xref: /titanic_52/usr/src/uts/common/fs/zfs/dbuf.c (revision f48205be61a214698b763ff550ab9e657525104c)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/zfs_context.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dbuf.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 
40 static void dbuf_destroy(dmu_buf_impl_t *db);
41 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
42 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, int checksum,
43     int compress, dmu_tx_t *tx);
44 static arc_done_func_t dbuf_write_ready;
45 static arc_done_func_t dbuf_write_done;
46 
47 int zfs_mdcomp_disable = 0;
48 
49 /*
50  * Global data structures and functions for the dbuf cache.
51  */
52 static kmem_cache_t *dbuf_cache;
53 
54 /* ARGSUSED */
55 static int
56 dbuf_cons(void *vdb, void *unused, int kmflag)
57 {
58 	dmu_buf_impl_t *db = vdb;
59 	bzero(db, sizeof (dmu_buf_impl_t));
60 
61 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
62 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
63 	refcount_create(&db->db_holds);
64 	return (0);
65 }
66 
67 /* ARGSUSED */
68 static void
69 dbuf_dest(void *vdb, void *unused)
70 {
71 	dmu_buf_impl_t *db = vdb;
72 	mutex_destroy(&db->db_mtx);
73 	cv_destroy(&db->db_changed);
74 	refcount_destroy(&db->db_holds);
75 }
76 
77 /*
78  * dbuf hash table routines
79  */
80 static dbuf_hash_table_t dbuf_hash_table;
81 
82 static uint64_t dbuf_hash_count;
83 
84 static uint64_t
85 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
86 {
87 	uintptr_t osv = (uintptr_t)os;
88 	uint64_t crc = -1ULL;
89 
90 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
91 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
92 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
93 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
94 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
95 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
96 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
97 
98 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
99 
100 	return (crc);
101 }
102 
103 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
104 
105 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
106 	((dbuf)->db.db_object == (obj) &&		\
107 	(dbuf)->db_objset == (os) &&			\
108 	(dbuf)->db_level == (level) &&			\
109 	(dbuf)->db_blkid == (blkid))
110 
111 dmu_buf_impl_t *
112 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
113 {
114 	dbuf_hash_table_t *h = &dbuf_hash_table;
115 	objset_impl_t *os = dn->dn_objset;
116 	uint64_t obj = dn->dn_object;
117 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
118 	uint64_t idx = hv & h->hash_table_mask;
119 	dmu_buf_impl_t *db;
120 
121 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
122 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
123 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
124 			mutex_enter(&db->db_mtx);
125 			if (db->db_state != DB_EVICTING) {
126 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
127 				return (db);
128 			}
129 			mutex_exit(&db->db_mtx);
130 		}
131 	}
132 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
133 	return (NULL);
134 }
135 
136 /*
137  * Insert an entry into the hash table.  If there is already an element
138  * equal to elem in the hash table, then the already existing element
139  * will be returned and the new element will not be inserted.
140  * Otherwise returns NULL.
141  */
142 static dmu_buf_impl_t *
143 dbuf_hash_insert(dmu_buf_impl_t *db)
144 {
145 	dbuf_hash_table_t *h = &dbuf_hash_table;
146 	objset_impl_t *os = db->db_objset;
147 	uint64_t obj = db->db.db_object;
148 	int level = db->db_level;
149 	uint64_t blkid = db->db_blkid;
150 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
151 	uint64_t idx = hv & h->hash_table_mask;
152 	dmu_buf_impl_t *dbf;
153 
154 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
155 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
156 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
157 			mutex_enter(&dbf->db_mtx);
158 			if (dbf->db_state != DB_EVICTING) {
159 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
160 				return (dbf);
161 			}
162 			mutex_exit(&dbf->db_mtx);
163 		}
164 	}
165 
166 	mutex_enter(&db->db_mtx);
167 	db->db_hash_next = h->hash_table[idx];
168 	h->hash_table[idx] = db;
169 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
170 	atomic_add_64(&dbuf_hash_count, 1);
171 
172 	return (NULL);
173 }
174 
175 /*
176  * Remove an entry from the hash table.  This operation will
177  * fail if there are any existing holds on the db.
178  */
179 static void
180 dbuf_hash_remove(dmu_buf_impl_t *db)
181 {
182 	dbuf_hash_table_t *h = &dbuf_hash_table;
183 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
184 	    db->db_level, db->db_blkid);
185 	uint64_t idx = hv & h->hash_table_mask;
186 	dmu_buf_impl_t *dbf, **dbp;
187 
188 	/*
189 	 * We musn't hold db_mtx to maintin lock ordering:
190 	 * DBUF_HASH_MUTEX > db_mtx.
191 	 */
192 	ASSERT(refcount_is_zero(&db->db_holds));
193 	ASSERT(db->db_state == DB_EVICTING);
194 	ASSERT(!MUTEX_HELD(&db->db_mtx));
195 
196 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
197 	dbp = &h->hash_table[idx];
198 	while ((dbf = *dbp) != db) {
199 		dbp = &dbf->db_hash_next;
200 		ASSERT(dbf != NULL);
201 	}
202 	*dbp = db->db_hash_next;
203 	db->db_hash_next = NULL;
204 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
205 	atomic_add_64(&dbuf_hash_count, -1);
206 }
207 
208 static arc_evict_func_t dbuf_do_evict;
209 
210 static void
211 dbuf_evict_user(dmu_buf_impl_t *db)
212 {
213 	ASSERT(MUTEX_HELD(&db->db_mtx));
214 
215 	if (db->db_level != 0 || db->db_evict_func == NULL)
216 		return;
217 
218 	if (db->db_user_data_ptr_ptr)
219 		*db->db_user_data_ptr_ptr = db->db.db_data;
220 	db->db_evict_func(&db->db, db->db_user_ptr);
221 	db->db_user_ptr = NULL;
222 	db->db_user_data_ptr_ptr = NULL;
223 	db->db_evict_func = NULL;
224 }
225 
226 void
227 dbuf_evict(dmu_buf_impl_t *db)
228 {
229 	ASSERT(MUTEX_HELD(&db->db_mtx));
230 	ASSERT(db->db_buf == NULL);
231 	ASSERT(db->db_data_pending == NULL);
232 
233 	dbuf_clear(db);
234 	dbuf_destroy(db);
235 }
236 
237 void
238 dbuf_init(void)
239 {
240 	uint64_t hsize = 1ULL << 16;
241 	dbuf_hash_table_t *h = &dbuf_hash_table;
242 	int i;
243 
244 	/*
245 	 * The hash table is big enough to fill all of physical memory
246 	 * with an average 4K block size.  The table will take up
247 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
248 	 */
249 	while (hsize * 4096 < physmem * PAGESIZE)
250 		hsize <<= 1;
251 
252 retry:
253 	h->hash_table_mask = hsize - 1;
254 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
255 	if (h->hash_table == NULL) {
256 		/* XXX - we should really return an error instead of assert */
257 		ASSERT(hsize > (1ULL << 10));
258 		hsize >>= 1;
259 		goto retry;
260 	}
261 
262 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
263 	    sizeof (dmu_buf_impl_t),
264 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
265 
266 	for (i = 0; i < DBUF_MUTEXES; i++)
267 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
268 }
269 
270 void
271 dbuf_fini(void)
272 {
273 	dbuf_hash_table_t *h = &dbuf_hash_table;
274 	int i;
275 
276 	for (i = 0; i < DBUF_MUTEXES; i++)
277 		mutex_destroy(&h->hash_mutexes[i]);
278 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
279 	kmem_cache_destroy(dbuf_cache);
280 }
281 
282 /*
283  * Other stuff.
284  */
285 
286 #ifdef ZFS_DEBUG
287 static void
288 dbuf_verify(dmu_buf_impl_t *db)
289 {
290 	dnode_t *dn = db->db_dnode;
291 
292 	ASSERT(MUTEX_HELD(&db->db_mtx));
293 
294 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
295 		return;
296 
297 	ASSERT(db->db_objset != NULL);
298 	if (dn == NULL) {
299 		ASSERT(db->db_parent == NULL);
300 		ASSERT(db->db_blkptr == NULL);
301 	} else {
302 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
303 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
304 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
305 		ASSERT(db->db_blkid == DB_BONUS_BLKID ||
306 		    list_head(&dn->dn_dbufs));
307 	}
308 	if (db->db_blkid == DB_BONUS_BLKID) {
309 		ASSERT(dn != NULL);
310 		ASSERT3U(db->db.db_size, ==, dn->dn_bonuslen);
311 		ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID);
312 	} else {
313 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
314 	}
315 
316 	if (db->db_level == 0) {
317 		/* we can be momentarily larger in dnode_set_blksz() */
318 		if (db->db_blkid != DB_BONUS_BLKID && dn) {
319 			ASSERT3U(db->db.db_size, >=, dn->dn_datablksz);
320 		}
321 		if (db->db.db_object == DMU_META_DNODE_OBJECT) {
322 			dbuf_dirty_record_t *dr = db->db_data_pending;
323 			/*
324 			 * it should only be modified in syncing
325 			 * context, so make sure we only have
326 			 * one copy of the data.
327 			 */
328 			ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
329 		}
330 	}
331 
332 	/* verify db->db_blkptr */
333 	if (db->db_blkptr) {
334 		if (db->db_parent == dn->dn_dbuf) {
335 			/* db is pointed to by the dnode */
336 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
337 			if (db->db.db_object == DMU_META_DNODE_OBJECT)
338 				ASSERT(db->db_parent == NULL);
339 			else
340 				ASSERT(db->db_parent != NULL);
341 			ASSERT3P(db->db_blkptr, ==,
342 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
343 		} else {
344 			/* db is pointed to by an indirect block */
345 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
346 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
347 			ASSERT3U(db->db_parent->db.db_object, ==,
348 			    db->db.db_object);
349 			/*
350 			 * dnode_grow_indblksz() can make this fail if we don't
351 			 * have the struct_rwlock.  XXX indblksz no longer
352 			 * grows.  safe to do this now?
353 			 */
354 			if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) {
355 				ASSERT3P(db->db_blkptr, ==,
356 				    ((blkptr_t *)db->db_parent->db.db_data +
357 				    db->db_blkid % epb));
358 			}
359 		}
360 	}
361 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
362 	    db->db.db_data && db->db_blkid != DB_BONUS_BLKID &&
363 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
364 		/*
365 		 * If the blkptr isn't set but they have nonzero data,
366 		 * it had better be dirty, otherwise we'll lose that
367 		 * data when we evict this buffer.
368 		 */
369 		if (db->db_dirtycnt == 0) {
370 			uint64_t *buf = db->db.db_data;
371 			int i;
372 
373 			for (i = 0; i < db->db.db_size >> 3; i++) {
374 				ASSERT(buf[i] == 0);
375 			}
376 		}
377 	}
378 }
379 #endif
380 
381 static void
382 dbuf_update_data(dmu_buf_impl_t *db)
383 {
384 	ASSERT(MUTEX_HELD(&db->db_mtx));
385 	if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
386 		ASSERT(!refcount_is_zero(&db->db_holds));
387 		*db->db_user_data_ptr_ptr = db->db.db_data;
388 	}
389 }
390 
391 static void
392 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
393 {
394 	ASSERT(MUTEX_HELD(&db->db_mtx));
395 	ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
396 	db->db_buf = buf;
397 	if (buf != NULL) {
398 		ASSERT(buf->b_data != NULL);
399 		db->db.db_data = buf->b_data;
400 		if (!arc_released(buf))
401 			arc_set_callback(buf, dbuf_do_evict, db);
402 		dbuf_update_data(db);
403 	} else {
404 		dbuf_evict_user(db);
405 		db->db.db_data = NULL;
406 		db->db_state = DB_UNCACHED;
407 	}
408 }
409 
410 uint64_t
411 dbuf_whichblock(dnode_t *dn, uint64_t offset)
412 {
413 	if (dn->dn_datablkshift) {
414 		return (offset >> dn->dn_datablkshift);
415 	} else {
416 		ASSERT3U(offset, <, dn->dn_datablksz);
417 		return (0);
418 	}
419 }
420 
421 static void
422 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
423 {
424 	dmu_buf_impl_t *db = vdb;
425 
426 	mutex_enter(&db->db_mtx);
427 	ASSERT3U(db->db_state, ==, DB_READ);
428 	/*
429 	 * All reads are synchronous, so we must have a hold on the dbuf
430 	 */
431 	ASSERT(refcount_count(&db->db_holds) > 0);
432 	ASSERT(db->db_buf == NULL);
433 	ASSERT(db->db.db_data == NULL);
434 	if (db->db_level == 0 && db->db_freed_in_flight) {
435 		/* we were freed in flight; disregard any error */
436 		arc_release(buf, db);
437 		bzero(buf->b_data, db->db.db_size);
438 		arc_buf_freeze(buf);
439 		db->db_freed_in_flight = FALSE;
440 		dbuf_set_data(db, buf);
441 		db->db_state = DB_CACHED;
442 	} else if (zio == NULL || zio->io_error == 0) {
443 		dbuf_set_data(db, buf);
444 		db->db_state = DB_CACHED;
445 	} else {
446 		ASSERT(db->db_blkid != DB_BONUS_BLKID);
447 		ASSERT3P(db->db_buf, ==, NULL);
448 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
449 		db->db_state = DB_UNCACHED;
450 	}
451 	cv_broadcast(&db->db_changed);
452 	mutex_exit(&db->db_mtx);
453 	dbuf_rele(db, NULL);
454 }
455 
456 static void
457 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
458 {
459 	blkptr_t *bp;
460 	zbookmark_t zb;
461 	uint32_t aflags = ARC_NOWAIT;
462 
463 	ASSERT(!refcount_is_zero(&db->db_holds));
464 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
465 	ASSERT(RW_LOCK_HELD(&db->db_dnode->dn_struct_rwlock));
466 	ASSERT(MUTEX_HELD(&db->db_mtx));
467 	ASSERT(db->db_state == DB_UNCACHED);
468 	ASSERT(db->db_buf == NULL);
469 
470 	if (db->db_blkid == DB_BONUS_BLKID) {
471 		ASSERT3U(db->db_dnode->dn_bonuslen, ==, db->db.db_size);
472 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
473 		arc_space_consume(512);
474 		if (db->db.db_size < DN_MAX_BONUSLEN)
475 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
476 		bcopy(DN_BONUS(db->db_dnode->dn_phys), db->db.db_data,
477 		    db->db.db_size);
478 		dbuf_update_data(db);
479 		db->db_state = DB_CACHED;
480 		mutex_exit(&db->db_mtx);
481 		return;
482 	}
483 
484 	if (db->db_level == 0 && dnode_block_freed(db->db_dnode, db->db_blkid))
485 		bp = NULL;
486 	else
487 		bp = db->db_blkptr;
488 
489 	if (bp == NULL)
490 		dprintf_dbuf(db, "blkptr: %s\n", "NULL");
491 	else
492 		dprintf_dbuf_bp(db, bp, "%s", "blkptr:");
493 
494 	if (bp == NULL || BP_IS_HOLE(bp)) {
495 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
496 
497 		ASSERT(bp == NULL || BP_IS_HOLE(bp));
498 		dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
499 		    db->db.db_size, db, type));
500 		bzero(db->db.db_data, db->db.db_size);
501 		db->db_state = DB_CACHED;
502 		*flags |= DB_RF_CACHED;
503 		mutex_exit(&db->db_mtx);
504 		return;
505 	}
506 
507 	db->db_state = DB_READ;
508 	mutex_exit(&db->db_mtx);
509 
510 	zb.zb_objset = db->db_objset->os_dsl_dataset ?
511 	    db->db_objset->os_dsl_dataset->ds_object : 0;
512 	zb.zb_object = db->db.db_object;
513 	zb.zb_level = db->db_level;
514 	zb.zb_blkid = db->db_blkid;
515 
516 	dbuf_add_ref(db, NULL);
517 	/* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
518 	ASSERT3U(db->db_dnode->dn_type, <, DMU_OT_NUMTYPES);
519 	(void) arc_read(zio, db->db_dnode->dn_objset->os_spa, bp,
520 	    db->db_level > 0 ? byteswap_uint64_array :
521 	    dmu_ot[db->db_dnode->dn_type].ot_byteswap,
522 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
523 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
524 	    &aflags, &zb);
525 	if (aflags & ARC_CACHED)
526 		*flags |= DB_RF_CACHED;
527 }
528 
529 int
530 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
531 {
532 	int err = 0;
533 	int havepzio = (zio != NULL);
534 	int prefetch;
535 
536 	/*
537 	 * We don't have to hold the mutex to check db_state because it
538 	 * can't be freed while we have a hold on the buffer.
539 	 */
540 	ASSERT(!refcount_is_zero(&db->db_holds));
541 
542 	if ((flags & DB_RF_HAVESTRUCT) == 0)
543 		rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER);
544 
545 	prefetch = db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID &&
546 	    (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL;
547 
548 	mutex_enter(&db->db_mtx);
549 	if (db->db_state == DB_CACHED) {
550 		mutex_exit(&db->db_mtx);
551 		if (prefetch)
552 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
553 			    db->db.db_size, TRUE);
554 		if ((flags & DB_RF_HAVESTRUCT) == 0)
555 			rw_exit(&db->db_dnode->dn_struct_rwlock);
556 	} else if (db->db_state == DB_UNCACHED) {
557 		if (zio == NULL) {
558 			zio = zio_root(db->db_dnode->dn_objset->os_spa,
559 			    NULL, NULL, ZIO_FLAG_CANFAIL);
560 		}
561 		dbuf_read_impl(db, zio, &flags);
562 
563 		/* dbuf_read_impl has dropped db_mtx for us */
564 
565 		if (prefetch)
566 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
567 			    db->db.db_size, flags & DB_RF_CACHED);
568 
569 		if ((flags & DB_RF_HAVESTRUCT) == 0)
570 			rw_exit(&db->db_dnode->dn_struct_rwlock);
571 
572 		if (!havepzio)
573 			err = zio_wait(zio);
574 	} else {
575 		mutex_exit(&db->db_mtx);
576 		if (prefetch)
577 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
578 			    db->db.db_size, TRUE);
579 		if ((flags & DB_RF_HAVESTRUCT) == 0)
580 			rw_exit(&db->db_dnode->dn_struct_rwlock);
581 
582 		mutex_enter(&db->db_mtx);
583 		if ((flags & DB_RF_NEVERWAIT) == 0) {
584 			while (db->db_state == DB_READ ||
585 			    db->db_state == DB_FILL) {
586 				ASSERT(db->db_state == DB_READ ||
587 				    (flags & DB_RF_HAVESTRUCT) == 0);
588 				cv_wait(&db->db_changed, &db->db_mtx);
589 			}
590 			if (db->db_state == DB_UNCACHED)
591 				err = EIO;
592 		}
593 		mutex_exit(&db->db_mtx);
594 	}
595 
596 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
597 	return (err);
598 }
599 
600 static void
601 dbuf_noread(dmu_buf_impl_t *db)
602 {
603 	ASSERT(!refcount_is_zero(&db->db_holds));
604 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
605 	mutex_enter(&db->db_mtx);
606 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
607 		cv_wait(&db->db_changed, &db->db_mtx);
608 	if (db->db_state == DB_UNCACHED) {
609 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
610 
611 		ASSERT(db->db_buf == NULL);
612 		ASSERT(db->db.db_data == NULL);
613 		dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
614 		    db->db.db_size, db, type));
615 		db->db_state = DB_FILL;
616 	} else {
617 		ASSERT3U(db->db_state, ==, DB_CACHED);
618 	}
619 	mutex_exit(&db->db_mtx);
620 }
621 
622 /*
623  * This is our just-in-time copy function.  It makes a copy of
624  * buffers, that have been modified in a previous transaction
625  * group, before we modify them in the current active group.
626  *
627  * This function is used in two places: when we are dirtying a
628  * buffer for the first time in a txg, and when we are freeing
629  * a range in a dnode that includes this buffer.
630  *
631  * Note that when we are called from dbuf_free_range() we do
632  * not put a hold on the buffer, we just traverse the active
633  * dbuf list for the dnode.
634  */
635 static void
636 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
637 {
638 	dbuf_dirty_record_t *dr = db->db_last_dirty;
639 
640 	ASSERT(MUTEX_HELD(&db->db_mtx));
641 	ASSERT(db->db.db_data != NULL);
642 	ASSERT(db->db_level == 0);
643 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
644 
645 	if (dr == NULL ||
646 	    (dr->dt.dl.dr_data !=
647 	    ((db->db_blkid  == DB_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
648 		return;
649 
650 	/*
651 	 * If the last dirty record for this dbuf has not yet synced
652 	 * and its referencing the dbuf data, either:
653 	 * 	reset the reference to point to a new copy,
654 	 * or (if there a no active holders)
655 	 *	just null out the current db_data pointer.
656 	 */
657 	ASSERT(dr->dr_txg >= txg - 2);
658 	if (db->db_blkid == DB_BONUS_BLKID) {
659 		/* Note that the data bufs here are zio_bufs */
660 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
661 		arc_space_consume(DN_MAX_BONUSLEN);
662 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
663 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
664 		int size = db->db.db_size;
665 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
666 		dr->dt.dl.dr_data = arc_buf_alloc(
667 		    db->db_dnode->dn_objset->os_spa, size, db, type);
668 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
669 	} else {
670 		dbuf_set_data(db, NULL);
671 	}
672 }
673 
674 void
675 dbuf_unoverride(dbuf_dirty_record_t *dr)
676 {
677 	dmu_buf_impl_t *db = dr->dr_dbuf;
678 	uint64_t txg = dr->dr_txg;
679 
680 	ASSERT(MUTEX_HELD(&db->db_mtx));
681 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
682 	ASSERT(db->db_level == 0);
683 
684 	if (db->db_blkid == DB_BONUS_BLKID ||
685 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
686 		return;
687 
688 	/* free this block */
689 	if (!BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) {
690 		/* XXX can get silent EIO here */
691 		(void) arc_free(NULL, db->db_dnode->dn_objset->os_spa,
692 		    txg, &dr->dt.dl.dr_overridden_by, NULL, NULL, ARC_WAIT);
693 	}
694 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
695 	/*
696 	 * Release the already-written buffer, so we leave it in
697 	 * a consistent dirty state.  Note that all callers are
698 	 * modifying the buffer, so they will immediately do
699 	 * another (redundant) arc_release().  Therefore, leave
700 	 * the buf thawed to save the effort of freezing &
701 	 * immediately re-thawing it.
702 	 */
703 	arc_release(dr->dt.dl.dr_data, db);
704 }
705 
706 void
707 dbuf_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
708 {
709 	dmu_buf_impl_t *db, *db_next;
710 	uint64_t txg = tx->tx_txg;
711 
712 	dprintf_dnode(dn, "blkid=%llu nblks=%llu\n", blkid, nblks);
713 	mutex_enter(&dn->dn_dbufs_mtx);
714 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
715 		db_next = list_next(&dn->dn_dbufs, db);
716 		ASSERT(db->db_blkid != DB_BONUS_BLKID);
717 		if (db->db_level != 0)
718 			continue;
719 		dprintf_dbuf(db, "found buf %s\n", "");
720 		if (db->db_blkid < blkid ||
721 		    db->db_blkid >= blkid+nblks)
722 			continue;
723 
724 		/* found a level 0 buffer in the range */
725 		if (dbuf_undirty(db, tx))
726 			continue;
727 
728 		mutex_enter(&db->db_mtx);
729 		if (db->db_state == DB_UNCACHED ||
730 		    db->db_state == DB_EVICTING) {
731 			ASSERT(db->db.db_data == NULL);
732 			mutex_exit(&db->db_mtx);
733 			continue;
734 		}
735 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
736 			/* will be handled in dbuf_read_done or dbuf_rele */
737 			db->db_freed_in_flight = TRUE;
738 			mutex_exit(&db->db_mtx);
739 			continue;
740 		}
741 		if (refcount_count(&db->db_holds) == 0) {
742 			ASSERT(db->db_buf);
743 			dbuf_clear(db);
744 			continue;
745 		}
746 		/* The dbuf is referenced */
747 
748 		if (db->db_last_dirty != NULL) {
749 			dbuf_dirty_record_t *dr = db->db_last_dirty;
750 
751 			if (dr->dr_txg == txg) {
752 				/*
753 				 * This buffer is "in-use", re-adjust the file
754 				 * size to reflect that this buffer may
755 				 * contain new data when we sync.
756 				 */
757 				if (db->db_blkid > dn->dn_maxblkid)
758 					dn->dn_maxblkid = db->db_blkid;
759 				dbuf_unoverride(dr);
760 			} else {
761 				/*
762 				 * This dbuf is not dirty in the open context.
763 				 * Either uncache it (if its not referenced in
764 				 * the open context) or reset its contents to
765 				 * empty.
766 				 */
767 				dbuf_fix_old_data(db, txg);
768 			}
769 		}
770 		/* clear the contents if its cached */
771 		if (db->db_state == DB_CACHED) {
772 			ASSERT(db->db.db_data != NULL);
773 			arc_release(db->db_buf, db);
774 			bzero(db->db.db_data, db->db.db_size);
775 			arc_buf_freeze(db->db_buf);
776 		}
777 
778 		mutex_exit(&db->db_mtx);
779 	}
780 	mutex_exit(&dn->dn_dbufs_mtx);
781 }
782 
783 static int
784 dbuf_new_block(dmu_buf_impl_t *db)
785 {
786 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
787 	uint64_t birth_txg = 0;
788 
789 	/* Don't count meta-objects */
790 	if (ds == NULL)
791 		return (FALSE);
792 
793 	/*
794 	 * We don't need any locking to protect db_blkptr:
795 	 * If it's syncing, then db_last_dirty will be set
796 	 * so we'll ignore db_blkptr.
797 	 */
798 	ASSERT(MUTEX_HELD(&db->db_mtx));
799 	/* If we have been dirtied since the last snapshot, its not new */
800 	if (db->db_last_dirty)
801 		birth_txg = db->db_last_dirty->dr_txg;
802 	else if (db->db_blkptr)
803 		birth_txg = db->db_blkptr->blk_birth;
804 
805 	if (birth_txg)
806 		return (!dsl_dataset_block_freeable(ds, birth_txg));
807 	else
808 		return (TRUE);
809 }
810 
811 void
812 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
813 {
814 	arc_buf_t *buf, *obuf;
815 	int osize = db->db.db_size;
816 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
817 
818 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
819 
820 	/* XXX does *this* func really need the lock? */
821 	ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock));
822 
823 	/*
824 	 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
825 	 * is OK, because there can be no other references to the db
826 	 * when we are changing its size, so no concurrent DB_FILL can
827 	 * be happening.
828 	 */
829 	/*
830 	 * XXX we should be doing a dbuf_read, checking the return
831 	 * value and returning that up to our callers
832 	 */
833 	dbuf_will_dirty(db, tx);
834 
835 	/* create the data buffer for the new block */
836 	buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type);
837 
838 	/* copy old block data to the new block */
839 	obuf = db->db_buf;
840 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
841 	/* zero the remainder */
842 	if (size > osize)
843 		bzero((uint8_t *)buf->b_data + osize, size - osize);
844 
845 	mutex_enter(&db->db_mtx);
846 	dbuf_set_data(db, buf);
847 	VERIFY(arc_buf_remove_ref(obuf, db) == 1);
848 	db->db.db_size = size;
849 
850 	if (db->db_level == 0) {
851 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
852 		db->db_last_dirty->dt.dl.dr_data = buf;
853 	}
854 	mutex_exit(&db->db_mtx);
855 
856 	dnode_willuse_space(db->db_dnode, size-osize, tx);
857 }
858 
859 dbuf_dirty_record_t *
860 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
861 {
862 	dnode_t *dn = db->db_dnode;
863 	objset_impl_t *os = dn->dn_objset;
864 	dbuf_dirty_record_t **drp, *dr;
865 	int drop_struct_lock = FALSE;
866 	int txgoff = tx->tx_txg & TXG_MASK;
867 
868 	ASSERT(tx->tx_txg != 0);
869 	ASSERT(!refcount_is_zero(&db->db_holds));
870 	DMU_TX_DIRTY_BUF(tx, db);
871 
872 	/*
873 	 * Shouldn't dirty a regular buffer in syncing context.  Private
874 	 * objects may be dirtied in syncing context, but only if they
875 	 * were already pre-dirtied in open context.
876 	 * XXX We may want to prohibit dirtying in syncing context even
877 	 * if they did pre-dirty.
878 	 */
879 	ASSERT(!dmu_tx_is_syncing(tx) ||
880 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
881 	    dn->dn_object == DMU_META_DNODE_OBJECT ||
882 	    dn->dn_objset->os_dsl_dataset == NULL ||
883 	    dsl_dir_is_private(dn->dn_objset->os_dsl_dataset->ds_dir));
884 
885 	/*
886 	 * We make this assert for private objects as well, but after we
887 	 * check if we're already dirty.  They are allowed to re-dirty
888 	 * in syncing context.
889 	 */
890 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
891 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
892 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
893 
894 	mutex_enter(&db->db_mtx);
895 	/*
896 	 * XXX make this true for indirects too?  The problem is that
897 	 * transactions created with dmu_tx_create_assigned() from
898 	 * syncing context don't bother holding ahead.
899 	 */
900 	ASSERT(db->db_level != 0 ||
901 	    db->db_state == DB_CACHED || db->db_state == DB_FILL);
902 
903 	mutex_enter(&dn->dn_mtx);
904 	/*
905 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
906 	 * initialize the objset.
907 	 */
908 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
909 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
910 		dn->dn_dirtyctx =
911 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
912 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
913 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
914 	}
915 	mutex_exit(&dn->dn_mtx);
916 
917 	/*
918 	 * If this buffer is already dirty, we're done.
919 	 */
920 	drp = &db->db_last_dirty;
921 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
922 	    db->db.db_object == DMU_META_DNODE_OBJECT);
923 	while (*drp && (*drp)->dr_txg > tx->tx_txg)
924 		drp = &(*drp)->dr_next;
925 	if (*drp && (*drp)->dr_txg == tx->tx_txg) {
926 		if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) {
927 			/*
928 			 * If this buffer has already been written out,
929 			 * we now need to reset its state.
930 			 */
931 			dbuf_unoverride(*drp);
932 			if (db->db.db_object != DMU_META_DNODE_OBJECT)
933 				arc_buf_thaw(db->db_buf);
934 		}
935 		mutex_exit(&db->db_mtx);
936 		return (*drp);
937 	}
938 
939 	/*
940 	 * Only valid if not already dirty.
941 	 */
942 	ASSERT(dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
943 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
944 
945 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
946 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
947 	    dn->dn_phys->dn_nlevels > db->db_level ||
948 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
949 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
950 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
951 
952 	/*
953 	 * We should only be dirtying in syncing context if it's the
954 	 * mos, a spa os, or we're initializing the os.  However, we are
955 	 * allowed to dirty in syncing context provided we already
956 	 * dirtied it in open context.  Hence we must make this
957 	 * assertion only if we're not already dirty.
958 	 */
959 	ASSERT(!dmu_tx_is_syncing(tx) ||
960 	    os->os_dsl_dataset == NULL ||
961 	    !dsl_dir_is_private(os->os_dsl_dataset->ds_dir) ||
962 	    !BP_IS_HOLE(os->os_rootbp));
963 	ASSERT(db->db.db_size != 0);
964 
965 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
966 
967 	/*
968 	 * If this buffer is dirty in an old transaction group we need
969 	 * to make a copy of it so that the changes we make in this
970 	 * transaction group won't leak out when we sync the older txg.
971 	 */
972 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
973 	if (db->db_level == 0) {
974 		void *data_old = db->db_buf;
975 
976 		if (db->db_blkid == DB_BONUS_BLKID) {
977 			dbuf_fix_old_data(db, tx->tx_txg);
978 			data_old = db->db.db_data;
979 		} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
980 			/*
981 			 * Release the data buffer from the cache so that we
982 			 * can modify it without impacting possible other users
983 			 * of this cached data block.  Note that indirect
984 			 * blocks and private objects are not released until the
985 			 * syncing state (since they are only modified then).
986 			 */
987 			arc_release(db->db_buf, db);
988 			dbuf_fix_old_data(db, tx->tx_txg);
989 			data_old = db->db_buf;
990 		}
991 		ASSERT(data_old != NULL);
992 		dr->dt.dl.dr_data = data_old;
993 	} else {
994 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
995 		list_create(&dr->dt.di.dr_children,
996 		    sizeof (dbuf_dirty_record_t),
997 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
998 	}
999 	dr->dr_dbuf = db;
1000 	dr->dr_txg = tx->tx_txg;
1001 	dr->dr_next = *drp;
1002 	*drp = dr;
1003 
1004 	/*
1005 	 * We could have been freed_in_flight between the dbuf_noread
1006 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1007 	 * happened after the free.
1008 	 */
1009 	if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) {
1010 		mutex_enter(&dn->dn_mtx);
1011 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1012 		mutex_exit(&dn->dn_mtx);
1013 		db->db_freed_in_flight = FALSE;
1014 	}
1015 
1016 	if (db->db_blkid != DB_BONUS_BLKID) {
1017 		/*
1018 		 * Update the accounting.
1019 		 */
1020 		if (!dbuf_new_block(db) && db->db_blkptr) {
1021 			/*
1022 			 * This is only a guess -- if the dbuf is dirty
1023 			 * in a previous txg, we don't know how much
1024 			 * space it will use on disk yet.  We should
1025 			 * really have the struct_rwlock to access
1026 			 * db_blkptr, but since this is just a guess,
1027 			 * it's OK if we get an odd answer.
1028 			 */
1029 			dnode_willuse_space(dn,
1030 			    -bp_get_dasize(os->os_spa, db->db_blkptr), tx);
1031 		}
1032 		dnode_willuse_space(dn, db->db.db_size, tx);
1033 	}
1034 
1035 	/*
1036 	 * This buffer is now part of this txg
1037 	 */
1038 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1039 	db->db_dirtycnt += 1;
1040 	ASSERT3U(db->db_dirtycnt, <=, 3);
1041 
1042 	mutex_exit(&db->db_mtx);
1043 
1044 	if (db->db_blkid == DB_BONUS_BLKID) {
1045 		mutex_enter(&dn->dn_mtx);
1046 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1047 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1048 		mutex_exit(&dn->dn_mtx);
1049 		dnode_setdirty(dn, tx);
1050 		return (dr);
1051 	}
1052 
1053 	if (db->db_level == 0) {
1054 		dnode_new_blkid(dn, db->db_blkid, tx);
1055 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1056 	}
1057 
1058 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1059 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1060 		drop_struct_lock = TRUE;
1061 	}
1062 
1063 	if (db->db_level+1 < dn->dn_nlevels) {
1064 		dmu_buf_impl_t *parent = db->db_parent;
1065 		dbuf_dirty_record_t *di;
1066 		int parent_held = FALSE;
1067 
1068 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1069 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1070 
1071 			parent = dbuf_hold_level(dn, db->db_level+1,
1072 			    db->db_blkid >> epbs, FTAG);
1073 			parent_held = TRUE;
1074 		}
1075 		if (drop_struct_lock)
1076 			rw_exit(&dn->dn_struct_rwlock);
1077 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1078 		di = dbuf_dirty(parent, tx);
1079 		if (parent_held)
1080 			dbuf_rele(parent, FTAG);
1081 
1082 		mutex_enter(&db->db_mtx);
1083 		/*  possible race with dbuf_undirty() */
1084 		if (db->db_last_dirty == dr ||
1085 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1086 			mutex_enter(&di->dt.di.dr_mtx);
1087 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1088 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1089 			list_insert_tail(&di->dt.di.dr_children, dr);
1090 			mutex_exit(&di->dt.di.dr_mtx);
1091 			dr->dr_parent = di;
1092 		}
1093 		mutex_exit(&db->db_mtx);
1094 	} else {
1095 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1096 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1097 		ASSERT(db->db_parent == NULL ||
1098 		    db->db_parent == db->db_dnode->dn_dbuf);
1099 		mutex_enter(&dn->dn_mtx);
1100 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1101 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1102 		mutex_exit(&dn->dn_mtx);
1103 		if (drop_struct_lock)
1104 			rw_exit(&dn->dn_struct_rwlock);
1105 	}
1106 
1107 	dnode_setdirty(dn, tx);
1108 	return (dr);
1109 }
1110 
1111 static int
1112 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1113 {
1114 	dnode_t *dn = db->db_dnode;
1115 	uint64_t txg = tx->tx_txg;
1116 	dbuf_dirty_record_t *dr;
1117 
1118 	ASSERT(txg != 0);
1119 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
1120 
1121 	mutex_enter(&db->db_mtx);
1122 
1123 	/*
1124 	 * If this buffer is not dirty, we're done.
1125 	 */
1126 	for (dr = db->db_last_dirty; dr; dr = dr->dr_next)
1127 		if (dr->dr_txg <= txg)
1128 			break;
1129 	if (dr == NULL || dr->dr_txg < txg) {
1130 		mutex_exit(&db->db_mtx);
1131 		return (0);
1132 	}
1133 	ASSERT(dr->dr_txg == txg);
1134 
1135 	/*
1136 	 * If this buffer is currently held, we cannot undirty
1137 	 * it, since one of the current holders may be in the
1138 	 * middle of an update.  Note that users of dbuf_undirty()
1139 	 * should not place a hold on the dbuf before the call.
1140 	 */
1141 	if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1142 		mutex_exit(&db->db_mtx);
1143 		/* Make sure we don't toss this buffer at sync phase */
1144 		mutex_enter(&dn->dn_mtx);
1145 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1146 		mutex_exit(&dn->dn_mtx);
1147 		return (0);
1148 	}
1149 
1150 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1151 
1152 	ASSERT(db->db.db_size != 0);
1153 
1154 	/* XXX would be nice to fix up dn_towrite_space[] */
1155 
1156 	db->db_last_dirty = dr->dr_next;
1157 
1158 	if (dr->dr_parent) {
1159 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1160 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1161 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1162 	} else if (db->db_level+1 == dn->dn_nlevels) {
1163 		ASSERT3P(db->db_parent, ==, dn->dn_dbuf);
1164 		mutex_enter(&dn->dn_mtx);
1165 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1166 		mutex_exit(&dn->dn_mtx);
1167 	}
1168 
1169 	if (db->db_level == 0) {
1170 		dbuf_unoverride(dr);
1171 
1172 		ASSERT(db->db_buf != NULL);
1173 		ASSERT(dr->dt.dl.dr_data != NULL);
1174 		if (dr->dt.dl.dr_data != db->db_buf)
1175 			VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1);
1176 	} else {
1177 		ASSERT(db->db_buf != NULL);
1178 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1179 		/* XXX - mutex and list destroy? */
1180 	}
1181 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1182 
1183 	ASSERT(db->db_dirtycnt > 0);
1184 	db->db_dirtycnt -= 1;
1185 
1186 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1187 		arc_buf_t *buf = db->db_buf;
1188 
1189 		ASSERT(arc_released(buf));
1190 		dbuf_set_data(db, NULL);
1191 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
1192 		dbuf_evict(db);
1193 		return (1);
1194 	}
1195 
1196 	mutex_exit(&db->db_mtx);
1197 	return (0);
1198 }
1199 
1200 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1201 void
1202 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1203 {
1204 	int rf = DB_RF_MUST_SUCCEED;
1205 
1206 	ASSERT(tx->tx_txg != 0);
1207 	ASSERT(!refcount_is_zero(&db->db_holds));
1208 
1209 	if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock))
1210 		rf |= DB_RF_HAVESTRUCT;
1211 	(void) dbuf_read(db, NULL, rf);
1212 	(void) dbuf_dirty(db, tx);
1213 }
1214 
1215 void
1216 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1217 {
1218 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1219 
1220 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
1221 	ASSERT(tx->tx_txg != 0);
1222 	ASSERT(db->db_level == 0);
1223 	ASSERT(!refcount_is_zero(&db->db_holds));
1224 
1225 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1226 	    dmu_tx_private_ok(tx));
1227 
1228 	dbuf_noread(db);
1229 	(void) dbuf_dirty(db, tx);
1230 }
1231 
1232 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1233 /* ARGSUSED */
1234 void
1235 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1236 {
1237 	mutex_enter(&db->db_mtx);
1238 	DBUF_VERIFY(db);
1239 
1240 	if (db->db_state == DB_FILL) {
1241 		if (db->db_level == 0 && db->db_freed_in_flight) {
1242 			ASSERT(db->db_blkid != DB_BONUS_BLKID);
1243 			/* we were freed while filling */
1244 			/* XXX dbuf_undirty? */
1245 			bzero(db->db.db_data, db->db.db_size);
1246 			db->db_freed_in_flight = FALSE;
1247 		}
1248 		db->db_state = DB_CACHED;
1249 		cv_broadcast(&db->db_changed);
1250 	}
1251 	mutex_exit(&db->db_mtx);
1252 }
1253 
1254 /*
1255  * "Clear" the contents of this dbuf.  This will mark the dbuf
1256  * EVICTING and clear *most* of its references.  Unfortunetely,
1257  * when we are not holding the dn_dbufs_mtx, we can't clear the
1258  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1259  * in this case.  For callers from the DMU we will usually see:
1260  *	dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1261  * For the arc callback, we will usually see:
1262  * 	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1263  * Sometimes, though, we will get a mix of these two:
1264  *	DMU: dbuf_clear()->arc_buf_evict()
1265  *	ARC: dbuf_do_evict()->dbuf_destroy()
1266  */
1267 void
1268 dbuf_clear(dmu_buf_impl_t *db)
1269 {
1270 	dnode_t *dn = db->db_dnode;
1271 	dmu_buf_impl_t *parent = db->db_parent;
1272 	dmu_buf_impl_t *dndb = dn->dn_dbuf;
1273 	int dbuf_gone = FALSE;
1274 
1275 	ASSERT(MUTEX_HELD(&db->db_mtx));
1276 	ASSERT(refcount_is_zero(&db->db_holds));
1277 
1278 	dbuf_evict_user(db);
1279 
1280 	if (db->db_state == DB_CACHED) {
1281 		ASSERT(db->db.db_data != NULL);
1282 		if (db->db_blkid == DB_BONUS_BLKID) {
1283 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1284 			arc_space_return(DN_MAX_BONUSLEN);
1285 		}
1286 		db->db.db_data = NULL;
1287 		db->db_state = DB_UNCACHED;
1288 	}
1289 
1290 	ASSERT3U(db->db_state, ==, DB_UNCACHED);
1291 	ASSERT(db->db_data_pending == NULL);
1292 
1293 	db->db_state = DB_EVICTING;
1294 	db->db_blkptr = NULL;
1295 
1296 	if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1297 		list_remove(&dn->dn_dbufs, db);
1298 		dnode_rele(dn, db);
1299 	}
1300 
1301 	if (db->db_buf)
1302 		dbuf_gone = arc_buf_evict(db->db_buf);
1303 
1304 	if (!dbuf_gone)
1305 		mutex_exit(&db->db_mtx);
1306 
1307 	/*
1308 	 * If this dbuf is referened from an indirect dbuf,
1309 	 * decrement the ref count on the indirect dbuf.
1310 	 */
1311 	if (parent && parent != dndb)
1312 		dbuf_rele(parent, db);
1313 }
1314 
1315 static int
1316 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1317     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1318 {
1319 	int nlevels, epbs;
1320 
1321 	*parentp = NULL;
1322 	*bpp = NULL;
1323 
1324 	ASSERT(blkid != DB_BONUS_BLKID);
1325 
1326 	if (dn->dn_phys->dn_nlevels == 0)
1327 		nlevels = 1;
1328 	else
1329 		nlevels = dn->dn_phys->dn_nlevels;
1330 
1331 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1332 
1333 	ASSERT3U(level * epbs, <, 64);
1334 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1335 	if (level >= nlevels ||
1336 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1337 		/* the buffer has no parent yet */
1338 		return (ENOENT);
1339 	} else if (level < nlevels-1) {
1340 		/* this block is referenced from an indirect block */
1341 		int err = dbuf_hold_impl(dn, level+1,
1342 		    blkid >> epbs, fail_sparse, NULL, parentp);
1343 		if (err)
1344 			return (err);
1345 		err = dbuf_read(*parentp, NULL,
1346 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1347 		if (err) {
1348 			dbuf_rele(*parentp, NULL);
1349 			*parentp = NULL;
1350 			return (err);
1351 		}
1352 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1353 		    (blkid & ((1ULL << epbs) - 1));
1354 		return (0);
1355 	} else {
1356 		/* the block is referenced from the dnode */
1357 		ASSERT3U(level, ==, nlevels-1);
1358 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1359 		    blkid < dn->dn_phys->dn_nblkptr);
1360 		if (dn->dn_dbuf) {
1361 			dbuf_add_ref(dn->dn_dbuf, NULL);
1362 			*parentp = dn->dn_dbuf;
1363 		}
1364 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1365 		return (0);
1366 	}
1367 }
1368 
1369 static dmu_buf_impl_t *
1370 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1371     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1372 {
1373 	objset_impl_t *os = dn->dn_objset;
1374 	dmu_buf_impl_t *db, *odb;
1375 
1376 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1377 	ASSERT(dn->dn_type != DMU_OT_NONE);
1378 
1379 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1380 
1381 	db->db_objset = os;
1382 	db->db.db_object = dn->dn_object;
1383 	db->db_level = level;
1384 	db->db_blkid = blkid;
1385 	db->db_last_dirty = NULL;
1386 	db->db_dirtycnt = 0;
1387 	db->db_dnode = dn;
1388 	db->db_parent = parent;
1389 	db->db_blkptr = blkptr;
1390 
1391 	db->db_user_ptr = NULL;
1392 	db->db_user_data_ptr_ptr = NULL;
1393 	db->db_evict_func = NULL;
1394 	db->db_immediate_evict = 0;
1395 	db->db_freed_in_flight = 0;
1396 
1397 	if (blkid == DB_BONUS_BLKID) {
1398 		ASSERT3P(parent, ==, dn->dn_dbuf);
1399 		db->db.db_size = dn->dn_bonuslen;
1400 		db->db.db_offset = DB_BONUS_BLKID;
1401 		db->db_state = DB_UNCACHED;
1402 		/* the bonus dbuf is not placed in the hash table */
1403 		arc_space_consume(sizeof (dmu_buf_impl_t));
1404 		return (db);
1405 	} else {
1406 		int blocksize =
1407 		    db->db_level ? 1<<dn->dn_indblkshift :  dn->dn_datablksz;
1408 		db->db.db_size = blocksize;
1409 		db->db.db_offset = db->db_blkid * blocksize;
1410 	}
1411 
1412 	/*
1413 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1414 	 * in the hash table *and* added to the dbufs list.
1415 	 * This prevents a possible deadlock with someone
1416 	 * trying to look up this dbuf before its added to the
1417 	 * dn_dbufs list.
1418 	 */
1419 	mutex_enter(&dn->dn_dbufs_mtx);
1420 	db->db_state = DB_EVICTING;
1421 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1422 		/* someone else inserted it first */
1423 		kmem_cache_free(dbuf_cache, db);
1424 		mutex_exit(&dn->dn_dbufs_mtx);
1425 		return (odb);
1426 	}
1427 	list_insert_head(&dn->dn_dbufs, db);
1428 	db->db_state = DB_UNCACHED;
1429 	mutex_exit(&dn->dn_dbufs_mtx);
1430 	arc_space_consume(sizeof (dmu_buf_impl_t));
1431 
1432 	if (parent && parent != dn->dn_dbuf)
1433 		dbuf_add_ref(parent, db);
1434 
1435 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1436 	    refcount_count(&dn->dn_holds) > 0);
1437 	(void) refcount_add(&dn->dn_holds, db);
1438 
1439 	dprintf_dbuf(db, "db=%p\n", db);
1440 
1441 	return (db);
1442 }
1443 
1444 static int
1445 dbuf_do_evict(void *private)
1446 {
1447 	arc_buf_t *buf = private;
1448 	dmu_buf_impl_t *db = buf->b_private;
1449 
1450 	if (!MUTEX_HELD(&db->db_mtx))
1451 		mutex_enter(&db->db_mtx);
1452 
1453 	ASSERT(refcount_is_zero(&db->db_holds));
1454 
1455 	if (db->db_state != DB_EVICTING) {
1456 		ASSERT(db->db_state == DB_CACHED);
1457 		DBUF_VERIFY(db);
1458 		db->db_buf = NULL;
1459 		dbuf_evict(db);
1460 	} else {
1461 		mutex_exit(&db->db_mtx);
1462 		dbuf_destroy(db);
1463 	}
1464 	return (0);
1465 }
1466 
1467 static void
1468 dbuf_destroy(dmu_buf_impl_t *db)
1469 {
1470 	ASSERT(refcount_is_zero(&db->db_holds));
1471 
1472 	if (db->db_blkid != DB_BONUS_BLKID) {
1473 		dnode_t *dn = db->db_dnode;
1474 		boolean_t need_mutex = !MUTEX_HELD(&dn->dn_dbufs_mtx);
1475 
1476 		if (need_mutex)
1477 			mutex_enter(&dn->dn_dbufs_mtx);
1478 
1479 		/*
1480 		 * If this dbuf is still on the dn_dbufs list,
1481 		 * remove it from that list.
1482 		 */
1483 		if (list_link_active(&db->db_link)) {
1484 			ASSERT(need_mutex);
1485 			list_remove(&dn->dn_dbufs, db);
1486 			mutex_exit(&dn->dn_dbufs_mtx);
1487 
1488 			dnode_rele(dn, db);
1489 		} else if (need_mutex) {
1490 			mutex_exit(&dn->dn_dbufs_mtx);
1491 		}
1492 		dbuf_hash_remove(db);
1493 	}
1494 	db->db_parent = NULL;
1495 	db->db_dnode = NULL;
1496 	db->db_buf = NULL;
1497 
1498 	ASSERT(!list_link_active(&db->db_link));
1499 	ASSERT(db->db.db_data == NULL);
1500 	ASSERT(db->db_hash_next == NULL);
1501 	ASSERT(db->db_blkptr == NULL);
1502 	ASSERT(db->db_data_pending == NULL);
1503 
1504 	kmem_cache_free(dbuf_cache, db);
1505 	arc_space_return(sizeof (dmu_buf_impl_t));
1506 }
1507 
1508 void
1509 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1510 {
1511 	dmu_buf_impl_t *db = NULL;
1512 	blkptr_t *bp = NULL;
1513 
1514 	ASSERT(blkid != DB_BONUS_BLKID);
1515 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1516 
1517 	if (dnode_block_freed(dn, blkid))
1518 		return;
1519 
1520 	/* dbuf_find() returns with db_mtx held */
1521 	if (db = dbuf_find(dn, 0, blkid)) {
1522 		if (refcount_count(&db->db_holds) > 0) {
1523 			/*
1524 			 * This dbuf is active.  We assume that it is
1525 			 * already CACHED, or else about to be either
1526 			 * read or filled.
1527 			 */
1528 			mutex_exit(&db->db_mtx);
1529 			return;
1530 		}
1531 		mutex_exit(&db->db_mtx);
1532 		db = NULL;
1533 	}
1534 
1535 	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1536 		if (bp && !BP_IS_HOLE(bp)) {
1537 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1538 			zbookmark_t zb;
1539 			zb.zb_objset = dn->dn_objset->os_dsl_dataset ?
1540 			    dn->dn_objset->os_dsl_dataset->ds_object : 0;
1541 			zb.zb_object = dn->dn_object;
1542 			zb.zb_level = 0;
1543 			zb.zb_blkid = blkid;
1544 
1545 			(void) arc_read(NULL, dn->dn_objset->os_spa, bp,
1546 			    dmu_ot[dn->dn_type].ot_byteswap,
1547 			    NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
1548 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1549 			    &aflags, &zb);
1550 		}
1551 		if (db)
1552 			dbuf_rele(db, NULL);
1553 	}
1554 }
1555 
1556 /*
1557  * Returns with db_holds incremented, and db_mtx not held.
1558  * Note: dn_struct_rwlock must be held.
1559  */
1560 int
1561 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1562     void *tag, dmu_buf_impl_t **dbp)
1563 {
1564 	dmu_buf_impl_t *db, *parent = NULL;
1565 
1566 	ASSERT(blkid != DB_BONUS_BLKID);
1567 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1568 	ASSERT3U(dn->dn_nlevels, >, level);
1569 
1570 	*dbp = NULL;
1571 top:
1572 	/* dbuf_find() returns with db_mtx held */
1573 	db = dbuf_find(dn, level, blkid);
1574 
1575 	if (db == NULL) {
1576 		blkptr_t *bp = NULL;
1577 		int err;
1578 
1579 		ASSERT3P(parent, ==, NULL);
1580 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1581 		if (fail_sparse) {
1582 			if (err == 0 && bp && BP_IS_HOLE(bp))
1583 				err = ENOENT;
1584 			if (err) {
1585 				if (parent)
1586 					dbuf_rele(parent, NULL);
1587 				return (err);
1588 			}
1589 		}
1590 		if (err && err != ENOENT)
1591 			return (err);
1592 		db = dbuf_create(dn, level, blkid, parent, bp);
1593 	}
1594 
1595 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1596 		arc_buf_add_ref(db->db_buf, db);
1597 		if (db->db_buf->b_data == NULL) {
1598 			dbuf_clear(db);
1599 			if (parent) {
1600 				dbuf_rele(parent, NULL);
1601 				parent = NULL;
1602 			}
1603 			goto top;
1604 		}
1605 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1606 	}
1607 
1608 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1609 
1610 	/*
1611 	 * If this buffer is currently syncing out, and we are are
1612 	 * still referencing it from db_data, we need to make a copy
1613 	 * of it in case we decide we want to dirty it again in this txg.
1614 	 */
1615 	if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID &&
1616 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
1617 	    db->db_state == DB_CACHED && db->db_data_pending) {
1618 		dbuf_dirty_record_t *dr = db->db_data_pending;
1619 
1620 		if (dr->dt.dl.dr_data == db->db_buf) {
1621 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1622 
1623 			dbuf_set_data(db,
1624 			    arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
1625 			    db->db.db_size, db, type));
1626 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1627 			    db->db.db_size);
1628 		}
1629 	}
1630 
1631 	(void) refcount_add(&db->db_holds, tag);
1632 	dbuf_update_data(db);
1633 	DBUF_VERIFY(db);
1634 	mutex_exit(&db->db_mtx);
1635 
1636 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
1637 	if (parent)
1638 		dbuf_rele(parent, NULL);
1639 
1640 	ASSERT3P(db->db_dnode, ==, dn);
1641 	ASSERT3U(db->db_blkid, ==, blkid);
1642 	ASSERT3U(db->db_level, ==, level);
1643 	*dbp = db;
1644 
1645 	return (0);
1646 }
1647 
1648 dmu_buf_impl_t *
1649 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1650 {
1651 	dmu_buf_impl_t *db;
1652 	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1653 	return (err ? NULL : db);
1654 }
1655 
1656 dmu_buf_impl_t *
1657 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1658 {
1659 	dmu_buf_impl_t *db;
1660 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1661 	return (err ? NULL : db);
1662 }
1663 
1664 dmu_buf_impl_t *
1665 dbuf_create_bonus(dnode_t *dn)
1666 {
1667 	dmu_buf_impl_t *db = dn->dn_bonus;
1668 
1669 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1670 
1671 	ASSERT(dn->dn_bonus == NULL);
1672 	db = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL);
1673 	return (db);
1674 }
1675 
1676 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1677 void
1678 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1679 {
1680 	int64_t holds = refcount_add(&db->db_holds, tag);
1681 	ASSERT(holds > 1);
1682 }
1683 
1684 #pragma weak dmu_buf_rele = dbuf_rele
1685 void
1686 dbuf_rele(dmu_buf_impl_t *db, void *tag)
1687 {
1688 	int64_t holds;
1689 
1690 	mutex_enter(&db->db_mtx);
1691 	DBUF_VERIFY(db);
1692 
1693 	holds = refcount_remove(&db->db_holds, tag);
1694 	ASSERT(holds >= 0);
1695 
1696 	/*
1697 	 * We can't freeze indirects if there is a possibility that they
1698 	 * may be modified in the current syncing context.
1699 	 */
1700 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
1701 		arc_buf_freeze(db->db_buf);
1702 
1703 	if (holds == db->db_dirtycnt &&
1704 	    db->db_level == 0 && db->db_immediate_evict)
1705 		dbuf_evict_user(db);
1706 
1707 	if (holds == 0) {
1708 		if (db->db_blkid == DB_BONUS_BLKID) {
1709 			mutex_exit(&db->db_mtx);
1710 			dnode_rele(db->db_dnode, db);
1711 		} else if (db->db_buf == NULL) {
1712 			/*
1713 			 * This is a special case: we never associated this
1714 			 * dbuf with any data allocated from the ARC.
1715 			 */
1716 			ASSERT3U(db->db_state, ==, DB_UNCACHED);
1717 			dbuf_evict(db);
1718 		} else if (arc_released(db->db_buf)) {
1719 			arc_buf_t *buf = db->db_buf;
1720 			/*
1721 			 * This dbuf has anonymous data associated with it.
1722 			 */
1723 			dbuf_set_data(db, NULL);
1724 			VERIFY(arc_buf_remove_ref(buf, db) == 1);
1725 			dbuf_evict(db);
1726 		} else {
1727 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0);
1728 			mutex_exit(&db->db_mtx);
1729 		}
1730 	} else {
1731 		mutex_exit(&db->db_mtx);
1732 	}
1733 }
1734 
1735 #pragma weak dmu_buf_refcount = dbuf_refcount
1736 uint64_t
1737 dbuf_refcount(dmu_buf_impl_t *db)
1738 {
1739 	return (refcount_count(&db->db_holds));
1740 }
1741 
1742 void *
1743 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1744     dmu_buf_evict_func_t *evict_func)
1745 {
1746 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1747 	    user_data_ptr_ptr, evict_func));
1748 }
1749 
1750 void *
1751 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1752     dmu_buf_evict_func_t *evict_func)
1753 {
1754 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1755 
1756 	db->db_immediate_evict = TRUE;
1757 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1758 	    user_data_ptr_ptr, evict_func));
1759 }
1760 
1761 void *
1762 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
1763     void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
1764 {
1765 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1766 	ASSERT(db->db_level == 0);
1767 
1768 	ASSERT((user_ptr == NULL) == (evict_func == NULL));
1769 
1770 	mutex_enter(&db->db_mtx);
1771 
1772 	if (db->db_user_ptr == old_user_ptr) {
1773 		db->db_user_ptr = user_ptr;
1774 		db->db_user_data_ptr_ptr = user_data_ptr_ptr;
1775 		db->db_evict_func = evict_func;
1776 
1777 		dbuf_update_data(db);
1778 	} else {
1779 		old_user_ptr = db->db_user_ptr;
1780 	}
1781 
1782 	mutex_exit(&db->db_mtx);
1783 	return (old_user_ptr);
1784 }
1785 
1786 void *
1787 dmu_buf_get_user(dmu_buf_t *db_fake)
1788 {
1789 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1790 	ASSERT(!refcount_is_zero(&db->db_holds));
1791 
1792 	return (db->db_user_ptr);
1793 }
1794 
1795 static void
1796 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
1797 {
1798 	/* ASSERT(dmu_tx_is_syncing(tx) */
1799 	ASSERT(MUTEX_HELD(&db->db_mtx));
1800 
1801 	if (db->db_blkptr != NULL)
1802 		return;
1803 
1804 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
1805 		/*
1806 		 * This buffer was allocated at a time when there was
1807 		 * no available blkptrs from the dnode, or it was
1808 		 * inappropriate to hook it in (i.e., nlevels mis-match).
1809 		 */
1810 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
1811 		ASSERT(db->db_parent == NULL);
1812 		db->db_parent = dn->dn_dbuf;
1813 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
1814 		DBUF_VERIFY(db);
1815 	} else {
1816 		dmu_buf_impl_t *parent = db->db_parent;
1817 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1818 
1819 		ASSERT(dn->dn_phys->dn_nlevels > 1);
1820 		if (parent == NULL) {
1821 			mutex_exit(&db->db_mtx);
1822 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
1823 			(void) dbuf_hold_impl(dn, db->db_level+1,
1824 			    db->db_blkid >> epbs, FALSE, db, &parent);
1825 			rw_exit(&dn->dn_struct_rwlock);
1826 			mutex_enter(&db->db_mtx);
1827 			db->db_parent = parent;
1828 		}
1829 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
1830 		    (db->db_blkid & ((1ULL << epbs) - 1));
1831 		DBUF_VERIFY(db);
1832 	}
1833 }
1834 
1835 static void
1836 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
1837 {
1838 	dmu_buf_impl_t *db = dr->dr_dbuf;
1839 	dnode_t *dn = db->db_dnode;
1840 	zio_t *zio;
1841 
1842 	ASSERT(dmu_tx_is_syncing(tx));
1843 
1844 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
1845 
1846 	mutex_enter(&db->db_mtx);
1847 
1848 	ASSERT(db->db_level > 0);
1849 	DBUF_VERIFY(db);
1850 
1851 	if (db->db_buf == NULL) {
1852 		mutex_exit(&db->db_mtx);
1853 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
1854 		mutex_enter(&db->db_mtx);
1855 	}
1856 	ASSERT3U(db->db_state, ==, DB_CACHED);
1857 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
1858 	ASSERT(db->db_buf != NULL);
1859 
1860 	dbuf_check_blkptr(dn, db);
1861 
1862 	db->db_data_pending = dr;
1863 
1864 	arc_release(db->db_buf, db);
1865 	mutex_exit(&db->db_mtx);
1866 
1867 	/*
1868 	 * XXX -- we should design a compression algorithm
1869 	 * that specializes in arrays of bps.
1870 	 */
1871 	dbuf_write(dr, db->db_buf, ZIO_CHECKSUM_FLETCHER_4,
1872 	    zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : ZIO_COMPRESS_LZJB, tx);
1873 
1874 	zio = dr->dr_zio;
1875 	mutex_enter(&dr->dt.di.dr_mtx);
1876 	dbuf_sync_list(&dr->dt.di.dr_children, tx);
1877 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1878 	mutex_exit(&dr->dt.di.dr_mtx);
1879 	zio_nowait(zio);
1880 }
1881 
1882 static void
1883 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
1884 {
1885 	arc_buf_t **datap = &dr->dt.dl.dr_data;
1886 	dmu_buf_impl_t *db = dr->dr_dbuf;
1887 	dnode_t *dn = db->db_dnode;
1888 	objset_impl_t *os = dn->dn_objset;
1889 	uint64_t txg = tx->tx_txg;
1890 	int checksum, compress;
1891 	int blksz;
1892 
1893 	ASSERT(dmu_tx_is_syncing(tx));
1894 
1895 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
1896 
1897 	mutex_enter(&db->db_mtx);
1898 	/*
1899 	 * To be synced, we must be dirtied.  But we
1900 	 * might have been freed after the dirty.
1901 	 */
1902 	if (db->db_state == DB_UNCACHED) {
1903 		/* This buffer has been freed since it was dirtied */
1904 		ASSERT(db->db.db_data == NULL);
1905 	} else if (db->db_state == DB_FILL) {
1906 		/* This buffer was freed and is now being re-filled */
1907 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
1908 	} else {
1909 		ASSERT3U(db->db_state, ==, DB_CACHED);
1910 	}
1911 	DBUF_VERIFY(db);
1912 
1913 	/*
1914 	 * If this is a bonus buffer, simply copy the bonus data into the
1915 	 * dnode.  It will be written out when the dnode is synced (and it
1916 	 * will be synced, since it must have been dirty for dbuf_sync to
1917 	 * be called).
1918 	 */
1919 	if (db->db_blkid == DB_BONUS_BLKID) {
1920 		dbuf_dirty_record_t **drp;
1921 		/*
1922 		 * Use dn_phys->dn_bonuslen since db.db_size is the length
1923 		 * of the bonus buffer in the open transaction rather than
1924 		 * the syncing transaction.
1925 		 */
1926 		ASSERT(*datap != NULL);
1927 		ASSERT3U(db->db_level, ==, 0);
1928 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
1929 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
1930 		if (*datap != db->db.db_data) {
1931 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
1932 			arc_space_return(DN_MAX_BONUSLEN);
1933 		}
1934 		db->db_data_pending = NULL;
1935 		drp = &db->db_last_dirty;
1936 		while (*drp != dr)
1937 			drp = &(*drp)->dr_next;
1938 		ASSERT((*drp)->dr_next == NULL);
1939 		*drp = NULL;
1940 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
1941 		ASSERT(db->db_dirtycnt > 0);
1942 		db->db_dirtycnt -= 1;
1943 		mutex_exit(&db->db_mtx);
1944 		dbuf_rele(db, (void *)(uintptr_t)txg);
1945 		return;
1946 	}
1947 
1948 	/*
1949 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
1950 	 * operation to sneak in. As a result, we need to ensure that we
1951 	 * don't check the dr_override_state until we have returned from
1952 	 * dbuf_check_blkptr.
1953 	 */
1954 	dbuf_check_blkptr(dn, db);
1955 
1956 	/*
1957 	 * If this buffer is in the middle of an immdiate write,
1958 	 * wait for the synchronous IO to complete.
1959 	 */
1960 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
1961 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
1962 		cv_wait(&db->db_changed, &db->db_mtx);
1963 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
1964 	}
1965 
1966 	/*
1967 	 * If this dbuf has already been written out via an immediate write,
1968 	 * just complete the write by copying over the new block pointer and
1969 	 * updating the accounting via the write-completion functions.
1970 	 */
1971 	if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1972 		zio_t zio_fake;
1973 
1974 		zio_fake.io_private = &db;
1975 		zio_fake.io_error = 0;
1976 		zio_fake.io_bp = db->db_blkptr;
1977 		zio_fake.io_bp_orig = *db->db_blkptr;
1978 		zio_fake.io_txg = txg;
1979 
1980 		*db->db_blkptr = dr->dt.dl.dr_overridden_by;
1981 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1982 		db->db_data_pending = dr;
1983 		dr->dr_zio = &zio_fake;
1984 		mutex_exit(&db->db_mtx);
1985 
1986 		if (BP_IS_OLDER(&zio_fake.io_bp_orig, txg))
1987 			dsl_dataset_block_kill(os->os_dsl_dataset,
1988 			    &zio_fake.io_bp_orig, dn->dn_zio, tx);
1989 
1990 		dbuf_write_ready(&zio_fake, db->db_buf, db);
1991 		dbuf_write_done(&zio_fake, db->db_buf, db);
1992 
1993 		return;
1994 	}
1995 
1996 	blksz = arc_buf_size(*datap);
1997 
1998 	if (dn->dn_object != DMU_META_DNODE_OBJECT) {
1999 		/*
2000 		 * If this buffer is currently "in use" (i.e., there are
2001 		 * active holds and db_data still references it), then make
2002 		 * a copy before we start the write so that any modifications
2003 		 * from the open txg will not leak into this write.
2004 		 *
2005 		 * NOTE: this copy does not need to be made for objects only
2006 		 * modified in the syncing context (e.g. DNONE_DNODE blocks).
2007 		 */
2008 		if (refcount_count(&db->db_holds) > 1 && *datap == db->db_buf) {
2009 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2010 			*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2011 			bcopy(db->db.db_data, (*datap)->b_data, blksz);
2012 		}
2013 	} else {
2014 		/*
2015 		 * Private object buffers are released here rather
2016 		 * than in dbuf_dirty() since they are only modified
2017 		 * in the syncing context and we don't want the
2018 		 * overhead of making multiple copies of the data.
2019 		 */
2020 		arc_release(db->db_buf, db);
2021 	}
2022 
2023 	ASSERT(*datap != NULL);
2024 	db->db_data_pending = dr;
2025 
2026 	mutex_exit(&db->db_mtx);
2027 
2028 	/*
2029 	 * Allow dnode settings to override objset settings,
2030 	 * except for metadata checksums.
2031 	 */
2032 	if (dmu_ot[dn->dn_type].ot_metadata) {
2033 		checksum = os->os_md_checksum;
2034 		compress = zio_compress_select(dn->dn_compress,
2035 		    os->os_md_compress);
2036 	} else {
2037 		checksum = zio_checksum_select(dn->dn_checksum,
2038 		    os->os_checksum);
2039 		compress = zio_compress_select(dn->dn_compress,
2040 		    os->os_compress);
2041 	}
2042 
2043 	dbuf_write(dr, *datap, checksum, compress, tx);
2044 
2045 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2046 	if (dn->dn_object == DMU_META_DNODE_OBJECT)
2047 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2048 	else
2049 		zio_nowait(dr->dr_zio);
2050 }
2051 
2052 void
2053 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2054 {
2055 	dbuf_dirty_record_t *dr;
2056 
2057 	while (dr = list_head(list)) {
2058 		if (dr->dr_zio != NULL) {
2059 			/*
2060 			 * If we find an already initialized zio then we
2061 			 * are processing the meta-dnode, and we have finished.
2062 			 * The dbufs for all dnodes are put back on the list
2063 			 * during processing, so that we can zio_wait()
2064 			 * these IOs after initiating all child IOs.
2065 			 */
2066 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2067 			    DMU_META_DNODE_OBJECT);
2068 			break;
2069 		}
2070 		list_remove(list, dr);
2071 		if (dr->dr_dbuf->db_level > 0)
2072 			dbuf_sync_indirect(dr, tx);
2073 		else
2074 			dbuf_sync_leaf(dr, tx);
2075 	}
2076 }
2077 
2078 static void
2079 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, int checksum,
2080     int compress, dmu_tx_t *tx)
2081 {
2082 	dmu_buf_impl_t *db = dr->dr_dbuf;
2083 	dnode_t *dn = db->db_dnode;
2084 	objset_impl_t *os = dn->dn_objset;
2085 	dmu_buf_impl_t *parent = db->db_parent;
2086 	uint64_t txg = tx->tx_txg;
2087 	zbookmark_t zb;
2088 	zio_t *zio;
2089 	int zio_flags;
2090 
2091 	if (parent != dn->dn_dbuf) {
2092 		ASSERT(parent && parent->db_data_pending);
2093 		ASSERT(db->db_level == parent->db_level-1);
2094 		ASSERT(arc_released(parent->db_buf));
2095 		zio = parent->db_data_pending->dr_zio;
2096 	} else {
2097 		ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1);
2098 		ASSERT3P(db->db_blkptr, ==,
2099 		    &dn->dn_phys->dn_blkptr[db->db_blkid]);
2100 		zio = dn->dn_zio;
2101 	}
2102 
2103 	ASSERT(db->db_level == 0 || data == db->db_buf);
2104 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2105 	ASSERT(zio);
2106 
2107 	zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0;
2108 	zb.zb_object = db->db.db_object;
2109 	zb.zb_level = db->db_level;
2110 	zb.zb_blkid = db->db_blkid;
2111 
2112 	zio_flags = ZIO_FLAG_MUSTSUCCEED;
2113 	if (dmu_ot[dn->dn_type].ot_metadata || zb.zb_level != 0)
2114 		zio_flags |= ZIO_FLAG_METADATA;
2115 	if (BP_IS_OLDER(db->db_blkptr, txg))
2116 		dsl_dataset_block_kill(
2117 		    os->os_dsl_dataset, db->db_blkptr, zio, tx);
2118 
2119 	dr->dr_zio = arc_write(zio, os->os_spa, checksum, compress,
2120 	    dmu_get_replication_level(os, &zb, dn->dn_type), txg,
2121 	    db->db_blkptr, data, dbuf_write_ready, dbuf_write_done, db,
2122 	    ZIO_PRIORITY_ASYNC_WRITE, zio_flags, &zb);
2123 }
2124 
2125 /* ARGSUSED */
2126 static void
2127 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2128 {
2129 	dmu_buf_impl_t *db = vdb;
2130 	dnode_t *dn = db->db_dnode;
2131 	objset_impl_t *os = dn->dn_objset;
2132 	blkptr_t *bp_orig = &zio->io_bp_orig;
2133 	uint64_t fill = 0;
2134 	int old_size, new_size, i;
2135 
2136 	dprintf_dbuf_bp(db, bp_orig, "bp_orig: %s", "");
2137 
2138 	old_size = bp_get_dasize(os->os_spa, bp_orig);
2139 	new_size = bp_get_dasize(os->os_spa, zio->io_bp);
2140 
2141 	dnode_diduse_space(dn, new_size-old_size);
2142 
2143 	if (BP_IS_HOLE(zio->io_bp)) {
2144 		dsl_dataset_t *ds = os->os_dsl_dataset;
2145 		dmu_tx_t *tx = os->os_synctx;
2146 
2147 		if (bp_orig->blk_birth == tx->tx_txg)
2148 			dsl_dataset_block_kill(ds, bp_orig, NULL, tx);
2149 		ASSERT3U(db->db_blkptr->blk_fill, ==, 0);
2150 		return;
2151 	}
2152 
2153 	mutex_enter(&db->db_mtx);
2154 
2155 	if (db->db_level == 0) {
2156 		mutex_enter(&dn->dn_mtx);
2157 		if (db->db_blkid > dn->dn_phys->dn_maxblkid)
2158 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2159 		mutex_exit(&dn->dn_mtx);
2160 
2161 		if (dn->dn_type == DMU_OT_DNODE) {
2162 			dnode_phys_t *dnp = db->db.db_data;
2163 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2164 			    i--, dnp++) {
2165 				if (dnp->dn_type != DMU_OT_NONE)
2166 					fill++;
2167 			}
2168 		} else {
2169 			fill = 1;
2170 		}
2171 	} else {
2172 		blkptr_t *bp = db->db.db_data;
2173 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2174 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, bp++) {
2175 			if (BP_IS_HOLE(bp))
2176 				continue;
2177 			ASSERT3U(BP_GET_LSIZE(bp), ==,
2178 			    db->db_level == 1 ? dn->dn_datablksz :
2179 			    (1<<dn->dn_phys->dn_indblkshift));
2180 			fill += bp->blk_fill;
2181 		}
2182 	}
2183 
2184 	db->db_blkptr->blk_fill = fill;
2185 	BP_SET_TYPE(db->db_blkptr, dn->dn_type);
2186 	BP_SET_LEVEL(db->db_blkptr, db->db_level);
2187 
2188 	mutex_exit(&db->db_mtx);
2189 
2190 	/* We must do this after we've set the bp's type and level */
2191 	if (!DVA_EQUAL(BP_IDENTITY(zio->io_bp), BP_IDENTITY(bp_orig))) {
2192 		dsl_dataset_t *ds = os->os_dsl_dataset;
2193 		dmu_tx_t *tx = os->os_synctx;
2194 
2195 		if (bp_orig->blk_birth == tx->tx_txg)
2196 			dsl_dataset_block_kill(ds, bp_orig, NULL, tx);
2197 		dsl_dataset_block_born(ds, zio->io_bp, tx);
2198 	}
2199 }
2200 
2201 /* ARGSUSED */
2202 static void
2203 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2204 {
2205 	dmu_buf_impl_t *db = vdb;
2206 	uint64_t txg = zio->io_txg;
2207 	dbuf_dirty_record_t **drp, *dr;
2208 
2209 	ASSERT3U(zio->io_error, ==, 0);
2210 
2211 	mutex_enter(&db->db_mtx);
2212 
2213 	drp = &db->db_last_dirty;
2214 	while (*drp != db->db_data_pending)
2215 		drp = &(*drp)->dr_next;
2216 	ASSERT(!list_link_active(&(*drp)->dr_dirty_node));
2217 	ASSERT((*drp)->dr_txg == txg);
2218 	ASSERT((*drp)->dr_next == NULL);
2219 	dr = *drp;
2220 	*drp = NULL;
2221 
2222 	if (db->db_level == 0) {
2223 		ASSERT(db->db_blkid != DB_BONUS_BLKID);
2224 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2225 
2226 		if (dr->dt.dl.dr_data != db->db_buf)
2227 			VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1);
2228 		else if (!BP_IS_HOLE(db->db_blkptr))
2229 			arc_set_callback(db->db_buf, dbuf_do_evict, db);
2230 		else
2231 			ASSERT(arc_released(db->db_buf));
2232 	} else {
2233 		dnode_t *dn = db->db_dnode;
2234 
2235 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2236 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2237 		if (!BP_IS_HOLE(db->db_blkptr)) {
2238 			int epbs =
2239 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2240 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2241 			    db->db.db_size);
2242 			ASSERT3U(dn->dn_phys->dn_maxblkid
2243 			    >> (db->db_level * epbs), >=, db->db_blkid);
2244 			arc_set_callback(db->db_buf, dbuf_do_evict, db);
2245 		}
2246 	}
2247 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
2248 
2249 	cv_broadcast(&db->db_changed);
2250 	ASSERT(db->db_dirtycnt > 0);
2251 	db->db_dirtycnt -= 1;
2252 	db->db_data_pending = NULL;
2253 	mutex_exit(&db->db_mtx);
2254 
2255 	dprintf_dbuf_bp(db, zio->io_bp, "bp: %s", "");
2256 
2257 	dbuf_rele(db, (void *)(uintptr_t)txg);
2258 }
2259