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 /* 221ab7f2deSmaybee * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 23fa9e4066Sahrens * Use is subject to license terms. 24fa9e4066Sahrens */ 25fa9e4066Sahrens 26fa9e4066Sahrens #include <sys/zfs_context.h> 27fa9e4066Sahrens #include <sys/dmu.h> 28fa9e4066Sahrens #include <sys/dmu_impl.h> 29fa9e4066Sahrens #include <sys/dbuf.h> 30fa9e4066Sahrens #include <sys/dmu_objset.h> 31fa9e4066Sahrens #include <sys/dsl_dataset.h> 32fa9e4066Sahrens #include <sys/dsl_dir.h> 33fa9e4066Sahrens #include <sys/dmu_tx.h> 34fa9e4066Sahrens #include <sys/spa.h> 35fa9e4066Sahrens #include <sys/zio.h> 36fa9e4066Sahrens #include <sys/dmu_zfetch.h> 37fa9e4066Sahrens 38fa9e4066Sahrens static void dbuf_destroy(dmu_buf_impl_t *db); 39fa9e4066Sahrens static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 40088f3894Sahrens static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 41c717a561Smaybee static arc_done_func_t dbuf_write_ready; 42fa9e4066Sahrens static arc_done_func_t dbuf_write_done; 43fa9e4066Sahrens 44fa9e4066Sahrens /* 45fa9e4066Sahrens * Global data structures and functions for the dbuf cache. 46fa9e4066Sahrens */ 47fa9e4066Sahrens static kmem_cache_t *dbuf_cache; 48fa9e4066Sahrens 49fa9e4066Sahrens /* ARGSUSED */ 50fa9e4066Sahrens static int 51fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag) 52fa9e4066Sahrens { 53fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 54fa9e4066Sahrens bzero(db, sizeof (dmu_buf_impl_t)); 55fa9e4066Sahrens 56fa9e4066Sahrens mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 57fa9e4066Sahrens cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 58fa9e4066Sahrens refcount_create(&db->db_holds); 59fa9e4066Sahrens return (0); 60fa9e4066Sahrens } 61fa9e4066Sahrens 62fa9e4066Sahrens /* ARGSUSED */ 63fa9e4066Sahrens static void 64fa9e4066Sahrens dbuf_dest(void *vdb, void *unused) 65fa9e4066Sahrens { 66fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 67fa9e4066Sahrens mutex_destroy(&db->db_mtx); 68fa9e4066Sahrens cv_destroy(&db->db_changed); 69fa9e4066Sahrens refcount_destroy(&db->db_holds); 70fa9e4066Sahrens } 71fa9e4066Sahrens 72fa9e4066Sahrens /* 73fa9e4066Sahrens * dbuf hash table routines 74fa9e4066Sahrens */ 75fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table; 76fa9e4066Sahrens 77fa9e4066Sahrens static uint64_t dbuf_hash_count; 78fa9e4066Sahrens 79fa9e4066Sahrens static uint64_t 80fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 81fa9e4066Sahrens { 82fa9e4066Sahrens uintptr_t osv = (uintptr_t)os; 83fa9e4066Sahrens uint64_t crc = -1ULL; 84fa9e4066Sahrens 85fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 86fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 87fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 88fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 89fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 90fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 91fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 92fa9e4066Sahrens 93fa9e4066Sahrens crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 94fa9e4066Sahrens 95fa9e4066Sahrens return (crc); 96fa9e4066Sahrens } 97fa9e4066Sahrens 98fa9e4066Sahrens #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 99fa9e4066Sahrens 100fa9e4066Sahrens #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 101fa9e4066Sahrens ((dbuf)->db.db_object == (obj) && \ 102fa9e4066Sahrens (dbuf)->db_objset == (os) && \ 103fa9e4066Sahrens (dbuf)->db_level == (level) && \ 104fa9e4066Sahrens (dbuf)->db_blkid == (blkid)) 105fa9e4066Sahrens 106fa9e4066Sahrens dmu_buf_impl_t * 107fa9e4066Sahrens dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 108fa9e4066Sahrens { 109fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 110fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 111fa9e4066Sahrens uint64_t obj = dn->dn_object; 112fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 113fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 114fa9e4066Sahrens dmu_buf_impl_t *db; 115fa9e4066Sahrens 116fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 117fa9e4066Sahrens for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 118fa9e4066Sahrens if (DBUF_EQUAL(db, os, obj, level, blkid)) { 119fa9e4066Sahrens mutex_enter(&db->db_mtx); 120ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 121fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 122fa9e4066Sahrens return (db); 123fa9e4066Sahrens } 124fa9e4066Sahrens mutex_exit(&db->db_mtx); 125fa9e4066Sahrens } 126fa9e4066Sahrens } 127fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 128fa9e4066Sahrens return (NULL); 129fa9e4066Sahrens } 130fa9e4066Sahrens 131fa9e4066Sahrens /* 132fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 133fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 134fa9e4066Sahrens * will be returned and the new element will not be inserted. 135fa9e4066Sahrens * Otherwise returns NULL. 136fa9e4066Sahrens */ 137fa9e4066Sahrens static dmu_buf_impl_t * 138fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db) 139fa9e4066Sahrens { 140fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 141fa9e4066Sahrens objset_impl_t *os = db->db_objset; 142fa9e4066Sahrens uint64_t obj = db->db.db_object; 143fa9e4066Sahrens int level = db->db_level; 144fa9e4066Sahrens uint64_t blkid = db->db_blkid; 145fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 146fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 147fa9e4066Sahrens dmu_buf_impl_t *dbf; 148fa9e4066Sahrens 149fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 150fa9e4066Sahrens for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 151fa9e4066Sahrens if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 152fa9e4066Sahrens mutex_enter(&dbf->db_mtx); 153ea8dc4b6Seschrock if (dbf->db_state != DB_EVICTING) { 154fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 155fa9e4066Sahrens return (dbf); 156fa9e4066Sahrens } 157fa9e4066Sahrens mutex_exit(&dbf->db_mtx); 158fa9e4066Sahrens } 159fa9e4066Sahrens } 160fa9e4066Sahrens 161fa9e4066Sahrens mutex_enter(&db->db_mtx); 162fa9e4066Sahrens db->db_hash_next = h->hash_table[idx]; 163fa9e4066Sahrens h->hash_table[idx] = db; 164fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 165fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, 1); 166fa9e4066Sahrens 167fa9e4066Sahrens return (NULL); 168fa9e4066Sahrens } 169fa9e4066Sahrens 170fa9e4066Sahrens /* 171fa9e4066Sahrens * Remove an entry from the hash table. This operation will 172fa9e4066Sahrens * fail if there are any existing holds on the db. 173fa9e4066Sahrens */ 174fa9e4066Sahrens static void 175fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db) 176fa9e4066Sahrens { 177fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 178fa9e4066Sahrens uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 179fa9e4066Sahrens db->db_level, db->db_blkid); 180fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 181fa9e4066Sahrens dmu_buf_impl_t *dbf, **dbp; 182fa9e4066Sahrens 183fa9e4066Sahrens /* 184fa9e4066Sahrens * We musn't hold db_mtx to maintin lock ordering: 185fa9e4066Sahrens * DBUF_HASH_MUTEX > db_mtx. 186fa9e4066Sahrens */ 187fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 188ea8dc4b6Seschrock ASSERT(db->db_state == DB_EVICTING); 189fa9e4066Sahrens ASSERT(!MUTEX_HELD(&db->db_mtx)); 190fa9e4066Sahrens 191fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 192fa9e4066Sahrens dbp = &h->hash_table[idx]; 193fa9e4066Sahrens while ((dbf = *dbp) != db) { 194fa9e4066Sahrens dbp = &dbf->db_hash_next; 195fa9e4066Sahrens ASSERT(dbf != NULL); 196fa9e4066Sahrens } 197fa9e4066Sahrens *dbp = db->db_hash_next; 198fa9e4066Sahrens db->db_hash_next = NULL; 199fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 200fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, -1); 201fa9e4066Sahrens } 202fa9e4066Sahrens 203ea8dc4b6Seschrock static arc_evict_func_t dbuf_do_evict; 204fa9e4066Sahrens 205fa9e4066Sahrens static void 206fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db) 207fa9e4066Sahrens { 208fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 209fa9e4066Sahrens 210c717a561Smaybee if (db->db_level != 0 || db->db_evict_func == NULL) 211fa9e4066Sahrens return; 212fa9e4066Sahrens 213c717a561Smaybee if (db->db_user_data_ptr_ptr) 214c717a561Smaybee *db->db_user_data_ptr_ptr = db->db.db_data; 215c717a561Smaybee db->db_evict_func(&db->db, db->db_user_ptr); 216c717a561Smaybee db->db_user_ptr = NULL; 217c717a561Smaybee db->db_user_data_ptr_ptr = NULL; 218c717a561Smaybee db->db_evict_func = NULL; 219fa9e4066Sahrens } 220fa9e4066Sahrens 221ea8dc4b6Seschrock void 222ea8dc4b6Seschrock dbuf_evict(dmu_buf_impl_t *db) 223ea8dc4b6Seschrock { 224ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 225ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 226c717a561Smaybee ASSERT(db->db_data_pending == NULL); 227ea8dc4b6Seschrock 228ea8dc4b6Seschrock dbuf_clear(db); 229ea8dc4b6Seschrock dbuf_destroy(db); 230ea8dc4b6Seschrock } 231ea8dc4b6Seschrock 232fa9e4066Sahrens void 233fa9e4066Sahrens dbuf_init(void) 234fa9e4066Sahrens { 235ea8dc4b6Seschrock uint64_t hsize = 1ULL << 16; 236fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 237fa9e4066Sahrens int i; 238fa9e4066Sahrens 239fa9e4066Sahrens /* 240fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 241ea8dc4b6Seschrock * with an average 4K block size. The table will take up 242ea8dc4b6Seschrock * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 243fa9e4066Sahrens */ 244ea8dc4b6Seschrock while (hsize * 4096 < physmem * PAGESIZE) 245fa9e4066Sahrens hsize <<= 1; 246fa9e4066Sahrens 247ea8dc4b6Seschrock retry: 248fa9e4066Sahrens h->hash_table_mask = hsize - 1; 249ea8dc4b6Seschrock h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 250ea8dc4b6Seschrock if (h->hash_table == NULL) { 251ea8dc4b6Seschrock /* XXX - we should really return an error instead of assert */ 252ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 10)); 253ea8dc4b6Seschrock hsize >>= 1; 254ea8dc4b6Seschrock goto retry; 255ea8dc4b6Seschrock } 256fa9e4066Sahrens 257fa9e4066Sahrens dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 258fa9e4066Sahrens sizeof (dmu_buf_impl_t), 259fa9e4066Sahrens 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 260fa9e4066Sahrens 261fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 262fa9e4066Sahrens mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 263fa9e4066Sahrens } 264fa9e4066Sahrens 265fa9e4066Sahrens void 266fa9e4066Sahrens dbuf_fini(void) 267fa9e4066Sahrens { 268fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 269fa9e4066Sahrens int i; 270fa9e4066Sahrens 271fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 272fa9e4066Sahrens mutex_destroy(&h->hash_mutexes[i]); 273fa9e4066Sahrens kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 274fa9e4066Sahrens kmem_cache_destroy(dbuf_cache); 275fa9e4066Sahrens } 276fa9e4066Sahrens 277fa9e4066Sahrens /* 278fa9e4066Sahrens * Other stuff. 279fa9e4066Sahrens */ 280fa9e4066Sahrens 2819c9dc39aSek #ifdef ZFS_DEBUG 282fa9e4066Sahrens static void 283fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db) 284fa9e4066Sahrens { 285fa9e4066Sahrens dnode_t *dn = db->db_dnode; 286fa9e4066Sahrens 287fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 288fa9e4066Sahrens 289fa9e4066Sahrens if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 290fa9e4066Sahrens return; 291fa9e4066Sahrens 292fa9e4066Sahrens ASSERT(db->db_objset != NULL); 293fa9e4066Sahrens if (dn == NULL) { 294fa9e4066Sahrens ASSERT(db->db_parent == NULL); 295fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 296fa9e4066Sahrens } else { 297fa9e4066Sahrens ASSERT3U(db->db.db_object, ==, dn->dn_object); 298fa9e4066Sahrens ASSERT3P(db->db_objset, ==, dn->dn_objset); 299fa9e4066Sahrens ASSERT3U(db->db_level, <, dn->dn_nlevels); 300ea8dc4b6Seschrock ASSERT(db->db_blkid == DB_BONUS_BLKID || 301ea8dc4b6Seschrock list_head(&dn->dn_dbufs)); 302fa9e4066Sahrens } 303fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 304fa9e4066Sahrens ASSERT(dn != NULL); 3051934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 306fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID); 307fa9e4066Sahrens } else { 308fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 309fa9e4066Sahrens } 310fa9e4066Sahrens 311*88b7b0f2SMatthew Ahrens /* 312*88b7b0f2SMatthew Ahrens * We can't assert that db_size matches dn_datablksz because it 313*88b7b0f2SMatthew Ahrens * can be momentarily different when another thread is doing 314*88b7b0f2SMatthew Ahrens * dnode_set_blksz(). 315*88b7b0f2SMatthew Ahrens */ 316*88b7b0f2SMatthew Ahrens if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 317*88b7b0f2SMatthew Ahrens dbuf_dirty_record_t *dr = db->db_data_pending; 318*88b7b0f2SMatthew Ahrens /* 319*88b7b0f2SMatthew Ahrens * It should only be modified in syncing context, so 320*88b7b0f2SMatthew Ahrens * make sure we only have one copy of the data. 321*88b7b0f2SMatthew Ahrens */ 322*88b7b0f2SMatthew Ahrens ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 323fa9e4066Sahrens } 324fa9e4066Sahrens 325fa9e4066Sahrens /* verify db->db_blkptr */ 326fa9e4066Sahrens if (db->db_blkptr) { 327fa9e4066Sahrens if (db->db_parent == dn->dn_dbuf) { 328fa9e4066Sahrens /* db is pointed to by the dnode */ 329fa9e4066Sahrens /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 330ea8dc4b6Seschrock if (db->db.db_object == DMU_META_DNODE_OBJECT) 331fa9e4066Sahrens ASSERT(db->db_parent == NULL); 332fa9e4066Sahrens else 333fa9e4066Sahrens ASSERT(db->db_parent != NULL); 334fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 335fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 336fa9e4066Sahrens } else { 337fa9e4066Sahrens /* db is pointed to by an indirect block */ 338fa9e4066Sahrens int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 339fa9e4066Sahrens ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 340fa9e4066Sahrens ASSERT3U(db->db_parent->db.db_object, ==, 341fa9e4066Sahrens db->db.db_object); 342fa9e4066Sahrens /* 343fa9e4066Sahrens * dnode_grow_indblksz() can make this fail if we don't 344fa9e4066Sahrens * have the struct_rwlock. XXX indblksz no longer 345fa9e4066Sahrens * grows. safe to do this now? 346fa9e4066Sahrens */ 347fa9e4066Sahrens if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) { 348fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 349fa9e4066Sahrens ((blkptr_t *)db->db_parent->db.db_data + 350fa9e4066Sahrens db->db_blkid % epb)); 351fa9e4066Sahrens } 352fa9e4066Sahrens } 353fa9e4066Sahrens } 354fa9e4066Sahrens if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 355fa9e4066Sahrens db->db.db_data && db->db_blkid != DB_BONUS_BLKID && 356fa9e4066Sahrens db->db_state != DB_FILL && !dn->dn_free_txg) { 357fa9e4066Sahrens /* 358fa9e4066Sahrens * If the blkptr isn't set but they have nonzero data, 359fa9e4066Sahrens * it had better be dirty, otherwise we'll lose that 360fa9e4066Sahrens * data when we evict this buffer. 361fa9e4066Sahrens */ 362fa9e4066Sahrens if (db->db_dirtycnt == 0) { 363fa9e4066Sahrens uint64_t *buf = db->db.db_data; 364fa9e4066Sahrens int i; 365fa9e4066Sahrens 366fa9e4066Sahrens for (i = 0; i < db->db.db_size >> 3; i++) { 367fa9e4066Sahrens ASSERT(buf[i] == 0); 368fa9e4066Sahrens } 369fa9e4066Sahrens } 370fa9e4066Sahrens } 371fa9e4066Sahrens } 3729c9dc39aSek #endif 373fa9e4066Sahrens 374fa9e4066Sahrens static void 375fa9e4066Sahrens dbuf_update_data(dmu_buf_impl_t *db) 376fa9e4066Sahrens { 377fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 378c717a561Smaybee if (db->db_level == 0 && db->db_user_data_ptr_ptr) { 379fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 380c717a561Smaybee *db->db_user_data_ptr_ptr = db->db.db_data; 381fa9e4066Sahrens } 382fa9e4066Sahrens } 383fa9e4066Sahrens 384fa9e4066Sahrens static void 385fa9e4066Sahrens dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 386fa9e4066Sahrens { 387fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 388ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 389fa9e4066Sahrens db->db_buf = buf; 390ea8dc4b6Seschrock if (buf != NULL) { 391ea8dc4b6Seschrock ASSERT(buf->b_data != NULL); 392ea8dc4b6Seschrock db->db.db_data = buf->b_data; 393ea8dc4b6Seschrock if (!arc_released(buf)) 394ea8dc4b6Seschrock arc_set_callback(buf, dbuf_do_evict, db); 395ea8dc4b6Seschrock dbuf_update_data(db); 396ea8dc4b6Seschrock } else { 397ea8dc4b6Seschrock dbuf_evict_user(db); 398ea8dc4b6Seschrock db->db.db_data = NULL; 399ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 400ea8dc4b6Seschrock } 401fa9e4066Sahrens } 402fa9e4066Sahrens 403fa9e4066Sahrens uint64_t 404fa9e4066Sahrens dbuf_whichblock(dnode_t *dn, uint64_t offset) 405fa9e4066Sahrens { 406fa9e4066Sahrens if (dn->dn_datablkshift) { 407fa9e4066Sahrens return (offset >> dn->dn_datablkshift); 408fa9e4066Sahrens } else { 409fa9e4066Sahrens ASSERT3U(offset, <, dn->dn_datablksz); 410fa9e4066Sahrens return (0); 411fa9e4066Sahrens } 412fa9e4066Sahrens } 413fa9e4066Sahrens 414fa9e4066Sahrens static void 415fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 416fa9e4066Sahrens { 417fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 418fa9e4066Sahrens 419fa9e4066Sahrens mutex_enter(&db->db_mtx); 420fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_READ); 421fa9e4066Sahrens /* 422fa9e4066Sahrens * All reads are synchronous, so we must have a hold on the dbuf 423fa9e4066Sahrens */ 424fa9e4066Sahrens ASSERT(refcount_count(&db->db_holds) > 0); 425ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 426fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 427c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 428fa9e4066Sahrens /* we were freed in flight; disregard any error */ 429fa9e4066Sahrens arc_release(buf, db); 430fa9e4066Sahrens bzero(buf->b_data, db->db.db_size); 4316b4acc8bSahrens arc_buf_freeze(buf); 432c717a561Smaybee db->db_freed_in_flight = FALSE; 433fa9e4066Sahrens dbuf_set_data(db, buf); 434fa9e4066Sahrens db->db_state = DB_CACHED; 435fa9e4066Sahrens } else if (zio == NULL || zio->io_error == 0) { 436fa9e4066Sahrens dbuf_set_data(db, buf); 437fa9e4066Sahrens db->db_state = DB_CACHED; 438fa9e4066Sahrens } else { 439fa9e4066Sahrens ASSERT(db->db_blkid != DB_BONUS_BLKID); 440fa9e4066Sahrens ASSERT3P(db->db_buf, ==, NULL); 441ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 442ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 443fa9e4066Sahrens } 444fa9e4066Sahrens cv_broadcast(&db->db_changed); 445fa9e4066Sahrens mutex_exit(&db->db_mtx); 446ea8dc4b6Seschrock dbuf_rele(db, NULL); 447fa9e4066Sahrens } 448fa9e4066Sahrens 449ea8dc4b6Seschrock static void 45013506d1eSmaybee dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 451fa9e4066Sahrens { 452088f3894Sahrens dnode_t *dn = db->db_dnode; 453ea8dc4b6Seschrock zbookmark_t zb; 45413506d1eSmaybee uint32_t aflags = ARC_NOWAIT; 455088f3894Sahrens arc_buf_t *pbuf; 456fa9e4066Sahrens 457fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 458fa9e4066Sahrens /* We need the struct_rwlock to prevent db_blkptr from changing. */ 459088f3894Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 460ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 461ea8dc4b6Seschrock ASSERT(db->db_state == DB_UNCACHED); 462ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 463fa9e4066Sahrens 464fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 465088f3894Sahrens int bonuslen = dn->dn_bonuslen; 4661934e92fSmaybee 4671934e92fSmaybee ASSERT3U(bonuslen, <=, db->db.db_size); 468ea8dc4b6Seschrock db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 46925c2057bSmaybee arc_space_consume(DN_MAX_BONUSLEN); 4701934e92fSmaybee if (bonuslen < DN_MAX_BONUSLEN) 471ea8dc4b6Seschrock bzero(db->db.db_data, DN_MAX_BONUSLEN); 472088f3894Sahrens bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, 4731934e92fSmaybee bonuslen); 474ea8dc4b6Seschrock dbuf_update_data(db); 475fa9e4066Sahrens db->db_state = DB_CACHED; 476fa9e4066Sahrens mutex_exit(&db->db_mtx); 477fa9e4066Sahrens return; 478fa9e4066Sahrens } 479fa9e4066Sahrens 4801c8564a7SMark Maybee /* 4811c8564a7SMark Maybee * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 4821c8564a7SMark Maybee * processes the delete record and clears the bp while we are waiting 4831c8564a7SMark Maybee * for the dn_mtx (resulting in a "no" from block_freed). 4841c8564a7SMark Maybee */ 485088f3894Sahrens if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 4861c8564a7SMark Maybee (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 4871c8564a7SMark Maybee BP_IS_HOLE(db->db_blkptr)))) { 488ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 489ad23a2dbSjohansen 490088f3894Sahrens dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa, 491ad23a2dbSjohansen db->db.db_size, db, type)); 492fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 493fa9e4066Sahrens db->db_state = DB_CACHED; 49413506d1eSmaybee *flags |= DB_RF_CACHED; 495fa9e4066Sahrens mutex_exit(&db->db_mtx); 496fa9e4066Sahrens return; 497fa9e4066Sahrens } 498fa9e4066Sahrens 499fa9e4066Sahrens db->db_state = DB_READ; 500fa9e4066Sahrens mutex_exit(&db->db_mtx); 501fa9e4066Sahrens 5023baa08fcSek if (DBUF_IS_L2CACHEABLE(db)) 5033baa08fcSek aflags |= ARC_L2CACHE; 5043baa08fcSek 505ea8dc4b6Seschrock zb.zb_objset = db->db_objset->os_dsl_dataset ? 506ea8dc4b6Seschrock db->db_objset->os_dsl_dataset->ds_object : 0; 507ea8dc4b6Seschrock zb.zb_object = db->db.db_object; 508ea8dc4b6Seschrock zb.zb_level = db->db_level; 509ea8dc4b6Seschrock zb.zb_blkid = db->db_blkid; 510ea8dc4b6Seschrock 511ea8dc4b6Seschrock dbuf_add_ref(db, NULL); 512fa9e4066Sahrens /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ 513088f3894Sahrens 514088f3894Sahrens if (db->db_parent) 515088f3894Sahrens pbuf = db->db_parent->db_buf; 516088f3894Sahrens else 517088f3894Sahrens pbuf = db->db_objset->os_phys_buf; 518088f3894Sahrens 519088f3894Sahrens (void) arc_read(zio, dn->dn_objset->os_spa, db->db_blkptr, pbuf, 520fa9e4066Sahrens dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 52113506d1eSmaybee (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 52213506d1eSmaybee &aflags, &zb); 52313506d1eSmaybee if (aflags & ARC_CACHED) 52413506d1eSmaybee *flags |= DB_RF_CACHED; 525fa9e4066Sahrens } 526fa9e4066Sahrens 527ea8dc4b6Seschrock int 528ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 529fa9e4066Sahrens { 530ea8dc4b6Seschrock int err = 0; 531ea8dc4b6Seschrock int havepzio = (zio != NULL); 53213506d1eSmaybee int prefetch; 533fa9e4066Sahrens 534fa9e4066Sahrens /* 535fa9e4066Sahrens * We don't have to hold the mutex to check db_state because it 536fa9e4066Sahrens * can't be freed while we have a hold on the buffer. 537fa9e4066Sahrens */ 538fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 539fa9e4066Sahrens 540ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 541ea8dc4b6Seschrock rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER); 542ea8dc4b6Seschrock 54313506d1eSmaybee prefetch = db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 5443baa08fcSek (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL && 5453baa08fcSek DBUF_IS_CACHEABLE(db); 54613506d1eSmaybee 547ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 548ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 549ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 55013506d1eSmaybee if (prefetch) 55113506d1eSmaybee dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 55213506d1eSmaybee db->db.db_size, TRUE); 553fa9e4066Sahrens if ((flags & DB_RF_HAVESTRUCT) == 0) 554ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 555ea8dc4b6Seschrock } else if (db->db_state == DB_UNCACHED) { 556ea8dc4b6Seschrock if (zio == NULL) { 557ea8dc4b6Seschrock zio = zio_root(db->db_dnode->dn_objset->os_spa, 558ea8dc4b6Seschrock NULL, NULL, ZIO_FLAG_CANFAIL); 559ea8dc4b6Seschrock } 56013506d1eSmaybee dbuf_read_impl(db, zio, &flags); 56113506d1eSmaybee 562ea8dc4b6Seschrock /* dbuf_read_impl has dropped db_mtx for us */ 563ea8dc4b6Seschrock 56413506d1eSmaybee if (prefetch) 565ea8dc4b6Seschrock dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 56613506d1eSmaybee db->db.db_size, flags & DB_RF_CACHED); 567ea8dc4b6Seschrock 568fa9e4066Sahrens if ((flags & DB_RF_HAVESTRUCT) == 0) 569fa9e4066Sahrens rw_exit(&db->db_dnode->dn_struct_rwlock); 570fa9e4066Sahrens 571ea8dc4b6Seschrock if (!havepzio) 572ea8dc4b6Seschrock err = zio_wait(zio); 573ea8dc4b6Seschrock } else { 57413506d1eSmaybee mutex_exit(&db->db_mtx); 57513506d1eSmaybee if (prefetch) 57613506d1eSmaybee dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 57713506d1eSmaybee db->db.db_size, TRUE); 578ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 579ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 58013506d1eSmaybee 58113506d1eSmaybee mutex_enter(&db->db_mtx); 582ea8dc4b6Seschrock if ((flags & DB_RF_NEVERWAIT) == 0) { 583ea8dc4b6Seschrock while (db->db_state == DB_READ || 584ea8dc4b6Seschrock db->db_state == DB_FILL) { 585ea8dc4b6Seschrock ASSERT(db->db_state == DB_READ || 586ea8dc4b6Seschrock (flags & DB_RF_HAVESTRUCT) == 0); 587ea8dc4b6Seschrock cv_wait(&db->db_changed, &db->db_mtx); 588ea8dc4b6Seschrock } 589ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED) 590ea8dc4b6Seschrock err = EIO; 591ea8dc4b6Seschrock } 592ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 593fa9e4066Sahrens } 594fa9e4066Sahrens 595ea8dc4b6Seschrock ASSERT(err || havepzio || db->db_state == DB_CACHED); 596ea8dc4b6Seschrock return (err); 597fa9e4066Sahrens } 598fa9e4066Sahrens 599fa9e4066Sahrens static void 600fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db) 601fa9e4066Sahrens { 602fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 603ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 604fa9e4066Sahrens mutex_enter(&db->db_mtx); 605fa9e4066Sahrens while (db->db_state == DB_READ || db->db_state == DB_FILL) 606fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 607fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 608ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 609ad23a2dbSjohansen 610ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 611fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 612fa9e4066Sahrens dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 613ad23a2dbSjohansen db->db.db_size, db, type)); 614fa9e4066Sahrens db->db_state = DB_FILL; 615fa9e4066Sahrens } else { 616fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 617fa9e4066Sahrens } 618fa9e4066Sahrens mutex_exit(&db->db_mtx); 619fa9e4066Sahrens } 620fa9e4066Sahrens 621fa9e4066Sahrens /* 622fa9e4066Sahrens * This is our just-in-time copy function. It makes a copy of 623fa9e4066Sahrens * buffers, that have been modified in a previous transaction 624fa9e4066Sahrens * group, before we modify them in the current active group. 625fa9e4066Sahrens * 626fa9e4066Sahrens * This function is used in two places: when we are dirtying a 627fa9e4066Sahrens * buffer for the first time in a txg, and when we are freeing 628fa9e4066Sahrens * a range in a dnode that includes this buffer. 629fa9e4066Sahrens * 630fa9e4066Sahrens * Note that when we are called from dbuf_free_range() we do 631fa9e4066Sahrens * not put a hold on the buffer, we just traverse the active 632fa9e4066Sahrens * dbuf list for the dnode. 633fa9e4066Sahrens */ 634fa9e4066Sahrens static void 635fa9e4066Sahrens dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 636fa9e4066Sahrens { 637c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 638fa9e4066Sahrens 639fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 640fa9e4066Sahrens ASSERT(db->db.db_data != NULL); 641c717a561Smaybee ASSERT(db->db_level == 0); 642c717a561Smaybee ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 643fa9e4066Sahrens 6444d31c452Smaybee if (dr == NULL || 6454d31c452Smaybee (dr->dt.dl.dr_data != 6464d31c452Smaybee ((db->db_blkid == DB_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 647fa9e4066Sahrens return; 648fa9e4066Sahrens 649fa9e4066Sahrens /* 650c717a561Smaybee * If the last dirty record for this dbuf has not yet synced 651c717a561Smaybee * and its referencing the dbuf data, either: 652c717a561Smaybee * reset the reference to point to a new copy, 653c717a561Smaybee * or (if there a no active holders) 654c717a561Smaybee * just null out the current db_data pointer. 655fa9e4066Sahrens */ 656c717a561Smaybee ASSERT(dr->dr_txg >= txg - 2); 657c717a561Smaybee if (db->db_blkid == DB_BONUS_BLKID) { 658c717a561Smaybee /* Note that the data bufs here are zio_bufs */ 659c717a561Smaybee dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 6600e8c6158Smaybee arc_space_consume(DN_MAX_BONUSLEN); 661c717a561Smaybee bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 662c717a561Smaybee } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 663c717a561Smaybee int size = db->db.db_size; 664c717a561Smaybee arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 665c717a561Smaybee dr->dt.dl.dr_data = arc_buf_alloc( 666c717a561Smaybee db->db_dnode->dn_objset->os_spa, size, db, type); 667c717a561Smaybee bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 668c717a561Smaybee } else { 669c717a561Smaybee dbuf_set_data(db, NULL); 670fa9e4066Sahrens } 671fa9e4066Sahrens } 672fa9e4066Sahrens 673c717a561Smaybee void 674c717a561Smaybee dbuf_unoverride(dbuf_dirty_record_t *dr) 675ea8dc4b6Seschrock { 676c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 677c717a561Smaybee uint64_t txg = dr->dr_txg; 678ea8dc4b6Seschrock 679ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 680c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 681c717a561Smaybee ASSERT(db->db_level == 0); 682ea8dc4b6Seschrock 683c717a561Smaybee if (db->db_blkid == DB_BONUS_BLKID || 684c717a561Smaybee dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 685c717a561Smaybee return; 686ea8dc4b6Seschrock 687c717a561Smaybee /* free this block */ 688c717a561Smaybee if (!BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) { 689c717a561Smaybee /* XXX can get silent EIO here */ 690088f3894Sahrens (void) dsl_free(NULL, 691088f3894Sahrens spa_get_dsl(db->db_dnode->dn_objset->os_spa), 692c717a561Smaybee txg, &dr->dt.dl.dr_overridden_by, NULL, NULL, ARC_WAIT); 693fa9e4066Sahrens } 694c717a561Smaybee dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 695c717a561Smaybee /* 696c717a561Smaybee * Release the already-written buffer, so we leave it in 697c717a561Smaybee * a consistent dirty state. Note that all callers are 698c717a561Smaybee * modifying the buffer, so they will immediately do 699c717a561Smaybee * another (redundant) arc_release(). Therefore, leave 700c717a561Smaybee * the buf thawed to save the effort of freezing & 701c717a561Smaybee * immediately re-thawing it. 702c717a561Smaybee */ 703c717a561Smaybee arc_release(dr->dt.dl.dr_data, db); 704fa9e4066Sahrens } 705fa9e4066Sahrens 706cdb0ab79Smaybee /* 707cdb0ab79Smaybee * Evict (if its unreferenced) or clear (if its referenced) any level-0 708cdb0ab79Smaybee * data blocks in the free range, so that any future readers will find 709cdb0ab79Smaybee * empty blocks. Also, if we happen accross any level-1 dbufs in the 710cdb0ab79Smaybee * range that have not already been marked dirty, mark them dirty so 711cdb0ab79Smaybee * they stay in memory. 712cdb0ab79Smaybee */ 713fa9e4066Sahrens void 714cdb0ab79Smaybee dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) 715fa9e4066Sahrens { 716fa9e4066Sahrens dmu_buf_impl_t *db, *db_next; 717fa9e4066Sahrens uint64_t txg = tx->tx_txg; 718cdb0ab79Smaybee int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 719cdb0ab79Smaybee uint64_t first_l1 = start >> epbs; 720cdb0ab79Smaybee uint64_t last_l1 = end >> epbs; 721fa9e4066Sahrens 722cdb0ab79Smaybee if (end > dn->dn_maxblkid) { 723cdb0ab79Smaybee end = dn->dn_maxblkid; 724cdb0ab79Smaybee last_l1 = end >> epbs; 725cdb0ab79Smaybee } 726cdb0ab79Smaybee dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); 727fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 728fa9e4066Sahrens for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 729fa9e4066Sahrens db_next = list_next(&dn->dn_dbufs, db); 730ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 731cdb0ab79Smaybee 732cdb0ab79Smaybee if (db->db_level == 1 && 733cdb0ab79Smaybee db->db_blkid >= first_l1 && db->db_blkid <= last_l1) { 734cdb0ab79Smaybee mutex_enter(&db->db_mtx); 735cdb0ab79Smaybee if (db->db_last_dirty && 736cdb0ab79Smaybee db->db_last_dirty->dr_txg < txg) { 737cdb0ab79Smaybee dbuf_add_ref(db, FTAG); 738cdb0ab79Smaybee mutex_exit(&db->db_mtx); 739cdb0ab79Smaybee dbuf_will_dirty(db, tx); 740cdb0ab79Smaybee dbuf_rele(db, FTAG); 741cdb0ab79Smaybee } else { 742cdb0ab79Smaybee mutex_exit(&db->db_mtx); 743cdb0ab79Smaybee } 744cdb0ab79Smaybee } 745cdb0ab79Smaybee 746ea8dc4b6Seschrock if (db->db_level != 0) 747fa9e4066Sahrens continue; 748fa9e4066Sahrens dprintf_dbuf(db, "found buf %s\n", ""); 749cdb0ab79Smaybee if (db->db_blkid < start || db->db_blkid > end) 750fa9e4066Sahrens continue; 751fa9e4066Sahrens 752fa9e4066Sahrens /* found a level 0 buffer in the range */ 753fa9e4066Sahrens if (dbuf_undirty(db, tx)) 754fa9e4066Sahrens continue; 755fa9e4066Sahrens 756fa9e4066Sahrens mutex_enter(&db->db_mtx); 757ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED || 758ea8dc4b6Seschrock db->db_state == DB_EVICTING) { 759fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 760fa9e4066Sahrens mutex_exit(&db->db_mtx); 761fa9e4066Sahrens continue; 762fa9e4066Sahrens } 763c543ec06Sahrens if (db->db_state == DB_READ || db->db_state == DB_FILL) { 764c543ec06Sahrens /* will be handled in dbuf_read_done or dbuf_rele */ 765c717a561Smaybee db->db_freed_in_flight = TRUE; 766fa9e4066Sahrens mutex_exit(&db->db_mtx); 767fa9e4066Sahrens continue; 768fa9e4066Sahrens } 769ea8dc4b6Seschrock if (refcount_count(&db->db_holds) == 0) { 770ea8dc4b6Seschrock ASSERT(db->db_buf); 771ea8dc4b6Seschrock dbuf_clear(db); 772ea8dc4b6Seschrock continue; 773ea8dc4b6Seschrock } 774c717a561Smaybee /* The dbuf is referenced */ 775fa9e4066Sahrens 776c717a561Smaybee if (db->db_last_dirty != NULL) { 777c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 778c717a561Smaybee 779c717a561Smaybee if (dr->dr_txg == txg) { 78044eda4d7Smaybee /* 781c717a561Smaybee * This buffer is "in-use", re-adjust the file 782c717a561Smaybee * size to reflect that this buffer may 783c717a561Smaybee * contain new data when we sync. 78444eda4d7Smaybee */ 785c717a561Smaybee if (db->db_blkid > dn->dn_maxblkid) 786c717a561Smaybee dn->dn_maxblkid = db->db_blkid; 787c717a561Smaybee dbuf_unoverride(dr); 788c717a561Smaybee } else { 789c717a561Smaybee /* 790c717a561Smaybee * This dbuf is not dirty in the open context. 791c717a561Smaybee * Either uncache it (if its not referenced in 792c717a561Smaybee * the open context) or reset its contents to 793c717a561Smaybee * empty. 794c717a561Smaybee */ 795c717a561Smaybee dbuf_fix_old_data(db, txg); 79644eda4d7Smaybee } 797ea8dc4b6Seschrock } 798c717a561Smaybee /* clear the contents if its cached */ 799ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 800ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 801fa9e4066Sahrens arc_release(db->db_buf, db); 802fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 8036b4acc8bSahrens arc_buf_freeze(db->db_buf); 804fa9e4066Sahrens } 805ea8dc4b6Seschrock 806fa9e4066Sahrens mutex_exit(&db->db_mtx); 807fa9e4066Sahrens } 808fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 809fa9e4066Sahrens } 810fa9e4066Sahrens 811fa9e4066Sahrens static int 8121934e92fSmaybee dbuf_block_freeable(dmu_buf_impl_t *db) 813fa9e4066Sahrens { 814fa9e4066Sahrens dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 815fa9e4066Sahrens uint64_t birth_txg = 0; 816fa9e4066Sahrens 817fa9e4066Sahrens /* 818fa9e4066Sahrens * We don't need any locking to protect db_blkptr: 819c717a561Smaybee * If it's syncing, then db_last_dirty will be set 820c717a561Smaybee * so we'll ignore db_blkptr. 821fa9e4066Sahrens */ 822c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 823c717a561Smaybee if (db->db_last_dirty) 824c717a561Smaybee birth_txg = db->db_last_dirty->dr_txg; 825fa9e4066Sahrens else if (db->db_blkptr) 826fa9e4066Sahrens birth_txg = db->db_blkptr->blk_birth; 827fa9e4066Sahrens 8281934e92fSmaybee /* If we don't exist or are in a snapshot, we can't be freed */ 829fa9e4066Sahrens if (birth_txg) 8301934e92fSmaybee return (ds == NULL || 8311934e92fSmaybee dsl_dataset_block_freeable(ds, birth_txg)); 832fa9e4066Sahrens else 8331934e92fSmaybee return (FALSE); 834fa9e4066Sahrens } 835fa9e4066Sahrens 836fa9e4066Sahrens void 837fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 838fa9e4066Sahrens { 839fa9e4066Sahrens arc_buf_t *buf, *obuf; 840fa9e4066Sahrens int osize = db->db.db_size; 841ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 842fa9e4066Sahrens 843ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 844ea8dc4b6Seschrock 845fa9e4066Sahrens /* XXX does *this* func really need the lock? */ 846fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)); 847fa9e4066Sahrens 848fa9e4066Sahrens /* 849fa9e4066Sahrens * This call to dbuf_will_dirty() with the dn_struct_rwlock held 850fa9e4066Sahrens * is OK, because there can be no other references to the db 851fa9e4066Sahrens * when we are changing its size, so no concurrent DB_FILL can 852fa9e4066Sahrens * be happening. 853fa9e4066Sahrens */ 854ea8dc4b6Seschrock /* 855ea8dc4b6Seschrock * XXX we should be doing a dbuf_read, checking the return 856ea8dc4b6Seschrock * value and returning that up to our callers 857ea8dc4b6Seschrock */ 858fa9e4066Sahrens dbuf_will_dirty(db, tx); 859fa9e4066Sahrens 860fa9e4066Sahrens /* create the data buffer for the new block */ 861ad23a2dbSjohansen buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type); 862fa9e4066Sahrens 863fa9e4066Sahrens /* copy old block data to the new block */ 864fa9e4066Sahrens obuf = db->db_buf; 865f65e61c0Sahrens bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 866fa9e4066Sahrens /* zero the remainder */ 867f65e61c0Sahrens if (size > osize) 868f65e61c0Sahrens bzero((uint8_t *)buf->b_data + osize, size - osize); 869fa9e4066Sahrens 870fa9e4066Sahrens mutex_enter(&db->db_mtx); 871fa9e4066Sahrens dbuf_set_data(db, buf); 872ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(obuf, db) == 1); 873fa9e4066Sahrens db->db.db_size = size; 874fa9e4066Sahrens 875c717a561Smaybee if (db->db_level == 0) { 876c717a561Smaybee ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 877c717a561Smaybee db->db_last_dirty->dt.dl.dr_data = buf; 878c717a561Smaybee } 879fa9e4066Sahrens mutex_exit(&db->db_mtx); 880fa9e4066Sahrens 881fa9e4066Sahrens dnode_willuse_space(db->db_dnode, size-osize, tx); 882fa9e4066Sahrens } 883fa9e4066Sahrens 884c717a561Smaybee dbuf_dirty_record_t * 885fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 886fa9e4066Sahrens { 887fa9e4066Sahrens dnode_t *dn = db->db_dnode; 888fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 889c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 890fa9e4066Sahrens int drop_struct_lock = FALSE; 891d3469faaSMark Maybee boolean_t do_free_accounting = B_FALSE; 892fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 893fa9e4066Sahrens 894fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 895fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 8969c9dc39aSek DMU_TX_DIRTY_BUF(tx, db); 897fa9e4066Sahrens 898fa9e4066Sahrens /* 899fa9e4066Sahrens * Shouldn't dirty a regular buffer in syncing context. Private 900fa9e4066Sahrens * objects may be dirtied in syncing context, but only if they 901fa9e4066Sahrens * were already pre-dirtied in open context. 902fa9e4066Sahrens * XXX We may want to prohibit dirtying in syncing context even 903fa9e4066Sahrens * if they did pre-dirty. 904fa9e4066Sahrens */ 905c717a561Smaybee ASSERT(!dmu_tx_is_syncing(tx) || 906c717a561Smaybee BP_IS_HOLE(dn->dn_objset->os_rootbp) || 907c717a561Smaybee dn->dn_object == DMU_META_DNODE_OBJECT || 908c717a561Smaybee dn->dn_objset->os_dsl_dataset == NULL || 909c717a561Smaybee dsl_dir_is_private(dn->dn_objset->os_dsl_dataset->ds_dir)); 910fa9e4066Sahrens 911fa9e4066Sahrens /* 912fa9e4066Sahrens * We make this assert for private objects as well, but after we 913fa9e4066Sahrens * check if we're already dirty. They are allowed to re-dirty 914fa9e4066Sahrens * in syncing context. 915fa9e4066Sahrens */ 916ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 917c717a561Smaybee dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 918fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 919fa9e4066Sahrens 920fa9e4066Sahrens mutex_enter(&db->db_mtx); 921fa9e4066Sahrens /* 922c717a561Smaybee * XXX make this true for indirects too? The problem is that 923c717a561Smaybee * transactions created with dmu_tx_create_assigned() from 924c717a561Smaybee * syncing context don't bother holding ahead. 925fa9e4066Sahrens */ 926c717a561Smaybee ASSERT(db->db_level != 0 || 927c717a561Smaybee db->db_state == DB_CACHED || db->db_state == DB_FILL); 928fa9e4066Sahrens 929fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 930fa9e4066Sahrens /* 931fa9e4066Sahrens * Don't set dirtyctx to SYNC if we're just modifying this as we 932fa9e4066Sahrens * initialize the objset. 933fa9e4066Sahrens */ 934fa9e4066Sahrens if (dn->dn_dirtyctx == DN_UNDIRTIED && 935c717a561Smaybee !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 936fa9e4066Sahrens dn->dn_dirtyctx = 937fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 938fa9e4066Sahrens ASSERT(dn->dn_dirtyctx_firstset == NULL); 939fa9e4066Sahrens dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 940fa9e4066Sahrens } 941fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 942fa9e4066Sahrens 943fa9e4066Sahrens /* 944fa9e4066Sahrens * If this buffer is already dirty, we're done. 945fa9e4066Sahrens */ 946c717a561Smaybee drp = &db->db_last_dirty; 947c717a561Smaybee ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 948c717a561Smaybee db->db.db_object == DMU_META_DNODE_OBJECT); 9497e2186e3Sbonwick while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 9507e2186e3Sbonwick drp = &dr->dr_next; 9517e2186e3Sbonwick if (dr && dr->dr_txg == tx->tx_txg) { 952c717a561Smaybee if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 953c717a561Smaybee /* 954c717a561Smaybee * If this buffer has already been written out, 955c717a561Smaybee * we now need to reset its state. 956c717a561Smaybee */ 9577e2186e3Sbonwick dbuf_unoverride(dr); 958c717a561Smaybee if (db->db.db_object != DMU_META_DNODE_OBJECT) 959c717a561Smaybee arc_buf_thaw(db->db_buf); 960c717a561Smaybee } 961fa9e4066Sahrens mutex_exit(&db->db_mtx); 9627e2186e3Sbonwick return (dr); 963fa9e4066Sahrens } 964fa9e4066Sahrens 965fa9e4066Sahrens /* 966fa9e4066Sahrens * Only valid if not already dirty. 967fa9e4066Sahrens */ 968fa9e4066Sahrens ASSERT(dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 969fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 970fa9e4066Sahrens 971fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, db->db_level); 972fa9e4066Sahrens ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 973fa9e4066Sahrens dn->dn_phys->dn_nlevels > db->db_level || 974fa9e4066Sahrens dn->dn_next_nlevels[txgoff] > db->db_level || 975fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 976fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 977fa9e4066Sahrens 978fa9e4066Sahrens /* 979fa9e4066Sahrens * We should only be dirtying in syncing context if it's the 980fa9e4066Sahrens * mos, a spa os, or we're initializing the os. However, we are 981fa9e4066Sahrens * allowed to dirty in syncing context provided we already 982fa9e4066Sahrens * dirtied it in open context. Hence we must make this 983fa9e4066Sahrens * assertion only if we're not already dirty. 984fa9e4066Sahrens */ 985fa9e4066Sahrens ASSERT(!dmu_tx_is_syncing(tx) || 986fa9e4066Sahrens os->os_dsl_dataset == NULL || 987fa9e4066Sahrens !dsl_dir_is_private(os->os_dsl_dataset->ds_dir) || 988c717a561Smaybee !BP_IS_HOLE(os->os_rootbp)); 989fa9e4066Sahrens ASSERT(db->db.db_size != 0); 990fa9e4066Sahrens 991fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 992fa9e4066Sahrens 9931934e92fSmaybee if (db->db_blkid != DB_BONUS_BLKID) { 9941934e92fSmaybee /* 9951934e92fSmaybee * Update the accounting. 996d3469faaSMark Maybee * Note: we delay "free accounting" until after we drop 997d3469faaSMark Maybee * the db_mtx. This keeps us from grabbing other locks 998d3469faaSMark Maybee * (and possibly deadlocking) in bp_get_dasize() while 999d3469faaSMark Maybee * also holding the db_mtx. 10001934e92fSmaybee */ 10011934e92fSmaybee dnode_willuse_space(dn, db->db.db_size, tx); 1002d3469faaSMark Maybee do_free_accounting = dbuf_block_freeable(db); 10031934e92fSmaybee } 10041934e92fSmaybee 1005ea8dc4b6Seschrock /* 1006ea8dc4b6Seschrock * If this buffer is dirty in an old transaction group we need 1007ea8dc4b6Seschrock * to make a copy of it so that the changes we make in this 1008ea8dc4b6Seschrock * transaction group won't leak out when we sync the older txg. 1009ea8dc4b6Seschrock */ 1010c717a561Smaybee dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1011c717a561Smaybee if (db->db_level == 0) { 1012c717a561Smaybee void *data_old = db->db_buf; 1013c717a561Smaybee 1014c717a561Smaybee if (db->db_blkid == DB_BONUS_BLKID) { 1015c717a561Smaybee dbuf_fix_old_data(db, tx->tx_txg); 1016c717a561Smaybee data_old = db->db.db_data; 1017c717a561Smaybee } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1018c717a561Smaybee /* 1019c717a561Smaybee * Release the data buffer from the cache so that we 1020c717a561Smaybee * can modify it without impacting possible other users 1021c717a561Smaybee * of this cached data block. Note that indirect 1022c717a561Smaybee * blocks and private objects are not released until the 1023c717a561Smaybee * syncing state (since they are only modified then). 1024c717a561Smaybee */ 1025fa9e4066Sahrens arc_release(db->db_buf, db); 1026fa9e4066Sahrens dbuf_fix_old_data(db, tx->tx_txg); 1027c717a561Smaybee data_old = db->db_buf; 1028fa9e4066Sahrens } 1029c717a561Smaybee ASSERT(data_old != NULL); 1030c717a561Smaybee dr->dt.dl.dr_data = data_old; 1031c717a561Smaybee } else { 1032c717a561Smaybee mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1033c717a561Smaybee list_create(&dr->dt.di.dr_children, 1034c717a561Smaybee sizeof (dbuf_dirty_record_t), 1035c717a561Smaybee offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1036fa9e4066Sahrens } 1037c717a561Smaybee dr->dr_dbuf = db; 1038c717a561Smaybee dr->dr_txg = tx->tx_txg; 1039c717a561Smaybee dr->dr_next = *drp; 1040c717a561Smaybee *drp = dr; 1041fa9e4066Sahrens 1042fa9e4066Sahrens /* 1043fa9e4066Sahrens * We could have been freed_in_flight between the dbuf_noread 1044fa9e4066Sahrens * and dbuf_dirty. We win, as though the dbuf_noread() had 1045fa9e4066Sahrens * happened after the free. 1046fa9e4066Sahrens */ 1047fa9e4066Sahrens if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 1048c717a561Smaybee mutex_enter(&dn->dn_mtx); 1049fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1050c717a561Smaybee mutex_exit(&dn->dn_mtx); 1051c717a561Smaybee db->db_freed_in_flight = FALSE; 1052fa9e4066Sahrens } 1053fa9e4066Sahrens 1054fa9e4066Sahrens /* 1055fa9e4066Sahrens * This buffer is now part of this txg 1056fa9e4066Sahrens */ 1057fa9e4066Sahrens dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1058fa9e4066Sahrens db->db_dirtycnt += 1; 1059fa9e4066Sahrens ASSERT3U(db->db_dirtycnt, <=, 3); 1060fa9e4066Sahrens 1061fa9e4066Sahrens mutex_exit(&db->db_mtx); 1062fa9e4066Sahrens 1063fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 1064c717a561Smaybee mutex_enter(&dn->dn_mtx); 1065c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1066c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1067c717a561Smaybee mutex_exit(&dn->dn_mtx); 1068fa9e4066Sahrens dnode_setdirty(dn, tx); 1069c717a561Smaybee return (dr); 1070d3469faaSMark Maybee } else if (do_free_accounting) { 1071d3469faaSMark Maybee blkptr_t *bp = db->db_blkptr; 1072d3469faaSMark Maybee int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1073d3469faaSMark Maybee bp_get_dasize(os->os_spa, bp) : db->db.db_size; 1074d3469faaSMark Maybee /* 1075d3469faaSMark Maybee * This is only a guess -- if the dbuf is dirty 1076d3469faaSMark Maybee * in a previous txg, we don't know how much 1077d3469faaSMark Maybee * space it will use on disk yet. We should 1078d3469faaSMark Maybee * really have the struct_rwlock to access 1079d3469faaSMark Maybee * db_blkptr, but since this is just a guess, 1080d3469faaSMark Maybee * it's OK if we get an odd answer. 1081d3469faaSMark Maybee */ 1082d3469faaSMark Maybee dnode_willuse_space(dn, -willfree, tx); 1083fa9e4066Sahrens } 1084fa9e4066Sahrens 1085fa9e4066Sahrens if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1086fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 1087fa9e4066Sahrens drop_struct_lock = TRUE; 1088fa9e4066Sahrens } 1089fa9e4066Sahrens 10908346f03fSJonathan W Adams if (db->db_level == 0) { 10918346f03fSJonathan W Adams dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 10928346f03fSJonathan W Adams ASSERT(dn->dn_maxblkid >= db->db_blkid); 10938346f03fSJonathan W Adams } 10948346f03fSJonathan W Adams 109544eda4d7Smaybee if (db->db_level+1 < dn->dn_nlevels) { 1096c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 1097c717a561Smaybee dbuf_dirty_record_t *di; 1098c717a561Smaybee int parent_held = FALSE; 1099c717a561Smaybee 1100c717a561Smaybee if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1101c717a561Smaybee int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1102c717a561Smaybee 1103c717a561Smaybee parent = dbuf_hold_level(dn, db->db_level+1, 1104c717a561Smaybee db->db_blkid >> epbs, FTAG); 1105c717a561Smaybee parent_held = TRUE; 1106c717a561Smaybee } 1107fa9e4066Sahrens if (drop_struct_lock) 1108fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1109c717a561Smaybee ASSERT3U(db->db_level+1, ==, parent->db_level); 1110c717a561Smaybee di = dbuf_dirty(parent, tx); 1111c717a561Smaybee if (parent_held) 1112c717a561Smaybee dbuf_rele(parent, FTAG); 1113c717a561Smaybee 1114c717a561Smaybee mutex_enter(&db->db_mtx); 1115c717a561Smaybee /* possible race with dbuf_undirty() */ 1116c717a561Smaybee if (db->db_last_dirty == dr || 1117c717a561Smaybee dn->dn_object == DMU_META_DNODE_OBJECT) { 1118c717a561Smaybee mutex_enter(&di->dt.di.dr_mtx); 1119c717a561Smaybee ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1120c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1121c717a561Smaybee list_insert_tail(&di->dt.di.dr_children, dr); 1122c717a561Smaybee mutex_exit(&di->dt.di.dr_mtx); 1123c717a561Smaybee dr->dr_parent = di; 1124c717a561Smaybee } 1125c717a561Smaybee mutex_exit(&db->db_mtx); 1126fa9e4066Sahrens } else { 1127c717a561Smaybee ASSERT(db->db_level+1 == dn->dn_nlevels); 1128c717a561Smaybee ASSERT(db->db_blkid < dn->dn_nblkptr); 1129c717a561Smaybee ASSERT(db->db_parent == NULL || 1130c717a561Smaybee db->db_parent == db->db_dnode->dn_dbuf); 1131c717a561Smaybee mutex_enter(&dn->dn_mtx); 1132c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1133c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1134c717a561Smaybee mutex_exit(&dn->dn_mtx); 1135fa9e4066Sahrens if (drop_struct_lock) 1136fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1137fa9e4066Sahrens } 1138fa9e4066Sahrens 1139fa9e4066Sahrens dnode_setdirty(dn, tx); 1140c717a561Smaybee return (dr); 1141fa9e4066Sahrens } 1142fa9e4066Sahrens 1143fa9e4066Sahrens static int 1144fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1145fa9e4066Sahrens { 1146fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1147c717a561Smaybee uint64_t txg = tx->tx_txg; 114817f17c2dSbonwick dbuf_dirty_record_t *dr, **drp; 1149fa9e4066Sahrens 1150c717a561Smaybee ASSERT(txg != 0); 1151ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1152fa9e4066Sahrens 1153fa9e4066Sahrens mutex_enter(&db->db_mtx); 1154fa9e4066Sahrens 1155fa9e4066Sahrens /* 1156fa9e4066Sahrens * If this buffer is not dirty, we're done. 1157fa9e4066Sahrens */ 115817f17c2dSbonwick for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1159c717a561Smaybee if (dr->dr_txg <= txg) 1160c717a561Smaybee break; 1161c717a561Smaybee if (dr == NULL || dr->dr_txg < txg) { 1162fa9e4066Sahrens mutex_exit(&db->db_mtx); 1163fa9e4066Sahrens return (0); 1164fa9e4066Sahrens } 1165c717a561Smaybee ASSERT(dr->dr_txg == txg); 1166fa9e4066Sahrens 1167fa9e4066Sahrens /* 1168fa9e4066Sahrens * If this buffer is currently held, we cannot undirty 1169fa9e4066Sahrens * it, since one of the current holders may be in the 1170fa9e4066Sahrens * middle of an update. Note that users of dbuf_undirty() 1171fa9e4066Sahrens * should not place a hold on the dbuf before the call. 1172fa9e4066Sahrens */ 1173fa9e4066Sahrens if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1174fa9e4066Sahrens mutex_exit(&db->db_mtx); 117544eda4d7Smaybee /* Make sure we don't toss this buffer at sync phase */ 1176fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1177fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1178fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1179fa9e4066Sahrens return (0); 1180fa9e4066Sahrens } 1181fa9e4066Sahrens 1182fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1183fa9e4066Sahrens 1184fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1185fa9e4066Sahrens 1186fa9e4066Sahrens /* XXX would be nice to fix up dn_towrite_space[] */ 1187fa9e4066Sahrens 118817f17c2dSbonwick *drp = dr->dr_next; 1189c717a561Smaybee 1190c717a561Smaybee if (dr->dr_parent) { 1191c717a561Smaybee mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1192c717a561Smaybee list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1193c717a561Smaybee mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1194c717a561Smaybee } else if (db->db_level+1 == dn->dn_nlevels) { 1195cdb0ab79Smaybee ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1196c717a561Smaybee mutex_enter(&dn->dn_mtx); 1197c717a561Smaybee list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1198c717a561Smaybee mutex_exit(&dn->dn_mtx); 1199c717a561Smaybee } 1200c717a561Smaybee 1201c717a561Smaybee if (db->db_level == 0) { 1202c717a561Smaybee dbuf_unoverride(dr); 1203c717a561Smaybee 1204c717a561Smaybee ASSERT(db->db_buf != NULL); 1205c717a561Smaybee ASSERT(dr->dt.dl.dr_data != NULL); 1206c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 1207c717a561Smaybee VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1); 1208c717a561Smaybee } else { 1209c717a561Smaybee ASSERT(db->db_buf != NULL); 1210c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1211c25056deSgw mutex_destroy(&dr->dt.di.dr_mtx); 1212c25056deSgw list_destroy(&dr->dt.di.dr_children); 1213c717a561Smaybee } 1214c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1215fa9e4066Sahrens 1216fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1217fa9e4066Sahrens db->db_dirtycnt -= 1; 1218fa9e4066Sahrens 1219c717a561Smaybee if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1220ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1221fa9e4066Sahrens 1222ea8dc4b6Seschrock ASSERT(arc_released(buf)); 1223ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1224ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1225fa9e4066Sahrens dbuf_evict(db); 1226fa9e4066Sahrens return (1); 1227fa9e4066Sahrens } 1228fa9e4066Sahrens 1229fa9e4066Sahrens mutex_exit(&db->db_mtx); 1230fa9e4066Sahrens return (0); 1231fa9e4066Sahrens } 1232fa9e4066Sahrens 1233fa9e4066Sahrens #pragma weak dmu_buf_will_dirty = dbuf_will_dirty 1234fa9e4066Sahrens void 1235fa9e4066Sahrens dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1236fa9e4066Sahrens { 12371ab7f2deSmaybee int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1238fa9e4066Sahrens 1239fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1240fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1241fa9e4066Sahrens 1242fa9e4066Sahrens if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) 1243fa9e4066Sahrens rf |= DB_RF_HAVESTRUCT; 1244ea8dc4b6Seschrock (void) dbuf_read(db, NULL, rf); 1245c717a561Smaybee (void) dbuf_dirty(db, tx); 1246fa9e4066Sahrens } 1247fa9e4066Sahrens 1248fa9e4066Sahrens void 1249ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1250fa9e4066Sahrens { 1251ea8dc4b6Seschrock dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1252ea8dc4b6Seschrock 1253ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1254fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1255fa9e4066Sahrens ASSERT(db->db_level == 0); 1256fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1257fa9e4066Sahrens 1258ea8dc4b6Seschrock ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1259fa9e4066Sahrens dmu_tx_private_ok(tx)); 1260fa9e4066Sahrens 1261fa9e4066Sahrens dbuf_noread(db); 1262c717a561Smaybee (void) dbuf_dirty(db, tx); 1263fa9e4066Sahrens } 1264fa9e4066Sahrens 1265fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done 1266fa9e4066Sahrens /* ARGSUSED */ 1267fa9e4066Sahrens void 1268fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1269fa9e4066Sahrens { 1270fa9e4066Sahrens mutex_enter(&db->db_mtx); 12719c9dc39aSek DBUF_VERIFY(db); 1272fa9e4066Sahrens 1273fa9e4066Sahrens if (db->db_state == DB_FILL) { 1274c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 1275ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1276fa9e4066Sahrens /* we were freed while filling */ 1277fa9e4066Sahrens /* XXX dbuf_undirty? */ 1278fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 1279c717a561Smaybee db->db_freed_in_flight = FALSE; 1280fa9e4066Sahrens } 1281fa9e4066Sahrens db->db_state = DB_CACHED; 1282fa9e4066Sahrens cv_broadcast(&db->db_changed); 1283fa9e4066Sahrens } 1284fa9e4066Sahrens mutex_exit(&db->db_mtx); 1285fa9e4066Sahrens } 1286fa9e4066Sahrens 1287ea8dc4b6Seschrock /* 1288ea8dc4b6Seschrock * "Clear" the contents of this dbuf. This will mark the dbuf 1289ea8dc4b6Seschrock * EVICTING and clear *most* of its references. Unfortunetely, 1290ea8dc4b6Seschrock * when we are not holding the dn_dbufs_mtx, we can't clear the 1291ea8dc4b6Seschrock * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1292ea8dc4b6Seschrock * in this case. For callers from the DMU we will usually see: 1293ea8dc4b6Seschrock * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1294ea8dc4b6Seschrock * For the arc callback, we will usually see: 1295ea8dc4b6Seschrock * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1296ea8dc4b6Seschrock * Sometimes, though, we will get a mix of these two: 1297ea8dc4b6Seschrock * DMU: dbuf_clear()->arc_buf_evict() 1298ea8dc4b6Seschrock * ARC: dbuf_do_evict()->dbuf_destroy() 1299ea8dc4b6Seschrock */ 1300ea8dc4b6Seschrock void 1301fa9e4066Sahrens dbuf_clear(dmu_buf_impl_t *db) 1302fa9e4066Sahrens { 1303fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1304ea8dc4b6Seschrock dmu_buf_impl_t *parent = db->db_parent; 1305c543ec06Sahrens dmu_buf_impl_t *dndb = dn->dn_dbuf; 1306ea8dc4b6Seschrock int dbuf_gone = FALSE; 1307fa9e4066Sahrens 1308fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 1309fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1310fa9e4066Sahrens 1311ea8dc4b6Seschrock dbuf_evict_user(db); 1312ea8dc4b6Seschrock 1313fa9e4066Sahrens if (db->db_state == DB_CACHED) { 1314ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 13150e8c6158Smaybee if (db->db_blkid == DB_BONUS_BLKID) { 1316ea8dc4b6Seschrock zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 13170e8c6158Smaybee arc_space_return(DN_MAX_BONUSLEN); 13180e8c6158Smaybee } 1319fa9e4066Sahrens db->db.db_data = NULL; 1320fa9e4066Sahrens db->db_state = DB_UNCACHED; 1321fa9e4066Sahrens } 1322fa9e4066Sahrens 1323fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_UNCACHED); 1324fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1325fa9e4066Sahrens 1326ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1327ea8dc4b6Seschrock db->db_blkptr = NULL; 1328ea8dc4b6Seschrock 1329ea8dc4b6Seschrock if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1330ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1331ea8dc4b6Seschrock dnode_rele(dn, db); 13321934e92fSmaybee db->db_dnode = NULL; 1333ea8dc4b6Seschrock } 1334ea8dc4b6Seschrock 1335ea8dc4b6Seschrock if (db->db_buf) 1336ea8dc4b6Seschrock dbuf_gone = arc_buf_evict(db->db_buf); 1337ea8dc4b6Seschrock 1338ea8dc4b6Seschrock if (!dbuf_gone) 1339ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1340fa9e4066Sahrens 1341fa9e4066Sahrens /* 1342fa9e4066Sahrens * If this dbuf is referened from an indirect dbuf, 1343fa9e4066Sahrens * decrement the ref count on the indirect dbuf. 1344fa9e4066Sahrens */ 1345c543ec06Sahrens if (parent && parent != dndb) 1346ea8dc4b6Seschrock dbuf_rele(parent, db); 1347fa9e4066Sahrens } 1348fa9e4066Sahrens 1349fa9e4066Sahrens static int 1350fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1351fa9e4066Sahrens dmu_buf_impl_t **parentp, blkptr_t **bpp) 1352fa9e4066Sahrens { 1353fa9e4066Sahrens int nlevels, epbs; 1354fa9e4066Sahrens 13550b69c2f0Sahrens *parentp = NULL; 13560b69c2f0Sahrens *bpp = NULL; 13570b69c2f0Sahrens 1358ea8dc4b6Seschrock ASSERT(blkid != DB_BONUS_BLKID); 1359ea8dc4b6Seschrock 1360fa9e4066Sahrens if (dn->dn_phys->dn_nlevels == 0) 1361fa9e4066Sahrens nlevels = 1; 1362fa9e4066Sahrens else 1363fa9e4066Sahrens nlevels = dn->dn_phys->dn_nlevels; 1364fa9e4066Sahrens 1365fa9e4066Sahrens epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1366fa9e4066Sahrens 1367fa9e4066Sahrens ASSERT3U(level * epbs, <, 64); 1368fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1369ea8dc4b6Seschrock if (level >= nlevels || 1370fa9e4066Sahrens (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1371fa9e4066Sahrens /* the buffer has no parent yet */ 1372fa9e4066Sahrens return (ENOENT); 1373fa9e4066Sahrens } else if (level < nlevels-1) { 1374fa9e4066Sahrens /* this block is referenced from an indirect block */ 1375fa9e4066Sahrens int err = dbuf_hold_impl(dn, level+1, 1376fa9e4066Sahrens blkid >> epbs, fail_sparse, NULL, parentp); 1377fa9e4066Sahrens if (err) 1378fa9e4066Sahrens return (err); 1379ea8dc4b6Seschrock err = dbuf_read(*parentp, NULL, 1380ea8dc4b6Seschrock (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1381c543ec06Sahrens if (err) { 1382c543ec06Sahrens dbuf_rele(*parentp, NULL); 1383c543ec06Sahrens *parentp = NULL; 1384c543ec06Sahrens return (err); 1385ea8dc4b6Seschrock } 1386c543ec06Sahrens *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1387c543ec06Sahrens (blkid & ((1ULL << epbs) - 1)); 1388c543ec06Sahrens return (0); 1389fa9e4066Sahrens } else { 1390fa9e4066Sahrens /* the block is referenced from the dnode */ 1391fa9e4066Sahrens ASSERT3U(level, ==, nlevels-1); 1392fa9e4066Sahrens ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1393fa9e4066Sahrens blkid < dn->dn_phys->dn_nblkptr); 1394c543ec06Sahrens if (dn->dn_dbuf) { 1395c543ec06Sahrens dbuf_add_ref(dn->dn_dbuf, NULL); 1396c543ec06Sahrens *parentp = dn->dn_dbuf; 1397c543ec06Sahrens } 1398fa9e4066Sahrens *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1399fa9e4066Sahrens return (0); 1400fa9e4066Sahrens } 1401fa9e4066Sahrens } 1402fa9e4066Sahrens 1403fa9e4066Sahrens static dmu_buf_impl_t * 1404fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1405fa9e4066Sahrens dmu_buf_impl_t *parent, blkptr_t *blkptr) 1406fa9e4066Sahrens { 1407fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 1408fa9e4066Sahrens dmu_buf_impl_t *db, *odb; 1409fa9e4066Sahrens 1410fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1411fa9e4066Sahrens ASSERT(dn->dn_type != DMU_OT_NONE); 1412fa9e4066Sahrens 1413fa9e4066Sahrens db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1414fa9e4066Sahrens 1415fa9e4066Sahrens db->db_objset = os; 1416fa9e4066Sahrens db->db.db_object = dn->dn_object; 1417fa9e4066Sahrens db->db_level = level; 1418fa9e4066Sahrens db->db_blkid = blkid; 1419c717a561Smaybee db->db_last_dirty = NULL; 1420ea8dc4b6Seschrock db->db_dirtycnt = 0; 1421ea8dc4b6Seschrock db->db_dnode = dn; 1422ea8dc4b6Seschrock db->db_parent = parent; 1423ea8dc4b6Seschrock db->db_blkptr = blkptr; 1424fa9e4066Sahrens 1425c717a561Smaybee db->db_user_ptr = NULL; 1426c717a561Smaybee db->db_user_data_ptr_ptr = NULL; 1427c717a561Smaybee db->db_evict_func = NULL; 1428c717a561Smaybee db->db_immediate_evict = 0; 1429c717a561Smaybee db->db_freed_in_flight = 0; 1430ea8dc4b6Seschrock 1431ea8dc4b6Seschrock if (blkid == DB_BONUS_BLKID) { 1432ea8dc4b6Seschrock ASSERT3P(parent, ==, dn->dn_dbuf); 14331934e92fSmaybee db->db.db_size = DN_MAX_BONUSLEN - 14341934e92fSmaybee (dn->dn_nblkptr-1) * sizeof (blkptr_t); 14351934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1436fa9e4066Sahrens db->db.db_offset = DB_BONUS_BLKID; 1437ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1438ea8dc4b6Seschrock /* the bonus dbuf is not placed in the hash table */ 14390e8c6158Smaybee arc_space_consume(sizeof (dmu_buf_impl_t)); 1440ea8dc4b6Seschrock return (db); 1441fa9e4066Sahrens } else { 1442fa9e4066Sahrens int blocksize = 1443fa9e4066Sahrens db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; 1444fa9e4066Sahrens db->db.db_size = blocksize; 1445fa9e4066Sahrens db->db.db_offset = db->db_blkid * blocksize; 1446fa9e4066Sahrens } 1447fa9e4066Sahrens 1448fa9e4066Sahrens /* 1449fa9e4066Sahrens * Hold the dn_dbufs_mtx while we get the new dbuf 1450fa9e4066Sahrens * in the hash table *and* added to the dbufs list. 1451fa9e4066Sahrens * This prevents a possible deadlock with someone 1452fa9e4066Sahrens * trying to look up this dbuf before its added to the 1453fa9e4066Sahrens * dn_dbufs list. 1454fa9e4066Sahrens */ 1455fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 1456ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1457fa9e4066Sahrens if ((odb = dbuf_hash_insert(db)) != NULL) { 1458fa9e4066Sahrens /* someone else inserted it first */ 1459fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 1460fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1461fa9e4066Sahrens return (odb); 1462fa9e4066Sahrens } 1463fa9e4066Sahrens list_insert_head(&dn->dn_dbufs, db); 1464ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1465fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 14660e8c6158Smaybee arc_space_consume(sizeof (dmu_buf_impl_t)); 1467fa9e4066Sahrens 1468fa9e4066Sahrens if (parent && parent != dn->dn_dbuf) 1469fa9e4066Sahrens dbuf_add_ref(parent, db); 1470fa9e4066Sahrens 1471ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1472ea8dc4b6Seschrock refcount_count(&dn->dn_holds) > 0); 1473fa9e4066Sahrens (void) refcount_add(&dn->dn_holds, db); 1474fa9e4066Sahrens 1475fa9e4066Sahrens dprintf_dbuf(db, "db=%p\n", db); 1476fa9e4066Sahrens 1477fa9e4066Sahrens return (db); 1478fa9e4066Sahrens } 1479fa9e4066Sahrens 1480fa9e4066Sahrens static int 1481ea8dc4b6Seschrock dbuf_do_evict(void *private) 1482fa9e4066Sahrens { 1483ea8dc4b6Seschrock arc_buf_t *buf = private; 1484ea8dc4b6Seschrock dmu_buf_impl_t *db = buf->b_private; 1485fa9e4066Sahrens 1486ea8dc4b6Seschrock if (!MUTEX_HELD(&db->db_mtx)) 1487ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 1488fa9e4066Sahrens 1489ea8dc4b6Seschrock ASSERT(refcount_is_zero(&db->db_holds)); 1490fa9e4066Sahrens 1491ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 1492ea8dc4b6Seschrock ASSERT(db->db_state == DB_CACHED); 1493ea8dc4b6Seschrock DBUF_VERIFY(db); 1494ea8dc4b6Seschrock db->db_buf = NULL; 1495ea8dc4b6Seschrock dbuf_evict(db); 1496ea8dc4b6Seschrock } else { 1497ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1498ea8dc4b6Seschrock dbuf_destroy(db); 1499fa9e4066Sahrens } 1500ea8dc4b6Seschrock return (0); 1501fa9e4066Sahrens } 1502fa9e4066Sahrens 1503fa9e4066Sahrens static void 1504fa9e4066Sahrens dbuf_destroy(dmu_buf_impl_t *db) 1505fa9e4066Sahrens { 1506fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1507fa9e4066Sahrens 1508ea8dc4b6Seschrock if (db->db_blkid != DB_BONUS_BLKID) { 1509ea8dc4b6Seschrock /* 1510ea8dc4b6Seschrock * If this dbuf is still on the dn_dbufs list, 1511ea8dc4b6Seschrock * remove it from that list. 1512ea8dc4b6Seschrock */ 15131934e92fSmaybee if (db->db_dnode) { 15141934e92fSmaybee dnode_t *dn = db->db_dnode; 15151934e92fSmaybee 15161934e92fSmaybee mutex_enter(&dn->dn_dbufs_mtx); 1517ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1518c543ec06Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1519ea8dc4b6Seschrock 1520ea8dc4b6Seschrock dnode_rele(dn, db); 15211934e92fSmaybee db->db_dnode = NULL; 1522ea8dc4b6Seschrock } 1523ea8dc4b6Seschrock dbuf_hash_remove(db); 1524ea8dc4b6Seschrock } 1525ea8dc4b6Seschrock db->db_parent = NULL; 1526ea8dc4b6Seschrock db->db_buf = NULL; 1527ea8dc4b6Seschrock 1528f82bfe17Sgw ASSERT(!list_link_active(&db->db_link)); 1529fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1530fa9e4066Sahrens ASSERT(db->db_hash_next == NULL); 1531fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 1532fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1533fa9e4066Sahrens 1534fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 15350e8c6158Smaybee arc_space_return(sizeof (dmu_buf_impl_t)); 1536fa9e4066Sahrens } 1537fa9e4066Sahrens 1538fa9e4066Sahrens void 1539fa9e4066Sahrens dbuf_prefetch(dnode_t *dn, uint64_t blkid) 1540fa9e4066Sahrens { 154113506d1eSmaybee dmu_buf_impl_t *db = NULL; 1542fa9e4066Sahrens blkptr_t *bp = NULL; 1543fa9e4066Sahrens 1544fa9e4066Sahrens ASSERT(blkid != DB_BONUS_BLKID); 1545fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1546fa9e4066Sahrens 1547fa9e4066Sahrens if (dnode_block_freed(dn, blkid)) 1548fa9e4066Sahrens return; 1549fa9e4066Sahrens 1550fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1551fa9e4066Sahrens if (db = dbuf_find(dn, 0, blkid)) { 155213506d1eSmaybee if (refcount_count(&db->db_holds) > 0) { 155313506d1eSmaybee /* 155413506d1eSmaybee * This dbuf is active. We assume that it is 155513506d1eSmaybee * already CACHED, or else about to be either 155613506d1eSmaybee * read or filled. 155713506d1eSmaybee */ 155813506d1eSmaybee mutex_exit(&db->db_mtx); 155913506d1eSmaybee return; 156013506d1eSmaybee } 1561fa9e4066Sahrens mutex_exit(&db->db_mtx); 15620b69c2f0Sahrens db = NULL; 1563fa9e4066Sahrens } 1564fa9e4066Sahrens 156513506d1eSmaybee if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1566fa9e4066Sahrens if (bp && !BP_IS_HOLE(bp)) { 1567088f3894Sahrens arc_buf_t *pbuf; 156813506d1eSmaybee uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; 1569ea8dc4b6Seschrock zbookmark_t zb; 1570ea8dc4b6Seschrock zb.zb_objset = dn->dn_objset->os_dsl_dataset ? 1571ea8dc4b6Seschrock dn->dn_objset->os_dsl_dataset->ds_object : 0; 1572ea8dc4b6Seschrock zb.zb_object = dn->dn_object; 1573ea8dc4b6Seschrock zb.zb_level = 0; 1574ea8dc4b6Seschrock zb.zb_blkid = blkid; 1575ea8dc4b6Seschrock 1576088f3894Sahrens if (db) 1577088f3894Sahrens pbuf = db->db_buf; 1578088f3894Sahrens else 1579088f3894Sahrens pbuf = dn->dn_objset->os_phys_buf; 1580088f3894Sahrens 1581088f3894Sahrens (void) arc_read(NULL, dn->dn_objset->os_spa, 1582088f3894Sahrens bp, pbuf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 1583fa9e4066Sahrens ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 158413506d1eSmaybee &aflags, &zb); 1585fa9e4066Sahrens } 158613506d1eSmaybee if (db) 158713506d1eSmaybee dbuf_rele(db, NULL); 1588fa9e4066Sahrens } 1589fa9e4066Sahrens } 1590fa9e4066Sahrens 1591fa9e4066Sahrens /* 1592fa9e4066Sahrens * Returns with db_holds incremented, and db_mtx not held. 1593fa9e4066Sahrens * Note: dn_struct_rwlock must be held. 1594fa9e4066Sahrens */ 1595fa9e4066Sahrens int 1596fa9e4066Sahrens dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1597fa9e4066Sahrens void *tag, dmu_buf_impl_t **dbp) 1598fa9e4066Sahrens { 1599fa9e4066Sahrens dmu_buf_impl_t *db, *parent = NULL; 1600fa9e4066Sahrens 1601ea8dc4b6Seschrock ASSERT(blkid != DB_BONUS_BLKID); 1602fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1603fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, level); 1604fa9e4066Sahrens 1605fa9e4066Sahrens *dbp = NULL; 1606ea8dc4b6Seschrock top: 1607fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1608fa9e4066Sahrens db = dbuf_find(dn, level, blkid); 1609fa9e4066Sahrens 1610fa9e4066Sahrens if (db == NULL) { 1611fa9e4066Sahrens blkptr_t *bp = NULL; 1612fa9e4066Sahrens int err; 1613fa9e4066Sahrens 1614c543ec06Sahrens ASSERT3P(parent, ==, NULL); 1615fa9e4066Sahrens err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1616fa9e4066Sahrens if (fail_sparse) { 1617fa9e4066Sahrens if (err == 0 && bp && BP_IS_HOLE(bp)) 1618fa9e4066Sahrens err = ENOENT; 1619fa9e4066Sahrens if (err) { 1620c543ec06Sahrens if (parent) 1621ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1622fa9e4066Sahrens return (err); 1623fa9e4066Sahrens } 1624fa9e4066Sahrens } 1625ea8dc4b6Seschrock if (err && err != ENOENT) 1626ea8dc4b6Seschrock return (err); 1627fa9e4066Sahrens db = dbuf_create(dn, level, blkid, parent, bp); 1628fa9e4066Sahrens } 1629fa9e4066Sahrens 1630ea8dc4b6Seschrock if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1631ea8dc4b6Seschrock arc_buf_add_ref(db->db_buf, db); 1632ea8dc4b6Seschrock if (db->db_buf->b_data == NULL) { 1633ea8dc4b6Seschrock dbuf_clear(db); 1634c543ec06Sahrens if (parent) { 1635c543ec06Sahrens dbuf_rele(parent, NULL); 1636c543ec06Sahrens parent = NULL; 1637c543ec06Sahrens } 1638ea8dc4b6Seschrock goto top; 1639ea8dc4b6Seschrock } 1640ea8dc4b6Seschrock ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1641ea8dc4b6Seschrock } 1642ea8dc4b6Seschrock 1643ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1644ea8dc4b6Seschrock 1645fa9e4066Sahrens /* 1646c717a561Smaybee * If this buffer is currently syncing out, and we are are 1647c717a561Smaybee * still referencing it from db_data, we need to make a copy 1648c717a561Smaybee * of it in case we decide we want to dirty it again in this txg. 1649fa9e4066Sahrens */ 1650c717a561Smaybee if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 1651ea8dc4b6Seschrock dn->dn_object != DMU_META_DNODE_OBJECT && 1652c717a561Smaybee db->db_state == DB_CACHED && db->db_data_pending) { 1653c717a561Smaybee dbuf_dirty_record_t *dr = db->db_data_pending; 1654fa9e4066Sahrens 1655c717a561Smaybee if (dr->dt.dl.dr_data == db->db_buf) { 1656c717a561Smaybee arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1657c717a561Smaybee 1658c717a561Smaybee dbuf_set_data(db, 1659c717a561Smaybee arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 1660c717a561Smaybee db->db.db_size, db, type)); 1661c717a561Smaybee bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1662c717a561Smaybee db->db.db_size); 1663c717a561Smaybee } 1664fa9e4066Sahrens } 1665fa9e4066Sahrens 1666ea8dc4b6Seschrock (void) refcount_add(&db->db_holds, tag); 1667fa9e4066Sahrens dbuf_update_data(db); 16689c9dc39aSek DBUF_VERIFY(db); 1669fa9e4066Sahrens mutex_exit(&db->db_mtx); 1670fa9e4066Sahrens 1671fa9e4066Sahrens /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1672c543ec06Sahrens if (parent) 1673ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1674fa9e4066Sahrens 1675fa9e4066Sahrens ASSERT3P(db->db_dnode, ==, dn); 1676fa9e4066Sahrens ASSERT3U(db->db_blkid, ==, blkid); 1677fa9e4066Sahrens ASSERT3U(db->db_level, ==, level); 1678fa9e4066Sahrens *dbp = db; 1679fa9e4066Sahrens 1680fa9e4066Sahrens return (0); 1681fa9e4066Sahrens } 1682fa9e4066Sahrens 1683fa9e4066Sahrens dmu_buf_impl_t * 1684ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1685fa9e4066Sahrens { 1686fa9e4066Sahrens dmu_buf_impl_t *db; 1687ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1688ea8dc4b6Seschrock return (err ? NULL : db); 1689fa9e4066Sahrens } 1690fa9e4066Sahrens 1691fa9e4066Sahrens dmu_buf_impl_t * 1692fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1693fa9e4066Sahrens { 1694fa9e4066Sahrens dmu_buf_impl_t *db; 1695ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1696ea8dc4b6Seschrock return (err ? NULL : db); 1697fa9e4066Sahrens } 1698fa9e4066Sahrens 16991934e92fSmaybee void 1700ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn) 1701fa9e4066Sahrens { 1702ea8dc4b6Seschrock ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1703ea8dc4b6Seschrock 1704ea8dc4b6Seschrock ASSERT(dn->dn_bonus == NULL); 17051934e92fSmaybee dn->dn_bonus = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL); 1706fa9e4066Sahrens } 1707fa9e4066Sahrens 1708ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref 1709fa9e4066Sahrens void 1710fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 1711fa9e4066Sahrens { 1712ea8dc4b6Seschrock int64_t holds = refcount_add(&db->db_holds, tag); 1713ea8dc4b6Seschrock ASSERT(holds > 1); 1714fa9e4066Sahrens } 1715fa9e4066Sahrens 1716ea8dc4b6Seschrock #pragma weak dmu_buf_rele = dbuf_rele 1717fa9e4066Sahrens void 1718ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag) 1719fa9e4066Sahrens { 1720fa9e4066Sahrens int64_t holds; 1721fa9e4066Sahrens 1722fa9e4066Sahrens mutex_enter(&db->db_mtx); 17239c9dc39aSek DBUF_VERIFY(db); 1724fa9e4066Sahrens 1725fa9e4066Sahrens holds = refcount_remove(&db->db_holds, tag); 1726ea8dc4b6Seschrock ASSERT(holds >= 0); 1727ea8dc4b6Seschrock 1728c717a561Smaybee /* 1729c717a561Smaybee * We can't freeze indirects if there is a possibility that they 1730c717a561Smaybee * may be modified in the current syncing context. 1731c717a561Smaybee */ 1732c717a561Smaybee if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 17336b4acc8bSahrens arc_buf_freeze(db->db_buf); 17346b4acc8bSahrens 1735ea8dc4b6Seschrock if (holds == db->db_dirtycnt && 1736c717a561Smaybee db->db_level == 0 && db->db_immediate_evict) 1737ea8dc4b6Seschrock dbuf_evict_user(db); 1738fa9e4066Sahrens 1739fa9e4066Sahrens if (holds == 0) { 1740ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 1741ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1742ea8dc4b6Seschrock dnode_rele(db->db_dnode, db); 1743ea8dc4b6Seschrock } else if (db->db_buf == NULL) { 1744ea8dc4b6Seschrock /* 1745ea8dc4b6Seschrock * This is a special case: we never associated this 1746ea8dc4b6Seschrock * dbuf with any data allocated from the ARC. 1747ea8dc4b6Seschrock */ 1748ea8dc4b6Seschrock ASSERT3U(db->db_state, ==, DB_UNCACHED); 1749ea8dc4b6Seschrock dbuf_evict(db); 17506b4acc8bSahrens } else if (arc_released(db->db_buf)) { 1751ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1752ea8dc4b6Seschrock /* 1753ea8dc4b6Seschrock * This dbuf has anonymous data associated with it. 1754ea8dc4b6Seschrock */ 1755ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1756ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1757ea8dc4b6Seschrock dbuf_evict(db); 1758ea8dc4b6Seschrock } else { 1759ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); 17603baa08fcSek if (!DBUF_IS_CACHEABLE(db)) 17613baa08fcSek dbuf_clear(db); 17623baa08fcSek else 17633baa08fcSek mutex_exit(&db->db_mtx); 1764ea8dc4b6Seschrock } 1765fa9e4066Sahrens } else { 1766fa9e4066Sahrens mutex_exit(&db->db_mtx); 1767fa9e4066Sahrens } 1768fa9e4066Sahrens } 1769fa9e4066Sahrens 1770fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount 1771fa9e4066Sahrens uint64_t 1772fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db) 1773fa9e4066Sahrens { 1774fa9e4066Sahrens return (refcount_count(&db->db_holds)); 1775fa9e4066Sahrens } 1776fa9e4066Sahrens 1777fa9e4066Sahrens void * 1778fa9e4066Sahrens dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1779fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 1780fa9e4066Sahrens { 1781fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1782fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 1783fa9e4066Sahrens } 1784fa9e4066Sahrens 1785fa9e4066Sahrens void * 1786fa9e4066Sahrens dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1787fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 1788fa9e4066Sahrens { 1789fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1790fa9e4066Sahrens 1791c717a561Smaybee db->db_immediate_evict = TRUE; 1792fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1793fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 1794fa9e4066Sahrens } 1795fa9e4066Sahrens 1796fa9e4066Sahrens void * 1797fa9e4066Sahrens dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 1798fa9e4066Sahrens void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 1799fa9e4066Sahrens { 1800fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1801fa9e4066Sahrens ASSERT(db->db_level == 0); 1802fa9e4066Sahrens 1803fa9e4066Sahrens ASSERT((user_ptr == NULL) == (evict_func == NULL)); 1804fa9e4066Sahrens 1805fa9e4066Sahrens mutex_enter(&db->db_mtx); 1806fa9e4066Sahrens 1807c717a561Smaybee if (db->db_user_ptr == old_user_ptr) { 1808c717a561Smaybee db->db_user_ptr = user_ptr; 1809c717a561Smaybee db->db_user_data_ptr_ptr = user_data_ptr_ptr; 1810c717a561Smaybee db->db_evict_func = evict_func; 1811fa9e4066Sahrens 1812fa9e4066Sahrens dbuf_update_data(db); 1813fa9e4066Sahrens } else { 1814c717a561Smaybee old_user_ptr = db->db_user_ptr; 1815fa9e4066Sahrens } 1816fa9e4066Sahrens 1817fa9e4066Sahrens mutex_exit(&db->db_mtx); 1818fa9e4066Sahrens return (old_user_ptr); 1819fa9e4066Sahrens } 1820fa9e4066Sahrens 1821fa9e4066Sahrens void * 1822fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake) 1823fa9e4066Sahrens { 1824fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1825fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1826fa9e4066Sahrens 1827c717a561Smaybee return (db->db_user_ptr); 1828fa9e4066Sahrens } 1829fa9e4066Sahrens 1830c717a561Smaybee static void 1831c717a561Smaybee dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 1832fa9e4066Sahrens { 1833c717a561Smaybee /* ASSERT(dmu_tx_is_syncing(tx) */ 1834c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 1835c717a561Smaybee 1836c717a561Smaybee if (db->db_blkptr != NULL) 1837c717a561Smaybee return; 1838c717a561Smaybee 1839c717a561Smaybee if (db->db_level == dn->dn_phys->dn_nlevels-1) { 1840c717a561Smaybee /* 1841c717a561Smaybee * This buffer was allocated at a time when there was 1842c717a561Smaybee * no available blkptrs from the dnode, or it was 1843c717a561Smaybee * inappropriate to hook it in (i.e., nlevels mis-match). 1844c717a561Smaybee */ 1845c717a561Smaybee ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 1846c717a561Smaybee ASSERT(db->db_parent == NULL); 1847c717a561Smaybee db->db_parent = dn->dn_dbuf; 1848c717a561Smaybee db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 1849c717a561Smaybee DBUF_VERIFY(db); 1850c717a561Smaybee } else { 1851c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 1852c717a561Smaybee int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 1853c717a561Smaybee 1854c717a561Smaybee ASSERT(dn->dn_phys->dn_nlevels > 1); 1855c717a561Smaybee if (parent == NULL) { 1856c717a561Smaybee mutex_exit(&db->db_mtx); 1857c717a561Smaybee rw_enter(&dn->dn_struct_rwlock, RW_READER); 1858c717a561Smaybee (void) dbuf_hold_impl(dn, db->db_level+1, 1859c717a561Smaybee db->db_blkid >> epbs, FALSE, db, &parent); 1860c717a561Smaybee rw_exit(&dn->dn_struct_rwlock); 1861c717a561Smaybee mutex_enter(&db->db_mtx); 1862c717a561Smaybee db->db_parent = parent; 1863c717a561Smaybee } 1864c717a561Smaybee db->db_blkptr = (blkptr_t *)parent->db.db_data + 1865c717a561Smaybee (db->db_blkid & ((1ULL << epbs) - 1)); 1866c717a561Smaybee DBUF_VERIFY(db); 1867c717a561Smaybee } 1868c717a561Smaybee } 1869c717a561Smaybee 1870c717a561Smaybee static void 1871c717a561Smaybee dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 1872c717a561Smaybee { 1873c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 1874c717a561Smaybee dnode_t *dn = db->db_dnode; 1875c717a561Smaybee zio_t *zio; 1876c717a561Smaybee 1877c717a561Smaybee ASSERT(dmu_tx_is_syncing(tx)); 1878c717a561Smaybee 1879c717a561Smaybee dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 1880c717a561Smaybee 1881c717a561Smaybee mutex_enter(&db->db_mtx); 1882c717a561Smaybee 1883c717a561Smaybee ASSERT(db->db_level > 0); 1884c717a561Smaybee DBUF_VERIFY(db); 1885c717a561Smaybee 1886c717a561Smaybee if (db->db_buf == NULL) { 1887c717a561Smaybee mutex_exit(&db->db_mtx); 1888c717a561Smaybee (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 1889c717a561Smaybee mutex_enter(&db->db_mtx); 1890c717a561Smaybee } 1891c717a561Smaybee ASSERT3U(db->db_state, ==, DB_CACHED); 1892c717a561Smaybee ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 1893c717a561Smaybee ASSERT(db->db_buf != NULL); 1894c717a561Smaybee 1895c717a561Smaybee dbuf_check_blkptr(dn, db); 1896c717a561Smaybee 1897c717a561Smaybee db->db_data_pending = dr; 1898c717a561Smaybee 1899af2c4821Smaybee mutex_exit(&db->db_mtx); 1900088f3894Sahrens dbuf_write(dr, db->db_buf, tx); 1901c717a561Smaybee 1902c717a561Smaybee zio = dr->dr_zio; 1903c717a561Smaybee mutex_enter(&dr->dt.di.dr_mtx); 1904c717a561Smaybee dbuf_sync_list(&dr->dt.di.dr_children, tx); 1905c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1906c717a561Smaybee mutex_exit(&dr->dt.di.dr_mtx); 1907c717a561Smaybee zio_nowait(zio); 1908c717a561Smaybee } 1909c717a561Smaybee 1910c717a561Smaybee static void 1911c717a561Smaybee dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 1912c717a561Smaybee { 1913c717a561Smaybee arc_buf_t **datap = &dr->dt.dl.dr_data; 1914c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 1915fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1916fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 1917c717a561Smaybee uint64_t txg = tx->tx_txg; 1918fa9e4066Sahrens int blksz; 1919fa9e4066Sahrens 1920fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 1921fa9e4066Sahrens 1922fa9e4066Sahrens dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 1923fa9e4066Sahrens 1924fa9e4066Sahrens mutex_enter(&db->db_mtx); 1925fa9e4066Sahrens /* 1926fa9e4066Sahrens * To be synced, we must be dirtied. But we 1927fa9e4066Sahrens * might have been freed after the dirty. 1928fa9e4066Sahrens */ 1929fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 1930fa9e4066Sahrens /* This buffer has been freed since it was dirtied */ 1931fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1932fa9e4066Sahrens } else if (db->db_state == DB_FILL) { 1933fa9e4066Sahrens /* This buffer was freed and is now being re-filled */ 1934c717a561Smaybee ASSERT(db->db.db_data != dr->dt.dl.dr_data); 1935fa9e4066Sahrens } else { 1936fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 1937fa9e4066Sahrens } 19389c9dc39aSek DBUF_VERIFY(db); 1939fa9e4066Sahrens 1940fa9e4066Sahrens /* 1941c717a561Smaybee * If this is a bonus buffer, simply copy the bonus data into the 1942c717a561Smaybee * dnode. It will be written out when the dnode is synced (and it 1943c717a561Smaybee * will be synced, since it must have been dirty for dbuf_sync to 1944c717a561Smaybee * be called). 1945fa9e4066Sahrens */ 1946ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 1947c717a561Smaybee dbuf_dirty_record_t **drp; 19481934e92fSmaybee 1949ea8dc4b6Seschrock ASSERT(*datap != NULL); 1950ea8dc4b6Seschrock ASSERT3U(db->db_level, ==, 0); 1951ea8dc4b6Seschrock ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 1952ea8dc4b6Seschrock bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 19530e8c6158Smaybee if (*datap != db->db.db_data) { 1954ea8dc4b6Seschrock zio_buf_free(*datap, DN_MAX_BONUSLEN); 19550e8c6158Smaybee arc_space_return(DN_MAX_BONUSLEN); 19560e8c6158Smaybee } 1957ea8dc4b6Seschrock db->db_data_pending = NULL; 1958c717a561Smaybee drp = &db->db_last_dirty; 1959c717a561Smaybee while (*drp != dr) 1960c717a561Smaybee drp = &(*drp)->dr_next; 196117f17c2dSbonwick ASSERT(dr->dr_next == NULL); 196217f17c2dSbonwick *drp = dr->dr_next; 1963c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1964ea8dc4b6Seschrock ASSERT(db->db_dirtycnt > 0); 1965ea8dc4b6Seschrock db->db_dirtycnt -= 1; 1966ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1967ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 1968ea8dc4b6Seschrock return; 1969ea8dc4b6Seschrock } 1970ea8dc4b6Seschrock 1971f82bfe17Sgw /* 1972f82bfe17Sgw * This function may have dropped the db_mtx lock allowing a dmu_sync 1973f82bfe17Sgw * operation to sneak in. As a result, we need to ensure that we 1974f82bfe17Sgw * don't check the dr_override_state until we have returned from 1975f82bfe17Sgw * dbuf_check_blkptr. 1976f82bfe17Sgw */ 1977f82bfe17Sgw dbuf_check_blkptr(dn, db); 1978f82bfe17Sgw 1979c717a561Smaybee /* 1980c717a561Smaybee * If this buffer is in the middle of an immdiate write, 1981c717a561Smaybee * wait for the synchronous IO to complete. 1982c717a561Smaybee */ 1983c717a561Smaybee while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 1984c717a561Smaybee ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 1985c717a561Smaybee cv_wait(&db->db_changed, &db->db_mtx); 1986c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 1987c717a561Smaybee } 1988c5c6ffa0Smaybee 1989c717a561Smaybee /* 1990c717a561Smaybee * If this dbuf has already been written out via an immediate write, 1991c717a561Smaybee * just complete the write by copying over the new block pointer and 1992c717a561Smaybee * updating the accounting via the write-completion functions. 1993c717a561Smaybee */ 1994c717a561Smaybee if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1995c717a561Smaybee zio_t zio_fake; 1996c717a561Smaybee 1997c717a561Smaybee zio_fake.io_private = &db; 1998c717a561Smaybee zio_fake.io_error = 0; 1999c717a561Smaybee zio_fake.io_bp = db->db_blkptr; 2000c717a561Smaybee zio_fake.io_bp_orig = *db->db_blkptr; 2001c717a561Smaybee zio_fake.io_txg = txg; 2002e14bb325SJeff Bonwick zio_fake.io_flags = 0; 2003c717a561Smaybee 2004c717a561Smaybee *db->db_blkptr = dr->dt.dl.dr_overridden_by; 2005c717a561Smaybee dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 2006c717a561Smaybee db->db_data_pending = dr; 2007c717a561Smaybee dr->dr_zio = &zio_fake; 2008c717a561Smaybee mutex_exit(&db->db_mtx); 2009c717a561Smaybee 2010e14bb325SJeff Bonwick ASSERT(!DVA_EQUAL(BP_IDENTITY(zio_fake.io_bp), 2011f4a72450SJeff Bonwick BP_IDENTITY(&zio_fake.io_bp_orig)) || 2012f4a72450SJeff Bonwick BP_IS_HOLE(zio_fake.io_bp)); 2013e14bb325SJeff Bonwick 2014c717a561Smaybee if (BP_IS_OLDER(&zio_fake.io_bp_orig, txg)) 2015cdb0ab79Smaybee (void) dsl_dataset_block_kill(os->os_dsl_dataset, 2016c717a561Smaybee &zio_fake.io_bp_orig, dn->dn_zio, tx); 2017c717a561Smaybee 2018c717a561Smaybee dbuf_write_ready(&zio_fake, db->db_buf, db); 2019c717a561Smaybee dbuf_write_done(&zio_fake, db->db_buf, db); 2020c717a561Smaybee 2021c717a561Smaybee return; 2022c717a561Smaybee } 2023c717a561Smaybee 2024c717a561Smaybee blksz = arc_buf_size(*datap); 2025c717a561Smaybee 2026c717a561Smaybee if (dn->dn_object != DMU_META_DNODE_OBJECT) { 2027fa9e4066Sahrens /* 2028fa9e4066Sahrens * If this buffer is currently "in use" (i.e., there are 2029fa9e4066Sahrens * active holds and db_data still references it), then make 2030fa9e4066Sahrens * a copy before we start the write so that any modifications 2031fa9e4066Sahrens * from the open txg will not leak into this write. 2032fa9e4066Sahrens * 2033fa9e4066Sahrens * NOTE: this copy does not need to be made for objects only 2034c717a561Smaybee * modified in the syncing context (e.g. DNONE_DNODE blocks). 2035fa9e4066Sahrens */ 2036c717a561Smaybee if (refcount_count(&db->db_holds) > 1 && *datap == db->db_buf) { 2037c717a561Smaybee arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2038c717a561Smaybee *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2039c717a561Smaybee bcopy(db->db.db_data, (*datap)->b_data, blksz); 2040fa9e4066Sahrens } 2041fa9e4066Sahrens } 2042fa9e4066Sahrens 2043c717a561Smaybee ASSERT(*datap != NULL); 2044c717a561Smaybee db->db_data_pending = dr; 2045fa9e4066Sahrens 2046c717a561Smaybee mutex_exit(&db->db_mtx); 2047fa9e4066Sahrens 2048088f3894Sahrens dbuf_write(dr, *datap, tx); 2049fa9e4066Sahrens 2050c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 2051c717a561Smaybee if (dn->dn_object == DMU_META_DNODE_OBJECT) 2052c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2053c717a561Smaybee else 2054c717a561Smaybee zio_nowait(dr->dr_zio); 2055c717a561Smaybee } 205623b11526Smaybee 2057c717a561Smaybee void 2058c717a561Smaybee dbuf_sync_list(list_t *list, dmu_tx_t *tx) 2059c717a561Smaybee { 2060c717a561Smaybee dbuf_dirty_record_t *dr; 2061c717a561Smaybee 2062c717a561Smaybee while (dr = list_head(list)) { 2063c717a561Smaybee if (dr->dr_zio != NULL) { 2064c717a561Smaybee /* 2065c717a561Smaybee * If we find an already initialized zio then we 2066c717a561Smaybee * are processing the meta-dnode, and we have finished. 2067c717a561Smaybee * The dbufs for all dnodes are put back on the list 2068c717a561Smaybee * during processing, so that we can zio_wait() 2069c717a561Smaybee * these IOs after initiating all child IOs. 2070c717a561Smaybee */ 2071c717a561Smaybee ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2072c717a561Smaybee DMU_META_DNODE_OBJECT); 2073c717a561Smaybee break; 207423b11526Smaybee } 2075c717a561Smaybee list_remove(list, dr); 2076c717a561Smaybee if (dr->dr_dbuf->db_level > 0) 2077c717a561Smaybee dbuf_sync_indirect(dr, tx); 2078c717a561Smaybee else 2079c717a561Smaybee dbuf_sync_leaf(dr, tx); 208023b11526Smaybee } 2081c717a561Smaybee } 208223b11526Smaybee 2083c717a561Smaybee static void 2084088f3894Sahrens dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 2085c717a561Smaybee { 2086c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 2087c717a561Smaybee dnode_t *dn = db->db_dnode; 2088c717a561Smaybee objset_impl_t *os = dn->dn_objset; 2089c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 2090c717a561Smaybee uint64_t txg = tx->tx_txg; 2091c717a561Smaybee zbookmark_t zb; 2092088f3894Sahrens writeprops_t wp = { 0 }; 2093c717a561Smaybee zio_t *zio; 2094fa9e4066Sahrens 2095088f3894Sahrens if (!BP_IS_HOLE(db->db_blkptr) && 2096088f3894Sahrens (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE)) { 2097088f3894Sahrens /* 2098088f3894Sahrens * Private object buffers are released here rather 2099088f3894Sahrens * than in dbuf_dirty() since they are only modified 2100088f3894Sahrens * in the syncing context and we don't want the 2101088f3894Sahrens * overhead of making multiple copies of the data. 2102088f3894Sahrens */ 2103088f3894Sahrens arc_release(data, db); 2104088f3894Sahrens } else { 2105088f3894Sahrens ASSERT(arc_released(data)); 2106088f3894Sahrens /* XXX why do we need to thaw here? */ 2107088f3894Sahrens arc_buf_thaw(data); 2108088f3894Sahrens } 2109088f3894Sahrens 2110c717a561Smaybee if (parent != dn->dn_dbuf) { 2111c717a561Smaybee ASSERT(parent && parent->db_data_pending); 2112fa9e4066Sahrens ASSERT(db->db_level == parent->db_level-1); 2113c717a561Smaybee ASSERT(arc_released(parent->db_buf)); 2114c717a561Smaybee zio = parent->db_data_pending->dr_zio; 2115ea8dc4b6Seschrock } else { 2116fa9e4066Sahrens ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1); 2117fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 2118fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 2119c717a561Smaybee zio = dn->dn_zio; 2120ea8dc4b6Seschrock } 2121fa9e4066Sahrens 2122c717a561Smaybee ASSERT(db->db_level == 0 || data == db->db_buf); 2123c717a561Smaybee ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2124c717a561Smaybee ASSERT(zio); 2125fa9e4066Sahrens 2126ea8dc4b6Seschrock zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0; 2127ea8dc4b6Seschrock zb.zb_object = db->db.db_object; 2128ea8dc4b6Seschrock zb.zb_level = db->db_level; 2129ea8dc4b6Seschrock zb.zb_blkid = db->db_blkid; 213044cd46caSbillm 2131088f3894Sahrens wp.wp_type = dn->dn_type; 2132088f3894Sahrens wp.wp_level = db->db_level; 2133088f3894Sahrens wp.wp_copies = os->os_copies; 2134088f3894Sahrens wp.wp_dncompress = dn->dn_compress; 2135088f3894Sahrens wp.wp_oscompress = os->os_compress; 2136088f3894Sahrens wp.wp_dnchecksum = dn->dn_checksum; 2137088f3894Sahrens wp.wp_oschecksum = os->os_checksum; 2138088f3894Sahrens 2139c717a561Smaybee if (BP_IS_OLDER(db->db_blkptr, txg)) 2140cdb0ab79Smaybee (void) dsl_dataset_block_kill( 2141c717a561Smaybee os->os_dsl_dataset, db->db_blkptr, zio, tx); 2142fa9e4066Sahrens 2143088f3894Sahrens dr->dr_zio = arc_write(zio, os->os_spa, &wp, 21443baa08fcSek DBUF_IS_L2CACHEABLE(db), txg, db->db_blkptr, 21453baa08fcSek data, dbuf_write_ready, dbuf_write_done, db, 2146e14bb325SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2147fa9e4066Sahrens } 2148fa9e4066Sahrens 2149fa9e4066Sahrens /* ARGSUSED */ 2150fa9e4066Sahrens static void 2151c717a561Smaybee dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2152fa9e4066Sahrens { 2153fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 2154fa9e4066Sahrens dnode_t *dn = db->db_dnode; 2155fa9e4066Sahrens objset_impl_t *os = dn->dn_objset; 2156e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2157c717a561Smaybee blkptr_t *bp_orig = &zio->io_bp_orig; 2158fa9e4066Sahrens uint64_t fill = 0; 2159c717a561Smaybee int old_size, new_size, i; 2160fa9e4066Sahrens 2161e14bb325SJeff Bonwick ASSERT(db->db_blkptr == bp); 2162e14bb325SJeff Bonwick 2163c717a561Smaybee dprintf_dbuf_bp(db, bp_orig, "bp_orig: %s", ""); 2164fa9e4066Sahrens 2165c717a561Smaybee old_size = bp_get_dasize(os->os_spa, bp_orig); 2166e14bb325SJeff Bonwick new_size = bp_get_dasize(os->os_spa, bp); 2167fa9e4066Sahrens 2168e14bb325SJeff Bonwick dnode_diduse_space(dn, new_size - old_size); 2169fa9e4066Sahrens 2170e14bb325SJeff Bonwick if (BP_IS_HOLE(bp)) { 2171c717a561Smaybee dsl_dataset_t *ds = os->os_dsl_dataset; 2172c717a561Smaybee dmu_tx_t *tx = os->os_synctx; 2173fa9e4066Sahrens 2174c717a561Smaybee if (bp_orig->blk_birth == tx->tx_txg) 2175e14bb325SJeff Bonwick (void) dsl_dataset_block_kill(ds, bp_orig, zio, tx); 2176e14bb325SJeff Bonwick ASSERT3U(bp->blk_fill, ==, 0); 2177c717a561Smaybee return; 2178c717a561Smaybee } 2179c5c6ffa0Smaybee 2180e14bb325SJeff Bonwick ASSERT(BP_GET_TYPE(bp) == dn->dn_type); 2181e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2182e14bb325SJeff Bonwick 2183c717a561Smaybee mutex_enter(&db->db_mtx); 2184fa9e4066Sahrens 2185fa9e4066Sahrens if (db->db_level == 0) { 2186fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 2187c717a561Smaybee if (db->db_blkid > dn->dn_phys->dn_maxblkid) 2188fa9e4066Sahrens dn->dn_phys->dn_maxblkid = db->db_blkid; 2189fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 2190fa9e4066Sahrens 2191fa9e4066Sahrens if (dn->dn_type == DMU_OT_DNODE) { 2192fa9e4066Sahrens dnode_phys_t *dnp = db->db.db_data; 2193fa9e4066Sahrens for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2194fa9e4066Sahrens i--, dnp++) { 2195fa9e4066Sahrens if (dnp->dn_type != DMU_OT_NONE) 2196fa9e4066Sahrens fill++; 2197fa9e4066Sahrens } 2198fa9e4066Sahrens } else { 2199c717a561Smaybee fill = 1; 2200fa9e4066Sahrens } 2201fa9e4066Sahrens } else { 2202e14bb325SJeff Bonwick blkptr_t *ibp = db->db.db_data; 2203fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2204e14bb325SJeff Bonwick for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2205e14bb325SJeff Bonwick if (BP_IS_HOLE(ibp)) 2206fa9e4066Sahrens continue; 2207e14bb325SJeff Bonwick ASSERT3U(BP_GET_LSIZE(ibp), ==, 2208fa9e4066Sahrens db->db_level == 1 ? dn->dn_datablksz : 2209fa9e4066Sahrens (1<<dn->dn_phys->dn_indblkshift)); 2210e14bb325SJeff Bonwick fill += ibp->blk_fill; 2211fa9e4066Sahrens } 2212fa9e4066Sahrens } 2213fa9e4066Sahrens 2214e14bb325SJeff Bonwick bp->blk_fill = fill; 2215c717a561Smaybee 2216c717a561Smaybee mutex_exit(&db->db_mtx); 2217c717a561Smaybee 2218e14bb325SJeff Bonwick if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 2219e14bb325SJeff Bonwick ASSERT(DVA_EQUAL(BP_IDENTITY(bp), BP_IDENTITY(bp_orig))); 2220e14bb325SJeff Bonwick } else { 2221c717a561Smaybee dsl_dataset_t *ds = os->os_dsl_dataset; 2222c717a561Smaybee dmu_tx_t *tx = os->os_synctx; 2223c717a561Smaybee 2224c717a561Smaybee if (bp_orig->blk_birth == tx->tx_txg) 2225e14bb325SJeff Bonwick (void) dsl_dataset_block_kill(ds, bp_orig, zio, tx); 2226e14bb325SJeff Bonwick dsl_dataset_block_born(ds, bp, tx); 2227fa9e4066Sahrens } 2228c717a561Smaybee } 2229fa9e4066Sahrens 2230c717a561Smaybee /* ARGSUSED */ 2231c717a561Smaybee static void 2232c717a561Smaybee dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2233c717a561Smaybee { 2234c717a561Smaybee dmu_buf_impl_t *db = vdb; 2235c717a561Smaybee uint64_t txg = zio->io_txg; 2236c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 2237c717a561Smaybee 2238c717a561Smaybee ASSERT3U(zio->io_error, ==, 0); 2239c717a561Smaybee 2240c717a561Smaybee mutex_enter(&db->db_mtx); 2241c717a561Smaybee 2242c717a561Smaybee drp = &db->db_last_dirty; 224317f17c2dSbonwick while ((dr = *drp) != db->db_data_pending) 224417f17c2dSbonwick drp = &dr->dr_next; 224517f17c2dSbonwick ASSERT(!list_link_active(&dr->dr_dirty_node)); 224617f17c2dSbonwick ASSERT(dr->dr_txg == txg); 224717f17c2dSbonwick ASSERT(dr->dr_next == NULL); 224817f17c2dSbonwick *drp = dr->dr_next; 2249c717a561Smaybee 2250c717a561Smaybee if (db->db_level == 0) { 2251c717a561Smaybee ASSERT(db->db_blkid != DB_BONUS_BLKID); 2252c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2253c717a561Smaybee 2254c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 2255c717a561Smaybee VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1); 2256c717a561Smaybee else if (!BP_IS_HOLE(db->db_blkptr)) 2257c717a561Smaybee arc_set_callback(db->db_buf, dbuf_do_evict, db); 2258c717a561Smaybee else 2259c717a561Smaybee ASSERT(arc_released(db->db_buf)); 2260c717a561Smaybee } else { 2261c717a561Smaybee dnode_t *dn = db->db_dnode; 2262c717a561Smaybee 2263c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2264c717a561Smaybee ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2265c717a561Smaybee if (!BP_IS_HOLE(db->db_blkptr)) { 2266c717a561Smaybee int epbs = 2267c717a561Smaybee dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2268c717a561Smaybee ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2269c717a561Smaybee db->db.db_size); 2270c717a561Smaybee ASSERT3U(dn->dn_phys->dn_maxblkid 2271c717a561Smaybee >> (db->db_level * epbs), >=, db->db_blkid); 2272c717a561Smaybee arc_set_callback(db->db_buf, dbuf_do_evict, db); 2273c717a561Smaybee } 2274c25056deSgw mutex_destroy(&dr->dt.di.dr_mtx); 2275c25056deSgw list_destroy(&dr->dt.di.dr_children); 2276c717a561Smaybee } 2277c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2278fa9e4066Sahrens 2279fa9e4066Sahrens cv_broadcast(&db->db_changed); 2280fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 2281fa9e4066Sahrens db->db_dirtycnt -= 1; 2282c717a561Smaybee db->db_data_pending = NULL; 2283fa9e4066Sahrens mutex_exit(&db->db_mtx); 2284fa9e4066Sahrens 2285c717a561Smaybee dprintf_dbuf_bp(db, zio->io_bp, "bp: %s", ""); 2286fa9e4066Sahrens 2287ea8dc4b6Seschrock dbuf_rele(db, (void *)(uintptr_t)txg); 2288fa9e4066Sahrens } 2289