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, 2015 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  */
  29 
  30 #include <sys/zfs_context.h>
  31 #include <sys/dmu.h>
  32 #include <sys/dmu_send.h>
  33 #include <sys/dmu_impl.h>
  34 #include <sys/dbuf.h>
  35 #include <sys/dmu_objset.h>
  36 #include <sys/dsl_dataset.h>
  37 #include <sys/dsl_dir.h>
  38 #include <sys/dmu_tx.h>
  39 #include <sys/spa.h>
  40 #include <sys/zio.h>
  41 #include <sys/dmu_zfetch.h>
  42 #include <sys/sa.h>
  43 #include <sys/sa_impl.h>
  44 #include <sys/zfeature.h>
  45 #include <sys/blkptr.h>
  46 #include <sys/range_tree.h>
  47 
  48 /*
  49  * Number of times that zfs_free_range() took the slow path while doing
  50  * a zfs receive.  A nonzero value indicates a potential performance problem.
  51  */
  52 uint64_t zfs_free_range_recv_miss;
  53 
  54 static void dbuf_destroy(dmu_buf_impl_t *db);
  55 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
  56 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
  57 
  58 #ifndef __lint
  59 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
  60     dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp);
  61 #endif /* ! __lint */
  62 
  63 /*
  64  * Global data structures and functions for the dbuf cache.
  65  */
  66 static kmem_cache_t *dbuf_cache;
  67 static taskq_t *dbu_evict_taskq;
  68 
  69 /* ARGSUSED */
  70 static int
  71 dbuf_cons(void *vdb, void *unused, int kmflag)
  72 {
  73         dmu_buf_impl_t *db = vdb;
  74         bzero(db, sizeof (dmu_buf_impl_t));
  75 
  76         mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
  77         cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
  78         refcount_create(&db->db_holds);
  79 
  80         return (0);
  81 }
  82 
  83 /* ARGSUSED */
  84 static void
  85 dbuf_dest(void *vdb, void *unused)
  86 {
  87         dmu_buf_impl_t *db = vdb;
  88         mutex_destroy(&db->db_mtx);
  89         cv_destroy(&db->db_changed);
  90         refcount_destroy(&db->db_holds);
  91 }
  92 
  93 /*
  94  * dbuf hash table routines
  95  */
  96 static dbuf_hash_table_t dbuf_hash_table;
  97 
  98 static uint64_t dbuf_hash_count;
  99 
 100 static uint64_t
 101 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
 102 {
 103         uintptr_t osv = (uintptr_t)os;
 104         uint64_t crc = -1ULL;
 105 
 106         ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
 107         crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
 108         crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
 109         crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
 110         crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
 111         crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
 112         crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
 113 
 114         crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
 115 
 116         return (crc);
 117 }
 118 
 119 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
 120 
 121 #define DBUF_EQUAL(dbuf, os, obj, level, blkid)         \
 122         ((dbuf)->db.db_object == (obj) &&            \
 123         (dbuf)->db_objset == (os) &&                 \
 124         (dbuf)->db_level == (level) &&                       \
 125         (dbuf)->db_blkid == (blkid))
 126 
 127 dmu_buf_impl_t *
 128 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
 129 {
 130         dbuf_hash_table_t *h = &dbuf_hash_table;
 131         uint64_t hv = DBUF_HASH(os, obj, level, blkid);
 132         uint64_t idx = hv & h->hash_table_mask;
 133         dmu_buf_impl_t *db;
 134 
 135         mutex_enter(DBUF_HASH_MUTEX(h, idx));
 136         for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
 137                 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
 138                         mutex_enter(&db->db_mtx);
 139                         if (db->db_state != DB_EVICTING) {
 140                                 mutex_exit(DBUF_HASH_MUTEX(h, idx));
 141                                 return (db);
 142                         }
 143                         mutex_exit(&db->db_mtx);
 144                 }
 145         }
 146         mutex_exit(DBUF_HASH_MUTEX(h, idx));
 147         return (NULL);
 148 }
 149 
 150 static dmu_buf_impl_t *
 151 dbuf_find_bonus(objset_t *os, uint64_t object)
 152 {
 153         dnode_t *dn;
 154         dmu_buf_impl_t *db = NULL;
 155 
 156         if (dnode_hold(os, object, FTAG, &dn) == 0) {
 157                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 158                 if (dn->dn_bonus != NULL) {
 159                         db = dn->dn_bonus;
 160                         mutex_enter(&db->db_mtx);
 161                 }
 162                 rw_exit(&dn->dn_struct_rwlock);
 163                 dnode_rele(dn, FTAG);
 164         }
 165         return (db);
 166 }
 167 
 168 /*
 169  * Insert an entry into the hash table.  If there is already an element
 170  * equal to elem in the hash table, then the already existing element
 171  * will be returned and the new element will not be inserted.
 172  * Otherwise returns NULL.
 173  */
 174 static dmu_buf_impl_t *
 175 dbuf_hash_insert(dmu_buf_impl_t *db)
 176 {
 177         dbuf_hash_table_t *h = &dbuf_hash_table;
 178         objset_t *os = db->db_objset;
 179         uint64_t obj = db->db.db_object;
 180         int level = db->db_level;
 181         uint64_t blkid = db->db_blkid;
 182         uint64_t hv = DBUF_HASH(os, obj, level, blkid);
 183         uint64_t idx = hv & h->hash_table_mask;
 184         dmu_buf_impl_t *dbf;
 185 
 186         mutex_enter(DBUF_HASH_MUTEX(h, idx));
 187         for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
 188                 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
 189                         mutex_enter(&dbf->db_mtx);
 190                         if (dbf->db_state != DB_EVICTING) {
 191                                 mutex_exit(DBUF_HASH_MUTEX(h, idx));
 192                                 return (dbf);
 193                         }
 194                         mutex_exit(&dbf->db_mtx);
 195                 }
 196         }
 197 
 198         mutex_enter(&db->db_mtx);
 199         db->db_hash_next = h->hash_table[idx];
 200         h->hash_table[idx] = db;
 201         mutex_exit(DBUF_HASH_MUTEX(h, idx));
 202         atomic_inc_64(&dbuf_hash_count);
 203 
 204         return (NULL);
 205 }
 206 
 207 /*
 208  * Remove an entry from the hash table.  It must be in the EVICTING state.
 209  */
 210 static void
 211 dbuf_hash_remove(dmu_buf_impl_t *db)
 212 {
 213         dbuf_hash_table_t *h = &dbuf_hash_table;
 214         uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
 215             db->db_level, db->db_blkid);
 216         uint64_t idx = hv & h->hash_table_mask;
 217         dmu_buf_impl_t *dbf, **dbp;
 218 
 219         /*
 220          * We musn't hold db_mtx to maintain lock ordering:
 221          * DBUF_HASH_MUTEX > db_mtx.
 222          */
 223         ASSERT(refcount_is_zero(&db->db_holds));
 224         ASSERT(db->db_state == DB_EVICTING);
 225         ASSERT(!MUTEX_HELD(&db->db_mtx));
 226 
 227         mutex_enter(DBUF_HASH_MUTEX(h, idx));
 228         dbp = &h->hash_table[idx];
 229         while ((dbf = *dbp) != db) {
 230                 dbp = &dbf->db_hash_next;
 231                 ASSERT(dbf != NULL);
 232         }
 233         *dbp = db->db_hash_next;
 234         db->db_hash_next = NULL;
 235         mutex_exit(DBUF_HASH_MUTEX(h, idx));
 236         atomic_dec_64(&dbuf_hash_count);
 237 }
 238 
 239 static arc_evict_func_t dbuf_do_evict;
 240 
 241 typedef enum {
 242         DBVU_EVICTING,
 243         DBVU_NOT_EVICTING
 244 } dbvu_verify_type_t;
 245 
 246 static void
 247 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
 248 {
 249 #ifdef ZFS_DEBUG
 250         int64_t holds;
 251 
 252         if (db->db_user == NULL)
 253                 return;
 254 
 255         /* Only data blocks support the attachment of user data. */
 256         ASSERT(db->db_level == 0);
 257 
 258         /* Clients must resolve a dbuf before attaching user data. */
 259         ASSERT(db->db.db_data != NULL);
 260         ASSERT3U(db->db_state, ==, DB_CACHED);
 261 
 262         holds = refcount_count(&db->db_holds);
 263         if (verify_type == DBVU_EVICTING) {
 264                 /*
 265                  * Immediate eviction occurs when holds == dirtycnt.
 266                  * For normal eviction buffers, holds is zero on
 267                  * eviction, except when dbuf_fix_old_data() calls
 268                  * dbuf_clear_data().  However, the hold count can grow
 269                  * during eviction even though db_mtx is held (see
 270                  * dmu_bonus_hold() for an example), so we can only
 271                  * test the generic invariant that holds >= dirtycnt.
 272                  */
 273                 ASSERT3U(holds, >=, db->db_dirtycnt);
 274         } else {
 275                 if (db->db_user_immediate_evict == TRUE)
 276                         ASSERT3U(holds, >=, db->db_dirtycnt);
 277                 else
 278                         ASSERT3U(holds, >, 0);
 279         }
 280 #endif
 281 }
 282 
 283 static void
 284 dbuf_evict_user(dmu_buf_impl_t *db)
 285 {
 286         dmu_buf_user_t *dbu = db->db_user;
 287 
 288         ASSERT(MUTEX_HELD(&db->db_mtx));
 289 
 290         if (dbu == NULL)
 291                 return;
 292 
 293         dbuf_verify_user(db, DBVU_EVICTING);
 294         db->db_user = NULL;
 295 
 296 #ifdef ZFS_DEBUG
 297         if (dbu->dbu_clear_on_evict_dbufp != NULL)
 298                 *dbu->dbu_clear_on_evict_dbufp = NULL;
 299 #endif
 300 
 301         /*
 302          * Invoke the callback from a taskq to avoid lock order reversals
 303          * and limit stack depth.
 304          */
 305         taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
 306             &dbu->dbu_tqent);
 307 }
 308 
 309 boolean_t
 310 dbuf_is_metadata(dmu_buf_impl_t *db)
 311 {
 312         if (db->db_level > 0) {
 313                 return (B_TRUE);
 314         } else {
 315                 boolean_t is_metadata;
 316 
 317                 DB_DNODE_ENTER(db);
 318                 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
 319                 DB_DNODE_EXIT(db);
 320 
 321                 return (is_metadata);
 322         }
 323 }
 324 
 325 void
 326 dbuf_evict(dmu_buf_impl_t *db)
 327 {
 328         ASSERT(MUTEX_HELD(&db->db_mtx));
 329         ASSERT(db->db_buf == NULL);
 330         ASSERT(db->db_data_pending == NULL);
 331 
 332         dbuf_clear(db);
 333         dbuf_destroy(db);
 334 }
 335 
 336 void
 337 dbuf_init(void)
 338 {
 339         uint64_t hsize = 1ULL << 16;
 340         dbuf_hash_table_t *h = &dbuf_hash_table;
 341         int i;
 342 
 343         /*
 344          * The hash table is big enough to fill all of physical memory
 345          * with an average 4K block size.  The table will take up
 346          * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
 347          */
 348         while (hsize * 4096 < physmem * PAGESIZE)
 349                 hsize <<= 1;
 350 
 351 retry:
 352         h->hash_table_mask = hsize - 1;
 353         h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
 354         if (h->hash_table == NULL) {
 355                 /* XXX - we should really return an error instead of assert */
 356                 ASSERT(hsize > (1ULL << 10));
 357                 hsize >>= 1;
 358                 goto retry;
 359         }
 360 
 361         dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
 362             sizeof (dmu_buf_impl_t),
 363             0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
 364 
 365         for (i = 0; i < DBUF_MUTEXES; i++)
 366                 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
 367 
 368         /*
 369          * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
 370          * configuration is not required.
 371          */
 372         dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
 373 }
 374 
 375 void
 376 dbuf_fini(void)
 377 {
 378         dbuf_hash_table_t *h = &dbuf_hash_table;
 379         int i;
 380 
 381         for (i = 0; i < DBUF_MUTEXES; i++)
 382                 mutex_destroy(&h->hash_mutexes[i]);
 383         kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
 384         kmem_cache_destroy(dbuf_cache);
 385         taskq_destroy(dbu_evict_taskq);
 386 }
 387 
 388 /*
 389  * Other stuff.
 390  */
 391 
 392 #ifdef ZFS_DEBUG
 393 static void
 394 dbuf_verify(dmu_buf_impl_t *db)
 395 {
 396         dnode_t *dn;
 397         dbuf_dirty_record_t *dr;
 398 
 399         ASSERT(MUTEX_HELD(&db->db_mtx));
 400 
 401         if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
 402                 return;
 403 
 404         ASSERT(db->db_objset != NULL);
 405         DB_DNODE_ENTER(db);
 406         dn = DB_DNODE(db);
 407         if (dn == NULL) {
 408                 ASSERT(db->db_parent == NULL);
 409                 ASSERT(db->db_blkptr == NULL);
 410         } else {
 411                 ASSERT3U(db->db.db_object, ==, dn->dn_object);
 412                 ASSERT3P(db->db_objset, ==, dn->dn_objset);
 413                 ASSERT3U(db->db_level, <, dn->dn_nlevels);
 414                 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
 415                     db->db_blkid == DMU_SPILL_BLKID ||
 416                     !avl_is_empty(&dn->dn_dbufs));
 417         }
 418         if (db->db_blkid == DMU_BONUS_BLKID) {
 419                 ASSERT(dn != NULL);
 420                 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
 421                 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
 422         } else if (db->db_blkid == DMU_SPILL_BLKID) {
 423                 ASSERT(dn != NULL);
 424                 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
 425                 ASSERT0(db->db.db_offset);
 426         } else {
 427                 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
 428         }
 429 
 430         for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
 431                 ASSERT(dr->dr_dbuf == db);
 432 
 433         for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
 434                 ASSERT(dr->dr_dbuf == db);
 435 
 436         /*
 437          * We can't assert that db_size matches dn_datablksz because it
 438          * can be momentarily different when another thread is doing
 439          * dnode_set_blksz().
 440          */
 441         if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
 442                 dr = db->db_data_pending;
 443                 /*
 444                  * It should only be modified in syncing context, so
 445                  * make sure we only have one copy of the data.
 446                  */
 447                 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
 448         }
 449 
 450         /* verify db->db_blkptr */
 451         if (db->db_blkptr) {
 452                 if (db->db_parent == dn->dn_dbuf) {
 453                         /* db is pointed to by the dnode */
 454                         /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
 455                         if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
 456                                 ASSERT(db->db_parent == NULL);
 457                         else
 458                                 ASSERT(db->db_parent != NULL);
 459                         if (db->db_blkid != DMU_SPILL_BLKID)
 460                                 ASSERT3P(db->db_blkptr, ==,
 461                                     &dn->dn_phys->dn_blkptr[db->db_blkid]);
 462                 } else {
 463                         /* db is pointed to by an indirect block */
 464                         int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
 465                         ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
 466                         ASSERT3U(db->db_parent->db.db_object, ==,
 467                             db->db.db_object);
 468                         /*
 469                          * dnode_grow_indblksz() can make this fail if we don't
 470                          * have the struct_rwlock.  XXX indblksz no longer
 471                          * grows.  safe to do this now?
 472                          */
 473                         if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
 474                                 ASSERT3P(db->db_blkptr, ==,
 475                                     ((blkptr_t *)db->db_parent->db.db_data +
 476                                     db->db_blkid % epb));
 477                         }
 478                 }
 479         }
 480         if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
 481             (db->db_buf == NULL || db->db_buf->b_data) &&
 482             db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
 483             db->db_state != DB_FILL && !dn->dn_free_txg) {
 484                 /*
 485                  * If the blkptr isn't set but they have nonzero data,
 486                  * it had better be dirty, otherwise we'll lose that
 487                  * data when we evict this buffer.
 488                  */
 489                 if (db->db_dirtycnt == 0) {
 490                         uint64_t *buf = db->db.db_data;
 491                         int i;
 492 
 493                         for (i = 0; i < db->db.db_size >> 3; i++) {
 494                                 ASSERT(buf[i] == 0);
 495                         }
 496                 }
 497         }
 498         DB_DNODE_EXIT(db);
 499 }
 500 #endif
 501 
 502 static void
 503 dbuf_clear_data(dmu_buf_impl_t *db)
 504 {
 505         ASSERT(MUTEX_HELD(&db->db_mtx));
 506         dbuf_evict_user(db);
 507         db->db_buf = NULL;
 508         db->db.db_data = NULL;
 509         if (db->db_state != DB_NOFILL)
 510                 db->db_state = DB_UNCACHED;
 511 }
 512 
 513 static void
 514 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
 515 {
 516         ASSERT(MUTEX_HELD(&db->db_mtx));
 517         ASSERT(buf != NULL);
 518 
 519         db->db_buf = buf;
 520         ASSERT(buf->b_data != NULL);
 521         db->db.db_data = buf->b_data;
 522         if (!arc_released(buf))
 523                 arc_set_callback(buf, dbuf_do_evict, db);
 524 }
 525 
 526 /*
 527  * Loan out an arc_buf for read.  Return the loaned arc_buf.
 528  */
 529 arc_buf_t *
 530 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
 531 {
 532         arc_buf_t *abuf;
 533 
 534         mutex_enter(&db->db_mtx);
 535         if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
 536                 int blksz = db->db.db_size;
 537                 spa_t *spa = db->db_objset->os_spa;
 538 
 539                 mutex_exit(&db->db_mtx);
 540                 abuf = arc_loan_buf(spa, blksz);
 541                 bcopy(db->db.db_data, abuf->b_data, blksz);
 542         } else {
 543                 abuf = db->db_buf;
 544                 arc_loan_inuse_buf(abuf, db);
 545                 dbuf_clear_data(db);
 546                 mutex_exit(&db->db_mtx);
 547         }
 548         return (abuf);
 549 }
 550 
 551 /*
 552  * Calculate which level n block references the data at the level 0 offset
 553  * provided.
 554  */
 555 uint64_t
 556 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
 557 {
 558         if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
 559                 /*
 560                  * The level n blkid is equal to the level 0 blkid divided by
 561                  * the number of level 0s in a level n block.
 562                  *
 563                  * The level 0 blkid is offset >> datablkshift =
 564                  * offset / 2^datablkshift.
 565                  *
 566                  * The number of level 0s in a level n is the number of block
 567                  * pointers in an indirect block, raised to the power of level.
 568                  * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
 569                  * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
 570                  *
 571                  * Thus, the level n blkid is: offset /
 572                  * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
 573                  * = offset / 2^(datablkshift + level *
 574                  *   (indblkshift - SPA_BLKPTRSHIFT))
 575                  * = offset >> (datablkshift + level *
 576                  *   (indblkshift - SPA_BLKPTRSHIFT))
 577                  */
 578                 return (offset >> (dn->dn_datablkshift + level *
 579                     (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
 580         } else {
 581                 ASSERT3U(offset, <, dn->dn_datablksz);
 582                 return (0);
 583         }
 584 }
 585 
 586 static void
 587 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
 588 {
 589         dmu_buf_impl_t *db = vdb;
 590 
 591         mutex_enter(&db->db_mtx);
 592         ASSERT3U(db->db_state, ==, DB_READ);
 593         /*
 594          * All reads are synchronous, so we must have a hold on the dbuf
 595          */
 596         ASSERT(refcount_count(&db->db_holds) > 0);
 597         ASSERT(db->db_buf == NULL);
 598         ASSERT(db->db.db_data == NULL);
 599         if (db->db_level == 0 && db->db_freed_in_flight) {
 600                 /* we were freed in flight; disregard any error */
 601                 arc_release(buf, db);
 602                 bzero(buf->b_data, db->db.db_size);
 603                 arc_buf_freeze(buf);
 604                 db->db_freed_in_flight = FALSE;
 605                 dbuf_set_data(db, buf);
 606                 db->db_state = DB_CACHED;
 607         } else if (zio == NULL || zio->io_error == 0) {
 608                 dbuf_set_data(db, buf);
 609                 db->db_state = DB_CACHED;
 610         } else {
 611                 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
 612                 ASSERT3P(db->db_buf, ==, NULL);
 613                 VERIFY(arc_buf_remove_ref(buf, db));
 614                 db->db_state = DB_UNCACHED;
 615         }
 616         cv_broadcast(&db->db_changed);
 617         dbuf_rele_and_unlock(db, NULL);
 618 }
 619 
 620 static void
 621 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 622 {
 623         dnode_t *dn;
 624         zbookmark_phys_t zb;
 625         arc_flags_t aflags = ARC_FLAG_NOWAIT;
 626 
 627         DB_DNODE_ENTER(db);
 628         dn = DB_DNODE(db);
 629         ASSERT(!refcount_is_zero(&db->db_holds));
 630         /* We need the struct_rwlock to prevent db_blkptr from changing. */
 631         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
 632         ASSERT(MUTEX_HELD(&db->db_mtx));
 633         ASSERT(db->db_state == DB_UNCACHED);
 634         ASSERT(db->db_buf == NULL);
 635 
 636         if (db->db_blkid == DMU_BONUS_BLKID) {
 637                 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
 638 
 639                 ASSERT3U(bonuslen, <=, db->db.db_size);
 640                 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
 641                 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
 642                 if (bonuslen < DN_MAX_BONUSLEN)
 643                         bzero(db->db.db_data, DN_MAX_BONUSLEN);
 644                 if (bonuslen)
 645                         bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
 646                 DB_DNODE_EXIT(db);
 647                 db->db_state = DB_CACHED;
 648                 mutex_exit(&db->db_mtx);
 649                 return;
 650         }
 651 
 652         /*
 653          * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
 654          * processes the delete record and clears the bp while we are waiting
 655          * for the dn_mtx (resulting in a "no" from block_freed).
 656          */
 657         if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
 658             (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
 659             BP_IS_HOLE(db->db_blkptr)))) {
 660                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
 661 
 662                 DB_DNODE_EXIT(db);
 663                 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
 664                     db->db.db_size, db, type));
 665                 bzero(db->db.db_data, db->db.db_size);
 666                 db->db_state = DB_CACHED;
 667                 mutex_exit(&db->db_mtx);
 668                 return;
 669         }
 670 
 671         DB_DNODE_EXIT(db);
 672 
 673         db->db_state = DB_READ;
 674         mutex_exit(&db->db_mtx);
 675 
 676         if (DBUF_IS_L2CACHEABLE(db))
 677                 aflags |= ARC_FLAG_L2CACHE;
 678         if (DBUF_IS_L2COMPRESSIBLE(db))
 679                 aflags |= ARC_FLAG_L2COMPRESS;
 680 
 681         SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
 682             db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
 683             db->db.db_object, db->db_level, db->db_blkid);
 684 
 685         dbuf_add_ref(db, NULL);
 686 
 687         (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
 688             dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
 689             (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
 690             &aflags, &zb);
 691 }
 692 
 693 int
 694 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 695 {
 696         int err = 0;
 697         boolean_t havepzio = (zio != NULL);
 698         boolean_t prefetch;
 699         dnode_t *dn;
 700 
 701         /*
 702          * We don't have to hold the mutex to check db_state because it
 703          * can't be freed while we have a hold on the buffer.
 704          */
 705         ASSERT(!refcount_is_zero(&db->db_holds));
 706 
 707         if (db->db_state == DB_NOFILL)
 708                 return (SET_ERROR(EIO));
 709 
 710         DB_DNODE_ENTER(db);
 711         dn = DB_DNODE(db);
 712         if ((flags & DB_RF_HAVESTRUCT) == 0)
 713                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 714 
 715         prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
 716             (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
 717             DBUF_IS_CACHEABLE(db);
 718 
 719         mutex_enter(&db->db_mtx);
 720         if (db->db_state == DB_CACHED) {
 721                 mutex_exit(&db->db_mtx);
 722                 if (prefetch)
 723                         dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
 724                 if ((flags & DB_RF_HAVESTRUCT) == 0)
 725                         rw_exit(&dn->dn_struct_rwlock);
 726                 DB_DNODE_EXIT(db);
 727         } else if (db->db_state == DB_UNCACHED) {
 728                 spa_t *spa = dn->dn_objset->os_spa;
 729 
 730                 if (zio == NULL)
 731                         zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
 732                 dbuf_read_impl(db, zio, flags);
 733 
 734                 /* dbuf_read_impl has dropped db_mtx for us */
 735 
 736                 if (prefetch)
 737                         dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
 738 
 739                 if ((flags & DB_RF_HAVESTRUCT) == 0)
 740                         rw_exit(&dn->dn_struct_rwlock);
 741                 DB_DNODE_EXIT(db);
 742 
 743                 if (!havepzio)
 744                         err = zio_wait(zio);
 745         } else {
 746                 /*
 747                  * Another reader came in while the dbuf was in flight
 748                  * between UNCACHED and CACHED.  Either a writer will finish
 749                  * writing the buffer (sending the dbuf to CACHED) or the
 750                  * first reader's request will reach the read_done callback
 751                  * and send the dbuf to CACHED.  Otherwise, a failure
 752                  * occurred and the dbuf went to UNCACHED.
 753                  */
 754                 mutex_exit(&db->db_mtx);
 755                 if (prefetch)
 756                         dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
 757                 if ((flags & DB_RF_HAVESTRUCT) == 0)
 758                         rw_exit(&dn->dn_struct_rwlock);
 759                 DB_DNODE_EXIT(db);
 760 
 761                 /* Skip the wait per the caller's request. */
 762                 mutex_enter(&db->db_mtx);
 763                 if ((flags & DB_RF_NEVERWAIT) == 0) {
 764                         while (db->db_state == DB_READ ||
 765                             db->db_state == DB_FILL) {
 766                                 ASSERT(db->db_state == DB_READ ||
 767                                     (flags & DB_RF_HAVESTRUCT) == 0);
 768                                 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
 769                                     db, zio_t *, zio);
 770                                 cv_wait(&db->db_changed, &db->db_mtx);
 771                         }
 772                         if (db->db_state == DB_UNCACHED)
 773                                 err = SET_ERROR(EIO);
 774                 }
 775                 mutex_exit(&db->db_mtx);
 776         }
 777 
 778         ASSERT(err || havepzio || db->db_state == DB_CACHED);
 779         return (err);
 780 }
 781 
 782 static void
 783 dbuf_noread(dmu_buf_impl_t *db)
 784 {
 785         ASSERT(!refcount_is_zero(&db->db_holds));
 786         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
 787         mutex_enter(&db->db_mtx);
 788         while (db->db_state == DB_READ || db->db_state == DB_FILL)
 789                 cv_wait(&db->db_changed, &db->db_mtx);
 790         if (db->db_state == DB_UNCACHED) {
 791                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
 792                 spa_t *spa = db->db_objset->os_spa;
 793 
 794                 ASSERT(db->db_buf == NULL);
 795                 ASSERT(db->db.db_data == NULL);
 796                 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
 797                 db->db_state = DB_FILL;
 798         } else if (db->db_state == DB_NOFILL) {
 799                 dbuf_clear_data(db);
 800         } else {
 801                 ASSERT3U(db->db_state, ==, DB_CACHED);
 802         }
 803         mutex_exit(&db->db_mtx);
 804 }
 805 
 806 /*
 807  * This is our just-in-time copy function.  It makes a copy of
 808  * buffers, that have been modified in a previous transaction
 809  * group, before we modify them in the current active group.
 810  *
 811  * This function is used in two places: when we are dirtying a
 812  * buffer for the first time in a txg, and when we are freeing
 813  * a range in a dnode that includes this buffer.
 814  *
 815  * Note that when we are called from dbuf_free_range() we do
 816  * not put a hold on the buffer, we just traverse the active
 817  * dbuf list for the dnode.
 818  */
 819 static void
 820 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
 821 {
 822         dbuf_dirty_record_t *dr = db->db_last_dirty;
 823 
 824         ASSERT(MUTEX_HELD(&db->db_mtx));
 825         ASSERT(db->db.db_data != NULL);
 826         ASSERT(db->db_level == 0);
 827         ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
 828 
 829         if (dr == NULL ||
 830             (dr->dt.dl.dr_data !=
 831             ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
 832                 return;
 833 
 834         /*
 835          * If the last dirty record for this dbuf has not yet synced
 836          * and its referencing the dbuf data, either:
 837          *      reset the reference to point to a new copy,
 838          * or (if there a no active holders)
 839          *      just null out the current db_data pointer.
 840          */
 841         ASSERT(dr->dr_txg >= txg - 2);
 842         if (db->db_blkid == DMU_BONUS_BLKID) {
 843                 /* Note that the data bufs here are zio_bufs */
 844                 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
 845                 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
 846                 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
 847         } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
 848                 int size = db->db.db_size;
 849                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
 850                 spa_t *spa = db->db_objset->os_spa;
 851 
 852                 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
 853                 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
 854         } else {
 855                 dbuf_clear_data(db);
 856         }
 857 }
 858 
 859 void
 860 dbuf_unoverride(dbuf_dirty_record_t *dr)
 861 {
 862         dmu_buf_impl_t *db = dr->dr_dbuf;
 863         blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
 864         uint64_t txg = dr->dr_txg;
 865 
 866         ASSERT(MUTEX_HELD(&db->db_mtx));
 867         ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
 868         ASSERT(db->db_level == 0);
 869 
 870         if (db->db_blkid == DMU_BONUS_BLKID ||
 871             dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
 872                 return;
 873 
 874         ASSERT(db->db_data_pending != dr);
 875 
 876         /* free this block */
 877         if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
 878                 zio_free(db->db_objset->os_spa, txg, bp);
 879 
 880         dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
 881         dr->dt.dl.dr_nopwrite = B_FALSE;
 882 
 883         /*
 884          * Release the already-written buffer, so we leave it in
 885          * a consistent dirty state.  Note that all callers are
 886          * modifying the buffer, so they will immediately do
 887          * another (redundant) arc_release().  Therefore, leave
 888          * the buf thawed to save the effort of freezing &
 889          * immediately re-thawing it.
 890          */
 891         arc_release(dr->dt.dl.dr_data, db);
 892 }
 893 
 894 /*
 895  * Evict (if its unreferenced) or clear (if its referenced) any level-0
 896  * data blocks in the free range, so that any future readers will find
 897  * empty blocks.
 898  *
 899  * This is a no-op if the dataset is in the middle of an incremental
 900  * receive; see comment below for details.
 901  */
 902 void
 903 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
 904     dmu_tx_t *tx)
 905 {
 906         dmu_buf_impl_t db_search;
 907         dmu_buf_impl_t *db, *db_next;
 908         uint64_t txg = tx->tx_txg;
 909         avl_index_t where;
 910 
 911         if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
 912                 end_blkid = dn->dn_maxblkid;
 913         dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
 914 
 915         db_search.db_level = 0;
 916         db_search.db_blkid = start_blkid;
 917         db_search.db_state = DB_SEARCH;
 918 
 919         mutex_enter(&dn->dn_dbufs_mtx);
 920         if (start_blkid >= dn->dn_unlisted_l0_blkid) {
 921                 /* There can't be any dbufs in this range; no need to search. */
 922 #ifdef DEBUG
 923                 db = avl_find(&dn->dn_dbufs, &db_search, &where);
 924                 ASSERT3P(db, ==, NULL);
 925                 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
 926                 ASSERT(db == NULL || db->db_level > 0);
 927 #endif
 928                 mutex_exit(&dn->dn_dbufs_mtx);
 929                 return;
 930         } else if (dmu_objset_is_receiving(dn->dn_objset)) {
 931                 /*
 932                  * If we are receiving, we expect there to be no dbufs in
 933                  * the range to be freed, because receive modifies each
 934                  * block at most once, and in offset order.  If this is
 935                  * not the case, it can lead to performance problems,
 936                  * so note that we unexpectedly took the slow path.
 937                  */
 938                 atomic_inc_64(&zfs_free_range_recv_miss);
 939         }
 940 
 941         db = avl_find(&dn->dn_dbufs, &db_search, &where);
 942         ASSERT3P(db, ==, NULL);
 943         db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
 944 
 945         for (; db != NULL; db = db_next) {
 946                 db_next = AVL_NEXT(&dn->dn_dbufs, db);
 947                 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
 948 
 949                 if (db->db_level != 0 || db->db_blkid > end_blkid) {
 950                         break;
 951                 }
 952                 ASSERT3U(db->db_blkid, >=, start_blkid);
 953 
 954                 /* found a level 0 buffer in the range */
 955                 mutex_enter(&db->db_mtx);
 956                 if (dbuf_undirty(db, tx)) {
 957                         /* mutex has been dropped and dbuf destroyed */
 958                         continue;
 959                 }
 960 
 961                 if (db->db_state == DB_UNCACHED ||
 962                     db->db_state == DB_NOFILL ||
 963                     db->db_state == DB_EVICTING) {
 964                         ASSERT(db->db.db_data == NULL);
 965                         mutex_exit(&db->db_mtx);
 966                         continue;
 967                 }
 968                 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
 969                         /* will be handled in dbuf_read_done or dbuf_rele */
 970                         db->db_freed_in_flight = TRUE;
 971                         mutex_exit(&db->db_mtx);
 972                         continue;
 973                 }
 974                 if (refcount_count(&db->db_holds) == 0) {
 975                         ASSERT(db->db_buf);
 976                         dbuf_clear(db);
 977                         continue;
 978                 }
 979                 /* The dbuf is referenced */
 980 
 981                 if (db->db_last_dirty != NULL) {
 982                         dbuf_dirty_record_t *dr = db->db_last_dirty;
 983 
 984                         if (dr->dr_txg == txg) {
 985                                 /*
 986                                  * This buffer is "in-use", re-adjust the file
 987                                  * size to reflect that this buffer may
 988                                  * contain new data when we sync.
 989                                  */
 990                                 if (db->db_blkid != DMU_SPILL_BLKID &&
 991                                     db->db_blkid > dn->dn_maxblkid)
 992                                         dn->dn_maxblkid = db->db_blkid;
 993                                 dbuf_unoverride(dr);
 994                         } else {
 995                                 /*
 996                                  * This dbuf is not dirty in the open context.
 997                                  * Either uncache it (if its not referenced in
 998                                  * the open context) or reset its contents to
 999                                  * empty.
1000                                  */
1001                                 dbuf_fix_old_data(db, txg);
1002                         }
1003                 }
1004                 /* clear the contents if its cached */
1005                 if (db->db_state == DB_CACHED) {
1006                         ASSERT(db->db.db_data != NULL);
1007                         arc_release(db->db_buf, db);
1008                         bzero(db->db.db_data, db->db.db_size);
1009                         arc_buf_freeze(db->db_buf);
1010                 }
1011 
1012                 mutex_exit(&db->db_mtx);
1013         }
1014         mutex_exit(&dn->dn_dbufs_mtx);
1015 }
1016 
1017 static int
1018 dbuf_block_freeable(dmu_buf_impl_t *db)
1019 {
1020         dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1021         uint64_t birth_txg = 0;
1022 
1023         /*
1024          * We don't need any locking to protect db_blkptr:
1025          * If it's syncing, then db_last_dirty will be set
1026          * so we'll ignore db_blkptr.
1027          *
1028          * This logic ensures that only block births for
1029          * filled blocks are considered.
1030          */
1031         ASSERT(MUTEX_HELD(&db->db_mtx));
1032         if (db->db_last_dirty && (db->db_blkptr == NULL ||
1033             !BP_IS_HOLE(db->db_blkptr))) {
1034                 birth_txg = db->db_last_dirty->dr_txg;
1035         } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1036                 birth_txg = db->db_blkptr->blk_birth;
1037         }
1038 
1039         /*
1040          * If this block don't exist or is in a snapshot, it can't be freed.
1041          * Don't pass the bp to dsl_dataset_block_freeable() since we
1042          * are holding the db_mtx lock and might deadlock if we are
1043          * prefetching a dedup-ed block.
1044          */
1045         if (birth_txg != 0)
1046                 return (ds == NULL ||
1047                     dsl_dataset_block_freeable(ds, NULL, birth_txg));
1048         else
1049                 return (B_FALSE);
1050 }
1051 
1052 void
1053 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1054 {
1055         arc_buf_t *buf, *obuf;
1056         int osize = db->db.db_size;
1057         arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1058         dnode_t *dn;
1059 
1060         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1061 
1062         DB_DNODE_ENTER(db);
1063         dn = DB_DNODE(db);
1064 
1065         /* XXX does *this* func really need the lock? */
1066         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1067 
1068         /*
1069          * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1070          * is OK, because there can be no other references to the db
1071          * when we are changing its size, so no concurrent DB_FILL can
1072          * be happening.
1073          */
1074         /*
1075          * XXX we should be doing a dbuf_read, checking the return
1076          * value and returning that up to our callers
1077          */
1078         dmu_buf_will_dirty(&db->db, tx);
1079 
1080         /* create the data buffer for the new block */
1081         buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1082 
1083         /* copy old block data to the new block */
1084         obuf = db->db_buf;
1085         bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1086         /* zero the remainder */
1087         if (size > osize)
1088                 bzero((uint8_t *)buf->b_data + osize, size - osize);
1089 
1090         mutex_enter(&db->db_mtx);
1091         dbuf_set_data(db, buf);
1092         VERIFY(arc_buf_remove_ref(obuf, db));
1093         db->db.db_size = size;
1094 
1095         if (db->db_level == 0) {
1096                 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1097                 db->db_last_dirty->dt.dl.dr_data = buf;
1098         }
1099         mutex_exit(&db->db_mtx);
1100 
1101         dnode_willuse_space(dn, size-osize, tx);
1102         DB_DNODE_EXIT(db);
1103 }
1104 
1105 void
1106 dbuf_release_bp(dmu_buf_impl_t *db)
1107 {
1108         objset_t *os = db->db_objset;
1109 
1110         ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1111         ASSERT(arc_released(os->os_phys_buf) ||
1112             list_link_active(&os->os_dsl_dataset->ds_synced_link));
1113         ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1114 
1115         (void) arc_release(db->db_buf, db);
1116 }
1117 
1118 /*
1119  * We already have a dirty record for this TXG, and we are being
1120  * dirtied again.
1121  */
1122 static void
1123 dbuf_redirty(dbuf_dirty_record_t *dr)
1124 {
1125         dmu_buf_impl_t *db = dr->dr_dbuf;
1126 
1127         ASSERT(MUTEX_HELD(&db->db_mtx));
1128 
1129         if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1130                 /*
1131                  * If this buffer has already been written out,
1132                  * we now need to reset its state.
1133                  */
1134                 dbuf_unoverride(dr);
1135                 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1136                     db->db_state != DB_NOFILL) {
1137                         /* Already released on initial dirty, so just thaw. */
1138                         ASSERT(arc_released(db->db_buf));
1139                         arc_buf_thaw(db->db_buf);
1140                 }
1141         }
1142 }
1143 
1144 dbuf_dirty_record_t *
1145 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1146 {
1147         dnode_t *dn;
1148         objset_t *os;
1149         dbuf_dirty_record_t **drp, *dr;
1150         int drop_struct_lock = FALSE;
1151         boolean_t do_free_accounting = B_FALSE;
1152         int txgoff = tx->tx_txg & TXG_MASK;
1153 
1154         ASSERT(tx->tx_txg != 0);
1155         ASSERT(!refcount_is_zero(&db->db_holds));
1156         DMU_TX_DIRTY_BUF(tx, db);
1157 
1158         DB_DNODE_ENTER(db);
1159         dn = DB_DNODE(db);
1160         /*
1161          * Shouldn't dirty a regular buffer in syncing context.  Private
1162          * objects may be dirtied in syncing context, but only if they
1163          * were already pre-dirtied in open context.
1164          */
1165         ASSERT(!dmu_tx_is_syncing(tx) ||
1166             BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1167             DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1168             dn->dn_objset->os_dsl_dataset == NULL);
1169         /*
1170          * We make this assert for private objects as well, but after we
1171          * check if we're already dirty.  They are allowed to re-dirty
1172          * in syncing context.
1173          */
1174         ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1175             dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1176             (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1177 
1178         mutex_enter(&db->db_mtx);
1179         /*
1180          * XXX make this true for indirects too?  The problem is that
1181          * transactions created with dmu_tx_create_assigned() from
1182          * syncing context don't bother holding ahead.
1183          */
1184         ASSERT(db->db_level != 0 ||
1185             db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1186             db->db_state == DB_NOFILL);
1187 
1188         mutex_enter(&dn->dn_mtx);
1189         /*
1190          * Don't set dirtyctx to SYNC if we're just modifying this as we
1191          * initialize the objset.
1192          */
1193         if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1194             !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1195                 dn->dn_dirtyctx =
1196                     (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1197                 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1198                 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1199         }
1200         mutex_exit(&dn->dn_mtx);
1201 
1202         if (db->db_blkid == DMU_SPILL_BLKID)
1203                 dn->dn_have_spill = B_TRUE;
1204 
1205         /*
1206          * If this buffer is already dirty, we're done.
1207          */
1208         drp = &db->db_last_dirty;
1209         ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1210             db->db.db_object == DMU_META_DNODE_OBJECT);
1211         while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1212                 drp = &dr->dr_next;
1213         if (dr && dr->dr_txg == tx->tx_txg) {
1214                 DB_DNODE_EXIT(db);
1215 
1216                 dbuf_redirty(dr);
1217                 mutex_exit(&db->db_mtx);
1218                 return (dr);
1219         }
1220 
1221         /*
1222          * Only valid if not already dirty.
1223          */
1224         ASSERT(dn->dn_object == 0 ||
1225             dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1226             (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1227 
1228         ASSERT3U(dn->dn_nlevels, >, db->db_level);
1229         ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1230             dn->dn_phys->dn_nlevels > db->db_level ||
1231             dn->dn_next_nlevels[txgoff] > db->db_level ||
1232             dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1233             dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1234 
1235         /*
1236          * We should only be dirtying in syncing context if it's the
1237          * mos or we're initializing the os or it's a special object.
1238          * However, we are allowed to dirty in syncing context provided
1239          * we already dirtied it in open context.  Hence we must make
1240          * this assertion only if we're not already dirty.
1241          */
1242         os = dn->dn_objset;
1243         ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1244             os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1245         ASSERT(db->db.db_size != 0);
1246 
1247         dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1248 
1249         if (db->db_blkid != DMU_BONUS_BLKID) {
1250                 /*
1251                  * Update the accounting.
1252                  * Note: we delay "free accounting" until after we drop
1253                  * the db_mtx.  This keeps us from grabbing other locks
1254                  * (and possibly deadlocking) in bp_get_dsize() while
1255                  * also holding the db_mtx.
1256                  */
1257                 dnode_willuse_space(dn, db->db.db_size, tx);
1258                 do_free_accounting = dbuf_block_freeable(db);
1259         }
1260 
1261         /*
1262          * If this buffer is dirty in an old transaction group we need
1263          * to make a copy of it so that the changes we make in this
1264          * transaction group won't leak out when we sync the older txg.
1265          */
1266         dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1267         if (db->db_level == 0) {
1268                 void *data_old = db->db_buf;
1269 
1270                 if (db->db_state != DB_NOFILL) {
1271                         if (db->db_blkid == DMU_BONUS_BLKID) {
1272                                 dbuf_fix_old_data(db, tx->tx_txg);
1273                                 data_old = db->db.db_data;
1274                         } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1275                                 /*
1276                                  * Release the data buffer from the cache so
1277                                  * that we can modify it without impacting
1278                                  * possible other users of this cached data
1279                                  * block.  Note that indirect blocks and
1280                                  * private objects are not released until the
1281                                  * syncing state (since they are only modified
1282                                  * then).
1283                                  */
1284                                 arc_release(db->db_buf, db);
1285                                 dbuf_fix_old_data(db, tx->tx_txg);
1286                                 data_old = db->db_buf;
1287                         }
1288                         ASSERT(data_old != NULL);
1289                 }
1290                 dr->dt.dl.dr_data = data_old;
1291         } else {
1292                 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1293                 list_create(&dr->dt.di.dr_children,
1294                     sizeof (dbuf_dirty_record_t),
1295                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
1296         }
1297         if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1298                 dr->dr_accounted = db->db.db_size;
1299         dr->dr_dbuf = db;
1300         dr->dr_txg = tx->tx_txg;
1301         dr->dr_next = *drp;
1302         *drp = dr;
1303 
1304         /*
1305          * We could have been freed_in_flight between the dbuf_noread
1306          * and dbuf_dirty.  We win, as though the dbuf_noread() had
1307          * happened after the free.
1308          */
1309         if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1310             db->db_blkid != DMU_SPILL_BLKID) {
1311                 mutex_enter(&dn->dn_mtx);
1312                 if (dn->dn_free_ranges[txgoff] != NULL) {
1313                         range_tree_clear(dn->dn_free_ranges[txgoff],
1314                             db->db_blkid, 1);
1315                 }
1316                 mutex_exit(&dn->dn_mtx);
1317                 db->db_freed_in_flight = FALSE;
1318         }
1319 
1320         /*
1321          * This buffer is now part of this txg
1322          */
1323         dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1324         db->db_dirtycnt += 1;
1325         ASSERT3U(db->db_dirtycnt, <=, 3);
1326 
1327         mutex_exit(&db->db_mtx);
1328 
1329         if (db->db_blkid == DMU_BONUS_BLKID ||
1330             db->db_blkid == DMU_SPILL_BLKID) {
1331                 mutex_enter(&dn->dn_mtx);
1332                 ASSERT(!list_link_active(&dr->dr_dirty_node));
1333                 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1334                 mutex_exit(&dn->dn_mtx);
1335                 dnode_setdirty(dn, tx);
1336                 DB_DNODE_EXIT(db);
1337                 return (dr);
1338         } else if (do_free_accounting) {
1339                 blkptr_t *bp = db->db_blkptr;
1340                 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1341                     bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1342                 /*
1343                  * This is only a guess -- if the dbuf is dirty
1344                  * in a previous txg, we don't know how much
1345                  * space it will use on disk yet.  We should
1346                  * really have the struct_rwlock to access
1347                  * db_blkptr, but since this is just a guess,
1348                  * it's OK if we get an odd answer.
1349                  */
1350                 ddt_prefetch(os->os_spa, bp);
1351                 dnode_willuse_space(dn, -willfree, tx);
1352         }
1353 
1354         if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1355                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1356                 drop_struct_lock = TRUE;
1357         }
1358 
1359         if (db->db_level == 0) {
1360                 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1361                 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1362         }
1363 
1364         if (db->db_level+1 < dn->dn_nlevels) {
1365                 dmu_buf_impl_t *parent = db->db_parent;
1366                 dbuf_dirty_record_t *di;
1367                 int parent_held = FALSE;
1368 
1369                 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1370                         int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1371 
1372                         parent = dbuf_hold_level(dn, db->db_level+1,
1373                             db->db_blkid >> epbs, FTAG);
1374                         ASSERT(parent != NULL);
1375                         parent_held = TRUE;
1376                 }
1377                 if (drop_struct_lock)
1378                         rw_exit(&dn->dn_struct_rwlock);
1379                 ASSERT3U(db->db_level+1, ==, parent->db_level);
1380                 di = dbuf_dirty(parent, tx);
1381                 if (parent_held)
1382                         dbuf_rele(parent, FTAG);
1383 
1384                 mutex_enter(&db->db_mtx);
1385                 /*
1386                  * Since we've dropped the mutex, it's possible that
1387                  * dbuf_undirty() might have changed this out from under us.
1388                  */
1389                 if (db->db_last_dirty == dr ||
1390                     dn->dn_object == DMU_META_DNODE_OBJECT) {
1391                         mutex_enter(&di->dt.di.dr_mtx);
1392                         ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1393                         ASSERT(!list_link_active(&dr->dr_dirty_node));
1394                         list_insert_tail(&di->dt.di.dr_children, dr);
1395                         mutex_exit(&di->dt.di.dr_mtx);
1396                         dr->dr_parent = di;
1397                 }
1398                 mutex_exit(&db->db_mtx);
1399         } else {
1400                 ASSERT(db->db_level+1 == dn->dn_nlevels);
1401                 ASSERT(db->db_blkid < dn->dn_nblkptr);
1402                 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1403                 mutex_enter(&dn->dn_mtx);
1404                 ASSERT(!list_link_active(&dr->dr_dirty_node));
1405                 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1406                 mutex_exit(&dn->dn_mtx);
1407                 if (drop_struct_lock)
1408                         rw_exit(&dn->dn_struct_rwlock);
1409         }
1410 
1411         dnode_setdirty(dn, tx);
1412         DB_DNODE_EXIT(db);
1413         return (dr);
1414 }
1415 
1416 /*
1417  * Undirty a buffer in the transaction group referenced by the given
1418  * transaction.  Return whether this evicted the dbuf.
1419  */
1420 static boolean_t
1421 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1422 {
1423         dnode_t *dn;
1424         uint64_t txg = tx->tx_txg;
1425         dbuf_dirty_record_t *dr, **drp;
1426 
1427         ASSERT(txg != 0);
1428 
1429         /*
1430          * Due to our use of dn_nlevels below, this can only be called
1431          * in open context, unless we are operating on the MOS.
1432          * From syncing context, dn_nlevels may be different from the
1433          * dn_nlevels used when dbuf was dirtied.
1434          */
1435         ASSERT(db->db_objset ==
1436             dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1437             txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1438         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1439         ASSERT0(db->db_level);
1440         ASSERT(MUTEX_HELD(&db->db_mtx));
1441 
1442         /*
1443          * If this buffer is not dirty, we're done.
1444          */
1445         for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1446                 if (dr->dr_txg <= txg)
1447                         break;
1448         if (dr == NULL || dr->dr_txg < txg)
1449                 return (B_FALSE);
1450         ASSERT(dr->dr_txg == txg);
1451         ASSERT(dr->dr_dbuf == db);
1452 
1453         DB_DNODE_ENTER(db);
1454         dn = DB_DNODE(db);
1455 
1456         dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1457 
1458         ASSERT(db->db.db_size != 0);
1459 
1460         dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1461             dr->dr_accounted, txg);
1462 
1463         *drp = dr->dr_next;
1464 
1465         /*
1466          * Note that there are three places in dbuf_dirty()
1467          * where this dirty record may be put on a list.
1468          * Make sure to do a list_remove corresponding to
1469          * every one of those list_insert calls.
1470          */
1471         if (dr->dr_parent) {
1472                 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1473                 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1474                 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1475         } else if (db->db_blkid == DMU_SPILL_BLKID ||
1476             db->db_level + 1 == dn->dn_nlevels) {
1477                 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1478                 mutex_enter(&dn->dn_mtx);
1479                 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1480                 mutex_exit(&dn->dn_mtx);
1481         }
1482         DB_DNODE_EXIT(db);
1483 
1484         if (db->db_state != DB_NOFILL) {
1485                 dbuf_unoverride(dr);
1486 
1487                 ASSERT(db->db_buf != NULL);
1488                 ASSERT(dr->dt.dl.dr_data != NULL);
1489                 if (dr->dt.dl.dr_data != db->db_buf)
1490                         VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1491         }
1492 
1493         kmem_free(dr, sizeof (dbuf_dirty_record_t));
1494 
1495         ASSERT(db->db_dirtycnt > 0);
1496         db->db_dirtycnt -= 1;
1497 
1498         if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1499                 arc_buf_t *buf = db->db_buf;
1500 
1501                 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1502                 dbuf_clear_data(db);
1503                 VERIFY(arc_buf_remove_ref(buf, db));
1504                 dbuf_evict(db);
1505                 return (B_TRUE);
1506         }
1507 
1508         return (B_FALSE);
1509 }
1510 
1511 void
1512 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1513 {
1514         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1515         int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1516 
1517         ASSERT(tx->tx_txg != 0);
1518         ASSERT(!refcount_is_zero(&db->db_holds));
1519 
1520         /*
1521          * Quick check for dirtyness.  For already dirty blocks, this
1522          * reduces runtime of this function by >90%, and overall performance
1523          * by 50% for some workloads (e.g. file deletion with indirect blocks
1524          * cached).
1525          */
1526         mutex_enter(&db->db_mtx);
1527         dbuf_dirty_record_t *dr;
1528         for (dr = db->db_last_dirty;
1529             dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1530                 /*
1531                  * It's possible that it is already dirty but not cached,
1532                  * because there are some calls to dbuf_dirty() that don't
1533                  * go through dmu_buf_will_dirty().
1534                  */
1535                 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1536                         /* This dbuf is already dirty and cached. */
1537                         dbuf_redirty(dr);
1538                         mutex_exit(&db->db_mtx);
1539                         return;
1540                 }
1541         }
1542         mutex_exit(&db->db_mtx);
1543 
1544         DB_DNODE_ENTER(db);
1545         if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1546                 rf |= DB_RF_HAVESTRUCT;
1547         DB_DNODE_EXIT(db);
1548         (void) dbuf_read(db, NULL, rf);
1549         (void) dbuf_dirty(db, tx);
1550 }
1551 
1552 void
1553 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1554 {
1555         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1556 
1557         db->db_state = DB_NOFILL;
1558 
1559         dmu_buf_will_fill(db_fake, tx);
1560 }
1561 
1562 void
1563 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1564 {
1565         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1566 
1567         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1568         ASSERT(tx->tx_txg != 0);
1569         ASSERT(db->db_level == 0);
1570         ASSERT(!refcount_is_zero(&db->db_holds));
1571 
1572         ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1573             dmu_tx_private_ok(tx));
1574 
1575         dbuf_noread(db);
1576         (void) dbuf_dirty(db, tx);
1577 }
1578 
1579 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1580 /* ARGSUSED */
1581 void
1582 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1583 {
1584         mutex_enter(&db->db_mtx);
1585         DBUF_VERIFY(db);
1586 
1587         if (db->db_state == DB_FILL) {
1588                 if (db->db_level == 0 && db->db_freed_in_flight) {
1589                         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1590                         /* we were freed while filling */
1591                         /* XXX dbuf_undirty? */
1592                         bzero(db->db.db_data, db->db.db_size);
1593                         db->db_freed_in_flight = FALSE;
1594                 }
1595                 db->db_state = DB_CACHED;
1596                 cv_broadcast(&db->db_changed);
1597         }
1598         mutex_exit(&db->db_mtx);
1599 }
1600 
1601 void
1602 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1603     bp_embedded_type_t etype, enum zio_compress comp,
1604     int uncompressed_size, int compressed_size, int byteorder,
1605     dmu_tx_t *tx)
1606 {
1607         dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1608         struct dirty_leaf *dl;
1609         dmu_object_type_t type;
1610 
1611         if (etype == BP_EMBEDDED_TYPE_DATA) {
1612                 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1613                     SPA_FEATURE_EMBEDDED_DATA));
1614         }
1615 
1616         DB_DNODE_ENTER(db);
1617         type = DB_DNODE(db)->dn_type;
1618         DB_DNODE_EXIT(db);
1619 
1620         ASSERT0(db->db_level);
1621         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1622 
1623         dmu_buf_will_not_fill(dbuf, tx);
1624 
1625         ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1626         dl = &db->db_last_dirty->dt.dl;
1627         encode_embedded_bp_compressed(&dl->dr_overridden_by,
1628             data, comp, uncompressed_size, compressed_size);
1629         BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1630         BP_SET_TYPE(&dl->dr_overridden_by, type);
1631         BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1632         BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1633 
1634         dl->dr_override_state = DR_OVERRIDDEN;
1635         dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1636 }
1637 
1638 /*
1639  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1640  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1641  */
1642 void
1643 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1644 {
1645         ASSERT(!refcount_is_zero(&db->db_holds));
1646         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1647         ASSERT(db->db_level == 0);
1648         ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1649         ASSERT(buf != NULL);
1650         ASSERT(arc_buf_size(buf) == db->db.db_size);
1651         ASSERT(tx->tx_txg != 0);
1652 
1653         arc_return_buf(buf, db);
1654         ASSERT(arc_released(buf));
1655 
1656         mutex_enter(&db->db_mtx);
1657 
1658         while (db->db_state == DB_READ || db->db_state == DB_FILL)
1659                 cv_wait(&db->db_changed, &db->db_mtx);
1660 
1661         ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1662 
1663         if (db->db_state == DB_CACHED &&
1664             refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1665                 mutex_exit(&db->db_mtx);
1666                 (void) dbuf_dirty(db, tx);
1667                 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1668                 VERIFY(arc_buf_remove_ref(buf, db));
1669                 xuio_stat_wbuf_copied();
1670                 return;
1671         }
1672 
1673         xuio_stat_wbuf_nocopy();
1674         if (db->db_state == DB_CACHED) {
1675                 dbuf_dirty_record_t *dr = db->db_last_dirty;
1676 
1677                 ASSERT(db->db_buf != NULL);
1678                 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1679                         ASSERT(dr->dt.dl.dr_data == db->db_buf);
1680                         if (!arc_released(db->db_buf)) {
1681                                 ASSERT(dr->dt.dl.dr_override_state ==
1682                                     DR_OVERRIDDEN);
1683                                 arc_release(db->db_buf, db);
1684                         }
1685                         dr->dt.dl.dr_data = buf;
1686                         VERIFY(arc_buf_remove_ref(db->db_buf, db));
1687                 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1688                         arc_release(db->db_buf, db);
1689                         VERIFY(arc_buf_remove_ref(db->db_buf, db));
1690                 }
1691                 db->db_buf = NULL;
1692         }
1693         ASSERT(db->db_buf == NULL);
1694         dbuf_set_data(db, buf);
1695         db->db_state = DB_FILL;
1696         mutex_exit(&db->db_mtx);
1697         (void) dbuf_dirty(db, tx);
1698         dmu_buf_fill_done(&db->db, tx);
1699 }
1700 
1701 /*
1702  * "Clear" the contents of this dbuf.  This will mark the dbuf
1703  * EVICTING and clear *most* of its references.  Unfortunately,
1704  * when we are not holding the dn_dbufs_mtx, we can't clear the
1705  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1706  * in this case.  For callers from the DMU we will usually see:
1707  *      dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1708  * For the arc callback, we will usually see:
1709  *      dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1710  * Sometimes, though, we will get a mix of these two:
1711  *      DMU: dbuf_clear()->arc_clear_callback()
1712  *      ARC: dbuf_do_evict()->dbuf_destroy()
1713  *
1714  * This routine will dissociate the dbuf from the arc, by calling
1715  * arc_clear_callback(), but will not evict the data from the ARC.
1716  */
1717 void
1718 dbuf_clear(dmu_buf_impl_t *db)
1719 {
1720         dnode_t *dn;
1721         dmu_buf_impl_t *parent = db->db_parent;
1722         dmu_buf_impl_t *dndb;
1723         boolean_t dbuf_gone = B_FALSE;
1724 
1725         ASSERT(MUTEX_HELD(&db->db_mtx));
1726         ASSERT(refcount_is_zero(&db->db_holds));
1727 
1728         dbuf_evict_user(db);
1729 
1730         if (db->db_state == DB_CACHED) {
1731                 ASSERT(db->db.db_data != NULL);
1732                 if (db->db_blkid == DMU_BONUS_BLKID) {
1733                         zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1734                         arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1735                 }
1736                 db->db.db_data = NULL;
1737                 db->db_state = DB_UNCACHED;
1738         }
1739 
1740         ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1741         ASSERT(db->db_data_pending == NULL);
1742 
1743         db->db_state = DB_EVICTING;
1744         db->db_blkptr = NULL;
1745 
1746         DB_DNODE_ENTER(db);
1747         dn = DB_DNODE(db);
1748         dndb = dn->dn_dbuf;
1749         if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1750                 avl_remove(&dn->dn_dbufs, db);
1751                 atomic_dec_32(&dn->dn_dbufs_count);
1752                 membar_producer();
1753                 DB_DNODE_EXIT(db);
1754                 /*
1755                  * Decrementing the dbuf count means that the hold corresponding
1756                  * to the removed dbuf is no longer discounted in dnode_move(),
1757                  * so the dnode cannot be moved until after we release the hold.
1758                  * The membar_producer() ensures visibility of the decremented
1759                  * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1760                  * release any lock.
1761                  */
1762                 dnode_rele(dn, db);
1763                 db->db_dnode_handle = NULL;
1764         } else {
1765                 DB_DNODE_EXIT(db);
1766         }
1767 
1768         if (db->db_buf)
1769                 dbuf_gone = arc_clear_callback(db->db_buf);
1770 
1771         if (!dbuf_gone)
1772                 mutex_exit(&db->db_mtx);
1773 
1774         /*
1775          * If this dbuf is referenced from an indirect dbuf,
1776          * decrement the ref count on the indirect dbuf.
1777          */
1778         if (parent && parent != dndb)
1779                 dbuf_rele(parent, db);
1780 }
1781 
1782 /*
1783  * Note: While bpp will always be updated if the function returns success,
1784  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1785  * this happens when the dnode is the meta-dnode, or a userused or groupused
1786  * object.
1787  */
1788 static int
1789 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1790     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1791 {
1792         int nlevels, epbs;
1793 
1794         *parentp = NULL;
1795         *bpp = NULL;
1796 
1797         ASSERT(blkid != DMU_BONUS_BLKID);
1798 
1799         if (blkid == DMU_SPILL_BLKID) {
1800                 mutex_enter(&dn->dn_mtx);
1801                 if (dn->dn_have_spill &&
1802                     (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1803                         *bpp = &dn->dn_phys->dn_spill;
1804                 else
1805                         *bpp = NULL;
1806                 dbuf_add_ref(dn->dn_dbuf, NULL);
1807                 *parentp = dn->dn_dbuf;
1808                 mutex_exit(&dn->dn_mtx);
1809                 return (0);
1810         }
1811 
1812         if (dn->dn_phys->dn_nlevels == 0)
1813                 nlevels = 1;
1814         else
1815                 nlevels = dn->dn_phys->dn_nlevels;
1816 
1817         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1818 
1819         ASSERT3U(level * epbs, <, 64);
1820         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1821         if (level >= nlevels ||
1822             (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1823                 /* the buffer has no parent yet */
1824                 return (SET_ERROR(ENOENT));
1825         } else if (level < nlevels-1) {
1826                 /* this block is referenced from an indirect block */
1827                 int err = dbuf_hold_impl(dn, level+1,
1828                     blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
1829                 if (err)
1830                         return (err);
1831                 err = dbuf_read(*parentp, NULL,
1832                     (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1833                 if (err) {
1834                         dbuf_rele(*parentp, NULL);
1835                         *parentp = NULL;
1836                         return (err);
1837                 }
1838                 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1839                     (blkid & ((1ULL << epbs) - 1));
1840                 return (0);
1841         } else {
1842                 /* the block is referenced from the dnode */
1843                 ASSERT3U(level, ==, nlevels-1);
1844                 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1845                     blkid < dn->dn_phys->dn_nblkptr);
1846                 if (dn->dn_dbuf) {
1847                         dbuf_add_ref(dn->dn_dbuf, NULL);
1848                         *parentp = dn->dn_dbuf;
1849                 }
1850                 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1851                 return (0);
1852         }
1853 }
1854 
1855 static dmu_buf_impl_t *
1856 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1857     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1858 {
1859         objset_t *os = dn->dn_objset;
1860         dmu_buf_impl_t *db, *odb;
1861 
1862         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1863         ASSERT(dn->dn_type != DMU_OT_NONE);
1864 
1865         db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1866 
1867         db->db_objset = os;
1868         db->db.db_object = dn->dn_object;
1869         db->db_level = level;
1870         db->db_blkid = blkid;
1871         db->db_last_dirty = NULL;
1872         db->db_dirtycnt = 0;
1873         db->db_dnode_handle = dn->dn_handle;
1874         db->db_parent = parent;
1875         db->db_blkptr = blkptr;
1876 
1877         db->db_user = NULL;
1878         db->db_user_immediate_evict = FALSE;
1879         db->db_freed_in_flight = FALSE;
1880         db->db_pending_evict = FALSE;
1881 
1882         if (blkid == DMU_BONUS_BLKID) {
1883                 ASSERT3P(parent, ==, dn->dn_dbuf);
1884                 db->db.db_size = DN_MAX_BONUSLEN -
1885                     (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1886                 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1887                 db->db.db_offset = DMU_BONUS_BLKID;
1888                 db->db_state = DB_UNCACHED;
1889                 /* the bonus dbuf is not placed in the hash table */
1890                 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1891                 return (db);
1892         } else if (blkid == DMU_SPILL_BLKID) {
1893                 db->db.db_size = (blkptr != NULL) ?
1894                     BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1895                 db->db.db_offset = 0;
1896         } else {
1897                 int blocksize =
1898                     db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1899                 db->db.db_size = blocksize;
1900                 db->db.db_offset = db->db_blkid * blocksize;
1901         }
1902 
1903         /*
1904          * Hold the dn_dbufs_mtx while we get the new dbuf
1905          * in the hash table *and* added to the dbufs list.
1906          * This prevents a possible deadlock with someone
1907          * trying to look up this dbuf before its added to the
1908          * dn_dbufs list.
1909          */
1910         mutex_enter(&dn->dn_dbufs_mtx);
1911         db->db_state = DB_EVICTING;
1912         if ((odb = dbuf_hash_insert(db)) != NULL) {
1913                 /* someone else inserted it first */
1914                 kmem_cache_free(dbuf_cache, db);
1915                 mutex_exit(&dn->dn_dbufs_mtx);
1916                 return (odb);
1917         }
1918         avl_add(&dn->dn_dbufs, db);
1919         if (db->db_level == 0 && db->db_blkid >=
1920             dn->dn_unlisted_l0_blkid)
1921                 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1922         db->db_state = DB_UNCACHED;
1923         mutex_exit(&dn->dn_dbufs_mtx);
1924         arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1925 
1926         if (parent && parent != dn->dn_dbuf)
1927                 dbuf_add_ref(parent, db);
1928 
1929         ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1930             refcount_count(&dn->dn_holds) > 0);
1931         (void) refcount_add(&dn->dn_holds, db);
1932         atomic_inc_32(&dn->dn_dbufs_count);
1933 
1934         dprintf_dbuf(db, "db=%p\n", db);
1935 
1936         return (db);
1937 }
1938 
1939 static int
1940 dbuf_do_evict(void *private)
1941 {
1942         dmu_buf_impl_t *db = private;
1943 
1944         if (!MUTEX_HELD(&db->db_mtx))
1945                 mutex_enter(&db->db_mtx);
1946 
1947         ASSERT(refcount_is_zero(&db->db_holds));
1948 
1949         if (db->db_state != DB_EVICTING) {
1950                 ASSERT(db->db_state == DB_CACHED);
1951                 DBUF_VERIFY(db);
1952                 db->db_buf = NULL;
1953                 dbuf_evict(db);
1954         } else {
1955                 mutex_exit(&db->db_mtx);
1956                 dbuf_destroy(db);
1957         }
1958         return (0);
1959 }
1960 
1961 static void
1962 dbuf_destroy(dmu_buf_impl_t *db)
1963 {
1964         ASSERT(refcount_is_zero(&db->db_holds));
1965 
1966         if (db->db_blkid != DMU_BONUS_BLKID) {
1967                 /*
1968                  * If this dbuf is still on the dn_dbufs list,
1969                  * remove it from that list.
1970                  */
1971                 if (db->db_dnode_handle != NULL) {
1972                         dnode_t *dn;
1973 
1974                         DB_DNODE_ENTER(db);
1975                         dn = DB_DNODE(db);
1976                         mutex_enter(&dn->dn_dbufs_mtx);
1977                         avl_remove(&dn->dn_dbufs, db);
1978                         atomic_dec_32(&dn->dn_dbufs_count);
1979                         mutex_exit(&dn->dn_dbufs_mtx);
1980                         DB_DNODE_EXIT(db);
1981                         /*
1982                          * Decrementing the dbuf count means that the hold
1983                          * corresponding to the removed dbuf is no longer
1984                          * discounted in dnode_move(), so the dnode cannot be
1985                          * moved until after we release the hold.
1986                          */
1987                         dnode_rele(dn, db);
1988                         db->db_dnode_handle = NULL;
1989                 }
1990                 dbuf_hash_remove(db);
1991         }
1992         db->db_parent = NULL;
1993         db->db_buf = NULL;
1994 
1995         ASSERT(db->db.db_data == NULL);
1996         ASSERT(db->db_hash_next == NULL);
1997         ASSERT(db->db_blkptr == NULL);
1998         ASSERT(db->db_data_pending == NULL);
1999 
2000         kmem_cache_free(dbuf_cache, db);
2001         arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2002 }
2003 
2004 typedef struct dbuf_prefetch_arg {
2005         spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2006         zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2007         int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2008         int dpa_curlevel; /* The current level that we're reading */
2009         zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2010         zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2011         arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2012 } dbuf_prefetch_arg_t;
2013 
2014 /*
2015  * Actually issue the prefetch read for the block given.
2016  */
2017 static void
2018 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2019 {
2020         if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2021                 return;
2022 
2023         arc_flags_t aflags =
2024             dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2025 
2026         ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2027         ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2028         ASSERT(dpa->dpa_zio != NULL);
2029         (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2030             dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2031             &aflags, &dpa->dpa_zb);
2032 }
2033 
2034 /*
2035  * Called when an indirect block above our prefetch target is read in.  This
2036  * will either read in the next indirect block down the tree or issue the actual
2037  * prefetch if the next block down is our target.
2038  */
2039 static void
2040 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2041 {
2042         dbuf_prefetch_arg_t *dpa = private;
2043 
2044         ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2045         ASSERT3S(dpa->dpa_curlevel, >, 0);
2046         if (zio != NULL) {
2047                 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2048                 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2049                 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2050         }
2051 
2052         dpa->dpa_curlevel--;
2053 
2054         uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2055             (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2056         blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2057             P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2058         if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2059                 kmem_free(dpa, sizeof (*dpa));
2060         } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2061                 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2062                 dbuf_issue_final_prefetch(dpa, bp);
2063                 kmem_free(dpa, sizeof (*dpa));
2064         } else {
2065                 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2066                 zbookmark_phys_t zb;
2067 
2068                 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2069 
2070                 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2071                     dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2072 
2073                 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2074                     bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2075                     ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2076                     &iter_aflags, &zb);
2077         }
2078         (void) arc_buf_remove_ref(abuf, private);
2079 }
2080 
2081 /*
2082  * Issue prefetch reads for the given block on the given level.  If the indirect
2083  * blocks above that block are not in memory, we will read them in
2084  * asynchronously.  As a result, this call never blocks waiting for a read to
2085  * complete.
2086  */
2087 void
2088 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2089     arc_flags_t aflags)
2090 {
2091         blkptr_t bp;
2092         int epbs, nlevels, curlevel;
2093         uint64_t curblkid;
2094 
2095         ASSERT(blkid != DMU_BONUS_BLKID);
2096         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2097 
2098         if (blkid > dn->dn_maxblkid)
2099                 return;
2100 
2101         if (dnode_block_freed(dn, blkid))
2102                 return;
2103 
2104         /*
2105          * This dnode hasn't been written to disk yet, so there's nothing to
2106          * prefetch.
2107          */
2108         nlevels = dn->dn_phys->dn_nlevels;
2109         if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2110                 return;
2111 
2112         epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2113         if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2114                 return;
2115 
2116         dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2117             level, blkid);
2118         if (db != NULL) {
2119                 mutex_exit(&db->db_mtx);
2120                 /*
2121                  * This dbuf already exists.  It is either CACHED, or
2122                  * (we assume) about to be read or filled.
2123                  */
2124                 return;
2125         }
2126 
2127         /*
2128          * Find the closest ancestor (indirect block) of the target block
2129          * that is present in the cache.  In this indirect block, we will
2130          * find the bp that is at curlevel, curblkid.
2131          */
2132         curlevel = level;
2133         curblkid = blkid;
2134         while (curlevel < nlevels - 1) {
2135                 int parent_level = curlevel + 1;
2136                 uint64_t parent_blkid = curblkid >> epbs;
2137                 dmu_buf_impl_t *db;
2138 
2139                 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2140                     FALSE, TRUE, FTAG, &db) == 0) {
2141                         blkptr_t *bpp = db->db_buf->b_data;
2142                         bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2143                         dbuf_rele(db, FTAG);
2144                         break;
2145                 }
2146 
2147                 curlevel = parent_level;
2148                 curblkid = parent_blkid;
2149         }
2150 
2151         if (curlevel == nlevels - 1) {
2152                 /* No cached indirect blocks found. */
2153                 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2154                 bp = dn->dn_phys->dn_blkptr[curblkid];
2155         }
2156         if (BP_IS_HOLE(&bp))
2157                 return;
2158 
2159         ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2160 
2161         zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2162             ZIO_FLAG_CANFAIL);
2163 
2164         dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2165         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2166         SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2167             dn->dn_object, level, blkid);
2168         dpa->dpa_curlevel = curlevel;
2169         dpa->dpa_prio = prio;
2170         dpa->dpa_aflags = aflags;
2171         dpa->dpa_spa = dn->dn_objset->os_spa;
2172         dpa->dpa_epbs = epbs;
2173         dpa->dpa_zio = pio;
2174 
2175         /*
2176          * If we have the indirect just above us, no need to do the asynchronous
2177          * prefetch chain; we'll just run the last step ourselves.  If we're at
2178          * a higher level, though, we want to issue the prefetches for all the
2179          * indirect blocks asynchronously, so we can go on with whatever we were
2180          * doing.
2181          */
2182         if (curlevel == level) {
2183                 ASSERT3U(curblkid, ==, blkid);
2184                 dbuf_issue_final_prefetch(dpa, &bp);
2185                 kmem_free(dpa, sizeof (*dpa));
2186         } else {
2187                 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2188                 zbookmark_phys_t zb;
2189 
2190                 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2191                     dn->dn_object, curlevel, curblkid);
2192                 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2193                     &bp, dbuf_prefetch_indirect_done, dpa, prio,
2194                     ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2195                     &iter_aflags, &zb);
2196         }
2197         /*
2198          * We use pio here instead of dpa_zio since it's possible that
2199          * dpa may have already been freed.
2200          */
2201         zio_nowait(pio);
2202 }
2203 
2204 /*
2205  * Returns with db_holds incremented, and db_mtx not held.
2206  * Note: dn_struct_rwlock must be held.
2207  */
2208 int
2209 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2210     boolean_t fail_sparse, boolean_t fail_uncached,
2211     void *tag, dmu_buf_impl_t **dbp)
2212 {
2213         dmu_buf_impl_t *db, *parent = NULL;
2214 
2215         ASSERT(blkid != DMU_BONUS_BLKID);
2216         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2217         ASSERT3U(dn->dn_nlevels, >, level);
2218 
2219         *dbp = NULL;
2220 top:
2221         /* dbuf_find() returns with db_mtx held */
2222         db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2223 
2224         if (db == NULL) {
2225                 blkptr_t *bp = NULL;
2226                 int err;
2227 
2228                 if (fail_uncached)
2229                         return (SET_ERROR(ENOENT));
2230 
2231                 ASSERT3P(parent, ==, NULL);
2232                 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2233                 if (fail_sparse) {
2234                         if (err == 0 && bp && BP_IS_HOLE(bp))
2235                                 err = SET_ERROR(ENOENT);
2236                         if (err) {
2237                                 if (parent)
2238                                         dbuf_rele(parent, NULL);
2239                                 return (err);
2240                         }
2241                 }
2242                 if (err && err != ENOENT)
2243                         return (err);
2244                 db = dbuf_create(dn, level, blkid, parent, bp);
2245         }
2246 
2247         if (fail_uncached && db->db_state != DB_CACHED) {
2248                 mutex_exit(&db->db_mtx);
2249                 return (SET_ERROR(ENOENT));
2250         }
2251 
2252         if (db->db_buf && refcount_is_zero(&db->db_holds)) {
2253                 arc_buf_add_ref(db->db_buf, db);
2254                 if (db->db_buf->b_data == NULL) {
2255                         dbuf_clear(db);
2256                         if (parent) {
2257                                 dbuf_rele(parent, NULL);
2258                                 parent = NULL;
2259                         }
2260                         goto top;
2261                 }
2262                 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2263         }
2264 
2265         ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2266 
2267         /*
2268          * If this buffer is currently syncing out, and we are are
2269          * still referencing it from db_data, we need to make a copy
2270          * of it in case we decide we want to dirty it again in this txg.
2271          */
2272         if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2273             dn->dn_object != DMU_META_DNODE_OBJECT &&
2274             db->db_state == DB_CACHED && db->db_data_pending) {
2275                 dbuf_dirty_record_t *dr = db->db_data_pending;
2276 
2277                 if (dr->dt.dl.dr_data == db->db_buf) {
2278                         arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2279 
2280                         dbuf_set_data(db,
2281                             arc_buf_alloc(dn->dn_objset->os_spa,
2282                             db->db.db_size, db, type));
2283                         bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2284                             db->db.db_size);
2285                 }
2286         }
2287 
2288         (void) refcount_add(&db->db_holds, tag);
2289         DBUF_VERIFY(db);
2290         mutex_exit(&db->db_mtx);
2291 
2292         /* NOTE: we can't rele the parent until after we drop the db_mtx */
2293         if (parent)
2294                 dbuf_rele(parent, NULL);
2295 
2296         ASSERT3P(DB_DNODE(db), ==, dn);
2297         ASSERT3U(db->db_blkid, ==, blkid);
2298         ASSERT3U(db->db_level, ==, level);
2299         *dbp = db;
2300 
2301         return (0);
2302 }
2303 
2304 dmu_buf_impl_t *
2305 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2306 {
2307         return (dbuf_hold_level(dn, 0, blkid, tag));
2308 }
2309 
2310 dmu_buf_impl_t *
2311 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2312 {
2313         dmu_buf_impl_t *db;
2314         int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2315         return (err ? NULL : db);
2316 }
2317 
2318 void
2319 dbuf_create_bonus(dnode_t *dn)
2320 {
2321         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2322 
2323         ASSERT(dn->dn_bonus == NULL);
2324         dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2325 }
2326 
2327 int
2328 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2329 {
2330         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2331         dnode_t *dn;
2332 
2333         if (db->db_blkid != DMU_SPILL_BLKID)
2334                 return (SET_ERROR(ENOTSUP));
2335         if (blksz == 0)
2336                 blksz = SPA_MINBLOCKSIZE;
2337         ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2338         blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2339 
2340         DB_DNODE_ENTER(db);
2341         dn = DB_DNODE(db);
2342         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2343         dbuf_new_size(db, blksz, tx);
2344         rw_exit(&dn->dn_struct_rwlock);
2345         DB_DNODE_EXIT(db);
2346 
2347         return (0);
2348 }
2349 
2350 void
2351 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2352 {
2353         dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2354 }
2355 
2356 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2357 void
2358 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2359 {
2360         int64_t holds = refcount_add(&db->db_holds, tag);
2361         ASSERT(holds > 1);
2362 }
2363 
2364 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2365 boolean_t
2366 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2367     void *tag)
2368 {
2369         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2370         dmu_buf_impl_t *found_db;
2371         boolean_t result = B_FALSE;
2372 
2373         if (db->db_blkid == DMU_BONUS_BLKID)
2374                 found_db = dbuf_find_bonus(os, obj);
2375         else
2376                 found_db = dbuf_find(os, obj, 0, blkid);
2377 
2378         if (found_db != NULL) {
2379                 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2380                         (void) refcount_add(&db->db_holds, tag);
2381                         result = B_TRUE;
2382                 }
2383                 mutex_exit(&db->db_mtx);
2384         }
2385         return (result);
2386 }
2387 
2388 /*
2389  * If you call dbuf_rele() you had better not be referencing the dnode handle
2390  * unless you have some other direct or indirect hold on the dnode. (An indirect
2391  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2392  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2393  * dnode's parent dbuf evicting its dnode handles.
2394  */
2395 void
2396 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2397 {
2398         mutex_enter(&db->db_mtx);
2399         dbuf_rele_and_unlock(db, tag);
2400 }
2401 
2402 void
2403 dmu_buf_rele(dmu_buf_t *db, void *tag)
2404 {
2405         dbuf_rele((dmu_buf_impl_t *)db, tag);
2406 }
2407 
2408 /*
2409  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2410  * db_dirtycnt and db_holds to be updated atomically.
2411  */
2412 void
2413 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2414 {
2415         int64_t holds;
2416 
2417         ASSERT(MUTEX_HELD(&db->db_mtx));
2418         DBUF_VERIFY(db);
2419 
2420         /*
2421          * Remove the reference to the dbuf before removing its hold on the
2422          * dnode so we can guarantee in dnode_move() that a referenced bonus
2423          * buffer has a corresponding dnode hold.
2424          */
2425         holds = refcount_remove(&db->db_holds, tag);
2426         ASSERT(holds >= 0);
2427 
2428         /*
2429          * We can't freeze indirects if there is a possibility that they
2430          * may be modified in the current syncing context.
2431          */
2432         if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2433                 arc_buf_freeze(db->db_buf);
2434 
2435         if (holds == db->db_dirtycnt &&
2436             db->db_level == 0 && db->db_user_immediate_evict)
2437                 dbuf_evict_user(db);
2438 
2439         if (holds == 0) {
2440                 if (db->db_blkid == DMU_BONUS_BLKID) {
2441                         dnode_t *dn;
2442                         boolean_t evict_dbuf = db->db_pending_evict;
2443 
2444                         /*
2445                          * If the dnode moves here, we cannot cross this
2446                          * barrier until the move completes.
2447                          */
2448                         DB_DNODE_ENTER(db);
2449 
2450                         dn = DB_DNODE(db);
2451                         atomic_dec_32(&dn->dn_dbufs_count);
2452 
2453                         /*
2454                          * Decrementing the dbuf count means that the bonus
2455                          * buffer's dnode hold is no longer discounted in
2456                          * dnode_move(). The dnode cannot move until after
2457                          * the dnode_rele() below.
2458                          */
2459                         DB_DNODE_EXIT(db);
2460 
2461                         /*
2462                          * Do not reference db after its lock is dropped.
2463                          * Another thread may evict it.
2464                          */
2465                         mutex_exit(&db->db_mtx);
2466 
2467                         if (evict_dbuf)
2468                                 dnode_evict_bonus(dn);
2469 
2470                         dnode_rele(dn, db);
2471                 } else if (db->db_buf == NULL) {
2472                         /*
2473                          * This is a special case: we never associated this
2474                          * dbuf with any data allocated from the ARC.
2475                          */
2476                         ASSERT(db->db_state == DB_UNCACHED ||
2477                             db->db_state == DB_NOFILL);
2478                         dbuf_evict(db);
2479                 } else if (arc_released(db->db_buf)) {
2480                         arc_buf_t *buf = db->db_buf;
2481                         /*
2482                          * This dbuf has anonymous data associated with it.
2483                          */
2484                         dbuf_clear_data(db);
2485                         VERIFY(arc_buf_remove_ref(buf, db));
2486                         dbuf_evict(db);
2487                 } else {
2488                         VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2489 
2490                         /*
2491                          * A dbuf will be eligible for eviction if either the
2492                          * 'primarycache' property is set or a duplicate
2493                          * copy of this buffer is already cached in the arc.
2494                          *
2495                          * In the case of the 'primarycache' a buffer
2496                          * is considered for eviction if it matches the
2497                          * criteria set in the property.
2498                          *
2499                          * To decide if our buffer is considered a
2500                          * duplicate, we must call into the arc to determine
2501                          * if multiple buffers are referencing the same
2502                          * block on-disk. If so, then we simply evict
2503                          * ourselves.
2504                          */
2505                         if (!DBUF_IS_CACHEABLE(db)) {
2506                                 if (db->db_blkptr != NULL &&
2507                                     !BP_IS_HOLE(db->db_blkptr) &&
2508                                     !BP_IS_EMBEDDED(db->db_blkptr)) {
2509                                         spa_t *spa =
2510                                             dmu_objset_spa(db->db_objset);
2511                                         blkptr_t bp = *db->db_blkptr;
2512                                         dbuf_clear(db);
2513                                         arc_freed(spa, &bp);
2514                                 } else {
2515                                         dbuf_clear(db);
2516                                 }
2517                         } else if (db->db_pending_evict ||
2518                             arc_buf_eviction_needed(db->db_buf)) {
2519                                 dbuf_clear(db);
2520                         } else {
2521                                 mutex_exit(&db->db_mtx);
2522                         }
2523                 }
2524         } else {
2525                 mutex_exit(&db->db_mtx);
2526         }
2527 }
2528 
2529 #pragma weak dmu_buf_refcount = dbuf_refcount
2530 uint64_t
2531 dbuf_refcount(dmu_buf_impl_t *db)
2532 {
2533         return (refcount_count(&db->db_holds));
2534 }
2535 
2536 void *
2537 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2538     dmu_buf_user_t *new_user)
2539 {
2540         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2541 
2542         mutex_enter(&db->db_mtx);
2543         dbuf_verify_user(db, DBVU_NOT_EVICTING);
2544         if (db->db_user == old_user)
2545                 db->db_user = new_user;
2546         else
2547                 old_user = db->db_user;
2548         dbuf_verify_user(db, DBVU_NOT_EVICTING);
2549         mutex_exit(&db->db_mtx);
2550 
2551         return (old_user);
2552 }
2553 
2554 void *
2555 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2556 {
2557         return (dmu_buf_replace_user(db_fake, NULL, user));
2558 }
2559 
2560 void *
2561 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2562 {
2563         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2564 
2565         db->db_user_immediate_evict = TRUE;
2566         return (dmu_buf_set_user(db_fake, user));
2567 }
2568 
2569 void *
2570 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2571 {
2572         return (dmu_buf_replace_user(db_fake, user, NULL));
2573 }
2574 
2575 void *
2576 dmu_buf_get_user(dmu_buf_t *db_fake)
2577 {
2578         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2579 
2580         dbuf_verify_user(db, DBVU_NOT_EVICTING);
2581         return (db->db_user);
2582 }
2583 
2584 void
2585 dmu_buf_user_evict_wait()
2586 {
2587         taskq_wait(dbu_evict_taskq);
2588 }
2589 
2590 boolean_t
2591 dmu_buf_freeable(dmu_buf_t *dbuf)
2592 {
2593         boolean_t res = B_FALSE;
2594         dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2595 
2596         if (db->db_blkptr)
2597                 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2598                     db->db_blkptr, db->db_blkptr->blk_birth);
2599 
2600         return (res);
2601 }
2602 
2603 blkptr_t *
2604 dmu_buf_get_blkptr(dmu_buf_t *db)
2605 {
2606         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2607         return (dbi->db_blkptr);
2608 }
2609 
2610 static void
2611 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2612 {
2613         /* ASSERT(dmu_tx_is_syncing(tx) */
2614         ASSERT(MUTEX_HELD(&db->db_mtx));
2615 
2616         if (db->db_blkptr != NULL)
2617                 return;
2618 
2619         if (db->db_blkid == DMU_SPILL_BLKID) {
2620                 db->db_blkptr = &dn->dn_phys->dn_spill;
2621                 BP_ZERO(db->db_blkptr);
2622                 return;
2623         }
2624         if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2625                 /*
2626                  * This buffer was allocated at a time when there was
2627                  * no available blkptrs from the dnode, or it was
2628                  * inappropriate to hook it in (i.e., nlevels mis-match).
2629                  */
2630                 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2631                 ASSERT(db->db_parent == NULL);
2632                 db->db_parent = dn->dn_dbuf;
2633                 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2634                 DBUF_VERIFY(db);
2635         } else {
2636                 dmu_buf_impl_t *parent = db->db_parent;
2637                 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2638 
2639                 ASSERT(dn->dn_phys->dn_nlevels > 1);
2640                 if (parent == NULL) {
2641                         mutex_exit(&db->db_mtx);
2642                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
2643                         parent = dbuf_hold_level(dn, db->db_level + 1,
2644                             db->db_blkid >> epbs, db);
2645                         rw_exit(&dn->dn_struct_rwlock);
2646                         mutex_enter(&db->db_mtx);
2647                         db->db_parent = parent;
2648                 }
2649                 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2650                     (db->db_blkid & ((1ULL << epbs) - 1));
2651                 DBUF_VERIFY(db);
2652         }
2653 }
2654 
2655 static void
2656 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2657 {
2658         dmu_buf_impl_t *db = dr->dr_dbuf;
2659         dnode_t *dn;
2660         zio_t *zio;
2661 
2662         ASSERT(dmu_tx_is_syncing(tx));
2663 
2664         dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2665 
2666         mutex_enter(&db->db_mtx);
2667 
2668         ASSERT(db->db_level > 0);
2669         DBUF_VERIFY(db);
2670 
2671         /* Read the block if it hasn't been read yet. */
2672         if (db->db_buf == NULL) {
2673                 mutex_exit(&db->db_mtx);
2674                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2675                 mutex_enter(&db->db_mtx);
2676         }
2677         ASSERT3U(db->db_state, ==, DB_CACHED);
2678         ASSERT(db->db_buf != NULL);
2679 
2680         DB_DNODE_ENTER(db);
2681         dn = DB_DNODE(db);
2682         /* Indirect block size must match what the dnode thinks it is. */
2683         ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2684         dbuf_check_blkptr(dn, db);
2685         DB_DNODE_EXIT(db);
2686 
2687         /* Provide the pending dirty record to child dbufs */
2688         db->db_data_pending = dr;
2689 
2690         mutex_exit(&db->db_mtx);
2691         dbuf_write(dr, db->db_buf, tx);
2692 
2693         zio = dr->dr_zio;
2694         mutex_enter(&dr->dt.di.dr_mtx);
2695         dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
2696         ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2697         mutex_exit(&dr->dt.di.dr_mtx);
2698         zio_nowait(zio);
2699 }
2700 
2701 static void
2702 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2703 {
2704         arc_buf_t **datap = &dr->dt.dl.dr_data;
2705         dmu_buf_impl_t *db = dr->dr_dbuf;
2706         dnode_t *dn;
2707         objset_t *os;
2708         uint64_t txg = tx->tx_txg;
2709 
2710         ASSERT(dmu_tx_is_syncing(tx));
2711 
2712         dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2713 
2714         mutex_enter(&db->db_mtx);
2715         /*
2716          * To be synced, we must be dirtied.  But we
2717          * might have been freed after the dirty.
2718          */
2719         if (db->db_state == DB_UNCACHED) {
2720                 /* This buffer has been freed since it was dirtied */
2721                 ASSERT(db->db.db_data == NULL);
2722         } else if (db->db_state == DB_FILL) {
2723                 /* This buffer was freed and is now being re-filled */
2724                 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2725         } else {
2726                 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2727         }
2728         DBUF_VERIFY(db);
2729 
2730         DB_DNODE_ENTER(db);
2731         dn = DB_DNODE(db);
2732 
2733         if (db->db_blkid == DMU_SPILL_BLKID) {
2734                 mutex_enter(&dn->dn_mtx);
2735                 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2736                 mutex_exit(&dn->dn_mtx);
2737         }
2738 
2739         /*
2740          * If this is a bonus buffer, simply copy the bonus data into the
2741          * dnode.  It will be written out when the dnode is synced (and it
2742          * will be synced, since it must have been dirty for dbuf_sync to
2743          * be called).
2744          */
2745         if (db->db_blkid == DMU_BONUS_BLKID) {
2746                 dbuf_dirty_record_t **drp;
2747 
2748                 ASSERT(*datap != NULL);
2749                 ASSERT0(db->db_level);
2750                 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2751                 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2752                 DB_DNODE_EXIT(db);
2753 
2754                 if (*datap != db->db.db_data) {
2755                         zio_buf_free(*datap, DN_MAX_BONUSLEN);
2756                         arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2757                 }
2758                 db->db_data_pending = NULL;
2759                 drp = &db->db_last_dirty;
2760                 while (*drp != dr)
2761                         drp = &(*drp)->dr_next;
2762                 ASSERT(dr->dr_next == NULL);
2763                 ASSERT(dr->dr_dbuf == db);
2764                 *drp = dr->dr_next;
2765                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2766                 ASSERT(db->db_dirtycnt > 0);
2767                 db->db_dirtycnt -= 1;
2768                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2769                 return;
2770         }
2771 
2772         os = dn->dn_objset;
2773 
2774         /*
2775          * This function may have dropped the db_mtx lock allowing a dmu_sync
2776          * operation to sneak in. As a result, we need to ensure that we
2777          * don't check the dr_override_state until we have returned from
2778          * dbuf_check_blkptr.
2779          */
2780         dbuf_check_blkptr(dn, db);
2781 
2782         /*
2783          * If this buffer is in the middle of an immediate write,
2784          * wait for the synchronous IO to complete.
2785          */
2786         while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2787                 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2788                 cv_wait(&db->db_changed, &db->db_mtx);
2789                 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2790         }
2791 
2792         if (db->db_state != DB_NOFILL &&
2793             dn->dn_object != DMU_META_DNODE_OBJECT &&
2794             refcount_count(&db->db_holds) > 1 &&
2795             dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2796             *datap == db->db_buf) {
2797                 /*
2798                  * If this buffer is currently "in use" (i.e., there
2799                  * are active holds and db_data still references it),
2800                  * then make a copy before we start the write so that
2801                  * any modifications from the open txg will not leak
2802                  * into this write.
2803                  *
2804                  * NOTE: this copy does not need to be made for
2805                  * objects only modified in the syncing context (e.g.
2806                  * DNONE_DNODE blocks).
2807                  */
2808                 int blksz = arc_buf_size(*datap);
2809                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2810                 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2811                 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2812         }
2813         db->db_data_pending = dr;
2814 
2815         mutex_exit(&db->db_mtx);
2816 
2817         dbuf_write(dr, *datap, tx);
2818 
2819         ASSERT(!list_link_active(&dr->dr_dirty_node));
2820         if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2821                 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2822                 DB_DNODE_EXIT(db);
2823         } else {
2824                 /*
2825                  * Although zio_nowait() does not "wait for an IO", it does
2826                  * initiate the IO. If this is an empty write it seems plausible
2827                  * that the IO could actually be completed before the nowait
2828                  * returns. We need to DB_DNODE_EXIT() first in case
2829                  * zio_nowait() invalidates the dbuf.
2830                  */
2831                 DB_DNODE_EXIT(db);
2832                 zio_nowait(dr->dr_zio);
2833         }
2834 }
2835 
2836 void
2837 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
2838 {
2839         dbuf_dirty_record_t *dr;
2840 
2841         while (dr = list_head(list)) {
2842                 if (dr->dr_zio != NULL) {
2843                         /*
2844                          * If we find an already initialized zio then we
2845                          * are processing the meta-dnode, and we have finished.
2846                          * The dbufs for all dnodes are put back on the list
2847                          * during processing, so that we can zio_wait()
2848                          * these IOs after initiating all child IOs.
2849                          */
2850                         ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2851                             DMU_META_DNODE_OBJECT);
2852                         break;
2853                 }
2854                 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
2855                     dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
2856                         VERIFY3U(dr->dr_dbuf->db_level, ==, level);
2857                 }
2858                 list_remove(list, dr);
2859                 if (dr->dr_dbuf->db_level > 0)
2860                         dbuf_sync_indirect(dr, tx);
2861                 else
2862                         dbuf_sync_leaf(dr, tx);
2863         }
2864 }
2865 
2866 /* ARGSUSED */
2867 static void
2868 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2869 {
2870         dmu_buf_impl_t *db = vdb;
2871         dnode_t *dn;
2872         blkptr_t *bp = zio->io_bp;
2873         blkptr_t *bp_orig = &zio->io_bp_orig;
2874         spa_t *spa = zio->io_spa;
2875         int64_t delta;
2876         uint64_t fill = 0;
2877         int i;
2878 
2879         ASSERT3P(db->db_blkptr, ==, bp);
2880 
2881         DB_DNODE_ENTER(db);
2882         dn = DB_DNODE(db);
2883         delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2884         dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2885         zio->io_prev_space_delta = delta;
2886 
2887         if (bp->blk_birth != 0) {
2888                 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2889                     BP_GET_TYPE(bp) == dn->dn_type) ||
2890                     (db->db_blkid == DMU_SPILL_BLKID &&
2891                     BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2892                     BP_IS_EMBEDDED(bp));
2893                 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2894         }
2895 
2896         mutex_enter(&db->db_mtx);
2897 
2898 #ifdef ZFS_DEBUG
2899         if (db->db_blkid == DMU_SPILL_BLKID) {
2900                 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2901                 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2902                     db->db_blkptr == &dn->dn_phys->dn_spill);
2903         }
2904 #endif
2905 
2906         if (db->db_level == 0) {
2907                 mutex_enter(&dn->dn_mtx);
2908                 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2909                     db->db_blkid != DMU_SPILL_BLKID)
2910                         dn->dn_phys->dn_maxblkid = db->db_blkid;
2911                 mutex_exit(&dn->dn_mtx);
2912 
2913                 if (dn->dn_type == DMU_OT_DNODE) {
2914                         dnode_phys_t *dnp = db->db.db_data;
2915                         for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2916                             i--, dnp++) {
2917                                 if (dnp->dn_type != DMU_OT_NONE)
2918                                         fill++;
2919                         }
2920                 } else {
2921                         if (BP_IS_HOLE(bp)) {
2922                                 fill = 0;
2923                         } else {
2924                                 fill = 1;
2925                         }
2926                 }
2927         } else {
2928                 blkptr_t *ibp = db->db.db_data;
2929                 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2930                 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2931                         if (BP_IS_HOLE(ibp))
2932                                 continue;
2933                         fill += BP_GET_FILL(ibp);
2934                 }
2935         }
2936         DB_DNODE_EXIT(db);
2937 
2938         if (!BP_IS_EMBEDDED(bp))
2939                 bp->blk_fill = fill;
2940 
2941         mutex_exit(&db->db_mtx);
2942 }
2943 
2944 /*
2945  * The SPA will call this callback several times for each zio - once
2946  * for every physical child i/o (zio->io_phys_children times).  This
2947  * allows the DMU to monitor the progress of each logical i/o.  For example,
2948  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2949  * block.  There may be a long delay before all copies/fragments are completed,
2950  * so this callback allows us to retire dirty space gradually, as the physical
2951  * i/os complete.
2952  */
2953 /* ARGSUSED */
2954 static void
2955 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2956 {
2957         dmu_buf_impl_t *db = arg;
2958         objset_t *os = db->db_objset;
2959         dsl_pool_t *dp = dmu_objset_pool(os);
2960         dbuf_dirty_record_t *dr;
2961         int delta = 0;
2962 
2963         dr = db->db_data_pending;
2964         ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2965 
2966         /*
2967          * The callback will be called io_phys_children times.  Retire one
2968          * portion of our dirty space each time we are called.  Any rounding
2969          * error will be cleaned up by dsl_pool_sync()'s call to
2970          * dsl_pool_undirty_space().
2971          */
2972         delta = dr->dr_accounted / zio->io_phys_children;
2973         dsl_pool_undirty_space(dp, delta, zio->io_txg);
2974 }
2975 
2976 /* ARGSUSED */
2977 static void
2978 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2979 {
2980         dmu_buf_impl_t *db = vdb;
2981         blkptr_t *bp_orig = &zio->io_bp_orig;
2982         blkptr_t *bp = db->db_blkptr;
2983         objset_t *os = db->db_objset;
2984         dmu_tx_t *tx = os->os_synctx;
2985         dbuf_dirty_record_t **drp, *dr;
2986 
2987         ASSERT0(zio->io_error);
2988         ASSERT(db->db_blkptr == bp);
2989 
2990         /*
2991          * For nopwrites and rewrites we ensure that the bp matches our
2992          * original and bypass all the accounting.
2993          */
2994         if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2995                 ASSERT(BP_EQUAL(bp, bp_orig));
2996         } else {
2997                 dsl_dataset_t *ds = os->os_dsl_dataset;
2998                 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2999                 dsl_dataset_block_born(ds, bp, tx);
3000         }
3001 
3002         mutex_enter(&db->db_mtx);
3003 
3004         DBUF_VERIFY(db);
3005 
3006         drp = &db->db_last_dirty;
3007         while ((dr = *drp) != db->db_data_pending)
3008                 drp = &dr->dr_next;
3009         ASSERT(!list_link_active(&dr->dr_dirty_node));
3010         ASSERT(dr->dr_dbuf == db);
3011         ASSERT(dr->dr_next == NULL);
3012         *drp = dr->dr_next;
3013 
3014 #ifdef ZFS_DEBUG
3015         if (db->db_blkid == DMU_SPILL_BLKID) {
3016                 dnode_t *dn;
3017 
3018                 DB_DNODE_ENTER(db);
3019                 dn = DB_DNODE(db);
3020                 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3021                 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3022                     db->db_blkptr == &dn->dn_phys->dn_spill);
3023                 DB_DNODE_EXIT(db);
3024         }
3025 #endif
3026 
3027         if (db->db_level == 0) {
3028                 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3029                 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3030                 if (db->db_state != DB_NOFILL) {
3031                         if (dr->dt.dl.dr_data != db->db_buf)
3032                                 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
3033                                     db));
3034                         else if (!arc_released(db->db_buf))
3035                                 arc_set_callback(db->db_buf, dbuf_do_evict, db);
3036                 }
3037         } else {
3038                 dnode_t *dn;
3039 
3040                 DB_DNODE_ENTER(db);
3041                 dn = DB_DNODE(db);
3042                 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3043                 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3044                 if (!BP_IS_HOLE(db->db_blkptr)) {
3045                         int epbs =
3046                             dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3047                         ASSERT3U(db->db_blkid, <=,
3048                             dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3049                         ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3050                             db->db.db_size);
3051                         if (!arc_released(db->db_buf))
3052                                 arc_set_callback(db->db_buf, dbuf_do_evict, db);
3053                 }
3054                 DB_DNODE_EXIT(db);
3055                 mutex_destroy(&dr->dt.di.dr_mtx);
3056                 list_destroy(&dr->dt.di.dr_children);
3057         }
3058         kmem_free(dr, sizeof (dbuf_dirty_record_t));
3059 
3060         cv_broadcast(&db->db_changed);
3061         ASSERT(db->db_dirtycnt > 0);
3062         db->db_dirtycnt -= 1;
3063         db->db_data_pending = NULL;
3064         dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3065 }
3066 
3067 static void
3068 dbuf_write_nofill_ready(zio_t *zio)
3069 {
3070         dbuf_write_ready(zio, NULL, zio->io_private);
3071 }
3072 
3073 static void
3074 dbuf_write_nofill_done(zio_t *zio)
3075 {
3076         dbuf_write_done(zio, NULL, zio->io_private);
3077 }
3078 
3079 static void
3080 dbuf_write_override_ready(zio_t *zio)
3081 {
3082         dbuf_dirty_record_t *dr = zio->io_private;
3083         dmu_buf_impl_t *db = dr->dr_dbuf;
3084 
3085         dbuf_write_ready(zio, NULL, db);
3086 }
3087 
3088 static void
3089 dbuf_write_override_done(zio_t *zio)
3090 {
3091         dbuf_dirty_record_t *dr = zio->io_private;
3092         dmu_buf_impl_t *db = dr->dr_dbuf;
3093         blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3094 
3095         mutex_enter(&db->db_mtx);
3096         if (!BP_EQUAL(zio->io_bp, obp)) {
3097                 if (!BP_IS_HOLE(obp))
3098                         dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3099                 arc_release(dr->dt.dl.dr_data, db);
3100         }
3101         mutex_exit(&db->db_mtx);
3102 
3103         dbuf_write_done(zio, NULL, db);
3104 }
3105 
3106 /* Issue I/O to commit a dirty buffer to disk. */
3107 static void
3108 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3109 {
3110         dmu_buf_impl_t *db = dr->dr_dbuf;
3111         dnode_t *dn;
3112         objset_t *os;
3113         dmu_buf_impl_t *parent = db->db_parent;
3114         uint64_t txg = tx->tx_txg;
3115         zbookmark_phys_t zb;
3116         zio_prop_t zp;
3117         zio_t *zio;
3118         int wp_flag = 0;
3119 
3120         DB_DNODE_ENTER(db);
3121         dn = DB_DNODE(db);
3122         os = dn->dn_objset;
3123 
3124         if (db->db_state != DB_NOFILL) {
3125                 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3126                         /*
3127                          * Private object buffers are released here rather
3128                          * than in dbuf_dirty() since they are only modified
3129                          * in the syncing context and we don't want the
3130                          * overhead of making multiple copies of the data.
3131                          */
3132                         if (BP_IS_HOLE(db->db_blkptr)) {
3133                                 arc_buf_thaw(data);
3134                         } else {
3135                                 dbuf_release_bp(db);
3136                         }
3137                 }
3138         }
3139 
3140         if (parent != dn->dn_dbuf) {
3141                 /* Our parent is an indirect block. */
3142                 /* We have a dirty parent that has been scheduled for write. */
3143                 ASSERT(parent && parent->db_data_pending);
3144                 /* Our parent's buffer is one level closer to the dnode. */
3145                 ASSERT(db->db_level == parent->db_level-1);
3146                 /*
3147                  * We're about to modify our parent's db_data by modifying
3148                  * our block pointer, so the parent must be released.
3149                  */
3150                 ASSERT(arc_released(parent->db_buf));
3151                 zio = parent->db_data_pending->dr_zio;
3152         } else {
3153                 /* Our parent is the dnode itself. */
3154                 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3155                     db->db_blkid != DMU_SPILL_BLKID) ||
3156                     (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3157                 if (db->db_blkid != DMU_SPILL_BLKID)
3158                         ASSERT3P(db->db_blkptr, ==,
3159                             &dn->dn_phys->dn_blkptr[db->db_blkid]);
3160                 zio = dn->dn_zio;
3161         }
3162 
3163         ASSERT(db->db_level == 0 || data == db->db_buf);
3164         ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3165         ASSERT(zio);
3166 
3167         SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3168             os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3169             db->db.db_object, db->db_level, db->db_blkid);
3170 
3171         if (db->db_blkid == DMU_SPILL_BLKID)
3172                 wp_flag = WP_SPILL;
3173         wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3174 
3175         dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3176         DB_DNODE_EXIT(db);
3177 
3178         if (db->db_level == 0 &&
3179             dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3180                 /*
3181                  * The BP for this block has been provided by open context
3182                  * (by dmu_sync() or dmu_buf_write_embedded()).
3183                  */
3184                 void *contents = (data != NULL) ? data->b_data : NULL;
3185 
3186                 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3187                     db->db_blkptr, contents, db->db.db_size, &zp,
3188                     dbuf_write_override_ready, NULL, dbuf_write_override_done,
3189                     dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3190                 mutex_enter(&db->db_mtx);
3191                 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3192                 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3193                     dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3194                 mutex_exit(&db->db_mtx);
3195         } else if (db->db_state == DB_NOFILL) {
3196                 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3197                     zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3198                 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3199                     db->db_blkptr, NULL, db->db.db_size, &zp,
3200                     dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
3201                     ZIO_PRIORITY_ASYNC_WRITE,
3202                     ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3203         } else {
3204                 ASSERT(arc_released(data));
3205                 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3206                     db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
3207                     DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
3208                     dbuf_write_physdone, dbuf_write_done, db,
3209                     ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3210         }
3211 }