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 /* 2234f18512Seschrock * Copyright 2006 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 26fa9e4066Sahrens #pragma ident "%Z%%M% %I% %E% SMI" 27fa9e4066Sahrens 28fa9e4066Sahrens #include <sys/zfs_context.h> 29fa9e4066Sahrens #include <sys/dmu.h> 30fa9e4066Sahrens #include <sys/dmu_impl.h> 31fa9e4066Sahrens #include <sys/dbuf.h> 32fa9e4066Sahrens #include <sys/dmu_objset.h> 33fa9e4066Sahrens #include <sys/dsl_dataset.h> 34fa9e4066Sahrens #include <sys/dsl_dir.h> 35fa9e4066Sahrens #include <sys/dmu_tx.h> 36fa9e4066Sahrens #include <sys/spa.h> 37fa9e4066Sahrens #include <sys/zio.h> 38fa9e4066Sahrens #include <sys/dmu_zfetch.h> 39fa9e4066Sahrens 40fa9e4066Sahrens static void dbuf_destroy(dmu_buf_impl_t *db); 41fa9e4066Sahrens static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 42fa9e4066Sahrens static arc_done_func_t dbuf_write_done; 43fa9e4066Sahrens 44fa9e4066Sahrens /* 45fa9e4066Sahrens * Global data structures and functions for the dbuf cache. 46fa9e4066Sahrens */ 47fa9e4066Sahrens taskq_t *dbuf_tq; 48fa9e4066Sahrens static kmem_cache_t *dbuf_cache; 49fa9e4066Sahrens 50fa9e4066Sahrens /* ARGSUSED */ 51fa9e4066Sahrens static int 52fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag) 53fa9e4066Sahrens { 54fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 55fa9e4066Sahrens bzero(db, sizeof (dmu_buf_impl_t)); 56fa9e4066Sahrens 57fa9e4066Sahrens mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 58fa9e4066Sahrens cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 59fa9e4066Sahrens refcount_create(&db->db_holds); 60fa9e4066Sahrens return (0); 61fa9e4066Sahrens } 62fa9e4066Sahrens 63fa9e4066Sahrens /* ARGSUSED */ 64fa9e4066Sahrens static void 65fa9e4066Sahrens dbuf_dest(void *vdb, void *unused) 66fa9e4066Sahrens { 67fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 68fa9e4066Sahrens mutex_destroy(&db->db_mtx); 69fa9e4066Sahrens cv_destroy(&db->db_changed); 70fa9e4066Sahrens refcount_destroy(&db->db_holds); 71fa9e4066Sahrens } 72fa9e4066Sahrens 73fa9e4066Sahrens /* 74fa9e4066Sahrens * dbuf hash table routines 75fa9e4066Sahrens */ 76fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table; 77fa9e4066Sahrens 78fa9e4066Sahrens static uint64_t dbuf_hash_count; 79fa9e4066Sahrens 80fa9e4066Sahrens static uint64_t 81fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 82fa9e4066Sahrens { 83fa9e4066Sahrens uintptr_t osv = (uintptr_t)os; 84fa9e4066Sahrens uint64_t crc = -1ULL; 85fa9e4066Sahrens 86fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 87fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 88fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 89fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 90fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 91fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 92fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 93fa9e4066Sahrens 94fa9e4066Sahrens crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 95fa9e4066Sahrens 96fa9e4066Sahrens return (crc); 97fa9e4066Sahrens } 98fa9e4066Sahrens 99fa9e4066Sahrens #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 100fa9e4066Sahrens 101fa9e4066Sahrens #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 102fa9e4066Sahrens ((dbuf)->db.db_object == (obj) && \ 103fa9e4066Sahrens (dbuf)->db_objset == (os) && \ 104fa9e4066Sahrens (dbuf)->db_level == (level) && \ 105fa9e4066Sahrens (dbuf)->db_blkid == (blkid)) 106fa9e4066Sahrens 107fa9e4066Sahrens dmu_buf_impl_t * 108fa9e4066Sahrens dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 109fa9e4066Sahrens { 110fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 111fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 112fa9e4066Sahrens uint64_t obj = dn->dn_object; 113fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 114fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 115fa9e4066Sahrens dmu_buf_impl_t *db; 116fa9e4066Sahrens 117fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 118fa9e4066Sahrens for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 119fa9e4066Sahrens if (DBUF_EQUAL(db, os, obj, level, blkid)) { 120fa9e4066Sahrens mutex_enter(&db->db_mtx); 121ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 122fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 123fa9e4066Sahrens return (db); 124fa9e4066Sahrens } 125fa9e4066Sahrens mutex_exit(&db->db_mtx); 126fa9e4066Sahrens } 127fa9e4066Sahrens } 128fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 129fa9e4066Sahrens return (NULL); 130fa9e4066Sahrens } 131fa9e4066Sahrens 132fa9e4066Sahrens /* 133fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 134fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 135fa9e4066Sahrens * will be returned and the new element will not be inserted. 136fa9e4066Sahrens * Otherwise returns NULL. 137fa9e4066Sahrens */ 138fa9e4066Sahrens static dmu_buf_impl_t * 139fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db) 140fa9e4066Sahrens { 141fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 142fa9e4066Sahrens objset_impl_t *os = db->db_objset; 143fa9e4066Sahrens uint64_t obj = db->db.db_object; 144fa9e4066Sahrens int level = db->db_level; 145fa9e4066Sahrens uint64_t blkid = db->db_blkid; 146fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 147fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 148fa9e4066Sahrens dmu_buf_impl_t *dbf; 149fa9e4066Sahrens 150fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 151fa9e4066Sahrens for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 152fa9e4066Sahrens if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 153fa9e4066Sahrens mutex_enter(&dbf->db_mtx); 154ea8dc4b6Seschrock if (dbf->db_state != DB_EVICTING) { 155fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 156fa9e4066Sahrens return (dbf); 157fa9e4066Sahrens } 158fa9e4066Sahrens mutex_exit(&dbf->db_mtx); 159fa9e4066Sahrens } 160fa9e4066Sahrens } 161fa9e4066Sahrens 162fa9e4066Sahrens mutex_enter(&db->db_mtx); 163fa9e4066Sahrens db->db_hash_next = h->hash_table[idx]; 164fa9e4066Sahrens h->hash_table[idx] = db; 165fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 166fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, 1); 167fa9e4066Sahrens 168fa9e4066Sahrens return (NULL); 169fa9e4066Sahrens } 170fa9e4066Sahrens 171fa9e4066Sahrens /* 172fa9e4066Sahrens * Remove an entry from the hash table. This operation will 173fa9e4066Sahrens * fail if there are any existing holds on the db. 174fa9e4066Sahrens */ 175fa9e4066Sahrens static void 176fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db) 177fa9e4066Sahrens { 178fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 179fa9e4066Sahrens uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 180fa9e4066Sahrens db->db_level, db->db_blkid); 181fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 182fa9e4066Sahrens dmu_buf_impl_t *dbf, **dbp; 183fa9e4066Sahrens 184fa9e4066Sahrens /* 185fa9e4066Sahrens * We musn't hold db_mtx to maintin lock ordering: 186fa9e4066Sahrens * DBUF_HASH_MUTEX > db_mtx. 187fa9e4066Sahrens */ 188fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 189ea8dc4b6Seschrock ASSERT(db->db_state == DB_EVICTING); 190fa9e4066Sahrens ASSERT(!MUTEX_HELD(&db->db_mtx)); 191fa9e4066Sahrens 192fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 193fa9e4066Sahrens dbp = &h->hash_table[idx]; 194fa9e4066Sahrens while ((dbf = *dbp) != db) { 195fa9e4066Sahrens dbp = &dbf->db_hash_next; 196fa9e4066Sahrens ASSERT(dbf != NULL); 197fa9e4066Sahrens } 198fa9e4066Sahrens *dbp = db->db_hash_next; 199fa9e4066Sahrens db->db_hash_next = NULL; 200fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 201fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, -1); 202fa9e4066Sahrens } 203fa9e4066Sahrens 204ea8dc4b6Seschrock static arc_evict_func_t dbuf_do_evict; 205fa9e4066Sahrens 206fa9e4066Sahrens static void 207fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db) 208fa9e4066Sahrens { 209fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 210fa9e4066Sahrens 211fa9e4066Sahrens if (db->db_level != 0 || db->db_d.db_evict_func == NULL) 212fa9e4066Sahrens return; 213fa9e4066Sahrens 214fa9e4066Sahrens if (db->db_d.db_user_data_ptr_ptr) 215fa9e4066Sahrens *db->db_d.db_user_data_ptr_ptr = db->db.db_data; 216fa9e4066Sahrens db->db_d.db_evict_func(&db->db, db->db_d.db_user_ptr); 217fa9e4066Sahrens db->db_d.db_user_ptr = NULL; 218fa9e4066Sahrens db->db_d.db_user_data_ptr_ptr = NULL; 219fa9e4066Sahrens db->db_d.db_evict_func = NULL; 220fa9e4066Sahrens } 221fa9e4066Sahrens 222fa9e4066Sahrens void 223ea8dc4b6Seschrock dbuf_evict(dmu_buf_impl_t *db) 224ea8dc4b6Seschrock { 225ea8dc4b6Seschrock int i; 226ea8dc4b6Seschrock 227ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 228ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 229ea8dc4b6Seschrock 230ea8dc4b6Seschrock #ifdef ZFS_DEBUG 231ea8dc4b6Seschrock for (i = 0; i < TXG_SIZE; i++) { 232ea8dc4b6Seschrock ASSERT(!list_link_active(&db->db_dirty_node[i])); 233ea8dc4b6Seschrock ASSERT(db->db_level != 0 || db->db_d.db_data_old[i] == NULL); 234ea8dc4b6Seschrock } 235ea8dc4b6Seschrock #endif 236ea8dc4b6Seschrock dbuf_clear(db); 237ea8dc4b6Seschrock dbuf_destroy(db); 238ea8dc4b6Seschrock } 239ea8dc4b6Seschrock 240ea8dc4b6Seschrock void 241fa9e4066Sahrens dbuf_init(void) 242fa9e4066Sahrens { 243ea8dc4b6Seschrock uint64_t hsize = 1ULL << 16; 244fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 245fa9e4066Sahrens int i; 246fa9e4066Sahrens 247fa9e4066Sahrens /* 248fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 249ea8dc4b6Seschrock * with an average 4K block size. The table will take up 250ea8dc4b6Seschrock * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 251fa9e4066Sahrens */ 252ea8dc4b6Seschrock while (hsize * 4096 < physmem * PAGESIZE) 253fa9e4066Sahrens hsize <<= 1; 254fa9e4066Sahrens 255ea8dc4b6Seschrock retry: 256fa9e4066Sahrens h->hash_table_mask = hsize - 1; 257ea8dc4b6Seschrock h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 258ea8dc4b6Seschrock if (h->hash_table == NULL) { 259ea8dc4b6Seschrock /* XXX - we should really return an error instead of assert */ 260ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 10)); 261ea8dc4b6Seschrock hsize >>= 1; 262ea8dc4b6Seschrock goto retry; 263ea8dc4b6Seschrock } 264fa9e4066Sahrens 265fa9e4066Sahrens dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 266fa9e4066Sahrens sizeof (dmu_buf_impl_t), 267fa9e4066Sahrens 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 268fa9e4066Sahrens dbuf_tq = taskq_create("dbuf_tq", 8, maxclsyspri, 50, INT_MAX, 269fa9e4066Sahrens TASKQ_PREPOPULATE); 270fa9e4066Sahrens 271fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 272fa9e4066Sahrens mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 273fa9e4066Sahrens } 274fa9e4066Sahrens 275fa9e4066Sahrens void 276fa9e4066Sahrens dbuf_fini(void) 277fa9e4066Sahrens { 278fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 279fa9e4066Sahrens int i; 280fa9e4066Sahrens 281fa9e4066Sahrens taskq_destroy(dbuf_tq); 282fa9e4066Sahrens dbuf_tq = NULL; 283fa9e4066Sahrens 284fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 285fa9e4066Sahrens mutex_destroy(&h->hash_mutexes[i]); 286fa9e4066Sahrens kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 287fa9e4066Sahrens kmem_cache_destroy(dbuf_cache); 288fa9e4066Sahrens } 289fa9e4066Sahrens 290fa9e4066Sahrens /* 291fa9e4066Sahrens * Other stuff. 292fa9e4066Sahrens */ 293fa9e4066Sahrens 2949c9dc39aSek110237 #ifdef ZFS_DEBUG 295fa9e4066Sahrens static void 296fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db) 297fa9e4066Sahrens { 298fa9e4066Sahrens int i; 299fa9e4066Sahrens dnode_t *dn = db->db_dnode; 300fa9e4066Sahrens 301fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 302fa9e4066Sahrens 303fa9e4066Sahrens if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 304fa9e4066Sahrens return; 305fa9e4066Sahrens 306fa9e4066Sahrens ASSERT(db->db_objset != NULL); 307fa9e4066Sahrens if (dn == NULL) { 308fa9e4066Sahrens ASSERT(db->db_parent == NULL); 309fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 310fa9e4066Sahrens } else { 311fa9e4066Sahrens ASSERT3U(db->db.db_object, ==, dn->dn_object); 312fa9e4066Sahrens ASSERT3P(db->db_objset, ==, dn->dn_objset); 313fa9e4066Sahrens ASSERT3U(db->db_level, <, dn->dn_nlevels); 314ea8dc4b6Seschrock ASSERT(db->db_blkid == DB_BONUS_BLKID || 315ea8dc4b6Seschrock list_head(&dn->dn_dbufs)); 316fa9e4066Sahrens } 317fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 318fa9e4066Sahrens ASSERT(dn != NULL); 319fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, dn->dn_bonuslen); 320fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID); 321fa9e4066Sahrens } else { 322fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 323fa9e4066Sahrens } 324fa9e4066Sahrens 325fa9e4066Sahrens if (db->db_level == 0) { 326fa9e4066Sahrens /* we can be momentarily larger in dnode_set_blksz() */ 327fa9e4066Sahrens if (db->db_blkid != DB_BONUS_BLKID && dn) { 328fa9e4066Sahrens ASSERT3U(db->db.db_size, >=, dn->dn_datablksz); 329fa9e4066Sahrens } 330ea8dc4b6Seschrock if (db->db.db_object == DMU_META_DNODE_OBJECT) { 331fa9e4066Sahrens for (i = 0; i < TXG_SIZE; i++) { 332fa9e4066Sahrens /* 333fa9e4066Sahrens * it should only be modified in syncing 334fa9e4066Sahrens * context, so make sure we only have 335fa9e4066Sahrens * one copy of the data. 336fa9e4066Sahrens */ 337fa9e4066Sahrens ASSERT(db->db_d.db_data_old[i] == NULL || 338fa9e4066Sahrens db->db_d.db_data_old[i] == db->db_buf); 339fa9e4066Sahrens } 340fa9e4066Sahrens } 341fa9e4066Sahrens } 342fa9e4066Sahrens 343fa9e4066Sahrens /* verify db->db_blkptr */ 344fa9e4066Sahrens if (db->db_blkptr) { 345fa9e4066Sahrens if (db->db_parent == dn->dn_dbuf) { 346fa9e4066Sahrens /* db is pointed to by the dnode */ 347fa9e4066Sahrens /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 348ea8dc4b6Seschrock if (db->db.db_object == DMU_META_DNODE_OBJECT) 349fa9e4066Sahrens ASSERT(db->db_parent == NULL); 350fa9e4066Sahrens else 351fa9e4066Sahrens ASSERT(db->db_parent != NULL); 352fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 353fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 354fa9e4066Sahrens } else { 355fa9e4066Sahrens /* db is pointed to by an indirect block */ 356fa9e4066Sahrens int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 357fa9e4066Sahrens ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 358fa9e4066Sahrens ASSERT3U(db->db_parent->db.db_object, ==, 359fa9e4066Sahrens db->db.db_object); 360fa9e4066Sahrens /* 361fa9e4066Sahrens * dnode_grow_indblksz() can make this fail if we don't 362fa9e4066Sahrens * have the struct_rwlock. XXX indblksz no longer 363fa9e4066Sahrens * grows. safe to do this now? 364fa9e4066Sahrens */ 365fa9e4066Sahrens if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) { 366fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 367fa9e4066Sahrens ((blkptr_t *)db->db_parent->db.db_data + 368fa9e4066Sahrens db->db_blkid % epb)); 369fa9e4066Sahrens } 370fa9e4066Sahrens } 371fa9e4066Sahrens } 372fa9e4066Sahrens if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 373fa9e4066Sahrens db->db.db_data && db->db_blkid != DB_BONUS_BLKID && 374fa9e4066Sahrens db->db_state != DB_FILL && !dn->dn_free_txg) { 375fa9e4066Sahrens /* 376fa9e4066Sahrens * If the blkptr isn't set but they have nonzero data, 377fa9e4066Sahrens * it had better be dirty, otherwise we'll lose that 378fa9e4066Sahrens * data when we evict this buffer. 379fa9e4066Sahrens */ 380fa9e4066Sahrens if (db->db_dirtycnt == 0) { 381fa9e4066Sahrens uint64_t *buf = db->db.db_data; 382fa9e4066Sahrens int i; 383fa9e4066Sahrens 384fa9e4066Sahrens for (i = 0; i < db->db.db_size >> 3; i++) { 385fa9e4066Sahrens ASSERT(buf[i] == 0); 386fa9e4066Sahrens } 387fa9e4066Sahrens } 388fa9e4066Sahrens } 389fa9e4066Sahrens } 3909c9dc39aSek110237 #endif 391fa9e4066Sahrens 392fa9e4066Sahrens static void 393fa9e4066Sahrens dbuf_update_data(dmu_buf_impl_t *db) 394fa9e4066Sahrens { 395fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 396fa9e4066Sahrens if (db->db_level == 0 && db->db_d.db_user_data_ptr_ptr) { 397fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 398fa9e4066Sahrens *db->db_d.db_user_data_ptr_ptr = db->db.db_data; 399fa9e4066Sahrens } 400fa9e4066Sahrens } 401fa9e4066Sahrens 402fa9e4066Sahrens static void 403fa9e4066Sahrens dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 404fa9e4066Sahrens { 405fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 406ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 407fa9e4066Sahrens db->db_buf = buf; 408ea8dc4b6Seschrock if (buf != NULL) { 409ea8dc4b6Seschrock ASSERT(buf->b_data != NULL); 410fa9e4066Sahrens db->db.db_data = buf->b_data; 411ea8dc4b6Seschrock if (!arc_released(buf)) 412ea8dc4b6Seschrock arc_set_callback(buf, dbuf_do_evict, db); 413fa9e4066Sahrens dbuf_update_data(db); 414ea8dc4b6Seschrock } else { 415ea8dc4b6Seschrock dbuf_evict_user(db); 416ea8dc4b6Seschrock db->db.db_data = NULL; 417ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 418ea8dc4b6Seschrock } 419fa9e4066Sahrens } 420fa9e4066Sahrens 421fa9e4066Sahrens uint64_t 422fa9e4066Sahrens dbuf_whichblock(dnode_t *dn, uint64_t offset) 423fa9e4066Sahrens { 424fa9e4066Sahrens if (dn->dn_datablkshift) { 425fa9e4066Sahrens return (offset >> dn->dn_datablkshift); 426fa9e4066Sahrens } else { 427fa9e4066Sahrens ASSERT3U(offset, <, dn->dn_datablksz); 428fa9e4066Sahrens return (0); 429fa9e4066Sahrens } 430fa9e4066Sahrens } 431fa9e4066Sahrens 432fa9e4066Sahrens static void 433fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 434fa9e4066Sahrens { 435fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 436fa9e4066Sahrens 437fa9e4066Sahrens mutex_enter(&db->db_mtx); 438fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_READ); 439fa9e4066Sahrens /* 440fa9e4066Sahrens * All reads are synchronous, so we must have a hold on the dbuf 441fa9e4066Sahrens */ 442fa9e4066Sahrens ASSERT(refcount_count(&db->db_holds) > 0); 443ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 444fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 445fa9e4066Sahrens if (db->db_level == 0 && db->db_d.db_freed_in_flight) { 446fa9e4066Sahrens /* we were freed in flight; disregard any error */ 447fa9e4066Sahrens arc_release(buf, db); 448fa9e4066Sahrens bzero(buf->b_data, db->db.db_size); 449fa9e4066Sahrens db->db_d.db_freed_in_flight = FALSE; 450fa9e4066Sahrens dbuf_set_data(db, buf); 451fa9e4066Sahrens db->db_state = DB_CACHED; 452fa9e4066Sahrens } else if (zio == NULL || zio->io_error == 0) { 453fa9e4066Sahrens dbuf_set_data(db, buf); 454fa9e4066Sahrens db->db_state = DB_CACHED; 455fa9e4066Sahrens } else { 456fa9e4066Sahrens ASSERT(db->db_blkid != DB_BONUS_BLKID); 457fa9e4066Sahrens ASSERT3P(db->db_buf, ==, NULL); 458ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 459ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 460fa9e4066Sahrens } 461fa9e4066Sahrens cv_broadcast(&db->db_changed); 462fa9e4066Sahrens mutex_exit(&db->db_mtx); 463ea8dc4b6Seschrock dbuf_rele(db, NULL); 464fa9e4066Sahrens } 465fa9e4066Sahrens 466ea8dc4b6Seschrock static void 467fa9e4066Sahrens dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 468fa9e4066Sahrens { 469fa9e4066Sahrens blkptr_t *bp; 470ea8dc4b6Seschrock zbookmark_t zb; 471fa9e4066Sahrens 472fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 473fa9e4066Sahrens /* We need the struct_rwlock to prevent db_blkptr from changing. */ 474fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&db->db_dnode->dn_struct_rwlock)); 475ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 476ea8dc4b6Seschrock ASSERT(db->db_state == DB_UNCACHED); 477ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 478fa9e4066Sahrens 479fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 480fa9e4066Sahrens ASSERT3U(db->db_dnode->dn_bonuslen, ==, db->db.db_size); 481ea8dc4b6Seschrock db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 482fa9e4066Sahrens if (db->db.db_size < DN_MAX_BONUSLEN) 483ea8dc4b6Seschrock bzero(db->db.db_data, DN_MAX_BONUSLEN); 484ea8dc4b6Seschrock bcopy(DN_BONUS(db->db_dnode->dn_phys), db->db.db_data, 485fa9e4066Sahrens db->db.db_size); 486ea8dc4b6Seschrock dbuf_update_data(db); 487fa9e4066Sahrens db->db_state = DB_CACHED; 488fa9e4066Sahrens mutex_exit(&db->db_mtx); 489fa9e4066Sahrens return; 490fa9e4066Sahrens } 491fa9e4066Sahrens 492fa9e4066Sahrens if (db->db_level == 0 && dnode_block_freed(db->db_dnode, db->db_blkid)) 493fa9e4066Sahrens bp = NULL; 494fa9e4066Sahrens else 495fa9e4066Sahrens bp = db->db_blkptr; 496fa9e4066Sahrens 497fa9e4066Sahrens if (bp == NULL) 498fa9e4066Sahrens dprintf_dbuf(db, "blkptr: %s\n", "NULL"); 499fa9e4066Sahrens else 500fa9e4066Sahrens dprintf_dbuf_bp(db, bp, "%s", "blkptr:"); 501fa9e4066Sahrens 502fa9e4066Sahrens if (bp == NULL || BP_IS_HOLE(bp)) { 503fa9e4066Sahrens ASSERT(bp == NULL || BP_IS_HOLE(bp)); 504fa9e4066Sahrens dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 505fa9e4066Sahrens db->db.db_size, db)); 506fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 507fa9e4066Sahrens db->db_state = DB_CACHED; 508fa9e4066Sahrens mutex_exit(&db->db_mtx); 509fa9e4066Sahrens return; 510fa9e4066Sahrens } 511fa9e4066Sahrens 512fa9e4066Sahrens db->db_state = DB_READ; 513fa9e4066Sahrens mutex_exit(&db->db_mtx); 514fa9e4066Sahrens 515ea8dc4b6Seschrock zb.zb_objset = db->db_objset->os_dsl_dataset ? 516ea8dc4b6Seschrock db->db_objset->os_dsl_dataset->ds_object : 0; 517ea8dc4b6Seschrock zb.zb_object = db->db.db_object; 518ea8dc4b6Seschrock zb.zb_level = db->db_level; 519ea8dc4b6Seschrock zb.zb_blkid = db->db_blkid; 520ea8dc4b6Seschrock 521ea8dc4b6Seschrock dbuf_add_ref(db, NULL); 522fa9e4066Sahrens /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ 523fa9e4066Sahrens (void) arc_read(zio, db->db_dnode->dn_objset->os_spa, bp, 524fa9e4066Sahrens db->db_level > 0 ? byteswap_uint64_array : 525fa9e4066Sahrens dmu_ot[db->db_dnode->dn_type].ot_byteswap, 526fa9e4066Sahrens dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 527fa9e4066Sahrens (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 528ea8dc4b6Seschrock ARC_NOWAIT, &zb); 529fa9e4066Sahrens } 530fa9e4066Sahrens 531ea8dc4b6Seschrock int 532ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 533fa9e4066Sahrens { 534ea8dc4b6Seschrock int err = 0; 535ea8dc4b6Seschrock int havepzio = (zio != NULL); 536fa9e4066Sahrens 537fa9e4066Sahrens /* 538fa9e4066Sahrens * We don't have to hold the mutex to check db_state because it 539fa9e4066Sahrens * can't be freed while we have a hold on the buffer. 540fa9e4066Sahrens */ 541fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 542fa9e4066Sahrens if (db->db_state == DB_CACHED) 543fa9e4066Sahrens return (0); 544fa9e4066Sahrens 545fa9e4066Sahrens if ((flags & DB_RF_HAVESTRUCT) == 0) 546fa9e4066Sahrens rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER); 547fa9e4066Sahrens 548fa9e4066Sahrens mutex_enter(&db->db_mtx); 549ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 550ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 551ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 552ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 553ea8dc4b6Seschrock } else if (db->db_state == DB_UNCACHED) { 554ea8dc4b6Seschrock if (zio == NULL) { 555ea8dc4b6Seschrock zio = zio_root(db->db_dnode->dn_objset->os_spa, 556ea8dc4b6Seschrock NULL, NULL, ZIO_FLAG_CANFAIL); 557ea8dc4b6Seschrock } 558ea8dc4b6Seschrock dbuf_read_impl(db, zio, flags); 559ea8dc4b6Seschrock /* dbuf_read_impl has dropped db_mtx for us */ 560ea8dc4b6Seschrock 561ea8dc4b6Seschrock if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 562ea8dc4b6Seschrock (flags & DB_RF_NOPREFETCH) == 0 && 563ea8dc4b6Seschrock db->db_dnode != NULL) { 564ea8dc4b6Seschrock dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 565ea8dc4b6Seschrock db->db.db_size); 566ea8dc4b6Seschrock } 567ea8dc4b6Seschrock 568ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 569ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 570ea8dc4b6Seschrock 571ea8dc4b6Seschrock if (!havepzio) 572ea8dc4b6Seschrock err = zio_wait(zio); 573ea8dc4b6Seschrock } else { 574ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 575ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 576ea8dc4b6Seschrock if ((flags & DB_RF_NEVERWAIT) == 0) { 577ea8dc4b6Seschrock while (db->db_state == DB_READ || 578ea8dc4b6Seschrock db->db_state == DB_FILL) { 579fa9e4066Sahrens ASSERT(db->db_state == DB_READ || 580fa9e4066Sahrens (flags & DB_RF_HAVESTRUCT) == 0); 581fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 582fa9e4066Sahrens } 583ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED) 584ea8dc4b6Seschrock err = EIO; 585ea8dc4b6Seschrock } 586fa9e4066Sahrens mutex_exit(&db->db_mtx); 587fa9e4066Sahrens } 588fa9e4066Sahrens 589ea8dc4b6Seschrock ASSERT(err || havepzio || db->db_state == DB_CACHED); 590ea8dc4b6Seschrock return (err); 591fa9e4066Sahrens } 592fa9e4066Sahrens 593fa9e4066Sahrens static void 594fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db) 595fa9e4066Sahrens { 596fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 597ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 598fa9e4066Sahrens mutex_enter(&db->db_mtx); 599fa9e4066Sahrens while (db->db_state == DB_READ || db->db_state == DB_FILL) 600fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 601fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 602ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 603fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 604fa9e4066Sahrens dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 605ea8dc4b6Seschrock db->db.db_size, db)); 606fa9e4066Sahrens db->db_state = DB_FILL; 607fa9e4066Sahrens } else { 608fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 609fa9e4066Sahrens } 610fa9e4066Sahrens mutex_exit(&db->db_mtx); 611fa9e4066Sahrens } 612fa9e4066Sahrens 613fa9e4066Sahrens /* 614fa9e4066Sahrens * This is our just-in-time copy function. It makes a copy of 615fa9e4066Sahrens * buffers, that have been modified in a previous transaction 616fa9e4066Sahrens * group, before we modify them in the current active group. 617fa9e4066Sahrens * 618fa9e4066Sahrens * This function is used in two places: when we are dirtying a 619fa9e4066Sahrens * buffer for the first time in a txg, and when we are freeing 620fa9e4066Sahrens * a range in a dnode that includes this buffer. 621fa9e4066Sahrens * 622fa9e4066Sahrens * Note that when we are called from dbuf_free_range() we do 623fa9e4066Sahrens * not put a hold on the buffer, we just traverse the active 624fa9e4066Sahrens * dbuf list for the dnode. 625fa9e4066Sahrens */ 626fa9e4066Sahrens static void 627fa9e4066Sahrens dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 628fa9e4066Sahrens { 629fa9e4066Sahrens arc_buf_t **quiescing, **syncing; 630fa9e4066Sahrens 631fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 632fa9e4066Sahrens ASSERT(db->db.db_data != NULL); 633ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 634fa9e4066Sahrens 635ea8dc4b6Seschrock quiescing = (arc_buf_t **)&db->db_d.db_data_old[(txg-1)&TXG_MASK]; 636ea8dc4b6Seschrock syncing = (arc_buf_t **)&db->db_d.db_data_old[(txg-2)&TXG_MASK]; 637fa9e4066Sahrens 638fa9e4066Sahrens /* 639fa9e4066Sahrens * If this buffer is referenced from the current quiescing 640fa9e4066Sahrens * transaction group: either make a copy and reset the reference 641fa9e4066Sahrens * to point to the copy, or (if there a no active holders) just 642fa9e4066Sahrens * null out the current db_data pointer. 643fa9e4066Sahrens */ 644fa9e4066Sahrens if (*quiescing == db->db_buf) { 645fa9e4066Sahrens /* 646fa9e4066Sahrens * If the quiescing txg is "dirty", then we better not 647fa9e4066Sahrens * be referencing the same buffer from the syncing txg. 648fa9e4066Sahrens */ 649fa9e4066Sahrens ASSERT(*syncing != db->db_buf); 650fa9e4066Sahrens if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 651ea8dc4b6Seschrock int size = db->db.db_size; 652fa9e4066Sahrens *quiescing = arc_buf_alloc( 653fa9e4066Sahrens db->db_dnode->dn_objset->os_spa, size, db); 654fa9e4066Sahrens bcopy(db->db.db_data, (*quiescing)->b_data, size); 655fa9e4066Sahrens } else { 656ea8dc4b6Seschrock dbuf_set_data(db, NULL); 657fa9e4066Sahrens } 658fa9e4066Sahrens return; 659fa9e4066Sahrens } 660fa9e4066Sahrens 661fa9e4066Sahrens /* 662fa9e4066Sahrens * If this buffer is referenced from the current syncing 663fa9e4066Sahrens * transaction group: either 664fa9e4066Sahrens * 1 - make a copy and reset the reference, or 665fa9e4066Sahrens * 2 - if there are no holders, just null the current db_data. 666fa9e4066Sahrens */ 667fa9e4066Sahrens if (*syncing == db->db_buf) { 668fa9e4066Sahrens ASSERT3P(*quiescing, ==, NULL); 669fa9e4066Sahrens ASSERT3U(db->db_dirtycnt, ==, 1); 670fa9e4066Sahrens if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 671ea8dc4b6Seschrock int size = db->db.db_size; 672fa9e4066Sahrens /* we can't copy if we have already started a write */ 673fa9e4066Sahrens ASSERT(*syncing != db->db_data_pending); 674fa9e4066Sahrens *syncing = arc_buf_alloc( 675fa9e4066Sahrens db->db_dnode->dn_objset->os_spa, size, db); 676fa9e4066Sahrens bcopy(db->db.db_data, (*syncing)->b_data, size); 677fa9e4066Sahrens } else { 678ea8dc4b6Seschrock dbuf_set_data(db, NULL); 679fa9e4066Sahrens } 680fa9e4066Sahrens } 681fa9e4066Sahrens } 682fa9e4066Sahrens 683ea8dc4b6Seschrock /* 684ea8dc4b6Seschrock * This is the "bonus buffer" version of the above routine 685ea8dc4b6Seschrock */ 686ea8dc4b6Seschrock static void 687ea8dc4b6Seschrock dbuf_fix_old_bonus_data(dmu_buf_impl_t *db, uint64_t txg) 688ea8dc4b6Seschrock { 689ea8dc4b6Seschrock void **quiescing, **syncing; 690ea8dc4b6Seschrock 691ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 692ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 693ea8dc4b6Seschrock ASSERT(db->db_blkid == DB_BONUS_BLKID); 694ea8dc4b6Seschrock 695ea8dc4b6Seschrock quiescing = &db->db_d.db_data_old[(txg-1)&TXG_MASK]; 696ea8dc4b6Seschrock syncing = &db->db_d.db_data_old[(txg-2)&TXG_MASK]; 697ea8dc4b6Seschrock 698ea8dc4b6Seschrock if (*quiescing == db->db.db_data) { 699ea8dc4b6Seschrock ASSERT(*syncing != db->db.db_data); 700ea8dc4b6Seschrock *quiescing = zio_buf_alloc(DN_MAX_BONUSLEN); 701ea8dc4b6Seschrock bcopy(db->db.db_data, *quiescing, DN_MAX_BONUSLEN); 702ea8dc4b6Seschrock } else if (*syncing == db->db.db_data) { 703ea8dc4b6Seschrock ASSERT3P(*quiescing, ==, NULL); 704ea8dc4b6Seschrock ASSERT3U(db->db_dirtycnt, ==, 1); 705ea8dc4b6Seschrock *syncing = zio_buf_alloc(DN_MAX_BONUSLEN); 706ea8dc4b6Seschrock bcopy(db->db.db_data, *syncing, DN_MAX_BONUSLEN); 707ea8dc4b6Seschrock } 708ea8dc4b6Seschrock } 709ea8dc4b6Seschrock 710fa9e4066Sahrens void 711fa9e4066Sahrens dbuf_unoverride(dmu_buf_impl_t *db, uint64_t txg) 712fa9e4066Sahrens { 713ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 714fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 715fa9e4066Sahrens if (db->db_d.db_overridden_by[txg&TXG_MASK] == IN_DMU_SYNC) { 716fa9e4066Sahrens db->db_d.db_overridden_by[txg&TXG_MASK] = NULL; 717fa9e4066Sahrens } else if (db->db_d.db_overridden_by[txg&TXG_MASK] != NULL) { 718fa9e4066Sahrens /* free this block */ 719fa9e4066Sahrens ASSERT(list_link_active(&db->db_dirty_node[txg&TXG_MASK]) || 720fa9e4066Sahrens db->db_dnode->dn_free_txg == txg); 721fa9e4066Sahrens if (!BP_IS_HOLE(db->db_d.db_overridden_by[txg&TXG_MASK])) { 722fa9e4066Sahrens /* XXX can get silent EIO here */ 723fa9e4066Sahrens (void) arc_free(NULL, db->db_dnode->dn_objset->os_spa, 724fa9e4066Sahrens txg, db->db_d.db_overridden_by[txg&TXG_MASK], 725fa9e4066Sahrens NULL, NULL, ARC_WAIT); 726fa9e4066Sahrens } 727fa9e4066Sahrens kmem_free(db->db_d.db_overridden_by[txg&TXG_MASK], 728fa9e4066Sahrens sizeof (blkptr_t)); 729fa9e4066Sahrens db->db_d.db_overridden_by[txg&TXG_MASK] = NULL; 730fa9e4066Sahrens /* release the already-written buffer */ 731fa9e4066Sahrens arc_release(db->db_d.db_data_old[txg&TXG_MASK], db); 732fa9e4066Sahrens } 733fa9e4066Sahrens } 734fa9e4066Sahrens 735fa9e4066Sahrens void 736fa9e4066Sahrens dbuf_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx) 737fa9e4066Sahrens { 738fa9e4066Sahrens dmu_buf_impl_t *db, *db_next; 739fa9e4066Sahrens uint64_t txg = tx->tx_txg; 740fa9e4066Sahrens 741fa9e4066Sahrens dprintf_dnode(dn, "blkid=%llu nblks=%llu\n", blkid, nblks); 742fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 743fa9e4066Sahrens for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 744fa9e4066Sahrens db_next = list_next(&dn->dn_dbufs, db); 745ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 746ea8dc4b6Seschrock if (db->db_level != 0) 747fa9e4066Sahrens continue; 748fa9e4066Sahrens dprintf_dbuf(db, "found buf %s\n", ""); 749fa9e4066Sahrens if (db->db_blkid < blkid || 750fa9e4066Sahrens db->db_blkid >= blkid+nblks) 751fa9e4066Sahrens continue; 752fa9e4066Sahrens 753fa9e4066Sahrens /* found a level 0 buffer in the range */ 754fa9e4066Sahrens if (dbuf_undirty(db, tx)) 755fa9e4066Sahrens continue; 756fa9e4066Sahrens 757fa9e4066Sahrens mutex_enter(&db->db_mtx); 758ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED || 759ea8dc4b6Seschrock db->db_state == DB_EVICTING) { 760fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 761fa9e4066Sahrens mutex_exit(&db->db_mtx); 762fa9e4066Sahrens continue; 763fa9e4066Sahrens } 764*c543ec06Sahrens if (db->db_state == DB_READ || db->db_state == DB_FILL) { 765*c543ec06Sahrens /* will be handled in dbuf_read_done or dbuf_rele */ 766fa9e4066Sahrens db->db_d.db_freed_in_flight = TRUE; 767fa9e4066Sahrens mutex_exit(&db->db_mtx); 768fa9e4066Sahrens continue; 769fa9e4066Sahrens } 770ea8dc4b6Seschrock if (refcount_count(&db->db_holds) == 0) { 771ea8dc4b6Seschrock ASSERT(db->db_buf); 772ea8dc4b6Seschrock dbuf_clear(db); 773ea8dc4b6Seschrock continue; 774ea8dc4b6Seschrock } 775ea8dc4b6Seschrock /* The dbuf is CACHED and referenced */ 776fa9e4066Sahrens 777ea8dc4b6Seschrock if (!list_link_active(&db->db_dirty_node[txg & TXG_MASK])) { 778ea8dc4b6Seschrock /* 779ea8dc4b6Seschrock * This dbuf is not currently dirty. We will either 780ea8dc4b6Seschrock * uncache it (if its not referenced in the open 781ea8dc4b6Seschrock * context) or reset its contents to empty. 782ea8dc4b6Seschrock */ 783fa9e4066Sahrens dbuf_fix_old_data(db, txg); 784ea8dc4b6Seschrock } else if (db->db_d.db_overridden_by[txg & TXG_MASK] != NULL) { 785ea8dc4b6Seschrock /* 786ea8dc4b6Seschrock * This dbuf is overridden. Clear that state. 787ea8dc4b6Seschrock */ 788ea8dc4b6Seschrock dbuf_unoverride(db, txg); 789ea8dc4b6Seschrock } 790fa9e4066Sahrens /* fill in with appropriate data */ 791ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 792ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 793fa9e4066Sahrens arc_release(db->db_buf, db); 794fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 795fa9e4066Sahrens } 796ea8dc4b6Seschrock 797fa9e4066Sahrens mutex_exit(&db->db_mtx); 798fa9e4066Sahrens } 799fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 800fa9e4066Sahrens } 801fa9e4066Sahrens 802fa9e4066Sahrens static int 803ea8dc4b6Seschrock dbuf_new_block(dmu_buf_impl_t *db) 804fa9e4066Sahrens { 805fa9e4066Sahrens dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 806fa9e4066Sahrens uint64_t birth_txg = 0; 807fa9e4066Sahrens 808fa9e4066Sahrens /* Don't count meta-objects */ 809fa9e4066Sahrens if (ds == NULL) 810fa9e4066Sahrens return (FALSE); 811fa9e4066Sahrens 812fa9e4066Sahrens /* 813fa9e4066Sahrens * We don't need any locking to protect db_blkptr: 814fa9e4066Sahrens * If it's syncing, then db_dirtied will be set so we'll 815fa9e4066Sahrens * ignore db_blkptr. 816fa9e4066Sahrens */ 817fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); /* XXX strictly necessary? */ 818fa9e4066Sahrens /* If we have been dirtied since the last snapshot, its not new */ 819fa9e4066Sahrens if (db->db_dirtied) 820fa9e4066Sahrens birth_txg = db->db_dirtied; 821fa9e4066Sahrens else if (db->db_blkptr) 822fa9e4066Sahrens birth_txg = db->db_blkptr->blk_birth; 823fa9e4066Sahrens 824fa9e4066Sahrens if (birth_txg) 825ea8dc4b6Seschrock return (!dsl_dataset_block_freeable(ds, birth_txg)); 826fa9e4066Sahrens else 827fa9e4066Sahrens return (TRUE); 828fa9e4066Sahrens } 829fa9e4066Sahrens 830fa9e4066Sahrens void 831fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 832fa9e4066Sahrens { 833fa9e4066Sahrens arc_buf_t *buf, *obuf; 834fa9e4066Sahrens int osize = db->db.db_size; 835fa9e4066Sahrens 836ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 837ea8dc4b6Seschrock 838fa9e4066Sahrens /* XXX does *this* func really need the lock? */ 839fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)); 840fa9e4066Sahrens 841fa9e4066Sahrens /* 842fa9e4066Sahrens * This call to dbuf_will_dirty() with the dn_struct_rwlock held 843fa9e4066Sahrens * is OK, because there can be no other references to the db 844fa9e4066Sahrens * when we are changing its size, so no concurrent DB_FILL can 845fa9e4066Sahrens * be happening. 846fa9e4066Sahrens */ 847ea8dc4b6Seschrock /* 848ea8dc4b6Seschrock * XXX we should be doing a dbuf_read, checking the return 849ea8dc4b6Seschrock * value and returning that up to our callers 850ea8dc4b6Seschrock */ 851fa9e4066Sahrens dbuf_will_dirty(db, tx); 852fa9e4066Sahrens 853fa9e4066Sahrens /* create the data buffer for the new block */ 854fa9e4066Sahrens buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db); 855fa9e4066Sahrens 856fa9e4066Sahrens /* copy old block data to the new block */ 857fa9e4066Sahrens obuf = db->db_buf; 858f65e61c0Sahrens bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 859fa9e4066Sahrens /* zero the remainder */ 860f65e61c0Sahrens if (size > osize) 861fa9e4066Sahrens bzero((uint8_t *)buf->b_data + osize, size - osize); 862fa9e4066Sahrens 863fa9e4066Sahrens mutex_enter(&db->db_mtx); 864fa9e4066Sahrens dbuf_set_data(db, buf); 865ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(obuf, db) == 1); 866fa9e4066Sahrens db->db.db_size = size; 867fa9e4066Sahrens 868fa9e4066Sahrens if (db->db_level == 0) 869fa9e4066Sahrens db->db_d.db_data_old[tx->tx_txg&TXG_MASK] = buf; 870fa9e4066Sahrens mutex_exit(&db->db_mtx); 871fa9e4066Sahrens 872fa9e4066Sahrens dnode_willuse_space(db->db_dnode, size-osize, tx); 873fa9e4066Sahrens } 874fa9e4066Sahrens 875fa9e4066Sahrens void 876fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 877fa9e4066Sahrens { 878fa9e4066Sahrens dnode_t *dn = db->db_dnode; 879fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 880fa9e4066Sahrens int drop_struct_lock = FALSE; 881fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 882fa9e4066Sahrens 883fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 884fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 8859c9dc39aSek110237 DMU_TX_DIRTY_BUF(tx, db); 886fa9e4066Sahrens 887fa9e4066Sahrens /* 888fa9e4066Sahrens * Shouldn't dirty a regular buffer in syncing context. Private 889fa9e4066Sahrens * objects may be dirtied in syncing context, but only if they 890fa9e4066Sahrens * were already pre-dirtied in open context. 891fa9e4066Sahrens * XXX We may want to prohibit dirtying in syncing context even 892fa9e4066Sahrens * if they did pre-dirty. 893fa9e4066Sahrens */ 894fa9e4066Sahrens ASSERT(!(dmu_tx_is_syncing(tx) && 895fa9e4066Sahrens !BP_IS_HOLE(&dn->dn_objset->os_rootbp) && 896ea8dc4b6Seschrock dn->dn_object != DMU_META_DNODE_OBJECT && 897fa9e4066Sahrens dn->dn_objset->os_dsl_dataset != NULL && 898fa9e4066Sahrens !dsl_dir_is_private( 899fa9e4066Sahrens dn->dn_objset->os_dsl_dataset->ds_dir))); 900fa9e4066Sahrens 901fa9e4066Sahrens /* 902fa9e4066Sahrens * We make this assert for private objects as well, but after we 903fa9e4066Sahrens * check if we're already dirty. They are allowed to re-dirty 904fa9e4066Sahrens * in syncing context. 905fa9e4066Sahrens */ 906ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 907fa9e4066Sahrens dn->dn_dirtyctx == DN_UNDIRTIED || 908fa9e4066Sahrens dn->dn_dirtyctx == 909fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 910fa9e4066Sahrens 911fa9e4066Sahrens mutex_enter(&db->db_mtx); 912fa9e4066Sahrens /* XXX make this true for indirects too? */ 913fa9e4066Sahrens ASSERT(db->db_level != 0 || db->db_state == DB_CACHED || 914fa9e4066Sahrens db->db_state == DB_FILL); 915fa9e4066Sahrens 916fa9e4066Sahrens /* 917fa9e4066Sahrens * If this buffer is currently part of an "overridden" region, 918fa9e4066Sahrens * we now need to remove it from that region. 919fa9e4066Sahrens */ 920fa9e4066Sahrens if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 921fa9e4066Sahrens db->db_d.db_overridden_by[txgoff] != NULL) { 922fa9e4066Sahrens dbuf_unoverride(db, tx->tx_txg); 923fa9e4066Sahrens } 924fa9e4066Sahrens 925fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 926fa9e4066Sahrens /* 927fa9e4066Sahrens * Don't set dirtyctx to SYNC if we're just modifying this as we 928fa9e4066Sahrens * initialize the objset. 929fa9e4066Sahrens */ 930fa9e4066Sahrens if (dn->dn_dirtyctx == DN_UNDIRTIED && 931fa9e4066Sahrens !BP_IS_HOLE(&dn->dn_objset->os_rootbp)) { 932fa9e4066Sahrens dn->dn_dirtyctx = 933fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 934fa9e4066Sahrens ASSERT(dn->dn_dirtyctx_firstset == NULL); 935fa9e4066Sahrens dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 936fa9e4066Sahrens } 937fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 938fa9e4066Sahrens 939fa9e4066Sahrens /* 940fa9e4066Sahrens * If this buffer is already dirty, we're done. 941fa9e4066Sahrens */ 942fa9e4066Sahrens if (list_link_active(&db->db_dirty_node[txgoff])) { 943fa9e4066Sahrens mutex_exit(&db->db_mtx); 944fa9e4066Sahrens return; 945fa9e4066Sahrens } 946fa9e4066Sahrens 947fa9e4066Sahrens /* 948fa9e4066Sahrens * Only valid if not already dirty. 949fa9e4066Sahrens */ 950fa9e4066Sahrens ASSERT(dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 951fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 952fa9e4066Sahrens 953fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, db->db_level); 954fa9e4066Sahrens ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 955fa9e4066Sahrens dn->dn_phys->dn_nlevels > db->db_level || 956fa9e4066Sahrens dn->dn_next_nlevels[txgoff] > db->db_level || 957fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 958fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 959fa9e4066Sahrens 960fa9e4066Sahrens /* 961fa9e4066Sahrens * We should only be dirtying in syncing context if it's the 962fa9e4066Sahrens * mos, a spa os, or we're initializing the os. However, we are 963fa9e4066Sahrens * allowed to dirty in syncing context provided we already 964fa9e4066Sahrens * dirtied it in open context. Hence we must make this 965fa9e4066Sahrens * assertion only if we're not already dirty. 966fa9e4066Sahrens */ 967fa9e4066Sahrens ASSERT(!dmu_tx_is_syncing(tx) || 968fa9e4066Sahrens os->os_dsl_dataset == NULL || 969fa9e4066Sahrens !dsl_dir_is_private(os->os_dsl_dataset->ds_dir) || 970fa9e4066Sahrens !BP_IS_HOLE(&os->os_rootbp)); 971fa9e4066Sahrens ASSERT(db->db.db_size != 0); 972fa9e4066Sahrens 973fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 974fa9e4066Sahrens 975ea8dc4b6Seschrock /* 976ea8dc4b6Seschrock * If this buffer is dirty in an old transaction group we need 977ea8dc4b6Seschrock * to make a copy of it so that the changes we make in this 978ea8dc4b6Seschrock * transaction group won't leak out when we sync the older txg. 979ea8dc4b6Seschrock */ 980ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 981ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 982ea8dc4b6Seschrock ASSERT(db->db_d.db_data_old[txgoff] == NULL); 983ea8dc4b6Seschrock dbuf_fix_old_bonus_data(db, tx->tx_txg); 984ea8dc4b6Seschrock db->db_d.db_data_old[txgoff] = db->db.db_data; 985ea8dc4b6Seschrock } else if (db->db_level == 0) { 986fa9e4066Sahrens /* 987fa9e4066Sahrens * Release the data buffer from the cache so that we 988fa9e4066Sahrens * can modify it without impacting possible other users 989fa9e4066Sahrens * of this cached data block. Note that indirect blocks 990fa9e4066Sahrens * and private objects are not released until the syncing 991fa9e4066Sahrens * state (since they are only modified then). 992fa9e4066Sahrens */ 993fa9e4066Sahrens ASSERT(db->db_buf != NULL); 994fa9e4066Sahrens ASSERT(db->db_d.db_data_old[txgoff] == NULL); 995ea8dc4b6Seschrock if (db->db.db_object != DMU_META_DNODE_OBJECT) { 996fa9e4066Sahrens arc_release(db->db_buf, db); 997fa9e4066Sahrens dbuf_fix_old_data(db, tx->tx_txg); 998fa9e4066Sahrens ASSERT(db->db_buf != NULL); 999fa9e4066Sahrens } 1000fa9e4066Sahrens db->db_d.db_data_old[txgoff] = db->db_buf; 1001fa9e4066Sahrens } 1002fa9e4066Sahrens 1003fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1004fa9e4066Sahrens /* 1005fa9e4066Sahrens * We could have been freed_in_flight between the dbuf_noread 1006fa9e4066Sahrens * and dbuf_dirty. We win, as though the dbuf_noread() had 1007fa9e4066Sahrens * happened after the free. 1008fa9e4066Sahrens */ 1009fa9e4066Sahrens if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 1010fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1011fa9e4066Sahrens db->db_d.db_freed_in_flight = FALSE; 1012fa9e4066Sahrens } 1013fa9e4066Sahrens 1014fa9e4066Sahrens db->db_dirtied = tx->tx_txg; 1015fa9e4066Sahrens list_insert_tail(&dn->dn_dirty_dbufs[txgoff], db); 1016fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1017fa9e4066Sahrens 1018fa9e4066Sahrens if (db->db_blkid != DB_BONUS_BLKID) { 1019ea8dc4b6Seschrock /* 1020ea8dc4b6Seschrock * Update the accounting. 1021ea8dc4b6Seschrock */ 1022ea8dc4b6Seschrock if (!dbuf_new_block(db) && db->db_blkptr) { 1023fa9e4066Sahrens /* 1024fa9e4066Sahrens * This is only a guess -- if the dbuf is dirty 1025fa9e4066Sahrens * in a previous txg, we don't know how much 1026fa9e4066Sahrens * space it will use on disk yet. We should 1027fa9e4066Sahrens * really have the struct_rwlock to access 1028fa9e4066Sahrens * db_blkptr, but since this is just a guess, 1029fa9e4066Sahrens * it's OK if we get an odd answer. 1030fa9e4066Sahrens */ 1031fa9e4066Sahrens dnode_willuse_space(dn, 1032fa9e4066Sahrens -BP_GET_ASIZE(db->db_blkptr), tx); 1033fa9e4066Sahrens } 1034fa9e4066Sahrens dnode_willuse_space(dn, db->db.db_size, tx); 1035fa9e4066Sahrens } 1036fa9e4066Sahrens 1037fa9e4066Sahrens /* 1038fa9e4066Sahrens * This buffer is now part of this txg 1039fa9e4066Sahrens */ 1040fa9e4066Sahrens dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1041fa9e4066Sahrens db->db_dirtycnt += 1; 1042fa9e4066Sahrens ASSERT3U(db->db_dirtycnt, <=, 3); 1043fa9e4066Sahrens 1044fa9e4066Sahrens mutex_exit(&db->db_mtx); 1045fa9e4066Sahrens 1046fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 1047fa9e4066Sahrens dnode_setdirty(dn, tx); 1048fa9e4066Sahrens return; 1049fa9e4066Sahrens } 1050fa9e4066Sahrens 1051fa9e4066Sahrens if (db->db_level == 0) 1052fa9e4066Sahrens dnode_new_blkid(dn, db->db_blkid, tx); 1053fa9e4066Sahrens 1054fa9e4066Sahrens if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1055fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 1056fa9e4066Sahrens drop_struct_lock = TRUE; 1057fa9e4066Sahrens } 1058fa9e4066Sahrens 1059fa9e4066Sahrens if (db->db_level < dn->dn_nlevels-1) { 1060fa9e4066Sahrens int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1061fa9e4066Sahrens dmu_buf_impl_t *parent; 1062fa9e4066Sahrens parent = dbuf_hold_level(dn, db->db_level+1, 1063fa9e4066Sahrens db->db_blkid >> epbs, FTAG); 1064fa9e4066Sahrens if (drop_struct_lock) 1065fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1066fa9e4066Sahrens dbuf_dirty(parent, tx); 1067ea8dc4b6Seschrock dbuf_rele(parent, FTAG); 1068fa9e4066Sahrens } else { 1069fa9e4066Sahrens if (drop_struct_lock) 1070fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1071fa9e4066Sahrens } 1072fa9e4066Sahrens 1073fa9e4066Sahrens dnode_setdirty(dn, tx); 1074fa9e4066Sahrens } 1075fa9e4066Sahrens 1076fa9e4066Sahrens static int 1077fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1078fa9e4066Sahrens { 1079fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1080fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 1081ea8dc4b6Seschrock int64_t holds; 1082fa9e4066Sahrens 1083fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1084ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1085fa9e4066Sahrens 1086fa9e4066Sahrens mutex_enter(&db->db_mtx); 1087fa9e4066Sahrens 1088fa9e4066Sahrens /* 1089fa9e4066Sahrens * If this buffer is not dirty, we're done. 1090fa9e4066Sahrens */ 1091fa9e4066Sahrens if (!list_link_active(&db->db_dirty_node[txgoff])) { 1092fa9e4066Sahrens mutex_exit(&db->db_mtx); 1093fa9e4066Sahrens return (0); 1094fa9e4066Sahrens } 1095fa9e4066Sahrens 1096fa9e4066Sahrens /* 1097fa9e4066Sahrens * If this buffer is currently held, we cannot undirty 1098fa9e4066Sahrens * it, since one of the current holders may be in the 1099fa9e4066Sahrens * middle of an update. Note that users of dbuf_undirty() 1100fa9e4066Sahrens * should not place a hold on the dbuf before the call. 1101fa9e4066Sahrens * XXX - this check assumes we are being called from 1102fa9e4066Sahrens * dbuf_free_range(), perhaps we should move it there? 1103fa9e4066Sahrens */ 1104fa9e4066Sahrens if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1105fa9e4066Sahrens mutex_exit(&db->db_mtx); 1106fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1107fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1108fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1109fa9e4066Sahrens return (0); 1110fa9e4066Sahrens } 1111fa9e4066Sahrens 1112fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1113fa9e4066Sahrens 1114fa9e4066Sahrens dbuf_unoverride(db, tx->tx_txg); 1115fa9e4066Sahrens 1116fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1117fa9e4066Sahrens if (db->db_level == 0) { 1118fa9e4066Sahrens ASSERT(db->db_buf != NULL); 1119fa9e4066Sahrens ASSERT(db->db_d.db_data_old[txgoff] != NULL); 1120fa9e4066Sahrens if (db->db_d.db_data_old[txgoff] != db->db_buf) 1121ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref( 1122ea8dc4b6Seschrock db->db_d.db_data_old[txgoff], db) == 1); 1123fa9e4066Sahrens db->db_d.db_data_old[txgoff] = NULL; 1124fa9e4066Sahrens } 1125fa9e4066Sahrens 1126fa9e4066Sahrens /* XXX would be nice to fix up dn_towrite_space[] */ 1127fa9e4066Sahrens /* XXX undo db_dirtied? but how? */ 1128fa9e4066Sahrens /* db->db_dirtied = tx->tx_txg; */ 1129fa9e4066Sahrens 1130fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1131fa9e4066Sahrens list_remove(&dn->dn_dirty_dbufs[txgoff], db); 1132fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1133fa9e4066Sahrens 1134fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1135fa9e4066Sahrens db->db_dirtycnt -= 1; 1136fa9e4066Sahrens 1137ea8dc4b6Seschrock if ((holds = refcount_remove(&db->db_holds, 1138ea8dc4b6Seschrock (void *)(uintptr_t)tx->tx_txg)) == 0) { 1139ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1140fa9e4066Sahrens 1141ea8dc4b6Seschrock ASSERT(arc_released(buf)); 1142ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1143ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1144fa9e4066Sahrens dbuf_evict(db); 1145fa9e4066Sahrens return (1); 1146fa9e4066Sahrens } 1147ea8dc4b6Seschrock ASSERT(holds > 0); 1148fa9e4066Sahrens 1149fa9e4066Sahrens mutex_exit(&db->db_mtx); 1150fa9e4066Sahrens return (0); 1151fa9e4066Sahrens } 1152fa9e4066Sahrens 1153fa9e4066Sahrens #pragma weak dmu_buf_will_dirty = dbuf_will_dirty 1154fa9e4066Sahrens void 1155fa9e4066Sahrens dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1156fa9e4066Sahrens { 1157fa9e4066Sahrens int rf = DB_RF_MUST_SUCCEED; 1158fa9e4066Sahrens 1159fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1160fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1161fa9e4066Sahrens 1162fa9e4066Sahrens if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) 1163fa9e4066Sahrens rf |= DB_RF_HAVESTRUCT; 1164ea8dc4b6Seschrock (void) dbuf_read(db, NULL, rf); 1165fa9e4066Sahrens dbuf_dirty(db, tx); 1166fa9e4066Sahrens } 1167fa9e4066Sahrens 1168fa9e4066Sahrens void 1169ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1170fa9e4066Sahrens { 1171ea8dc4b6Seschrock dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1172ea8dc4b6Seschrock 1173ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1174fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1175fa9e4066Sahrens ASSERT(db->db_level == 0); 1176fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1177fa9e4066Sahrens 1178ea8dc4b6Seschrock ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1179fa9e4066Sahrens dmu_tx_private_ok(tx)); 1180fa9e4066Sahrens 1181fa9e4066Sahrens dbuf_noread(db); 1182fa9e4066Sahrens dbuf_dirty(db, tx); 1183fa9e4066Sahrens } 1184fa9e4066Sahrens 1185fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done 1186fa9e4066Sahrens /* ARGSUSED */ 1187fa9e4066Sahrens void 1188fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1189fa9e4066Sahrens { 1190fa9e4066Sahrens mutex_enter(&db->db_mtx); 11919c9dc39aSek110237 DBUF_VERIFY(db); 1192fa9e4066Sahrens 1193fa9e4066Sahrens if (db->db_state == DB_FILL) { 1194fa9e4066Sahrens if (db->db_level == 0 && db->db_d.db_freed_in_flight) { 1195ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1196fa9e4066Sahrens /* we were freed while filling */ 1197fa9e4066Sahrens /* XXX dbuf_undirty? */ 1198fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 1199fa9e4066Sahrens db->db_d.db_freed_in_flight = FALSE; 1200fa9e4066Sahrens } 1201fa9e4066Sahrens db->db_state = DB_CACHED; 1202fa9e4066Sahrens cv_broadcast(&db->db_changed); 1203fa9e4066Sahrens } 1204fa9e4066Sahrens mutex_exit(&db->db_mtx); 1205fa9e4066Sahrens } 1206fa9e4066Sahrens 1207ea8dc4b6Seschrock /* 1208ea8dc4b6Seschrock * "Clear" the contents of this dbuf. This will mark the dbuf 1209ea8dc4b6Seschrock * EVICTING and clear *most* of its references. Unfortunetely, 1210ea8dc4b6Seschrock * when we are not holding the dn_dbufs_mtx, we can't clear the 1211ea8dc4b6Seschrock * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1212ea8dc4b6Seschrock * in this case. For callers from the DMU we will usually see: 1213ea8dc4b6Seschrock * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1214ea8dc4b6Seschrock * For the arc callback, we will usually see: 1215ea8dc4b6Seschrock * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1216ea8dc4b6Seschrock * Sometimes, though, we will get a mix of these two: 1217ea8dc4b6Seschrock * DMU: dbuf_clear()->arc_buf_evict() 1218ea8dc4b6Seschrock * ARC: dbuf_do_evict()->dbuf_destroy() 1219ea8dc4b6Seschrock */ 1220ea8dc4b6Seschrock void 1221fa9e4066Sahrens dbuf_clear(dmu_buf_impl_t *db) 1222fa9e4066Sahrens { 1223fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1224ea8dc4b6Seschrock dmu_buf_impl_t *parent = db->db_parent; 1225*c543ec06Sahrens dmu_buf_impl_t *dndb = dn->dn_dbuf; 1226ea8dc4b6Seschrock int dbuf_gone = FALSE; 1227fa9e4066Sahrens 1228fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 1229fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1230fa9e4066Sahrens 1231ea8dc4b6Seschrock dbuf_evict_user(db); 1232ea8dc4b6Seschrock 1233fa9e4066Sahrens if (db->db_state == DB_CACHED) { 1234ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 1235ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) 1236ea8dc4b6Seschrock zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1237fa9e4066Sahrens db->db.db_data = NULL; 1238fa9e4066Sahrens db->db_state = DB_UNCACHED; 1239fa9e4066Sahrens } 1240fa9e4066Sahrens 1241fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_UNCACHED); 1242fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1243fa9e4066Sahrens 1244ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1245ea8dc4b6Seschrock db->db_blkptr = NULL; 1246ea8dc4b6Seschrock 1247ea8dc4b6Seschrock if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1248ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1249ea8dc4b6Seschrock dnode_rele(dn, db); 1250ea8dc4b6Seschrock } 1251ea8dc4b6Seschrock 1252ea8dc4b6Seschrock if (db->db_buf) 1253ea8dc4b6Seschrock dbuf_gone = arc_buf_evict(db->db_buf); 1254ea8dc4b6Seschrock 1255ea8dc4b6Seschrock if (!dbuf_gone) 1256fa9e4066Sahrens mutex_exit(&db->db_mtx); 1257fa9e4066Sahrens 1258fa9e4066Sahrens /* 1259fa9e4066Sahrens * If this dbuf is referened from an indirect dbuf, 1260fa9e4066Sahrens * decrement the ref count on the indirect dbuf. 1261fa9e4066Sahrens */ 1262*c543ec06Sahrens if (parent && parent != dndb) 1263ea8dc4b6Seschrock dbuf_rele(parent, db); 1264fa9e4066Sahrens } 1265fa9e4066Sahrens 1266fa9e4066Sahrens static int 1267fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1268fa9e4066Sahrens dmu_buf_impl_t **parentp, blkptr_t **bpp) 1269fa9e4066Sahrens { 1270fa9e4066Sahrens int nlevels, epbs; 1271fa9e4066Sahrens 1272ea8dc4b6Seschrock ASSERT(blkid != DB_BONUS_BLKID); 1273ea8dc4b6Seschrock 1274fa9e4066Sahrens if (dn->dn_phys->dn_nlevels == 0) 1275fa9e4066Sahrens nlevels = 1; 1276fa9e4066Sahrens else 1277fa9e4066Sahrens nlevels = dn->dn_phys->dn_nlevels; 1278fa9e4066Sahrens 1279fa9e4066Sahrens epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1280fa9e4066Sahrens 1281fa9e4066Sahrens ASSERT3U(level * epbs, <, 64); 1282fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1283ea8dc4b6Seschrock if (level >= nlevels || 1284fa9e4066Sahrens (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1285fa9e4066Sahrens /* the buffer has no parent yet */ 1286fa9e4066Sahrens *parentp = NULL; 1287fa9e4066Sahrens *bpp = NULL; 1288fa9e4066Sahrens return (ENOENT); 1289fa9e4066Sahrens } else if (level < nlevels-1) { 1290fa9e4066Sahrens /* this block is referenced from an indirect block */ 1291fa9e4066Sahrens int err = dbuf_hold_impl(dn, level+1, 1292fa9e4066Sahrens blkid >> epbs, fail_sparse, NULL, parentp); 1293fa9e4066Sahrens if (err) 1294fa9e4066Sahrens return (err); 1295ea8dc4b6Seschrock err = dbuf_read(*parentp, NULL, 1296ea8dc4b6Seschrock (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1297*c543ec06Sahrens if (err) { 1298*c543ec06Sahrens dbuf_rele(*parentp, NULL); 1299*c543ec06Sahrens *parentp = NULL; 1300*c543ec06Sahrens return (err); 1301*c543ec06Sahrens } 1302fa9e4066Sahrens *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1303fa9e4066Sahrens (blkid & ((1ULL << epbs) - 1)); 1304*c543ec06Sahrens return (0); 1305fa9e4066Sahrens } else { 1306fa9e4066Sahrens /* the block is referenced from the dnode */ 1307fa9e4066Sahrens ASSERT3U(level, ==, nlevels-1); 1308fa9e4066Sahrens ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1309fa9e4066Sahrens blkid < dn->dn_phys->dn_nblkptr); 1310*c543ec06Sahrens if (dn->dn_dbuf) { 1311*c543ec06Sahrens dbuf_add_ref(dn->dn_dbuf, NULL); 1312fa9e4066Sahrens *parentp = dn->dn_dbuf; 1313*c543ec06Sahrens } 1314fa9e4066Sahrens *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1315fa9e4066Sahrens return (0); 1316fa9e4066Sahrens } 1317fa9e4066Sahrens } 1318fa9e4066Sahrens 1319fa9e4066Sahrens static dmu_buf_impl_t * 1320fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1321fa9e4066Sahrens dmu_buf_impl_t *parent, blkptr_t *blkptr) 1322fa9e4066Sahrens { 1323fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 1324fa9e4066Sahrens dmu_buf_impl_t *db, *odb; 1325fa9e4066Sahrens 1326fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1327fa9e4066Sahrens ASSERT(dn->dn_type != DMU_OT_NONE); 1328fa9e4066Sahrens 1329fa9e4066Sahrens db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1330fa9e4066Sahrens 1331fa9e4066Sahrens db->db_objset = os; 1332fa9e4066Sahrens db->db.db_object = dn->dn_object; 1333fa9e4066Sahrens db->db_level = level; 1334fa9e4066Sahrens db->db_blkid = blkid; 1335ea8dc4b6Seschrock db->db_dirtied = 0; 1336ea8dc4b6Seschrock db->db_dirtycnt = 0; 1337ea8dc4b6Seschrock db->db_dnode = dn; 1338ea8dc4b6Seschrock db->db_parent = parent; 1339ea8dc4b6Seschrock db->db_blkptr = blkptr; 1340fa9e4066Sahrens 1341ea8dc4b6Seschrock bzero(&db->db_d, sizeof (db->db_d)); 1342ea8dc4b6Seschrock 1343ea8dc4b6Seschrock if (blkid == DB_BONUS_BLKID) { 1344ea8dc4b6Seschrock ASSERT3P(parent, ==, dn->dn_dbuf); 1345fa9e4066Sahrens db->db.db_size = dn->dn_bonuslen; 1346fa9e4066Sahrens db->db.db_offset = DB_BONUS_BLKID; 1347ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1348ea8dc4b6Seschrock /* the bonus dbuf is not placed in the hash table */ 1349ea8dc4b6Seschrock return (db); 1350fa9e4066Sahrens } else { 1351fa9e4066Sahrens int blocksize = 1352fa9e4066Sahrens db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; 1353fa9e4066Sahrens db->db.db_size = blocksize; 1354fa9e4066Sahrens db->db.db_offset = db->db_blkid * blocksize; 1355fa9e4066Sahrens } 1356fa9e4066Sahrens 1357fa9e4066Sahrens /* 1358fa9e4066Sahrens * Hold the dn_dbufs_mtx while we get the new dbuf 1359fa9e4066Sahrens * in the hash table *and* added to the dbufs list. 1360fa9e4066Sahrens * This prevents a possible deadlock with someone 1361fa9e4066Sahrens * trying to look up this dbuf before its added to the 1362fa9e4066Sahrens * dn_dbufs list. 1363fa9e4066Sahrens */ 1364fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 1365ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1366fa9e4066Sahrens if ((odb = dbuf_hash_insert(db)) != NULL) { 1367fa9e4066Sahrens /* someone else inserted it first */ 1368fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 1369fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1370fa9e4066Sahrens return (odb); 1371fa9e4066Sahrens } 1372fa9e4066Sahrens list_insert_head(&dn->dn_dbufs, db); 1373ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1374fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1375fa9e4066Sahrens 1376fa9e4066Sahrens if (parent && parent != dn->dn_dbuf) 1377fa9e4066Sahrens dbuf_add_ref(parent, db); 1378fa9e4066Sahrens 1379ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1380ea8dc4b6Seschrock refcount_count(&dn->dn_holds) > 0); 1381fa9e4066Sahrens (void) refcount_add(&dn->dn_holds, db); 1382fa9e4066Sahrens 1383fa9e4066Sahrens dprintf_dbuf(db, "db=%p\n", db); 1384fa9e4066Sahrens 1385fa9e4066Sahrens return (db); 1386fa9e4066Sahrens } 1387fa9e4066Sahrens 1388fa9e4066Sahrens static int 1389ea8dc4b6Seschrock dbuf_do_evict(void *private) 1390fa9e4066Sahrens { 1391ea8dc4b6Seschrock arc_buf_t *buf = private; 1392ea8dc4b6Seschrock dmu_buf_impl_t *db = buf->b_private; 1393fa9e4066Sahrens 1394ea8dc4b6Seschrock if (!MUTEX_HELD(&db->db_mtx)) 1395ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 1396ea8dc4b6Seschrock 1397ea8dc4b6Seschrock ASSERT(db->db_buf == buf); 1398ea8dc4b6Seschrock ASSERT(refcount_is_zero(&db->db_holds)); 1399ea8dc4b6Seschrock 1400ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 1401ea8dc4b6Seschrock ASSERT(db->db_state == DB_CACHED); 14029c9dc39aSek110237 DBUF_VERIFY(db); 1403ea8dc4b6Seschrock db->db_buf = NULL; 1404ea8dc4b6Seschrock dbuf_evict(db); 1405ea8dc4b6Seschrock } else { 1406ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1407ea8dc4b6Seschrock dbuf_destroy(db); 1408fa9e4066Sahrens } 1409ea8dc4b6Seschrock return (0); 1410fa9e4066Sahrens } 1411fa9e4066Sahrens 1412fa9e4066Sahrens static void 1413fa9e4066Sahrens dbuf_destroy(dmu_buf_impl_t *db) 1414fa9e4066Sahrens { 1415fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1416fa9e4066Sahrens 1417ea8dc4b6Seschrock if (db->db_blkid != DB_BONUS_BLKID) { 1418ea8dc4b6Seschrock dnode_t *dn = db->db_dnode; 1419ea8dc4b6Seschrock 1420ea8dc4b6Seschrock /* 1421ea8dc4b6Seschrock * If this dbuf is still on the dn_dbufs list, 1422ea8dc4b6Seschrock * remove it from that list. 1423ea8dc4b6Seschrock */ 1424ea8dc4b6Seschrock if (list_link_active(&db->db_link)) { 1425ea8dc4b6Seschrock mutex_enter(&dn->dn_dbufs_mtx); 1426ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1427ea8dc4b6Seschrock mutex_exit(&dn->dn_dbufs_mtx); 1428ea8dc4b6Seschrock 1429ea8dc4b6Seschrock dnode_rele(dn, db); 1430ea8dc4b6Seschrock } 1431ea8dc4b6Seschrock dbuf_hash_remove(db); 1432ea8dc4b6Seschrock } 1433ea8dc4b6Seschrock db->db_parent = NULL; 1434ea8dc4b6Seschrock db->db_dnode = NULL; 1435ea8dc4b6Seschrock db->db_buf = NULL; 1436ea8dc4b6Seschrock 1437fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1438fa9e4066Sahrens ASSERT(db->db_hash_next == NULL); 1439fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 1440fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1441fa9e4066Sahrens 1442fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 1443fa9e4066Sahrens } 1444fa9e4066Sahrens 1445fa9e4066Sahrens void 1446fa9e4066Sahrens dbuf_prefetch(dnode_t *dn, uint64_t blkid) 1447fa9e4066Sahrens { 1448fa9e4066Sahrens dmu_buf_impl_t *db, *parent = NULL; 1449fa9e4066Sahrens blkptr_t *bp = NULL; 1450fa9e4066Sahrens 1451fa9e4066Sahrens ASSERT(blkid != DB_BONUS_BLKID); 1452fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1453fa9e4066Sahrens 1454fa9e4066Sahrens if (dnode_block_freed(dn, blkid)) 1455fa9e4066Sahrens return; 1456fa9e4066Sahrens 1457fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1458fa9e4066Sahrens if (db = dbuf_find(dn, 0, blkid)) { 1459fa9e4066Sahrens /* 1460fa9e4066Sahrens * This dbuf is already in the cache. We assume that 1461fa9e4066Sahrens * it is already CACHED, or else about to be either 1462fa9e4066Sahrens * read or filled. 1463fa9e4066Sahrens */ 1464fa9e4066Sahrens mutex_exit(&db->db_mtx); 1465fa9e4066Sahrens return; 1466fa9e4066Sahrens } 1467fa9e4066Sahrens 1468fa9e4066Sahrens if (dbuf_findbp(dn, 0, blkid, TRUE, &parent, &bp) == 0) { 1469fa9e4066Sahrens if (bp && !BP_IS_HOLE(bp)) { 1470ea8dc4b6Seschrock zbookmark_t zb; 1471ea8dc4b6Seschrock zb.zb_objset = dn->dn_objset->os_dsl_dataset ? 1472ea8dc4b6Seschrock dn->dn_objset->os_dsl_dataset->ds_object : 0; 1473ea8dc4b6Seschrock zb.zb_object = dn->dn_object; 1474ea8dc4b6Seschrock zb.zb_level = 0; 1475ea8dc4b6Seschrock zb.zb_blkid = blkid; 1476ea8dc4b6Seschrock 1477fa9e4066Sahrens (void) arc_read(NULL, dn->dn_objset->os_spa, bp, 1478fa9e4066Sahrens dmu_ot[dn->dn_type].ot_byteswap, 1479fa9e4066Sahrens NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 1480fa9e4066Sahrens ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1481ea8dc4b6Seschrock (ARC_NOWAIT | ARC_PREFETCH), &zb); 1482fa9e4066Sahrens } 1483*c543ec06Sahrens if (parent) 1484ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1485fa9e4066Sahrens } 1486fa9e4066Sahrens } 1487fa9e4066Sahrens 1488fa9e4066Sahrens /* 1489fa9e4066Sahrens * Returns with db_holds incremented, and db_mtx not held. 1490fa9e4066Sahrens * Note: dn_struct_rwlock must be held. 1491fa9e4066Sahrens */ 1492fa9e4066Sahrens int 1493fa9e4066Sahrens dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1494fa9e4066Sahrens void *tag, dmu_buf_impl_t **dbp) 1495fa9e4066Sahrens { 1496fa9e4066Sahrens dmu_buf_impl_t *db, *parent = NULL; 1497fa9e4066Sahrens 1498ea8dc4b6Seschrock ASSERT(blkid != DB_BONUS_BLKID); 1499fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1500fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, level); 1501fa9e4066Sahrens 1502fa9e4066Sahrens *dbp = NULL; 1503ea8dc4b6Seschrock top: 1504fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1505fa9e4066Sahrens db = dbuf_find(dn, level, blkid); 1506fa9e4066Sahrens 1507fa9e4066Sahrens if (db == NULL) { 1508fa9e4066Sahrens blkptr_t *bp = NULL; 1509fa9e4066Sahrens int err; 1510fa9e4066Sahrens 1511*c543ec06Sahrens ASSERT3P(parent, ==, NULL); 1512fa9e4066Sahrens err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1513fa9e4066Sahrens if (fail_sparse) { 1514fa9e4066Sahrens if (err == 0 && bp && BP_IS_HOLE(bp)) 1515fa9e4066Sahrens err = ENOENT; 1516fa9e4066Sahrens if (err) { 1517*c543ec06Sahrens if (parent) 1518ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1519fa9e4066Sahrens return (err); 1520fa9e4066Sahrens } 1521fa9e4066Sahrens } 1522ea8dc4b6Seschrock if (err && err != ENOENT) 1523ea8dc4b6Seschrock return (err); 1524fa9e4066Sahrens db = dbuf_create(dn, level, blkid, parent, bp); 1525fa9e4066Sahrens } 1526fa9e4066Sahrens 1527ea8dc4b6Seschrock if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1528ea8dc4b6Seschrock arc_buf_add_ref(db->db_buf, db); 1529ea8dc4b6Seschrock if (db->db_buf->b_data == NULL) { 1530ea8dc4b6Seschrock dbuf_clear(db); 1531*c543ec06Sahrens if (parent) { 1532*c543ec06Sahrens dbuf_rele(parent, NULL); 1533*c543ec06Sahrens parent = NULL; 1534*c543ec06Sahrens } 1535ea8dc4b6Seschrock goto top; 1536ea8dc4b6Seschrock } 1537ea8dc4b6Seschrock ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1538ea8dc4b6Seschrock } 1539ea8dc4b6Seschrock 1540ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1541ea8dc4b6Seschrock 1542fa9e4066Sahrens /* 1543fa9e4066Sahrens * If this buffer is currently syncing out, and we are 1544fa9e4066Sahrens * are still referencing it from db_data, we need to make 1545fa9e4066Sahrens * a copy of it in case we decide we want to dirty it 1546fa9e4066Sahrens * again in this txg. 1547fa9e4066Sahrens */ 1548fa9e4066Sahrens if (db->db_level == 0 && db->db_state == DB_CACHED && 1549ea8dc4b6Seschrock dn->dn_object != DMU_META_DNODE_OBJECT && 1550fa9e4066Sahrens db->db_data_pending == db->db_buf) { 1551fa9e4066Sahrens int size = (db->db_blkid == DB_BONUS_BLKID) ? 1552fa9e4066Sahrens DN_MAX_BONUSLEN : db->db.db_size; 1553fa9e4066Sahrens 1554fa9e4066Sahrens dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 1555fa9e4066Sahrens size, db)); 1556fa9e4066Sahrens bcopy(db->db_data_pending->b_data, db->db.db_data, 1557fa9e4066Sahrens db->db.db_size); 1558fa9e4066Sahrens } 1559fa9e4066Sahrens 1560ea8dc4b6Seschrock (void) refcount_add(&db->db_holds, tag); 1561fa9e4066Sahrens dbuf_update_data(db); 15629c9dc39aSek110237 DBUF_VERIFY(db); 1563fa9e4066Sahrens mutex_exit(&db->db_mtx); 1564fa9e4066Sahrens 1565fa9e4066Sahrens /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1566*c543ec06Sahrens if (parent) 1567ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1568fa9e4066Sahrens 1569fa9e4066Sahrens ASSERT3P(db->db_dnode, ==, dn); 1570fa9e4066Sahrens ASSERT3U(db->db_blkid, ==, blkid); 1571fa9e4066Sahrens ASSERT3U(db->db_level, ==, level); 1572fa9e4066Sahrens *dbp = db; 1573fa9e4066Sahrens 1574fa9e4066Sahrens return (0); 1575fa9e4066Sahrens } 1576fa9e4066Sahrens 1577fa9e4066Sahrens dmu_buf_impl_t * 1578ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1579fa9e4066Sahrens { 1580fa9e4066Sahrens dmu_buf_impl_t *db; 1581ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1582ea8dc4b6Seschrock return (err ? NULL : db); 1583fa9e4066Sahrens } 1584fa9e4066Sahrens 1585fa9e4066Sahrens dmu_buf_impl_t * 1586fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1587fa9e4066Sahrens { 1588fa9e4066Sahrens dmu_buf_impl_t *db; 1589ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1590ea8dc4b6Seschrock return (err ? NULL : db); 1591fa9e4066Sahrens } 1592fa9e4066Sahrens 1593fa9e4066Sahrens dmu_buf_impl_t * 1594ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn) 1595fa9e4066Sahrens { 1596ea8dc4b6Seschrock dmu_buf_impl_t *db = dn->dn_bonus; 1597ea8dc4b6Seschrock 1598ea8dc4b6Seschrock ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1599ea8dc4b6Seschrock 1600ea8dc4b6Seschrock ASSERT(dn->dn_bonus == NULL); 1601ea8dc4b6Seschrock db = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL); 1602fa9e4066Sahrens return (db); 1603fa9e4066Sahrens } 1604fa9e4066Sahrens 1605ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref 1606fa9e4066Sahrens void 1607fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 1608fa9e4066Sahrens { 1609ea8dc4b6Seschrock int64_t holds = refcount_add(&db->db_holds, tag); 1610ea8dc4b6Seschrock ASSERT(holds > 1); 1611fa9e4066Sahrens } 1612fa9e4066Sahrens 1613ea8dc4b6Seschrock #pragma weak dmu_buf_rele = dbuf_rele 1614fa9e4066Sahrens void 1615ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag) 1616fa9e4066Sahrens { 1617fa9e4066Sahrens int64_t holds; 1618fa9e4066Sahrens 1619fa9e4066Sahrens mutex_enter(&db->db_mtx); 16209c9dc39aSek110237 DBUF_VERIFY(db); 1621fa9e4066Sahrens 1622fa9e4066Sahrens holds = refcount_remove(&db->db_holds, tag); 1623ea8dc4b6Seschrock ASSERT(holds >= 0); 1624fa9e4066Sahrens 1625fa9e4066Sahrens if (holds == db->db_dirtycnt && 1626fa9e4066Sahrens db->db_level == 0 && db->db_d.db_immediate_evict) 1627fa9e4066Sahrens dbuf_evict_user(db); 1628ea8dc4b6Seschrock 1629ea8dc4b6Seschrock if (holds == 0) { 1630ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 1631ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1632ea8dc4b6Seschrock dnode_rele(db->db_dnode, db); 1633ea8dc4b6Seschrock } else if (db->db_buf == NULL) { 1634ea8dc4b6Seschrock /* 1635ea8dc4b6Seschrock * This is a special case: we never associated this 1636ea8dc4b6Seschrock * dbuf with any data allocated from the ARC. 1637ea8dc4b6Seschrock */ 1638ea8dc4b6Seschrock ASSERT3U(db->db_state, ==, DB_UNCACHED); 1639ea8dc4b6Seschrock dbuf_evict(db); 1640ea8dc4b6Seschrock } else if (arc_released(db->db_buf)) { 1641ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1642ea8dc4b6Seschrock /* 1643ea8dc4b6Seschrock * This dbuf has anonymous data associated with it. 1644ea8dc4b6Seschrock */ 1645ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1646ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1647ea8dc4b6Seschrock dbuf_evict(db); 1648ea8dc4b6Seschrock } else { 1649ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); 1650fa9e4066Sahrens mutex_exit(&db->db_mtx); 1651fa9e4066Sahrens } 1652ea8dc4b6Seschrock } else { 1653ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1654fa9e4066Sahrens } 1655fa9e4066Sahrens } 1656fa9e4066Sahrens 1657fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount 1658fa9e4066Sahrens uint64_t 1659fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db) 1660fa9e4066Sahrens { 1661fa9e4066Sahrens return (refcount_count(&db->db_holds)); 1662fa9e4066Sahrens } 1663fa9e4066Sahrens 1664fa9e4066Sahrens void * 1665fa9e4066Sahrens dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1666fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 1667fa9e4066Sahrens { 1668fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1669fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 1670fa9e4066Sahrens } 1671fa9e4066Sahrens 1672fa9e4066Sahrens void * 1673fa9e4066Sahrens dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1674fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 1675fa9e4066Sahrens { 1676fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1677fa9e4066Sahrens 1678fa9e4066Sahrens db->db_d.db_immediate_evict = TRUE; 1679fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1680fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 1681fa9e4066Sahrens } 1682fa9e4066Sahrens 1683fa9e4066Sahrens void * 1684fa9e4066Sahrens dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 1685fa9e4066Sahrens void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 1686fa9e4066Sahrens { 1687fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1688fa9e4066Sahrens ASSERT(db->db_level == 0); 1689fa9e4066Sahrens 1690fa9e4066Sahrens ASSERT((user_ptr == NULL) == (evict_func == NULL)); 1691fa9e4066Sahrens 1692fa9e4066Sahrens mutex_enter(&db->db_mtx); 1693fa9e4066Sahrens 1694fa9e4066Sahrens if (db->db_d.db_user_ptr == old_user_ptr) { 1695fa9e4066Sahrens db->db_d.db_user_ptr = user_ptr; 1696fa9e4066Sahrens db->db_d.db_user_data_ptr_ptr = user_data_ptr_ptr; 1697fa9e4066Sahrens db->db_d.db_evict_func = evict_func; 1698fa9e4066Sahrens 1699fa9e4066Sahrens dbuf_update_data(db); 1700fa9e4066Sahrens } else { 1701fa9e4066Sahrens old_user_ptr = db->db_d.db_user_ptr; 1702fa9e4066Sahrens } 1703fa9e4066Sahrens 1704fa9e4066Sahrens mutex_exit(&db->db_mtx); 1705fa9e4066Sahrens return (old_user_ptr); 1706fa9e4066Sahrens } 1707fa9e4066Sahrens 1708fa9e4066Sahrens void * 1709fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake) 1710fa9e4066Sahrens { 1711fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1712fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1713fa9e4066Sahrens 1714fa9e4066Sahrens return (db->db_d.db_user_ptr); 1715fa9e4066Sahrens } 1716fa9e4066Sahrens 1717fa9e4066Sahrens void 1718fa9e4066Sahrens dbuf_sync(dmu_buf_impl_t *db, zio_t *zio, dmu_tx_t *tx) 1719fa9e4066Sahrens { 1720fa9e4066Sahrens arc_buf_t **data; 1721fa9e4066Sahrens uint64_t txg = tx->tx_txg; 1722fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1723fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 172423b11526Smaybee int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 1725ea8dc4b6Seschrock int checksum, compress; 1726ea8dc4b6Seschrock zbookmark_t zb; 1727fa9e4066Sahrens int blksz; 1728fa9e4066Sahrens 1729fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 1730fa9e4066Sahrens 1731fa9e4066Sahrens dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 1732fa9e4066Sahrens 1733fa9e4066Sahrens mutex_enter(&db->db_mtx); 1734fa9e4066Sahrens /* 1735fa9e4066Sahrens * To be synced, we must be dirtied. But we 1736fa9e4066Sahrens * might have been freed after the dirty. 1737fa9e4066Sahrens */ 1738fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 1739fa9e4066Sahrens /* This buffer has been freed since it was dirtied */ 1740fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1741fa9e4066Sahrens } else if (db->db_state == DB_FILL) { 1742fa9e4066Sahrens /* This buffer was freed and is now being re-filled */ 1743fa9e4066Sahrens ASSERT(db->db.db_data != db->db_d.db_data_old[txg&TXG_MASK]); 1744fa9e4066Sahrens } else { 1745fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 1746fa9e4066Sahrens } 17479c9dc39aSek110237 DBUF_VERIFY(db); 1748fa9e4066Sahrens 1749fa9e4066Sahrens /* 1750fa9e4066Sahrens * Don't need a lock on db_dirty (dn_mtx), because it can't 1751fa9e4066Sahrens * be modified yet. 1752fa9e4066Sahrens */ 1753fa9e4066Sahrens 1754ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 1755ea8dc4b6Seschrock void **datap = &db->db_d.db_data_old[txg&TXG_MASK]; 1756ea8dc4b6Seschrock /* 1757ea8dc4b6Seschrock * Simply copy the bonus data into the dnode. It will 1758ea8dc4b6Seschrock * be written out when the dnode is synced (and it will 1759ea8dc4b6Seschrock * be synced, since it must have been dirty for dbuf_sync 1760ea8dc4b6Seschrock * to be called). 1761ea8dc4b6Seschrock */ 1762ea8dc4b6Seschrock /* 1763ea8dc4b6Seschrock * Use dn_phys->dn_bonuslen since db.db_size is the length 1764ea8dc4b6Seschrock * of the bonus buffer in the open transaction rather than 1765ea8dc4b6Seschrock * the syncing transaction. 1766ea8dc4b6Seschrock */ 1767ea8dc4b6Seschrock ASSERT(*datap != NULL); 1768ea8dc4b6Seschrock ASSERT3U(db->db_level, ==, 0); 1769ea8dc4b6Seschrock ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 1770ea8dc4b6Seschrock bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 1771ea8dc4b6Seschrock if (*datap != db->db.db_data) 1772ea8dc4b6Seschrock zio_buf_free(*datap, DN_MAX_BONUSLEN); 1773ea8dc4b6Seschrock db->db_d.db_data_old[txg&TXG_MASK] = NULL; 1774ea8dc4b6Seschrock db->db_data_pending = NULL; 1775ea8dc4b6Seschrock if (db->db_dirtied == txg) 1776ea8dc4b6Seschrock db->db_dirtied = 0; 1777ea8dc4b6Seschrock ASSERT(db->db_dirtycnt > 0); 1778ea8dc4b6Seschrock db->db_dirtycnt -= 1; 1779ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1780ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 1781ea8dc4b6Seschrock return; 1782ea8dc4b6Seschrock } 1783ea8dc4b6Seschrock 1784fa9e4066Sahrens if (db->db_level == 0) { 1785ea8dc4b6Seschrock data = (arc_buf_t **)&db->db_d.db_data_old[txg&TXG_MASK]; 1786fa9e4066Sahrens blksz = arc_buf_size(*data); 1787fa9e4066Sahrens /* 1788fa9e4066Sahrens * If this buffer is currently "in use" (i.e., there are 1789fa9e4066Sahrens * active holds and db_data still references it), then make 1790fa9e4066Sahrens * a copy before we start the write so that any modifications 1791fa9e4066Sahrens * from the open txg will not leak into this write. 1792fa9e4066Sahrens * 1793fa9e4066Sahrens * NOTE: this copy does not need to be made for objects only 1794fa9e4066Sahrens * modified in the syncing context (e.g. DNONE_DNODE blocks) 1795fa9e4066Sahrens * or if there is no actual write involved (bonus blocks). 1796fa9e4066Sahrens */ 1797ea8dc4b6Seschrock if (dn->dn_object != DMU_META_DNODE_OBJECT && 1798ea8dc4b6Seschrock db->db_d.db_overridden_by[txg&TXG_MASK] == NULL) { 1799fa9e4066Sahrens if (refcount_count(&db->db_holds) > 1 && 1800fa9e4066Sahrens *data == db->db_buf) { 1801ea8dc4b6Seschrock *data = arc_buf_alloc(os->os_spa, blksz, db); 1802fa9e4066Sahrens bcopy(db->db.db_data, (*data)->b_data, blksz); 1803fa9e4066Sahrens } 1804fa9e4066Sahrens db->db_data_pending = *data; 1805ea8dc4b6Seschrock } else if (dn->dn_object == DMU_META_DNODE_OBJECT) { 1806fa9e4066Sahrens /* 1807fa9e4066Sahrens * Private object buffers are released here rather 1808fa9e4066Sahrens * than in dbuf_dirty() since they are only modified 1809fa9e4066Sahrens * in the syncing context and we don't want the 1810fa9e4066Sahrens * overhead of making multiple copies of the data. 1811fa9e4066Sahrens */ 1812fa9e4066Sahrens arc_release(db->db_buf, db); 1813fa9e4066Sahrens } 1814fa9e4066Sahrens } else { 1815fa9e4066Sahrens data = &db->db_buf; 1816fa9e4066Sahrens if (*data == NULL) { 1817fa9e4066Sahrens /* 1818fa9e4066Sahrens * This can happen if we dirty and then free 1819fa9e4066Sahrens * the level-0 data blocks in the same txg. So 1820fa9e4066Sahrens * this indirect remains unchanged. 1821fa9e4066Sahrens */ 1822fa9e4066Sahrens if (db->db_dirtied == txg) 1823fa9e4066Sahrens db->db_dirtied = 0; 1824fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1825fa9e4066Sahrens db->db_dirtycnt -= 1; 1826fa9e4066Sahrens mutex_exit(&db->db_mtx); 1827ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 1828fa9e4066Sahrens return; 1829fa9e4066Sahrens } 1830fa9e4066Sahrens blksz = db->db.db_size; 1831fa9e4066Sahrens ASSERT3U(blksz, ==, 1<<dn->dn_phys->dn_indblkshift); 1832fa9e4066Sahrens } 1833fa9e4066Sahrens 1834fa9e4066Sahrens ASSERT(*data != NULL); 1835fa9e4066Sahrens 1836ea8dc4b6Seschrock if (db->db_level > 0 && !arc_released(db->db_buf)) { 1837fa9e4066Sahrens /* 1838fa9e4066Sahrens * This indirect buffer was marked dirty, but 1839fa9e4066Sahrens * never modified (if it had been modified, then 1840fa9e4066Sahrens * we would have released the buffer). There is 1841fa9e4066Sahrens * no reason to write anything. 1842fa9e4066Sahrens */ 1843fa9e4066Sahrens db->db_data_pending = NULL; 1844fa9e4066Sahrens if (db->db_dirtied == txg) 1845fa9e4066Sahrens db->db_dirtied = 0; 1846fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1847fa9e4066Sahrens db->db_dirtycnt -= 1; 1848fa9e4066Sahrens mutex_exit(&db->db_mtx); 1849ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 1850fa9e4066Sahrens return; 1851fa9e4066Sahrens } else if (db->db_blkptr == NULL && 1852fa9e4066Sahrens db->db_level == dn->dn_phys->dn_nlevels-1 && 1853fa9e4066Sahrens db->db_blkid < dn->dn_phys->dn_nblkptr) { 1854fa9e4066Sahrens /* 1855fa9e4066Sahrens * This buffer was allocated at a time when there was 1856fa9e4066Sahrens * no available blkptrs from the dnode, or it was 1857fa9e4066Sahrens * inappropriate to hook it in (i.e., nlevels mis-match). 1858fa9e4066Sahrens */ 1859fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 1860fa9e4066Sahrens ASSERT(db->db_parent == NULL); 1861fa9e4066Sahrens db->db_parent = dn->dn_dbuf; 1862fa9e4066Sahrens db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 18639c9dc39aSek110237 DBUF_VERIFY(db); 1864fa9e4066Sahrens mutex_exit(&db->db_mtx); 1865fa9e4066Sahrens } else if (db->db_blkptr == NULL) { 1866fa9e4066Sahrens dmu_buf_impl_t *parent = db->db_parent; 1867fa9e4066Sahrens 1868fa9e4066Sahrens mutex_exit(&db->db_mtx); 1869fa9e4066Sahrens ASSERT(dn->dn_phys->dn_nlevels > 1); 1870fa9e4066Sahrens if (parent == NULL) { 1871fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 1872fa9e4066Sahrens (void) dbuf_hold_impl(dn, db->db_level+1, 1873ea8dc4b6Seschrock db->db_blkid >> epbs, FALSE, FTAG, &parent); 1874fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1875fa9e4066Sahrens dbuf_add_ref(parent, db); 1876fa9e4066Sahrens db->db_parent = parent; 1877ea8dc4b6Seschrock dbuf_rele(parent, FTAG); 1878fa9e4066Sahrens } 1879ea8dc4b6Seschrock (void) dbuf_read(parent, NULL, DB_RF_MUST_SUCCEED); 1880fa9e4066Sahrens } else { 1881fa9e4066Sahrens mutex_exit(&db->db_mtx); 1882fa9e4066Sahrens } 1883fa9e4066Sahrens 1884ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || db->db_parent != NULL); 1885fa9e4066Sahrens 188623b11526Smaybee if (db->db_level > 0 && 188723b11526Smaybee db->db_blkid > dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)) { 188823b11526Smaybee /* 188923b11526Smaybee * Don't write indirect blocks past EOF. 189023b11526Smaybee * We get these when we truncate a file *after* dirtying 189123b11526Smaybee * blocks in the truncate range (we undirty the level 0 189223b11526Smaybee * blocks in dbuf_free_range(), but not the indirects). 189323b11526Smaybee */ 189423b11526Smaybee #ifdef ZFS_DEBUG 189523b11526Smaybee /* 189623b11526Smaybee * Verify that this indirect block is empty. 189723b11526Smaybee */ 189823b11526Smaybee blkptr_t *bplist; 189923b11526Smaybee int i; 190023b11526Smaybee 190123b11526Smaybee mutex_enter(&db->db_mtx); 190223b11526Smaybee bplist = db->db.db_data; 190323b11526Smaybee for (i = 0; i < (1 << epbs); i++) { 190423b11526Smaybee if (!BP_IS_HOLE(&bplist[i])) { 190523b11526Smaybee panic("data past EOF: " 190634f18512Seschrock "db=%p level=%d id=%llu i=%d\n", 190734f18512Seschrock db, db->db_level, 190834f18512Seschrock (u_longlong_t)db->db_blkid, i); 190923b11526Smaybee } 191023b11526Smaybee } 191123b11526Smaybee mutex_exit(&db->db_mtx); 191223b11526Smaybee #endif 191323b11526Smaybee ASSERT(db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)); 191423b11526Smaybee mutex_enter(&db->db_mtx); 191523b11526Smaybee db->db_dirtycnt -= 1; 191623b11526Smaybee mutex_exit(&db->db_mtx); 1917ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 191823b11526Smaybee return; 191923b11526Smaybee } 192023b11526Smaybee 1921fa9e4066Sahrens if (db->db_parent != dn->dn_dbuf) { 1922fa9e4066Sahrens dmu_buf_impl_t *parent = db->db_parent; 1923fa9e4066Sahrens 1924fa9e4066Sahrens mutex_enter(&db->db_mtx); 1925fa9e4066Sahrens ASSERT(db->db_level == parent->db_level-1); 1926fa9e4066Sahrens ASSERT(list_link_active(&parent->db_dirty_node[txg&TXG_MASK])); 1927fa9e4066Sahrens /* 1928ea8dc4b6Seschrock * We may have read this indirect block after we dirtied it, 1929fa9e4066Sahrens * so never released it from the cache. 1930fa9e4066Sahrens */ 1931ea8dc4b6Seschrock arc_release(parent->db_buf, db->db_parent); 1932fa9e4066Sahrens 1933fa9e4066Sahrens db->db_blkptr = (blkptr_t *)parent->db.db_data + 1934fa9e4066Sahrens (db->db_blkid & ((1ULL << epbs) - 1)); 19359c9dc39aSek110237 DBUF_VERIFY(db); 1936fa9e4066Sahrens mutex_exit(&db->db_mtx); 1937fa9e4066Sahrens #ifdef ZFS_DEBUG 1938ea8dc4b6Seschrock } else { 1939fa9e4066Sahrens /* 1940fa9e4066Sahrens * We don't need to dnode_setdirty(dn) because if we got 1941fa9e4066Sahrens * here then the parent is already dirty. 1942fa9e4066Sahrens */ 1943fa9e4066Sahrens ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1); 1944fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 1945fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 1946fa9e4066Sahrens #endif 1947ea8dc4b6Seschrock } 1948ea8dc4b6Seschrock ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1949ea8dc4b6Seschrock 1950fa9e4066Sahrens if (db->db_level == 0 && 1951fa9e4066Sahrens db->db_d.db_overridden_by[txg&TXG_MASK] != NULL) { 1952ea8dc4b6Seschrock arc_buf_t **old = 1953ea8dc4b6Seschrock (arc_buf_t **)&db->db_d.db_data_old[txg&TXG_MASK]; 1954fa9e4066Sahrens blkptr_t **bpp = &db->db_d.db_overridden_by[txg&TXG_MASK]; 1955fa9e4066Sahrens int old_size = BP_GET_ASIZE(db->db_blkptr); 1956fa9e4066Sahrens int new_size = BP_GET_ASIZE(*bpp); 1957fa9e4066Sahrens 1958fa9e4066Sahrens ASSERT(db->db_blkid != DB_BONUS_BLKID); 1959fa9e4066Sahrens 1960fa9e4066Sahrens dnode_diduse_space(dn, new_size-old_size); 1961fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1962fa9e4066Sahrens if (db->db_blkid > dn->dn_phys->dn_maxblkid) 1963fa9e4066Sahrens dn->dn_phys->dn_maxblkid = db->db_blkid; 1964fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1965fa9e4066Sahrens 1966fa9e4066Sahrens dsl_dataset_block_born(os->os_dsl_dataset, *bpp, tx); 1967fa9e4066Sahrens if (!BP_IS_HOLE(db->db_blkptr)) 1968fa9e4066Sahrens dsl_dataset_block_kill(os->os_dsl_dataset, 1969fa9e4066Sahrens db->db_blkptr, os->os_synctx); 1970fa9e4066Sahrens 1971fa9e4066Sahrens mutex_enter(&db->db_mtx); 1972fa9e4066Sahrens *db->db_blkptr = **bpp; 1973fa9e4066Sahrens kmem_free(*bpp, sizeof (blkptr_t)); 1974fa9e4066Sahrens *bpp = NULL; 1975fa9e4066Sahrens 1976fa9e4066Sahrens if (*old != db->db_buf) 1977ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(*old, db) == 1); 1978ea8dc4b6Seschrock else if (!BP_IS_HOLE(db->db_blkptr)) 1979ea8dc4b6Seschrock arc_set_callback(db->db_buf, dbuf_do_evict, db); 1980ea8dc4b6Seschrock else 1981ea8dc4b6Seschrock ASSERT(arc_released(db->db_buf)); 1982fa9e4066Sahrens *old = NULL; 1983fa9e4066Sahrens db->db_data_pending = NULL; 1984fa9e4066Sahrens 1985fa9e4066Sahrens cv_broadcast(&db->db_changed); 1986fa9e4066Sahrens 1987fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1988fa9e4066Sahrens db->db_dirtycnt -= 1; 1989fa9e4066Sahrens mutex_exit(&db->db_mtx); 1990ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 1991ea8dc4b6Seschrock return; 1992ea8dc4b6Seschrock } 1993fa9e4066Sahrens 1994fa9e4066Sahrens if (db->db_level > 0) { 1995fa9e4066Sahrens /* 1996fa9e4066Sahrens * XXX -- we should design a compression algorithm 1997fa9e4066Sahrens * that specializes in arrays of bps. 1998fa9e4066Sahrens */ 1999fa9e4066Sahrens checksum = ZIO_CHECKSUM_FLETCHER_4; 2000ea8dc4b6Seschrock compress = ZIO_COMPRESS_LZJB; 2001fa9e4066Sahrens } else { 2002fa9e4066Sahrens /* 2003fa9e4066Sahrens * Allow dnode settings to override objset settings, 2004fa9e4066Sahrens * except for metadata checksums. 2005fa9e4066Sahrens */ 2006fa9e4066Sahrens if (dmu_ot[dn->dn_type].ot_metadata) { 2007fa9e4066Sahrens checksum = os->os_md_checksum; 2008fa9e4066Sahrens compress = zio_compress_select(dn->dn_compress, 2009fa9e4066Sahrens os->os_md_compress); 2010fa9e4066Sahrens } else { 2011fa9e4066Sahrens checksum = zio_checksum_select(dn->dn_checksum, 2012fa9e4066Sahrens os->os_checksum); 2013fa9e4066Sahrens compress = zio_compress_select(dn->dn_compress, 2014fa9e4066Sahrens os->os_compress); 2015fa9e4066Sahrens } 2016fa9e4066Sahrens } 2017fa9e4066Sahrens #ifdef ZFS_DEBUG 2018fa9e4066Sahrens if (db->db_parent) { 2019fa9e4066Sahrens ASSERT(list_link_active( 2020fa9e4066Sahrens &db->db_parent->db_dirty_node[txg&TXG_MASK])); 2021fa9e4066Sahrens ASSERT(db->db_parent == dn->dn_dbuf || 2022fa9e4066Sahrens db->db_parent->db_level > 0); 2023ea8dc4b6Seschrock if (dn->dn_object == DMU_META_DNODE_OBJECT || db->db_level > 0) 2024fa9e4066Sahrens ASSERT(*data == db->db_buf); 2025fa9e4066Sahrens } 2026fa9e4066Sahrens #endif 2027fa9e4066Sahrens ASSERT3U(db->db_blkptr->blk_birth, <=, tx->tx_txg); 2028ea8dc4b6Seschrock zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0; 2029ea8dc4b6Seschrock zb.zb_object = db->db.db_object; 2030ea8dc4b6Seschrock zb.zb_level = db->db_level; 2031ea8dc4b6Seschrock zb.zb_blkid = db->db_blkid; 2032fa9e4066Sahrens (void) arc_write(zio, os->os_spa, checksum, compress, txg, 2033fa9e4066Sahrens db->db_blkptr, *data, dbuf_write_done, db, 2034ea8dc4b6Seschrock ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, ARC_NOWAIT, &zb); 2035fa9e4066Sahrens /* 2036fa9e4066Sahrens * We can't access db after arc_write, since it could finish 2037fa9e4066Sahrens * and be freed, and we have no locks on it. 2038fa9e4066Sahrens */ 2039fa9e4066Sahrens } 2040fa9e4066Sahrens 2041fa9e4066Sahrens struct dbuf_arg { 2042fa9e4066Sahrens objset_impl_t *os; 2043fa9e4066Sahrens blkptr_t bp; 2044fa9e4066Sahrens }; 2045fa9e4066Sahrens 2046fa9e4066Sahrens static void 2047fa9e4066Sahrens dbuf_do_born(void *arg) 2048fa9e4066Sahrens { 2049fa9e4066Sahrens struct dbuf_arg *da = arg; 2050fa9e4066Sahrens dsl_dataset_block_born(da->os->os_dsl_dataset, 2051fa9e4066Sahrens &da->bp, da->os->os_synctx); 2052fa9e4066Sahrens kmem_free(da, sizeof (struct dbuf_arg)); 2053fa9e4066Sahrens } 2054fa9e4066Sahrens 2055fa9e4066Sahrens static void 2056fa9e4066Sahrens dbuf_do_kill(void *arg) 2057fa9e4066Sahrens { 2058fa9e4066Sahrens struct dbuf_arg *da = arg; 2059fa9e4066Sahrens dsl_dataset_block_kill(da->os->os_dsl_dataset, 2060fa9e4066Sahrens &da->bp, da->os->os_synctx); 2061fa9e4066Sahrens kmem_free(da, sizeof (struct dbuf_arg)); 2062fa9e4066Sahrens } 2063fa9e4066Sahrens 2064fa9e4066Sahrens /* ARGSUSED */ 2065fa9e4066Sahrens static void 2066fa9e4066Sahrens dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2067fa9e4066Sahrens { 2068fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 2069fa9e4066Sahrens dnode_t *dn = db->db_dnode; 2070fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 2071fa9e4066Sahrens uint64_t txg = zio->io_txg; 2072fa9e4066Sahrens uint64_t fill = 0; 2073fa9e4066Sahrens int i; 2074fa9e4066Sahrens int old_size, new_size; 2075fa9e4066Sahrens 2076fa9e4066Sahrens ASSERT3U(zio->io_error, ==, 0); 2077fa9e4066Sahrens 2078fa9e4066Sahrens dprintf_dbuf_bp(db, &zio->io_bp_orig, "bp_orig: %s", ""); 2079fa9e4066Sahrens 2080fa9e4066Sahrens old_size = BP_GET_ASIZE(&zio->io_bp_orig); 2081fa9e4066Sahrens new_size = BP_GET_ASIZE(zio->io_bp); 2082fa9e4066Sahrens 2083fa9e4066Sahrens dnode_diduse_space(dn, new_size-old_size); 2084fa9e4066Sahrens 2085fa9e4066Sahrens mutex_enter(&db->db_mtx); 2086fa9e4066Sahrens 2087fa9e4066Sahrens if (db->db_dirtied == txg) 2088fa9e4066Sahrens db->db_dirtied = 0; 2089fa9e4066Sahrens 2090fa9e4066Sahrens if (db->db_level == 0) { 2091ea8dc4b6Seschrock arc_buf_t **old = 2092ea8dc4b6Seschrock (arc_buf_t **)&db->db_d.db_data_old[txg&TXG_MASK]; 2093fa9e4066Sahrens 2094fa9e4066Sahrens ASSERT(db->db_blkid != DB_BONUS_BLKID); 2095fa9e4066Sahrens 2096fa9e4066Sahrens if (*old != db->db_buf) 2097ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(*old, db) == 1); 2098ea8dc4b6Seschrock else if (!BP_IS_HOLE(db->db_blkptr)) 2099ea8dc4b6Seschrock arc_set_callback(db->db_buf, dbuf_do_evict, db); 2100ea8dc4b6Seschrock else 2101ea8dc4b6Seschrock ASSERT(arc_released(db->db_buf)); 2102fa9e4066Sahrens *old = NULL; 2103fa9e4066Sahrens db->db_data_pending = NULL; 2104fa9e4066Sahrens 2105fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 2106fa9e4066Sahrens if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2107fa9e4066Sahrens !BP_IS_HOLE(db->db_blkptr)) 2108fa9e4066Sahrens dn->dn_phys->dn_maxblkid = db->db_blkid; 2109fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 2110fa9e4066Sahrens 2111fa9e4066Sahrens if (dn->dn_type == DMU_OT_DNODE) { 2112fa9e4066Sahrens dnode_phys_t *dnp = db->db.db_data; 2113fa9e4066Sahrens for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2114fa9e4066Sahrens i--, dnp++) { 2115fa9e4066Sahrens if (dnp->dn_type != DMU_OT_NONE) 2116fa9e4066Sahrens fill++; 2117fa9e4066Sahrens } 2118fa9e4066Sahrens } else { 2119fa9e4066Sahrens if (!BP_IS_HOLE(db->db_blkptr)) 2120fa9e4066Sahrens fill = 1; 2121fa9e4066Sahrens } 2122fa9e4066Sahrens } else { 2123fa9e4066Sahrens blkptr_t *bp = db->db.db_data; 2124fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2125fa9e4066Sahrens if (!BP_IS_HOLE(db->db_blkptr)) { 212623b11526Smaybee int epbs = 212723b11526Smaybee dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2128fa9e4066Sahrens ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, db->db.db_size); 2129fa9e4066Sahrens ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2130fa9e4066Sahrens db->db.db_size); 213123b11526Smaybee ASSERT3U(dn->dn_phys->dn_maxblkid 213223b11526Smaybee >> (db->db_level * epbs), >=, db->db_blkid); 2133ea8dc4b6Seschrock arc_set_callback(db->db_buf, dbuf_do_evict, db); 2134fa9e4066Sahrens } 2135fa9e4066Sahrens for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, bp++) { 2136fa9e4066Sahrens if (BP_IS_HOLE(bp)) 2137fa9e4066Sahrens continue; 2138fa9e4066Sahrens ASSERT3U(BP_GET_LSIZE(bp), ==, 2139fa9e4066Sahrens db->db_level == 1 ? dn->dn_datablksz : 2140fa9e4066Sahrens (1<<dn->dn_phys->dn_indblkshift)); 2141fa9e4066Sahrens fill += bp->blk_fill; 2142fa9e4066Sahrens } 2143fa9e4066Sahrens } 2144fa9e4066Sahrens 2145fa9e4066Sahrens if (!BP_IS_HOLE(db->db_blkptr)) { 2146fa9e4066Sahrens db->db_blkptr->blk_fill = fill; 2147fa9e4066Sahrens BP_SET_TYPE(db->db_blkptr, dn->dn_type); 2148fa9e4066Sahrens BP_SET_LEVEL(db->db_blkptr, db->db_level); 2149fa9e4066Sahrens } else { 2150fa9e4066Sahrens ASSERT3U(fill, ==, 0); 2151fa9e4066Sahrens ASSERT3U(db->db_blkptr->blk_fill, ==, 0); 2152fa9e4066Sahrens } 2153fa9e4066Sahrens 2154fa9e4066Sahrens dprintf_dbuf_bp(db, db->db_blkptr, 2155fa9e4066Sahrens "wrote %llu bytes to blkptr:", zio->io_size); 2156fa9e4066Sahrens 2157fa9e4066Sahrens ASSERT(db->db_parent == NULL || 2158fa9e4066Sahrens list_link_active(&db->db_parent->db_dirty_node[txg&TXG_MASK])); 2159fa9e4066Sahrens cv_broadcast(&db->db_changed); 2160fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 2161fa9e4066Sahrens db->db_dirtycnt -= 1; 2162fa9e4066Sahrens mutex_exit(&db->db_mtx); 2163fa9e4066Sahrens 2164fa9e4066Sahrens /* We must do this after we've set the bp's type and level */ 2165fa9e4066Sahrens if (!DVA_EQUAL(BP_IDENTITY(zio->io_bp), 2166fa9e4066Sahrens BP_IDENTITY(&zio->io_bp_orig))) { 2167fa9e4066Sahrens struct dbuf_arg *da; 2168fa9e4066Sahrens da = kmem_alloc(sizeof (struct dbuf_arg), KM_SLEEP); 2169fa9e4066Sahrens da->os = os; 2170fa9e4066Sahrens da->bp = *zio->io_bp; 2171fa9e4066Sahrens (void) taskq_dispatch(dbuf_tq, dbuf_do_born, da, 0); 2172fa9e4066Sahrens if (!BP_IS_HOLE(&zio->io_bp_orig)) { 2173fa9e4066Sahrens da = kmem_alloc(sizeof (struct dbuf_arg), KM_SLEEP); 2174fa9e4066Sahrens da->os = os; 2175fa9e4066Sahrens da->bp = zio->io_bp_orig; 2176fa9e4066Sahrens (void) taskq_dispatch(dbuf_tq, dbuf_do_kill, da, 0); 2177fa9e4066Sahrens } 2178fa9e4066Sahrens } 2179fa9e4066Sahrens 2180ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 2181fa9e4066Sahrens } 2182