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 /* 225a98e54bSBrendan Gregg - Sun Microsystems * Copyright 2009 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); 41fa9e4066Sahrens 42fa9e4066Sahrens /* 43fa9e4066Sahrens * Global data structures and functions for the dbuf cache. 44fa9e4066Sahrens */ 45fa9e4066Sahrens static kmem_cache_t *dbuf_cache; 46fa9e4066Sahrens 47fa9e4066Sahrens /* ARGSUSED */ 48fa9e4066Sahrens static int 49fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag) 50fa9e4066Sahrens { 51fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 52fa9e4066Sahrens bzero(db, sizeof (dmu_buf_impl_t)); 53fa9e4066Sahrens 54fa9e4066Sahrens mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 55fa9e4066Sahrens cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 56fa9e4066Sahrens refcount_create(&db->db_holds); 57fa9e4066Sahrens return (0); 58fa9e4066Sahrens } 59fa9e4066Sahrens 60fa9e4066Sahrens /* ARGSUSED */ 61fa9e4066Sahrens static void 62fa9e4066Sahrens dbuf_dest(void *vdb, void *unused) 63fa9e4066Sahrens { 64fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 65fa9e4066Sahrens mutex_destroy(&db->db_mtx); 66fa9e4066Sahrens cv_destroy(&db->db_changed); 67fa9e4066Sahrens refcount_destroy(&db->db_holds); 68fa9e4066Sahrens } 69fa9e4066Sahrens 70fa9e4066Sahrens /* 71fa9e4066Sahrens * dbuf hash table routines 72fa9e4066Sahrens */ 73fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table; 74fa9e4066Sahrens 75fa9e4066Sahrens static uint64_t dbuf_hash_count; 76fa9e4066Sahrens 77fa9e4066Sahrens static uint64_t 78fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 79fa9e4066Sahrens { 80fa9e4066Sahrens uintptr_t osv = (uintptr_t)os; 81fa9e4066Sahrens uint64_t crc = -1ULL; 82fa9e4066Sahrens 83fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 84fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 85fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 86fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 87fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 88fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 89fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 90fa9e4066Sahrens 91fa9e4066Sahrens crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 92fa9e4066Sahrens 93fa9e4066Sahrens return (crc); 94fa9e4066Sahrens } 95fa9e4066Sahrens 96fa9e4066Sahrens #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 97fa9e4066Sahrens 98fa9e4066Sahrens #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 99fa9e4066Sahrens ((dbuf)->db.db_object == (obj) && \ 100fa9e4066Sahrens (dbuf)->db_objset == (os) && \ 101fa9e4066Sahrens (dbuf)->db_level == (level) && \ 102fa9e4066Sahrens (dbuf)->db_blkid == (blkid)) 103fa9e4066Sahrens 104fa9e4066Sahrens dmu_buf_impl_t * 105fa9e4066Sahrens dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 106fa9e4066Sahrens { 107fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 108503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 109fa9e4066Sahrens uint64_t obj = dn->dn_object; 110fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 111fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 112fa9e4066Sahrens dmu_buf_impl_t *db; 113fa9e4066Sahrens 114fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 115fa9e4066Sahrens for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 116fa9e4066Sahrens if (DBUF_EQUAL(db, os, obj, level, blkid)) { 117fa9e4066Sahrens mutex_enter(&db->db_mtx); 118ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 119fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 120fa9e4066Sahrens return (db); 121fa9e4066Sahrens } 122fa9e4066Sahrens mutex_exit(&db->db_mtx); 123fa9e4066Sahrens } 124fa9e4066Sahrens } 125fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 126fa9e4066Sahrens return (NULL); 127fa9e4066Sahrens } 128fa9e4066Sahrens 129fa9e4066Sahrens /* 130fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 131fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 132fa9e4066Sahrens * will be returned and the new element will not be inserted. 133fa9e4066Sahrens * Otherwise returns NULL. 134fa9e4066Sahrens */ 135fa9e4066Sahrens static dmu_buf_impl_t * 136fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db) 137fa9e4066Sahrens { 138fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 139503ad85cSMatthew Ahrens objset_t *os = db->db_objset; 140fa9e4066Sahrens uint64_t obj = db->db.db_object; 141fa9e4066Sahrens int level = db->db_level; 142fa9e4066Sahrens uint64_t blkid = db->db_blkid; 143fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 144fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 145fa9e4066Sahrens dmu_buf_impl_t *dbf; 146fa9e4066Sahrens 147fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 148fa9e4066Sahrens for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 149fa9e4066Sahrens if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 150fa9e4066Sahrens mutex_enter(&dbf->db_mtx); 151ea8dc4b6Seschrock if (dbf->db_state != DB_EVICTING) { 152fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 153fa9e4066Sahrens return (dbf); 154fa9e4066Sahrens } 155fa9e4066Sahrens mutex_exit(&dbf->db_mtx); 156fa9e4066Sahrens } 157fa9e4066Sahrens } 158fa9e4066Sahrens 159fa9e4066Sahrens mutex_enter(&db->db_mtx); 160fa9e4066Sahrens db->db_hash_next = h->hash_table[idx]; 161fa9e4066Sahrens h->hash_table[idx] = db; 162fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 163fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, 1); 164fa9e4066Sahrens 165fa9e4066Sahrens return (NULL); 166fa9e4066Sahrens } 167fa9e4066Sahrens 168fa9e4066Sahrens /* 169fa9e4066Sahrens * Remove an entry from the hash table. This operation will 170fa9e4066Sahrens * fail if there are any existing holds on the db. 171fa9e4066Sahrens */ 172fa9e4066Sahrens static void 173fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db) 174fa9e4066Sahrens { 175fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 176fa9e4066Sahrens uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 177fa9e4066Sahrens db->db_level, db->db_blkid); 178fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 179fa9e4066Sahrens dmu_buf_impl_t *dbf, **dbp; 180fa9e4066Sahrens 181fa9e4066Sahrens /* 182fa9e4066Sahrens * We musn't hold db_mtx to maintin lock ordering: 183fa9e4066Sahrens * DBUF_HASH_MUTEX > db_mtx. 184fa9e4066Sahrens */ 185fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 186ea8dc4b6Seschrock ASSERT(db->db_state == DB_EVICTING); 187fa9e4066Sahrens ASSERT(!MUTEX_HELD(&db->db_mtx)); 188fa9e4066Sahrens 189fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 190fa9e4066Sahrens dbp = &h->hash_table[idx]; 191fa9e4066Sahrens while ((dbf = *dbp) != db) { 192fa9e4066Sahrens dbp = &dbf->db_hash_next; 193fa9e4066Sahrens ASSERT(dbf != NULL); 194fa9e4066Sahrens } 195fa9e4066Sahrens *dbp = db->db_hash_next; 196fa9e4066Sahrens db->db_hash_next = NULL; 197fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 198fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, -1); 199fa9e4066Sahrens } 200fa9e4066Sahrens 201ea8dc4b6Seschrock static arc_evict_func_t dbuf_do_evict; 202fa9e4066Sahrens 203fa9e4066Sahrens static void 204fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db) 205fa9e4066Sahrens { 206fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 207fa9e4066Sahrens 208c717a561Smaybee if (db->db_level != 0 || db->db_evict_func == NULL) 209fa9e4066Sahrens return; 210fa9e4066Sahrens 211c717a561Smaybee if (db->db_user_data_ptr_ptr) 212c717a561Smaybee *db->db_user_data_ptr_ptr = db->db.db_data; 213c717a561Smaybee db->db_evict_func(&db->db, db->db_user_ptr); 214c717a561Smaybee db->db_user_ptr = NULL; 215c717a561Smaybee db->db_user_data_ptr_ptr = NULL; 216c717a561Smaybee db->db_evict_func = NULL; 217fa9e4066Sahrens } 218fa9e4066Sahrens 219fa9e4066Sahrens void 220ea8dc4b6Seschrock dbuf_evict(dmu_buf_impl_t *db) 221ea8dc4b6Seschrock { 222ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 223ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 224c717a561Smaybee ASSERT(db->db_data_pending == NULL); 225ea8dc4b6Seschrock 226ea8dc4b6Seschrock dbuf_clear(db); 227ea8dc4b6Seschrock dbuf_destroy(db); 228ea8dc4b6Seschrock } 229ea8dc4b6Seschrock 230ea8dc4b6Seschrock void 231fa9e4066Sahrens dbuf_init(void) 232fa9e4066Sahrens { 233ea8dc4b6Seschrock uint64_t hsize = 1ULL << 16; 234fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 235fa9e4066Sahrens int i; 236fa9e4066Sahrens 237fa9e4066Sahrens /* 238fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 239ea8dc4b6Seschrock * with an average 4K block size. The table will take up 240ea8dc4b6Seschrock * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 241fa9e4066Sahrens */ 242ea8dc4b6Seschrock while (hsize * 4096 < physmem * PAGESIZE) 243fa9e4066Sahrens hsize <<= 1; 244fa9e4066Sahrens 245ea8dc4b6Seschrock retry: 246fa9e4066Sahrens h->hash_table_mask = hsize - 1; 247ea8dc4b6Seschrock h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 248ea8dc4b6Seschrock if (h->hash_table == NULL) { 249ea8dc4b6Seschrock /* XXX - we should really return an error instead of assert */ 250ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 10)); 251ea8dc4b6Seschrock hsize >>= 1; 252ea8dc4b6Seschrock goto retry; 253ea8dc4b6Seschrock } 254fa9e4066Sahrens 255fa9e4066Sahrens dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 256fa9e4066Sahrens sizeof (dmu_buf_impl_t), 257fa9e4066Sahrens 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 258fa9e4066Sahrens 259fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 260fa9e4066Sahrens mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 261fa9e4066Sahrens } 262fa9e4066Sahrens 263fa9e4066Sahrens void 264fa9e4066Sahrens dbuf_fini(void) 265fa9e4066Sahrens { 266fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 267fa9e4066Sahrens int i; 268fa9e4066Sahrens 269fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 270fa9e4066Sahrens mutex_destroy(&h->hash_mutexes[i]); 271fa9e4066Sahrens kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 272fa9e4066Sahrens kmem_cache_destroy(dbuf_cache); 273fa9e4066Sahrens } 274fa9e4066Sahrens 275fa9e4066Sahrens /* 276fa9e4066Sahrens * Other stuff. 277fa9e4066Sahrens */ 278fa9e4066Sahrens 2799c9dc39aSek110237 #ifdef ZFS_DEBUG 280fa9e4066Sahrens static void 281fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db) 282fa9e4066Sahrens { 283fa9e4066Sahrens dnode_t *dn = db->db_dnode; 284*b24ab676SJeff Bonwick dbuf_dirty_record_t *dr; 285fa9e4066Sahrens 286fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 287fa9e4066Sahrens 288fa9e4066Sahrens if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 289fa9e4066Sahrens return; 290fa9e4066Sahrens 291fa9e4066Sahrens ASSERT(db->db_objset != NULL); 292fa9e4066Sahrens if (dn == NULL) { 293fa9e4066Sahrens ASSERT(db->db_parent == NULL); 294fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 295fa9e4066Sahrens } else { 296fa9e4066Sahrens ASSERT3U(db->db.db_object, ==, dn->dn_object); 297fa9e4066Sahrens ASSERT3P(db->db_objset, ==, dn->dn_objset); 298fa9e4066Sahrens ASSERT3U(db->db_level, <, dn->dn_nlevels); 299ea8dc4b6Seschrock ASSERT(db->db_blkid == DB_BONUS_BLKID || 300ea8dc4b6Seschrock list_head(&dn->dn_dbufs)); 301fa9e4066Sahrens } 302fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 303fa9e4066Sahrens ASSERT(dn != NULL); 3041934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 305fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID); 306fa9e4066Sahrens } else { 307fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 308fa9e4066Sahrens } 309fa9e4066Sahrens 310*b24ab676SJeff Bonwick for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 311*b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 312*b24ab676SJeff Bonwick 313*b24ab676SJeff Bonwick for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 314*b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 315*b24ab676SJeff Bonwick 31688b7b0f2SMatthew Ahrens /* 31788b7b0f2SMatthew Ahrens * We can't assert that db_size matches dn_datablksz because it 31888b7b0f2SMatthew Ahrens * can be momentarily different when another thread is doing 31988b7b0f2SMatthew Ahrens * dnode_set_blksz(). 32088b7b0f2SMatthew Ahrens */ 32188b7b0f2SMatthew Ahrens if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 322*b24ab676SJeff Bonwick dr = db->db_data_pending; 323fa9e4066Sahrens /* 32488b7b0f2SMatthew Ahrens * It should only be modified in syncing context, so 32588b7b0f2SMatthew Ahrens * make sure we only have one copy of the data. 326fa9e4066Sahrens */ 327c717a561Smaybee ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 328fa9e4066Sahrens } 329fa9e4066Sahrens 330fa9e4066Sahrens /* verify db->db_blkptr */ 331fa9e4066Sahrens if (db->db_blkptr) { 332fa9e4066Sahrens if (db->db_parent == dn->dn_dbuf) { 333fa9e4066Sahrens /* db is pointed to by the dnode */ 334fa9e4066Sahrens /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 33514843421SMatthew Ahrens if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 336fa9e4066Sahrens ASSERT(db->db_parent == NULL); 337fa9e4066Sahrens else 338fa9e4066Sahrens ASSERT(db->db_parent != NULL); 339fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 340fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 341fa9e4066Sahrens } else { 342fa9e4066Sahrens /* db is pointed to by an indirect block */ 343fa9e4066Sahrens int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 344fa9e4066Sahrens ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 345fa9e4066Sahrens ASSERT3U(db->db_parent->db.db_object, ==, 346fa9e4066Sahrens db->db.db_object); 347fa9e4066Sahrens /* 348fa9e4066Sahrens * dnode_grow_indblksz() can make this fail if we don't 349fa9e4066Sahrens * have the struct_rwlock. XXX indblksz no longer 350fa9e4066Sahrens * grows. safe to do this now? 351fa9e4066Sahrens */ 352fa9e4066Sahrens if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) { 353fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 354fa9e4066Sahrens ((blkptr_t *)db->db_parent->db.db_data + 355fa9e4066Sahrens db->db_blkid % epb)); 356fa9e4066Sahrens } 357fa9e4066Sahrens } 358fa9e4066Sahrens } 359fa9e4066Sahrens if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 360fa9e4066Sahrens db->db.db_data && db->db_blkid != DB_BONUS_BLKID && 361fa9e4066Sahrens db->db_state != DB_FILL && !dn->dn_free_txg) { 362fa9e4066Sahrens /* 363fa9e4066Sahrens * If the blkptr isn't set but they have nonzero data, 364fa9e4066Sahrens * it had better be dirty, otherwise we'll lose that 365fa9e4066Sahrens * data when we evict this buffer. 366fa9e4066Sahrens */ 367fa9e4066Sahrens if (db->db_dirtycnt == 0) { 368fa9e4066Sahrens uint64_t *buf = db->db.db_data; 369fa9e4066Sahrens int i; 370fa9e4066Sahrens 371fa9e4066Sahrens for (i = 0; i < db->db.db_size >> 3; i++) { 372fa9e4066Sahrens ASSERT(buf[i] == 0); 373fa9e4066Sahrens } 374fa9e4066Sahrens } 375fa9e4066Sahrens } 376fa9e4066Sahrens } 3779c9dc39aSek110237 #endif 378fa9e4066Sahrens 379fa9e4066Sahrens static void 380fa9e4066Sahrens dbuf_update_data(dmu_buf_impl_t *db) 381fa9e4066Sahrens { 382fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 383c717a561Smaybee if (db->db_level == 0 && db->db_user_data_ptr_ptr) { 384fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 385c717a561Smaybee *db->db_user_data_ptr_ptr = db->db.db_data; 386fa9e4066Sahrens } 387fa9e4066Sahrens } 388fa9e4066Sahrens 389fa9e4066Sahrens static void 390fa9e4066Sahrens dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 391fa9e4066Sahrens { 392fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 393ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 394fa9e4066Sahrens db->db_buf = buf; 395ea8dc4b6Seschrock if (buf != NULL) { 396ea8dc4b6Seschrock ASSERT(buf->b_data != NULL); 397fa9e4066Sahrens db->db.db_data = buf->b_data; 398ea8dc4b6Seschrock if (!arc_released(buf)) 399ea8dc4b6Seschrock arc_set_callback(buf, dbuf_do_evict, db); 400fa9e4066Sahrens dbuf_update_data(db); 401ea8dc4b6Seschrock } else { 402ea8dc4b6Seschrock dbuf_evict_user(db); 403ea8dc4b6Seschrock db->db.db_data = NULL; 40482c9918fSTim Haley if (db->db_state != DB_NOFILL) 405ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 406ea8dc4b6Seschrock } 407fa9e4066Sahrens } 408fa9e4066Sahrens 409fa9e4066Sahrens uint64_t 410fa9e4066Sahrens dbuf_whichblock(dnode_t *dn, uint64_t offset) 411fa9e4066Sahrens { 412fa9e4066Sahrens if (dn->dn_datablkshift) { 413fa9e4066Sahrens return (offset >> dn->dn_datablkshift); 414fa9e4066Sahrens } else { 415fa9e4066Sahrens ASSERT3U(offset, <, dn->dn_datablksz); 416fa9e4066Sahrens return (0); 417fa9e4066Sahrens } 418fa9e4066Sahrens } 419fa9e4066Sahrens 420fa9e4066Sahrens static void 421fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 422fa9e4066Sahrens { 423fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 424fa9e4066Sahrens 425fa9e4066Sahrens mutex_enter(&db->db_mtx); 426fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_READ); 427fa9e4066Sahrens /* 428fa9e4066Sahrens * All reads are synchronous, so we must have a hold on the dbuf 429fa9e4066Sahrens */ 430fa9e4066Sahrens ASSERT(refcount_count(&db->db_holds) > 0); 431ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 432fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 433c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 434fa9e4066Sahrens /* we were freed in flight; disregard any error */ 435fa9e4066Sahrens arc_release(buf, db); 436fa9e4066Sahrens bzero(buf->b_data, db->db.db_size); 4376b4acc8bSahrens arc_buf_freeze(buf); 438c717a561Smaybee db->db_freed_in_flight = FALSE; 439fa9e4066Sahrens dbuf_set_data(db, buf); 440fa9e4066Sahrens db->db_state = DB_CACHED; 441fa9e4066Sahrens } else if (zio == NULL || zio->io_error == 0) { 442fa9e4066Sahrens dbuf_set_data(db, buf); 443fa9e4066Sahrens db->db_state = DB_CACHED; 444fa9e4066Sahrens } else { 445fa9e4066Sahrens ASSERT(db->db_blkid != DB_BONUS_BLKID); 446fa9e4066Sahrens ASSERT3P(db->db_buf, ==, NULL); 447ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 448ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 449fa9e4066Sahrens } 450fa9e4066Sahrens cv_broadcast(&db->db_changed); 451fa9e4066Sahrens mutex_exit(&db->db_mtx); 452ea8dc4b6Seschrock dbuf_rele(db, NULL); 453fa9e4066Sahrens } 454fa9e4066Sahrens 455ea8dc4b6Seschrock static void 45613506d1eSmaybee dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 457fa9e4066Sahrens { 458088f3894Sahrens dnode_t *dn = db->db_dnode; 459ea8dc4b6Seschrock zbookmark_t zb; 46013506d1eSmaybee uint32_t aflags = ARC_NOWAIT; 461088f3894Sahrens arc_buf_t *pbuf; 462fa9e4066Sahrens 463fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 464fa9e4066Sahrens /* We need the struct_rwlock to prevent db_blkptr from changing. */ 465088f3894Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 466ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 467ea8dc4b6Seschrock ASSERT(db->db_state == DB_UNCACHED); 468ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 469fa9e4066Sahrens 470fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 471cf04dda1SMark Maybee int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 4721934e92fSmaybee 4731934e92fSmaybee ASSERT3U(bonuslen, <=, db->db.db_size); 474ea8dc4b6Seschrock db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 4755a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 4761934e92fSmaybee if (bonuslen < DN_MAX_BONUSLEN) 477ea8dc4b6Seschrock bzero(db->db.db_data, DN_MAX_BONUSLEN); 478cf04dda1SMark Maybee if (bonuslen) 479cf04dda1SMark Maybee bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 480ea8dc4b6Seschrock dbuf_update_data(db); 481fa9e4066Sahrens db->db_state = DB_CACHED; 482fa9e4066Sahrens mutex_exit(&db->db_mtx); 483fa9e4066Sahrens return; 484fa9e4066Sahrens } 485fa9e4066Sahrens 4861c8564a7SMark Maybee /* 4871c8564a7SMark Maybee * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 4881c8564a7SMark Maybee * processes the delete record and clears the bp while we are waiting 4891c8564a7SMark Maybee * for the dn_mtx (resulting in a "no" from block_freed). 4901c8564a7SMark Maybee */ 491088f3894Sahrens if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 4921c8564a7SMark Maybee (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 4931c8564a7SMark Maybee BP_IS_HOLE(db->db_blkptr)))) { 494ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 495ad23a2dbSjohansen 496088f3894Sahrens dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa, 497ad23a2dbSjohansen db->db.db_size, db, type)); 498fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 499fa9e4066Sahrens db->db_state = DB_CACHED; 50013506d1eSmaybee *flags |= DB_RF_CACHED; 501fa9e4066Sahrens mutex_exit(&db->db_mtx); 502fa9e4066Sahrens return; 503fa9e4066Sahrens } 504fa9e4066Sahrens 505fa9e4066Sahrens db->db_state = DB_READ; 506fa9e4066Sahrens mutex_exit(&db->db_mtx); 507fa9e4066Sahrens 5083baa08fcSek110237 if (DBUF_IS_L2CACHEABLE(db)) 5093baa08fcSek110237 aflags |= ARC_L2CACHE; 5103baa08fcSek110237 511*b24ab676SJeff Bonwick SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 512*b24ab676SJeff Bonwick db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 513*b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 514ea8dc4b6Seschrock 515ea8dc4b6Seschrock dbuf_add_ref(db, NULL); 516fa9e4066Sahrens /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ 517088f3894Sahrens 518088f3894Sahrens if (db->db_parent) 519088f3894Sahrens pbuf = db->db_parent->db_buf; 520088f3894Sahrens else 521088f3894Sahrens pbuf = db->db_objset->os_phys_buf; 522088f3894Sahrens 523088f3894Sahrens (void) arc_read(zio, dn->dn_objset->os_spa, db->db_blkptr, pbuf, 524fa9e4066Sahrens dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 52513506d1eSmaybee (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 52613506d1eSmaybee &aflags, &zb); 52713506d1eSmaybee if (aflags & ARC_CACHED) 52813506d1eSmaybee *flags |= DB_RF_CACHED; 529fa9e4066Sahrens } 530fa9e4066Sahrens 531ea8dc4b6Seschrock int 532ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 533fa9e4066Sahrens { 534ea8dc4b6Seschrock int err = 0; 535ea8dc4b6Seschrock int havepzio = (zio != NULL); 53613506d1eSmaybee int prefetch; 537fa9e4066Sahrens 538fa9e4066Sahrens /* 539fa9e4066Sahrens * We don't have to hold the mutex to check db_state because it 540fa9e4066Sahrens * can't be freed while we have a hold on the buffer. 541fa9e4066Sahrens */ 542fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 543fa9e4066Sahrens 54482c9918fSTim Haley if (db->db_state == DB_NOFILL) 54582c9918fSTim Haley return (EIO); 54682c9918fSTim Haley 547fa9e4066Sahrens if ((flags & DB_RF_HAVESTRUCT) == 0) 548fa9e4066Sahrens rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER); 549fa9e4066Sahrens 55013506d1eSmaybee prefetch = db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 5513baa08fcSek110237 (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL && 5523baa08fcSek110237 DBUF_IS_CACHEABLE(db); 55313506d1eSmaybee 554fa9e4066Sahrens mutex_enter(&db->db_mtx); 555ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 556ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 55713506d1eSmaybee if (prefetch) 55813506d1eSmaybee dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 55913506d1eSmaybee db->db.db_size, TRUE); 560ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 561ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 562ea8dc4b6Seschrock } else if (db->db_state == DB_UNCACHED) { 563ea8dc4b6Seschrock if (zio == NULL) { 564ea8dc4b6Seschrock zio = zio_root(db->db_dnode->dn_objset->os_spa, 565ea8dc4b6Seschrock NULL, NULL, ZIO_FLAG_CANFAIL); 566ea8dc4b6Seschrock } 56713506d1eSmaybee dbuf_read_impl(db, zio, &flags); 56813506d1eSmaybee 569ea8dc4b6Seschrock /* dbuf_read_impl has dropped db_mtx for us */ 570ea8dc4b6Seschrock 57113506d1eSmaybee if (prefetch) 572ea8dc4b6Seschrock dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 57313506d1eSmaybee db->db.db_size, flags & DB_RF_CACHED); 574ea8dc4b6Seschrock 575ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 576ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 577ea8dc4b6Seschrock 578ea8dc4b6Seschrock if (!havepzio) 579ea8dc4b6Seschrock err = zio_wait(zio); 580ea8dc4b6Seschrock } else { 58113506d1eSmaybee mutex_exit(&db->db_mtx); 58213506d1eSmaybee if (prefetch) 58313506d1eSmaybee dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 58413506d1eSmaybee db->db.db_size, TRUE); 585ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 586ea8dc4b6Seschrock rw_exit(&db->db_dnode->dn_struct_rwlock); 58713506d1eSmaybee 58813506d1eSmaybee mutex_enter(&db->db_mtx); 589ea8dc4b6Seschrock if ((flags & DB_RF_NEVERWAIT) == 0) { 590ea8dc4b6Seschrock while (db->db_state == DB_READ || 591ea8dc4b6Seschrock db->db_state == DB_FILL) { 592fa9e4066Sahrens ASSERT(db->db_state == DB_READ || 593fa9e4066Sahrens (flags & DB_RF_HAVESTRUCT) == 0); 594fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 595fa9e4066Sahrens } 596ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED) 597ea8dc4b6Seschrock err = EIO; 598ea8dc4b6Seschrock } 599fa9e4066Sahrens mutex_exit(&db->db_mtx); 600fa9e4066Sahrens } 601fa9e4066Sahrens 602ea8dc4b6Seschrock ASSERT(err || havepzio || db->db_state == DB_CACHED); 603ea8dc4b6Seschrock return (err); 604fa9e4066Sahrens } 605fa9e4066Sahrens 606fa9e4066Sahrens static void 607fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db) 608fa9e4066Sahrens { 609fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 610ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 611fa9e4066Sahrens mutex_enter(&db->db_mtx); 612fa9e4066Sahrens while (db->db_state == DB_READ || db->db_state == DB_FILL) 613fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 614fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 615ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 616ad23a2dbSjohansen 617ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 618fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 619fa9e4066Sahrens dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 620ad23a2dbSjohansen db->db.db_size, db, type)); 621fa9e4066Sahrens db->db_state = DB_FILL; 62282c9918fSTim Haley } else if (db->db_state == DB_NOFILL) { 62382c9918fSTim Haley dbuf_set_data(db, NULL); 624fa9e4066Sahrens } else { 625fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 626fa9e4066Sahrens } 627fa9e4066Sahrens mutex_exit(&db->db_mtx); 628fa9e4066Sahrens } 629fa9e4066Sahrens 630fa9e4066Sahrens /* 631fa9e4066Sahrens * This is our just-in-time copy function. It makes a copy of 632fa9e4066Sahrens * buffers, that have been modified in a previous transaction 633fa9e4066Sahrens * group, before we modify them in the current active group. 634fa9e4066Sahrens * 635fa9e4066Sahrens * This function is used in two places: when we are dirtying a 636fa9e4066Sahrens * buffer for the first time in a txg, and when we are freeing 637fa9e4066Sahrens * a range in a dnode that includes this buffer. 638fa9e4066Sahrens * 639fa9e4066Sahrens * Note that when we are called from dbuf_free_range() we do 640fa9e4066Sahrens * not put a hold on the buffer, we just traverse the active 641fa9e4066Sahrens * dbuf list for the dnode. 642fa9e4066Sahrens */ 643fa9e4066Sahrens static void 644fa9e4066Sahrens dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 645fa9e4066Sahrens { 646c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 647fa9e4066Sahrens 648fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 649fa9e4066Sahrens ASSERT(db->db.db_data != NULL); 650c717a561Smaybee ASSERT(db->db_level == 0); 651c717a561Smaybee ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 652fa9e4066Sahrens 6534d31c452Smaybee if (dr == NULL || 6544d31c452Smaybee (dr->dt.dl.dr_data != 6554d31c452Smaybee ((db->db_blkid == DB_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 656fa9e4066Sahrens return; 657fa9e4066Sahrens 658fa9e4066Sahrens /* 659c717a561Smaybee * If the last dirty record for this dbuf has not yet synced 660c717a561Smaybee * and its referencing the dbuf data, either: 661c717a561Smaybee * reset the reference to point to a new copy, 662c717a561Smaybee * or (if there a no active holders) 663c717a561Smaybee * just null out the current db_data pointer. 664fa9e4066Sahrens */ 665c717a561Smaybee ASSERT(dr->dr_txg >= txg - 2); 666c717a561Smaybee if (db->db_blkid == DB_BONUS_BLKID) { 667c717a561Smaybee /* Note that the data bufs here are zio_bufs */ 668c717a561Smaybee dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 6695a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 670c717a561Smaybee bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 671c717a561Smaybee } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 672ea8dc4b6Seschrock int size = db->db.db_size; 673c717a561Smaybee arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 674c717a561Smaybee dr->dt.dl.dr_data = arc_buf_alloc( 675ad23a2dbSjohansen db->db_dnode->dn_objset->os_spa, size, db, type); 676c717a561Smaybee bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 677fa9e4066Sahrens } else { 678ea8dc4b6Seschrock dbuf_set_data(db, NULL); 679fa9e4066Sahrens } 680fa9e4066Sahrens } 681ea8dc4b6Seschrock 682fa9e4066Sahrens void 683c717a561Smaybee dbuf_unoverride(dbuf_dirty_record_t *dr) 684fa9e4066Sahrens { 685c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 686*b24ab676SJeff Bonwick blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 687c717a561Smaybee uint64_t txg = dr->dr_txg; 688c5c6ffa0Smaybee 689c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 690c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 691c717a561Smaybee ASSERT(db->db_level == 0); 692c717a561Smaybee 693c717a561Smaybee if (db->db_blkid == DB_BONUS_BLKID || 694c717a561Smaybee dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 695c717a561Smaybee return; 696c717a561Smaybee 697*b24ab676SJeff Bonwick ASSERT(db->db_data_pending != dr); 698*b24ab676SJeff Bonwick 699fa9e4066Sahrens /* free this block */ 700*b24ab676SJeff Bonwick if (!BP_IS_HOLE(bp)) 701*b24ab676SJeff Bonwick dsl_free(spa_get_dsl(db->db_dnode->dn_objset->os_spa), txg, bp); 702*b24ab676SJeff Bonwick 703c717a561Smaybee dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 7046b4acc8bSahrens /* 7056b4acc8bSahrens * Release the already-written buffer, so we leave it in 7066b4acc8bSahrens * a consistent dirty state. Note that all callers are 7076b4acc8bSahrens * modifying the buffer, so they will immediately do 7086b4acc8bSahrens * another (redundant) arc_release(). Therefore, leave 7096b4acc8bSahrens * the buf thawed to save the effort of freezing & 7106b4acc8bSahrens * immediately re-thawing it. 7116b4acc8bSahrens */ 712c717a561Smaybee arc_release(dr->dt.dl.dr_data, db); 713fa9e4066Sahrens } 714fa9e4066Sahrens 715cdb0ab79Smaybee /* 716cdb0ab79Smaybee * Evict (if its unreferenced) or clear (if its referenced) any level-0 717cdb0ab79Smaybee * data blocks in the free range, so that any future readers will find 718cdb0ab79Smaybee * empty blocks. Also, if we happen accross any level-1 dbufs in the 719cdb0ab79Smaybee * range that have not already been marked dirty, mark them dirty so 720cdb0ab79Smaybee * they stay in memory. 721cdb0ab79Smaybee */ 722fa9e4066Sahrens void 723cdb0ab79Smaybee dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) 724fa9e4066Sahrens { 725fa9e4066Sahrens dmu_buf_impl_t *db, *db_next; 726fa9e4066Sahrens uint64_t txg = tx->tx_txg; 727cdb0ab79Smaybee int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 728cdb0ab79Smaybee uint64_t first_l1 = start >> epbs; 729cdb0ab79Smaybee uint64_t last_l1 = end >> epbs; 730fa9e4066Sahrens 731cdb0ab79Smaybee if (end > dn->dn_maxblkid) { 732cdb0ab79Smaybee end = dn->dn_maxblkid; 733cdb0ab79Smaybee last_l1 = end >> epbs; 734cdb0ab79Smaybee } 735cdb0ab79Smaybee dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); 736fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 737fa9e4066Sahrens for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 738fa9e4066Sahrens db_next = list_next(&dn->dn_dbufs, db); 739ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 740cdb0ab79Smaybee 741cdb0ab79Smaybee if (db->db_level == 1 && 742cdb0ab79Smaybee db->db_blkid >= first_l1 && db->db_blkid <= last_l1) { 743cdb0ab79Smaybee mutex_enter(&db->db_mtx); 744cdb0ab79Smaybee if (db->db_last_dirty && 745cdb0ab79Smaybee db->db_last_dirty->dr_txg < txg) { 746cdb0ab79Smaybee dbuf_add_ref(db, FTAG); 747cdb0ab79Smaybee mutex_exit(&db->db_mtx); 748cdb0ab79Smaybee dbuf_will_dirty(db, tx); 749cdb0ab79Smaybee dbuf_rele(db, FTAG); 750cdb0ab79Smaybee } else { 751cdb0ab79Smaybee mutex_exit(&db->db_mtx); 752cdb0ab79Smaybee } 753cdb0ab79Smaybee } 754cdb0ab79Smaybee 755ea8dc4b6Seschrock if (db->db_level != 0) 756fa9e4066Sahrens continue; 757fa9e4066Sahrens dprintf_dbuf(db, "found buf %s\n", ""); 758cdb0ab79Smaybee if (db->db_blkid < start || db->db_blkid > end) 759fa9e4066Sahrens continue; 760fa9e4066Sahrens 761fa9e4066Sahrens /* found a level 0 buffer in the range */ 762fa9e4066Sahrens if (dbuf_undirty(db, tx)) 763fa9e4066Sahrens continue; 764fa9e4066Sahrens 765fa9e4066Sahrens mutex_enter(&db->db_mtx); 766ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED || 76782c9918fSTim Haley db->db_state == DB_NOFILL || 768ea8dc4b6Seschrock db->db_state == DB_EVICTING) { 769fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 770fa9e4066Sahrens mutex_exit(&db->db_mtx); 771fa9e4066Sahrens continue; 772fa9e4066Sahrens } 773c543ec06Sahrens if (db->db_state == DB_READ || db->db_state == DB_FILL) { 774c543ec06Sahrens /* will be handled in dbuf_read_done or dbuf_rele */ 775c717a561Smaybee db->db_freed_in_flight = TRUE; 776fa9e4066Sahrens mutex_exit(&db->db_mtx); 777fa9e4066Sahrens continue; 778fa9e4066Sahrens } 779ea8dc4b6Seschrock if (refcount_count(&db->db_holds) == 0) { 780ea8dc4b6Seschrock ASSERT(db->db_buf); 781ea8dc4b6Seschrock dbuf_clear(db); 782ea8dc4b6Seschrock continue; 783ea8dc4b6Seschrock } 784c717a561Smaybee /* The dbuf is referenced */ 785fa9e4066Sahrens 786c717a561Smaybee if (db->db_last_dirty != NULL) { 787c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 788c717a561Smaybee 789c717a561Smaybee if (dr->dr_txg == txg) { 790ea8dc4b6Seschrock /* 791c717a561Smaybee * This buffer is "in-use", re-adjust the file 792c717a561Smaybee * size to reflect that this buffer may 793c717a561Smaybee * contain new data when we sync. 794ea8dc4b6Seschrock */ 79544eda4d7Smaybee if (db->db_blkid > dn->dn_maxblkid) 79644eda4d7Smaybee dn->dn_maxblkid = db->db_blkid; 797c717a561Smaybee dbuf_unoverride(dr); 798c717a561Smaybee } else { 799c717a561Smaybee /* 800c717a561Smaybee * This dbuf is not dirty in the open context. 801c717a561Smaybee * Either uncache it (if its not referenced in 802c717a561Smaybee * the open context) or reset its contents to 803c717a561Smaybee * empty. 804c717a561Smaybee */ 805c717a561Smaybee dbuf_fix_old_data(db, txg); 80644eda4d7Smaybee } 807c717a561Smaybee } 808c717a561Smaybee /* clear the contents if its cached */ 809ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 810ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 811fa9e4066Sahrens arc_release(db->db_buf, db); 812fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 8136b4acc8bSahrens arc_buf_freeze(db->db_buf); 814fa9e4066Sahrens } 815ea8dc4b6Seschrock 816fa9e4066Sahrens mutex_exit(&db->db_mtx); 817fa9e4066Sahrens } 818fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 819fa9e4066Sahrens } 820fa9e4066Sahrens 821fa9e4066Sahrens static int 8221934e92fSmaybee dbuf_block_freeable(dmu_buf_impl_t *db) 823fa9e4066Sahrens { 824fa9e4066Sahrens dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 825fa9e4066Sahrens uint64_t birth_txg = 0; 826fa9e4066Sahrens 827fa9e4066Sahrens /* 828fa9e4066Sahrens * We don't need any locking to protect db_blkptr: 829c717a561Smaybee * If it's syncing, then db_last_dirty will be set 830c717a561Smaybee * so we'll ignore db_blkptr. 831fa9e4066Sahrens */ 832c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 833c717a561Smaybee if (db->db_last_dirty) 834c717a561Smaybee birth_txg = db->db_last_dirty->dr_txg; 835fa9e4066Sahrens else if (db->db_blkptr) 836fa9e4066Sahrens birth_txg = db->db_blkptr->blk_birth; 837fa9e4066Sahrens 8381934e92fSmaybee /* If we don't exist or are in a snapshot, we can't be freed */ 839fa9e4066Sahrens if (birth_txg) 8401934e92fSmaybee return (ds == NULL || 8411934e92fSmaybee dsl_dataset_block_freeable(ds, birth_txg)); 842fa9e4066Sahrens else 8431934e92fSmaybee return (FALSE); 844fa9e4066Sahrens } 845fa9e4066Sahrens 846fa9e4066Sahrens void 847fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 848fa9e4066Sahrens { 849fa9e4066Sahrens arc_buf_t *buf, *obuf; 850fa9e4066Sahrens int osize = db->db.db_size; 851ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 852fa9e4066Sahrens 853ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 854ea8dc4b6Seschrock 855fa9e4066Sahrens /* XXX does *this* func really need the lock? */ 856fa9e4066Sahrens ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)); 857fa9e4066Sahrens 858fa9e4066Sahrens /* 859fa9e4066Sahrens * This call to dbuf_will_dirty() with the dn_struct_rwlock held 860fa9e4066Sahrens * is OK, because there can be no other references to the db 861fa9e4066Sahrens * when we are changing its size, so no concurrent DB_FILL can 862fa9e4066Sahrens * be happening. 863fa9e4066Sahrens */ 864ea8dc4b6Seschrock /* 865ea8dc4b6Seschrock * XXX we should be doing a dbuf_read, checking the return 866ea8dc4b6Seschrock * value and returning that up to our callers 867ea8dc4b6Seschrock */ 868fa9e4066Sahrens dbuf_will_dirty(db, tx); 869fa9e4066Sahrens 870fa9e4066Sahrens /* create the data buffer for the new block */ 871ad23a2dbSjohansen buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type); 872fa9e4066Sahrens 873fa9e4066Sahrens /* copy old block data to the new block */ 874fa9e4066Sahrens obuf = db->db_buf; 875f65e61c0Sahrens bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 876fa9e4066Sahrens /* zero the remainder */ 877f65e61c0Sahrens if (size > osize) 878fa9e4066Sahrens bzero((uint8_t *)buf->b_data + osize, size - osize); 879fa9e4066Sahrens 880fa9e4066Sahrens mutex_enter(&db->db_mtx); 881fa9e4066Sahrens dbuf_set_data(db, buf); 882ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(obuf, db) == 1); 883fa9e4066Sahrens db->db.db_size = size; 884fa9e4066Sahrens 885c717a561Smaybee if (db->db_level == 0) { 886c717a561Smaybee ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 887c717a561Smaybee db->db_last_dirty->dt.dl.dr_data = buf; 888c717a561Smaybee } 889fa9e4066Sahrens mutex_exit(&db->db_mtx); 890fa9e4066Sahrens 891fa9e4066Sahrens dnode_willuse_space(db->db_dnode, size-osize, tx); 892fa9e4066Sahrens } 893fa9e4066Sahrens 894c717a561Smaybee dbuf_dirty_record_t * 895fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 896fa9e4066Sahrens { 897fa9e4066Sahrens dnode_t *dn = db->db_dnode; 898503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 899c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 900fa9e4066Sahrens int drop_struct_lock = FALSE; 901d3469faaSMark Maybee boolean_t do_free_accounting = B_FALSE; 902fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 903fa9e4066Sahrens 904fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 905fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 9069c9dc39aSek110237 DMU_TX_DIRTY_BUF(tx, db); 907fa9e4066Sahrens 908fa9e4066Sahrens /* 909fa9e4066Sahrens * Shouldn't dirty a regular buffer in syncing context. Private 910fa9e4066Sahrens * objects may be dirtied in syncing context, but only if they 911fa9e4066Sahrens * were already pre-dirtied in open context. 912fa9e4066Sahrens */ 913c717a561Smaybee ASSERT(!dmu_tx_is_syncing(tx) || 914c717a561Smaybee BP_IS_HOLE(dn->dn_objset->os_rootbp) || 91514843421SMatthew Ahrens DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 91614843421SMatthew Ahrens dn->dn_objset->os_dsl_dataset == NULL); 917fa9e4066Sahrens /* 918fa9e4066Sahrens * We make this assert for private objects as well, but after we 919fa9e4066Sahrens * check if we're already dirty. They are allowed to re-dirty 920fa9e4066Sahrens * in syncing context. 921fa9e4066Sahrens */ 922ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 923c717a561Smaybee dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 924fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 925fa9e4066Sahrens 926fa9e4066Sahrens mutex_enter(&db->db_mtx); 927fa9e4066Sahrens /* 928c717a561Smaybee * XXX make this true for indirects too? The problem is that 929c717a561Smaybee * transactions created with dmu_tx_create_assigned() from 930c717a561Smaybee * syncing context don't bother holding ahead. 931fa9e4066Sahrens */ 932c717a561Smaybee ASSERT(db->db_level != 0 || 93382c9918fSTim Haley db->db_state == DB_CACHED || db->db_state == DB_FILL || 93482c9918fSTim Haley db->db_state == DB_NOFILL); 935fa9e4066Sahrens 936fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 937fa9e4066Sahrens /* 938fa9e4066Sahrens * Don't set dirtyctx to SYNC if we're just modifying this as we 939fa9e4066Sahrens * initialize the objset. 940fa9e4066Sahrens */ 941fa9e4066Sahrens if (dn->dn_dirtyctx == DN_UNDIRTIED && 942c717a561Smaybee !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 943fa9e4066Sahrens dn->dn_dirtyctx = 944fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 945fa9e4066Sahrens ASSERT(dn->dn_dirtyctx_firstset == NULL); 946fa9e4066Sahrens dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 947fa9e4066Sahrens } 948fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 949fa9e4066Sahrens 950fa9e4066Sahrens /* 951fa9e4066Sahrens * If this buffer is already dirty, we're done. 952fa9e4066Sahrens */ 953c717a561Smaybee drp = &db->db_last_dirty; 954c717a561Smaybee ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 955c717a561Smaybee db->db.db_object == DMU_META_DNODE_OBJECT); 9567e2186e3Sbonwick while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 9577e2186e3Sbonwick drp = &dr->dr_next; 9587e2186e3Sbonwick if (dr && dr->dr_txg == tx->tx_txg) { 959c717a561Smaybee if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 960c717a561Smaybee /* 961c717a561Smaybee * If this buffer has already been written out, 962c717a561Smaybee * we now need to reset its state. 963c717a561Smaybee */ 9647e2186e3Sbonwick dbuf_unoverride(dr); 965*b24ab676SJeff Bonwick if (db->db.db_object != DMU_META_DNODE_OBJECT && 966*b24ab676SJeff Bonwick db->db_state != DB_NOFILL) 9676b4acc8bSahrens arc_buf_thaw(db->db_buf); 968c717a561Smaybee } 969fa9e4066Sahrens mutex_exit(&db->db_mtx); 9707e2186e3Sbonwick return (dr); 971fa9e4066Sahrens } 972fa9e4066Sahrens 973fa9e4066Sahrens /* 974fa9e4066Sahrens * Only valid if not already dirty. 975fa9e4066Sahrens */ 97614843421SMatthew Ahrens ASSERT(dn->dn_object == 0 || 97714843421SMatthew Ahrens dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 978fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 979fa9e4066Sahrens 980fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, db->db_level); 981fa9e4066Sahrens ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 982fa9e4066Sahrens dn->dn_phys->dn_nlevels > db->db_level || 983fa9e4066Sahrens dn->dn_next_nlevels[txgoff] > db->db_level || 984fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 985fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 986fa9e4066Sahrens 987fa9e4066Sahrens /* 988fa9e4066Sahrens * We should only be dirtying in syncing context if it's the 98914843421SMatthew Ahrens * mos or we're initializing the os or it's a special object. 99014843421SMatthew Ahrens * However, we are allowed to dirty in syncing context provided 99114843421SMatthew Ahrens * we already dirtied it in open context. Hence we must make 99214843421SMatthew Ahrens * this assertion only if we're not already dirty. 993fa9e4066Sahrens */ 99414843421SMatthew Ahrens ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 99514843421SMatthew Ahrens os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 996fa9e4066Sahrens ASSERT(db->db.db_size != 0); 997fa9e4066Sahrens 998fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 999fa9e4066Sahrens 10001934e92fSmaybee if (db->db_blkid != DB_BONUS_BLKID) { 10011934e92fSmaybee /* 10021934e92fSmaybee * Update the accounting. 1003d3469faaSMark Maybee * Note: we delay "free accounting" until after we drop 1004d3469faaSMark Maybee * the db_mtx. This keeps us from grabbing other locks 1005*b24ab676SJeff Bonwick * (and possibly deadlocking) in bp_get_dsize() while 1006d3469faaSMark Maybee * also holding the db_mtx. 10071934e92fSmaybee */ 10081934e92fSmaybee dnode_willuse_space(dn, db->db.db_size, tx); 1009d3469faaSMark Maybee do_free_accounting = dbuf_block_freeable(db); 10101934e92fSmaybee } 10111934e92fSmaybee 1012ea8dc4b6Seschrock /* 1013ea8dc4b6Seschrock * If this buffer is dirty in an old transaction group we need 1014ea8dc4b6Seschrock * to make a copy of it so that the changes we make in this 1015ea8dc4b6Seschrock * transaction group won't leak out when we sync the older txg. 1016ea8dc4b6Seschrock */ 1017c717a561Smaybee dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1018c717a561Smaybee if (db->db_level == 0) { 1019c717a561Smaybee void *data_old = db->db_buf; 1020c717a561Smaybee 102182c9918fSTim Haley if (db->db_state != DB_NOFILL) { 1022ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 1023c717a561Smaybee dbuf_fix_old_data(db, tx->tx_txg); 1024c717a561Smaybee data_old = db->db.db_data; 1025c717a561Smaybee } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1026fa9e4066Sahrens /* 102782c9918fSTim Haley * Release the data buffer from the cache so 102882c9918fSTim Haley * that we can modify it without impacting 102982c9918fSTim Haley * possible other users of this cached data 103082c9918fSTim Haley * block. Note that indirect blocks and 103182c9918fSTim Haley * private objects are not released until the 103282c9918fSTim Haley * syncing state (since they are only modified 103382c9918fSTim Haley * then). 1034fa9e4066Sahrens */ 1035fa9e4066Sahrens arc_release(db->db_buf, db); 1036fa9e4066Sahrens dbuf_fix_old_data(db, tx->tx_txg); 1037c717a561Smaybee data_old = db->db_buf; 1038fa9e4066Sahrens } 1039c717a561Smaybee ASSERT(data_old != NULL); 104082c9918fSTim Haley } 1041c717a561Smaybee dr->dt.dl.dr_data = data_old; 1042c717a561Smaybee } else { 1043c717a561Smaybee mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1044c717a561Smaybee list_create(&dr->dt.di.dr_children, 1045c717a561Smaybee sizeof (dbuf_dirty_record_t), 1046c717a561Smaybee offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1047fa9e4066Sahrens } 1048c717a561Smaybee dr->dr_dbuf = db; 1049c717a561Smaybee dr->dr_txg = tx->tx_txg; 1050c717a561Smaybee dr->dr_next = *drp; 1051c717a561Smaybee *drp = dr; 1052fa9e4066Sahrens 1053fa9e4066Sahrens /* 1054fa9e4066Sahrens * We could have been freed_in_flight between the dbuf_noread 1055fa9e4066Sahrens * and dbuf_dirty. We win, as though the dbuf_noread() had 1056fa9e4066Sahrens * happened after the free. 1057fa9e4066Sahrens */ 1058fa9e4066Sahrens if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 1059c717a561Smaybee mutex_enter(&dn->dn_mtx); 1060fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1061fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1062c717a561Smaybee db->db_freed_in_flight = FALSE; 1063c717a561Smaybee } 1064fa9e4066Sahrens 1065fa9e4066Sahrens /* 1066fa9e4066Sahrens * This buffer is now part of this txg 1067fa9e4066Sahrens */ 1068fa9e4066Sahrens dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1069fa9e4066Sahrens db->db_dirtycnt += 1; 1070fa9e4066Sahrens ASSERT3U(db->db_dirtycnt, <=, 3); 1071fa9e4066Sahrens 1072fa9e4066Sahrens mutex_exit(&db->db_mtx); 1073fa9e4066Sahrens 1074fa9e4066Sahrens if (db->db_blkid == DB_BONUS_BLKID) { 1075c717a561Smaybee mutex_enter(&dn->dn_mtx); 1076c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1077c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1078c717a561Smaybee mutex_exit(&dn->dn_mtx); 1079fa9e4066Sahrens dnode_setdirty(dn, tx); 1080c717a561Smaybee return (dr); 1081d3469faaSMark Maybee } else if (do_free_accounting) { 1082d3469faaSMark Maybee blkptr_t *bp = db->db_blkptr; 1083d3469faaSMark Maybee int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1084*b24ab676SJeff Bonwick bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1085d3469faaSMark Maybee /* 1086d3469faaSMark Maybee * This is only a guess -- if the dbuf is dirty 1087d3469faaSMark Maybee * in a previous txg, we don't know how much 1088d3469faaSMark Maybee * space it will use on disk yet. We should 1089d3469faaSMark Maybee * really have the struct_rwlock to access 1090d3469faaSMark Maybee * db_blkptr, but since this is just a guess, 1091d3469faaSMark Maybee * it's OK if we get an odd answer. 1092d3469faaSMark Maybee */ 1093d3469faaSMark Maybee dnode_willuse_space(dn, -willfree, tx); 1094fa9e4066Sahrens } 1095fa9e4066Sahrens 1096fa9e4066Sahrens if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1097fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 1098fa9e4066Sahrens drop_struct_lock = TRUE; 1099fa9e4066Sahrens } 1100fa9e4066Sahrens 11018346f03fSJonathan W Adams if (db->db_level == 0) { 11028346f03fSJonathan W Adams dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 11038346f03fSJonathan W Adams ASSERT(dn->dn_maxblkid >= db->db_blkid); 11048346f03fSJonathan W Adams } 11058346f03fSJonathan W Adams 110644eda4d7Smaybee if (db->db_level+1 < dn->dn_nlevels) { 1107c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 1108c717a561Smaybee dbuf_dirty_record_t *di; 1109c717a561Smaybee int parent_held = FALSE; 1110c717a561Smaybee 1111c717a561Smaybee if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1112fa9e4066Sahrens int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1113c717a561Smaybee 1114fa9e4066Sahrens parent = dbuf_hold_level(dn, db->db_level+1, 1115fa9e4066Sahrens db->db_blkid >> epbs, FTAG); 1116c717a561Smaybee parent_held = TRUE; 1117c717a561Smaybee } 1118fa9e4066Sahrens if (drop_struct_lock) 1119fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1120c717a561Smaybee ASSERT3U(db->db_level+1, ==, parent->db_level); 1121c717a561Smaybee di = dbuf_dirty(parent, tx); 1122c717a561Smaybee if (parent_held) 1123ea8dc4b6Seschrock dbuf_rele(parent, FTAG); 1124c717a561Smaybee 1125c717a561Smaybee mutex_enter(&db->db_mtx); 1126c717a561Smaybee /* possible race with dbuf_undirty() */ 1127c717a561Smaybee if (db->db_last_dirty == dr || 1128c717a561Smaybee dn->dn_object == DMU_META_DNODE_OBJECT) { 1129c717a561Smaybee mutex_enter(&di->dt.di.dr_mtx); 1130c717a561Smaybee ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1131c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1132c717a561Smaybee list_insert_tail(&di->dt.di.dr_children, dr); 1133c717a561Smaybee mutex_exit(&di->dt.di.dr_mtx); 1134c717a561Smaybee dr->dr_parent = di; 1135c717a561Smaybee } 1136c717a561Smaybee mutex_exit(&db->db_mtx); 1137fa9e4066Sahrens } else { 1138c717a561Smaybee ASSERT(db->db_level+1 == dn->dn_nlevels); 1139c717a561Smaybee ASSERT(db->db_blkid < dn->dn_nblkptr); 1140c717a561Smaybee ASSERT(db->db_parent == NULL || 1141c717a561Smaybee db->db_parent == db->db_dnode->dn_dbuf); 1142c717a561Smaybee mutex_enter(&dn->dn_mtx); 1143c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1144c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1145c717a561Smaybee mutex_exit(&dn->dn_mtx); 1146fa9e4066Sahrens if (drop_struct_lock) 1147fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1148fa9e4066Sahrens } 1149fa9e4066Sahrens 1150fa9e4066Sahrens dnode_setdirty(dn, tx); 1151c717a561Smaybee return (dr); 1152fa9e4066Sahrens } 1153fa9e4066Sahrens 1154fa9e4066Sahrens static int 1155fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1156fa9e4066Sahrens { 1157fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1158c717a561Smaybee uint64_t txg = tx->tx_txg; 115917f17c2dSbonwick dbuf_dirty_record_t *dr, **drp; 1160fa9e4066Sahrens 1161c717a561Smaybee ASSERT(txg != 0); 1162ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1163fa9e4066Sahrens 1164fa9e4066Sahrens mutex_enter(&db->db_mtx); 1165fa9e4066Sahrens 1166fa9e4066Sahrens /* 1167fa9e4066Sahrens * If this buffer is not dirty, we're done. 1168fa9e4066Sahrens */ 116917f17c2dSbonwick for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1170c717a561Smaybee if (dr->dr_txg <= txg) 1171c717a561Smaybee break; 1172c717a561Smaybee if (dr == NULL || dr->dr_txg < txg) { 1173fa9e4066Sahrens mutex_exit(&db->db_mtx); 1174fa9e4066Sahrens return (0); 1175fa9e4066Sahrens } 1176c717a561Smaybee ASSERT(dr->dr_txg == txg); 1177*b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 1178fa9e4066Sahrens 1179fa9e4066Sahrens /* 1180fa9e4066Sahrens * If this buffer is currently held, we cannot undirty 1181fa9e4066Sahrens * it, since one of the current holders may be in the 1182fa9e4066Sahrens * middle of an update. Note that users of dbuf_undirty() 1183fa9e4066Sahrens * should not place a hold on the dbuf before the call. 1184fa9e4066Sahrens */ 1185fa9e4066Sahrens if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1186fa9e4066Sahrens mutex_exit(&db->db_mtx); 118744eda4d7Smaybee /* Make sure we don't toss this buffer at sync phase */ 1188fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1189fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1190fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1191fa9e4066Sahrens return (0); 1192fa9e4066Sahrens } 1193fa9e4066Sahrens 1194fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1195fa9e4066Sahrens 1196fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1197fa9e4066Sahrens 1198fa9e4066Sahrens /* XXX would be nice to fix up dn_towrite_space[] */ 1199fa9e4066Sahrens 120017f17c2dSbonwick *drp = dr->dr_next; 1201c717a561Smaybee 1202c717a561Smaybee if (dr->dr_parent) { 1203c717a561Smaybee mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1204c717a561Smaybee list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1205c717a561Smaybee mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1206c717a561Smaybee } else if (db->db_level+1 == dn->dn_nlevels) { 1207cdb0ab79Smaybee ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1208fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1209c717a561Smaybee list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1210fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1211c717a561Smaybee } 1212c717a561Smaybee 1213c717a561Smaybee if (db->db_level == 0) { 121482c9918fSTim Haley if (db->db_state != DB_NOFILL) { 1215c717a561Smaybee dbuf_unoverride(dr); 1216c717a561Smaybee 1217c717a561Smaybee ASSERT(db->db_buf != NULL); 1218c717a561Smaybee ASSERT(dr->dt.dl.dr_data != NULL); 1219c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 122082c9918fSTim Haley VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 122182c9918fSTim Haley db) == 1); 122282c9918fSTim Haley } 1223c717a561Smaybee } else { 1224c717a561Smaybee ASSERT(db->db_buf != NULL); 1225c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1226c25056deSgw25295 mutex_destroy(&dr->dt.di.dr_mtx); 1227c25056deSgw25295 list_destroy(&dr->dt.di.dr_children); 1228c717a561Smaybee } 1229c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1230fa9e4066Sahrens 1231fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1232fa9e4066Sahrens db->db_dirtycnt -= 1; 1233fa9e4066Sahrens 1234c717a561Smaybee if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1235ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1236fa9e4066Sahrens 1237*b24ab676SJeff Bonwick ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1238ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1239ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1240fa9e4066Sahrens dbuf_evict(db); 1241fa9e4066Sahrens return (1); 1242fa9e4066Sahrens } 1243fa9e4066Sahrens 1244fa9e4066Sahrens mutex_exit(&db->db_mtx); 1245fa9e4066Sahrens return (0); 1246fa9e4066Sahrens } 1247fa9e4066Sahrens 1248fa9e4066Sahrens #pragma weak dmu_buf_will_dirty = dbuf_will_dirty 1249fa9e4066Sahrens void 1250fa9e4066Sahrens dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1251fa9e4066Sahrens { 12521ab7f2deSmaybee int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1253fa9e4066Sahrens 1254fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1255fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1256fa9e4066Sahrens 1257fa9e4066Sahrens if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) 1258fa9e4066Sahrens rf |= DB_RF_HAVESTRUCT; 1259ea8dc4b6Seschrock (void) dbuf_read(db, NULL, rf); 1260c717a561Smaybee (void) dbuf_dirty(db, tx); 1261fa9e4066Sahrens } 1262fa9e4066Sahrens 1263fa9e4066Sahrens void 126482c9918fSTim Haley dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 126582c9918fSTim Haley { 126682c9918fSTim Haley dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 126782c9918fSTim Haley 126882c9918fSTim Haley db->db_state = DB_NOFILL; 126982c9918fSTim Haley 127082c9918fSTim Haley dmu_buf_will_fill(db_fake, tx); 127182c9918fSTim Haley } 127282c9918fSTim Haley 127382c9918fSTim Haley void 1274ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1275fa9e4066Sahrens { 1276ea8dc4b6Seschrock dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1277ea8dc4b6Seschrock 1278ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1279fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1280fa9e4066Sahrens ASSERT(db->db_level == 0); 1281fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1282fa9e4066Sahrens 1283ea8dc4b6Seschrock ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1284fa9e4066Sahrens dmu_tx_private_ok(tx)); 1285fa9e4066Sahrens 1286fa9e4066Sahrens dbuf_noread(db); 1287c717a561Smaybee (void) dbuf_dirty(db, tx); 1288fa9e4066Sahrens } 1289fa9e4066Sahrens 1290fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done 1291fa9e4066Sahrens /* ARGSUSED */ 1292fa9e4066Sahrens void 1293fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1294fa9e4066Sahrens { 1295fa9e4066Sahrens mutex_enter(&db->db_mtx); 12969c9dc39aSek110237 DBUF_VERIFY(db); 1297fa9e4066Sahrens 1298fa9e4066Sahrens if (db->db_state == DB_FILL) { 1299c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 1300ea8dc4b6Seschrock ASSERT(db->db_blkid != DB_BONUS_BLKID); 1301fa9e4066Sahrens /* we were freed while filling */ 1302fa9e4066Sahrens /* XXX dbuf_undirty? */ 1303fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 1304c717a561Smaybee db->db_freed_in_flight = FALSE; 1305fa9e4066Sahrens } 1306fa9e4066Sahrens db->db_state = DB_CACHED; 1307fa9e4066Sahrens cv_broadcast(&db->db_changed); 1308fa9e4066Sahrens } 1309fa9e4066Sahrens mutex_exit(&db->db_mtx); 1310fa9e4066Sahrens } 1311fa9e4066Sahrens 1312ea8dc4b6Seschrock /* 13132fdbea25SAleksandr Guzovskiy * Directly assign a provided arc buf to a given dbuf if it's not referenced 13142fdbea25SAleksandr Guzovskiy * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 13152fdbea25SAleksandr Guzovskiy */ 13162fdbea25SAleksandr Guzovskiy void 13172fdbea25SAleksandr Guzovskiy dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 13182fdbea25SAleksandr Guzovskiy { 13192fdbea25SAleksandr Guzovskiy ASSERT(!refcount_is_zero(&db->db_holds)); 13202fdbea25SAleksandr Guzovskiy ASSERT(db->db_dnode->dn_object != DMU_META_DNODE_OBJECT); 13212fdbea25SAleksandr Guzovskiy ASSERT(db->db_blkid != DB_BONUS_BLKID); 13222fdbea25SAleksandr Guzovskiy ASSERT(db->db_level == 0); 13232fdbea25SAleksandr Guzovskiy ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 13242fdbea25SAleksandr Guzovskiy ASSERT(buf != NULL); 13252fdbea25SAleksandr Guzovskiy ASSERT(arc_buf_size(buf) == db->db.db_size); 13262fdbea25SAleksandr Guzovskiy ASSERT(tx->tx_txg != 0); 13272fdbea25SAleksandr Guzovskiy 13282fdbea25SAleksandr Guzovskiy arc_return_buf(buf, db); 13292fdbea25SAleksandr Guzovskiy ASSERT(arc_released(buf)); 13302fdbea25SAleksandr Guzovskiy 13312fdbea25SAleksandr Guzovskiy mutex_enter(&db->db_mtx); 13322fdbea25SAleksandr Guzovskiy 13332fdbea25SAleksandr Guzovskiy while (db->db_state == DB_READ || db->db_state == DB_FILL) 13342fdbea25SAleksandr Guzovskiy cv_wait(&db->db_changed, &db->db_mtx); 13352fdbea25SAleksandr Guzovskiy 13362fdbea25SAleksandr Guzovskiy ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 13372fdbea25SAleksandr Guzovskiy 13382fdbea25SAleksandr Guzovskiy if (db->db_state == DB_CACHED && 13392fdbea25SAleksandr Guzovskiy refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 13402fdbea25SAleksandr Guzovskiy mutex_exit(&db->db_mtx); 13412fdbea25SAleksandr Guzovskiy (void) dbuf_dirty(db, tx); 13422fdbea25SAleksandr Guzovskiy bcopy(buf->b_data, db->db.db_data, db->db.db_size); 13432fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(buf, db) == 1); 13442fdbea25SAleksandr Guzovskiy return; 13452fdbea25SAleksandr Guzovskiy } 13462fdbea25SAleksandr Guzovskiy 13472fdbea25SAleksandr Guzovskiy if (db->db_state == DB_CACHED) { 13482fdbea25SAleksandr Guzovskiy dbuf_dirty_record_t *dr = db->db_last_dirty; 13492fdbea25SAleksandr Guzovskiy 13502fdbea25SAleksandr Guzovskiy ASSERT(db->db_buf != NULL); 13512fdbea25SAleksandr Guzovskiy if (dr != NULL && dr->dr_txg == tx->tx_txg) { 13522fdbea25SAleksandr Guzovskiy ASSERT(dr->dt.dl.dr_data == db->db_buf); 13532fdbea25SAleksandr Guzovskiy if (!arc_released(db->db_buf)) { 13542fdbea25SAleksandr Guzovskiy ASSERT(dr->dt.dl.dr_override_state == 13552fdbea25SAleksandr Guzovskiy DR_OVERRIDDEN); 13562fdbea25SAleksandr Guzovskiy arc_release(db->db_buf, db); 13572fdbea25SAleksandr Guzovskiy } 13582fdbea25SAleksandr Guzovskiy dr->dt.dl.dr_data = buf; 13592fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); 13602fdbea25SAleksandr Guzovskiy } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 13612fdbea25SAleksandr Guzovskiy arc_release(db->db_buf, db); 13622fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); 13632fdbea25SAleksandr Guzovskiy } 13642fdbea25SAleksandr Guzovskiy db->db_buf = NULL; 13652fdbea25SAleksandr Guzovskiy } 13662fdbea25SAleksandr Guzovskiy ASSERT(db->db_buf == NULL); 13672fdbea25SAleksandr Guzovskiy dbuf_set_data(db, buf); 13682fdbea25SAleksandr Guzovskiy db->db_state = DB_FILL; 13692fdbea25SAleksandr Guzovskiy mutex_exit(&db->db_mtx); 13702fdbea25SAleksandr Guzovskiy (void) dbuf_dirty(db, tx); 13712fdbea25SAleksandr Guzovskiy dbuf_fill_done(db, tx); 13722fdbea25SAleksandr Guzovskiy } 13732fdbea25SAleksandr Guzovskiy 13742fdbea25SAleksandr Guzovskiy /* 1375ea8dc4b6Seschrock * "Clear" the contents of this dbuf. This will mark the dbuf 1376ea8dc4b6Seschrock * EVICTING and clear *most* of its references. Unfortunetely, 1377ea8dc4b6Seschrock * when we are not holding the dn_dbufs_mtx, we can't clear the 1378ea8dc4b6Seschrock * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1379ea8dc4b6Seschrock * in this case. For callers from the DMU we will usually see: 1380ea8dc4b6Seschrock * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1381ea8dc4b6Seschrock * For the arc callback, we will usually see: 1382ea8dc4b6Seschrock * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1383ea8dc4b6Seschrock * Sometimes, though, we will get a mix of these two: 1384ea8dc4b6Seschrock * DMU: dbuf_clear()->arc_buf_evict() 1385ea8dc4b6Seschrock * ARC: dbuf_do_evict()->dbuf_destroy() 1386ea8dc4b6Seschrock */ 1387ea8dc4b6Seschrock void 1388fa9e4066Sahrens dbuf_clear(dmu_buf_impl_t *db) 1389fa9e4066Sahrens { 1390fa9e4066Sahrens dnode_t *dn = db->db_dnode; 1391ea8dc4b6Seschrock dmu_buf_impl_t *parent = db->db_parent; 1392c543ec06Sahrens dmu_buf_impl_t *dndb = dn->dn_dbuf; 1393ea8dc4b6Seschrock int dbuf_gone = FALSE; 1394fa9e4066Sahrens 1395fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 1396fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1397fa9e4066Sahrens 1398ea8dc4b6Seschrock dbuf_evict_user(db); 1399ea8dc4b6Seschrock 1400fa9e4066Sahrens if (db->db_state == DB_CACHED) { 1401ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 14020e8c6158Smaybee if (db->db_blkid == DB_BONUS_BLKID) { 1403ea8dc4b6Seschrock zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 14045a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 14050e8c6158Smaybee } 1406fa9e4066Sahrens db->db.db_data = NULL; 1407fa9e4066Sahrens db->db_state = DB_UNCACHED; 1408fa9e4066Sahrens } 1409fa9e4066Sahrens 141082c9918fSTim Haley ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1411fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1412fa9e4066Sahrens 1413ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1414ea8dc4b6Seschrock db->db_blkptr = NULL; 1415ea8dc4b6Seschrock 1416ea8dc4b6Seschrock if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1417ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1418ea8dc4b6Seschrock dnode_rele(dn, db); 14191934e92fSmaybee db->db_dnode = NULL; 1420ea8dc4b6Seschrock } 1421ea8dc4b6Seschrock 1422ea8dc4b6Seschrock if (db->db_buf) 1423ea8dc4b6Seschrock dbuf_gone = arc_buf_evict(db->db_buf); 1424ea8dc4b6Seschrock 1425ea8dc4b6Seschrock if (!dbuf_gone) 1426fa9e4066Sahrens mutex_exit(&db->db_mtx); 1427fa9e4066Sahrens 1428fa9e4066Sahrens /* 1429fa9e4066Sahrens * If this dbuf is referened from an indirect dbuf, 1430fa9e4066Sahrens * decrement the ref count on the indirect dbuf. 1431fa9e4066Sahrens */ 1432c543ec06Sahrens if (parent && parent != dndb) 1433ea8dc4b6Seschrock dbuf_rele(parent, db); 1434fa9e4066Sahrens } 1435fa9e4066Sahrens 1436fa9e4066Sahrens static int 1437fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1438fa9e4066Sahrens dmu_buf_impl_t **parentp, blkptr_t **bpp) 1439fa9e4066Sahrens { 1440fa9e4066Sahrens int nlevels, epbs; 1441fa9e4066Sahrens 14420b69c2f0Sahrens *parentp = NULL; 14430b69c2f0Sahrens *bpp = NULL; 14440b69c2f0Sahrens 1445ea8dc4b6Seschrock ASSERT(blkid != DB_BONUS_BLKID); 1446ea8dc4b6Seschrock 1447fa9e4066Sahrens if (dn->dn_phys->dn_nlevels == 0) 1448fa9e4066Sahrens nlevels = 1; 1449fa9e4066Sahrens else 1450fa9e4066Sahrens nlevels = dn->dn_phys->dn_nlevels; 1451fa9e4066Sahrens 1452fa9e4066Sahrens epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1453fa9e4066Sahrens 1454fa9e4066Sahrens ASSERT3U(level * epbs, <, 64); 1455fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1456ea8dc4b6Seschrock if (level >= nlevels || 1457fa9e4066Sahrens (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1458fa9e4066Sahrens /* the buffer has no parent yet */ 1459fa9e4066Sahrens return (ENOENT); 1460fa9e4066Sahrens } else if (level < nlevels-1) { 1461fa9e4066Sahrens /* this block is referenced from an indirect block */ 1462fa9e4066Sahrens int err = dbuf_hold_impl(dn, level+1, 1463fa9e4066Sahrens blkid >> epbs, fail_sparse, NULL, parentp); 1464fa9e4066Sahrens if (err) 1465fa9e4066Sahrens return (err); 1466ea8dc4b6Seschrock err = dbuf_read(*parentp, NULL, 1467ea8dc4b6Seschrock (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1468c543ec06Sahrens if (err) { 1469c543ec06Sahrens dbuf_rele(*parentp, NULL); 1470c543ec06Sahrens *parentp = NULL; 1471c543ec06Sahrens return (err); 1472c543ec06Sahrens } 1473fa9e4066Sahrens *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1474fa9e4066Sahrens (blkid & ((1ULL << epbs) - 1)); 1475c543ec06Sahrens return (0); 1476fa9e4066Sahrens } else { 1477fa9e4066Sahrens /* the block is referenced from the dnode */ 1478fa9e4066Sahrens ASSERT3U(level, ==, nlevels-1); 1479fa9e4066Sahrens ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1480fa9e4066Sahrens blkid < dn->dn_phys->dn_nblkptr); 1481c543ec06Sahrens if (dn->dn_dbuf) { 1482c543ec06Sahrens dbuf_add_ref(dn->dn_dbuf, NULL); 1483fa9e4066Sahrens *parentp = dn->dn_dbuf; 1484c543ec06Sahrens } 1485fa9e4066Sahrens *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1486fa9e4066Sahrens return (0); 1487fa9e4066Sahrens } 1488fa9e4066Sahrens } 1489fa9e4066Sahrens 1490fa9e4066Sahrens static dmu_buf_impl_t * 1491fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1492fa9e4066Sahrens dmu_buf_impl_t *parent, blkptr_t *blkptr) 1493fa9e4066Sahrens { 1494503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 1495fa9e4066Sahrens dmu_buf_impl_t *db, *odb; 1496fa9e4066Sahrens 1497fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1498fa9e4066Sahrens ASSERT(dn->dn_type != DMU_OT_NONE); 1499fa9e4066Sahrens 1500fa9e4066Sahrens db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1501fa9e4066Sahrens 1502fa9e4066Sahrens db->db_objset = os; 1503fa9e4066Sahrens db->db.db_object = dn->dn_object; 1504fa9e4066Sahrens db->db_level = level; 1505fa9e4066Sahrens db->db_blkid = blkid; 1506c717a561Smaybee db->db_last_dirty = NULL; 1507ea8dc4b6Seschrock db->db_dirtycnt = 0; 1508ea8dc4b6Seschrock db->db_dnode = dn; 1509ea8dc4b6Seschrock db->db_parent = parent; 1510ea8dc4b6Seschrock db->db_blkptr = blkptr; 1511fa9e4066Sahrens 1512c717a561Smaybee db->db_user_ptr = NULL; 1513c717a561Smaybee db->db_user_data_ptr_ptr = NULL; 1514c717a561Smaybee db->db_evict_func = NULL; 1515c717a561Smaybee db->db_immediate_evict = 0; 1516c717a561Smaybee db->db_freed_in_flight = 0; 1517ea8dc4b6Seschrock 1518ea8dc4b6Seschrock if (blkid == DB_BONUS_BLKID) { 1519ea8dc4b6Seschrock ASSERT3P(parent, ==, dn->dn_dbuf); 15201934e92fSmaybee db->db.db_size = DN_MAX_BONUSLEN - 15211934e92fSmaybee (dn->dn_nblkptr-1) * sizeof (blkptr_t); 15221934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1523fa9e4066Sahrens db->db.db_offset = DB_BONUS_BLKID; 1524ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1525ea8dc4b6Seschrock /* the bonus dbuf is not placed in the hash table */ 15265a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1527ea8dc4b6Seschrock return (db); 1528fa9e4066Sahrens } else { 1529fa9e4066Sahrens int blocksize = 1530fa9e4066Sahrens db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; 1531fa9e4066Sahrens db->db.db_size = blocksize; 1532fa9e4066Sahrens db->db.db_offset = db->db_blkid * blocksize; 1533fa9e4066Sahrens } 1534fa9e4066Sahrens 1535fa9e4066Sahrens /* 1536fa9e4066Sahrens * Hold the dn_dbufs_mtx while we get the new dbuf 1537fa9e4066Sahrens * in the hash table *and* added to the dbufs list. 1538fa9e4066Sahrens * This prevents a possible deadlock with someone 1539fa9e4066Sahrens * trying to look up this dbuf before its added to the 1540fa9e4066Sahrens * dn_dbufs list. 1541fa9e4066Sahrens */ 1542fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 1543ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1544fa9e4066Sahrens if ((odb = dbuf_hash_insert(db)) != NULL) { 1545fa9e4066Sahrens /* someone else inserted it first */ 1546fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 1547fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1548fa9e4066Sahrens return (odb); 1549fa9e4066Sahrens } 1550fa9e4066Sahrens list_insert_head(&dn->dn_dbufs, db); 1551ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1552fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 15535a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1554fa9e4066Sahrens 1555fa9e4066Sahrens if (parent && parent != dn->dn_dbuf) 1556fa9e4066Sahrens dbuf_add_ref(parent, db); 1557fa9e4066Sahrens 1558ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1559ea8dc4b6Seschrock refcount_count(&dn->dn_holds) > 0); 1560fa9e4066Sahrens (void) refcount_add(&dn->dn_holds, db); 1561fa9e4066Sahrens 1562fa9e4066Sahrens dprintf_dbuf(db, "db=%p\n", db); 1563fa9e4066Sahrens 1564fa9e4066Sahrens return (db); 1565fa9e4066Sahrens } 1566fa9e4066Sahrens 1567fa9e4066Sahrens static int 1568ea8dc4b6Seschrock dbuf_do_evict(void *private) 1569fa9e4066Sahrens { 1570ea8dc4b6Seschrock arc_buf_t *buf = private; 1571ea8dc4b6Seschrock dmu_buf_impl_t *db = buf->b_private; 1572fa9e4066Sahrens 1573ea8dc4b6Seschrock if (!MUTEX_HELD(&db->db_mtx)) 1574ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 1575ea8dc4b6Seschrock 1576ea8dc4b6Seschrock ASSERT(refcount_is_zero(&db->db_holds)); 1577ea8dc4b6Seschrock 1578ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 1579ea8dc4b6Seschrock ASSERT(db->db_state == DB_CACHED); 15809c9dc39aSek110237 DBUF_VERIFY(db); 1581ea8dc4b6Seschrock db->db_buf = NULL; 1582ea8dc4b6Seschrock dbuf_evict(db); 1583ea8dc4b6Seschrock } else { 1584ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1585ea8dc4b6Seschrock dbuf_destroy(db); 1586fa9e4066Sahrens } 1587ea8dc4b6Seschrock return (0); 1588fa9e4066Sahrens } 1589fa9e4066Sahrens 1590fa9e4066Sahrens static void 1591fa9e4066Sahrens dbuf_destroy(dmu_buf_impl_t *db) 1592fa9e4066Sahrens { 1593fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1594fa9e4066Sahrens 1595ea8dc4b6Seschrock if (db->db_blkid != DB_BONUS_BLKID) { 1596ea8dc4b6Seschrock /* 1597ea8dc4b6Seschrock * If this dbuf is still on the dn_dbufs list, 1598ea8dc4b6Seschrock * remove it from that list. 1599ea8dc4b6Seschrock */ 16001934e92fSmaybee if (db->db_dnode) { 16011934e92fSmaybee dnode_t *dn = db->db_dnode; 16021934e92fSmaybee 16031934e92fSmaybee mutex_enter(&dn->dn_dbufs_mtx); 1604ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1605ea8dc4b6Seschrock mutex_exit(&dn->dn_dbufs_mtx); 1606ea8dc4b6Seschrock 1607ea8dc4b6Seschrock dnode_rele(dn, db); 16081934e92fSmaybee db->db_dnode = NULL; 1609ea8dc4b6Seschrock } 1610ea8dc4b6Seschrock dbuf_hash_remove(db); 1611ea8dc4b6Seschrock } 1612ea8dc4b6Seschrock db->db_parent = NULL; 1613ea8dc4b6Seschrock db->db_buf = NULL; 1614ea8dc4b6Seschrock 1615f82bfe17Sgw25295 ASSERT(!list_link_active(&db->db_link)); 1616fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1617fa9e4066Sahrens ASSERT(db->db_hash_next == NULL); 1618fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 1619fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1620fa9e4066Sahrens 1621fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 16225a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1623fa9e4066Sahrens } 1624fa9e4066Sahrens 1625fa9e4066Sahrens void 1626fa9e4066Sahrens dbuf_prefetch(dnode_t *dn, uint64_t blkid) 1627fa9e4066Sahrens { 162813506d1eSmaybee dmu_buf_impl_t *db = NULL; 1629fa9e4066Sahrens blkptr_t *bp = NULL; 1630fa9e4066Sahrens 1631fa9e4066Sahrens ASSERT(blkid != DB_BONUS_BLKID); 1632fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1633fa9e4066Sahrens 1634fa9e4066Sahrens if (dnode_block_freed(dn, blkid)) 1635fa9e4066Sahrens return; 1636fa9e4066Sahrens 1637fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1638fa9e4066Sahrens if (db = dbuf_find(dn, 0, blkid)) { 163913506d1eSmaybee if (refcount_count(&db->db_holds) > 0) { 1640fa9e4066Sahrens /* 164113506d1eSmaybee * This dbuf is active. We assume that it is 164213506d1eSmaybee * already CACHED, or else about to be either 1643fa9e4066Sahrens * read or filled. 1644fa9e4066Sahrens */ 1645fa9e4066Sahrens mutex_exit(&db->db_mtx); 1646fa9e4066Sahrens return; 1647fa9e4066Sahrens } 164813506d1eSmaybee mutex_exit(&db->db_mtx); 16490b69c2f0Sahrens db = NULL; 165013506d1eSmaybee } 1651fa9e4066Sahrens 165213506d1eSmaybee if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1653fa9e4066Sahrens if (bp && !BP_IS_HOLE(bp)) { 1654088f3894Sahrens arc_buf_t *pbuf; 1655*b24ab676SJeff Bonwick dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 165613506d1eSmaybee uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; 1657ea8dc4b6Seschrock zbookmark_t zb; 1658*b24ab676SJeff Bonwick 1659*b24ab676SJeff Bonwick SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 1660*b24ab676SJeff Bonwick dn->dn_object, 0, blkid); 1661ea8dc4b6Seschrock 1662088f3894Sahrens if (db) 1663088f3894Sahrens pbuf = db->db_buf; 1664088f3894Sahrens else 1665088f3894Sahrens pbuf = dn->dn_objset->os_phys_buf; 1666088f3894Sahrens 1667088f3894Sahrens (void) arc_read(NULL, dn->dn_objset->os_spa, 1668088f3894Sahrens bp, pbuf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 1669fa9e4066Sahrens ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 167013506d1eSmaybee &aflags, &zb); 1671fa9e4066Sahrens } 167213506d1eSmaybee if (db) 167313506d1eSmaybee dbuf_rele(db, NULL); 1674fa9e4066Sahrens } 1675fa9e4066Sahrens } 1676fa9e4066Sahrens 1677fa9e4066Sahrens /* 1678fa9e4066Sahrens * Returns with db_holds incremented, and db_mtx not held. 1679fa9e4066Sahrens * Note: dn_struct_rwlock must be held. 1680fa9e4066Sahrens */ 1681fa9e4066Sahrens int 1682fa9e4066Sahrens dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1683fa9e4066Sahrens void *tag, dmu_buf_impl_t **dbp) 1684fa9e4066Sahrens { 1685fa9e4066Sahrens dmu_buf_impl_t *db, *parent = NULL; 1686fa9e4066Sahrens 1687ea8dc4b6Seschrock ASSERT(blkid != DB_BONUS_BLKID); 1688fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1689fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, level); 1690fa9e4066Sahrens 1691fa9e4066Sahrens *dbp = NULL; 1692ea8dc4b6Seschrock top: 1693fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1694fa9e4066Sahrens db = dbuf_find(dn, level, blkid); 1695fa9e4066Sahrens 1696fa9e4066Sahrens if (db == NULL) { 1697fa9e4066Sahrens blkptr_t *bp = NULL; 1698fa9e4066Sahrens int err; 1699fa9e4066Sahrens 1700c543ec06Sahrens ASSERT3P(parent, ==, NULL); 1701fa9e4066Sahrens err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1702fa9e4066Sahrens if (fail_sparse) { 1703fa9e4066Sahrens if (err == 0 && bp && BP_IS_HOLE(bp)) 1704fa9e4066Sahrens err = ENOENT; 1705fa9e4066Sahrens if (err) { 1706c543ec06Sahrens if (parent) 1707ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1708fa9e4066Sahrens return (err); 1709fa9e4066Sahrens } 1710fa9e4066Sahrens } 1711ea8dc4b6Seschrock if (err && err != ENOENT) 1712ea8dc4b6Seschrock return (err); 1713fa9e4066Sahrens db = dbuf_create(dn, level, blkid, parent, bp); 1714fa9e4066Sahrens } 1715fa9e4066Sahrens 1716ea8dc4b6Seschrock if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1717ea8dc4b6Seschrock arc_buf_add_ref(db->db_buf, db); 1718ea8dc4b6Seschrock if (db->db_buf->b_data == NULL) { 1719ea8dc4b6Seschrock dbuf_clear(db); 1720c543ec06Sahrens if (parent) { 1721c543ec06Sahrens dbuf_rele(parent, NULL); 1722c543ec06Sahrens parent = NULL; 1723c543ec06Sahrens } 1724ea8dc4b6Seschrock goto top; 1725ea8dc4b6Seschrock } 1726ea8dc4b6Seschrock ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1727ea8dc4b6Seschrock } 1728ea8dc4b6Seschrock 1729ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1730ea8dc4b6Seschrock 1731fa9e4066Sahrens /* 1732c717a561Smaybee * If this buffer is currently syncing out, and we are are 1733c717a561Smaybee * still referencing it from db_data, we need to make a copy 1734c717a561Smaybee * of it in case we decide we want to dirty it again in this txg. 1735fa9e4066Sahrens */ 1736c717a561Smaybee if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 1737ea8dc4b6Seschrock dn->dn_object != DMU_META_DNODE_OBJECT && 1738c717a561Smaybee db->db_state == DB_CACHED && db->db_data_pending) { 1739c717a561Smaybee dbuf_dirty_record_t *dr = db->db_data_pending; 1740c717a561Smaybee 1741c717a561Smaybee if (dr->dt.dl.dr_data == db->db_buf) { 1742ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1743fa9e4066Sahrens 1744c717a561Smaybee dbuf_set_data(db, 1745c717a561Smaybee arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 1746c717a561Smaybee db->db.db_size, db, type)); 1747c717a561Smaybee bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1748fa9e4066Sahrens db->db.db_size); 1749fa9e4066Sahrens } 1750c717a561Smaybee } 1751fa9e4066Sahrens 1752ea8dc4b6Seschrock (void) refcount_add(&db->db_holds, tag); 1753fa9e4066Sahrens dbuf_update_data(db); 17549c9dc39aSek110237 DBUF_VERIFY(db); 1755fa9e4066Sahrens mutex_exit(&db->db_mtx); 1756fa9e4066Sahrens 1757fa9e4066Sahrens /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1758c543ec06Sahrens if (parent) 1759ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1760fa9e4066Sahrens 1761fa9e4066Sahrens ASSERT3P(db->db_dnode, ==, dn); 1762fa9e4066Sahrens ASSERT3U(db->db_blkid, ==, blkid); 1763fa9e4066Sahrens ASSERT3U(db->db_level, ==, level); 1764fa9e4066Sahrens *dbp = db; 1765fa9e4066Sahrens 1766fa9e4066Sahrens return (0); 1767fa9e4066Sahrens } 1768fa9e4066Sahrens 1769fa9e4066Sahrens dmu_buf_impl_t * 1770ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1771fa9e4066Sahrens { 1772fa9e4066Sahrens dmu_buf_impl_t *db; 1773ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1774ea8dc4b6Seschrock return (err ? NULL : db); 1775fa9e4066Sahrens } 1776fa9e4066Sahrens 1777fa9e4066Sahrens dmu_buf_impl_t * 1778fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1779fa9e4066Sahrens { 1780fa9e4066Sahrens dmu_buf_impl_t *db; 1781ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1782ea8dc4b6Seschrock return (err ? NULL : db); 1783fa9e4066Sahrens } 1784fa9e4066Sahrens 17851934e92fSmaybee void 1786ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn) 1787fa9e4066Sahrens { 1788ea8dc4b6Seschrock ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1789ea8dc4b6Seschrock 1790ea8dc4b6Seschrock ASSERT(dn->dn_bonus == NULL); 17911934e92fSmaybee dn->dn_bonus = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL); 1792fa9e4066Sahrens } 1793fa9e4066Sahrens 1794ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref 1795fa9e4066Sahrens void 1796fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 1797fa9e4066Sahrens { 1798ea8dc4b6Seschrock int64_t holds = refcount_add(&db->db_holds, tag); 1799ea8dc4b6Seschrock ASSERT(holds > 1); 1800fa9e4066Sahrens } 1801fa9e4066Sahrens 1802ea8dc4b6Seschrock #pragma weak dmu_buf_rele = dbuf_rele 1803fa9e4066Sahrens void 1804ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag) 1805fa9e4066Sahrens { 1806*b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 1807*b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, tag); 1808*b24ab676SJeff Bonwick } 1809*b24ab676SJeff Bonwick 1810*b24ab676SJeff Bonwick /* 1811*b24ab676SJeff Bonwick * dbuf_rele() for an already-locked dbuf. This is necessary to allow 1812*b24ab676SJeff Bonwick * db_dirtycnt and db_holds to be updated atomically. 1813*b24ab676SJeff Bonwick */ 1814*b24ab676SJeff Bonwick void 1815*b24ab676SJeff Bonwick dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 1816*b24ab676SJeff Bonwick { 1817fa9e4066Sahrens int64_t holds; 1818fa9e4066Sahrens 1819*b24ab676SJeff Bonwick ASSERT(MUTEX_HELD(&db->db_mtx)); 18209c9dc39aSek110237 DBUF_VERIFY(db); 1821fa9e4066Sahrens 1822fa9e4066Sahrens holds = refcount_remove(&db->db_holds, tag); 1823ea8dc4b6Seschrock ASSERT(holds >= 0); 1824fa9e4066Sahrens 1825c717a561Smaybee /* 1826c717a561Smaybee * We can't freeze indirects if there is a possibility that they 1827c717a561Smaybee * may be modified in the current syncing context. 1828c717a561Smaybee */ 1829c717a561Smaybee if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 18306b4acc8bSahrens arc_buf_freeze(db->db_buf); 18316b4acc8bSahrens 1832fa9e4066Sahrens if (holds == db->db_dirtycnt && 1833c717a561Smaybee db->db_level == 0 && db->db_immediate_evict) 1834fa9e4066Sahrens dbuf_evict_user(db); 1835ea8dc4b6Seschrock 1836ea8dc4b6Seschrock if (holds == 0) { 1837ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 1838ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1839ea8dc4b6Seschrock dnode_rele(db->db_dnode, db); 1840ea8dc4b6Seschrock } else if (db->db_buf == NULL) { 1841ea8dc4b6Seschrock /* 1842ea8dc4b6Seschrock * This is a special case: we never associated this 1843ea8dc4b6Seschrock * dbuf with any data allocated from the ARC. 1844ea8dc4b6Seschrock */ 184582c9918fSTim Haley ASSERT(db->db_state == DB_UNCACHED || 184682c9918fSTim Haley db->db_state == DB_NOFILL); 1847ea8dc4b6Seschrock dbuf_evict(db); 1848ea8dc4b6Seschrock } else if (arc_released(db->db_buf)) { 1849ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1850ea8dc4b6Seschrock /* 1851ea8dc4b6Seschrock * This dbuf has anonymous data associated with it. 1852ea8dc4b6Seschrock */ 1853ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1854ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1855ea8dc4b6Seschrock dbuf_evict(db); 1856ea8dc4b6Seschrock } else { 1857ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); 18583baa08fcSek110237 if (!DBUF_IS_CACHEABLE(db)) 18593baa08fcSek110237 dbuf_clear(db); 18603baa08fcSek110237 else 1861fa9e4066Sahrens mutex_exit(&db->db_mtx); 1862fa9e4066Sahrens } 1863ea8dc4b6Seschrock } else { 1864ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1865fa9e4066Sahrens } 1866fa9e4066Sahrens } 1867fa9e4066Sahrens 1868fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount 1869fa9e4066Sahrens uint64_t 1870fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db) 1871fa9e4066Sahrens { 1872fa9e4066Sahrens return (refcount_count(&db->db_holds)); 1873fa9e4066Sahrens } 1874fa9e4066Sahrens 1875fa9e4066Sahrens void * 1876fa9e4066Sahrens dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1877fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 1878fa9e4066Sahrens { 1879fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1880fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 1881fa9e4066Sahrens } 1882fa9e4066Sahrens 1883fa9e4066Sahrens void * 1884fa9e4066Sahrens dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1885fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 1886fa9e4066Sahrens { 1887fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1888fa9e4066Sahrens 1889c717a561Smaybee db->db_immediate_evict = TRUE; 1890fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1891fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 1892fa9e4066Sahrens } 1893fa9e4066Sahrens 1894fa9e4066Sahrens void * 1895fa9e4066Sahrens dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 1896fa9e4066Sahrens void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 1897fa9e4066Sahrens { 1898fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1899fa9e4066Sahrens ASSERT(db->db_level == 0); 1900fa9e4066Sahrens 1901fa9e4066Sahrens ASSERT((user_ptr == NULL) == (evict_func == NULL)); 1902fa9e4066Sahrens 1903fa9e4066Sahrens mutex_enter(&db->db_mtx); 1904fa9e4066Sahrens 1905c717a561Smaybee if (db->db_user_ptr == old_user_ptr) { 1906c717a561Smaybee db->db_user_ptr = user_ptr; 1907c717a561Smaybee db->db_user_data_ptr_ptr = user_data_ptr_ptr; 1908c717a561Smaybee db->db_evict_func = evict_func; 1909fa9e4066Sahrens 1910fa9e4066Sahrens dbuf_update_data(db); 1911fa9e4066Sahrens } else { 1912c717a561Smaybee old_user_ptr = db->db_user_ptr; 1913fa9e4066Sahrens } 1914fa9e4066Sahrens 1915fa9e4066Sahrens mutex_exit(&db->db_mtx); 1916fa9e4066Sahrens return (old_user_ptr); 1917fa9e4066Sahrens } 1918fa9e4066Sahrens 1919fa9e4066Sahrens void * 1920fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake) 1921fa9e4066Sahrens { 1922fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1923fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1924fa9e4066Sahrens 1925c717a561Smaybee return (db->db_user_ptr); 1926fa9e4066Sahrens } 1927fa9e4066Sahrens 19283d692628SSanjeev Bagewadi boolean_t 19293d692628SSanjeev Bagewadi dmu_buf_freeable(dmu_buf_t *dbuf) 19303d692628SSanjeev Bagewadi { 19313d692628SSanjeev Bagewadi boolean_t res = B_FALSE; 19323d692628SSanjeev Bagewadi dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 19333d692628SSanjeev Bagewadi 19343d692628SSanjeev Bagewadi if (db->db_blkptr) 19353d692628SSanjeev Bagewadi res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 19363d692628SSanjeev Bagewadi db->db_blkptr->blk_birth); 19373d692628SSanjeev Bagewadi 19383d692628SSanjeev Bagewadi return (res); 19393d692628SSanjeev Bagewadi } 19403d692628SSanjeev Bagewadi 1941c717a561Smaybee static void 1942c717a561Smaybee dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 1943fa9e4066Sahrens { 1944c717a561Smaybee /* ASSERT(dmu_tx_is_syncing(tx) */ 1945c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 1946c717a561Smaybee 1947c717a561Smaybee if (db->db_blkptr != NULL) 1948c717a561Smaybee return; 1949c717a561Smaybee 1950c717a561Smaybee if (db->db_level == dn->dn_phys->dn_nlevels-1) { 1951c717a561Smaybee /* 1952c717a561Smaybee * This buffer was allocated at a time when there was 1953c717a561Smaybee * no available blkptrs from the dnode, or it was 1954c717a561Smaybee * inappropriate to hook it in (i.e., nlevels mis-match). 1955c717a561Smaybee */ 1956c717a561Smaybee ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 1957c717a561Smaybee ASSERT(db->db_parent == NULL); 1958c717a561Smaybee db->db_parent = dn->dn_dbuf; 1959c717a561Smaybee db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 1960c717a561Smaybee DBUF_VERIFY(db); 1961c717a561Smaybee } else { 1962c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 1963c717a561Smaybee int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 1964c717a561Smaybee 1965c717a561Smaybee ASSERT(dn->dn_phys->dn_nlevels > 1); 1966c717a561Smaybee if (parent == NULL) { 1967c717a561Smaybee mutex_exit(&db->db_mtx); 1968c717a561Smaybee rw_enter(&dn->dn_struct_rwlock, RW_READER); 1969c717a561Smaybee (void) dbuf_hold_impl(dn, db->db_level+1, 1970c717a561Smaybee db->db_blkid >> epbs, FALSE, db, &parent); 1971c717a561Smaybee rw_exit(&dn->dn_struct_rwlock); 1972c717a561Smaybee mutex_enter(&db->db_mtx); 1973c717a561Smaybee db->db_parent = parent; 1974c717a561Smaybee } 1975c717a561Smaybee db->db_blkptr = (blkptr_t *)parent->db.db_data + 1976c717a561Smaybee (db->db_blkid & ((1ULL << epbs) - 1)); 1977c717a561Smaybee DBUF_VERIFY(db); 1978c717a561Smaybee } 1979c717a561Smaybee } 1980c717a561Smaybee 1981c717a561Smaybee static void 1982c717a561Smaybee dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 1983c717a561Smaybee { 1984c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 1985c717a561Smaybee dnode_t *dn = db->db_dnode; 1986c717a561Smaybee zio_t *zio; 1987c717a561Smaybee 1988c717a561Smaybee ASSERT(dmu_tx_is_syncing(tx)); 1989c717a561Smaybee 1990c717a561Smaybee dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 1991c717a561Smaybee 1992c717a561Smaybee mutex_enter(&db->db_mtx); 1993c717a561Smaybee 1994c717a561Smaybee ASSERT(db->db_level > 0); 1995c717a561Smaybee DBUF_VERIFY(db); 1996c717a561Smaybee 1997c717a561Smaybee if (db->db_buf == NULL) { 1998c717a561Smaybee mutex_exit(&db->db_mtx); 1999c717a561Smaybee (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2000c717a561Smaybee mutex_enter(&db->db_mtx); 2001c717a561Smaybee } 2002c717a561Smaybee ASSERT3U(db->db_state, ==, DB_CACHED); 2003c717a561Smaybee ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2004c717a561Smaybee ASSERT(db->db_buf != NULL); 2005c717a561Smaybee 2006c717a561Smaybee dbuf_check_blkptr(dn, db); 2007c717a561Smaybee 2008c717a561Smaybee db->db_data_pending = dr; 2009c717a561Smaybee 2010af2c4821Smaybee mutex_exit(&db->db_mtx); 2011088f3894Sahrens dbuf_write(dr, db->db_buf, tx); 2012c717a561Smaybee 2013c717a561Smaybee zio = dr->dr_zio; 2014c717a561Smaybee mutex_enter(&dr->dt.di.dr_mtx); 2015c717a561Smaybee dbuf_sync_list(&dr->dt.di.dr_children, tx); 2016c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2017c717a561Smaybee mutex_exit(&dr->dt.di.dr_mtx); 2018c717a561Smaybee zio_nowait(zio); 2019c717a561Smaybee } 2020c717a561Smaybee 2021c717a561Smaybee static void 2022c717a561Smaybee dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2023c717a561Smaybee { 2024c717a561Smaybee arc_buf_t **datap = &dr->dt.dl.dr_data; 2025c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 2026fa9e4066Sahrens dnode_t *dn = db->db_dnode; 2027503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 2028c717a561Smaybee uint64_t txg = tx->tx_txg; 2029fa9e4066Sahrens 2030fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 2031fa9e4066Sahrens 2032fa9e4066Sahrens dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2033fa9e4066Sahrens 2034fa9e4066Sahrens mutex_enter(&db->db_mtx); 2035fa9e4066Sahrens /* 2036fa9e4066Sahrens * To be synced, we must be dirtied. But we 2037fa9e4066Sahrens * might have been freed after the dirty. 2038fa9e4066Sahrens */ 2039fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 2040fa9e4066Sahrens /* This buffer has been freed since it was dirtied */ 2041fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 2042fa9e4066Sahrens } else if (db->db_state == DB_FILL) { 2043fa9e4066Sahrens /* This buffer was freed and is now being re-filled */ 2044c717a561Smaybee ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2045fa9e4066Sahrens } else { 204682c9918fSTim Haley ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2047fa9e4066Sahrens } 20489c9dc39aSek110237 DBUF_VERIFY(db); 2049fa9e4066Sahrens 2050fa9e4066Sahrens /* 2051c717a561Smaybee * If this is a bonus buffer, simply copy the bonus data into the 2052c717a561Smaybee * dnode. It will be written out when the dnode is synced (and it 2053c717a561Smaybee * will be synced, since it must have been dirty for dbuf_sync to 2054c717a561Smaybee * be called). 2055fa9e4066Sahrens */ 2056ea8dc4b6Seschrock if (db->db_blkid == DB_BONUS_BLKID) { 2057c717a561Smaybee dbuf_dirty_record_t **drp; 20581934e92fSmaybee 2059ea8dc4b6Seschrock ASSERT(*datap != NULL); 2060ea8dc4b6Seschrock ASSERT3U(db->db_level, ==, 0); 2061ea8dc4b6Seschrock ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2062ea8dc4b6Seschrock bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 20630e8c6158Smaybee if (*datap != db->db.db_data) { 2064ea8dc4b6Seschrock zio_buf_free(*datap, DN_MAX_BONUSLEN); 20655a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 20660e8c6158Smaybee } 2067ea8dc4b6Seschrock db->db_data_pending = NULL; 2068c717a561Smaybee drp = &db->db_last_dirty; 2069c717a561Smaybee while (*drp != dr) 2070c717a561Smaybee drp = &(*drp)->dr_next; 207117f17c2dSbonwick ASSERT(dr->dr_next == NULL); 2072*b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 207317f17c2dSbonwick *drp = dr->dr_next; 2074c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2075ea8dc4b6Seschrock ASSERT(db->db_dirtycnt > 0); 2076ea8dc4b6Seschrock db->db_dirtycnt -= 1; 2077*b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2078ea8dc4b6Seschrock return; 2079ea8dc4b6Seschrock } 2080ea8dc4b6Seschrock 2081c5c6ffa0Smaybee /* 2082f82bfe17Sgw25295 * This function may have dropped the db_mtx lock allowing a dmu_sync 2083f82bfe17Sgw25295 * operation to sneak in. As a result, we need to ensure that we 2084f82bfe17Sgw25295 * don't check the dr_override_state until we have returned from 2085f82bfe17Sgw25295 * dbuf_check_blkptr. 2086f82bfe17Sgw25295 */ 2087f82bfe17Sgw25295 dbuf_check_blkptr(dn, db); 2088f82bfe17Sgw25295 2089f82bfe17Sgw25295 /* 2090c717a561Smaybee * If this buffer is in the middle of an immdiate write, 2091c717a561Smaybee * wait for the synchronous IO to complete. 2092c5c6ffa0Smaybee */ 2093c717a561Smaybee while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2094c5c6ffa0Smaybee ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2095c5c6ffa0Smaybee cv_wait(&db->db_changed, &db->db_mtx); 2096c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2097c5c6ffa0Smaybee } 2098c717a561Smaybee 2099ab69d62fSMatthew Ahrens if (db->db_state != DB_NOFILL && 2100ab69d62fSMatthew Ahrens dn->dn_object != DMU_META_DNODE_OBJECT && 2101ab69d62fSMatthew Ahrens refcount_count(&db->db_holds) > 1 && 2102*b24ab676SJeff Bonwick dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2103ab69d62fSMatthew Ahrens *datap == db->db_buf) { 2104fa9e4066Sahrens /* 210582c9918fSTim Haley * If this buffer is currently "in use" (i.e., there 210682c9918fSTim Haley * are active holds and db_data still references it), 210782c9918fSTim Haley * then make a copy before we start the write so that 210882c9918fSTim Haley * any modifications from the open txg will not leak 210982c9918fSTim Haley * into this write. 2110fa9e4066Sahrens * 211182c9918fSTim Haley * NOTE: this copy does not need to be made for 211282c9918fSTim Haley * objects only modified in the syncing context (e.g. 211382c9918fSTim Haley * DNONE_DNODE blocks). 2114fa9e4066Sahrens */ 2115ab69d62fSMatthew Ahrens int blksz = arc_buf_size(*datap); 2116ab69d62fSMatthew Ahrens arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2117ab69d62fSMatthew Ahrens *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2118c717a561Smaybee bcopy(db->db.db_data, (*datap)->b_data, blksz); 2119fa9e4066Sahrens } 2120c717a561Smaybee db->db_data_pending = dr; 2121fa9e4066Sahrens 2122fa9e4066Sahrens mutex_exit(&db->db_mtx); 2123fa9e4066Sahrens 2124088f3894Sahrens dbuf_write(dr, *datap, tx); 2125c717a561Smaybee 2126c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 2127c717a561Smaybee if (dn->dn_object == DMU_META_DNODE_OBJECT) 2128c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2129c717a561Smaybee else 2130c717a561Smaybee zio_nowait(dr->dr_zio); 2131fa9e4066Sahrens } 2132c717a561Smaybee 2133c717a561Smaybee void 2134c717a561Smaybee dbuf_sync_list(list_t *list, dmu_tx_t *tx) 2135c717a561Smaybee { 2136c717a561Smaybee dbuf_dirty_record_t *dr; 2137c717a561Smaybee 2138c717a561Smaybee while (dr = list_head(list)) { 2139c717a561Smaybee if (dr->dr_zio != NULL) { 2140c717a561Smaybee /* 2141c717a561Smaybee * If we find an already initialized zio then we 2142c717a561Smaybee * are processing the meta-dnode, and we have finished. 2143c717a561Smaybee * The dbufs for all dnodes are put back on the list 2144c717a561Smaybee * during processing, so that we can zio_wait() 2145c717a561Smaybee * these IOs after initiating all child IOs. 2146c717a561Smaybee */ 2147c717a561Smaybee ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2148c717a561Smaybee DMU_META_DNODE_OBJECT); 2149c717a561Smaybee break; 2150fa9e4066Sahrens } 2151c717a561Smaybee list_remove(list, dr); 2152c717a561Smaybee if (dr->dr_dbuf->db_level > 0) 2153c717a561Smaybee dbuf_sync_indirect(dr, tx); 2154c717a561Smaybee else 2155c717a561Smaybee dbuf_sync_leaf(dr, tx); 2156c717a561Smaybee } 2157c717a561Smaybee } 2158c717a561Smaybee 2159fa9e4066Sahrens /* ARGSUSED */ 2160fa9e4066Sahrens static void 2161c717a561Smaybee dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2162fa9e4066Sahrens { 2163fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 2164e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2165c717a561Smaybee blkptr_t *bp_orig = &zio->io_bp_orig; 2166*b24ab676SJeff Bonwick dnode_t *dn = db->db_dnode; 2167*b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2168*b24ab676SJeff Bonwick int64_t delta; 2169fa9e4066Sahrens uint64_t fill = 0; 2170*b24ab676SJeff Bonwick int i; 2171fa9e4066Sahrens 2172e14bb325SJeff Bonwick ASSERT(db->db_blkptr == bp); 2173e14bb325SJeff Bonwick 2174*b24ab676SJeff Bonwick delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2175*b24ab676SJeff Bonwick dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2176*b24ab676SJeff Bonwick zio->io_prev_space_delta = delta; 2177fa9e4066Sahrens 2178e14bb325SJeff Bonwick if (BP_IS_HOLE(bp)) { 2179*b24ab676SJeff Bonwick ASSERT(bp->blk_fill == 0); 2180c717a561Smaybee return; 2181c717a561Smaybee } 2182c717a561Smaybee 2183e14bb325SJeff Bonwick ASSERT(BP_GET_TYPE(bp) == dn->dn_type); 2184e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2185e14bb325SJeff Bonwick 2186fa9e4066Sahrens mutex_enter(&db->db_mtx); 2187fa9e4066Sahrens 2188fa9e4066Sahrens if (db->db_level == 0) { 2189fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 2190c717a561Smaybee if (db->db_blkid > dn->dn_phys->dn_maxblkid) 2191fa9e4066Sahrens dn->dn_phys->dn_maxblkid = db->db_blkid; 2192fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 2193fa9e4066Sahrens 2194fa9e4066Sahrens if (dn->dn_type == DMU_OT_DNODE) { 2195fa9e4066Sahrens dnode_phys_t *dnp = db->db.db_data; 2196fa9e4066Sahrens for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2197fa9e4066Sahrens i--, dnp++) { 2198fa9e4066Sahrens if (dnp->dn_type != DMU_OT_NONE) 2199fa9e4066Sahrens fill++; 2200fa9e4066Sahrens } 2201fa9e4066Sahrens } else { 2202fa9e4066Sahrens fill = 1; 2203fa9e4066Sahrens } 2204fa9e4066Sahrens } else { 2205e14bb325SJeff Bonwick blkptr_t *ibp = db->db.db_data; 2206fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2207e14bb325SJeff Bonwick for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2208e14bb325SJeff Bonwick if (BP_IS_HOLE(ibp)) 2209fa9e4066Sahrens continue; 2210e14bb325SJeff Bonwick ASSERT3U(BP_GET_LSIZE(ibp), ==, 2211fa9e4066Sahrens db->db_level == 1 ? dn->dn_datablksz : 2212fa9e4066Sahrens (1<<dn->dn_phys->dn_indblkshift)); 2213e14bb325SJeff Bonwick fill += ibp->blk_fill; 2214fa9e4066Sahrens } 2215fa9e4066Sahrens } 2216fa9e4066Sahrens 2217e14bb325SJeff Bonwick bp->blk_fill = fill; 2218fa9e4066Sahrens 2219fa9e4066Sahrens mutex_exit(&db->db_mtx); 2220fa9e4066Sahrens } 2221fa9e4066Sahrens 2222c717a561Smaybee /* ARGSUSED */ 2223c717a561Smaybee static void 2224c717a561Smaybee dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2225c717a561Smaybee { 2226c717a561Smaybee dmu_buf_impl_t *db = vdb; 2227*b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2228*b24ab676SJeff Bonwick blkptr_t *bp_orig = &zio->io_bp_orig; 2229*b24ab676SJeff Bonwick dnode_t *dn = db->db_dnode; 2230*b24ab676SJeff Bonwick objset_t *os = dn->dn_objset; 2231c717a561Smaybee uint64_t txg = zio->io_txg; 2232c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 2233c717a561Smaybee 2234c717a561Smaybee ASSERT3U(zio->io_error, ==, 0); 2235*b24ab676SJeff Bonwick ASSERT(db->db_blkptr == bp); 2236*b24ab676SJeff Bonwick 2237*b24ab676SJeff Bonwick if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 2238*b24ab676SJeff Bonwick ASSERT(BP_EQUAL(bp, bp_orig)); 2239*b24ab676SJeff Bonwick } else { 2240*b24ab676SJeff Bonwick dsl_dataset_t *ds = os->os_dsl_dataset; 2241*b24ab676SJeff Bonwick dmu_tx_t *tx = os->os_synctx; 2242*b24ab676SJeff Bonwick 2243*b24ab676SJeff Bonwick (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2244*b24ab676SJeff Bonwick dsl_dataset_block_born(ds, bp, tx); 2245*b24ab676SJeff Bonwick } 2246c717a561Smaybee 2247c717a561Smaybee mutex_enter(&db->db_mtx); 2248c717a561Smaybee 2249*b24ab676SJeff Bonwick DBUF_VERIFY(db); 2250*b24ab676SJeff Bonwick 2251c717a561Smaybee drp = &db->db_last_dirty; 225217f17c2dSbonwick while ((dr = *drp) != db->db_data_pending) 225317f17c2dSbonwick drp = &dr->dr_next; 225417f17c2dSbonwick ASSERT(!list_link_active(&dr->dr_dirty_node)); 225517f17c2dSbonwick ASSERT(dr->dr_txg == txg); 2256*b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 225717f17c2dSbonwick ASSERT(dr->dr_next == NULL); 225817f17c2dSbonwick *drp = dr->dr_next; 2259c717a561Smaybee 2260c717a561Smaybee if (db->db_level == 0) { 2261c717a561Smaybee ASSERT(db->db_blkid != DB_BONUS_BLKID); 2262c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 226382c9918fSTim Haley if (db->db_state != DB_NOFILL) { 2264c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 226582c9918fSTim Haley VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 226682c9918fSTim Haley db) == 1); 2267*b24ab676SJeff Bonwick else if (!arc_released(db->db_buf)) 2268c717a561Smaybee arc_set_callback(db->db_buf, dbuf_do_evict, db); 226982c9918fSTim Haley } 2270c717a561Smaybee } else { 2271c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2272c717a561Smaybee ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2273c717a561Smaybee if (!BP_IS_HOLE(db->db_blkptr)) { 2274c717a561Smaybee int epbs = 2275c717a561Smaybee dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2276c717a561Smaybee ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2277c717a561Smaybee db->db.db_size); 2278c717a561Smaybee ASSERT3U(dn->dn_phys->dn_maxblkid 2279c717a561Smaybee >> (db->db_level * epbs), >=, db->db_blkid); 2280c717a561Smaybee arc_set_callback(db->db_buf, dbuf_do_evict, db); 2281c717a561Smaybee } 2282c25056deSgw25295 mutex_destroy(&dr->dt.di.dr_mtx); 2283c25056deSgw25295 list_destroy(&dr->dt.di.dr_children); 2284c717a561Smaybee } 2285c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2286c717a561Smaybee 2287c717a561Smaybee cv_broadcast(&db->db_changed); 2288c717a561Smaybee ASSERT(db->db_dirtycnt > 0); 2289c717a561Smaybee db->db_dirtycnt -= 1; 2290c717a561Smaybee db->db_data_pending = NULL; 2291*b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2292*b24ab676SJeff Bonwick } 2293*b24ab676SJeff Bonwick 2294*b24ab676SJeff Bonwick static void 2295*b24ab676SJeff Bonwick dbuf_write_nofill_ready(zio_t *zio) 2296*b24ab676SJeff Bonwick { 2297*b24ab676SJeff Bonwick dbuf_write_ready(zio, NULL, zio->io_private); 2298*b24ab676SJeff Bonwick } 2299*b24ab676SJeff Bonwick 2300*b24ab676SJeff Bonwick static void 2301*b24ab676SJeff Bonwick dbuf_write_nofill_done(zio_t *zio) 2302*b24ab676SJeff Bonwick { 2303*b24ab676SJeff Bonwick dbuf_write_done(zio, NULL, zio->io_private); 2304*b24ab676SJeff Bonwick } 2305*b24ab676SJeff Bonwick 2306*b24ab676SJeff Bonwick static void 2307*b24ab676SJeff Bonwick dbuf_write_override_ready(zio_t *zio) 2308*b24ab676SJeff Bonwick { 2309*b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = zio->io_private; 2310*b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 2311*b24ab676SJeff Bonwick 2312*b24ab676SJeff Bonwick dbuf_write_ready(zio, NULL, db); 2313*b24ab676SJeff Bonwick } 2314*b24ab676SJeff Bonwick 2315*b24ab676SJeff Bonwick static void 2316*b24ab676SJeff Bonwick dbuf_write_override_done(zio_t *zio) 2317*b24ab676SJeff Bonwick { 2318*b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = zio->io_private; 2319*b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 2320*b24ab676SJeff Bonwick blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 2321*b24ab676SJeff Bonwick 2322*b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 2323*b24ab676SJeff Bonwick if (!BP_EQUAL(zio->io_bp, obp)) { 2324*b24ab676SJeff Bonwick if (!BP_IS_HOLE(obp)) 2325*b24ab676SJeff Bonwick dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 2326*b24ab676SJeff Bonwick arc_release(dr->dt.dl.dr_data, db); 2327*b24ab676SJeff Bonwick } 2328c717a561Smaybee mutex_exit(&db->db_mtx); 2329c717a561Smaybee 2330*b24ab676SJeff Bonwick dbuf_write_done(zio, NULL, db); 2331*b24ab676SJeff Bonwick } 2332c717a561Smaybee 2333*b24ab676SJeff Bonwick static void 2334*b24ab676SJeff Bonwick dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 2335*b24ab676SJeff Bonwick { 2336*b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 2337*b24ab676SJeff Bonwick dnode_t *dn = db->db_dnode; 2338*b24ab676SJeff Bonwick objset_t *os = dn->dn_objset; 2339*b24ab676SJeff Bonwick dmu_buf_impl_t *parent = db->db_parent; 2340*b24ab676SJeff Bonwick uint64_t txg = tx->tx_txg; 2341*b24ab676SJeff Bonwick zbookmark_t zb; 2342*b24ab676SJeff Bonwick zio_prop_t zp; 2343*b24ab676SJeff Bonwick zio_t *zio; 2344*b24ab676SJeff Bonwick 2345*b24ab676SJeff Bonwick if (db->db_state != DB_NOFILL) { 2346*b24ab676SJeff Bonwick if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 2347*b24ab676SJeff Bonwick /* 2348*b24ab676SJeff Bonwick * Private object buffers are released here rather 2349*b24ab676SJeff Bonwick * than in dbuf_dirty() since they are only modified 2350*b24ab676SJeff Bonwick * in the syncing context and we don't want the 2351*b24ab676SJeff Bonwick * overhead of making multiple copies of the data. 2352*b24ab676SJeff Bonwick */ 2353*b24ab676SJeff Bonwick if (BP_IS_HOLE(db->db_blkptr)) { 2354*b24ab676SJeff Bonwick arc_buf_thaw(data); 2355*b24ab676SJeff Bonwick } else { 2356*b24ab676SJeff Bonwick arc_release(data, db); 2357*b24ab676SJeff Bonwick } 2358*b24ab676SJeff Bonwick } 2359*b24ab676SJeff Bonwick } 2360*b24ab676SJeff Bonwick 2361*b24ab676SJeff Bonwick if (parent != dn->dn_dbuf) { 2362*b24ab676SJeff Bonwick ASSERT(parent && parent->db_data_pending); 2363*b24ab676SJeff Bonwick ASSERT(db->db_level == parent->db_level-1); 2364*b24ab676SJeff Bonwick ASSERT(arc_released(parent->db_buf)); 2365*b24ab676SJeff Bonwick zio = parent->db_data_pending->dr_zio; 2366*b24ab676SJeff Bonwick } else { 2367*b24ab676SJeff Bonwick ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1); 2368*b24ab676SJeff Bonwick ASSERT3P(db->db_blkptr, ==, 2369*b24ab676SJeff Bonwick &dn->dn_phys->dn_blkptr[db->db_blkid]); 2370*b24ab676SJeff Bonwick zio = dn->dn_zio; 2371*b24ab676SJeff Bonwick } 2372*b24ab676SJeff Bonwick 2373*b24ab676SJeff Bonwick ASSERT(db->db_level == 0 || data == db->db_buf); 2374*b24ab676SJeff Bonwick ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2375*b24ab676SJeff Bonwick ASSERT(zio); 2376*b24ab676SJeff Bonwick 2377*b24ab676SJeff Bonwick SET_BOOKMARK(&zb, os->os_dsl_dataset ? 2378*b24ab676SJeff Bonwick os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 2379*b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 2380*b24ab676SJeff Bonwick 2381*b24ab676SJeff Bonwick dmu_write_policy(os, dn, db->db_level, 2382*b24ab676SJeff Bonwick db->db_state == DB_NOFILL ? WP_NOFILL : 0, &zp); 2383*b24ab676SJeff Bonwick 2384*b24ab676SJeff Bonwick if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2385*b24ab676SJeff Bonwick ASSERT(db->db_state != DB_NOFILL); 2386*b24ab676SJeff Bonwick dr->dr_zio = zio_write(zio, os->os_spa, txg, 2387*b24ab676SJeff Bonwick db->db_blkptr, data->b_data, arc_buf_size(data), &zp, 2388*b24ab676SJeff Bonwick dbuf_write_override_ready, dbuf_write_override_done, dr, 2389*b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2390*b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 2391*b24ab676SJeff Bonwick dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 2392*b24ab676SJeff Bonwick zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 2393*b24ab676SJeff Bonwick dr->dt.dl.dr_copies); 2394*b24ab676SJeff Bonwick mutex_exit(&db->db_mtx); 2395*b24ab676SJeff Bonwick } else if (db->db_state == DB_NOFILL) { 2396*b24ab676SJeff Bonwick ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); 2397*b24ab676SJeff Bonwick dr->dr_zio = zio_write(zio, os->os_spa, txg, 2398*b24ab676SJeff Bonwick db->db_blkptr, NULL, db->db.db_size, &zp, 2399*b24ab676SJeff Bonwick dbuf_write_nofill_ready, dbuf_write_nofill_done, db, 2400*b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, 2401*b24ab676SJeff Bonwick ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 2402*b24ab676SJeff Bonwick } else { 2403*b24ab676SJeff Bonwick ASSERT(arc_released(data)); 2404*b24ab676SJeff Bonwick dr->dr_zio = arc_write(zio, os->os_spa, txg, 2405*b24ab676SJeff Bonwick db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp, 2406*b24ab676SJeff Bonwick dbuf_write_ready, dbuf_write_done, db, 2407*b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2408*b24ab676SJeff Bonwick } 2409fa9e4066Sahrens } 2410