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, 2018 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 #include <sys/abd.h> 50 #include <sys/vdev.h> 51 #include <sys/cityhash.h> 52 #include <sys/spa_impl.h> 53 54 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 55 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 56 57 #ifndef __lint 58 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, 59 dmu_buf_evict_func_t *evict_func_sync, 60 dmu_buf_evict_func_t *evict_func_async, 61 dmu_buf_t **clear_on_evict_dbufp); 62 #endif /* ! __lint */ 63 64 /* 65 * Global data structures and functions for the dbuf cache. 66 */ 67 static kmem_cache_t *dbuf_kmem_cache; 68 static taskq_t *dbu_evict_taskq; 69 70 static kthread_t *dbuf_cache_evict_thread; 71 static kmutex_t dbuf_evict_lock; 72 static kcondvar_t dbuf_evict_cv; 73 static boolean_t dbuf_evict_thread_exit; 74 75 /* 76 * There are two dbuf caches; each dbuf can only be in one of them at a time. 77 * 78 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands 79 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs 80 * that represent the metadata that describes filesystems/snapshots/ 81 * bookmarks/properties/etc. We only evict from this cache when we export a 82 * pool, to short-circuit as much I/O as possible for all administrative 83 * commands that need the metadata. There is no eviction policy for this 84 * cache, because we try to only include types in it which would occupy a 85 * very small amount of space per object but create a large impact on the 86 * performance of these commands. Instead, after it reaches a maximum size 87 * (which should only happen on very small memory systems with a very large 88 * number of filesystem objects), we stop taking new dbufs into the 89 * metadata cache, instead putting them in the normal dbuf cache. 90 * 91 * 2. LRU cache of dbufs. The "dbuf cache" maintains a list of dbufs that 92 * are not currently held but have been recently released. These dbufs 93 * are not eligible for arc eviction until they are aged out of the cache. 94 * Dbufs that are aged out of the cache will be immediately destroyed and 95 * become eligible for arc eviction. 96 * 97 * Dbufs are added to these caches once the last hold is released. If a dbuf is 98 * later accessed and still exists in the dbuf cache, then it will be removed 99 * from the cache and later re-added to the head of the cache. 100 * 101 * If a given dbuf meets the requirements for the metadata cache, it will go 102 * there, otherwise it will be considered for the generic LRU dbuf cache. The 103 * caches and the refcounts tracking their sizes are stored in an array indexed 104 * by those caches' matching enum values (from dbuf_cached_state_t). 105 */ 106 typedef struct dbuf_cache { 107 multilist_t *cache; 108 zfs_refcount_t size; 109 } dbuf_cache_t; 110 dbuf_cache_t dbuf_caches[DB_CACHE_MAX]; 111 112 /* Size limits for the caches */ 113 uint64_t dbuf_cache_max_bytes = 0; 114 uint64_t dbuf_metadata_cache_max_bytes = 0; 115 /* Set the default sizes of the caches to log2 fraction of arc size */ 116 int dbuf_cache_shift = 5; 117 int dbuf_metadata_cache_shift = 6; 118 119 /* 120 * For diagnostic purposes, this is incremented whenever we can't add 121 * something to the metadata cache because it's full, and instead put 122 * the data in the regular dbuf cache. 123 */ 124 uint64_t dbuf_metadata_cache_overflow; 125 126 /* 127 * The LRU dbuf cache uses a three-stage eviction policy: 128 * - A low water marker designates when the dbuf eviction thread 129 * should stop evicting from the dbuf cache. 130 * - When we reach the maximum size (aka mid water mark), we 131 * signal the eviction thread to run. 132 * - The high water mark indicates when the eviction thread 133 * is unable to keep up with the incoming load and eviction must 134 * happen in the context of the calling thread. 135 * 136 * The dbuf cache: 137 * (max size) 138 * low water mid water hi water 139 * +----------------------------------------+----------+----------+ 140 * | | | | 141 * | | | | 142 * | | | | 143 * | | | | 144 * +----------------------------------------+----------+----------+ 145 * stop signal evict 146 * evicting eviction directly 147 * thread 148 * 149 * The high and low water marks indicate the operating range for the eviction 150 * thread. The low water mark is, by default, 90% of the total size of the 151 * cache and the high water mark is at 110% (both of these percentages can be 152 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, 153 * respectively). The eviction thread will try to ensure that the cache remains 154 * within this range by waking up every second and checking if the cache is 155 * above the low water mark. The thread can also be woken up by callers adding 156 * elements into the cache if the cache is larger than the mid water (i.e max 157 * cache size). Once the eviction thread is woken up and eviction is required, 158 * it will continue evicting buffers until it's able to reduce the cache size 159 * to the low water mark. If the cache size continues to grow and hits the high 160 * water mark, then callers adding elements to the cache will begin to evict 161 * directly from the cache until the cache is no longer above the high water 162 * mark. 163 */ 164 165 /* 166 * The percentage above and below the maximum cache size. 167 */ 168 uint_t dbuf_cache_hiwater_pct = 10; 169 uint_t dbuf_cache_lowater_pct = 10; 170 171 /* ARGSUSED */ 172 static int 173 dbuf_cons(void *vdb, void *unused, int kmflag) 174 { 175 dmu_buf_impl_t *db = vdb; 176 bzero(db, sizeof (dmu_buf_impl_t)); 177 178 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 179 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 180 multilist_link_init(&db->db_cache_link); 181 zfs_refcount_create(&db->db_holds); 182 183 return (0); 184 } 185 186 /* ARGSUSED */ 187 static void 188 dbuf_dest(void *vdb, void *unused) 189 { 190 dmu_buf_impl_t *db = vdb; 191 mutex_destroy(&db->db_mtx); 192 cv_destroy(&db->db_changed); 193 ASSERT(!multilist_link_active(&db->db_cache_link)); 194 zfs_refcount_destroy(&db->db_holds); 195 } 196 197 /* 198 * dbuf hash table routines 199 */ 200 static dbuf_hash_table_t dbuf_hash_table; 201 202 static uint64_t dbuf_hash_count; 203 204 /* 205 * We use Cityhash for this. It's fast, and has good hash properties without 206 * requiring any large static buffers. 207 */ 208 static uint64_t 209 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 210 { 211 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid)); 212 } 213 214 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 215 ((dbuf)->db.db_object == (obj) && \ 216 (dbuf)->db_objset == (os) && \ 217 (dbuf)->db_level == (level) && \ 218 (dbuf)->db_blkid == (blkid)) 219 220 dmu_buf_impl_t * 221 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 222 { 223 dbuf_hash_table_t *h = &dbuf_hash_table; 224 uint64_t hv = dbuf_hash(os, obj, level, blkid); 225 uint64_t idx = hv & h->hash_table_mask; 226 dmu_buf_impl_t *db; 227 228 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 229 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 230 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 231 mutex_enter(&db->db_mtx); 232 if (db->db_state != DB_EVICTING) { 233 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 234 return (db); 235 } 236 mutex_exit(&db->db_mtx); 237 } 238 } 239 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 240 return (NULL); 241 } 242 243 static dmu_buf_impl_t * 244 dbuf_find_bonus(objset_t *os, uint64_t object) 245 { 246 dnode_t *dn; 247 dmu_buf_impl_t *db = NULL; 248 249 if (dnode_hold(os, object, FTAG, &dn) == 0) { 250 rw_enter(&dn->dn_struct_rwlock, RW_READER); 251 if (dn->dn_bonus != NULL) { 252 db = dn->dn_bonus; 253 mutex_enter(&db->db_mtx); 254 } 255 rw_exit(&dn->dn_struct_rwlock); 256 dnode_rele(dn, FTAG); 257 } 258 return (db); 259 } 260 261 /* 262 * Insert an entry into the hash table. If there is already an element 263 * equal to elem in the hash table, then the already existing element 264 * will be returned and the new element will not be inserted. 265 * Otherwise returns NULL. 266 */ 267 static dmu_buf_impl_t * 268 dbuf_hash_insert(dmu_buf_impl_t *db) 269 { 270 dbuf_hash_table_t *h = &dbuf_hash_table; 271 objset_t *os = db->db_objset; 272 uint64_t obj = db->db.db_object; 273 int level = db->db_level; 274 uint64_t blkid = db->db_blkid; 275 uint64_t hv = dbuf_hash(os, obj, level, blkid); 276 uint64_t idx = hv & h->hash_table_mask; 277 dmu_buf_impl_t *dbf; 278 279 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 280 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 281 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 282 mutex_enter(&dbf->db_mtx); 283 if (dbf->db_state != DB_EVICTING) { 284 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 285 return (dbf); 286 } 287 mutex_exit(&dbf->db_mtx); 288 } 289 } 290 291 mutex_enter(&db->db_mtx); 292 db->db_hash_next = h->hash_table[idx]; 293 h->hash_table[idx] = db; 294 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 295 atomic_inc_64(&dbuf_hash_count); 296 297 return (NULL); 298 } 299 300 /* 301 * Remove an entry from the hash table. It must be in the EVICTING state. 302 */ 303 static void 304 dbuf_hash_remove(dmu_buf_impl_t *db) 305 { 306 dbuf_hash_table_t *h = &dbuf_hash_table; 307 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object, 308 db->db_level, db->db_blkid); 309 uint64_t idx = hv & h->hash_table_mask; 310 dmu_buf_impl_t *dbf, **dbp; 311 312 /* 313 * We mustn't hold db_mtx to maintain lock ordering: 314 * DBUF_HASH_MUTEX > db_mtx. 315 */ 316 ASSERT(zfs_refcount_is_zero(&db->db_holds)); 317 ASSERT(db->db_state == DB_EVICTING); 318 ASSERT(!MUTEX_HELD(&db->db_mtx)); 319 320 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 321 dbp = &h->hash_table[idx]; 322 while ((dbf = *dbp) != db) { 323 dbp = &dbf->db_hash_next; 324 ASSERT(dbf != NULL); 325 } 326 *dbp = db->db_hash_next; 327 db->db_hash_next = NULL; 328 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 329 atomic_dec_64(&dbuf_hash_count); 330 } 331 332 typedef enum { 333 DBVU_EVICTING, 334 DBVU_NOT_EVICTING 335 } dbvu_verify_type_t; 336 337 static void 338 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 339 { 340 #ifdef ZFS_DEBUG 341 int64_t holds; 342 343 if (db->db_user == NULL) 344 return; 345 346 /* Only data blocks support the attachment of user data. */ 347 ASSERT(db->db_level == 0); 348 349 /* Clients must resolve a dbuf before attaching user data. */ 350 ASSERT(db->db.db_data != NULL); 351 ASSERT3U(db->db_state, ==, DB_CACHED); 352 353 holds = zfs_refcount_count(&db->db_holds); 354 if (verify_type == DBVU_EVICTING) { 355 /* 356 * Immediate eviction occurs when holds == dirtycnt. 357 * For normal eviction buffers, holds is zero on 358 * eviction, except when dbuf_fix_old_data() calls 359 * dbuf_clear_data(). However, the hold count can grow 360 * during eviction even though db_mtx is held (see 361 * dmu_bonus_hold() for an example), so we can only 362 * test the generic invariant that holds >= dirtycnt. 363 */ 364 ASSERT3U(holds, >=, db->db_dirtycnt); 365 } else { 366 if (db->db_user_immediate_evict == TRUE) 367 ASSERT3U(holds, >=, db->db_dirtycnt); 368 else 369 ASSERT3U(holds, >, 0); 370 } 371 #endif 372 } 373 374 static void 375 dbuf_evict_user(dmu_buf_impl_t *db) 376 { 377 dmu_buf_user_t *dbu = db->db_user; 378 379 ASSERT(MUTEX_HELD(&db->db_mtx)); 380 381 if (dbu == NULL) 382 return; 383 384 dbuf_verify_user(db, DBVU_EVICTING); 385 db->db_user = NULL; 386 387 #ifdef ZFS_DEBUG 388 if (dbu->dbu_clear_on_evict_dbufp != NULL) 389 *dbu->dbu_clear_on_evict_dbufp = NULL; 390 #endif 391 392 /* 393 * There are two eviction callbacks - one that we call synchronously 394 * and one that we invoke via a taskq. The async one is useful for 395 * avoiding lock order reversals and limiting stack depth. 396 * 397 * Note that if we have a sync callback but no async callback, 398 * it's likely that the sync callback will free the structure 399 * containing the dbu. In that case we need to take care to not 400 * dereference dbu after calling the sync evict func. 401 */ 402 boolean_t has_async = (dbu->dbu_evict_func_async != NULL); 403 404 if (dbu->dbu_evict_func_sync != NULL) 405 dbu->dbu_evict_func_sync(dbu); 406 407 if (has_async) { 408 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, 409 dbu, 0, &dbu->dbu_tqent); 410 } 411 } 412 413 boolean_t 414 dbuf_is_metadata(dmu_buf_impl_t *db) 415 { 416 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { 417 return (B_TRUE); 418 } else { 419 boolean_t is_metadata; 420 421 DB_DNODE_ENTER(db); 422 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 423 DB_DNODE_EXIT(db); 424 425 return (is_metadata); 426 } 427 } 428 429 /* 430 * This returns whether this dbuf should be stored in the metadata cache, which 431 * is based on whether it's from one of the dnode types that store data related 432 * to traversing dataset hierarchies. 433 */ 434 static boolean_t 435 dbuf_include_in_metadata_cache(dmu_buf_impl_t *db) 436 { 437 DB_DNODE_ENTER(db); 438 dmu_object_type_t type = DB_DNODE(db)->dn_type; 439 DB_DNODE_EXIT(db); 440 441 /* Check if this dbuf is one of the types we care about */ 442 if (DMU_OT_IS_METADATA_CACHED(type)) { 443 /* If we hit this, then we set something up wrong in dmu_ot */ 444 ASSERT(DMU_OT_IS_METADATA(type)); 445 446 /* 447 * Sanity check for small-memory systems: don't allocate too 448 * much memory for this purpose. 449 */ 450 if (zfs_refcount_count( 451 &dbuf_caches[DB_DBUF_METADATA_CACHE].size) > 452 dbuf_metadata_cache_max_bytes) { 453 dbuf_metadata_cache_overflow++; 454 DTRACE_PROBE1(dbuf__metadata__cache__overflow, 455 dmu_buf_impl_t *, db); 456 return (B_FALSE); 457 } 458 459 return (B_TRUE); 460 } 461 462 return (B_FALSE); 463 } 464 465 /* 466 * This function *must* return indices evenly distributed between all 467 * sublists of the multilist. This is needed due to how the dbuf eviction 468 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly 469 * distributed between all sublists and uses this assumption when 470 * deciding which sublist to evict from and how much to evict from it. 471 */ 472 unsigned int 473 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) 474 { 475 dmu_buf_impl_t *db = obj; 476 477 /* 478 * The assumption here, is the hash value for a given 479 * dmu_buf_impl_t will remain constant throughout it's lifetime 480 * (i.e. it's objset, object, level and blkid fields don't change). 481 * Thus, we don't need to store the dbuf's sublist index 482 * on insertion, as this index can be recalculated on removal. 483 * 484 * Also, the low order bits of the hash value are thought to be 485 * distributed evenly. Otherwise, in the case that the multilist 486 * has a power of two number of sublists, each sublists' usage 487 * would not be evenly distributed. 488 */ 489 return (dbuf_hash(db->db_objset, db->db.db_object, 490 db->db_level, db->db_blkid) % 491 multilist_get_num_sublists(ml)); 492 } 493 494 static inline boolean_t 495 dbuf_cache_above_hiwater(void) 496 { 497 uint64_t dbuf_cache_hiwater_bytes = 498 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100; 499 500 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > 501 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes); 502 } 503 504 static inline boolean_t 505 dbuf_cache_above_lowater(void) 506 { 507 uint64_t dbuf_cache_lowater_bytes = 508 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100; 509 510 return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > 511 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes); 512 } 513 514 /* 515 * Evict the oldest eligible dbuf from the dbuf cache. 516 */ 517 static void 518 dbuf_evict_one(void) 519 { 520 int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache); 521 multilist_sublist_t *mls = multilist_sublist_lock( 522 dbuf_caches[DB_DBUF_CACHE].cache, idx); 523 524 ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); 525 526 dmu_buf_impl_t *db = multilist_sublist_tail(mls); 527 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { 528 db = multilist_sublist_prev(mls, db); 529 } 530 531 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, 532 multilist_sublist_t *, mls); 533 534 if (db != NULL) { 535 multilist_sublist_remove(mls, db); 536 multilist_sublist_unlock(mls); 537 (void) zfs_refcount_remove_many( 538 &dbuf_caches[DB_DBUF_CACHE].size, 539 db->db.db_size, db); 540 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE); 541 db->db_caching_status = DB_NO_CACHE; 542 dbuf_destroy(db); 543 } else { 544 multilist_sublist_unlock(mls); 545 } 546 } 547 548 /* 549 * The dbuf evict thread is responsible for aging out dbufs from the 550 * cache. Once the cache has reached it's maximum size, dbufs are removed 551 * and destroyed. The eviction thread will continue running until the size 552 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged 553 * out of the cache it is destroyed and becomes eligible for arc eviction. 554 */ 555 /* ARGSUSED */ 556 static void 557 dbuf_evict_thread(void *unused) 558 { 559 callb_cpr_t cpr; 560 561 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); 562 563 mutex_enter(&dbuf_evict_lock); 564 while (!dbuf_evict_thread_exit) { 565 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 566 CALLB_CPR_SAFE_BEGIN(&cpr); 567 (void) cv_timedwait_hires(&dbuf_evict_cv, 568 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 569 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); 570 } 571 mutex_exit(&dbuf_evict_lock); 572 573 /* 574 * Keep evicting as long as we're above the low water mark 575 * for the cache. We do this without holding the locks to 576 * minimize lock contention. 577 */ 578 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 579 dbuf_evict_one(); 580 } 581 582 mutex_enter(&dbuf_evict_lock); 583 } 584 585 dbuf_evict_thread_exit = B_FALSE; 586 cv_broadcast(&dbuf_evict_cv); 587 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ 588 thread_exit(); 589 } 590 591 /* 592 * Wake up the dbuf eviction thread if the dbuf cache is at its max size. 593 * If the dbuf cache is at its high water mark, then evict a dbuf from the 594 * dbuf cache using the callers context. 595 */ 596 static void 597 dbuf_evict_notify(void) 598 { 599 /* 600 * We check if we should evict without holding the dbuf_evict_lock, 601 * because it's OK to occasionally make the wrong decision here, 602 * and grabbing the lock results in massive lock contention. 603 */ 604 if (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > 605 dbuf_cache_max_bytes) { 606 if (dbuf_cache_above_hiwater()) 607 dbuf_evict_one(); 608 cv_signal(&dbuf_evict_cv); 609 } 610 } 611 612 void 613 dbuf_init(void) 614 { 615 uint64_t hsize = 1ULL << 16; 616 dbuf_hash_table_t *h = &dbuf_hash_table; 617 int i; 618 619 /* 620 * The hash table is big enough to fill all of physical memory 621 * with an average 4K block size. The table will take up 622 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 623 */ 624 while (hsize * 4096 < physmem * PAGESIZE) 625 hsize <<= 1; 626 627 retry: 628 h->hash_table_mask = hsize - 1; 629 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 630 if (h->hash_table == NULL) { 631 /* XXX - we should really return an error instead of assert */ 632 ASSERT(hsize > (1ULL << 10)); 633 hsize >>= 1; 634 goto retry; 635 } 636 637 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", 638 sizeof (dmu_buf_impl_t), 639 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 640 641 for (i = 0; i < DBUF_MUTEXES; i++) 642 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 643 644 /* 645 * Setup the parameters for the dbuf caches. We set the sizes of the 646 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default) 647 * of the size of the ARC, respectively. If the values are set in 648 * /etc/system and they're not greater than the size of the ARC, then 649 * we honor that value. 650 */ 651 if (dbuf_cache_max_bytes == 0 || 652 dbuf_cache_max_bytes >= arc_max_bytes()) { 653 dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift; 654 } 655 if (dbuf_metadata_cache_max_bytes == 0 || 656 dbuf_metadata_cache_max_bytes >= arc_max_bytes()) { 657 dbuf_metadata_cache_max_bytes = 658 arc_max_bytes() >> dbuf_metadata_cache_shift; 659 } 660 661 /* 662 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 663 * configuration is not required. 664 */ 665 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 666 667 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { 668 dbuf_caches[dcs].cache = 669 multilist_create(sizeof (dmu_buf_impl_t), 670 offsetof(dmu_buf_impl_t, db_cache_link), 671 dbuf_cache_multilist_index_func); 672 zfs_refcount_create(&dbuf_caches[dcs].size); 673 } 674 675 dbuf_evict_thread_exit = B_FALSE; 676 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); 677 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); 678 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, 679 NULL, 0, &p0, TS_RUN, minclsyspri); 680 } 681 682 void 683 dbuf_fini(void) 684 { 685 dbuf_hash_table_t *h = &dbuf_hash_table; 686 int i; 687 688 for (i = 0; i < DBUF_MUTEXES; i++) 689 mutex_destroy(&h->hash_mutexes[i]); 690 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 691 kmem_cache_destroy(dbuf_kmem_cache); 692 taskq_destroy(dbu_evict_taskq); 693 694 mutex_enter(&dbuf_evict_lock); 695 dbuf_evict_thread_exit = B_TRUE; 696 while (dbuf_evict_thread_exit) { 697 cv_signal(&dbuf_evict_cv); 698 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); 699 } 700 mutex_exit(&dbuf_evict_lock); 701 702 mutex_destroy(&dbuf_evict_lock); 703 cv_destroy(&dbuf_evict_cv); 704 705 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { 706 zfs_refcount_destroy(&dbuf_caches[dcs].size); 707 multilist_destroy(dbuf_caches[dcs].cache); 708 } 709 } 710 711 /* 712 * Other stuff. 713 */ 714 715 #ifdef ZFS_DEBUG 716 static void 717 dbuf_verify(dmu_buf_impl_t *db) 718 { 719 dnode_t *dn; 720 dbuf_dirty_record_t *dr; 721 722 ASSERT(MUTEX_HELD(&db->db_mtx)); 723 724 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 725 return; 726 727 ASSERT(db->db_objset != NULL); 728 DB_DNODE_ENTER(db); 729 dn = DB_DNODE(db); 730 if (dn == NULL) { 731 ASSERT(db->db_parent == NULL); 732 ASSERT(db->db_blkptr == NULL); 733 } else { 734 ASSERT3U(db->db.db_object, ==, dn->dn_object); 735 ASSERT3P(db->db_objset, ==, dn->dn_objset); 736 ASSERT3U(db->db_level, <, dn->dn_nlevels); 737 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 738 db->db_blkid == DMU_SPILL_BLKID || 739 !avl_is_empty(&dn->dn_dbufs)); 740 } 741 if (db->db_blkid == DMU_BONUS_BLKID) { 742 ASSERT(dn != NULL); 743 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 744 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 745 } else if (db->db_blkid == DMU_SPILL_BLKID) { 746 ASSERT(dn != NULL); 747 ASSERT0(db->db.db_offset); 748 } else { 749 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 750 } 751 752 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 753 ASSERT(dr->dr_dbuf == db); 754 755 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 756 ASSERT(dr->dr_dbuf == db); 757 758 /* 759 * We can't assert that db_size matches dn_datablksz because it 760 * can be momentarily different when another thread is doing 761 * dnode_set_blksz(). 762 */ 763 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 764 dr = db->db_data_pending; 765 /* 766 * It should only be modified in syncing context, so 767 * make sure we only have one copy of the data. 768 */ 769 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 770 } 771 772 /* verify db->db_blkptr */ 773 if (db->db_blkptr) { 774 if (db->db_parent == dn->dn_dbuf) { 775 /* db is pointed to by the dnode */ 776 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 777 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 778 ASSERT(db->db_parent == NULL); 779 else 780 ASSERT(db->db_parent != NULL); 781 if (db->db_blkid != DMU_SPILL_BLKID) 782 ASSERT3P(db->db_blkptr, ==, 783 &dn->dn_phys->dn_blkptr[db->db_blkid]); 784 } else { 785 /* db is pointed to by an indirect block */ 786 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 787 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 788 ASSERT3U(db->db_parent->db.db_object, ==, 789 db->db.db_object); 790 /* 791 * dnode_grow_indblksz() can make this fail if we don't 792 * have the struct_rwlock. XXX indblksz no longer 793 * grows. safe to do this now? 794 */ 795 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 796 ASSERT3P(db->db_blkptr, ==, 797 ((blkptr_t *)db->db_parent->db.db_data + 798 db->db_blkid % epb)); 799 } 800 } 801 } 802 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 803 (db->db_buf == NULL || db->db_buf->b_data) && 804 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 805 db->db_state != DB_FILL && !dn->dn_free_txg) { 806 /* 807 * If the blkptr isn't set but they have nonzero data, 808 * it had better be dirty, otherwise we'll lose that 809 * data when we evict this buffer. 810 * 811 * There is an exception to this rule for indirect blocks; in 812 * this case, if the indirect block is a hole, we fill in a few 813 * fields on each of the child blocks (importantly, birth time) 814 * to prevent hole birth times from being lost when you 815 * partially fill in a hole. 816 */ 817 if (db->db_dirtycnt == 0) { 818 if (db->db_level == 0) { 819 uint64_t *buf = db->db.db_data; 820 int i; 821 822 for (i = 0; i < db->db.db_size >> 3; i++) { 823 ASSERT(buf[i] == 0); 824 } 825 } else { 826 blkptr_t *bps = db->db.db_data; 827 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, 828 db->db.db_size); 829 /* 830 * We want to verify that all the blkptrs in the 831 * indirect block are holes, but we may have 832 * automatically set up a few fields for them. 833 * We iterate through each blkptr and verify 834 * they only have those fields set. 835 */ 836 for (int i = 0; 837 i < db->db.db_size / sizeof (blkptr_t); 838 i++) { 839 blkptr_t *bp = &bps[i]; 840 ASSERT(ZIO_CHECKSUM_IS_ZERO( 841 &bp->blk_cksum)); 842 ASSERT( 843 DVA_IS_EMPTY(&bp->blk_dva[0]) && 844 DVA_IS_EMPTY(&bp->blk_dva[1]) && 845 DVA_IS_EMPTY(&bp->blk_dva[2])); 846 ASSERT0(bp->blk_fill); 847 ASSERT0(bp->blk_pad[0]); 848 ASSERT0(bp->blk_pad[1]); 849 ASSERT(!BP_IS_EMBEDDED(bp)); 850 ASSERT(BP_IS_HOLE(bp)); 851 ASSERT0(bp->blk_phys_birth); 852 } 853 } 854 } 855 } 856 DB_DNODE_EXIT(db); 857 } 858 #endif 859 860 static void 861 dbuf_clear_data(dmu_buf_impl_t *db) 862 { 863 ASSERT(MUTEX_HELD(&db->db_mtx)); 864 dbuf_evict_user(db); 865 ASSERT3P(db->db_buf, ==, NULL); 866 db->db.db_data = NULL; 867 if (db->db_state != DB_NOFILL) 868 db->db_state = DB_UNCACHED; 869 } 870 871 static void 872 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 873 { 874 ASSERT(MUTEX_HELD(&db->db_mtx)); 875 ASSERT(buf != NULL); 876 877 db->db_buf = buf; 878 ASSERT(buf->b_data != NULL); 879 db->db.db_data = buf->b_data; 880 } 881 882 /* 883 * Loan out an arc_buf for read. Return the loaned arc_buf. 884 */ 885 arc_buf_t * 886 dbuf_loan_arcbuf(dmu_buf_impl_t *db) 887 { 888 arc_buf_t *abuf; 889 890 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 891 mutex_enter(&db->db_mtx); 892 if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) { 893 int blksz = db->db.db_size; 894 spa_t *spa = db->db_objset->os_spa; 895 896 mutex_exit(&db->db_mtx); 897 abuf = arc_loan_buf(spa, B_FALSE, blksz); 898 bcopy(db->db.db_data, abuf->b_data, blksz); 899 } else { 900 abuf = db->db_buf; 901 arc_loan_inuse_buf(abuf, db); 902 db->db_buf = NULL; 903 dbuf_clear_data(db); 904 mutex_exit(&db->db_mtx); 905 } 906 return (abuf); 907 } 908 909 /* 910 * Calculate which level n block references the data at the level 0 offset 911 * provided. 912 */ 913 uint64_t 914 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 915 { 916 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 917 /* 918 * The level n blkid is equal to the level 0 blkid divided by 919 * the number of level 0s in a level n block. 920 * 921 * The level 0 blkid is offset >> datablkshift = 922 * offset / 2^datablkshift. 923 * 924 * The number of level 0s in a level n is the number of block 925 * pointers in an indirect block, raised to the power of level. 926 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 927 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 928 * 929 * Thus, the level n blkid is: offset / 930 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 931 * = offset / 2^(datablkshift + level * 932 * (indblkshift - SPA_BLKPTRSHIFT)) 933 * = offset >> (datablkshift + level * 934 * (indblkshift - SPA_BLKPTRSHIFT)) 935 */ 936 return (offset >> (dn->dn_datablkshift + level * 937 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 938 } else { 939 ASSERT3U(offset, <, dn->dn_datablksz); 940 return (0); 941 } 942 } 943 944 /* ARGSUSED */ 945 static void 946 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, 947 arc_buf_t *buf, void *vdb) 948 { 949 dmu_buf_impl_t *db = vdb; 950 951 mutex_enter(&db->db_mtx); 952 ASSERT3U(db->db_state, ==, DB_READ); 953 /* 954 * All reads are synchronous, so we must have a hold on the dbuf 955 */ 956 ASSERT(zfs_refcount_count(&db->db_holds) > 0); 957 ASSERT(db->db_buf == NULL); 958 ASSERT(db->db.db_data == NULL); 959 if (buf == NULL) { 960 /* i/o error */ 961 ASSERT(zio == NULL || zio->io_error != 0); 962 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 963 ASSERT3P(db->db_buf, ==, NULL); 964 db->db_state = DB_UNCACHED; 965 } else if (db->db_level == 0 && db->db_freed_in_flight) { 966 /* we were freed in flight; disregard any error */ 967 ASSERT(zio == NULL || zio->io_error == 0); 968 if (buf == NULL) { 969 buf = arc_alloc_buf(db->db_objset->os_spa, 970 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size); 971 } 972 arc_release(buf, db); 973 bzero(buf->b_data, db->db.db_size); 974 arc_buf_freeze(buf); 975 db->db_freed_in_flight = FALSE; 976 dbuf_set_data(db, buf); 977 db->db_state = DB_CACHED; 978 } else if (buf != NULL) { 979 /* success */ 980 ASSERT(zio == NULL || zio->io_error == 0); 981 dbuf_set_data(db, buf); 982 db->db_state = DB_CACHED; 983 } 984 cv_broadcast(&db->db_changed); 985 dbuf_rele_and_unlock(db, NULL, B_FALSE); 986 } 987 988 989 /* 990 * This function ensures that, when doing a decrypting read of a block, 991 * we make sure we have decrypted the dnode associated with it. We must do 992 * this so that we ensure we are fully authenticating the checksum-of-MACs 993 * tree from the root of the objset down to this block. Indirect blocks are 994 * always verified against their secure checksum-of-MACs assuming that the 995 * dnode containing them is correct. Now that we are doing a decrypting read, 996 * we can be sure that the key is loaded and verify that assumption. This is 997 * especially important considering that we always read encrypted dnode 998 * blocks as raw data (without verifying their MACs) to start, and 999 * decrypt / authenticate them when we need to read an encrypted bonus buffer. 1000 */ 1001 static int 1002 dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags) 1003 { 1004 int err = 0; 1005 objset_t *os = db->db_objset; 1006 arc_buf_t *dnode_abuf; 1007 dnode_t *dn; 1008 zbookmark_phys_t zb; 1009 1010 ASSERT(MUTEX_HELD(&db->db_mtx)); 1011 1012 if (!os->os_encrypted || os->os_raw_receive || 1013 (flags & DB_RF_NO_DECRYPT) != 0) 1014 return (0); 1015 1016 DB_DNODE_ENTER(db); 1017 dn = DB_DNODE(db); 1018 dnode_abuf = (dn->dn_dbuf != NULL) ? dn->dn_dbuf->db_buf : NULL; 1019 1020 if (dnode_abuf == NULL || !arc_is_encrypted(dnode_abuf)) { 1021 DB_DNODE_EXIT(db); 1022 return (0); 1023 } 1024 1025 SET_BOOKMARK(&zb, dmu_objset_id(os), 1026 DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid); 1027 err = arc_untransform(dnode_abuf, os->os_spa, &zb, B_TRUE); 1028 1029 /* 1030 * An error code of EACCES tells us that the key is still not 1031 * available. This is ok if we are only reading authenticated 1032 * (and therefore non-encrypted) blocks. 1033 */ 1034 if (err == EACCES && ((db->db_blkid != DMU_BONUS_BLKID && 1035 !DMU_OT_IS_ENCRYPTED(dn->dn_type)) || 1036 (db->db_blkid == DMU_BONUS_BLKID && 1037 !DMU_OT_IS_ENCRYPTED(dn->dn_bonustype)))) 1038 err = 0; 1039 1040 1041 DB_DNODE_EXIT(db); 1042 1043 return (err); 1044 } 1045 1046 static int 1047 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 1048 { 1049 dnode_t *dn; 1050 zbookmark_phys_t zb; 1051 arc_flags_t aflags = ARC_FLAG_NOWAIT; 1052 int err, zio_flags = 0; 1053 1054 DB_DNODE_ENTER(db); 1055 dn = DB_DNODE(db); 1056 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 1057 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 1058 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1059 ASSERT(MUTEX_HELD(&db->db_mtx)); 1060 ASSERT(db->db_state == DB_UNCACHED); 1061 ASSERT(db->db_buf == NULL); 1062 1063 if (db->db_blkid == DMU_BONUS_BLKID) { 1064 /* 1065 * The bonus length stored in the dnode may be less than 1066 * the maximum available space in the bonus buffer. 1067 */ 1068 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 1069 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); 1070 1071 /* if the underlying dnode block is encrypted, decrypt it */ 1072 err = dbuf_read_verify_dnode_crypt(db, flags); 1073 if (err != 0) { 1074 DB_DNODE_EXIT(db); 1075 mutex_exit(&db->db_mtx); 1076 return (err); 1077 } 1078 1079 ASSERT3U(bonuslen, <=, db->db.db_size); 1080 db->db.db_data = zio_buf_alloc(max_bonuslen); 1081 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS); 1082 if (bonuslen < max_bonuslen) 1083 bzero(db->db.db_data, max_bonuslen); 1084 if (bonuslen) 1085 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 1086 DB_DNODE_EXIT(db); 1087 db->db_state = DB_CACHED; 1088 mutex_exit(&db->db_mtx); 1089 return (0); 1090 } 1091 1092 /* 1093 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 1094 * processes the delete record and clears the bp while we are waiting 1095 * for the dn_mtx (resulting in a "no" from block_freed). 1096 */ 1097 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 1098 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 1099 BP_IS_HOLE(db->db_blkptr)))) { 1100 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1101 1102 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, 1103 db->db.db_size)); 1104 bzero(db->db.db_data, db->db.db_size); 1105 1106 if (db->db_blkptr != NULL && db->db_level > 0 && 1107 BP_IS_HOLE(db->db_blkptr) && 1108 db->db_blkptr->blk_birth != 0) { 1109 blkptr_t *bps = db->db.db_data; 1110 for (int i = 0; i < ((1 << 1111 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); 1112 i++) { 1113 blkptr_t *bp = &bps[i]; 1114 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 1115 1 << dn->dn_indblkshift); 1116 BP_SET_LSIZE(bp, 1117 BP_GET_LEVEL(db->db_blkptr) == 1 ? 1118 dn->dn_datablksz : 1119 BP_GET_LSIZE(db->db_blkptr)); 1120 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); 1121 BP_SET_LEVEL(bp, 1122 BP_GET_LEVEL(db->db_blkptr) - 1); 1123 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); 1124 } 1125 } 1126 DB_DNODE_EXIT(db); 1127 db->db_state = DB_CACHED; 1128 mutex_exit(&db->db_mtx); 1129 return (0); 1130 } 1131 1132 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), 1133 db->db.db_object, db->db_level, db->db_blkid); 1134 1135 /* 1136 * All bps of an encrypted os should have the encryption bit set. 1137 * If this is not true it indicates tampering and we report an error. 1138 */ 1139 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) { 1140 spa_log_error(db->db_objset->os_spa, &zb); 1141 zfs_panic_recover("unencrypted block in encrypted " 1142 "object set %llu", dmu_objset_id(db->db_objset)); 1143 DB_DNODE_EXIT(db); 1144 mutex_exit(&db->db_mtx); 1145 return (SET_ERROR(EIO)); 1146 } 1147 1148 err = dbuf_read_verify_dnode_crypt(db, flags); 1149 if (err != 0) { 1150 DB_DNODE_EXIT(db); 1151 mutex_exit(&db->db_mtx); 1152 return (err); 1153 } 1154 1155 DB_DNODE_EXIT(db); 1156 1157 db->db_state = DB_READ; 1158 mutex_exit(&db->db_mtx); 1159 1160 if (DBUF_IS_L2CACHEABLE(db)) 1161 aflags |= ARC_FLAG_L2CACHE; 1162 1163 dbuf_add_ref(db, NULL); 1164 1165 zio_flags = (flags & DB_RF_CANFAIL) ? 1166 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED; 1167 1168 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr)) 1169 zio_flags |= ZIO_FLAG_RAW; 1170 1171 err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 1172 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags, 1173 &aflags, &zb); 1174 1175 return (err); 1176 } 1177 1178 /* 1179 * This is our just-in-time copy function. It makes a copy of buffers that 1180 * have been modified in a previous transaction group before we access them in 1181 * the current active group. 1182 * 1183 * This function is used in three places: when we are dirtying a buffer for the 1184 * first time in a txg, when we are freeing a range in a dnode that includes 1185 * this buffer, and when we are accessing a buffer which was received compressed 1186 * and later referenced in a WRITE_BYREF record. 1187 * 1188 * Note that when we are called from dbuf_free_range() we do not put a hold on 1189 * the buffer, we just traverse the active dbuf list for the dnode. 1190 */ 1191 static void 1192 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 1193 { 1194 dbuf_dirty_record_t *dr = db->db_last_dirty; 1195 1196 ASSERT(MUTEX_HELD(&db->db_mtx)); 1197 ASSERT(db->db.db_data != NULL); 1198 ASSERT(db->db_level == 0); 1199 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 1200 1201 if (dr == NULL || 1202 (dr->dt.dl.dr_data != 1203 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 1204 return; 1205 1206 /* 1207 * If the last dirty record for this dbuf has not yet synced 1208 * and its referencing the dbuf data, either: 1209 * reset the reference to point to a new copy, 1210 * or (if there a no active holders) 1211 * just null out the current db_data pointer. 1212 */ 1213 ASSERT3U(dr->dr_txg, >=, txg - 2); 1214 if (db->db_blkid == DMU_BONUS_BLKID) { 1215 /* Note that the data bufs here are zio_bufs */ 1216 dnode_t *dn = DB_DNODE(db); 1217 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); 1218 dr->dt.dl.dr_data = zio_buf_alloc(bonuslen); 1219 arc_space_consume(bonuslen, ARC_SPACE_BONUS); 1220 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen); 1221 } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) { 1222 dnode_t *dn = DB_DNODE(db); 1223 int size = arc_buf_size(db->db_buf); 1224 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1225 spa_t *spa = db->db_objset->os_spa; 1226 enum zio_compress compress_type = 1227 arc_get_compression(db->db_buf); 1228 1229 if (arc_is_encrypted(db->db_buf)) { 1230 boolean_t byteorder; 1231 uint8_t salt[ZIO_DATA_SALT_LEN]; 1232 uint8_t iv[ZIO_DATA_IV_LEN]; 1233 uint8_t mac[ZIO_DATA_MAC_LEN]; 1234 1235 arc_get_raw_params(db->db_buf, &byteorder, salt, 1236 iv, mac); 1237 dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db, 1238 dmu_objset_id(dn->dn_objset), byteorder, salt, iv, 1239 mac, dn->dn_type, size, arc_buf_lsize(db->db_buf), 1240 compress_type); 1241 } else if (compress_type != ZIO_COMPRESS_OFF) { 1242 ASSERT3U(type, ==, ARC_BUFC_DATA); 1243 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, 1244 size, arc_buf_lsize(db->db_buf), compress_type); 1245 } else { 1246 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); 1247 } 1248 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 1249 } else { 1250 db->db_buf = NULL; 1251 dbuf_clear_data(db); 1252 } 1253 } 1254 1255 int 1256 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 1257 { 1258 int err = 0; 1259 boolean_t prefetch; 1260 dnode_t *dn; 1261 1262 /* 1263 * We don't have to hold the mutex to check db_state because it 1264 * can't be freed while we have a hold on the buffer. 1265 */ 1266 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 1267 1268 if (db->db_state == DB_NOFILL) 1269 return (SET_ERROR(EIO)); 1270 1271 DB_DNODE_ENTER(db); 1272 dn = DB_DNODE(db); 1273 if ((flags & DB_RF_HAVESTRUCT) == 0) 1274 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1275 1276 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1277 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 1278 DBUF_IS_CACHEABLE(db); 1279 1280 mutex_enter(&db->db_mtx); 1281 if (db->db_state == DB_CACHED) { 1282 spa_t *spa = dn->dn_objset->os_spa; 1283 1284 /* 1285 * Ensure that this block's dnode has been decrypted if 1286 * the caller has requested decrypted data. 1287 */ 1288 err = dbuf_read_verify_dnode_crypt(db, flags); 1289 1290 /* 1291 * If the arc buf is compressed or encrypted and the caller 1292 * requested uncompressed data, we need to untransform it 1293 * before returning. We also call arc_untransform() on any 1294 * unauthenticated blocks, which will verify their MAC if 1295 * the key is now available. 1296 */ 1297 if (err == 0 && db->db_buf != NULL && 1298 (flags & DB_RF_NO_DECRYPT) == 0 && 1299 (arc_is_encrypted(db->db_buf) || 1300 arc_is_unauthenticated(db->db_buf) || 1301 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) { 1302 zbookmark_phys_t zb; 1303 1304 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), 1305 db->db.db_object, db->db_level, db->db_blkid); 1306 dbuf_fix_old_data(db, spa_syncing_txg(spa)); 1307 err = arc_untransform(db->db_buf, spa, &zb, B_FALSE); 1308 dbuf_set_data(db, db->db_buf); 1309 } 1310 mutex_exit(&db->db_mtx); 1311 if (err == 0 && prefetch) 1312 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1313 if ((flags & DB_RF_HAVESTRUCT) == 0) 1314 rw_exit(&dn->dn_struct_rwlock); 1315 DB_DNODE_EXIT(db); 1316 } else if (db->db_state == DB_UNCACHED) { 1317 spa_t *spa = dn->dn_objset->os_spa; 1318 boolean_t need_wait = B_FALSE; 1319 1320 if (zio == NULL && 1321 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1322 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 1323 need_wait = B_TRUE; 1324 } 1325 err = dbuf_read_impl(db, zio, flags); 1326 1327 /* dbuf_read_impl has dropped db_mtx for us */ 1328 1329 if (!err && prefetch) 1330 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1331 1332 if ((flags & DB_RF_HAVESTRUCT) == 0) 1333 rw_exit(&dn->dn_struct_rwlock); 1334 DB_DNODE_EXIT(db); 1335 1336 if (!err && need_wait) 1337 err = zio_wait(zio); 1338 } else { 1339 /* 1340 * Another reader came in while the dbuf was in flight 1341 * between UNCACHED and CACHED. Either a writer will finish 1342 * writing the buffer (sending the dbuf to CACHED) or the 1343 * first reader's request will reach the read_done callback 1344 * and send the dbuf to CACHED. Otherwise, a failure 1345 * occurred and the dbuf went to UNCACHED. 1346 */ 1347 mutex_exit(&db->db_mtx); 1348 if (prefetch) 1349 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1350 if ((flags & DB_RF_HAVESTRUCT) == 0) 1351 rw_exit(&dn->dn_struct_rwlock); 1352 DB_DNODE_EXIT(db); 1353 1354 /* Skip the wait per the caller's request. */ 1355 mutex_enter(&db->db_mtx); 1356 if ((flags & DB_RF_NEVERWAIT) == 0) { 1357 while (db->db_state == DB_READ || 1358 db->db_state == DB_FILL) { 1359 ASSERT(db->db_state == DB_READ || 1360 (flags & DB_RF_HAVESTRUCT) == 0); 1361 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 1362 db, zio_t *, zio); 1363 cv_wait(&db->db_changed, &db->db_mtx); 1364 } 1365 if (db->db_state == DB_UNCACHED) 1366 err = SET_ERROR(EIO); 1367 } 1368 mutex_exit(&db->db_mtx); 1369 } 1370 1371 return (err); 1372 } 1373 1374 static void 1375 dbuf_noread(dmu_buf_impl_t *db) 1376 { 1377 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 1378 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1379 mutex_enter(&db->db_mtx); 1380 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1381 cv_wait(&db->db_changed, &db->db_mtx); 1382 if (db->db_state == DB_UNCACHED) { 1383 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1384 spa_t *spa = db->db_objset->os_spa; 1385 1386 ASSERT(db->db_buf == NULL); 1387 ASSERT(db->db.db_data == NULL); 1388 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); 1389 db->db_state = DB_FILL; 1390 } else if (db->db_state == DB_NOFILL) { 1391 dbuf_clear_data(db); 1392 } else { 1393 ASSERT3U(db->db_state, ==, DB_CACHED); 1394 } 1395 mutex_exit(&db->db_mtx); 1396 } 1397 1398 void 1399 dbuf_unoverride(dbuf_dirty_record_t *dr) 1400 { 1401 dmu_buf_impl_t *db = dr->dr_dbuf; 1402 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 1403 uint64_t txg = dr->dr_txg; 1404 1405 ASSERT(MUTEX_HELD(&db->db_mtx)); 1406 /* 1407 * This assert is valid because dmu_sync() expects to be called by 1408 * a zilog's get_data while holding a range lock. This call only 1409 * comes from dbuf_dirty() callers who must also hold a range lock. 1410 */ 1411 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 1412 ASSERT(db->db_level == 0); 1413 1414 if (db->db_blkid == DMU_BONUS_BLKID || 1415 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 1416 return; 1417 1418 ASSERT(db->db_data_pending != dr); 1419 1420 /* free this block */ 1421 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 1422 zio_free(db->db_objset->os_spa, txg, bp); 1423 1424 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1425 dr->dt.dl.dr_nopwrite = B_FALSE; 1426 dr->dt.dl.dr_has_raw_params = B_FALSE; 1427 1428 /* 1429 * Release the already-written buffer, so we leave it in 1430 * a consistent dirty state. Note that all callers are 1431 * modifying the buffer, so they will immediately do 1432 * another (redundant) arc_release(). Therefore, leave 1433 * the buf thawed to save the effort of freezing & 1434 * immediately re-thawing it. 1435 */ 1436 arc_release(dr->dt.dl.dr_data, db); 1437 } 1438 1439 /* 1440 * Evict (if its unreferenced) or clear (if its referenced) any level-0 1441 * data blocks in the free range, so that any future readers will find 1442 * empty blocks. 1443 */ 1444 void 1445 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 1446 dmu_tx_t *tx) 1447 { 1448 dmu_buf_impl_t db_search; 1449 dmu_buf_impl_t *db, *db_next; 1450 uint64_t txg = tx->tx_txg; 1451 avl_index_t where; 1452 1453 if (end_blkid > dn->dn_maxblkid && 1454 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) 1455 end_blkid = dn->dn_maxblkid; 1456 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 1457 1458 db_search.db_level = 0; 1459 db_search.db_blkid = start_blkid; 1460 db_search.db_state = DB_SEARCH; 1461 1462 mutex_enter(&dn->dn_dbufs_mtx); 1463 db = avl_find(&dn->dn_dbufs, &db_search, &where); 1464 ASSERT3P(db, ==, NULL); 1465 1466 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 1467 1468 for (; db != NULL; db = db_next) { 1469 db_next = AVL_NEXT(&dn->dn_dbufs, db); 1470 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1471 1472 if (db->db_level != 0 || db->db_blkid > end_blkid) { 1473 break; 1474 } 1475 ASSERT3U(db->db_blkid, >=, start_blkid); 1476 1477 /* found a level 0 buffer in the range */ 1478 mutex_enter(&db->db_mtx); 1479 if (dbuf_undirty(db, tx)) { 1480 /* mutex has been dropped and dbuf destroyed */ 1481 continue; 1482 } 1483 1484 if (db->db_state == DB_UNCACHED || 1485 db->db_state == DB_NOFILL || 1486 db->db_state == DB_EVICTING) { 1487 ASSERT(db->db.db_data == NULL); 1488 mutex_exit(&db->db_mtx); 1489 continue; 1490 } 1491 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 1492 /* will be handled in dbuf_read_done or dbuf_rele */ 1493 db->db_freed_in_flight = TRUE; 1494 mutex_exit(&db->db_mtx); 1495 continue; 1496 } 1497 if (zfs_refcount_count(&db->db_holds) == 0) { 1498 ASSERT(db->db_buf); 1499 dbuf_destroy(db); 1500 continue; 1501 } 1502 /* The dbuf is referenced */ 1503 1504 if (db->db_last_dirty != NULL) { 1505 dbuf_dirty_record_t *dr = db->db_last_dirty; 1506 1507 if (dr->dr_txg == txg) { 1508 /* 1509 * This buffer is "in-use", re-adjust the file 1510 * size to reflect that this buffer may 1511 * contain new data when we sync. 1512 */ 1513 if (db->db_blkid != DMU_SPILL_BLKID && 1514 db->db_blkid > dn->dn_maxblkid) 1515 dn->dn_maxblkid = db->db_blkid; 1516 dbuf_unoverride(dr); 1517 } else { 1518 /* 1519 * This dbuf is not dirty in the open context. 1520 * Either uncache it (if its not referenced in 1521 * the open context) or reset its contents to 1522 * empty. 1523 */ 1524 dbuf_fix_old_data(db, txg); 1525 } 1526 } 1527 /* clear the contents if its cached */ 1528 if (db->db_state == DB_CACHED) { 1529 ASSERT(db->db.db_data != NULL); 1530 arc_release(db->db_buf, db); 1531 bzero(db->db.db_data, db->db.db_size); 1532 arc_buf_freeze(db->db_buf); 1533 } 1534 1535 mutex_exit(&db->db_mtx); 1536 } 1537 mutex_exit(&dn->dn_dbufs_mtx); 1538 } 1539 1540 void 1541 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1542 { 1543 arc_buf_t *buf, *obuf; 1544 int osize = db->db.db_size; 1545 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1546 dnode_t *dn; 1547 1548 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1549 1550 DB_DNODE_ENTER(db); 1551 dn = DB_DNODE(db); 1552 1553 /* XXX does *this* func really need the lock? */ 1554 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1555 1556 /* 1557 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1558 * is OK, because there can be no other references to the db 1559 * when we are changing its size, so no concurrent DB_FILL can 1560 * be happening. 1561 */ 1562 /* 1563 * XXX we should be doing a dbuf_read, checking the return 1564 * value and returning that up to our callers 1565 */ 1566 dmu_buf_will_dirty(&db->db, tx); 1567 1568 /* create the data buffer for the new block */ 1569 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); 1570 1571 /* copy old block data to the new block */ 1572 obuf = db->db_buf; 1573 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1574 /* zero the remainder */ 1575 if (size > osize) 1576 bzero((uint8_t *)buf->b_data + osize, size - osize); 1577 1578 mutex_enter(&db->db_mtx); 1579 dbuf_set_data(db, buf); 1580 arc_buf_destroy(obuf, db); 1581 db->db.db_size = size; 1582 1583 if (db->db_level == 0) { 1584 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1585 db->db_last_dirty->dt.dl.dr_data = buf; 1586 } 1587 mutex_exit(&db->db_mtx); 1588 1589 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); 1590 DB_DNODE_EXIT(db); 1591 } 1592 1593 void 1594 dbuf_release_bp(dmu_buf_impl_t *db) 1595 { 1596 objset_t *os = db->db_objset; 1597 1598 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1599 ASSERT(arc_released(os->os_phys_buf) || 1600 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1601 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1602 1603 (void) arc_release(db->db_buf, db); 1604 } 1605 1606 /* 1607 * We already have a dirty record for this TXG, and we are being 1608 * dirtied again. 1609 */ 1610 static void 1611 dbuf_redirty(dbuf_dirty_record_t *dr) 1612 { 1613 dmu_buf_impl_t *db = dr->dr_dbuf; 1614 1615 ASSERT(MUTEX_HELD(&db->db_mtx)); 1616 1617 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1618 /* 1619 * If this buffer has already been written out, 1620 * we now need to reset its state. 1621 */ 1622 dbuf_unoverride(dr); 1623 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1624 db->db_state != DB_NOFILL) { 1625 /* Already released on initial dirty, so just thaw. */ 1626 ASSERT(arc_released(db->db_buf)); 1627 arc_buf_thaw(db->db_buf); 1628 } 1629 } 1630 } 1631 1632 dbuf_dirty_record_t * 1633 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1634 { 1635 dnode_t *dn; 1636 objset_t *os; 1637 dbuf_dirty_record_t **drp, *dr; 1638 int drop_struct_lock = FALSE; 1639 int txgoff = tx->tx_txg & TXG_MASK; 1640 1641 ASSERT(tx->tx_txg != 0); 1642 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 1643 DMU_TX_DIRTY_BUF(tx, db); 1644 1645 DB_DNODE_ENTER(db); 1646 dn = DB_DNODE(db); 1647 /* 1648 * Shouldn't dirty a regular buffer in syncing context. Private 1649 * objects may be dirtied in syncing context, but only if they 1650 * were already pre-dirtied in open context. 1651 */ 1652 #ifdef DEBUG 1653 if (dn->dn_objset->os_dsl_dataset != NULL) { 1654 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1655 RW_READER, FTAG); 1656 } 1657 ASSERT(!dmu_tx_is_syncing(tx) || 1658 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1659 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1660 dn->dn_objset->os_dsl_dataset == NULL); 1661 if (dn->dn_objset->os_dsl_dataset != NULL) 1662 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); 1663 #endif 1664 /* 1665 * We make this assert for private objects as well, but after we 1666 * check if we're already dirty. They are allowed to re-dirty 1667 * in syncing context. 1668 */ 1669 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1670 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1671 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1672 1673 mutex_enter(&db->db_mtx); 1674 /* 1675 * XXX make this true for indirects too? The problem is that 1676 * transactions created with dmu_tx_create_assigned() from 1677 * syncing context don't bother holding ahead. 1678 */ 1679 ASSERT(db->db_level != 0 || 1680 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1681 db->db_state == DB_NOFILL); 1682 1683 mutex_enter(&dn->dn_mtx); 1684 /* 1685 * Don't set dirtyctx to SYNC if we're just modifying this as we 1686 * initialize the objset. 1687 */ 1688 if (dn->dn_dirtyctx == DN_UNDIRTIED) { 1689 if (dn->dn_objset->os_dsl_dataset != NULL) { 1690 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1691 RW_READER, FTAG); 1692 } 1693 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1694 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? 1695 DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1696 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1697 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1698 } 1699 if (dn->dn_objset->os_dsl_dataset != NULL) { 1700 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1701 FTAG); 1702 } 1703 } 1704 1705 if (tx->tx_txg > dn->dn_dirty_txg) 1706 dn->dn_dirty_txg = tx->tx_txg; 1707 mutex_exit(&dn->dn_mtx); 1708 1709 if (db->db_blkid == DMU_SPILL_BLKID) 1710 dn->dn_have_spill = B_TRUE; 1711 1712 /* 1713 * If this buffer is already dirty, we're done. 1714 */ 1715 drp = &db->db_last_dirty; 1716 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1717 db->db.db_object == DMU_META_DNODE_OBJECT); 1718 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1719 drp = &dr->dr_next; 1720 if (dr && dr->dr_txg == tx->tx_txg) { 1721 DB_DNODE_EXIT(db); 1722 1723 dbuf_redirty(dr); 1724 mutex_exit(&db->db_mtx); 1725 return (dr); 1726 } 1727 1728 /* 1729 * Only valid if not already dirty. 1730 */ 1731 ASSERT(dn->dn_object == 0 || 1732 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1733 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1734 1735 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1736 1737 /* 1738 * We should only be dirtying in syncing context if it's the 1739 * mos or we're initializing the os or it's a special object. 1740 * However, we are allowed to dirty in syncing context provided 1741 * we already dirtied it in open context. Hence we must make 1742 * this assertion only if we're not already dirty. 1743 */ 1744 os = dn->dn_objset; 1745 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); 1746 #ifdef DEBUG 1747 if (dn->dn_objset->os_dsl_dataset != NULL) 1748 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); 1749 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1750 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1751 if (dn->dn_objset->os_dsl_dataset != NULL) 1752 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); 1753 #endif 1754 ASSERT(db->db.db_size != 0); 1755 1756 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1757 1758 if (db->db_blkid != DMU_BONUS_BLKID) { 1759 dmu_objset_willuse_space(os, db->db.db_size, tx); 1760 } 1761 1762 /* 1763 * If this buffer is dirty in an old transaction group we need 1764 * to make a copy of it so that the changes we make in this 1765 * transaction group won't leak out when we sync the older txg. 1766 */ 1767 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1768 if (db->db_level == 0) { 1769 void *data_old = db->db_buf; 1770 1771 if (db->db_state != DB_NOFILL) { 1772 if (db->db_blkid == DMU_BONUS_BLKID) { 1773 dbuf_fix_old_data(db, tx->tx_txg); 1774 data_old = db->db.db_data; 1775 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1776 /* 1777 * Release the data buffer from the cache so 1778 * that we can modify it without impacting 1779 * possible other users of this cached data 1780 * block. Note that indirect blocks and 1781 * private objects are not released until the 1782 * syncing state (since they are only modified 1783 * then). 1784 */ 1785 arc_release(db->db_buf, db); 1786 dbuf_fix_old_data(db, tx->tx_txg); 1787 data_old = db->db_buf; 1788 } 1789 ASSERT(data_old != NULL); 1790 } 1791 dr->dt.dl.dr_data = data_old; 1792 } else { 1793 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1794 list_create(&dr->dt.di.dr_children, 1795 sizeof (dbuf_dirty_record_t), 1796 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1797 } 1798 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1799 dr->dr_accounted = db->db.db_size; 1800 dr->dr_dbuf = db; 1801 dr->dr_txg = tx->tx_txg; 1802 dr->dr_next = *drp; 1803 *drp = dr; 1804 1805 /* 1806 * We could have been freed_in_flight between the dbuf_noread 1807 * and dbuf_dirty. We win, as though the dbuf_noread() had 1808 * happened after the free. 1809 */ 1810 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1811 db->db_blkid != DMU_SPILL_BLKID) { 1812 mutex_enter(&dn->dn_mtx); 1813 if (dn->dn_free_ranges[txgoff] != NULL) { 1814 range_tree_clear(dn->dn_free_ranges[txgoff], 1815 db->db_blkid, 1); 1816 } 1817 mutex_exit(&dn->dn_mtx); 1818 db->db_freed_in_flight = FALSE; 1819 } 1820 1821 /* 1822 * This buffer is now part of this txg 1823 */ 1824 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1825 db->db_dirtycnt += 1; 1826 ASSERT3U(db->db_dirtycnt, <=, 3); 1827 1828 mutex_exit(&db->db_mtx); 1829 1830 if (db->db_blkid == DMU_BONUS_BLKID || 1831 db->db_blkid == DMU_SPILL_BLKID) { 1832 mutex_enter(&dn->dn_mtx); 1833 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1834 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1835 mutex_exit(&dn->dn_mtx); 1836 dnode_setdirty(dn, tx); 1837 DB_DNODE_EXIT(db); 1838 return (dr); 1839 } 1840 1841 /* 1842 * The dn_struct_rwlock prevents db_blkptr from changing 1843 * due to a write from syncing context completing 1844 * while we are running, so we want to acquire it before 1845 * looking at db_blkptr. 1846 */ 1847 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1848 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1849 drop_struct_lock = TRUE; 1850 } 1851 1852 /* 1853 * We need to hold the dn_struct_rwlock to make this assertion, 1854 * because it protects dn_phys / dn_next_nlevels from changing. 1855 */ 1856 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1857 dn->dn_phys->dn_nlevels > db->db_level || 1858 dn->dn_next_nlevels[txgoff] > db->db_level || 1859 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1860 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1861 1862 /* 1863 * If we are overwriting a dedup BP, then unless it is snapshotted, 1864 * when we get to syncing context we will need to decrement its 1865 * refcount in the DDT. Prefetch the relevant DDT block so that 1866 * syncing context won't have to wait for the i/o. 1867 */ 1868 ddt_prefetch(os->os_spa, db->db_blkptr); 1869 1870 if (db->db_level == 0) { 1871 ASSERT(!db->db_objset->os_raw_receive || 1872 dn->dn_maxblkid >= db->db_blkid); 1873 dnode_new_blkid(dn, db->db_blkid, tx, 1874 drop_struct_lock, B_FALSE); 1875 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1876 } 1877 1878 if (db->db_level+1 < dn->dn_nlevels) { 1879 dmu_buf_impl_t *parent = db->db_parent; 1880 dbuf_dirty_record_t *di; 1881 int parent_held = FALSE; 1882 1883 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1884 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1885 1886 parent = dbuf_hold_level(dn, db->db_level+1, 1887 db->db_blkid >> epbs, FTAG); 1888 ASSERT(parent != NULL); 1889 parent_held = TRUE; 1890 } 1891 if (drop_struct_lock) 1892 rw_exit(&dn->dn_struct_rwlock); 1893 ASSERT3U(db->db_level+1, ==, parent->db_level); 1894 di = dbuf_dirty(parent, tx); 1895 if (parent_held) 1896 dbuf_rele(parent, FTAG); 1897 1898 mutex_enter(&db->db_mtx); 1899 /* 1900 * Since we've dropped the mutex, it's possible that 1901 * dbuf_undirty() might have changed this out from under us. 1902 */ 1903 if (db->db_last_dirty == dr || 1904 dn->dn_object == DMU_META_DNODE_OBJECT) { 1905 mutex_enter(&di->dt.di.dr_mtx); 1906 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1907 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1908 list_insert_tail(&di->dt.di.dr_children, dr); 1909 mutex_exit(&di->dt.di.dr_mtx); 1910 dr->dr_parent = di; 1911 } 1912 mutex_exit(&db->db_mtx); 1913 } else { 1914 ASSERT(db->db_level+1 == dn->dn_nlevels); 1915 ASSERT(db->db_blkid < dn->dn_nblkptr); 1916 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1917 mutex_enter(&dn->dn_mtx); 1918 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1919 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1920 mutex_exit(&dn->dn_mtx); 1921 if (drop_struct_lock) 1922 rw_exit(&dn->dn_struct_rwlock); 1923 } 1924 1925 dnode_setdirty(dn, tx); 1926 DB_DNODE_EXIT(db); 1927 return (dr); 1928 } 1929 1930 /* 1931 * Undirty a buffer in the transaction group referenced by the given 1932 * transaction. Return whether this evicted the dbuf. 1933 */ 1934 static boolean_t 1935 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1936 { 1937 dnode_t *dn; 1938 uint64_t txg = tx->tx_txg; 1939 dbuf_dirty_record_t *dr, **drp; 1940 1941 ASSERT(txg != 0); 1942 1943 /* 1944 * Due to our use of dn_nlevels below, this can only be called 1945 * in open context, unless we are operating on the MOS. 1946 * From syncing context, dn_nlevels may be different from the 1947 * dn_nlevels used when dbuf was dirtied. 1948 */ 1949 ASSERT(db->db_objset == 1950 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1951 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1952 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1953 ASSERT0(db->db_level); 1954 ASSERT(MUTEX_HELD(&db->db_mtx)); 1955 1956 /* 1957 * If this buffer is not dirty, we're done. 1958 */ 1959 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1960 if (dr->dr_txg <= txg) 1961 break; 1962 if (dr == NULL || dr->dr_txg < txg) 1963 return (B_FALSE); 1964 ASSERT(dr->dr_txg == txg); 1965 ASSERT(dr->dr_dbuf == db); 1966 1967 DB_DNODE_ENTER(db); 1968 dn = DB_DNODE(db); 1969 1970 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1971 1972 ASSERT(db->db.db_size != 0); 1973 1974 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1975 dr->dr_accounted, txg); 1976 1977 *drp = dr->dr_next; 1978 1979 /* 1980 * Note that there are three places in dbuf_dirty() 1981 * where this dirty record may be put on a list. 1982 * Make sure to do a list_remove corresponding to 1983 * every one of those list_insert calls. 1984 */ 1985 if (dr->dr_parent) { 1986 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1987 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1988 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1989 } else if (db->db_blkid == DMU_SPILL_BLKID || 1990 db->db_level + 1 == dn->dn_nlevels) { 1991 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1992 mutex_enter(&dn->dn_mtx); 1993 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1994 mutex_exit(&dn->dn_mtx); 1995 } 1996 DB_DNODE_EXIT(db); 1997 1998 if (db->db_state != DB_NOFILL) { 1999 dbuf_unoverride(dr); 2000 2001 ASSERT(db->db_buf != NULL); 2002 ASSERT(dr->dt.dl.dr_data != NULL); 2003 if (dr->dt.dl.dr_data != db->db_buf) 2004 arc_buf_destroy(dr->dt.dl.dr_data, db); 2005 } 2006 2007 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2008 2009 ASSERT(db->db_dirtycnt > 0); 2010 db->db_dirtycnt -= 1; 2011 2012 if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 2013 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); 2014 dbuf_destroy(db); 2015 return (B_TRUE); 2016 } 2017 2018 return (B_FALSE); 2019 } 2020 2021 static void 2022 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx) 2023 { 2024 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2025 2026 ASSERT(tx->tx_txg != 0); 2027 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 2028 2029 /* 2030 * Quick check for dirtyness. For already dirty blocks, this 2031 * reduces runtime of this function by >90%, and overall performance 2032 * by 50% for some workloads (e.g. file deletion with indirect blocks 2033 * cached). 2034 */ 2035 mutex_enter(&db->db_mtx); 2036 dbuf_dirty_record_t *dr; 2037 for (dr = db->db_last_dirty; 2038 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 2039 /* 2040 * It's possible that it is already dirty but not cached, 2041 * because there are some calls to dbuf_dirty() that don't 2042 * go through dmu_buf_will_dirty(). 2043 */ 2044 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 2045 /* This dbuf is already dirty and cached. */ 2046 dbuf_redirty(dr); 2047 mutex_exit(&db->db_mtx); 2048 return; 2049 } 2050 } 2051 mutex_exit(&db->db_mtx); 2052 2053 DB_DNODE_ENTER(db); 2054 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 2055 flags |= DB_RF_HAVESTRUCT; 2056 DB_DNODE_EXIT(db); 2057 (void) dbuf_read(db, NULL, flags); 2058 (void) dbuf_dirty(db, tx); 2059 } 2060 2061 void 2062 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 2063 { 2064 dmu_buf_will_dirty_impl(db_fake, 2065 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx); 2066 } 2067 2068 void 2069 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 2070 { 2071 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2072 2073 db->db_state = DB_NOFILL; 2074 2075 dmu_buf_will_fill(db_fake, tx); 2076 } 2077 2078 void 2079 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 2080 { 2081 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2082 2083 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2084 ASSERT(tx->tx_txg != 0); 2085 ASSERT(db->db_level == 0); 2086 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 2087 2088 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 2089 dmu_tx_private_ok(tx)); 2090 2091 dbuf_noread(db); 2092 (void) dbuf_dirty(db, tx); 2093 } 2094 2095 /* 2096 * This function is effectively the same as dmu_buf_will_dirty(), but 2097 * indicates the caller expects raw encrypted data in the db, and provides 2098 * the crypt params (byteorder, salt, iv, mac) which should be stored in the 2099 * blkptr_t when this dbuf is written. This is only used for blocks of 2100 * dnodes during a raw receive. 2101 */ 2102 void 2103 dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder, 2104 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx) 2105 { 2106 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2107 dbuf_dirty_record_t *dr; 2108 2109 /* 2110 * dr_has_raw_params is only processed for blocks of dnodes 2111 * (see dbuf_sync_dnode_leaf_crypt()). 2112 */ 2113 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); 2114 ASSERT3U(db->db_level, ==, 0); 2115 2116 dmu_buf_will_dirty_impl(db_fake, 2117 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx); 2118 2119 dr = db->db_last_dirty; 2120 while (dr != NULL && dr->dr_txg > tx->tx_txg) 2121 dr = dr->dr_next; 2122 2123 ASSERT3P(dr, !=, NULL); 2124 ASSERT3U(dr->dr_txg, ==, tx->tx_txg); 2125 2126 dr->dt.dl.dr_has_raw_params = B_TRUE; 2127 dr->dt.dl.dr_byteorder = byteorder; 2128 bcopy(salt, dr->dt.dl.dr_salt, ZIO_DATA_SALT_LEN); 2129 bcopy(iv, dr->dt.dl.dr_iv, ZIO_DATA_IV_LEN); 2130 bcopy(mac, dr->dt.dl.dr_mac, ZIO_DATA_MAC_LEN); 2131 } 2132 2133 #pragma weak dmu_buf_fill_done = dbuf_fill_done 2134 /* ARGSUSED */ 2135 void 2136 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 2137 { 2138 mutex_enter(&db->db_mtx); 2139 DBUF_VERIFY(db); 2140 2141 if (db->db_state == DB_FILL) { 2142 if (db->db_level == 0 && db->db_freed_in_flight) { 2143 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2144 /* we were freed while filling */ 2145 /* XXX dbuf_undirty? */ 2146 bzero(db->db.db_data, db->db.db_size); 2147 db->db_freed_in_flight = FALSE; 2148 } 2149 db->db_state = DB_CACHED; 2150 cv_broadcast(&db->db_changed); 2151 } 2152 mutex_exit(&db->db_mtx); 2153 } 2154 2155 void 2156 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 2157 bp_embedded_type_t etype, enum zio_compress comp, 2158 int uncompressed_size, int compressed_size, int byteorder, 2159 dmu_tx_t *tx) 2160 { 2161 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2162 struct dirty_leaf *dl; 2163 dmu_object_type_t type; 2164 2165 if (etype == BP_EMBEDDED_TYPE_DATA) { 2166 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 2167 SPA_FEATURE_EMBEDDED_DATA)); 2168 } 2169 2170 DB_DNODE_ENTER(db); 2171 type = DB_DNODE(db)->dn_type; 2172 DB_DNODE_EXIT(db); 2173 2174 ASSERT0(db->db_level); 2175 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2176 2177 dmu_buf_will_not_fill(dbuf, tx); 2178 2179 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 2180 dl = &db->db_last_dirty->dt.dl; 2181 encode_embedded_bp_compressed(&dl->dr_overridden_by, 2182 data, comp, uncompressed_size, compressed_size); 2183 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 2184 BP_SET_TYPE(&dl->dr_overridden_by, type); 2185 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 2186 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 2187 2188 dl->dr_override_state = DR_OVERRIDDEN; 2189 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 2190 } 2191 2192 /* 2193 * Directly assign a provided arc buf to a given dbuf if it's not referenced 2194 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 2195 */ 2196 void 2197 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 2198 { 2199 ASSERT(!zfs_refcount_is_zero(&db->db_holds)); 2200 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2201 ASSERT(db->db_level == 0); 2202 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); 2203 ASSERT(buf != NULL); 2204 ASSERT(arc_buf_lsize(buf) == db->db.db_size); 2205 ASSERT(tx->tx_txg != 0); 2206 2207 arc_return_buf(buf, db); 2208 ASSERT(arc_released(buf)); 2209 2210 mutex_enter(&db->db_mtx); 2211 2212 while (db->db_state == DB_READ || db->db_state == DB_FILL) 2213 cv_wait(&db->db_changed, &db->db_mtx); 2214 2215 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 2216 2217 if (db->db_state == DB_CACHED && 2218 zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 2219 /* 2220 * In practice, we will never have a case where we have an 2221 * encrypted arc buffer while additional holds exist on the 2222 * dbuf. We don't handle this here so we simply assert that 2223 * fact instead. 2224 */ 2225 ASSERT(!arc_is_encrypted(buf)); 2226 mutex_exit(&db->db_mtx); 2227 (void) dbuf_dirty(db, tx); 2228 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 2229 arc_buf_destroy(buf, db); 2230 xuio_stat_wbuf_copied(); 2231 return; 2232 } 2233 2234 xuio_stat_wbuf_nocopy(); 2235 if (db->db_state == DB_CACHED) { 2236 dbuf_dirty_record_t *dr = db->db_last_dirty; 2237 2238 ASSERT(db->db_buf != NULL); 2239 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 2240 ASSERT(dr->dt.dl.dr_data == db->db_buf); 2241 2242 if (!arc_released(db->db_buf)) { 2243 ASSERT(dr->dt.dl.dr_override_state == 2244 DR_OVERRIDDEN); 2245 arc_release(db->db_buf, db); 2246 } 2247 dr->dt.dl.dr_data = buf; 2248 arc_buf_destroy(db->db_buf, db); 2249 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 2250 arc_release(db->db_buf, db); 2251 arc_buf_destroy(db->db_buf, db); 2252 } 2253 db->db_buf = NULL; 2254 } 2255 ASSERT(db->db_buf == NULL); 2256 dbuf_set_data(db, buf); 2257 db->db_state = DB_FILL; 2258 mutex_exit(&db->db_mtx); 2259 (void) dbuf_dirty(db, tx); 2260 dmu_buf_fill_done(&db->db, tx); 2261 } 2262 2263 void 2264 dbuf_destroy(dmu_buf_impl_t *db) 2265 { 2266 dnode_t *dn; 2267 dmu_buf_impl_t *parent = db->db_parent; 2268 dmu_buf_impl_t *dndb; 2269 2270 ASSERT(MUTEX_HELD(&db->db_mtx)); 2271 ASSERT(zfs_refcount_is_zero(&db->db_holds)); 2272 2273 if (db->db_buf != NULL) { 2274 arc_buf_destroy(db->db_buf, db); 2275 db->db_buf = NULL; 2276 } 2277 2278 if (db->db_blkid == DMU_BONUS_BLKID) { 2279 int slots = DB_DNODE(db)->dn_num_slots; 2280 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); 2281 if (db->db.db_data != NULL) { 2282 zio_buf_free(db->db.db_data, bonuslen); 2283 arc_space_return(bonuslen, ARC_SPACE_BONUS); 2284 db->db_state = DB_UNCACHED; 2285 } 2286 } 2287 2288 dbuf_clear_data(db); 2289 2290 if (multilist_link_active(&db->db_cache_link)) { 2291 ASSERT(db->db_caching_status == DB_DBUF_CACHE || 2292 db->db_caching_status == DB_DBUF_METADATA_CACHE); 2293 2294 multilist_remove(dbuf_caches[db->db_caching_status].cache, db); 2295 (void) zfs_refcount_remove_many( 2296 &dbuf_caches[db->db_caching_status].size, 2297 db->db.db_size, db); 2298 2299 db->db_caching_status = DB_NO_CACHE; 2300 } 2301 2302 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 2303 ASSERT(db->db_data_pending == NULL); 2304 2305 db->db_state = DB_EVICTING; 2306 db->db_blkptr = NULL; 2307 2308 /* 2309 * Now that db_state is DB_EVICTING, nobody else can find this via 2310 * the hash table. We can now drop db_mtx, which allows us to 2311 * acquire the dn_dbufs_mtx. 2312 */ 2313 mutex_exit(&db->db_mtx); 2314 2315 DB_DNODE_ENTER(db); 2316 dn = DB_DNODE(db); 2317 dndb = dn->dn_dbuf; 2318 if (db->db_blkid != DMU_BONUS_BLKID) { 2319 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); 2320 if (needlock) 2321 mutex_enter(&dn->dn_dbufs_mtx); 2322 avl_remove(&dn->dn_dbufs, db); 2323 atomic_dec_32(&dn->dn_dbufs_count); 2324 membar_producer(); 2325 DB_DNODE_EXIT(db); 2326 if (needlock) 2327 mutex_exit(&dn->dn_dbufs_mtx); 2328 /* 2329 * Decrementing the dbuf count means that the hold corresponding 2330 * to the removed dbuf is no longer discounted in dnode_move(), 2331 * so the dnode cannot be moved until after we release the hold. 2332 * The membar_producer() ensures visibility of the decremented 2333 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 2334 * release any lock. 2335 */ 2336 mutex_enter(&dn->dn_mtx); 2337 dnode_rele_and_unlock(dn, db, B_TRUE); 2338 db->db_dnode_handle = NULL; 2339 2340 dbuf_hash_remove(db); 2341 } else { 2342 DB_DNODE_EXIT(db); 2343 } 2344 2345 ASSERT(zfs_refcount_is_zero(&db->db_holds)); 2346 2347 db->db_parent = NULL; 2348 2349 ASSERT(db->db_buf == NULL); 2350 ASSERT(db->db.db_data == NULL); 2351 ASSERT(db->db_hash_next == NULL); 2352 ASSERT(db->db_blkptr == NULL); 2353 ASSERT(db->db_data_pending == NULL); 2354 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE); 2355 ASSERT(!multilist_link_active(&db->db_cache_link)); 2356 2357 kmem_cache_free(dbuf_kmem_cache, db); 2358 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2359 2360 /* 2361 * If this dbuf is referenced from an indirect dbuf, 2362 * decrement the ref count on the indirect dbuf. 2363 */ 2364 if (parent && parent != dndb) { 2365 mutex_enter(&parent->db_mtx); 2366 dbuf_rele_and_unlock(parent, db, B_TRUE); 2367 } 2368 } 2369 2370 /* 2371 * Note: While bpp will always be updated if the function returns success, 2372 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 2373 * this happens when the dnode is the meta-dnode, or a userused or groupused 2374 * object. 2375 */ 2376 static int 2377 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 2378 dmu_buf_impl_t **parentp, blkptr_t **bpp) 2379 { 2380 *parentp = NULL; 2381 *bpp = NULL; 2382 2383 ASSERT(blkid != DMU_BONUS_BLKID); 2384 2385 if (blkid == DMU_SPILL_BLKID) { 2386 mutex_enter(&dn->dn_mtx); 2387 if (dn->dn_have_spill && 2388 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 2389 *bpp = DN_SPILL_BLKPTR(dn->dn_phys); 2390 else 2391 *bpp = NULL; 2392 dbuf_add_ref(dn->dn_dbuf, NULL); 2393 *parentp = dn->dn_dbuf; 2394 mutex_exit(&dn->dn_mtx); 2395 return (0); 2396 } 2397 2398 int nlevels = 2399 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; 2400 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2401 2402 ASSERT3U(level * epbs, <, 64); 2403 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2404 /* 2405 * This assertion shouldn't trip as long as the max indirect block size 2406 * is less than 1M. The reason for this is that up to that point, 2407 * the number of levels required to address an entire object with blocks 2408 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In 2409 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 2410 * (i.e. we can address the entire object), objects will all use at most 2411 * N-1 levels and the assertion won't overflow. However, once epbs is 2412 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be 2413 * enough to address an entire object, so objects will have 5 levels, 2414 * but then this assertion will overflow. 2415 * 2416 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we 2417 * need to redo this logic to handle overflows. 2418 */ 2419 ASSERT(level >= nlevels || 2420 ((nlevels - level - 1) * epbs) + 2421 highbit64(dn->dn_phys->dn_nblkptr) <= 64); 2422 if (level >= nlevels || 2423 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << 2424 ((nlevels - level - 1) * epbs)) || 2425 (fail_sparse && 2426 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 2427 /* the buffer has no parent yet */ 2428 return (SET_ERROR(ENOENT)); 2429 } else if (level < nlevels-1) { 2430 /* this block is referenced from an indirect block */ 2431 int err = dbuf_hold_impl(dn, level+1, 2432 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 2433 if (err) 2434 return (err); 2435 err = dbuf_read(*parentp, NULL, 2436 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 2437 if (err) { 2438 dbuf_rele(*parentp, NULL); 2439 *parentp = NULL; 2440 return (err); 2441 } 2442 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 2443 (blkid & ((1ULL << epbs) - 1)); 2444 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) 2445 ASSERT(BP_IS_HOLE(*bpp)); 2446 return (0); 2447 } else { 2448 /* the block is referenced from the dnode */ 2449 ASSERT3U(level, ==, nlevels-1); 2450 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 2451 blkid < dn->dn_phys->dn_nblkptr); 2452 if (dn->dn_dbuf) { 2453 dbuf_add_ref(dn->dn_dbuf, NULL); 2454 *parentp = dn->dn_dbuf; 2455 } 2456 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 2457 return (0); 2458 } 2459 } 2460 2461 static dmu_buf_impl_t * 2462 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 2463 dmu_buf_impl_t *parent, blkptr_t *blkptr) 2464 { 2465 objset_t *os = dn->dn_objset; 2466 dmu_buf_impl_t *db, *odb; 2467 2468 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2469 ASSERT(dn->dn_type != DMU_OT_NONE); 2470 2471 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); 2472 2473 db->db_objset = os; 2474 db->db.db_object = dn->dn_object; 2475 db->db_level = level; 2476 db->db_blkid = blkid; 2477 db->db_last_dirty = NULL; 2478 db->db_dirtycnt = 0; 2479 db->db_dnode_handle = dn->dn_handle; 2480 db->db_parent = parent; 2481 db->db_blkptr = blkptr; 2482 2483 db->db_user = NULL; 2484 db->db_user_immediate_evict = FALSE; 2485 db->db_freed_in_flight = FALSE; 2486 db->db_pending_evict = FALSE; 2487 2488 if (blkid == DMU_BONUS_BLKID) { 2489 ASSERT3P(parent, ==, dn->dn_dbuf); 2490 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - 2491 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 2492 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 2493 db->db.db_offset = DMU_BONUS_BLKID; 2494 db->db_state = DB_UNCACHED; 2495 db->db_caching_status = DB_NO_CACHE; 2496 /* the bonus dbuf is not placed in the hash table */ 2497 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2498 return (db); 2499 } else if (blkid == DMU_SPILL_BLKID) { 2500 db->db.db_size = (blkptr != NULL) ? 2501 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 2502 db->db.db_offset = 0; 2503 } else { 2504 int blocksize = 2505 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 2506 db->db.db_size = blocksize; 2507 db->db.db_offset = db->db_blkid * blocksize; 2508 } 2509 2510 /* 2511 * Hold the dn_dbufs_mtx while we get the new dbuf 2512 * in the hash table *and* added to the dbufs list. 2513 * This prevents a possible deadlock with someone 2514 * trying to look up this dbuf before its added to the 2515 * dn_dbufs list. 2516 */ 2517 mutex_enter(&dn->dn_dbufs_mtx); 2518 db->db_state = DB_EVICTING; 2519 if ((odb = dbuf_hash_insert(db)) != NULL) { 2520 /* someone else inserted it first */ 2521 kmem_cache_free(dbuf_kmem_cache, db); 2522 mutex_exit(&dn->dn_dbufs_mtx); 2523 return (odb); 2524 } 2525 avl_add(&dn->dn_dbufs, db); 2526 2527 db->db_state = DB_UNCACHED; 2528 db->db_caching_status = DB_NO_CACHE; 2529 mutex_exit(&dn->dn_dbufs_mtx); 2530 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2531 2532 if (parent && parent != dn->dn_dbuf) 2533 dbuf_add_ref(parent, db); 2534 2535 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 2536 zfs_refcount_count(&dn->dn_holds) > 0); 2537 (void) zfs_refcount_add(&dn->dn_holds, db); 2538 atomic_inc_32(&dn->dn_dbufs_count); 2539 2540 dprintf_dbuf(db, "db=%p\n", db); 2541 2542 return (db); 2543 } 2544 2545 typedef struct dbuf_prefetch_arg { 2546 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2547 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2548 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2549 int dpa_curlevel; /* The current level that we're reading */ 2550 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ 2551 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2552 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2553 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2554 } dbuf_prefetch_arg_t; 2555 2556 /* 2557 * Actually issue the prefetch read for the block given. 2558 */ 2559 static void 2560 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2561 { 2562 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2563 return; 2564 2565 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; 2566 arc_flags_t aflags = 2567 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2568 2569 /* dnodes are always read as raw and then converted later */ 2570 if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) && 2571 dpa->dpa_curlevel == 0) 2572 zio_flags |= ZIO_FLAG_RAW; 2573 2574 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2575 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2576 ASSERT(dpa->dpa_zio != NULL); 2577 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2578 dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb); 2579 } 2580 2581 /* 2582 * Called when an indirect block above our prefetch target is read in. This 2583 * will either read in the next indirect block down the tree or issue the actual 2584 * prefetch if the next block down is our target. 2585 */ 2586 /* ARGSUSED */ 2587 static void 2588 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb, 2589 const blkptr_t *iobp, arc_buf_t *abuf, void *private) 2590 { 2591 dbuf_prefetch_arg_t *dpa = private; 2592 2593 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2594 ASSERT3S(dpa->dpa_curlevel, >, 0); 2595 2596 if (abuf == NULL) { 2597 ASSERT(zio == NULL || zio->io_error != 0); 2598 kmem_free(dpa, sizeof (*dpa)); 2599 return; 2600 } 2601 ASSERT(zio == NULL || zio->io_error == 0); 2602 2603 /* 2604 * The dpa_dnode is only valid if we are called with a NULL 2605 * zio. This indicates that the arc_read() returned without 2606 * first calling zio_read() to issue a physical read. Once 2607 * a physical read is made the dpa_dnode must be invalidated 2608 * as the locks guarding it may have been dropped. If the 2609 * dpa_dnode is still valid, then we want to add it to the dbuf 2610 * cache. To do so, we must hold the dbuf associated with the block 2611 * we just prefetched, read its contents so that we associate it 2612 * with an arc_buf_t, and then release it. 2613 */ 2614 if (zio != NULL) { 2615 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2616 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) { 2617 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); 2618 } else { 2619 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2620 } 2621 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2622 2623 dpa->dpa_dnode = NULL; 2624 } else if (dpa->dpa_dnode != NULL) { 2625 uint64_t curblkid = dpa->dpa_zb.zb_blkid >> 2626 (dpa->dpa_epbs * (dpa->dpa_curlevel - 2627 dpa->dpa_zb.zb_level)); 2628 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, 2629 dpa->dpa_curlevel, curblkid, FTAG); 2630 (void) dbuf_read(db, NULL, 2631 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); 2632 dbuf_rele(db, FTAG); 2633 } 2634 2635 dpa->dpa_curlevel--; 2636 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2637 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2638 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2639 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2640 2641 if (BP_IS_HOLE(bp)) { 2642 kmem_free(dpa, sizeof (*dpa)); 2643 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2644 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2645 dbuf_issue_final_prefetch(dpa, bp); 2646 kmem_free(dpa, sizeof (*dpa)); 2647 } else { 2648 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2649 zbookmark_phys_t zb; 2650 2651 /* flag if L2ARC eligible, l2arc_noprefetch then decides */ 2652 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) 2653 iter_aflags |= ARC_FLAG_L2CACHE; 2654 2655 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2656 2657 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2658 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2659 2660 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2661 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2662 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2663 &iter_aflags, &zb); 2664 } 2665 2666 arc_buf_destroy(abuf, private); 2667 } 2668 2669 /* 2670 * Issue prefetch reads for the given block on the given level. If the indirect 2671 * blocks above that block are not in memory, we will read them in 2672 * asynchronously. As a result, this call never blocks waiting for a read to 2673 * complete. Note that the prefetch might fail if the dataset is encrypted and 2674 * the encryption key is unmapped before the IO completes. 2675 */ 2676 void 2677 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2678 arc_flags_t aflags) 2679 { 2680 blkptr_t bp; 2681 int epbs, nlevels, curlevel; 2682 uint64_t curblkid; 2683 2684 ASSERT(blkid != DMU_BONUS_BLKID); 2685 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2686 2687 if (blkid > dn->dn_maxblkid) 2688 return; 2689 2690 if (dnode_block_freed(dn, blkid)) 2691 return; 2692 2693 /* 2694 * This dnode hasn't been written to disk yet, so there's nothing to 2695 * prefetch. 2696 */ 2697 nlevels = dn->dn_phys->dn_nlevels; 2698 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2699 return; 2700 2701 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2702 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2703 return; 2704 2705 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2706 level, blkid); 2707 if (db != NULL) { 2708 mutex_exit(&db->db_mtx); 2709 /* 2710 * This dbuf already exists. It is either CACHED, or 2711 * (we assume) about to be read or filled. 2712 */ 2713 return; 2714 } 2715 2716 /* 2717 * Find the closest ancestor (indirect block) of the target block 2718 * that is present in the cache. In this indirect block, we will 2719 * find the bp that is at curlevel, curblkid. 2720 */ 2721 curlevel = level; 2722 curblkid = blkid; 2723 while (curlevel < nlevels - 1) { 2724 int parent_level = curlevel + 1; 2725 uint64_t parent_blkid = curblkid >> epbs; 2726 dmu_buf_impl_t *db; 2727 2728 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2729 FALSE, TRUE, FTAG, &db) == 0) { 2730 blkptr_t *bpp = db->db_buf->b_data; 2731 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2732 dbuf_rele(db, FTAG); 2733 break; 2734 } 2735 2736 curlevel = parent_level; 2737 curblkid = parent_blkid; 2738 } 2739 2740 if (curlevel == nlevels - 1) { 2741 /* No cached indirect blocks found. */ 2742 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2743 bp = dn->dn_phys->dn_blkptr[curblkid]; 2744 } 2745 if (BP_IS_HOLE(&bp)) 2746 return; 2747 2748 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2749 2750 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2751 ZIO_FLAG_CANFAIL); 2752 2753 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2754 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2755 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2756 dn->dn_object, level, blkid); 2757 dpa->dpa_curlevel = curlevel; 2758 dpa->dpa_prio = prio; 2759 dpa->dpa_aflags = aflags; 2760 dpa->dpa_spa = dn->dn_objset->os_spa; 2761 dpa->dpa_dnode = dn; 2762 dpa->dpa_epbs = epbs; 2763 dpa->dpa_zio = pio; 2764 2765 /* flag if L2ARC eligible, l2arc_noprefetch then decides */ 2766 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) 2767 dpa->dpa_aflags |= ARC_FLAG_L2CACHE; 2768 2769 /* 2770 * If we have the indirect just above us, no need to do the asynchronous 2771 * prefetch chain; we'll just run the last step ourselves. If we're at 2772 * a higher level, though, we want to issue the prefetches for all the 2773 * indirect blocks asynchronously, so we can go on with whatever we were 2774 * doing. 2775 */ 2776 if (curlevel == level) { 2777 ASSERT3U(curblkid, ==, blkid); 2778 dbuf_issue_final_prefetch(dpa, &bp); 2779 kmem_free(dpa, sizeof (*dpa)); 2780 } else { 2781 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2782 zbookmark_phys_t zb; 2783 2784 /* flag if L2ARC eligible, l2arc_noprefetch then decides */ 2785 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) 2786 iter_aflags |= ARC_FLAG_L2CACHE; 2787 2788 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2789 dn->dn_object, curlevel, curblkid); 2790 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2791 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2792 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2793 &iter_aflags, &zb); 2794 } 2795 /* 2796 * We use pio here instead of dpa_zio since it's possible that 2797 * dpa may have already been freed. 2798 */ 2799 zio_nowait(pio); 2800 } 2801 2802 /* 2803 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles 2804 * the case of encrypted, compressed and uncompressed buffers by 2805 * allocating the new buffer, respectively, with arc_alloc_raw_buf(), 2806 * arc_alloc_compressed_buf() or arc_alloc_buf().* 2807 * 2808 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl(). 2809 */ 2810 static void 2811 dbuf_hold_copy(dnode_t *dn, dmu_buf_impl_t *db, dbuf_dirty_record_t *dr) 2812 { 2813 arc_buf_t *data = dr->dt.dl.dr_data; 2814 enum zio_compress compress_type = arc_get_compression(data); 2815 2816 if (arc_is_encrypted(data)) { 2817 boolean_t byteorder; 2818 uint8_t salt[ZIO_DATA_SALT_LEN]; 2819 uint8_t iv[ZIO_DATA_IV_LEN]; 2820 uint8_t mac[ZIO_DATA_MAC_LEN]; 2821 2822 arc_get_raw_params(data, &byteorder, salt, iv, mac); 2823 dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db, 2824 dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac, 2825 dn->dn_type, arc_buf_size(data), arc_buf_lsize(data), 2826 compress_type)); 2827 } else if (compress_type != ZIO_COMPRESS_OFF) { 2828 dbuf_set_data(db, arc_alloc_compressed_buf( 2829 dn->dn_objset->os_spa, db, arc_buf_size(data), 2830 arc_buf_lsize(data), compress_type)); 2831 } else { 2832 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db, 2833 DBUF_GET_BUFC_TYPE(db), db->db.db_size)); 2834 } 2835 2836 bcopy(data->b_data, db->db.db_data, arc_buf_size(data)); 2837 } 2838 2839 /* 2840 * Returns with db_holds incremented, and db_mtx not held. 2841 * Note: dn_struct_rwlock must be held. 2842 */ 2843 int 2844 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2845 boolean_t fail_sparse, boolean_t fail_uncached, 2846 void *tag, dmu_buf_impl_t **dbp) 2847 { 2848 dmu_buf_impl_t *db, *parent = NULL; 2849 2850 ASSERT(blkid != DMU_BONUS_BLKID); 2851 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2852 ASSERT3U(dn->dn_nlevels, >, level); 2853 2854 *dbp = NULL; 2855 top: 2856 /* dbuf_find() returns with db_mtx held */ 2857 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2858 2859 if (db == NULL) { 2860 blkptr_t *bp = NULL; 2861 int err; 2862 2863 if (fail_uncached) 2864 return (SET_ERROR(ENOENT)); 2865 2866 ASSERT3P(parent, ==, NULL); 2867 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2868 if (fail_sparse) { 2869 if (err == 0 && bp && BP_IS_HOLE(bp)) 2870 err = SET_ERROR(ENOENT); 2871 if (err) { 2872 if (parent) 2873 dbuf_rele(parent, NULL); 2874 return (err); 2875 } 2876 } 2877 if (err && err != ENOENT) 2878 return (err); 2879 db = dbuf_create(dn, level, blkid, parent, bp); 2880 } 2881 2882 if (fail_uncached && db->db_state != DB_CACHED) { 2883 mutex_exit(&db->db_mtx); 2884 return (SET_ERROR(ENOENT)); 2885 } 2886 2887 if (db->db_buf != NULL) { 2888 arc_buf_access(db->db_buf); 2889 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2890 } 2891 2892 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2893 2894 /* 2895 * If this buffer is currently syncing out, and we are are 2896 * still referencing it from db_data, we need to make a copy 2897 * of it in case we decide we want to dirty it again in this txg. 2898 */ 2899 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2900 dn->dn_object != DMU_META_DNODE_OBJECT && 2901 db->db_state == DB_CACHED && db->db_data_pending) { 2902 dbuf_dirty_record_t *dr = db->db_data_pending; 2903 if (dr->dt.dl.dr_data == db->db_buf) 2904 dbuf_hold_copy(dn, db, dr); 2905 } 2906 2907 if (multilist_link_active(&db->db_cache_link)) { 2908 ASSERT(zfs_refcount_is_zero(&db->db_holds)); 2909 ASSERT(db->db_caching_status == DB_DBUF_CACHE || 2910 db->db_caching_status == DB_DBUF_METADATA_CACHE); 2911 2912 multilist_remove(dbuf_caches[db->db_caching_status].cache, db); 2913 (void) zfs_refcount_remove_many( 2914 &dbuf_caches[db->db_caching_status].size, 2915 db->db.db_size, db); 2916 2917 db->db_caching_status = DB_NO_CACHE; 2918 } 2919 (void) zfs_refcount_add(&db->db_holds, tag); 2920 DBUF_VERIFY(db); 2921 mutex_exit(&db->db_mtx); 2922 2923 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2924 if (parent) 2925 dbuf_rele(parent, NULL); 2926 2927 ASSERT3P(DB_DNODE(db), ==, dn); 2928 ASSERT3U(db->db_blkid, ==, blkid); 2929 ASSERT3U(db->db_level, ==, level); 2930 *dbp = db; 2931 2932 return (0); 2933 } 2934 2935 dmu_buf_impl_t * 2936 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2937 { 2938 return (dbuf_hold_level(dn, 0, blkid, tag)); 2939 } 2940 2941 dmu_buf_impl_t * 2942 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2943 { 2944 dmu_buf_impl_t *db; 2945 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2946 return (err ? NULL : db); 2947 } 2948 2949 void 2950 dbuf_create_bonus(dnode_t *dn) 2951 { 2952 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2953 2954 ASSERT(dn->dn_bonus == NULL); 2955 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2956 } 2957 2958 int 2959 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2960 { 2961 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2962 dnode_t *dn; 2963 2964 if (db->db_blkid != DMU_SPILL_BLKID) 2965 return (SET_ERROR(ENOTSUP)); 2966 if (blksz == 0) 2967 blksz = SPA_MINBLOCKSIZE; 2968 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2969 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2970 2971 DB_DNODE_ENTER(db); 2972 dn = DB_DNODE(db); 2973 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2974 dbuf_new_size(db, blksz, tx); 2975 rw_exit(&dn->dn_struct_rwlock); 2976 DB_DNODE_EXIT(db); 2977 2978 return (0); 2979 } 2980 2981 void 2982 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2983 { 2984 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2985 } 2986 2987 #pragma weak dmu_buf_add_ref = dbuf_add_ref 2988 void 2989 dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2990 { 2991 int64_t holds = zfs_refcount_add(&db->db_holds, tag); 2992 ASSERT3S(holds, >, 1); 2993 } 2994 2995 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2996 boolean_t 2997 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2998 void *tag) 2999 { 3000 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 3001 dmu_buf_impl_t *found_db; 3002 boolean_t result = B_FALSE; 3003 3004 if (db->db_blkid == DMU_BONUS_BLKID) 3005 found_db = dbuf_find_bonus(os, obj); 3006 else 3007 found_db = dbuf_find(os, obj, 0, blkid); 3008 3009 if (found_db != NULL) { 3010 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 3011 (void) zfs_refcount_add(&db->db_holds, tag); 3012 result = B_TRUE; 3013 } 3014 mutex_exit(&db->db_mtx); 3015 } 3016 return (result); 3017 } 3018 3019 /* 3020 * If you call dbuf_rele() you had better not be referencing the dnode handle 3021 * unless you have some other direct or indirect hold on the dnode. (An indirect 3022 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 3023 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 3024 * dnode's parent dbuf evicting its dnode handles. 3025 */ 3026 void 3027 dbuf_rele(dmu_buf_impl_t *db, void *tag) 3028 { 3029 mutex_enter(&db->db_mtx); 3030 dbuf_rele_and_unlock(db, tag, B_FALSE); 3031 } 3032 3033 void 3034 dmu_buf_rele(dmu_buf_t *db, void *tag) 3035 { 3036 dbuf_rele((dmu_buf_impl_t *)db, tag); 3037 } 3038 3039 /* 3040 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 3041 * db_dirtycnt and db_holds to be updated atomically. The 'evicting' 3042 * argument should be set if we are already in the dbuf-evicting code 3043 * path, in which case we don't want to recursively evict. This allows us to 3044 * avoid deeply nested stacks that would have a call flow similar to this: 3045 * 3046 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() 3047 * ^ | 3048 * | | 3049 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ 3050 * 3051 */ 3052 void 3053 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting) 3054 { 3055 int64_t holds; 3056 3057 ASSERT(MUTEX_HELD(&db->db_mtx)); 3058 DBUF_VERIFY(db); 3059 3060 /* 3061 * Remove the reference to the dbuf before removing its hold on the 3062 * dnode so we can guarantee in dnode_move() that a referenced bonus 3063 * buffer has a corresponding dnode hold. 3064 */ 3065 holds = zfs_refcount_remove(&db->db_holds, tag); 3066 ASSERT(holds >= 0); 3067 3068 /* 3069 * We can't freeze indirects if there is a possibility that they 3070 * may be modified in the current syncing context. 3071 */ 3072 if (db->db_buf != NULL && 3073 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { 3074 arc_buf_freeze(db->db_buf); 3075 } 3076 3077 if (holds == db->db_dirtycnt && 3078 db->db_level == 0 && db->db_user_immediate_evict) 3079 dbuf_evict_user(db); 3080 3081 if (holds == 0) { 3082 if (db->db_blkid == DMU_BONUS_BLKID) { 3083 dnode_t *dn; 3084 boolean_t evict_dbuf = db->db_pending_evict; 3085 3086 /* 3087 * If the dnode moves here, we cannot cross this 3088 * barrier until the move completes. 3089 */ 3090 DB_DNODE_ENTER(db); 3091 3092 dn = DB_DNODE(db); 3093 atomic_dec_32(&dn->dn_dbufs_count); 3094 3095 /* 3096 * Decrementing the dbuf count means that the bonus 3097 * buffer's dnode hold is no longer discounted in 3098 * dnode_move(). The dnode cannot move until after 3099 * the dnode_rele() below. 3100 */ 3101 DB_DNODE_EXIT(db); 3102 3103 /* 3104 * Do not reference db after its lock is dropped. 3105 * Another thread may evict it. 3106 */ 3107 mutex_exit(&db->db_mtx); 3108 3109 if (evict_dbuf) 3110 dnode_evict_bonus(dn); 3111 3112 dnode_rele(dn, db); 3113 } else if (db->db_buf == NULL) { 3114 /* 3115 * This is a special case: we never associated this 3116 * dbuf with any data allocated from the ARC. 3117 */ 3118 ASSERT(db->db_state == DB_UNCACHED || 3119 db->db_state == DB_NOFILL); 3120 dbuf_destroy(db); 3121 } else if (arc_released(db->db_buf)) { 3122 /* 3123 * This dbuf has anonymous data associated with it. 3124 */ 3125 dbuf_destroy(db); 3126 } else { 3127 boolean_t do_arc_evict = B_FALSE; 3128 blkptr_t bp; 3129 spa_t *spa = dmu_objset_spa(db->db_objset); 3130 3131 if (!DBUF_IS_CACHEABLE(db) && 3132 db->db_blkptr != NULL && 3133 !BP_IS_HOLE(db->db_blkptr) && 3134 !BP_IS_EMBEDDED(db->db_blkptr)) { 3135 do_arc_evict = B_TRUE; 3136 bp = *db->db_blkptr; 3137 } 3138 3139 if (!DBUF_IS_CACHEABLE(db) || 3140 db->db_pending_evict) { 3141 dbuf_destroy(db); 3142 } else if (!multilist_link_active(&db->db_cache_link)) { 3143 ASSERT3U(db->db_caching_status, ==, 3144 DB_NO_CACHE); 3145 3146 dbuf_cached_state_t dcs = 3147 dbuf_include_in_metadata_cache(db) ? 3148 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE; 3149 db->db_caching_status = dcs; 3150 3151 multilist_insert(dbuf_caches[dcs].cache, db); 3152 (void) zfs_refcount_add_many( 3153 &dbuf_caches[dcs].size, db->db.db_size, db); 3154 mutex_exit(&db->db_mtx); 3155 3156 if (db->db_caching_status == DB_DBUF_CACHE && 3157 !evicting) { 3158 dbuf_evict_notify(); 3159 } 3160 } 3161 3162 if (do_arc_evict) 3163 arc_freed(spa, &bp); 3164 } 3165 } else { 3166 mutex_exit(&db->db_mtx); 3167 } 3168 3169 } 3170 3171 #pragma weak dmu_buf_refcount = dbuf_refcount 3172 uint64_t 3173 dbuf_refcount(dmu_buf_impl_t *db) 3174 { 3175 return (zfs_refcount_count(&db->db_holds)); 3176 } 3177 3178 uint64_t 3179 dmu_buf_user_refcount(dmu_buf_t *db_fake) 3180 { 3181 uint64_t holds; 3182 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 3183 3184 mutex_enter(&db->db_mtx); 3185 ASSERT3U(zfs_refcount_count(&db->db_holds), >=, db->db_dirtycnt); 3186 holds = zfs_refcount_count(&db->db_holds) - db->db_dirtycnt; 3187 mutex_exit(&db->db_mtx); 3188 3189 return (holds); 3190 } 3191 3192 void * 3193 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 3194 dmu_buf_user_t *new_user) 3195 { 3196 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 3197 3198 mutex_enter(&db->db_mtx); 3199 dbuf_verify_user(db, DBVU_NOT_EVICTING); 3200 if (db->db_user == old_user) 3201 db->db_user = new_user; 3202 else 3203 old_user = db->db_user; 3204 dbuf_verify_user(db, DBVU_NOT_EVICTING); 3205 mutex_exit(&db->db_mtx); 3206 3207 return (old_user); 3208 } 3209 3210 void * 3211 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 3212 { 3213 return (dmu_buf_replace_user(db_fake, NULL, user)); 3214 } 3215 3216 void * 3217 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 3218 { 3219 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 3220 3221 db->db_user_immediate_evict = TRUE; 3222 return (dmu_buf_set_user(db_fake, user)); 3223 } 3224 3225 void * 3226 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 3227 { 3228 return (dmu_buf_replace_user(db_fake, user, NULL)); 3229 } 3230 3231 void * 3232 dmu_buf_get_user(dmu_buf_t *db_fake) 3233 { 3234 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 3235 3236 dbuf_verify_user(db, DBVU_NOT_EVICTING); 3237 return (db->db_user); 3238 } 3239 3240 void 3241 dmu_buf_user_evict_wait() 3242 { 3243 taskq_wait(dbu_evict_taskq); 3244 } 3245 3246 blkptr_t * 3247 dmu_buf_get_blkptr(dmu_buf_t *db) 3248 { 3249 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 3250 return (dbi->db_blkptr); 3251 } 3252 3253 objset_t * 3254 dmu_buf_get_objset(dmu_buf_t *db) 3255 { 3256 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 3257 return (dbi->db_objset); 3258 } 3259 3260 dnode_t * 3261 dmu_buf_dnode_enter(dmu_buf_t *db) 3262 { 3263 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 3264 DB_DNODE_ENTER(dbi); 3265 return (DB_DNODE(dbi)); 3266 } 3267 3268 void 3269 dmu_buf_dnode_exit(dmu_buf_t *db) 3270 { 3271 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 3272 DB_DNODE_EXIT(dbi); 3273 } 3274 3275 static void 3276 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 3277 { 3278 /* ASSERT(dmu_tx_is_syncing(tx) */ 3279 ASSERT(MUTEX_HELD(&db->db_mtx)); 3280 3281 if (db->db_blkptr != NULL) 3282 return; 3283 3284 if (db->db_blkid == DMU_SPILL_BLKID) { 3285 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys); 3286 BP_ZERO(db->db_blkptr); 3287 return; 3288 } 3289 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 3290 /* 3291 * This buffer was allocated at a time when there was 3292 * no available blkptrs from the dnode, or it was 3293 * inappropriate to hook it in (i.e., nlevels mis-match). 3294 */ 3295 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 3296 ASSERT(db->db_parent == NULL); 3297 db->db_parent = dn->dn_dbuf; 3298 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 3299 DBUF_VERIFY(db); 3300 } else { 3301 dmu_buf_impl_t *parent = db->db_parent; 3302 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3303 3304 ASSERT(dn->dn_phys->dn_nlevels > 1); 3305 if (parent == NULL) { 3306 mutex_exit(&db->db_mtx); 3307 rw_enter(&dn->dn_struct_rwlock, RW_READER); 3308 parent = dbuf_hold_level(dn, db->db_level + 1, 3309 db->db_blkid >> epbs, db); 3310 rw_exit(&dn->dn_struct_rwlock); 3311 mutex_enter(&db->db_mtx); 3312 db->db_parent = parent; 3313 } 3314 db->db_blkptr = (blkptr_t *)parent->db.db_data + 3315 (db->db_blkid & ((1ULL << epbs) - 1)); 3316 DBUF_VERIFY(db); 3317 } 3318 } 3319 3320 /* 3321 * When syncing out blocks of dnodes, adjust the block to deal with 3322 * encryption. Normally, we make sure the block is decrypted before writing 3323 * it. If we have crypt params, then we are writing a raw (encrypted) block, 3324 * from a raw receive. In this case, set the ARC buf's crypt params so 3325 * that the BP will be filled with the correct byteorder, salt, iv, and mac. 3326 * 3327 * XXX we should handle decrypting the dnode block in dbuf_dirty(). 3328 */ 3329 static void 3330 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr) 3331 { 3332 int err; 3333 dmu_buf_impl_t *db = dr->dr_dbuf; 3334 3335 ASSERT(MUTEX_HELD(&db->db_mtx)); 3336 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); 3337 ASSERT3U(db->db_level, ==, 0); 3338 3339 if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) { 3340 zbookmark_phys_t zb; 3341 3342 /* 3343 * Unfortunately, there is currently no mechanism for 3344 * syncing context to handle decryption errors. An error 3345 * here is only possible if an attacker maliciously 3346 * changed a dnode block and updated the associated 3347 * checksums going up the block tree. 3348 */ 3349 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), 3350 db->db.db_object, db->db_level, db->db_blkid); 3351 err = arc_untransform(db->db_buf, db->db_objset->os_spa, 3352 &zb, B_TRUE); 3353 if (err) 3354 panic("Invalid dnode block MAC"); 3355 } else if (dr->dt.dl.dr_has_raw_params) { 3356 (void) arc_release(dr->dt.dl.dr_data, db); 3357 arc_convert_to_raw(dr->dt.dl.dr_data, 3358 dmu_objset_id(db->db_objset), 3359 dr->dt.dl.dr_byteorder, DMU_OT_DNODE, 3360 dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac); 3361 } 3362 } 3363 3364 static void 3365 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3366 { 3367 dmu_buf_impl_t *db = dr->dr_dbuf; 3368 dnode_t *dn; 3369 zio_t *zio; 3370 3371 ASSERT(dmu_tx_is_syncing(tx)); 3372 3373 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3374 3375 mutex_enter(&db->db_mtx); 3376 3377 ASSERT(db->db_level > 0); 3378 DBUF_VERIFY(db); 3379 3380 /* Read the block if it hasn't been read yet. */ 3381 if (db->db_buf == NULL) { 3382 mutex_exit(&db->db_mtx); 3383 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 3384 mutex_enter(&db->db_mtx); 3385 } 3386 ASSERT3U(db->db_state, ==, DB_CACHED); 3387 ASSERT(db->db_buf != NULL); 3388 3389 DB_DNODE_ENTER(db); 3390 dn = DB_DNODE(db); 3391 /* Indirect block size must match what the dnode thinks it is. */ 3392 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3393 dbuf_check_blkptr(dn, db); 3394 DB_DNODE_EXIT(db); 3395 3396 /* Provide the pending dirty record to child dbufs */ 3397 db->db_data_pending = dr; 3398 3399 mutex_exit(&db->db_mtx); 3400 3401 dbuf_write(dr, db->db_buf, tx); 3402 3403 zio = dr->dr_zio; 3404 mutex_enter(&dr->dt.di.dr_mtx); 3405 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 3406 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3407 mutex_exit(&dr->dt.di.dr_mtx); 3408 zio_nowait(zio); 3409 } 3410 3411 static void 3412 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3413 { 3414 arc_buf_t **datap = &dr->dt.dl.dr_data; 3415 dmu_buf_impl_t *db = dr->dr_dbuf; 3416 dnode_t *dn; 3417 objset_t *os; 3418 uint64_t txg = tx->tx_txg; 3419 3420 ASSERT(dmu_tx_is_syncing(tx)); 3421 3422 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3423 3424 mutex_enter(&db->db_mtx); 3425 /* 3426 * To be synced, we must be dirtied. But we 3427 * might have been freed after the dirty. 3428 */ 3429 if (db->db_state == DB_UNCACHED) { 3430 /* This buffer has been freed since it was dirtied */ 3431 ASSERT(db->db.db_data == NULL); 3432 } else if (db->db_state == DB_FILL) { 3433 /* This buffer was freed and is now being re-filled */ 3434 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 3435 } else { 3436 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 3437 } 3438 DBUF_VERIFY(db); 3439 3440 DB_DNODE_ENTER(db); 3441 dn = DB_DNODE(db); 3442 3443 if (db->db_blkid == DMU_SPILL_BLKID) { 3444 mutex_enter(&dn->dn_mtx); 3445 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 3446 mutex_exit(&dn->dn_mtx); 3447 } 3448 3449 /* 3450 * If this is a bonus buffer, simply copy the bonus data into the 3451 * dnode. It will be written out when the dnode is synced (and it 3452 * will be synced, since it must have been dirty for dbuf_sync to 3453 * be called). 3454 */ 3455 if (db->db_blkid == DMU_BONUS_BLKID) { 3456 dbuf_dirty_record_t **drp; 3457 3458 ASSERT(*datap != NULL); 3459 ASSERT0(db->db_level); 3460 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=, 3461 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1)); 3462 bcopy(*datap, DN_BONUS(dn->dn_phys), 3463 DN_MAX_BONUS_LEN(dn->dn_phys)); 3464 DB_DNODE_EXIT(db); 3465 3466 if (*datap != db->db.db_data) { 3467 int slots = DB_DNODE(db)->dn_num_slots; 3468 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); 3469 zio_buf_free(*datap, bonuslen); 3470 arc_space_return(bonuslen, ARC_SPACE_BONUS); 3471 } 3472 db->db_data_pending = NULL; 3473 drp = &db->db_last_dirty; 3474 while (*drp != dr) 3475 drp = &(*drp)->dr_next; 3476 ASSERT(dr->dr_next == NULL); 3477 ASSERT(dr->dr_dbuf == db); 3478 *drp = dr->dr_next; 3479 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3480 ASSERT(db->db_dirtycnt > 0); 3481 db->db_dirtycnt -= 1; 3482 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE); 3483 return; 3484 } 3485 3486 os = dn->dn_objset; 3487 3488 /* 3489 * This function may have dropped the db_mtx lock allowing a dmu_sync 3490 * operation to sneak in. As a result, we need to ensure that we 3491 * don't check the dr_override_state until we have returned from 3492 * dbuf_check_blkptr. 3493 */ 3494 dbuf_check_blkptr(dn, db); 3495 3496 /* 3497 * If this buffer is in the middle of an immediate write, 3498 * wait for the synchronous IO to complete. 3499 */ 3500 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 3501 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 3502 cv_wait(&db->db_changed, &db->db_mtx); 3503 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 3504 } 3505 3506 /* 3507 * If this is a dnode block, ensure it is appropriately encrypted 3508 * or decrypted, depending on what we are writing to it this txg. 3509 */ 3510 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT) 3511 dbuf_prepare_encrypted_dnode_leaf(dr); 3512 3513 if (db->db_state != DB_NOFILL && 3514 dn->dn_object != DMU_META_DNODE_OBJECT && 3515 zfs_refcount_count(&db->db_holds) > 1 && 3516 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 3517 *datap == db->db_buf) { 3518 /* 3519 * If this buffer is currently "in use" (i.e., there 3520 * are active holds and db_data still references it), 3521 * then make a copy before we start the write so that 3522 * any modifications from the open txg will not leak 3523 * into this write. 3524 * 3525 * NOTE: this copy does not need to be made for 3526 * objects only modified in the syncing context (e.g. 3527 * DNONE_DNODE blocks). 3528 */ 3529 int psize = arc_buf_size(*datap); 3530 int lsize = arc_buf_lsize(*datap); 3531 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 3532 enum zio_compress compress_type = arc_get_compression(*datap); 3533 3534 if (arc_is_encrypted(*datap)) { 3535 boolean_t byteorder; 3536 uint8_t salt[ZIO_DATA_SALT_LEN]; 3537 uint8_t iv[ZIO_DATA_IV_LEN]; 3538 uint8_t mac[ZIO_DATA_MAC_LEN]; 3539 3540 arc_get_raw_params(*datap, &byteorder, salt, iv, mac); 3541 *datap = arc_alloc_raw_buf(os->os_spa, db, 3542 dmu_objset_id(os), byteorder, salt, iv, mac, 3543 dn->dn_type, psize, lsize, compress_type); 3544 } else if (compress_type != ZIO_COMPRESS_OFF) { 3545 ASSERT3U(type, ==, ARC_BUFC_DATA); 3546 *datap = arc_alloc_compressed_buf(os->os_spa, db, 3547 psize, lsize, compress_type); 3548 } else { 3549 *datap = arc_alloc_buf(os->os_spa, db, type, psize); 3550 } 3551 bcopy(db->db.db_data, (*datap)->b_data, psize); 3552 } 3553 db->db_data_pending = dr; 3554 3555 mutex_exit(&db->db_mtx); 3556 3557 dbuf_write(dr, *datap, tx); 3558 3559 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3560 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 3561 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 3562 DB_DNODE_EXIT(db); 3563 } else { 3564 /* 3565 * Although zio_nowait() does not "wait for an IO", it does 3566 * initiate the IO. If this is an empty write it seems plausible 3567 * that the IO could actually be completed before the nowait 3568 * returns. We need to DB_DNODE_EXIT() first in case 3569 * zio_nowait() invalidates the dbuf. 3570 */ 3571 DB_DNODE_EXIT(db); 3572 zio_nowait(dr->dr_zio); 3573 } 3574 } 3575 3576 void 3577 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 3578 { 3579 dbuf_dirty_record_t *dr; 3580 3581 while (dr = list_head(list)) { 3582 if (dr->dr_zio != NULL) { 3583 /* 3584 * If we find an already initialized zio then we 3585 * are processing the meta-dnode, and we have finished. 3586 * The dbufs for all dnodes are put back on the list 3587 * during processing, so that we can zio_wait() 3588 * these IOs after initiating all child IOs. 3589 */ 3590 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 3591 DMU_META_DNODE_OBJECT); 3592 break; 3593 } 3594 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 3595 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 3596 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 3597 } 3598 list_remove(list, dr); 3599 if (dr->dr_dbuf->db_level > 0) 3600 dbuf_sync_indirect(dr, tx); 3601 else 3602 dbuf_sync_leaf(dr, tx); 3603 } 3604 } 3605 3606 /* ARGSUSED */ 3607 static void 3608 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3609 { 3610 dmu_buf_impl_t *db = vdb; 3611 dnode_t *dn; 3612 blkptr_t *bp = zio->io_bp; 3613 blkptr_t *bp_orig = &zio->io_bp_orig; 3614 spa_t *spa = zio->io_spa; 3615 int64_t delta; 3616 uint64_t fill = 0; 3617 int i; 3618 3619 ASSERT3P(db->db_blkptr, !=, NULL); 3620 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); 3621 3622 DB_DNODE_ENTER(db); 3623 dn = DB_DNODE(db); 3624 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 3625 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 3626 zio->io_prev_space_delta = delta; 3627 3628 if (bp->blk_birth != 0) { 3629 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 3630 BP_GET_TYPE(bp) == dn->dn_type) || 3631 (db->db_blkid == DMU_SPILL_BLKID && 3632 BP_GET_TYPE(bp) == dn->dn_bonustype) || 3633 BP_IS_EMBEDDED(bp)); 3634 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 3635 } 3636 3637 mutex_enter(&db->db_mtx); 3638 3639 #ifdef ZFS_DEBUG 3640 if (db->db_blkid == DMU_SPILL_BLKID) { 3641 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3642 ASSERT(!(BP_IS_HOLE(bp)) && 3643 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); 3644 } 3645 #endif 3646 3647 if (db->db_level == 0) { 3648 mutex_enter(&dn->dn_mtx); 3649 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 3650 db->db_blkid != DMU_SPILL_BLKID) { 3651 ASSERT0(db->db_objset->os_raw_receive); 3652 dn->dn_phys->dn_maxblkid = db->db_blkid; 3653 } 3654 mutex_exit(&dn->dn_mtx); 3655 3656 if (dn->dn_type == DMU_OT_DNODE) { 3657 i = 0; 3658 while (i < db->db.db_size) { 3659 dnode_phys_t *dnp = 3660 (void *)(((char *)db->db.db_data) + i); 3661 3662 i += DNODE_MIN_SIZE; 3663 if (dnp->dn_type != DMU_OT_NONE) { 3664 fill++; 3665 i += dnp->dn_extra_slots * 3666 DNODE_MIN_SIZE; 3667 } 3668 } 3669 } else { 3670 if (BP_IS_HOLE(bp)) { 3671 fill = 0; 3672 } else { 3673 fill = 1; 3674 } 3675 } 3676 } else { 3677 blkptr_t *ibp = db->db.db_data; 3678 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3679 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 3680 if (BP_IS_HOLE(ibp)) 3681 continue; 3682 fill += BP_GET_FILL(ibp); 3683 } 3684 } 3685 DB_DNODE_EXIT(db); 3686 3687 if (!BP_IS_EMBEDDED(bp)) 3688 BP_SET_FILL(bp, fill); 3689 3690 mutex_exit(&db->db_mtx); 3691 3692 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3693 *db->db_blkptr = *bp; 3694 rw_exit(&dn->dn_struct_rwlock); 3695 } 3696 3697 /* ARGSUSED */ 3698 /* 3699 * This function gets called just prior to running through the compression 3700 * stage of the zio pipeline. If we're an indirect block comprised of only 3701 * holes, then we want this indirect to be compressed away to a hole. In 3702 * order to do that we must zero out any information about the holes that 3703 * this indirect points to prior to before we try to compress it. 3704 */ 3705 static void 3706 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3707 { 3708 dmu_buf_impl_t *db = vdb; 3709 dnode_t *dn; 3710 blkptr_t *bp; 3711 unsigned int epbs, i; 3712 3713 ASSERT3U(db->db_level, >, 0); 3714 DB_DNODE_ENTER(db); 3715 dn = DB_DNODE(db); 3716 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3717 ASSERT3U(epbs, <, 31); 3718 3719 /* Determine if all our children are holes */ 3720 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { 3721 if (!BP_IS_HOLE(bp)) 3722 break; 3723 } 3724 3725 /* 3726 * If all the children are holes, then zero them all out so that 3727 * we may get compressed away. 3728 */ 3729 if (i == 1 << epbs) { 3730 /* 3731 * We only found holes. Grab the rwlock to prevent 3732 * anybody from reading the blocks we're about to 3733 * zero out. 3734 */ 3735 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3736 bzero(db->db.db_data, db->db.db_size); 3737 rw_exit(&dn->dn_struct_rwlock); 3738 } 3739 DB_DNODE_EXIT(db); 3740 } 3741 3742 /* 3743 * The SPA will call this callback several times for each zio - once 3744 * for every physical child i/o (zio->io_phys_children times). This 3745 * allows the DMU to monitor the progress of each logical i/o. For example, 3746 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 3747 * block. There may be a long delay before all copies/fragments are completed, 3748 * so this callback allows us to retire dirty space gradually, as the physical 3749 * i/os complete. 3750 */ 3751 /* ARGSUSED */ 3752 static void 3753 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 3754 { 3755 dmu_buf_impl_t *db = arg; 3756 objset_t *os = db->db_objset; 3757 dsl_pool_t *dp = dmu_objset_pool(os); 3758 dbuf_dirty_record_t *dr; 3759 int delta = 0; 3760 3761 dr = db->db_data_pending; 3762 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 3763 3764 /* 3765 * The callback will be called io_phys_children times. Retire one 3766 * portion of our dirty space each time we are called. Any rounding 3767 * error will be cleaned up by dsl_pool_sync()'s call to 3768 * dsl_pool_undirty_space(). 3769 */ 3770 delta = dr->dr_accounted / zio->io_phys_children; 3771 dsl_pool_undirty_space(dp, delta, zio->io_txg); 3772 } 3773 3774 /* ARGSUSED */ 3775 static void 3776 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3777 { 3778 dmu_buf_impl_t *db = vdb; 3779 blkptr_t *bp_orig = &zio->io_bp_orig; 3780 blkptr_t *bp = db->db_blkptr; 3781 objset_t *os = db->db_objset; 3782 dmu_tx_t *tx = os->os_synctx; 3783 dbuf_dirty_record_t **drp, *dr; 3784 3785 ASSERT0(zio->io_error); 3786 ASSERT(db->db_blkptr == bp); 3787 3788 /* 3789 * For nopwrites and rewrites we ensure that the bp matches our 3790 * original and bypass all the accounting. 3791 */ 3792 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3793 ASSERT(BP_EQUAL(bp, bp_orig)); 3794 } else { 3795 dsl_dataset_t *ds = os->os_dsl_dataset; 3796 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3797 dsl_dataset_block_born(ds, bp, tx); 3798 } 3799 3800 mutex_enter(&db->db_mtx); 3801 3802 DBUF_VERIFY(db); 3803 3804 drp = &db->db_last_dirty; 3805 while ((dr = *drp) != db->db_data_pending) 3806 drp = &dr->dr_next; 3807 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3808 ASSERT(dr->dr_dbuf == db); 3809 ASSERT(dr->dr_next == NULL); 3810 *drp = dr->dr_next; 3811 3812 #ifdef ZFS_DEBUG 3813 if (db->db_blkid == DMU_SPILL_BLKID) { 3814 dnode_t *dn; 3815 3816 DB_DNODE_ENTER(db); 3817 dn = DB_DNODE(db); 3818 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3819 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3820 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); 3821 DB_DNODE_EXIT(db); 3822 } 3823 #endif 3824 3825 if (db->db_level == 0) { 3826 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3827 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3828 if (db->db_state != DB_NOFILL) { 3829 if (dr->dt.dl.dr_data != db->db_buf) 3830 arc_buf_destroy(dr->dt.dl.dr_data, db); 3831 } 3832 } else { 3833 dnode_t *dn; 3834 3835 DB_DNODE_ENTER(db); 3836 dn = DB_DNODE(db); 3837 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3838 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3839 if (!BP_IS_HOLE(db->db_blkptr)) { 3840 int epbs = 3841 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3842 ASSERT3U(db->db_blkid, <=, 3843 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3844 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3845 db->db.db_size); 3846 } 3847 DB_DNODE_EXIT(db); 3848 mutex_destroy(&dr->dt.di.dr_mtx); 3849 list_destroy(&dr->dt.di.dr_children); 3850 } 3851 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3852 3853 cv_broadcast(&db->db_changed); 3854 ASSERT(db->db_dirtycnt > 0); 3855 db->db_dirtycnt -= 1; 3856 db->db_data_pending = NULL; 3857 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE); 3858 } 3859 3860 static void 3861 dbuf_write_nofill_ready(zio_t *zio) 3862 { 3863 dbuf_write_ready(zio, NULL, zio->io_private); 3864 } 3865 3866 static void 3867 dbuf_write_nofill_done(zio_t *zio) 3868 { 3869 dbuf_write_done(zio, NULL, zio->io_private); 3870 } 3871 3872 static void 3873 dbuf_write_override_ready(zio_t *zio) 3874 { 3875 dbuf_dirty_record_t *dr = zio->io_private; 3876 dmu_buf_impl_t *db = dr->dr_dbuf; 3877 3878 dbuf_write_ready(zio, NULL, db); 3879 } 3880 3881 static void 3882 dbuf_write_override_done(zio_t *zio) 3883 { 3884 dbuf_dirty_record_t *dr = zio->io_private; 3885 dmu_buf_impl_t *db = dr->dr_dbuf; 3886 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3887 3888 mutex_enter(&db->db_mtx); 3889 if (!BP_EQUAL(zio->io_bp, obp)) { 3890 if (!BP_IS_HOLE(obp)) 3891 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3892 arc_release(dr->dt.dl.dr_data, db); 3893 } 3894 mutex_exit(&db->db_mtx); 3895 dbuf_write_done(zio, NULL, db); 3896 3897 if (zio->io_abd != NULL) 3898 abd_put(zio->io_abd); 3899 } 3900 3901 typedef struct dbuf_remap_impl_callback_arg { 3902 objset_t *drica_os; 3903 uint64_t drica_blk_birth; 3904 dmu_tx_t *drica_tx; 3905 } dbuf_remap_impl_callback_arg_t; 3906 3907 static void 3908 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size, 3909 void *arg) 3910 { 3911 dbuf_remap_impl_callback_arg_t *drica = arg; 3912 objset_t *os = drica->drica_os; 3913 spa_t *spa = dmu_objset_spa(os); 3914 dmu_tx_t *tx = drica->drica_tx; 3915 3916 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); 3917 3918 if (os == spa_meta_objset(spa)) { 3919 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx); 3920 } else { 3921 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset, 3922 size, drica->drica_blk_birth, tx); 3923 } 3924 } 3925 3926 static void 3927 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx) 3928 { 3929 blkptr_t bp_copy = *bp; 3930 spa_t *spa = dmu_objset_spa(dn->dn_objset); 3931 dbuf_remap_impl_callback_arg_t drica; 3932 3933 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); 3934 3935 drica.drica_os = dn->dn_objset; 3936 drica.drica_blk_birth = bp->blk_birth; 3937 drica.drica_tx = tx; 3938 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback, 3939 &drica)) { 3940 /* 3941 * The struct_rwlock prevents dbuf_read_impl() from 3942 * dereferencing the BP while we are changing it. To 3943 * avoid lock contention, only grab it when we are actually 3944 * changing the BP. 3945 */ 3946 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3947 *bp = bp_copy; 3948 rw_exit(&dn->dn_struct_rwlock); 3949 } 3950 } 3951 3952 /* 3953 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting 3954 * to remap a copy of every bp in the dbuf. 3955 */ 3956 boolean_t 3957 dbuf_can_remap(const dmu_buf_impl_t *db) 3958 { 3959 spa_t *spa = dmu_objset_spa(db->db_objset); 3960 blkptr_t *bp = db->db.db_data; 3961 boolean_t ret = B_FALSE; 3962 3963 ASSERT3U(db->db_level, >, 0); 3964 ASSERT3S(db->db_state, ==, DB_CACHED); 3965 3966 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)); 3967 3968 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); 3969 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { 3970 blkptr_t bp_copy = bp[i]; 3971 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) { 3972 ret = B_TRUE; 3973 break; 3974 } 3975 } 3976 spa_config_exit(spa, SCL_VDEV, FTAG); 3977 3978 return (ret); 3979 } 3980 3981 boolean_t 3982 dnode_needs_remap(const dnode_t *dn) 3983 { 3984 spa_t *spa = dmu_objset_spa(dn->dn_objset); 3985 boolean_t ret = B_FALSE; 3986 3987 if (dn->dn_phys->dn_nlevels == 0) { 3988 return (B_FALSE); 3989 } 3990 3991 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)); 3992 3993 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); 3994 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) { 3995 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j]; 3996 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) { 3997 ret = B_TRUE; 3998 break; 3999 } 4000 } 4001 spa_config_exit(spa, SCL_VDEV, FTAG); 4002 4003 return (ret); 4004 } 4005 4006 /* 4007 * Remap any existing BP's to concrete vdevs, if possible. 4008 */ 4009 static void 4010 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx) 4011 { 4012 spa_t *spa = dmu_objset_spa(db->db_objset); 4013 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); 4014 4015 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)) 4016 return; 4017 4018 if (db->db_level > 0) { 4019 blkptr_t *bp = db->db.db_data; 4020 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { 4021 dbuf_remap_impl(dn, &bp[i], tx); 4022 } 4023 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) { 4024 dnode_phys_t *dnp = db->db.db_data; 4025 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==, 4026 DMU_OT_DNODE); 4027 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; i++) { 4028 for (int j = 0; j < dnp[i].dn_nblkptr; j++) { 4029 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx); 4030 } 4031 } 4032 } 4033 } 4034 4035 4036 /* Issue I/O to commit a dirty buffer to disk. */ 4037 static void 4038 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 4039 { 4040 dmu_buf_impl_t *db = dr->dr_dbuf; 4041 dnode_t *dn; 4042 objset_t *os; 4043 dmu_buf_impl_t *parent = db->db_parent; 4044 uint64_t txg = tx->tx_txg; 4045 zbookmark_phys_t zb; 4046 zio_prop_t zp; 4047 zio_t *zio; 4048 int wp_flag = 0; 4049 4050 ASSERT(dmu_tx_is_syncing(tx)); 4051 4052 DB_DNODE_ENTER(db); 4053 dn = DB_DNODE(db); 4054 os = dn->dn_objset; 4055 4056 if (db->db_state != DB_NOFILL) { 4057 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 4058 /* 4059 * Private object buffers are released here rather 4060 * than in dbuf_dirty() since they are only modified 4061 * in the syncing context and we don't want the 4062 * overhead of making multiple copies of the data. 4063 */ 4064 if (BP_IS_HOLE(db->db_blkptr)) { 4065 arc_buf_thaw(data); 4066 } else { 4067 dbuf_release_bp(db); 4068 } 4069 dbuf_remap(dn, db, tx); 4070 } 4071 } 4072 4073 if (parent != dn->dn_dbuf) { 4074 /* Our parent is an indirect block. */ 4075 /* We have a dirty parent that has been scheduled for write. */ 4076 ASSERT(parent && parent->db_data_pending); 4077 /* Our parent's buffer is one level closer to the dnode. */ 4078 ASSERT(db->db_level == parent->db_level-1); 4079 /* 4080 * We're about to modify our parent's db_data by modifying 4081 * our block pointer, so the parent must be released. 4082 */ 4083 ASSERT(arc_released(parent->db_buf)); 4084 zio = parent->db_data_pending->dr_zio; 4085 } else { 4086 /* Our parent is the dnode itself. */ 4087 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 4088 db->db_blkid != DMU_SPILL_BLKID) || 4089 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 4090 if (db->db_blkid != DMU_SPILL_BLKID) 4091 ASSERT3P(db->db_blkptr, ==, 4092 &dn->dn_phys->dn_blkptr[db->db_blkid]); 4093 zio = dn->dn_zio; 4094 } 4095 4096 ASSERT(db->db_level == 0 || data == db->db_buf); 4097 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 4098 ASSERT(zio); 4099 4100 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 4101 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 4102 db->db.db_object, db->db_level, db->db_blkid); 4103 4104 if (db->db_blkid == DMU_SPILL_BLKID) 4105 wp_flag = WP_SPILL; 4106 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 4107 4108 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 4109 4110 DB_DNODE_EXIT(db); 4111 4112 /* 4113 * We copy the blkptr now (rather than when we instantiate the dirty 4114 * record), because its value can change between open context and 4115 * syncing context. We do not need to hold dn_struct_rwlock to read 4116 * db_blkptr because we are in syncing context. 4117 */ 4118 dr->dr_bp_copy = *db->db_blkptr; 4119 4120 if (db->db_level == 0 && 4121 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 4122 /* 4123 * The BP for this block has been provided by open context 4124 * (by dmu_sync() or dmu_buf_write_embedded()). 4125 */ 4126 abd_t *contents = (data != NULL) ? 4127 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; 4128 4129 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy, 4130 contents, db->db.db_size, db->db.db_size, &zp, 4131 dbuf_write_override_ready, NULL, NULL, 4132 dbuf_write_override_done, 4133 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 4134 mutex_enter(&db->db_mtx); 4135 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 4136 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 4137 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 4138 mutex_exit(&db->db_mtx); 4139 } else if (db->db_state == DB_NOFILL) { 4140 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 4141 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 4142 dr->dr_zio = zio_write(zio, os->os_spa, txg, 4143 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, 4144 dbuf_write_nofill_ready, NULL, NULL, 4145 dbuf_write_nofill_done, db, 4146 ZIO_PRIORITY_ASYNC_WRITE, 4147 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 4148 } else { 4149 ASSERT(arc_released(data)); 4150 4151 /* 4152 * For indirect blocks, we want to setup the children 4153 * ready callback so that we can properly handle an indirect 4154 * block that only contains holes. 4155 */ 4156 arc_write_done_func_t *children_ready_cb = NULL; 4157 if (db->db_level != 0) 4158 children_ready_cb = dbuf_write_children_ready; 4159 4160 dr->dr_zio = arc_write(zio, os->os_spa, txg, 4161 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), 4162 &zp, dbuf_write_ready, children_ready_cb, 4163 dbuf_write_physdone, dbuf_write_done, db, 4164 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 4165 } 4166 } 4167