1fa9e4066Sahrens /* 2fa9e4066Sahrens * CDDL HEADER START 3fa9e4066Sahrens * 4fa9e4066Sahrens * The contents of this file are subject to the terms of the 5f65e61c0Sahrens * Common Development and Distribution License (the "License"). 6f65e61c0Sahrens * You may not use this file except in compliance with the License. 7fa9e4066Sahrens * 8fa9e4066Sahrens * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9fa9e4066Sahrens * or http://www.opensolaris.org/os/licensing. 10fa9e4066Sahrens * See the License for the specific language governing permissions 11fa9e4066Sahrens * and limitations under the License. 12fa9e4066Sahrens * 13fa9e4066Sahrens * When distributing Covered Code, include this CDDL HEADER in each 14fa9e4066Sahrens * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15fa9e4066Sahrens * If applicable, add the following below this CDDL HEADER, with the 16fa9e4066Sahrens * fields enclosed by brackets "[]" replaced with your own identifying 17fa9e4066Sahrens * information: Portions Copyright [yyyy] [name of copyright owner] 18fa9e4066Sahrens * 19fa9e4066Sahrens * CDDL HEADER END 20fa9e4066Sahrens */ 21fa9e4066Sahrens /* 2206e0070dSMark Shellenbaum * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 233f2366c2SGordon Ross * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24ad135b5dSChristopher Siden * Copyright (c) 2012 by Delphix. All rights reserved. 25fa9e4066Sahrens */ 26fa9e4066Sahrens 27fa9e4066Sahrens #include <sys/zfs_context.h> 28fa9e4066Sahrens #include <sys/dmu.h> 29fa9e4066Sahrens #include <sys/dmu_impl.h> 30fa9e4066Sahrens #include <sys/dbuf.h> 31fa9e4066Sahrens #include <sys/dmu_objset.h> 32fa9e4066Sahrens #include <sys/dsl_dataset.h> 33fa9e4066Sahrens #include <sys/dsl_dir.h> 34fa9e4066Sahrens #include <sys/dmu_tx.h> 35fa9e4066Sahrens #include <sys/spa.h> 36fa9e4066Sahrens #include <sys/zio.h> 37fa9e4066Sahrens #include <sys/dmu_zfetch.h> 380a586ceaSMark Shellenbaum #include <sys/sa.h> 390a586ceaSMark Shellenbaum #include <sys/sa_impl.h> 40fa9e4066Sahrens 41fa9e4066Sahrens static void dbuf_destroy(dmu_buf_impl_t *db); 42fa9e4066Sahrens static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 43088f3894Sahrens static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 44fa9e4066Sahrens 45fa9e4066Sahrens /* 46fa9e4066Sahrens * Global data structures and functions for the dbuf cache. 47fa9e4066Sahrens */ 48fa9e4066Sahrens static kmem_cache_t *dbuf_cache; 49fa9e4066Sahrens 50fa9e4066Sahrens /* ARGSUSED */ 51fa9e4066Sahrens static int 52fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag) 53fa9e4066Sahrens { 54fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 55fa9e4066Sahrens bzero(db, sizeof (dmu_buf_impl_t)); 56fa9e4066Sahrens 57fa9e4066Sahrens mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 58fa9e4066Sahrens cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 59fa9e4066Sahrens refcount_create(&db->db_holds); 60fa9e4066Sahrens return (0); 61fa9e4066Sahrens } 62fa9e4066Sahrens 63fa9e4066Sahrens /* ARGSUSED */ 64fa9e4066Sahrens static void 65fa9e4066Sahrens dbuf_dest(void *vdb, void *unused) 66fa9e4066Sahrens { 67fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 68fa9e4066Sahrens mutex_destroy(&db->db_mtx); 69fa9e4066Sahrens cv_destroy(&db->db_changed); 70fa9e4066Sahrens refcount_destroy(&db->db_holds); 71fa9e4066Sahrens } 72fa9e4066Sahrens 73fa9e4066Sahrens /* 74fa9e4066Sahrens * dbuf hash table routines 75fa9e4066Sahrens */ 76fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table; 77fa9e4066Sahrens 78fa9e4066Sahrens static uint64_t dbuf_hash_count; 79fa9e4066Sahrens 80fa9e4066Sahrens static uint64_t 81fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 82fa9e4066Sahrens { 83fa9e4066Sahrens uintptr_t osv = (uintptr_t)os; 84fa9e4066Sahrens uint64_t crc = -1ULL; 85fa9e4066Sahrens 86fa9e4066Sahrens ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 87fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 88fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 89fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 90fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 91fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 92fa9e4066Sahrens crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 93fa9e4066Sahrens 94fa9e4066Sahrens crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 95fa9e4066Sahrens 96fa9e4066Sahrens return (crc); 97fa9e4066Sahrens } 98fa9e4066Sahrens 99fa9e4066Sahrens #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 100fa9e4066Sahrens 101fa9e4066Sahrens #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 102fa9e4066Sahrens ((dbuf)->db.db_object == (obj) && \ 103fa9e4066Sahrens (dbuf)->db_objset == (os) && \ 104fa9e4066Sahrens (dbuf)->db_level == (level) && \ 105fa9e4066Sahrens (dbuf)->db_blkid == (blkid)) 106fa9e4066Sahrens 107fa9e4066Sahrens dmu_buf_impl_t * 108fa9e4066Sahrens dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 109fa9e4066Sahrens { 110fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 111503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 112fa9e4066Sahrens uint64_t obj = dn->dn_object; 113fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 114fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 115fa9e4066Sahrens dmu_buf_impl_t *db; 116fa9e4066Sahrens 117fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 118fa9e4066Sahrens for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 119fa9e4066Sahrens if (DBUF_EQUAL(db, os, obj, level, blkid)) { 120fa9e4066Sahrens mutex_enter(&db->db_mtx); 121ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 122fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 123fa9e4066Sahrens return (db); 124fa9e4066Sahrens } 125fa9e4066Sahrens mutex_exit(&db->db_mtx); 126fa9e4066Sahrens } 127fa9e4066Sahrens } 128fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 129fa9e4066Sahrens return (NULL); 130fa9e4066Sahrens } 131fa9e4066Sahrens 132fa9e4066Sahrens /* 133fa9e4066Sahrens * Insert an entry into the hash table. If there is already an element 134fa9e4066Sahrens * equal to elem in the hash table, then the already existing element 135fa9e4066Sahrens * will be returned and the new element will not be inserted. 136fa9e4066Sahrens * Otherwise returns NULL. 137fa9e4066Sahrens */ 138fa9e4066Sahrens static dmu_buf_impl_t * 139fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db) 140fa9e4066Sahrens { 141fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 142503ad85cSMatthew Ahrens objset_t *os = db->db_objset; 143fa9e4066Sahrens uint64_t obj = db->db.db_object; 144fa9e4066Sahrens int level = db->db_level; 145fa9e4066Sahrens uint64_t blkid = db->db_blkid; 146fa9e4066Sahrens uint64_t hv = DBUF_HASH(os, obj, level, blkid); 147fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 148fa9e4066Sahrens dmu_buf_impl_t *dbf; 149fa9e4066Sahrens 150fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 151fa9e4066Sahrens for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 152fa9e4066Sahrens if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 153fa9e4066Sahrens mutex_enter(&dbf->db_mtx); 154ea8dc4b6Seschrock if (dbf->db_state != DB_EVICTING) { 155fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 156fa9e4066Sahrens return (dbf); 157fa9e4066Sahrens } 158fa9e4066Sahrens mutex_exit(&dbf->db_mtx); 159fa9e4066Sahrens } 160fa9e4066Sahrens } 161fa9e4066Sahrens 162fa9e4066Sahrens mutex_enter(&db->db_mtx); 163fa9e4066Sahrens db->db_hash_next = h->hash_table[idx]; 164fa9e4066Sahrens h->hash_table[idx] = db; 165fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 166fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, 1); 167fa9e4066Sahrens 168fa9e4066Sahrens return (NULL); 169fa9e4066Sahrens } 170fa9e4066Sahrens 171fa9e4066Sahrens /* 172fa9e4066Sahrens * Remove an entry from the hash table. This operation will 173fa9e4066Sahrens * fail if there are any existing holds on the db. 174fa9e4066Sahrens */ 175fa9e4066Sahrens static void 176fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db) 177fa9e4066Sahrens { 178fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 179fa9e4066Sahrens uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 180fa9e4066Sahrens db->db_level, db->db_blkid); 181fa9e4066Sahrens uint64_t idx = hv & h->hash_table_mask; 182fa9e4066Sahrens dmu_buf_impl_t *dbf, **dbp; 183fa9e4066Sahrens 184fa9e4066Sahrens /* 185fa9e4066Sahrens * We musn't hold db_mtx to maintin lock ordering: 186fa9e4066Sahrens * DBUF_HASH_MUTEX > db_mtx. 187fa9e4066Sahrens */ 188fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 189ea8dc4b6Seschrock ASSERT(db->db_state == DB_EVICTING); 190fa9e4066Sahrens ASSERT(!MUTEX_HELD(&db->db_mtx)); 191fa9e4066Sahrens 192fa9e4066Sahrens mutex_enter(DBUF_HASH_MUTEX(h, idx)); 193fa9e4066Sahrens dbp = &h->hash_table[idx]; 194fa9e4066Sahrens while ((dbf = *dbp) != db) { 195fa9e4066Sahrens dbp = &dbf->db_hash_next; 196fa9e4066Sahrens ASSERT(dbf != NULL); 197fa9e4066Sahrens } 198fa9e4066Sahrens *dbp = db->db_hash_next; 199fa9e4066Sahrens db->db_hash_next = NULL; 200fa9e4066Sahrens mutex_exit(DBUF_HASH_MUTEX(h, idx)); 201fa9e4066Sahrens atomic_add_64(&dbuf_hash_count, -1); 202fa9e4066Sahrens } 203fa9e4066Sahrens 204ea8dc4b6Seschrock static arc_evict_func_t dbuf_do_evict; 205fa9e4066Sahrens 206fa9e4066Sahrens static void 207fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db) 208fa9e4066Sahrens { 209fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 210fa9e4066Sahrens 211c717a561Smaybee if (db->db_level != 0 || db->db_evict_func == NULL) 212fa9e4066Sahrens return; 213fa9e4066Sahrens 214c717a561Smaybee if (db->db_user_data_ptr_ptr) 215c717a561Smaybee *db->db_user_data_ptr_ptr = db->db.db_data; 216c717a561Smaybee db->db_evict_func(&db->db, db->db_user_ptr); 217c717a561Smaybee db->db_user_ptr = NULL; 218c717a561Smaybee db->db_user_data_ptr_ptr = NULL; 219c717a561Smaybee db->db_evict_func = NULL; 220fa9e4066Sahrens } 221fa9e4066Sahrens 222744947dcSTom Erickson boolean_t 223744947dcSTom Erickson dbuf_is_metadata(dmu_buf_impl_t *db) 224744947dcSTom Erickson { 225744947dcSTom Erickson if (db->db_level > 0) { 226744947dcSTom Erickson return (B_TRUE); 227744947dcSTom Erickson } else { 228744947dcSTom Erickson boolean_t is_metadata; 229744947dcSTom Erickson 230744947dcSTom Erickson DB_DNODE_ENTER(db); 231ad135b5dSChristopher Siden is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 232744947dcSTom Erickson DB_DNODE_EXIT(db); 233744947dcSTom Erickson 234744947dcSTom Erickson return (is_metadata); 235744947dcSTom Erickson } 236744947dcSTom Erickson } 237744947dcSTom Erickson 238fa9e4066Sahrens void 239ea8dc4b6Seschrock dbuf_evict(dmu_buf_impl_t *db) 240ea8dc4b6Seschrock { 241ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 242ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 243c717a561Smaybee ASSERT(db->db_data_pending == NULL); 244ea8dc4b6Seschrock 245ea8dc4b6Seschrock dbuf_clear(db); 246ea8dc4b6Seschrock dbuf_destroy(db); 247ea8dc4b6Seschrock } 248ea8dc4b6Seschrock 249ea8dc4b6Seschrock void 250fa9e4066Sahrens dbuf_init(void) 251fa9e4066Sahrens { 252ea8dc4b6Seschrock uint64_t hsize = 1ULL << 16; 253fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 254fa9e4066Sahrens int i; 255fa9e4066Sahrens 256fa9e4066Sahrens /* 257fa9e4066Sahrens * The hash table is big enough to fill all of physical memory 258ea8dc4b6Seschrock * with an average 4K block size. The table will take up 259ea8dc4b6Seschrock * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 260fa9e4066Sahrens */ 261ea8dc4b6Seschrock while (hsize * 4096 < physmem * PAGESIZE) 262fa9e4066Sahrens hsize <<= 1; 263fa9e4066Sahrens 264ea8dc4b6Seschrock retry: 265fa9e4066Sahrens h->hash_table_mask = hsize - 1; 266ea8dc4b6Seschrock h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 267ea8dc4b6Seschrock if (h->hash_table == NULL) { 268ea8dc4b6Seschrock /* XXX - we should really return an error instead of assert */ 269ea8dc4b6Seschrock ASSERT(hsize > (1ULL << 10)); 270ea8dc4b6Seschrock hsize >>= 1; 271ea8dc4b6Seschrock goto retry; 272ea8dc4b6Seschrock } 273fa9e4066Sahrens 274fa9e4066Sahrens dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 275fa9e4066Sahrens sizeof (dmu_buf_impl_t), 276fa9e4066Sahrens 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 277fa9e4066Sahrens 278fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 279fa9e4066Sahrens mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 280fa9e4066Sahrens } 281fa9e4066Sahrens 282fa9e4066Sahrens void 283fa9e4066Sahrens dbuf_fini(void) 284fa9e4066Sahrens { 285fa9e4066Sahrens dbuf_hash_table_t *h = &dbuf_hash_table; 286fa9e4066Sahrens int i; 287fa9e4066Sahrens 288fa9e4066Sahrens for (i = 0; i < DBUF_MUTEXES; i++) 289fa9e4066Sahrens mutex_destroy(&h->hash_mutexes[i]); 290fa9e4066Sahrens kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 291fa9e4066Sahrens kmem_cache_destroy(dbuf_cache); 292fa9e4066Sahrens } 293fa9e4066Sahrens 294fa9e4066Sahrens /* 295fa9e4066Sahrens * Other stuff. 296fa9e4066Sahrens */ 297fa9e4066Sahrens 2989c9dc39aSek110237 #ifdef ZFS_DEBUG 299fa9e4066Sahrens static void 300fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db) 301fa9e4066Sahrens { 302744947dcSTom Erickson dnode_t *dn; 303b24ab676SJeff Bonwick dbuf_dirty_record_t *dr; 304fa9e4066Sahrens 305fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 306fa9e4066Sahrens 307fa9e4066Sahrens if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 308fa9e4066Sahrens return; 309fa9e4066Sahrens 310fa9e4066Sahrens ASSERT(db->db_objset != NULL); 311744947dcSTom Erickson DB_DNODE_ENTER(db); 312744947dcSTom Erickson dn = DB_DNODE(db); 313fa9e4066Sahrens if (dn == NULL) { 314fa9e4066Sahrens ASSERT(db->db_parent == NULL); 315fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 316fa9e4066Sahrens } else { 317fa9e4066Sahrens ASSERT3U(db->db.db_object, ==, dn->dn_object); 318fa9e4066Sahrens ASSERT3P(db->db_objset, ==, dn->dn_objset); 319fa9e4066Sahrens ASSERT3U(db->db_level, <, dn->dn_nlevels); 320744947dcSTom Erickson ASSERT(db->db_blkid == DMU_BONUS_BLKID || 321744947dcSTom Erickson db->db_blkid == DMU_SPILL_BLKID || 322744947dcSTom Erickson !list_is_empty(&dn->dn_dbufs)); 323fa9e4066Sahrens } 3240a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 325fa9e4066Sahrens ASSERT(dn != NULL); 3261934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 3270a586ceaSMark Shellenbaum ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 3280a586ceaSMark Shellenbaum } else if (db->db_blkid == DMU_SPILL_BLKID) { 3290a586ceaSMark Shellenbaum ASSERT(dn != NULL); 3300a586ceaSMark Shellenbaum ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 331fb09f5aaSMadhav Suresh ASSERT0(db->db.db_offset); 332fa9e4066Sahrens } else { 333fa9e4066Sahrens ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 334fa9e4066Sahrens } 335fa9e4066Sahrens 336b24ab676SJeff Bonwick for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 337b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 338b24ab676SJeff Bonwick 339b24ab676SJeff Bonwick for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 340b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 341b24ab676SJeff Bonwick 34288b7b0f2SMatthew Ahrens /* 34388b7b0f2SMatthew Ahrens * We can't assert that db_size matches dn_datablksz because it 34488b7b0f2SMatthew Ahrens * can be momentarily different when another thread is doing 34588b7b0f2SMatthew Ahrens * dnode_set_blksz(). 34688b7b0f2SMatthew Ahrens */ 34788b7b0f2SMatthew Ahrens if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 348b24ab676SJeff Bonwick dr = db->db_data_pending; 349fa9e4066Sahrens /* 35088b7b0f2SMatthew Ahrens * It should only be modified in syncing context, so 35188b7b0f2SMatthew Ahrens * make sure we only have one copy of the data. 352fa9e4066Sahrens */ 353c717a561Smaybee ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 354fa9e4066Sahrens } 355fa9e4066Sahrens 356fa9e4066Sahrens /* verify db->db_blkptr */ 357fa9e4066Sahrens if (db->db_blkptr) { 358fa9e4066Sahrens if (db->db_parent == dn->dn_dbuf) { 359fa9e4066Sahrens /* db is pointed to by the dnode */ 360fa9e4066Sahrens /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 36114843421SMatthew Ahrens if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 362fa9e4066Sahrens ASSERT(db->db_parent == NULL); 363fa9e4066Sahrens else 364fa9e4066Sahrens ASSERT(db->db_parent != NULL); 3650a586ceaSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID) 366fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 367fa9e4066Sahrens &dn->dn_phys->dn_blkptr[db->db_blkid]); 368fa9e4066Sahrens } else { 369fa9e4066Sahrens /* db is pointed to by an indirect block */ 370fa9e4066Sahrens int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 371fa9e4066Sahrens ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 372fa9e4066Sahrens ASSERT3U(db->db_parent->db.db_object, ==, 373fa9e4066Sahrens db->db.db_object); 374fa9e4066Sahrens /* 375fa9e4066Sahrens * dnode_grow_indblksz() can make this fail if we don't 376fa9e4066Sahrens * have the struct_rwlock. XXX indblksz no longer 377fa9e4066Sahrens * grows. safe to do this now? 378fa9e4066Sahrens */ 379744947dcSTom Erickson if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 380fa9e4066Sahrens ASSERT3P(db->db_blkptr, ==, 381fa9e4066Sahrens ((blkptr_t *)db->db_parent->db.db_data + 382fa9e4066Sahrens db->db_blkid % epb)); 383fa9e4066Sahrens } 384fa9e4066Sahrens } 385fa9e4066Sahrens } 386fa9e4066Sahrens if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 3873f9d6ad7SLin Ling (db->db_buf == NULL || db->db_buf->b_data) && 3880a586ceaSMark Shellenbaum db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 389fa9e4066Sahrens db->db_state != DB_FILL && !dn->dn_free_txg) { 390fa9e4066Sahrens /* 391fa9e4066Sahrens * If the blkptr isn't set but they have nonzero data, 392fa9e4066Sahrens * it had better be dirty, otherwise we'll lose that 393fa9e4066Sahrens * data when we evict this buffer. 394fa9e4066Sahrens */ 395fa9e4066Sahrens if (db->db_dirtycnt == 0) { 396fa9e4066Sahrens uint64_t *buf = db->db.db_data; 397fa9e4066Sahrens int i; 398fa9e4066Sahrens 399fa9e4066Sahrens for (i = 0; i < db->db.db_size >> 3; i++) { 400fa9e4066Sahrens ASSERT(buf[i] == 0); 401fa9e4066Sahrens } 402fa9e4066Sahrens } 403fa9e4066Sahrens } 404744947dcSTom Erickson DB_DNODE_EXIT(db); 405fa9e4066Sahrens } 4069c9dc39aSek110237 #endif 407fa9e4066Sahrens 408fa9e4066Sahrens static void 409fa9e4066Sahrens dbuf_update_data(dmu_buf_impl_t *db) 410fa9e4066Sahrens { 411fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 412c717a561Smaybee if (db->db_level == 0 && db->db_user_data_ptr_ptr) { 413fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 414c717a561Smaybee *db->db_user_data_ptr_ptr = db->db.db_data; 415fa9e4066Sahrens } 416fa9e4066Sahrens } 417fa9e4066Sahrens 418fa9e4066Sahrens static void 419fa9e4066Sahrens dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 420fa9e4066Sahrens { 421fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 422ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 423fa9e4066Sahrens db->db_buf = buf; 424ea8dc4b6Seschrock if (buf != NULL) { 425ea8dc4b6Seschrock ASSERT(buf->b_data != NULL); 426fa9e4066Sahrens db->db.db_data = buf->b_data; 427ea8dc4b6Seschrock if (!arc_released(buf)) 428ea8dc4b6Seschrock arc_set_callback(buf, dbuf_do_evict, db); 429fa9e4066Sahrens dbuf_update_data(db); 430ea8dc4b6Seschrock } else { 431ea8dc4b6Seschrock dbuf_evict_user(db); 432ea8dc4b6Seschrock db->db.db_data = NULL; 43382c9918fSTim Haley if (db->db_state != DB_NOFILL) 434ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 435ea8dc4b6Seschrock } 436fa9e4066Sahrens } 437fa9e4066Sahrens 438c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /* 439c242f9a0Schunli zhang - Sun Microsystems - Irvine United States * Loan out an arc_buf for read. Return the loaned arc_buf. 440c242f9a0Schunli zhang - Sun Microsystems - Irvine United States */ 441c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t * 442c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dbuf_loan_arcbuf(dmu_buf_impl_t *db) 443c242f9a0Schunli zhang - Sun Microsystems - Irvine United States { 444c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t *abuf; 445c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 446c242f9a0Schunli zhang - Sun Microsystems - Irvine United States mutex_enter(&db->db_mtx); 447c242f9a0Schunli zhang - Sun Microsystems - Irvine United States if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 448c242f9a0Schunli zhang - Sun Microsystems - Irvine United States int blksz = db->db.db_size; 449744947dcSTom Erickson spa_t *spa; 450744947dcSTom Erickson 451c242f9a0Schunli zhang - Sun Microsystems - Irvine United States mutex_exit(&db->db_mtx); 452744947dcSTom Erickson DB_GET_SPA(&spa, db); 453744947dcSTom Erickson abuf = arc_loan_buf(spa, blksz); 454c242f9a0Schunli zhang - Sun Microsystems - Irvine United States bcopy(db->db.db_data, abuf->b_data, blksz); 455c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } else { 456c242f9a0Schunli zhang - Sun Microsystems - Irvine United States abuf = db->db_buf; 457c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_loan_inuse_buf(abuf, db); 458c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dbuf_set_data(db, NULL); 459c242f9a0Schunli zhang - Sun Microsystems - Irvine United States mutex_exit(&db->db_mtx); 460c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 461c242f9a0Schunli zhang - Sun Microsystems - Irvine United States return (abuf); 462c242f9a0Schunli zhang - Sun Microsystems - Irvine United States } 463c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 464fa9e4066Sahrens uint64_t 465fa9e4066Sahrens dbuf_whichblock(dnode_t *dn, uint64_t offset) 466fa9e4066Sahrens { 467fa9e4066Sahrens if (dn->dn_datablkshift) { 468fa9e4066Sahrens return (offset >> dn->dn_datablkshift); 469fa9e4066Sahrens } else { 470fa9e4066Sahrens ASSERT3U(offset, <, dn->dn_datablksz); 471fa9e4066Sahrens return (0); 472fa9e4066Sahrens } 473fa9e4066Sahrens } 474fa9e4066Sahrens 475fa9e4066Sahrens static void 476fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 477fa9e4066Sahrens { 478fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 479fa9e4066Sahrens 480fa9e4066Sahrens mutex_enter(&db->db_mtx); 481fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_READ); 482fa9e4066Sahrens /* 483fa9e4066Sahrens * All reads are synchronous, so we must have a hold on the dbuf 484fa9e4066Sahrens */ 485fa9e4066Sahrens ASSERT(refcount_count(&db->db_holds) > 0); 486ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 487fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 488c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 489fa9e4066Sahrens /* we were freed in flight; disregard any error */ 490fa9e4066Sahrens arc_release(buf, db); 491fa9e4066Sahrens bzero(buf->b_data, db->db.db_size); 4926b4acc8bSahrens arc_buf_freeze(buf); 493c717a561Smaybee db->db_freed_in_flight = FALSE; 494fa9e4066Sahrens dbuf_set_data(db, buf); 495fa9e4066Sahrens db->db_state = DB_CACHED; 496fa9e4066Sahrens } else if (zio == NULL || zio->io_error == 0) { 497fa9e4066Sahrens dbuf_set_data(db, buf); 498fa9e4066Sahrens db->db_state = DB_CACHED; 499fa9e4066Sahrens } else { 5000a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 501fa9e4066Sahrens ASSERT3P(db->db_buf, ==, NULL); 502ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 503ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 504fa9e4066Sahrens } 505fa9e4066Sahrens cv_broadcast(&db->db_changed); 5063f9d6ad7SLin Ling dbuf_rele_and_unlock(db, NULL); 507fa9e4066Sahrens } 508fa9e4066Sahrens 509ea8dc4b6Seschrock static void 51013506d1eSmaybee dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 511fa9e4066Sahrens { 512744947dcSTom Erickson dnode_t *dn; 513744947dcSTom Erickson spa_t *spa; 514ea8dc4b6Seschrock zbookmark_t zb; 51513506d1eSmaybee uint32_t aflags = ARC_NOWAIT; 516088f3894Sahrens arc_buf_t *pbuf; 517fa9e4066Sahrens 518744947dcSTom Erickson DB_DNODE_ENTER(db); 519744947dcSTom Erickson dn = DB_DNODE(db); 520fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 521fa9e4066Sahrens /* We need the struct_rwlock to prevent db_blkptr from changing. */ 522088f3894Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 523ea8dc4b6Seschrock ASSERT(MUTEX_HELD(&db->db_mtx)); 524ea8dc4b6Seschrock ASSERT(db->db_state == DB_UNCACHED); 525ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 526fa9e4066Sahrens 5270a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 528cf04dda1SMark Maybee int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 5291934e92fSmaybee 5301934e92fSmaybee ASSERT3U(bonuslen, <=, db->db.db_size); 531ea8dc4b6Seschrock db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 5325a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 5331934e92fSmaybee if (bonuslen < DN_MAX_BONUSLEN) 534ea8dc4b6Seschrock bzero(db->db.db_data, DN_MAX_BONUSLEN); 535cf04dda1SMark Maybee if (bonuslen) 536cf04dda1SMark Maybee bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 537744947dcSTom Erickson DB_DNODE_EXIT(db); 538ea8dc4b6Seschrock dbuf_update_data(db); 539fa9e4066Sahrens db->db_state = DB_CACHED; 540fa9e4066Sahrens mutex_exit(&db->db_mtx); 541fa9e4066Sahrens return; 542fa9e4066Sahrens } 543fa9e4066Sahrens 5441c8564a7SMark Maybee /* 5451c8564a7SMark Maybee * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 5461c8564a7SMark Maybee * processes the delete record and clears the bp while we are waiting 5471c8564a7SMark Maybee * for the dn_mtx (resulting in a "no" from block_freed). 5481c8564a7SMark Maybee */ 549088f3894Sahrens if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 5501c8564a7SMark Maybee (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 5511c8564a7SMark Maybee BP_IS_HOLE(db->db_blkptr)))) { 552ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 553ad23a2dbSjohansen 554088f3894Sahrens dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa, 555ad23a2dbSjohansen db->db.db_size, db, type)); 556744947dcSTom Erickson DB_DNODE_EXIT(db); 557fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 558fa9e4066Sahrens db->db_state = DB_CACHED; 55913506d1eSmaybee *flags |= DB_RF_CACHED; 560fa9e4066Sahrens mutex_exit(&db->db_mtx); 561fa9e4066Sahrens return; 562fa9e4066Sahrens } 563fa9e4066Sahrens 564744947dcSTom Erickson spa = dn->dn_objset->os_spa; 565744947dcSTom Erickson DB_DNODE_EXIT(db); 566744947dcSTom Erickson 567fa9e4066Sahrens db->db_state = DB_READ; 568fa9e4066Sahrens mutex_exit(&db->db_mtx); 569fa9e4066Sahrens 5703baa08fcSek110237 if (DBUF_IS_L2CACHEABLE(db)) 5713baa08fcSek110237 aflags |= ARC_L2CACHE; 5723baa08fcSek110237 573b24ab676SJeff Bonwick SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 574b24ab676SJeff Bonwick db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 575b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 576ea8dc4b6Seschrock 577ea8dc4b6Seschrock dbuf_add_ref(db, NULL); 578fa9e4066Sahrens /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ 579088f3894Sahrens 580088f3894Sahrens if (db->db_parent) 581088f3894Sahrens pbuf = db->db_parent->db_buf; 582088f3894Sahrens else 583088f3894Sahrens pbuf = db->db_objset->os_phys_buf; 584088f3894Sahrens 585744947dcSTom Erickson (void) dsl_read(zio, spa, db->db_blkptr, pbuf, 586fa9e4066Sahrens dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 58713506d1eSmaybee (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 58813506d1eSmaybee &aflags, &zb); 58913506d1eSmaybee if (aflags & ARC_CACHED) 59013506d1eSmaybee *flags |= DB_RF_CACHED; 591fa9e4066Sahrens } 592fa9e4066Sahrens 593ea8dc4b6Seschrock int 594ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 595fa9e4066Sahrens { 596ea8dc4b6Seschrock int err = 0; 597ea8dc4b6Seschrock int havepzio = (zio != NULL); 59813506d1eSmaybee int prefetch; 599744947dcSTom Erickson dnode_t *dn; 600fa9e4066Sahrens 601fa9e4066Sahrens /* 602fa9e4066Sahrens * We don't have to hold the mutex to check db_state because it 603fa9e4066Sahrens * can't be freed while we have a hold on the buffer. 604fa9e4066Sahrens */ 605fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 606fa9e4066Sahrens 60782c9918fSTim Haley if (db->db_state == DB_NOFILL) 60882c9918fSTim Haley return (EIO); 60982c9918fSTim Haley 610744947dcSTom Erickson DB_DNODE_ENTER(db); 611744947dcSTom Erickson dn = DB_DNODE(db); 612fa9e4066Sahrens if ((flags & DB_RF_HAVESTRUCT) == 0) 613744947dcSTom Erickson rw_enter(&dn->dn_struct_rwlock, RW_READER); 614fa9e4066Sahrens 6150a586ceaSMark Shellenbaum prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 616744947dcSTom Erickson (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 6173baa08fcSek110237 DBUF_IS_CACHEABLE(db); 61813506d1eSmaybee 619fa9e4066Sahrens mutex_enter(&db->db_mtx); 620ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 621ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 62213506d1eSmaybee if (prefetch) 623744947dcSTom Erickson dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 62413506d1eSmaybee db->db.db_size, TRUE); 625ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 626744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 627744947dcSTom Erickson DB_DNODE_EXIT(db); 628ea8dc4b6Seschrock } else if (db->db_state == DB_UNCACHED) { 629744947dcSTom Erickson spa_t *spa = dn->dn_objset->os_spa; 630744947dcSTom Erickson 631744947dcSTom Erickson if (zio == NULL) 632744947dcSTom Erickson zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 63313506d1eSmaybee dbuf_read_impl(db, zio, &flags); 63413506d1eSmaybee 635ea8dc4b6Seschrock /* dbuf_read_impl has dropped db_mtx for us */ 636ea8dc4b6Seschrock 63713506d1eSmaybee if (prefetch) 638744947dcSTom Erickson dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 63913506d1eSmaybee db->db.db_size, flags & DB_RF_CACHED); 640ea8dc4b6Seschrock 641ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 642744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 643744947dcSTom Erickson DB_DNODE_EXIT(db); 644ea8dc4b6Seschrock 645ea8dc4b6Seschrock if (!havepzio) 646ea8dc4b6Seschrock err = zio_wait(zio); 647ea8dc4b6Seschrock } else { 64813506d1eSmaybee mutex_exit(&db->db_mtx); 64913506d1eSmaybee if (prefetch) 650744947dcSTom Erickson dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 65113506d1eSmaybee db->db.db_size, TRUE); 652ea8dc4b6Seschrock if ((flags & DB_RF_HAVESTRUCT) == 0) 653744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 654744947dcSTom Erickson DB_DNODE_EXIT(db); 65513506d1eSmaybee 65613506d1eSmaybee mutex_enter(&db->db_mtx); 657ea8dc4b6Seschrock if ((flags & DB_RF_NEVERWAIT) == 0) { 658ea8dc4b6Seschrock while (db->db_state == DB_READ || 659ea8dc4b6Seschrock db->db_state == DB_FILL) { 660fa9e4066Sahrens ASSERT(db->db_state == DB_READ || 661fa9e4066Sahrens (flags & DB_RF_HAVESTRUCT) == 0); 662fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 663fa9e4066Sahrens } 664ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED) 665ea8dc4b6Seschrock err = EIO; 666ea8dc4b6Seschrock } 667fa9e4066Sahrens mutex_exit(&db->db_mtx); 668fa9e4066Sahrens } 669fa9e4066Sahrens 670ea8dc4b6Seschrock ASSERT(err || havepzio || db->db_state == DB_CACHED); 671ea8dc4b6Seschrock return (err); 672fa9e4066Sahrens } 673fa9e4066Sahrens 674fa9e4066Sahrens static void 675fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db) 676fa9e4066Sahrens { 677fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 6780a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 679fa9e4066Sahrens mutex_enter(&db->db_mtx); 680fa9e4066Sahrens while (db->db_state == DB_READ || db->db_state == DB_FILL) 681fa9e4066Sahrens cv_wait(&db->db_changed, &db->db_mtx); 682fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 683ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 684744947dcSTom Erickson spa_t *spa; 685ad23a2dbSjohansen 686ea8dc4b6Seschrock ASSERT(db->db_buf == NULL); 687fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 688744947dcSTom Erickson DB_GET_SPA(&spa, db); 689744947dcSTom Erickson dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 690fa9e4066Sahrens db->db_state = DB_FILL; 69182c9918fSTim Haley } else if (db->db_state == DB_NOFILL) { 69282c9918fSTim Haley dbuf_set_data(db, NULL); 693fa9e4066Sahrens } else { 694fa9e4066Sahrens ASSERT3U(db->db_state, ==, DB_CACHED); 695fa9e4066Sahrens } 696fa9e4066Sahrens mutex_exit(&db->db_mtx); 697fa9e4066Sahrens } 698fa9e4066Sahrens 699fa9e4066Sahrens /* 700fa9e4066Sahrens * This is our just-in-time copy function. It makes a copy of 701fa9e4066Sahrens * buffers, that have been modified in a previous transaction 702fa9e4066Sahrens * group, before we modify them in the current active group. 703fa9e4066Sahrens * 704fa9e4066Sahrens * This function is used in two places: when we are dirtying a 705fa9e4066Sahrens * buffer for the first time in a txg, and when we are freeing 706fa9e4066Sahrens * a range in a dnode that includes this buffer. 707fa9e4066Sahrens * 708fa9e4066Sahrens * Note that when we are called from dbuf_free_range() we do 709fa9e4066Sahrens * not put a hold on the buffer, we just traverse the active 710fa9e4066Sahrens * dbuf list for the dnode. 711fa9e4066Sahrens */ 712fa9e4066Sahrens static void 713fa9e4066Sahrens dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 714fa9e4066Sahrens { 715c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 716fa9e4066Sahrens 717fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 718fa9e4066Sahrens ASSERT(db->db.db_data != NULL); 719c717a561Smaybee ASSERT(db->db_level == 0); 720c717a561Smaybee ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 721fa9e4066Sahrens 7224d31c452Smaybee if (dr == NULL || 7234d31c452Smaybee (dr->dt.dl.dr_data != 7240a586ceaSMark Shellenbaum ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 725fa9e4066Sahrens return; 726fa9e4066Sahrens 727fa9e4066Sahrens /* 728c717a561Smaybee * If the last dirty record for this dbuf has not yet synced 729c717a561Smaybee * and its referencing the dbuf data, either: 730c717a561Smaybee * reset the reference to point to a new copy, 731c717a561Smaybee * or (if there a no active holders) 732c717a561Smaybee * just null out the current db_data pointer. 733fa9e4066Sahrens */ 734c717a561Smaybee ASSERT(dr->dr_txg >= txg - 2); 7350a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 736c717a561Smaybee /* Note that the data bufs here are zio_bufs */ 737c717a561Smaybee dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 7385a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 739c717a561Smaybee bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 740c717a561Smaybee } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 741ea8dc4b6Seschrock int size = db->db.db_size; 742c717a561Smaybee arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 743744947dcSTom Erickson spa_t *spa; 744744947dcSTom Erickson 745744947dcSTom Erickson DB_GET_SPA(&spa, db); 746744947dcSTom Erickson dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 747c717a561Smaybee bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 748fa9e4066Sahrens } else { 749ea8dc4b6Seschrock dbuf_set_data(db, NULL); 750fa9e4066Sahrens } 751fa9e4066Sahrens } 752ea8dc4b6Seschrock 753fa9e4066Sahrens void 754c717a561Smaybee dbuf_unoverride(dbuf_dirty_record_t *dr) 755fa9e4066Sahrens { 756c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 757b24ab676SJeff Bonwick blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 758c717a561Smaybee uint64_t txg = dr->dr_txg; 759c5c6ffa0Smaybee 760c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 761c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 762c717a561Smaybee ASSERT(db->db_level == 0); 763c717a561Smaybee 7640a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID || 765c717a561Smaybee dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 766c717a561Smaybee return; 767c717a561Smaybee 768b24ab676SJeff Bonwick ASSERT(db->db_data_pending != dr); 769b24ab676SJeff Bonwick 770fa9e4066Sahrens /* free this block */ 771*80901aeaSGeorge Wilson if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) { 772744947dcSTom Erickson spa_t *spa; 773b24ab676SJeff Bonwick 774744947dcSTom Erickson DB_GET_SPA(&spa, db); 775744947dcSTom Erickson zio_free(spa, txg, bp); 776744947dcSTom Erickson } 777c717a561Smaybee dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 778*80901aeaSGeorge Wilson dr->dt.dl.dr_nopwrite = B_FALSE; 779*80901aeaSGeorge Wilson 7806b4acc8bSahrens /* 7816b4acc8bSahrens * Release the already-written buffer, so we leave it in 7826b4acc8bSahrens * a consistent dirty state. Note that all callers are 7836b4acc8bSahrens * modifying the buffer, so they will immediately do 7846b4acc8bSahrens * another (redundant) arc_release(). Therefore, leave 7856b4acc8bSahrens * the buf thawed to save the effort of freezing & 7866b4acc8bSahrens * immediately re-thawing it. 7876b4acc8bSahrens */ 788c717a561Smaybee arc_release(dr->dt.dl.dr_data, db); 789fa9e4066Sahrens } 790fa9e4066Sahrens 791cdb0ab79Smaybee /* 792cdb0ab79Smaybee * Evict (if its unreferenced) or clear (if its referenced) any level-0 793cdb0ab79Smaybee * data blocks in the free range, so that any future readers will find 794cdb0ab79Smaybee * empty blocks. Also, if we happen accross any level-1 dbufs in the 795cdb0ab79Smaybee * range that have not already been marked dirty, mark them dirty so 796cdb0ab79Smaybee * they stay in memory. 797cdb0ab79Smaybee */ 798fa9e4066Sahrens void 799cdb0ab79Smaybee dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) 800fa9e4066Sahrens { 801fa9e4066Sahrens dmu_buf_impl_t *db, *db_next; 802fa9e4066Sahrens uint64_t txg = tx->tx_txg; 803cdb0ab79Smaybee int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 804cdb0ab79Smaybee uint64_t first_l1 = start >> epbs; 805cdb0ab79Smaybee uint64_t last_l1 = end >> epbs; 806fa9e4066Sahrens 8070a586ceaSMark Shellenbaum if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) { 808cdb0ab79Smaybee end = dn->dn_maxblkid; 809cdb0ab79Smaybee last_l1 = end >> epbs; 810cdb0ab79Smaybee } 811cdb0ab79Smaybee dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); 812fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 813fa9e4066Sahrens for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 814fa9e4066Sahrens db_next = list_next(&dn->dn_dbufs, db); 8150a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 816cdb0ab79Smaybee 817cdb0ab79Smaybee if (db->db_level == 1 && 818cdb0ab79Smaybee db->db_blkid >= first_l1 && db->db_blkid <= last_l1) { 819cdb0ab79Smaybee mutex_enter(&db->db_mtx); 820cdb0ab79Smaybee if (db->db_last_dirty && 821cdb0ab79Smaybee db->db_last_dirty->dr_txg < txg) { 822cdb0ab79Smaybee dbuf_add_ref(db, FTAG); 823cdb0ab79Smaybee mutex_exit(&db->db_mtx); 824cdb0ab79Smaybee dbuf_will_dirty(db, tx); 825cdb0ab79Smaybee dbuf_rele(db, FTAG); 826cdb0ab79Smaybee } else { 827cdb0ab79Smaybee mutex_exit(&db->db_mtx); 828cdb0ab79Smaybee } 829cdb0ab79Smaybee } 830cdb0ab79Smaybee 831ea8dc4b6Seschrock if (db->db_level != 0) 832fa9e4066Sahrens continue; 833fa9e4066Sahrens dprintf_dbuf(db, "found buf %s\n", ""); 834cdb0ab79Smaybee if (db->db_blkid < start || db->db_blkid > end) 835fa9e4066Sahrens continue; 836fa9e4066Sahrens 837fa9e4066Sahrens /* found a level 0 buffer in the range */ 838fa9e4066Sahrens if (dbuf_undirty(db, tx)) 839fa9e4066Sahrens continue; 840fa9e4066Sahrens 841fa9e4066Sahrens mutex_enter(&db->db_mtx); 842ea8dc4b6Seschrock if (db->db_state == DB_UNCACHED || 84382c9918fSTim Haley db->db_state == DB_NOFILL || 844ea8dc4b6Seschrock db->db_state == DB_EVICTING) { 845fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 846fa9e4066Sahrens mutex_exit(&db->db_mtx); 847fa9e4066Sahrens continue; 848fa9e4066Sahrens } 849c543ec06Sahrens if (db->db_state == DB_READ || db->db_state == DB_FILL) { 850c543ec06Sahrens /* will be handled in dbuf_read_done or dbuf_rele */ 851c717a561Smaybee db->db_freed_in_flight = TRUE; 852fa9e4066Sahrens mutex_exit(&db->db_mtx); 853fa9e4066Sahrens continue; 854fa9e4066Sahrens } 855ea8dc4b6Seschrock if (refcount_count(&db->db_holds) == 0) { 856ea8dc4b6Seschrock ASSERT(db->db_buf); 857ea8dc4b6Seschrock dbuf_clear(db); 858ea8dc4b6Seschrock continue; 859ea8dc4b6Seschrock } 860c717a561Smaybee /* The dbuf is referenced */ 861fa9e4066Sahrens 862c717a561Smaybee if (db->db_last_dirty != NULL) { 863c717a561Smaybee dbuf_dirty_record_t *dr = db->db_last_dirty; 864c717a561Smaybee 865c717a561Smaybee if (dr->dr_txg == txg) { 866ea8dc4b6Seschrock /* 867c717a561Smaybee * This buffer is "in-use", re-adjust the file 868c717a561Smaybee * size to reflect that this buffer may 869c717a561Smaybee * contain new data when we sync. 870ea8dc4b6Seschrock */ 87106e0070dSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID && 87206e0070dSMark Shellenbaum db->db_blkid > dn->dn_maxblkid) 87344eda4d7Smaybee dn->dn_maxblkid = db->db_blkid; 874c717a561Smaybee dbuf_unoverride(dr); 875c717a561Smaybee } else { 876c717a561Smaybee /* 877c717a561Smaybee * This dbuf is not dirty in the open context. 878c717a561Smaybee * Either uncache it (if its not referenced in 879c717a561Smaybee * the open context) or reset its contents to 880c717a561Smaybee * empty. 881c717a561Smaybee */ 882c717a561Smaybee dbuf_fix_old_data(db, txg); 88344eda4d7Smaybee } 884c717a561Smaybee } 885c717a561Smaybee /* clear the contents if its cached */ 886ea8dc4b6Seschrock if (db->db_state == DB_CACHED) { 887ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 888fa9e4066Sahrens arc_release(db->db_buf, db); 889fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 8906b4acc8bSahrens arc_buf_freeze(db->db_buf); 891fa9e4066Sahrens } 892ea8dc4b6Seschrock 893fa9e4066Sahrens mutex_exit(&db->db_mtx); 894fa9e4066Sahrens } 895fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 896fa9e4066Sahrens } 897fa9e4066Sahrens 898fa9e4066Sahrens static int 8991934e92fSmaybee dbuf_block_freeable(dmu_buf_impl_t *db) 900fa9e4066Sahrens { 901fa9e4066Sahrens dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 902fa9e4066Sahrens uint64_t birth_txg = 0; 903fa9e4066Sahrens 904fa9e4066Sahrens /* 905fa9e4066Sahrens * We don't need any locking to protect db_blkptr: 906c717a561Smaybee * If it's syncing, then db_last_dirty will be set 907c717a561Smaybee * so we'll ignore db_blkptr. 908fa9e4066Sahrens */ 909c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 910c717a561Smaybee if (db->db_last_dirty) 911c717a561Smaybee birth_txg = db->db_last_dirty->dr_txg; 912fa9e4066Sahrens else if (db->db_blkptr) 913fa9e4066Sahrens birth_txg = db->db_blkptr->blk_birth; 914fa9e4066Sahrens 915837b568bSGeorge Wilson /* 916837b568bSGeorge Wilson * If we don't exist or are in a snapshot, we can't be freed. 917837b568bSGeorge Wilson * Don't pass the bp to dsl_dataset_block_freeable() since we 918837b568bSGeorge Wilson * are holding the db_mtx lock and might deadlock if we are 919837b568bSGeorge Wilson * prefetching a dedup-ed block. 920837b568bSGeorge Wilson */ 921fa9e4066Sahrens if (birth_txg) 9221934e92fSmaybee return (ds == NULL || 923837b568bSGeorge Wilson dsl_dataset_block_freeable(ds, NULL, birth_txg)); 924fa9e4066Sahrens else 9251934e92fSmaybee return (FALSE); 926fa9e4066Sahrens } 927fa9e4066Sahrens 928fa9e4066Sahrens void 929fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 930fa9e4066Sahrens { 931fa9e4066Sahrens arc_buf_t *buf, *obuf; 932fa9e4066Sahrens int osize = db->db.db_size; 933ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 934744947dcSTom Erickson dnode_t *dn; 935fa9e4066Sahrens 9360a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 937ea8dc4b6Seschrock 938744947dcSTom Erickson DB_DNODE_ENTER(db); 939744947dcSTom Erickson dn = DB_DNODE(db); 940744947dcSTom Erickson 941fa9e4066Sahrens /* XXX does *this* func really need the lock? */ 942744947dcSTom Erickson ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 943fa9e4066Sahrens 944fa9e4066Sahrens /* 945fa9e4066Sahrens * This call to dbuf_will_dirty() with the dn_struct_rwlock held 946fa9e4066Sahrens * is OK, because there can be no other references to the db 947fa9e4066Sahrens * when we are changing its size, so no concurrent DB_FILL can 948fa9e4066Sahrens * be happening. 949fa9e4066Sahrens */ 950ea8dc4b6Seschrock /* 951ea8dc4b6Seschrock * XXX we should be doing a dbuf_read, checking the return 952ea8dc4b6Seschrock * value and returning that up to our callers 953ea8dc4b6Seschrock */ 954fa9e4066Sahrens dbuf_will_dirty(db, tx); 955fa9e4066Sahrens 956fa9e4066Sahrens /* create the data buffer for the new block */ 957744947dcSTom Erickson buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 958fa9e4066Sahrens 959fa9e4066Sahrens /* copy old block data to the new block */ 960fa9e4066Sahrens obuf = db->db_buf; 961f65e61c0Sahrens bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 962fa9e4066Sahrens /* zero the remainder */ 963f65e61c0Sahrens if (size > osize) 964fa9e4066Sahrens bzero((uint8_t *)buf->b_data + osize, size - osize); 965fa9e4066Sahrens 966fa9e4066Sahrens mutex_enter(&db->db_mtx); 967fa9e4066Sahrens dbuf_set_data(db, buf); 968ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(obuf, db) == 1); 969fa9e4066Sahrens db->db.db_size = size; 970fa9e4066Sahrens 971c717a561Smaybee if (db->db_level == 0) { 972c717a561Smaybee ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 973c717a561Smaybee db->db_last_dirty->dt.dl.dr_data = buf; 974c717a561Smaybee } 975fa9e4066Sahrens mutex_exit(&db->db_mtx); 976fa9e4066Sahrens 977744947dcSTom Erickson dnode_willuse_space(dn, size-osize, tx); 978744947dcSTom Erickson DB_DNODE_EXIT(db); 979fa9e4066Sahrens } 980fa9e4066Sahrens 9813f9d6ad7SLin Ling void 9823f9d6ad7SLin Ling dbuf_release_bp(dmu_buf_impl_t *db) 9833f9d6ad7SLin Ling { 984744947dcSTom Erickson objset_t *os; 9853f9d6ad7SLin Ling zbookmark_t zb; 9863f9d6ad7SLin Ling 987744947dcSTom Erickson DB_GET_OBJSET(&os, db); 9883f9d6ad7SLin Ling ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 9893f9d6ad7SLin Ling ASSERT(arc_released(os->os_phys_buf) || 9903f9d6ad7SLin Ling list_link_active(&os->os_dsl_dataset->ds_synced_link)); 9913f9d6ad7SLin Ling ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 9923f9d6ad7SLin Ling 9933f9d6ad7SLin Ling zb.zb_objset = os->os_dsl_dataset ? 9943f9d6ad7SLin Ling os->os_dsl_dataset->ds_object : 0; 9953f9d6ad7SLin Ling zb.zb_object = db->db.db_object; 9963f9d6ad7SLin Ling zb.zb_level = db->db_level; 9973f9d6ad7SLin Ling zb.zb_blkid = db->db_blkid; 9983f9d6ad7SLin Ling (void) arc_release_bp(db->db_buf, db, 9993f9d6ad7SLin Ling db->db_blkptr, os->os_spa, &zb); 10003f9d6ad7SLin Ling } 10013f9d6ad7SLin Ling 1002c717a561Smaybee dbuf_dirty_record_t * 1003fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1004fa9e4066Sahrens { 1005744947dcSTom Erickson dnode_t *dn; 1006744947dcSTom Erickson objset_t *os; 1007c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 1008fa9e4066Sahrens int drop_struct_lock = FALSE; 1009d3469faaSMark Maybee boolean_t do_free_accounting = B_FALSE; 1010fa9e4066Sahrens int txgoff = tx->tx_txg & TXG_MASK; 1011fa9e4066Sahrens 1012fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1013fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 10149c9dc39aSek110237 DMU_TX_DIRTY_BUF(tx, db); 1015fa9e4066Sahrens 1016744947dcSTom Erickson DB_DNODE_ENTER(db); 1017744947dcSTom Erickson dn = DB_DNODE(db); 1018fa9e4066Sahrens /* 1019fa9e4066Sahrens * Shouldn't dirty a regular buffer in syncing context. Private 1020fa9e4066Sahrens * objects may be dirtied in syncing context, but only if they 1021fa9e4066Sahrens * were already pre-dirtied in open context. 1022fa9e4066Sahrens */ 1023c717a561Smaybee ASSERT(!dmu_tx_is_syncing(tx) || 1024c717a561Smaybee BP_IS_HOLE(dn->dn_objset->os_rootbp) || 102514843421SMatthew Ahrens DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 102614843421SMatthew Ahrens dn->dn_objset->os_dsl_dataset == NULL); 1027fa9e4066Sahrens /* 1028fa9e4066Sahrens * We make this assert for private objects as well, but after we 1029fa9e4066Sahrens * check if we're already dirty. They are allowed to re-dirty 1030fa9e4066Sahrens * in syncing context. 1031fa9e4066Sahrens */ 1032ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1033c717a561Smaybee dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1034fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1035fa9e4066Sahrens 1036fa9e4066Sahrens mutex_enter(&db->db_mtx); 1037fa9e4066Sahrens /* 1038c717a561Smaybee * XXX make this true for indirects too? The problem is that 1039c717a561Smaybee * transactions created with dmu_tx_create_assigned() from 1040c717a561Smaybee * syncing context don't bother holding ahead. 1041fa9e4066Sahrens */ 1042c717a561Smaybee ASSERT(db->db_level != 0 || 104382c9918fSTim Haley db->db_state == DB_CACHED || db->db_state == DB_FILL || 104482c9918fSTim Haley db->db_state == DB_NOFILL); 1045fa9e4066Sahrens 1046fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1047fa9e4066Sahrens /* 1048fa9e4066Sahrens * Don't set dirtyctx to SYNC if we're just modifying this as we 1049fa9e4066Sahrens * initialize the objset. 1050fa9e4066Sahrens */ 1051fa9e4066Sahrens if (dn->dn_dirtyctx == DN_UNDIRTIED && 1052c717a561Smaybee !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1053fa9e4066Sahrens dn->dn_dirtyctx = 1054fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1055fa9e4066Sahrens ASSERT(dn->dn_dirtyctx_firstset == NULL); 1056fa9e4066Sahrens dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1057fa9e4066Sahrens } 1058fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1059fa9e4066Sahrens 10600a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) 10610a586ceaSMark Shellenbaum dn->dn_have_spill = B_TRUE; 10620a586ceaSMark Shellenbaum 1063fa9e4066Sahrens /* 1064fa9e4066Sahrens * If this buffer is already dirty, we're done. 1065fa9e4066Sahrens */ 1066c717a561Smaybee drp = &db->db_last_dirty; 1067c717a561Smaybee ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1068c717a561Smaybee db->db.db_object == DMU_META_DNODE_OBJECT); 10697e2186e3Sbonwick while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 10707e2186e3Sbonwick drp = &dr->dr_next; 10717e2186e3Sbonwick if (dr && dr->dr_txg == tx->tx_txg) { 1072744947dcSTom Erickson DB_DNODE_EXIT(db); 1073744947dcSTom Erickson 10740a586ceaSMark Shellenbaum if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1075c717a561Smaybee /* 1076c717a561Smaybee * If this buffer has already been written out, 1077c717a561Smaybee * we now need to reset its state. 1078c717a561Smaybee */ 10797e2186e3Sbonwick dbuf_unoverride(dr); 1080b24ab676SJeff Bonwick if (db->db.db_object != DMU_META_DNODE_OBJECT && 1081b24ab676SJeff Bonwick db->db_state != DB_NOFILL) 10826b4acc8bSahrens arc_buf_thaw(db->db_buf); 1083c717a561Smaybee } 1084fa9e4066Sahrens mutex_exit(&db->db_mtx); 10857e2186e3Sbonwick return (dr); 1086fa9e4066Sahrens } 1087fa9e4066Sahrens 1088fa9e4066Sahrens /* 1089fa9e4066Sahrens * Only valid if not already dirty. 1090fa9e4066Sahrens */ 109114843421SMatthew Ahrens ASSERT(dn->dn_object == 0 || 109214843421SMatthew Ahrens dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1093fa9e4066Sahrens (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1094fa9e4066Sahrens 1095fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, db->db_level); 1096fa9e4066Sahrens ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1097fa9e4066Sahrens dn->dn_phys->dn_nlevels > db->db_level || 1098fa9e4066Sahrens dn->dn_next_nlevels[txgoff] > db->db_level || 1099fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1100fa9e4066Sahrens dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1101fa9e4066Sahrens 1102fa9e4066Sahrens /* 1103fa9e4066Sahrens * We should only be dirtying in syncing context if it's the 110414843421SMatthew Ahrens * mos or we're initializing the os or it's a special object. 110514843421SMatthew Ahrens * However, we are allowed to dirty in syncing context provided 110614843421SMatthew Ahrens * we already dirtied it in open context. Hence we must make 110714843421SMatthew Ahrens * this assertion only if we're not already dirty. 1108fa9e4066Sahrens */ 1109744947dcSTom Erickson os = dn->dn_objset; 111014843421SMatthew Ahrens ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 111114843421SMatthew Ahrens os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1112fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1113fa9e4066Sahrens 1114fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1115fa9e4066Sahrens 11160a586ceaSMark Shellenbaum if (db->db_blkid != DMU_BONUS_BLKID) { 11171934e92fSmaybee /* 11181934e92fSmaybee * Update the accounting. 1119d3469faaSMark Maybee * Note: we delay "free accounting" until after we drop 1120d3469faaSMark Maybee * the db_mtx. This keeps us from grabbing other locks 1121b24ab676SJeff Bonwick * (and possibly deadlocking) in bp_get_dsize() while 1122d3469faaSMark Maybee * also holding the db_mtx. 11231934e92fSmaybee */ 11241934e92fSmaybee dnode_willuse_space(dn, db->db.db_size, tx); 1125d3469faaSMark Maybee do_free_accounting = dbuf_block_freeable(db); 11261934e92fSmaybee } 11271934e92fSmaybee 1128ea8dc4b6Seschrock /* 1129ea8dc4b6Seschrock * If this buffer is dirty in an old transaction group we need 1130ea8dc4b6Seschrock * to make a copy of it so that the changes we make in this 1131ea8dc4b6Seschrock * transaction group won't leak out when we sync the older txg. 1132ea8dc4b6Seschrock */ 1133c717a561Smaybee dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1134c717a561Smaybee if (db->db_level == 0) { 1135c717a561Smaybee void *data_old = db->db_buf; 1136c717a561Smaybee 113782c9918fSTim Haley if (db->db_state != DB_NOFILL) { 11380a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 1139c717a561Smaybee dbuf_fix_old_data(db, tx->tx_txg); 1140c717a561Smaybee data_old = db->db.db_data; 1141c717a561Smaybee } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1142fa9e4066Sahrens /* 114382c9918fSTim Haley * Release the data buffer from the cache so 114482c9918fSTim Haley * that we can modify it without impacting 114582c9918fSTim Haley * possible other users of this cached data 114682c9918fSTim Haley * block. Note that indirect blocks and 114782c9918fSTim Haley * private objects are not released until the 114882c9918fSTim Haley * syncing state (since they are only modified 114982c9918fSTim Haley * then). 1150fa9e4066Sahrens */ 1151fa9e4066Sahrens arc_release(db->db_buf, db); 1152fa9e4066Sahrens dbuf_fix_old_data(db, tx->tx_txg); 1153c717a561Smaybee data_old = db->db_buf; 1154fa9e4066Sahrens } 1155c717a561Smaybee ASSERT(data_old != NULL); 115682c9918fSTim Haley } 1157c717a561Smaybee dr->dt.dl.dr_data = data_old; 1158c717a561Smaybee } else { 1159c717a561Smaybee mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1160c717a561Smaybee list_create(&dr->dt.di.dr_children, 1161c717a561Smaybee sizeof (dbuf_dirty_record_t), 1162c717a561Smaybee offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1163fa9e4066Sahrens } 1164c717a561Smaybee dr->dr_dbuf = db; 1165c717a561Smaybee dr->dr_txg = tx->tx_txg; 1166c717a561Smaybee dr->dr_next = *drp; 1167c717a561Smaybee *drp = dr; 1168fa9e4066Sahrens 1169fa9e4066Sahrens /* 1170fa9e4066Sahrens * We could have been freed_in_flight between the dbuf_noread 1171fa9e4066Sahrens * and dbuf_dirty. We win, as though the dbuf_noread() had 1172fa9e4066Sahrens * happened after the free. 1173fa9e4066Sahrens */ 11740a586ceaSMark Shellenbaum if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 11750a586ceaSMark Shellenbaum db->db_blkid != DMU_SPILL_BLKID) { 1176c717a561Smaybee mutex_enter(&dn->dn_mtx); 1177fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1178fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1179c717a561Smaybee db->db_freed_in_flight = FALSE; 1180c717a561Smaybee } 1181fa9e4066Sahrens 1182fa9e4066Sahrens /* 1183fa9e4066Sahrens * This buffer is now part of this txg 1184fa9e4066Sahrens */ 1185fa9e4066Sahrens dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1186fa9e4066Sahrens db->db_dirtycnt += 1; 1187fa9e4066Sahrens ASSERT3U(db->db_dirtycnt, <=, 3); 1188fa9e4066Sahrens 1189fa9e4066Sahrens mutex_exit(&db->db_mtx); 1190fa9e4066Sahrens 11910a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID || 11920a586ceaSMark Shellenbaum db->db_blkid == DMU_SPILL_BLKID) { 1193c717a561Smaybee mutex_enter(&dn->dn_mtx); 1194c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1195c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1196c717a561Smaybee mutex_exit(&dn->dn_mtx); 1197fa9e4066Sahrens dnode_setdirty(dn, tx); 1198744947dcSTom Erickson DB_DNODE_EXIT(db); 1199c717a561Smaybee return (dr); 1200d3469faaSMark Maybee } else if (do_free_accounting) { 1201d3469faaSMark Maybee blkptr_t *bp = db->db_blkptr; 1202d3469faaSMark Maybee int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1203b24ab676SJeff Bonwick bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1204d3469faaSMark Maybee /* 1205d3469faaSMark Maybee * This is only a guess -- if the dbuf is dirty 1206d3469faaSMark Maybee * in a previous txg, we don't know how much 1207d3469faaSMark Maybee * space it will use on disk yet. We should 1208d3469faaSMark Maybee * really have the struct_rwlock to access 1209d3469faaSMark Maybee * db_blkptr, but since this is just a guess, 1210d3469faaSMark Maybee * it's OK if we get an odd answer. 1211d3469faaSMark Maybee */ 1212837b568bSGeorge Wilson ddt_prefetch(os->os_spa, bp); 1213d3469faaSMark Maybee dnode_willuse_space(dn, -willfree, tx); 1214fa9e4066Sahrens } 1215fa9e4066Sahrens 1216fa9e4066Sahrens if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1217fa9e4066Sahrens rw_enter(&dn->dn_struct_rwlock, RW_READER); 1218fa9e4066Sahrens drop_struct_lock = TRUE; 1219fa9e4066Sahrens } 1220fa9e4066Sahrens 12218346f03fSJonathan W Adams if (db->db_level == 0) { 12228346f03fSJonathan W Adams dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 12238346f03fSJonathan W Adams ASSERT(dn->dn_maxblkid >= db->db_blkid); 12248346f03fSJonathan W Adams } 12258346f03fSJonathan W Adams 122644eda4d7Smaybee if (db->db_level+1 < dn->dn_nlevels) { 1227c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 1228c717a561Smaybee dbuf_dirty_record_t *di; 1229c717a561Smaybee int parent_held = FALSE; 1230c717a561Smaybee 1231c717a561Smaybee if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1232fa9e4066Sahrens int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1233c717a561Smaybee 1234fa9e4066Sahrens parent = dbuf_hold_level(dn, db->db_level+1, 1235fa9e4066Sahrens db->db_blkid >> epbs, FTAG); 123601025c89SJohn Harres ASSERT(parent != NULL); 1237c717a561Smaybee parent_held = TRUE; 1238c717a561Smaybee } 1239fa9e4066Sahrens if (drop_struct_lock) 1240fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1241c717a561Smaybee ASSERT3U(db->db_level+1, ==, parent->db_level); 1242c717a561Smaybee di = dbuf_dirty(parent, tx); 1243c717a561Smaybee if (parent_held) 1244ea8dc4b6Seschrock dbuf_rele(parent, FTAG); 1245c717a561Smaybee 1246c717a561Smaybee mutex_enter(&db->db_mtx); 1247c717a561Smaybee /* possible race with dbuf_undirty() */ 1248c717a561Smaybee if (db->db_last_dirty == dr || 1249c717a561Smaybee dn->dn_object == DMU_META_DNODE_OBJECT) { 1250c717a561Smaybee mutex_enter(&di->dt.di.dr_mtx); 1251c717a561Smaybee ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1252c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1253c717a561Smaybee list_insert_tail(&di->dt.di.dr_children, dr); 1254c717a561Smaybee mutex_exit(&di->dt.di.dr_mtx); 1255c717a561Smaybee dr->dr_parent = di; 1256c717a561Smaybee } 1257c717a561Smaybee mutex_exit(&db->db_mtx); 1258fa9e4066Sahrens } else { 1259c717a561Smaybee ASSERT(db->db_level+1 == dn->dn_nlevels); 1260c717a561Smaybee ASSERT(db->db_blkid < dn->dn_nblkptr); 1261744947dcSTom Erickson ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1262c717a561Smaybee mutex_enter(&dn->dn_mtx); 1263c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 1264c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1265c717a561Smaybee mutex_exit(&dn->dn_mtx); 1266fa9e4066Sahrens if (drop_struct_lock) 1267fa9e4066Sahrens rw_exit(&dn->dn_struct_rwlock); 1268fa9e4066Sahrens } 1269fa9e4066Sahrens 1270fa9e4066Sahrens dnode_setdirty(dn, tx); 1271744947dcSTom Erickson DB_DNODE_EXIT(db); 1272c717a561Smaybee return (dr); 1273fa9e4066Sahrens } 1274fa9e4066Sahrens 1275fa9e4066Sahrens static int 1276fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1277fa9e4066Sahrens { 1278744947dcSTom Erickson dnode_t *dn; 1279c717a561Smaybee uint64_t txg = tx->tx_txg; 128017f17c2dSbonwick dbuf_dirty_record_t *dr, **drp; 1281fa9e4066Sahrens 1282c717a561Smaybee ASSERT(txg != 0); 12830a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1284fa9e4066Sahrens 1285fa9e4066Sahrens mutex_enter(&db->db_mtx); 1286fa9e4066Sahrens /* 1287fa9e4066Sahrens * If this buffer is not dirty, we're done. 1288fa9e4066Sahrens */ 128917f17c2dSbonwick for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1290c717a561Smaybee if (dr->dr_txg <= txg) 1291c717a561Smaybee break; 1292c717a561Smaybee if (dr == NULL || dr->dr_txg < txg) { 1293fa9e4066Sahrens mutex_exit(&db->db_mtx); 1294fa9e4066Sahrens return (0); 1295fa9e4066Sahrens } 1296c717a561Smaybee ASSERT(dr->dr_txg == txg); 1297b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 1298fa9e4066Sahrens 1299744947dcSTom Erickson DB_DNODE_ENTER(db); 1300744947dcSTom Erickson dn = DB_DNODE(db); 1301744947dcSTom Erickson 1302fa9e4066Sahrens /* 1303fa9e4066Sahrens * If this buffer is currently held, we cannot undirty 1304fa9e4066Sahrens * it, since one of the current holders may be in the 1305fa9e4066Sahrens * middle of an update. Note that users of dbuf_undirty() 1306fa9e4066Sahrens * should not place a hold on the dbuf before the call. 13073f2366c2SGordon Ross * Also note: we can get here with a spill block, so 13083f2366c2SGordon Ross * test for that similar to how dbuf_dirty does. 1309fa9e4066Sahrens */ 1310fa9e4066Sahrens if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1311fa9e4066Sahrens mutex_exit(&db->db_mtx); 131244eda4d7Smaybee /* Make sure we don't toss this buffer at sync phase */ 13133f2366c2SGordon Ross if (db->db_blkid != DMU_SPILL_BLKID) { 1314fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1315fa9e4066Sahrens dnode_clear_range(dn, db->db_blkid, 1, tx); 1316fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 13173f2366c2SGordon Ross } 1318744947dcSTom Erickson DB_DNODE_EXIT(db); 1319fa9e4066Sahrens return (0); 1320fa9e4066Sahrens } 1321fa9e4066Sahrens 1322fa9e4066Sahrens dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1323fa9e4066Sahrens 1324fa9e4066Sahrens ASSERT(db->db.db_size != 0); 1325fa9e4066Sahrens 1326fa9e4066Sahrens /* XXX would be nice to fix up dn_towrite_space[] */ 1327fa9e4066Sahrens 132817f17c2dSbonwick *drp = dr->dr_next; 1329c717a561Smaybee 13303f2366c2SGordon Ross /* 13313f2366c2SGordon Ross * Note that there are three places in dbuf_dirty() 13323f2366c2SGordon Ross * where this dirty record may be put on a list. 13333f2366c2SGordon Ross * Make sure to do a list_remove corresponding to 13343f2366c2SGordon Ross * every one of those list_insert calls. 13353f2366c2SGordon Ross */ 1336c717a561Smaybee if (dr->dr_parent) { 1337c717a561Smaybee mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1338c717a561Smaybee list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1339c717a561Smaybee mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 13403f2366c2SGordon Ross } else if (db->db_blkid == DMU_SPILL_BLKID || 13413f2366c2SGordon Ross db->db_level+1 == dn->dn_nlevels) { 1342cdb0ab79Smaybee ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1343fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 1344c717a561Smaybee list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1345fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 1346c717a561Smaybee } 1347744947dcSTom Erickson DB_DNODE_EXIT(db); 1348c717a561Smaybee 1349c717a561Smaybee if (db->db_level == 0) { 135082c9918fSTim Haley if (db->db_state != DB_NOFILL) { 1351c717a561Smaybee dbuf_unoverride(dr); 1352c717a561Smaybee 1353c717a561Smaybee ASSERT(db->db_buf != NULL); 1354c717a561Smaybee ASSERT(dr->dt.dl.dr_data != NULL); 1355c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 135682c9918fSTim Haley VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 135782c9918fSTim Haley db) == 1); 135882c9918fSTim Haley } 1359c717a561Smaybee } else { 1360c717a561Smaybee ASSERT(db->db_buf != NULL); 1361c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1362c25056deSgw25295 mutex_destroy(&dr->dt.di.dr_mtx); 1363c25056deSgw25295 list_destroy(&dr->dt.di.dr_children); 1364c717a561Smaybee } 1365c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1366fa9e4066Sahrens 1367fa9e4066Sahrens ASSERT(db->db_dirtycnt > 0); 1368fa9e4066Sahrens db->db_dirtycnt -= 1; 1369fa9e4066Sahrens 1370c717a561Smaybee if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1371ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 1372fa9e4066Sahrens 1373b24ab676SJeff Bonwick ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1374ea8dc4b6Seschrock dbuf_set_data(db, NULL); 1375ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 1376fa9e4066Sahrens dbuf_evict(db); 1377fa9e4066Sahrens return (1); 1378fa9e4066Sahrens } 1379fa9e4066Sahrens 1380fa9e4066Sahrens mutex_exit(&db->db_mtx); 1381fa9e4066Sahrens return (0); 1382fa9e4066Sahrens } 1383fa9e4066Sahrens 1384fa9e4066Sahrens #pragma weak dmu_buf_will_dirty = dbuf_will_dirty 1385fa9e4066Sahrens void 1386fa9e4066Sahrens dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1387fa9e4066Sahrens { 13881ab7f2deSmaybee int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1389fa9e4066Sahrens 1390fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1391fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1392fa9e4066Sahrens 1393744947dcSTom Erickson DB_DNODE_ENTER(db); 1394744947dcSTom Erickson if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1395fa9e4066Sahrens rf |= DB_RF_HAVESTRUCT; 1396744947dcSTom Erickson DB_DNODE_EXIT(db); 1397ea8dc4b6Seschrock (void) dbuf_read(db, NULL, rf); 1398c717a561Smaybee (void) dbuf_dirty(db, tx); 1399fa9e4066Sahrens } 1400fa9e4066Sahrens 1401fa9e4066Sahrens void 140282c9918fSTim Haley dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 140382c9918fSTim Haley { 140482c9918fSTim Haley dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 140582c9918fSTim Haley 140682c9918fSTim Haley db->db_state = DB_NOFILL; 140782c9918fSTim Haley 140882c9918fSTim Haley dmu_buf_will_fill(db_fake, tx); 140982c9918fSTim Haley } 141082c9918fSTim Haley 141182c9918fSTim Haley void 1412ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1413fa9e4066Sahrens { 1414ea8dc4b6Seschrock dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1415ea8dc4b6Seschrock 14160a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1417fa9e4066Sahrens ASSERT(tx->tx_txg != 0); 1418fa9e4066Sahrens ASSERT(db->db_level == 0); 1419fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 1420fa9e4066Sahrens 1421ea8dc4b6Seschrock ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1422fa9e4066Sahrens dmu_tx_private_ok(tx)); 1423fa9e4066Sahrens 1424fa9e4066Sahrens dbuf_noread(db); 1425c717a561Smaybee (void) dbuf_dirty(db, tx); 1426fa9e4066Sahrens } 1427fa9e4066Sahrens 1428fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done 1429fa9e4066Sahrens /* ARGSUSED */ 1430fa9e4066Sahrens void 1431fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1432fa9e4066Sahrens { 1433fa9e4066Sahrens mutex_enter(&db->db_mtx); 14349c9dc39aSek110237 DBUF_VERIFY(db); 1435fa9e4066Sahrens 1436fa9e4066Sahrens if (db->db_state == DB_FILL) { 1437c717a561Smaybee if (db->db_level == 0 && db->db_freed_in_flight) { 14380a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1439fa9e4066Sahrens /* we were freed while filling */ 1440fa9e4066Sahrens /* XXX dbuf_undirty? */ 1441fa9e4066Sahrens bzero(db->db.db_data, db->db.db_size); 1442c717a561Smaybee db->db_freed_in_flight = FALSE; 1443fa9e4066Sahrens } 1444fa9e4066Sahrens db->db_state = DB_CACHED; 1445fa9e4066Sahrens cv_broadcast(&db->db_changed); 1446fa9e4066Sahrens } 1447fa9e4066Sahrens mutex_exit(&db->db_mtx); 1448fa9e4066Sahrens } 1449fa9e4066Sahrens 1450ea8dc4b6Seschrock /* 14512fdbea25SAleksandr Guzovskiy * Directly assign a provided arc buf to a given dbuf if it's not referenced 14522fdbea25SAleksandr Guzovskiy * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 14532fdbea25SAleksandr Guzovskiy */ 14542fdbea25SAleksandr Guzovskiy void 14552fdbea25SAleksandr Guzovskiy dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 14562fdbea25SAleksandr Guzovskiy { 14572fdbea25SAleksandr Guzovskiy ASSERT(!refcount_is_zero(&db->db_holds)); 14580a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 14592fdbea25SAleksandr Guzovskiy ASSERT(db->db_level == 0); 14602fdbea25SAleksandr Guzovskiy ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 14612fdbea25SAleksandr Guzovskiy ASSERT(buf != NULL); 14622fdbea25SAleksandr Guzovskiy ASSERT(arc_buf_size(buf) == db->db.db_size); 14632fdbea25SAleksandr Guzovskiy ASSERT(tx->tx_txg != 0); 14642fdbea25SAleksandr Guzovskiy 14652fdbea25SAleksandr Guzovskiy arc_return_buf(buf, db); 14662fdbea25SAleksandr Guzovskiy ASSERT(arc_released(buf)); 14672fdbea25SAleksandr Guzovskiy 14682fdbea25SAleksandr Guzovskiy mutex_enter(&db->db_mtx); 14692fdbea25SAleksandr Guzovskiy 14702fdbea25SAleksandr Guzovskiy while (db->db_state == DB_READ || db->db_state == DB_FILL) 14712fdbea25SAleksandr Guzovskiy cv_wait(&db->db_changed, &db->db_mtx); 14722fdbea25SAleksandr Guzovskiy 14732fdbea25SAleksandr Guzovskiy ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 14742fdbea25SAleksandr Guzovskiy 14752fdbea25SAleksandr Guzovskiy if (db->db_state == DB_CACHED && 14762fdbea25SAleksandr Guzovskiy refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 14772fdbea25SAleksandr Guzovskiy mutex_exit(&db->db_mtx); 14782fdbea25SAleksandr Guzovskiy (void) dbuf_dirty(db, tx); 14792fdbea25SAleksandr Guzovskiy bcopy(buf->b_data, db->db.db_data, db->db.db_size); 14802fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(buf, db) == 1); 1481c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_wbuf_copied(); 14822fdbea25SAleksandr Guzovskiy return; 14832fdbea25SAleksandr Guzovskiy } 14842fdbea25SAleksandr Guzovskiy 1485c242f9a0Schunli zhang - Sun Microsystems - Irvine United States xuio_stat_wbuf_nocopy(); 14862fdbea25SAleksandr Guzovskiy if (db->db_state == DB_CACHED) { 14872fdbea25SAleksandr Guzovskiy dbuf_dirty_record_t *dr = db->db_last_dirty; 14882fdbea25SAleksandr Guzovskiy 14892fdbea25SAleksandr Guzovskiy ASSERT(db->db_buf != NULL); 14902fdbea25SAleksandr Guzovskiy if (dr != NULL && dr->dr_txg == tx->tx_txg) { 14912fdbea25SAleksandr Guzovskiy ASSERT(dr->dt.dl.dr_data == db->db_buf); 14922fdbea25SAleksandr Guzovskiy if (!arc_released(db->db_buf)) { 14932fdbea25SAleksandr Guzovskiy ASSERT(dr->dt.dl.dr_override_state == 14942fdbea25SAleksandr Guzovskiy DR_OVERRIDDEN); 14952fdbea25SAleksandr Guzovskiy arc_release(db->db_buf, db); 14962fdbea25SAleksandr Guzovskiy } 14972fdbea25SAleksandr Guzovskiy dr->dt.dl.dr_data = buf; 14982fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); 14992fdbea25SAleksandr Guzovskiy } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 15002fdbea25SAleksandr Guzovskiy arc_release(db->db_buf, db); 15012fdbea25SAleksandr Guzovskiy VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); 15022fdbea25SAleksandr Guzovskiy } 15032fdbea25SAleksandr Guzovskiy db->db_buf = NULL; 15042fdbea25SAleksandr Guzovskiy } 15052fdbea25SAleksandr Guzovskiy ASSERT(db->db_buf == NULL); 15062fdbea25SAleksandr Guzovskiy dbuf_set_data(db, buf); 15072fdbea25SAleksandr Guzovskiy db->db_state = DB_FILL; 15082fdbea25SAleksandr Guzovskiy mutex_exit(&db->db_mtx); 15092fdbea25SAleksandr Guzovskiy (void) dbuf_dirty(db, tx); 15102fdbea25SAleksandr Guzovskiy dbuf_fill_done(db, tx); 15112fdbea25SAleksandr Guzovskiy } 15122fdbea25SAleksandr Guzovskiy 15132fdbea25SAleksandr Guzovskiy /* 1514ea8dc4b6Seschrock * "Clear" the contents of this dbuf. This will mark the dbuf 1515ea8dc4b6Seschrock * EVICTING and clear *most* of its references. Unfortunetely, 1516ea8dc4b6Seschrock * when we are not holding the dn_dbufs_mtx, we can't clear the 1517ea8dc4b6Seschrock * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1518ea8dc4b6Seschrock * in this case. For callers from the DMU we will usually see: 1519ea8dc4b6Seschrock * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1520ea8dc4b6Seschrock * For the arc callback, we will usually see: 1521ea8dc4b6Seschrock * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1522ea8dc4b6Seschrock * Sometimes, though, we will get a mix of these two: 1523ea8dc4b6Seschrock * DMU: dbuf_clear()->arc_buf_evict() 1524ea8dc4b6Seschrock * ARC: dbuf_do_evict()->dbuf_destroy() 1525ea8dc4b6Seschrock */ 1526ea8dc4b6Seschrock void 1527fa9e4066Sahrens dbuf_clear(dmu_buf_impl_t *db) 1528fa9e4066Sahrens { 1529744947dcSTom Erickson dnode_t *dn; 1530ea8dc4b6Seschrock dmu_buf_impl_t *parent = db->db_parent; 1531744947dcSTom Erickson dmu_buf_impl_t *dndb; 1532ea8dc4b6Seschrock int dbuf_gone = FALSE; 1533fa9e4066Sahrens 1534fa9e4066Sahrens ASSERT(MUTEX_HELD(&db->db_mtx)); 1535fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1536fa9e4066Sahrens 1537ea8dc4b6Seschrock dbuf_evict_user(db); 1538ea8dc4b6Seschrock 1539fa9e4066Sahrens if (db->db_state == DB_CACHED) { 1540ea8dc4b6Seschrock ASSERT(db->db.db_data != NULL); 15410a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 1542ea8dc4b6Seschrock zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 15435a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 15440e8c6158Smaybee } 1545fa9e4066Sahrens db->db.db_data = NULL; 1546fa9e4066Sahrens db->db_state = DB_UNCACHED; 1547fa9e4066Sahrens } 1548fa9e4066Sahrens 154982c9918fSTim Haley ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1550fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1551fa9e4066Sahrens 1552ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1553ea8dc4b6Seschrock db->db_blkptr = NULL; 1554ea8dc4b6Seschrock 1555744947dcSTom Erickson DB_DNODE_ENTER(db); 1556744947dcSTom Erickson dn = DB_DNODE(db); 1557744947dcSTom Erickson dndb = dn->dn_dbuf; 15580a586ceaSMark Shellenbaum if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1559ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1560744947dcSTom Erickson (void) atomic_dec_32_nv(&dn->dn_dbufs_count); 1561744947dcSTom Erickson membar_producer(); 1562744947dcSTom Erickson DB_DNODE_EXIT(db); 1563744947dcSTom Erickson /* 1564744947dcSTom Erickson * Decrementing the dbuf count means that the hold corresponding 1565744947dcSTom Erickson * to the removed dbuf is no longer discounted in dnode_move(), 1566744947dcSTom Erickson * so the dnode cannot be moved until after we release the hold. 1567744947dcSTom Erickson * The membar_producer() ensures visibility of the decremented 1568744947dcSTom Erickson * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1569744947dcSTom Erickson * release any lock. 1570744947dcSTom Erickson */ 1571ea8dc4b6Seschrock dnode_rele(dn, db); 1572744947dcSTom Erickson db->db_dnode_handle = NULL; 1573744947dcSTom Erickson } else { 1574744947dcSTom Erickson DB_DNODE_EXIT(db); 1575ea8dc4b6Seschrock } 1576ea8dc4b6Seschrock 1577ea8dc4b6Seschrock if (db->db_buf) 1578ea8dc4b6Seschrock dbuf_gone = arc_buf_evict(db->db_buf); 1579ea8dc4b6Seschrock 1580ea8dc4b6Seschrock if (!dbuf_gone) 1581fa9e4066Sahrens mutex_exit(&db->db_mtx); 1582fa9e4066Sahrens 1583fa9e4066Sahrens /* 1584744947dcSTom Erickson * If this dbuf is referenced from an indirect dbuf, 1585fa9e4066Sahrens * decrement the ref count on the indirect dbuf. 1586fa9e4066Sahrens */ 1587c543ec06Sahrens if (parent && parent != dndb) 1588ea8dc4b6Seschrock dbuf_rele(parent, db); 1589fa9e4066Sahrens } 1590fa9e4066Sahrens 1591fa9e4066Sahrens static int 1592fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1593fa9e4066Sahrens dmu_buf_impl_t **parentp, blkptr_t **bpp) 1594fa9e4066Sahrens { 1595fa9e4066Sahrens int nlevels, epbs; 1596fa9e4066Sahrens 15970b69c2f0Sahrens *parentp = NULL; 15980b69c2f0Sahrens *bpp = NULL; 15990b69c2f0Sahrens 16000a586ceaSMark Shellenbaum ASSERT(blkid != DMU_BONUS_BLKID); 16010a586ceaSMark Shellenbaum 16020a586ceaSMark Shellenbaum if (blkid == DMU_SPILL_BLKID) { 16030a586ceaSMark Shellenbaum mutex_enter(&dn->dn_mtx); 160406e0070dSMark Shellenbaum if (dn->dn_have_spill && 160506e0070dSMark Shellenbaum (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 16060a586ceaSMark Shellenbaum *bpp = &dn->dn_phys->dn_spill; 16070a586ceaSMark Shellenbaum else 16080a586ceaSMark Shellenbaum *bpp = NULL; 16090a586ceaSMark Shellenbaum dbuf_add_ref(dn->dn_dbuf, NULL); 16100a586ceaSMark Shellenbaum *parentp = dn->dn_dbuf; 16110a586ceaSMark Shellenbaum mutex_exit(&dn->dn_mtx); 16120a586ceaSMark Shellenbaum return (0); 16130a586ceaSMark Shellenbaum } 1614ea8dc4b6Seschrock 1615fa9e4066Sahrens if (dn->dn_phys->dn_nlevels == 0) 1616fa9e4066Sahrens nlevels = 1; 1617fa9e4066Sahrens else 1618fa9e4066Sahrens nlevels = dn->dn_phys->dn_nlevels; 1619fa9e4066Sahrens 1620fa9e4066Sahrens epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1621fa9e4066Sahrens 1622fa9e4066Sahrens ASSERT3U(level * epbs, <, 64); 1623fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1624ea8dc4b6Seschrock if (level >= nlevels || 1625fa9e4066Sahrens (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1626fa9e4066Sahrens /* the buffer has no parent yet */ 1627fa9e4066Sahrens return (ENOENT); 1628fa9e4066Sahrens } else if (level < nlevels-1) { 1629fa9e4066Sahrens /* this block is referenced from an indirect block */ 1630fa9e4066Sahrens int err = dbuf_hold_impl(dn, level+1, 1631fa9e4066Sahrens blkid >> epbs, fail_sparse, NULL, parentp); 1632fa9e4066Sahrens if (err) 1633fa9e4066Sahrens return (err); 1634ea8dc4b6Seschrock err = dbuf_read(*parentp, NULL, 1635ea8dc4b6Seschrock (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1636c543ec06Sahrens if (err) { 1637c543ec06Sahrens dbuf_rele(*parentp, NULL); 1638c543ec06Sahrens *parentp = NULL; 1639c543ec06Sahrens return (err); 1640c543ec06Sahrens } 1641fa9e4066Sahrens *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1642fa9e4066Sahrens (blkid & ((1ULL << epbs) - 1)); 1643c543ec06Sahrens return (0); 1644fa9e4066Sahrens } else { 1645fa9e4066Sahrens /* the block is referenced from the dnode */ 1646fa9e4066Sahrens ASSERT3U(level, ==, nlevels-1); 1647fa9e4066Sahrens ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1648fa9e4066Sahrens blkid < dn->dn_phys->dn_nblkptr); 1649c543ec06Sahrens if (dn->dn_dbuf) { 1650c543ec06Sahrens dbuf_add_ref(dn->dn_dbuf, NULL); 1651fa9e4066Sahrens *parentp = dn->dn_dbuf; 1652c543ec06Sahrens } 1653fa9e4066Sahrens *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1654fa9e4066Sahrens return (0); 1655fa9e4066Sahrens } 1656fa9e4066Sahrens } 1657fa9e4066Sahrens 1658fa9e4066Sahrens static dmu_buf_impl_t * 1659fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1660fa9e4066Sahrens dmu_buf_impl_t *parent, blkptr_t *blkptr) 1661fa9e4066Sahrens { 1662503ad85cSMatthew Ahrens objset_t *os = dn->dn_objset; 1663fa9e4066Sahrens dmu_buf_impl_t *db, *odb; 1664fa9e4066Sahrens 1665fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1666fa9e4066Sahrens ASSERT(dn->dn_type != DMU_OT_NONE); 1667fa9e4066Sahrens 1668fa9e4066Sahrens db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1669fa9e4066Sahrens 1670fa9e4066Sahrens db->db_objset = os; 1671fa9e4066Sahrens db->db.db_object = dn->dn_object; 1672fa9e4066Sahrens db->db_level = level; 1673fa9e4066Sahrens db->db_blkid = blkid; 1674c717a561Smaybee db->db_last_dirty = NULL; 1675ea8dc4b6Seschrock db->db_dirtycnt = 0; 1676744947dcSTom Erickson db->db_dnode_handle = dn->dn_handle; 1677ea8dc4b6Seschrock db->db_parent = parent; 1678ea8dc4b6Seschrock db->db_blkptr = blkptr; 1679fa9e4066Sahrens 1680c717a561Smaybee db->db_user_ptr = NULL; 1681c717a561Smaybee db->db_user_data_ptr_ptr = NULL; 1682c717a561Smaybee db->db_evict_func = NULL; 1683c717a561Smaybee db->db_immediate_evict = 0; 1684c717a561Smaybee db->db_freed_in_flight = 0; 1685ea8dc4b6Seschrock 16860a586ceaSMark Shellenbaum if (blkid == DMU_BONUS_BLKID) { 1687ea8dc4b6Seschrock ASSERT3P(parent, ==, dn->dn_dbuf); 16881934e92fSmaybee db->db.db_size = DN_MAX_BONUSLEN - 16891934e92fSmaybee (dn->dn_nblkptr-1) * sizeof (blkptr_t); 16901934e92fSmaybee ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 16910a586ceaSMark Shellenbaum db->db.db_offset = DMU_BONUS_BLKID; 1692ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1693ea8dc4b6Seschrock /* the bonus dbuf is not placed in the hash table */ 16945a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1695ea8dc4b6Seschrock return (db); 16960a586ceaSMark Shellenbaum } else if (blkid == DMU_SPILL_BLKID) { 16970a586ceaSMark Shellenbaum db->db.db_size = (blkptr != NULL) ? 16980a586ceaSMark Shellenbaum BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 16990a586ceaSMark Shellenbaum db->db.db_offset = 0; 1700fa9e4066Sahrens } else { 1701fa9e4066Sahrens int blocksize = 1702fa9e4066Sahrens db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; 1703fa9e4066Sahrens db->db.db_size = blocksize; 1704fa9e4066Sahrens db->db.db_offset = db->db_blkid * blocksize; 1705fa9e4066Sahrens } 1706fa9e4066Sahrens 1707fa9e4066Sahrens /* 1708fa9e4066Sahrens * Hold the dn_dbufs_mtx while we get the new dbuf 1709fa9e4066Sahrens * in the hash table *and* added to the dbufs list. 1710fa9e4066Sahrens * This prevents a possible deadlock with someone 1711fa9e4066Sahrens * trying to look up this dbuf before its added to the 1712fa9e4066Sahrens * dn_dbufs list. 1713fa9e4066Sahrens */ 1714fa9e4066Sahrens mutex_enter(&dn->dn_dbufs_mtx); 1715ea8dc4b6Seschrock db->db_state = DB_EVICTING; 1716fa9e4066Sahrens if ((odb = dbuf_hash_insert(db)) != NULL) { 1717fa9e4066Sahrens /* someone else inserted it first */ 1718fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 1719fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 1720fa9e4066Sahrens return (odb); 1721fa9e4066Sahrens } 1722fa9e4066Sahrens list_insert_head(&dn->dn_dbufs, db); 1723ea8dc4b6Seschrock db->db_state = DB_UNCACHED; 1724fa9e4066Sahrens mutex_exit(&dn->dn_dbufs_mtx); 17255a98e54bSBrendan Gregg - Sun Microsystems arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1726fa9e4066Sahrens 1727fa9e4066Sahrens if (parent && parent != dn->dn_dbuf) 1728fa9e4066Sahrens dbuf_add_ref(parent, db); 1729fa9e4066Sahrens 1730ea8dc4b6Seschrock ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1731ea8dc4b6Seschrock refcount_count(&dn->dn_holds) > 0); 1732fa9e4066Sahrens (void) refcount_add(&dn->dn_holds, db); 1733744947dcSTom Erickson (void) atomic_inc_32_nv(&dn->dn_dbufs_count); 1734fa9e4066Sahrens 1735fa9e4066Sahrens dprintf_dbuf(db, "db=%p\n", db); 1736fa9e4066Sahrens 1737fa9e4066Sahrens return (db); 1738fa9e4066Sahrens } 1739fa9e4066Sahrens 1740fa9e4066Sahrens static int 1741ea8dc4b6Seschrock dbuf_do_evict(void *private) 1742fa9e4066Sahrens { 1743ea8dc4b6Seschrock arc_buf_t *buf = private; 1744ea8dc4b6Seschrock dmu_buf_impl_t *db = buf->b_private; 1745fa9e4066Sahrens 1746ea8dc4b6Seschrock if (!MUTEX_HELD(&db->db_mtx)) 1747ea8dc4b6Seschrock mutex_enter(&db->db_mtx); 1748ea8dc4b6Seschrock 1749ea8dc4b6Seschrock ASSERT(refcount_is_zero(&db->db_holds)); 1750ea8dc4b6Seschrock 1751ea8dc4b6Seschrock if (db->db_state != DB_EVICTING) { 1752ea8dc4b6Seschrock ASSERT(db->db_state == DB_CACHED); 17539c9dc39aSek110237 DBUF_VERIFY(db); 1754ea8dc4b6Seschrock db->db_buf = NULL; 1755ea8dc4b6Seschrock dbuf_evict(db); 1756ea8dc4b6Seschrock } else { 1757ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 1758ea8dc4b6Seschrock dbuf_destroy(db); 1759fa9e4066Sahrens } 1760ea8dc4b6Seschrock return (0); 1761fa9e4066Sahrens } 1762fa9e4066Sahrens 1763fa9e4066Sahrens static void 1764fa9e4066Sahrens dbuf_destroy(dmu_buf_impl_t *db) 1765fa9e4066Sahrens { 1766fa9e4066Sahrens ASSERT(refcount_is_zero(&db->db_holds)); 1767fa9e4066Sahrens 17680a586ceaSMark Shellenbaum if (db->db_blkid != DMU_BONUS_BLKID) { 1769ea8dc4b6Seschrock /* 1770ea8dc4b6Seschrock * If this dbuf is still on the dn_dbufs list, 1771ea8dc4b6Seschrock * remove it from that list. 1772ea8dc4b6Seschrock */ 1773744947dcSTom Erickson if (db->db_dnode_handle != NULL) { 1774744947dcSTom Erickson dnode_t *dn; 17751934e92fSmaybee 1776744947dcSTom Erickson DB_DNODE_ENTER(db); 1777744947dcSTom Erickson dn = DB_DNODE(db); 17781934e92fSmaybee mutex_enter(&dn->dn_dbufs_mtx); 1779ea8dc4b6Seschrock list_remove(&dn->dn_dbufs, db); 1780744947dcSTom Erickson (void) atomic_dec_32_nv(&dn->dn_dbufs_count); 1781ea8dc4b6Seschrock mutex_exit(&dn->dn_dbufs_mtx); 1782744947dcSTom Erickson DB_DNODE_EXIT(db); 1783744947dcSTom Erickson /* 1784744947dcSTom Erickson * Decrementing the dbuf count means that the hold 1785744947dcSTom Erickson * corresponding to the removed dbuf is no longer 1786744947dcSTom Erickson * discounted in dnode_move(), so the dnode cannot be 1787744947dcSTom Erickson * moved until after we release the hold. 1788744947dcSTom Erickson */ 1789ea8dc4b6Seschrock dnode_rele(dn, db); 1790744947dcSTom Erickson db->db_dnode_handle = NULL; 1791ea8dc4b6Seschrock } 1792ea8dc4b6Seschrock dbuf_hash_remove(db); 1793ea8dc4b6Seschrock } 1794ea8dc4b6Seschrock db->db_parent = NULL; 1795ea8dc4b6Seschrock db->db_buf = NULL; 1796ea8dc4b6Seschrock 1797f82bfe17Sgw25295 ASSERT(!list_link_active(&db->db_link)); 1798fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 1799fa9e4066Sahrens ASSERT(db->db_hash_next == NULL); 1800fa9e4066Sahrens ASSERT(db->db_blkptr == NULL); 1801fa9e4066Sahrens ASSERT(db->db_data_pending == NULL); 1802fa9e4066Sahrens 1803fa9e4066Sahrens kmem_cache_free(dbuf_cache, db); 18045a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1805fa9e4066Sahrens } 1806fa9e4066Sahrens 1807fa9e4066Sahrens void 1808fa9e4066Sahrens dbuf_prefetch(dnode_t *dn, uint64_t blkid) 1809fa9e4066Sahrens { 181013506d1eSmaybee dmu_buf_impl_t *db = NULL; 1811fa9e4066Sahrens blkptr_t *bp = NULL; 1812fa9e4066Sahrens 18130a586ceaSMark Shellenbaum ASSERT(blkid != DMU_BONUS_BLKID); 1814fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1815fa9e4066Sahrens 1816fa9e4066Sahrens if (dnode_block_freed(dn, blkid)) 1817fa9e4066Sahrens return; 1818fa9e4066Sahrens 1819fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1820fa9e4066Sahrens if (db = dbuf_find(dn, 0, blkid)) { 1821fa9e4066Sahrens /* 18229e9c486fSGeorge Wilson * This dbuf is already in the cache. We assume that 18239e9c486fSGeorge Wilson * it is already CACHED, or else about to be either 1824fa9e4066Sahrens * read or filled. 1825fa9e4066Sahrens */ 1826fa9e4066Sahrens mutex_exit(&db->db_mtx); 1827fa9e4066Sahrens return; 1828fa9e4066Sahrens } 1829fa9e4066Sahrens 183013506d1eSmaybee if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1831fa9e4066Sahrens if (bp && !BP_IS_HOLE(bp)) { 1832c7cd2421SGeorge Wilson int priority = dn->dn_type == DMU_OT_DDT_ZAP ? 1833c7cd2421SGeorge Wilson ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ; 1834088f3894Sahrens arc_buf_t *pbuf; 1835b24ab676SJeff Bonwick dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 183613506d1eSmaybee uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; 1837ea8dc4b6Seschrock zbookmark_t zb; 1838b24ab676SJeff Bonwick 1839b24ab676SJeff Bonwick SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 1840b24ab676SJeff Bonwick dn->dn_object, 0, blkid); 1841ea8dc4b6Seschrock 1842088f3894Sahrens if (db) 1843088f3894Sahrens pbuf = db->db_buf; 1844088f3894Sahrens else 1845088f3894Sahrens pbuf = dn->dn_objset->os_phys_buf; 1846088f3894Sahrens 18473f9d6ad7SLin Ling (void) dsl_read(NULL, dn->dn_objset->os_spa, 1848c7cd2421SGeorge Wilson bp, pbuf, NULL, NULL, priority, 1849fa9e4066Sahrens ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 185013506d1eSmaybee &aflags, &zb); 1851fa9e4066Sahrens } 185213506d1eSmaybee if (db) 185313506d1eSmaybee dbuf_rele(db, NULL); 1854fa9e4066Sahrens } 1855fa9e4066Sahrens } 1856fa9e4066Sahrens 1857fa9e4066Sahrens /* 1858fa9e4066Sahrens * Returns with db_holds incremented, and db_mtx not held. 1859fa9e4066Sahrens * Note: dn_struct_rwlock must be held. 1860fa9e4066Sahrens */ 1861fa9e4066Sahrens int 1862fa9e4066Sahrens dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1863fa9e4066Sahrens void *tag, dmu_buf_impl_t **dbp) 1864fa9e4066Sahrens { 1865fa9e4066Sahrens dmu_buf_impl_t *db, *parent = NULL; 1866fa9e4066Sahrens 18670a586ceaSMark Shellenbaum ASSERT(blkid != DMU_BONUS_BLKID); 1868fa9e4066Sahrens ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1869fa9e4066Sahrens ASSERT3U(dn->dn_nlevels, >, level); 1870fa9e4066Sahrens 1871fa9e4066Sahrens *dbp = NULL; 1872ea8dc4b6Seschrock top: 1873fa9e4066Sahrens /* dbuf_find() returns with db_mtx held */ 1874fa9e4066Sahrens db = dbuf_find(dn, level, blkid); 1875fa9e4066Sahrens 1876fa9e4066Sahrens if (db == NULL) { 1877fa9e4066Sahrens blkptr_t *bp = NULL; 1878fa9e4066Sahrens int err; 1879fa9e4066Sahrens 1880c543ec06Sahrens ASSERT3P(parent, ==, NULL); 1881fa9e4066Sahrens err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1882fa9e4066Sahrens if (fail_sparse) { 1883fa9e4066Sahrens if (err == 0 && bp && BP_IS_HOLE(bp)) 1884fa9e4066Sahrens err = ENOENT; 1885fa9e4066Sahrens if (err) { 1886c543ec06Sahrens if (parent) 1887ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1888fa9e4066Sahrens return (err); 1889fa9e4066Sahrens } 1890fa9e4066Sahrens } 1891ea8dc4b6Seschrock if (err && err != ENOENT) 1892ea8dc4b6Seschrock return (err); 1893fa9e4066Sahrens db = dbuf_create(dn, level, blkid, parent, bp); 1894fa9e4066Sahrens } 1895fa9e4066Sahrens 1896ea8dc4b6Seschrock if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1897ea8dc4b6Seschrock arc_buf_add_ref(db->db_buf, db); 1898ea8dc4b6Seschrock if (db->db_buf->b_data == NULL) { 1899ea8dc4b6Seschrock dbuf_clear(db); 1900c543ec06Sahrens if (parent) { 1901c543ec06Sahrens dbuf_rele(parent, NULL); 1902c543ec06Sahrens parent = NULL; 1903c543ec06Sahrens } 1904ea8dc4b6Seschrock goto top; 1905ea8dc4b6Seschrock } 1906ea8dc4b6Seschrock ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1907ea8dc4b6Seschrock } 1908ea8dc4b6Seschrock 1909ea8dc4b6Seschrock ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1910ea8dc4b6Seschrock 1911fa9e4066Sahrens /* 1912c717a561Smaybee * If this buffer is currently syncing out, and we are are 1913c717a561Smaybee * still referencing it from db_data, we need to make a copy 1914c717a561Smaybee * of it in case we decide we want to dirty it again in this txg. 1915fa9e4066Sahrens */ 19160a586ceaSMark Shellenbaum if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1917ea8dc4b6Seschrock dn->dn_object != DMU_META_DNODE_OBJECT && 1918c717a561Smaybee db->db_state == DB_CACHED && db->db_data_pending) { 1919c717a561Smaybee dbuf_dirty_record_t *dr = db->db_data_pending; 1920c717a561Smaybee 1921c717a561Smaybee if (dr->dt.dl.dr_data == db->db_buf) { 1922ad23a2dbSjohansen arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1923fa9e4066Sahrens 1924c717a561Smaybee dbuf_set_data(db, 1925744947dcSTom Erickson arc_buf_alloc(dn->dn_objset->os_spa, 1926c717a561Smaybee db->db.db_size, db, type)); 1927c717a561Smaybee bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1928fa9e4066Sahrens db->db.db_size); 1929fa9e4066Sahrens } 1930c717a561Smaybee } 1931fa9e4066Sahrens 1932ea8dc4b6Seschrock (void) refcount_add(&db->db_holds, tag); 1933fa9e4066Sahrens dbuf_update_data(db); 19349c9dc39aSek110237 DBUF_VERIFY(db); 1935fa9e4066Sahrens mutex_exit(&db->db_mtx); 1936fa9e4066Sahrens 1937fa9e4066Sahrens /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1938c543ec06Sahrens if (parent) 1939ea8dc4b6Seschrock dbuf_rele(parent, NULL); 1940fa9e4066Sahrens 1941744947dcSTom Erickson ASSERT3P(DB_DNODE(db), ==, dn); 1942fa9e4066Sahrens ASSERT3U(db->db_blkid, ==, blkid); 1943fa9e4066Sahrens ASSERT3U(db->db_level, ==, level); 1944fa9e4066Sahrens *dbp = db; 1945fa9e4066Sahrens 1946fa9e4066Sahrens return (0); 1947fa9e4066Sahrens } 1948fa9e4066Sahrens 1949fa9e4066Sahrens dmu_buf_impl_t * 1950ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1951fa9e4066Sahrens { 1952fa9e4066Sahrens dmu_buf_impl_t *db; 1953ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1954ea8dc4b6Seschrock return (err ? NULL : db); 1955fa9e4066Sahrens } 1956fa9e4066Sahrens 1957fa9e4066Sahrens dmu_buf_impl_t * 1958fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1959fa9e4066Sahrens { 1960fa9e4066Sahrens dmu_buf_impl_t *db; 1961ea8dc4b6Seschrock int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1962ea8dc4b6Seschrock return (err ? NULL : db); 1963fa9e4066Sahrens } 1964fa9e4066Sahrens 19651934e92fSmaybee void 1966ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn) 1967fa9e4066Sahrens { 1968ea8dc4b6Seschrock ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1969ea8dc4b6Seschrock 1970ea8dc4b6Seschrock ASSERT(dn->dn_bonus == NULL); 19710a586ceaSMark Shellenbaum dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 19720a586ceaSMark Shellenbaum } 19730a586ceaSMark Shellenbaum 19740a586ceaSMark Shellenbaum int 19750a586ceaSMark Shellenbaum dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 19760a586ceaSMark Shellenbaum { 19770a586ceaSMark Shellenbaum dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1978744947dcSTom Erickson dnode_t *dn; 1979744947dcSTom Erickson 19800a586ceaSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID) 19810a586ceaSMark Shellenbaum return (ENOTSUP); 19820a586ceaSMark Shellenbaum if (blksz == 0) 19830a586ceaSMark Shellenbaum blksz = SPA_MINBLOCKSIZE; 19840a586ceaSMark Shellenbaum if (blksz > SPA_MAXBLOCKSIZE) 19850a586ceaSMark Shellenbaum blksz = SPA_MAXBLOCKSIZE; 19860a586ceaSMark Shellenbaum else 19870a586ceaSMark Shellenbaum blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 19880a586ceaSMark Shellenbaum 1989744947dcSTom Erickson DB_DNODE_ENTER(db); 1990744947dcSTom Erickson dn = DB_DNODE(db); 1991744947dcSTom Erickson rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 19920a586ceaSMark Shellenbaum dbuf_new_size(db, blksz, tx); 1993744947dcSTom Erickson rw_exit(&dn->dn_struct_rwlock); 1994744947dcSTom Erickson DB_DNODE_EXIT(db); 19950a586ceaSMark Shellenbaum 19960a586ceaSMark Shellenbaum return (0); 19970a586ceaSMark Shellenbaum } 19980a586ceaSMark Shellenbaum 19990a586ceaSMark Shellenbaum void 20000a586ceaSMark Shellenbaum dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 20010a586ceaSMark Shellenbaum { 20020a586ceaSMark Shellenbaum dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2003fa9e4066Sahrens } 2004fa9e4066Sahrens 2005ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref 2006fa9e4066Sahrens void 2007fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2008fa9e4066Sahrens { 2009ea8dc4b6Seschrock int64_t holds = refcount_add(&db->db_holds, tag); 2010ea8dc4b6Seschrock ASSERT(holds > 1); 2011fa9e4066Sahrens } 2012fa9e4066Sahrens 2013744947dcSTom Erickson /* 2014744947dcSTom Erickson * If you call dbuf_rele() you had better not be referencing the dnode handle 2015744947dcSTom Erickson * unless you have some other direct or indirect hold on the dnode. (An indirect 2016744947dcSTom Erickson * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2017744947dcSTom Erickson * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2018744947dcSTom Erickson * dnode's parent dbuf evicting its dnode handles. 2019744947dcSTom Erickson */ 2020ea8dc4b6Seschrock #pragma weak dmu_buf_rele = dbuf_rele 2021fa9e4066Sahrens void 2022ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag) 2023fa9e4066Sahrens { 2024b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 2025b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, tag); 2026b24ab676SJeff Bonwick } 2027b24ab676SJeff Bonwick 2028b24ab676SJeff Bonwick /* 2029b24ab676SJeff Bonwick * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2030b24ab676SJeff Bonwick * db_dirtycnt and db_holds to be updated atomically. 2031b24ab676SJeff Bonwick */ 2032b24ab676SJeff Bonwick void 2033b24ab676SJeff Bonwick dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2034b24ab676SJeff Bonwick { 2035fa9e4066Sahrens int64_t holds; 2036fa9e4066Sahrens 2037b24ab676SJeff Bonwick ASSERT(MUTEX_HELD(&db->db_mtx)); 20389c9dc39aSek110237 DBUF_VERIFY(db); 2039fa9e4066Sahrens 2040744947dcSTom Erickson /* 2041744947dcSTom Erickson * Remove the reference to the dbuf before removing its hold on the 2042744947dcSTom Erickson * dnode so we can guarantee in dnode_move() that a referenced bonus 2043744947dcSTom Erickson * buffer has a corresponding dnode hold. 2044744947dcSTom Erickson */ 2045fa9e4066Sahrens holds = refcount_remove(&db->db_holds, tag); 2046ea8dc4b6Seschrock ASSERT(holds >= 0); 2047fa9e4066Sahrens 2048c717a561Smaybee /* 2049c717a561Smaybee * We can't freeze indirects if there is a possibility that they 2050c717a561Smaybee * may be modified in the current syncing context. 2051c717a561Smaybee */ 2052c717a561Smaybee if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 20536b4acc8bSahrens arc_buf_freeze(db->db_buf); 20546b4acc8bSahrens 2055fa9e4066Sahrens if (holds == db->db_dirtycnt && 2056c717a561Smaybee db->db_level == 0 && db->db_immediate_evict) 2057fa9e4066Sahrens dbuf_evict_user(db); 2058ea8dc4b6Seschrock 2059ea8dc4b6Seschrock if (holds == 0) { 20600a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 2061ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 2062744947dcSTom Erickson 2063744947dcSTom Erickson /* 2064744947dcSTom Erickson * If the dnode moves here, we cannot cross this barrier 2065744947dcSTom Erickson * until the move completes. 2066744947dcSTom Erickson */ 2067744947dcSTom Erickson DB_DNODE_ENTER(db); 2068744947dcSTom Erickson (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count); 2069744947dcSTom Erickson DB_DNODE_EXIT(db); 2070744947dcSTom Erickson /* 2071744947dcSTom Erickson * The bonus buffer's dnode hold is no longer discounted 2072744947dcSTom Erickson * in dnode_move(). The dnode cannot move until after 2073744947dcSTom Erickson * the dnode_rele(). 2074744947dcSTom Erickson */ 2075744947dcSTom Erickson dnode_rele(DB_DNODE(db), db); 2076ea8dc4b6Seschrock } else if (db->db_buf == NULL) { 2077ea8dc4b6Seschrock /* 2078ea8dc4b6Seschrock * This is a special case: we never associated this 2079ea8dc4b6Seschrock * dbuf with any data allocated from the ARC. 2080ea8dc4b6Seschrock */ 208182c9918fSTim Haley ASSERT(db->db_state == DB_UNCACHED || 208282c9918fSTim Haley db->db_state == DB_NOFILL); 2083ea8dc4b6Seschrock dbuf_evict(db); 2084ea8dc4b6Seschrock } else if (arc_released(db->db_buf)) { 2085ea8dc4b6Seschrock arc_buf_t *buf = db->db_buf; 2086ea8dc4b6Seschrock /* 2087ea8dc4b6Seschrock * This dbuf has anonymous data associated with it. 2088ea8dc4b6Seschrock */ 2089ea8dc4b6Seschrock dbuf_set_data(db, NULL); 2090ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(buf, db) == 1); 2091ea8dc4b6Seschrock dbuf_evict(db); 2092ea8dc4b6Seschrock } else { 2093ea8dc4b6Seschrock VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); 20949253d63dSGeorge Wilson 20959253d63dSGeorge Wilson /* 20969253d63dSGeorge Wilson * A dbuf will be eligible for eviction if either the 20979253d63dSGeorge Wilson * 'primarycache' property is set or a duplicate 20989253d63dSGeorge Wilson * copy of this buffer is already cached in the arc. 20999253d63dSGeorge Wilson * 21009253d63dSGeorge Wilson * In the case of the 'primarycache' a buffer 21019253d63dSGeorge Wilson * is considered for eviction if it matches the 21029253d63dSGeorge Wilson * criteria set in the property. 21039253d63dSGeorge Wilson * 21049253d63dSGeorge Wilson * To decide if our buffer is considered a 21059253d63dSGeorge Wilson * duplicate, we must call into the arc to determine 21069253d63dSGeorge Wilson * if multiple buffers are referencing the same 21079253d63dSGeorge Wilson * block on-disk. If so, then we simply evict 21089253d63dSGeorge Wilson * ourselves. 21099253d63dSGeorge Wilson */ 21109253d63dSGeorge Wilson if (!DBUF_IS_CACHEABLE(db) || 21119253d63dSGeorge Wilson arc_buf_eviction_needed(db->db_buf)) 21123baa08fcSek110237 dbuf_clear(db); 21133baa08fcSek110237 else 2114fa9e4066Sahrens mutex_exit(&db->db_mtx); 2115fa9e4066Sahrens } 2116ea8dc4b6Seschrock } else { 2117ea8dc4b6Seschrock mutex_exit(&db->db_mtx); 2118fa9e4066Sahrens } 2119fa9e4066Sahrens } 2120fa9e4066Sahrens 2121fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount 2122fa9e4066Sahrens uint64_t 2123fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db) 2124fa9e4066Sahrens { 2125fa9e4066Sahrens return (refcount_count(&db->db_holds)); 2126fa9e4066Sahrens } 2127fa9e4066Sahrens 2128fa9e4066Sahrens void * 2129fa9e4066Sahrens dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 2130fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 2131fa9e4066Sahrens { 2132fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 2133fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 2134fa9e4066Sahrens } 2135fa9e4066Sahrens 2136fa9e4066Sahrens void * 2137fa9e4066Sahrens dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 2138fa9e4066Sahrens dmu_buf_evict_func_t *evict_func) 2139fa9e4066Sahrens { 2140fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2141fa9e4066Sahrens 2142c717a561Smaybee db->db_immediate_evict = TRUE; 2143fa9e4066Sahrens return (dmu_buf_update_user(db_fake, NULL, user_ptr, 2144fa9e4066Sahrens user_data_ptr_ptr, evict_func)); 2145fa9e4066Sahrens } 2146fa9e4066Sahrens 2147fa9e4066Sahrens void * 2148fa9e4066Sahrens dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 2149fa9e4066Sahrens void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 2150fa9e4066Sahrens { 2151fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2152fa9e4066Sahrens ASSERT(db->db_level == 0); 2153fa9e4066Sahrens 2154fa9e4066Sahrens ASSERT((user_ptr == NULL) == (evict_func == NULL)); 2155fa9e4066Sahrens 2156fa9e4066Sahrens mutex_enter(&db->db_mtx); 2157fa9e4066Sahrens 2158c717a561Smaybee if (db->db_user_ptr == old_user_ptr) { 2159c717a561Smaybee db->db_user_ptr = user_ptr; 2160c717a561Smaybee db->db_user_data_ptr_ptr = user_data_ptr_ptr; 2161c717a561Smaybee db->db_evict_func = evict_func; 2162fa9e4066Sahrens 2163fa9e4066Sahrens dbuf_update_data(db); 2164fa9e4066Sahrens } else { 2165c717a561Smaybee old_user_ptr = db->db_user_ptr; 2166fa9e4066Sahrens } 2167fa9e4066Sahrens 2168fa9e4066Sahrens mutex_exit(&db->db_mtx); 2169fa9e4066Sahrens return (old_user_ptr); 2170fa9e4066Sahrens } 2171fa9e4066Sahrens 2172fa9e4066Sahrens void * 2173fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake) 2174fa9e4066Sahrens { 2175fa9e4066Sahrens dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2176fa9e4066Sahrens ASSERT(!refcount_is_zero(&db->db_holds)); 2177fa9e4066Sahrens 2178c717a561Smaybee return (db->db_user_ptr); 2179fa9e4066Sahrens } 2180fa9e4066Sahrens 21813d692628SSanjeev Bagewadi boolean_t 21823d692628SSanjeev Bagewadi dmu_buf_freeable(dmu_buf_t *dbuf) 21833d692628SSanjeev Bagewadi { 21843d692628SSanjeev Bagewadi boolean_t res = B_FALSE; 21853d692628SSanjeev Bagewadi dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 21863d692628SSanjeev Bagewadi 21873d692628SSanjeev Bagewadi if (db->db_blkptr) 21883d692628SSanjeev Bagewadi res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2189c7cd2421SGeorge Wilson db->db_blkptr, db->db_blkptr->blk_birth); 21903d692628SSanjeev Bagewadi 21913d692628SSanjeev Bagewadi return (res); 21923d692628SSanjeev Bagewadi } 21933d692628SSanjeev Bagewadi 2194*80901aeaSGeorge Wilson blkptr_t * 2195*80901aeaSGeorge Wilson dmu_buf_get_blkptr(dmu_buf_t *db) 2196*80901aeaSGeorge Wilson { 2197*80901aeaSGeorge Wilson dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2198*80901aeaSGeorge Wilson return (dbi->db_blkptr); 2199*80901aeaSGeorge Wilson } 2200*80901aeaSGeorge Wilson 2201c717a561Smaybee static void 2202c717a561Smaybee dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2203fa9e4066Sahrens { 2204c717a561Smaybee /* ASSERT(dmu_tx_is_syncing(tx) */ 2205c717a561Smaybee ASSERT(MUTEX_HELD(&db->db_mtx)); 2206c717a561Smaybee 2207c717a561Smaybee if (db->db_blkptr != NULL) 2208c717a561Smaybee return; 2209c717a561Smaybee 22100a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 22110a586ceaSMark Shellenbaum db->db_blkptr = &dn->dn_phys->dn_spill; 22120a586ceaSMark Shellenbaum BP_ZERO(db->db_blkptr); 22130a586ceaSMark Shellenbaum return; 22140a586ceaSMark Shellenbaum } 2215c717a561Smaybee if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2216c717a561Smaybee /* 2217c717a561Smaybee * This buffer was allocated at a time when there was 2218c717a561Smaybee * no available blkptrs from the dnode, or it was 2219c717a561Smaybee * inappropriate to hook it in (i.e., nlevels mis-match). 2220c717a561Smaybee */ 2221c717a561Smaybee ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2222c717a561Smaybee ASSERT(db->db_parent == NULL); 2223c717a561Smaybee db->db_parent = dn->dn_dbuf; 2224c717a561Smaybee db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2225c717a561Smaybee DBUF_VERIFY(db); 2226c717a561Smaybee } else { 2227c717a561Smaybee dmu_buf_impl_t *parent = db->db_parent; 2228c717a561Smaybee int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2229c717a561Smaybee 2230c717a561Smaybee ASSERT(dn->dn_phys->dn_nlevels > 1); 2231c717a561Smaybee if (parent == NULL) { 2232c717a561Smaybee mutex_exit(&db->db_mtx); 2233c717a561Smaybee rw_enter(&dn->dn_struct_rwlock, RW_READER); 2234c717a561Smaybee (void) dbuf_hold_impl(dn, db->db_level+1, 2235c717a561Smaybee db->db_blkid >> epbs, FALSE, db, &parent); 2236c717a561Smaybee rw_exit(&dn->dn_struct_rwlock); 2237c717a561Smaybee mutex_enter(&db->db_mtx); 2238c717a561Smaybee db->db_parent = parent; 2239c717a561Smaybee } 2240c717a561Smaybee db->db_blkptr = (blkptr_t *)parent->db.db_data + 2241c717a561Smaybee (db->db_blkid & ((1ULL << epbs) - 1)); 2242c717a561Smaybee DBUF_VERIFY(db); 2243c717a561Smaybee } 2244c717a561Smaybee } 2245c717a561Smaybee 2246c717a561Smaybee static void 2247c717a561Smaybee dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2248c717a561Smaybee { 2249c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 2250744947dcSTom Erickson dnode_t *dn; 2251c717a561Smaybee zio_t *zio; 2252c717a561Smaybee 2253c717a561Smaybee ASSERT(dmu_tx_is_syncing(tx)); 2254c717a561Smaybee 2255c717a561Smaybee dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2256c717a561Smaybee 2257c717a561Smaybee mutex_enter(&db->db_mtx); 2258c717a561Smaybee 2259c717a561Smaybee ASSERT(db->db_level > 0); 2260c717a561Smaybee DBUF_VERIFY(db); 2261c717a561Smaybee 2262c717a561Smaybee if (db->db_buf == NULL) { 2263c717a561Smaybee mutex_exit(&db->db_mtx); 2264c717a561Smaybee (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2265c717a561Smaybee mutex_enter(&db->db_mtx); 2266c717a561Smaybee } 2267c717a561Smaybee ASSERT3U(db->db_state, ==, DB_CACHED); 2268c717a561Smaybee ASSERT(db->db_buf != NULL); 2269c717a561Smaybee 2270744947dcSTom Erickson DB_DNODE_ENTER(db); 2271744947dcSTom Erickson dn = DB_DNODE(db); 2272744947dcSTom Erickson ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2273c717a561Smaybee dbuf_check_blkptr(dn, db); 2274744947dcSTom Erickson DB_DNODE_EXIT(db); 2275c717a561Smaybee 2276c717a561Smaybee db->db_data_pending = dr; 2277c717a561Smaybee 2278af2c4821Smaybee mutex_exit(&db->db_mtx); 2279088f3894Sahrens dbuf_write(dr, db->db_buf, tx); 2280c717a561Smaybee 2281c717a561Smaybee zio = dr->dr_zio; 2282c717a561Smaybee mutex_enter(&dr->dt.di.dr_mtx); 2283c717a561Smaybee dbuf_sync_list(&dr->dt.di.dr_children, tx); 2284c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2285c717a561Smaybee mutex_exit(&dr->dt.di.dr_mtx); 2286c717a561Smaybee zio_nowait(zio); 2287c717a561Smaybee } 2288c717a561Smaybee 2289c717a561Smaybee static void 2290c717a561Smaybee dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2291c717a561Smaybee { 2292c717a561Smaybee arc_buf_t **datap = &dr->dt.dl.dr_data; 2293c717a561Smaybee dmu_buf_impl_t *db = dr->dr_dbuf; 2294744947dcSTom Erickson dnode_t *dn; 2295744947dcSTom Erickson objset_t *os; 2296c717a561Smaybee uint64_t txg = tx->tx_txg; 2297fa9e4066Sahrens 2298fa9e4066Sahrens ASSERT(dmu_tx_is_syncing(tx)); 2299fa9e4066Sahrens 2300fa9e4066Sahrens dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2301fa9e4066Sahrens 2302fa9e4066Sahrens mutex_enter(&db->db_mtx); 2303fa9e4066Sahrens /* 2304fa9e4066Sahrens * To be synced, we must be dirtied. But we 2305fa9e4066Sahrens * might have been freed after the dirty. 2306fa9e4066Sahrens */ 2307fa9e4066Sahrens if (db->db_state == DB_UNCACHED) { 2308fa9e4066Sahrens /* This buffer has been freed since it was dirtied */ 2309fa9e4066Sahrens ASSERT(db->db.db_data == NULL); 2310fa9e4066Sahrens } else if (db->db_state == DB_FILL) { 2311fa9e4066Sahrens /* This buffer was freed and is now being re-filled */ 2312c717a561Smaybee ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2313fa9e4066Sahrens } else { 231482c9918fSTim Haley ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2315fa9e4066Sahrens } 23169c9dc39aSek110237 DBUF_VERIFY(db); 2317fa9e4066Sahrens 2318744947dcSTom Erickson DB_DNODE_ENTER(db); 2319744947dcSTom Erickson dn = DB_DNODE(db); 2320744947dcSTom Erickson 23210a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 23220a586ceaSMark Shellenbaum mutex_enter(&dn->dn_mtx); 23230a586ceaSMark Shellenbaum dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 23240a586ceaSMark Shellenbaum mutex_exit(&dn->dn_mtx); 23250a586ceaSMark Shellenbaum } 23260a586ceaSMark Shellenbaum 2327fa9e4066Sahrens /* 2328c717a561Smaybee * If this is a bonus buffer, simply copy the bonus data into the 2329c717a561Smaybee * dnode. It will be written out when the dnode is synced (and it 2330c717a561Smaybee * will be synced, since it must have been dirty for dbuf_sync to 2331c717a561Smaybee * be called). 2332fa9e4066Sahrens */ 23330a586ceaSMark Shellenbaum if (db->db_blkid == DMU_BONUS_BLKID) { 2334c717a561Smaybee dbuf_dirty_record_t **drp; 23351934e92fSmaybee 2336ea8dc4b6Seschrock ASSERT(*datap != NULL); 2337fb09f5aaSMadhav Suresh ASSERT0(db->db_level); 2338ea8dc4b6Seschrock ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2339ea8dc4b6Seschrock bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2340744947dcSTom Erickson DB_DNODE_EXIT(db); 2341744947dcSTom Erickson 23420e8c6158Smaybee if (*datap != db->db.db_data) { 2343ea8dc4b6Seschrock zio_buf_free(*datap, DN_MAX_BONUSLEN); 23445a98e54bSBrendan Gregg - Sun Microsystems arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 23450e8c6158Smaybee } 2346ea8dc4b6Seschrock db->db_data_pending = NULL; 2347c717a561Smaybee drp = &db->db_last_dirty; 2348c717a561Smaybee while (*drp != dr) 2349c717a561Smaybee drp = &(*drp)->dr_next; 235017f17c2dSbonwick ASSERT(dr->dr_next == NULL); 2351b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 235217f17c2dSbonwick *drp = dr->dr_next; 2353c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2354ea8dc4b6Seschrock ASSERT(db->db_dirtycnt > 0); 2355ea8dc4b6Seschrock db->db_dirtycnt -= 1; 2356b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2357ea8dc4b6Seschrock return; 2358ea8dc4b6Seschrock } 2359ea8dc4b6Seschrock 2360744947dcSTom Erickson os = dn->dn_objset; 2361744947dcSTom Erickson 2362c5c6ffa0Smaybee /* 2363f82bfe17Sgw25295 * This function may have dropped the db_mtx lock allowing a dmu_sync 2364f82bfe17Sgw25295 * operation to sneak in. As a result, we need to ensure that we 2365f82bfe17Sgw25295 * don't check the dr_override_state until we have returned from 2366f82bfe17Sgw25295 * dbuf_check_blkptr. 2367f82bfe17Sgw25295 */ 2368f82bfe17Sgw25295 dbuf_check_blkptr(dn, db); 2369f82bfe17Sgw25295 2370f82bfe17Sgw25295 /* 2371744947dcSTom Erickson * If this buffer is in the middle of an immediate write, 2372c717a561Smaybee * wait for the synchronous IO to complete. 2373c5c6ffa0Smaybee */ 2374c717a561Smaybee while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2375c5c6ffa0Smaybee ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2376c5c6ffa0Smaybee cv_wait(&db->db_changed, &db->db_mtx); 2377c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2378c5c6ffa0Smaybee } 2379c717a561Smaybee 2380ab69d62fSMatthew Ahrens if (db->db_state != DB_NOFILL && 2381ab69d62fSMatthew Ahrens dn->dn_object != DMU_META_DNODE_OBJECT && 2382ab69d62fSMatthew Ahrens refcount_count(&db->db_holds) > 1 && 2383b24ab676SJeff Bonwick dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2384ab69d62fSMatthew Ahrens *datap == db->db_buf) { 2385fa9e4066Sahrens /* 238682c9918fSTim Haley * If this buffer is currently "in use" (i.e., there 238782c9918fSTim Haley * are active holds and db_data still references it), 238882c9918fSTim Haley * then make a copy before we start the write so that 238982c9918fSTim Haley * any modifications from the open txg will not leak 239082c9918fSTim Haley * into this write. 2391fa9e4066Sahrens * 239282c9918fSTim Haley * NOTE: this copy does not need to be made for 239382c9918fSTim Haley * objects only modified in the syncing context (e.g. 239482c9918fSTim Haley * DNONE_DNODE blocks). 2395fa9e4066Sahrens */ 2396ab69d62fSMatthew Ahrens int blksz = arc_buf_size(*datap); 2397ab69d62fSMatthew Ahrens arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2398ab69d62fSMatthew Ahrens *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2399c717a561Smaybee bcopy(db->db.db_data, (*datap)->b_data, blksz); 2400fa9e4066Sahrens } 2401c717a561Smaybee db->db_data_pending = dr; 2402fa9e4066Sahrens 2403fa9e4066Sahrens mutex_exit(&db->db_mtx); 2404fa9e4066Sahrens 2405088f3894Sahrens dbuf_write(dr, *datap, tx); 2406c717a561Smaybee 2407c717a561Smaybee ASSERT(!list_link_active(&dr->dr_dirty_node)); 2408744947dcSTom Erickson if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2409c717a561Smaybee list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2410744947dcSTom Erickson DB_DNODE_EXIT(db); 2411744947dcSTom Erickson } else { 2412744947dcSTom Erickson /* 2413744947dcSTom Erickson * Although zio_nowait() does not "wait for an IO", it does 2414744947dcSTom Erickson * initiate the IO. If this is an empty write it seems plausible 2415744947dcSTom Erickson * that the IO could actually be completed before the nowait 2416744947dcSTom Erickson * returns. We need to DB_DNODE_EXIT() first in case 2417744947dcSTom Erickson * zio_nowait() invalidates the dbuf. 2418744947dcSTom Erickson */ 2419744947dcSTom Erickson DB_DNODE_EXIT(db); 2420c717a561Smaybee zio_nowait(dr->dr_zio); 2421fa9e4066Sahrens } 2422744947dcSTom Erickson } 2423c717a561Smaybee 2424c717a561Smaybee void 2425c717a561Smaybee dbuf_sync_list(list_t *list, dmu_tx_t *tx) 2426c717a561Smaybee { 2427c717a561Smaybee dbuf_dirty_record_t *dr; 2428c717a561Smaybee 2429c717a561Smaybee while (dr = list_head(list)) { 2430c717a561Smaybee if (dr->dr_zio != NULL) { 2431c717a561Smaybee /* 2432c717a561Smaybee * If we find an already initialized zio then we 2433c717a561Smaybee * are processing the meta-dnode, and we have finished. 2434c717a561Smaybee * The dbufs for all dnodes are put back on the list 2435c717a561Smaybee * during processing, so that we can zio_wait() 2436c717a561Smaybee * these IOs after initiating all child IOs. 2437c717a561Smaybee */ 2438c717a561Smaybee ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2439c717a561Smaybee DMU_META_DNODE_OBJECT); 2440c717a561Smaybee break; 2441fa9e4066Sahrens } 2442c717a561Smaybee list_remove(list, dr); 2443c717a561Smaybee if (dr->dr_dbuf->db_level > 0) 2444c717a561Smaybee dbuf_sync_indirect(dr, tx); 2445c717a561Smaybee else 2446c717a561Smaybee dbuf_sync_leaf(dr, tx); 2447c717a561Smaybee } 2448c717a561Smaybee } 2449c717a561Smaybee 2450fa9e4066Sahrens /* ARGSUSED */ 2451fa9e4066Sahrens static void 2452c717a561Smaybee dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2453fa9e4066Sahrens { 2454fa9e4066Sahrens dmu_buf_impl_t *db = vdb; 2455744947dcSTom Erickson dnode_t *dn; 2456e14bb325SJeff Bonwick blkptr_t *bp = zio->io_bp; 2457c717a561Smaybee blkptr_t *bp_orig = &zio->io_bp_orig; 2458b24ab676SJeff Bonwick spa_t *spa = zio->io_spa; 2459b24ab676SJeff Bonwick int64_t delta; 2460fa9e4066Sahrens uint64_t fill = 0; 2461b24ab676SJeff Bonwick int i; 2462fa9e4066Sahrens 2463e14bb325SJeff Bonwick ASSERT(db->db_blkptr == bp); 2464e14bb325SJeff Bonwick 2465744947dcSTom Erickson DB_DNODE_ENTER(db); 2466744947dcSTom Erickson dn = DB_DNODE(db); 2467b24ab676SJeff Bonwick delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2468b24ab676SJeff Bonwick dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2469b24ab676SJeff Bonwick zio->io_prev_space_delta = delta; 2470fa9e4066Sahrens 2471e14bb325SJeff Bonwick if (BP_IS_HOLE(bp)) { 2472b24ab676SJeff Bonwick ASSERT(bp->blk_fill == 0); 2473744947dcSTom Erickson DB_DNODE_EXIT(db); 2474c717a561Smaybee return; 2475c717a561Smaybee } 2476c717a561Smaybee 24770a586ceaSMark Shellenbaum ASSERT((db->db_blkid != DMU_SPILL_BLKID && 24780a586ceaSMark Shellenbaum BP_GET_TYPE(bp) == dn->dn_type) || 24790a586ceaSMark Shellenbaum (db->db_blkid == DMU_SPILL_BLKID && 24800a586ceaSMark Shellenbaum BP_GET_TYPE(bp) == dn->dn_bonustype)); 2481e14bb325SJeff Bonwick ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2482e14bb325SJeff Bonwick 2483fa9e4066Sahrens mutex_enter(&db->db_mtx); 2484fa9e4066Sahrens 24850a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG 24860a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 24870a586ceaSMark Shellenbaum ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 24880a586ceaSMark Shellenbaum ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 24890a586ceaSMark Shellenbaum db->db_blkptr == &dn->dn_phys->dn_spill); 24900a586ceaSMark Shellenbaum } 24910a586ceaSMark Shellenbaum #endif 24920a586ceaSMark Shellenbaum 2493fa9e4066Sahrens if (db->db_level == 0) { 2494fa9e4066Sahrens mutex_enter(&dn->dn_mtx); 24950a586ceaSMark Shellenbaum if (db->db_blkid > dn->dn_phys->dn_maxblkid && 24960a586ceaSMark Shellenbaum db->db_blkid != DMU_SPILL_BLKID) 2497fa9e4066Sahrens dn->dn_phys->dn_maxblkid = db->db_blkid; 2498fa9e4066Sahrens mutex_exit(&dn->dn_mtx); 2499fa9e4066Sahrens 2500fa9e4066Sahrens if (dn->dn_type == DMU_OT_DNODE) { 2501fa9e4066Sahrens dnode_phys_t *dnp = db->db.db_data; 2502fa9e4066Sahrens for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2503fa9e4066Sahrens i--, dnp++) { 2504fa9e4066Sahrens if (dnp->dn_type != DMU_OT_NONE) 2505fa9e4066Sahrens fill++; 2506fa9e4066Sahrens } 2507fa9e4066Sahrens } else { 2508fa9e4066Sahrens fill = 1; 2509fa9e4066Sahrens } 2510fa9e4066Sahrens } else { 2511e14bb325SJeff Bonwick blkptr_t *ibp = db->db.db_data; 2512fa9e4066Sahrens ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2513e14bb325SJeff Bonwick for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2514e14bb325SJeff Bonwick if (BP_IS_HOLE(ibp)) 2515fa9e4066Sahrens continue; 2516e14bb325SJeff Bonwick fill += ibp->blk_fill; 2517fa9e4066Sahrens } 2518fa9e4066Sahrens } 2519744947dcSTom Erickson DB_DNODE_EXIT(db); 2520fa9e4066Sahrens 2521e14bb325SJeff Bonwick bp->blk_fill = fill; 2522fa9e4066Sahrens 2523fa9e4066Sahrens mutex_exit(&db->db_mtx); 2524fa9e4066Sahrens } 2525fa9e4066Sahrens 2526c717a561Smaybee /* ARGSUSED */ 2527c717a561Smaybee static void 2528c717a561Smaybee dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2529c717a561Smaybee { 2530c717a561Smaybee dmu_buf_impl_t *db = vdb; 2531b24ab676SJeff Bonwick blkptr_t *bp = zio->io_bp; 2532b24ab676SJeff Bonwick blkptr_t *bp_orig = &zio->io_bp_orig; 2533c717a561Smaybee uint64_t txg = zio->io_txg; 2534c717a561Smaybee dbuf_dirty_record_t **drp, *dr; 2535c717a561Smaybee 2536fb09f5aaSMadhav Suresh ASSERT0(zio->io_error); 2537b24ab676SJeff Bonwick ASSERT(db->db_blkptr == bp); 2538b24ab676SJeff Bonwick 2539*80901aeaSGeorge Wilson /* 2540*80901aeaSGeorge Wilson * For nopwrites and rewrites we ensure that the bp matches our 2541*80901aeaSGeorge Wilson * original and bypass all the accounting. 2542*80901aeaSGeorge Wilson */ 2543*80901aeaSGeorge Wilson if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 2544b24ab676SJeff Bonwick ASSERT(BP_EQUAL(bp, bp_orig)); 2545b24ab676SJeff Bonwick } else { 2546744947dcSTom Erickson objset_t *os; 2547744947dcSTom Erickson dsl_dataset_t *ds; 2548744947dcSTom Erickson dmu_tx_t *tx; 2549744947dcSTom Erickson 2550744947dcSTom Erickson DB_GET_OBJSET(&os, db); 2551744947dcSTom Erickson ds = os->os_dsl_dataset; 2552744947dcSTom Erickson tx = os->os_synctx; 2553b24ab676SJeff Bonwick 2554b24ab676SJeff Bonwick (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2555b24ab676SJeff Bonwick dsl_dataset_block_born(ds, bp, tx); 2556b24ab676SJeff Bonwick } 2557c717a561Smaybee 2558c717a561Smaybee mutex_enter(&db->db_mtx); 2559c717a561Smaybee 2560b24ab676SJeff Bonwick DBUF_VERIFY(db); 2561b24ab676SJeff Bonwick 2562c717a561Smaybee drp = &db->db_last_dirty; 256317f17c2dSbonwick while ((dr = *drp) != db->db_data_pending) 256417f17c2dSbonwick drp = &dr->dr_next; 256517f17c2dSbonwick ASSERT(!list_link_active(&dr->dr_dirty_node)); 256617f17c2dSbonwick ASSERT(dr->dr_txg == txg); 2567b24ab676SJeff Bonwick ASSERT(dr->dr_dbuf == db); 256817f17c2dSbonwick ASSERT(dr->dr_next == NULL); 256917f17c2dSbonwick *drp = dr->dr_next; 2570c717a561Smaybee 25710a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG 25720a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) { 2573744947dcSTom Erickson dnode_t *dn; 2574744947dcSTom Erickson 2575744947dcSTom Erickson DB_DNODE_ENTER(db); 2576744947dcSTom Erickson dn = DB_DNODE(db); 25770a586ceaSMark Shellenbaum ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 25780a586ceaSMark Shellenbaum ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 25790a586ceaSMark Shellenbaum db->db_blkptr == &dn->dn_phys->dn_spill); 2580744947dcSTom Erickson DB_DNODE_EXIT(db); 25810a586ceaSMark Shellenbaum } 25820a586ceaSMark Shellenbaum #endif 25830a586ceaSMark Shellenbaum 2584c717a561Smaybee if (db->db_level == 0) { 25850a586ceaSMark Shellenbaum ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2586c717a561Smaybee ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 258782c9918fSTim Haley if (db->db_state != DB_NOFILL) { 2588c717a561Smaybee if (dr->dt.dl.dr_data != db->db_buf) 258982c9918fSTim Haley VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 259082c9918fSTim Haley db) == 1); 2591b24ab676SJeff Bonwick else if (!arc_released(db->db_buf)) 2592c717a561Smaybee arc_set_callback(db->db_buf, dbuf_do_evict, db); 259382c9918fSTim Haley } 2594c717a561Smaybee } else { 2595744947dcSTom Erickson dnode_t *dn; 2596744947dcSTom Erickson 2597744947dcSTom Erickson DB_DNODE_ENTER(db); 2598744947dcSTom Erickson dn = DB_DNODE(db); 2599c717a561Smaybee ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2600c717a561Smaybee ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2601c717a561Smaybee if (!BP_IS_HOLE(db->db_blkptr)) { 2602c717a561Smaybee int epbs = 2603c717a561Smaybee dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2604c717a561Smaybee ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2605c717a561Smaybee db->db.db_size); 2606c717a561Smaybee ASSERT3U(dn->dn_phys->dn_maxblkid 2607c717a561Smaybee >> (db->db_level * epbs), >=, db->db_blkid); 2608c717a561Smaybee arc_set_callback(db->db_buf, dbuf_do_evict, db); 2609c717a561Smaybee } 2610744947dcSTom Erickson DB_DNODE_EXIT(db); 2611c25056deSgw25295 mutex_destroy(&dr->dt.di.dr_mtx); 2612c25056deSgw25295 list_destroy(&dr->dt.di.dr_children); 2613c717a561Smaybee } 2614c717a561Smaybee kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2615c717a561Smaybee 2616c717a561Smaybee cv_broadcast(&db->db_changed); 2617c717a561Smaybee ASSERT(db->db_dirtycnt > 0); 2618c717a561Smaybee db->db_dirtycnt -= 1; 2619c717a561Smaybee db->db_data_pending = NULL; 2620b24ab676SJeff Bonwick dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2621b24ab676SJeff Bonwick } 2622b24ab676SJeff Bonwick 2623b24ab676SJeff Bonwick static void 2624b24ab676SJeff Bonwick dbuf_write_nofill_ready(zio_t *zio) 2625b24ab676SJeff Bonwick { 2626b24ab676SJeff Bonwick dbuf_write_ready(zio, NULL, zio->io_private); 2627b24ab676SJeff Bonwick } 2628b24ab676SJeff Bonwick 2629b24ab676SJeff Bonwick static void 2630b24ab676SJeff Bonwick dbuf_write_nofill_done(zio_t *zio) 2631b24ab676SJeff Bonwick { 2632b24ab676SJeff Bonwick dbuf_write_done(zio, NULL, zio->io_private); 2633b24ab676SJeff Bonwick } 2634b24ab676SJeff Bonwick 2635b24ab676SJeff Bonwick static void 2636b24ab676SJeff Bonwick dbuf_write_override_ready(zio_t *zio) 2637b24ab676SJeff Bonwick { 2638b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = zio->io_private; 2639b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 2640b24ab676SJeff Bonwick 2641b24ab676SJeff Bonwick dbuf_write_ready(zio, NULL, db); 2642b24ab676SJeff Bonwick } 2643b24ab676SJeff Bonwick 2644b24ab676SJeff Bonwick static void 2645b24ab676SJeff Bonwick dbuf_write_override_done(zio_t *zio) 2646b24ab676SJeff Bonwick { 2647b24ab676SJeff Bonwick dbuf_dirty_record_t *dr = zio->io_private; 2648b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 2649b24ab676SJeff Bonwick blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 2650b24ab676SJeff Bonwick 2651b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 2652b24ab676SJeff Bonwick if (!BP_EQUAL(zio->io_bp, obp)) { 2653b24ab676SJeff Bonwick if (!BP_IS_HOLE(obp)) 2654b24ab676SJeff Bonwick dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 2655b24ab676SJeff Bonwick arc_release(dr->dt.dl.dr_data, db); 2656b24ab676SJeff Bonwick } 2657c717a561Smaybee mutex_exit(&db->db_mtx); 2658c717a561Smaybee 2659b24ab676SJeff Bonwick dbuf_write_done(zio, NULL, db); 2660b24ab676SJeff Bonwick } 2661c717a561Smaybee 2662b24ab676SJeff Bonwick static void 2663b24ab676SJeff Bonwick dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 2664b24ab676SJeff Bonwick { 2665b24ab676SJeff Bonwick dmu_buf_impl_t *db = dr->dr_dbuf; 2666744947dcSTom Erickson dnode_t *dn; 2667744947dcSTom Erickson objset_t *os; 2668b24ab676SJeff Bonwick dmu_buf_impl_t *parent = db->db_parent; 2669b24ab676SJeff Bonwick uint64_t txg = tx->tx_txg; 2670b24ab676SJeff Bonwick zbookmark_t zb; 2671b24ab676SJeff Bonwick zio_prop_t zp; 2672b24ab676SJeff Bonwick zio_t *zio; 26730a586ceaSMark Shellenbaum int wp_flag = 0; 2674b24ab676SJeff Bonwick 2675744947dcSTom Erickson DB_DNODE_ENTER(db); 2676744947dcSTom Erickson dn = DB_DNODE(db); 2677744947dcSTom Erickson os = dn->dn_objset; 2678744947dcSTom Erickson 2679b24ab676SJeff Bonwick if (db->db_state != DB_NOFILL) { 2680b24ab676SJeff Bonwick if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 2681b24ab676SJeff Bonwick /* 2682b24ab676SJeff Bonwick * Private object buffers are released here rather 2683b24ab676SJeff Bonwick * than in dbuf_dirty() since they are only modified 2684b24ab676SJeff Bonwick * in the syncing context and we don't want the 2685b24ab676SJeff Bonwick * overhead of making multiple copies of the data. 2686b24ab676SJeff Bonwick */ 2687b24ab676SJeff Bonwick if (BP_IS_HOLE(db->db_blkptr)) { 2688b24ab676SJeff Bonwick arc_buf_thaw(data); 2689b24ab676SJeff Bonwick } else { 26903f9d6ad7SLin Ling dbuf_release_bp(db); 2691b24ab676SJeff Bonwick } 2692b24ab676SJeff Bonwick } 2693b24ab676SJeff Bonwick } 2694b24ab676SJeff Bonwick 2695b24ab676SJeff Bonwick if (parent != dn->dn_dbuf) { 2696b24ab676SJeff Bonwick ASSERT(parent && parent->db_data_pending); 2697b24ab676SJeff Bonwick ASSERT(db->db_level == parent->db_level-1); 2698b24ab676SJeff Bonwick ASSERT(arc_released(parent->db_buf)); 2699b24ab676SJeff Bonwick zio = parent->db_data_pending->dr_zio; 2700b24ab676SJeff Bonwick } else { 27010a586ceaSMark Shellenbaum ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 27020a586ceaSMark Shellenbaum db->db_blkid != DMU_SPILL_BLKID) || 27030a586ceaSMark Shellenbaum (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 27040a586ceaSMark Shellenbaum if (db->db_blkid != DMU_SPILL_BLKID) 2705b24ab676SJeff Bonwick ASSERT3P(db->db_blkptr, ==, 2706b24ab676SJeff Bonwick &dn->dn_phys->dn_blkptr[db->db_blkid]); 2707b24ab676SJeff Bonwick zio = dn->dn_zio; 2708b24ab676SJeff Bonwick } 2709b24ab676SJeff Bonwick 2710b24ab676SJeff Bonwick ASSERT(db->db_level == 0 || data == db->db_buf); 2711b24ab676SJeff Bonwick ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2712b24ab676SJeff Bonwick ASSERT(zio); 2713b24ab676SJeff Bonwick 2714b24ab676SJeff Bonwick SET_BOOKMARK(&zb, os->os_dsl_dataset ? 2715b24ab676SJeff Bonwick os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 2716b24ab676SJeff Bonwick db->db.db_object, db->db_level, db->db_blkid); 2717b24ab676SJeff Bonwick 27180a586ceaSMark Shellenbaum if (db->db_blkid == DMU_SPILL_BLKID) 27190a586ceaSMark Shellenbaum wp_flag = WP_SPILL; 27200a586ceaSMark Shellenbaum wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 27210a586ceaSMark Shellenbaum 27220a586ceaSMark Shellenbaum dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 2723744947dcSTom Erickson DB_DNODE_EXIT(db); 2724b24ab676SJeff Bonwick 2725b24ab676SJeff Bonwick if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2726b24ab676SJeff Bonwick ASSERT(db->db_state != DB_NOFILL); 2727b24ab676SJeff Bonwick dr->dr_zio = zio_write(zio, os->os_spa, txg, 2728b24ab676SJeff Bonwick db->db_blkptr, data->b_data, arc_buf_size(data), &zp, 2729b24ab676SJeff Bonwick dbuf_write_override_ready, dbuf_write_override_done, dr, 2730b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2731b24ab676SJeff Bonwick mutex_enter(&db->db_mtx); 2732b24ab676SJeff Bonwick dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 2733b24ab676SJeff Bonwick zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 2734*80901aeaSGeorge Wilson dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 2735b24ab676SJeff Bonwick mutex_exit(&db->db_mtx); 2736b24ab676SJeff Bonwick } else if (db->db_state == DB_NOFILL) { 2737b24ab676SJeff Bonwick ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); 2738b24ab676SJeff Bonwick dr->dr_zio = zio_write(zio, os->os_spa, txg, 2739b24ab676SJeff Bonwick db->db_blkptr, NULL, db->db.db_size, &zp, 2740b24ab676SJeff Bonwick dbuf_write_nofill_ready, dbuf_write_nofill_done, db, 2741b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, 2742b24ab676SJeff Bonwick ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 2743b24ab676SJeff Bonwick } else { 2744b24ab676SJeff Bonwick ASSERT(arc_released(data)); 2745b24ab676SJeff Bonwick dr->dr_zio = arc_write(zio, os->os_spa, txg, 2746b24ab676SJeff Bonwick db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp, 2747b24ab676SJeff Bonwick dbuf_write_ready, dbuf_write_done, db, 2748b24ab676SJeff Bonwick ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2749b24ab676SJeff Bonwick } 2750fa9e4066Sahrens } 2751