dbuf.c revision 307292
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, 2016 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 50uint_t zfs_dbuf_evict_key; 51 52/* 53 * Number of times that zfs_free_range() took the slow path while doing 54 * a zfs receive. A nonzero value indicates a potential performance problem. 55 */ 56uint64_t zfs_free_range_recv_miss; 57 58static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 59static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 60 61/* 62 * Global data structures and functions for the dbuf cache. 63 */ 64static kmem_cache_t *dbuf_kmem_cache; 65static taskq_t *dbu_evict_taskq; 66 67static kthread_t *dbuf_cache_evict_thread; 68static kmutex_t dbuf_evict_lock; 69static kcondvar_t dbuf_evict_cv; 70static boolean_t dbuf_evict_thread_exit; 71 72/* 73 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that 74 * are not currently held but have been recently released. These dbufs 75 * are not eligible for arc eviction until they are aged out of the cache. 76 * Dbufs are added to the dbuf cache once the last hold is released. If a 77 * dbuf is later accessed and still exists in the dbuf cache, then it will 78 * be removed from the cache and later re-added to the head of the cache. 79 * Dbufs that are aged out of the cache will be immediately destroyed and 80 * become eligible for arc eviction. 81 */ 82static multilist_t dbuf_cache; 83static refcount_t dbuf_cache_size; 84uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024; 85 86/* Cap the size of the dbuf cache to log2 fraction of arc size. */ 87int dbuf_cache_max_shift = 5; 88 89/* 90 * The dbuf cache uses a three-stage eviction policy: 91 * - A low water marker designates when the dbuf eviction thread 92 * should stop evicting from the dbuf cache. 93 * - When we reach the maximum size (aka mid water mark), we 94 * signal the eviction thread to run. 95 * - The high water mark indicates when the eviction thread 96 * is unable to keep up with the incoming load and eviction must 97 * happen in the context of the calling thread. 98 * 99 * The dbuf cache: 100 * (max size) 101 * low water mid water hi water 102 * +----------------------------------------+----------+----------+ 103 * | | | | 104 * | | | | 105 * | | | | 106 * | | | | 107 * +----------------------------------------+----------+----------+ 108 * stop signal evict 109 * evicting eviction directly 110 * thread 111 * 112 * The high and low water marks indicate the operating range for the eviction 113 * thread. The low water mark is, by default, 90% of the total size of the 114 * cache and the high water mark is at 110% (both of these percentages can be 115 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, 116 * respectively). The eviction thread will try to ensure that the cache remains 117 * within this range by waking up every second and checking if the cache is 118 * above the low water mark. The thread can also be woken up by callers adding 119 * elements into the cache if the cache is larger than the mid water (i.e max 120 * cache size). Once the eviction thread is woken up and eviction is required, 121 * it will continue evicting buffers until it's able to reduce the cache size 122 * to the low water mark. If the cache size continues to grow and hits the high 123 * water mark, then callers adding elments to the cache will begin to evict 124 * directly from the cache until the cache is no longer above the high water 125 * mark. 126 */ 127 128/* 129 * The percentage above and below the maximum cache size. 130 */ 131uint_t dbuf_cache_hiwater_pct = 10; 132uint_t dbuf_cache_lowater_pct = 10; 133 134/* ARGSUSED */ 135static int 136dbuf_cons(void *vdb, void *unused, int kmflag) 137{ 138 dmu_buf_impl_t *db = vdb; 139 bzero(db, sizeof (dmu_buf_impl_t)); 140 141 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 142 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 143 multilist_link_init(&db->db_cache_link); 144 refcount_create(&db->db_holds); 145 146 return (0); 147} 148 149/* ARGSUSED */ 150static void 151dbuf_dest(void *vdb, void *unused) 152{ 153 dmu_buf_impl_t *db = vdb; 154 mutex_destroy(&db->db_mtx); 155 cv_destroy(&db->db_changed); 156 ASSERT(!multilist_link_active(&db->db_cache_link)); 157 refcount_destroy(&db->db_holds); 158} 159 160/* 161 * dbuf hash table routines 162 */ 163static dbuf_hash_table_t dbuf_hash_table; 164 165static uint64_t dbuf_hash_count; 166 167static uint64_t 168dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 169{ 170 uintptr_t osv = (uintptr_t)os; 171 uint64_t crc = -1ULL; 172 173 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 174 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 175 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 176 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 177 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 178 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 179 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 180 181 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 182 183 return (crc); 184} 185 186#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 187 ((dbuf)->db.db_object == (obj) && \ 188 (dbuf)->db_objset == (os) && \ 189 (dbuf)->db_level == (level) && \ 190 (dbuf)->db_blkid == (blkid)) 191 192dmu_buf_impl_t * 193dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 194{ 195 dbuf_hash_table_t *h = &dbuf_hash_table; 196 uint64_t hv = dbuf_hash(os, obj, level, blkid); 197 uint64_t idx = hv & h->hash_table_mask; 198 dmu_buf_impl_t *db; 199 200 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 201 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 202 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 203 mutex_enter(&db->db_mtx); 204 if (db->db_state != DB_EVICTING) { 205 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 206 return (db); 207 } 208 mutex_exit(&db->db_mtx); 209 } 210 } 211 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 212 return (NULL); 213} 214 215static dmu_buf_impl_t * 216dbuf_find_bonus(objset_t *os, uint64_t object) 217{ 218 dnode_t *dn; 219 dmu_buf_impl_t *db = NULL; 220 221 if (dnode_hold(os, object, FTAG, &dn) == 0) { 222 rw_enter(&dn->dn_struct_rwlock, RW_READER); 223 if (dn->dn_bonus != NULL) { 224 db = dn->dn_bonus; 225 mutex_enter(&db->db_mtx); 226 } 227 rw_exit(&dn->dn_struct_rwlock); 228 dnode_rele(dn, FTAG); 229 } 230 return (db); 231} 232 233/* 234 * Insert an entry into the hash table. If there is already an element 235 * equal to elem in the hash table, then the already existing element 236 * will be returned and the new element will not be inserted. 237 * Otherwise returns NULL. 238 */ 239static dmu_buf_impl_t * 240dbuf_hash_insert(dmu_buf_impl_t *db) 241{ 242 dbuf_hash_table_t *h = &dbuf_hash_table; 243 objset_t *os = db->db_objset; 244 uint64_t obj = db->db.db_object; 245 int level = db->db_level; 246 uint64_t blkid = db->db_blkid; 247 uint64_t hv = dbuf_hash(os, obj, level, blkid); 248 uint64_t idx = hv & h->hash_table_mask; 249 dmu_buf_impl_t *dbf; 250 251 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 252 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 253 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 254 mutex_enter(&dbf->db_mtx); 255 if (dbf->db_state != DB_EVICTING) { 256 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 257 return (dbf); 258 } 259 mutex_exit(&dbf->db_mtx); 260 } 261 } 262 263 mutex_enter(&db->db_mtx); 264 db->db_hash_next = h->hash_table[idx]; 265 h->hash_table[idx] = db; 266 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 267 atomic_inc_64(&dbuf_hash_count); 268 269 return (NULL); 270} 271 272/* 273 * Remove an entry from the hash table. It must be in the EVICTING state. 274 */ 275static void 276dbuf_hash_remove(dmu_buf_impl_t *db) 277{ 278 dbuf_hash_table_t *h = &dbuf_hash_table; 279 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object, 280 db->db_level, db->db_blkid); 281 uint64_t idx = hv & h->hash_table_mask; 282 dmu_buf_impl_t *dbf, **dbp; 283 284 /* 285 * We musn't hold db_mtx to maintain lock ordering: 286 * DBUF_HASH_MUTEX > db_mtx. 287 */ 288 ASSERT(refcount_is_zero(&db->db_holds)); 289 ASSERT(db->db_state == DB_EVICTING); 290 ASSERT(!MUTEX_HELD(&db->db_mtx)); 291 292 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 293 dbp = &h->hash_table[idx]; 294 while ((dbf = *dbp) != db) { 295 dbp = &dbf->db_hash_next; 296 ASSERT(dbf != NULL); 297 } 298 *dbp = db->db_hash_next; 299 db->db_hash_next = NULL; 300 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 301 atomic_dec_64(&dbuf_hash_count); 302} 303 304typedef enum { 305 DBVU_EVICTING, 306 DBVU_NOT_EVICTING 307} dbvu_verify_type_t; 308 309static void 310dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 311{ 312#ifdef ZFS_DEBUG 313 int64_t holds; 314 315 if (db->db_user == NULL) 316 return; 317 318 /* Only data blocks support the attachment of user data. */ 319 ASSERT(db->db_level == 0); 320 321 /* Clients must resolve a dbuf before attaching user data. */ 322 ASSERT(db->db.db_data != NULL); 323 ASSERT3U(db->db_state, ==, DB_CACHED); 324 325 holds = refcount_count(&db->db_holds); 326 if (verify_type == DBVU_EVICTING) { 327 /* 328 * Immediate eviction occurs when holds == dirtycnt. 329 * For normal eviction buffers, holds is zero on 330 * eviction, except when dbuf_fix_old_data() calls 331 * dbuf_clear_data(). However, the hold count can grow 332 * during eviction even though db_mtx is held (see 333 * dmu_bonus_hold() for an example), so we can only 334 * test the generic invariant that holds >= dirtycnt. 335 */ 336 ASSERT3U(holds, >=, db->db_dirtycnt); 337 } else { 338 if (db->db_user_immediate_evict == TRUE) 339 ASSERT3U(holds, >=, db->db_dirtycnt); 340 else 341 ASSERT3U(holds, >, 0); 342 } 343#endif 344} 345 346static void 347dbuf_evict_user(dmu_buf_impl_t *db) 348{ 349 dmu_buf_user_t *dbu = db->db_user; 350 351 ASSERT(MUTEX_HELD(&db->db_mtx)); 352 353 if (dbu == NULL) 354 return; 355 356 dbuf_verify_user(db, DBVU_EVICTING); 357 db->db_user = NULL; 358 359#ifdef ZFS_DEBUG 360 if (dbu->dbu_clear_on_evict_dbufp != NULL) 361 *dbu->dbu_clear_on_evict_dbufp = NULL; 362#endif 363 364 /* 365 * Invoke the callback from a taskq to avoid lock order reversals 366 * and limit stack depth. 367 */ 368 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, 369 &dbu->dbu_tqent); 370} 371 372boolean_t 373dbuf_is_metadata(dmu_buf_impl_t *db) 374{ 375 if (db->db_level > 0) { 376 return (B_TRUE); 377 } else { 378 boolean_t is_metadata; 379 380 DB_DNODE_ENTER(db); 381 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 382 DB_DNODE_EXIT(db); 383 384 return (is_metadata); 385 } 386} 387 388/* 389 * This function *must* return indices evenly distributed between all 390 * sublists of the multilist. This is needed due to how the dbuf eviction 391 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly 392 * distributed between all sublists and uses this assumption when 393 * deciding which sublist to evict from and how much to evict from it. 394 */ 395unsigned int 396dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) 397{ 398 dmu_buf_impl_t *db = obj; 399 400 /* 401 * The assumption here, is the hash value for a given 402 * dmu_buf_impl_t will remain constant throughout it's lifetime 403 * (i.e. it's objset, object, level and blkid fields don't change). 404 * Thus, we don't need to store the dbuf's sublist index 405 * on insertion, as this index can be recalculated on removal. 406 * 407 * Also, the low order bits of the hash value are thought to be 408 * distributed evenly. Otherwise, in the case that the multilist 409 * has a power of two number of sublists, each sublists' usage 410 * would not be evenly distributed. 411 */ 412 return (dbuf_hash(db->db_objset, db->db.db_object, 413 db->db_level, db->db_blkid) % 414 multilist_get_num_sublists(ml)); 415} 416 417static inline boolean_t 418dbuf_cache_above_hiwater(void) 419{ 420 uint64_t dbuf_cache_hiwater_bytes = 421 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100; 422 423 return (refcount_count(&dbuf_cache_size) > 424 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes); 425} 426 427static inline boolean_t 428dbuf_cache_above_lowater(void) 429{ 430 uint64_t dbuf_cache_lowater_bytes = 431 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100; 432 433 return (refcount_count(&dbuf_cache_size) > 434 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes); 435} 436 437/* 438 * Evict the oldest eligible dbuf from the dbuf cache. 439 */ 440static void 441dbuf_evict_one(void) 442{ 443 int idx = multilist_get_random_index(&dbuf_cache); 444 multilist_sublist_t *mls = multilist_sublist_lock(&dbuf_cache, idx); 445 446 ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); 447 448 /* 449 * Set the thread's tsd to indicate that it's processing evictions. 450 * Once a thread stops evicting from the dbuf cache it will 451 * reset its tsd to NULL. 452 */ 453 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL); 454 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE); 455 456 dmu_buf_impl_t *db = multilist_sublist_tail(mls); 457 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { 458 db = multilist_sublist_prev(mls, db); 459 } 460 461 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, 462 multilist_sublist_t *, mls); 463 464 if (db != NULL) { 465 multilist_sublist_remove(mls, db); 466 multilist_sublist_unlock(mls); 467 (void) refcount_remove_many(&dbuf_cache_size, 468 db->db.db_size, db); 469 dbuf_destroy(db); 470 } else { 471 multilist_sublist_unlock(mls); 472 } 473 (void) tsd_set(zfs_dbuf_evict_key, NULL); 474} 475 476/* 477 * The dbuf evict thread is responsible for aging out dbufs from the 478 * cache. Once the cache has reached it's maximum size, dbufs are removed 479 * and destroyed. The eviction thread will continue running until the size 480 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged 481 * out of the cache it is destroyed and becomes eligible for arc eviction. 482 */ 483static void 484dbuf_evict_thread(void *dummy __unused) 485{ 486 callb_cpr_t cpr; 487 488 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); 489 490 mutex_enter(&dbuf_evict_lock); 491 while (!dbuf_evict_thread_exit) { 492 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 493 CALLB_CPR_SAFE_BEGIN(&cpr); 494 (void) cv_timedwait_hires(&dbuf_evict_cv, 495 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 496 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); 497 } 498 mutex_exit(&dbuf_evict_lock); 499 500 /* 501 * Keep evicting as long as we're above the low water mark 502 * for the cache. We do this without holding the locks to 503 * minimize lock contention. 504 */ 505 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { 506 dbuf_evict_one(); 507 } 508 509 mutex_enter(&dbuf_evict_lock); 510 } 511 512 dbuf_evict_thread_exit = B_FALSE; 513 cv_broadcast(&dbuf_evict_cv); 514 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ 515 thread_exit(); 516} 517 518/* 519 * Wake up the dbuf eviction thread if the dbuf cache is at its max size. 520 * If the dbuf cache is at its high water mark, then evict a dbuf from the 521 * dbuf cache using the callers context. 522 */ 523static void 524dbuf_evict_notify(void) 525{ 526 527 /* 528 * We use thread specific data to track when a thread has 529 * started processing evictions. This allows us to avoid deeply 530 * nested stacks that would have a call flow similar to this: 531 * 532 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() 533 * ^ | 534 * | | 535 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ 536 * 537 * The dbuf_eviction_thread will always have its tsd set until 538 * that thread exits. All other threads will only set their tsd 539 * if they are participating in the eviction process. This only 540 * happens if the eviction thread is unable to process evictions 541 * fast enough. To keep the dbuf cache size in check, other threads 542 * can evict from the dbuf cache directly. Those threads will set 543 * their tsd values so that we ensure that they only evict one dbuf 544 * from the dbuf cache. 545 */ 546 if (tsd_get(zfs_dbuf_evict_key) != NULL) 547 return; 548 549 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) { 550 boolean_t evict_now = B_FALSE; 551 552 mutex_enter(&dbuf_evict_lock); 553 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) { 554 evict_now = dbuf_cache_above_hiwater(); 555 cv_signal(&dbuf_evict_cv); 556 } 557 mutex_exit(&dbuf_evict_lock); 558 559 if (evict_now) { 560 dbuf_evict_one(); 561 } 562 } 563} 564 565void 566dbuf_init(void) 567{ 568 uint64_t hsize = 1ULL << 16; 569 dbuf_hash_table_t *h = &dbuf_hash_table; 570 int i; 571 572 /* 573 * The hash table is big enough to fill all of physical memory 574 * with an average 4K block size. The table will take up 575 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 576 */ 577 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 578 hsize <<= 1; 579 580retry: 581 h->hash_table_mask = hsize - 1; 582 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 583 if (h->hash_table == NULL) { 584 /* XXX - we should really return an error instead of assert */ 585 ASSERT(hsize > (1ULL << 10)); 586 hsize >>= 1; 587 goto retry; 588 } 589 590 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", 591 sizeof (dmu_buf_impl_t), 592 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 593 594 for (i = 0; i < DBUF_MUTEXES; i++) 595 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 596 597 /* 598 * Setup the parameters for the dbuf cache. We cap the size of the 599 * dbuf cache to 1/32nd (default) of the size of the ARC. 600 */ 601 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes, 602 arc_max_bytes() >> dbuf_cache_max_shift); 603 604 /* 605 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 606 * configuration is not required. 607 */ 608 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 609 610 multilist_create(&dbuf_cache, sizeof (dmu_buf_impl_t), 611 offsetof(dmu_buf_impl_t, db_cache_link), 612 zfs_arc_num_sublists_per_state, 613 dbuf_cache_multilist_index_func); 614 refcount_create(&dbuf_cache_size); 615 616 tsd_create(&zfs_dbuf_evict_key, NULL); 617 dbuf_evict_thread_exit = B_FALSE; 618 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); 619 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); 620 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, 621 NULL, 0, &p0, TS_RUN, minclsyspri); 622} 623 624void 625dbuf_fini(void) 626{ 627 dbuf_hash_table_t *h = &dbuf_hash_table; 628 int i; 629 630 for (i = 0; i < DBUF_MUTEXES; i++) 631 mutex_destroy(&h->hash_mutexes[i]); 632 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 633 kmem_cache_destroy(dbuf_kmem_cache); 634 taskq_destroy(dbu_evict_taskq); 635 636 mutex_enter(&dbuf_evict_lock); 637 dbuf_evict_thread_exit = B_TRUE; 638 while (dbuf_evict_thread_exit) { 639 cv_signal(&dbuf_evict_cv); 640 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); 641 } 642 mutex_exit(&dbuf_evict_lock); 643 tsd_destroy(&zfs_dbuf_evict_key); 644 645 mutex_destroy(&dbuf_evict_lock); 646 cv_destroy(&dbuf_evict_cv); 647 648 refcount_destroy(&dbuf_cache_size); 649 multilist_destroy(&dbuf_cache); 650} 651 652/* 653 * Other stuff. 654 */ 655 656#ifdef ZFS_DEBUG 657static void 658dbuf_verify(dmu_buf_impl_t *db) 659{ 660 dnode_t *dn; 661 dbuf_dirty_record_t *dr; 662 663 ASSERT(MUTEX_HELD(&db->db_mtx)); 664 665 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 666 return; 667 668 ASSERT(db->db_objset != NULL); 669 DB_DNODE_ENTER(db); 670 dn = DB_DNODE(db); 671 if (dn == NULL) { 672 ASSERT(db->db_parent == NULL); 673 ASSERT(db->db_blkptr == NULL); 674 } else { 675 ASSERT3U(db->db.db_object, ==, dn->dn_object); 676 ASSERT3P(db->db_objset, ==, dn->dn_objset); 677 ASSERT3U(db->db_level, <, dn->dn_nlevels); 678 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 679 db->db_blkid == DMU_SPILL_BLKID || 680 !avl_is_empty(&dn->dn_dbufs)); 681 } 682 if (db->db_blkid == DMU_BONUS_BLKID) { 683 ASSERT(dn != NULL); 684 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 685 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 686 } else if (db->db_blkid == DMU_SPILL_BLKID) { 687 ASSERT(dn != NULL); 688 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 689 ASSERT0(db->db.db_offset); 690 } else { 691 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 692 } 693 694 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 695 ASSERT(dr->dr_dbuf == db); 696 697 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 698 ASSERT(dr->dr_dbuf == db); 699 700 /* 701 * We can't assert that db_size matches dn_datablksz because it 702 * can be momentarily different when another thread is doing 703 * dnode_set_blksz(). 704 */ 705 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 706 dr = db->db_data_pending; 707 /* 708 * It should only be modified in syncing context, so 709 * make sure we only have one copy of the data. 710 */ 711 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 712 } 713 714 /* verify db->db_blkptr */ 715 if (db->db_blkptr) { 716 if (db->db_parent == dn->dn_dbuf) { 717 /* db is pointed to by the dnode */ 718 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 719 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 720 ASSERT(db->db_parent == NULL); 721 else 722 ASSERT(db->db_parent != NULL); 723 if (db->db_blkid != DMU_SPILL_BLKID) 724 ASSERT3P(db->db_blkptr, ==, 725 &dn->dn_phys->dn_blkptr[db->db_blkid]); 726 } else { 727 /* db is pointed to by an indirect block */ 728 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 729 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 730 ASSERT3U(db->db_parent->db.db_object, ==, 731 db->db.db_object); 732 /* 733 * dnode_grow_indblksz() can make this fail if we don't 734 * have the struct_rwlock. XXX indblksz no longer 735 * grows. safe to do this now? 736 */ 737 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 738 ASSERT3P(db->db_blkptr, ==, 739 ((blkptr_t *)db->db_parent->db.db_data + 740 db->db_blkid % epb)); 741 } 742 } 743 } 744 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 745 (db->db_buf == NULL || db->db_buf->b_data) && 746 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 747 db->db_state != DB_FILL && !dn->dn_free_txg) { 748 /* 749 * If the blkptr isn't set but they have nonzero data, 750 * it had better be dirty, otherwise we'll lose that 751 * data when we evict this buffer. 752 * 753 * There is an exception to this rule for indirect blocks; in 754 * this case, if the indirect block is a hole, we fill in a few 755 * fields on each of the child blocks (importantly, birth time) 756 * to prevent hole birth times from being lost when you 757 * partially fill in a hole. 758 */ 759 if (db->db_dirtycnt == 0) { 760 if (db->db_level == 0) { 761 uint64_t *buf = db->db.db_data; 762 int i; 763 764 for (i = 0; i < db->db.db_size >> 3; i++) { 765 ASSERT(buf[i] == 0); 766 } 767 } else { 768 blkptr_t *bps = db->db.db_data; 769 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, 770 db->db.db_size); 771 /* 772 * We want to verify that all the blkptrs in the 773 * indirect block are holes, but we may have 774 * automatically set up a few fields for them. 775 * We iterate through each blkptr and verify 776 * they only have those fields set. 777 */ 778 for (int i = 0; 779 i < db->db.db_size / sizeof (blkptr_t); 780 i++) { 781 blkptr_t *bp = &bps[i]; 782 ASSERT(ZIO_CHECKSUM_IS_ZERO( 783 &bp->blk_cksum)); 784 ASSERT( 785 DVA_IS_EMPTY(&bp->blk_dva[0]) && 786 DVA_IS_EMPTY(&bp->blk_dva[1]) && 787 DVA_IS_EMPTY(&bp->blk_dva[2])); 788 ASSERT0(bp->blk_fill); 789 ASSERT0(bp->blk_pad[0]); 790 ASSERT0(bp->blk_pad[1]); 791 ASSERT(!BP_IS_EMBEDDED(bp)); 792 ASSERT(BP_IS_HOLE(bp)); 793 ASSERT0(bp->blk_phys_birth); 794 } 795 } 796 } 797 } 798 DB_DNODE_EXIT(db); 799} 800#endif 801 802static void 803dbuf_clear_data(dmu_buf_impl_t *db) 804{ 805 ASSERT(MUTEX_HELD(&db->db_mtx)); 806 dbuf_evict_user(db); 807 ASSERT3P(db->db_buf, ==, NULL); 808 db->db.db_data = NULL; 809 if (db->db_state != DB_NOFILL) 810 db->db_state = DB_UNCACHED; 811} 812 813static void 814dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 815{ 816 ASSERT(MUTEX_HELD(&db->db_mtx)); 817 ASSERT(buf != NULL); 818 819 db->db_buf = buf; 820 ASSERT(buf->b_data != NULL); 821 db->db.db_data = buf->b_data; 822} 823 824/* 825 * Loan out an arc_buf for read. Return the loaned arc_buf. 826 */ 827arc_buf_t * 828dbuf_loan_arcbuf(dmu_buf_impl_t *db) 829{ 830 arc_buf_t *abuf; 831 832 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 833 mutex_enter(&db->db_mtx); 834 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 835 int blksz = db->db.db_size; 836 spa_t *spa = db->db_objset->os_spa; 837 838 mutex_exit(&db->db_mtx); 839 abuf = arc_loan_buf(spa, blksz); 840 bcopy(db->db.db_data, abuf->b_data, blksz); 841 } else { 842 abuf = db->db_buf; 843 arc_loan_inuse_buf(abuf, db); 844 db->db_buf = NULL; 845 dbuf_clear_data(db); 846 mutex_exit(&db->db_mtx); 847 } 848 return (abuf); 849} 850 851/* 852 * Calculate which level n block references the data at the level 0 offset 853 * provided. 854 */ 855uint64_t 856dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 857{ 858 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 859 /* 860 * The level n blkid is equal to the level 0 blkid divided by 861 * the number of level 0s in a level n block. 862 * 863 * The level 0 blkid is offset >> datablkshift = 864 * offset / 2^datablkshift. 865 * 866 * The number of level 0s in a level n is the number of block 867 * pointers in an indirect block, raised to the power of level. 868 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 869 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 870 * 871 * Thus, the level n blkid is: offset / 872 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 873 * = offset / 2^(datablkshift + level * 874 * (indblkshift - SPA_BLKPTRSHIFT)) 875 * = offset >> (datablkshift + level * 876 * (indblkshift - SPA_BLKPTRSHIFT)) 877 */ 878 return (offset >> (dn->dn_datablkshift + level * 879 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 880 } else { 881 ASSERT3U(offset, <, dn->dn_datablksz); 882 return (0); 883 } 884} 885 886static void 887dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 888{ 889 dmu_buf_impl_t *db = vdb; 890 891 mutex_enter(&db->db_mtx); 892 ASSERT3U(db->db_state, ==, DB_READ); 893 /* 894 * All reads are synchronous, so we must have a hold on the dbuf 895 */ 896 ASSERT(refcount_count(&db->db_holds) > 0); 897 ASSERT(db->db_buf == NULL); 898 ASSERT(db->db.db_data == NULL); 899 if (db->db_level == 0 && db->db_freed_in_flight) { 900 /* we were freed in flight; disregard any error */ 901 arc_release(buf, db); 902 bzero(buf->b_data, db->db.db_size); 903 arc_buf_freeze(buf); 904 db->db_freed_in_flight = FALSE; 905 dbuf_set_data(db, buf); 906 db->db_state = DB_CACHED; 907 } else if (zio == NULL || zio->io_error == 0) { 908 dbuf_set_data(db, buf); 909 db->db_state = DB_CACHED; 910 } else { 911 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 912 ASSERT3P(db->db_buf, ==, NULL); 913 arc_buf_destroy(buf, db); 914 db->db_state = DB_UNCACHED; 915 } 916 cv_broadcast(&db->db_changed); 917 dbuf_rele_and_unlock(db, NULL); 918} 919 920static void 921dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 922{ 923 dnode_t *dn; 924 zbookmark_phys_t zb; 925 arc_flags_t aflags = ARC_FLAG_NOWAIT; 926 927 DB_DNODE_ENTER(db); 928 dn = DB_DNODE(db); 929 ASSERT(!refcount_is_zero(&db->db_holds)); 930 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 931 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 932 ASSERT(MUTEX_HELD(&db->db_mtx)); 933 ASSERT(db->db_state == DB_UNCACHED); 934 ASSERT(db->db_buf == NULL); 935 936 if (db->db_blkid == DMU_BONUS_BLKID) { 937 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 938 939 ASSERT3U(bonuslen, <=, db->db.db_size); 940 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 941 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 942 if (bonuslen < DN_MAX_BONUSLEN) 943 bzero(db->db.db_data, DN_MAX_BONUSLEN); 944 if (bonuslen) 945 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 946 DB_DNODE_EXIT(db); 947 db->db_state = DB_CACHED; 948 mutex_exit(&db->db_mtx); 949 return; 950 } 951 952 /* 953 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 954 * processes the delete record and clears the bp while we are waiting 955 * for the dn_mtx (resulting in a "no" from block_freed). 956 */ 957 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 958 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 959 BP_IS_HOLE(db->db_blkptr)))) { 960 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 961 962 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, 963 db->db.db_size, db, type)); 964 bzero(db->db.db_data, db->db.db_size); 965 966 if (db->db_blkptr != NULL && db->db_level > 0 && 967 BP_IS_HOLE(db->db_blkptr) && 968 db->db_blkptr->blk_birth != 0) { 969 blkptr_t *bps = db->db.db_data; 970 for (int i = 0; i < ((1 << 971 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); 972 i++) { 973 blkptr_t *bp = &bps[i]; 974 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 975 1 << dn->dn_indblkshift); 976 BP_SET_LSIZE(bp, 977 BP_GET_LEVEL(db->db_blkptr) == 1 ? 978 dn->dn_datablksz : 979 BP_GET_LSIZE(db->db_blkptr)); 980 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); 981 BP_SET_LEVEL(bp, 982 BP_GET_LEVEL(db->db_blkptr) - 1); 983 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); 984 } 985 } 986 DB_DNODE_EXIT(db); 987 db->db_state = DB_CACHED; 988 mutex_exit(&db->db_mtx); 989 return; 990 } 991 992 DB_DNODE_EXIT(db); 993 994 db->db_state = DB_READ; 995 mutex_exit(&db->db_mtx); 996 997 if (DBUF_IS_L2CACHEABLE(db)) 998 aflags |= ARC_FLAG_L2CACHE; 999 1000 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 1001 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 1002 db->db.db_object, db->db_level, db->db_blkid); 1003 1004 dbuf_add_ref(db, NULL); 1005 1006 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 1007 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 1008 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 1009 &aflags, &zb); 1010} 1011 1012int 1013dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 1014{ 1015 int err = 0; 1016 boolean_t havepzio = (zio != NULL); 1017 boolean_t prefetch; 1018 dnode_t *dn; 1019 1020 /* 1021 * We don't have to hold the mutex to check db_state because it 1022 * can't be freed while we have a hold on the buffer. 1023 */ 1024 ASSERT(!refcount_is_zero(&db->db_holds)); 1025 1026 if (db->db_state == DB_NOFILL) 1027 return (SET_ERROR(EIO)); 1028 1029 DB_DNODE_ENTER(db); 1030 dn = DB_DNODE(db); 1031 if ((flags & DB_RF_HAVESTRUCT) == 0) 1032 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1033 1034 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1035 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 1036 DBUF_IS_CACHEABLE(db); 1037 1038 mutex_enter(&db->db_mtx); 1039 if (db->db_state == DB_CACHED) { 1040 mutex_exit(&db->db_mtx); 1041 if (prefetch) 1042 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1043 if ((flags & DB_RF_HAVESTRUCT) == 0) 1044 rw_exit(&dn->dn_struct_rwlock); 1045 DB_DNODE_EXIT(db); 1046 } else if (db->db_state == DB_UNCACHED) { 1047 spa_t *spa = dn->dn_objset->os_spa; 1048 1049 if (zio == NULL) 1050 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 1051 dbuf_read_impl(db, zio, flags); 1052 1053 /* dbuf_read_impl has dropped db_mtx for us */ 1054 1055 if (prefetch) 1056 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1057 1058 if ((flags & DB_RF_HAVESTRUCT) == 0) 1059 rw_exit(&dn->dn_struct_rwlock); 1060 DB_DNODE_EXIT(db); 1061 1062 if (!havepzio) 1063 err = zio_wait(zio); 1064 } else { 1065 /* 1066 * Another reader came in while the dbuf was in flight 1067 * between UNCACHED and CACHED. Either a writer will finish 1068 * writing the buffer (sending the dbuf to CACHED) or the 1069 * first reader's request will reach the read_done callback 1070 * and send the dbuf to CACHED. Otherwise, a failure 1071 * occurred and the dbuf went to UNCACHED. 1072 */ 1073 mutex_exit(&db->db_mtx); 1074 if (prefetch) 1075 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 1076 if ((flags & DB_RF_HAVESTRUCT) == 0) 1077 rw_exit(&dn->dn_struct_rwlock); 1078 DB_DNODE_EXIT(db); 1079 1080 /* Skip the wait per the caller's request. */ 1081 mutex_enter(&db->db_mtx); 1082 if ((flags & DB_RF_NEVERWAIT) == 0) { 1083 while (db->db_state == DB_READ || 1084 db->db_state == DB_FILL) { 1085 ASSERT(db->db_state == DB_READ || 1086 (flags & DB_RF_HAVESTRUCT) == 0); 1087 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 1088 db, zio_t *, zio); 1089 cv_wait(&db->db_changed, &db->db_mtx); 1090 } 1091 if (db->db_state == DB_UNCACHED) 1092 err = SET_ERROR(EIO); 1093 } 1094 mutex_exit(&db->db_mtx); 1095 } 1096 1097 ASSERT(err || havepzio || db->db_state == DB_CACHED); 1098 return (err); 1099} 1100 1101static void 1102dbuf_noread(dmu_buf_impl_t *db) 1103{ 1104 ASSERT(!refcount_is_zero(&db->db_holds)); 1105 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1106 mutex_enter(&db->db_mtx); 1107 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1108 cv_wait(&db->db_changed, &db->db_mtx); 1109 if (db->db_state == DB_UNCACHED) { 1110 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1111 spa_t *spa = db->db_objset->os_spa; 1112 1113 ASSERT(db->db_buf == NULL); 1114 ASSERT(db->db.db_data == NULL); 1115 dbuf_set_data(db, arc_alloc_buf(spa, db->db.db_size, db, type)); 1116 db->db_state = DB_FILL; 1117 } else if (db->db_state == DB_NOFILL) { 1118 dbuf_clear_data(db); 1119 } else { 1120 ASSERT3U(db->db_state, ==, DB_CACHED); 1121 } 1122 mutex_exit(&db->db_mtx); 1123} 1124 1125/* 1126 * This is our just-in-time copy function. It makes a copy of 1127 * buffers, that have been modified in a previous transaction 1128 * group, before we modify them in the current active group. 1129 * 1130 * This function is used in two places: when we are dirtying a 1131 * buffer for the first time in a txg, and when we are freeing 1132 * a range in a dnode that includes this buffer. 1133 * 1134 * Note that when we are called from dbuf_free_range() we do 1135 * not put a hold on the buffer, we just traverse the active 1136 * dbuf list for the dnode. 1137 */ 1138static void 1139dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 1140{ 1141 dbuf_dirty_record_t *dr = db->db_last_dirty; 1142 1143 ASSERT(MUTEX_HELD(&db->db_mtx)); 1144 ASSERT(db->db.db_data != NULL); 1145 ASSERT(db->db_level == 0); 1146 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 1147 1148 if (dr == NULL || 1149 (dr->dt.dl.dr_data != 1150 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 1151 return; 1152 1153 /* 1154 * If the last dirty record for this dbuf has not yet synced 1155 * and its referencing the dbuf data, either: 1156 * reset the reference to point to a new copy, 1157 * or (if there a no active holders) 1158 * just null out the current db_data pointer. 1159 */ 1160 ASSERT(dr->dr_txg >= txg - 2); 1161 if (db->db_blkid == DMU_BONUS_BLKID) { 1162 /* Note that the data bufs here are zio_bufs */ 1163 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 1164 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1165 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 1166 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1167 int size = db->db.db_size; 1168 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1169 spa_t *spa = db->db_objset->os_spa; 1170 1171 dr->dt.dl.dr_data = arc_alloc_buf(spa, size, db, type); 1172 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 1173 } else { 1174 db->db_buf = NULL; 1175 dbuf_clear_data(db); 1176 } 1177} 1178 1179void 1180dbuf_unoverride(dbuf_dirty_record_t *dr) 1181{ 1182 dmu_buf_impl_t *db = dr->dr_dbuf; 1183 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 1184 uint64_t txg = dr->dr_txg; 1185 1186 ASSERT(MUTEX_HELD(&db->db_mtx)); 1187 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 1188 ASSERT(db->db_level == 0); 1189 1190 if (db->db_blkid == DMU_BONUS_BLKID || 1191 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 1192 return; 1193 1194 ASSERT(db->db_data_pending != dr); 1195 1196 /* free this block */ 1197 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 1198 zio_free(db->db_objset->os_spa, txg, bp); 1199 1200 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1201 dr->dt.dl.dr_nopwrite = B_FALSE; 1202 1203 /* 1204 * Release the already-written buffer, so we leave it in 1205 * a consistent dirty state. Note that all callers are 1206 * modifying the buffer, so they will immediately do 1207 * another (redundant) arc_release(). Therefore, leave 1208 * the buf thawed to save the effort of freezing & 1209 * immediately re-thawing it. 1210 */ 1211 arc_release(dr->dt.dl.dr_data, db); 1212} 1213 1214/* 1215 * Evict (if its unreferenced) or clear (if its referenced) any level-0 1216 * data blocks in the free range, so that any future readers will find 1217 * empty blocks. 1218 * 1219 * This is a no-op if the dataset is in the middle of an incremental 1220 * receive; see comment below for details. 1221 */ 1222void 1223dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 1224 dmu_tx_t *tx) 1225{ 1226 dmu_buf_impl_t db_search; 1227 dmu_buf_impl_t *db, *db_next; 1228 uint64_t txg = tx->tx_txg; 1229 avl_index_t where; 1230 1231 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 1232 end_blkid = dn->dn_maxblkid; 1233 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 1234 1235 db_search.db_level = 0; 1236 db_search.db_blkid = start_blkid; 1237 db_search.db_state = DB_SEARCH; 1238 1239 mutex_enter(&dn->dn_dbufs_mtx); 1240 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 1241 /* There can't be any dbufs in this range; no need to search. */ 1242#ifdef DEBUG 1243 db = avl_find(&dn->dn_dbufs, &db_search, &where); 1244 ASSERT3P(db, ==, NULL); 1245 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 1246 ASSERT(db == NULL || db->db_level > 0); 1247#endif 1248 mutex_exit(&dn->dn_dbufs_mtx); 1249 return; 1250 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 1251 /* 1252 * If we are receiving, we expect there to be no dbufs in 1253 * the range to be freed, because receive modifies each 1254 * block at most once, and in offset order. If this is 1255 * not the case, it can lead to performance problems, 1256 * so note that we unexpectedly took the slow path. 1257 */ 1258 atomic_inc_64(&zfs_free_range_recv_miss); 1259 } 1260 1261 db = avl_find(&dn->dn_dbufs, &db_search, &where); 1262 ASSERT3P(db, ==, NULL); 1263 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 1264 1265 for (; db != NULL; db = db_next) { 1266 db_next = AVL_NEXT(&dn->dn_dbufs, db); 1267 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1268 1269 if (db->db_level != 0 || db->db_blkid > end_blkid) { 1270 break; 1271 } 1272 ASSERT3U(db->db_blkid, >=, start_blkid); 1273 1274 /* found a level 0 buffer in the range */ 1275 mutex_enter(&db->db_mtx); 1276 if (dbuf_undirty(db, tx)) { 1277 /* mutex has been dropped and dbuf destroyed */ 1278 continue; 1279 } 1280 1281 if (db->db_state == DB_UNCACHED || 1282 db->db_state == DB_NOFILL || 1283 db->db_state == DB_EVICTING) { 1284 ASSERT(db->db.db_data == NULL); 1285 mutex_exit(&db->db_mtx); 1286 continue; 1287 } 1288 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 1289 /* will be handled in dbuf_read_done or dbuf_rele */ 1290 db->db_freed_in_flight = TRUE; 1291 mutex_exit(&db->db_mtx); 1292 continue; 1293 } 1294 if (refcount_count(&db->db_holds) == 0) { 1295 ASSERT(db->db_buf); 1296 dbuf_destroy(db); 1297 continue; 1298 } 1299 /* The dbuf is referenced */ 1300 1301 if (db->db_last_dirty != NULL) { 1302 dbuf_dirty_record_t *dr = db->db_last_dirty; 1303 1304 if (dr->dr_txg == txg) { 1305 /* 1306 * This buffer is "in-use", re-adjust the file 1307 * size to reflect that this buffer may 1308 * contain new data when we sync. 1309 */ 1310 if (db->db_blkid != DMU_SPILL_BLKID && 1311 db->db_blkid > dn->dn_maxblkid) 1312 dn->dn_maxblkid = db->db_blkid; 1313 dbuf_unoverride(dr); 1314 } else { 1315 /* 1316 * This dbuf is not dirty in the open context. 1317 * Either uncache it (if its not referenced in 1318 * the open context) or reset its contents to 1319 * empty. 1320 */ 1321 dbuf_fix_old_data(db, txg); 1322 } 1323 } 1324 /* clear the contents if its cached */ 1325 if (db->db_state == DB_CACHED) { 1326 ASSERT(db->db.db_data != NULL); 1327 arc_release(db->db_buf, db); 1328 bzero(db->db.db_data, db->db.db_size); 1329 arc_buf_freeze(db->db_buf); 1330 } 1331 1332 mutex_exit(&db->db_mtx); 1333 } 1334 mutex_exit(&dn->dn_dbufs_mtx); 1335} 1336 1337static int 1338dbuf_block_freeable(dmu_buf_impl_t *db) 1339{ 1340 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 1341 uint64_t birth_txg = 0; 1342 1343 /* 1344 * We don't need any locking to protect db_blkptr: 1345 * If it's syncing, then db_last_dirty will be set 1346 * so we'll ignore db_blkptr. 1347 * 1348 * This logic ensures that only block births for 1349 * filled blocks are considered. 1350 */ 1351 ASSERT(MUTEX_HELD(&db->db_mtx)); 1352 if (db->db_last_dirty && (db->db_blkptr == NULL || 1353 !BP_IS_HOLE(db->db_blkptr))) { 1354 birth_txg = db->db_last_dirty->dr_txg; 1355 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1356 birth_txg = db->db_blkptr->blk_birth; 1357 } 1358 1359 /* 1360 * If this block don't exist or is in a snapshot, it can't be freed. 1361 * Don't pass the bp to dsl_dataset_block_freeable() since we 1362 * are holding the db_mtx lock and might deadlock if we are 1363 * prefetching a dedup-ed block. 1364 */ 1365 if (birth_txg != 0) 1366 return (ds == NULL || 1367 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 1368 else 1369 return (B_FALSE); 1370} 1371 1372void 1373dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1374{ 1375 arc_buf_t *buf, *obuf; 1376 int osize = db->db.db_size; 1377 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1378 dnode_t *dn; 1379 1380 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1381 1382 DB_DNODE_ENTER(db); 1383 dn = DB_DNODE(db); 1384 1385 /* XXX does *this* func really need the lock? */ 1386 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1387 1388 /* 1389 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1390 * is OK, because there can be no other references to the db 1391 * when we are changing its size, so no concurrent DB_FILL can 1392 * be happening. 1393 */ 1394 /* 1395 * XXX we should be doing a dbuf_read, checking the return 1396 * value and returning that up to our callers 1397 */ 1398 dmu_buf_will_dirty(&db->db, tx); 1399 1400 /* create the data buffer for the new block */ 1401 buf = arc_alloc_buf(dn->dn_objset->os_spa, size, db, type); 1402 1403 /* copy old block data to the new block */ 1404 obuf = db->db_buf; 1405 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1406 /* zero the remainder */ 1407 if (size > osize) 1408 bzero((uint8_t *)buf->b_data + osize, size - osize); 1409 1410 mutex_enter(&db->db_mtx); 1411 dbuf_set_data(db, buf); 1412 arc_buf_destroy(obuf, db); 1413 db->db.db_size = size; 1414 1415 if (db->db_level == 0) { 1416 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1417 db->db_last_dirty->dt.dl.dr_data = buf; 1418 } 1419 mutex_exit(&db->db_mtx); 1420 1421 dnode_willuse_space(dn, size-osize, tx); 1422 DB_DNODE_EXIT(db); 1423} 1424 1425void 1426dbuf_release_bp(dmu_buf_impl_t *db) 1427{ 1428 objset_t *os = db->db_objset; 1429 1430 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1431 ASSERT(arc_released(os->os_phys_buf) || 1432 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1433 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1434 1435 (void) arc_release(db->db_buf, db); 1436} 1437 1438/* 1439 * We already have a dirty record for this TXG, and we are being 1440 * dirtied again. 1441 */ 1442static void 1443dbuf_redirty(dbuf_dirty_record_t *dr) 1444{ 1445 dmu_buf_impl_t *db = dr->dr_dbuf; 1446 1447 ASSERT(MUTEX_HELD(&db->db_mtx)); 1448 1449 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1450 /* 1451 * If this buffer has already been written out, 1452 * we now need to reset its state. 1453 */ 1454 dbuf_unoverride(dr); 1455 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1456 db->db_state != DB_NOFILL) { 1457 /* Already released on initial dirty, so just thaw. */ 1458 ASSERT(arc_released(db->db_buf)); 1459 arc_buf_thaw(db->db_buf); 1460 } 1461 } 1462} 1463 1464dbuf_dirty_record_t * 1465dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1466{ 1467 dnode_t *dn; 1468 objset_t *os; 1469 dbuf_dirty_record_t **drp, *dr; 1470 int drop_struct_lock = FALSE; 1471 boolean_t do_free_accounting = B_FALSE; 1472 int txgoff = tx->tx_txg & TXG_MASK; 1473 1474 ASSERT(tx->tx_txg != 0); 1475 ASSERT(!refcount_is_zero(&db->db_holds)); 1476 DMU_TX_DIRTY_BUF(tx, db); 1477 1478 DB_DNODE_ENTER(db); 1479 dn = DB_DNODE(db); 1480 /* 1481 * Shouldn't dirty a regular buffer in syncing context. Private 1482 * objects may be dirtied in syncing context, but only if they 1483 * were already pre-dirtied in open context. 1484 */ 1485 ASSERT(!dmu_tx_is_syncing(tx) || 1486 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1487 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1488 dn->dn_objset->os_dsl_dataset == NULL); 1489 /* 1490 * We make this assert for private objects as well, but after we 1491 * check if we're already dirty. They are allowed to re-dirty 1492 * in syncing context. 1493 */ 1494 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1495 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1496 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1497 1498 mutex_enter(&db->db_mtx); 1499 /* 1500 * XXX make this true for indirects too? The problem is that 1501 * transactions created with dmu_tx_create_assigned() from 1502 * syncing context don't bother holding ahead. 1503 */ 1504 ASSERT(db->db_level != 0 || 1505 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1506 db->db_state == DB_NOFILL); 1507 1508 mutex_enter(&dn->dn_mtx); 1509 /* 1510 * Don't set dirtyctx to SYNC if we're just modifying this as we 1511 * initialize the objset. 1512 */ 1513 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1514 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1515 dn->dn_dirtyctx = 1516 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1517 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1518 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1519 } 1520 mutex_exit(&dn->dn_mtx); 1521 1522 if (db->db_blkid == DMU_SPILL_BLKID) 1523 dn->dn_have_spill = B_TRUE; 1524 1525 /* 1526 * If this buffer is already dirty, we're done. 1527 */ 1528 drp = &db->db_last_dirty; 1529 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1530 db->db.db_object == DMU_META_DNODE_OBJECT); 1531 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1532 drp = &dr->dr_next; 1533 if (dr && dr->dr_txg == tx->tx_txg) { 1534 DB_DNODE_EXIT(db); 1535 1536 dbuf_redirty(dr); 1537 mutex_exit(&db->db_mtx); 1538 return (dr); 1539 } 1540 1541 /* 1542 * Only valid if not already dirty. 1543 */ 1544 ASSERT(dn->dn_object == 0 || 1545 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1546 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1547 1548 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1549 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1550 dn->dn_phys->dn_nlevels > db->db_level || 1551 dn->dn_next_nlevels[txgoff] > db->db_level || 1552 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1553 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1554 1555 /* 1556 * We should only be dirtying in syncing context if it's the 1557 * mos or we're initializing the os or it's a special object. 1558 * However, we are allowed to dirty in syncing context provided 1559 * we already dirtied it in open context. Hence we must make 1560 * this assertion only if we're not already dirty. 1561 */ 1562 os = dn->dn_objset; 1563 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1564 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1565 ASSERT(db->db.db_size != 0); 1566 1567 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1568 1569 if (db->db_blkid != DMU_BONUS_BLKID) { 1570 /* 1571 * Update the accounting. 1572 * Note: we delay "free accounting" until after we drop 1573 * the db_mtx. This keeps us from grabbing other locks 1574 * (and possibly deadlocking) in bp_get_dsize() while 1575 * also holding the db_mtx. 1576 */ 1577 dnode_willuse_space(dn, db->db.db_size, tx); 1578 do_free_accounting = dbuf_block_freeable(db); 1579 } 1580 1581 /* 1582 * If this buffer is dirty in an old transaction group we need 1583 * to make a copy of it so that the changes we make in this 1584 * transaction group won't leak out when we sync the older txg. 1585 */ 1586 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1587 if (db->db_level == 0) { 1588 void *data_old = db->db_buf; 1589 1590 if (db->db_state != DB_NOFILL) { 1591 if (db->db_blkid == DMU_BONUS_BLKID) { 1592 dbuf_fix_old_data(db, tx->tx_txg); 1593 data_old = db->db.db_data; 1594 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1595 /* 1596 * Release the data buffer from the cache so 1597 * that we can modify it without impacting 1598 * possible other users of this cached data 1599 * block. Note that indirect blocks and 1600 * private objects are not released until the 1601 * syncing state (since they are only modified 1602 * then). 1603 */ 1604 arc_release(db->db_buf, db); 1605 dbuf_fix_old_data(db, tx->tx_txg); 1606 data_old = db->db_buf; 1607 } 1608 ASSERT(data_old != NULL); 1609 } 1610 dr->dt.dl.dr_data = data_old; 1611 } else { 1612 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1613 list_create(&dr->dt.di.dr_children, 1614 sizeof (dbuf_dirty_record_t), 1615 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1616 } 1617 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1618 dr->dr_accounted = db->db.db_size; 1619 dr->dr_dbuf = db; 1620 dr->dr_txg = tx->tx_txg; 1621 dr->dr_next = *drp; 1622 *drp = dr; 1623 1624 /* 1625 * We could have been freed_in_flight between the dbuf_noread 1626 * and dbuf_dirty. We win, as though the dbuf_noread() had 1627 * happened after the free. 1628 */ 1629 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1630 db->db_blkid != DMU_SPILL_BLKID) { 1631 mutex_enter(&dn->dn_mtx); 1632 if (dn->dn_free_ranges[txgoff] != NULL) { 1633 range_tree_clear(dn->dn_free_ranges[txgoff], 1634 db->db_blkid, 1); 1635 } 1636 mutex_exit(&dn->dn_mtx); 1637 db->db_freed_in_flight = FALSE; 1638 } 1639 1640 /* 1641 * This buffer is now part of this txg 1642 */ 1643 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1644 db->db_dirtycnt += 1; 1645 ASSERT3U(db->db_dirtycnt, <=, 3); 1646 1647 mutex_exit(&db->db_mtx); 1648 1649 if (db->db_blkid == DMU_BONUS_BLKID || 1650 db->db_blkid == DMU_SPILL_BLKID) { 1651 mutex_enter(&dn->dn_mtx); 1652 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1653 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1654 mutex_exit(&dn->dn_mtx); 1655 dnode_setdirty(dn, tx); 1656 DB_DNODE_EXIT(db); 1657 return (dr); 1658 } 1659 1660 /* 1661 * The dn_struct_rwlock prevents db_blkptr from changing 1662 * due to a write from syncing context completing 1663 * while we are running, so we want to acquire it before 1664 * looking at db_blkptr. 1665 */ 1666 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1667 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1668 drop_struct_lock = TRUE; 1669 } 1670 1671 if (do_free_accounting) { 1672 blkptr_t *bp = db->db_blkptr; 1673 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1674 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1675 /* 1676 * This is only a guess -- if the dbuf is dirty 1677 * in a previous txg, we don't know how much 1678 * space it will use on disk yet. We should 1679 * really have the struct_rwlock to access 1680 * db_blkptr, but since this is just a guess, 1681 * it's OK if we get an odd answer. 1682 */ 1683 ddt_prefetch(os->os_spa, bp); 1684 dnode_willuse_space(dn, -willfree, tx); 1685 } 1686 1687 if (db->db_level == 0) { 1688 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1689 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1690 } 1691 1692 if (db->db_level+1 < dn->dn_nlevels) { 1693 dmu_buf_impl_t *parent = db->db_parent; 1694 dbuf_dirty_record_t *di; 1695 int parent_held = FALSE; 1696 1697 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1698 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1699 1700 parent = dbuf_hold_level(dn, db->db_level+1, 1701 db->db_blkid >> epbs, FTAG); 1702 ASSERT(parent != NULL); 1703 parent_held = TRUE; 1704 } 1705 if (drop_struct_lock) 1706 rw_exit(&dn->dn_struct_rwlock); 1707 ASSERT3U(db->db_level+1, ==, parent->db_level); 1708 di = dbuf_dirty(parent, tx); 1709 if (parent_held) 1710 dbuf_rele(parent, FTAG); 1711 1712 mutex_enter(&db->db_mtx); 1713 /* 1714 * Since we've dropped the mutex, it's possible that 1715 * dbuf_undirty() might have changed this out from under us. 1716 */ 1717 if (db->db_last_dirty == dr || 1718 dn->dn_object == DMU_META_DNODE_OBJECT) { 1719 mutex_enter(&di->dt.di.dr_mtx); 1720 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1721 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1722 list_insert_tail(&di->dt.di.dr_children, dr); 1723 mutex_exit(&di->dt.di.dr_mtx); 1724 dr->dr_parent = di; 1725 } 1726 mutex_exit(&db->db_mtx); 1727 } else { 1728 ASSERT(db->db_level+1 == dn->dn_nlevels); 1729 ASSERT(db->db_blkid < dn->dn_nblkptr); 1730 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1731 mutex_enter(&dn->dn_mtx); 1732 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1733 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1734 mutex_exit(&dn->dn_mtx); 1735 if (drop_struct_lock) 1736 rw_exit(&dn->dn_struct_rwlock); 1737 } 1738 1739 dnode_setdirty(dn, tx); 1740 DB_DNODE_EXIT(db); 1741 return (dr); 1742} 1743 1744/* 1745 * Undirty a buffer in the transaction group referenced by the given 1746 * transaction. Return whether this evicted the dbuf. 1747 */ 1748static boolean_t 1749dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1750{ 1751 dnode_t *dn; 1752 uint64_t txg = tx->tx_txg; 1753 dbuf_dirty_record_t *dr, **drp; 1754 1755 ASSERT(txg != 0); 1756 1757 /* 1758 * Due to our use of dn_nlevels below, this can only be called 1759 * in open context, unless we are operating on the MOS. 1760 * From syncing context, dn_nlevels may be different from the 1761 * dn_nlevels used when dbuf was dirtied. 1762 */ 1763 ASSERT(db->db_objset == 1764 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1765 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1766 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1767 ASSERT0(db->db_level); 1768 ASSERT(MUTEX_HELD(&db->db_mtx)); 1769 1770 /* 1771 * If this buffer is not dirty, we're done. 1772 */ 1773 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1774 if (dr->dr_txg <= txg) 1775 break; 1776 if (dr == NULL || dr->dr_txg < txg) 1777 return (B_FALSE); 1778 ASSERT(dr->dr_txg == txg); 1779 ASSERT(dr->dr_dbuf == db); 1780 1781 DB_DNODE_ENTER(db); 1782 dn = DB_DNODE(db); 1783 1784 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1785 1786 ASSERT(db->db.db_size != 0); 1787 1788 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1789 dr->dr_accounted, txg); 1790 1791 *drp = dr->dr_next; 1792 1793 /* 1794 * Note that there are three places in dbuf_dirty() 1795 * where this dirty record may be put on a list. 1796 * Make sure to do a list_remove corresponding to 1797 * every one of those list_insert calls. 1798 */ 1799 if (dr->dr_parent) { 1800 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1801 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1802 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1803 } else if (db->db_blkid == DMU_SPILL_BLKID || 1804 db->db_level + 1 == dn->dn_nlevels) { 1805 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1806 mutex_enter(&dn->dn_mtx); 1807 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1808 mutex_exit(&dn->dn_mtx); 1809 } 1810 DB_DNODE_EXIT(db); 1811 1812 if (db->db_state != DB_NOFILL) { 1813 dbuf_unoverride(dr); 1814 1815 ASSERT(db->db_buf != NULL); 1816 ASSERT(dr->dt.dl.dr_data != NULL); 1817 if (dr->dt.dl.dr_data != db->db_buf) 1818 arc_buf_destroy(dr->dt.dl.dr_data, db); 1819 } 1820 1821 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1822 1823 ASSERT(db->db_dirtycnt > 0); 1824 db->db_dirtycnt -= 1; 1825 1826 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1827 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); 1828 dbuf_destroy(db); 1829 return (B_TRUE); 1830 } 1831 1832 return (B_FALSE); 1833} 1834 1835void 1836dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1837{ 1838 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1839 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1840 1841 ASSERT(tx->tx_txg != 0); 1842 ASSERT(!refcount_is_zero(&db->db_holds)); 1843 1844 /* 1845 * Quick check for dirtyness. For already dirty blocks, this 1846 * reduces runtime of this function by >90%, and overall performance 1847 * by 50% for some workloads (e.g. file deletion with indirect blocks 1848 * cached). 1849 */ 1850 mutex_enter(&db->db_mtx); 1851 dbuf_dirty_record_t *dr; 1852 for (dr = db->db_last_dirty; 1853 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1854 /* 1855 * It's possible that it is already dirty but not cached, 1856 * because there are some calls to dbuf_dirty() that don't 1857 * go through dmu_buf_will_dirty(). 1858 */ 1859 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1860 /* This dbuf is already dirty and cached. */ 1861 dbuf_redirty(dr); 1862 mutex_exit(&db->db_mtx); 1863 return; 1864 } 1865 } 1866 mutex_exit(&db->db_mtx); 1867 1868 DB_DNODE_ENTER(db); 1869 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1870 rf |= DB_RF_HAVESTRUCT; 1871 DB_DNODE_EXIT(db); 1872 (void) dbuf_read(db, NULL, rf); 1873 (void) dbuf_dirty(db, tx); 1874} 1875 1876void 1877dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1878{ 1879 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1880 1881 db->db_state = DB_NOFILL; 1882 1883 dmu_buf_will_fill(db_fake, tx); 1884} 1885 1886void 1887dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1888{ 1889 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1890 1891 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1892 ASSERT(tx->tx_txg != 0); 1893 ASSERT(db->db_level == 0); 1894 ASSERT(!refcount_is_zero(&db->db_holds)); 1895 1896 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1897 dmu_tx_private_ok(tx)); 1898 1899 dbuf_noread(db); 1900 (void) dbuf_dirty(db, tx); 1901} 1902 1903#pragma weak dmu_buf_fill_done = dbuf_fill_done 1904/* ARGSUSED */ 1905void 1906dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1907{ 1908 mutex_enter(&db->db_mtx); 1909 DBUF_VERIFY(db); 1910 1911 if (db->db_state == DB_FILL) { 1912 if (db->db_level == 0 && db->db_freed_in_flight) { 1913 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1914 /* we were freed while filling */ 1915 /* XXX dbuf_undirty? */ 1916 bzero(db->db.db_data, db->db.db_size); 1917 db->db_freed_in_flight = FALSE; 1918 } 1919 db->db_state = DB_CACHED; 1920 cv_broadcast(&db->db_changed); 1921 } 1922 mutex_exit(&db->db_mtx); 1923} 1924 1925void 1926dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1927 bp_embedded_type_t etype, enum zio_compress comp, 1928 int uncompressed_size, int compressed_size, int byteorder, 1929 dmu_tx_t *tx) 1930{ 1931 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1932 struct dirty_leaf *dl; 1933 dmu_object_type_t type; 1934 1935 if (etype == BP_EMBEDDED_TYPE_DATA) { 1936 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1937 SPA_FEATURE_EMBEDDED_DATA)); 1938 } 1939 1940 DB_DNODE_ENTER(db); 1941 type = DB_DNODE(db)->dn_type; 1942 DB_DNODE_EXIT(db); 1943 1944 ASSERT0(db->db_level); 1945 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1946 1947 dmu_buf_will_not_fill(dbuf, tx); 1948 1949 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1950 dl = &db->db_last_dirty->dt.dl; 1951 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1952 data, comp, uncompressed_size, compressed_size); 1953 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1954 BP_SET_TYPE(&dl->dr_overridden_by, type); 1955 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1956 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1957 1958 dl->dr_override_state = DR_OVERRIDDEN; 1959 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1960} 1961 1962/* 1963 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1964 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1965 */ 1966void 1967dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1968{ 1969 ASSERT(!refcount_is_zero(&db->db_holds)); 1970 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1971 ASSERT(db->db_level == 0); 1972 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1973 ASSERT(buf != NULL); 1974 ASSERT(arc_buf_size(buf) == db->db.db_size); 1975 ASSERT(tx->tx_txg != 0); 1976 1977 arc_return_buf(buf, db); 1978 ASSERT(arc_released(buf)); 1979 1980 mutex_enter(&db->db_mtx); 1981 1982 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1983 cv_wait(&db->db_changed, &db->db_mtx); 1984 1985 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1986 1987 if (db->db_state == DB_CACHED && 1988 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1989 mutex_exit(&db->db_mtx); 1990 (void) dbuf_dirty(db, tx); 1991 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1992 arc_buf_destroy(buf, db); 1993 xuio_stat_wbuf_copied(); 1994 return; 1995 } 1996 1997 xuio_stat_wbuf_nocopy(); 1998 if (db->db_state == DB_CACHED) { 1999 dbuf_dirty_record_t *dr = db->db_last_dirty; 2000 2001 ASSERT(db->db_buf != NULL); 2002 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 2003 ASSERT(dr->dt.dl.dr_data == db->db_buf); 2004 if (!arc_released(db->db_buf)) { 2005 ASSERT(dr->dt.dl.dr_override_state == 2006 DR_OVERRIDDEN); 2007 arc_release(db->db_buf, db); 2008 } 2009 dr->dt.dl.dr_data = buf; 2010 arc_buf_destroy(db->db_buf, db); 2011 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 2012 arc_release(db->db_buf, db); 2013 arc_buf_destroy(db->db_buf, db); 2014 } 2015 db->db_buf = NULL; 2016 } 2017 ASSERT(db->db_buf == NULL); 2018 dbuf_set_data(db, buf); 2019 db->db_state = DB_FILL; 2020 mutex_exit(&db->db_mtx); 2021 (void) dbuf_dirty(db, tx); 2022 dmu_buf_fill_done(&db->db, tx); 2023} 2024 2025void 2026dbuf_destroy(dmu_buf_impl_t *db) 2027{ 2028 dnode_t *dn; 2029 dmu_buf_impl_t *parent = db->db_parent; 2030 dmu_buf_impl_t *dndb; 2031 2032 ASSERT(MUTEX_HELD(&db->db_mtx)); 2033 ASSERT(refcount_is_zero(&db->db_holds)); 2034 2035 if (db->db_buf != NULL) { 2036 arc_buf_destroy(db->db_buf, db); 2037 db->db_buf = NULL; 2038 } 2039 2040 if (db->db_blkid == DMU_BONUS_BLKID) { 2041 ASSERT(db->db.db_data != NULL); 2042 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 2043 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2044 db->db_state = DB_UNCACHED; 2045 } 2046 2047 dbuf_clear_data(db); 2048 2049 if (multilist_link_active(&db->db_cache_link)) { 2050 multilist_remove(&dbuf_cache, db); 2051 (void) refcount_remove_many(&dbuf_cache_size, 2052 db->db.db_size, db); 2053 } 2054 2055 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 2056 ASSERT(db->db_data_pending == NULL); 2057 2058 db->db_state = DB_EVICTING; 2059 db->db_blkptr = NULL; 2060 2061 /* 2062 * Now that db_state is DB_EVICTING, nobody else can find this via 2063 * the hash table. We can now drop db_mtx, which allows us to 2064 * acquire the dn_dbufs_mtx. 2065 */ 2066 mutex_exit(&db->db_mtx); 2067 2068 DB_DNODE_ENTER(db); 2069 dn = DB_DNODE(db); 2070 dndb = dn->dn_dbuf; 2071 if (db->db_blkid != DMU_BONUS_BLKID) { 2072 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); 2073 if (needlock) 2074 mutex_enter(&dn->dn_dbufs_mtx); 2075 avl_remove(&dn->dn_dbufs, db); 2076 atomic_dec_32(&dn->dn_dbufs_count); 2077 membar_producer(); 2078 DB_DNODE_EXIT(db); 2079 if (needlock) 2080 mutex_exit(&dn->dn_dbufs_mtx); 2081 /* 2082 * Decrementing the dbuf count means that the hold corresponding 2083 * to the removed dbuf is no longer discounted in dnode_move(), 2084 * so the dnode cannot be moved until after we release the hold. 2085 * The membar_producer() ensures visibility of the decremented 2086 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 2087 * release any lock. 2088 */ 2089 dnode_rele(dn, db); 2090 db->db_dnode_handle = NULL; 2091 2092 dbuf_hash_remove(db); 2093 } else { 2094 DB_DNODE_EXIT(db); 2095 } 2096 2097 ASSERT(refcount_is_zero(&db->db_holds)); 2098 2099 db->db_parent = NULL; 2100 2101 ASSERT(db->db_buf == NULL); 2102 ASSERT(db->db.db_data == NULL); 2103 ASSERT(db->db_hash_next == NULL); 2104 ASSERT(db->db_blkptr == NULL); 2105 ASSERT(db->db_data_pending == NULL); 2106 ASSERT(!multilist_link_active(&db->db_cache_link)); 2107 2108 kmem_cache_free(dbuf_kmem_cache, db); 2109 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2110 2111 /* 2112 * If this dbuf is referenced from an indirect dbuf, 2113 * decrement the ref count on the indirect dbuf. 2114 */ 2115 if (parent && parent != dndb) 2116 dbuf_rele(parent, db); 2117} 2118 2119/* 2120 * Note: While bpp will always be updated if the function returns success, 2121 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 2122 * this happens when the dnode is the meta-dnode, or a userused or groupused 2123 * object. 2124 */ 2125static int 2126dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 2127 dmu_buf_impl_t **parentp, blkptr_t **bpp) 2128{ 2129 int nlevels, epbs; 2130 2131 *parentp = NULL; 2132 *bpp = NULL; 2133 2134 ASSERT(blkid != DMU_BONUS_BLKID); 2135 2136 if (blkid == DMU_SPILL_BLKID) { 2137 mutex_enter(&dn->dn_mtx); 2138 if (dn->dn_have_spill && 2139 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 2140 *bpp = &dn->dn_phys->dn_spill; 2141 else 2142 *bpp = NULL; 2143 dbuf_add_ref(dn->dn_dbuf, NULL); 2144 *parentp = dn->dn_dbuf; 2145 mutex_exit(&dn->dn_mtx); 2146 return (0); 2147 } 2148 2149 if (dn->dn_phys->dn_nlevels == 0) 2150 nlevels = 1; 2151 else 2152 nlevels = dn->dn_phys->dn_nlevels; 2153 2154 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2155 2156 ASSERT3U(level * epbs, <, 64); 2157 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2158 if (level >= nlevels || 2159 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 2160 /* the buffer has no parent yet */ 2161 return (SET_ERROR(ENOENT)); 2162 } else if (level < nlevels-1) { 2163 /* this block is referenced from an indirect block */ 2164 int err = dbuf_hold_impl(dn, level+1, 2165 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 2166 if (err) 2167 return (err); 2168 err = dbuf_read(*parentp, NULL, 2169 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 2170 if (err) { 2171 dbuf_rele(*parentp, NULL); 2172 *parentp = NULL; 2173 return (err); 2174 } 2175 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 2176 (blkid & ((1ULL << epbs) - 1)); 2177 return (0); 2178 } else { 2179 /* the block is referenced from the dnode */ 2180 ASSERT3U(level, ==, nlevels-1); 2181 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 2182 blkid < dn->dn_phys->dn_nblkptr); 2183 if (dn->dn_dbuf) { 2184 dbuf_add_ref(dn->dn_dbuf, NULL); 2185 *parentp = dn->dn_dbuf; 2186 } 2187 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 2188 return (0); 2189 } 2190} 2191 2192static dmu_buf_impl_t * 2193dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 2194 dmu_buf_impl_t *parent, blkptr_t *blkptr) 2195{ 2196 objset_t *os = dn->dn_objset; 2197 dmu_buf_impl_t *db, *odb; 2198 2199 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2200 ASSERT(dn->dn_type != DMU_OT_NONE); 2201 2202 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); 2203 2204 db->db_objset = os; 2205 db->db.db_object = dn->dn_object; 2206 db->db_level = level; 2207 db->db_blkid = blkid; 2208 db->db_last_dirty = NULL; 2209 db->db_dirtycnt = 0; 2210 db->db_dnode_handle = dn->dn_handle; 2211 db->db_parent = parent; 2212 db->db_blkptr = blkptr; 2213 2214 db->db_user = NULL; 2215 db->db_user_immediate_evict = FALSE; 2216 db->db_freed_in_flight = FALSE; 2217 db->db_pending_evict = FALSE; 2218 2219 if (blkid == DMU_BONUS_BLKID) { 2220 ASSERT3P(parent, ==, dn->dn_dbuf); 2221 db->db.db_size = DN_MAX_BONUSLEN - 2222 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 2223 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 2224 db->db.db_offset = DMU_BONUS_BLKID; 2225 db->db_state = DB_UNCACHED; 2226 /* the bonus dbuf is not placed in the hash table */ 2227 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2228 return (db); 2229 } else if (blkid == DMU_SPILL_BLKID) { 2230 db->db.db_size = (blkptr != NULL) ? 2231 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 2232 db->db.db_offset = 0; 2233 } else { 2234 int blocksize = 2235 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 2236 db->db.db_size = blocksize; 2237 db->db.db_offset = db->db_blkid * blocksize; 2238 } 2239 2240 /* 2241 * Hold the dn_dbufs_mtx while we get the new dbuf 2242 * in the hash table *and* added to the dbufs list. 2243 * This prevents a possible deadlock with someone 2244 * trying to look up this dbuf before its added to the 2245 * dn_dbufs list. 2246 */ 2247 mutex_enter(&dn->dn_dbufs_mtx); 2248 db->db_state = DB_EVICTING; 2249 if ((odb = dbuf_hash_insert(db)) != NULL) { 2250 /* someone else inserted it first */ 2251 kmem_cache_free(dbuf_kmem_cache, db); 2252 mutex_exit(&dn->dn_dbufs_mtx); 2253 return (odb); 2254 } 2255 avl_add(&dn->dn_dbufs, db); 2256 if (db->db_level == 0 && db->db_blkid >= 2257 dn->dn_unlisted_l0_blkid) 2258 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 2259 db->db_state = DB_UNCACHED; 2260 mutex_exit(&dn->dn_dbufs_mtx); 2261 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2262 2263 if (parent && parent != dn->dn_dbuf) 2264 dbuf_add_ref(parent, db); 2265 2266 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 2267 refcount_count(&dn->dn_holds) > 0); 2268 (void) refcount_add(&dn->dn_holds, db); 2269 atomic_inc_32(&dn->dn_dbufs_count); 2270 2271 dprintf_dbuf(db, "db=%p\n", db); 2272 2273 return (db); 2274} 2275 2276typedef struct dbuf_prefetch_arg { 2277 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2278 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2279 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2280 int dpa_curlevel; /* The current level that we're reading */ 2281 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ 2282 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2283 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2284 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2285} dbuf_prefetch_arg_t; 2286 2287/* 2288 * Actually issue the prefetch read for the block given. 2289 */ 2290static void 2291dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2292{ 2293 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2294 return; 2295 2296 arc_flags_t aflags = 2297 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2298 2299 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2300 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2301 ASSERT(dpa->dpa_zio != NULL); 2302 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2303 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2304 &aflags, &dpa->dpa_zb); 2305} 2306 2307/* 2308 * Called when an indirect block above our prefetch target is read in. This 2309 * will either read in the next indirect block down the tree or issue the actual 2310 * prefetch if the next block down is our target. 2311 */ 2312static void 2313dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2314{ 2315 dbuf_prefetch_arg_t *dpa = private; 2316 2317 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2318 ASSERT3S(dpa->dpa_curlevel, >, 0); 2319 2320 /* 2321 * The dpa_dnode is only valid if we are called with a NULL 2322 * zio. This indicates that the arc_read() returned without 2323 * first calling zio_read() to issue a physical read. Once 2324 * a physical read is made the dpa_dnode must be invalidated 2325 * as the locks guarding it may have been dropped. If the 2326 * dpa_dnode is still valid, then we want to add it to the dbuf 2327 * cache. To do so, we must hold the dbuf associated with the block 2328 * we just prefetched, read its contents so that we associate it 2329 * with an arc_buf_t, and then release it. 2330 */ 2331 if (zio != NULL) { 2332 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2333 if (zio->io_flags & ZIO_FLAG_RAW) { 2334 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); 2335 } else { 2336 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2337 } 2338 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2339 2340 dpa->dpa_dnode = NULL; 2341 } else if (dpa->dpa_dnode != NULL) { 2342 uint64_t curblkid = dpa->dpa_zb.zb_blkid >> 2343 (dpa->dpa_epbs * (dpa->dpa_curlevel - 2344 dpa->dpa_zb.zb_level)); 2345 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, 2346 dpa->dpa_curlevel, curblkid, FTAG); 2347 (void) dbuf_read(db, NULL, 2348 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); 2349 dbuf_rele(db, FTAG); 2350 } 2351 2352 dpa->dpa_curlevel--; 2353 2354 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2355 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2356 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2357 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2358 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2359 kmem_free(dpa, sizeof (*dpa)); 2360 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2361 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2362 dbuf_issue_final_prefetch(dpa, bp); 2363 kmem_free(dpa, sizeof (*dpa)); 2364 } else { 2365 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2366 zbookmark_phys_t zb; 2367 2368 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2369 2370 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2371 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2372 2373 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2374 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2375 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2376 &iter_aflags, &zb); 2377 } 2378 2379 arc_buf_destroy(abuf, private); 2380} 2381 2382/* 2383 * Issue prefetch reads for the given block on the given level. If the indirect 2384 * blocks above that block are not in memory, we will read them in 2385 * asynchronously. As a result, this call never blocks waiting for a read to 2386 * complete. 2387 */ 2388void 2389dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2390 arc_flags_t aflags) 2391{ 2392 blkptr_t bp; 2393 int epbs, nlevels, curlevel; 2394 uint64_t curblkid; 2395 2396 ASSERT(blkid != DMU_BONUS_BLKID); 2397 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2398 2399 if (blkid > dn->dn_maxblkid) 2400 return; 2401 2402 if (dnode_block_freed(dn, blkid)) 2403 return; 2404 2405 /* 2406 * This dnode hasn't been written to disk yet, so there's nothing to 2407 * prefetch. 2408 */ 2409 nlevels = dn->dn_phys->dn_nlevels; 2410 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2411 return; 2412 2413 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2414 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2415 return; 2416 2417 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2418 level, blkid); 2419 if (db != NULL) { 2420 mutex_exit(&db->db_mtx); 2421 /* 2422 * This dbuf already exists. It is either CACHED, or 2423 * (we assume) about to be read or filled. 2424 */ 2425 return; 2426 } 2427 2428 /* 2429 * Find the closest ancestor (indirect block) of the target block 2430 * that is present in the cache. In this indirect block, we will 2431 * find the bp that is at curlevel, curblkid. 2432 */ 2433 curlevel = level; 2434 curblkid = blkid; 2435 while (curlevel < nlevels - 1) { 2436 int parent_level = curlevel + 1; 2437 uint64_t parent_blkid = curblkid >> epbs; 2438 dmu_buf_impl_t *db; 2439 2440 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2441 FALSE, TRUE, FTAG, &db) == 0) { 2442 blkptr_t *bpp = db->db_buf->b_data; 2443 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2444 dbuf_rele(db, FTAG); 2445 break; 2446 } 2447 2448 curlevel = parent_level; 2449 curblkid = parent_blkid; 2450 } 2451 2452 if (curlevel == nlevels - 1) { 2453 /* No cached indirect blocks found. */ 2454 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2455 bp = dn->dn_phys->dn_blkptr[curblkid]; 2456 } 2457 if (BP_IS_HOLE(&bp)) 2458 return; 2459 2460 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2461 2462 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2463 ZIO_FLAG_CANFAIL); 2464 2465 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2466 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2467 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2468 dn->dn_object, level, blkid); 2469 dpa->dpa_curlevel = curlevel; 2470 dpa->dpa_prio = prio; 2471 dpa->dpa_aflags = aflags; 2472 dpa->dpa_spa = dn->dn_objset->os_spa; 2473 dpa->dpa_dnode = dn; 2474 dpa->dpa_epbs = epbs; 2475 dpa->dpa_zio = pio; 2476 2477 /* 2478 * If we have the indirect just above us, no need to do the asynchronous 2479 * prefetch chain; we'll just run the last step ourselves. If we're at 2480 * a higher level, though, we want to issue the prefetches for all the 2481 * indirect blocks asynchronously, so we can go on with whatever we were 2482 * doing. 2483 */ 2484 if (curlevel == level) { 2485 ASSERT3U(curblkid, ==, blkid); 2486 dbuf_issue_final_prefetch(dpa, &bp); 2487 kmem_free(dpa, sizeof (*dpa)); 2488 } else { 2489 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2490 zbookmark_phys_t zb; 2491 2492 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2493 dn->dn_object, curlevel, curblkid); 2494 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2495 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2496 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2497 &iter_aflags, &zb); 2498 } 2499 /* 2500 * We use pio here instead of dpa_zio since it's possible that 2501 * dpa may have already been freed. 2502 */ 2503 zio_nowait(pio); 2504} 2505 2506/* 2507 * Returns with db_holds incremented, and db_mtx not held. 2508 * Note: dn_struct_rwlock must be held. 2509 */ 2510int 2511dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2512 boolean_t fail_sparse, boolean_t fail_uncached, 2513 void *tag, dmu_buf_impl_t **dbp) 2514{ 2515 dmu_buf_impl_t *db, *parent = NULL; 2516 2517 ASSERT(blkid != DMU_BONUS_BLKID); 2518 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2519 ASSERT3U(dn->dn_nlevels, >, level); 2520 2521 *dbp = NULL; 2522top: 2523 /* dbuf_find() returns with db_mtx held */ 2524 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2525 2526 if (db == NULL) { 2527 blkptr_t *bp = NULL; 2528 int err; 2529 2530 if (fail_uncached) 2531 return (SET_ERROR(ENOENT)); 2532 2533 ASSERT3P(parent, ==, NULL); 2534 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2535 if (fail_sparse) { 2536 if (err == 0 && bp && BP_IS_HOLE(bp)) 2537 err = SET_ERROR(ENOENT); 2538 if (err) { 2539 if (parent) 2540 dbuf_rele(parent, NULL); 2541 return (err); 2542 } 2543 } 2544 if (err && err != ENOENT) 2545 return (err); 2546 db = dbuf_create(dn, level, blkid, parent, bp); 2547 } 2548 2549 if (fail_uncached && db->db_state != DB_CACHED) { 2550 mutex_exit(&db->db_mtx); 2551 return (SET_ERROR(ENOENT)); 2552 } 2553 2554 if (db->db_buf != NULL) 2555 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2556 2557 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2558 2559 /* 2560 * If this buffer is currently syncing out, and we are are 2561 * still referencing it from db_data, we need to make a copy 2562 * of it in case we decide we want to dirty it again in this txg. 2563 */ 2564 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2565 dn->dn_object != DMU_META_DNODE_OBJECT && 2566 db->db_state == DB_CACHED && db->db_data_pending) { 2567 dbuf_dirty_record_t *dr = db->db_data_pending; 2568 2569 if (dr->dt.dl.dr_data == db->db_buf) { 2570 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2571 2572 dbuf_set_data(db, 2573 arc_alloc_buf(dn->dn_objset->os_spa, 2574 db->db.db_size, db, type)); 2575 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2576 db->db.db_size); 2577 } 2578 } 2579 2580 if (multilist_link_active(&db->db_cache_link)) { 2581 ASSERT(refcount_is_zero(&db->db_holds)); 2582 multilist_remove(&dbuf_cache, db); 2583 (void) refcount_remove_many(&dbuf_cache_size, 2584 db->db.db_size, db); 2585 } 2586 (void) refcount_add(&db->db_holds, tag); 2587 DBUF_VERIFY(db); 2588 mutex_exit(&db->db_mtx); 2589 2590 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2591 if (parent) 2592 dbuf_rele(parent, NULL); 2593 2594 ASSERT3P(DB_DNODE(db), ==, dn); 2595 ASSERT3U(db->db_blkid, ==, blkid); 2596 ASSERT3U(db->db_level, ==, level); 2597 *dbp = db; 2598 2599 return (0); 2600} 2601 2602dmu_buf_impl_t * 2603dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2604{ 2605 return (dbuf_hold_level(dn, 0, blkid, tag)); 2606} 2607 2608dmu_buf_impl_t * 2609dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2610{ 2611 dmu_buf_impl_t *db; 2612 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2613 return (err ? NULL : db); 2614} 2615 2616void 2617dbuf_create_bonus(dnode_t *dn) 2618{ 2619 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2620 2621 ASSERT(dn->dn_bonus == NULL); 2622 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2623} 2624 2625int 2626dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2627{ 2628 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2629 dnode_t *dn; 2630 2631 if (db->db_blkid != DMU_SPILL_BLKID) 2632 return (SET_ERROR(ENOTSUP)); 2633 if (blksz == 0) 2634 blksz = SPA_MINBLOCKSIZE; 2635 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2636 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2637 2638 DB_DNODE_ENTER(db); 2639 dn = DB_DNODE(db); 2640 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2641 dbuf_new_size(db, blksz, tx); 2642 rw_exit(&dn->dn_struct_rwlock); 2643 DB_DNODE_EXIT(db); 2644 2645 return (0); 2646} 2647 2648void 2649dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2650{ 2651 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2652} 2653 2654#pragma weak dmu_buf_add_ref = dbuf_add_ref 2655void 2656dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2657{ 2658 int64_t holds = refcount_add(&db->db_holds, tag); 2659 ASSERT3S(holds, >, 1); 2660} 2661 2662#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2663boolean_t 2664dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2665 void *tag) 2666{ 2667 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2668 dmu_buf_impl_t *found_db; 2669 boolean_t result = B_FALSE; 2670 2671 if (db->db_blkid == DMU_BONUS_BLKID) 2672 found_db = dbuf_find_bonus(os, obj); 2673 else 2674 found_db = dbuf_find(os, obj, 0, blkid); 2675 2676 if (found_db != NULL) { 2677 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2678 (void) refcount_add(&db->db_holds, tag); 2679 result = B_TRUE; 2680 } 2681 mutex_exit(&db->db_mtx); 2682 } 2683 return (result); 2684} 2685 2686/* 2687 * If you call dbuf_rele() you had better not be referencing the dnode handle 2688 * unless you have some other direct or indirect hold on the dnode. (An indirect 2689 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2690 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2691 * dnode's parent dbuf evicting its dnode handles. 2692 */ 2693void 2694dbuf_rele(dmu_buf_impl_t *db, void *tag) 2695{ 2696 mutex_enter(&db->db_mtx); 2697 dbuf_rele_and_unlock(db, tag); 2698} 2699 2700void 2701dmu_buf_rele(dmu_buf_t *db, void *tag) 2702{ 2703 dbuf_rele((dmu_buf_impl_t *)db, tag); 2704} 2705 2706/* 2707 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2708 * db_dirtycnt and db_holds to be updated atomically. 2709 */ 2710void 2711dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2712{ 2713 int64_t holds; 2714 2715 ASSERT(MUTEX_HELD(&db->db_mtx)); 2716 DBUF_VERIFY(db); 2717 2718 /* 2719 * Remove the reference to the dbuf before removing its hold on the 2720 * dnode so we can guarantee in dnode_move() that a referenced bonus 2721 * buffer has a corresponding dnode hold. 2722 */ 2723 holds = refcount_remove(&db->db_holds, tag); 2724 ASSERT(holds >= 0); 2725 2726 /* 2727 * We can't freeze indirects if there is a possibility that they 2728 * may be modified in the current syncing context. 2729 */ 2730 if (db->db_buf != NULL && 2731 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { 2732 arc_buf_freeze(db->db_buf); 2733 } 2734 2735 if (holds == db->db_dirtycnt && 2736 db->db_level == 0 && db->db_user_immediate_evict) 2737 dbuf_evict_user(db); 2738 2739 if (holds == 0) { 2740 if (db->db_blkid == DMU_BONUS_BLKID) { 2741 dnode_t *dn; 2742 boolean_t evict_dbuf = db->db_pending_evict; 2743 2744 /* 2745 * If the dnode moves here, we cannot cross this 2746 * barrier until the move completes. 2747 */ 2748 DB_DNODE_ENTER(db); 2749 2750 dn = DB_DNODE(db); 2751 atomic_dec_32(&dn->dn_dbufs_count); 2752 2753 /* 2754 * Decrementing the dbuf count means that the bonus 2755 * buffer's dnode hold is no longer discounted in 2756 * dnode_move(). The dnode cannot move until after 2757 * the dnode_rele() below. 2758 */ 2759 DB_DNODE_EXIT(db); 2760 2761 /* 2762 * Do not reference db after its lock is dropped. 2763 * Another thread may evict it. 2764 */ 2765 mutex_exit(&db->db_mtx); 2766 2767 if (evict_dbuf) 2768 dnode_evict_bonus(dn); 2769 2770 dnode_rele(dn, db); 2771 } else if (db->db_buf == NULL) { 2772 /* 2773 * This is a special case: we never associated this 2774 * dbuf with any data allocated from the ARC. 2775 */ 2776 ASSERT(db->db_state == DB_UNCACHED || 2777 db->db_state == DB_NOFILL); 2778 dbuf_destroy(db); 2779 } else if (arc_released(db->db_buf)) { 2780 /* 2781 * This dbuf has anonymous data associated with it. 2782 */ 2783 dbuf_destroy(db); 2784 } else { 2785 boolean_t do_arc_evict = B_FALSE; 2786 blkptr_t bp; 2787 spa_t *spa = dmu_objset_spa(db->db_objset); 2788 2789 if (!DBUF_IS_CACHEABLE(db) && 2790 db->db_blkptr != NULL && 2791 !BP_IS_HOLE(db->db_blkptr) && 2792 !BP_IS_EMBEDDED(db->db_blkptr)) { 2793 do_arc_evict = B_TRUE; 2794 bp = *db->db_blkptr; 2795 } 2796 2797 if (!DBUF_IS_CACHEABLE(db) || 2798 db->db_pending_evict) { 2799 dbuf_destroy(db); 2800 } else if (!multilist_link_active(&db->db_cache_link)) { 2801 multilist_insert(&dbuf_cache, db); 2802 (void) refcount_add_many(&dbuf_cache_size, 2803 db->db.db_size, db); 2804 mutex_exit(&db->db_mtx); 2805 2806 dbuf_evict_notify(); 2807 } 2808 2809 if (do_arc_evict) 2810 arc_freed(spa, &bp); 2811 } 2812 } else { 2813 mutex_exit(&db->db_mtx); 2814 } 2815 2816} 2817 2818#pragma weak dmu_buf_refcount = dbuf_refcount 2819uint64_t 2820dbuf_refcount(dmu_buf_impl_t *db) 2821{ 2822 return (refcount_count(&db->db_holds)); 2823} 2824 2825void * 2826dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2827 dmu_buf_user_t *new_user) 2828{ 2829 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2830 2831 mutex_enter(&db->db_mtx); 2832 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2833 if (db->db_user == old_user) 2834 db->db_user = new_user; 2835 else 2836 old_user = db->db_user; 2837 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2838 mutex_exit(&db->db_mtx); 2839 2840 return (old_user); 2841} 2842 2843void * 2844dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2845{ 2846 return (dmu_buf_replace_user(db_fake, NULL, user)); 2847} 2848 2849void * 2850dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2851{ 2852 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2853 2854 db->db_user_immediate_evict = TRUE; 2855 return (dmu_buf_set_user(db_fake, user)); 2856} 2857 2858void * 2859dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2860{ 2861 return (dmu_buf_replace_user(db_fake, user, NULL)); 2862} 2863 2864void * 2865dmu_buf_get_user(dmu_buf_t *db_fake) 2866{ 2867 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2868 2869 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2870 return (db->db_user); 2871} 2872 2873void 2874dmu_buf_user_evict_wait() 2875{ 2876 taskq_wait(dbu_evict_taskq); 2877} 2878 2879boolean_t 2880dmu_buf_freeable(dmu_buf_t *dbuf) 2881{ 2882 boolean_t res = B_FALSE; 2883 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2884 2885 if (db->db_blkptr) 2886 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2887 db->db_blkptr, db->db_blkptr->blk_birth); 2888 2889 return (res); 2890} 2891 2892blkptr_t * 2893dmu_buf_get_blkptr(dmu_buf_t *db) 2894{ 2895 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2896 return (dbi->db_blkptr); 2897} 2898 2899objset_t * 2900dmu_buf_get_objset(dmu_buf_t *db) 2901{ 2902 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2903 return (dbi->db_objset); 2904} 2905 2906dnode_t * 2907dmu_buf_dnode_enter(dmu_buf_t *db) 2908{ 2909 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2910 DB_DNODE_ENTER(dbi); 2911 return (DB_DNODE(dbi)); 2912} 2913 2914void 2915dmu_buf_dnode_exit(dmu_buf_t *db) 2916{ 2917 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2918 DB_DNODE_EXIT(dbi); 2919} 2920 2921static void 2922dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2923{ 2924 /* ASSERT(dmu_tx_is_syncing(tx) */ 2925 ASSERT(MUTEX_HELD(&db->db_mtx)); 2926 2927 if (db->db_blkptr != NULL) 2928 return; 2929 2930 if (db->db_blkid == DMU_SPILL_BLKID) { 2931 db->db_blkptr = &dn->dn_phys->dn_spill; 2932 BP_ZERO(db->db_blkptr); 2933 return; 2934 } 2935 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2936 /* 2937 * This buffer was allocated at a time when there was 2938 * no available blkptrs from the dnode, or it was 2939 * inappropriate to hook it in (i.e., nlevels mis-match). 2940 */ 2941 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2942 ASSERT(db->db_parent == NULL); 2943 db->db_parent = dn->dn_dbuf; 2944 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2945 DBUF_VERIFY(db); 2946 } else { 2947 dmu_buf_impl_t *parent = db->db_parent; 2948 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2949 2950 ASSERT(dn->dn_phys->dn_nlevels > 1); 2951 if (parent == NULL) { 2952 mutex_exit(&db->db_mtx); 2953 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2954 parent = dbuf_hold_level(dn, db->db_level + 1, 2955 db->db_blkid >> epbs, db); 2956 rw_exit(&dn->dn_struct_rwlock); 2957 mutex_enter(&db->db_mtx); 2958 db->db_parent = parent; 2959 } 2960 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2961 (db->db_blkid & ((1ULL << epbs) - 1)); 2962 DBUF_VERIFY(db); 2963 } 2964} 2965 2966static void 2967dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2968{ 2969 dmu_buf_impl_t *db = dr->dr_dbuf; 2970 dnode_t *dn; 2971 zio_t *zio; 2972 2973 ASSERT(dmu_tx_is_syncing(tx)); 2974 2975 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2976 2977 mutex_enter(&db->db_mtx); 2978 2979 ASSERT(db->db_level > 0); 2980 DBUF_VERIFY(db); 2981 2982 /* Read the block if it hasn't been read yet. */ 2983 if (db->db_buf == NULL) { 2984 mutex_exit(&db->db_mtx); 2985 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2986 mutex_enter(&db->db_mtx); 2987 } 2988 ASSERT3U(db->db_state, ==, DB_CACHED); 2989 ASSERT(db->db_buf != NULL); 2990 2991 DB_DNODE_ENTER(db); 2992 dn = DB_DNODE(db); 2993 /* Indirect block size must match what the dnode thinks it is. */ 2994 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2995 dbuf_check_blkptr(dn, db); 2996 DB_DNODE_EXIT(db); 2997 2998 /* Provide the pending dirty record to child dbufs */ 2999 db->db_data_pending = dr; 3000 3001 mutex_exit(&db->db_mtx); 3002 dbuf_write(dr, db->db_buf, tx); 3003 3004 zio = dr->dr_zio; 3005 mutex_enter(&dr->dt.di.dr_mtx); 3006 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 3007 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3008 mutex_exit(&dr->dt.di.dr_mtx); 3009 zio_nowait(zio); 3010} 3011 3012static void 3013dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3014{ 3015 arc_buf_t **datap = &dr->dt.dl.dr_data; 3016 dmu_buf_impl_t *db = dr->dr_dbuf; 3017 dnode_t *dn; 3018 objset_t *os; 3019 uint64_t txg = tx->tx_txg; 3020 3021 ASSERT(dmu_tx_is_syncing(tx)); 3022 3023 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3024 3025 mutex_enter(&db->db_mtx); 3026 /* 3027 * To be synced, we must be dirtied. But we 3028 * might have been freed after the dirty. 3029 */ 3030 if (db->db_state == DB_UNCACHED) { 3031 /* This buffer has been freed since it was dirtied */ 3032 ASSERT(db->db.db_data == NULL); 3033 } else if (db->db_state == DB_FILL) { 3034 /* This buffer was freed and is now being re-filled */ 3035 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 3036 } else { 3037 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 3038 } 3039 DBUF_VERIFY(db); 3040 3041 DB_DNODE_ENTER(db); 3042 dn = DB_DNODE(db); 3043 3044 if (db->db_blkid == DMU_SPILL_BLKID) { 3045 mutex_enter(&dn->dn_mtx); 3046 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 3047 mutex_exit(&dn->dn_mtx); 3048 } 3049 3050 /* 3051 * If this is a bonus buffer, simply copy the bonus data into the 3052 * dnode. It will be written out when the dnode is synced (and it 3053 * will be synced, since it must have been dirty for dbuf_sync to 3054 * be called). 3055 */ 3056 if (db->db_blkid == DMU_BONUS_BLKID) { 3057 dbuf_dirty_record_t **drp; 3058 3059 ASSERT(*datap != NULL); 3060 ASSERT0(db->db_level); 3061 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 3062 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 3063 DB_DNODE_EXIT(db); 3064 3065 if (*datap != db->db.db_data) { 3066 zio_buf_free(*datap, DN_MAX_BONUSLEN); 3067 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 3068 } 3069 db->db_data_pending = NULL; 3070 drp = &db->db_last_dirty; 3071 while (*drp != dr) 3072 drp = &(*drp)->dr_next; 3073 ASSERT(dr->dr_next == NULL); 3074 ASSERT(dr->dr_dbuf == db); 3075 *drp = dr->dr_next; 3076 if (dr->dr_dbuf->db_level != 0) { 3077 list_destroy(&dr->dt.di.dr_children); 3078 mutex_destroy(&dr->dt.di.dr_mtx); 3079 } 3080 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3081 ASSERT(db->db_dirtycnt > 0); 3082 db->db_dirtycnt -= 1; 3083 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 3084 return; 3085 } 3086 3087 os = dn->dn_objset; 3088 3089 /* 3090 * This function may have dropped the db_mtx lock allowing a dmu_sync 3091 * operation to sneak in. As a result, we need to ensure that we 3092 * don't check the dr_override_state until we have returned from 3093 * dbuf_check_blkptr. 3094 */ 3095 dbuf_check_blkptr(dn, db); 3096 3097 /* 3098 * If this buffer is in the middle of an immediate write, 3099 * wait for the synchronous IO to complete. 3100 */ 3101 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 3102 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 3103 cv_wait(&db->db_changed, &db->db_mtx); 3104 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 3105 } 3106 3107 if (db->db_state != DB_NOFILL && 3108 dn->dn_object != DMU_META_DNODE_OBJECT && 3109 refcount_count(&db->db_holds) > 1 && 3110 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 3111 *datap == db->db_buf) { 3112 /* 3113 * If this buffer is currently "in use" (i.e., there 3114 * are active holds and db_data still references it), 3115 * then make a copy before we start the write so that 3116 * any modifications from the open txg will not leak 3117 * into this write. 3118 * 3119 * NOTE: this copy does not need to be made for 3120 * objects only modified in the syncing context (e.g. 3121 * DNONE_DNODE blocks). 3122 */ 3123 int blksz = arc_buf_size(*datap); 3124 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 3125 *datap = arc_alloc_buf(os->os_spa, blksz, db, type); 3126 bcopy(db->db.db_data, (*datap)->b_data, blksz); 3127 } 3128 db->db_data_pending = dr; 3129 3130 mutex_exit(&db->db_mtx); 3131 3132 dbuf_write(dr, *datap, tx); 3133 3134 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3135 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 3136 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 3137 DB_DNODE_EXIT(db); 3138 } else { 3139 /* 3140 * Although zio_nowait() does not "wait for an IO", it does 3141 * initiate the IO. If this is an empty write it seems plausible 3142 * that the IO could actually be completed before the nowait 3143 * returns. We need to DB_DNODE_EXIT() first in case 3144 * zio_nowait() invalidates the dbuf. 3145 */ 3146 DB_DNODE_EXIT(db); 3147 zio_nowait(dr->dr_zio); 3148 } 3149} 3150 3151void 3152dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 3153{ 3154 dbuf_dirty_record_t *dr; 3155 3156 while (dr = list_head(list)) { 3157 if (dr->dr_zio != NULL) { 3158 /* 3159 * If we find an already initialized zio then we 3160 * are processing the meta-dnode, and we have finished. 3161 * The dbufs for all dnodes are put back on the list 3162 * during processing, so that we can zio_wait() 3163 * these IOs after initiating all child IOs. 3164 */ 3165 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 3166 DMU_META_DNODE_OBJECT); 3167 break; 3168 } 3169 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 3170 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 3171 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 3172 } 3173 list_remove(list, dr); 3174 if (dr->dr_dbuf->db_level > 0) 3175 dbuf_sync_indirect(dr, tx); 3176 else 3177 dbuf_sync_leaf(dr, tx); 3178 } 3179} 3180 3181/* ARGSUSED */ 3182static void 3183dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3184{ 3185 dmu_buf_impl_t *db = vdb; 3186 dnode_t *dn; 3187 blkptr_t *bp = zio->io_bp; 3188 blkptr_t *bp_orig = &zio->io_bp_orig; 3189 spa_t *spa = zio->io_spa; 3190 int64_t delta; 3191 uint64_t fill = 0; 3192 int i; 3193 3194 ASSERT3P(db->db_blkptr, !=, NULL); 3195 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); 3196 3197 DB_DNODE_ENTER(db); 3198 dn = DB_DNODE(db); 3199 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 3200 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 3201 zio->io_prev_space_delta = delta; 3202 3203 if (bp->blk_birth != 0) { 3204 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 3205 BP_GET_TYPE(bp) == dn->dn_type) || 3206 (db->db_blkid == DMU_SPILL_BLKID && 3207 BP_GET_TYPE(bp) == dn->dn_bonustype) || 3208 BP_IS_EMBEDDED(bp)); 3209 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 3210 } 3211 3212 mutex_enter(&db->db_mtx); 3213 3214#ifdef ZFS_DEBUG 3215 if (db->db_blkid == DMU_SPILL_BLKID) { 3216 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3217 ASSERT(!(BP_IS_HOLE(bp)) && 3218 db->db_blkptr == &dn->dn_phys->dn_spill); 3219 } 3220#endif 3221 3222 if (db->db_level == 0) { 3223 mutex_enter(&dn->dn_mtx); 3224 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 3225 db->db_blkid != DMU_SPILL_BLKID) 3226 dn->dn_phys->dn_maxblkid = db->db_blkid; 3227 mutex_exit(&dn->dn_mtx); 3228 3229 if (dn->dn_type == DMU_OT_DNODE) { 3230 dnode_phys_t *dnp = db->db.db_data; 3231 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 3232 i--, dnp++) { 3233 if (dnp->dn_type != DMU_OT_NONE) 3234 fill++; 3235 } 3236 } else { 3237 if (BP_IS_HOLE(bp)) { 3238 fill = 0; 3239 } else { 3240 fill = 1; 3241 } 3242 } 3243 } else { 3244 blkptr_t *ibp = db->db.db_data; 3245 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3246 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 3247 if (BP_IS_HOLE(ibp)) 3248 continue; 3249 fill += BP_GET_FILL(ibp); 3250 } 3251 } 3252 DB_DNODE_EXIT(db); 3253 3254 if (!BP_IS_EMBEDDED(bp)) 3255 bp->blk_fill = fill; 3256 3257 mutex_exit(&db->db_mtx); 3258 3259 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3260 *db->db_blkptr = *bp; 3261 rw_exit(&dn->dn_struct_rwlock); 3262} 3263 3264/* ARGSUSED */ 3265/* 3266 * This function gets called just prior to running through the compression 3267 * stage of the zio pipeline. If we're an indirect block comprised of only 3268 * holes, then we want this indirect to be compressed away to a hole. In 3269 * order to do that we must zero out any information about the holes that 3270 * this indirect points to prior to before we try to compress it. 3271 */ 3272static void 3273dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3274{ 3275 dmu_buf_impl_t *db = vdb; 3276 dnode_t *dn; 3277 blkptr_t *bp; 3278 uint64_t i; 3279 int epbs; 3280 3281 ASSERT3U(db->db_level, >, 0); 3282 DB_DNODE_ENTER(db); 3283 dn = DB_DNODE(db); 3284 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3285 3286 /* Determine if all our children are holes */ 3287 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { 3288 if (!BP_IS_HOLE(bp)) 3289 break; 3290 } 3291 3292 /* 3293 * If all the children are holes, then zero them all out so that 3294 * we may get compressed away. 3295 */ 3296 if (i == 1 << epbs) { 3297 /* didn't find any non-holes */ 3298 bzero(db->db.db_data, db->db.db_size); 3299 } 3300 DB_DNODE_EXIT(db); 3301} 3302 3303/* 3304 * The SPA will call this callback several times for each zio - once 3305 * for every physical child i/o (zio->io_phys_children times). This 3306 * allows the DMU to monitor the progress of each logical i/o. For example, 3307 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 3308 * block. There may be a long delay before all copies/fragments are completed, 3309 * so this callback allows us to retire dirty space gradually, as the physical 3310 * i/os complete. 3311 */ 3312/* ARGSUSED */ 3313static void 3314dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 3315{ 3316 dmu_buf_impl_t *db = arg; 3317 objset_t *os = db->db_objset; 3318 dsl_pool_t *dp = dmu_objset_pool(os); 3319 dbuf_dirty_record_t *dr; 3320 int delta = 0; 3321 3322 dr = db->db_data_pending; 3323 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 3324 3325 /* 3326 * The callback will be called io_phys_children times. Retire one 3327 * portion of our dirty space each time we are called. Any rounding 3328 * error will be cleaned up by dsl_pool_sync()'s call to 3329 * dsl_pool_undirty_space(). 3330 */ 3331 delta = dr->dr_accounted / zio->io_phys_children; 3332 dsl_pool_undirty_space(dp, delta, zio->io_txg); 3333} 3334 3335/* ARGSUSED */ 3336static void 3337dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3338{ 3339 dmu_buf_impl_t *db = vdb; 3340 blkptr_t *bp_orig = &zio->io_bp_orig; 3341 blkptr_t *bp = db->db_blkptr; 3342 objset_t *os = db->db_objset; 3343 dmu_tx_t *tx = os->os_synctx; 3344 dbuf_dirty_record_t **drp, *dr; 3345 3346 ASSERT0(zio->io_error); 3347 ASSERT(db->db_blkptr == bp); 3348 3349 /* 3350 * For nopwrites and rewrites we ensure that the bp matches our 3351 * original and bypass all the accounting. 3352 */ 3353 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3354 ASSERT(BP_EQUAL(bp, bp_orig)); 3355 } else { 3356 dsl_dataset_t *ds = os->os_dsl_dataset; 3357 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3358 dsl_dataset_block_born(ds, bp, tx); 3359 } 3360 3361 mutex_enter(&db->db_mtx); 3362 3363 DBUF_VERIFY(db); 3364 3365 drp = &db->db_last_dirty; 3366 while ((dr = *drp) != db->db_data_pending) 3367 drp = &dr->dr_next; 3368 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3369 ASSERT(dr->dr_dbuf == db); 3370 ASSERT(dr->dr_next == NULL); 3371 *drp = dr->dr_next; 3372 3373#ifdef ZFS_DEBUG 3374 if (db->db_blkid == DMU_SPILL_BLKID) { 3375 dnode_t *dn; 3376 3377 DB_DNODE_ENTER(db); 3378 dn = DB_DNODE(db); 3379 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3380 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3381 db->db_blkptr == &dn->dn_phys->dn_spill); 3382 DB_DNODE_EXIT(db); 3383 } 3384#endif 3385 3386 if (db->db_level == 0) { 3387 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3388 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3389 if (db->db_state != DB_NOFILL) { 3390 if (dr->dt.dl.dr_data != db->db_buf) 3391 arc_buf_destroy(dr->dt.dl.dr_data, db); 3392 } 3393 } else { 3394 dnode_t *dn; 3395 3396 DB_DNODE_ENTER(db); 3397 dn = DB_DNODE(db); 3398 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3399 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3400 if (!BP_IS_HOLE(db->db_blkptr)) { 3401 int epbs = 3402 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3403 ASSERT3U(db->db_blkid, <=, 3404 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3405 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3406 db->db.db_size); 3407 } 3408 DB_DNODE_EXIT(db); 3409 mutex_destroy(&dr->dt.di.dr_mtx); 3410 list_destroy(&dr->dt.di.dr_children); 3411 } 3412 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3413 3414 cv_broadcast(&db->db_changed); 3415 ASSERT(db->db_dirtycnt > 0); 3416 db->db_dirtycnt -= 1; 3417 db->db_data_pending = NULL; 3418 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3419} 3420 3421static void 3422dbuf_write_nofill_ready(zio_t *zio) 3423{ 3424 dbuf_write_ready(zio, NULL, zio->io_private); 3425} 3426 3427static void 3428dbuf_write_nofill_done(zio_t *zio) 3429{ 3430 dbuf_write_done(zio, NULL, zio->io_private); 3431} 3432 3433static void 3434dbuf_write_override_ready(zio_t *zio) 3435{ 3436 dbuf_dirty_record_t *dr = zio->io_private; 3437 dmu_buf_impl_t *db = dr->dr_dbuf; 3438 3439 dbuf_write_ready(zio, NULL, db); 3440} 3441 3442static void 3443dbuf_write_override_done(zio_t *zio) 3444{ 3445 dbuf_dirty_record_t *dr = zio->io_private; 3446 dmu_buf_impl_t *db = dr->dr_dbuf; 3447 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3448 3449 mutex_enter(&db->db_mtx); 3450 if (!BP_EQUAL(zio->io_bp, obp)) { 3451 if (!BP_IS_HOLE(obp)) 3452 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3453 arc_release(dr->dt.dl.dr_data, db); 3454 } 3455 mutex_exit(&db->db_mtx); 3456 3457 dbuf_write_done(zio, NULL, db); 3458} 3459 3460/* Issue I/O to commit a dirty buffer to disk. */ 3461static void 3462dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3463{ 3464 dmu_buf_impl_t *db = dr->dr_dbuf; 3465 dnode_t *dn; 3466 objset_t *os; 3467 dmu_buf_impl_t *parent = db->db_parent; 3468 uint64_t txg = tx->tx_txg; 3469 zbookmark_phys_t zb; 3470 zio_prop_t zp; 3471 zio_t *zio; 3472 int wp_flag = 0; 3473 3474 ASSERT(dmu_tx_is_syncing(tx)); 3475 3476 DB_DNODE_ENTER(db); 3477 dn = DB_DNODE(db); 3478 os = dn->dn_objset; 3479 3480 if (db->db_state != DB_NOFILL) { 3481 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3482 /* 3483 * Private object buffers are released here rather 3484 * than in dbuf_dirty() since they are only modified 3485 * in the syncing context and we don't want the 3486 * overhead of making multiple copies of the data. 3487 */ 3488 if (BP_IS_HOLE(db->db_blkptr)) { 3489 arc_buf_thaw(data); 3490 } else { 3491 dbuf_release_bp(db); 3492 } 3493 } 3494 } 3495 3496 if (parent != dn->dn_dbuf) { 3497 /* Our parent is an indirect block. */ 3498 /* We have a dirty parent that has been scheduled for write. */ 3499 ASSERT(parent && parent->db_data_pending); 3500 /* Our parent's buffer is one level closer to the dnode. */ 3501 ASSERT(db->db_level == parent->db_level-1); 3502 /* 3503 * We're about to modify our parent's db_data by modifying 3504 * our block pointer, so the parent must be released. 3505 */ 3506 ASSERT(arc_released(parent->db_buf)); 3507 zio = parent->db_data_pending->dr_zio; 3508 } else { 3509 /* Our parent is the dnode itself. */ 3510 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3511 db->db_blkid != DMU_SPILL_BLKID) || 3512 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3513 if (db->db_blkid != DMU_SPILL_BLKID) 3514 ASSERT3P(db->db_blkptr, ==, 3515 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3516 zio = dn->dn_zio; 3517 } 3518 3519 ASSERT(db->db_level == 0 || data == db->db_buf); 3520 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3521 ASSERT(zio); 3522 3523 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3524 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3525 db->db.db_object, db->db_level, db->db_blkid); 3526 3527 if (db->db_blkid == DMU_SPILL_BLKID) 3528 wp_flag = WP_SPILL; 3529 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3530 3531 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3532 DB_DNODE_EXIT(db); 3533 3534 /* 3535 * We copy the blkptr now (rather than when we instantiate the dirty 3536 * record), because its value can change between open context and 3537 * syncing context. We do not need to hold dn_struct_rwlock to read 3538 * db_blkptr because we are in syncing context. 3539 */ 3540 dr->dr_bp_copy = *db->db_blkptr; 3541 3542 if (db->db_level == 0 && 3543 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3544 /* 3545 * The BP for this block has been provided by open context 3546 * (by dmu_sync() or dmu_buf_write_embedded()). 3547 */ 3548 void *contents = (data != NULL) ? data->b_data : NULL; 3549 3550 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3551 &dr->dr_bp_copy, contents, db->db.db_size, &zp, 3552 dbuf_write_override_ready, NULL, NULL, 3553 dbuf_write_override_done, 3554 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3555 mutex_enter(&db->db_mtx); 3556 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3557 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3558 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3559 mutex_exit(&db->db_mtx); 3560 } else if (db->db_state == DB_NOFILL) { 3561 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3562 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3563 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3564 &dr->dr_bp_copy, NULL, db->db.db_size, &zp, 3565 dbuf_write_nofill_ready, NULL, NULL, 3566 dbuf_write_nofill_done, db, 3567 ZIO_PRIORITY_ASYNC_WRITE, 3568 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3569 } else { 3570 ASSERT(arc_released(data)); 3571 3572 /* 3573 * For indirect blocks, we want to setup the children 3574 * ready callback so that we can properly handle an indirect 3575 * block that only contains holes. 3576 */ 3577 arc_done_func_t *children_ready_cb = NULL; 3578 if (db->db_level != 0) 3579 children_ready_cb = dbuf_write_children_ready; 3580 3581 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3582 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), 3583 &zp, dbuf_write_ready, children_ready_cb, 3584 dbuf_write_physdone, dbuf_write_done, db, 3585 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3586 } 3587} 3588