1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 */ 24 25 #include <sys/zfs_context.h> 26 #include <sys/dmu.h> 27 #include <sys/dmu_impl.h> 28 #include <sys/dbuf.h> 29 #include <sys/dmu_objset.h> 30 #include <sys/dsl_dataset.h> 31 #include <sys/dsl_dir.h> 32 #include <sys/dmu_tx.h> 33 #include <sys/spa.h> 34 #include <sys/zio.h> 35 #include <sys/dmu_zfetch.h> 36 #include <sys/sa.h> 37 #include <sys/sa_impl.h> 38 39 static void dbuf_destroy(dmu_buf_impl_t *db); 40 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 41 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 42 43 /* 44 * Global data structures and functions for the dbuf cache. 45 */ 46 static kmem_cache_t *dbuf_cache; 47 48 /* ARGSUSED */ 49 static int 50 dbuf_cons(void *vdb, void *unused, int kmflag) 51 { 52 dmu_buf_impl_t *db = vdb; 53 bzero(db, sizeof (dmu_buf_impl_t)); 54 55 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 56 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 57 refcount_create(&db->db_holds); 58 return (0); 59 } 60 61 /* ARGSUSED */ 62 static void 63 dbuf_dest(void *vdb, void *unused) 64 { 65 dmu_buf_impl_t *db = vdb; 66 mutex_destroy(&db->db_mtx); 67 cv_destroy(&db->db_changed); 68 refcount_destroy(&db->db_holds); 69 } 70 71 /* 72 * dbuf hash table routines 73 */ 74 static dbuf_hash_table_t dbuf_hash_table; 75 76 static uint64_t dbuf_hash_count; 77 78 static uint64_t 79 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 80 { 81 uintptr_t osv = (uintptr_t)os; 82 uint64_t crc = -1ULL; 83 84 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 85 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 86 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 87 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 88 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 89 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 90 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 91 92 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 93 94 return (crc); 95 } 96 97 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 98 99 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 100 ((dbuf)->db.db_object == (obj) && \ 101 (dbuf)->db_objset == (os) && \ 102 (dbuf)->db_level == (level) && \ 103 (dbuf)->db_blkid == (blkid)) 104 105 dmu_buf_impl_t * 106 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 107 { 108 dbuf_hash_table_t *h = &dbuf_hash_table; 109 objset_t *os = dn->dn_objset; 110 uint64_t obj = dn->dn_object; 111 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 112 uint64_t idx = hv & h->hash_table_mask; 113 dmu_buf_impl_t *db; 114 115 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 116 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 117 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 118 mutex_enter(&db->db_mtx); 119 if (db->db_state != DB_EVICTING) { 120 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 121 return (db); 122 } 123 mutex_exit(&db->db_mtx); 124 } 125 } 126 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 127 return (NULL); 128 } 129 130 /* 131 * Insert an entry into the hash table. If there is already an element 132 * equal to elem in the hash table, then the already existing element 133 * will be returned and the new element will not be inserted. 134 * Otherwise returns NULL. 135 */ 136 static dmu_buf_impl_t * 137 dbuf_hash_insert(dmu_buf_impl_t *db) 138 { 139 dbuf_hash_table_t *h = &dbuf_hash_table; 140 objset_t *os = db->db_objset; 141 uint64_t obj = db->db.db_object; 142 int level = db->db_level; 143 uint64_t blkid = db->db_blkid; 144 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 145 uint64_t idx = hv & h->hash_table_mask; 146 dmu_buf_impl_t *dbf; 147 148 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 149 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 150 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 151 mutex_enter(&dbf->db_mtx); 152 if (dbf->db_state != DB_EVICTING) { 153 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 154 return (dbf); 155 } 156 mutex_exit(&dbf->db_mtx); 157 } 158 } 159 160 mutex_enter(&db->db_mtx); 161 db->db_hash_next = h->hash_table[idx]; 162 h->hash_table[idx] = db; 163 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 164 atomic_add_64(&dbuf_hash_count, 1); 165 166 return (NULL); 167 } 168 169 /* 170 * Remove an entry from the hash table. This operation will 171 * fail if there are any existing holds on the db. 172 */ 173 static void 174 dbuf_hash_remove(dmu_buf_impl_t *db) 175 { 176 dbuf_hash_table_t *h = &dbuf_hash_table; 177 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 178 db->db_level, db->db_blkid); 179 uint64_t idx = hv & h->hash_table_mask; 180 dmu_buf_impl_t *dbf, **dbp; 181 182 /* 183 * We musn't hold db_mtx to maintin lock ordering: 184 * DBUF_HASH_MUTEX > db_mtx. 185 */ 186 ASSERT(refcount_is_zero(&db->db_holds)); 187 ASSERT(db->db_state == DB_EVICTING); 188 ASSERT(!MUTEX_HELD(&db->db_mtx)); 189 190 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 191 dbp = &h->hash_table[idx]; 192 while ((dbf = *dbp) != db) { 193 dbp = &dbf->db_hash_next; 194 ASSERT(dbf != NULL); 195 } 196 *dbp = db->db_hash_next; 197 db->db_hash_next = NULL; 198 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 199 atomic_add_64(&dbuf_hash_count, -1); 200 } 201 202 static arc_evict_func_t dbuf_do_evict; 203 204 static void 205 dbuf_evict_user(dmu_buf_impl_t *db) 206 { 207 ASSERT(MUTEX_HELD(&db->db_mtx)); 208 209 if (db->db_level != 0 || db->db_evict_func == NULL) 210 return; 211 212 if (db->db_user_data_ptr_ptr) 213 *db->db_user_data_ptr_ptr = db->db.db_data; 214 db->db_evict_func(&db->db, db->db_user_ptr); 215 db->db_user_ptr = NULL; 216 db->db_user_data_ptr_ptr = NULL; 217 db->db_evict_func = NULL; 218 } 219 220 void 221 dbuf_evict(dmu_buf_impl_t *db) 222 { 223 ASSERT(MUTEX_HELD(&db->db_mtx)); 224 ASSERT(db->db_buf == NULL); 225 ASSERT(db->db_data_pending == NULL); 226 227 dbuf_clear(db); 228 dbuf_destroy(db); 229 } 230 231 void 232 dbuf_init(void) 233 { 234 uint64_t hsize = 1ULL << 16; 235 dbuf_hash_table_t *h = &dbuf_hash_table; 236 int i; 237 238 /* 239 * The hash table is big enough to fill all of physical memory 240 * with an average 4K block size. The table will take up 241 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 242 */ 243 while (hsize * 4096 < physmem * PAGESIZE) 244 hsize <<= 1; 245 246 retry: 247 h->hash_table_mask = hsize - 1; 248 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 249 if (h->hash_table == NULL) { 250 /* XXX - we should really return an error instead of assert */ 251 ASSERT(hsize > (1ULL << 10)); 252 hsize >>= 1; 253 goto retry; 254 } 255 256 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 257 sizeof (dmu_buf_impl_t), 258 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 259 260 for (i = 0; i < DBUF_MUTEXES; i++) 261 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 262 } 263 264 void 265 dbuf_fini(void) 266 { 267 dbuf_hash_table_t *h = &dbuf_hash_table; 268 int i; 269 270 for (i = 0; i < DBUF_MUTEXES; i++) 271 mutex_destroy(&h->hash_mutexes[i]); 272 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 273 kmem_cache_destroy(dbuf_cache); 274 } 275 276 /* 277 * Other stuff. 278 */ 279 280 #ifdef ZFS_DEBUG 281 static void 282 dbuf_verify(dmu_buf_impl_t *db) 283 { 284 dnode_t *dn = db->db_dnode; 285 dbuf_dirty_record_t *dr; 286 287 ASSERT(MUTEX_HELD(&db->db_mtx)); 288 289 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 290 return; 291 292 ASSERT(db->db_objset != NULL); 293 if (dn == NULL) { 294 ASSERT(db->db_parent == NULL); 295 ASSERT(db->db_blkptr == NULL); 296 } else { 297 ASSERT3U(db->db.db_object, ==, dn->dn_object); 298 ASSERT3P(db->db_objset, ==, dn->dn_objset); 299 ASSERT3U(db->db_level, <, dn->dn_nlevels); 300 ASSERT(db->db_blkid == DMU_BONUS_BLKID || db->db_blkid == 301 DMU_SPILL_BLKID || list_head(&dn->dn_dbufs)); 302 } 303 if (db->db_blkid == DMU_BONUS_BLKID) { 304 ASSERT(dn != NULL); 305 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 306 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 307 } else if (db->db_blkid == DMU_SPILL_BLKID) { 308 ASSERT(dn != NULL); 309 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 310 ASSERT3U(db->db.db_offset, ==, 0); 311 } else { 312 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 313 } 314 315 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 316 ASSERT(dr->dr_dbuf == db); 317 318 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 319 ASSERT(dr->dr_dbuf == db); 320 321 /* 322 * We can't assert that db_size matches dn_datablksz because it 323 * can be momentarily different when another thread is doing 324 * dnode_set_blksz(). 325 */ 326 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 327 dr = db->db_data_pending; 328 /* 329 * It should only be modified in syncing context, so 330 * make sure we only have one copy of the data. 331 */ 332 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 333 } 334 335 /* verify db->db_blkptr */ 336 if (db->db_blkptr) { 337 if (db->db_parent == dn->dn_dbuf) { 338 /* db is pointed to by the dnode */ 339 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 340 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 341 ASSERT(db->db_parent == NULL); 342 else 343 ASSERT(db->db_parent != NULL); 344 if (db->db_blkid != DMU_SPILL_BLKID) 345 ASSERT3P(db->db_blkptr, ==, 346 &dn->dn_phys->dn_blkptr[db->db_blkid]); 347 } else { 348 /* db is pointed to by an indirect block */ 349 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 350 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 351 ASSERT3U(db->db_parent->db.db_object, ==, 352 db->db.db_object); 353 /* 354 * dnode_grow_indblksz() can make this fail if we don't 355 * have the struct_rwlock. XXX indblksz no longer 356 * grows. safe to do this now? 357 */ 358 if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) { 359 ASSERT3P(db->db_blkptr, ==, 360 ((blkptr_t *)db->db_parent->db.db_data + 361 db->db_blkid % epb)); 362 } 363 } 364 } 365 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 366 (db->db_buf == NULL || db->db_buf->b_data) && 367 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 368 db->db_state != DB_FILL && !dn->dn_free_txg) { 369 /* 370 * If the blkptr isn't set but they have nonzero data, 371 * it had better be dirty, otherwise we'll lose that 372 * data when we evict this buffer. 373 */ 374 if (db->db_dirtycnt == 0) { 375 uint64_t *buf = db->db.db_data; 376 int i; 377 378 for (i = 0; i < db->db.db_size >> 3; i++) { 379 ASSERT(buf[i] == 0); 380 } 381 } 382 } 383 } 384 #endif 385 386 static void 387 dbuf_update_data(dmu_buf_impl_t *db) 388 { 389 ASSERT(MUTEX_HELD(&db->db_mtx)); 390 if (db->db_level == 0 && db->db_user_data_ptr_ptr) { 391 ASSERT(!refcount_is_zero(&db->db_holds)); 392 *db->db_user_data_ptr_ptr = db->db.db_data; 393 } 394 } 395 396 static void 397 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 398 { 399 ASSERT(MUTEX_HELD(&db->db_mtx)); 400 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 401 db->db_buf = buf; 402 if (buf != NULL) { 403 ASSERT(buf->b_data != NULL); 404 db->db.db_data = buf->b_data; 405 if (!arc_released(buf)) 406 arc_set_callback(buf, dbuf_do_evict, db); 407 dbuf_update_data(db); 408 } else { 409 dbuf_evict_user(db); 410 db->db.db_data = NULL; 411 if (db->db_state != DB_NOFILL) 412 db->db_state = DB_UNCACHED; 413 } 414 } 415 416 /* 417 * Loan out an arc_buf for read. Return the loaned arc_buf. 418 */ 419 arc_buf_t * 420 dbuf_loan_arcbuf(dmu_buf_impl_t *db) 421 { 422 arc_buf_t *abuf; 423 424 mutex_enter(&db->db_mtx); 425 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 426 int blksz = db->db.db_size; 427 mutex_exit(&db->db_mtx); 428 abuf = arc_loan_buf(db->db_dnode->dn_objset->os_spa, blksz); 429 bcopy(db->db.db_data, abuf->b_data, blksz); 430 } else { 431 abuf = db->db_buf; 432 arc_loan_inuse_buf(abuf, db); 433 dbuf_set_data(db, NULL); 434 mutex_exit(&db->db_mtx); 435 } 436 return (abuf); 437 } 438 439 uint64_t 440 dbuf_whichblock(dnode_t *dn, uint64_t offset) 441 { 442 if (dn->dn_datablkshift) { 443 return (offset >> dn->dn_datablkshift); 444 } else { 445 ASSERT3U(offset, <, dn->dn_datablksz); 446 return (0); 447 } 448 } 449 450 static void 451 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 452 { 453 dmu_buf_impl_t *db = vdb; 454 455 mutex_enter(&db->db_mtx); 456 ASSERT3U(db->db_state, ==, DB_READ); 457 /* 458 * All reads are synchronous, so we must have a hold on the dbuf 459 */ 460 ASSERT(refcount_count(&db->db_holds) > 0); 461 ASSERT(db->db_buf == NULL); 462 ASSERT(db->db.db_data == NULL); 463 if (db->db_level == 0 && db->db_freed_in_flight) { 464 /* we were freed in flight; disregard any error */ 465 arc_release(buf, db); 466 bzero(buf->b_data, db->db.db_size); 467 arc_buf_freeze(buf); 468 db->db_freed_in_flight = FALSE; 469 dbuf_set_data(db, buf); 470 db->db_state = DB_CACHED; 471 } else if (zio == NULL || zio->io_error == 0) { 472 dbuf_set_data(db, buf); 473 db->db_state = DB_CACHED; 474 } else { 475 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 476 ASSERT3P(db->db_buf, ==, NULL); 477 VERIFY(arc_buf_remove_ref(buf, db) == 1); 478 db->db_state = DB_UNCACHED; 479 } 480 cv_broadcast(&db->db_changed); 481 dbuf_rele_and_unlock(db, NULL); 482 } 483 484 static void 485 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 486 { 487 dnode_t *dn = db->db_dnode; 488 zbookmark_t zb; 489 uint32_t aflags = ARC_NOWAIT; 490 arc_buf_t *pbuf; 491 492 ASSERT(!refcount_is_zero(&db->db_holds)); 493 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 494 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 495 ASSERT(MUTEX_HELD(&db->db_mtx)); 496 ASSERT(db->db_state == DB_UNCACHED); 497 ASSERT(db->db_buf == NULL); 498 499 if (db->db_blkid == DMU_BONUS_BLKID) { 500 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 501 502 ASSERT3U(bonuslen, <=, db->db.db_size); 503 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 504 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 505 if (bonuslen < DN_MAX_BONUSLEN) 506 bzero(db->db.db_data, DN_MAX_BONUSLEN); 507 if (bonuslen) 508 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 509 dbuf_update_data(db); 510 db->db_state = DB_CACHED; 511 mutex_exit(&db->db_mtx); 512 return; 513 } 514 515 /* 516 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 517 * processes the delete record and clears the bp while we are waiting 518 * for the dn_mtx (resulting in a "no" from block_freed). 519 */ 520 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 521 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 522 BP_IS_HOLE(db->db_blkptr)))) { 523 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 524 525 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa, 526 db->db.db_size, db, type)); 527 bzero(db->db.db_data, db->db.db_size); 528 db->db_state = DB_CACHED; 529 *flags |= DB_RF_CACHED; 530 mutex_exit(&db->db_mtx); 531 return; 532 } 533 534 db->db_state = DB_READ; 535 mutex_exit(&db->db_mtx); 536 537 if (DBUF_IS_L2CACHEABLE(db)) 538 aflags |= ARC_L2CACHE; 539 540 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 541 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 542 db->db.db_object, db->db_level, db->db_blkid); 543 544 dbuf_add_ref(db, NULL); 545 /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ 546 547 if (db->db_parent) 548 pbuf = db->db_parent->db_buf; 549 else 550 pbuf = db->db_objset->os_phys_buf; 551 552 (void) dsl_read(zio, dn->dn_objset->os_spa, db->db_blkptr, pbuf, 553 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 554 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 555 &aflags, &zb); 556 if (aflags & ARC_CACHED) 557 *flags |= DB_RF_CACHED; 558 } 559 560 int 561 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 562 { 563 int err = 0; 564 int havepzio = (zio != NULL); 565 int prefetch; 566 567 /* 568 * We don't have to hold the mutex to check db_state because it 569 * can't be freed while we have a hold on the buffer. 570 */ 571 ASSERT(!refcount_is_zero(&db->db_holds)); 572 573 if (db->db_state == DB_NOFILL) 574 return (EIO); 575 576 if ((flags & DB_RF_HAVESTRUCT) == 0) 577 rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER); 578 579 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 580 (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL && 581 DBUF_IS_CACHEABLE(db); 582 583 mutex_enter(&db->db_mtx); 584 if (db->db_state == DB_CACHED) { 585 mutex_exit(&db->db_mtx); 586 if (prefetch) 587 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 588 db->db.db_size, TRUE); 589 if ((flags & DB_RF_HAVESTRUCT) == 0) 590 rw_exit(&db->db_dnode->dn_struct_rwlock); 591 } else if (db->db_state == DB_UNCACHED) { 592 if (zio == NULL) { 593 zio = zio_root(db->db_dnode->dn_objset->os_spa, 594 NULL, NULL, ZIO_FLAG_CANFAIL); 595 } 596 dbuf_read_impl(db, zio, &flags); 597 598 /* dbuf_read_impl has dropped db_mtx for us */ 599 600 if (prefetch) 601 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 602 db->db.db_size, flags & DB_RF_CACHED); 603 604 if ((flags & DB_RF_HAVESTRUCT) == 0) 605 rw_exit(&db->db_dnode->dn_struct_rwlock); 606 607 if (!havepzio) 608 err = zio_wait(zio); 609 } else { 610 mutex_exit(&db->db_mtx); 611 if (prefetch) 612 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 613 db->db.db_size, TRUE); 614 if ((flags & DB_RF_HAVESTRUCT) == 0) 615 rw_exit(&db->db_dnode->dn_struct_rwlock); 616 617 mutex_enter(&db->db_mtx); 618 if ((flags & DB_RF_NEVERWAIT) == 0) { 619 while (db->db_state == DB_READ || 620 db->db_state == DB_FILL) { 621 ASSERT(db->db_state == DB_READ || 622 (flags & DB_RF_HAVESTRUCT) == 0); 623 cv_wait(&db->db_changed, &db->db_mtx); 624 } 625 if (db->db_state == DB_UNCACHED) 626 err = EIO; 627 } 628 mutex_exit(&db->db_mtx); 629 } 630 631 ASSERT(err || havepzio || db->db_state == DB_CACHED); 632 return (err); 633 } 634 635 static void 636 dbuf_noread(dmu_buf_impl_t *db) 637 { 638 ASSERT(!refcount_is_zero(&db->db_holds)); 639 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 640 mutex_enter(&db->db_mtx); 641 while (db->db_state == DB_READ || db->db_state == DB_FILL) 642 cv_wait(&db->db_changed, &db->db_mtx); 643 if (db->db_state == DB_UNCACHED) { 644 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 645 646 ASSERT(db->db_buf == NULL); 647 ASSERT(db->db.db_data == NULL); 648 dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 649 db->db.db_size, db, type)); 650 db->db_state = DB_FILL; 651 } else if (db->db_state == DB_NOFILL) { 652 dbuf_set_data(db, NULL); 653 } else { 654 ASSERT3U(db->db_state, ==, DB_CACHED); 655 } 656 mutex_exit(&db->db_mtx); 657 } 658 659 /* 660 * This is our just-in-time copy function. It makes a copy of 661 * buffers, that have been modified in a previous transaction 662 * group, before we modify them in the current active group. 663 * 664 * This function is used in two places: when we are dirtying a 665 * buffer for the first time in a txg, and when we are freeing 666 * a range in a dnode that includes this buffer. 667 * 668 * Note that when we are called from dbuf_free_range() we do 669 * not put a hold on the buffer, we just traverse the active 670 * dbuf list for the dnode. 671 */ 672 static void 673 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 674 { 675 dbuf_dirty_record_t *dr = db->db_last_dirty; 676 677 ASSERT(MUTEX_HELD(&db->db_mtx)); 678 ASSERT(db->db.db_data != NULL); 679 ASSERT(db->db_level == 0); 680 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 681 682 if (dr == NULL || 683 (dr->dt.dl.dr_data != 684 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 685 return; 686 687 /* 688 * If the last dirty record for this dbuf has not yet synced 689 * and its referencing the dbuf data, either: 690 * reset the reference to point to a new copy, 691 * or (if there a no active holders) 692 * just null out the current db_data pointer. 693 */ 694 ASSERT(dr->dr_txg >= txg - 2); 695 if (db->db_blkid == DMU_BONUS_BLKID) { 696 /* Note that the data bufs here are zio_bufs */ 697 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 698 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 699 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 700 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 701 int size = db->db.db_size; 702 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 703 dr->dt.dl.dr_data = arc_buf_alloc( 704 db->db_dnode->dn_objset->os_spa, size, db, type); 705 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 706 } else { 707 dbuf_set_data(db, NULL); 708 } 709 } 710 711 void 712 dbuf_unoverride(dbuf_dirty_record_t *dr) 713 { 714 dmu_buf_impl_t *db = dr->dr_dbuf; 715 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 716 uint64_t txg = dr->dr_txg; 717 718 ASSERT(MUTEX_HELD(&db->db_mtx)); 719 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 720 ASSERT(db->db_level == 0); 721 722 if (db->db_blkid == DMU_BONUS_BLKID || 723 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 724 return; 725 726 ASSERT(db->db_data_pending != dr); 727 728 /* free this block */ 729 if (!BP_IS_HOLE(bp)) 730 zio_free(db->db_dnode->dn_objset->os_spa, txg, bp); 731 732 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 733 /* 734 * Release the already-written buffer, so we leave it in 735 * a consistent dirty state. Note that all callers are 736 * modifying the buffer, so they will immediately do 737 * another (redundant) arc_release(). Therefore, leave 738 * the buf thawed to save the effort of freezing & 739 * immediately re-thawing it. 740 */ 741 arc_release(dr->dt.dl.dr_data, db); 742 } 743 744 /* 745 * Evict (if its unreferenced) or clear (if its referenced) any level-0 746 * data blocks in the free range, so that any future readers will find 747 * empty blocks. Also, if we happen accross any level-1 dbufs in the 748 * range that have not already been marked dirty, mark them dirty so 749 * they stay in memory. 750 */ 751 void 752 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) 753 { 754 dmu_buf_impl_t *db, *db_next; 755 uint64_t txg = tx->tx_txg; 756 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 757 uint64_t first_l1 = start >> epbs; 758 uint64_t last_l1 = end >> epbs; 759 760 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) { 761 end = dn->dn_maxblkid; 762 last_l1 = end >> epbs; 763 } 764 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); 765 mutex_enter(&dn->dn_dbufs_mtx); 766 for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 767 db_next = list_next(&dn->dn_dbufs, db); 768 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 769 770 if (db->db_level == 1 && 771 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) { 772 mutex_enter(&db->db_mtx); 773 if (db->db_last_dirty && 774 db->db_last_dirty->dr_txg < txg) { 775 dbuf_add_ref(db, FTAG); 776 mutex_exit(&db->db_mtx); 777 dbuf_will_dirty(db, tx); 778 dbuf_rele(db, FTAG); 779 } else { 780 mutex_exit(&db->db_mtx); 781 } 782 } 783 784 if (db->db_level != 0) 785 continue; 786 dprintf_dbuf(db, "found buf %s\n", ""); 787 if (db->db_blkid < start || db->db_blkid > end) 788 continue; 789 790 /* found a level 0 buffer in the range */ 791 if (dbuf_undirty(db, tx)) 792 continue; 793 794 mutex_enter(&db->db_mtx); 795 if (db->db_state == DB_UNCACHED || 796 db->db_state == DB_NOFILL || 797 db->db_state == DB_EVICTING) { 798 ASSERT(db->db.db_data == NULL); 799 mutex_exit(&db->db_mtx); 800 continue; 801 } 802 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 803 /* will be handled in dbuf_read_done or dbuf_rele */ 804 db->db_freed_in_flight = TRUE; 805 mutex_exit(&db->db_mtx); 806 continue; 807 } 808 if (refcount_count(&db->db_holds) == 0) { 809 ASSERT(db->db_buf); 810 dbuf_clear(db); 811 continue; 812 } 813 /* The dbuf is referenced */ 814 815 if (db->db_last_dirty != NULL) { 816 dbuf_dirty_record_t *dr = db->db_last_dirty; 817 818 if (dr->dr_txg == txg) { 819 /* 820 * This buffer is "in-use", re-adjust the file 821 * size to reflect that this buffer may 822 * contain new data when we sync. 823 */ 824 if (db->db_blkid != DMU_SPILL_BLKID && 825 db->db_blkid > dn->dn_maxblkid) 826 dn->dn_maxblkid = db->db_blkid; 827 dbuf_unoverride(dr); 828 } else { 829 /* 830 * This dbuf is not dirty in the open context. 831 * Either uncache it (if its not referenced in 832 * the open context) or reset its contents to 833 * empty. 834 */ 835 dbuf_fix_old_data(db, txg); 836 } 837 } 838 /* clear the contents if its cached */ 839 if (db->db_state == DB_CACHED) { 840 ASSERT(db->db.db_data != NULL); 841 arc_release(db->db_buf, db); 842 bzero(db->db.db_data, db->db.db_size); 843 arc_buf_freeze(db->db_buf); 844 } 845 846 mutex_exit(&db->db_mtx); 847 } 848 mutex_exit(&dn->dn_dbufs_mtx); 849 } 850 851 static int 852 dbuf_block_freeable(dmu_buf_impl_t *db) 853 { 854 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 855 uint64_t birth_txg = 0; 856 857 /* 858 * We don't need any locking to protect db_blkptr: 859 * If it's syncing, then db_last_dirty will be set 860 * so we'll ignore db_blkptr. 861 */ 862 ASSERT(MUTEX_HELD(&db->db_mtx)); 863 if (db->db_last_dirty) 864 birth_txg = db->db_last_dirty->dr_txg; 865 else if (db->db_blkptr) 866 birth_txg = db->db_blkptr->blk_birth; 867 868 /* 869 * If we don't exist or are in a snapshot, we can't be freed. 870 * Don't pass the bp to dsl_dataset_block_freeable() since we 871 * are holding the db_mtx lock and might deadlock if we are 872 * prefetching a dedup-ed block. 873 */ 874 if (birth_txg) 875 return (ds == NULL || 876 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 877 else 878 return (FALSE); 879 } 880 881 void 882 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 883 { 884 arc_buf_t *buf, *obuf; 885 int osize = db->db.db_size; 886 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 887 888 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 889 890 /* XXX does *this* func really need the lock? */ 891 ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)); 892 893 /* 894 * This call to dbuf_will_dirty() with the dn_struct_rwlock held 895 * is OK, because there can be no other references to the db 896 * when we are changing its size, so no concurrent DB_FILL can 897 * be happening. 898 */ 899 /* 900 * XXX we should be doing a dbuf_read, checking the return 901 * value and returning that up to our callers 902 */ 903 dbuf_will_dirty(db, tx); 904 905 /* create the data buffer for the new block */ 906 buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type); 907 908 /* copy old block data to the new block */ 909 obuf = db->db_buf; 910 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 911 /* zero the remainder */ 912 if (size > osize) 913 bzero((uint8_t *)buf->b_data + osize, size - osize); 914 915 mutex_enter(&db->db_mtx); 916 dbuf_set_data(db, buf); 917 VERIFY(arc_buf_remove_ref(obuf, db) == 1); 918 db->db.db_size = size; 919 920 if (db->db_level == 0) { 921 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 922 db->db_last_dirty->dt.dl.dr_data = buf; 923 } 924 mutex_exit(&db->db_mtx); 925 926 dnode_willuse_space(db->db_dnode, size-osize, tx); 927 } 928 929 void 930 dbuf_release_bp(dmu_buf_impl_t *db) 931 { 932 objset_t *os = db->db_dnode->dn_objset; 933 zbookmark_t zb; 934 935 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 936 ASSERT(arc_released(os->os_phys_buf) || 937 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 938 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 939 940 zb.zb_objset = os->os_dsl_dataset ? 941 os->os_dsl_dataset->ds_object : 0; 942 zb.zb_object = db->db.db_object; 943 zb.zb_level = db->db_level; 944 zb.zb_blkid = db->db_blkid; 945 (void) arc_release_bp(db->db_buf, db, 946 db->db_blkptr, os->os_spa, &zb); 947 } 948 949 dbuf_dirty_record_t * 950 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 951 { 952 dnode_t *dn = db->db_dnode; 953 objset_t *os = dn->dn_objset; 954 dbuf_dirty_record_t **drp, *dr; 955 int drop_struct_lock = FALSE; 956 boolean_t do_free_accounting = B_FALSE; 957 int txgoff = tx->tx_txg & TXG_MASK; 958 959 ASSERT(tx->tx_txg != 0); 960 ASSERT(!refcount_is_zero(&db->db_holds)); 961 DMU_TX_DIRTY_BUF(tx, db); 962 963 /* 964 * Shouldn't dirty a regular buffer in syncing context. Private 965 * objects may be dirtied in syncing context, but only if they 966 * were already pre-dirtied in open context. 967 */ 968 ASSERT(!dmu_tx_is_syncing(tx) || 969 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 970 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 971 dn->dn_objset->os_dsl_dataset == NULL); 972 /* 973 * We make this assert for private objects as well, but after we 974 * check if we're already dirty. They are allowed to re-dirty 975 * in syncing context. 976 */ 977 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 978 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 979 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 980 981 mutex_enter(&db->db_mtx); 982 /* 983 * XXX make this true for indirects too? The problem is that 984 * transactions created with dmu_tx_create_assigned() from 985 * syncing context don't bother holding ahead. 986 */ 987 ASSERT(db->db_level != 0 || 988 db->db_state == DB_CACHED || db->db_state == DB_FILL || 989 db->db_state == DB_NOFILL); 990 991 mutex_enter(&dn->dn_mtx); 992 /* 993 * Don't set dirtyctx to SYNC if we're just modifying this as we 994 * initialize the objset. 995 */ 996 if (dn->dn_dirtyctx == DN_UNDIRTIED && 997 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 998 dn->dn_dirtyctx = 999 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1000 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1001 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1002 } 1003 mutex_exit(&dn->dn_mtx); 1004 1005 if (db->db_blkid == DMU_SPILL_BLKID) 1006 dn->dn_have_spill = B_TRUE; 1007 1008 /* 1009 * If this buffer is already dirty, we're done. 1010 */ 1011 drp = &db->db_last_dirty; 1012 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1013 db->db.db_object == DMU_META_DNODE_OBJECT); 1014 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1015 drp = &dr->dr_next; 1016 if (dr && dr->dr_txg == tx->tx_txg) { 1017 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1018 /* 1019 * If this buffer has already been written out, 1020 * we now need to reset its state. 1021 */ 1022 dbuf_unoverride(dr); 1023 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1024 db->db_state != DB_NOFILL) 1025 arc_buf_thaw(db->db_buf); 1026 } 1027 mutex_exit(&db->db_mtx); 1028 return (dr); 1029 } 1030 1031 /* 1032 * Only valid if not already dirty. 1033 */ 1034 ASSERT(dn->dn_object == 0 || 1035 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1036 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1037 1038 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1039 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1040 dn->dn_phys->dn_nlevels > db->db_level || 1041 dn->dn_next_nlevels[txgoff] > db->db_level || 1042 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1043 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1044 1045 /* 1046 * We should only be dirtying in syncing context if it's the 1047 * mos or we're initializing the os or it's a special object. 1048 * However, we are allowed to dirty in syncing context provided 1049 * we already dirtied it in open context. Hence we must make 1050 * this assertion only if we're not already dirty. 1051 */ 1052 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1053 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1054 ASSERT(db->db.db_size != 0); 1055 1056 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1057 1058 if (db->db_blkid != DMU_BONUS_BLKID) { 1059 /* 1060 * Update the accounting. 1061 * Note: we delay "free accounting" until after we drop 1062 * the db_mtx. This keeps us from grabbing other locks 1063 * (and possibly deadlocking) in bp_get_dsize() while 1064 * also holding the db_mtx. 1065 */ 1066 dnode_willuse_space(dn, db->db.db_size, tx); 1067 do_free_accounting = dbuf_block_freeable(db); 1068 } 1069 1070 /* 1071 * If this buffer is dirty in an old transaction group we need 1072 * to make a copy of it so that the changes we make in this 1073 * transaction group won't leak out when we sync the older txg. 1074 */ 1075 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1076 if (db->db_level == 0) { 1077 void *data_old = db->db_buf; 1078 1079 if (db->db_state != DB_NOFILL) { 1080 if (db->db_blkid == DMU_BONUS_BLKID) { 1081 dbuf_fix_old_data(db, tx->tx_txg); 1082 data_old = db->db.db_data; 1083 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1084 /* 1085 * Release the data buffer from the cache so 1086 * that we can modify it without impacting 1087 * possible other users of this cached data 1088 * block. Note that indirect blocks and 1089 * private objects are not released until the 1090 * syncing state (since they are only modified 1091 * then). 1092 */ 1093 arc_release(db->db_buf, db); 1094 dbuf_fix_old_data(db, tx->tx_txg); 1095 data_old = db->db_buf; 1096 } 1097 ASSERT(data_old != NULL); 1098 } 1099 dr->dt.dl.dr_data = data_old; 1100 } else { 1101 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1102 list_create(&dr->dt.di.dr_children, 1103 sizeof (dbuf_dirty_record_t), 1104 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1105 } 1106 dr->dr_dbuf = db; 1107 dr->dr_txg = tx->tx_txg; 1108 dr->dr_next = *drp; 1109 *drp = dr; 1110 1111 /* 1112 * We could have been freed_in_flight between the dbuf_noread 1113 * and dbuf_dirty. We win, as though the dbuf_noread() had 1114 * happened after the free. 1115 */ 1116 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1117 db->db_blkid != DMU_SPILL_BLKID) { 1118 mutex_enter(&dn->dn_mtx); 1119 dnode_clear_range(dn, db->db_blkid, 1, tx); 1120 mutex_exit(&dn->dn_mtx); 1121 db->db_freed_in_flight = FALSE; 1122 } 1123 1124 /* 1125 * This buffer is now part of this txg 1126 */ 1127 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1128 db->db_dirtycnt += 1; 1129 ASSERT3U(db->db_dirtycnt, <=, 3); 1130 1131 mutex_exit(&db->db_mtx); 1132 1133 if (db->db_blkid == DMU_BONUS_BLKID || 1134 db->db_blkid == DMU_SPILL_BLKID) { 1135 mutex_enter(&dn->dn_mtx); 1136 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1137 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1138 mutex_exit(&dn->dn_mtx); 1139 dnode_setdirty(dn, tx); 1140 return (dr); 1141 } else if (do_free_accounting) { 1142 blkptr_t *bp = db->db_blkptr; 1143 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1144 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1145 /* 1146 * This is only a guess -- if the dbuf is dirty 1147 * in a previous txg, we don't know how much 1148 * space it will use on disk yet. We should 1149 * really have the struct_rwlock to access 1150 * db_blkptr, but since this is just a guess, 1151 * it's OK if we get an odd answer. 1152 */ 1153 ddt_prefetch(os->os_spa, bp); 1154 dnode_willuse_space(dn, -willfree, tx); 1155 } 1156 1157 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1158 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1159 drop_struct_lock = TRUE; 1160 } 1161 1162 if (db->db_level == 0) { 1163 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1164 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1165 } 1166 1167 if (db->db_level+1 < dn->dn_nlevels) { 1168 dmu_buf_impl_t *parent = db->db_parent; 1169 dbuf_dirty_record_t *di; 1170 int parent_held = FALSE; 1171 1172 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1173 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1174 1175 parent = dbuf_hold_level(dn, db->db_level+1, 1176 db->db_blkid >> epbs, FTAG); 1177 ASSERT(parent != NULL); 1178 parent_held = TRUE; 1179 } 1180 if (drop_struct_lock) 1181 rw_exit(&dn->dn_struct_rwlock); 1182 ASSERT3U(db->db_level+1, ==, parent->db_level); 1183 di = dbuf_dirty(parent, tx); 1184 if (parent_held) 1185 dbuf_rele(parent, FTAG); 1186 1187 mutex_enter(&db->db_mtx); 1188 /* possible race with dbuf_undirty() */ 1189 if (db->db_last_dirty == dr || 1190 dn->dn_object == DMU_META_DNODE_OBJECT) { 1191 mutex_enter(&di->dt.di.dr_mtx); 1192 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1193 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1194 list_insert_tail(&di->dt.di.dr_children, dr); 1195 mutex_exit(&di->dt.di.dr_mtx); 1196 dr->dr_parent = di; 1197 } 1198 mutex_exit(&db->db_mtx); 1199 } else { 1200 ASSERT(db->db_level+1 == dn->dn_nlevels); 1201 ASSERT(db->db_blkid < dn->dn_nblkptr); 1202 ASSERT(db->db_parent == NULL || 1203 db->db_parent == db->db_dnode->dn_dbuf); 1204 mutex_enter(&dn->dn_mtx); 1205 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1206 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1207 mutex_exit(&dn->dn_mtx); 1208 if (drop_struct_lock) 1209 rw_exit(&dn->dn_struct_rwlock); 1210 } 1211 1212 dnode_setdirty(dn, tx); 1213 return (dr); 1214 } 1215 1216 static int 1217 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1218 { 1219 dnode_t *dn = db->db_dnode; 1220 uint64_t txg = tx->tx_txg; 1221 dbuf_dirty_record_t *dr, **drp; 1222 1223 ASSERT(txg != 0); 1224 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1225 1226 mutex_enter(&db->db_mtx); 1227 /* 1228 * If this buffer is not dirty, we're done. 1229 */ 1230 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1231 if (dr->dr_txg <= txg) 1232 break; 1233 if (dr == NULL || dr->dr_txg < txg) { 1234 mutex_exit(&db->db_mtx); 1235 return (0); 1236 } 1237 ASSERT(dr->dr_txg == txg); 1238 ASSERT(dr->dr_dbuf == db); 1239 1240 /* 1241 * If this buffer is currently held, we cannot undirty 1242 * it, since one of the current holders may be in the 1243 * middle of an update. Note that users of dbuf_undirty() 1244 * should not place a hold on the dbuf before the call. 1245 */ 1246 if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1247 mutex_exit(&db->db_mtx); 1248 /* Make sure we don't toss this buffer at sync phase */ 1249 mutex_enter(&dn->dn_mtx); 1250 dnode_clear_range(dn, db->db_blkid, 1, tx); 1251 mutex_exit(&dn->dn_mtx); 1252 return (0); 1253 } 1254 1255 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1256 1257 ASSERT(db->db.db_size != 0); 1258 1259 /* XXX would be nice to fix up dn_towrite_space[] */ 1260 1261 *drp = dr->dr_next; 1262 1263 if (dr->dr_parent) { 1264 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1265 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1266 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1267 } else if (db->db_level+1 == dn->dn_nlevels) { 1268 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1269 mutex_enter(&dn->dn_mtx); 1270 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1271 mutex_exit(&dn->dn_mtx); 1272 } 1273 1274 if (db->db_level == 0) { 1275 if (db->db_state != DB_NOFILL) { 1276 dbuf_unoverride(dr); 1277 1278 ASSERT(db->db_buf != NULL); 1279 ASSERT(dr->dt.dl.dr_data != NULL); 1280 if (dr->dt.dl.dr_data != db->db_buf) 1281 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 1282 db) == 1); 1283 } 1284 } else { 1285 ASSERT(db->db_buf != NULL); 1286 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1287 mutex_destroy(&dr->dt.di.dr_mtx); 1288 list_destroy(&dr->dt.di.dr_children); 1289 } 1290 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1291 1292 ASSERT(db->db_dirtycnt > 0); 1293 db->db_dirtycnt -= 1; 1294 1295 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1296 arc_buf_t *buf = db->db_buf; 1297 1298 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1299 dbuf_set_data(db, NULL); 1300 VERIFY(arc_buf_remove_ref(buf, db) == 1); 1301 dbuf_evict(db); 1302 return (1); 1303 } 1304 1305 mutex_exit(&db->db_mtx); 1306 return (0); 1307 } 1308 1309 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty 1310 void 1311 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1312 { 1313 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1314 1315 ASSERT(tx->tx_txg != 0); 1316 ASSERT(!refcount_is_zero(&db->db_holds)); 1317 1318 if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) 1319 rf |= DB_RF_HAVESTRUCT; 1320 (void) dbuf_read(db, NULL, rf); 1321 (void) dbuf_dirty(db, tx); 1322 } 1323 1324 void 1325 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1326 { 1327 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1328 1329 db->db_state = DB_NOFILL; 1330 1331 dmu_buf_will_fill(db_fake, tx); 1332 } 1333 1334 void 1335 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1336 { 1337 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1338 1339 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1340 ASSERT(tx->tx_txg != 0); 1341 ASSERT(db->db_level == 0); 1342 ASSERT(!refcount_is_zero(&db->db_holds)); 1343 1344 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1345 dmu_tx_private_ok(tx)); 1346 1347 dbuf_noread(db); 1348 (void) dbuf_dirty(db, tx); 1349 } 1350 1351 #pragma weak dmu_buf_fill_done = dbuf_fill_done 1352 /* ARGSUSED */ 1353 void 1354 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1355 { 1356 mutex_enter(&db->db_mtx); 1357 DBUF_VERIFY(db); 1358 1359 if (db->db_state == DB_FILL) { 1360 if (db->db_level == 0 && db->db_freed_in_flight) { 1361 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1362 /* we were freed while filling */ 1363 /* XXX dbuf_undirty? */ 1364 bzero(db->db.db_data, db->db.db_size); 1365 db->db_freed_in_flight = FALSE; 1366 } 1367 db->db_state = DB_CACHED; 1368 cv_broadcast(&db->db_changed); 1369 } 1370 mutex_exit(&db->db_mtx); 1371 } 1372 1373 /* 1374 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1375 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1376 */ 1377 void 1378 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1379 { 1380 ASSERT(!refcount_is_zero(&db->db_holds)); 1381 ASSERT(db->db_dnode->dn_object != DMU_META_DNODE_OBJECT); 1382 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1383 ASSERT(db->db_level == 0); 1384 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1385 ASSERT(buf != NULL); 1386 ASSERT(arc_buf_size(buf) == db->db.db_size); 1387 ASSERT(tx->tx_txg != 0); 1388 1389 arc_return_buf(buf, db); 1390 ASSERT(arc_released(buf)); 1391 1392 mutex_enter(&db->db_mtx); 1393 1394 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1395 cv_wait(&db->db_changed, &db->db_mtx); 1396 1397 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1398 1399 if (db->db_state == DB_CACHED && 1400 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1401 mutex_exit(&db->db_mtx); 1402 (void) dbuf_dirty(db, tx); 1403 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1404 VERIFY(arc_buf_remove_ref(buf, db) == 1); 1405 xuio_stat_wbuf_copied(); 1406 return; 1407 } 1408 1409 xuio_stat_wbuf_nocopy(); 1410 if (db->db_state == DB_CACHED) { 1411 dbuf_dirty_record_t *dr = db->db_last_dirty; 1412 1413 ASSERT(db->db_buf != NULL); 1414 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1415 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1416 if (!arc_released(db->db_buf)) { 1417 ASSERT(dr->dt.dl.dr_override_state == 1418 DR_OVERRIDDEN); 1419 arc_release(db->db_buf, db); 1420 } 1421 dr->dt.dl.dr_data = buf; 1422 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); 1423 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1424 arc_release(db->db_buf, db); 1425 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); 1426 } 1427 db->db_buf = NULL; 1428 } 1429 ASSERT(db->db_buf == NULL); 1430 dbuf_set_data(db, buf); 1431 db->db_state = DB_FILL; 1432 mutex_exit(&db->db_mtx); 1433 (void) dbuf_dirty(db, tx); 1434 dbuf_fill_done(db, tx); 1435 } 1436 1437 /* 1438 * "Clear" the contents of this dbuf. This will mark the dbuf 1439 * EVICTING and clear *most* of its references. Unfortunetely, 1440 * when we are not holding the dn_dbufs_mtx, we can't clear the 1441 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1442 * in this case. For callers from the DMU we will usually see: 1443 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1444 * For the arc callback, we will usually see: 1445 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1446 * Sometimes, though, we will get a mix of these two: 1447 * DMU: dbuf_clear()->arc_buf_evict() 1448 * ARC: dbuf_do_evict()->dbuf_destroy() 1449 */ 1450 void 1451 dbuf_clear(dmu_buf_impl_t *db) 1452 { 1453 dnode_t *dn = db->db_dnode; 1454 dmu_buf_impl_t *parent = db->db_parent; 1455 dmu_buf_impl_t *dndb = dn->dn_dbuf; 1456 int dbuf_gone = FALSE; 1457 1458 ASSERT(MUTEX_HELD(&db->db_mtx)); 1459 ASSERT(refcount_is_zero(&db->db_holds)); 1460 1461 dbuf_evict_user(db); 1462 1463 if (db->db_state == DB_CACHED) { 1464 ASSERT(db->db.db_data != NULL); 1465 if (db->db_blkid == DMU_BONUS_BLKID) { 1466 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1467 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1468 } 1469 db->db.db_data = NULL; 1470 db->db_state = DB_UNCACHED; 1471 } 1472 1473 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1474 ASSERT(db->db_data_pending == NULL); 1475 1476 db->db_state = DB_EVICTING; 1477 db->db_blkptr = NULL; 1478 1479 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1480 list_remove(&dn->dn_dbufs, db); 1481 dnode_rele(dn, db); 1482 db->db_dnode = NULL; 1483 } 1484 1485 if (db->db_buf) 1486 dbuf_gone = arc_buf_evict(db->db_buf); 1487 1488 if (!dbuf_gone) 1489 mutex_exit(&db->db_mtx); 1490 1491 /* 1492 * If this dbuf is referened from an indirect dbuf, 1493 * decrement the ref count on the indirect dbuf. 1494 */ 1495 if (parent && parent != dndb) 1496 dbuf_rele(parent, db); 1497 } 1498 1499 static int 1500 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1501 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1502 { 1503 int nlevels, epbs; 1504 1505 *parentp = NULL; 1506 *bpp = NULL; 1507 1508 ASSERT(blkid != DMU_BONUS_BLKID); 1509 1510 if (blkid == DMU_SPILL_BLKID) { 1511 mutex_enter(&dn->dn_mtx); 1512 if (dn->dn_have_spill && 1513 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1514 *bpp = &dn->dn_phys->dn_spill; 1515 else 1516 *bpp = NULL; 1517 dbuf_add_ref(dn->dn_dbuf, NULL); 1518 *parentp = dn->dn_dbuf; 1519 mutex_exit(&dn->dn_mtx); 1520 return (0); 1521 } 1522 1523 if (dn->dn_phys->dn_nlevels == 0) 1524 nlevels = 1; 1525 else 1526 nlevels = dn->dn_phys->dn_nlevels; 1527 1528 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1529 1530 ASSERT3U(level * epbs, <, 64); 1531 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1532 if (level >= nlevels || 1533 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1534 /* the buffer has no parent yet */ 1535 return (ENOENT); 1536 } else if (level < nlevels-1) { 1537 /* this block is referenced from an indirect block */ 1538 int err = dbuf_hold_impl(dn, level+1, 1539 blkid >> epbs, fail_sparse, NULL, parentp); 1540 if (err) 1541 return (err); 1542 err = dbuf_read(*parentp, NULL, 1543 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1544 if (err) { 1545 dbuf_rele(*parentp, NULL); 1546 *parentp = NULL; 1547 return (err); 1548 } 1549 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1550 (blkid & ((1ULL << epbs) - 1)); 1551 return (0); 1552 } else { 1553 /* the block is referenced from the dnode */ 1554 ASSERT3U(level, ==, nlevels-1); 1555 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1556 blkid < dn->dn_phys->dn_nblkptr); 1557 if (dn->dn_dbuf) { 1558 dbuf_add_ref(dn->dn_dbuf, NULL); 1559 *parentp = dn->dn_dbuf; 1560 } 1561 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1562 return (0); 1563 } 1564 } 1565 1566 static dmu_buf_impl_t * 1567 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1568 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1569 { 1570 objset_t *os = dn->dn_objset; 1571 dmu_buf_impl_t *db, *odb; 1572 1573 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1574 ASSERT(dn->dn_type != DMU_OT_NONE); 1575 1576 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1577 1578 db->db_objset = os; 1579 db->db.db_object = dn->dn_object; 1580 db->db_level = level; 1581 db->db_blkid = blkid; 1582 db->db_last_dirty = NULL; 1583 db->db_dirtycnt = 0; 1584 db->db_dnode = dn; 1585 db->db_parent = parent; 1586 db->db_blkptr = blkptr; 1587 1588 db->db_user_ptr = NULL; 1589 db->db_user_data_ptr_ptr = NULL; 1590 db->db_evict_func = NULL; 1591 db->db_immediate_evict = 0; 1592 db->db_freed_in_flight = 0; 1593 1594 if (blkid == DMU_BONUS_BLKID) { 1595 ASSERT3P(parent, ==, dn->dn_dbuf); 1596 db->db.db_size = DN_MAX_BONUSLEN - 1597 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1598 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1599 db->db.db_offset = DMU_BONUS_BLKID; 1600 db->db_state = DB_UNCACHED; 1601 /* the bonus dbuf is not placed in the hash table */ 1602 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1603 return (db); 1604 } else if (blkid == DMU_SPILL_BLKID) { 1605 db->db.db_size = (blkptr != NULL) ? 1606 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1607 db->db.db_offset = 0; 1608 } else { 1609 int blocksize = 1610 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; 1611 db->db.db_size = blocksize; 1612 db->db.db_offset = db->db_blkid * blocksize; 1613 } 1614 1615 /* 1616 * Hold the dn_dbufs_mtx while we get the new dbuf 1617 * in the hash table *and* added to the dbufs list. 1618 * This prevents a possible deadlock with someone 1619 * trying to look up this dbuf before its added to the 1620 * dn_dbufs list. 1621 */ 1622 mutex_enter(&dn->dn_dbufs_mtx); 1623 db->db_state = DB_EVICTING; 1624 if ((odb = dbuf_hash_insert(db)) != NULL) { 1625 /* someone else inserted it first */ 1626 kmem_cache_free(dbuf_cache, db); 1627 mutex_exit(&dn->dn_dbufs_mtx); 1628 return (odb); 1629 } 1630 list_insert_head(&dn->dn_dbufs, db); 1631 db->db_state = DB_UNCACHED; 1632 mutex_exit(&dn->dn_dbufs_mtx); 1633 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1634 1635 if (parent && parent != dn->dn_dbuf) 1636 dbuf_add_ref(parent, db); 1637 1638 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1639 refcount_count(&dn->dn_holds) > 0); 1640 (void) refcount_add(&dn->dn_holds, db); 1641 1642 dprintf_dbuf(db, "db=%p\n", db); 1643 1644 return (db); 1645 } 1646 1647 static int 1648 dbuf_do_evict(void *private) 1649 { 1650 arc_buf_t *buf = private; 1651 dmu_buf_impl_t *db = buf->b_private; 1652 1653 if (!MUTEX_HELD(&db->db_mtx)) 1654 mutex_enter(&db->db_mtx); 1655 1656 ASSERT(refcount_is_zero(&db->db_holds)); 1657 1658 if (db->db_state != DB_EVICTING) { 1659 ASSERT(db->db_state == DB_CACHED); 1660 DBUF_VERIFY(db); 1661 db->db_buf = NULL; 1662 dbuf_evict(db); 1663 } else { 1664 mutex_exit(&db->db_mtx); 1665 dbuf_destroy(db); 1666 } 1667 return (0); 1668 } 1669 1670 static void 1671 dbuf_destroy(dmu_buf_impl_t *db) 1672 { 1673 ASSERT(refcount_is_zero(&db->db_holds)); 1674 1675 if (db->db_blkid != DMU_BONUS_BLKID) { 1676 /* 1677 * If this dbuf is still on the dn_dbufs list, 1678 * remove it from that list. 1679 */ 1680 if (db->db_dnode) { 1681 dnode_t *dn = db->db_dnode; 1682 1683 mutex_enter(&dn->dn_dbufs_mtx); 1684 list_remove(&dn->dn_dbufs, db); 1685 mutex_exit(&dn->dn_dbufs_mtx); 1686 1687 dnode_rele(dn, db); 1688 db->db_dnode = NULL; 1689 } 1690 dbuf_hash_remove(db); 1691 } 1692 db->db_parent = NULL; 1693 db->db_buf = NULL; 1694 1695 ASSERT(!list_link_active(&db->db_link)); 1696 ASSERT(db->db.db_data == NULL); 1697 ASSERT(db->db_hash_next == NULL); 1698 ASSERT(db->db_blkptr == NULL); 1699 ASSERT(db->db_data_pending == NULL); 1700 1701 kmem_cache_free(dbuf_cache, db); 1702 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1703 } 1704 1705 void 1706 dbuf_prefetch(dnode_t *dn, uint64_t blkid) 1707 { 1708 dmu_buf_impl_t *db = NULL; 1709 blkptr_t *bp = NULL; 1710 1711 ASSERT(blkid != DMU_BONUS_BLKID); 1712 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1713 1714 if (dnode_block_freed(dn, blkid)) 1715 return; 1716 1717 /* dbuf_find() returns with db_mtx held */ 1718 if (db = dbuf_find(dn, 0, blkid)) { 1719 if (refcount_count(&db->db_holds) > 0) { 1720 /* 1721 * This dbuf is active. We assume that it is 1722 * already CACHED, or else about to be either 1723 * read or filled. 1724 */ 1725 mutex_exit(&db->db_mtx); 1726 return; 1727 } 1728 mutex_exit(&db->db_mtx); 1729 db = NULL; 1730 } 1731 1732 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1733 if (bp && !BP_IS_HOLE(bp)) { 1734 int priority = dn->dn_type == DMU_OT_DDT_ZAP ? 1735 ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ; 1736 arc_buf_t *pbuf; 1737 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 1738 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; 1739 zbookmark_t zb; 1740 1741 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 1742 dn->dn_object, 0, blkid); 1743 1744 if (db) 1745 pbuf = db->db_buf; 1746 else 1747 pbuf = dn->dn_objset->os_phys_buf; 1748 1749 (void) dsl_read(NULL, dn->dn_objset->os_spa, 1750 bp, pbuf, NULL, NULL, priority, 1751 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1752 &aflags, &zb); 1753 } 1754 if (db) 1755 dbuf_rele(db, NULL); 1756 } 1757 } 1758 1759 /* 1760 * Returns with db_holds incremented, and db_mtx not held. 1761 * Note: dn_struct_rwlock must be held. 1762 */ 1763 int 1764 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1765 void *tag, dmu_buf_impl_t **dbp) 1766 { 1767 dmu_buf_impl_t *db, *parent = NULL; 1768 1769 ASSERT(blkid != DMU_BONUS_BLKID); 1770 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1771 ASSERT3U(dn->dn_nlevels, >, level); 1772 1773 *dbp = NULL; 1774 top: 1775 /* dbuf_find() returns with db_mtx held */ 1776 db = dbuf_find(dn, level, blkid); 1777 1778 if (db == NULL) { 1779 blkptr_t *bp = NULL; 1780 int err; 1781 1782 ASSERT3P(parent, ==, NULL); 1783 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1784 if (fail_sparse) { 1785 if (err == 0 && bp && BP_IS_HOLE(bp)) 1786 err = ENOENT; 1787 if (err) { 1788 if (parent) 1789 dbuf_rele(parent, NULL); 1790 return (err); 1791 } 1792 } 1793 if (err && err != ENOENT) 1794 return (err); 1795 db = dbuf_create(dn, level, blkid, parent, bp); 1796 } 1797 1798 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1799 arc_buf_add_ref(db->db_buf, db); 1800 if (db->db_buf->b_data == NULL) { 1801 dbuf_clear(db); 1802 if (parent) { 1803 dbuf_rele(parent, NULL); 1804 parent = NULL; 1805 } 1806 goto top; 1807 } 1808 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1809 } 1810 1811 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1812 1813 /* 1814 * If this buffer is currently syncing out, and we are are 1815 * still referencing it from db_data, we need to make a copy 1816 * of it in case we decide we want to dirty it again in this txg. 1817 */ 1818 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1819 dn->dn_object != DMU_META_DNODE_OBJECT && 1820 db->db_state == DB_CACHED && db->db_data_pending) { 1821 dbuf_dirty_record_t *dr = db->db_data_pending; 1822 1823 if (dr->dt.dl.dr_data == db->db_buf) { 1824 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1825 1826 dbuf_set_data(db, 1827 arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 1828 db->db.db_size, db, type)); 1829 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1830 db->db.db_size); 1831 } 1832 } 1833 1834 (void) refcount_add(&db->db_holds, tag); 1835 dbuf_update_data(db); 1836 DBUF_VERIFY(db); 1837 mutex_exit(&db->db_mtx); 1838 1839 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1840 if (parent) 1841 dbuf_rele(parent, NULL); 1842 1843 ASSERT3P(db->db_dnode, ==, dn); 1844 ASSERT3U(db->db_blkid, ==, blkid); 1845 ASSERT3U(db->db_level, ==, level); 1846 *dbp = db; 1847 1848 return (0); 1849 } 1850 1851 dmu_buf_impl_t * 1852 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1853 { 1854 dmu_buf_impl_t *db; 1855 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1856 return (err ? NULL : db); 1857 } 1858 1859 dmu_buf_impl_t * 1860 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1861 { 1862 dmu_buf_impl_t *db; 1863 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1864 return (err ? NULL : db); 1865 } 1866 1867 void 1868 dbuf_create_bonus(dnode_t *dn) 1869 { 1870 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1871 1872 ASSERT(dn->dn_bonus == NULL); 1873 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 1874 } 1875 1876 int 1877 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 1878 { 1879 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1880 if (db->db_blkid != DMU_SPILL_BLKID) 1881 return (ENOTSUP); 1882 if (blksz == 0) 1883 blksz = SPA_MINBLOCKSIZE; 1884 if (blksz > SPA_MAXBLOCKSIZE) 1885 blksz = SPA_MAXBLOCKSIZE; 1886 else 1887 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 1888 1889 rw_enter(&db->db_dnode->dn_struct_rwlock, RW_WRITER); 1890 dbuf_new_size(db, blksz, tx); 1891 rw_exit(&db->db_dnode->dn_struct_rwlock); 1892 1893 return (0); 1894 } 1895 1896 void 1897 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 1898 { 1899 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 1900 } 1901 1902 #pragma weak dmu_buf_add_ref = dbuf_add_ref 1903 void 1904 dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 1905 { 1906 int64_t holds = refcount_add(&db->db_holds, tag); 1907 ASSERT(holds > 1); 1908 } 1909 1910 #pragma weak dmu_buf_rele = dbuf_rele 1911 void 1912 dbuf_rele(dmu_buf_impl_t *db, void *tag) 1913 { 1914 mutex_enter(&db->db_mtx); 1915 dbuf_rele_and_unlock(db, tag); 1916 } 1917 1918 /* 1919 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 1920 * db_dirtycnt and db_holds to be updated atomically. 1921 */ 1922 void 1923 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 1924 { 1925 int64_t holds; 1926 1927 ASSERT(MUTEX_HELD(&db->db_mtx)); 1928 DBUF_VERIFY(db); 1929 1930 holds = refcount_remove(&db->db_holds, tag); 1931 ASSERT(holds >= 0); 1932 1933 /* 1934 * We can't freeze indirects if there is a possibility that they 1935 * may be modified in the current syncing context. 1936 */ 1937 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 1938 arc_buf_freeze(db->db_buf); 1939 1940 if (holds == db->db_dirtycnt && 1941 db->db_level == 0 && db->db_immediate_evict) 1942 dbuf_evict_user(db); 1943 1944 if (holds == 0) { 1945 if (db->db_blkid == DMU_BONUS_BLKID) { 1946 mutex_exit(&db->db_mtx); 1947 dnode_rele(db->db_dnode, db); 1948 } else if (db->db_buf == NULL) { 1949 /* 1950 * This is a special case: we never associated this 1951 * dbuf with any data allocated from the ARC. 1952 */ 1953 ASSERT(db->db_state == DB_UNCACHED || 1954 db->db_state == DB_NOFILL); 1955 dbuf_evict(db); 1956 } else if (arc_released(db->db_buf)) { 1957 arc_buf_t *buf = db->db_buf; 1958 /* 1959 * This dbuf has anonymous data associated with it. 1960 */ 1961 dbuf_set_data(db, NULL); 1962 VERIFY(arc_buf_remove_ref(buf, db) == 1); 1963 dbuf_evict(db); 1964 } else { 1965 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); 1966 if (!DBUF_IS_CACHEABLE(db)) 1967 dbuf_clear(db); 1968 else 1969 mutex_exit(&db->db_mtx); 1970 } 1971 } else { 1972 mutex_exit(&db->db_mtx); 1973 } 1974 } 1975 1976 #pragma weak dmu_buf_refcount = dbuf_refcount 1977 uint64_t 1978 dbuf_refcount(dmu_buf_impl_t *db) 1979 { 1980 return (refcount_count(&db->db_holds)); 1981 } 1982 1983 void * 1984 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1985 dmu_buf_evict_func_t *evict_func) 1986 { 1987 return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1988 user_data_ptr_ptr, evict_func)); 1989 } 1990 1991 void * 1992 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1993 dmu_buf_evict_func_t *evict_func) 1994 { 1995 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1996 1997 db->db_immediate_evict = TRUE; 1998 return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1999 user_data_ptr_ptr, evict_func)); 2000 } 2001 2002 void * 2003 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 2004 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 2005 { 2006 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2007 ASSERT(db->db_level == 0); 2008 2009 ASSERT((user_ptr == NULL) == (evict_func == NULL)); 2010 2011 mutex_enter(&db->db_mtx); 2012 2013 if (db->db_user_ptr == old_user_ptr) { 2014 db->db_user_ptr = user_ptr; 2015 db->db_user_data_ptr_ptr = user_data_ptr_ptr; 2016 db->db_evict_func = evict_func; 2017 2018 dbuf_update_data(db); 2019 } else { 2020 old_user_ptr = db->db_user_ptr; 2021 } 2022 2023 mutex_exit(&db->db_mtx); 2024 return (old_user_ptr); 2025 } 2026 2027 void * 2028 dmu_buf_get_user(dmu_buf_t *db_fake) 2029 { 2030 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2031 ASSERT(!refcount_is_zero(&db->db_holds)); 2032 2033 return (db->db_user_ptr); 2034 } 2035 2036 boolean_t 2037 dmu_buf_freeable(dmu_buf_t *dbuf) 2038 { 2039 boolean_t res = B_FALSE; 2040 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2041 2042 if (db->db_blkptr) 2043 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2044 db->db_blkptr, db->db_blkptr->blk_birth); 2045 2046 return (res); 2047 } 2048 2049 static void 2050 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2051 { 2052 /* ASSERT(dmu_tx_is_syncing(tx) */ 2053 ASSERT(MUTEX_HELD(&db->db_mtx)); 2054 2055 if (db->db_blkptr != NULL) 2056 return; 2057 2058 if (db->db_blkid == DMU_SPILL_BLKID) { 2059 db->db_blkptr = &dn->dn_phys->dn_spill; 2060 BP_ZERO(db->db_blkptr); 2061 return; 2062 } 2063 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2064 /* 2065 * This buffer was allocated at a time when there was 2066 * no available blkptrs from the dnode, or it was 2067 * inappropriate to hook it in (i.e., nlevels mis-match). 2068 */ 2069 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2070 ASSERT(db->db_parent == NULL); 2071 db->db_parent = dn->dn_dbuf; 2072 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2073 DBUF_VERIFY(db); 2074 } else { 2075 dmu_buf_impl_t *parent = db->db_parent; 2076 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2077 2078 ASSERT(dn->dn_phys->dn_nlevels > 1); 2079 if (parent == NULL) { 2080 mutex_exit(&db->db_mtx); 2081 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2082 (void) dbuf_hold_impl(dn, db->db_level+1, 2083 db->db_blkid >> epbs, FALSE, db, &parent); 2084 rw_exit(&dn->dn_struct_rwlock); 2085 mutex_enter(&db->db_mtx); 2086 db->db_parent = parent; 2087 } 2088 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2089 (db->db_blkid & ((1ULL << epbs) - 1)); 2090 DBUF_VERIFY(db); 2091 } 2092 } 2093 2094 static void 2095 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2096 { 2097 dmu_buf_impl_t *db = dr->dr_dbuf; 2098 dnode_t *dn = db->db_dnode; 2099 zio_t *zio; 2100 2101 ASSERT(dmu_tx_is_syncing(tx)); 2102 2103 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2104 2105 mutex_enter(&db->db_mtx); 2106 2107 ASSERT(db->db_level > 0); 2108 DBUF_VERIFY(db); 2109 2110 if (db->db_buf == NULL) { 2111 mutex_exit(&db->db_mtx); 2112 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2113 mutex_enter(&db->db_mtx); 2114 } 2115 ASSERT3U(db->db_state, ==, DB_CACHED); 2116 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2117 ASSERT(db->db_buf != NULL); 2118 2119 dbuf_check_blkptr(dn, db); 2120 2121 db->db_data_pending = dr; 2122 2123 mutex_exit(&db->db_mtx); 2124 dbuf_write(dr, db->db_buf, tx); 2125 2126 zio = dr->dr_zio; 2127 mutex_enter(&dr->dt.di.dr_mtx); 2128 dbuf_sync_list(&dr->dt.di.dr_children, tx); 2129 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2130 mutex_exit(&dr->dt.di.dr_mtx); 2131 zio_nowait(zio); 2132 } 2133 2134 static void 2135 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2136 { 2137 arc_buf_t **datap = &dr->dt.dl.dr_data; 2138 dmu_buf_impl_t *db = dr->dr_dbuf; 2139 dnode_t *dn = db->db_dnode; 2140 objset_t *os = dn->dn_objset; 2141 uint64_t txg = tx->tx_txg; 2142 2143 ASSERT(dmu_tx_is_syncing(tx)); 2144 2145 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2146 2147 mutex_enter(&db->db_mtx); 2148 /* 2149 * To be synced, we must be dirtied. But we 2150 * might have been freed after the dirty. 2151 */ 2152 if (db->db_state == DB_UNCACHED) { 2153 /* This buffer has been freed since it was dirtied */ 2154 ASSERT(db->db.db_data == NULL); 2155 } else if (db->db_state == DB_FILL) { 2156 /* This buffer was freed and is now being re-filled */ 2157 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2158 } else { 2159 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2160 } 2161 DBUF_VERIFY(db); 2162 2163 if (db->db_blkid == DMU_SPILL_BLKID) { 2164 mutex_enter(&dn->dn_mtx); 2165 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2166 mutex_exit(&dn->dn_mtx); 2167 } 2168 2169 /* 2170 * If this is a bonus buffer, simply copy the bonus data into the 2171 * dnode. It will be written out when the dnode is synced (and it 2172 * will be synced, since it must have been dirty for dbuf_sync to 2173 * be called). 2174 */ 2175 if (db->db_blkid == DMU_BONUS_BLKID) { 2176 dbuf_dirty_record_t **drp; 2177 2178 ASSERT(*datap != NULL); 2179 ASSERT3U(db->db_level, ==, 0); 2180 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2181 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2182 if (*datap != db->db.db_data) { 2183 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2184 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2185 } 2186 db->db_data_pending = NULL; 2187 drp = &db->db_last_dirty; 2188 while (*drp != dr) 2189 drp = &(*drp)->dr_next; 2190 ASSERT(dr->dr_next == NULL); 2191 ASSERT(dr->dr_dbuf == db); 2192 *drp = dr->dr_next; 2193 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2194 ASSERT(db->db_dirtycnt > 0); 2195 db->db_dirtycnt -= 1; 2196 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2197 return; 2198 } 2199 2200 /* 2201 * This function may have dropped the db_mtx lock allowing a dmu_sync 2202 * operation to sneak in. As a result, we need to ensure that we 2203 * don't check the dr_override_state until we have returned from 2204 * dbuf_check_blkptr. 2205 */ 2206 dbuf_check_blkptr(dn, db); 2207 2208 /* 2209 * If this buffer is in the middle of an immdiate write, 2210 * wait for the synchronous IO to complete. 2211 */ 2212 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2213 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2214 cv_wait(&db->db_changed, &db->db_mtx); 2215 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2216 } 2217 2218 if (db->db_state != DB_NOFILL && 2219 dn->dn_object != DMU_META_DNODE_OBJECT && 2220 refcount_count(&db->db_holds) > 1 && 2221 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2222 *datap == db->db_buf) { 2223 /* 2224 * If this buffer is currently "in use" (i.e., there 2225 * are active holds and db_data still references it), 2226 * then make a copy before we start the write so that 2227 * any modifications from the open txg will not leak 2228 * into this write. 2229 * 2230 * NOTE: this copy does not need to be made for 2231 * objects only modified in the syncing context (e.g. 2232 * DNONE_DNODE blocks). 2233 */ 2234 int blksz = arc_buf_size(*datap); 2235 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2236 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2237 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2238 } 2239 db->db_data_pending = dr; 2240 2241 mutex_exit(&db->db_mtx); 2242 2243 dbuf_write(dr, *datap, tx); 2244 2245 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2246 if (dn->dn_object == DMU_META_DNODE_OBJECT) 2247 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2248 else 2249 zio_nowait(dr->dr_zio); 2250 } 2251 2252 void 2253 dbuf_sync_list(list_t *list, dmu_tx_t *tx) 2254 { 2255 dbuf_dirty_record_t *dr; 2256 2257 while (dr = list_head(list)) { 2258 if (dr->dr_zio != NULL) { 2259 /* 2260 * If we find an already initialized zio then we 2261 * are processing the meta-dnode, and we have finished. 2262 * The dbufs for all dnodes are put back on the list 2263 * during processing, so that we can zio_wait() 2264 * these IOs after initiating all child IOs. 2265 */ 2266 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2267 DMU_META_DNODE_OBJECT); 2268 break; 2269 } 2270 list_remove(list, dr); 2271 if (dr->dr_dbuf->db_level > 0) 2272 dbuf_sync_indirect(dr, tx); 2273 else 2274 dbuf_sync_leaf(dr, tx); 2275 } 2276 } 2277 2278 /* ARGSUSED */ 2279 static void 2280 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2281 { 2282 dmu_buf_impl_t *db = vdb; 2283 blkptr_t *bp = zio->io_bp; 2284 blkptr_t *bp_orig = &zio->io_bp_orig; 2285 dnode_t *dn = db->db_dnode; 2286 spa_t *spa = zio->io_spa; 2287 int64_t delta; 2288 uint64_t fill = 0; 2289 int i; 2290 2291 ASSERT(db->db_blkptr == bp); 2292 2293 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2294 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2295 zio->io_prev_space_delta = delta; 2296 2297 if (BP_IS_HOLE(bp)) { 2298 ASSERT(bp->blk_fill == 0); 2299 return; 2300 } 2301 2302 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2303 BP_GET_TYPE(bp) == dn->dn_type) || 2304 (db->db_blkid == DMU_SPILL_BLKID && 2305 BP_GET_TYPE(bp) == dn->dn_bonustype)); 2306 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2307 2308 mutex_enter(&db->db_mtx); 2309 2310 #ifdef ZFS_DEBUG 2311 if (db->db_blkid == DMU_SPILL_BLKID) { 2312 dnode_t *dn = db->db_dnode; 2313 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2314 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2315 db->db_blkptr == &dn->dn_phys->dn_spill); 2316 } 2317 #endif 2318 2319 if (db->db_level == 0) { 2320 mutex_enter(&dn->dn_mtx); 2321 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2322 db->db_blkid != DMU_SPILL_BLKID) 2323 dn->dn_phys->dn_maxblkid = db->db_blkid; 2324 mutex_exit(&dn->dn_mtx); 2325 2326 if (dn->dn_type == DMU_OT_DNODE) { 2327 dnode_phys_t *dnp = db->db.db_data; 2328 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2329 i--, dnp++) { 2330 if (dnp->dn_type != DMU_OT_NONE) 2331 fill++; 2332 } 2333 } else { 2334 fill = 1; 2335 } 2336 } else { 2337 blkptr_t *ibp = db->db.db_data; 2338 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2339 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2340 if (BP_IS_HOLE(ibp)) 2341 continue; 2342 fill += ibp->blk_fill; 2343 } 2344 } 2345 2346 bp->blk_fill = fill; 2347 2348 mutex_exit(&db->db_mtx); 2349 } 2350 2351 /* ARGSUSED */ 2352 static void 2353 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2354 { 2355 dmu_buf_impl_t *db = vdb; 2356 blkptr_t *bp = zio->io_bp; 2357 blkptr_t *bp_orig = &zio->io_bp_orig; 2358 dnode_t *dn = db->db_dnode; 2359 objset_t *os = dn->dn_objset; 2360 uint64_t txg = zio->io_txg; 2361 dbuf_dirty_record_t **drp, *dr; 2362 2363 ASSERT3U(zio->io_error, ==, 0); 2364 ASSERT(db->db_blkptr == bp); 2365 2366 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 2367 ASSERT(BP_EQUAL(bp, bp_orig)); 2368 } else { 2369 dsl_dataset_t *ds = os->os_dsl_dataset; 2370 dmu_tx_t *tx = os->os_synctx; 2371 2372 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2373 dsl_dataset_block_born(ds, bp, tx); 2374 } 2375 2376 mutex_enter(&db->db_mtx); 2377 2378 DBUF_VERIFY(db); 2379 2380 drp = &db->db_last_dirty; 2381 while ((dr = *drp) != db->db_data_pending) 2382 drp = &dr->dr_next; 2383 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2384 ASSERT(dr->dr_txg == txg); 2385 ASSERT(dr->dr_dbuf == db); 2386 ASSERT(dr->dr_next == NULL); 2387 *drp = dr->dr_next; 2388 2389 #ifdef ZFS_DEBUG 2390 if (db->db_blkid == DMU_SPILL_BLKID) { 2391 dnode_t *dn = db->db_dnode; 2392 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2393 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2394 db->db_blkptr == &dn->dn_phys->dn_spill); 2395 } 2396 #endif 2397 2398 if (db->db_level == 0) { 2399 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2400 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2401 if (db->db_state != DB_NOFILL) { 2402 if (dr->dt.dl.dr_data != db->db_buf) 2403 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 2404 db) == 1); 2405 else if (!arc_released(db->db_buf)) 2406 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2407 } 2408 } else { 2409 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2410 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2411 if (!BP_IS_HOLE(db->db_blkptr)) { 2412 int epbs = 2413 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2414 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2415 db->db.db_size); 2416 ASSERT3U(dn->dn_phys->dn_maxblkid 2417 >> (db->db_level * epbs), >=, db->db_blkid); 2418 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2419 } 2420 mutex_destroy(&dr->dt.di.dr_mtx); 2421 list_destroy(&dr->dt.di.dr_children); 2422 } 2423 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2424 2425 cv_broadcast(&db->db_changed); 2426 ASSERT(db->db_dirtycnt > 0); 2427 db->db_dirtycnt -= 1; 2428 db->db_data_pending = NULL; 2429 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2430 } 2431 2432 static void 2433 dbuf_write_nofill_ready(zio_t *zio) 2434 { 2435 dbuf_write_ready(zio, NULL, zio->io_private); 2436 } 2437 2438 static void 2439 dbuf_write_nofill_done(zio_t *zio) 2440 { 2441 dbuf_write_done(zio, NULL, zio->io_private); 2442 } 2443 2444 static void 2445 dbuf_write_override_ready(zio_t *zio) 2446 { 2447 dbuf_dirty_record_t *dr = zio->io_private; 2448 dmu_buf_impl_t *db = dr->dr_dbuf; 2449 2450 dbuf_write_ready(zio, NULL, db); 2451 } 2452 2453 static void 2454 dbuf_write_override_done(zio_t *zio) 2455 { 2456 dbuf_dirty_record_t *dr = zio->io_private; 2457 dmu_buf_impl_t *db = dr->dr_dbuf; 2458 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 2459 2460 mutex_enter(&db->db_mtx); 2461 if (!BP_EQUAL(zio->io_bp, obp)) { 2462 if (!BP_IS_HOLE(obp)) 2463 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 2464 arc_release(dr->dt.dl.dr_data, db); 2465 } 2466 mutex_exit(&db->db_mtx); 2467 2468 dbuf_write_done(zio, NULL, db); 2469 } 2470 2471 static void 2472 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 2473 { 2474 dmu_buf_impl_t *db = dr->dr_dbuf; 2475 dnode_t *dn = db->db_dnode; 2476 objset_t *os = dn->dn_objset; 2477 dmu_buf_impl_t *parent = db->db_parent; 2478 uint64_t txg = tx->tx_txg; 2479 zbookmark_t zb; 2480 zio_prop_t zp; 2481 zio_t *zio; 2482 int wp_flag = 0; 2483 2484 if (db->db_state != DB_NOFILL) { 2485 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 2486 /* 2487 * Private object buffers are released here rather 2488 * than in dbuf_dirty() since they are only modified 2489 * in the syncing context and we don't want the 2490 * overhead of making multiple copies of the data. 2491 */ 2492 if (BP_IS_HOLE(db->db_blkptr)) { 2493 arc_buf_thaw(data); 2494 } else { 2495 dbuf_release_bp(db); 2496 } 2497 } 2498 } 2499 2500 if (parent != dn->dn_dbuf) { 2501 ASSERT(parent && parent->db_data_pending); 2502 ASSERT(db->db_level == parent->db_level-1); 2503 ASSERT(arc_released(parent->db_buf)); 2504 zio = parent->db_data_pending->dr_zio; 2505 } else { 2506 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 2507 db->db_blkid != DMU_SPILL_BLKID) || 2508 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 2509 if (db->db_blkid != DMU_SPILL_BLKID) 2510 ASSERT3P(db->db_blkptr, ==, 2511 &dn->dn_phys->dn_blkptr[db->db_blkid]); 2512 zio = dn->dn_zio; 2513 } 2514 2515 ASSERT(db->db_level == 0 || data == db->db_buf); 2516 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2517 ASSERT(zio); 2518 2519 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 2520 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 2521 db->db.db_object, db->db_level, db->db_blkid); 2522 2523 if (db->db_blkid == DMU_SPILL_BLKID) 2524 wp_flag = WP_SPILL; 2525 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 2526 2527 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 2528 2529 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2530 ASSERT(db->db_state != DB_NOFILL); 2531 dr->dr_zio = zio_write(zio, os->os_spa, txg, 2532 db->db_blkptr, data->b_data, arc_buf_size(data), &zp, 2533 dbuf_write_override_ready, dbuf_write_override_done, dr, 2534 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2535 mutex_enter(&db->db_mtx); 2536 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 2537 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 2538 dr->dt.dl.dr_copies); 2539 mutex_exit(&db->db_mtx); 2540 } else if (db->db_state == DB_NOFILL) { 2541 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); 2542 dr->dr_zio = zio_write(zio, os->os_spa, txg, 2543 db->db_blkptr, NULL, db->db.db_size, &zp, 2544 dbuf_write_nofill_ready, dbuf_write_nofill_done, db, 2545 ZIO_PRIORITY_ASYNC_WRITE, 2546 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 2547 } else { 2548 ASSERT(arc_released(data)); 2549 dr->dr_zio = arc_write(zio, os->os_spa, txg, 2550 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp, 2551 dbuf_write_ready, dbuf_write_done, db, 2552 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2553 } 2554 } 2555