dn->dn_dbufs, db); 2265653af1b8SStephen Blinick 2266ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 2267fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 22685a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2269fa9e4066Sahrens 2270fa9e4066Sahrens if (parent && parent != dn->dn_dbuf) 2271fa9e4066Sahrens dbuf_add_ref(parent, db); 2272fa9e4066Sahrens 2273ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 2274ea8dc4b6Seschrock refcount_count(&dn->dn_holds) > 0); 2275fa9e4066Sahrens (void) refcount_add(&dn->dn_holds, db); 2276640c1670SJosef 'Jeff' Sipek atomic_inc_32(&dn->dn_dbufs_count); 2277fa9e4066Sahrens 2278fa9e4066Sahrens dprintf_dbuf(db, "db=%p\n", db); 2279fa9e4066Sahrens 2280fa9e4066Sahrens return (db); 2281fa9e4066Sahrens } 2282fa9e4066Sahrens 2283a2cdcdd2SPaul Dagnelie typedef struct dbuf_prefetch_arg { 2284a2cdcdd2SPaul Dagnelie spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2285a2cdcdd2SPaul Dagnelie zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2286a2cdcdd2SPaul Dagnelie int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2287a2cdcdd2SPaul Dagnelie int dpa_curlevel; /* The current level that we're reading */ 2288dcbf3bd6SGeorge Wilson dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ 2289a2cdcdd2SPaul Dagnelie zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2290a2cdcdd2SPaul Dagnelie zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2291a2cdcdd2SPaul Dagnelie arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2292a2cdcdd2SPaul Dagnelie } dbuf_prefetch_arg_t; 2293a2cdcdd2SPaul Dagnelie 2294a2cdcdd2SPaul Dagnelie /* 2295a2cdcdd2SPaul Dagnelie * Actually issue the prefetch read for the block given. 2296a2cdcdd2SPaul Dagnelie */ 2297a2cdcdd2SPaul Dagnelie static void 2298a2cdcdd2SPaul Dagnelie dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2299fa9e4066Sahrens { 2300a2cdcdd2SPaul Dagnelie if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2301a2cdcdd2SPaul Dagnelie return; 2302a2cdcdd2SPaul Dagnelie 2303a2cdcdd2SPaul Dagnelie arc_flags_t aflags = 2304a2cdcdd2SPaul Dagnelie dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2305a2cdcdd2SPaul Dagnelie 2306a2cdcdd2SPaul Dagnelie ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2307a2cdcdd2SPaul Dagnelie ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2308a2cdcdd2SPaul Dagnelie ASSERT(dpa->dpa_zio != NULL); 2309a2cdcdd2SPaul Dagnelie (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2310a2cdcdd2SPaul Dagnelie dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2311a2cdcdd2SPaul Dagnelie &aflags, &dpa->dpa_zb); 2312a2cdcdd2SPaul Dagnelie } 2313a2cdcdd2SPaul Dagnelie 2314a2cdcdd2SPaul Dagnelie /* 2315a2cdcdd2SPaul Dagnelie * Called when an indirect block above our prefetch target is read in. This 2316a2cdcdd2SPaul Dagnelie * will either read in the next indirect block down the tree or issue the actual 2317a2cdcdd2SPaul Dagnelie * prefetch if the next block down is our target. 2318a2cdcdd2SPaul Dagnelie */ 2319a2cdcdd2SPaul Dagnelie static void 2320a2cdcdd2SPaul Dagnelie dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2321a2cdcdd2SPaul Dagnelie { 2322a2cdcdd2SPaul Dagnelie dbuf_prefetch_arg_t *dpa = private; 2323a2cdcdd2SPaul Dagnelie 2324a2cdcdd2SPaul Dagnelie ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2325a2cdcdd2SPaul Dagnelie ASSERT3S(dpa->dpa_curlevel, >, 0); 2326dcbf3bd6SGeorge Wilson 2327dcbf3bd6SGeorge Wilson /* 2328dcbf3bd6SGeorge Wilson * The dpa_dnode is only valid if we are called with a NULL 2329dcbf3bd6SGeorge Wilson * zio. This indicates that the arc_read() returned without 2330dcbf3bd6SGeorge Wilson * first calling zio_read() to issue a physical read. Once 2331dcbf3bd6SGeorge Wilson * a physical read is made the dpa_dnode must be invalidated 2332dcbf3bd6SGeorge Wilson * as the locks guarding it may have been dropped. If the 2333dcbf3bd6SGeorge Wilson * dpa_dnode is still valid, then we want to add it to the dbuf 2334dcbf3bd6SGeorge Wilson * cache. To do so, we must hold the dbuf associated with the block 2335dcbf3bd6SGeorge Wilson * we just prefetched, read its contents so that we associate it 2336dcbf3bd6SGeorge Wilson * with an arc_buf_t, and then release it. 2337dcbf3bd6SGeorge Wilson */ 2338a2cdcdd2SPaul Dagnelie if (zio != NULL) { 2339a2cdcdd2SPaul Dagnelie ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2340dcbf3bd6SGeorge Wilson if (zio->io_flags & ZIO_FLAG_RAW) { 2341dcbf3bd6SGeorge Wilson ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); 2342dcbf3bd6SGeorge Wilson } else { 2343a2cdcdd2SPaul Dagnelie ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2344dcbf3bd6SGeorge Wilson } 2345a2cdcdd2SPaul Dagnelie ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2346dcbf3bd6SGeorge Wilson 2347dcbf3bd6SGeorge Wilson dpa->dpa_dnode = NULL; 2348dcbf3bd6SGeorge Wilson } else if (dpa->dpa_dnode != NULL) { 2349dcbf3bd6SGeorge Wilson uint64_t curblkid = dpa->dpa_zb.zb_blkid >> 2350dcbf3bd6SGeorge Wilson (dpa->dpa_epbs * (dpa->dpa_curlevel - 2351dcbf3bd6SGeorge Wilson dpa->dpa_zb.zb_level)); 2352dcbf3bd6SGeorge Wilson dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, 2353dcbf3bd6SGeorge Wilson dpa->dpa_curlevel, curblkid, FTAG); 2354dcbf3bd6SGeorge Wilson (void) dbuf_read(db, NULL, 2355dcbf3bd6SGeorge Wilson DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); 2356dcbf3bd6SGeorge Wilson dbuf_rele(db, FTAG); 2357a2cdcdd2SPaul Dagnelie } 2358a2cdcdd2SPaul Dagnelie 2359a2cdcdd2SPaul Dagnelie dpa->dpa_curlevel--; 2360a2cdcdd2SPaul Dagnelie 2361a2cdcdd2SPaul Dagnelie uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2362a2cdcdd2SPaul Dagnelie (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2363a2cdcdd2SPaul Dagnelie blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2364a2cdcdd2SPaul Dagnelie P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2365a2cdcdd2SPaul Dagnelie if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2366a2cdcdd2SPaul Dagnelie kmem_free(dpa, sizeof (*dpa)); 2367a2cdcdd2SPaul Dagnelie } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2368a2cdcdd2SPaul Dagnelie ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2369a2cdcdd2SPaul Dagnelie dbuf_issue_final_prefetch(dpa, bp); 2370a2cdcdd2SPaul Dagnelie kmem_free(dpa, sizeof (*dpa)); 2371a2cdcdd2SPaul Dagnelie } else { 2372a2cdcdd2SPaul Dagnelie arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2373a2cdcdd2SPaul Dagnelie zbookmark_phys_t zb; 2374a2cdcdd2SPaul Dagnelie 2375a2cdcdd2SPaul Dagnelie ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2376a2cdcdd2SPaul Dagnelie 2377a2cdcdd2SPaul Dagnelie SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2378a2cdcdd2SPaul Dagnelie dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2379a2cdcdd2SPaul Dagnelie 2380a2cdcdd2SPaul Dagnelie (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2381a2cdcdd2SPaul Dagnelie bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2382a2cdcdd2SPaul Dagnelie ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2383a2cdcdd2SPaul Dagnelie &iter_aflags, &zb); 2384a2cdcdd2SPaul Dagnelie } 2385dcbf3bd6SGeorge Wilson 2386dcbf3bd6SGeorge Wilson arc_buf_destroy(abuf, private); 2387a2cdcdd2SPaul Dagnelie } 2388a2cdcdd2SPaul Dagnelie 2389a2cdcdd2SPaul Dagnelie /* 2390a2cdcdd2SPaul Dagnelie * Issue prefetch reads for the given block on the given level. If the indirect 2391a2cdcdd2SPaul Dagnelie * blocks above that block are not in memory, we will read them in 2392a2cdcdd2SPaul Dagnelie * asynchronously. As a result, this call never blocks waiting for a read to 2393a2cdcdd2SPaul Dagnelie * complete. 2394a2cdcdd2SPaul Dagnelie */ 2395a2cdcdd2SPaul Dagnelie void 2396a2cdcdd2SPaul Dagnelie dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2397a2cdcdd2SPaul Dagnelie arc_flags_t aflags) 2398a2cdcdd2SPaul Dagnelie { 2399a2cdcdd2SPaul Dagnelie blkptr_t bp; 2400a2cdcdd2SPaul Dagnelie int epbs, nlevels, curlevel; 2401a2cdcdd2SPaul Dagnelie uint64_t curblkid; 2402fa9e4066Sahrens 24030a586ceaSMark Shellenbaum ASSERT(blkid != DMU_BONUS_BLKID); 2404fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2405fa9e4066Sahrens 2406cf6106c8SMatthew Ahrens if (blkid > dn->dn_maxblkid) 2407cf6106c8SMatthew Ahrens return; 2408cf6106c8SMatthew Ahrens 2409fa9e4066Sahrens if (dnode_block_freed(dn, blkid)) 2410fa9e4066Sahrens return; 2411fa9e4066Sahrens 2412fa9e4066Sahrens /* 2413a2cdcdd2SPaul Dagnelie * This dnode hasn't been written to disk yet, so there's nothing to 2414a2cdcdd2SPaul Dagnelie * prefetch. 2415fa9e4066Sahrens */ 2416a2cdcdd2SPaul Dagnelie nlevels = dn->dn_phys->dn_nlevels; 2417a2cdcdd2SPaul Dagnelie if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2418a2cdcdd2SPaul Dagnelie return; 2419a2cdcdd2SPaul Dagnelie 2420a2cdcdd2SPaul Dagnelie epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2421a2cdcdd2SPaul Dagnelie if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2422a2cdcdd2SPaul Dagnelie return; 2423a2cdcdd2SPaul Dagnelie 2424a2cdcdd2SPaul Dagnelie dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2425a2cdcdd2SPaul Dagnelie level, blkid); 2426a2cdcdd2SPaul Dagnelie if (db != NULL) { 2427fa9e4066Sahrens mutex_exit(&db->db_mtx); 2428a2cdcdd2SPaul Dagnelie /* 2429a2cdcdd2SPaul Dagnelie * This dbuf already exists. It is either CACHED, or 2430a2cdcdd2SPaul Dagnelie * (we assume) about to be read or filled. 2431a2cdcdd2SPaul Dagnelie */ 2432fa9e4066Sahrens return; 2433fa9e4066Sahrens } 2434fa9e4066Sahrens 2435a2cdcdd2SPaul Dagnelie /* 2436a2cdcdd2SPaul Dagnelie * Find the closest ancestor (indirect block) of the target block 2437a2cdcdd2SPaul Dagnelie * that is present in the cache. In this indirect block, we will 2438a2cdcdd2SPaul Dagnelie * find the bp that is at curlevel, curblkid. 2439a2cdcdd2SPaul Dagnelie */ 2440a2cdcdd2SPaul Dagnelie curlevel = level; 2441a2cdcdd2SPaul Dagnelie curblkid = blkid; 2442a2cdcdd2SPaul Dagnelie while (curlevel < nlevels - 1) { 2443a2cdcdd2SPaul Dagnelie int parent_level = curlevel + 1; 2444a2cdcdd2SPaul Dagnelie uint64_t parent_blkid = curblkid >> epbs; 2445a2cdcdd2SPaul Dagnelie dmu_buf_impl_t *db; 2446a2cdcdd2SPaul Dagnelie 2447a2cdcdd2SPaul Dagnelie if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2448a2cdcdd2SPaul Dagnelie FALSE, TRUE, FTAG, &db) == 0) { 2449a2cdcdd2SPaul Dagnelie blkptr_t *bpp = db->db_buf->b_data; 2450a2cdcdd2SPaul Dagnelie bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2451a2cdcdd2SPaul Dagnelie dbuf_rele(db, FTAG); 2452a2cdcdd2SPaul Dagnelie break; 2453a2cdcdd2SPaul Dagnelie } 2454a2cdcdd2SPaul Dagnelie 2455a2cdcdd2SPaul Dagnelie curlevel = parent_level; 2456a2cdcdd2SPaul Dagnelie curblkid = parent_blkid; 2457a2cdcdd2SPaul Dagnelie } 2458a2cdcdd2SPaul Dagnelie 2459a2cdcdd2SPaul Dagnelie if (curlevel == nlevels - 1) { 2460a2cdcdd2SPaul Dagnelie /* No cached indirect blocks found. */ 2461a2cdcdd2SPaul Dagnelie ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2462a2cdcdd2SPaul Dagnelie bp = dn->dn_phys->dn_blkptr[curblkid]; 2463a2cdcdd2SPaul Dagnelie } 2464a2cdcdd2SPaul Dagnelie if (BP_IS_HOLE(&bp)) 2465a2cdcdd2SPaul Dagnelie return; 2466a2cdcdd2SPaul Dagnelie 2467a2cdcdd2SPaul Dagnelie ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2468a2cdcdd2SPaul Dagnelie 2469a2cdcdd2SPaul Dagnelie zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2470a2cdcdd2SPaul Dagnelie ZIO_FLAG_CANFAIL); 2471a2cdcdd2SPaul Dagnelie 2472a2cdcdd2SPaul Dagnelie dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2473b24ab676SJeff Bonwick dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2474a2cdcdd2SPaul Dagnelie SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2475a2cdcdd2SPaul Dagnelie dn->dn_object, level, blkid); 2476a2cdcdd2SPaul Dagnelie dpa->dpa_curlevel = curlevel; 2477a2cdcdd2SPaul Dagnelie dpa->dpa_prio = prio; 2478a2cdcdd2SPaul Dagnelie dpa->dpa_aflags = aflags; 2479a2cdcdd2SPaul Dagnelie dpa->dpa_spa = dn->dn_objset->os_spa; 2480dcbf3bd6SGeorge Wilson dpa->dpa_dnode = dn; 2481a2cdcdd2SPaul Dagnelie dpa->dpa_epbs = epbs; 2482a2cdcdd2SPaul Dagnelie dpa->dpa_zio = pio; 2483a2cdcdd2SPaul Dagnelie 2484a2cdcdd2SPaul Dagnelie /* 2485a2cdcdd2SPaul Dagnelie * If we have the indirect just above us, no need to do the asynchronous 2486a2cdcdd2SPaul Dagnelie * prefetch chain; we'll just run the last step ourselves. If we're at 2487a2cdcdd2SPaul Dagnelie * a higher level, though, we want to issue the prefetches for all the 2488a2cdcdd2SPaul Dagnelie * indirect blocks asynchronously, so we can go on with whatever we were 2489a2cdcdd2SPaul Dagnelie * doing. 2490a2cdcdd2SPaul Dagnelie */ 2491a2cdcdd2SPaul Dagnelie if (curlevel == level) { 2492a2cdcdd2SPaul Dagnelie ASSERT3U(curblkid, ==, blkid); 2493a2cdcdd2SPaul Dagnelie dbuf_issue_final_prefetch(dpa, &bp); 2494a2cdcdd2SPaul Dagnelie kmem_free(dpa, sizeof (*dpa)); 2495a2cdcdd2SPaul Dagnelie } else { 2496a2cdcdd2SPaul Dagnelie arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 24977802d7bfSMatthew Ahrens zbookmark_phys_t zb; 2498b24ab676SJeff Bonwick 2499a2cdcdd2SPaul Dagnelie SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2500a2cdcdd2SPaul Dagnelie dn->dn_object, curlevel, curblkid); 2501a2cdcdd2SPaul Dagnelie (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2502a2cdcdd2SPaul Dagnelie &bp, dbuf_prefetch_indirect_done, dpa, prio, 2503fa9e4066Sahrens ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2504a2cdcdd2SPaul Dagnelie &iter_aflags, &zb); 2505fa9e4066Sahrens } 2506a2cdcdd2SPaul Dagnelie /* 2507a2cdcdd2SPaul Dagnelie * We use pio here instead of dpa_zio since it's possible that 2508a2cdcdd2SPaul Dagnelie * dpa may have already been freed. 2509a2cdcdd2SPaul Dagnelie */ 2510a2cdcdd2SPaul Dagnelie zio_nowait(pio); 2511fa9e4066Sahrens } 2512fa9e4066Sahrens 2513fa9e4066Sahrens /* 2514fa9e4066Sahrens * Returns with db_holds incremented, and db_mtx not held. 2515fa9e4066Sahrens * Note: dn_struct_rwlock must be held. 2516fa9e4066Sahrens */ 2517fa9e4066Sahrens int 2518a2cdcdd2SPaul Dagnelie dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2519a2cdcdd2SPaul Dagnelie boolean_t fail_sparse, boolean_t fail_uncached, 2520fa9e4066Sahrens void *tag, dmu_buf_impl_t **dbp) 2521fa9e4066Sahrens { 2522fa9e4066Sahrens dmu_buf_impl_t *db, *parent = NULL; 2523fa9e4066Sahrens 25240a586ceaSMark Shellenbaum ASSERT(blkid != DMU_BONUS_BLKID); 2525fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2526fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, level); 2527fa9e4066Sahrens 2528fa9e4066Sahrens *dbp = NULL; 2529ea8dc4b6Seschrock top: 2530fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 2531e57a022bSJustin T. Gibbs db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2532fa9e4066Sahrens 2533fa9e4066Sahrens if (db == NULL) { 2534fa9e4066Sahrens blkptr_t *bp = NULL; 2535fa9e4066Sahrens int err; 2536fa9e4066Sahrens 2537a2cdcdd2SPaul Dagnelie if (fail_uncached) 2538a2cdcdd2SPaul Dagnelie return (SET_ERROR(ENOENT)); 2539a2cdcdd2SPaul Dagnelie 2540c543ec06Sahrens ASSERT3P(parent, ==, NULL); 2541fa9e4066Sahrens err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2542fa9e4066Sahrens if (fail_sparse) { 2543fa9e4066Sahrens if (err == 0 && bp && BP_IS_HOLE(bp)) 2544be6fd75aSMatthew Ahrens err = SET_ERROR(ENOENT); 2545fa9e4066Sahrens if (err) { 2546c543ec06Sahrens if (parent) 2547ea8dc4b6Seschrock dbuf_rele(parent, NULL); 2548fa9e4066Sahrens return (err); 2549fa9e4066Sahrens } 2550fa9e4066Sahrens } 2551ea8dc4b6Seschrock if (err && err != ENOENT) 2552ea8dc4b6Seschrock return (err); 2553fa9e4066Sahrens db = dbuf_create(dn, level, blkid, parent, bp); 2554fa9e4066Sahrens } 2555fa9e4066Sahrens 2556a2cdcdd2SPaul Dagnelie if (fail_uncached && db->db_state != DB_CACHED) { 2557a2cdcdd2SPaul Dagnelie mutex_exit(&db->db_mtx); 2558a2cdcdd2SPaul Dagnelie return (SET_ERROR(ENOENT)); 2559a2cdcdd2SPaul Dagnelie } 2560a2cdcdd2SPaul Dagnelie 2561dcbf3bd6SGeorge Wilson if (db->db_buf != NULL) 2562ea8dc4b6Seschrock ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2563ea8dc4b6Seschrock 2564ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2565ea8dc4b6Seschrock 2566fa9e4066Sahrens /* 2567c717a561Smaybee * If this buffer is currently syncing out, and we are are 2568c717a561Smaybee * still referencing it from db_data, we need to make a copy 2569c717a561Smaybee * of it in case we decide we want to dirty it again in this txg. 2570fa9e4066Sahrens */ 25710a586ceaSMark Shellenbaum if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2572ea8dc4b6Seschrock dn->dn_object != DMU_META_DNODE_OBJECT && 2573c717a561Smaybee db->db_state == DB_CACHED && db->db_data_pending) { 2574c717a561Smaybee dbuf_dirty_record_t *dr = db->db_data_pending; 2575c717a561Smaybee 2576c717a561Smaybee if (dr->dt.dl.dr_data == db->db_buf) { 2577ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2578fa9e4066Sahrens 2579c717a561Smaybee dbuf_set_data(db, 25805602294fSDan Kimmel arc_alloc_buf(dn->dn_objset->os_spa, db, type, 25815602294fSDan Kimmel db->db.db_size)); 2582c717a561Smaybee bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2583fa9e4066Sahrens db->db.db_size); 2584fa9e4066Sahrens } 2585c717a561Smaybee } 2586fa9e4066Sahrens 2587dcbf3bd6SGeorge Wilson if (multilist_link_active(&db->db_cache_link)) { 2588dcbf3bd6SGeorge Wilson ASSERT(refcount_is_zero(&db->db_holds)); 258994c2d0ebSMatthew Ahrens multilist_remove(dbuf_cache, db); 2590dcbf3bd6SGeorge Wilson (void) refcount_remove_many(&dbuf_cache_size, 2591dcbf3bd6SGeorge Wilson db->db.db_size, db); 2592dcbf3bd6SGeorge Wilson } 2593ea8dc4b6Seschrock (void) refcount_add(&db->db_holds, tag); 25949c9dc39aSek110237 DBUF_VERIFY(db); 2595fa9e4066Sahrens mutex_exit(&db->db_mtx); 2596fa9e4066Sahrens 2597fa9e4066Sahrens /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2598c543ec06Sahrens if (parent) 2599ea8dc4b6Seschrock dbuf_rele(parent, NULL); 2600fa9e4066Sahrens 2601744947dcSTom Erickson ASSERT3P(DB_DNODE(db), ==, dn); 2602fa9e4066Sahrens ASSERT3U(db->db_blkid, ==, blkid); 2603fa9e4066Sahrens ASSERT3U(db->db_level, ==, level); 2604fa9e4066Sahrens *dbp = db; 2605fa9e4066Sahrens 2606fa9e4066Sahrens return (0); 2607fa9e4066Sahrens } 2608fa9e4066Sahrens 2609fa9e4066Sahrens dmu_buf_impl_t * 2610ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2611fa9e4066Sahrens { 2612a2cdcdd2SPaul Dagnelie return (dbuf_hold_level(dn, 0, blkid, tag)); 2613fa9e4066Sahrens } 2614fa9e4066Sahrens 2615fa9e4066Sahrens dmu_buf_impl_t * 2616fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2617fa9e4066Sahrens { 2618fa9e4066Sahrens dmu_buf_impl_t *db; 2619a2cdcdd2SPaul Dagnelie int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2620ea8dc4b6Seschrock return (err ? NULL : db); 2621fa9e4066Sahrens } 2622fa9e4066Sahrens 26231934e92fSmaybee void 2624ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn) 2625fa9e4066Sahrens { 2626ea8dc4b6Seschrock ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2627ea8dc4b6Seschrock 2628ea8dc4b6Seschrock ASSERT(dn->dn_bonus == NULL); 26290a586ceaSMark Shellenbaum dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 26300a586ceaSMark Shellenbaum } 26310a586ceaSMark Shellenbaum 26320a586ceaSMark Shellenbaum int 26330a586ceaSMark Shellenbaum dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 26340a586ceaSMark Shellenbaum { 26350a586ceaSMark Shellenbaum dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2636744947dcSTom Erickson dnode_t *dn; 2637744947dcSTom Erickson 26380a586ceaSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID) 2639be6fd75aSMatthew Ahrens return (SET_ERROR(ENOTSUP)); 26400a586ceaSMark Shellenbaum if (blksz == 0) 26410a586ceaSMark Shellenbaum blksz = SPA_MINBLOCKSIZE; 2642b5152584SMatthew Ahrens ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 26430a586ceaSMark Shellenbaum blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 26440a586ceaSMark Shellenbaum 2645744947dcSTom Erickson DB_DNODE_ENTER(db); 2646744947dcSTom Erickson dn = DB_DNODE(db); 2647744947dcSTom Erickson rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 26480a586ceaSMark Shellenbaum dbuf_new_size(db, blksz, tx); 2649744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 2650744947dcSTom Erickson DB_DNODE_EXIT(db); 26510a586ceaSMark Shellenbaum 26520a586ceaSMark Shellenbaum return (0); 26530a586ceaSMark Shellenbaum } 26540a586ceaSMark Shellenbaum 26550a586ceaSMark Shellenbaum void 26560a586ceaSMark Shellenbaum dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 26570a586ceaSMark Shellenbaum { 26580a586ceaSMark Shellenbaum dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2659fa9e4066Sahrens } 2660fa9e4066Sahrens 2661ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref 2662fa9e4066Sahrens void 2663fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2664fa9e4066Sahrens { 2665ea8dc4b6Seschrock int64_t holds = refcount_add(&db->db_holds, tag); 2666dcbf3bd6SGeorge Wilson ASSERT3S(holds, >, 1); 2667fa9e4066Sahrens } 2668fa9e4066Sahrens 2669e57a022bSJustin T. Gibbs #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2670e57a022bSJustin T. Gibbs boolean_t 2671e57a022bSJustin T. Gibbs dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2672e57a022bSJustin T. Gibbs void *tag) 2673e57a022bSJustin T. Gibbs { 2674e57a022bSJustin T. Gibbs dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2675e57a022bSJustin T. Gibbs dmu_buf_impl_t *found_db; 2676e57a022bSJustin T. Gibbs boolean_t result = B_FALSE; 2677e57a022bSJustin T. Gibbs 2678e57a022bSJustin T. Gibbs if (db->db_blkid == DMU_BONUS_BLKID) 2679e57a022bSJustin T. Gibbs found_db = dbuf_find_bonus(os, obj); 2680e57a022bSJustin T. Gibbs else 2681e57a022bSJustin T. Gibbs found_db = dbuf_find(os, obj, 0, blkid); 2682e57a022bSJustin T. Gibbs 2683e57a022bSJustin T. Gibbs if (found_db != NULL) { 2684e57a022bSJustin T. Gibbs if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2685e57a022bSJustin T. Gibbs (void) refcount_add(&db->db_holds, tag); 2686e57a022bSJustin T. Gibbs result = B_TRUE; 2687e57a022bSJustin T. Gibbs } 2688e57a022bSJustin T. Gibbs mutex_exit(&db->db_mtx); 2689e57a022bSJustin T. Gibbs } 2690e57a022bSJustin T. Gibbs return (result); 2691e57a022bSJustin T. Gibbs } 2692e57a022bSJustin T. Gibbs 2693744947dcSTom Erickson /* 2694744947dcSTom Erickson * If you call dbuf_rele() you had better not be referencing the dnode handle 2695744947dcSTom Erickson * unless you have some other direct or indirect hold on the dnode. (An indirect 2696744947dcSTom Erickson * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2697744947dcSTom Erickson * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2698744947dcSTom Erickson * dnode's parent dbuf evicting its dnode handles. 2699744947dcSTom Erickson */ 2700fa9e4066Sahrens void 2701ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag) 2702fa9e4066Sahrens { 2703b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 2704b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, tag); 2705b24ab676SJeff Bonwick } 2706b24ab676SJeff Bonwick 270743466aaeSMax Grossman void 270843466aaeSMax Grossman dmu_buf_rele(dmu_buf_t *db, void *tag) 270943466aaeSMax Grossman { 271043466aaeSMax Grossman dbuf_rele((dmu_buf_impl_t *)db, tag); 271143466aaeSMax Grossman } 271243466aaeSMax Grossman 2713b24ab676SJeff Bonwick /* 2714b24ab676SJeff Bonwick * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2715b24ab676SJeff Bonwick * db_dirtycnt and db_holds to be updated atomically. 2716b24ab676SJeff Bonwick */ 2717b24ab676SJeff Bonwick void 2718b24ab676SJeff Bonwick dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2719b24ab676SJeff Bonwick { 2720fa9e4066Sahrens int64_t holds; 2721fa9e4066Sahrens 2722b24ab676SJeff Bonwick ASSERT(MUTEX_HELD(&db->db_mtx)); 27239c9dc39aSek110237 DBUF_VERIFY(db); 2724fa9e4066Sahrens 2725744947dcSTom Erickson /* 2726744947dcSTom Erickson * Remove the reference to the dbuf before removing its hold on the 2727744947dcSTom Erickson * dnode so we can guarantee in dnode_move() that a referenced bonus 2728744947dcSTom Erickson * buffer has a corresponding dnode hold. 2729744947dcSTom Erickson */ 2730fa9e4066Sahrens holds = refcount_remove(&db->db_holds, tag); 2731ea8dc4b6Seschrock ASSERT(holds >= 0); 2732fa9e4066Sahrens 2733c717a561Smaybee /* 2734c717a561Smaybee * We can't freeze indirects if there is a possibility that they 2735c717a561Smaybee * may be modified in the current syncing context. 2736c717a561Smaybee */ 2737dcbf3bd6SGeorge Wilson if (db->db_buf != NULL && 2738dcbf3bd6SGeorge Wilson holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { 27396b4acc8bSahrens arc_buf_freeze(db->db_buf); 2740dcbf3bd6SGeorge Wilson } 27416b4acc8bSahrens 2742fa9e4066Sahrens if (holds == db->db_dirtycnt && 2743d2058105SJustin T. Gibbs db->db_level == 0 && db->db_user_immediate_evict) 2744fa9e4066Sahrens dbuf_evict_user(db); 2745ea8dc4b6Seschrock 2746ea8dc4b6Seschrock if (holds == 0) { 27470a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 2748cd485b49SJustin T. Gibbs dnode_t *dn; 2749d2058105SJustin T. Gibbs boolean_t evict_dbuf = db->db_pending_evict; 2750cd485b49SJustin T. Gibbs 2751cd485b49SJustin T. Gibbs /* 2752cd485b49SJustin T. Gibbs * If the dnode moves here, we cannot cross this 2753cd485b49SJustin T. Gibbs * barrier until the move completes. 2754cd485b49SJustin T. Gibbs */ 2755cd485b49SJustin T. Gibbs DB_DNODE_ENTER(db); 2756cd485b49SJustin T. Gibbs 2757cd485b49SJustin T. Gibbs dn = DB_DNODE(db); 2758cd485b49SJustin T. Gibbs atomic_dec_32(&dn->dn_dbufs_count); 2759cd485b49SJustin T. Gibbs 2760cd485b49SJustin T. Gibbs /* 2761cd485b49SJustin T. Gibbs * Decrementing the dbuf count means that the bonus 2762cd485b49SJustin T. Gibbs * buffer's dnode hold is no longer discounted in 2763cd485b49SJustin T. Gibbs * dnode_move(). The dnode cannot move until after 2764d2058105SJustin T. Gibbs * the dnode_rele() below. 2765cd485b49SJustin T. Gibbs */ 2766cd485b49SJustin T. Gibbs DB_DNODE_EXIT(db); 2767cd485b49SJustin T. Gibbs 2768cd485b49SJustin T. Gibbs /* 2769cd485b49SJustin T. Gibbs * Do not reference db after its lock is dropped. 2770cd485b49SJustin T. Gibbs * Another thread may evict it. 2771cd485b49SJustin T. Gibbs */ 2772ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 2773744947dcSTom Erickson 2774d2058105SJustin T. Gibbs if (evict_dbuf) 2775cd485b49SJustin T. Gibbs dnode_evict_bonus(dn); 2776d2058105SJustin T. Gibbs 2777d2058105SJustin T. Gibbs dnode_rele(dn, db); 2778ea8dc4b6Seschrock } else if (db->db_buf == NULL) { 2779ea8dc4b6Seschrock /* 2780ea8dc4b6Seschrock * This is a special case: we never associated this 2781ea8dc4b6Seschrock * dbuf with any data allocated from the ARC. 2782ea8dc4b6Seschrock */ 278382c9918fSTim Haley ASSERT(db->db_state == DB_UNCACHED || 278482c9918fSTim Haley db->db_state == DB_NOFILL); 2785dcbf3bd6SGeorge Wilson dbuf_destroy(db); 2786ea8dc4b6Seschrock } else if (arc_released(db->db_buf)) { 2787ea8dc4b6Seschrock /* 2788ea8dc4b6Seschrock * This dbuf has anonymous data associated with it. 2789ea8dc4b6Seschrock */ 2790dcbf3bd6SGeorge Wilson dbuf_destroy(db); 2791ea8dc4b6Seschrock } else { 2792dcbf3bd6SGeorge Wilson boolean_t do_arc_evict = B_FALSE; 2793dcbf3bd6SGeorge Wilson blkptr_t bp; 2794dcbf3bd6SGeorge Wilson spa_t *spa = dmu_objset_spa(db->db_objset); 27959253d63dSGeorge Wilson 2796dcbf3bd6SGeorge Wilson if (!DBUF_IS_CACHEABLE(db) && 2797dcbf3bd6SGeorge Wilson db->db_blkptr != NULL && 2798bbfa8ea8SMatthew Ahrens !BP_IS_HOLE(db->db_blkptr) && 2799bbfa8ea8SMatthew Ahrens !BP_IS_EMBEDDED(db->db_blkptr)) { 2800dcbf3bd6SGeorge Wilson do_arc_evict = B_TRUE; 2801dcbf3bd6SGeorge Wilson bp = *db->db_blkptr; 2802dcbf3bd6SGeorge Wilson } 2803dcbf3bd6SGeorge Wilson 2804dcbf3bd6SGeorge Wilson if (!DBUF_IS_CACHEABLE(db) || 2805dcbf3bd6SGeorge Wilson db->db_pending_evict) { 2806dcbf3bd6SGeorge Wilson dbuf_destroy(db); 2807dcbf3bd6SGeorge Wilson } else if (!multilist_link_active(&db->db_cache_link)) { 280894c2d0ebSMatthew Ahrens multilist_insert(dbuf_cache, db); 2809dcbf3bd6SGeorge Wilson (void) refcount_add_many(&dbuf_cache_size, 2810dcbf3bd6SGeorge Wilson db->db.db_size, db); 2811dcbf3bd6SGeorge Wilson mutex_exit(&db->db_mtx); 2812dcbf3bd6SGeorge Wilson 2813dcbf3bd6SGeorge Wilson dbuf_evict_notify(); 2814dcbf3bd6SGeorge Wilson } 2815dcbf3bd6SGeorge Wilson 2816dcbf3bd6SGeorge Wilson if (do_arc_evict) 2817bbfa8ea8SMatthew Ahrens arc_freed(spa, &bp); 2818bbfa8ea8SMatthew Ahrens } 2819ea8dc4b6Seschrock } else { 2820ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 2821fa9e4066Sahrens } 2822dcbf3bd6SGeorge Wilson 2823fa9e4066Sahrens } 2824fa9e4066Sahrens 2825fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount 2826fa9e4066Sahrens uint64_t 2827fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db) 2828fa9e4066Sahrens { 2829fa9e4066Sahrens return (refcount_count(&db->db_holds)); 2830fa9e4066Sahrens } 2831fa9e4066Sahrens 2832fa9e4066Sahrens void * 2833bc9014e6SJustin Gibbs dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2834bc9014e6SJustin Gibbs dmu_buf_user_t *new_user) 2835fa9e4066Sahrens { 2836bc9014e6SJustin Gibbs dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2837bc9014e6SJustin Gibbs 2838bc9014e6SJustin Gibbs mutex_enter(&db->db_mtx); 2839bc9014e6SJustin Gibbs dbuf_verify_user(db, DBVU_NOT_EVICTING); 2840bc9014e6SJustin Gibbs if (db->db_user == old_user) 2841bc9014e6SJustin Gibbs db->db_user = new_user; 2842bc9014e6SJustin Gibbs else 2843bc9014e6SJustin Gibbs old_user = db->db_user; 2844bc9014e6SJustin Gibbs dbuf_verify_user(db, DBVU_NOT_EVICTING); 2845bc9014e6SJustin Gibbs mutex_exit(&db->db_mtx); 2846bc9014e6SJustin Gibbs 2847bc9014e6SJustin Gibbs return (old_user); 2848fa9e4066Sahrens } 2849fa9e4066Sahrens 2850fa9e4066Sahrens void * 2851bc9014e6SJustin Gibbs dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2852bc9014e6SJustin Gibbs { 2853bc9014e6SJustin Gibbs return (dmu_buf_replace_user(db_fake, NULL, user)); 2854bc9014e6SJustin Gibbs } 2855bc9014e6SJustin Gibbs 2856bc9014e6SJustin Gibbs void * 2857bc9014e6SJustin Gibbs dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2858fa9e4066Sahrens { 2859fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2860fa9e4066Sahrens 2861d2058105SJustin T. Gibbs db->db_user_immediate_evict = TRUE; 2862bc9014e6SJustin Gibbs return (dmu_buf_set_user(db_fake, user)); 2863fa9e4066Sahrens } 2864fa9e4066Sahrens 2865fa9e4066Sahrens void * 2866bc9014e6SJustin Gibbs dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2867fa9e4066Sahrens { 2868bc9014e6SJustin Gibbs return (dmu_buf_replace_user(db_fake, user, NULL)); 2869fa9e4066Sahrens } 2870fa9e4066Sahrens 2871fa9e4066Sahrens void * 2872fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake) 2873fa9e4066Sahrens { 2874fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2875fa9e4066Sahrens 2876bc9014e6SJustin Gibbs dbuf_verify_user(db, DBVU_NOT_EVICTING); 2877bc9014e6SJustin Gibbs return (db->db_user); 2878bc9014e6SJustin Gibbs } 2879bc9014e6SJustin Gibbs 2880bc9014e6SJustin Gibbs void 2881bc9014e6SJustin Gibbs dmu_buf_user_evict_wait() 2882bc9014e6SJustin Gibbs { 2883bc9014e6SJustin Gibbs taskq_wait(dbu_evict_taskq); 2884fa9e4066Sahrens } 2885fa9e4066Sahrens 288680901aeaSGeorge Wilson blkptr_t * 288780901aeaSGeorge Wilson dmu_buf_get_blkptr(dmu_buf_t *db) 288880901aeaSGeorge Wilson { 288980901aeaSGeorge Wilson dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 289080901aeaSGeorge Wilson return (dbi->db_blkptr); 289180901aeaSGeorge Wilson } 289280901aeaSGeorge Wilson 2893ae972795SMatthew Ahrens objset_t * 2894ae972795SMatthew Ahrens dmu_buf_get_objset(dmu_buf_t *db) 2895ae972795SMatthew Ahrens { 2896ae972795SMatthew Ahrens dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2897ae972795SMatthew Ahrens return (dbi->db_objset); 2898ae972795SMatthew Ahrens } 2899ae972795SMatthew Ahrens 290079d72832SMatthew Ahrens dnode_t * 290179d72832SMatthew Ahrens dmu_buf_dnode_enter(dmu_buf_t *db) 290279d72832SMatthew Ahrens { 290379d72832SMatthew Ahrens dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 290479d72832SMatthew Ahrens DB_DNODE_ENTER(dbi); 290579d72832SMatthew Ahrens return (DB_DNODE(dbi)); 290679d72832SMatthew Ahrens } 290779d72832SMatthew Ahrens 290879d72832SMatthew Ahrens void 290979d72832SMatthew Ahrens dmu_buf_dnode_exit(dmu_buf_t *db) 291079d72832SMatthew Ahrens { 291179d72832SMatthew Ahrens dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 291279d72832SMatthew Ahrens DB_DNODE_EXIT(dbi); 291379d72832SMatthew Ahrens } 291479d72832SMatthew Ahrens 2915c717a561Smaybee static void 2916c717a561Smaybee dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2917fa9e4066Sahrens { 2918c717a561Smaybee /* ASSERT(dmu_tx_is_syncing(tx) */ 2919c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 2920c717a561Smaybee 2921c717a561Smaybee if (db->db_blkptr != NULL) 2922c717a561Smaybee return; 2923c717a561Smaybee 29240a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 29250a586ceaSMark Shellenbaum db->db_blkptr = &dn->dn_phys->dn_spill; 29260a586ceaSMark Shellenbaum BP_ZERO(db->db_blkptr); 29270a586ceaSMark Shellenbaum return; 29280a586ceaSMark Shellenbaum } 2929c717a561Smaybee if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2930c717a561Smaybee /* 2931c717a561Smaybee * This buffer was allocated at a time when there was 2932c717a561Smaybee * no available blkptrs from the dnode, or it was 2933c717a561Smaybee * inappropriate to hook it in (i.e., nlevels mis-match). 2934c717a561Smaybee */ 2935c717a561Smaybee ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2936c717a561Smaybee ASSERT(db->db_parent == NULL); 2937c717a561Smaybee db->db_parent = dn->dn_dbuf; 2938c717a561Smaybee db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2939c717a561Smaybee DBUF_VERIFY(db); 2940c717a561Smaybee } else { 2941c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 2942c717a561Smaybee int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2943c717a561Smaybee 2944c717a561Smaybee ASSERT(dn->dn_phys->dn_nlevels > 1); 2945c717a561Smaybee if (parent == NULL) { 2946c717a561Smaybee mutex_exit(&db->db_mtx); 2947c717a561Smaybee rw_enter(&dn->dn_struct_rwlock, RW_READER); 2948a2cdcdd2SPaul Dagnelie parent = dbuf_hold_level(dn, db->db_level + 1, 2949a2cdcdd2SPaul Dagnelie db->db_blkid >> epbs, db); 2950c717a561Smaybee rw_exit(&dn->dn_struct_rwlock); 2951c717a561Smaybee mutex_enter(&db->db_mtx); 2952c717a561Smaybee db->db_parent = parent; 2953c717a561Smaybee } 2954c717a561Smaybee db->db_blkptr = (blkptr_t *)parent->db.db_data + 2955c717a561Smaybee (db->db_blkid & ((1ULL << epbs) - 1)); 2956c717a561Smaybee DBUF_VERIFY(db); 2957c717a561Smaybee } 2958c717a561Smaybee } 2959c717a561Smaybee 2960c717a561Smaybee static void 2961c717a561Smaybee dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2962c717a561Smaybee { 2963c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 2964744947dcSTom Erickson dnode_t *dn; 2965c717a561Smaybee zio_t *zio; 2966c717a561Smaybee 2967c717a561Smaybee ASSERT(dmu_tx_is_syncing(tx)); 2968c717a561Smaybee 2969c717a561Smaybee dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2970c717a561Smaybee 2971c717a561Smaybee mutex_enter(&db->db_mtx); 2972c717a561Smaybee 2973c717a561Smaybee ASSERT(db->db_level > 0); 2974c717a561Smaybee DBUF_VERIFY(db); 2975c717a561Smaybee 29763e30c24aSWill Andrews /* Read the block if it hasn't been read yet. */ 2977c717a561Smaybee if (db->db_buf == NULL) { 2978c717a561Smaybee mutex_exit(&db->db_mtx); 2979c717a561Smaybee (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2980c717a561Smaybee mutex_enter(&db->db_mtx); 2981c717a561Smaybee } 2982c717a561Smaybee ASSERT3U(db->db_state, ==, DB_CACHED); 2983c717a561Smaybee ASSERT(db->db_buf != NULL); 2984c717a561Smaybee 2985744947dcSTom Erickson DB_DNODE_ENTER(db); 2986744947dcSTom Erickson dn = DB_DNODE(db); 29873e30c24aSWill Andrews /* Indirect block size must match what the dnode thinks it is. */ 2988744947dcSTom Erickson ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2989c717a561Smaybee dbuf_check_blkptr(dn, db); 2990744947dcSTom Erickson DB_DNODE_EXIT(db); 2991c717a561Smaybee 29923e30c24aSWill Andrews /* Provide the pending dirty record to child dbufs */ 2993c717a561Smaybee db->db_data_pending = dr; 2994c717a561Smaybee 2995af2c4821Smaybee mutex_exit(&db->db_mtx); 2996088f3894Sahrens dbuf_write(dr, db->db_buf, tx); 2997c717a561Smaybee 2998c717a561Smaybee zio = dr->dr_zio; 2999c717a561Smaybee mutex_enter(&dr->dt.di.dr_mtx); 300046e1baa6SMatthew Ahrens dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 3001c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3002c717a561Smaybee mutex_exit(&dr->dt.di.dr_mtx); 3003c717a561Smaybee zio_nowait(zio); 3004c717a561Smaybee } 3005c717a561Smaybee 3006c717a561Smaybee static void 3007c717a561Smaybee dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3008c717a561Smaybee { 3009c717a561Smaybee arc_buf_t **datap = &dr->dt.dl.dr_data; 3010c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 3011744947dcSTom Erickson dnode_t *dn; 3012744947dcSTom Erickson objset_t *os; 3013c717a561Smaybee uint64_t txg = tx->tx_txg; 3014fa9e4066Sahrens 3015fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 3016fa9e4066Sahrens 3017fa9e4066Sahrens dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3018fa9e4066Sahrens 3019fa9e4066Sahrens mutex_enter(&db->db_mtx); 3020fa9e4066Sahrens /* 3021fa9e4066Sahrens * To be synced, we must be dirtied. But we 3022fa9e4066Sahrens * might have been freed after the dirty. 3023fa9e4066Sahrens */ 3024fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 3025fa9e4066Sahrens /* This buffer has been freed since it was dirtied */ 3026fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 3027fa9e4066Sahrens } else if (db->db_state == DB_FILL) { 3028fa9e4066Sahrens /* This buffer was freed and is now being re-filled */ 3029c717a561Smaybee ASSERT(db->db.db_data != dr->dt.dl.dr_data); 3030fa9e4066Sahrens } else { 303182c9918fSTim Haley ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 3032fa9e4066Sahrens } 30339c9dc39aSek110237 DBUF_VERIFY(db); 3034fa9e4066Sahrens 3035744947dcSTom Erickson DB_DNODE_ENTER(db); 3036744947dcSTom Erickson dn = DB_DNODE(db); 3037744947dcSTom Erickson 30380a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 30390a586ceaSMark Shellenbaum mutex_enter(&dn->dn_mtx); 30400a586ceaSMark Shellenbaum dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 30410a586ceaSMark Shellenbaum mutex_exit(&dn->dn_mtx); 30420a586ceaSMark Shellenbaum } 30430a586ceaSMark Shellenbaum 3044fa9e4066Sahrens /* 3045c717a561Smaybee * If this is a bonus buffer, simply copy the bonus data into the 3046c717a561Smaybee * dnode. It will be written out when the dnode is synced (and it 3047c717a561Smaybee * will be synced, since it must have been dirty for dbuf_sync to 3048c717a561Smaybee * be called). 3049fa9e4066Sahrens */ 30500a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 3051c717a561Smaybee dbuf_dirty_record_t **drp; 30521934e92fSmaybee 3053ea8dc4b6Seschrock ASSERT(*datap != NULL); 3054fb09f5aaSMadhav Suresh ASSERT0(db->db_level); 3055ea8dc4b6Seschrock ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 3056ea8dc4b6Seschrock bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 3057744947dcSTom Erickson DB_DNODE_EXIT(db); 3058744947dcSTom Erickson 30590e8c6158Smaybee if (*datap != db->db.db_data) { 3060ea8dc4b6Seschrock zio_buf_free(*datap, DN_MAX_BONUSLEN); 30615a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 30620e8c6158Smaybee } 3063ea8dc4b6Seschrock db->db_data_pending = NULL; 3064c717a561Smaybee drp = &db->db_last_dirty; 3065c717a561Smaybee while (*drp != dr) 3066c717a561Smaybee drp = &(*drp)->dr_next; 306717f17c2dSbonwick ASSERT(dr->dr_next == NULL); 3068b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 306917f17c2dSbonwick *drp = dr->dr_next; 3070c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3071ea8dc4b6Seschrock ASSERT(db->db_dirtycnt > 0); 3072ea8dc4b6Seschrock db->db_dirtycnt -= 1; 3073b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 3074ea8dc4b6Seschrock return; 3075ea8dc4b6Seschrock } 3076ea8dc4b6Seschrock 3077744947dcSTom Erickson os = dn->dn_objset; 3078744947dcSTom Erickson 3079c5c6ffa0Smaybee /* 3080f82bfe17Sgw25295 * This function may have dropped the db_mtx lock allowing a dmu_sync 3081f82bfe17Sgw25295 * operation to sneak in. As a result, we need to ensure that we 3082f82bfe17Sgw25295 * don't check the dr_override_state until we have returned from 3083f82bfe17Sgw25295 * dbuf_check_blkptr. 3084f82bfe17Sgw25295 */ 3085f82bfe17Sgw25295 dbuf_check_blkptr(dn, db); 3086f82bfe17Sgw25295 3087f82bfe17Sgw25295 /* 3088744947dcSTom Erickson * If this buffer is in the middle of an immediate write, 3089c717a561Smaybee * wait for the synchronous IO to complete. 3090c5c6ffa0Smaybee */ 3091c717a561Smaybee while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 3092c5c6ffa0Smaybee ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 3093c5c6ffa0Smaybee cv_wait(&db->db_changed, &db->db_mtx); 3094c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 3095c5c6ffa0Smaybee } 3096c717a561Smaybee 3097ab69d62fSMatthew Ahrens if (db->db_state != DB_NOFILL && 3098ab69d62fSMatthew Ahrens dn->dn_object != DMU_META_DNODE_OBJECT && 3099ab69d62fSMatthew Ahrens refcount_count(&db->db_holds) > 1 && 3100b24ab676SJeff Bonwick dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 3101ab69d62fSMatthew Ahrens *datap == db->db_buf) { 3102fa9e4066Sahrens /* 310382c9918fSTim Haley * If this buffer is currently "in use" (i.e., there 310482c9918fSTim Haley * are active holds and db_data still references it), 310582c9918fSTim Haley * then make a copy before we start the write so that 310682c9918fSTim Haley * any modifications from the open txg will not leak 310782c9918fSTim Haley * into this write. 3108fa9e4066Sahrens * 310982c9918fSTim Haley * NOTE: this copy does not need to be made for 311082c9918fSTim Haley * objects only modified in the syncing context (e.g. 311182c9918fSTim Haley * DNONE_DNODE blocks). 3112fa9e4066Sahrens */ 31135602294fSDan Kimmel int psize = arc_buf_size(*datap); 3114ab69d62fSMatthew Ahrens arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 31155602294fSDan Kimmel enum zio_compress compress_type = arc_get_compression(*datap); 31165602294fSDan Kimmel 31175602294fSDan Kimmel if (compress_type == ZIO_COMPRESS_OFF) { 31185602294fSDan Kimmel *datap = arc_alloc_buf(os->os_spa, db, type, psize); 31195602294fSDan Kimmel } else { 31205602294fSDan Kimmel ASSERT3U(type, ==, ARC_BUFC_DATA); 31215602294fSDan Kimmel int lsize = arc_buf_lsize(*datap); 31225602294fSDan Kimmel *datap = arc_alloc_compressed_buf(os->os_spa, db, 31235602294fSDan Kimmel psize, lsize, compress_type); 31245602294fSDan Kimmel } 31255602294fSDan Kimmel bcopy(db->db.db_data, (*datap)->b_data, psize); 3126fa9e4066Sahrens } 3127c717a561Smaybee db->db_data_pending = dr; 3128fa9e4066Sahrens 3129fa9e4066Sahrens mutex_exit(&db->db_mtx); 3130fa9e4066Sahrens 3131088f3894Sahrens dbuf_write(dr, *datap, tx); 3132c717a561Smaybee 3133c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 3134744947dcSTom Erickson if (dn->dn_object == DMU_META_DNODE_OBJECT) { 3135c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 3136744947dcSTom Erickson DB_DNODE_EXIT(db); 3137744947dcSTom Erickson } else { 3138744947dcSTom Erickson /* 3139744947dcSTom Erickson * Although zio_nowait() does not "wait for an IO", it does 3140744947dcSTom Erickson * initiate the IO. If this is an empty write it seems plausible 3141744947dcSTom Erickson * that the IO could actually be completed before the nowait 3142744947dcSTom Erickson * returns. We need to DB_DNODE_EXIT() first in case 3143744947dcSTom Erickson * zio_nowait() invalidates the dbuf. 3144744947dcSTom Erickson */ 3145744947dcSTom Erickson DB_DNODE_EXIT(db); 3146c717a561Smaybee zio_nowait(dr->dr_zio); 3147fa9e4066Sahrens } 3148744947dcSTom Erickson } 3149c717a561Smaybee 3150c717a561Smaybee void 315146e1baa6SMatthew Ahrens dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 3152c717a561Smaybee { 3153c717a561Smaybee dbuf_dirty_record_t *dr; 3154c717a561Smaybee 3155c717a561Smaybee while (dr = list_head(list)) { 3156c717a561Smaybee if (dr->dr_zio != NULL) { 3157c717a561Smaybee /* 3158c717a561Smaybee * If we find an already initialized zio then we 3159c717a561Smaybee * are processing the meta-dnode, and we have finished. 3160c717a561Smaybee * The dbufs for all dnodes are put back on the list 3161c717a561Smaybee * during processing, so that we can zio_wait() 3162c717a561Smaybee * these IOs after initiating all child IOs. 3163c717a561Smaybee */ 3164c717a561Smaybee ASSERT3U(dr->dr_dbuf->db.db_object, ==, 3165c717a561Smaybee DMU_META_DNODE_OBJECT); 3166c717a561Smaybee break; 3167fa9e4066Sahrens } 316846e1baa6SMatthew Ahrens if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 316946e1baa6SMatthew Ahrens dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 317046e1baa6SMatthew Ahrens VERIFY3U(dr->dr_dbuf->db_level, ==, level); 317146e1baa6SMatthew Ahrens } 3172c717a561Smaybee list_remove(list, dr); 3173c717a561Smaybee if (dr->dr_dbuf->db_level > 0) 3174c717a561Smaybee dbuf_sync_indirect(dr, tx); 3175c717a561Smaybee else 3176c717a561Smaybee dbuf_sync_leaf(dr, tx); 3177c717a561Smaybee } 3178c717a561Smaybee } 3179c717a561Smaybee 3180fa9e4066Sahrens /* ARGSUSED */ 3181fa9e4066Sahrens static void 3182c717a561Smaybee dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3183fa9e4066Sahrens { 3184fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 3185744947dcSTom Erickson dnode_t *dn; 3186e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 3187c717a561Smaybee blkptr_t *bp_orig = &zio->io_bp_orig; 3188b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 3189b24ab676SJeff Bonwick int64_t delta; 3190fa9e4066Sahrens uint64_t fill = 0; 3191b24ab676SJeff Bonwick int i; 3192fa9e4066Sahrens 319311ceac77SAlex Reece ASSERT3P(db->db_blkptr, !=, NULL); 319411ceac77SAlex Reece ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); 3195e14bb325SJeff Bonwick 3196744947dcSTom Erickson DB_DNODE_ENTER(db); 3197744947dcSTom Erickson dn = DB_DNODE(db); 3198b24ab676SJeff Bonwick delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 3199b24ab676SJeff Bonwick dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 3200b24ab676SJeff Bonwick zio->io_prev_space_delta = delta; 3201fa9e4066Sahrens 320243466aaeSMax Grossman if (bp->blk_birth != 0) { 32030a586ceaSMark Shellenbaum ASSERT((db->db_blkid != DMU_SPILL_BLKID && 32040a586ceaSMark Shellenbaum BP_GET_TYPE(bp) == dn->dn_type) || 32050a586ceaSMark Shellenbaum (db->db_blkid == DMU_SPILL_BLKID && 32065d7b4d43SMatthew Ahrens BP_GET_TYPE(bp) == dn->dn_bonustype) || 32075d7b4d43SMatthew Ahrens BP_IS_EMBEDDED(bp)); 3208e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(bp) == db->db_level); 320943466aaeSMax Grossman } 3210e14bb325SJeff Bonwick 3211fa9e4066Sahrens mutex_enter(&db->db_mtx); 3212fa9e4066Sahrens 32130a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG 32140a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 32150a586ceaSMark Shellenbaum ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 321611ceac77SAlex Reece ASSERT(!(BP_IS_HOLE(bp)) && 32170a586ceaSMark Shellenbaum db->db_blkptr == &dn->dn_phys->dn_spill); 32180a586ceaSMark Shellenbaum } 32190a586ceaSMark Shellenbaum #endif 32200a586ceaSMark Shellenbaum 3221fa9e4066Sahrens if (db->db_level == 0) { 3222fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 32230a586ceaSMark Shellenbaum if (db->db_blkid > dn->dn_phys->dn_maxblkid && 32240a586ceaSMark Shellenbaum db->db_blkid != DMU_SPILL_BLKID) 3225fa9e4066Sahrens dn->dn_phys->dn_maxblkid = db->db_blkid; 3226fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 3227fa9e4066Sahrens 3228fa9e4066Sahrens if (dn->dn_type == DMU_OT_DNODE) { 3229fa9e4066Sahrens dnode_phys_t *dnp = db->db.db_data; 3230fa9e4066Sahrens for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 3231fa9e4066Sahrens i--, dnp++) { 3232fa9e4066Sahrens if (dnp->dn_type != DMU_OT_NONE) 3233fa9e4066Sahrens fill++; 3234fa9e4066Sahrens } 3235fa9e4066Sahrens } else { 323643466aaeSMax Grossman if (BP_IS_HOLE(bp)) { 323743466aaeSMax Grossman fill = 0; 323843466aaeSMax Grossman } else { 3239fa9e4066Sahrens fill = 1; 3240fa9e4066Sahrens } 324143466aaeSMax Grossman } 3242fa9e4066Sahrens } else { 3243e14bb325SJeff Bonwick blkptr_t *ibp = db->db.db_data; 3244fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3245e14bb325SJeff Bonwick for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 3246e14bb325SJeff Bonwick if (BP_IS_HOLE(ibp)) 3247fa9e4066Sahrens continue; 32485d7b4d43SMatthew Ahrens fill += BP_GET_FILL(ibp); 3249fa9e4066Sahrens } 3250fa9e4066Sahrens } 3251744947dcSTom Erickson DB_DNODE_EXIT(db); 3252fa9e4066Sahrens 32535d7b4d43SMatthew Ahrens if (!BP_IS_EMBEDDED(bp)) 3254e14bb325SJeff Bonwick bp->blk_fill = fill; 3255fa9e4066Sahrens 3256fa9e4066Sahrens mutex_exit(&db->db_mtx); 325711ceac77SAlex Reece 325811ceac77SAlex Reece rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 325911ceac77SAlex Reece *db->db_blkptr = *bp; 326011ceac77SAlex Reece rw_exit(&dn->dn_struct_rwlock); 3261fa9e4066Sahrens } 3262fa9e4066Sahrens 32638df0bcf0SPaul Dagnelie /* ARGSUSED */ 32648df0bcf0SPaul Dagnelie /* 32658df0bcf0SPaul Dagnelie * This function gets called just prior to running through the compression 32668df0bcf0SPaul Dagnelie * stage of the zio pipeline. If we're an indirect block comprised of only 32678df0bcf0SPaul Dagnelie * holes, then we want this indirect to be compressed away to a hole. In 32688df0bcf0SPaul Dagnelie * order to do that we must zero out any information about the holes that 32698df0bcf0SPaul Dagnelie * this indirect points to prior to before we try to compress it. 32708df0bcf0SPaul Dagnelie */ 32718df0bcf0SPaul Dagnelie static void 32728df0bcf0SPaul Dagnelie dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 32738df0bcf0SPaul Dagnelie { 32748df0bcf0SPaul Dagnelie dmu_buf_impl_t *db = vdb; 32758df0bcf0SPaul Dagnelie dnode_t *dn; 32768df0bcf0SPaul Dagnelie blkptr_t *bp; 32771a01181fSGeorge Wilson unsigned int epbs, i; 32788df0bcf0SPaul Dagnelie 32798df0bcf0SPaul Dagnelie ASSERT3U(db->db_level, >, 0); 32808df0bcf0SPaul Dagnelie DB_DNODE_ENTER(db); 32818df0bcf0SPaul Dagnelie dn = DB_DNODE(db); 32828df0bcf0SPaul Dagnelie epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 32831a01181fSGeorge Wilson ASSERT3U(epbs, <, 31); 32848df0bcf0SPaul Dagnelie 32858df0bcf0SPaul Dagnelie /* Determine if all our children are holes */ 32868df0bcf0SPaul Dagnelie for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { 32878df0bcf0SPaul Dagnelie if (!BP_IS_HOLE(bp)) 32888df0bcf0SPaul Dagnelie break; 32898df0bcf0SPaul Dagnelie } 32908df0bcf0SPaul Dagnelie 32918df0bcf0SPaul Dagnelie /* 32928df0bcf0SPaul Dagnelie * If all the children are holes, then zero them all out so that 32938df0bcf0SPaul Dagnelie * we may get compressed away. 32948df0bcf0SPaul Dagnelie */ 32958df0bcf0SPaul Dagnelie if (i == 1 << epbs) { 32961a01181fSGeorge Wilson /* 32971a01181fSGeorge Wilson * We only found holes. Grab the rwlock to prevent 32981a01181fSGeorge Wilson * anybody from reading the blocks we're about to 32991a01181fSGeorge Wilson * zero out. 33001a01181fSGeorge Wilson */ 33011a01181fSGeorge Wilson rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 33028df0bcf0SPaul Dagnelie bzero(db->db.db_data, db->db.db_size); 33031a01181fSGeorge Wilson rw_exit(&dn->dn_struct_rwlock); 33048df0bcf0SPaul Dagnelie } 33058df0bcf0SPaul Dagnelie DB_DNODE_EXIT(db); 33068df0bcf0SPaul Dagnelie } 33078df0bcf0SPaul Dagnelie 330869962b56SMatthew Ahrens /* 330969962b56SMatthew Ahrens * The SPA will call this callback several times for each zio - once 331069962b56SMatthew Ahrens * for every physical child i/o (zio->io_phys_children times). This 331169962b56SMatthew Ahrens * allows the DMU to monitor the progress of each logical i/o. For example, 331269962b56SMatthew Ahrens * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 331369962b56SMatthew Ahrens * block. There may be a long delay before all copies/fragments are completed, 331469962b56SMatthew Ahrens * so this callback allows us to retire dirty space gradually, as the physical 331569962b56SMatthew Ahrens * i/os complete. 331669962b56SMatthew Ahrens */ 331769962b56SMatthew Ahrens /* ARGSUSED */ 331869962b56SMatthew Ahrens static void 331969962b56SMatthew Ahrens dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 332069962b56SMatthew Ahrens { 332169962b56SMatthew Ahrens dmu_buf_impl_t *db = arg; 332269962b56SMatthew Ahrens objset_t *os = db->db_objset; 332369962b56SMatthew Ahrens dsl_pool_t *dp = dmu_objset_pool(os); 332469962b56SMatthew Ahrens dbuf_dirty_record_t *dr; 332569962b56SMatthew Ahrens int delta = 0; 332669962b56SMatthew Ahrens 332769962b56SMatthew Ahrens dr = db->db_data_pending; 332869962b56SMatthew Ahrens ASSERT3U(dr->dr_txg, ==, zio->io_txg); 332969962b56SMatthew Ahrens 333069962b56SMatthew Ahrens /* 333169962b56SMatthew Ahrens * The callback will be called io_phys_children times. Retire one 333269962b56SMatthew Ahrens * portion of our dirty space each time we are called. Any rounding 333369962b56SMatthew Ahrens * error will be cleaned up by dsl_pool_sync()'s call to 333469962b56SMatthew Ahrens * dsl_pool_undirty_space(). 333569962b56SMatthew Ahrens */ 333669962b56SMatthew Ahrens delta = dr->dr_accounted / zio->io_phys_children; 333769962b56SMatthew Ahrens dsl_pool_undirty_space(dp, delta, zio->io_txg); 333869962b56SMatthew Ahrens } 333969962b56SMatthew Ahrens 3340c717a561Smaybee /* ARGSUSED */ 3341c717a561Smaybee static void 3342c717a561Smaybee dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3343c717a561Smaybee { 3344c717a561Smaybee dmu_buf_impl_t *db = vdb; 3345b24ab676SJeff Bonwick blkptr_t *bp_orig = &zio->io_bp_orig; 334643466aaeSMax Grossman blkptr_t *bp = db->db_blkptr; 334743466aaeSMax Grossman objset_t *os = db->db_objset; 334843466aaeSMax Grossman dmu_tx_t *tx = os->os_synctx; 3349c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 3350c717a561Smaybee 3351fb09f5aaSMadhav Suresh ASSERT0(zio->io_error); 3352b24ab676SJeff Bonwick ASSERT(db->db_blkptr == bp); 3353b24ab676SJeff Bonwick 335480901aeaSGeorge Wilson /* 335580901aeaSGeorge Wilson * For nopwrites and rewrites we ensure that the bp matches our 335680901aeaSGeorge Wilson * original and bypass all the accounting. 335780901aeaSGeorge Wilson */ 335880901aeaSGeorge Wilson if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3359b24ab676SJeff Bonwick ASSERT(BP_EQUAL(bp, bp_orig)); 3360b24ab676SJeff Bonwick } else { 336143466aaeSMax Grossman dsl_dataset_t *ds = os->os_dsl_dataset; 3362b24ab676SJeff Bonwick (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3363b24ab676SJeff Bonwick dsl_dataset_block_born(ds, bp, tx); 3364b24ab676SJeff Bonwick } 3365c717a561Smaybee 3366c717a561Smaybee mutex_enter(&db->db_mtx); 3367c717a561Smaybee 3368b24ab676SJeff Bonwick DBUF_VERIFY(db); 3369b24ab676SJeff Bonwick 3370c717a561Smaybee drp = &db->db_last_dirty; 337117f17c2dSbonwick while ((dr = *drp) != db->db_data_pending) 337217f17c2dSbonwick drp = &dr->dr_next; 337317f17c2dSbonwick ASSERT(!list_link_active(&dr->dr_dirty_node)); 3374b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 337517f17c2dSbonwick ASSERT(dr->dr_next == NULL); 337617f17c2dSbonwick *drp = dr->dr_next; 3377c717a561Smaybee 33780a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG 33790a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 3380744947dcSTom Erickson dnode_t *dn; 3381744947dcSTom Erickson 3382744947dcSTom Erickson DB_DNODE_ENTER(db); 3383744947dcSTom Erickson dn = DB_DNODE(db); 33840a586ceaSMark Shellenbaum ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 33850a586ceaSMark Shellenbaum ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 33860a586ceaSMark Shellenbaum db->db_blkptr == &dn->dn_phys->dn_spill); 3387744947dcSTom Erickson DB_DNODE_EXIT(db); 33880a586ceaSMark Shellenbaum } 33890a586ceaSMark Shellenbaum #endif 33900a586ceaSMark Shellenbaum 3391c717a561Smaybee if (db->db_level == 0) { 33920a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3393c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 339482c9918fSTim Haley if (db->db_state != DB_NOFILL) { 3395c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 3396dcbf3bd6SGeorge Wilson arc_buf_destroy(dr->dt.dl.dr_data, db); 339782c9918fSTim Haley } 3398c717a561Smaybee } else { 3399744947dcSTom Erickson dnode_t *dn; 3400744947dcSTom Erickson 3401744947dcSTom Erickson DB_DNODE_ENTER(db); 3402744947dcSTom Erickson dn = DB_DNODE(db); 3403c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3404c717a561Smaybee ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3405c717a561Smaybee if (!BP_IS_HOLE(db->db_blkptr)) { 3406c717a561Smaybee int epbs = 3407c717a561Smaybee dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 340843466aaeSMax Grossman ASSERT3U(db->db_blkid, <=, 340943466aaeSMax Grossman dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3410c717a561Smaybee ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3411c717a561Smaybee db->db.db_size); 3412c717a561Smaybee } 3413744947dcSTom Erickson DB_DNODE_EXIT(db); 3414c25056deSgw25295 mutex_destroy(&dr->dt.di.dr_mtx); 3415c25056deSgw25295 list_destroy(&dr->dt.di.dr_children); 3416c717a561Smaybee } 3417c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3418c717a561Smaybee 3419c717a561Smaybee cv_broadcast(&db->db_changed); 3420c717a561Smaybee ASSERT(db->db_dirtycnt > 0); 3421c717a561Smaybee db->db_dirtycnt -= 1; 3422c717a561Smaybee db->db_data_pending = NULL; 342343466aaeSMax Grossman dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3424b24ab676SJeff Bonwick } 3425b24ab676SJeff Bonwick 3426b24ab676SJeff Bonwick static void 3427b24ab676SJeff Bonwick dbuf_write_nofill_ready(zio_t *zio) 3428b24ab676SJeff Bonwick { 3429b24ab676SJeff Bonwick dbuf_write_ready(zio, NULL, zio->io_private); 3430b24ab676SJeff Bonwick } 3431b24ab676SJeff Bonwick 3432b24ab676SJeff Bonwick static void 3433b24ab676SJeff Bonwick dbuf_write_nofill_done(zio_t *zio) 3434b24ab676SJeff Bonwick { 3435b24ab676SJeff Bonwick dbuf_write_done(zio, NULL, zio->io_private); 3436b24ab676SJeff Bonwick } 3437b24ab676SJeff Bonwick 3438b24ab676SJeff Bonwick static void 3439b24ab676SJeff Bonwick dbuf_write_override_ready(zio_t *zio) 3440b24ab676SJeff Bonwick { 3441b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = zio->io_private; 3442b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 3443b24ab676SJeff Bonwick 3444b24ab676SJeff Bonwick dbuf_write_ready(zio, NULL, db); 3445b24ab676SJeff Bonwick } 3446b24ab676SJeff Bonwick 3447b24ab676SJeff Bonwick static void 3448b24ab676SJeff Bonwick dbuf_write_override_done(zio_t *zio) 3449b24ab676SJeff Bonwick { 3450b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = zio->io_private; 3451b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 3452b24ab676SJeff Bonwick blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3453b24ab676SJeff Bonwick 3454b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 3455b24ab676SJeff Bonwick if (!BP_EQUAL(zio->io_bp, obp)) { 3456b24ab676SJeff Bonwick if (!BP_IS_HOLE(obp)) 3457b24ab676SJeff Bonwick dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3458b24ab676SJeff Bonwick arc_release(dr->dt.dl.dr_data, db); 3459b24ab676SJeff Bonwick } 3460c717a561Smaybee mutex_exit(&db->db_mtx); 34614ee0199eSRobert Mustacchi dbuf_write_done(zio, NULL, db); 3462770499e1SDan Kimmel 3463770499e1SDan Kimmel if (zio->io_abd != NULL) 3464770499e1SDan Kimmel abd_put(zio->io_abd); 3465b24ab676SJeff Bonwick } 3466c717a561Smaybee 34673e30c24aSWill Andrews /* Issue I/O to commit a dirty buffer to disk. */ 3468b24ab676SJeff Bonwick static void 3469b24ab676SJeff Bonwick dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3470b24ab676SJeff Bonwick { 3471b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 3472744947dcSTom Erickson dnode_t *dn; 3473744947dcSTom Erickson objset_t *os; 3474b24ab676SJeff Bonwick dmu_buf_impl_t *parent = db->db_parent; 3475b24ab676SJeff Bonwick uint64_t txg = tx->tx_txg; 34767802d7bfSMatthew Ahrens zbookmark_phys_t zb; 3477b24ab676SJeff Bonwick zio_prop_t zp; 3478b24ab676SJeff Bonwick zio_t *zio; 34790a586ceaSMark Shellenbaum int wp_flag = 0; 3480b24ab676SJeff Bonwick 348111ceac77SAlex Reece ASSERT(dmu_tx_is_syncing(tx)); 348211ceac77SAlex Reece 3483744947dcSTom Erickson DB_DNODE_ENTER(db); 3484744947dcSTom Erickson dn = DB_DNODE(db); 3485744947dcSTom Erickson os = dn->dn_objset; 3486744947dcSTom Erickson 3487b24ab676SJeff Bonwick if (db->db_state != DB_NOFILL) { 3488b24ab676SJeff Bonwick if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3489b24ab676SJeff Bonwick /* 3490b24ab676SJeff Bonwick * Private object buffers are released here rather 3491b24ab676SJeff Bonwick * than in dbuf_dirty() since they are only modified 3492b24ab676SJeff Bonwick * in the syncing context and we don't want the 3493b24ab676SJeff Bonwick * overhead of making multiple copies of the data. 3494b24ab676SJeff Bonwick */ 3495b24ab676SJeff Bonwick if (BP_IS_HOLE(db->db_blkptr)) { 3496b24ab676SJeff Bonwick arc_buf_thaw(data); 3497b24ab676SJeff Bonwick } else { 34983f9d6ad7SLin Ling dbuf_release_bp(db); 3499b24ab676SJeff Bonwick } 3500b24ab676SJeff Bonwick } 3501b24ab676SJeff Bonwick } 3502b24ab676SJeff Bonwick 3503b24ab676SJeff Bonwick if (parent != dn->dn_dbuf) { 35043e30c24aSWill Andrews /* Our parent is an indirect block. */ 35053e30c24aSWill Andrews /* We have a dirty parent that has been scheduled for write. */ 3506b24ab676SJeff Bonwick ASSERT(parent && parent->db_data_pending); 35073e30c24aSWill Andrews /* Our parent's buffer is one level closer to the dnode. */ 3508b24ab676SJeff Bonwick ASSERT(db->db_level == parent->db_level-1); 35093e30c24aSWill Andrews /* 35103e30c24aSWill Andrews * We're about to modify our parent's db_data by modifying 35113e30c24aSWill Andrews * our block pointer, so the parent must be released. 35123e30c24aSWill Andrews */ 3513b24ab676SJeff Bonwick ASSERT(arc_released(parent->db_buf)); 3514b24ab676SJeff Bonwick zio = parent->db_data_pending->dr_zio; 3515b24ab676SJeff Bonwick } else { 35163e30c24aSWill Andrews /* Our parent is the dnode itself. */ 35170a586ceaSMark Shellenbaum ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 35180a586ceaSMark Shellenbaum db->db_blkid != DMU_SPILL_BLKID) || 35190a586ceaSMark Shellenbaum (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 35200a586ceaSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID) 3521b24ab676SJeff Bonwick ASSERT3P(db->db_blkptr, ==, 3522b24ab676SJeff Bonwick &dn->dn_phys->dn_blkptr[db->db_blkid]); 3523b24ab676SJeff Bonwick zio = dn->dn_zio; 3524b24ab676SJeff Bonwick } 3525b24ab676SJeff Bonwick 3526b24ab676SJeff Bonwick ASSERT(db->db_level == 0 || data == db->db_buf); 3527b24ab676SJeff Bonwick ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3528b24ab676SJeff Bonwick ASSERT(zio); 3529b24ab676SJeff Bonwick 3530b24ab676SJeff Bonwick SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3531b24ab676SJeff Bonwick os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3532b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 3533b24ab676SJeff Bonwick 35340a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) 35350a586ceaSMark Shellenbaum wp_flag = WP_SPILL; 35360a586ceaSMark Shellenbaum wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 35370a586ceaSMark Shellenbaum 3538*adaec86aSMatthew Ahrens dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3539744947dcSTom Erickson DB_DNODE_EXIT(db); 3540b24ab676SJeff Bonwick 354111ceac77SAlex Reece /* 354211ceac77SAlex Reece * We copy the blkptr now (rather than when we instantiate the dirty 354311ceac77SAlex Reece * record), because its value can change between open context and 354411ceac77SAlex Reece * syncing context. We do not need to hold dn_struct_rwlock to read 354511ceac77SAlex Reece * db_blkptr because we are in syncing context. 354611ceac77SAlex Reece */ 354711ceac77SAlex Reece dr->dr_bp_copy = *db->db_blkptr; 354811ceac77SAlex Reece 35495d7b4d43SMatthew Ahrens if (db->db_level == 0 && 35505d7b4d43SMatthew Ahrens dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 35515d7b4d43SMatthew Ahrens /* 35525d7b4d43SMatthew Ahrens * The BP for this block has been provided by open context 35535d7b4d43SMatthew Ahrens * (by dmu_sync() or dmu_buf_write_embedded()). 35545d7b4d43SMatthew Ahrens */ 3555770499e1SDan Kimmel abd_t *contents = (data != NULL) ? 3556770499e1SDan Kimmel abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; 35575d7b4d43SMatthew Ahrens 35585602294fSDan Kimmel dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy, 35595602294fSDan Kimmel contents, db->db.db_size, db->db.db_size, &zp, 35608df0bcf0SPaul Dagnelie dbuf_write_override_ready, NULL, NULL, 35618df0bcf0SPaul Dagnelie dbuf_write_override_done, 356269962b56SMatthew Ahrens dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3563b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 3564b24ab676SJeff Bonwick dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3565b24ab676SJeff Bonwick zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 356680901aeaSGeorge Wilson dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3567b24ab676SJeff Bonwick mutex_exit(&db->db_mtx); 3568b24ab676SJeff Bonwick } else if (db->db_state == DB_NOFILL) { 3569810e43b2SBill Pijewski ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3570810e43b2SBill Pijewski zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3571b24ab676SJeff Bonwick dr->dr_zio = zio_write(zio, os->os_spa, txg, 35725602294fSDan Kimmel &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, 35738df0bcf0SPaul Dagnelie dbuf_write_nofill_ready, NULL, NULL, 35748df0bcf0SPaul Dagnelie dbuf_write_nofill_done, db, 3575b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, 3576b24ab676SJeff Bonwick ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3577b24ab676SJeff Bonwick } else { 3578b24ab676SJeff Bonwick ASSERT(arc_released(data)); 35798df0bcf0SPaul Dagnelie 35808df0bcf0SPaul Dagnelie /* 35818df0bcf0SPaul Dagnelie * For indirect blocks, we want to setup the children 35828df0bcf0SPaul Dagnelie * ready callback so that we can properly handle an indirect 35838df0bcf0SPaul Dagnelie * block that only contains holes. 35848df0bcf0SPaul Dagnelie */ 35858df0bcf0SPaul Dagnelie arc_done_func_t *children_ready_cb = NULL; 35868df0bcf0SPaul Dagnelie if (db->db_level != 0) 35878df0bcf0SPaul Dagnelie children_ready_cb = dbuf_write_children_ready; 35888df0bcf0SPaul Dagnelie 3589b24ab676SJeff Bonwick dr->dr_zio = arc_write(zio, os->os_spa, txg, 359011ceac77SAlex Reece &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), 3591dcbf3bd6SGeorge Wilson &zp, dbuf_write_ready, children_ready_cb, 359269962b56SMatthew Ahrens dbuf_write_physdone, dbuf_write_done, db, 359369962b56SMatthew Ahrens ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3594b24ab676SJeff Bonwick } 3595fa9e4066Sahrens } 3596