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 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 */ 30 31 #include <sys/zfs_context.h> 32 #include <sys/dmu.h> 33 #include <sys/dmu_send.h> 34 #include <sys/dmu_impl.h> 35 #include <sys/dbuf.h> 36 #include <sys/dmu_objset.h> 37 #include <sys/dsl_dataset.h> 38 #include <sys/dsl_dir.h> 39 #include <sys/dmu_tx.h> 40 #include <sys/spa.h> 41 #include <sys/zio.h> 42 #include <sys/dmu_zfetch.h> 43 #include <sys/sa.h> 44 #include <sys/sa_impl.h> 45 #include <sys/zfeature.h> 46 #include <sys/blkptr.h> 47 #include <sys/range_tree.h> 48 #include <sys/callb.h> 49 50 uint_t zfs_dbuf_evict_key; 51 52 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 53 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 54 55 #ifndef __lint 56 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, 57 dmu_buf_evict_func_t *evict_func_sync, 58 dmu_buf_evict_func_t *evict_func_async, 59 dmu_buf_t **clear_on_evict_dbufp); 60 #endif /* ! __lint */ 61 62 /* 63 * Global data structures and functions for the dbuf cache. 64 */ 65 static kmem_cache_t *dbuf_kmem_cache; 66 static taskq_t *dbu_evict_taskq; 67 68 static kthread_t *dbuf_cache_evict_thread; 69 static kmutex_t dbuf_evict_lock; 70 static kcondvar_t dbuf_evict_cv; 71 static boolean_t dbuf_evict_thread_exit; 72 73 /* 74 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that 75 * are not currently held but have been recently released. These dbufs 76 * are not eligible for arc eviction until they are aged out of the cache. 77 * Dbufs are added to the dbuf cache once the last hold is released. If a 78 * dbuf is later accessed and still exists in the dbuf cache, then it will 79 * be removed from the cache and later re-added to the head of the cache. 80 * Dbufs that are aged out of the cache will be immediately destroyed and 81 * become eligible for arc eviction. 82 */ 83 static multilist_t *dbuf_cache; 84 static refcount_t dbuf_cache_size; 85 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024; 86 87 /* Cap the size of the dbuf cache to log2 fraction of arc size. */ 88 int dbuf_cache_max_shift = 5; 89 90 /* 91 * The dbuf cache uses a three-stage eviction policy: 92 * - A low water marker designates when the dbuf eviction thread 93 * should stop evicting from the dbuf cache. 94 * - When we reach the maximum size (aka mid water mark), we 95 * signal the eviction thread to run. 96 * - The high water mark indicates when the eviction thread 97 * is unable to keep up with the incoming load and eviction must 98 * happen in the context of the calling thread. 99 * 100 * The dbuf cache: 101 * (max size) 102 * low water mid water hi water 103 * +----------------------------------------+----------+----------+ 104 * | | | | 105 * | | | | 106 * | | | | 107 * | | | | 108 * +----------------------------------------+----------+----------+ 109 * stop signal evict 110 * evicting eviction directly 111 * thread 112 * 113 * The high and low water marks indicate the operating range for the eviction 114 * thread. The low water mark is, by default, 90% of the total size of the 115 * cache and the high water mark is at 110% (both of these percentages can be 116 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, 117 * respectively). The eviction thread will try to ensure that the cache remains 118 * within this range by waking up every second and checking if the cache is 119 * above the low water mark. The thread can also be woken up by callers adding 120 * elements into the cache if the cache is larger than the mid water (i.e max 121 * cache size). Once the eviction thread is woken up and eviction is required, 122 * it will continue evicting buffers until it's able to reduce the cache size 123 * to the low water mark. If the cache size continues to grow and hits the high 124 * water mark, then callers adding elments to the cache will begin to evict 125 * directly from the cache until the cache is no longer above the high water 126 * mark. 127 */ 128 129 /* 130 * The percentage above and below the maximum cache size. 131 */ 132 uint_t dbuf_cache_hiwater_pct = 10; 133 uint_t dbuf_cache_lowater_pct = 10; 134 135 /* ARGSUSED */ 136 static int 137 dbuf_cons(void *vdb, void *unused, int kmflag) 138 { 139 dmu_buf_impl_t *db = vdb; 140 bzero(db, sizeof (dmu_buf_impl_t)); 141 142 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 143 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 144 multilist_link_init(&db->db_cache_link); 145 refcount_create(&db->db_holds); 146 147 return (0); 148 } 149 150 /* ARGSUSED */ 151 static void 152 dbuf_dest(void *vdb, void *unused) 153 { 154 dmu_buf_impl_t *db = vdb; 155 mutex_destroy(&db->db_mtx); 156 cv_destroy(&db->db_changed); 157 ASSERT(!multilist_link_active(&db->db_cache_link)); 158 refcount_destroy(&db->db_holds); 159 } 160 161 /* 162 * dbuf hash table routines 163 */ 164 static dbuf_hash_table_t dbuf_hash_table; 165 166 static uint64_t dbuf_hash_count; 167 168 static uint64_t 169 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 170 { 171 uintptr_t osv = (uintptr_t)os; 172 uint64_t crc = -1ULL; 173 174 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 175 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 176 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 177 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 178 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 179 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 180 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 181 182 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 183 184 return (crc); 185 } 186 187 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 188 ((dbuf)->db.db_object == (obj) && \ 189 (dbuf)->db_objset == (os) && \ 190 (dbuf)->db_level == (level) && \ 191 (dbuf)->db_blkid == (blkid)) 192 193 dmu_buf_impl_t * 194 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 195 { 196 dbuf_hash_table_t *h = &dbuf_hash_table; 197 uint64_t hv = dbuf_hash(os, obj, level, blkid); 198 uint64_t idx = hv & h->hash_table_mask; 199 dmu_buf_impl_t *db; 200 201 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 202 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 203 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 204 mutex_enter(&db->db_mtx); 205 if (db->db_state != DB_EVICTING) { 206 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 207 return (db); 208 } 209 mutex_exit(&db->db_mtx); 210 } 211 } 212 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 213 return (NULL); 214 } 215 216 static dmu_buf_impl_t * 217 dbuf_find_bonus(objset_t *os, uint64_t object) 218 { 219 dnode_t *dn; 220 dmu_buf_impl_t *db = NULL; 221 222 if (dnode_hold(os, object, FTAG, &dn) == 0) { 223 rw_enter(&dn->dn_struct_rwlock, RW_READER); 224 if (dn->dn_bonus != NULL) { 225 db = dn->dn_bonus; 226 mutex_enter(&db->db_mtx); 227 } 228 rw_exit(&dn->dn_struct_rwlock); 229 dnode_rele(dn, FTAG); 230 } 231 return (db); 232 } 233 234 /* 235 * Insert an entry into the hash table. If there is already an element 236 * equal to elem in the hash table, then the already existing element 237 * will be returned and the new element will not be inserted. 238 * Otherwise returns NULL. 239 */ 240 static dmu_buf_impl_t * 241 dbuf_hash_insert(dmu_buf_impl_t *db) 242 { 243 dbuf_hash_table_t *h = &dbuf_hash_table; 244 objset_t *os = db->db_objset; 245 uint64_t obj = db->db.db_object; 246 int level = db->db_level; 247 uint64_t blkid = db->db_blkid; 248 uint64_t hv = dbuf_hash(os, obj, level, blkid); 249 uint64_t idx = hv & h->hash_table_mask; 250 dmu_buf_impl_t *dbf; 251 252 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 253 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 254 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 255 mutex_enter(&dbf->db_mtx); 256 if (dbf->db_state != DB_EVICTING) { 257 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 258 return (dbf); 259 } 260 mutex_exit(&dbf->db_mtx); 261 } 262 } 263 264 mutex_enter(&db->db_mtx); 265 db->db_hash_next = h->hash_table[idx]; 266 h->hash_table[idx] = db; 267 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 268 atomic_inc_64(&dbuf_hash_count); 269 270 return (NULL); 271 } 272 273 /* 274 * Remove an entry from the hash table. It must be in the EVICTING state. 275 */ 276 static void 277 dbuf_hash_remove(dmu_buf_impl_t *db) 278 { 279 dbuf_hash_table_t *h = &dbuf_hash_table; 280 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object, 281 db->db_level, db->db_blkid); 282 uint64_t idx = hv & h->hash_table_mask; 283 dmu_buf_impl_t *dbf, **dbp; 284 285 /* 286 * We musn't hold db_mtx to maintain lock ordering: 287 * DBUF_HASH_MUTEX > db_mtx. 288 */ 289 ASSERT(refcount_is_zero(&db->db_holds)); 290 ASSERT(db->db_state == DB_EVICTING); 291 ASSERT(!MUTEX_HELD(&db->db_mtx)); 292 293 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 294 dbp = &h->hash_table[idx]; 295 while ((dbf = *dbp) != db) { 296 dbp = &dbf->db_hash_next; 297 ASSERT(dbf != NULL); 298 } 299 *dbp = db->db_hash_next; 300 db->db_hash_next = NULL; 301 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 302 atomic_dec_64(&dbuf_hash_count); 303 } 304 305 typedef enum { 306 DBVU_EVICTING, 307 DBVU_NOT_EVICTING 308 } dbvu_verify_type_t; 309 310 static void 311 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 312 { 313 #ifdef ZFS_DEBUG 314 int64_t holds; 315 316 if (db->db_user == NULL) 317 return; 318 319 /* Only data blocks support the attachment of user data. */ 320 ASSERT(db->db_level == 0); 321 322 /* Clients must resolve a dbuf before attaching user data. */ 323 ASSERT(db->db.db_data != NULL); 324 ASSERT3U(db->db_state, ==, DB_CACHED); 325 326 holds = refcount_count(&db->db_holds); 327 if (verify_type == DBVU_EVICTING) { 328 /* 329 * Immediate eviction occurs when holds == dirtycnt. 330 * For normal eviction buffers, holds is zero on 331 * eviction, except when dbuf_fix_old_data() calls 332 * dbuf_clear_data(). However, the hold count can grow 333 * during eviction even though db_mtx is held (see 334 * dmu_bonus_hold() for an example), so we can only 335 * test the generic invariant that holds >= dirtycnt. 336 */ 337 ASSERT3U(holds, >=, db->db_dirtycnt); 338 } else { 339 if (db->db_user_immediate_evict == TRUE) 340 ASSERT3U(holds, >=, db->db_dirtycnt); 341 else 342 ASSERT3U(holds, >, 0); 343 } 344 #endif 345 } 346 347 static void 348 dbuf_evict_user(dmu_buf_impl_t *db) 349 { 350 dmu_buf_user_t *dbu = db->db_user; 351 352 ASSERT(MUTEX_HELD(&db->db_mtx)); 353 354 if (dbu == NULL) 355 return; 356 357 dbuf_verify_user(db, DBVU_EVICTING); 358 db->db_user = NULL; 359 360 #ifdef ZFS_DEBUG 361 if (dbu->dbu_clear_on_evict_dbufp != NULL) 362 *dbu->dbu_clear_on_evict_dbufp = NULL; 363 #endif 364 365 /* 366 * There are two eviction callbacks - one that we call synchronously 367 * and one that we invoke via a taskq. The async one is useful for 368 * avoiding lock order reversals and limiting stack depth. 369 * 370 * Note that if we have a sync callback but no async callback, 371 * it's likely that the sync callback will free the structure 372 * containing the dbu. In that case we need to take care to not 373 * dereference dbu after calling the sync evict func. 374 */ 375 boolean_t has_async = (dbu->dbu_evict_func_async != NULL); 376 377 if (dbu->dbu_evict_func_sync != NULL) 378 dbu->dbu_evict_func_sync(dbu); 379 380 if (has_async) { 381 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, 382 dbu, 0, &dbu->dbu_tqent); 383 } 384 } 385 386 boolean_t 387 dbuf_is_metadata(dmu_buf_impl_t *db) 388 { 389 if (db->db_level > 0) { 390 return (B_TRUE); 391 } else { 392 boolean_t is_metadata; 393 394 DB_DNODE_ENTER(db); 395 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 396 DB_DNODE_EXIT(db); 397 398 return (is_metadata); 399 } 400 } 401 402 /* 403 * This function *must* return indices evenly distributed between all 404 * sublists of the multilist. This is needed due to how the dbuf eviction 405 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly 406 * distributed between all sublists and uses this assumption when 407 * deciding which sublist to evict from and how much to evict from it. 408 */ 409 unsigned int 410 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) 411 { 412 dmu_buf_impl_t *db = obj; 413 414 /* 415 * The assumption here, is the hash value for a given 416 * dmu_buf_impl_t will remain constant throughout it's lifetime 417 * (i.e. it's objset, object, level and blkid fields don't change). 418 * Thus, we don't need to store the dbuf's sublist index 419 * on insertion, as this index can be recalculated on removal. 420 * 421 * Also, the low order bits of the hash value are thought to be 422 * distributed evenly. Otherwise, in the case that the multilist 423 * has a power of two number of sublists, each sublists' usage 424 * would not be evenly distributed. 425 */ 426 return (dbuf_hash(db->db_objset, db->db.db_object, 427 db->db_level, db->db_blkid) % 428 multilist_get_num_sublists(ml)); 429 } 430 431 static inline boolean_t 432 dbuf_cache_above_hiwater(void) 433 { 434 uint64_t dbuf_cache_hiwater_bytes = 435 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100; 436 437 return (refcount_count(&dbuf_cache_size) > 438 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes); 439 } 440 441 static inline boolean_t 442 dbuf_cache_above_lowater(void) 443 { 444 uint64_t dbuf_cache_lowater_bytes = 445 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100; 446 447 return (refcount_count(&dbuf_cache_size) > 448 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes); 449 } 450 451 /* 452 * Evict the oldest eligible dbuf from the dbuf cache. 453 */ 454 static void 455 dbuf_evict_one(void) 456 { 457 int idx = multilist_get_random_index(dbuf_cache); 458 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx); 459 460 ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); 461 462 /* 463 * Set the thread's tsd to indicate that it's processing evictions. 464 * Once a thread stops evicting from the dbuf cache it will 465 * reset its tsd to NULL. 466 */ 467 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL); 468 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE); 469 470 dmu_buf_impl_t *db = multilist_sublist_tail(mls); 471 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { 472 db = multilist_sublist_prev(mls, db); 473 } 474 475 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, 476 multilist_sublist_t *, mls); 477 478 if (db != NULL) { 479 multilist_sublist_remove(mls, db); 480 multilist_sublist_unlock(mls); 481 (void) refcount_remove_many(&dbuf_cache_size, 482 db->db.db_size, db); 483 dbuf_destroy(db); 484 } else { 485 multilist_sublist_unlock(mls); 486 } 487 (void) tsd_set(zfs_dbuf_evict_key, NULL); 488 } 489 490 /* 491 * The dbuf evict thread is responsible for aging out dbufs from the 492 * cache. Once the cache has reached it's maximum size, dbufs are removed 493 * and destroyed. The eviction thread will continue running until the size 494 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged 495 * out of the cache it is destroyed and becomes eligible for arc eviction. 496 */ 497 static void 498 dbuf_evict_thread(void) 499 { 500 callb_cpr_t cpr; 501 502 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); 503 504 mutex_enter(&dbuf_evict_lock); 505 while (!dbuf_evict_thread_exit) { 506 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 507 CALLB_CPR_SAFE_BEGIN(&cpr); 508 (void) cv_timedwait_hires(&dbuf_evict_cv, 509 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 510 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); 511 } 512 mutex_exit(&dbuf_evict_lock); 513 514 /* 515 * Keep evicting as long as we're above the low water mark 516 * for the cache. We do this without holding the locks to 517 * minimize lock contention. 518 */ 519 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 520 dbuf_evict_one(); 521 } 522 523 mutex_enter(&dbuf_evict_lock); 524 } 525 526 dbuf_evict_thread_exit = B_FALSE; 527 cv_broadcast(&dbuf_evict_cv); 528 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ 529 thread_exit(); 530 } 531 532 /* 533 * Wake up the dbuf eviction thread if the dbuf cache is at its max size. 534 * If the dbuf cache is at its high water mark, then evict a dbuf from the 535 * dbuf cache using the callers context. 536 */ 537 static void 538 dbuf_evict_notify(void) 539 { 540 541 /* 542 * We use thread specific data to track when a thread has 543 * started processing evictions. This allows us to avoid deeply 544 * nested stacks that would have a call flow similar to this: 545 * 546 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() 547 * ^ | 548 * | | 549 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ 550 * 551 * The dbuf_eviction_thread will always have its tsd set until 552 * that thread exits. All other threads will only set their tsd 553 * if they are participating in the eviction process. This only 554 * happens if the eviction thread is unable to process evictions 555 * fast enough. To keep the dbuf cache size in check, other threads 556 * can evict from the dbuf cache directly. Those threads will set 557 * their tsd values so that we ensure that they only evict one dbuf 558 * from the dbuf cache. 559 */ 560 if (tsd_get(zfs_dbuf_evict_key) != NULL) 561 return; 562 563 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) { 564 boolean_t evict_now = B_FALSE; 565 566 mutex_enter(&dbuf_evict_lock); 567 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) { 568 evict_now = dbuf_cache_above_hiwater(); 569 cv_signal(&dbuf_evict_cv); 570 } 571 mutex_exit(&dbuf_evict_lock); 572 573 if (evict_now) { 574 dbuf_evict_one(); 575 } 576 } 577 } 578 579 void 580 dbuf_init(void) 581 { 582 uint64_t hsize = 1ULL << 16; 583 dbuf_hash_table_t *h = &dbuf_hash_table; 584 int i; 585 586 /* 587 * The hash table is big enough to fill all of physical memory 588 * with an average 4K block size. The table will take up 589 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 590 */ 591 while (hsize * 4096 < physmem * PAGESIZE) 592 hsize <<= 1; 593 594 retry: 595 h->hash_table_mask = hsize - 1; 596 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 597 if (h->hash_table == NULL) { 598 /* XXX - we should really return an error instead of assert */ 599 ASSERT(hsize > (1ULL << 10)); 600 hsize >>= 1; 601 goto retry; 602 } 603 604 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", 605 sizeof (dmu_buf_impl_t), 606 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 607 608 for (i = 0; i < DBUF_MUTEXES; i++) 609 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 610 611 /* 612 * Setup the parameters for the dbuf cache. We cap the size of the 613 * dbuf cache to 1/32nd (default) of the size of the ARC. 614 */ 615 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes, 616 arc_max_bytes() >> dbuf_cache_max_shift); 617 618 /* 619 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 620 * configuration is not required. 621 */ 622 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 623 624 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t), 625 offsetof(dmu_buf_impl_t, db_cache_link), 626 dbuf_cache_multilist_index_func); 627 refcount_create(&dbuf_cache_size); 628 629 tsd_create(&zfs_dbuf_evict_key, NULL); 630 dbuf_evict_thread_exit = B_FALSE; 631 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); 632 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); 633 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, 634 NULL, 0, &p0, TS_RUN, minclsyspri); 635 } 636 637 void 638 dbuf_fini(void) 639 { 640 dbuf_hash_table_t *h = &dbuf_hash_table; 641 int i; 642 643 for (i = 0; i < DBUF_MUTEXES; i++) 644 mutex_destroy(&h->hash_mutexes[i]); 645 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 646 kmem_cache_destroy(dbuf_kmem_cache); 647 taskq_destroy(dbu_evict_taskq); 648 649 mutex_enter(&dbuf_evict_lock); 650 dbuf_evict_thread_exit = B_TRUE; 651 while (dbuf_evict_thread_exit) { 652 cv_signal(&dbuf_evict_cv); 653 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); 654 } 655 mutex_exit(&dbuf_evict_lock); 656 tsd_destroy(&zfs_dbuf_evict_key); 657 658 mutex_destroy(&dbuf_evict_lock); 659 cv_destroy(&dbuf_evict_cv); 660 661 refcount_destroy(&dbuf_cache_size); 662 multilist_destroy(dbuf_cache); 663 } 664 665 /* 666 * Other stuff. 667 */ 668 669 #ifdef ZFS_DEBUG 670 static void 671 dbuf_verify(dmu_buf_impl_t *db) 672 { 673 dnode_t *dn; 674 dbuf_dirty_record_t *dr; 675 676 ASSERT(MUTEX_HELD(&db->db_mtx)); 677 678 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 679 return; 680 681 ASSERT(db->db_objset != NULL); 682 DB_DNODE_ENTER(db); 683 dn = DB_DNODE(db); 684 if (dn == NULL) { 685 ASSERT(db->db_parent == NULL); 686 ASSERT(db->db_blkptr == NULL); 687 } else { 688 ASSERT3U(db->db.db_object, ==, dn->dn_object); 689 ASSERT3P(db->db_objset, ==, dn->dn_objset); 690 ASSERT3U(db->db_level, <, dn->dn_nlevels); 691 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 692 db->db_blkid == DMU_SPILL_BLKID || 693 !avl_is_empty(&dn->dn_dbufs)); 694 } 695 if (db->db_blkid == DMU_BONUS_BLKID) { 696 ASSERT(dn != NULL); 697 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 698 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 699 } else if (db->db_blkid == DMU_SPILL_BLKID) { 700 ASSERT(dn != NULL); 701 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 702 ASSERT0(db->db.db_offset); 703 } else { 704 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 705 } 706 707 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 708 ASSERT(dr->dr_dbuf == db); 709 710 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 711 ASSERT(dr->dr_dbuf == db); 712 713 /* 714 * We can't assert that db_size matches dn_datablksz because it 715 * can be momentarily different when another thread is doing 716 * dnode_set_blksz(). 717 */ 718 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 719 dr = db->db_data_pending; 720 /* 721 * It should only be modified in syncing context, so 722 * make sure we only have one copy of the data. 723 */ 724 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 725 } 726 727 /* verify db->db_blkptr */ 728 if (db->db_blkptr) { 729 if (db->db_parent == dn->dn_dbuf) { 730 /* db is pointed to by the dnode */ 731 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 732 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 733 ASSERT(db->db_parent == NULL); 734 else 735 ASSERT(db->db_parent != NULL); 736 if (db->db_blkid != DMU_SPILL_BLKID) 737 ASSERT3P(db->db_blkptr, ==, 738 &dn->dn_phys->dn_blkptr[db->db_blkid]); 739 } else { 740 /* db is pointed to by an indirect block */ 741 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 742 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 743 ASSERT3U(db->db_parent->db.db_object, ==, 744 db->db.db_object); 745 /* 746 * dnode_grow_indblksz() can make this fail if we don't 747 * have the struct_rwlock. XXX indblksz no longer 748 * grows. safe to do this now? 749 */ 750 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 751 ASSERT3P(db->db_blkptr, ==, 752 ((blkptr_t *)db->db_parent->db.db_data + 753 db->db_blkid % epb)); 754 } 755 } 756 } 757 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 758 (db->db_buf == NULL || db->db_buf->b_data) && 759 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 760 db->db_state != DB_FILL && !dn->dn_free_txg) { 761 /* 762 * If the blkptr isn't set but they have nonzero data, 763 * it had better be dirty, otherwise we'll lose that 764 * data when we evict this buffer. 765 * 766 * There is an exception to this rule for indirect blocks; in 767 * this case, if the indirect block is a hole, we fill in a few 768 * fields on each of the child blocks (importantly, birth time) 769 * to prevent hole birth times from being lost when you 770 * partially fill in a hole. 771 */ 772 if (db->db_dirtycnt == 0) { 773 if (db->db_level == 0) { 774 uint64_t *buf = db->db.db_data; 775 int i; 776 777 for (i = 0; i < db->db.db_size >> 3; i++) { 778 ASSERT(buf[i] == 0); 779 } 780 } else { 781 blkptr_t *bps = db->db.db_data; 782 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, 783 db->db.db_size); 784 /* 785 * We want to verify that all the blkptrs in the 786 * indirect block are holes, but we may have 787 * automatically set up a few fields for them. 788 * We iterate through each blkptr and verify 789 * they only have those fields set. 790 */ 791 for (int i = 0; 792 i < db->db.db_size / sizeof (blkptr_t); 793 i++) { 794 blkptr_t *bp = &bps[i]; 795 ASSERT(ZIO_CHECKSUM_IS_ZERO( 796 &bp->blk_cksum)); 797 ASSERT( 798 DVA_IS_EMPTY(&bp->blk_dva[0]) && 799 DVA_IS_EMPTY(&bp->blk_dva[1]) && 800 DVA_IS_EMPTY(&bp->blk_dva[2])); 801 ASSERT0(bp->blk_fill); 802 ASSERT0(bp->blk_pad[0]); 803 ASSERT0(bp->blk_pad[1]); 804 ASSERT(!BP_IS_EMBEDDED(bp)); 805 ASSERT(BP_IS_HOLE(bp)); 806 ASSERT0(bp->blk_phys_birth); 807 } 808 } 809 } 810 } 811 DB_DNODE_EXIT(db); 812 } 813 #endif 814 815 static void 816 dbuf_clear_data(dmu_buf_impl_t *db) 817 { 818 ASSERT(MUTEX_HELD(&db->db_mtx)); 819 dbuf_evict_user(db); 820 ASSERT3P(db->db_buf, ==, NULL); 821 db->db.db_data = NULL; 822 if (db->db_state != DB_NOFILL) 823 db->db_state = DB_UNCACHED; 824 } 825 826 static void 827 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 828 { 829 ASSERT(MUTEX_HELD(&db->db_mtx)); 830 ASSERT(buf != NULL); 831 832 db->db_buf = buf; 833 ASSERT(buf->b_data != NULL); 834 db->db.db_data = buf->b_data; 835 } 836 837 /* 838 * Loan out an arc_buf for read. Return the loaned arc_buf. 839 */ 840 arc_buf_t * 841 dbuf_loan_arcbuf(dmu_buf_impl_t *db) 842 { 843 arc_buf_t *abuf; 844 845 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 846 mutex_enter(&db->db_mtx); 847 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 848 int blksz = db->db.db_size; 849 spa_t *spa = db->db_objset->os_spa; 850 851 mutex_exit(&db->db_mtx); 852 abuf = arc_loan_buf(spa, B_FALSE, blksz); 853 bcopy(db->db.db_data, abuf->b_data, blksz); 854 } else { 855 abuf = db->db_buf; 856 arc_loan_inuse_buf(abuf, db); 857 db->db_buf = NULL; 858 dbuf_clear_data(db); 859 mutex_exit(&db->db_mtx); 860 } 861 return (abuf); 862 } 863 864 /* 865 * Calculate which level n block references the data at the level 0 offset 866 * provided. 867 */ 868 uint64_t 869 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 870 { 871 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 872 /* 873 * The level n blkid is equal to the level 0 blkid divided by 874 * the number of level 0s in a level n block. 875 * 876 * The level 0 blkid is offset >> datablkshift = 877 * offset / 2^datablkshift. 878 * 879 * The number of level 0s in a level n is the number of block 880 * pointers in an indirect block, raised to the power of level. 881 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 882 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 883 * 884 * Thus, the level n blkid is: offset / 885 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 886 * = offset / 2^(datablkshift + level * 887 * (indblkshift - SPA_BLKPTRSHIFT)) 888 * = offset >> (datablkshift + level * 889 * (indblkshift - SPA_BLKPTRSHIFT)) 890 */ 891 return (offset >> (dn->dn_datablkshift + level * 892 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 893 } else { 894 ASSERT3U(offset, <, dn->dn_datablksz); 895 return (0); 896 } 897 } 898 899 static void 900 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 901 { 902 dmu_buf_impl_t *db = vdb; 903 904 mutex_enter(&db->db_mtx); 905 ASSERT3U(db->db_state, ==, DB_READ); 906 /* 907 * All reads are synchronous, so we must have a hold on the dbuf 908 */ 909 ASSERT(refcount_count(&db->db_holds) > 0); 910 ASSERT(db->db_buf == NULL); 911 ASSERT(db->db.db_data == NULL); 912 if (db->db_level == 0 && db->db_freed_in_flight) { 913 /* we were freed in flight; disregard any error */ 914 arc_release(buf, db); 915 bzero(buf->b_data, db->db.db_size); 916 arc_buf_freeze(buf); 917 db->db_freed_in_flight = FALSE; 918 dbuf_set_data(db, buf); 919 db->db_state = DB_CACHED; 920 } else if (zio == NULL || zio->io_error == 0) { 921 dbuf_set_data(db, buf); 922 db->db_state = DB_CACHED; 923 } else { 924 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 925 ASSERT3P(db->db_buf, ==, NULL); 926 arc_buf_destroy(buf, db); 927 db->db_state = DB_UNCACHED; 928 } 929 cv_broadcast(&db->db_changed); 930 dbuf_rele_and_unlock(db, NULL); 931 } 932 933 static void 934 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 935 { 936 dnode_t *dn; 937 zbookmark_phys_t zb; 938 arc_flags_t aflags = ARC_FLAG_NOWAIT; 939 940 DB_DNODE_ENTER(db); 941 dn = DB_DNODE(db); 942 ASSERT(!refcount_is_zero(&db->db_holds)); 943 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 944 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 945 ASSERT(MUTEX_HELD(&db->db_mtx)); 946 ASSERT(db->db_state == DB_UNCACHED); 947 ASSERT(db->db_buf == NULL); 948 949 if (db->db_blkid == DMU_BONUS_BLKID) { 950 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 951 952 ASSERT3U(bonuslen, <=, db->db.db_size); 953 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 954 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 955 if (bonuslen < DN_MAX_BONUSLEN) 956 bzero(db->db.db_data, DN_MAX_BONUSLEN); 957 if (bonuslen) 958 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 959 DB_DNODE_EXIT(db); 960 db->db_state = DB_CACHED; 961 mutex_exit(&db->db_mtx); 962 return; 963 } 964 965 /* 966 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 967 * processes the delete record and clears the bp while we are waiting 968 * for the dn_mtx (resulting in a "no" from block_freed). 969 */ 970 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 971 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 972 BP_IS_HOLE(db->db_blkptr)))) { 973 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 974 975 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, 976 db->db.db_size)); 977 bzero(db->db.db_data, db->db.db_size); 978 979 if (db->db_blkptr != NULL && db->db_level > 0 && 980 BP_IS_HOLE(db->db_blkptr) && 981 db->db_blkptr->blk_birth != 0) { 982 blkptr_t *bps = db->db.db_data; 983 for (int i = 0; i < ((1 << 984 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); 985 i++) { 986 blkptr_t *bp = &bps[i]; 987 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 988 1 << dn->dn_indblkshift); 989 BP_SET_LSIZE(bp, 990 BP_GET_LEVEL(db->db_blkptr) == 1 ? 991 dn->dn_datablksz : 992 BP_GET_LSIZE(db->db_blkptr)); 993 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); 994 BP_SET_LEVEL(bp, 995 BP_GET_LEVEL(db->db_blkptr) - 1); 996 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); 997 } 998 } 999 DB_DNODE_EXIT(db); 1000 db->db_state = DB_CACHED; 1001 mutex_exit(&db->db_mtx); 1002 return; 1003 } 1004 1005 DB_DNODE_EXIT(db); 1006 1007 db->db_state = DB_READ; 1008 mutex_exit(&db->db_mtx); 1009 1010 if (DBUF_IS_L2CACHEABLE(db)) 1011 aflags |= ARC_FLAG_L2CACHE; 1012 1013 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 1014 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 1015 db->db.db_object, db->db_level, db->db_blkid); 1016 1017 dbuf_add_ref(db, NULL); 1018 1019 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 1020 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 1021 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 1022 &aflags, &zb); 1023 } 1024 1025 /* 1026 * This is our just-in-time copy function. It makes a copy of buffers that 1027 * have been modified in a previous transaction group before we access them in 1028 * the current active group. 1029 * 1030 * This function is used in three places: when we are dirtying a buffer for the 1031 * first time in a txg, when we are freeing a range in a dnode that includes 1032 * this buffer, and when we are accessing a buffer which was received compressed 1033 * and later referenced in a WRITE_BYREF record. 1034 * 1035 * Note that when we are called from dbuf_free_range() we do not put a hold on 1036 * the buffer, we just traverse the active dbuf list for the dnode. 1037 */ 1038 static void 1039 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 1040 { 1041 dbuf_dirty_record_t *dr = db->db_last_dirty; 1042 1043 ASSERT(MUTEX_HELD(&db->db_mtx)); 1044 ASSERT(db->db.db_data != NULL); 1045 ASSERT(db->db_level == 0); 1046 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 1047 1048 if (dr == NULL || 1049 (dr->dt.dl.dr_data != 1050 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 1051 return; 1052 1053 /* 1054 * If the last dirty record for this dbuf has not yet synced 1055 * and its referencing the dbuf data, either: 1056 * reset the reference to point to a new copy, 1057 * or (if there a no active holders) 1058 * just null out the current db_data pointer. 1059 */ 1060 ASSERT(dr->dr_txg >= txg - 2); 1061 if (db->db_blkid == DMU_BONUS_BLKID) { 1062 /* Note that the data bufs here are zio_bufs */ 1063 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 1064 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1065 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 1066 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1067 int size = arc_buf_size(db->db_buf); 1068 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1069 spa_t *spa = db->db_objset->os_spa; 1070 enum zio_compress compress_type = 1071 arc_get_compression(db->db_buf); 1072 1073 if (compress_type == ZIO_COMPRESS_OFF) { 1074 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); 1075 } else { 1076 ASSERT3U(type, ==, ARC_BUFC_DATA); 1077 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, 1078 size, arc_buf_lsize(db->db_buf), compress_type); 1079 } 1080 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 1081 } else { 1082 db->db_buf = NULL; 1083 dbuf_clear_data(db); 1084 } 1085 } 1086 1087 int 1088 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 1089 { 1090 int err = 0; 1091 boolean_t prefetch; 1092 dnode_t *dn; 1093 1094 /* 1095 * We don't have to hold the mutex to check db_state because it 1096 * can't be freed while we have a hold on the buffer. 1097 */ 1098 ASSERT(!refcount_is_zero(&db->db_holds)); 1099 1100 if (db->db_state == DB_NOFILL) 1101 return (SET_ERROR(EIO)); 1102 1103 DB_DNODE_ENTER(db); 1104 dn = DB_DNODE(db); 1105 if ((flags & DB_RF_HAVESTRUCT) == 0) 1106 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1107 1108 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1109 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 1110 DBUF_IS_CACHEABLE(db); 1111 1112 mutex_enter(&db->db_mtx); 1113 if (db->db_state == DB_CACHED) { 1114 /* 1115 * If the arc buf is compressed, we need to decompress it to 1116 * read the data. This could happen during the "zfs receive" of 1117 * a stream which is compressed and deduplicated. 1118 */ 1119 if (db->db_buf != NULL && 1120 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) { 1121 dbuf_fix_old_data(db, 1122 spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1123 err = arc_decompress(db->db_buf); 1124 dbuf_set_data(db, db->db_buf); 1125 } 1126 mutex_exit(&db->db_mtx); 1127 if (prefetch) 1128 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1129 if ((flags & DB_RF_HAVESTRUCT) == 0) 1130 rw_exit(&dn->dn_struct_rwlock); 1131 DB_DNODE_EXIT(db); 1132 } else if (db->db_state == DB_UNCACHED) { 1133 spa_t *spa = dn->dn_objset->os_spa; 1134 boolean_t need_wait = B_FALSE; 1135 1136 if (zio == NULL && 1137 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1138 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 1139 need_wait = B_TRUE; 1140 } 1141 dbuf_read_impl(db, zio, flags); 1142 1143 /* dbuf_read_impl has dropped db_mtx for us */ 1144 1145 if (prefetch) 1146 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1147 1148 if ((flags & DB_RF_HAVESTRUCT) == 0) 1149 rw_exit(&dn->dn_struct_rwlock); 1150 DB_DNODE_EXIT(db); 1151 1152 if (need_wait) 1153 err = zio_wait(zio); 1154 } else { 1155 /* 1156 * Another reader came in while the dbuf was in flight 1157 * between UNCACHED and CACHED. Either a writer will finish 1158 * writing the buffer (sending the dbuf to CACHED) or the 1159 * first reader's request will reach the read_done callback 1160 * and send the dbuf to CACHED. Otherwise, a failure 1161 * occurred and the dbuf went to UNCACHED. 1162 */ 1163 mutex_exit(&db->db_mtx); 1164 if (prefetch) 1165 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1166 if ((flags & DB_RF_HAVESTRUCT) == 0) 1167 rw_exit(&dn->dn_struct_rwlock); 1168 DB_DNODE_EXIT(db); 1169 1170 /* Skip the wait per the caller's request. */ 1171 mutex_enter(&db->db_mtx); 1172 if ((flags & DB_RF_NEVERWAIT) == 0) { 1173 while (db->db_state == DB_READ || 1174 db->db_state == DB_FILL) { 1175 ASSERT(db->db_state == DB_READ || 1176 (flags & DB_RF_HAVESTRUCT) == 0); 1177 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 1178 db, zio_t *, zio); 1179 cv_wait(&db->db_changed, &db->db_mtx); 1180 } 1181 if (db->db_state == DB_UNCACHED) 1182 err = SET_ERROR(EIO); 1183 } 1184 mutex_exit(&db->db_mtx); 1185 } 1186 1187 return (err); 1188 } 1189 1190 static void 1191 dbuf_noread(dmu_buf_impl_t *db) 1192 { 1193 ASSERT(!refcount_is_zero(&db->db_holds)); 1194 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1195 mutex_enter(&db->db_mtx); 1196 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1197 cv_wait(&db->db_changed, &db->db_mtx); 1198 if (db->db_state == DB_UNCACHED) { 1199 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1200 spa_t *spa = db->db_objset->os_spa; 1201 1202 ASSERT(db->db_buf == NULL); 1203 ASSERT(db->db.db_data == NULL); 1204 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); 1205 db->db_state = DB_FILL; 1206 } else if (db->db_state == DB_NOFILL) { 1207 dbuf_clear_data(db); 1208 } else { 1209 ASSERT3U(db->db_state, ==, DB_CACHED); 1210 } 1211 mutex_exit(&db->db_mtx); 1212 } 1213 1214 void 1215 dbuf_unoverride(dbuf_dirty_record_t *dr) 1216 { 1217 dmu_buf_impl_t *db = dr->dr_dbuf; 1218 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 1219 uint64_t txg = dr->dr_txg; 1220 1221 ASSERT(MUTEX_HELD(&db->db_mtx)); 1222 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 1223 ASSERT(db->db_level == 0); 1224 1225 if (db->db_blkid == DMU_BONUS_BLKID || 1226 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 1227 return; 1228 1229 ASSERT(db->db_data_pending != dr); 1230 1231 /* free this block */ 1232 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 1233 zio_free(db->db_objset->os_spa, txg, bp); 1234 1235 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1236 dr->dt.dl.dr_nopwrite = B_FALSE; 1237 1238 /* 1239 * Release the already-written buffer, so we leave it in 1240 * a consistent dirty state. Note that all callers are 1241 * modifying the buffer, so they will immediately do 1242 * another (redundant) arc_release(). Therefore, leave 1243 * the buf thawed to save the effort of freezing & 1244 * immediately re-thawing it. 1245 */ 1246 arc_release(dr->dt.dl.dr_data, db); 1247 } 1248 1249 /* 1250 * Evict (if its unreferenced) or clear (if its referenced) any level-0 1251 * data blocks in the free range, so that any future readers will find 1252 * empty blocks. 1253 */ 1254 void 1255 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 1256 dmu_tx_t *tx) 1257 { 1258 dmu_buf_impl_t db_search; 1259 dmu_buf_impl_t *db, *db_next; 1260 uint64_t txg = tx->tx_txg; 1261 avl_index_t where; 1262 1263 if (end_blkid > dn->dn_maxblkid && 1264 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) 1265 end_blkid = dn->dn_maxblkid; 1266 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 1267 1268 db_search.db_level = 0; 1269 db_search.db_blkid = start_blkid; 1270 db_search.db_state = DB_SEARCH; 1271 1272 mutex_enter(&dn->dn_dbufs_mtx); 1273 db = avl_find(&dn->dn_dbufs, &db_search, &where); 1274 ASSERT3P(db, ==, NULL); 1275 1276 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 1277 1278 for (; db != NULL; db = db_next) { 1279 db_next = AVL_NEXT(&dn->dn_dbufs, db); 1280 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1281 1282 if (db->db_level != 0 || db->db_blkid > end_blkid) { 1283 break; 1284 } 1285 ASSERT3U(db->db_blkid, >=, start_blkid); 1286 1287 /* found a level 0 buffer in the range */ 1288 mutex_enter(&db->db_mtx); 1289 if (dbuf_undirty(db, tx)) { 1290 /* mutex has been dropped and dbuf destroyed */ 1291 continue; 1292 } 1293 1294 if (db->db_state == DB_UNCACHED || 1295 db->db_state == DB_NOFILL || 1296 db->db_state == DB_EVICTING) { 1297 ASSERT(db->db.db_data == NULL); 1298 mutex_exit(&db->db_mtx); 1299 continue; 1300 } 1301 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 1302 /* will be handled in dbuf_read_done or dbuf_rele */ 1303 db->db_freed_in_flight = TRUE; 1304 mutex_exit(&db->db_mtx); 1305 continue; 1306 } 1307 if (refcount_count(&db->db_holds) == 0) { 1308 ASSERT(db->db_buf); 1309 dbuf_destroy(db); 1310 continue; 1311 } 1312 /* The dbuf is referenced */ 1313 1314 if (db->db_last_dirty != NULL) { 1315 dbuf_dirty_record_t *dr = db->db_last_dirty; 1316 1317 if (dr->dr_txg == txg) { 1318 /* 1319 * This buffer is "in-use", re-adjust the file 1320 * size to reflect that this buffer may 1321 * contain new data when we sync. 1322 */ 1323 if (db->db_blkid != DMU_SPILL_BLKID && 1324 db->db_blkid > dn->dn_maxblkid) 1325 dn->dn_maxblkid = db->db_blkid; 1326 dbuf_unoverride(dr); 1327 } else { 1328 /* 1329 * This dbuf is not dirty in the open context. 1330 * Either uncache it (if its not referenced in 1331 * the open context) or reset its contents to 1332 * empty. 1333 */ 1334 dbuf_fix_old_data(db, txg); 1335 } 1336 } 1337 /* clear the contents if its cached */ 1338 if (db->db_state == DB_CACHED) { 1339 ASSERT(db->db.db_data != NULL); 1340 arc_release(db->db_buf, db); 1341 bzero(db->db.db_data, db->db.db_size); 1342 arc_buf_freeze(db->db_buf); 1343 } 1344 1345 mutex_exit(&db->db_mtx); 1346 } 1347 mutex_exit(&dn->dn_dbufs_mtx); 1348 } 1349 1350 void 1351 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1352 { 1353 arc_buf_t *buf, *obuf; 1354 int osize = db->db.db_size; 1355 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1356 dnode_t *dn; 1357 1358 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1359 1360 DB_DNODE_ENTER(db); 1361 dn = DB_DNODE(db); 1362 1363 /* XXX does *this* func really need the lock? */ 1364 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1365 1366 /* 1367 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1368 * is OK, because there can be no other references to the db 1369 * when we are changing its size, so no concurrent DB_FILL can 1370 * be happening. 1371 */ 1372 /* 1373 * XXX we should be doing a dbuf_read, checking the return 1374 * value and returning that up to our callers 1375 */ 1376 dmu_buf_will_dirty(&db->db, tx); 1377 1378 /* create the data buffer for the new block */ 1379 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); 1380 1381 /* copy old block data to the new block */ 1382 obuf = db->db_buf; 1383 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1384 /* zero the remainder */ 1385 if (size > osize) 1386 bzero((uint8_t *)buf->b_data + osize, size - osize); 1387 1388 mutex_enter(&db->db_mtx); 1389 dbuf_set_data(db, buf); 1390 arc_buf_destroy(obuf, db); 1391 db->db.db_size = size; 1392 1393 if (db->db_level == 0) { 1394 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1395 db->db_last_dirty->dt.dl.dr_data = buf; 1396 } 1397 mutex_exit(&db->db_mtx); 1398 1399 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); 1400 DB_DNODE_EXIT(db); 1401 } 1402 1403 void 1404 dbuf_release_bp(dmu_buf_impl_t *db) 1405 { 1406 objset_t *os = db->db_objset; 1407 1408 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1409 ASSERT(arc_released(os->os_phys_buf) || 1410 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1411 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1412 1413 (void) arc_release(db->db_buf, db); 1414 } 1415 1416 /* 1417 * We already have a dirty record for this TXG, and we are being 1418 * dirtied again. 1419 */ 1420 static void 1421 dbuf_redirty(dbuf_dirty_record_t *dr) 1422 { 1423 dmu_buf_impl_t *db = dr->dr_dbuf; 1424 1425 ASSERT(MUTEX_HELD(&db->db_mtx)); 1426 1427 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1428 /* 1429 * If this buffer has already been written out, 1430 * we now need to reset its state. 1431 */ 1432 dbuf_unoverride(dr); 1433 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1434 db->db_state != DB_NOFILL) { 1435 /* Already released on initial dirty, so just thaw. */ 1436 ASSERT(arc_released(db->db_buf)); 1437 arc_buf_thaw(db->db_buf); 1438 } 1439 } 1440 } 1441 1442 dbuf_dirty_record_t * 1443 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1444 { 1445 dnode_t *dn; 1446 objset_t *os; 1447 dbuf_dirty_record_t **drp, *dr; 1448 int drop_struct_lock = FALSE; 1449 int txgoff = tx->tx_txg & TXG_MASK; 1450 1451 ASSERT(tx->tx_txg != 0); 1452 ASSERT(!refcount_is_zero(&db->db_holds)); 1453 DMU_TX_DIRTY_BUF(tx, db); 1454 1455 DB_DNODE_ENTER(db); 1456 dn = DB_DNODE(db); 1457 /* 1458 * Shouldn't dirty a regular buffer in syncing context. Private 1459 * objects may be dirtied in syncing context, but only if they 1460 * were already pre-dirtied in open context. 1461 */ 1462 #ifdef DEBUG 1463 if (dn->dn_objset->os_dsl_dataset != NULL) { 1464 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1465 RW_READER, FTAG); 1466 } 1467 ASSERT(!dmu_tx_is_syncing(tx) || 1468 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1469 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1470 dn->dn_objset->os_dsl_dataset == NULL); 1471 if (dn->dn_objset->os_dsl_dataset != NULL) 1472 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); 1473 #endif 1474 /* 1475 * We make this assert for private objects as well, but after we 1476 * check if we're already dirty. They are allowed to re-dirty 1477 * in syncing context. 1478 */ 1479 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1480 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1481 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1482 1483 mutex_enter(&db->db_mtx); 1484 /* 1485 * XXX make this true for indirects too? The problem is that 1486 * transactions created with dmu_tx_create_assigned() from 1487 * syncing context don't bother holding ahead. 1488 */ 1489 ASSERT(db->db_level != 0 || 1490 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1491 db->db_state == DB_NOFILL); 1492 1493 mutex_enter(&dn->dn_mtx); 1494 /* 1495 * Don't set dirtyctx to SYNC if we're just modifying this as we 1496 * initialize the objset. 1497 */ 1498 if (dn->dn_dirtyctx == DN_UNDIRTIED) { 1499 if (dn->dn_objset->os_dsl_dataset != NULL) { 1500 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1501 RW_READER, FTAG); 1502 } 1503 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1504 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? 1505 DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1506 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1507 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1508 } 1509 if (dn->dn_objset->os_dsl_dataset != NULL) { 1510 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1511 FTAG); 1512 } 1513 } 1514 mutex_exit(&dn->dn_mtx); 1515 1516 if (db->db_blkid == DMU_SPILL_BLKID) 1517 dn->dn_have_spill = B_TRUE; 1518 1519 /* 1520 * If this buffer is already dirty, we're done. 1521 */ 1522 drp = &db->db_last_dirty; 1523 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1524 db->db.db_object == DMU_META_DNODE_OBJECT); 1525 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1526 drp = &dr->dr_next; 1527 if (dr && dr->dr_txg == tx->tx_txg) { 1528 DB_DNODE_EXIT(db); 1529 1530 dbuf_redirty(dr); 1531 mutex_exit(&db->db_mtx); 1532 return (dr); 1533 } 1534 1535 /* 1536 * Only valid if not already dirty. 1537 */ 1538 ASSERT(dn->dn_object == 0 || 1539 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1540 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1541 1542 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1543 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1544 dn->dn_phys->dn_nlevels > db->db_level || 1545 dn->dn_next_nlevels[txgoff] > db->db_level || 1546 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1547 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1548 1549 /* 1550 * We should only be dirtying in syncing context if it's the 1551 * mos or we're initializing the os or it's a special object. 1552 * However, we are allowed to dirty in syncing context provided 1553 * we already dirtied it in open context. Hence we must make 1554 * this assertion only if we're not already dirty. 1555 */ 1556 os = dn->dn_objset; 1557 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); 1558 #ifdef DEBUG 1559 if (dn->dn_objset->os_dsl_dataset != NULL) 1560 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); 1561 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1562 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1563 if (dn->dn_objset->os_dsl_dataset != NULL) 1564 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); 1565 #endif 1566 ASSERT(db->db.db_size != 0); 1567 1568 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1569 1570 if (db->db_blkid != DMU_BONUS_BLKID) { 1571 dmu_objset_willuse_space(os, db->db.db_size, tx); 1572 } 1573 1574 /* 1575 * If this buffer is dirty in an old transaction group we need 1576 * to make a copy of it so that the changes we make in this 1577 * transaction group won't leak out when we sync the older txg. 1578 */ 1579 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1580 if (db->db_level == 0) { 1581 void *data_old = db->db_buf; 1582 1583 if (db->db_state != DB_NOFILL) { 1584 if (db->db_blkid == DMU_BONUS_BLKID) { 1585 dbuf_fix_old_data(db, tx->tx_txg); 1586 data_old = db->db.db_data; 1587 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1588 /* 1589 * Release the data buffer from the cache so 1590 * that we can modify it without impacting 1591 * possible other users of this cached data 1592 * block. Note that indirect blocks and 1593 * private objects are not released until the 1594 * syncing state (since they are only modified 1595 * then). 1596 */ 1597 arc_release(db->db_buf, db); 1598 dbuf_fix_old_data(db, tx->tx_txg); 1599 data_old = db->db_buf; 1600 } 1601 ASSERT(data_old != NULL); 1602 } 1603 dr->dt.dl.dr_data = data_old; 1604 } else { 1605 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1606 list_create(&dr->dt.di.dr_children, 1607 sizeof (dbuf_dirty_record_t), 1608 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1609 } 1610 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1611 dr->dr_accounted = db->db.db_size; 1612 dr->dr_dbuf = db; 1613 dr->dr_txg = tx->tx_txg; 1614 dr->dr_next = *drp; 1615 *drp = dr; 1616 1617 /* 1618 * We could have been freed_in_flight between the dbuf_noread 1619 * and dbuf_dirty. We win, as though the dbuf_noread() had 1620 * happened after the free. 1621 */ 1622 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1623 db->db_blkid != DMU_SPILL_BLKID) { 1624 mutex_enter(&dn->dn_mtx); 1625 if (dn->dn_free_ranges[txgoff] != NULL) { 1626 range_tree_clear(dn->dn_free_ranges[txgoff], 1627 db->db_blkid, 1); 1628 } 1629 mutex_exit(&dn->dn_mtx); 1630 db->db_freed_in_flight = FALSE; 1631 } 1632 1633 /* 1634 * This buffer is now part of this txg 1635 */ 1636 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1637 db->db_dirtycnt += 1; 1638 ASSERT3U(db->db_dirtycnt, <=, 3); 1639 1640 mutex_exit(&db->db_mtx); 1641 1642 if (db->db_blkid == DMU_BONUS_BLKID || 1643 db->db_blkid == DMU_SPILL_BLKID) { 1644 mutex_enter(&dn->dn_mtx); 1645 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1646 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1647 mutex_exit(&dn->dn_mtx); 1648 dnode_setdirty(dn, tx); 1649 DB_DNODE_EXIT(db); 1650 return (dr); 1651 } 1652 1653 /* 1654 * The dn_struct_rwlock prevents db_blkptr from changing 1655 * due to a write from syncing context completing 1656 * while we are running, so we want to acquire it before 1657 * looking at db_blkptr. 1658 */ 1659 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1660 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1661 drop_struct_lock = TRUE; 1662 } 1663 1664 /* 1665 * If we are overwriting a dedup BP, then unless it is snapshotted, 1666 * when we get to syncing context we will need to decrement its 1667 * refcount in the DDT. Prefetch the relevant DDT block so that 1668 * syncing context won't have to wait for the i/o. 1669 */ 1670 ddt_prefetch(os->os_spa, db->db_blkptr); 1671 1672 if (db->db_level == 0) { 1673 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1674 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1675 } 1676 1677 if (db->db_level+1 < dn->dn_nlevels) { 1678 dmu_buf_impl_t *parent = db->db_parent; 1679 dbuf_dirty_record_t *di; 1680 int parent_held = FALSE; 1681 1682 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1683 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1684 1685 parent = dbuf_hold_level(dn, db->db_level+1, 1686 db->db_blkid >> epbs, FTAG); 1687 ASSERT(parent != NULL); 1688 parent_held = TRUE; 1689 } 1690 if (drop_struct_lock) 1691 rw_exit(&dn->dn_struct_rwlock); 1692 ASSERT3U(db->db_level+1, ==, parent->db_level); 1693 di = dbuf_dirty(parent, tx); 1694 if (parent_held) 1695 dbuf_rele(parent, FTAG); 1696 1697 mutex_enter(&db->db_mtx); 1698 /* 1699 * Since we've dropped the mutex, it's possible that 1700 * dbuf_undirty() might have changed this out from under us. 1701 */ 1702 if (db->db_last_dirty == dr || 1703 dn->dn_object == DMU_META_DNODE_OBJECT) { 1704 mutex_enter(&di->dt.di.dr_mtx); 1705 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1706 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1707 list_insert_tail(&di->dt.di.dr_children, dr); 1708 mutex_exit(&di->dt.di.dr_mtx); 1709 dr->dr_parent = di; 1710 } 1711 mutex_exit(&db->db_mtx); 1712 } else { 1713 ASSERT(db->db_level+1 == dn->dn_nlevels); 1714 ASSERT(db->db_blkid < dn->dn_nblkptr); 1715 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1716 mutex_enter(&dn->dn_mtx); 1717 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1718 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1719 mutex_exit(&dn->dn_mtx); 1720 if (drop_struct_lock) 1721 rw_exit(&dn->dn_struct_rwlock); 1722 } 1723 1724 dnode_setdirty(dn, tx); 1725 DB_DNODE_EXIT(db); 1726 return (dr); 1727 } 1728 1729 /* 1730 * Undirty a buffer in the transaction group referenced by the given 1731 * transaction. Return whether this evicted the dbuf. 1732 */ 1733 static boolean_t 1734 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1735 { 1736 dnode_t *dn; 1737 uint64_t txg = tx->tx_txg; 1738 dbuf_dirty_record_t *dr, **drp; 1739 1740 ASSERT(txg != 0); 1741 1742 /* 1743 * Due to our use of dn_nlevels below, this can only be called 1744 * in open context, unless we are operating on the MOS. 1745 * From syncing context, dn_nlevels may be different from the 1746 * dn_nlevels used when dbuf was dirtied. 1747 */ 1748 ASSERT(db->db_objset == 1749 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1750 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1751 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1752 ASSERT0(db->db_level); 1753 ASSERT(MUTEX_HELD(&db->db_mtx)); 1754 1755 /* 1756 * If this buffer is not dirty, we're done. 1757 */ 1758 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1759 if (dr->dr_txg <= txg) 1760 break; 1761 if (dr == NULL || dr->dr_txg < txg) 1762 return (B_FALSE); 1763 ASSERT(dr->dr_txg == txg); 1764 ASSERT(dr->dr_dbuf == db); 1765 1766 DB_DNODE_ENTER(db); 1767 dn = DB_DNODE(db); 1768 1769 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1770 1771 ASSERT(db->db.db_size != 0); 1772 1773 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1774 dr->dr_accounted, txg); 1775 1776 *drp = dr->dr_next; 1777 1778 /* 1779 * Note that there are three places in dbuf_dirty() 1780 * where this dirty record may be put on a list. 1781 * Make sure to do a list_remove corresponding to 1782 * every one of those list_insert calls. 1783 */ 1784 if (dr->dr_parent) { 1785 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1786 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1787 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1788 } else if (db->db_blkid == DMU_SPILL_BLKID || 1789 db->db_level + 1 == dn->dn_nlevels) { 1790 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1791 mutex_enter(&dn->dn_mtx); 1792 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1793 mutex_exit(&dn->dn_mtx); 1794 } 1795 DB_DNODE_EXIT(db); 1796 1797 if (db->db_state != DB_NOFILL) { 1798 dbuf_unoverride(dr); 1799 1800 ASSERT(db->db_buf != NULL); 1801 ASSERT(dr->dt.dl.dr_data != NULL); 1802 if (dr->dt.dl.dr_data != db->db_buf) 1803 arc_buf_destroy(dr->dt.dl.dr_data, db); 1804 } 1805 1806 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1807 1808 ASSERT(db->db_dirtycnt > 0); 1809 db->db_dirtycnt -= 1; 1810 1811 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1812 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); 1813 dbuf_destroy(db); 1814 return (B_TRUE); 1815 } 1816 1817 return (B_FALSE); 1818 } 1819 1820 void 1821 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1822 { 1823 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1824 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1825 1826 ASSERT(tx->tx_txg != 0); 1827 ASSERT(!refcount_is_zero(&db->db_holds)); 1828 1829 /* 1830 * Quick check for dirtyness. For already dirty blocks, this 1831 * reduces runtime of this function by >90%, and overall performance 1832 * by 50% for some workloads (e.g. file deletion with indirect blocks 1833 * cached). 1834 */ 1835 mutex_enter(&db->db_mtx); 1836 dbuf_dirty_record_t *dr; 1837 for (dr = db->db_last_dirty; 1838 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1839 /* 1840 * It's possible that it is already dirty but not cached, 1841 * because there are some calls to dbuf_dirty() that don't 1842 * go through dmu_buf_will_dirty(). 1843 */ 1844 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1845 /* This dbuf is already dirty and cached. */ 1846 dbuf_redirty(dr); 1847 mutex_exit(&db->db_mtx); 1848 return; 1849 } 1850 } 1851 mutex_exit(&db->db_mtx); 1852 1853 DB_DNODE_ENTER(db); 1854 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1855 rf |= DB_RF_HAVESTRUCT; 1856 DB_DNODE_EXIT(db); 1857 (void) dbuf_read(db, NULL, rf); 1858 (void) dbuf_dirty(db, tx); 1859 } 1860 1861 void 1862 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1863 { 1864 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1865 1866 db->db_state = DB_NOFILL; 1867 1868 dmu_buf_will_fill(db_fake, tx); 1869 } 1870 1871 void 1872 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1873 { 1874 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1875 1876 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1877 ASSERT(tx->tx_txg != 0); 1878 ASSERT(db->db_level == 0); 1879 ASSERT(!refcount_is_zero(&db->db_holds)); 1880 1881 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1882 dmu_tx_private_ok(tx)); 1883 1884 dbuf_noread(db); 1885 (void) dbuf_dirty(db, tx); 1886 } 1887 1888 #pragma weak dmu_buf_fill_done = dbuf_fill_done 1889 /* ARGSUSED */ 1890 void 1891 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1892 { 1893 mutex_enter(&db->db_mtx); 1894 DBUF_VERIFY(db); 1895 1896 if (db->db_state == DB_FILL) { 1897 if (db->db_level == 0 && db->db_freed_in_flight) { 1898 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1899 /* we were freed while filling */ 1900 /* XXX dbuf_undirty? */ 1901 bzero(db->db.db_data, db->db.db_size); 1902 db->db_freed_in_flight = FALSE; 1903 } 1904 db->db_state = DB_CACHED; 1905 cv_broadcast(&db->db_changed); 1906 } 1907 mutex_exit(&db->db_mtx); 1908 } 1909 1910 void 1911 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1912 bp_embedded_type_t etype, enum zio_compress comp, 1913 int uncompressed_size, int compressed_size, int byteorder, 1914 dmu_tx_t *tx) 1915 { 1916 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1917 struct dirty_leaf *dl; 1918 dmu_object_type_t type; 1919 1920 if (etype == BP_EMBEDDED_TYPE_DATA) { 1921 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1922 SPA_FEATURE_EMBEDDED_DATA)); 1923 } 1924 1925 DB_DNODE_ENTER(db); 1926 type = DB_DNODE(db)->dn_type; 1927 DB_DNODE_EXIT(db); 1928 1929 ASSERT0(db->db_level); 1930 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1931 1932 dmu_buf_will_not_fill(dbuf, tx); 1933 1934 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1935 dl = &db->db_last_dirty->dt.dl; 1936 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1937 data, comp, uncompressed_size, compressed_size); 1938 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1939 BP_SET_TYPE(&dl->dr_overridden_by, type); 1940 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1941 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1942 1943 dl->dr_override_state = DR_OVERRIDDEN; 1944 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1945 } 1946 1947 /* 1948 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1949 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1950 */ 1951 void 1952 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1953 { 1954 ASSERT(!refcount_is_zero(&db->db_holds)); 1955 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1956 ASSERT(db->db_level == 0); 1957 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); 1958 ASSERT(buf != NULL); 1959 ASSERT(arc_buf_lsize(buf) == db->db.db_size); 1960 ASSERT(tx->tx_txg != 0); 1961 1962 arc_return_buf(buf, db); 1963 ASSERT(arc_released(buf)); 1964 1965 mutex_enter(&db->db_mtx); 1966 1967 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1968 cv_wait(&db->db_changed, &db->db_mtx); 1969 1970 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1971 1972 if (db->db_state == DB_CACHED && 1973 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1974 mutex_exit(&db->db_mtx); 1975 (void) dbuf_dirty(db, tx); 1976 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1977 arc_buf_destroy(buf, db); 1978 xuio_stat_wbuf_copied(); 1979 return; 1980 } 1981 1982 xuio_stat_wbuf_nocopy(); 1983 if (db->db_state == DB_CACHED) { 1984 dbuf_dirty_record_t *dr = db->db_last_dirty; 1985 1986 ASSERT(db->db_buf != NULL); 1987 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1988 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1989 if (!arc_released(db->db_buf)) { 1990 ASSERT(dr->dt.dl.dr_override_state == 1991 DR_OVERRIDDEN); 1992 arc_release(db->db_buf, db); 1993 } 1994 dr->dt.dl.dr_data = buf; 1995 arc_buf_destroy(db->db_buf, db); 1996 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1997 arc_release(db->db_buf, db); 1998 arc_buf_destroy(db->db_buf, db); 1999 } 2000 db->db_buf = NULL; 2001 } 2002 ASSERT(db->db_buf == NULL); 2003 dbuf_set_data(db, buf); 2004 db->db_state = DB_FILL; 2005 mutex_exit(&db->db_mtx); 2006 (void) dbuf_dirty(db, tx); 2007 dmu_buf_fill_done(&db->db, tx); 2008 } 2009 2010 void 2011 dbuf_destroy(dmu_buf_impl_t *db) 2012 { 2013 dnode_t *dn; 2014 dmu_buf_impl_t *parent = db->db_parent; 2015 dmu_buf_impl_t *dndb; 2016 2017 ASSERT(MUTEX_HELD(&db->db_mtx)); 2018 ASSERT(refcount_is_zero(&db->db_holds)); 2019 2020 if (db->db_buf != NULL) { 2021 arc_buf_destroy(db->db_buf, db); 2022 db->db_buf = NULL; 2023 } 2024 2025 if (db->db_blkid == DMU_BONUS_BLKID) { 2026 ASSERT(db->db.db_data != NULL); 2027 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 2028 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2029 db->db_state = DB_UNCACHED; 2030 } 2031 2032 dbuf_clear_data(db); 2033 2034 if (multilist_link_active(&db->db_cache_link)) { 2035 multilist_remove(dbuf_cache, db); 2036 (void) refcount_remove_many(&dbuf_cache_size, 2037 db->db.db_size, db); 2038 } 2039 2040 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 2041 ASSERT(db->db_data_pending == NULL); 2042 2043 db->db_state = DB_EVICTING; 2044 db->db_blkptr = NULL; 2045 2046 /* 2047 * Now that db_state is DB_EVICTING, nobody else can find this via 2048 * the hash table. We can now drop db_mtx, which allows us to 2049 * acquire the dn_dbufs_mtx. 2050 */ 2051 mutex_exit(&db->db_mtx); 2052 2053 DB_DNODE_ENTER(db); 2054 dn = DB_DNODE(db); 2055 dndb = dn->dn_dbuf; 2056 if (db->db_blkid != DMU_BONUS_BLKID) { 2057 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); 2058 if (needlock) 2059 mutex_enter(&dn->dn_dbufs_mtx); 2060 avl_remove(&dn->dn_dbufs, db); 2061 atomic_dec_32(&dn->dn_dbufs_count); 2062 membar_producer(); 2063 DB_DNODE_EXIT(db); 2064 if (needlock) 2065 mutex_exit(&dn->dn_dbufs_mtx); 2066 /* 2067 * Decrementing the dbuf count means that the hold corresponding 2068 * to the removed dbuf is no longer discounted in dnode_move(), 2069 * so the dnode cannot be moved until after we release the hold. 2070 * The membar_producer() ensures visibility of the decremented 2071 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 2072 * release any lock. 2073 */ 2074 dnode_rele(dn, db); 2075 db->db_dnode_handle = NULL; 2076 2077 dbuf_hash_remove(db); 2078 } else { 2079 DB_DNODE_EXIT(db); 2080 } 2081 2082 ASSERT(refcount_is_zero(&db->db_holds)); 2083 2084 db->db_parent = NULL; 2085 2086 ASSERT(db->db_buf == NULL); 2087 ASSERT(db->db.db_data == NULL); 2088 ASSERT(db->db_hash_next == NULL); 2089 ASSERT(db->db_blkptr == NULL); 2090 ASSERT(db->db_data_pending == NULL); 2091 ASSERT(!multilist_link_active(&db->db_cache_link)); 2092 2093 kmem_cache_free(dbuf_kmem_cache, db); 2094 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2095 2096 /* 2097 * If this dbuf is referenced from an indirect dbuf, 2098 * decrement the ref count on the indirect dbuf. 2099 */ 2100 if (parent && parent != dndb) 2101 dbuf_rele(parent, db); 2102 } 2103 2104 /* 2105 * Note: While bpp will always be updated if the function returns success, 2106 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 2107 * this happens when the dnode is the meta-dnode, or a userused or groupused 2108 * object. 2109 */ 2110 static int 2111 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 2112 dmu_buf_impl_t **parentp, blkptr_t **bpp) 2113 { 2114 *parentp = NULL; 2115 *bpp = NULL; 2116 2117 ASSERT(blkid != DMU_BONUS_BLKID); 2118 2119 if (blkid == DMU_SPILL_BLKID) { 2120 mutex_enter(&dn->dn_mtx); 2121 if (dn->dn_have_spill && 2122 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 2123 *bpp = &dn->dn_phys->dn_spill; 2124 else 2125 *bpp = NULL; 2126 dbuf_add_ref(dn->dn_dbuf, NULL); 2127 *parentp = dn->dn_dbuf; 2128 mutex_exit(&dn->dn_mtx); 2129 return (0); 2130 } 2131 2132 int nlevels = 2133 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; 2134 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2135 2136 ASSERT3U(level * epbs, <, 64); 2137 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2138 /* 2139 * This assertion shouldn't trip as long as the max indirect block size 2140 * is less than 1M. The reason for this is that up to that point, 2141 * the number of levels required to address an entire object with blocks 2142 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In 2143 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 2144 * (i.e. we can address the entire object), objects will all use at most 2145 * N-1 levels and the assertion won't overflow. However, once epbs is 2146 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be 2147 * enough to address an entire object, so objects will have 5 levels, 2148 * but then this assertion will overflow. 2149 * 2150 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we 2151 * need to redo this logic to handle overflows. 2152 */ 2153 ASSERT(level >= nlevels || 2154 ((nlevels - level - 1) * epbs) + 2155 highbit64(dn->dn_phys->dn_nblkptr) <= 64); 2156 if (level >= nlevels || 2157 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << 2158 ((nlevels - level - 1) * epbs)) || 2159 (fail_sparse && 2160 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 2161 /* the buffer has no parent yet */ 2162 return (SET_ERROR(ENOENT)); 2163 } else if (level < nlevels-1) { 2164 /* this block is referenced from an indirect block */ 2165 int err = dbuf_hold_impl(dn, level+1, 2166 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 2167 if (err) 2168 return (err); 2169 err = dbuf_read(*parentp, NULL, 2170 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 2171 if (err) { 2172 dbuf_rele(*parentp, NULL); 2173 *parentp = NULL; 2174 return (err); 2175 } 2176 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 2177 (blkid & ((1ULL << epbs) - 1)); 2178 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) 2179 ASSERT(BP_IS_HOLE(*bpp)); 2180 return (0); 2181 } else { 2182 /* the block is referenced from the dnode */ 2183 ASSERT3U(level, ==, nlevels-1); 2184 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 2185 blkid < dn->dn_phys->dn_nblkptr); 2186 if (dn->dn_dbuf) { 2187 dbuf_add_ref(dn->dn_dbuf, NULL); 2188 *parentp = dn->dn_dbuf; 2189 } 2190 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 2191 return (0); 2192 } 2193 } 2194 2195 static dmu_buf_impl_t * 2196 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 2197 dmu_buf_impl_t *parent, blkptr_t *blkptr) 2198 { 2199 objset_t *os = dn->dn_objset; 2200 dmu_buf_impl_t *db, *odb; 2201 2202 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2203 ASSERT(dn->dn_type != DMU_OT_NONE); 2204 2205 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); 2206 2207 db->db_objset = os; 2208 db->db.db_object = dn->dn_object; 2209 db->db_level = level; 2210 db->db_blkid = blkid; 2211 db->db_last_dirty = NULL; 2212 db->db_dirtycnt = 0; 2213 db->db_dnode_handle = dn->dn_handle; 2214 db->db_parent = parent; 2215 db->db_blkptr = blkptr; 2216 2217 db->db_user = NULL; 2218 db->db_user_immediate_evict = FALSE; 2219 db->db_freed_in_flight = FALSE; 2220 db->db_pending_evict = FALSE; 2221 2222 if (blkid == DMU_BONUS_BLKID) { 2223 ASSERT3P(parent, ==, dn->dn_dbuf); 2224 db->db.db_size = DN_MAX_BONUSLEN - 2225 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 2226 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 2227 db->db.db_offset = DMU_BONUS_BLKID; 2228 db->db_state = DB_UNCACHED; 2229 /* the bonus dbuf is not placed in the hash table */ 2230 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2231 return (db); 2232 } else if (blkid == DMU_SPILL_BLKID) { 2233 db->db.db_size = (blkptr != NULL) ? 2234 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 2235 db->db.db_offset = 0; 2236 } else { 2237 int blocksize = 2238 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 2239 db->db.db_size = blocksize; 2240 db->db.db_offset = db->db_blkid * blocksize; 2241 } 2242 2243 /* 2244 * Hold the dn_dbufs_mtx while we get the new dbuf 2245 * in the hash table *and* added to the dbufs list. 2246 * This prevents a possible deadlock with someone 2247 * trying to look up this dbuf before its added to the 2248 * dn_dbufs list. 2249 */ 2250 mutex_enter(&dn->dn_dbufs_mtx); 2251 db->db_state = DB_EVICTING; 2252 if ((odb = dbuf_hash_insert(db)) != NULL) { 2253 /* someone else inserted it first */ 2254 kmem_cache_free(dbuf_kmem_cache, db); 2255 mutex_exit(&dn->dn_dbufs_mtx); 2256 return (odb); 2257 } 2258 avl_add(&dn->dn_dbufs, db); 2259 2260 db->db_state = DB_UNCACHED; 2261 mutex_exit(&dn->dn_dbufs_mtx); 2262 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2263 2264 if (parent && parent != dn->dn_dbuf) 2265 dbuf_add_ref(parent, db); 2266 2267 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 2268 refcount_count(&dn->dn_holds) > 0); 2269 (void) refcount_add(&dn->dn_holds, db); 2270 atomic_inc_32(&dn->dn_dbufs_count); 2271 2272 dprintf_dbuf(db, "db=%p\n", db); 2273 2274 return (db); 2275 } 2276 2277 typedef struct dbuf_prefetch_arg { 2278 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2279 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2280 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2281 int dpa_curlevel; /* The current level that we're reading */ 2282 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ 2283 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2284 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2285 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2286 } dbuf_prefetch_arg_t; 2287 2288 /* 2289 * Actually issue the prefetch read for the block given. 2290 */ 2291 static void 2292 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2293 { 2294 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2295 return; 2296 2297 arc_flags_t aflags = 2298 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2299 2300 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2301 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2302 ASSERT(dpa->dpa_zio != NULL); 2303 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2304 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2305 &aflags, &dpa->dpa_zb); 2306 } 2307 2308 /* 2309 * Called when an indirect block above our prefetch target is read in. This 2310 * will either read in the next indirect block down the tree or issue the actual 2311 * prefetch if the next block down is our target. 2312 */ 2313 static void 2314 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2315 { 2316 dbuf_prefetch_arg_t *dpa = private; 2317 2318 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2319 ASSERT3S(dpa->dpa_curlevel, >, 0); 2320 2321 /* 2322 * The dpa_dnode is only valid if we are called with a NULL 2323 * zio. This indicates that the arc_read() returned without 2324 * first calling zio_read() to issue a physical read. Once 2325 * a physical read is made the dpa_dnode must be invalidated 2326 * as the locks guarding it may have been dropped. If the 2327 * dpa_dnode is still valid, then we want to add it to the dbuf 2328 * cache. To do so, we must hold the dbuf associated with the block 2329 * we just prefetched, read its contents so that we associate it 2330 * with an arc_buf_t, and then release it. 2331 */ 2332 if (zio != NULL) { 2333 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2334 if (zio->io_flags & ZIO_FLAG_RAW) { 2335 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); 2336 } else { 2337 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2338 } 2339 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2340 2341 dpa->dpa_dnode = NULL; 2342 } else if (dpa->dpa_dnode != NULL) { 2343 uint64_t curblkid = dpa->dpa_zb.zb_blkid >> 2344 (dpa->dpa_epbs * (dpa->dpa_curlevel - 2345 dpa->dpa_zb.zb_level)); 2346 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, 2347 dpa->dpa_curlevel, curblkid, FTAG); 2348 (void) dbuf_read(db, NULL, 2349 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); 2350 dbuf_rele(db, FTAG); 2351 } 2352 2353 dpa->dpa_curlevel--; 2354 2355 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2356 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2357 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2358 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2359 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2360 kmem_free(dpa, sizeof (*dpa)); 2361 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2362 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2363 dbuf_issue_final_prefetch(dpa, bp); 2364 kmem_free(dpa, sizeof (*dpa)); 2365 } else { 2366 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2367 zbookmark_phys_t zb; 2368 2369 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2370 2371 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2372 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2373 2374 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2375 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2376 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2377 &iter_aflags, &zb); 2378 } 2379 2380 arc_buf_destroy(abuf, private); 2381 } 2382 2383 /* 2384 * Issue prefetch reads for the given block on the given level. If the indirect 2385 * blocks above that block are not in memory, we will read them in 2386 * asynchronously. As a result, this call never blocks waiting for a read to 2387 * complete. 2388 */ 2389 void 2390 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2391 arc_flags_t aflags) 2392 { 2393 blkptr_t bp; 2394 int epbs, nlevels, curlevel; 2395 uint64_t curblkid; 2396 2397 ASSERT(blkid != DMU_BONUS_BLKID); 2398 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2399 2400 if (blkid > dn->dn_maxblkid) 2401 return; 2402 2403 if (dnode_block_freed(dn, blkid)) 2404 return; 2405 2406 /* 2407 * This dnode hasn't been written to disk yet, so there's nothing to 2408 * prefetch. 2409 */ 2410 nlevels = dn->dn_phys->dn_nlevels; 2411 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2412 return; 2413 2414 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2415 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2416 return; 2417 2418 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2419 level, blkid); 2420 if (db != NULL) { 2421 mutex_exit(&db->db_mtx); 2422 /* 2423 * This dbuf already exists. It is either CACHED, or 2424 * (we assume) about to be read or filled. 2425 */ 2426 return; 2427 } 2428 2429 /* 2430 * Find the closest ancestor (indirect block) of the target block 2431 * that is present in the cache. In this indirect block, we will 2432 * find the bp that is at curlevel, curblkid. 2433 */ 2434 curlevel = level; 2435 curblkid = blkid; 2436 while (curlevel < nlevels - 1) { 2437 int parent_level = curlevel + 1; 2438 uint64_t parent_blkid = curblkid >> epbs; 2439 dmu_buf_impl_t *db; 2440 2441 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2442 FALSE, TRUE, FTAG, &db) == 0) { 2443 blkptr_t *bpp = db->db_buf->b_data; 2444 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2445 dbuf_rele(db, FTAG); 2446 break; 2447 } 2448 2449 curlevel = parent_level; 2450 curblkid = parent_blkid; 2451 } 2452 2453 if (curlevel == nlevels - 1) { 2454 /* No cached indirect blocks found. */ 2455 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2456 bp = dn->dn_phys->dn_blkptr[curblkid]; 2457 } 2458 if (BP_IS_HOLE(&bp)) 2459 return; 2460 2461 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2462 2463 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2464 ZIO_FLAG_CANFAIL); 2465 2466 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2467 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2468 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2469 dn->dn_object, level, blkid); 2470 dpa->dpa_curlevel = curlevel; 2471 dpa->dpa_prio = prio; 2472 dpa->dpa_aflags = aflags; 2473 dpa->dpa_spa = dn->dn_objset->os_spa; 2474 dpa->dpa_dnode = dn; 2475 dpa->dpa_epbs = epbs; 2476 dpa->dpa_zio = pio; 2477 2478 /* 2479 * If we have the indirect just above us, no need to do the asynchronous 2480 * prefetch chain; we'll just run the last step ourselves. If we're at 2481 * a higher level, though, we want to issue the prefetches for all the 2482 * indirect blocks asynchronously, so we can go on with whatever we were 2483 * doing. 2484 */ 2485 if (curlevel == level) { 2486 ASSERT3U(curblkid, ==, blkid); 2487 dbuf_issue_final_prefetch(dpa, &bp); 2488 kmem_free(dpa, sizeof (*dpa)); 2489 } else { 2490 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2491 zbookmark_phys_t zb; 2492 2493 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2494 dn->dn_object, curlevel, curblkid); 2495 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2496 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2497 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2498 &iter_aflags, &zb); 2499 } 2500 /* 2501 * We use pio here instead of dpa_zio since it's possible that 2502 * dpa may have already been freed. 2503 */ 2504 zio_nowait(pio); 2505 } 2506 2507 /* 2508 * Returns with db_holds incremented, and db_mtx not held. 2509 * Note: dn_struct_rwlock must be held. 2510 */ 2511 int 2512 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2513 boolean_t fail_sparse, boolean_t fail_uncached, 2514 void *tag, dmu_buf_impl_t **dbp) 2515 { 2516 dmu_buf_impl_t *db, *parent = NULL; 2517 2518 ASSERT(blkid != DMU_BONUS_BLKID); 2519 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2520 ASSERT3U(dn->dn_nlevels, >, level); 2521 2522 *dbp = NULL; 2523 top: 2524 /* dbuf_find() returns with db_mtx held */ 2525 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2526 2527 if (db == NULL) { 2528 blkptr_t *bp = NULL; 2529 int err; 2530 2531 if (fail_uncached) 2532 return (SET_ERROR(ENOENT)); 2533 2534 ASSERT3P(parent, ==, NULL); 2535 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2536 if (fail_sparse) { 2537 if (err == 0 && bp && BP_IS_HOLE(bp)) 2538 err = SET_ERROR(ENOENT); 2539 if (err) { 2540 if (parent) 2541 dbuf_rele(parent, NULL); 2542 return (err); 2543 } 2544 } 2545 if (err && err != ENOENT) 2546 return (err); 2547 db = dbuf_create(dn, level, blkid, parent, bp); 2548 } 2549 2550 if (fail_uncached && db->db_state != DB_CACHED) { 2551 mutex_exit(&db->db_mtx); 2552 return (SET_ERROR(ENOENT)); 2553 } 2554 2555 if (db->db_buf != NULL) 2556 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2557 2558 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2559 2560 /* 2561 * If this buffer is currently syncing out, and we are are 2562 * still referencing it from db_data, we need to make a copy 2563 * of it in case we decide we want to dirty it again in this txg. 2564 */ 2565 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2566 dn->dn_object != DMU_META_DNODE_OBJECT && 2567 db->db_state == DB_CACHED && db->db_data_pending) { 2568 dbuf_dirty_record_t *dr = db->db_data_pending; 2569 2570 if (dr->dt.dl.dr_data == db->db_buf) { 2571 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2572 2573 dbuf_set_data(db, 2574 arc_alloc_buf(dn->dn_objset->os_spa, db, type, 2575 db->db.db_size)); 2576 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2577 db->db.db_size); 2578 } 2579 } 2580 2581 if (multilist_link_active(&db->db_cache_link)) { 2582 ASSERT(refcount_is_zero(&db->db_holds)); 2583 multilist_remove(dbuf_cache, db); 2584 (void) refcount_remove_many(&dbuf_cache_size, 2585 db->db.db_size, db); 2586 } 2587 (void) refcount_add(&db->db_holds, tag); 2588 DBUF_VERIFY(db); 2589 mutex_exit(&db->db_mtx); 2590 2591 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2592 if (parent) 2593 dbuf_rele(parent, NULL); 2594 2595 ASSERT3P(DB_DNODE(db), ==, dn); 2596 ASSERT3U(db->db_blkid, ==, blkid); 2597 ASSERT3U(db->db_level, ==, level); 2598 *dbp = db; 2599 2600 return (0); 2601 } 2602 2603 dmu_buf_impl_t * 2604 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2605 { 2606 return (dbuf_hold_level(dn, 0, blkid, tag)); 2607 } 2608 2609 dmu_buf_impl_t * 2610 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2611 { 2612 dmu_buf_impl_t *db; 2613 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2614 return (err ? NULL : db); 2615 } 2616 2617 void 2618 dbuf_create_bonus(dnode_t *dn) 2619 { 2620 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2621 2622 ASSERT(dn->dn_bonus == NULL); 2623 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2624 } 2625 2626 int 2627 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2628 { 2629 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2630 dnode_t *dn; 2631 2632 if (db->db_blkid != DMU_SPILL_BLKID) 2633 return (SET_ERROR(ENOTSUP)); 2634 if (blksz == 0) 2635 blksz = SPA_MINBLOCKSIZE; 2636 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2637 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2638 2639 DB_DNODE_ENTER(db); 2640 dn = DB_DNODE(db); 2641 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2642 dbuf_new_size(db, blksz, tx); 2643 rw_exit(&dn->dn_struct_rwlock); 2644 DB_DNODE_EXIT(db); 2645 2646 return (0); 2647 } 2648 2649 void 2650 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2651 { 2652 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2653 } 2654 2655 #pragma weak dmu_buf_add_ref = dbuf_add_ref 2656 void 2657 dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2658 { 2659 int64_t holds = refcount_add(&db->db_holds, tag); 2660 ASSERT3S(holds, >, 1); 2661 } 2662 2663 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2664 boolean_t 2665 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2666 void *tag) 2667 { 2668 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2669 dmu_buf_impl_t *found_db; 2670 boolean_t result = B_FALSE; 2671 2672 if (db->db_blkid == DMU_BONUS_BLKID) 2673 found_db = dbuf_find_bonus(os, obj); 2674 else 2675 found_db = dbuf_find(os, obj, 0, blkid); 2676 2677 if (found_db != NULL) { 2678 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2679 (void) refcount_add(&db->db_holds, tag); 2680 result = B_TRUE; 2681 } 2682 mutex_exit(&db->db_mtx); 2683 } 2684 return (result); 2685 } 2686 2687 /* 2688 * If you call dbuf_rele() you had better not be referencing the dnode handle 2689 * unless you have some other direct or indirect hold on the dnode. (An indirect 2690 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2691 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2692 * dnode's parent dbuf evicting its dnode handles. 2693 */ 2694 void 2695 dbuf_rele(dmu_buf_impl_t *db, void *tag) 2696 { 2697 mutex_enter(&db->db_mtx); 2698 dbuf_rele_and_unlock(db, tag); 2699 } 2700 2701 void 2702 dmu_buf_rele(dmu_buf_t *db, void *tag) 2703 { 2704 dbuf_rele((dmu_buf_impl_t *)db, tag); 2705 } 2706 2707 /* 2708 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2709 * db_dirtycnt and db_holds to be updated atomically. 2710 */ 2711 void 2712 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2713 { 2714 int64_t holds; 2715 2716 ASSERT(MUTEX_HELD(&db->db_mtx)); 2717 DBUF_VERIFY(db); 2718 2719 /* 2720 * Remove the reference to the dbuf before removing its hold on the 2721 * dnode so we can guarantee in dnode_move() that a referenced bonus 2722 * buffer has a corresponding dnode hold. 2723 */ 2724 holds = refcount_remove(&db->db_holds, tag); 2725 ASSERT(holds >= 0); 2726 2727 /* 2728 * We can't freeze indirects if there is a possibility that they 2729 * may be modified in the current syncing context. 2730 */ 2731 if (db->db_buf != NULL && 2732 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { 2733 arc_buf_freeze(db->db_buf); 2734 } 2735 2736 if (holds == db->db_dirtycnt && 2737 db->db_level == 0 && db->db_user_immediate_evict) 2738 dbuf_evict_user(db); 2739 2740 if (holds == 0) { 2741 if (db->db_blkid == DMU_BONUS_BLKID) { 2742 dnode_t *dn; 2743 boolean_t evict_dbuf = db->db_pending_evict; 2744 2745 /* 2746 * If the dnode moves here, we cannot cross this 2747 * barrier until the move completes. 2748 */ 2749 DB_DNODE_ENTER(db); 2750 2751 dn = DB_DNODE(db); 2752 atomic_dec_32(&dn->dn_dbufs_count); 2753 2754 /* 2755 * Decrementing the dbuf count means that the bonus 2756 * buffer's dnode hold is no longer discounted in 2757 * dnode_move(). The dnode cannot move until after 2758 * the dnode_rele() below. 2759 */ 2760 DB_DNODE_EXIT(db); 2761 2762 /* 2763 * Do not reference db after its lock is dropped. 2764 * Another thread may evict it. 2765 */ 2766 mutex_exit(&db->db_mtx); 2767 2768 if (evict_dbuf) 2769 dnode_evict_bonus(dn); 2770 2771 dnode_rele(dn, db); 2772 } else if (db->db_buf == NULL) { 2773 /* 2774 * This is a special case: we never associated this 2775 * dbuf with any data allocated from the ARC. 2776 */ 2777 ASSERT(db->db_state == DB_UNCACHED || 2778 db->db_state == DB_NOFILL); 2779 dbuf_destroy(db); 2780 } else if (arc_released(db->db_buf)) { 2781 /* 2782 * This dbuf has anonymous data associated with it. 2783 */ 2784 dbuf_destroy(db); 2785 } else { 2786 boolean_t do_arc_evict = B_FALSE; 2787 blkptr_t bp; 2788 spa_t *spa = dmu_objset_spa(db->db_objset); 2789 2790 if (!DBUF_IS_CACHEABLE(db) && 2791 db->db_blkptr != NULL && 2792 !BP_IS_HOLE(db->db_blkptr) && 2793 !BP_IS_EMBEDDED(db->db_blkptr)) { 2794 do_arc_evict = B_TRUE; 2795 bp = *db->db_blkptr; 2796 } 2797 2798 if (!DBUF_IS_CACHEABLE(db) || 2799 db->db_pending_evict) { 2800 dbuf_destroy(db); 2801 } else if (!multilist_link_active(&db->db_cache_link)) { 2802 multilist_insert(dbuf_cache, db); 2803 (void) refcount_add_many(&dbuf_cache_size, 2804 db->db.db_size, db); 2805 mutex_exit(&db->db_mtx); 2806 2807 dbuf_evict_notify(); 2808 } 2809 2810 if (do_arc_evict) 2811 arc_freed(spa, &bp); 2812 } 2813 } else { 2814 mutex_exit(&db->db_mtx); 2815 } 2816 2817 } 2818 2819 #pragma weak dmu_buf_refcount = dbuf_refcount 2820 uint64_t 2821 dbuf_refcount(dmu_buf_impl_t *db) 2822 { 2823 return (refcount_count(&db->db_holds)); 2824 } 2825 2826 void * 2827 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2828 dmu_buf_user_t *new_user) 2829 { 2830 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2831 2832 mutex_enter(&db->db_mtx); 2833 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2834 if (db->db_user == old_user) 2835 db->db_user = new_user; 2836 else 2837 old_user = db->db_user; 2838 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2839 mutex_exit(&db->db_mtx); 2840 2841 return (old_user); 2842 } 2843 2844 void * 2845 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2846 { 2847 return (dmu_buf_replace_user(db_fake, NULL, user)); 2848 } 2849 2850 void * 2851 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2852 { 2853 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2854 2855 db->db_user_immediate_evict = TRUE; 2856 return (dmu_buf_set_user(db_fake, user)); 2857 } 2858 2859 void * 2860 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2861 { 2862 return (dmu_buf_replace_user(db_fake, user, NULL)); 2863 } 2864 2865 void * 2866 dmu_buf_get_user(dmu_buf_t *db_fake) 2867 { 2868 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2869 2870 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2871 return (db->db_user); 2872 } 2873 2874 void 2875 dmu_buf_user_evict_wait() 2876 { 2877 taskq_wait(dbu_evict_taskq); 2878 } 2879 2880 blkptr_t * 2881 dmu_buf_get_blkptr(dmu_buf_t *db) 2882 { 2883 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2884 return (dbi->db_blkptr); 2885 } 2886 2887 objset_t * 2888 dmu_buf_get_objset(dmu_buf_t *db) 2889 { 2890 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2891 return (dbi->db_objset); 2892 } 2893 2894 dnode_t * 2895 dmu_buf_dnode_enter(dmu_buf_t *db) 2896 { 2897 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2898 DB_DNODE_ENTER(dbi); 2899 return (DB_DNODE(dbi)); 2900 } 2901 2902 void 2903 dmu_buf_dnode_exit(dmu_buf_t *db) 2904 { 2905 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2906 DB_DNODE_EXIT(dbi); 2907 } 2908 2909 static void 2910 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2911 { 2912 /* ASSERT(dmu_tx_is_syncing(tx) */ 2913 ASSERT(MUTEX_HELD(&db->db_mtx)); 2914 2915 if (db->db_blkptr != NULL) 2916 return; 2917 2918 if (db->db_blkid == DMU_SPILL_BLKID) { 2919 db->db_blkptr = &dn->dn_phys->dn_spill; 2920 BP_ZERO(db->db_blkptr); 2921 return; 2922 } 2923 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2924 /* 2925 * This buffer was allocated at a time when there was 2926 * no available blkptrs from the dnode, or it was 2927 * inappropriate to hook it in (i.e., nlevels mis-match). 2928 */ 2929 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2930 ASSERT(db->db_parent == NULL); 2931 db->db_parent = dn->dn_dbuf; 2932 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2933 DBUF_VERIFY(db); 2934 } else { 2935 dmu_buf_impl_t *parent = db->db_parent; 2936 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2937 2938 ASSERT(dn->dn_phys->dn_nlevels > 1); 2939 if (parent == NULL) { 2940 mutex_exit(&db->db_mtx); 2941 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2942 parent = dbuf_hold_level(dn, db->db_level + 1, 2943 db->db_blkid >> epbs, db); 2944 rw_exit(&dn->dn_struct_rwlock); 2945 mutex_enter(&db->db_mtx); 2946 db->db_parent = parent; 2947 } 2948 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2949 (db->db_blkid & ((1ULL << epbs) - 1)); 2950 DBUF_VERIFY(db); 2951 } 2952 } 2953 2954 static void 2955 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2956 { 2957 dmu_buf_impl_t *db = dr->dr_dbuf; 2958 dnode_t *dn; 2959 zio_t *zio; 2960 2961 ASSERT(dmu_tx_is_syncing(tx)); 2962 2963 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2964 2965 mutex_enter(&db->db_mtx); 2966 2967 ASSERT(db->db_level > 0); 2968 DBUF_VERIFY(db); 2969 2970 /* Read the block if it hasn't been read yet. */ 2971 if (db->db_buf == NULL) { 2972 mutex_exit(&db->db_mtx); 2973 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2974 mutex_enter(&db->db_mtx); 2975 } 2976 ASSERT3U(db->db_state, ==, DB_CACHED); 2977 ASSERT(db->db_buf != NULL); 2978 2979 DB_DNODE_ENTER(db); 2980 dn = DB_DNODE(db); 2981 /* Indirect block size must match what the dnode thinks it is. */ 2982 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2983 dbuf_check_blkptr(dn, db); 2984 DB_DNODE_EXIT(db); 2985 2986 /* Provide the pending dirty record to child dbufs */ 2987 db->db_data_pending = dr; 2988 2989 mutex_exit(&db->db_mtx); 2990 dbuf_write(dr, db->db_buf, tx); 2991 2992 zio = dr->dr_zio; 2993 mutex_enter(&dr->dt.di.dr_mtx); 2994 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2995 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2996 mutex_exit(&dr->dt.di.dr_mtx); 2997 zio_nowait(zio); 2998 } 2999 3000 static void 3001 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3002 { 3003 arc_buf_t **datap = &dr->dt.dl.dr_data; 3004 dmu_buf_impl_t *db = dr->dr_dbuf; 3005 dnode_t *dn; 3006 objset_t *os; 3007 uint64_t txg = tx->tx_txg; 3008 3009 ASSERT(dmu_tx_is_syncing(tx)); 3010 3011 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3012 3013 mutex_enter(&db->db_mtx); 3014 /* 3015 * To be synced, we must be dirtied. But we 3016 * might have been freed after the dirty. 3017 */ 3018 if (db->db_state == DB_UNCACHED) { 3019 /* This buffer has been freed since it was dirtied */ 3020 ASSERT(db->db.db_data == NULL); 3021 } else if (db->db_state == DB_FILL) { 3022 /* This buffer was freed and is now being re-filled */ 3023 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 3024 } else { 3025 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 3026 } 3027 DBUF_VERIFY(db); 3028 3029 DB_DNODE_ENTER(db); 3030 dn = DB_DNODE(db); 3031 3032 if (db->db_blkid == DMU_SPILL_BLKID) { 3033 mutex_enter(&dn->dn_mtx); 3034 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 3035 mutex_exit(&dn->dn_mtx); 3036 } 3037 3038 /* 3039 * If this is a bonus buffer, simply copy the bonus data into the 3040 * dnode. It will be written out when the dnode is synced (and it 3041 * will be synced, since it must have been dirty for dbuf_sync to 3042 * be called). 3043 */ 3044 if (db->db_blkid == DMU_BONUS_BLKID) { 3045 dbuf_dirty_record_t **drp; 3046 3047 ASSERT(*datap != NULL); 3048 ASSERT0(db->db_level); 3049 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 3050 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 3051 DB_DNODE_EXIT(db); 3052 3053 if (*datap != db->db.db_data) { 3054 zio_buf_free(*datap, DN_MAX_BONUSLEN); 3055 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 3056 } 3057 db->db_data_pending = NULL; 3058 drp = &db->db_last_dirty; 3059 while (*drp != dr) 3060 drp = &(*drp)->dr_next; 3061 ASSERT(dr->dr_next == NULL); 3062 ASSERT(dr->dr_dbuf == db); 3063 *drp = dr->dr_next; 3064 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3065 ASSERT(db->db_dirtycnt > 0); 3066 db->db_dirtycnt -= 1; 3067 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 3068 return; 3069 } 3070 3071 os = dn->dn_objset; 3072 3073 /* 3074 * This function may have dropped the db_mtx lock allowing a dmu_sync 3075 * operation to sneak in. As a result, we need to ensure that we 3076 * don't check the dr_override_state until we have returned from 3077 * dbuf_check_blkptr. 3078 */ 3079 dbuf_check_blkptr(dn, db); 3080 3081 /* 3082 * If this buffer is in the middle of an immediate write, 3083 * wait for the synchronous IO to complete. 3084 */ 3085 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 3086 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 3087 cv_wait(&db->db_changed, &db->db_mtx); 3088 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 3089 } 3090 3091 if (db->db_state != DB_NOFILL && 3092 dn->dn_object != DMU_META_DNODE_OBJECT && 3093 refcount_count(&db->db_holds) > 1 && 3094 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 3095 *datap == db->db_buf) { 3096 /* 3097 * If this buffer is currently "in use" (i.e., there 3098 * are active holds and db_data still references it), 3099 * then make a copy before we start the write so that 3100 * any modifications from the open txg will not leak 3101 * into this write. 3102 * 3103 * NOTE: this copy does not need to be made for 3104 * objects only modified in the syncing context (e.g. 3105 * DNONE_DNODE blocks). 3106 */ 3107 int psize = arc_buf_size(*datap); 3108 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 3109 enum zio_compress compress_type = arc_get_compression(*datap); 3110 3111 if (compress_type == ZIO_COMPRESS_OFF) { 3112 *datap = arc_alloc_buf(os->os_spa, db, type, psize); 3113 } else { 3114 ASSERT3U(type, ==, ARC_BUFC_DATA); 3115 int lsize = arc_buf_lsize(*datap); 3116 *datap = arc_alloc_compressed_buf(os->os_spa, db, 3117 psize, lsize, compress_type); 3118 } 3119 bcopy(db->db.db_data, (*datap)->b_data, psize); 3120 } 3121 db->db_data_pending = dr; 3122 3123 mutex_exit(&db->db_mtx); 3124 3125 dbuf_write(dr, *datap, tx); 3126 3127 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3128 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 3129 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 3130 DB_DNODE_EXIT(db); 3131 } else { 3132 /* 3133 * Although zio_nowait() does not "wait for an IO", it does 3134 * initiate the IO. If this is an empty write it seems plausible 3135 * that the IO could actually be completed before the nowait 3136 * returns. We need to DB_DNODE_EXIT() first in case 3137 * zio_nowait() invalidates the dbuf. 3138 */ 3139 DB_DNODE_EXIT(db); 3140 zio_nowait(dr->dr_zio); 3141 } 3142 } 3143 3144 void 3145 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 3146 { 3147 dbuf_dirty_record_t *dr; 3148 3149 while (dr = list_head(list)) { 3150 if (dr->dr_zio != NULL) { 3151 /* 3152 * If we find an already initialized zio then we 3153 * are processing the meta-dnode, and we have finished. 3154 * The dbufs for all dnodes are put back on the list 3155 * during processing, so that we can zio_wait() 3156 * these IOs after initiating all child IOs. 3157 */ 3158 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 3159 DMU_META_DNODE_OBJECT); 3160 break; 3161 } 3162 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 3163 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 3164 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 3165 } 3166 list_remove(list, dr); 3167 if (dr->dr_dbuf->db_level > 0) 3168 dbuf_sync_indirect(dr, tx); 3169 else 3170 dbuf_sync_leaf(dr, tx); 3171 } 3172 } 3173 3174 /* ARGSUSED */ 3175 static void 3176 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3177 { 3178 dmu_buf_impl_t *db = vdb; 3179 dnode_t *dn; 3180 blkptr_t *bp = zio->io_bp; 3181 blkptr_t *bp_orig = &zio->io_bp_orig; 3182 spa_t *spa = zio->io_spa; 3183 int64_t delta; 3184 uint64_t fill = 0; 3185 int i; 3186 3187 ASSERT3P(db->db_blkptr, !=, NULL); 3188 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); 3189 3190 DB_DNODE_ENTER(db); 3191 dn = DB_DNODE(db); 3192 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 3193 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 3194 zio->io_prev_space_delta = delta; 3195 3196 if (bp->blk_birth != 0) { 3197 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 3198 BP_GET_TYPE(bp) == dn->dn_type) || 3199 (db->db_blkid == DMU_SPILL_BLKID && 3200 BP_GET_TYPE(bp) == dn->dn_bonustype) || 3201 BP_IS_EMBEDDED(bp)); 3202 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 3203 } 3204 3205 mutex_enter(&db->db_mtx); 3206 3207 #ifdef ZFS_DEBUG 3208 if (db->db_blkid == DMU_SPILL_BLKID) { 3209 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3210 ASSERT(!(BP_IS_HOLE(bp)) && 3211 db->db_blkptr == &dn->dn_phys->dn_spill); 3212 } 3213 #endif 3214 3215 if (db->db_level == 0) { 3216 mutex_enter(&dn->dn_mtx); 3217 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 3218 db->db_blkid != DMU_SPILL_BLKID) 3219 dn->dn_phys->dn_maxblkid = db->db_blkid; 3220 mutex_exit(&dn->dn_mtx); 3221 3222 if (dn->dn_type == DMU_OT_DNODE) { 3223 dnode_phys_t *dnp = db->db.db_data; 3224 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 3225 i--, dnp++) { 3226 if (dnp->dn_type != DMU_OT_NONE) 3227 fill++; 3228 } 3229 } else { 3230 if (BP_IS_HOLE(bp)) { 3231 fill = 0; 3232 } else { 3233 fill = 1; 3234 } 3235 } 3236 } else { 3237 blkptr_t *ibp = db->db.db_data; 3238 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3239 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 3240 if (BP_IS_HOLE(ibp)) 3241 continue; 3242 fill += BP_GET_FILL(ibp); 3243 } 3244 } 3245 DB_DNODE_EXIT(db); 3246 3247 if (!BP_IS_EMBEDDED(bp)) 3248 bp->blk_fill = fill; 3249 3250 mutex_exit(&db->db_mtx); 3251 3252 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3253 *db->db_blkptr = *bp; 3254 rw_exit(&dn->dn_struct_rwlock); 3255 } 3256 3257 /* ARGSUSED */ 3258 /* 3259 * This function gets called just prior to running through the compression 3260 * stage of the zio pipeline. If we're an indirect block comprised of only 3261 * holes, then we want this indirect to be compressed away to a hole. In 3262 * order to do that we must zero out any information about the holes that 3263 * this indirect points to prior to before we try to compress it. 3264 */ 3265 static void 3266 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3267 { 3268 dmu_buf_impl_t *db = vdb; 3269 dnode_t *dn; 3270 blkptr_t *bp; 3271 unsigned int epbs, i; 3272 3273 ASSERT3U(db->db_level, >, 0); 3274 DB_DNODE_ENTER(db); 3275 dn = DB_DNODE(db); 3276 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3277 ASSERT3U(epbs, <, 31); 3278 3279 /* Determine if all our children are holes */ 3280 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { 3281 if (!BP_IS_HOLE(bp)) 3282 break; 3283 } 3284 3285 /* 3286 * If all the children are holes, then zero them all out so that 3287 * we may get compressed away. 3288 */ 3289 if (i == 1 << epbs) { 3290 /* 3291 * We only found holes. Grab the rwlock to prevent 3292 * anybody from reading the blocks we're about to 3293 * zero out. 3294 */ 3295 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3296 bzero(db->db.db_data, db->db.db_size); 3297 rw_exit(&dn->dn_struct_rwlock); 3298 } 3299 DB_DNODE_EXIT(db); 3300 } 3301 3302 /* 3303 * The SPA will call this callback several times for each zio - once 3304 * for every physical child i/o (zio->io_phys_children times). This 3305 * allows the DMU to monitor the progress of each logical i/o. For example, 3306 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 3307 * block. There may be a long delay before all copies/fragments are completed, 3308 * so this callback allows us to retire dirty space gradually, as the physical 3309 * i/os complete. 3310 */ 3311 /* ARGSUSED */ 3312 static void 3313 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 3314 { 3315 dmu_buf_impl_t *db = arg; 3316 objset_t *os = db->db_objset; 3317 dsl_pool_t *dp = dmu_objset_pool(os); 3318 dbuf_dirty_record_t *dr; 3319 int delta = 0; 3320 3321 dr = db->db_data_pending; 3322 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 3323 3324 /* 3325 * The callback will be called io_phys_children times. Retire one 3326 * portion of our dirty space each time we are called. Any rounding 3327 * error will be cleaned up by dsl_pool_sync()'s call to 3328 * dsl_pool_undirty_space(). 3329 */ 3330 delta = dr->dr_accounted / zio->io_phys_children; 3331 dsl_pool_undirty_space(dp, delta, zio->io_txg); 3332 } 3333 3334 /* ARGSUSED */ 3335 static void 3336 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3337 { 3338 dmu_buf_impl_t *db = vdb; 3339 blkptr_t *bp_orig = &zio->io_bp_orig; 3340 blkptr_t *bp = db->db_blkptr; 3341 objset_t *os = db->db_objset; 3342 dmu_tx_t *tx = os->os_synctx; 3343 dbuf_dirty_record_t **drp, *dr; 3344 3345 ASSERT0(zio->io_error); 3346 ASSERT(db->db_blkptr == bp); 3347 3348 /* 3349 * For nopwrites and rewrites we ensure that the bp matches our 3350 * original and bypass all the accounting. 3351 */ 3352 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3353 ASSERT(BP_EQUAL(bp, bp_orig)); 3354 } else { 3355 dsl_dataset_t *ds = os->os_dsl_dataset; 3356 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3357 dsl_dataset_block_born(ds, bp, tx); 3358 } 3359 3360 mutex_enter(&db->db_mtx); 3361 3362 DBUF_VERIFY(db); 3363 3364 drp = &db->db_last_dirty; 3365 while ((dr = *drp) != db->db_data_pending) 3366 drp = &dr->dr_next; 3367 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3368 ASSERT(dr->dr_dbuf == db); 3369 ASSERT(dr->dr_next == NULL); 3370 *drp = dr->dr_next; 3371 3372 #ifdef ZFS_DEBUG 3373 if (db->db_blkid == DMU_SPILL_BLKID) { 3374 dnode_t *dn; 3375 3376 DB_DNODE_ENTER(db); 3377 dn = DB_DNODE(db); 3378 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3379 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3380 db->db_blkptr == &dn->dn_phys->dn_spill); 3381 DB_DNODE_EXIT(db); 3382 } 3383 #endif 3384 3385 if (db->db_level == 0) { 3386 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3387 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3388 if (db->db_state != DB_NOFILL) { 3389 if (dr->dt.dl.dr_data != db->db_buf) 3390 arc_buf_destroy(dr->dt.dl.dr_data, db); 3391 } 3392 } else { 3393 dnode_t *dn; 3394 3395 DB_DNODE_ENTER(db); 3396 dn = DB_DNODE(db); 3397 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3398 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3399 if (!BP_IS_HOLE(db->db_blkptr)) { 3400 int epbs = 3401 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3402 ASSERT3U(db->db_blkid, <=, 3403 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3404 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3405 db->db.db_size); 3406 } 3407 DB_DNODE_EXIT(db); 3408 mutex_destroy(&dr->dt.di.dr_mtx); 3409 list_destroy(&dr->dt.di.dr_children); 3410 } 3411 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3412 3413 cv_broadcast(&db->db_changed); 3414 ASSERT(db->db_dirtycnt > 0); 3415 db->db_dirtycnt -= 1; 3416 db->db_data_pending = NULL; 3417 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3418 } 3419 3420 static void 3421 dbuf_write_nofill_ready(zio_t *zio) 3422 { 3423 dbuf_write_ready(zio, NULL, zio->io_private); 3424 } 3425 3426 static void 3427 dbuf_write_nofill_done(zio_t *zio) 3428 { 3429 dbuf_write_done(zio, NULL, zio->io_private); 3430 } 3431 3432 static void 3433 dbuf_write_override_ready(zio_t *zio) 3434 { 3435 dbuf_dirty_record_t *dr = zio->io_private; 3436 dmu_buf_impl_t *db = dr->dr_dbuf; 3437 3438 dbuf_write_ready(zio, NULL, db); 3439 } 3440 3441 static void 3442 dbuf_write_override_done(zio_t *zio) 3443 { 3444 dbuf_dirty_record_t *dr = zio->io_private; 3445 dmu_buf_impl_t *db = dr->dr_dbuf; 3446 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3447 3448 mutex_enter(&db->db_mtx); 3449 if (!BP_EQUAL(zio->io_bp, obp)) { 3450 if (!BP_IS_HOLE(obp)) 3451 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3452 arc_release(dr->dt.dl.dr_data, db); 3453 } 3454 mutex_exit(&db->db_mtx); 3455 3456 dbuf_write_done(zio, NULL, db); 3457 } 3458 3459 /* Issue I/O to commit a dirty buffer to disk. */ 3460 static void 3461 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3462 { 3463 dmu_buf_impl_t *db = dr->dr_dbuf; 3464 dnode_t *dn; 3465 objset_t *os; 3466 dmu_buf_impl_t *parent = db->db_parent; 3467 uint64_t txg = tx->tx_txg; 3468 zbookmark_phys_t zb; 3469 zio_prop_t zp; 3470 zio_t *zio; 3471 int wp_flag = 0; 3472 3473 ASSERT(dmu_tx_is_syncing(tx)); 3474 3475 DB_DNODE_ENTER(db); 3476 dn = DB_DNODE(db); 3477 os = dn->dn_objset; 3478 3479 if (db->db_state != DB_NOFILL) { 3480 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3481 /* 3482 * Private object buffers are released here rather 3483 * than in dbuf_dirty() since they are only modified 3484 * in the syncing context and we don't want the 3485 * overhead of making multiple copies of the data. 3486 */ 3487 if (BP_IS_HOLE(db->db_blkptr)) { 3488 arc_buf_thaw(data); 3489 } else { 3490 dbuf_release_bp(db); 3491 } 3492 } 3493 } 3494 3495 if (parent != dn->dn_dbuf) { 3496 /* Our parent is an indirect block. */ 3497 /* We have a dirty parent that has been scheduled for write. */ 3498 ASSERT(parent && parent->db_data_pending); 3499 /* Our parent's buffer is one level closer to the dnode. */ 3500 ASSERT(db->db_level == parent->db_level-1); 3501 /* 3502 * We're about to modify our parent's db_data by modifying 3503 * our block pointer, so the parent must be released. 3504 */ 3505 ASSERT(arc_released(parent->db_buf)); 3506 zio = parent->db_data_pending->dr_zio; 3507 } else { 3508 /* Our parent is the dnode itself. */ 3509 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3510 db->db_blkid != DMU_SPILL_BLKID) || 3511 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3512 if (db->db_blkid != DMU_SPILL_BLKID) 3513 ASSERT3P(db->db_blkptr, ==, 3514 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3515 zio = dn->dn_zio; 3516 } 3517 3518 ASSERT(db->db_level == 0 || data == db->db_buf); 3519 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3520 ASSERT(zio); 3521 3522 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3523 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3524 db->db.db_object, db->db_level, db->db_blkid); 3525 3526 if (db->db_blkid == DMU_SPILL_BLKID) 3527 wp_flag = WP_SPILL; 3528 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3529 3530 dmu_write_policy(os, dn, db->db_level, wp_flag, 3531 (data != NULL && arc_get_compression(data) != ZIO_COMPRESS_OFF) ? 3532 arc_get_compression(data) : ZIO_COMPRESS_INHERIT, &zp); 3533 DB_DNODE_EXIT(db); 3534 3535 /* 3536 * We copy the blkptr now (rather than when we instantiate the dirty 3537 * record), because its value can change between open context and 3538 * syncing context. We do not need to hold dn_struct_rwlock to read 3539 * db_blkptr because we are in syncing context. 3540 */ 3541 dr->dr_bp_copy = *db->db_blkptr; 3542 3543 if (db->db_level == 0 && 3544 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3545 /* 3546 * The BP for this block has been provided by open context 3547 * (by dmu_sync() or dmu_buf_write_embedded()). 3548 */ 3549 void *contents = (data != NULL) ? data->b_data : NULL; 3550 3551 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy, 3552 contents, db->db.db_size, db->db.db_size, &zp, 3553 dbuf_write_override_ready, NULL, NULL, 3554 dbuf_write_override_done, 3555 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3556 mutex_enter(&db->db_mtx); 3557 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3558 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3559 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3560 mutex_exit(&db->db_mtx); 3561 } else if (db->db_state == DB_NOFILL) { 3562 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3563 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3564 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3565 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, 3566 dbuf_write_nofill_ready, NULL, NULL, 3567 dbuf_write_nofill_done, db, 3568 ZIO_PRIORITY_ASYNC_WRITE, 3569 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3570 } else { 3571 ASSERT(arc_released(data)); 3572 3573 /* 3574 * For indirect blocks, we want to setup the children 3575 * ready callback so that we can properly handle an indirect 3576 * block that only contains holes. 3577 */ 3578 arc_done_func_t *children_ready_cb = NULL; 3579 if (db->db_level != 0) 3580 children_ready_cb = dbuf_write_children_ready; 3581 3582 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3583 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), 3584 &zp, dbuf_write_ready, children_ready_cb, 3585 dbuf_write_physdone, dbuf_write_done, db, 3586 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3587 } 3588 } 3589