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, 2014 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 uint64_t
 552 dbuf_whichblock(dnode_t *dn, uint64_t offset)
 553 {
 554         if (dn->dn_datablkshift) {
 555                 return (offset >> dn->dn_datablkshift);
 556         } else {
 557                 ASSERT3U(offset, <, dn->dn_datablksz);
 558                 return (0);
 559         }
 560 }
 561 
 562 static void
 563 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
 564 {
 565         dmu_buf_impl_t *db = vdb;
 566 
 567         mutex_enter(&db->db_mtx);
 568         ASSERT3U(db->db_state, ==, DB_READ);
 569         /*
 570          * All reads are synchronous, so we must have a hold on the dbuf
 571          */
 572         ASSERT(refcount_count(&db->db_holds) > 0);
 573         ASSERT(db->db_buf == NULL);
 574         ASSERT(db->db.db_data == NULL);
 575         if (db->db_level == 0 && db->db_freed_in_flight) {
 576                 /* we were freed in flight; disregard any error */
 577                 arc_release(buf, db);
 578                 bzero(buf->b_data, db->db.db_size);
 579                 arc_buf_freeze(buf);
 580                 db->db_freed_in_flight = FALSE;
 581                 dbuf_set_data(db, buf);
 582                 db->db_state = DB_CACHED;
 583         } else if (zio == NULL || zio->io_error == 0) {
 584                 dbuf_set_data(db, buf);
 585                 db->db_state = DB_CACHED;
 586         } else {
 587                 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
 588                 ASSERT3P(db->db_buf, ==, NULL);
 589                 VERIFY(arc_buf_remove_ref(buf, db));
 590                 db->db_state = DB_UNCACHED;
 591         }
 592         cv_broadcast(&db->db_changed);
 593         dbuf_rele_and_unlock(db, NULL);
 594 }
 595 
 596 static void
 597 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
 598 {
 599         dnode_t *dn;
 600         zbookmark_phys_t zb;
 601         arc_flags_t aflags = ARC_FLAG_NOWAIT;
 602 
 603         DB_DNODE_ENTER(db);
 604         dn = DB_DNODE(db);
 605         ASSERT(!refcount_is_zero(&db->db_holds));
 606         /* We need the struct_rwlock to prevent db_blkptr from changing. */
 607         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
 608         ASSERT(MUTEX_HELD(&db->db_mtx));
 609         ASSERT(db->db_state == DB_UNCACHED);
 610         ASSERT(db->db_buf == NULL);
 611 
 612         if (db->db_blkid == DMU_BONUS_BLKID) {
 613                 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
 614 
 615                 ASSERT3U(bonuslen, <=, db->db.db_size);
 616                 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
 617                 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
 618                 if (bonuslen < DN_MAX_BONUSLEN)
 619                         bzero(db->db.db_data, DN_MAX_BONUSLEN);
 620                 if (bonuslen)
 621                         bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
 622                 DB_DNODE_EXIT(db);
 623                 db->db_state = DB_CACHED;
 624                 mutex_exit(&db->db_mtx);
 625                 return;
 626         }
 627 
 628         /*
 629          * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
 630          * processes the delete record and clears the bp while we are waiting
 631          * for the dn_mtx (resulting in a "no" from block_freed).
 632          */
 633         if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
 634             (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
 635             BP_IS_HOLE(db->db_blkptr)))) {
 636                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
 637 
 638                 DB_DNODE_EXIT(db);
 639                 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
 640                     db->db.db_size, db, type));
 641                 bzero(db->db.db_data, db->db.db_size);
 642                 db->db_state = DB_CACHED;
 643                 *flags |= DB_RF_CACHED;
 644                 mutex_exit(&db->db_mtx);
 645                 return;
 646         }
 647 
 648         DB_DNODE_EXIT(db);
 649 
 650         db->db_state = DB_READ;
 651         mutex_exit(&db->db_mtx);
 652 
 653         if (DBUF_IS_L2CACHEABLE(db))
 654                 aflags |= ARC_FLAG_L2CACHE;
 655         if (DBUF_IS_L2COMPRESSIBLE(db))
 656                 aflags |= ARC_FLAG_L2COMPRESS;
 657 
 658         SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
 659             db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
 660             db->db.db_object, db->db_level, db->db_blkid);
 661 
 662         dbuf_add_ref(db, NULL);
 663 
 664         (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
 665             dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
 666             (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
 667             &aflags, &zb);
 668         if (aflags & ARC_FLAG_CACHED)
 669                 *flags |= DB_RF_CACHED;
 670 }
 671 
 672 int
 673 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 674 {
 675         int err = 0;
 676         boolean_t havepzio = (zio != NULL);
 677         boolean_t prefetch;
 678         dnode_t *dn;
 679 
 680         /*
 681          * We don't have to hold the mutex to check db_state because it
 682          * can't be freed while we have a hold on the buffer.
 683          */
 684         ASSERT(!refcount_is_zero(&db->db_holds));
 685 
 686         if (db->db_state == DB_NOFILL)
 687                 return (SET_ERROR(EIO));
 688 
 689         DB_DNODE_ENTER(db);
 690         dn = DB_DNODE(db);
 691         if ((flags & DB_RF_HAVESTRUCT) == 0)
 692                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
 693 
 694         prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
 695             (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
 696             DBUF_IS_CACHEABLE(db);
 697 
 698         mutex_enter(&db->db_mtx);
 699         if (db->db_state == DB_CACHED) {
 700                 mutex_exit(&db->db_mtx);
 701                 if (prefetch)
 702                         dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
 703                             db->db.db_size, TRUE);
 704                 if ((flags & DB_RF_HAVESTRUCT) == 0)
 705                         rw_exit(&dn->dn_struct_rwlock);
 706                 DB_DNODE_EXIT(db);
 707         } else if (db->db_state == DB_UNCACHED) {
 708                 spa_t *spa = dn->dn_objset->os_spa;
 709 
 710                 if (zio == NULL)
 711                         zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
 712                 dbuf_read_impl(db, zio, &flags);
 713 
 714                 /* dbuf_read_impl has dropped db_mtx for us */
 715 
 716                 if (prefetch)
 717                         dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
 718                             db->db.db_size, flags & DB_RF_CACHED);
 719 
 720                 if ((flags & DB_RF_HAVESTRUCT) == 0)
 721                         rw_exit(&dn->dn_struct_rwlock);
 722                 DB_DNODE_EXIT(db);
 723 
 724                 if (!havepzio)
 725                         err = zio_wait(zio);
 726         } else {
 727                 /*
 728                  * Another reader came in while the dbuf was in flight
 729                  * between UNCACHED and CACHED.  Either a writer will finish
 730                  * writing the buffer (sending the dbuf to CACHED) or the
 731                  * first reader's request will reach the read_done callback
 732                  * and send the dbuf to CACHED.  Otherwise, a failure
 733                  * occurred and the dbuf went to UNCACHED.
 734                  */
 735                 mutex_exit(&db->db_mtx);
 736                 if (prefetch)
 737                         dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
 738                             db->db.db_size, TRUE);
 739                 if ((flags & DB_RF_HAVESTRUCT) == 0)
 740                         rw_exit(&dn->dn_struct_rwlock);
 741                 DB_DNODE_EXIT(db);
 742 
 743                 /* Skip the wait per the caller's request. */
 744                 mutex_enter(&db->db_mtx);
 745                 if ((flags & DB_RF_NEVERWAIT) == 0) {
 746                         while (db->db_state == DB_READ ||
 747                             db->db_state == DB_FILL) {
 748                                 ASSERT(db->db_state == DB_READ ||
 749                                     (flags & DB_RF_HAVESTRUCT) == 0);
 750                                 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
 751                                     db, zio_t *, zio);
 752                                 cv_wait(&db->db_changed, &db->db_mtx);
 753                         }
 754                         if (db->db_state == DB_UNCACHED)
 755                                 err = SET_ERROR(EIO);
 756                 }
 757                 mutex_exit(&db->db_mtx);
 758         }
 759 
 760         ASSERT(err || havepzio || db->db_state == DB_CACHED);
 761         return (err);
 762 }
 763 
 764 static void
 765 dbuf_noread(dmu_buf_impl_t *db)
 766 {
 767         ASSERT(!refcount_is_zero(&db->db_holds));
 768         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
 769         mutex_enter(&db->db_mtx);
 770         while (db->db_state == DB_READ || db->db_state == DB_FILL)
 771                 cv_wait(&db->db_changed, &db->db_mtx);
 772         if (db->db_state == DB_UNCACHED) {
 773                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
 774                 spa_t *spa = db->db_objset->os_spa;
 775 
 776                 ASSERT(db->db_buf == NULL);
 777                 ASSERT(db->db.db_data == NULL);
 778                 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
 779                 db->db_state = DB_FILL;
 780         } else if (db->db_state == DB_NOFILL) {
 781                 dbuf_clear_data(db);
 782         } else {
 783                 ASSERT3U(db->db_state, ==, DB_CACHED);
 784         }
 785         mutex_exit(&db->db_mtx);
 786 }
 787 
 788 /*
 789  * This is our just-in-time copy function.  It makes a copy of
 790  * buffers, that have been modified in a previous transaction
 791  * group, before we modify them in the current active group.
 792  *
 793  * This function is used in two places: when we are dirtying a
 794  * buffer for the first time in a txg, and when we are freeing
 795  * a range in a dnode that includes this buffer.
 796  *
 797  * Note that when we are called from dbuf_free_range() we do
 798  * not put a hold on the buffer, we just traverse the active
 799  * dbuf list for the dnode.
 800  */
 801 static void
 802 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
 803 {
 804         dbuf_dirty_record_t *dr = db->db_last_dirty;
 805 
 806         ASSERT(MUTEX_HELD(&db->db_mtx));
 807         ASSERT(db->db.db_data != NULL);
 808         ASSERT(db->db_level == 0);
 809         ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
 810 
 811         if (dr == NULL ||
 812             (dr->dt.dl.dr_data !=
 813             ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
 814                 return;
 815 
 816         /*
 817          * If the last dirty record for this dbuf has not yet synced
 818          * and its referencing the dbuf data, either:
 819          *      reset the reference to point to a new copy,
 820          * or (if there a no active holders)
 821          *      just null out the current db_data pointer.
 822          */
 823         ASSERT(dr->dr_txg >= txg - 2);
 824         if (db->db_blkid == DMU_BONUS_BLKID) {
 825                 /* Note that the data bufs here are zio_bufs */
 826                 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
 827                 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
 828                 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
 829         } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
 830                 int size = db->db.db_size;
 831                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
 832                 spa_t *spa = db->db_objset->os_spa;
 833 
 834                 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
 835                 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
 836         } else {
 837                 dbuf_clear_data(db);
 838         }
 839 }
 840 
 841 void
 842 dbuf_unoverride(dbuf_dirty_record_t *dr)
 843 {
 844         dmu_buf_impl_t *db = dr->dr_dbuf;
 845         blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
 846         uint64_t txg = dr->dr_txg;
 847 
 848         ASSERT(MUTEX_HELD(&db->db_mtx));
 849         ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
 850         ASSERT(db->db_level == 0);
 851 
 852         if (db->db_blkid == DMU_BONUS_BLKID ||
 853             dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
 854                 return;
 855 
 856         ASSERT(db->db_data_pending != dr);
 857 
 858         /* free this block */
 859         if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
 860                 zio_free(db->db_objset->os_spa, txg, bp);
 861 
 862         dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
 863         dr->dt.dl.dr_nopwrite = B_FALSE;
 864 
 865         /*
 866          * Release the already-written buffer, so we leave it in
 867          * a consistent dirty state.  Note that all callers are
 868          * modifying the buffer, so they will immediately do
 869          * another (redundant) arc_release().  Therefore, leave
 870          * the buf thawed to save the effort of freezing &
 871          * immediately re-thawing it.
 872          */
 873         arc_release(dr->dt.dl.dr_data, db);
 874 }
 875 
 876 /*
 877  * Evict (if its unreferenced) or clear (if its referenced) any level-0
 878  * data blocks in the free range, so that any future readers will find
 879  * empty blocks.
 880  *
 881  * This is a no-op if the dataset is in the middle of an incremental
 882  * receive; see comment below for details.
 883  */
 884 void
 885 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
 886     dmu_tx_t *tx)
 887 {
 888         dmu_buf_impl_t db_search;
 889         dmu_buf_impl_t *db, *db_next;
 890         uint64_t txg = tx->tx_txg;
 891         avl_index_t where;
 892 
 893         if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
 894                 end_blkid = dn->dn_maxblkid;
 895         dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
 896 
 897         db_search.db_level = 0;
 898         db_search.db_blkid = start_blkid;
 899         db_search.db_state = DB_SEARCH;
 900 
 901         mutex_enter(&dn->dn_dbufs_mtx);
 902         if (start_blkid >= dn->dn_unlisted_l0_blkid) {
 903                 /* There can't be any dbufs in this range; no need to search. */
 904 #ifdef DEBUG
 905                 db = avl_find(&dn->dn_dbufs, &db_search, &where);
 906                 ASSERT3P(db, ==, NULL);
 907                 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
 908                 ASSERT(db == NULL || db->db_level > 0);
 909 #endif
 910                 mutex_exit(&dn->dn_dbufs_mtx);
 911                 return;
 912         } else if (dmu_objset_is_receiving(dn->dn_objset)) {
 913                 /*
 914                  * If we are receiving, we expect there to be no dbufs in
 915                  * the range to be freed, because receive modifies each
 916                  * block at most once, and in offset order.  If this is
 917                  * not the case, it can lead to performance problems,
 918                  * so note that we unexpectedly took the slow path.
 919                  */
 920                 atomic_inc_64(&zfs_free_range_recv_miss);
 921         }
 922 
 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 
 927         for (; db != NULL; db = db_next) {
 928                 db_next = AVL_NEXT(&dn->dn_dbufs, db);
 929                 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
 930 
 931                 if (db->db_level != 0 || db->db_blkid > end_blkid) {
 932                         break;
 933                 }
 934                 ASSERT3U(db->db_blkid, >=, start_blkid);
 935 
 936                 /* found a level 0 buffer in the range */
 937                 mutex_enter(&db->db_mtx);
 938                 if (dbuf_undirty(db, tx)) {
 939                         /* mutex has been dropped and dbuf destroyed */
 940                         continue;
 941                 }
 942 
 943                 if (db->db_state == DB_UNCACHED ||
 944                     db->db_state == DB_NOFILL ||
 945                     db->db_state == DB_EVICTING) {
 946                         ASSERT(db->db.db_data == NULL);
 947                         mutex_exit(&db->db_mtx);
 948                         continue;
 949                 }
 950                 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
 951                         /* will be handled in dbuf_read_done or dbuf_rele */
 952                         db->db_freed_in_flight = TRUE;
 953                         mutex_exit(&db->db_mtx);
 954                         continue;
 955                 }
 956                 if (refcount_count(&db->db_holds) == 0) {
 957                         ASSERT(db->db_buf);
 958                         dbuf_clear(db);
 959                         continue;
 960                 }
 961                 /* The dbuf is referenced */
 962 
 963                 if (db->db_last_dirty != NULL) {
 964                         dbuf_dirty_record_t *dr = db->db_last_dirty;
 965 
 966                         if (dr->dr_txg == txg) {
 967                                 /*
 968                                  * This buffer is "in-use", re-adjust the file
 969                                  * size to reflect that this buffer may
 970                                  * contain new data when we sync.
 971                                  */
 972                                 if (db->db_blkid != DMU_SPILL_BLKID &&
 973                                     db->db_blkid > dn->dn_maxblkid)
 974                                         dn->dn_maxblkid = db->db_blkid;
 975                                 dbuf_unoverride(dr);
 976                         } else {
 977                                 /*
 978                                  * This dbuf is not dirty in the open context.
 979                                  * Either uncache it (if its not referenced in
 980                                  * the open context) or reset its contents to
 981                                  * empty.
 982                                  */
 983                                 dbuf_fix_old_data(db, txg);
 984                         }
 985                 }
 986                 /* clear the contents if its cached */
 987                 if (db->db_state == DB_CACHED) {
 988                         ASSERT(db->db.db_data != NULL);
 989                         arc_release(db->db_buf, db);
 990                         bzero(db->db.db_data, db->db.db_size);
 991                         arc_buf_freeze(db->db_buf);
 992                 }
 993 
 994                 mutex_exit(&db->db_mtx);
 995         }
 996         mutex_exit(&dn->dn_dbufs_mtx);
 997 }
 998 
 999 static int
1000 dbuf_block_freeable(dmu_buf_impl_t *db)
1001 {
1002         dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1003         uint64_t birth_txg = 0;
1004 
1005         /*
1006          * We don't need any locking to protect db_blkptr:
1007          * If it's syncing, then db_last_dirty will be set
1008          * so we'll ignore db_blkptr.
1009          *
1010          * This logic ensures that only block births for
1011          * filled blocks are considered.
1012          */
1013         ASSERT(MUTEX_HELD(&db->db_mtx));
1014         if (db->db_last_dirty && (db->db_blkptr == NULL ||
1015             !BP_IS_HOLE(db->db_blkptr))) {
1016                 birth_txg = db->db_last_dirty->dr_txg;
1017         } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1018                 birth_txg = db->db_blkptr->blk_birth;
1019         }
1020 
1021         /*
1022          * If this block don't exist or is in a snapshot, it can't be freed.
1023          * Don't pass the bp to dsl_dataset_block_freeable() since we
1024          * are holding the db_mtx lock and might deadlock if we are
1025          * prefetching a dedup-ed block.
1026          */
1027         if (birth_txg != 0)
1028                 return (ds == NULL ||
1029                     dsl_dataset_block_freeable(ds, NULL, birth_txg));
1030         else
1031                 return (B_FALSE);
1032 }
1033 
1034 void
1035 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1036 {
1037         arc_buf_t *buf, *obuf;
1038         int osize = db->db.db_size;
1039         arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1040         dnode_t *dn;
1041 
1042         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1043 
1044         DB_DNODE_ENTER(db);
1045         dn = DB_DNODE(db);
1046 
1047         /* XXX does *this* func really need the lock? */
1048         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1049 
1050         /*
1051          * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1052          * is OK, because there can be no other references to the db
1053          * when we are changing its size, so no concurrent DB_FILL can
1054          * be happening.
1055          */
1056         /*
1057          * XXX we should be doing a dbuf_read, checking the return
1058          * value and returning that up to our callers
1059          */
1060         dmu_buf_will_dirty(&db->db, tx);
1061 
1062         /* create the data buffer for the new block */
1063         buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1064 
1065         /* copy old block data to the new block */
1066         obuf = db->db_buf;
1067         bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1068         /* zero the remainder */
1069         if (size > osize)
1070                 bzero((uint8_t *)buf->b_data + osize, size - osize);
1071 
1072         mutex_enter(&db->db_mtx);
1073         dbuf_set_data(db, buf);
1074         VERIFY(arc_buf_remove_ref(obuf, db));
1075         db->db.db_size = size;
1076 
1077         if (db->db_level == 0) {
1078                 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1079                 db->db_last_dirty->dt.dl.dr_data = buf;
1080         }
1081         mutex_exit(&db->db_mtx);
1082 
1083         dnode_willuse_space(dn, size-osize, tx);
1084         DB_DNODE_EXIT(db);
1085 }
1086 
1087 void
1088 dbuf_release_bp(dmu_buf_impl_t *db)
1089 {
1090         objset_t *os = db->db_objset;
1091 
1092         ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1093         ASSERT(arc_released(os->os_phys_buf) ||
1094             list_link_active(&os->os_dsl_dataset->ds_synced_link));
1095         ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1096 
1097         (void) arc_release(db->db_buf, db);
1098 }
1099 
1100 /*
1101  * We already have a dirty record for this TXG, and we are being
1102  * dirtied again.
1103  */
1104 static void
1105 dbuf_redirty(dbuf_dirty_record_t *dr)
1106 {
1107         dmu_buf_impl_t *db = dr->dr_dbuf;
1108 
1109         ASSERT(MUTEX_HELD(&db->db_mtx));
1110 
1111         if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1112                 /*
1113                  * If this buffer has already been written out,
1114                  * we now need to reset its state.
1115                  */
1116                 dbuf_unoverride(dr);
1117                 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1118                     db->db_state != DB_NOFILL) {
1119                         /* Already released on initial dirty, so just thaw. */
1120                         ASSERT(arc_released(db->db_buf));
1121                         arc_buf_thaw(db->db_buf);
1122                 }
1123         }
1124 }
1125 
1126 dbuf_dirty_record_t *
1127 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1128 {
1129         dnode_t *dn;
1130         objset_t *os;
1131         dbuf_dirty_record_t **drp, *dr;
1132         int drop_struct_lock = FALSE;
1133         boolean_t do_free_accounting = B_FALSE;
1134         int txgoff = tx->tx_txg & TXG_MASK;
1135 
1136         ASSERT(tx->tx_txg != 0);
1137         ASSERT(!refcount_is_zero(&db->db_holds));
1138         DMU_TX_DIRTY_BUF(tx, db);
1139 
1140         DB_DNODE_ENTER(db);
1141         dn = DB_DNODE(db);
1142         /*
1143          * Shouldn't dirty a regular buffer in syncing context.  Private
1144          * objects may be dirtied in syncing context, but only if they
1145          * were already pre-dirtied in open context.
1146          */
1147         ASSERT(!dmu_tx_is_syncing(tx) ||
1148             BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1149             DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1150             dn->dn_objset->os_dsl_dataset == NULL);
1151         /*
1152          * We make this assert for private objects as well, but after we
1153          * check if we're already dirty.  They are allowed to re-dirty
1154          * in syncing context.
1155          */
1156         ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1157             dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1158             (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1159 
1160         mutex_enter(&db->db_mtx);
1161         /*
1162          * XXX make this true for indirects too?  The problem is that
1163          * transactions created with dmu_tx_create_assigned() from
1164          * syncing context don't bother holding ahead.
1165          */
1166         ASSERT(db->db_level != 0 ||
1167             db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1168             db->db_state == DB_NOFILL);
1169 
1170         mutex_enter(&dn->dn_mtx);
1171         /*
1172          * Don't set dirtyctx to SYNC if we're just modifying this as we
1173          * initialize the objset.
1174          */
1175         if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1176             !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1177                 dn->dn_dirtyctx =
1178                     (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1179                 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1180                 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1181         }
1182         mutex_exit(&dn->dn_mtx);
1183 
1184         if (db->db_blkid == DMU_SPILL_BLKID)
1185                 dn->dn_have_spill = B_TRUE;
1186 
1187         /*
1188          * If this buffer is already dirty, we're done.
1189          */
1190         drp = &db->db_last_dirty;
1191         ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1192             db->db.db_object == DMU_META_DNODE_OBJECT);
1193         while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1194                 drp = &dr->dr_next;
1195         if (dr && dr->dr_txg == tx->tx_txg) {
1196                 DB_DNODE_EXIT(db);
1197 
1198                 dbuf_redirty(dr);
1199                 mutex_exit(&db->db_mtx);
1200                 return (dr);
1201         }
1202 
1203         /*
1204          * Only valid if not already dirty.
1205          */
1206         ASSERT(dn->dn_object == 0 ||
1207             dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1208             (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1209 
1210         ASSERT3U(dn->dn_nlevels, >, db->db_level);
1211         ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1212             dn->dn_phys->dn_nlevels > db->db_level ||
1213             dn->dn_next_nlevels[txgoff] > db->db_level ||
1214             dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1215             dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1216 
1217         /*
1218          * We should only be dirtying in syncing context if it's the
1219          * mos or we're initializing the os or it's a special object.
1220          * However, we are allowed to dirty in syncing context provided
1221          * we already dirtied it in open context.  Hence we must make
1222          * this assertion only if we're not already dirty.
1223          */
1224         os = dn->dn_objset;
1225         ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1226             os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1227         ASSERT(db->db.db_size != 0);
1228 
1229         dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1230 
1231         if (db->db_blkid != DMU_BONUS_BLKID) {
1232                 /*
1233                  * Update the accounting.
1234                  * Note: we delay "free accounting" until after we drop
1235                  * the db_mtx.  This keeps us from grabbing other locks
1236                  * (and possibly deadlocking) in bp_get_dsize() while
1237                  * also holding the db_mtx.
1238                  */
1239                 dnode_willuse_space(dn, db->db.db_size, tx);
1240                 do_free_accounting = dbuf_block_freeable(db);
1241         }
1242 
1243         /*
1244          * If this buffer is dirty in an old transaction group we need
1245          * to make a copy of it so that the changes we make in this
1246          * transaction group won't leak out when we sync the older txg.
1247          */
1248         dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1249         if (db->db_level == 0) {
1250                 void *data_old = db->db_buf;
1251 
1252                 if (db->db_state != DB_NOFILL) {
1253                         if (db->db_blkid == DMU_BONUS_BLKID) {
1254                                 dbuf_fix_old_data(db, tx->tx_txg);
1255                                 data_old = db->db.db_data;
1256                         } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1257                                 /*
1258                                  * Release the data buffer from the cache so
1259                                  * that we can modify it without impacting
1260                                  * possible other users of this cached data
1261                                  * block.  Note that indirect blocks and
1262                                  * private objects are not released until the
1263                                  * syncing state (since they are only modified
1264                                  * then).
1265                                  */
1266                                 arc_release(db->db_buf, db);
1267                                 dbuf_fix_old_data(db, tx->tx_txg);
1268                                 data_old = db->db_buf;
1269                         }
1270                         ASSERT(data_old != NULL);
1271                 }
1272                 dr->dt.dl.dr_data = data_old;
1273         } else {
1274                 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1275                 list_create(&dr->dt.di.dr_children,
1276                     sizeof (dbuf_dirty_record_t),
1277                     offsetof(dbuf_dirty_record_t, dr_dirty_node));
1278         }
1279         if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1280                 dr->dr_accounted = db->db.db_size;
1281         dr->dr_dbuf = db;
1282         dr->dr_txg = tx->tx_txg;
1283         dr->dr_next = *drp;
1284         *drp = dr;
1285 
1286         /*
1287          * We could have been freed_in_flight between the dbuf_noread
1288          * and dbuf_dirty.  We win, as though the dbuf_noread() had
1289          * happened after the free.
1290          */
1291         if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1292             db->db_blkid != DMU_SPILL_BLKID) {
1293                 mutex_enter(&dn->dn_mtx);
1294                 if (dn->dn_free_ranges[txgoff] != NULL) {
1295                         range_tree_clear(dn->dn_free_ranges[txgoff],
1296                             db->db_blkid, 1);
1297                 }
1298                 mutex_exit(&dn->dn_mtx);
1299                 db->db_freed_in_flight = FALSE;
1300         }
1301 
1302         /*
1303          * This buffer is now part of this txg
1304          */
1305         dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1306         db->db_dirtycnt += 1;
1307         ASSERT3U(db->db_dirtycnt, <=, 3);
1308 
1309         mutex_exit(&db->db_mtx);
1310 
1311         if (db->db_blkid == DMU_BONUS_BLKID ||
1312             db->db_blkid == DMU_SPILL_BLKID) {
1313                 mutex_enter(&dn->dn_mtx);
1314                 ASSERT(!list_link_active(&dr->dr_dirty_node));
1315                 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1316                 mutex_exit(&dn->dn_mtx);
1317                 dnode_setdirty(dn, tx);
1318                 DB_DNODE_EXIT(db);
1319                 return (dr);
1320         }
1321 
1322         /*
1323          * The dn_struct_rwlock prevents db_blkptr from changing
1324          * due to a write from syncing context completing
1325          * while we are running, so we want to acquire it before
1326          * looking at db_blkptr.
1327          */
1328         if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1329                 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1330                 drop_struct_lock = TRUE;
1331         }
1332 
1333         if (do_free_accounting) {
1334                 blkptr_t *bp = db->db_blkptr;
1335                 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1336                     bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1337                 /*
1338                  * This is only a guess -- if the dbuf is dirty
1339                  * in a previous txg, we don't know how much
1340                  * space it will use on disk yet.  We should
1341                  * really have the struct_rwlock to access
1342                  * db_blkptr, but since this is just a guess,
1343                  * it's OK if we get an odd answer.
1344                  */
1345                 ddt_prefetch(os->os_spa, bp);
1346                 dnode_willuse_space(dn, -willfree, tx);
1347         }
1348 
1349         if (db->db_level == 0) {
1350                 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1351                 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1352         }
1353 
1354         if (db->db_level+1 < dn->dn_nlevels) {
1355                 dmu_buf_impl_t *parent = db->db_parent;
1356                 dbuf_dirty_record_t *di;
1357                 int parent_held = FALSE;
1358 
1359                 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1360                         int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1361 
1362                         parent = dbuf_hold_level(dn, db->db_level+1,
1363                             db->db_blkid >> epbs, FTAG);
1364                         ASSERT(parent != NULL);
1365                         parent_held = TRUE;
1366                 }
1367                 if (drop_struct_lock)
1368                         rw_exit(&dn->dn_struct_rwlock);
1369                 ASSERT3U(db->db_level+1, ==, parent->db_level);
1370                 di = dbuf_dirty(parent, tx);
1371                 if (parent_held)
1372                         dbuf_rele(parent, FTAG);
1373 
1374                 mutex_enter(&db->db_mtx);
1375                 /*
1376                  * Since we've dropped the mutex, it's possible that
1377                  * dbuf_undirty() might have changed this out from under us.
1378                  */
1379                 if (db->db_last_dirty == dr ||
1380                     dn->dn_object == DMU_META_DNODE_OBJECT) {
1381                         mutex_enter(&di->dt.di.dr_mtx);
1382                         ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1383                         ASSERT(!list_link_active(&dr->dr_dirty_node));
1384                         list_insert_tail(&di->dt.di.dr_children, dr);
1385                         mutex_exit(&di->dt.di.dr_mtx);
1386                         dr->dr_parent = di;
1387                 }
1388                 mutex_exit(&db->db_mtx);
1389         } else {
1390                 ASSERT(db->db_level+1 == dn->dn_nlevels);
1391                 ASSERT(db->db_blkid < dn->dn_nblkptr);
1392                 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1393                 mutex_enter(&dn->dn_mtx);
1394                 ASSERT(!list_link_active(&dr->dr_dirty_node));
1395                 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1396                 mutex_exit(&dn->dn_mtx);
1397                 if (drop_struct_lock)
1398                         rw_exit(&dn->dn_struct_rwlock);
1399         }
1400 
1401         dnode_setdirty(dn, tx);
1402         DB_DNODE_EXIT(db);
1403         return (dr);
1404 }
1405 
1406 /*
1407  * Undirty a buffer in the transaction group referenced by the given
1408  * transaction.  Return whether this evicted the dbuf.
1409  */
1410 static boolean_t
1411 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1412 {
1413         dnode_t *dn;
1414         uint64_t txg = tx->tx_txg;
1415         dbuf_dirty_record_t *dr, **drp;
1416 
1417         ASSERT(txg != 0);
1418         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1419         ASSERT0(db->db_level);
1420         ASSERT(MUTEX_HELD(&db->db_mtx));
1421 
1422         /*
1423          * If this buffer is not dirty, we're done.
1424          */
1425         for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1426                 if (dr->dr_txg <= txg)
1427                         break;
1428         if (dr == NULL || dr->dr_txg < txg)
1429                 return (B_FALSE);
1430         ASSERT(dr->dr_txg == txg);
1431         ASSERT(dr->dr_dbuf == db);
1432 
1433         DB_DNODE_ENTER(db);
1434         dn = DB_DNODE(db);
1435 
1436         dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1437 
1438         ASSERT(db->db.db_size != 0);
1439 
1440         /*
1441          * Any space we accounted for in dp_dirty_* will be cleaned up by
1442          * dsl_pool_sync().  This is relatively rare so the discrepancy
1443          * is not a big deal.
1444          */
1445 
1446         *drp = dr->dr_next;
1447 
1448         /*
1449          * Note that there are three places in dbuf_dirty()
1450          * where this dirty record may be put on a list.
1451          * Make sure to do a list_remove corresponding to
1452          * every one of those list_insert calls.
1453          */
1454         if (dr->dr_parent) {
1455                 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1456                 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1457                 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1458         } else if (db->db_blkid == DMU_SPILL_BLKID ||
1459             db->db_level+1 == dn->dn_nlevels) {
1460                 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1461                 mutex_enter(&dn->dn_mtx);
1462                 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1463                 mutex_exit(&dn->dn_mtx);
1464         }
1465         DB_DNODE_EXIT(db);
1466 
1467         if (db->db_state != DB_NOFILL) {
1468                 dbuf_unoverride(dr);
1469 
1470                 ASSERT(db->db_buf != NULL);
1471                 ASSERT(dr->dt.dl.dr_data != NULL);
1472                 if (dr->dt.dl.dr_data != db->db_buf)
1473                         VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1474         }
1475 
1476         if (db->db_level != 0) {
1477                 mutex_destroy(&dr->dt.di.dr_mtx);
1478                 list_destroy(&dr->dt.di.dr_children);
1479         }
1480 
1481         kmem_free(dr, sizeof (dbuf_dirty_record_t));
1482 
1483         ASSERT(db->db_dirtycnt > 0);
1484         db->db_dirtycnt -= 1;
1485 
1486         if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1487                 arc_buf_t *buf = db->db_buf;
1488 
1489                 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1490                 dbuf_clear_data(db);
1491                 VERIFY(arc_buf_remove_ref(buf, db));
1492                 dbuf_evict(db);
1493                 return (B_TRUE);
1494         }
1495 
1496         return (B_FALSE);
1497 }
1498 
1499 void
1500 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1501 {
1502         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1503         int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1504 
1505         ASSERT(tx->tx_txg != 0);
1506         ASSERT(!refcount_is_zero(&db->db_holds));
1507 
1508         /*
1509          * Quick check for dirtyness.  For already dirty blocks, this
1510          * reduces runtime of this function by >90%, and overall performance
1511          * by 50% for some workloads (e.g. file deletion with indirect blocks
1512          * cached).
1513          */
1514         mutex_enter(&db->db_mtx);
1515         dbuf_dirty_record_t *dr;
1516         for (dr = db->db_last_dirty;
1517             dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1518                 /*
1519                  * It's possible that it is already dirty but not cached,
1520                  * because there are some calls to dbuf_dirty() that don't
1521                  * go through dmu_buf_will_dirty().
1522                  */
1523                 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1524                         /* This dbuf is already dirty and cached. */
1525                         dbuf_redirty(dr);
1526                         mutex_exit(&db->db_mtx);
1527                         return;
1528                 }
1529         }
1530         mutex_exit(&db->db_mtx);
1531 
1532         DB_DNODE_ENTER(db);
1533         if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1534                 rf |= DB_RF_HAVESTRUCT;
1535         DB_DNODE_EXIT(db);
1536         (void) dbuf_read(db, NULL, rf);
1537         (void) dbuf_dirty(db, tx);
1538 }
1539 
1540 void
1541 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1542 {
1543         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1544 
1545         db->db_state = DB_NOFILL;
1546 
1547         dmu_buf_will_fill(db_fake, tx);
1548 }
1549 
1550 void
1551 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1552 {
1553         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1554 
1555         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1556         ASSERT(tx->tx_txg != 0);
1557         ASSERT(db->db_level == 0);
1558         ASSERT(!refcount_is_zero(&db->db_holds));
1559 
1560         ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1561             dmu_tx_private_ok(tx));
1562 
1563         dbuf_noread(db);
1564         (void) dbuf_dirty(db, tx);
1565 }
1566 
1567 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1568 /* ARGSUSED */
1569 void
1570 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1571 {
1572         mutex_enter(&db->db_mtx);
1573         DBUF_VERIFY(db);
1574 
1575         if (db->db_state == DB_FILL) {
1576                 if (db->db_level == 0 && db->db_freed_in_flight) {
1577                         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1578                         /* we were freed while filling */
1579                         /* XXX dbuf_undirty? */
1580                         bzero(db->db.db_data, db->db.db_size);
1581                         db->db_freed_in_flight = FALSE;
1582                 }
1583                 db->db_state = DB_CACHED;
1584                 cv_broadcast(&db->db_changed);
1585         }
1586         mutex_exit(&db->db_mtx);
1587 }
1588 
1589 void
1590 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1591     bp_embedded_type_t etype, enum zio_compress comp,
1592     int uncompressed_size, int compressed_size, int byteorder,
1593     dmu_tx_t *tx)
1594 {
1595         dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1596         struct dirty_leaf *dl;
1597         dmu_object_type_t type;
1598 
1599         DB_DNODE_ENTER(db);
1600         type = DB_DNODE(db)->dn_type;
1601         DB_DNODE_EXIT(db);
1602 
1603         ASSERT0(db->db_level);
1604         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1605 
1606         dmu_buf_will_not_fill(dbuf, tx);
1607 
1608         ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1609         dl = &db->db_last_dirty->dt.dl;
1610         encode_embedded_bp_compressed(&dl->dr_overridden_by,
1611             data, comp, uncompressed_size, compressed_size);
1612         BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1613         BP_SET_TYPE(&dl->dr_overridden_by, type);
1614         BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1615         BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1616 
1617         dl->dr_override_state = DR_OVERRIDDEN;
1618         dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1619 }
1620 
1621 /*
1622  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1623  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1624  */
1625 void
1626 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1627 {
1628         ASSERT(!refcount_is_zero(&db->db_holds));
1629         ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1630         ASSERT(db->db_level == 0);
1631         ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1632         ASSERT(buf != NULL);
1633         ASSERT(arc_buf_size(buf) == db->db.db_size);
1634         ASSERT(tx->tx_txg != 0);
1635 
1636         arc_return_buf(buf, db);
1637         ASSERT(arc_released(buf));
1638 
1639         mutex_enter(&db->db_mtx);
1640 
1641         while (db->db_state == DB_READ || db->db_state == DB_FILL)
1642                 cv_wait(&db->db_changed, &db->db_mtx);
1643 
1644         ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1645 
1646         if (db->db_state == DB_CACHED &&
1647             refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1648                 mutex_exit(&db->db_mtx);
1649                 (void) dbuf_dirty(db, tx);
1650                 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1651                 VERIFY(arc_buf_remove_ref(buf, db));
1652                 xuio_stat_wbuf_copied();
1653                 return;
1654         }
1655 
1656         xuio_stat_wbuf_nocopy();
1657         if (db->db_state == DB_CACHED) {
1658                 dbuf_dirty_record_t *dr = db->db_last_dirty;
1659 
1660                 ASSERT(db->db_buf != NULL);
1661                 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1662                         ASSERT(dr->dt.dl.dr_data == db->db_buf);
1663                         if (!arc_released(db->db_buf)) {
1664                                 ASSERT(dr->dt.dl.dr_override_state ==
1665                                     DR_OVERRIDDEN);
1666                                 arc_release(db->db_buf, db);
1667                         }
1668                         dr->dt.dl.dr_data = buf;
1669                         VERIFY(arc_buf_remove_ref(db->db_buf, db));
1670                 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1671                         arc_release(db->db_buf, db);
1672                         VERIFY(arc_buf_remove_ref(db->db_buf, db));
1673                 }
1674                 db->db_buf = NULL;
1675         }
1676         ASSERT(db->db_buf == NULL);
1677         dbuf_set_data(db, buf);
1678         db->db_state = DB_FILL;
1679         mutex_exit(&db->db_mtx);
1680         (void) dbuf_dirty(db, tx);
1681         dmu_buf_fill_done(&db->db, tx);
1682 }
1683 
1684 /*
1685  * "Clear" the contents of this dbuf.  This will mark the dbuf
1686  * EVICTING and clear *most* of its references.  Unfortunately,
1687  * when we are not holding the dn_dbufs_mtx, we can't clear the
1688  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1689  * in this case.  For callers from the DMU we will usually see:
1690  *      dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1691  * For the arc callback, we will usually see:
1692  *      dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1693  * Sometimes, though, we will get a mix of these two:
1694  *      DMU: dbuf_clear()->arc_clear_callback()
1695  *      ARC: dbuf_do_evict()->dbuf_destroy()
1696  *
1697  * This routine will dissociate the dbuf from the arc, by calling
1698  * arc_clear_callback(), but will not evict the data from the ARC.
1699  */
1700 void
1701 dbuf_clear(dmu_buf_impl_t *db)
1702 {
1703         dnode_t *dn;
1704         dmu_buf_impl_t *parent = db->db_parent;
1705         dmu_buf_impl_t *dndb;
1706         boolean_t dbuf_gone = B_FALSE;
1707 
1708         ASSERT(MUTEX_HELD(&db->db_mtx));
1709         ASSERT(refcount_is_zero(&db->db_holds));
1710 
1711         dbuf_evict_user(db);
1712 
1713         if (db->db_state == DB_CACHED) {
1714                 ASSERT(db->db.db_data != NULL);
1715                 if (db->db_blkid == DMU_BONUS_BLKID) {
1716                         zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1717                         arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1718                 }
1719                 db->db.db_data = NULL;
1720                 db->db_state = DB_UNCACHED;
1721         }
1722 
1723         ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1724         ASSERT(db->db_data_pending == NULL);
1725 
1726         db->db_state = DB_EVICTING;
1727         db->db_blkptr = NULL;
1728 
1729         DB_DNODE_ENTER(db);
1730         dn = DB_DNODE(db);
1731         dndb = dn->dn_dbuf;
1732         if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1733                 avl_remove(&dn->dn_dbufs, db);
1734                 atomic_dec_32(&dn->dn_dbufs_count);
1735                 membar_producer();
1736                 DB_DNODE_EXIT(db);
1737                 /*
1738                  * Decrementing the dbuf count means that the hold corresponding
1739                  * to the removed dbuf is no longer discounted in dnode_move(),
1740                  * so the dnode cannot be moved until after we release the hold.
1741                  * The membar_producer() ensures visibility of the decremented
1742                  * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1743                  * release any lock.
1744                  */
1745                 dnode_rele(dn, db);
1746                 db->db_dnode_handle = NULL;
1747         } else {
1748                 DB_DNODE_EXIT(db);
1749         }
1750 
1751         if (db->db_buf)
1752                 dbuf_gone = arc_clear_callback(db->db_buf);
1753 
1754         if (!dbuf_gone)
1755                 mutex_exit(&db->db_mtx);
1756 
1757         /*
1758          * If this dbuf is referenced from an indirect dbuf,
1759          * decrement the ref count on the indirect dbuf.
1760          */
1761         if (parent && parent != dndb)
1762                 dbuf_rele(parent, db);
1763 }
1764 
1765 static int
1766 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1767     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1768 {
1769         int nlevels, epbs;
1770 
1771         *parentp = NULL;
1772         *bpp = NULL;
1773 
1774         ASSERT(blkid != DMU_BONUS_BLKID);
1775 
1776         if (blkid == DMU_SPILL_BLKID) {
1777                 mutex_enter(&dn->dn_mtx);
1778                 if (dn->dn_have_spill &&
1779                     (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1780                         *bpp = &dn->dn_phys->dn_spill;
1781                 else
1782                         *bpp = NULL;
1783                 dbuf_add_ref(dn->dn_dbuf, NULL);
1784                 *parentp = dn->dn_dbuf;
1785                 mutex_exit(&dn->dn_mtx);
1786                 return (0);
1787         }
1788 
1789         if (dn->dn_phys->dn_nlevels == 0)
1790                 nlevels = 1;
1791         else
1792                 nlevels = dn->dn_phys->dn_nlevels;
1793 
1794         epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1795 
1796         ASSERT3U(level * epbs, <, 64);
1797         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1798         if (level >= nlevels ||
1799             (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1800                 /* the buffer has no parent yet */
1801                 return (SET_ERROR(ENOENT));
1802         } else if (level < nlevels-1) {
1803                 /* this block is referenced from an indirect block */
1804                 int err = dbuf_hold_impl(dn, level+1,
1805                     blkid >> epbs, fail_sparse, NULL, parentp);
1806                 if (err)
1807                         return (err);
1808                 err = dbuf_read(*parentp, NULL,
1809                     (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1810                 if (err) {
1811                         dbuf_rele(*parentp, NULL);
1812                         *parentp = NULL;
1813                         return (err);
1814                 }
1815                 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1816                     (blkid & ((1ULL << epbs) - 1));
1817                 return (0);
1818         } else {
1819                 /* the block is referenced from the dnode */
1820                 ASSERT3U(level, ==, nlevels-1);
1821                 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1822                     blkid < dn->dn_phys->dn_nblkptr);
1823                 if (dn->dn_dbuf) {
1824                         dbuf_add_ref(dn->dn_dbuf, NULL);
1825                         *parentp = dn->dn_dbuf;
1826                 }
1827                 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1828                 return (0);
1829         }
1830 }
1831 
1832 static dmu_buf_impl_t *
1833 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1834     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1835 {
1836         objset_t *os = dn->dn_objset;
1837         dmu_buf_impl_t *db, *odb;
1838 
1839         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1840         ASSERT(dn->dn_type != DMU_OT_NONE);
1841 
1842         db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1843 
1844         db->db_objset = os;
1845         db->db.db_object = dn->dn_object;
1846         db->db_level = level;
1847         db->db_blkid = blkid;
1848         db->db_last_dirty = NULL;
1849         db->db_dirtycnt = 0;
1850         db->db_dnode_handle = dn->dn_handle;
1851         db->db_parent = parent;
1852         db->db_blkptr = blkptr;
1853 
1854         db->db_user = NULL;
1855         db->db_user_immediate_evict = FALSE;
1856         db->db_freed_in_flight = FALSE;
1857         db->db_pending_evict = FALSE;
1858 
1859         if (blkid == DMU_BONUS_BLKID) {
1860                 ASSERT3P(parent, ==, dn->dn_dbuf);
1861                 db->db.db_size = DN_MAX_BONUSLEN -
1862                     (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1863                 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1864                 db->db.db_offset = DMU_BONUS_BLKID;
1865                 db->db_state = DB_UNCACHED;
1866                 /* the bonus dbuf is not placed in the hash table */
1867                 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1868                 return (db);
1869         } else if (blkid == DMU_SPILL_BLKID) {
1870                 db->db.db_size = (blkptr != NULL) ?
1871                     BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1872                 db->db.db_offset = 0;
1873         } else {
1874                 int blocksize =
1875                     db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1876                 db->db.db_size = blocksize;
1877                 db->db.db_offset = db->db_blkid * blocksize;
1878         }
1879 
1880         /*
1881          * Hold the dn_dbufs_mtx while we get the new dbuf
1882          * in the hash table *and* added to the dbufs list.
1883          * This prevents a possible deadlock with someone
1884          * trying to look up this dbuf before its added to the
1885          * dn_dbufs list.
1886          */
1887         mutex_enter(&dn->dn_dbufs_mtx);
1888         db->db_state = DB_EVICTING;
1889         if ((odb = dbuf_hash_insert(db)) != NULL) {
1890                 /* someone else inserted it first */
1891                 kmem_cache_free(dbuf_cache, db);
1892                 mutex_exit(&dn->dn_dbufs_mtx);
1893                 return (odb);
1894         }
1895         avl_add(&dn->dn_dbufs, db);
1896         if (db->db_level == 0 && db->db_blkid >=
1897             dn->dn_unlisted_l0_blkid)
1898                 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1899         db->db_state = DB_UNCACHED;
1900         mutex_exit(&dn->dn_dbufs_mtx);
1901         arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1902 
1903         if (parent && parent != dn->dn_dbuf)
1904                 dbuf_add_ref(parent, db);
1905 
1906         ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1907             refcount_count(&dn->dn_holds) > 0);
1908         (void) refcount_add(&dn->dn_holds, db);
1909         atomic_inc_32(&dn->dn_dbufs_count);
1910 
1911         dprintf_dbuf(db, "db=%p\n", db);
1912 
1913         return (db);
1914 }
1915 
1916 static int
1917 dbuf_do_evict(void *private)
1918 {
1919         dmu_buf_impl_t *db = private;
1920 
1921         if (!MUTEX_HELD(&db->db_mtx))
1922                 mutex_enter(&db->db_mtx);
1923 
1924         ASSERT(refcount_is_zero(&db->db_holds));
1925 
1926         if (db->db_state != DB_EVICTING) {
1927                 ASSERT(db->db_state == DB_CACHED);
1928                 DBUF_VERIFY(db);
1929                 db->db_buf = NULL;
1930                 dbuf_evict(db);
1931         } else {
1932                 mutex_exit(&db->db_mtx);
1933                 dbuf_destroy(db);
1934         }
1935         return (0);
1936 }
1937 
1938 static void
1939 dbuf_destroy(dmu_buf_impl_t *db)
1940 {
1941         ASSERT(refcount_is_zero(&db->db_holds));
1942 
1943         if (db->db_blkid != DMU_BONUS_BLKID) {
1944                 /*
1945                  * If this dbuf is still on the dn_dbufs list,
1946                  * remove it from that list.
1947                  */
1948                 if (db->db_dnode_handle != NULL) {
1949                         dnode_t *dn;
1950 
1951                         DB_DNODE_ENTER(db);
1952                         dn = DB_DNODE(db);
1953                         mutex_enter(&dn->dn_dbufs_mtx);
1954                         avl_remove(&dn->dn_dbufs, db);
1955                         atomic_dec_32(&dn->dn_dbufs_count);
1956                         mutex_exit(&dn->dn_dbufs_mtx);
1957                         DB_DNODE_EXIT(db);
1958                         /*
1959                          * Decrementing the dbuf count means that the hold
1960                          * corresponding to the removed dbuf is no longer
1961                          * discounted in dnode_move(), so the dnode cannot be
1962                          * moved until after we release the hold.
1963                          */
1964                         dnode_rele(dn, db);
1965                         db->db_dnode_handle = NULL;
1966                 }
1967                 dbuf_hash_remove(db);
1968         }
1969         db->db_parent = NULL;
1970         db->db_buf = NULL;
1971 
1972         ASSERT(db->db.db_data == NULL);
1973         ASSERT(db->db_hash_next == NULL);
1974         ASSERT(db->db_blkptr == NULL);
1975         ASSERT(db->db_data_pending == NULL);
1976 
1977         kmem_cache_free(dbuf_cache, db);
1978         arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1979 }
1980 
1981 void
1982 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1983 {
1984         dmu_buf_impl_t *db = NULL;
1985         blkptr_t *bp = NULL;
1986 
1987         ASSERT(blkid != DMU_BONUS_BLKID);
1988         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1989 
1990         if (dnode_block_freed(dn, blkid))
1991                 return;
1992 
1993         /* dbuf_find() returns with db_mtx held */
1994         if (db = dbuf_find(dn->dn_objset, dn->dn_object, 0, blkid)) {
1995                 /*
1996                  * This dbuf is already in the cache.  We assume that
1997                  * it is already CACHED, or else about to be either
1998                  * read or filled.
1999                  */
2000                 mutex_exit(&db->db_mtx);
2001                 return;
2002         }
2003 
2004         if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
2005                 if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2006                         dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2007                         arc_flags_t aflags =
2008                             ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2009                         zbookmark_phys_t zb;
2010 
2011                         SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
2012                             dn->dn_object, 0, blkid);
2013 
2014                         (void) arc_read(NULL, dn->dn_objset->os_spa,
2015                             bp, NULL, NULL, prio,
2016                             ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2017                             &aflags, &zb);
2018                 }
2019                 if (db)
2020                         dbuf_rele(db, NULL);
2021         }
2022 }
2023 
2024 /*
2025  * Returns with db_holds incremented, and db_mtx not held.
2026  * Note: dn_struct_rwlock must be held.
2027  */
2028 int
2029 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
2030     void *tag, dmu_buf_impl_t **dbp)
2031 {
2032         dmu_buf_impl_t *db, *parent = NULL;
2033 
2034         ASSERT(blkid != DMU_BONUS_BLKID);
2035         ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2036         ASSERT3U(dn->dn_nlevels, >, level);
2037 
2038         *dbp = NULL;
2039 top:
2040         /* dbuf_find() returns with db_mtx held */
2041         db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2042 
2043         if (db == NULL) {
2044                 blkptr_t *bp = NULL;
2045                 int err;
2046 
2047                 ASSERT3P(parent, ==, NULL);
2048                 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2049                 if (fail_sparse) {
2050                         if (err == 0 && bp && BP_IS_HOLE(bp))
2051                                 err = SET_ERROR(ENOENT);
2052                         if (err) {
2053                                 if (parent)
2054                                         dbuf_rele(parent, NULL);
2055                                 return (err);
2056                         }
2057                 }
2058                 if (err && err != ENOENT)
2059                         return (err);
2060                 db = dbuf_create(dn, level, blkid, parent, bp);
2061         }
2062 
2063         if (db->db_buf && refcount_is_zero(&db->db_holds)) {
2064                 arc_buf_add_ref(db->db_buf, db);
2065                 if (db->db_buf->b_data == NULL) {
2066                         dbuf_clear(db);
2067                         if (parent) {
2068                                 dbuf_rele(parent, NULL);
2069                                 parent = NULL;
2070                         }
2071                         goto top;
2072                 }
2073                 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2074         }
2075 
2076         ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2077 
2078         /*
2079          * If this buffer is currently syncing out, and we are are
2080          * still referencing it from db_data, we need to make a copy
2081          * of it in case we decide we want to dirty it again in this txg.
2082          */
2083         if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2084             dn->dn_object != DMU_META_DNODE_OBJECT &&
2085             db->db_state == DB_CACHED && db->db_data_pending) {
2086                 dbuf_dirty_record_t *dr = db->db_data_pending;
2087 
2088                 if (dr->dt.dl.dr_data == db->db_buf) {
2089                         arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2090 
2091                         dbuf_set_data(db,
2092                             arc_buf_alloc(dn->dn_objset->os_spa,
2093                             db->db.db_size, db, type));
2094                         bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2095                             db->db.db_size);
2096                 }
2097         }
2098 
2099         (void) refcount_add(&db->db_holds, tag);
2100         DBUF_VERIFY(db);
2101         mutex_exit(&db->db_mtx);
2102 
2103         /* NOTE: we can't rele the parent until after we drop the db_mtx */
2104         if (parent)
2105                 dbuf_rele(parent, NULL);
2106 
2107         ASSERT3P(DB_DNODE(db), ==, dn);
2108         ASSERT3U(db->db_blkid, ==, blkid);
2109         ASSERT3U(db->db_level, ==, level);
2110         *dbp = db;
2111 
2112         return (0);
2113 }
2114 
2115 dmu_buf_impl_t *
2116 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2117 {
2118         dmu_buf_impl_t *db;
2119         int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
2120         return (err ? NULL : db);
2121 }
2122 
2123 dmu_buf_impl_t *
2124 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2125 {
2126         dmu_buf_impl_t *db;
2127         int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2128         return (err ? NULL : db);
2129 }
2130 
2131 void
2132 dbuf_create_bonus(dnode_t *dn)
2133 {
2134         ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2135 
2136         ASSERT(dn->dn_bonus == NULL);
2137         dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2138 }
2139 
2140 int
2141 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2142 {
2143         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2144         dnode_t *dn;
2145 
2146         if (db->db_blkid != DMU_SPILL_BLKID)
2147                 return (SET_ERROR(ENOTSUP));
2148         if (blksz == 0)
2149                 blksz = SPA_MINBLOCKSIZE;
2150         ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2151         blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2152 
2153         DB_DNODE_ENTER(db);
2154         dn = DB_DNODE(db);
2155         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2156         dbuf_new_size(db, blksz, tx);
2157         rw_exit(&dn->dn_struct_rwlock);
2158         DB_DNODE_EXIT(db);
2159 
2160         return (0);
2161 }
2162 
2163 void
2164 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2165 {
2166         dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2167 }
2168 
2169 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2170 void
2171 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2172 {
2173         int64_t holds = refcount_add(&db->db_holds, tag);
2174         ASSERT(holds > 1);
2175 }
2176 
2177 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2178 boolean_t
2179 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2180     void *tag)
2181 {
2182         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2183         dmu_buf_impl_t *found_db;
2184         boolean_t result = B_FALSE;
2185 
2186         if (db->db_blkid == DMU_BONUS_BLKID)
2187                 found_db = dbuf_find_bonus(os, obj);
2188         else
2189                 found_db = dbuf_find(os, obj, 0, blkid);
2190 
2191         if (found_db != NULL) {
2192                 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2193                         (void) refcount_add(&db->db_holds, tag);
2194                         result = B_TRUE;
2195                 }
2196                 mutex_exit(&db->db_mtx);
2197         }
2198         return (result);
2199 }
2200 
2201 /*
2202  * If you call dbuf_rele() you had better not be referencing the dnode handle
2203  * unless you have some other direct or indirect hold on the dnode. (An indirect
2204  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2205  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2206  * dnode's parent dbuf evicting its dnode handles.
2207  */
2208 void
2209 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2210 {
2211         mutex_enter(&db->db_mtx);
2212         dbuf_rele_and_unlock(db, tag);
2213 }
2214 
2215 void
2216 dmu_buf_rele(dmu_buf_t *db, void *tag)
2217 {
2218         dbuf_rele((dmu_buf_impl_t *)db, tag);
2219 }
2220 
2221 /*
2222  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2223  * db_dirtycnt and db_holds to be updated atomically.
2224  */
2225 void
2226 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2227 {
2228         int64_t holds;
2229 
2230         ASSERT(MUTEX_HELD(&db->db_mtx));
2231         DBUF_VERIFY(db);
2232 
2233         /*
2234          * Remove the reference to the dbuf before removing its hold on the
2235          * dnode so we can guarantee in dnode_move() that a referenced bonus
2236          * buffer has a corresponding dnode hold.
2237          */
2238         holds = refcount_remove(&db->db_holds, tag);
2239         ASSERT(holds >= 0);
2240 
2241         /*
2242          * We can't freeze indirects if there is a possibility that they
2243          * may be modified in the current syncing context.
2244          */
2245         if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2246                 arc_buf_freeze(db->db_buf);
2247 
2248         if (holds == db->db_dirtycnt &&
2249             db->db_level == 0 && db->db_user_immediate_evict)
2250                 dbuf_evict_user(db);
2251 
2252         if (holds == 0) {
2253                 if (db->db_blkid == DMU_BONUS_BLKID) {
2254                         dnode_t *dn;
2255                         boolean_t evict_dbuf = db->db_pending_evict;
2256 
2257                         /*
2258                          * If the dnode moves here, we cannot cross this
2259                          * barrier until the move completes.
2260                          */
2261                         DB_DNODE_ENTER(db);
2262 
2263                         dn = DB_DNODE(db);
2264                         atomic_dec_32(&dn->dn_dbufs_count);
2265 
2266                         /*
2267                          * Decrementing the dbuf count means that the bonus
2268                          * buffer's dnode hold is no longer discounted in
2269                          * dnode_move(). The dnode cannot move until after
2270                          * the dnode_rele() below.
2271                          */
2272                         DB_DNODE_EXIT(db);
2273 
2274                         /*
2275                          * Do not reference db after its lock is dropped.
2276                          * Another thread may evict it.
2277                          */
2278                         mutex_exit(&db->db_mtx);
2279 
2280                         if (evict_dbuf)
2281                                 dnode_evict_bonus(dn);
2282 
2283                         dnode_rele(dn, db);
2284                 } else if (db->db_buf == NULL) {
2285                         /*
2286                          * This is a special case: we never associated this
2287                          * dbuf with any data allocated from the ARC.
2288                          */
2289                         ASSERT(db->db_state == DB_UNCACHED ||
2290                             db->db_state == DB_NOFILL);
2291                         dbuf_evict(db);
2292                 } else if (arc_released(db->db_buf)) {
2293                         arc_buf_t *buf = db->db_buf;
2294                         /*
2295                          * This dbuf has anonymous data associated with it.
2296                          */
2297                         dbuf_clear_data(db);
2298                         VERIFY(arc_buf_remove_ref(buf, db));
2299                         dbuf_evict(db);
2300                 } else {
2301                         VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2302 
2303                         /*
2304                          * A dbuf will be eligible for eviction if either the
2305                          * 'primarycache' property is set or a duplicate
2306                          * copy of this buffer is already cached in the arc.
2307                          *
2308                          * In the case of the 'primarycache' a buffer
2309                          * is considered for eviction if it matches the
2310                          * criteria set in the property.
2311                          *
2312                          * To decide if our buffer is considered a
2313                          * duplicate, we must call into the arc to determine
2314                          * if multiple buffers are referencing the same
2315                          * block on-disk. If so, then we simply evict
2316                          * ourselves.
2317                          */
2318                         if (!DBUF_IS_CACHEABLE(db)) {
2319                                 if (db->db_blkptr != NULL &&
2320                                     !BP_IS_HOLE(db->db_blkptr) &&
2321                                     !BP_IS_EMBEDDED(db->db_blkptr)) {
2322                                         spa_t *spa =
2323                                             dmu_objset_spa(db->db_objset);
2324                                         blkptr_t bp = *db->db_blkptr;
2325                                         dbuf_clear(db);
2326                                         arc_freed(spa, &bp);
2327                                 } else {
2328                                         dbuf_clear(db);
2329                                 }
2330                         } else if (db->db_pending_evict ||
2331                             arc_buf_eviction_needed(db->db_buf)) {
2332                                 dbuf_clear(db);
2333                         } else {
2334                                 mutex_exit(&db->db_mtx);
2335                         }
2336                 }
2337         } else {
2338                 mutex_exit(&db->db_mtx);
2339         }
2340 }
2341 
2342 #pragma weak dmu_buf_refcount = dbuf_refcount
2343 uint64_t
2344 dbuf_refcount(dmu_buf_impl_t *db)
2345 {
2346         return (refcount_count(&db->db_holds));
2347 }
2348 
2349 void *
2350 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2351     dmu_buf_user_t *new_user)
2352 {
2353         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2354 
2355         mutex_enter(&db->db_mtx);
2356         dbuf_verify_user(db, DBVU_NOT_EVICTING);
2357         if (db->db_user == old_user)
2358                 db->db_user = new_user;
2359         else
2360                 old_user = db->db_user;
2361         dbuf_verify_user(db, DBVU_NOT_EVICTING);
2362         mutex_exit(&db->db_mtx);
2363 
2364         return (old_user);
2365 }
2366 
2367 void *
2368 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2369 {
2370         return (dmu_buf_replace_user(db_fake, NULL, user));
2371 }
2372 
2373 void *
2374 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2375 {
2376         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2377 
2378         db->db_user_immediate_evict = TRUE;
2379         return (dmu_buf_set_user(db_fake, user));
2380 }
2381 
2382 void *
2383 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2384 {
2385         return (dmu_buf_replace_user(db_fake, user, NULL));
2386 }
2387 
2388 void *
2389 dmu_buf_get_user(dmu_buf_t *db_fake)
2390 {
2391         dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2392 
2393         dbuf_verify_user(db, DBVU_NOT_EVICTING);
2394         return (db->db_user);
2395 }
2396 
2397 void
2398 dmu_buf_user_evict_wait()
2399 {
2400         taskq_wait(dbu_evict_taskq);
2401 }
2402 
2403 boolean_t
2404 dmu_buf_freeable(dmu_buf_t *dbuf)
2405 {
2406         boolean_t res = B_FALSE;
2407         dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2408 
2409         if (db->db_blkptr)
2410                 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2411                     db->db_blkptr, db->db_blkptr->blk_birth);
2412 
2413         return (res);
2414 }
2415 
2416 blkptr_t *
2417 dmu_buf_get_blkptr(dmu_buf_t *db)
2418 {
2419         dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2420         return (dbi->db_blkptr);
2421 }
2422 
2423 static void
2424 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2425 {
2426         /* ASSERT(dmu_tx_is_syncing(tx) */
2427         ASSERT(MUTEX_HELD(&db->db_mtx));
2428 
2429         if (db->db_blkptr != NULL)
2430                 return;
2431 
2432         if (db->db_blkid == DMU_SPILL_BLKID) {
2433                 db->db_blkptr = &dn->dn_phys->dn_spill;
2434                 BP_ZERO(db->db_blkptr);
2435                 return;
2436         }
2437         if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2438                 /*
2439                  * This buffer was allocated at a time when there was
2440                  * no available blkptrs from the dnode, or it was
2441                  * inappropriate to hook it in (i.e., nlevels mis-match).
2442                  */
2443                 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2444                 ASSERT(db->db_parent == NULL);
2445                 db->db_parent = dn->dn_dbuf;
2446                 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2447                 DBUF_VERIFY(db);
2448         } else {
2449                 dmu_buf_impl_t *parent = db->db_parent;
2450                 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2451 
2452                 ASSERT(dn->dn_phys->dn_nlevels > 1);
2453                 if (parent == NULL) {
2454                         mutex_exit(&db->db_mtx);
2455                         rw_enter(&dn->dn_struct_rwlock, RW_READER);
2456                         (void) dbuf_hold_impl(dn, db->db_level+1,
2457                             db->db_blkid >> epbs, FALSE, db, &parent);
2458                         rw_exit(&dn->dn_struct_rwlock);
2459                         mutex_enter(&db->db_mtx);
2460                         db->db_parent = parent;
2461                 }
2462                 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2463                     (db->db_blkid & ((1ULL << epbs) - 1));
2464                 DBUF_VERIFY(db);
2465         }
2466 }
2467 
2468 static void
2469 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2470 {
2471         dmu_buf_impl_t *db = dr->dr_dbuf;
2472         dnode_t *dn;
2473         zio_t *zio;
2474 
2475         ASSERT(dmu_tx_is_syncing(tx));
2476 
2477         dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2478 
2479         mutex_enter(&db->db_mtx);
2480 
2481         ASSERT(db->db_level > 0);
2482         DBUF_VERIFY(db);
2483 
2484         /* Read the block if it hasn't been read yet. */
2485         if (db->db_buf == NULL) {
2486                 mutex_exit(&db->db_mtx);
2487                 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2488                 mutex_enter(&db->db_mtx);
2489         }
2490         ASSERT3U(db->db_state, ==, DB_CACHED);
2491         ASSERT(db->db_buf != NULL);
2492 
2493         DB_DNODE_ENTER(db);
2494         dn = DB_DNODE(db);
2495         /* Indirect block size must match what the dnode thinks it is. */
2496         ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2497         dbuf_check_blkptr(dn, db);
2498         DB_DNODE_EXIT(db);
2499 
2500         /* Provide the pending dirty record to child dbufs */
2501         db->db_data_pending = dr;
2502 
2503         mutex_exit(&db->db_mtx);
2504         dbuf_write(dr, db->db_buf, tx);
2505 
2506         zio = dr->dr_zio;
2507         mutex_enter(&dr->dt.di.dr_mtx);
2508         dbuf_sync_list(&dr->dt.di.dr_children, tx);
2509         ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2510         mutex_exit(&dr->dt.di.dr_mtx);
2511         zio_nowait(zio);
2512 }
2513 
2514 static void
2515 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2516 {
2517         arc_buf_t **datap = &dr->dt.dl.dr_data;
2518         dmu_buf_impl_t *db = dr->dr_dbuf;
2519         dnode_t *dn;
2520         objset_t *os;
2521         uint64_t txg = tx->tx_txg;
2522 
2523         ASSERT(dmu_tx_is_syncing(tx));
2524 
2525         dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2526 
2527         mutex_enter(&db->db_mtx);
2528         /*
2529          * To be synced, we must be dirtied.  But we
2530          * might have been freed after the dirty.
2531          */
2532         if (db->db_state == DB_UNCACHED) {
2533                 /* This buffer has been freed since it was dirtied */
2534                 ASSERT(db->db.db_data == NULL);
2535         } else if (db->db_state == DB_FILL) {
2536                 /* This buffer was freed and is now being re-filled */
2537                 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2538         } else {
2539                 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2540         }
2541         DBUF_VERIFY(db);
2542 
2543         DB_DNODE_ENTER(db);
2544         dn = DB_DNODE(db);
2545 
2546         if (db->db_blkid == DMU_SPILL_BLKID) {
2547                 mutex_enter(&dn->dn_mtx);
2548                 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2549                 mutex_exit(&dn->dn_mtx);
2550         }
2551 
2552         /*
2553          * If this is a bonus buffer, simply copy the bonus data into the
2554          * dnode.  It will be written out when the dnode is synced (and it
2555          * will be synced, since it must have been dirty for dbuf_sync to
2556          * be called).
2557          */
2558         if (db->db_blkid == DMU_BONUS_BLKID) {
2559                 dbuf_dirty_record_t **drp;
2560 
2561                 ASSERT(*datap != NULL);
2562                 ASSERT0(db->db_level);
2563                 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2564                 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2565                 DB_DNODE_EXIT(db);
2566 
2567                 if (*datap != db->db.db_data) {
2568                         zio_buf_free(*datap, DN_MAX_BONUSLEN);
2569                         arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2570                 }
2571                 db->db_data_pending = NULL;
2572                 drp = &db->db_last_dirty;
2573                 while (*drp != dr)
2574                         drp = &(*drp)->dr_next;
2575                 ASSERT(dr->dr_next == NULL);
2576                 ASSERT(dr->dr_dbuf == db);
2577                 *drp = dr->dr_next;
2578                 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2579                 ASSERT(db->db_dirtycnt > 0);
2580                 db->db_dirtycnt -= 1;
2581                 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2582                 return;
2583         }
2584 
2585         os = dn->dn_objset;
2586 
2587         /*
2588          * This function may have dropped the db_mtx lock allowing a dmu_sync
2589          * operation to sneak in. As a result, we need to ensure that we
2590          * don't check the dr_override_state until we have returned from
2591          * dbuf_check_blkptr.
2592          */
2593         dbuf_check_blkptr(dn, db);
2594 
2595         /*
2596          * If this buffer is in the middle of an immediate write,
2597          * wait for the synchronous IO to complete.
2598          */
2599         while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2600                 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2601                 cv_wait(&db->db_changed, &db->db_mtx);
2602                 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2603         }
2604 
2605         if (db->db_state != DB_NOFILL &&
2606             dn->dn_object != DMU_META_DNODE_OBJECT &&
2607             refcount_count(&db->db_holds) > 1 &&
2608             dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2609             *datap == db->db_buf) {
2610                 /*
2611                  * If this buffer is currently "in use" (i.e., there
2612                  * are active holds and db_data still references it),
2613                  * then make a copy before we start the write so that
2614                  * any modifications from the open txg will not leak
2615                  * into this write.
2616                  *
2617                  * NOTE: this copy does not need to be made for
2618                  * objects only modified in the syncing context (e.g.
2619                  * DNONE_DNODE blocks).
2620                  */
2621                 int blksz = arc_buf_size(*datap);
2622                 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2623                 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2624                 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2625         }
2626         db->db_data_pending = dr;
2627 
2628         mutex_exit(&db->db_mtx);
2629 
2630         dbuf_write(dr, *datap, tx);
2631 
2632         ASSERT(!list_link_active(&dr->dr_dirty_node));
2633         if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2634                 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2635                 DB_DNODE_EXIT(db);
2636         } else {
2637                 /*
2638                  * Although zio_nowait() does not "wait for an IO", it does
2639                  * initiate the IO. If this is an empty write it seems plausible
2640                  * that the IO could actually be completed before the nowait
2641                  * returns. We need to DB_DNODE_EXIT() first in case
2642                  * zio_nowait() invalidates the dbuf.
2643                  */
2644                 DB_DNODE_EXIT(db);
2645                 zio_nowait(dr->dr_zio);
2646         }
2647 }
2648 
2649 void
2650 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2651 {
2652         dbuf_dirty_record_t *dr;
2653 
2654         while (dr = list_head(list)) {
2655                 if (dr->dr_zio != NULL) {
2656                         /*
2657                          * If we find an already initialized zio then we
2658                          * are processing the meta-dnode, and we have finished.
2659                          * The dbufs for all dnodes are put back on the list
2660                          * during processing, so that we can zio_wait()
2661                          * these IOs after initiating all child IOs.
2662                          */
2663                         ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2664                             DMU_META_DNODE_OBJECT);
2665                         break;
2666                 }
2667                 list_remove(list, dr);
2668                 if (dr->dr_dbuf->db_level > 0)
2669                         dbuf_sync_indirect(dr, tx);
2670                 else
2671                         dbuf_sync_leaf(dr, tx);
2672         }
2673 }
2674 
2675 /* ARGSUSED */
2676 static void
2677 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2678 {
2679         dmu_buf_impl_t *db = vdb;
2680         dnode_t *dn;
2681         blkptr_t *bp = zio->io_bp;
2682         blkptr_t *bp_orig = &zio->io_bp_orig;
2683         spa_t *spa = zio->io_spa;
2684         int64_t delta;
2685         uint64_t fill = 0;
2686         int i;
2687 
2688         ASSERT3P(db->db_blkptr, !=, NULL);
2689         ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
2690 
2691         DB_DNODE_ENTER(db);
2692         dn = DB_DNODE(db);
2693         delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2694         dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2695         zio->io_prev_space_delta = delta;
2696 
2697         if (bp->blk_birth != 0) {
2698                 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2699                     BP_GET_TYPE(bp) == dn->dn_type) ||
2700                     (db->db_blkid == DMU_SPILL_BLKID &&
2701                     BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2702                     BP_IS_EMBEDDED(bp));
2703                 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2704         }
2705 
2706         mutex_enter(&db->db_mtx);
2707 
2708 #ifdef ZFS_DEBUG
2709         if (db->db_blkid == DMU_SPILL_BLKID) {
2710                 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2711                 ASSERT(!(BP_IS_HOLE(bp)) &&
2712                     db->db_blkptr == &dn->dn_phys->dn_spill);
2713         }
2714 #endif
2715 
2716         if (db->db_level == 0) {
2717                 mutex_enter(&dn->dn_mtx);
2718                 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2719                     db->db_blkid != DMU_SPILL_BLKID)
2720                         dn->dn_phys->dn_maxblkid = db->db_blkid;
2721                 mutex_exit(&dn->dn_mtx);
2722 
2723                 if (dn->dn_type == DMU_OT_DNODE) {
2724                         dnode_phys_t *dnp = db->db.db_data;
2725                         for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2726                             i--, dnp++) {
2727                                 if (dnp->dn_type != DMU_OT_NONE)
2728                                         fill++;
2729                         }
2730                 } else {
2731                         if (BP_IS_HOLE(bp)) {
2732                                 fill = 0;
2733                         } else {
2734                                 fill = 1;
2735                         }
2736                 }
2737         } else {
2738                 blkptr_t *ibp = db->db.db_data;
2739                 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2740                 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2741                         if (BP_IS_HOLE(ibp))
2742                                 continue;
2743                         fill += BP_GET_FILL(ibp);
2744                 }
2745         }
2746         DB_DNODE_EXIT(db);
2747 
2748         if (!BP_IS_EMBEDDED(bp))
2749                 bp->blk_fill = fill;
2750 
2751         mutex_exit(&db->db_mtx);
2752 
2753         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2754         *db->db_blkptr = *bp;
2755         rw_exit(&dn->dn_struct_rwlock);
2756 }
2757 
2758 /*
2759  * The SPA will call this callback several times for each zio - once
2760  * for every physical child i/o (zio->io_phys_children times).  This
2761  * allows the DMU to monitor the progress of each logical i/o.  For example,
2762  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2763  * block.  There may be a long delay before all copies/fragments are completed,
2764  * so this callback allows us to retire dirty space gradually, as the physical
2765  * i/os complete.
2766  */
2767 /* ARGSUSED */
2768 static void
2769 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2770 {
2771         dmu_buf_impl_t *db = arg;
2772         objset_t *os = db->db_objset;
2773         dsl_pool_t *dp = dmu_objset_pool(os);
2774         dbuf_dirty_record_t *dr;
2775         int delta = 0;
2776 
2777         dr = db->db_data_pending;
2778         ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2779 
2780         /*
2781          * The callback will be called io_phys_children times.  Retire one
2782          * portion of our dirty space each time we are called.  Any rounding
2783          * error will be cleaned up by dsl_pool_sync()'s call to
2784          * dsl_pool_undirty_space().
2785          */
2786         delta = dr->dr_accounted / zio->io_phys_children;
2787         dsl_pool_undirty_space(dp, delta, zio->io_txg);
2788 }
2789 
2790 /* ARGSUSED */
2791 static void
2792 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2793 {
2794         dmu_buf_impl_t *db = vdb;
2795         blkptr_t *bp_orig = &zio->io_bp_orig;
2796         blkptr_t *bp = db->db_blkptr;
2797         objset_t *os = db->db_objset;
2798         dmu_tx_t *tx = os->os_synctx;
2799         dbuf_dirty_record_t **drp, *dr;
2800 
2801         ASSERT0(zio->io_error);
2802         ASSERT(db->db_blkptr == bp);
2803 
2804         /*
2805          * For nopwrites and rewrites we ensure that the bp matches our
2806          * original and bypass all the accounting.
2807          */
2808         if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2809                 ASSERT(BP_EQUAL(bp, bp_orig));
2810         } else {
2811                 dsl_dataset_t *ds = os->os_dsl_dataset;
2812                 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2813                 dsl_dataset_block_born(ds, bp, tx);
2814         }
2815 
2816         mutex_enter(&db->db_mtx);
2817 
2818         DBUF_VERIFY(db);
2819 
2820         drp = &db->db_last_dirty;
2821         while ((dr = *drp) != db->db_data_pending)
2822                 drp = &dr->dr_next;
2823         ASSERT(!list_link_active(&dr->dr_dirty_node));
2824         ASSERT(dr->dr_dbuf == db);
2825         ASSERT(dr->dr_next == NULL);
2826         *drp = dr->dr_next;
2827 
2828 #ifdef ZFS_DEBUG
2829         if (db->db_blkid == DMU_SPILL_BLKID) {
2830                 dnode_t *dn;
2831 
2832                 DB_DNODE_ENTER(db);
2833                 dn = DB_DNODE(db);
2834                 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2835                 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2836                     db->db_blkptr == &dn->dn_phys->dn_spill);
2837                 DB_DNODE_EXIT(db);
2838         }
2839 #endif
2840 
2841         if (db->db_level == 0) {
2842                 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2843                 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2844                 if (db->db_state != DB_NOFILL) {
2845                         if (dr->dt.dl.dr_data != db->db_buf)
2846                                 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2847                                     db));
2848                         else if (!arc_released(db->db_buf))
2849                                 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2850                 }
2851         } else {
2852                 dnode_t *dn;
2853 
2854                 DB_DNODE_ENTER(db);
2855                 dn = DB_DNODE(db);
2856                 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2857                 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2858                 if (!BP_IS_HOLE(db->db_blkptr)) {
2859                         int epbs =
2860                             dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2861                         ASSERT3U(db->db_blkid, <=,
2862                             dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2863                         ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2864                             db->db.db_size);
2865                         if (!arc_released(db->db_buf))
2866                                 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2867                 }
2868                 DB_DNODE_EXIT(db);
2869                 mutex_destroy(&dr->dt.di.dr_mtx);
2870                 list_destroy(&dr->dt.di.dr_children);
2871         }
2872         kmem_free(dr, sizeof (dbuf_dirty_record_t));
2873 
2874         cv_broadcast(&db->db_changed);
2875         ASSERT(db->db_dirtycnt > 0);
2876         db->db_dirtycnt -= 1;
2877         db->db_data_pending = NULL;
2878         dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2879 }
2880 
2881 static void
2882 dbuf_write_nofill_ready(zio_t *zio)
2883 {
2884         dbuf_write_ready(zio, NULL, zio->io_private);
2885 }
2886 
2887 static void
2888 dbuf_write_nofill_done(zio_t *zio)
2889 {
2890         dbuf_write_done(zio, NULL, zio->io_private);
2891 }
2892 
2893 static void
2894 dbuf_write_override_ready(zio_t *zio)
2895 {
2896         dbuf_dirty_record_t *dr = zio->io_private;
2897         dmu_buf_impl_t *db = dr->dr_dbuf;
2898 
2899         dbuf_write_ready(zio, NULL, db);
2900 }
2901 
2902 static void
2903 dbuf_write_override_done(zio_t *zio)
2904 {
2905         dbuf_dirty_record_t *dr = zio->io_private;
2906         dmu_buf_impl_t *db = dr->dr_dbuf;
2907         blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2908 
2909         mutex_enter(&db->db_mtx);
2910         if (!BP_EQUAL(zio->io_bp, obp)) {
2911                 if (!BP_IS_HOLE(obp))
2912                         dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2913                 arc_release(dr->dt.dl.dr_data, db);
2914         }
2915         mutex_exit(&db->db_mtx);
2916 
2917         dbuf_write_done(zio, NULL, db);
2918 }
2919 
2920 /* Issue I/O to commit a dirty buffer to disk. */
2921 static void
2922 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2923 {
2924         dmu_buf_impl_t *db = dr->dr_dbuf;
2925         dnode_t *dn;
2926         objset_t *os;
2927         dmu_buf_impl_t *parent = db->db_parent;
2928         uint64_t txg = tx->tx_txg;
2929         zbookmark_phys_t zb;
2930         zio_prop_t zp;
2931         zio_t *zio;
2932         int wp_flag = 0;
2933 
2934         ASSERT(dmu_tx_is_syncing(tx));
2935 
2936         DB_DNODE_ENTER(db);
2937         dn = DB_DNODE(db);
2938         os = dn->dn_objset;
2939 
2940         if (db->db_state != DB_NOFILL) {
2941                 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2942                         /*
2943                          * Private object buffers are released here rather
2944                          * than in dbuf_dirty() since they are only modified
2945                          * in the syncing context and we don't want the
2946                          * overhead of making multiple copies of the data.
2947                          */
2948                         if (BP_IS_HOLE(db->db_blkptr)) {
2949                                 arc_buf_thaw(data);
2950                         } else {
2951                                 dbuf_release_bp(db);
2952                         }
2953                 }
2954         }
2955 
2956         if (parent != dn->dn_dbuf) {
2957                 /* Our parent is an indirect block. */
2958                 /* We have a dirty parent that has been scheduled for write. */
2959                 ASSERT(parent && parent->db_data_pending);
2960                 /* Our parent's buffer is one level closer to the dnode. */
2961                 ASSERT(db->db_level == parent->db_level-1);
2962                 /*
2963                  * We're about to modify our parent's db_data by modifying
2964                  * our block pointer, so the parent must be released.
2965                  */
2966                 ASSERT(arc_released(parent->db_buf));
2967                 zio = parent->db_data_pending->dr_zio;
2968         } else {
2969                 /* Our parent is the dnode itself. */
2970                 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2971                     db->db_blkid != DMU_SPILL_BLKID) ||
2972                     (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2973                 if (db->db_blkid != DMU_SPILL_BLKID)
2974                         ASSERT3P(db->db_blkptr, ==,
2975                             &dn->dn_phys->dn_blkptr[db->db_blkid]);
2976                 zio = dn->dn_zio;
2977         }
2978 
2979         ASSERT(db->db_level == 0 || data == db->db_buf);
2980         ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2981         ASSERT(zio);
2982 
2983         SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2984             os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2985             db->db.db_object, db->db_level, db->db_blkid);
2986 
2987         if (db->db_blkid == DMU_SPILL_BLKID)
2988                 wp_flag = WP_SPILL;
2989         wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2990 
2991         dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2992         DB_DNODE_EXIT(db);
2993 
2994         /*
2995          * We copy the blkptr now (rather than when we instantiate the dirty
2996          * record), because its value can change between open context and
2997          * syncing context. We do not need to hold dn_struct_rwlock to read
2998          * db_blkptr because we are in syncing context.
2999          */
3000         dr->dr_bp_copy = *db->db_blkptr;
3001 
3002         if (db->db_level == 0 &&
3003             dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3004                 /*
3005                  * The BP for this block has been provided by open context
3006                  * (by dmu_sync() or dmu_buf_write_embedded()).
3007                  */
3008                 void *contents = (data != NULL) ? data->b_data : NULL;
3009 
3010                 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3011                     &dr->dr_bp_copy, contents, db->db.db_size, &zp,
3012                     dbuf_write_override_ready, NULL, dbuf_write_override_done,
3013                     dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3014                 mutex_enter(&db->db_mtx);
3015                 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3016                 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3017                     dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3018                 mutex_exit(&db->db_mtx);
3019         } else if (db->db_state == DB_NOFILL) {
3020                 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3021                     zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3022                 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3023                     &dr->dr_bp_copy, NULL, db->db.db_size, &zp,
3024                     dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
3025                     ZIO_PRIORITY_ASYNC_WRITE,
3026                     ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3027         } else {
3028                 ASSERT(arc_released(data));
3029                 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3030                     &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3031                     DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
3032                     dbuf_write_physdone, dbuf_write_done, db,
3033                     ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3034         }
3035 }