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#ifdef DEBUG 1486 if (dn->dn_objset->os_dsl_dataset != NULL) { 1487 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1488 RW_READER, FTAG); 1489 } 1490 ASSERT(!dmu_tx_is_syncing(tx) || 1491 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1492 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1493 dn->dn_objset->os_dsl_dataset == NULL); 1494 if (dn->dn_objset->os_dsl_dataset != NULL) 1495 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); 1496#endif 1497 /* 1498 * We make this assert for private objects as well, but after we 1499 * check if we're already dirty. They are allowed to re-dirty 1500 * in syncing context. 1501 */ 1502 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1503 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1504 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1505 1506 mutex_enter(&db->db_mtx); 1507 /* 1508 * XXX make this true for indirects too? The problem is that 1509 * transactions created with dmu_tx_create_assigned() from 1510 * syncing context don't bother holding ahead. 1511 */ 1512 ASSERT(db->db_level != 0 || 1513 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1514 db->db_state == DB_NOFILL); 1515 1516 mutex_enter(&dn->dn_mtx); 1517 /* 1518 * Don't set dirtyctx to SYNC if we're just modifying this as we 1519 * initialize the objset. 1520 */ 1521 if (dn->dn_dirtyctx == DN_UNDIRTIED) { 1522 if (dn->dn_objset->os_dsl_dataset != NULL) { 1523 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1524 RW_READER, FTAG); 1525 } 1526 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1527 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? 1528 DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1529 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1530 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1531 } 1532 if (dn->dn_objset->os_dsl_dataset != NULL) { 1533 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, 1534 FTAG); 1535 } 1536 } 1537 mutex_exit(&dn->dn_mtx); 1538 1539 if (db->db_blkid == DMU_SPILL_BLKID) 1540 dn->dn_have_spill = B_TRUE; 1541 1542 /* 1543 * If this buffer is already dirty, we're done. 1544 */ 1545 drp = &db->db_last_dirty; 1546 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1547 db->db.db_object == DMU_META_DNODE_OBJECT); 1548 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1549 drp = &dr->dr_next; 1550 if (dr && dr->dr_txg == tx->tx_txg) { 1551 DB_DNODE_EXIT(db); 1552 1553 dbuf_redirty(dr); 1554 mutex_exit(&db->db_mtx); 1555 return (dr); 1556 } 1557 1558 /* 1559 * Only valid if not already dirty. 1560 */ 1561 ASSERT(dn->dn_object == 0 || 1562 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1563 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1564 1565 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1566 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1567 dn->dn_phys->dn_nlevels > db->db_level || 1568 dn->dn_next_nlevels[txgoff] > db->db_level || 1569 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1570 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1571 1572 /* 1573 * We should only be dirtying in syncing context if it's the 1574 * mos or we're initializing the os or it's a special object. 1575 * However, we are allowed to dirty in syncing context provided 1576 * we already dirtied it in open context. Hence we must make 1577 * this assertion only if we're not already dirty. 1578 */ 1579 os = dn->dn_objset; 1580#ifdef DEBUG 1581 if (dn->dn_objset->os_dsl_dataset != NULL) 1582 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); 1583 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1584 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1585 if (dn->dn_objset->os_dsl_dataset != NULL) 1586 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); 1587#endif 1588 ASSERT(db->db.db_size != 0); 1589 1590 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1591 1592 if (db->db_blkid != DMU_BONUS_BLKID) { 1593 /* 1594 * Update the accounting. 1595 * Note: we delay "free accounting" until after we drop 1596 * the db_mtx. This keeps us from grabbing other locks 1597 * (and possibly deadlocking) in bp_get_dsize() while 1598 * also holding the db_mtx. 1599 */ 1600 dnode_willuse_space(dn, db->db.db_size, tx); 1601 do_free_accounting = dbuf_block_freeable(db); 1602 } 1603 1604 /* 1605 * If this buffer is dirty in an old transaction group we need 1606 * to make a copy of it so that the changes we make in this 1607 * transaction group won't leak out when we sync the older txg. 1608 */ 1609 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1610 if (db->db_level == 0) { 1611 void *data_old = db->db_buf; 1612 1613 if (db->db_state != DB_NOFILL) { 1614 if (db->db_blkid == DMU_BONUS_BLKID) { 1615 dbuf_fix_old_data(db, tx->tx_txg); 1616 data_old = db->db.db_data; 1617 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1618 /* 1619 * Release the data buffer from the cache so 1620 * that we can modify it without impacting 1621 * possible other users of this cached data 1622 * block. Note that indirect blocks and 1623 * private objects are not released until the 1624 * syncing state (since they are only modified 1625 * then). 1626 */ 1627 arc_release(db->db_buf, db); 1628 dbuf_fix_old_data(db, tx->tx_txg); 1629 data_old = db->db_buf; 1630 } 1631 ASSERT(data_old != NULL); 1632 } 1633 dr->dt.dl.dr_data = data_old; 1634 } else { 1635 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1636 list_create(&dr->dt.di.dr_children, 1637 sizeof (dbuf_dirty_record_t), 1638 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1639 } 1640 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1641 dr->dr_accounted = db->db.db_size; 1642 dr->dr_dbuf = db; 1643 dr->dr_txg = tx->tx_txg; 1644 dr->dr_next = *drp; 1645 *drp = dr; 1646 1647 /* 1648 * We could have been freed_in_flight between the dbuf_noread 1649 * and dbuf_dirty. We win, as though the dbuf_noread() had 1650 * happened after the free. 1651 */ 1652 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1653 db->db_blkid != DMU_SPILL_BLKID) { 1654 mutex_enter(&dn->dn_mtx); 1655 if (dn->dn_free_ranges[txgoff] != NULL) { 1656 range_tree_clear(dn->dn_free_ranges[txgoff], 1657 db->db_blkid, 1); 1658 } 1659 mutex_exit(&dn->dn_mtx); 1660 db->db_freed_in_flight = FALSE; 1661 } 1662 1663 /* 1664 * This buffer is now part of this txg 1665 */ 1666 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1667 db->db_dirtycnt += 1; 1668 ASSERT3U(db->db_dirtycnt, <=, 3); 1669 1670 mutex_exit(&db->db_mtx); 1671 1672 if (db->db_blkid == DMU_BONUS_BLKID || 1673 db->db_blkid == DMU_SPILL_BLKID) { 1674 mutex_enter(&dn->dn_mtx); 1675 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1676 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1677 mutex_exit(&dn->dn_mtx); 1678 dnode_setdirty(dn, tx); 1679 DB_DNODE_EXIT(db); 1680 return (dr); 1681 } 1682 1683 /* 1684 * The dn_struct_rwlock prevents db_blkptr from changing 1685 * due to a write from syncing context completing 1686 * while we are running, so we want to acquire it before 1687 * looking at db_blkptr. 1688 */ 1689 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1690 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1691 drop_struct_lock = TRUE; 1692 } 1693 1694 if (do_free_accounting) { 1695 blkptr_t *bp = db->db_blkptr; 1696 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1697 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1698 /* 1699 * This is only a guess -- if the dbuf is dirty 1700 * in a previous txg, we don't know how much 1701 * space it will use on disk yet. We should 1702 * really have the struct_rwlock to access 1703 * db_blkptr, but since this is just a guess, 1704 * it's OK if we get an odd answer. 1705 */ 1706 ddt_prefetch(os->os_spa, bp); 1707 dnode_willuse_space(dn, -willfree, tx); 1708 } 1709 1710 if (db->db_level == 0) { 1711 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1712 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1713 } 1714 1715 if (db->db_level+1 < dn->dn_nlevels) { 1716 dmu_buf_impl_t *parent = db->db_parent; 1717 dbuf_dirty_record_t *di; 1718 int parent_held = FALSE; 1719 1720 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1721 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1722 1723 parent = dbuf_hold_level(dn, db->db_level+1, 1724 db->db_blkid >> epbs, FTAG); 1725 ASSERT(parent != NULL); 1726 parent_held = TRUE; 1727 } 1728 if (drop_struct_lock) 1729 rw_exit(&dn->dn_struct_rwlock); 1730 ASSERT3U(db->db_level+1, ==, parent->db_level); 1731 di = dbuf_dirty(parent, tx); 1732 if (parent_held) 1733 dbuf_rele(parent, FTAG); 1734 1735 mutex_enter(&db->db_mtx); 1736 /* 1737 * Since we've dropped the mutex, it's possible that 1738 * dbuf_undirty() might have changed this out from under us. 1739 */ 1740 if (db->db_last_dirty == dr || 1741 dn->dn_object == DMU_META_DNODE_OBJECT) { 1742 mutex_enter(&di->dt.di.dr_mtx); 1743 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1744 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1745 list_insert_tail(&di->dt.di.dr_children, dr); 1746 mutex_exit(&di->dt.di.dr_mtx); 1747 dr->dr_parent = di; 1748 } 1749 mutex_exit(&db->db_mtx); 1750 } else { 1751 ASSERT(db->db_level+1 == dn->dn_nlevels); 1752 ASSERT(db->db_blkid < dn->dn_nblkptr); 1753 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1754 mutex_enter(&dn->dn_mtx); 1755 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1756 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1757 mutex_exit(&dn->dn_mtx); 1758 if (drop_struct_lock) 1759 rw_exit(&dn->dn_struct_rwlock); 1760 } 1761 1762 dnode_setdirty(dn, tx); 1763 DB_DNODE_EXIT(db); 1764 return (dr); 1765} 1766 1767/* 1768 * Undirty a buffer in the transaction group referenced by the given 1769 * transaction. Return whether this evicted the dbuf. 1770 */ 1771static boolean_t 1772dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1773{ 1774 dnode_t *dn; 1775 uint64_t txg = tx->tx_txg; 1776 dbuf_dirty_record_t *dr, **drp; 1777 1778 ASSERT(txg != 0); 1779 1780 /* 1781 * Due to our use of dn_nlevels below, this can only be called 1782 * in open context, unless we are operating on the MOS. 1783 * From syncing context, dn_nlevels may be different from the 1784 * dn_nlevels used when dbuf was dirtied. 1785 */ 1786 ASSERT(db->db_objset == 1787 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1788 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1789 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1790 ASSERT0(db->db_level); 1791 ASSERT(MUTEX_HELD(&db->db_mtx)); 1792 1793 /* 1794 * If this buffer is not dirty, we're done. 1795 */ 1796 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1797 if (dr->dr_txg <= txg) 1798 break; 1799 if (dr == NULL || dr->dr_txg < txg) 1800 return (B_FALSE); 1801 ASSERT(dr->dr_txg == txg); 1802 ASSERT(dr->dr_dbuf == db); 1803 1804 DB_DNODE_ENTER(db); 1805 dn = DB_DNODE(db); 1806 1807 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1808 1809 ASSERT(db->db.db_size != 0); 1810 1811 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1812 dr->dr_accounted, txg); 1813 1814 *drp = dr->dr_next; 1815 1816 /* 1817 * Note that there are three places in dbuf_dirty() 1818 * where this dirty record may be put on a list. 1819 * Make sure to do a list_remove corresponding to 1820 * every one of those list_insert calls. 1821 */ 1822 if (dr->dr_parent) { 1823 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1824 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1825 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1826 } else if (db->db_blkid == DMU_SPILL_BLKID || 1827 db->db_level + 1 == dn->dn_nlevels) { 1828 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1829 mutex_enter(&dn->dn_mtx); 1830 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1831 mutex_exit(&dn->dn_mtx); 1832 } 1833 DB_DNODE_EXIT(db); 1834 1835 if (db->db_state != DB_NOFILL) { 1836 dbuf_unoverride(dr); 1837 1838 ASSERT(db->db_buf != NULL); 1839 ASSERT(dr->dt.dl.dr_data != NULL); 1840 if (dr->dt.dl.dr_data != db->db_buf) 1841 arc_buf_destroy(dr->dt.dl.dr_data, db); 1842 } 1843 1844 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1845 1846 ASSERT(db->db_dirtycnt > 0); 1847 db->db_dirtycnt -= 1; 1848 1849 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1850 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); 1851 dbuf_destroy(db); 1852 return (B_TRUE); 1853 } 1854 1855 return (B_FALSE); 1856} 1857 1858void 1859dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1860{ 1861 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1862 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1863 1864 ASSERT(tx->tx_txg != 0); 1865 ASSERT(!refcount_is_zero(&db->db_holds)); 1866 1867 /* 1868 * Quick check for dirtyness. For already dirty blocks, this 1869 * reduces runtime of this function by >90%, and overall performance 1870 * by 50% for some workloads (e.g. file deletion with indirect blocks 1871 * cached). 1872 */ 1873 mutex_enter(&db->db_mtx); 1874 dbuf_dirty_record_t *dr; 1875 for (dr = db->db_last_dirty; 1876 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1877 /* 1878 * It's possible that it is already dirty but not cached, 1879 * because there are some calls to dbuf_dirty() that don't 1880 * go through dmu_buf_will_dirty(). 1881 */ 1882 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1883 /* This dbuf is already dirty and cached. */ 1884 dbuf_redirty(dr); 1885 mutex_exit(&db->db_mtx); 1886 return; 1887 } 1888 } 1889 mutex_exit(&db->db_mtx); 1890 1891 DB_DNODE_ENTER(db); 1892 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1893 rf |= DB_RF_HAVESTRUCT; 1894 DB_DNODE_EXIT(db); 1895 (void) dbuf_read(db, NULL, rf); 1896 (void) dbuf_dirty(db, tx); 1897} 1898 1899void 1900dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1901{ 1902 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1903 1904 db->db_state = DB_NOFILL; 1905 1906 dmu_buf_will_fill(db_fake, tx); 1907} 1908 1909void 1910dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1911{ 1912 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1913 1914 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1915 ASSERT(tx->tx_txg != 0); 1916 ASSERT(db->db_level == 0); 1917 ASSERT(!refcount_is_zero(&db->db_holds)); 1918 1919 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1920 dmu_tx_private_ok(tx)); 1921 1922 dbuf_noread(db); 1923 (void) dbuf_dirty(db, tx); 1924} 1925 1926#pragma weak dmu_buf_fill_done = dbuf_fill_done 1927/* ARGSUSED */ 1928void 1929dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1930{ 1931 mutex_enter(&db->db_mtx); 1932 DBUF_VERIFY(db); 1933 1934 if (db->db_state == DB_FILL) { 1935 if (db->db_level == 0 && db->db_freed_in_flight) { 1936 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1937 /* we were freed while filling */ 1938 /* XXX dbuf_undirty? */ 1939 bzero(db->db.db_data, db->db.db_size); 1940 db->db_freed_in_flight = FALSE; 1941 } 1942 db->db_state = DB_CACHED; 1943 cv_broadcast(&db->db_changed); 1944 } 1945 mutex_exit(&db->db_mtx); 1946} 1947 1948void 1949dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1950 bp_embedded_type_t etype, enum zio_compress comp, 1951 int uncompressed_size, int compressed_size, int byteorder, 1952 dmu_tx_t *tx) 1953{ 1954 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1955 struct dirty_leaf *dl; 1956 dmu_object_type_t type; 1957 1958 if (etype == BP_EMBEDDED_TYPE_DATA) { 1959 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1960 SPA_FEATURE_EMBEDDED_DATA)); 1961 } 1962 1963 DB_DNODE_ENTER(db); 1964 type = DB_DNODE(db)->dn_type; 1965 DB_DNODE_EXIT(db); 1966 1967 ASSERT0(db->db_level); 1968 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1969 1970 dmu_buf_will_not_fill(dbuf, tx); 1971 1972 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1973 dl = &db->db_last_dirty->dt.dl; 1974 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1975 data, comp, uncompressed_size, compressed_size); 1976 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1977 BP_SET_TYPE(&dl->dr_overridden_by, type); 1978 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1979 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1980 1981 dl->dr_override_state = DR_OVERRIDDEN; 1982 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1983} 1984 1985/* 1986 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1987 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1988 */ 1989void 1990dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1991{ 1992 ASSERT(!refcount_is_zero(&db->db_holds)); 1993 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1994 ASSERT(db->db_level == 0); 1995 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1996 ASSERT(buf != NULL); 1997 ASSERT(arc_buf_size(buf) == db->db.db_size); 1998 ASSERT(tx->tx_txg != 0); 1999 2000 arc_return_buf(buf, db); 2001 ASSERT(arc_released(buf)); 2002 2003 mutex_enter(&db->db_mtx); 2004 2005 while (db->db_state == DB_READ || db->db_state == DB_FILL) 2006 cv_wait(&db->db_changed, &db->db_mtx); 2007 2008 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 2009 2010 if (db->db_state == DB_CACHED && 2011 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 2012 mutex_exit(&db->db_mtx); 2013 (void) dbuf_dirty(db, tx); 2014 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 2015 arc_buf_destroy(buf, db); 2016 xuio_stat_wbuf_copied(); 2017 return; 2018 } 2019 2020 xuio_stat_wbuf_nocopy(); 2021 if (db->db_state == DB_CACHED) { 2022 dbuf_dirty_record_t *dr = db->db_last_dirty; 2023 2024 ASSERT(db->db_buf != NULL); 2025 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 2026 ASSERT(dr->dt.dl.dr_data == db->db_buf); 2027 if (!arc_released(db->db_buf)) { 2028 ASSERT(dr->dt.dl.dr_override_state == 2029 DR_OVERRIDDEN); 2030 arc_release(db->db_buf, db); 2031 } 2032 dr->dt.dl.dr_data = buf; 2033 arc_buf_destroy(db->db_buf, db); 2034 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 2035 arc_release(db->db_buf, db); 2036 arc_buf_destroy(db->db_buf, db); 2037 } 2038 db->db_buf = NULL; 2039 } 2040 ASSERT(db->db_buf == NULL); 2041 dbuf_set_data(db, buf); 2042 db->db_state = DB_FILL; 2043 mutex_exit(&db->db_mtx); 2044 (void) dbuf_dirty(db, tx); 2045 dmu_buf_fill_done(&db->db, tx); 2046} 2047 2048void 2049dbuf_destroy(dmu_buf_impl_t *db) 2050{ 2051 dnode_t *dn; 2052 dmu_buf_impl_t *parent = db->db_parent; 2053 dmu_buf_impl_t *dndb; 2054 2055 ASSERT(MUTEX_HELD(&db->db_mtx)); 2056 ASSERT(refcount_is_zero(&db->db_holds)); 2057 2058 if (db->db_buf != NULL) { 2059 arc_buf_destroy(db->db_buf, db); 2060 db->db_buf = NULL; 2061 } 2062 2063 if (db->db_blkid == DMU_BONUS_BLKID) { 2064 ASSERT(db->db.db_data != NULL); 2065 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 2066 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2067 db->db_state = DB_UNCACHED; 2068 } 2069 2070 dbuf_clear_data(db); 2071 2072 if (multilist_link_active(&db->db_cache_link)) { 2073 multilist_remove(&dbuf_cache, db); 2074 (void) refcount_remove_many(&dbuf_cache_size, 2075 db->db.db_size, db); 2076 } 2077 2078 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 2079 ASSERT(db->db_data_pending == NULL); 2080 2081 db->db_state = DB_EVICTING; 2082 db->db_blkptr = NULL; 2083 2084 /* 2085 * Now that db_state is DB_EVICTING, nobody else can find this via 2086 * the hash table. We can now drop db_mtx, which allows us to 2087 * acquire the dn_dbufs_mtx. 2088 */ 2089 mutex_exit(&db->db_mtx); 2090 2091 DB_DNODE_ENTER(db); 2092 dn = DB_DNODE(db); 2093 dndb = dn->dn_dbuf; 2094 if (db->db_blkid != DMU_BONUS_BLKID) { 2095 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); 2096 if (needlock) 2097 mutex_enter(&dn->dn_dbufs_mtx); 2098 avl_remove(&dn->dn_dbufs, db); 2099 atomic_dec_32(&dn->dn_dbufs_count); 2100 membar_producer(); 2101 DB_DNODE_EXIT(db); 2102 if (needlock) 2103 mutex_exit(&dn->dn_dbufs_mtx); 2104 /* 2105 * Decrementing the dbuf count means that the hold corresponding 2106 * to the removed dbuf is no longer discounted in dnode_move(), 2107 * so the dnode cannot be moved until after we release the hold. 2108 * The membar_producer() ensures visibility of the decremented 2109 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 2110 * release any lock. 2111 */ 2112 dnode_rele(dn, db); 2113 db->db_dnode_handle = NULL; 2114 2115 dbuf_hash_remove(db); 2116 } else { 2117 DB_DNODE_EXIT(db); 2118 } 2119 2120 ASSERT(refcount_is_zero(&db->db_holds)); 2121 2122 db->db_parent = NULL; 2123 2124 ASSERT(db->db_buf == NULL); 2125 ASSERT(db->db.db_data == NULL); 2126 ASSERT(db->db_hash_next == NULL); 2127 ASSERT(db->db_blkptr == NULL); 2128 ASSERT(db->db_data_pending == NULL); 2129 ASSERT(!multilist_link_active(&db->db_cache_link)); 2130 2131 kmem_cache_free(dbuf_kmem_cache, db); 2132 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2133 2134 /* 2135 * If this dbuf is referenced from an indirect dbuf, 2136 * decrement the ref count on the indirect dbuf. 2137 */ 2138 if (parent && parent != dndb) 2139 dbuf_rele(parent, db); 2140} 2141 2142/* 2143 * Note: While bpp will always be updated if the function returns success, 2144 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 2145 * this happens when the dnode is the meta-dnode, or a userused or groupused 2146 * object. 2147 */ 2148static int 2149dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 2150 dmu_buf_impl_t **parentp, blkptr_t **bpp) 2151{ 2152 int nlevels, epbs; 2153 2154 *parentp = NULL; 2155 *bpp = NULL; 2156 2157 ASSERT(blkid != DMU_BONUS_BLKID); 2158 2159 if (blkid == DMU_SPILL_BLKID) { 2160 mutex_enter(&dn->dn_mtx); 2161 if (dn->dn_have_spill && 2162 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 2163 *bpp = &dn->dn_phys->dn_spill; 2164 else 2165 *bpp = NULL; 2166 dbuf_add_ref(dn->dn_dbuf, NULL); 2167 *parentp = dn->dn_dbuf; 2168 mutex_exit(&dn->dn_mtx); 2169 return (0); 2170 } 2171 2172 if (dn->dn_phys->dn_nlevels == 0) 2173 nlevels = 1; 2174 else 2175 nlevels = dn->dn_phys->dn_nlevels; 2176 2177 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2178 2179 ASSERT3U(level * epbs, <, 64); 2180 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2181 if (level >= nlevels || 2182 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 2183 /* the buffer has no parent yet */ 2184 return (SET_ERROR(ENOENT)); 2185 } else if (level < nlevels-1) { 2186 /* this block is referenced from an indirect block */ 2187 int err = dbuf_hold_impl(dn, level+1, 2188 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 2189 if (err) 2190 return (err); 2191 err = dbuf_read(*parentp, NULL, 2192 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 2193 if (err) { 2194 dbuf_rele(*parentp, NULL); 2195 *parentp = NULL; 2196 return (err); 2197 } 2198 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 2199 (blkid & ((1ULL << epbs) - 1)); 2200 return (0); 2201 } else { 2202 /* the block is referenced from the dnode */ 2203 ASSERT3U(level, ==, nlevels-1); 2204 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 2205 blkid < dn->dn_phys->dn_nblkptr); 2206 if (dn->dn_dbuf) { 2207 dbuf_add_ref(dn->dn_dbuf, NULL); 2208 *parentp = dn->dn_dbuf; 2209 } 2210 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 2211 return (0); 2212 } 2213} 2214 2215static dmu_buf_impl_t * 2216dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 2217 dmu_buf_impl_t *parent, blkptr_t *blkptr) 2218{ 2219 objset_t *os = dn->dn_objset; 2220 dmu_buf_impl_t *db, *odb; 2221 2222 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2223 ASSERT(dn->dn_type != DMU_OT_NONE); 2224 2225 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); 2226 2227 db->db_objset = os; 2228 db->db.db_object = dn->dn_object; 2229 db->db_level = level; 2230 db->db_blkid = blkid; 2231 db->db_last_dirty = NULL; 2232 db->db_dirtycnt = 0; 2233 db->db_dnode_handle = dn->dn_handle; 2234 db->db_parent = parent; 2235 db->db_blkptr = blkptr; 2236 2237 db->db_user = NULL; 2238 db->db_user_immediate_evict = FALSE; 2239 db->db_freed_in_flight = FALSE; 2240 db->db_pending_evict = FALSE; 2241 2242 if (blkid == DMU_BONUS_BLKID) { 2243 ASSERT3P(parent, ==, dn->dn_dbuf); 2244 db->db.db_size = DN_MAX_BONUSLEN - 2245 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 2246 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 2247 db->db.db_offset = DMU_BONUS_BLKID; 2248 db->db_state = DB_UNCACHED; 2249 /* the bonus dbuf is not placed in the hash table */ 2250 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2251 return (db); 2252 } else if (blkid == DMU_SPILL_BLKID) { 2253 db->db.db_size = (blkptr != NULL) ? 2254 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 2255 db->db.db_offset = 0; 2256 } else { 2257 int blocksize = 2258 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 2259 db->db.db_size = blocksize; 2260 db->db.db_offset = db->db_blkid * blocksize; 2261 } 2262 2263 /* 2264 * Hold the dn_dbufs_mtx while we get the new dbuf 2265 * in the hash table *and* added to the dbufs list. 2266 * This prevents a possible deadlock with someone 2267 * trying to look up this dbuf before its added to the 2268 * dn_dbufs list. 2269 */ 2270 mutex_enter(&dn->dn_dbufs_mtx); 2271 db->db_state = DB_EVICTING; 2272 if ((odb = dbuf_hash_insert(db)) != NULL) { 2273 /* someone else inserted it first */ 2274 kmem_cache_free(dbuf_kmem_cache, db); 2275 mutex_exit(&dn->dn_dbufs_mtx); 2276 return (odb); 2277 } 2278 avl_add(&dn->dn_dbufs, db); 2279 if (db->db_level == 0 && db->db_blkid >= 2280 dn->dn_unlisted_l0_blkid) 2281 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 2282 db->db_state = DB_UNCACHED; 2283 mutex_exit(&dn->dn_dbufs_mtx); 2284 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2285 2286 if (parent && parent != dn->dn_dbuf) 2287 dbuf_add_ref(parent, db); 2288 2289 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 2290 refcount_count(&dn->dn_holds) > 0); 2291 (void) refcount_add(&dn->dn_holds, db); 2292 atomic_inc_32(&dn->dn_dbufs_count); 2293 2294 dprintf_dbuf(db, "db=%p\n", db); 2295 2296 return (db); 2297} 2298 2299typedef struct dbuf_prefetch_arg { 2300 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2301 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2302 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2303 int dpa_curlevel; /* The current level that we're reading */ 2304 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ 2305 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2306 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2307 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2308} dbuf_prefetch_arg_t; 2309 2310/* 2311 * Actually issue the prefetch read for the block given. 2312 */ 2313static void 2314dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2315{ 2316 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2317 return; 2318 2319 arc_flags_t aflags = 2320 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2321 2322 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2323 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2324 ASSERT(dpa->dpa_zio != NULL); 2325 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2326 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2327 &aflags, &dpa->dpa_zb); 2328} 2329 2330/* 2331 * Called when an indirect block above our prefetch target is read in. This 2332 * will either read in the next indirect block down the tree or issue the actual 2333 * prefetch if the next block down is our target. 2334 */ 2335static void 2336dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2337{ 2338 dbuf_prefetch_arg_t *dpa = private; 2339 2340 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2341 ASSERT3S(dpa->dpa_curlevel, >, 0); 2342 2343 /* 2344 * The dpa_dnode is only valid if we are called with a NULL 2345 * zio. This indicates that the arc_read() returned without 2346 * first calling zio_read() to issue a physical read. Once 2347 * a physical read is made the dpa_dnode must be invalidated 2348 * as the locks guarding it may have been dropped. If the 2349 * dpa_dnode is still valid, then we want to add it to the dbuf 2350 * cache. To do so, we must hold the dbuf associated with the block 2351 * we just prefetched, read its contents so that we associate it 2352 * with an arc_buf_t, and then release it. 2353 */ 2354 if (zio != NULL) { 2355 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2356 if (zio->io_flags & ZIO_FLAG_RAW) { 2357 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); 2358 } else { 2359 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2360 } 2361 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2362 2363 dpa->dpa_dnode = NULL; 2364 } else if (dpa->dpa_dnode != NULL) { 2365 uint64_t curblkid = dpa->dpa_zb.zb_blkid >> 2366 (dpa->dpa_epbs * (dpa->dpa_curlevel - 2367 dpa->dpa_zb.zb_level)); 2368 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, 2369 dpa->dpa_curlevel, curblkid, FTAG); 2370 (void) dbuf_read(db, NULL, 2371 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); 2372 dbuf_rele(db, FTAG); 2373 } 2374 2375 dpa->dpa_curlevel--; 2376 2377 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2378 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2379 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2380 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2381 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2382 kmem_free(dpa, sizeof (*dpa)); 2383 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2384 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2385 dbuf_issue_final_prefetch(dpa, bp); 2386 kmem_free(dpa, sizeof (*dpa)); 2387 } else { 2388 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2389 zbookmark_phys_t zb; 2390 2391 /* flag if L2ARC eligible, l2arc_noprefetch then decides */ 2392 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) 2393 iter_aflags |= ARC_FLAG_L2CACHE; 2394 2395 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2396 2397 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2398 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2399 2400 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2401 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2402 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2403 &iter_aflags, &zb); 2404 } 2405 2406 arc_buf_destroy(abuf, private); 2407} 2408 2409/* 2410 * Issue prefetch reads for the given block on the given level. If the indirect 2411 * blocks above that block are not in memory, we will read them in 2412 * asynchronously. As a result, this call never blocks waiting for a read to 2413 * complete. 2414 */ 2415void 2416dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2417 arc_flags_t aflags) 2418{ 2419 blkptr_t bp; 2420 int epbs, nlevels, curlevel; 2421 uint64_t curblkid; 2422 2423 ASSERT(blkid != DMU_BONUS_BLKID); 2424 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2425 2426 if (blkid > dn->dn_maxblkid) 2427 return; 2428 2429 if (dnode_block_freed(dn, blkid)) 2430 return; 2431 2432 /* 2433 * This dnode hasn't been written to disk yet, so there's nothing to 2434 * prefetch. 2435 */ 2436 nlevels = dn->dn_phys->dn_nlevels; 2437 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2438 return; 2439 2440 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2441 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2442 return; 2443 2444 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2445 level, blkid); 2446 if (db != NULL) { 2447 mutex_exit(&db->db_mtx); 2448 /* 2449 * This dbuf already exists. It is either CACHED, or 2450 * (we assume) about to be read or filled. 2451 */ 2452 return; 2453 } 2454 2455 /* 2456 * Find the closest ancestor (indirect block) of the target block 2457 * that is present in the cache. In this indirect block, we will 2458 * find the bp that is at curlevel, curblkid. 2459 */ 2460 curlevel = level; 2461 curblkid = blkid; 2462 while (curlevel < nlevels - 1) { 2463 int parent_level = curlevel + 1; 2464 uint64_t parent_blkid = curblkid >> epbs; 2465 dmu_buf_impl_t *db; 2466 2467 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2468 FALSE, TRUE, FTAG, &db) == 0) { 2469 blkptr_t *bpp = db->db_buf->b_data; 2470 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2471 dbuf_rele(db, FTAG); 2472 break; 2473 } 2474 2475 curlevel = parent_level; 2476 curblkid = parent_blkid; 2477 } 2478 2479 if (curlevel == nlevels - 1) { 2480 /* No cached indirect blocks found. */ 2481 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2482 bp = dn->dn_phys->dn_blkptr[curblkid]; 2483 } 2484 if (BP_IS_HOLE(&bp)) 2485 return; 2486 2487 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2488 2489 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2490 ZIO_FLAG_CANFAIL); 2491 2492 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2493 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2494 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2495 dn->dn_object, level, blkid); 2496 dpa->dpa_curlevel = curlevel; 2497 dpa->dpa_prio = prio; 2498 dpa->dpa_aflags = aflags; 2499 dpa->dpa_spa = dn->dn_objset->os_spa; 2500 dpa->dpa_dnode = dn; 2501 dpa->dpa_epbs = epbs; 2502 dpa->dpa_zio = pio; 2503 2504 /* flag if L2ARC eligible, l2arc_noprefetch then decides */ 2505 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) 2506 dpa->dpa_aflags |= ARC_FLAG_L2CACHE; 2507 2508 /* 2509 * If we have the indirect just above us, no need to do the asynchronous 2510 * prefetch chain; we'll just run the last step ourselves. If we're at 2511 * a higher level, though, we want to issue the prefetches for all the 2512 * indirect blocks asynchronously, so we can go on with whatever we were 2513 * doing. 2514 */ 2515 if (curlevel == level) { 2516 ASSERT3U(curblkid, ==, blkid); 2517 dbuf_issue_final_prefetch(dpa, &bp); 2518 kmem_free(dpa, sizeof (*dpa)); 2519 } else { 2520 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2521 zbookmark_phys_t zb; 2522 2523 /* flag if L2ARC eligible, l2arc_noprefetch then decides */ 2524 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) 2525 iter_aflags |= ARC_FLAG_L2CACHE; 2526 2527 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2528 dn->dn_object, curlevel, curblkid); 2529 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2530 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2531 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2532 &iter_aflags, &zb); 2533 } 2534 /* 2535 * We use pio here instead of dpa_zio since it's possible that 2536 * dpa may have already been freed. 2537 */ 2538 zio_nowait(pio); 2539} 2540 2541/* 2542 * Returns with db_holds incremented, and db_mtx not held. 2543 * Note: dn_struct_rwlock must be held. 2544 */ 2545int 2546dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2547 boolean_t fail_sparse, boolean_t fail_uncached, 2548 void *tag, dmu_buf_impl_t **dbp) 2549{ 2550 dmu_buf_impl_t *db, *parent = NULL; 2551 2552 ASSERT(blkid != DMU_BONUS_BLKID); 2553 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2554 ASSERT3U(dn->dn_nlevels, >, level); 2555 2556 *dbp = NULL; 2557top: 2558 /* dbuf_find() returns with db_mtx held */ 2559 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2560 2561 if (db == NULL) { 2562 blkptr_t *bp = NULL; 2563 int err; 2564 2565 if (fail_uncached) 2566 return (SET_ERROR(ENOENT)); 2567 2568 ASSERT3P(parent, ==, NULL); 2569 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2570 if (fail_sparse) { 2571 if (err == 0 && bp && BP_IS_HOLE(bp)) 2572 err = SET_ERROR(ENOENT); 2573 if (err) { 2574 if (parent) 2575 dbuf_rele(parent, NULL); 2576 return (err); 2577 } 2578 } 2579 if (err && err != ENOENT) 2580 return (err); 2581 db = dbuf_create(dn, level, blkid, parent, bp); 2582 } 2583 2584 if (fail_uncached && db->db_state != DB_CACHED) { 2585 mutex_exit(&db->db_mtx); 2586 return (SET_ERROR(ENOENT)); 2587 } 2588 2589 if (db->db_buf != NULL) 2590 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2591 2592 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2593 2594 /* 2595 * If this buffer is currently syncing out, and we are are 2596 * still referencing it from db_data, we need to make a copy 2597 * of it in case we decide we want to dirty it again in this txg. 2598 */ 2599 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2600 dn->dn_object != DMU_META_DNODE_OBJECT && 2601 db->db_state == DB_CACHED && db->db_data_pending) { 2602 dbuf_dirty_record_t *dr = db->db_data_pending; 2603 2604 if (dr->dt.dl.dr_data == db->db_buf) { 2605 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2606 2607 dbuf_set_data(db, 2608 arc_alloc_buf(dn->dn_objset->os_spa, 2609 db->db.db_size, db, type)); 2610 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2611 db->db.db_size); 2612 } 2613 } 2614 2615 if (multilist_link_active(&db->db_cache_link)) { 2616 ASSERT(refcount_is_zero(&db->db_holds)); 2617 multilist_remove(&dbuf_cache, db); 2618 (void) refcount_remove_many(&dbuf_cache_size, 2619 db->db.db_size, db); 2620 } 2621 (void) refcount_add(&db->db_holds, tag); 2622 DBUF_VERIFY(db); 2623 mutex_exit(&db->db_mtx); 2624 2625 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2626 if (parent) 2627 dbuf_rele(parent, NULL); 2628 2629 ASSERT3P(DB_DNODE(db), ==, dn); 2630 ASSERT3U(db->db_blkid, ==, blkid); 2631 ASSERT3U(db->db_level, ==, level); 2632 *dbp = db; 2633 2634 return (0); 2635} 2636 2637dmu_buf_impl_t * 2638dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2639{ 2640 return (dbuf_hold_level(dn, 0, blkid, tag)); 2641} 2642 2643dmu_buf_impl_t * 2644dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2645{ 2646 dmu_buf_impl_t *db; 2647 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2648 return (err ? NULL : db); 2649} 2650 2651void 2652dbuf_create_bonus(dnode_t *dn) 2653{ 2654 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2655 2656 ASSERT(dn->dn_bonus == NULL); 2657 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2658} 2659 2660int 2661dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2662{ 2663 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2664 dnode_t *dn; 2665 2666 if (db->db_blkid != DMU_SPILL_BLKID) 2667 return (SET_ERROR(ENOTSUP)); 2668 if (blksz == 0) 2669 blksz = SPA_MINBLOCKSIZE; 2670 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2671 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2672 2673 DB_DNODE_ENTER(db); 2674 dn = DB_DNODE(db); 2675 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2676 dbuf_new_size(db, blksz, tx); 2677 rw_exit(&dn->dn_struct_rwlock); 2678 DB_DNODE_EXIT(db); 2679 2680 return (0); 2681} 2682 2683void 2684dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2685{ 2686 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2687} 2688 2689#pragma weak dmu_buf_add_ref = dbuf_add_ref 2690void 2691dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2692{ 2693 int64_t holds = refcount_add(&db->db_holds, tag); 2694 ASSERT3S(holds, >, 1); 2695} 2696 2697#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2698boolean_t 2699dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2700 void *tag) 2701{ 2702 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2703 dmu_buf_impl_t *found_db; 2704 boolean_t result = B_FALSE; 2705 2706 if (db->db_blkid == DMU_BONUS_BLKID) 2707 found_db = dbuf_find_bonus(os, obj); 2708 else 2709 found_db = dbuf_find(os, obj, 0, blkid); 2710 2711 if (found_db != NULL) { 2712 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2713 (void) refcount_add(&db->db_holds, tag); 2714 result = B_TRUE; 2715 } 2716 mutex_exit(&db->db_mtx); 2717 } 2718 return (result); 2719} 2720 2721/* 2722 * If you call dbuf_rele() you had better not be referencing the dnode handle 2723 * unless you have some other direct or indirect hold on the dnode. (An indirect 2724 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2725 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2726 * dnode's parent dbuf evicting its dnode handles. 2727 */ 2728void 2729dbuf_rele(dmu_buf_impl_t *db, void *tag) 2730{ 2731 mutex_enter(&db->db_mtx); 2732 dbuf_rele_and_unlock(db, tag); 2733} 2734 2735void 2736dmu_buf_rele(dmu_buf_t *db, void *tag) 2737{ 2738 dbuf_rele((dmu_buf_impl_t *)db, tag); 2739} 2740 2741/* 2742 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2743 * db_dirtycnt and db_holds to be updated atomically. 2744 */ 2745void 2746dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2747{ 2748 int64_t holds; 2749 2750 ASSERT(MUTEX_HELD(&db->db_mtx)); 2751 DBUF_VERIFY(db); 2752 2753 /* 2754 * Remove the reference to the dbuf before removing its hold on the 2755 * dnode so we can guarantee in dnode_move() that a referenced bonus 2756 * buffer has a corresponding dnode hold. 2757 */ 2758 holds = refcount_remove(&db->db_holds, tag); 2759 ASSERT(holds >= 0); 2760 2761 /* 2762 * We can't freeze indirects if there is a possibility that they 2763 * may be modified in the current syncing context. 2764 */ 2765 if (db->db_buf != NULL && 2766 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { 2767 arc_buf_freeze(db->db_buf); 2768 } 2769 2770 if (holds == db->db_dirtycnt && 2771 db->db_level == 0 && db->db_user_immediate_evict) 2772 dbuf_evict_user(db); 2773 2774 if (holds == 0) { 2775 if (db->db_blkid == DMU_BONUS_BLKID) { 2776 dnode_t *dn; 2777 boolean_t evict_dbuf = db->db_pending_evict; 2778 2779 /* 2780 * If the dnode moves here, we cannot cross this 2781 * barrier until the move completes. 2782 */ 2783 DB_DNODE_ENTER(db); 2784 2785 dn = DB_DNODE(db); 2786 atomic_dec_32(&dn->dn_dbufs_count); 2787 2788 /* 2789 * Decrementing the dbuf count means that the bonus 2790 * buffer's dnode hold is no longer discounted in 2791 * dnode_move(). The dnode cannot move until after 2792 * the dnode_rele() below. 2793 */ 2794 DB_DNODE_EXIT(db); 2795 2796 /* 2797 * Do not reference db after its lock is dropped. 2798 * Another thread may evict it. 2799 */ 2800 mutex_exit(&db->db_mtx); 2801 2802 if (evict_dbuf) 2803 dnode_evict_bonus(dn); 2804 2805 dnode_rele(dn, db); 2806 } else if (db->db_buf == NULL) { 2807 /* 2808 * This is a special case: we never associated this 2809 * dbuf with any data allocated from the ARC. 2810 */ 2811 ASSERT(db->db_state == DB_UNCACHED || 2812 db->db_state == DB_NOFILL); 2813 dbuf_destroy(db); 2814 } else if (arc_released(db->db_buf)) { 2815 /* 2816 * This dbuf has anonymous data associated with it. 2817 */ 2818 dbuf_destroy(db); 2819 } else { 2820 boolean_t do_arc_evict = B_FALSE; 2821 blkptr_t bp; 2822 spa_t *spa = dmu_objset_spa(db->db_objset); 2823 2824 if (!DBUF_IS_CACHEABLE(db) && 2825 db->db_blkptr != NULL && 2826 !BP_IS_HOLE(db->db_blkptr) && 2827 !BP_IS_EMBEDDED(db->db_blkptr)) { 2828 do_arc_evict = B_TRUE; 2829 bp = *db->db_blkptr; 2830 } 2831 2832 if (!DBUF_IS_CACHEABLE(db) || 2833 db->db_pending_evict) { 2834 dbuf_destroy(db); 2835 } else if (!multilist_link_active(&db->db_cache_link)) { 2836 multilist_insert(&dbuf_cache, db); 2837 (void) refcount_add_many(&dbuf_cache_size, 2838 db->db.db_size, db); 2839 mutex_exit(&db->db_mtx); 2840 2841 dbuf_evict_notify(); 2842 } 2843 2844 if (do_arc_evict) 2845 arc_freed(spa, &bp); 2846 } 2847 } else { 2848 mutex_exit(&db->db_mtx); 2849 } 2850 2851} 2852 2853#pragma weak dmu_buf_refcount = dbuf_refcount 2854uint64_t 2855dbuf_refcount(dmu_buf_impl_t *db) 2856{ 2857 return (refcount_count(&db->db_holds)); 2858} 2859 2860void * 2861dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2862 dmu_buf_user_t *new_user) 2863{ 2864 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2865 2866 mutex_enter(&db->db_mtx); 2867 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2868 if (db->db_user == old_user) 2869 db->db_user = new_user; 2870 else 2871 old_user = db->db_user; 2872 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2873 mutex_exit(&db->db_mtx); 2874 2875 return (old_user); 2876} 2877 2878void * 2879dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2880{ 2881 return (dmu_buf_replace_user(db_fake, NULL, user)); 2882} 2883 2884void * 2885dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2886{ 2887 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2888 2889 db->db_user_immediate_evict = TRUE; 2890 return (dmu_buf_set_user(db_fake, user)); 2891} 2892 2893void * 2894dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2895{ 2896 return (dmu_buf_replace_user(db_fake, user, NULL)); 2897} 2898 2899void * 2900dmu_buf_get_user(dmu_buf_t *db_fake) 2901{ 2902 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2903 2904 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2905 return (db->db_user); 2906} 2907 2908void 2909dmu_buf_user_evict_wait() 2910{ 2911 taskq_wait(dbu_evict_taskq); 2912} 2913 2914boolean_t 2915dmu_buf_freeable(dmu_buf_t *dbuf) 2916{ 2917 boolean_t res = B_FALSE; 2918 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2919 2920 if (db->db_blkptr) 2921 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2922 db->db_blkptr, db->db_blkptr->blk_birth); 2923 2924 return (res); 2925} 2926 2927blkptr_t * 2928dmu_buf_get_blkptr(dmu_buf_t *db) 2929{ 2930 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2931 return (dbi->db_blkptr); 2932} 2933 2934objset_t * 2935dmu_buf_get_objset(dmu_buf_t *db) 2936{ 2937 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2938 return (dbi->db_objset); 2939} 2940 2941dnode_t * 2942dmu_buf_dnode_enter(dmu_buf_t *db) 2943{ 2944 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2945 DB_DNODE_ENTER(dbi); 2946 return (DB_DNODE(dbi)); 2947} 2948 2949void 2950dmu_buf_dnode_exit(dmu_buf_t *db) 2951{ 2952 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2953 DB_DNODE_EXIT(dbi); 2954} 2955 2956static void 2957dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2958{ 2959 /* ASSERT(dmu_tx_is_syncing(tx) */ 2960 ASSERT(MUTEX_HELD(&db->db_mtx)); 2961 2962 if (db->db_blkptr != NULL) 2963 return; 2964 2965 if (db->db_blkid == DMU_SPILL_BLKID) { 2966 db->db_blkptr = &dn->dn_phys->dn_spill; 2967 BP_ZERO(db->db_blkptr); 2968 return; 2969 } 2970 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2971 /* 2972 * This buffer was allocated at a time when there was 2973 * no available blkptrs from the dnode, or it was 2974 * inappropriate to hook it in (i.e., nlevels mis-match). 2975 */ 2976 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2977 ASSERT(db->db_parent == NULL); 2978 db->db_parent = dn->dn_dbuf; 2979 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2980 DBUF_VERIFY(db); 2981 } else { 2982 dmu_buf_impl_t *parent = db->db_parent; 2983 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2984 2985 ASSERT(dn->dn_phys->dn_nlevels > 1); 2986 if (parent == NULL) { 2987 mutex_exit(&db->db_mtx); 2988 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2989 parent = dbuf_hold_level(dn, db->db_level + 1, 2990 db->db_blkid >> epbs, db); 2991 rw_exit(&dn->dn_struct_rwlock); 2992 mutex_enter(&db->db_mtx); 2993 db->db_parent = parent; 2994 } 2995 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2996 (db->db_blkid & ((1ULL << epbs) - 1)); 2997 DBUF_VERIFY(db); 2998 } 2999} 3000 3001static void 3002dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3003{ 3004 dmu_buf_impl_t *db = dr->dr_dbuf; 3005 dnode_t *dn; 3006 zio_t *zio; 3007 3008 ASSERT(dmu_tx_is_syncing(tx)); 3009 3010 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3011 3012 mutex_enter(&db->db_mtx); 3013 3014 ASSERT(db->db_level > 0); 3015 DBUF_VERIFY(db); 3016 3017 /* Read the block if it hasn't been read yet. */ 3018 if (db->db_buf == NULL) { 3019 mutex_exit(&db->db_mtx); 3020 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 3021 mutex_enter(&db->db_mtx); 3022 } 3023 ASSERT3U(db->db_state, ==, DB_CACHED); 3024 ASSERT(db->db_buf != NULL); 3025 3026 DB_DNODE_ENTER(db); 3027 dn = DB_DNODE(db); 3028 /* Indirect block size must match what the dnode thinks it is. */ 3029 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3030 dbuf_check_blkptr(dn, db); 3031 DB_DNODE_EXIT(db); 3032 3033 /* Provide the pending dirty record to child dbufs */ 3034 db->db_data_pending = dr; 3035 3036 mutex_exit(&db->db_mtx); 3037 dbuf_write(dr, db->db_buf, tx); 3038 3039 zio = dr->dr_zio; 3040 mutex_enter(&dr->dt.di.dr_mtx); 3041 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 3042 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3043 mutex_exit(&dr->dt.di.dr_mtx); 3044 zio_nowait(zio); 3045} 3046 3047static void 3048dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 3049{ 3050 arc_buf_t **datap = &dr->dt.dl.dr_data; 3051 dmu_buf_impl_t *db = dr->dr_dbuf; 3052 dnode_t *dn; 3053 objset_t *os; 3054 uint64_t txg = tx->tx_txg; 3055 3056 ASSERT(dmu_tx_is_syncing(tx)); 3057 3058 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 3059 3060 mutex_enter(&db->db_mtx); 3061 /* 3062 * To be synced, we must be dirtied. But we 3063 * might have been freed after the dirty. 3064 */ 3065 if (db->db_state == DB_UNCACHED) { 3066 /* This buffer has been freed since it was dirtied */ 3067 ASSERT(db->db.db_data == NULL); 3068 } else if (db->db_state == DB_FILL) { 3069 /* This buffer was freed and is now being re-filled */ 3070 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 3071 } else { 3072 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 3073 } 3074 DBUF_VERIFY(db); 3075 3076 DB_DNODE_ENTER(db); 3077 dn = DB_DNODE(db); 3078 3079 if (db->db_blkid == DMU_SPILL_BLKID) { 3080 mutex_enter(&dn->dn_mtx); 3081 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 3082 mutex_exit(&dn->dn_mtx); 3083 } 3084 3085 /* 3086 * If this is a bonus buffer, simply copy the bonus data into the 3087 * dnode. It will be written out when the dnode is synced (and it 3088 * will be synced, since it must have been dirty for dbuf_sync to 3089 * be called). 3090 */ 3091 if (db->db_blkid == DMU_BONUS_BLKID) { 3092 dbuf_dirty_record_t **drp; 3093 3094 ASSERT(*datap != NULL); 3095 ASSERT0(db->db_level); 3096 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 3097 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 3098 DB_DNODE_EXIT(db); 3099 3100 if (*datap != db->db.db_data) { 3101 zio_buf_free(*datap, DN_MAX_BONUSLEN); 3102 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 3103 } 3104 db->db_data_pending = NULL; 3105 drp = &db->db_last_dirty; 3106 while (*drp != dr) 3107 drp = &(*drp)->dr_next; 3108 ASSERT(dr->dr_next == NULL); 3109 ASSERT(dr->dr_dbuf == db); 3110 *drp = dr->dr_next; 3111 if (dr->dr_dbuf->db_level != 0) { 3112 list_destroy(&dr->dt.di.dr_children); 3113 mutex_destroy(&dr->dt.di.dr_mtx); 3114 } 3115 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3116 ASSERT(db->db_dirtycnt > 0); 3117 db->db_dirtycnt -= 1; 3118 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 3119 return; 3120 } 3121 3122 os = dn->dn_objset; 3123 3124 /* 3125 * This function may have dropped the db_mtx lock allowing a dmu_sync 3126 * operation to sneak in. As a result, we need to ensure that we 3127 * don't check the dr_override_state until we have returned from 3128 * dbuf_check_blkptr. 3129 */ 3130 dbuf_check_blkptr(dn, db); 3131 3132 /* 3133 * If this buffer is in the middle of an immediate write, 3134 * wait for the synchronous IO to complete. 3135 */ 3136 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 3137 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 3138 cv_wait(&db->db_changed, &db->db_mtx); 3139 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 3140 } 3141 3142 if (db->db_state != DB_NOFILL && 3143 dn->dn_object != DMU_META_DNODE_OBJECT && 3144 refcount_count(&db->db_holds) > 1 && 3145 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 3146 *datap == db->db_buf) { 3147 /* 3148 * If this buffer is currently "in use" (i.e., there 3149 * are active holds and db_data still references it), 3150 * then make a copy before we start the write so that 3151 * any modifications from the open txg will not leak 3152 * into this write. 3153 * 3154 * NOTE: this copy does not need to be made for 3155 * objects only modified in the syncing context (e.g. 3156 * DNONE_DNODE blocks). 3157 */ 3158 int blksz = arc_buf_size(*datap); 3159 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 3160 *datap = arc_alloc_buf(os->os_spa, blksz, db, type); 3161 bcopy(db->db.db_data, (*datap)->b_data, blksz); 3162 } 3163 db->db_data_pending = dr; 3164 3165 mutex_exit(&db->db_mtx); 3166 3167 dbuf_write(dr, *datap, tx); 3168 3169 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3170 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 3171 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 3172 DB_DNODE_EXIT(db); 3173 } else { 3174 /* 3175 * Although zio_nowait() does not "wait for an IO", it does 3176 * initiate the IO. If this is an empty write it seems plausible 3177 * that the IO could actually be completed before the nowait 3178 * returns. We need to DB_DNODE_EXIT() first in case 3179 * zio_nowait() invalidates the dbuf. 3180 */ 3181 DB_DNODE_EXIT(db); 3182 zio_nowait(dr->dr_zio); 3183 } 3184} 3185 3186void 3187dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 3188{ 3189 dbuf_dirty_record_t *dr; 3190 3191 while (dr = list_head(list)) { 3192 if (dr->dr_zio != NULL) { 3193 /* 3194 * If we find an already initialized zio then we 3195 * are processing the meta-dnode, and we have finished. 3196 * The dbufs for all dnodes are put back on the list 3197 * during processing, so that we can zio_wait() 3198 * these IOs after initiating all child IOs. 3199 */ 3200 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 3201 DMU_META_DNODE_OBJECT); 3202 break; 3203 } 3204 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 3205 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 3206 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 3207 } 3208 list_remove(list, dr); 3209 if (dr->dr_dbuf->db_level > 0) 3210 dbuf_sync_indirect(dr, tx); 3211 else 3212 dbuf_sync_leaf(dr, tx); 3213 } 3214} 3215 3216/* ARGSUSED */ 3217static void 3218dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3219{ 3220 dmu_buf_impl_t *db = vdb; 3221 dnode_t *dn; 3222 blkptr_t *bp = zio->io_bp; 3223 blkptr_t *bp_orig = &zio->io_bp_orig; 3224 spa_t *spa = zio->io_spa; 3225 int64_t delta; 3226 uint64_t fill = 0; 3227 int i; 3228 3229 ASSERT3P(db->db_blkptr, !=, NULL); 3230 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); 3231 3232 DB_DNODE_ENTER(db); 3233 dn = DB_DNODE(db); 3234 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 3235 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 3236 zio->io_prev_space_delta = delta; 3237 3238 if (bp->blk_birth != 0) { 3239 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 3240 BP_GET_TYPE(bp) == dn->dn_type) || 3241 (db->db_blkid == DMU_SPILL_BLKID && 3242 BP_GET_TYPE(bp) == dn->dn_bonustype) || 3243 BP_IS_EMBEDDED(bp)); 3244 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 3245 } 3246 3247 mutex_enter(&db->db_mtx); 3248 3249#ifdef ZFS_DEBUG 3250 if (db->db_blkid == DMU_SPILL_BLKID) { 3251 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3252 ASSERT(!(BP_IS_HOLE(bp)) && 3253 db->db_blkptr == &dn->dn_phys->dn_spill); 3254 } 3255#endif 3256 3257 if (db->db_level == 0) { 3258 mutex_enter(&dn->dn_mtx); 3259 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 3260 db->db_blkid != DMU_SPILL_BLKID) 3261 dn->dn_phys->dn_maxblkid = db->db_blkid; 3262 mutex_exit(&dn->dn_mtx); 3263 3264 if (dn->dn_type == DMU_OT_DNODE) { 3265 dnode_phys_t *dnp = db->db.db_data; 3266 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 3267 i--, dnp++) { 3268 if (dnp->dn_type != DMU_OT_NONE) 3269 fill++; 3270 } 3271 } else { 3272 if (BP_IS_HOLE(bp)) { 3273 fill = 0; 3274 } else { 3275 fill = 1; 3276 } 3277 } 3278 } else { 3279 blkptr_t *ibp = db->db.db_data; 3280 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 3281 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 3282 if (BP_IS_HOLE(ibp)) 3283 continue; 3284 fill += BP_GET_FILL(ibp); 3285 } 3286 } 3287 DB_DNODE_EXIT(db); 3288 3289 if (!BP_IS_EMBEDDED(bp)) 3290 bp->blk_fill = fill; 3291 3292 mutex_exit(&db->db_mtx); 3293 3294 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3295 *db->db_blkptr = *bp; 3296 rw_exit(&dn->dn_struct_rwlock); 3297} 3298 3299/* ARGSUSED */ 3300/* 3301 * This function gets called just prior to running through the compression 3302 * stage of the zio pipeline. If we're an indirect block comprised of only 3303 * holes, then we want this indirect to be compressed away to a hole. In 3304 * order to do that we must zero out any information about the holes that 3305 * this indirect points to prior to before we try to compress it. 3306 */ 3307static void 3308dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3309{ 3310 dmu_buf_impl_t *db = vdb; 3311 dnode_t *dn; 3312 blkptr_t *bp; 3313 uint64_t i; 3314 int epbs; 3315 3316 ASSERT3U(db->db_level, >, 0); 3317 DB_DNODE_ENTER(db); 3318 dn = DB_DNODE(db); 3319 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3320 3321 /* Determine if all our children are holes */ 3322 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { 3323 if (!BP_IS_HOLE(bp)) 3324 break; 3325 } 3326 3327 /* 3328 * If all the children are holes, then zero them all out so that 3329 * we may get compressed away. 3330 */ 3331 if (i == 1 << epbs) { 3332 /* didn't find any non-holes */ 3333 bzero(db->db.db_data, db->db.db_size); 3334 } 3335 DB_DNODE_EXIT(db); 3336} 3337 3338/* 3339 * The SPA will call this callback several times for each zio - once 3340 * for every physical child i/o (zio->io_phys_children times). This 3341 * allows the DMU to monitor the progress of each logical i/o. For example, 3342 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 3343 * block. There may be a long delay before all copies/fragments are completed, 3344 * so this callback allows us to retire dirty space gradually, as the physical 3345 * i/os complete. 3346 */ 3347/* ARGSUSED */ 3348static void 3349dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 3350{ 3351 dmu_buf_impl_t *db = arg; 3352 objset_t *os = db->db_objset; 3353 dsl_pool_t *dp = dmu_objset_pool(os); 3354 dbuf_dirty_record_t *dr; 3355 int delta = 0; 3356 3357 dr = db->db_data_pending; 3358 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 3359 3360 /* 3361 * The callback will be called io_phys_children times. Retire one 3362 * portion of our dirty space each time we are called. Any rounding 3363 * error will be cleaned up by dsl_pool_sync()'s call to 3364 * dsl_pool_undirty_space(). 3365 */ 3366 delta = dr->dr_accounted / zio->io_phys_children; 3367 dsl_pool_undirty_space(dp, delta, zio->io_txg); 3368} 3369 3370/* ARGSUSED */ 3371static void 3372dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3373{ 3374 dmu_buf_impl_t *db = vdb; 3375 blkptr_t *bp_orig = &zio->io_bp_orig; 3376 blkptr_t *bp = db->db_blkptr; 3377 objset_t *os = db->db_objset; 3378 dmu_tx_t *tx = os->os_synctx; 3379 dbuf_dirty_record_t **drp, *dr; 3380 3381 ASSERT0(zio->io_error); 3382 ASSERT(db->db_blkptr == bp); 3383 3384 /* 3385 * For nopwrites and rewrites we ensure that the bp matches our 3386 * original and bypass all the accounting. 3387 */ 3388 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3389 ASSERT(BP_EQUAL(bp, bp_orig)); 3390 } else { 3391 dsl_dataset_t *ds = os->os_dsl_dataset; 3392 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3393 dsl_dataset_block_born(ds, bp, tx); 3394 } 3395 3396 mutex_enter(&db->db_mtx); 3397 3398 DBUF_VERIFY(db); 3399 3400 drp = &db->db_last_dirty; 3401 while ((dr = *drp) != db->db_data_pending) 3402 drp = &dr->dr_next; 3403 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3404 ASSERT(dr->dr_dbuf == db); 3405 ASSERT(dr->dr_next == NULL); 3406 *drp = dr->dr_next; 3407 3408#ifdef ZFS_DEBUG 3409 if (db->db_blkid == DMU_SPILL_BLKID) { 3410 dnode_t *dn; 3411 3412 DB_DNODE_ENTER(db); 3413 dn = DB_DNODE(db); 3414 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3415 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3416 db->db_blkptr == &dn->dn_phys->dn_spill); 3417 DB_DNODE_EXIT(db); 3418 } 3419#endif 3420 3421 if (db->db_level == 0) { 3422 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3423 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3424 if (db->db_state != DB_NOFILL) { 3425 if (dr->dt.dl.dr_data != db->db_buf) 3426 arc_buf_destroy(dr->dt.dl.dr_data, db); 3427 } 3428 } else { 3429 dnode_t *dn; 3430 3431 DB_DNODE_ENTER(db); 3432 dn = DB_DNODE(db); 3433 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3434 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3435 if (!BP_IS_HOLE(db->db_blkptr)) { 3436 int epbs = 3437 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3438 ASSERT3U(db->db_blkid, <=, 3439 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3440 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3441 db->db.db_size); 3442 } 3443 DB_DNODE_EXIT(db); 3444 mutex_destroy(&dr->dt.di.dr_mtx); 3445 list_destroy(&dr->dt.di.dr_children); 3446 } 3447 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3448 3449 cv_broadcast(&db->db_changed); 3450 ASSERT(db->db_dirtycnt > 0); 3451 db->db_dirtycnt -= 1; 3452 db->db_data_pending = NULL; 3453 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3454} 3455 3456static void 3457dbuf_write_nofill_ready(zio_t *zio) 3458{ 3459 dbuf_write_ready(zio, NULL, zio->io_private); 3460} 3461 3462static void 3463dbuf_write_nofill_done(zio_t *zio) 3464{ 3465 dbuf_write_done(zio, NULL, zio->io_private); 3466} 3467 3468static void 3469dbuf_write_override_ready(zio_t *zio) 3470{ 3471 dbuf_dirty_record_t *dr = zio->io_private; 3472 dmu_buf_impl_t *db = dr->dr_dbuf; 3473 3474 dbuf_write_ready(zio, NULL, db); 3475} 3476 3477static void 3478dbuf_write_override_done(zio_t *zio) 3479{ 3480 dbuf_dirty_record_t *dr = zio->io_private; 3481 dmu_buf_impl_t *db = dr->dr_dbuf; 3482 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3483 3484 mutex_enter(&db->db_mtx); 3485 if (!BP_EQUAL(zio->io_bp, obp)) { 3486 if (!BP_IS_HOLE(obp)) 3487 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3488 arc_release(dr->dt.dl.dr_data, db); 3489 } 3490 mutex_exit(&db->db_mtx); 3491 3492 dbuf_write_done(zio, NULL, db); 3493} 3494 3495/* Issue I/O to commit a dirty buffer to disk. */ 3496static void 3497dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3498{ 3499 dmu_buf_impl_t *db = dr->dr_dbuf; 3500 dnode_t *dn; 3501 objset_t *os; 3502 dmu_buf_impl_t *parent = db->db_parent; 3503 uint64_t txg = tx->tx_txg; 3504 zbookmark_phys_t zb; 3505 zio_prop_t zp; 3506 zio_t *zio; 3507 int wp_flag = 0; 3508 3509 ASSERT(dmu_tx_is_syncing(tx)); 3510 3511 DB_DNODE_ENTER(db); 3512 dn = DB_DNODE(db); 3513 os = dn->dn_objset; 3514 3515 if (db->db_state != DB_NOFILL) { 3516 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3517 /* 3518 * Private object buffers are released here rather 3519 * than in dbuf_dirty() since they are only modified 3520 * in the syncing context and we don't want the 3521 * overhead of making multiple copies of the data. 3522 */ 3523 if (BP_IS_HOLE(db->db_blkptr)) { 3524 arc_buf_thaw(data); 3525 } else { 3526 dbuf_release_bp(db); 3527 } 3528 } 3529 } 3530 3531 if (parent != dn->dn_dbuf) { 3532 /* Our parent is an indirect block. */ 3533 /* We have a dirty parent that has been scheduled for write. */ 3534 ASSERT(parent && parent->db_data_pending); 3535 /* Our parent's buffer is one level closer to the dnode. */ 3536 ASSERT(db->db_level == parent->db_level-1); 3537 /* 3538 * We're about to modify our parent's db_data by modifying 3539 * our block pointer, so the parent must be released. 3540 */ 3541 ASSERT(arc_released(parent->db_buf)); 3542 zio = parent->db_data_pending->dr_zio; 3543 } else { 3544 /* Our parent is the dnode itself. */ 3545 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3546 db->db_blkid != DMU_SPILL_BLKID) || 3547 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3548 if (db->db_blkid != DMU_SPILL_BLKID) 3549 ASSERT3P(db->db_blkptr, ==, 3550 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3551 zio = dn->dn_zio; 3552 } 3553 3554 ASSERT(db->db_level == 0 || data == db->db_buf); 3555 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3556 ASSERT(zio); 3557 3558 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3559 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3560 db->db.db_object, db->db_level, db->db_blkid); 3561 3562 if (db->db_blkid == DMU_SPILL_BLKID) 3563 wp_flag = WP_SPILL; 3564 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3565 3566 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3567 DB_DNODE_EXIT(db); 3568 3569 /* 3570 * We copy the blkptr now (rather than when we instantiate the dirty 3571 * record), because its value can change between open context and 3572 * syncing context. We do not need to hold dn_struct_rwlock to read 3573 * db_blkptr because we are in syncing context. 3574 */ 3575 dr->dr_bp_copy = *db->db_blkptr; 3576 3577 if (db->db_level == 0 && 3578 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3579 /* 3580 * The BP for this block has been provided by open context 3581 * (by dmu_sync() or dmu_buf_write_embedded()). 3582 */ 3583 void *contents = (data != NULL) ? data->b_data : NULL; 3584 3585 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3586 &dr->dr_bp_copy, contents, db->db.db_size, &zp, 3587 dbuf_write_override_ready, NULL, NULL, 3588 dbuf_write_override_done, 3589 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3590 mutex_enter(&db->db_mtx); 3591 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3592 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3593 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3594 mutex_exit(&db->db_mtx); 3595 } else if (db->db_state == DB_NOFILL) { 3596 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3597 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3598 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3599 &dr->dr_bp_copy, NULL, db->db.db_size, &zp, 3600 dbuf_write_nofill_ready, NULL, NULL, 3601 dbuf_write_nofill_done, db, 3602 ZIO_PRIORITY_ASYNC_WRITE, 3603 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3604 } else { 3605 ASSERT(arc_released(data)); 3606 3607 /* 3608 * For indirect blocks, we want to setup the children 3609 * ready callback so that we can properly handle an indirect 3610 * block that only contains holes. 3611 */ 3612 arc_done_func_t *children_ready_cb = NULL; 3613 if (db->db_level != 0) 3614 children_ready_cb = dbuf_write_children_ready; 3615 3616 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3617 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), 3618 &zp, dbuf_write_ready, children_ready_cb, 3619 dbuf_write_physdone, dbuf_write_done, db, 3620 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3621 } 3622} 3623