xref: /titanic_44/usr/src/uts/common/fs/zfs/dbuf.c (revision 1a5e258f5471356ca102c7176637cdce45bac147)
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  */
28 
29 #include <sys/zfs_context.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_send.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dbuf.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/spa.h>
39 #include <sys/zio.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/sa.h>
42 #include <sys/sa_impl.h>
43 #include <sys/zfeature.h>
44 #include <sys/blkptr.h>
45 #include <sys/range_tree.h>
46 
47 /*
48  * Number of times that zfs_free_range() took the slow path while doing
49  * a zfs receive.  A nonzero value indicates a potential performance problem.
50  */
51 uint64_t zfs_free_range_recv_miss;
52 
53 static void dbuf_destroy(dmu_buf_impl_t *db);
54 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
55 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
56 
57 /*
58  * Global data structures and functions for the dbuf cache.
59  */
60 static kmem_cache_t *dbuf_cache;
61 
62 /* ARGSUSED */
63 static int
64 dbuf_cons(void *vdb, void *unused, int kmflag)
65 {
66 	dmu_buf_impl_t *db = vdb;
67 	bzero(db, sizeof (dmu_buf_impl_t));
68 
69 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
70 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
71 	refcount_create(&db->db_holds);
72 
73 	db->db_creation = gethrtime();
74 
75 	return (0);
76 }
77 
78 /* ARGSUSED */
79 static void
80 dbuf_dest(void *vdb, void *unused)
81 {
82 	dmu_buf_impl_t *db = vdb;
83 	mutex_destroy(&db->db_mtx);
84 	cv_destroy(&db->db_changed);
85 	refcount_destroy(&db->db_holds);
86 }
87 
88 /*
89  * dbuf hash table routines
90  */
91 static dbuf_hash_table_t dbuf_hash_table;
92 
93 static uint64_t dbuf_hash_count;
94 
95 static uint64_t
96 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
97 {
98 	uintptr_t osv = (uintptr_t)os;
99 	uint64_t crc = -1ULL;
100 
101 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
102 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
103 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
104 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
105 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
106 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
107 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
108 
109 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
110 
111 	return (crc);
112 }
113 
114 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
115 
116 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
117 	((dbuf)->db.db_object == (obj) &&		\
118 	(dbuf)->db_objset == (os) &&			\
119 	(dbuf)->db_level == (level) &&			\
120 	(dbuf)->db_blkid == (blkid))
121 
122 dmu_buf_impl_t *
123 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
124 {
125 	dbuf_hash_table_t *h = &dbuf_hash_table;
126 	objset_t *os = dn->dn_objset;
127 	uint64_t obj = dn->dn_object;
128 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
129 	uint64_t idx = hv & h->hash_table_mask;
130 	dmu_buf_impl_t *db;
131 
132 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
133 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
134 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
135 			mutex_enter(&db->db_mtx);
136 			if (db->db_state != DB_EVICTING) {
137 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
138 				return (db);
139 			}
140 			mutex_exit(&db->db_mtx);
141 		}
142 	}
143 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
144 	return (NULL);
145 }
146 
147 /*
148  * Insert an entry into the hash table.  If there is already an element
149  * equal to elem in the hash table, then the already existing element
150  * will be returned and the new element will not be inserted.
151  * Otherwise returns NULL.
152  */
153 static dmu_buf_impl_t *
154 dbuf_hash_insert(dmu_buf_impl_t *db)
155 {
156 	dbuf_hash_table_t *h = &dbuf_hash_table;
157 	objset_t *os = db->db_objset;
158 	uint64_t obj = db->db.db_object;
159 	int level = db->db_level;
160 	uint64_t blkid = db->db_blkid;
161 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
162 	uint64_t idx = hv & h->hash_table_mask;
163 	dmu_buf_impl_t *dbf;
164 
165 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
166 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
167 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
168 			mutex_enter(&dbf->db_mtx);
169 			if (dbf->db_state != DB_EVICTING) {
170 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
171 				return (dbf);
172 			}
173 			mutex_exit(&dbf->db_mtx);
174 		}
175 	}
176 
177 	mutex_enter(&db->db_mtx);
178 	db->db_hash_next = h->hash_table[idx];
179 	h->hash_table[idx] = db;
180 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
181 	atomic_inc_64(&dbuf_hash_count);
182 
183 	return (NULL);
184 }
185 
186 /*
187  * Remove an entry from the hash table.  It must be in the EVICTING state.
188  */
189 static void
190 dbuf_hash_remove(dmu_buf_impl_t *db)
191 {
192 	dbuf_hash_table_t *h = &dbuf_hash_table;
193 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
194 	    db->db_level, db->db_blkid);
195 	uint64_t idx = hv & h->hash_table_mask;
196 	dmu_buf_impl_t *dbf, **dbp;
197 
198 	/*
199 	 * We musn't hold db_mtx to maintain lock ordering:
200 	 * DBUF_HASH_MUTEX > db_mtx.
201 	 */
202 	ASSERT(refcount_is_zero(&db->db_holds));
203 	ASSERT(db->db_state == DB_EVICTING);
204 	ASSERT(!MUTEX_HELD(&db->db_mtx));
205 
206 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
207 	dbp = &h->hash_table[idx];
208 	while ((dbf = *dbp) != db) {
209 		dbp = &dbf->db_hash_next;
210 		ASSERT(dbf != NULL);
211 	}
212 	*dbp = db->db_hash_next;
213 	db->db_hash_next = NULL;
214 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
215 	atomic_dec_64(&dbuf_hash_count);
216 }
217 
218 static arc_evict_func_t dbuf_do_evict;
219 
220 static void
221 dbuf_evict_user(dmu_buf_impl_t *db)
222 {
223 	ASSERT(MUTEX_HELD(&db->db_mtx));
224 
225 	if (db->db_level != 0 || db->db_evict_func == NULL)
226 		return;
227 
228 	if (db->db_user_data_ptr_ptr)
229 		*db->db_user_data_ptr_ptr = db->db.db_data;
230 	db->db_evict_func(&db->db, db->db_user_ptr);
231 	db->db_user_ptr = NULL;
232 	db->db_user_data_ptr_ptr = NULL;
233 	db->db_evict_func = NULL;
234 }
235 
236 boolean_t
237 dbuf_is_metadata(dmu_buf_impl_t *db)
238 {
239 	if (db->db_level > 0) {
240 		return (B_TRUE);
241 	} else {
242 		boolean_t is_metadata;
243 
244 		DB_DNODE_ENTER(db);
245 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
246 		DB_DNODE_EXIT(db);
247 
248 		return (is_metadata);
249 	}
250 }
251 
252 void
253 dbuf_evict(dmu_buf_impl_t *db)
254 {
255 	ASSERT(MUTEX_HELD(&db->db_mtx));
256 	ASSERT(db->db_buf == NULL);
257 	ASSERT(db->db_data_pending == NULL);
258 
259 	dbuf_clear(db);
260 	dbuf_destroy(db);
261 }
262 
263 void
264 dbuf_init(void)
265 {
266 	uint64_t hsize = 1ULL << 16;
267 	dbuf_hash_table_t *h = &dbuf_hash_table;
268 	int i;
269 
270 	/*
271 	 * The hash table is big enough to fill all of physical memory
272 	 * with an average 4K block size.  The table will take up
273 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
274 	 */
275 	while (hsize * 4096 < physmem * PAGESIZE)
276 		hsize <<= 1;
277 
278 retry:
279 	h->hash_table_mask = hsize - 1;
280 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
281 	if (h->hash_table == NULL) {
282 		/* XXX - we should really return an error instead of assert */
283 		ASSERT(hsize > (1ULL << 10));
284 		hsize >>= 1;
285 		goto retry;
286 	}
287 
288 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
289 	    sizeof (dmu_buf_impl_t),
290 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
291 
292 	for (i = 0; i < DBUF_MUTEXES; i++)
293 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
294 }
295 
296 void
297 dbuf_fini(void)
298 {
299 	dbuf_hash_table_t *h = &dbuf_hash_table;
300 	int i;
301 
302 	for (i = 0; i < DBUF_MUTEXES; i++)
303 		mutex_destroy(&h->hash_mutexes[i]);
304 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
305 	kmem_cache_destroy(dbuf_cache);
306 }
307 
308 /*
309  * Other stuff.
310  */
311 
312 #ifdef ZFS_DEBUG
313 static void
314 dbuf_verify(dmu_buf_impl_t *db)
315 {
316 	dnode_t *dn;
317 	dbuf_dirty_record_t *dr;
318 
319 	ASSERT(MUTEX_HELD(&db->db_mtx));
320 
321 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
322 		return;
323 
324 	ASSERT(db->db_objset != NULL);
325 	DB_DNODE_ENTER(db);
326 	dn = DB_DNODE(db);
327 	if (dn == NULL) {
328 		ASSERT(db->db_parent == NULL);
329 		ASSERT(db->db_blkptr == NULL);
330 	} else {
331 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
332 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
333 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
334 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
335 		    db->db_blkid == DMU_SPILL_BLKID ||
336 		    !avl_is_empty(&dn->dn_dbufs));
337 	}
338 	if (db->db_blkid == DMU_BONUS_BLKID) {
339 		ASSERT(dn != NULL);
340 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
341 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
342 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
343 		ASSERT(dn != NULL);
344 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
345 		ASSERT0(db->db.db_offset);
346 	} else {
347 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
348 	}
349 
350 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
351 		ASSERT(dr->dr_dbuf == db);
352 
353 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
354 		ASSERT(dr->dr_dbuf == db);
355 
356 	/*
357 	 * We can't assert that db_size matches dn_datablksz because it
358 	 * can be momentarily different when another thread is doing
359 	 * dnode_set_blksz().
360 	 */
361 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
362 		dr = db->db_data_pending;
363 		/*
364 		 * It should only be modified in syncing context, so
365 		 * make sure we only have one copy of the data.
366 		 */
367 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
368 	}
369 
370 	/* verify db->db_blkptr */
371 	if (db->db_blkptr) {
372 		if (db->db_parent == dn->dn_dbuf) {
373 			/* db is pointed to by the dnode */
374 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
375 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
376 				ASSERT(db->db_parent == NULL);
377 			else
378 				ASSERT(db->db_parent != NULL);
379 			if (db->db_blkid != DMU_SPILL_BLKID)
380 				ASSERT3P(db->db_blkptr, ==,
381 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
382 		} else {
383 			/* db is pointed to by an indirect block */
384 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
385 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
386 			ASSERT3U(db->db_parent->db.db_object, ==,
387 			    db->db.db_object);
388 			/*
389 			 * dnode_grow_indblksz() can make this fail if we don't
390 			 * have the struct_rwlock.  XXX indblksz no longer
391 			 * grows.  safe to do this now?
392 			 */
393 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
394 				ASSERT3P(db->db_blkptr, ==,
395 				    ((blkptr_t *)db->db_parent->db.db_data +
396 				    db->db_blkid % epb));
397 			}
398 		}
399 	}
400 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
401 	    (db->db_buf == NULL || db->db_buf->b_data) &&
402 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
403 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
404 		/*
405 		 * If the blkptr isn't set but they have nonzero data,
406 		 * it had better be dirty, otherwise we'll lose that
407 		 * data when we evict this buffer.
408 		 */
409 		if (db->db_dirtycnt == 0) {
410 			uint64_t *buf = db->db.db_data;
411 			int i;
412 
413 			for (i = 0; i < db->db.db_size >> 3; i++) {
414 				ASSERT(buf[i] == 0);
415 			}
416 		}
417 	}
418 	DB_DNODE_EXIT(db);
419 }
420 #endif
421 
422 static void
423 dbuf_update_data(dmu_buf_impl_t *db)
424 {
425 	ASSERT(MUTEX_HELD(&db->db_mtx));
426 	if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
427 		ASSERT(!refcount_is_zero(&db->db_holds));
428 		*db->db_user_data_ptr_ptr = db->db.db_data;
429 	}
430 }
431 
432 static void
433 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
434 {
435 	ASSERT(MUTEX_HELD(&db->db_mtx));
436 	db->db_buf = buf;
437 	if (buf != NULL) {
438 		ASSERT(buf->b_data != NULL);
439 		db->db.db_data = buf->b_data;
440 		if (!arc_released(buf))
441 			arc_set_callback(buf, dbuf_do_evict, db);
442 		dbuf_update_data(db);
443 	} else {
444 		dbuf_evict_user(db);
445 		db->db.db_data = NULL;
446 		if (db->db_state != DB_NOFILL)
447 			db->db_state = DB_UNCACHED;
448 	}
449 }
450 
451 /*
452  * Loan out an arc_buf for read.  Return the loaned arc_buf.
453  */
454 arc_buf_t *
455 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
456 {
457 	arc_buf_t *abuf;
458 
459 	mutex_enter(&db->db_mtx);
460 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
461 		int blksz = db->db.db_size;
462 		spa_t *spa = db->db_objset->os_spa;
463 
464 		mutex_exit(&db->db_mtx);
465 		abuf = arc_loan_buf(spa, blksz);
466 		bcopy(db->db.db_data, abuf->b_data, blksz);
467 	} else {
468 		abuf = db->db_buf;
469 		arc_loan_inuse_buf(abuf, db);
470 		dbuf_set_data(db, NULL);
471 		mutex_exit(&db->db_mtx);
472 	}
473 	return (abuf);
474 }
475 
476 uint64_t
477 dbuf_whichblock(dnode_t *dn, uint64_t offset)
478 {
479 	if (dn->dn_datablkshift) {
480 		return (offset >> dn->dn_datablkshift);
481 	} else {
482 		ASSERT3U(offset, <, dn->dn_datablksz);
483 		return (0);
484 	}
485 }
486 
487 static void
488 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
489 {
490 	dmu_buf_impl_t *db = vdb;
491 
492 	mutex_enter(&db->db_mtx);
493 	ASSERT3U(db->db_state, ==, DB_READ);
494 	/*
495 	 * All reads are synchronous, so we must have a hold on the dbuf
496 	 */
497 	ASSERT(refcount_count(&db->db_holds) > 0);
498 	ASSERT(db->db_buf == NULL);
499 	ASSERT(db->db.db_data == NULL);
500 	if (db->db_level == 0 && db->db_freed_in_flight) {
501 		/* we were freed in flight; disregard any error */
502 		arc_release(buf, db);
503 		bzero(buf->b_data, db->db.db_size);
504 		arc_buf_freeze(buf);
505 		db->db_freed_in_flight = FALSE;
506 		dbuf_set_data(db, buf);
507 		db->db_state = DB_CACHED;
508 	} else if (zio == NULL || zio->io_error == 0) {
509 		dbuf_set_data(db, buf);
510 		db->db_state = DB_CACHED;
511 	} else {
512 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
513 		ASSERT3P(db->db_buf, ==, NULL);
514 		VERIFY(arc_buf_remove_ref(buf, db));
515 		db->db_state = DB_UNCACHED;
516 	}
517 	cv_broadcast(&db->db_changed);
518 	dbuf_rele_and_unlock(db, NULL);
519 }
520 
521 static void
522 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
523 {
524 	dnode_t *dn;
525 	zbookmark_phys_t zb;
526 	uint32_t aflags = ARC_NOWAIT;
527 
528 	DB_DNODE_ENTER(db);
529 	dn = DB_DNODE(db);
530 	ASSERT(!refcount_is_zero(&db->db_holds));
531 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
532 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
533 	ASSERT(MUTEX_HELD(&db->db_mtx));
534 	ASSERT(db->db_state == DB_UNCACHED);
535 	ASSERT(db->db_buf == NULL);
536 
537 	if (db->db_blkid == DMU_BONUS_BLKID) {
538 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
539 
540 		ASSERT3U(bonuslen, <=, db->db.db_size);
541 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
542 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
543 		if (bonuslen < DN_MAX_BONUSLEN)
544 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
545 		if (bonuslen)
546 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
547 		DB_DNODE_EXIT(db);
548 		dbuf_update_data(db);
549 		db->db_state = DB_CACHED;
550 		mutex_exit(&db->db_mtx);
551 		return;
552 	}
553 
554 	/*
555 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
556 	 * processes the delete record and clears the bp while we are waiting
557 	 * for the dn_mtx (resulting in a "no" from block_freed).
558 	 */
559 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
560 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
561 	    BP_IS_HOLE(db->db_blkptr)))) {
562 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
563 
564 		DB_DNODE_EXIT(db);
565 		dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
566 		    db->db.db_size, db, type));
567 		bzero(db->db.db_data, db->db.db_size);
568 		db->db_state = DB_CACHED;
569 		*flags |= DB_RF_CACHED;
570 		mutex_exit(&db->db_mtx);
571 		return;
572 	}
573 
574 	DB_DNODE_EXIT(db);
575 
576 	db->db_state = DB_READ;
577 	mutex_exit(&db->db_mtx);
578 
579 	if (DBUF_IS_L2CACHEABLE(db))
580 		aflags |= ARC_L2CACHE;
581 	if (DBUF_IS_L2COMPRESSIBLE(db))
582 		aflags |= ARC_L2COMPRESS;
583 
584 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
585 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
586 	    db->db.db_object, db->db_level, db->db_blkid);
587 
588 	dbuf_add_ref(db, NULL);
589 
590 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
591 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
592 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
593 	    &aflags, &zb);
594 	if (aflags & ARC_CACHED)
595 		*flags |= DB_RF_CACHED;
596 }
597 
598 int
599 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
600 {
601 	int err = 0;
602 	boolean_t havepzio = (zio != NULL);
603 	boolean_t prefetch;
604 	dnode_t *dn;
605 
606 	/*
607 	 * We don't have to hold the mutex to check db_state because it
608 	 * can't be freed while we have a hold on the buffer.
609 	 */
610 	ASSERT(!refcount_is_zero(&db->db_holds));
611 
612 	if (db->db_state == DB_NOFILL)
613 		return (SET_ERROR(EIO));
614 
615 	DB_DNODE_ENTER(db);
616 	dn = DB_DNODE(db);
617 	if ((flags & DB_RF_HAVESTRUCT) == 0)
618 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
619 
620 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
621 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
622 	    DBUF_IS_CACHEABLE(db);
623 
624 	mutex_enter(&db->db_mtx);
625 	if (db->db_state == DB_CACHED) {
626 		mutex_exit(&db->db_mtx);
627 		if (prefetch)
628 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
629 			    db->db.db_size, TRUE);
630 		if ((flags & DB_RF_HAVESTRUCT) == 0)
631 			rw_exit(&dn->dn_struct_rwlock);
632 		DB_DNODE_EXIT(db);
633 	} else if (db->db_state == DB_UNCACHED) {
634 		spa_t *spa = dn->dn_objset->os_spa;
635 
636 		if (zio == NULL)
637 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
638 		dbuf_read_impl(db, zio, &flags);
639 
640 		/* dbuf_read_impl has dropped db_mtx for us */
641 
642 		if (prefetch)
643 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
644 			    db->db.db_size, flags & DB_RF_CACHED);
645 
646 		if ((flags & DB_RF_HAVESTRUCT) == 0)
647 			rw_exit(&dn->dn_struct_rwlock);
648 		DB_DNODE_EXIT(db);
649 
650 		if (!havepzio)
651 			err = zio_wait(zio);
652 	} else {
653 		/*
654 		 * Another reader came in while the dbuf was in flight
655 		 * between UNCACHED and CACHED.  Either a writer will finish
656 		 * writing the buffer (sending the dbuf to CACHED) or the
657 		 * first reader's request will reach the read_done callback
658 		 * and send the dbuf to CACHED.  Otherwise, a failure
659 		 * occurred and the dbuf went to UNCACHED.
660 		 */
661 		mutex_exit(&db->db_mtx);
662 		if (prefetch)
663 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
664 			    db->db.db_size, TRUE);
665 		if ((flags & DB_RF_HAVESTRUCT) == 0)
666 			rw_exit(&dn->dn_struct_rwlock);
667 		DB_DNODE_EXIT(db);
668 
669 		/* Skip the wait per the caller's request. */
670 		mutex_enter(&db->db_mtx);
671 		if ((flags & DB_RF_NEVERWAIT) == 0) {
672 			while (db->db_state == DB_READ ||
673 			    db->db_state == DB_FILL) {
674 				ASSERT(db->db_state == DB_READ ||
675 				    (flags & DB_RF_HAVESTRUCT) == 0);
676 				cv_wait(&db->db_changed, &db->db_mtx);
677 			}
678 			if (db->db_state == DB_UNCACHED)
679 				err = SET_ERROR(EIO);
680 		}
681 		mutex_exit(&db->db_mtx);
682 	}
683 
684 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
685 	return (err);
686 }
687 
688 static void
689 dbuf_noread(dmu_buf_impl_t *db)
690 {
691 	ASSERT(!refcount_is_zero(&db->db_holds));
692 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
693 	mutex_enter(&db->db_mtx);
694 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
695 		cv_wait(&db->db_changed, &db->db_mtx);
696 	if (db->db_state == DB_UNCACHED) {
697 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
698 		spa_t *spa = db->db_objset->os_spa;
699 
700 		ASSERT(db->db_buf == NULL);
701 		ASSERT(db->db.db_data == NULL);
702 		dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
703 		db->db_state = DB_FILL;
704 	} else if (db->db_state == DB_NOFILL) {
705 		dbuf_set_data(db, NULL);
706 	} else {
707 		ASSERT3U(db->db_state, ==, DB_CACHED);
708 	}
709 	mutex_exit(&db->db_mtx);
710 }
711 
712 /*
713  * This is our just-in-time copy function.  It makes a copy of
714  * buffers, that have been modified in a previous transaction
715  * group, before we modify them in the current active group.
716  *
717  * This function is used in two places: when we are dirtying a
718  * buffer for the first time in a txg, and when we are freeing
719  * a range in a dnode that includes this buffer.
720  *
721  * Note that when we are called from dbuf_free_range() we do
722  * not put a hold on the buffer, we just traverse the active
723  * dbuf list for the dnode.
724  */
725 static void
726 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
727 {
728 	dbuf_dirty_record_t *dr = db->db_last_dirty;
729 
730 	ASSERT(MUTEX_HELD(&db->db_mtx));
731 	ASSERT(db->db.db_data != NULL);
732 	ASSERT(db->db_level == 0);
733 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
734 
735 	if (dr == NULL ||
736 	    (dr->dt.dl.dr_data !=
737 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
738 		return;
739 
740 	/*
741 	 * If the last dirty record for this dbuf has not yet synced
742 	 * and its referencing the dbuf data, either:
743 	 *	reset the reference to point to a new copy,
744 	 * or (if there a no active holders)
745 	 *	just null out the current db_data pointer.
746 	 */
747 	ASSERT(dr->dr_txg >= txg - 2);
748 	if (db->db_blkid == DMU_BONUS_BLKID) {
749 		/* Note that the data bufs here are zio_bufs */
750 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
751 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
752 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
753 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
754 		int size = db->db.db_size;
755 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
756 		spa_t *spa = db->db_objset->os_spa;
757 
758 		dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
759 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
760 	} else {
761 		dbuf_set_data(db, NULL);
762 	}
763 }
764 
765 void
766 dbuf_unoverride(dbuf_dirty_record_t *dr)
767 {
768 	dmu_buf_impl_t *db = dr->dr_dbuf;
769 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
770 	uint64_t txg = dr->dr_txg;
771 
772 	ASSERT(MUTEX_HELD(&db->db_mtx));
773 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
774 	ASSERT(db->db_level == 0);
775 
776 	if (db->db_blkid == DMU_BONUS_BLKID ||
777 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
778 		return;
779 
780 	ASSERT(db->db_data_pending != dr);
781 
782 	/* free this block */
783 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
784 		zio_free(db->db_objset->os_spa, txg, bp);
785 
786 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
787 	dr->dt.dl.dr_nopwrite = B_FALSE;
788 
789 	/*
790 	 * Release the already-written buffer, so we leave it in
791 	 * a consistent dirty state.  Note that all callers are
792 	 * modifying the buffer, so they will immediately do
793 	 * another (redundant) arc_release().  Therefore, leave
794 	 * the buf thawed to save the effort of freezing &
795 	 * immediately re-thawing it.
796 	 */
797 	arc_release(dr->dt.dl.dr_data, db);
798 }
799 
800 /*
801  * Evict (if its unreferenced) or clear (if its referenced) any level-0
802  * data blocks in the free range, so that any future readers will find
803  * empty blocks.
804  *
805  * This is a no-op if the dataset is in the middle of an incremental
806  * receive; see comment below for details.
807  */
808 void
809 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
810     dmu_tx_t *tx)
811 {
812 	dmu_buf_impl_t *db, *db_next, db_search;
813 	uint64_t txg = tx->tx_txg;
814 	avl_index_t where;
815 
816 	if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
817 		end_blkid = dn->dn_maxblkid;
818 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
819 
820 	db_search.db_level = 0;
821 	db_search.db_blkid = start_blkid;
822 	db_search.db_creation = 0;
823 
824 	mutex_enter(&dn->dn_dbufs_mtx);
825 	if (start_blkid >= dn->dn_unlisted_l0_blkid) {
826 		/* There can't be any dbufs in this range; no need to search. */
827 #ifdef DEBUG
828 		db = avl_find(&dn->dn_dbufs, &db_search, &where);
829 		ASSERT3P(db, ==, NULL);
830 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
831 		ASSERT(db == NULL || db->db_level > 0);
832 #endif
833 		mutex_exit(&dn->dn_dbufs_mtx);
834 		return;
835 	} else if (dmu_objset_is_receiving(dn->dn_objset)) {
836 		/*
837 		 * If we are receiving, we expect there to be no dbufs in
838 		 * the range to be freed, because receive modifies each
839 		 * block at most once, and in offset order.  If this is
840 		 * not the case, it can lead to performance problems,
841 		 * so note that we unexpectedly took the slow path.
842 		 */
843 		atomic_inc_64(&zfs_free_range_recv_miss);
844 	}
845 
846 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
847 	ASSERT3P(db, ==, NULL);
848 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
849 
850 	for (; db != NULL; db = db_next) {
851 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
852 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
853 
854 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
855 			break;
856 		}
857 		ASSERT3U(db->db_blkid, >=, start_blkid);
858 
859 		/* found a level 0 buffer in the range */
860 		mutex_enter(&db->db_mtx);
861 		if (dbuf_undirty(db, tx)) {
862 			/* mutex has been dropped and dbuf destroyed */
863 			continue;
864 		}
865 
866 		if (db->db_state == DB_UNCACHED ||
867 		    db->db_state == DB_NOFILL ||
868 		    db->db_state == DB_EVICTING) {
869 			ASSERT(db->db.db_data == NULL);
870 			mutex_exit(&db->db_mtx);
871 			continue;
872 		}
873 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
874 			/* will be handled in dbuf_read_done or dbuf_rele */
875 			db->db_freed_in_flight = TRUE;
876 			mutex_exit(&db->db_mtx);
877 			continue;
878 		}
879 		if (refcount_count(&db->db_holds) == 0) {
880 			ASSERT(db->db_buf);
881 			dbuf_clear(db);
882 			continue;
883 		}
884 		/* The dbuf is referenced */
885 
886 		if (db->db_last_dirty != NULL) {
887 			dbuf_dirty_record_t *dr = db->db_last_dirty;
888 
889 			if (dr->dr_txg == txg) {
890 				/*
891 				 * This buffer is "in-use", re-adjust the file
892 				 * size to reflect that this buffer may
893 				 * contain new data when we sync.
894 				 */
895 				if (db->db_blkid != DMU_SPILL_BLKID &&
896 				    db->db_blkid > dn->dn_maxblkid)
897 					dn->dn_maxblkid = db->db_blkid;
898 				dbuf_unoverride(dr);
899 			} else {
900 				/*
901 				 * This dbuf is not dirty in the open context.
902 				 * Either uncache it (if its not referenced in
903 				 * the open context) or reset its contents to
904 				 * empty.
905 				 */
906 				dbuf_fix_old_data(db, txg);
907 			}
908 		}
909 		/* clear the contents if its cached */
910 		if (db->db_state == DB_CACHED) {
911 			ASSERT(db->db.db_data != NULL);
912 			arc_release(db->db_buf, db);
913 			bzero(db->db.db_data, db->db.db_size);
914 			arc_buf_freeze(db->db_buf);
915 		}
916 
917 		mutex_exit(&db->db_mtx);
918 	}
919 	mutex_exit(&dn->dn_dbufs_mtx);
920 }
921 
922 static int
923 dbuf_block_freeable(dmu_buf_impl_t *db)
924 {
925 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
926 	uint64_t birth_txg = 0;
927 
928 	/*
929 	 * We don't need any locking to protect db_blkptr:
930 	 * If it's syncing, then db_last_dirty will be set
931 	 * so we'll ignore db_blkptr.
932 	 *
933 	 * This logic ensures that only block births for
934 	 * filled blocks are considered.
935 	 */
936 	ASSERT(MUTEX_HELD(&db->db_mtx));
937 	if (db->db_last_dirty && (db->db_blkptr == NULL ||
938 	    !BP_IS_HOLE(db->db_blkptr))) {
939 		birth_txg = db->db_last_dirty->dr_txg;
940 	} else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
941 		birth_txg = db->db_blkptr->blk_birth;
942 	}
943 
944 	/*
945 	 * If this block don't exist or is in a snapshot, it can't be freed.
946 	 * Don't pass the bp to dsl_dataset_block_freeable() since we
947 	 * are holding the db_mtx lock and might deadlock if we are
948 	 * prefetching a dedup-ed block.
949 	 */
950 	if (birth_txg != 0)
951 		return (ds == NULL ||
952 		    dsl_dataset_block_freeable(ds, NULL, birth_txg));
953 	else
954 		return (B_FALSE);
955 }
956 
957 void
958 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
959 {
960 	arc_buf_t *buf, *obuf;
961 	int osize = db->db.db_size;
962 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
963 	dnode_t *dn;
964 
965 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
966 
967 	DB_DNODE_ENTER(db);
968 	dn = DB_DNODE(db);
969 
970 	/* XXX does *this* func really need the lock? */
971 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
972 
973 	/*
974 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
975 	 * is OK, because there can be no other references to the db
976 	 * when we are changing its size, so no concurrent DB_FILL can
977 	 * be happening.
978 	 */
979 	/*
980 	 * XXX we should be doing a dbuf_read, checking the return
981 	 * value and returning that up to our callers
982 	 */
983 	dmu_buf_will_dirty(&db->db, tx);
984 
985 	/* create the data buffer for the new block */
986 	buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
987 
988 	/* copy old block data to the new block */
989 	obuf = db->db_buf;
990 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
991 	/* zero the remainder */
992 	if (size > osize)
993 		bzero((uint8_t *)buf->b_data + osize, size - osize);
994 
995 	mutex_enter(&db->db_mtx);
996 	dbuf_set_data(db, buf);
997 	VERIFY(arc_buf_remove_ref(obuf, db));
998 	db->db.db_size = size;
999 
1000 	if (db->db_level == 0) {
1001 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1002 		db->db_last_dirty->dt.dl.dr_data = buf;
1003 	}
1004 	mutex_exit(&db->db_mtx);
1005 
1006 	dnode_willuse_space(dn, size-osize, tx);
1007 	DB_DNODE_EXIT(db);
1008 }
1009 
1010 void
1011 dbuf_release_bp(dmu_buf_impl_t *db)
1012 {
1013 	objset_t *os = db->db_objset;
1014 
1015 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1016 	ASSERT(arc_released(os->os_phys_buf) ||
1017 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
1018 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1019 
1020 	(void) arc_release(db->db_buf, db);
1021 }
1022 
1023 dbuf_dirty_record_t *
1024 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1025 {
1026 	dnode_t *dn;
1027 	objset_t *os;
1028 	dbuf_dirty_record_t **drp, *dr;
1029 	int drop_struct_lock = FALSE;
1030 	boolean_t do_free_accounting = B_FALSE;
1031 	int txgoff = tx->tx_txg & TXG_MASK;
1032 
1033 	ASSERT(tx->tx_txg != 0);
1034 	ASSERT(!refcount_is_zero(&db->db_holds));
1035 	DMU_TX_DIRTY_BUF(tx, db);
1036 
1037 	DB_DNODE_ENTER(db);
1038 	dn = DB_DNODE(db);
1039 	/*
1040 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1041 	 * objects may be dirtied in syncing context, but only if they
1042 	 * were already pre-dirtied in open context.
1043 	 */
1044 	ASSERT(!dmu_tx_is_syncing(tx) ||
1045 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1046 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1047 	    dn->dn_objset->os_dsl_dataset == NULL);
1048 	/*
1049 	 * We make this assert for private objects as well, but after we
1050 	 * check if we're already dirty.  They are allowed to re-dirty
1051 	 * in syncing context.
1052 	 */
1053 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1054 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1055 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1056 
1057 	mutex_enter(&db->db_mtx);
1058 	/*
1059 	 * XXX make this true for indirects too?  The problem is that
1060 	 * transactions created with dmu_tx_create_assigned() from
1061 	 * syncing context don't bother holding ahead.
1062 	 */
1063 	ASSERT(db->db_level != 0 ||
1064 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1065 	    db->db_state == DB_NOFILL);
1066 
1067 	mutex_enter(&dn->dn_mtx);
1068 	/*
1069 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1070 	 * initialize the objset.
1071 	 */
1072 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1073 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1074 		dn->dn_dirtyctx =
1075 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1076 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
1077 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1078 	}
1079 	mutex_exit(&dn->dn_mtx);
1080 
1081 	if (db->db_blkid == DMU_SPILL_BLKID)
1082 		dn->dn_have_spill = B_TRUE;
1083 
1084 	/*
1085 	 * If this buffer is already dirty, we're done.
1086 	 */
1087 	drp = &db->db_last_dirty;
1088 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1089 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1090 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1091 		drp = &dr->dr_next;
1092 	if (dr && dr->dr_txg == tx->tx_txg) {
1093 		DB_DNODE_EXIT(db);
1094 
1095 		if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1096 			/*
1097 			 * If this buffer has already been written out,
1098 			 * we now need to reset its state.
1099 			 */
1100 			dbuf_unoverride(dr);
1101 			if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1102 			    db->db_state != DB_NOFILL)
1103 				arc_buf_thaw(db->db_buf);
1104 		}
1105 		mutex_exit(&db->db_mtx);
1106 		return (dr);
1107 	}
1108 
1109 	/*
1110 	 * Only valid if not already dirty.
1111 	 */
1112 	ASSERT(dn->dn_object == 0 ||
1113 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1114 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1115 
1116 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1117 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1118 	    dn->dn_phys->dn_nlevels > db->db_level ||
1119 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1120 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1121 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1122 
1123 	/*
1124 	 * We should only be dirtying in syncing context if it's the
1125 	 * mos or we're initializing the os or it's a special object.
1126 	 * However, we are allowed to dirty in syncing context provided
1127 	 * we already dirtied it in open context.  Hence we must make
1128 	 * this assertion only if we're not already dirty.
1129 	 */
1130 	os = dn->dn_objset;
1131 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1132 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1133 	ASSERT(db->db.db_size != 0);
1134 
1135 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1136 
1137 	if (db->db_blkid != DMU_BONUS_BLKID) {
1138 		/*
1139 		 * Update the accounting.
1140 		 * Note: we delay "free accounting" until after we drop
1141 		 * the db_mtx.  This keeps us from grabbing other locks
1142 		 * (and possibly deadlocking) in bp_get_dsize() while
1143 		 * also holding the db_mtx.
1144 		 */
1145 		dnode_willuse_space(dn, db->db.db_size, tx);
1146 		do_free_accounting = dbuf_block_freeable(db);
1147 	}
1148 
1149 	/*
1150 	 * If this buffer is dirty in an old transaction group we need
1151 	 * to make a copy of it so that the changes we make in this
1152 	 * transaction group won't leak out when we sync the older txg.
1153 	 */
1154 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1155 	if (db->db_level == 0) {
1156 		void *data_old = db->db_buf;
1157 
1158 		if (db->db_state != DB_NOFILL) {
1159 			if (db->db_blkid == DMU_BONUS_BLKID) {
1160 				dbuf_fix_old_data(db, tx->tx_txg);
1161 				data_old = db->db.db_data;
1162 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1163 				/*
1164 				 * Release the data buffer from the cache so
1165 				 * that we can modify it without impacting
1166 				 * possible other users of this cached data
1167 				 * block.  Note that indirect blocks and
1168 				 * private objects are not released until the
1169 				 * syncing state (since they are only modified
1170 				 * then).
1171 				 */
1172 				arc_release(db->db_buf, db);
1173 				dbuf_fix_old_data(db, tx->tx_txg);
1174 				data_old = db->db_buf;
1175 			}
1176 			ASSERT(data_old != NULL);
1177 		}
1178 		dr->dt.dl.dr_data = data_old;
1179 	} else {
1180 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1181 		list_create(&dr->dt.di.dr_children,
1182 		    sizeof (dbuf_dirty_record_t),
1183 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1184 	}
1185 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1186 		dr->dr_accounted = db->db.db_size;
1187 	dr->dr_dbuf = db;
1188 	dr->dr_txg = tx->tx_txg;
1189 	dr->dr_next = *drp;
1190 	*drp = dr;
1191 
1192 	/*
1193 	 * We could have been freed_in_flight between the dbuf_noread
1194 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1195 	 * happened after the free.
1196 	 */
1197 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1198 	    db->db_blkid != DMU_SPILL_BLKID) {
1199 		mutex_enter(&dn->dn_mtx);
1200 		if (dn->dn_free_ranges[txgoff] != NULL) {
1201 			range_tree_clear(dn->dn_free_ranges[txgoff],
1202 			    db->db_blkid, 1);
1203 		}
1204 		mutex_exit(&dn->dn_mtx);
1205 		db->db_freed_in_flight = FALSE;
1206 	}
1207 
1208 	/*
1209 	 * This buffer is now part of this txg
1210 	 */
1211 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1212 	db->db_dirtycnt += 1;
1213 	ASSERT3U(db->db_dirtycnt, <=, 3);
1214 
1215 	mutex_exit(&db->db_mtx);
1216 
1217 	if (db->db_blkid == DMU_BONUS_BLKID ||
1218 	    db->db_blkid == DMU_SPILL_BLKID) {
1219 		mutex_enter(&dn->dn_mtx);
1220 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1221 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1222 		mutex_exit(&dn->dn_mtx);
1223 		dnode_setdirty(dn, tx);
1224 		DB_DNODE_EXIT(db);
1225 		return (dr);
1226 	} else if (do_free_accounting) {
1227 		blkptr_t *bp = db->db_blkptr;
1228 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1229 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1230 		/*
1231 		 * This is only a guess -- if the dbuf is dirty
1232 		 * in a previous txg, we don't know how much
1233 		 * space it will use on disk yet.  We should
1234 		 * really have the struct_rwlock to access
1235 		 * db_blkptr, but since this is just a guess,
1236 		 * it's OK if we get an odd answer.
1237 		 */
1238 		ddt_prefetch(os->os_spa, bp);
1239 		dnode_willuse_space(dn, -willfree, tx);
1240 	}
1241 
1242 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1243 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1244 		drop_struct_lock = TRUE;
1245 	}
1246 
1247 	if (db->db_level == 0) {
1248 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1249 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1250 	}
1251 
1252 	if (db->db_level+1 < dn->dn_nlevels) {
1253 		dmu_buf_impl_t *parent = db->db_parent;
1254 		dbuf_dirty_record_t *di;
1255 		int parent_held = FALSE;
1256 
1257 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1258 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1259 
1260 			parent = dbuf_hold_level(dn, db->db_level+1,
1261 			    db->db_blkid >> epbs, FTAG);
1262 			ASSERT(parent != NULL);
1263 			parent_held = TRUE;
1264 		}
1265 		if (drop_struct_lock)
1266 			rw_exit(&dn->dn_struct_rwlock);
1267 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1268 		di = dbuf_dirty(parent, tx);
1269 		if (parent_held)
1270 			dbuf_rele(parent, FTAG);
1271 
1272 		mutex_enter(&db->db_mtx);
1273 		/*
1274 		 * Since we've dropped the mutex, it's possible that
1275 		 * dbuf_undirty() might have changed this out from under us.
1276 		 */
1277 		if (db->db_last_dirty == dr ||
1278 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1279 			mutex_enter(&di->dt.di.dr_mtx);
1280 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1281 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1282 			list_insert_tail(&di->dt.di.dr_children, dr);
1283 			mutex_exit(&di->dt.di.dr_mtx);
1284 			dr->dr_parent = di;
1285 		}
1286 		mutex_exit(&db->db_mtx);
1287 	} else {
1288 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1289 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1290 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1291 		mutex_enter(&dn->dn_mtx);
1292 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1293 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1294 		mutex_exit(&dn->dn_mtx);
1295 		if (drop_struct_lock)
1296 			rw_exit(&dn->dn_struct_rwlock);
1297 	}
1298 
1299 	dnode_setdirty(dn, tx);
1300 	DB_DNODE_EXIT(db);
1301 	return (dr);
1302 }
1303 
1304 /*
1305  * Undirty a buffer in the transaction group referenced by the given
1306  * transaction.  Return whether this evicted the dbuf.
1307  */
1308 static boolean_t
1309 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1310 {
1311 	dnode_t *dn;
1312 	uint64_t txg = tx->tx_txg;
1313 	dbuf_dirty_record_t *dr, **drp;
1314 
1315 	ASSERT(txg != 0);
1316 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1317 	ASSERT0(db->db_level);
1318 	ASSERT(MUTEX_HELD(&db->db_mtx));
1319 
1320 	/*
1321 	 * If this buffer is not dirty, we're done.
1322 	 */
1323 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1324 		if (dr->dr_txg <= txg)
1325 			break;
1326 	if (dr == NULL || dr->dr_txg < txg)
1327 		return (B_FALSE);
1328 	ASSERT(dr->dr_txg == txg);
1329 	ASSERT(dr->dr_dbuf == db);
1330 
1331 	DB_DNODE_ENTER(db);
1332 	dn = DB_DNODE(db);
1333 
1334 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1335 
1336 	ASSERT(db->db.db_size != 0);
1337 
1338 	/*
1339 	 * Any space we accounted for in dp_dirty_* will be cleaned up by
1340 	 * dsl_pool_sync().  This is relatively rare so the discrepancy
1341 	 * is not a big deal.
1342 	 */
1343 
1344 	*drp = dr->dr_next;
1345 
1346 	/*
1347 	 * Note that there are three places in dbuf_dirty()
1348 	 * where this dirty record may be put on a list.
1349 	 * Make sure to do a list_remove corresponding to
1350 	 * every one of those list_insert calls.
1351 	 */
1352 	if (dr->dr_parent) {
1353 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1354 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1355 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1356 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1357 	    db->db_level+1 == dn->dn_nlevels) {
1358 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1359 		mutex_enter(&dn->dn_mtx);
1360 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1361 		mutex_exit(&dn->dn_mtx);
1362 	}
1363 	DB_DNODE_EXIT(db);
1364 
1365 	if (db->db_state != DB_NOFILL) {
1366 		dbuf_unoverride(dr);
1367 
1368 		ASSERT(db->db_buf != NULL);
1369 		ASSERT(dr->dt.dl.dr_data != NULL);
1370 		if (dr->dt.dl.dr_data != db->db_buf)
1371 			VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1372 	}
1373 
1374 	if (db->db_level != 0) {
1375 		mutex_destroy(&dr->dt.di.dr_mtx);
1376 		list_destroy(&dr->dt.di.dr_children);
1377 	}
1378 
1379 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1380 
1381 	ASSERT(db->db_dirtycnt > 0);
1382 	db->db_dirtycnt -= 1;
1383 
1384 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1385 		arc_buf_t *buf = db->db_buf;
1386 
1387 		ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1388 		dbuf_set_data(db, NULL);
1389 		VERIFY(arc_buf_remove_ref(buf, db));
1390 		dbuf_evict(db);
1391 		return (B_TRUE);
1392 	}
1393 
1394 	return (B_FALSE);
1395 }
1396 
1397 void
1398 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1399 {
1400 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1401 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1402 
1403 	ASSERT(tx->tx_txg != 0);
1404 	ASSERT(!refcount_is_zero(&db->db_holds));
1405 
1406 	DB_DNODE_ENTER(db);
1407 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1408 		rf |= DB_RF_HAVESTRUCT;
1409 	DB_DNODE_EXIT(db);
1410 	(void) dbuf_read(db, NULL, rf);
1411 	(void) dbuf_dirty(db, tx);
1412 }
1413 
1414 void
1415 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1416 {
1417 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1418 
1419 	db->db_state = DB_NOFILL;
1420 
1421 	dmu_buf_will_fill(db_fake, tx);
1422 }
1423 
1424 void
1425 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1426 {
1427 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1428 
1429 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1430 	ASSERT(tx->tx_txg != 0);
1431 	ASSERT(db->db_level == 0);
1432 	ASSERT(!refcount_is_zero(&db->db_holds));
1433 
1434 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1435 	    dmu_tx_private_ok(tx));
1436 
1437 	dbuf_noread(db);
1438 	(void) dbuf_dirty(db, tx);
1439 }
1440 
1441 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1442 /* ARGSUSED */
1443 void
1444 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1445 {
1446 	mutex_enter(&db->db_mtx);
1447 	DBUF_VERIFY(db);
1448 
1449 	if (db->db_state == DB_FILL) {
1450 		if (db->db_level == 0 && db->db_freed_in_flight) {
1451 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1452 			/* we were freed while filling */
1453 			/* XXX dbuf_undirty? */
1454 			bzero(db->db.db_data, db->db.db_size);
1455 			db->db_freed_in_flight = FALSE;
1456 		}
1457 		db->db_state = DB_CACHED;
1458 		cv_broadcast(&db->db_changed);
1459 	}
1460 	mutex_exit(&db->db_mtx);
1461 }
1462 
1463 void
1464 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1465     bp_embedded_type_t etype, enum zio_compress comp,
1466     int uncompressed_size, int compressed_size, int byteorder,
1467     dmu_tx_t *tx)
1468 {
1469 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1470 	struct dirty_leaf *dl;
1471 	dmu_object_type_t type;
1472 
1473 	DB_DNODE_ENTER(db);
1474 	type = DB_DNODE(db)->dn_type;
1475 	DB_DNODE_EXIT(db);
1476 
1477 	ASSERT0(db->db_level);
1478 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1479 
1480 	dmu_buf_will_not_fill(dbuf, tx);
1481 
1482 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1483 	dl = &db->db_last_dirty->dt.dl;
1484 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
1485 	    data, comp, uncompressed_size, compressed_size);
1486 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1487 	BP_SET_TYPE(&dl->dr_overridden_by, type);
1488 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1489 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1490 
1491 	dl->dr_override_state = DR_OVERRIDDEN;
1492 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1493 }
1494 
1495 /*
1496  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1497  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1498  */
1499 void
1500 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1501 {
1502 	ASSERT(!refcount_is_zero(&db->db_holds));
1503 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1504 	ASSERT(db->db_level == 0);
1505 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1506 	ASSERT(buf != NULL);
1507 	ASSERT(arc_buf_size(buf) == db->db.db_size);
1508 	ASSERT(tx->tx_txg != 0);
1509 
1510 	arc_return_buf(buf, db);
1511 	ASSERT(arc_released(buf));
1512 
1513 	mutex_enter(&db->db_mtx);
1514 
1515 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1516 		cv_wait(&db->db_changed, &db->db_mtx);
1517 
1518 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1519 
1520 	if (db->db_state == DB_CACHED &&
1521 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1522 		mutex_exit(&db->db_mtx);
1523 		(void) dbuf_dirty(db, tx);
1524 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1525 		VERIFY(arc_buf_remove_ref(buf, db));
1526 		xuio_stat_wbuf_copied();
1527 		return;
1528 	}
1529 
1530 	xuio_stat_wbuf_nocopy();
1531 	if (db->db_state == DB_CACHED) {
1532 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1533 
1534 		ASSERT(db->db_buf != NULL);
1535 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1536 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1537 			if (!arc_released(db->db_buf)) {
1538 				ASSERT(dr->dt.dl.dr_override_state ==
1539 				    DR_OVERRIDDEN);
1540 				arc_release(db->db_buf, db);
1541 			}
1542 			dr->dt.dl.dr_data = buf;
1543 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1544 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1545 			arc_release(db->db_buf, db);
1546 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1547 		}
1548 		db->db_buf = NULL;
1549 	}
1550 	ASSERT(db->db_buf == NULL);
1551 	dbuf_set_data(db, buf);
1552 	db->db_state = DB_FILL;
1553 	mutex_exit(&db->db_mtx);
1554 	(void) dbuf_dirty(db, tx);
1555 	dmu_buf_fill_done(&db->db, tx);
1556 }
1557 
1558 /*
1559  * "Clear" the contents of this dbuf.  This will mark the dbuf
1560  * EVICTING and clear *most* of its references.  Unfortunately,
1561  * when we are not holding the dn_dbufs_mtx, we can't clear the
1562  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1563  * in this case.  For callers from the DMU we will usually see:
1564  *	dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1565  * For the arc callback, we will usually see:
1566  *	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1567  * Sometimes, though, we will get a mix of these two:
1568  *	DMU: dbuf_clear()->arc_clear_callback()
1569  *	ARC: dbuf_do_evict()->dbuf_destroy()
1570  *
1571  * This routine will dissociate the dbuf from the arc, by calling
1572  * arc_clear_callback(), but will not evict the data from the ARC.
1573  */
1574 void
1575 dbuf_clear(dmu_buf_impl_t *db)
1576 {
1577 	dnode_t *dn;
1578 	dmu_buf_impl_t *parent = db->db_parent;
1579 	dmu_buf_impl_t *dndb;
1580 	boolean_t dbuf_gone = B_FALSE;
1581 
1582 	ASSERT(MUTEX_HELD(&db->db_mtx));
1583 	ASSERT(refcount_is_zero(&db->db_holds));
1584 
1585 	dbuf_evict_user(db);
1586 
1587 	if (db->db_state == DB_CACHED) {
1588 		ASSERT(db->db.db_data != NULL);
1589 		if (db->db_blkid == DMU_BONUS_BLKID) {
1590 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1591 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1592 		}
1593 		db->db.db_data = NULL;
1594 		db->db_state = DB_UNCACHED;
1595 	}
1596 
1597 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1598 	ASSERT(db->db_data_pending == NULL);
1599 
1600 	db->db_state = DB_EVICTING;
1601 	db->db_blkptr = NULL;
1602 
1603 	DB_DNODE_ENTER(db);
1604 	dn = DB_DNODE(db);
1605 	dndb = dn->dn_dbuf;
1606 	if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1607 		avl_remove(&dn->dn_dbufs, db);
1608 		(void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1609 		membar_producer();
1610 		DB_DNODE_EXIT(db);
1611 		/*
1612 		 * Decrementing the dbuf count means that the hold corresponding
1613 		 * to the removed dbuf is no longer discounted in dnode_move(),
1614 		 * so the dnode cannot be moved until after we release the hold.
1615 		 * The membar_producer() ensures visibility of the decremented
1616 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1617 		 * release any lock.
1618 		 */
1619 		dnode_rele(dn, db);
1620 		db->db_dnode_handle = NULL;
1621 	} else {
1622 		DB_DNODE_EXIT(db);
1623 	}
1624 
1625 	if (db->db_buf)
1626 		dbuf_gone = arc_clear_callback(db->db_buf);
1627 
1628 	if (!dbuf_gone)
1629 		mutex_exit(&db->db_mtx);
1630 
1631 	/*
1632 	 * If this dbuf is referenced from an indirect dbuf,
1633 	 * decrement the ref count on the indirect dbuf.
1634 	 */
1635 	if (parent && parent != dndb)
1636 		dbuf_rele(parent, db);
1637 }
1638 
1639 static int
1640 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1641     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1642 {
1643 	int nlevels, epbs;
1644 
1645 	*parentp = NULL;
1646 	*bpp = NULL;
1647 
1648 	ASSERT(blkid != DMU_BONUS_BLKID);
1649 
1650 	if (blkid == DMU_SPILL_BLKID) {
1651 		mutex_enter(&dn->dn_mtx);
1652 		if (dn->dn_have_spill &&
1653 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1654 			*bpp = &dn->dn_phys->dn_spill;
1655 		else
1656 			*bpp = NULL;
1657 		dbuf_add_ref(dn->dn_dbuf, NULL);
1658 		*parentp = dn->dn_dbuf;
1659 		mutex_exit(&dn->dn_mtx);
1660 		return (0);
1661 	}
1662 
1663 	if (dn->dn_phys->dn_nlevels == 0)
1664 		nlevels = 1;
1665 	else
1666 		nlevels = dn->dn_phys->dn_nlevels;
1667 
1668 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1669 
1670 	ASSERT3U(level * epbs, <, 64);
1671 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1672 	if (level >= nlevels ||
1673 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1674 		/* the buffer has no parent yet */
1675 		return (SET_ERROR(ENOENT));
1676 	} else if (level < nlevels-1) {
1677 		/* this block is referenced from an indirect block */
1678 		int err = dbuf_hold_impl(dn, level+1,
1679 		    blkid >> epbs, fail_sparse, NULL, parentp);
1680 		if (err)
1681 			return (err);
1682 		err = dbuf_read(*parentp, NULL,
1683 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1684 		if (err) {
1685 			dbuf_rele(*parentp, NULL);
1686 			*parentp = NULL;
1687 			return (err);
1688 		}
1689 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1690 		    (blkid & ((1ULL << epbs) - 1));
1691 		return (0);
1692 	} else {
1693 		/* the block is referenced from the dnode */
1694 		ASSERT3U(level, ==, nlevels-1);
1695 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1696 		    blkid < dn->dn_phys->dn_nblkptr);
1697 		if (dn->dn_dbuf) {
1698 			dbuf_add_ref(dn->dn_dbuf, NULL);
1699 			*parentp = dn->dn_dbuf;
1700 		}
1701 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1702 		return (0);
1703 	}
1704 }
1705 
1706 static dmu_buf_impl_t *
1707 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1708     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1709 {
1710 	objset_t *os = dn->dn_objset;
1711 	dmu_buf_impl_t *db, *odb;
1712 
1713 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1714 	ASSERT(dn->dn_type != DMU_OT_NONE);
1715 
1716 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1717 
1718 	db->db_objset = os;
1719 	db->db.db_object = dn->dn_object;
1720 	db->db_level = level;
1721 	db->db_blkid = blkid;
1722 	db->db_last_dirty = NULL;
1723 	db->db_dirtycnt = 0;
1724 	db->db_dnode_handle = dn->dn_handle;
1725 	db->db_parent = parent;
1726 	db->db_blkptr = blkptr;
1727 
1728 	db->db_user_ptr = NULL;
1729 	db->db_user_data_ptr_ptr = NULL;
1730 	db->db_evict_func = NULL;
1731 	db->db_immediate_evict = 0;
1732 	db->db_freed_in_flight = 0;
1733 
1734 	if (blkid == DMU_BONUS_BLKID) {
1735 		ASSERT3P(parent, ==, dn->dn_dbuf);
1736 		db->db.db_size = DN_MAX_BONUSLEN -
1737 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1738 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1739 		db->db.db_offset = DMU_BONUS_BLKID;
1740 		db->db_state = DB_UNCACHED;
1741 		/* the bonus dbuf is not placed in the hash table */
1742 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1743 		return (db);
1744 	} else if (blkid == DMU_SPILL_BLKID) {
1745 		db->db.db_size = (blkptr != NULL) ?
1746 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1747 		db->db.db_offset = 0;
1748 	} else {
1749 		int blocksize =
1750 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1751 		db->db.db_size = blocksize;
1752 		db->db.db_offset = db->db_blkid * blocksize;
1753 	}
1754 
1755 	/*
1756 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1757 	 * in the hash table *and* added to the dbufs list.
1758 	 * This prevents a possible deadlock with someone
1759 	 * trying to look up this dbuf before its added to the
1760 	 * dn_dbufs list.
1761 	 */
1762 	mutex_enter(&dn->dn_dbufs_mtx);
1763 	db->db_state = DB_EVICTING;
1764 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1765 		/* someone else inserted it first */
1766 		kmem_cache_free(dbuf_cache, db);
1767 		mutex_exit(&dn->dn_dbufs_mtx);
1768 		return (odb);
1769 	}
1770 	avl_add(&dn->dn_dbufs, db);
1771 	if (db->db_level == 0 && db->db_blkid >=
1772 	    dn->dn_unlisted_l0_blkid)
1773 		dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1774 	db->db_state = DB_UNCACHED;
1775 	mutex_exit(&dn->dn_dbufs_mtx);
1776 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1777 
1778 	if (parent && parent != dn->dn_dbuf)
1779 		dbuf_add_ref(parent, db);
1780 
1781 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1782 	    refcount_count(&dn->dn_holds) > 0);
1783 	(void) refcount_add(&dn->dn_holds, db);
1784 	(void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1785 
1786 	dprintf_dbuf(db, "db=%p\n", db);
1787 
1788 	return (db);
1789 }
1790 
1791 static int
1792 dbuf_do_evict(void *private)
1793 {
1794 	dmu_buf_impl_t *db = private;
1795 
1796 	if (!MUTEX_HELD(&db->db_mtx))
1797 		mutex_enter(&db->db_mtx);
1798 
1799 	ASSERT(refcount_is_zero(&db->db_holds));
1800 
1801 	if (db->db_state != DB_EVICTING) {
1802 		ASSERT(db->db_state == DB_CACHED);
1803 		DBUF_VERIFY(db);
1804 		db->db_buf = NULL;
1805 		dbuf_evict(db);
1806 	} else {
1807 		mutex_exit(&db->db_mtx);
1808 		dbuf_destroy(db);
1809 	}
1810 	return (0);
1811 }
1812 
1813 static void
1814 dbuf_destroy(dmu_buf_impl_t *db)
1815 {
1816 	ASSERT(refcount_is_zero(&db->db_holds));
1817 
1818 	if (db->db_blkid != DMU_BONUS_BLKID) {
1819 		/*
1820 		 * If this dbuf is still on the dn_dbufs list,
1821 		 * remove it from that list.
1822 		 */
1823 		if (db->db_dnode_handle != NULL) {
1824 			dnode_t *dn;
1825 
1826 			DB_DNODE_ENTER(db);
1827 			dn = DB_DNODE(db);
1828 			mutex_enter(&dn->dn_dbufs_mtx);
1829 			avl_remove(&dn->dn_dbufs, db);
1830 			(void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1831 			mutex_exit(&dn->dn_dbufs_mtx);
1832 			DB_DNODE_EXIT(db);
1833 			/*
1834 			 * Decrementing the dbuf count means that the hold
1835 			 * corresponding to the removed dbuf is no longer
1836 			 * discounted in dnode_move(), so the dnode cannot be
1837 			 * moved until after we release the hold.
1838 			 */
1839 			dnode_rele(dn, db);
1840 			db->db_dnode_handle = NULL;
1841 		}
1842 		dbuf_hash_remove(db);
1843 	}
1844 	db->db_parent = NULL;
1845 	db->db_buf = NULL;
1846 
1847 	ASSERT(db->db.db_data == NULL);
1848 	ASSERT(db->db_hash_next == NULL);
1849 	ASSERT(db->db_blkptr == NULL);
1850 	ASSERT(db->db_data_pending == NULL);
1851 
1852 	kmem_cache_free(dbuf_cache, db);
1853 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1854 }
1855 
1856 void
1857 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1858 {
1859 	dmu_buf_impl_t *db = NULL;
1860 	blkptr_t *bp = NULL;
1861 
1862 	ASSERT(blkid != DMU_BONUS_BLKID);
1863 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1864 
1865 	if (dnode_block_freed(dn, blkid))
1866 		return;
1867 
1868 	/* dbuf_find() returns with db_mtx held */
1869 	if (db = dbuf_find(dn, 0, blkid)) {
1870 		/*
1871 		 * This dbuf is already in the cache.  We assume that
1872 		 * it is already CACHED, or else about to be either
1873 		 * read or filled.
1874 		 */
1875 		mutex_exit(&db->db_mtx);
1876 		return;
1877 	}
1878 
1879 	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1880 		if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
1881 			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1882 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1883 			zbookmark_phys_t zb;
1884 
1885 			SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1886 			    dn->dn_object, 0, blkid);
1887 
1888 			(void) arc_read(NULL, dn->dn_objset->os_spa,
1889 			    bp, NULL, NULL, prio,
1890 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1891 			    &aflags, &zb);
1892 		}
1893 		if (db)
1894 			dbuf_rele(db, NULL);
1895 	}
1896 }
1897 
1898 /*
1899  * Returns with db_holds incremented, and db_mtx not held.
1900  * Note: dn_struct_rwlock must be held.
1901  */
1902 int
1903 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1904     void *tag, dmu_buf_impl_t **dbp)
1905 {
1906 	dmu_buf_impl_t *db, *parent = NULL;
1907 
1908 	ASSERT(blkid != DMU_BONUS_BLKID);
1909 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1910 	ASSERT3U(dn->dn_nlevels, >, level);
1911 
1912 	*dbp = NULL;
1913 top:
1914 	/* dbuf_find() returns with db_mtx held */
1915 	db = dbuf_find(dn, level, blkid);
1916 
1917 	if (db == NULL) {
1918 		blkptr_t *bp = NULL;
1919 		int err;
1920 
1921 		ASSERT3P(parent, ==, NULL);
1922 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1923 		if (fail_sparse) {
1924 			if (err == 0 && bp && BP_IS_HOLE(bp))
1925 				err = SET_ERROR(ENOENT);
1926 			if (err) {
1927 				if (parent)
1928 					dbuf_rele(parent, NULL);
1929 				return (err);
1930 			}
1931 		}
1932 		if (err && err != ENOENT)
1933 			return (err);
1934 		db = dbuf_create(dn, level, blkid, parent, bp);
1935 	}
1936 
1937 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1938 		arc_buf_add_ref(db->db_buf, db);
1939 		if (db->db_buf->b_data == NULL) {
1940 			dbuf_clear(db);
1941 			if (parent) {
1942 				dbuf_rele(parent, NULL);
1943 				parent = NULL;
1944 			}
1945 			goto top;
1946 		}
1947 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1948 	}
1949 
1950 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1951 
1952 	/*
1953 	 * If this buffer is currently syncing out, and we are are
1954 	 * still referencing it from db_data, we need to make a copy
1955 	 * of it in case we decide we want to dirty it again in this txg.
1956 	 */
1957 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1958 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
1959 	    db->db_state == DB_CACHED && db->db_data_pending) {
1960 		dbuf_dirty_record_t *dr = db->db_data_pending;
1961 
1962 		if (dr->dt.dl.dr_data == db->db_buf) {
1963 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1964 
1965 			dbuf_set_data(db,
1966 			    arc_buf_alloc(dn->dn_objset->os_spa,
1967 			    db->db.db_size, db, type));
1968 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1969 			    db->db.db_size);
1970 		}
1971 	}
1972 
1973 	(void) refcount_add(&db->db_holds, tag);
1974 	dbuf_update_data(db);
1975 	DBUF_VERIFY(db);
1976 	mutex_exit(&db->db_mtx);
1977 
1978 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
1979 	if (parent)
1980 		dbuf_rele(parent, NULL);
1981 
1982 	ASSERT3P(DB_DNODE(db), ==, dn);
1983 	ASSERT3U(db->db_blkid, ==, blkid);
1984 	ASSERT3U(db->db_level, ==, level);
1985 	*dbp = db;
1986 
1987 	return (0);
1988 }
1989 
1990 dmu_buf_impl_t *
1991 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1992 {
1993 	dmu_buf_impl_t *db;
1994 	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1995 	return (err ? NULL : db);
1996 }
1997 
1998 dmu_buf_impl_t *
1999 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2000 {
2001 	dmu_buf_impl_t *db;
2002 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2003 	return (err ? NULL : db);
2004 }
2005 
2006 void
2007 dbuf_create_bonus(dnode_t *dn)
2008 {
2009 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2010 
2011 	ASSERT(dn->dn_bonus == NULL);
2012 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2013 }
2014 
2015 int
2016 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2017 {
2018 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2019 	dnode_t *dn;
2020 
2021 	if (db->db_blkid != DMU_SPILL_BLKID)
2022 		return (SET_ERROR(ENOTSUP));
2023 	if (blksz == 0)
2024 		blksz = SPA_MINBLOCKSIZE;
2025 	if (blksz > SPA_MAXBLOCKSIZE)
2026 		blksz = SPA_MAXBLOCKSIZE;
2027 	else
2028 		blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2029 
2030 	DB_DNODE_ENTER(db);
2031 	dn = DB_DNODE(db);
2032 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2033 	dbuf_new_size(db, blksz, tx);
2034 	rw_exit(&dn->dn_struct_rwlock);
2035 	DB_DNODE_EXIT(db);
2036 
2037 	return (0);
2038 }
2039 
2040 void
2041 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2042 {
2043 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2044 }
2045 
2046 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2047 void
2048 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2049 {
2050 	int64_t holds = refcount_add(&db->db_holds, tag);
2051 	ASSERT(holds > 1);
2052 }
2053 
2054 /*
2055  * If you call dbuf_rele() you had better not be referencing the dnode handle
2056  * unless you have some other direct or indirect hold on the dnode. (An indirect
2057  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2058  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2059  * dnode's parent dbuf evicting its dnode handles.
2060  */
2061 void
2062 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2063 {
2064 	mutex_enter(&db->db_mtx);
2065 	dbuf_rele_and_unlock(db, tag);
2066 }
2067 
2068 void
2069 dmu_buf_rele(dmu_buf_t *db, void *tag)
2070 {
2071 	dbuf_rele((dmu_buf_impl_t *)db, tag);
2072 }
2073 
2074 /*
2075  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2076  * db_dirtycnt and db_holds to be updated atomically.
2077  */
2078 void
2079 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2080 {
2081 	int64_t holds;
2082 
2083 	ASSERT(MUTEX_HELD(&db->db_mtx));
2084 	DBUF_VERIFY(db);
2085 
2086 	/*
2087 	 * Remove the reference to the dbuf before removing its hold on the
2088 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2089 	 * buffer has a corresponding dnode hold.
2090 	 */
2091 	holds = refcount_remove(&db->db_holds, tag);
2092 	ASSERT(holds >= 0);
2093 
2094 	/*
2095 	 * We can't freeze indirects if there is a possibility that they
2096 	 * may be modified in the current syncing context.
2097 	 */
2098 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2099 		arc_buf_freeze(db->db_buf);
2100 
2101 	if (holds == db->db_dirtycnt &&
2102 	    db->db_level == 0 && db->db_immediate_evict)
2103 		dbuf_evict_user(db);
2104 
2105 	if (holds == 0) {
2106 		if (db->db_blkid == DMU_BONUS_BLKID) {
2107 			mutex_exit(&db->db_mtx);
2108 
2109 			/*
2110 			 * If the dnode moves here, we cannot cross this barrier
2111 			 * until the move completes.
2112 			 */
2113 			DB_DNODE_ENTER(db);
2114 			(void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2115 			DB_DNODE_EXIT(db);
2116 			/*
2117 			 * The bonus buffer's dnode hold is no longer discounted
2118 			 * in dnode_move(). The dnode cannot move until after
2119 			 * the dnode_rele().
2120 			 */
2121 			dnode_rele(DB_DNODE(db), db);
2122 		} else if (db->db_buf == NULL) {
2123 			/*
2124 			 * This is a special case: we never associated this
2125 			 * dbuf with any data allocated from the ARC.
2126 			 */
2127 			ASSERT(db->db_state == DB_UNCACHED ||
2128 			    db->db_state == DB_NOFILL);
2129 			dbuf_evict(db);
2130 		} else if (arc_released(db->db_buf)) {
2131 			arc_buf_t *buf = db->db_buf;
2132 			/*
2133 			 * This dbuf has anonymous data associated with it.
2134 			 */
2135 			dbuf_set_data(db, NULL);
2136 			VERIFY(arc_buf_remove_ref(buf, db));
2137 			dbuf_evict(db);
2138 		} else {
2139 			VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2140 
2141 			/*
2142 			 * A dbuf will be eligible for eviction if either the
2143 			 * 'primarycache' property is set or a duplicate
2144 			 * copy of this buffer is already cached in the arc.
2145 			 *
2146 			 * In the case of the 'primarycache' a buffer
2147 			 * is considered for eviction if it matches the
2148 			 * criteria set in the property.
2149 			 *
2150 			 * To decide if our buffer is considered a
2151 			 * duplicate, we must call into the arc to determine
2152 			 * if multiple buffers are referencing the same
2153 			 * block on-disk. If so, then we simply evict
2154 			 * ourselves.
2155 			 */
2156 			if (!DBUF_IS_CACHEABLE(db)) {
2157 				if (db->db_blkptr != NULL &&
2158 				    !BP_IS_HOLE(db->db_blkptr) &&
2159 				    !BP_IS_EMBEDDED(db->db_blkptr)) {
2160 					spa_t *spa =
2161 					    dmu_objset_spa(db->db_objset);
2162 					blkptr_t bp = *db->db_blkptr;
2163 					dbuf_clear(db);
2164 					arc_freed(spa, &bp);
2165 				} else {
2166 					dbuf_clear(db);
2167 				}
2168 			} else if (arc_buf_eviction_needed(db->db_buf)) {
2169 				dbuf_clear(db);
2170 			} else {
2171 				mutex_exit(&db->db_mtx);
2172 			}
2173 		}
2174 	} else {
2175 		mutex_exit(&db->db_mtx);
2176 	}
2177 }
2178 
2179 #pragma weak dmu_buf_refcount = dbuf_refcount
2180 uint64_t
2181 dbuf_refcount(dmu_buf_impl_t *db)
2182 {
2183 	return (refcount_count(&db->db_holds));
2184 }
2185 
2186 void *
2187 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2188     dmu_buf_evict_func_t *evict_func)
2189 {
2190 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2191 	    user_data_ptr_ptr, evict_func));
2192 }
2193 
2194 void *
2195 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2196     dmu_buf_evict_func_t *evict_func)
2197 {
2198 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2199 
2200 	db->db_immediate_evict = TRUE;
2201 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2202 	    user_data_ptr_ptr, evict_func));
2203 }
2204 
2205 void *
2206 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2207     void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2208 {
2209 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2210 	ASSERT(db->db_level == 0);
2211 
2212 	ASSERT((user_ptr == NULL) == (evict_func == NULL));
2213 
2214 	mutex_enter(&db->db_mtx);
2215 
2216 	if (db->db_user_ptr == old_user_ptr) {
2217 		db->db_user_ptr = user_ptr;
2218 		db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2219 		db->db_evict_func = evict_func;
2220 
2221 		dbuf_update_data(db);
2222 	} else {
2223 		old_user_ptr = db->db_user_ptr;
2224 	}
2225 
2226 	mutex_exit(&db->db_mtx);
2227 	return (old_user_ptr);
2228 }
2229 
2230 void *
2231 dmu_buf_get_user(dmu_buf_t *db_fake)
2232 {
2233 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2234 	ASSERT(!refcount_is_zero(&db->db_holds));
2235 
2236 	return (db->db_user_ptr);
2237 }
2238 
2239 boolean_t
2240 dmu_buf_freeable(dmu_buf_t *dbuf)
2241 {
2242 	boolean_t res = B_FALSE;
2243 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2244 
2245 	if (db->db_blkptr)
2246 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2247 		    db->db_blkptr, db->db_blkptr->blk_birth);
2248 
2249 	return (res);
2250 }
2251 
2252 blkptr_t *
2253 dmu_buf_get_blkptr(dmu_buf_t *db)
2254 {
2255 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2256 	return (dbi->db_blkptr);
2257 }
2258 
2259 static void
2260 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2261 {
2262 	/* ASSERT(dmu_tx_is_syncing(tx) */
2263 	ASSERT(MUTEX_HELD(&db->db_mtx));
2264 
2265 	if (db->db_blkptr != NULL)
2266 		return;
2267 
2268 	if (db->db_blkid == DMU_SPILL_BLKID) {
2269 		db->db_blkptr = &dn->dn_phys->dn_spill;
2270 		BP_ZERO(db->db_blkptr);
2271 		return;
2272 	}
2273 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2274 		/*
2275 		 * This buffer was allocated at a time when there was
2276 		 * no available blkptrs from the dnode, or it was
2277 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2278 		 */
2279 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2280 		ASSERT(db->db_parent == NULL);
2281 		db->db_parent = dn->dn_dbuf;
2282 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2283 		DBUF_VERIFY(db);
2284 	} else {
2285 		dmu_buf_impl_t *parent = db->db_parent;
2286 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2287 
2288 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2289 		if (parent == NULL) {
2290 			mutex_exit(&db->db_mtx);
2291 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2292 			(void) dbuf_hold_impl(dn, db->db_level+1,
2293 			    db->db_blkid >> epbs, FALSE, db, &parent);
2294 			rw_exit(&dn->dn_struct_rwlock);
2295 			mutex_enter(&db->db_mtx);
2296 			db->db_parent = parent;
2297 		}
2298 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2299 		    (db->db_blkid & ((1ULL << epbs) - 1));
2300 		DBUF_VERIFY(db);
2301 	}
2302 }
2303 
2304 static void
2305 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2306 {
2307 	dmu_buf_impl_t *db = dr->dr_dbuf;
2308 	dnode_t *dn;
2309 	zio_t *zio;
2310 
2311 	ASSERT(dmu_tx_is_syncing(tx));
2312 
2313 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2314 
2315 	mutex_enter(&db->db_mtx);
2316 
2317 	ASSERT(db->db_level > 0);
2318 	DBUF_VERIFY(db);
2319 
2320 	/* Read the block if it hasn't been read yet. */
2321 	if (db->db_buf == NULL) {
2322 		mutex_exit(&db->db_mtx);
2323 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2324 		mutex_enter(&db->db_mtx);
2325 	}
2326 	ASSERT3U(db->db_state, ==, DB_CACHED);
2327 	ASSERT(db->db_buf != NULL);
2328 
2329 	DB_DNODE_ENTER(db);
2330 	dn = DB_DNODE(db);
2331 	/* Indirect block size must match what the dnode thinks it is. */
2332 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2333 	dbuf_check_blkptr(dn, db);
2334 	DB_DNODE_EXIT(db);
2335 
2336 	/* Provide the pending dirty record to child dbufs */
2337 	db->db_data_pending = dr;
2338 
2339 	mutex_exit(&db->db_mtx);
2340 	dbuf_write(dr, db->db_buf, tx);
2341 
2342 	zio = dr->dr_zio;
2343 	mutex_enter(&dr->dt.di.dr_mtx);
2344 	dbuf_sync_list(&dr->dt.di.dr_children, tx);
2345 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2346 	mutex_exit(&dr->dt.di.dr_mtx);
2347 	zio_nowait(zio);
2348 }
2349 
2350 static void
2351 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2352 {
2353 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2354 	dmu_buf_impl_t *db = dr->dr_dbuf;
2355 	dnode_t *dn;
2356 	objset_t *os;
2357 	uint64_t txg = tx->tx_txg;
2358 
2359 	ASSERT(dmu_tx_is_syncing(tx));
2360 
2361 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2362 
2363 	mutex_enter(&db->db_mtx);
2364 	/*
2365 	 * To be synced, we must be dirtied.  But we
2366 	 * might have been freed after the dirty.
2367 	 */
2368 	if (db->db_state == DB_UNCACHED) {
2369 		/* This buffer has been freed since it was dirtied */
2370 		ASSERT(db->db.db_data == NULL);
2371 	} else if (db->db_state == DB_FILL) {
2372 		/* This buffer was freed and is now being re-filled */
2373 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2374 	} else {
2375 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2376 	}
2377 	DBUF_VERIFY(db);
2378 
2379 	DB_DNODE_ENTER(db);
2380 	dn = DB_DNODE(db);
2381 
2382 	if (db->db_blkid == DMU_SPILL_BLKID) {
2383 		mutex_enter(&dn->dn_mtx);
2384 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2385 		mutex_exit(&dn->dn_mtx);
2386 	}
2387 
2388 	/*
2389 	 * If this is a bonus buffer, simply copy the bonus data into the
2390 	 * dnode.  It will be written out when the dnode is synced (and it
2391 	 * will be synced, since it must have been dirty for dbuf_sync to
2392 	 * be called).
2393 	 */
2394 	if (db->db_blkid == DMU_BONUS_BLKID) {
2395 		dbuf_dirty_record_t **drp;
2396 
2397 		ASSERT(*datap != NULL);
2398 		ASSERT0(db->db_level);
2399 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2400 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2401 		DB_DNODE_EXIT(db);
2402 
2403 		if (*datap != db->db.db_data) {
2404 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
2405 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2406 		}
2407 		db->db_data_pending = NULL;
2408 		drp = &db->db_last_dirty;
2409 		while (*drp != dr)
2410 			drp = &(*drp)->dr_next;
2411 		ASSERT(dr->dr_next == NULL);
2412 		ASSERT(dr->dr_dbuf == db);
2413 		*drp = dr->dr_next;
2414 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
2415 		ASSERT(db->db_dirtycnt > 0);
2416 		db->db_dirtycnt -= 1;
2417 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2418 		return;
2419 	}
2420 
2421 	os = dn->dn_objset;
2422 
2423 	/*
2424 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
2425 	 * operation to sneak in. As a result, we need to ensure that we
2426 	 * don't check the dr_override_state until we have returned from
2427 	 * dbuf_check_blkptr.
2428 	 */
2429 	dbuf_check_blkptr(dn, db);
2430 
2431 	/*
2432 	 * If this buffer is in the middle of an immediate write,
2433 	 * wait for the synchronous IO to complete.
2434 	 */
2435 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2436 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2437 		cv_wait(&db->db_changed, &db->db_mtx);
2438 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2439 	}
2440 
2441 	if (db->db_state != DB_NOFILL &&
2442 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2443 	    refcount_count(&db->db_holds) > 1 &&
2444 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2445 	    *datap == db->db_buf) {
2446 		/*
2447 		 * If this buffer is currently "in use" (i.e., there
2448 		 * are active holds and db_data still references it),
2449 		 * then make a copy before we start the write so that
2450 		 * any modifications from the open txg will not leak
2451 		 * into this write.
2452 		 *
2453 		 * NOTE: this copy does not need to be made for
2454 		 * objects only modified in the syncing context (e.g.
2455 		 * DNONE_DNODE blocks).
2456 		 */
2457 		int blksz = arc_buf_size(*datap);
2458 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2459 		*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2460 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
2461 	}
2462 	db->db_data_pending = dr;
2463 
2464 	mutex_exit(&db->db_mtx);
2465 
2466 	dbuf_write(dr, *datap, tx);
2467 
2468 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2469 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2470 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2471 		DB_DNODE_EXIT(db);
2472 	} else {
2473 		/*
2474 		 * Although zio_nowait() does not "wait for an IO", it does
2475 		 * initiate the IO. If this is an empty write it seems plausible
2476 		 * that the IO could actually be completed before the nowait
2477 		 * returns. We need to DB_DNODE_EXIT() first in case
2478 		 * zio_nowait() invalidates the dbuf.
2479 		 */
2480 		DB_DNODE_EXIT(db);
2481 		zio_nowait(dr->dr_zio);
2482 	}
2483 }
2484 
2485 void
2486 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2487 {
2488 	dbuf_dirty_record_t *dr;
2489 
2490 	while (dr = list_head(list)) {
2491 		if (dr->dr_zio != NULL) {
2492 			/*
2493 			 * If we find an already initialized zio then we
2494 			 * are processing the meta-dnode, and we have finished.
2495 			 * The dbufs for all dnodes are put back on the list
2496 			 * during processing, so that we can zio_wait()
2497 			 * these IOs after initiating all child IOs.
2498 			 */
2499 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2500 			    DMU_META_DNODE_OBJECT);
2501 			break;
2502 		}
2503 		list_remove(list, dr);
2504 		if (dr->dr_dbuf->db_level > 0)
2505 			dbuf_sync_indirect(dr, tx);
2506 		else
2507 			dbuf_sync_leaf(dr, tx);
2508 	}
2509 }
2510 
2511 /* ARGSUSED */
2512 static void
2513 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2514 {
2515 	dmu_buf_impl_t *db = vdb;
2516 	dnode_t *dn;
2517 	blkptr_t *bp = zio->io_bp;
2518 	blkptr_t *bp_orig = &zio->io_bp_orig;
2519 	spa_t *spa = zio->io_spa;
2520 	int64_t delta;
2521 	uint64_t fill = 0;
2522 	int i;
2523 
2524 	ASSERT3P(db->db_blkptr, ==, bp);
2525 
2526 	DB_DNODE_ENTER(db);
2527 	dn = DB_DNODE(db);
2528 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2529 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2530 	zio->io_prev_space_delta = delta;
2531 
2532 	if (bp->blk_birth != 0) {
2533 		ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2534 		    BP_GET_TYPE(bp) == dn->dn_type) ||
2535 		    (db->db_blkid == DMU_SPILL_BLKID &&
2536 		    BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2537 		    BP_IS_EMBEDDED(bp));
2538 		ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2539 	}
2540 
2541 	mutex_enter(&db->db_mtx);
2542 
2543 #ifdef ZFS_DEBUG
2544 	if (db->db_blkid == DMU_SPILL_BLKID) {
2545 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2546 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2547 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2548 	}
2549 #endif
2550 
2551 	if (db->db_level == 0) {
2552 		mutex_enter(&dn->dn_mtx);
2553 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2554 		    db->db_blkid != DMU_SPILL_BLKID)
2555 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2556 		mutex_exit(&dn->dn_mtx);
2557 
2558 		if (dn->dn_type == DMU_OT_DNODE) {
2559 			dnode_phys_t *dnp = db->db.db_data;
2560 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2561 			    i--, dnp++) {
2562 				if (dnp->dn_type != DMU_OT_NONE)
2563 					fill++;
2564 			}
2565 		} else {
2566 			if (BP_IS_HOLE(bp)) {
2567 				fill = 0;
2568 			} else {
2569 				fill = 1;
2570 			}
2571 		}
2572 	} else {
2573 		blkptr_t *ibp = db->db.db_data;
2574 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2575 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2576 			if (BP_IS_HOLE(ibp))
2577 				continue;
2578 			fill += BP_GET_FILL(ibp);
2579 		}
2580 	}
2581 	DB_DNODE_EXIT(db);
2582 
2583 	if (!BP_IS_EMBEDDED(bp))
2584 		bp->blk_fill = fill;
2585 
2586 	mutex_exit(&db->db_mtx);
2587 }
2588 
2589 /*
2590  * The SPA will call this callback several times for each zio - once
2591  * for every physical child i/o (zio->io_phys_children times).  This
2592  * allows the DMU to monitor the progress of each logical i/o.  For example,
2593  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2594  * block.  There may be a long delay before all copies/fragments are completed,
2595  * so this callback allows us to retire dirty space gradually, as the physical
2596  * i/os complete.
2597  */
2598 /* ARGSUSED */
2599 static void
2600 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2601 {
2602 	dmu_buf_impl_t *db = arg;
2603 	objset_t *os = db->db_objset;
2604 	dsl_pool_t *dp = dmu_objset_pool(os);
2605 	dbuf_dirty_record_t *dr;
2606 	int delta = 0;
2607 
2608 	dr = db->db_data_pending;
2609 	ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2610 
2611 	/*
2612 	 * The callback will be called io_phys_children times.  Retire one
2613 	 * portion of our dirty space each time we are called.  Any rounding
2614 	 * error will be cleaned up by dsl_pool_sync()'s call to
2615 	 * dsl_pool_undirty_space().
2616 	 */
2617 	delta = dr->dr_accounted / zio->io_phys_children;
2618 	dsl_pool_undirty_space(dp, delta, zio->io_txg);
2619 }
2620 
2621 /* ARGSUSED */
2622 static void
2623 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2624 {
2625 	dmu_buf_impl_t *db = vdb;
2626 	blkptr_t *bp_orig = &zio->io_bp_orig;
2627 	blkptr_t *bp = db->db_blkptr;
2628 	objset_t *os = db->db_objset;
2629 	dmu_tx_t *tx = os->os_synctx;
2630 	dbuf_dirty_record_t **drp, *dr;
2631 
2632 	ASSERT0(zio->io_error);
2633 	ASSERT(db->db_blkptr == bp);
2634 
2635 	/*
2636 	 * For nopwrites and rewrites we ensure that the bp matches our
2637 	 * original and bypass all the accounting.
2638 	 */
2639 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2640 		ASSERT(BP_EQUAL(bp, bp_orig));
2641 	} else {
2642 		dsl_dataset_t *ds = os->os_dsl_dataset;
2643 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2644 		dsl_dataset_block_born(ds, bp, tx);
2645 	}
2646 
2647 	mutex_enter(&db->db_mtx);
2648 
2649 	DBUF_VERIFY(db);
2650 
2651 	drp = &db->db_last_dirty;
2652 	while ((dr = *drp) != db->db_data_pending)
2653 		drp = &dr->dr_next;
2654 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2655 	ASSERT(dr->dr_dbuf == db);
2656 	ASSERT(dr->dr_next == NULL);
2657 	*drp = dr->dr_next;
2658 
2659 #ifdef ZFS_DEBUG
2660 	if (db->db_blkid == DMU_SPILL_BLKID) {
2661 		dnode_t *dn;
2662 
2663 		DB_DNODE_ENTER(db);
2664 		dn = DB_DNODE(db);
2665 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2666 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2667 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2668 		DB_DNODE_EXIT(db);
2669 	}
2670 #endif
2671 
2672 	if (db->db_level == 0) {
2673 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2674 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2675 		if (db->db_state != DB_NOFILL) {
2676 			if (dr->dt.dl.dr_data != db->db_buf)
2677 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2678 				    db));
2679 			else if (!arc_released(db->db_buf))
2680 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2681 		}
2682 	} else {
2683 		dnode_t *dn;
2684 
2685 		DB_DNODE_ENTER(db);
2686 		dn = DB_DNODE(db);
2687 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2688 		ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2689 		if (!BP_IS_HOLE(db->db_blkptr)) {
2690 			int epbs =
2691 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2692 			ASSERT3U(db->db_blkid, <=,
2693 			    dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2694 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2695 			    db->db.db_size);
2696 			if (!arc_released(db->db_buf))
2697 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2698 		}
2699 		DB_DNODE_EXIT(db);
2700 		mutex_destroy(&dr->dt.di.dr_mtx);
2701 		list_destroy(&dr->dt.di.dr_children);
2702 	}
2703 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
2704 
2705 	cv_broadcast(&db->db_changed);
2706 	ASSERT(db->db_dirtycnt > 0);
2707 	db->db_dirtycnt -= 1;
2708 	db->db_data_pending = NULL;
2709 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2710 }
2711 
2712 static void
2713 dbuf_write_nofill_ready(zio_t *zio)
2714 {
2715 	dbuf_write_ready(zio, NULL, zio->io_private);
2716 }
2717 
2718 static void
2719 dbuf_write_nofill_done(zio_t *zio)
2720 {
2721 	dbuf_write_done(zio, NULL, zio->io_private);
2722 }
2723 
2724 static void
2725 dbuf_write_override_ready(zio_t *zio)
2726 {
2727 	dbuf_dirty_record_t *dr = zio->io_private;
2728 	dmu_buf_impl_t *db = dr->dr_dbuf;
2729 
2730 	dbuf_write_ready(zio, NULL, db);
2731 }
2732 
2733 static void
2734 dbuf_write_override_done(zio_t *zio)
2735 {
2736 	dbuf_dirty_record_t *dr = zio->io_private;
2737 	dmu_buf_impl_t *db = dr->dr_dbuf;
2738 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2739 
2740 	mutex_enter(&db->db_mtx);
2741 	if (!BP_EQUAL(zio->io_bp, obp)) {
2742 		if (!BP_IS_HOLE(obp))
2743 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2744 		arc_release(dr->dt.dl.dr_data, db);
2745 	}
2746 	mutex_exit(&db->db_mtx);
2747 
2748 	dbuf_write_done(zio, NULL, db);
2749 }
2750 
2751 /* Issue I/O to commit a dirty buffer to disk. */
2752 static void
2753 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2754 {
2755 	dmu_buf_impl_t *db = dr->dr_dbuf;
2756 	dnode_t *dn;
2757 	objset_t *os;
2758 	dmu_buf_impl_t *parent = db->db_parent;
2759 	uint64_t txg = tx->tx_txg;
2760 	zbookmark_phys_t zb;
2761 	zio_prop_t zp;
2762 	zio_t *zio;
2763 	int wp_flag = 0;
2764 
2765 	DB_DNODE_ENTER(db);
2766 	dn = DB_DNODE(db);
2767 	os = dn->dn_objset;
2768 
2769 	if (db->db_state != DB_NOFILL) {
2770 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2771 			/*
2772 			 * Private object buffers are released here rather
2773 			 * than in dbuf_dirty() since they are only modified
2774 			 * in the syncing context and we don't want the
2775 			 * overhead of making multiple copies of the data.
2776 			 */
2777 			if (BP_IS_HOLE(db->db_blkptr)) {
2778 				arc_buf_thaw(data);
2779 			} else {
2780 				dbuf_release_bp(db);
2781 			}
2782 		}
2783 	}
2784 
2785 	if (parent != dn->dn_dbuf) {
2786 		/* Our parent is an indirect block. */
2787 		/* We have a dirty parent that has been scheduled for write. */
2788 		ASSERT(parent && parent->db_data_pending);
2789 		/* Our parent's buffer is one level closer to the dnode. */
2790 		ASSERT(db->db_level == parent->db_level-1);
2791 		/*
2792 		 * We're about to modify our parent's db_data by modifying
2793 		 * our block pointer, so the parent must be released.
2794 		 */
2795 		ASSERT(arc_released(parent->db_buf));
2796 		zio = parent->db_data_pending->dr_zio;
2797 	} else {
2798 		/* Our parent is the dnode itself. */
2799 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2800 		    db->db_blkid != DMU_SPILL_BLKID) ||
2801 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2802 		if (db->db_blkid != DMU_SPILL_BLKID)
2803 			ASSERT3P(db->db_blkptr, ==,
2804 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
2805 		zio = dn->dn_zio;
2806 	}
2807 
2808 	ASSERT(db->db_level == 0 || data == db->db_buf);
2809 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2810 	ASSERT(zio);
2811 
2812 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2813 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2814 	    db->db.db_object, db->db_level, db->db_blkid);
2815 
2816 	if (db->db_blkid == DMU_SPILL_BLKID)
2817 		wp_flag = WP_SPILL;
2818 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2819 
2820 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2821 	DB_DNODE_EXIT(db);
2822 
2823 	if (db->db_level == 0 &&
2824 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2825 		/*
2826 		 * The BP for this block has been provided by open context
2827 		 * (by dmu_sync() or dmu_buf_write_embedded()).
2828 		 */
2829 		void *contents = (data != NULL) ? data->b_data : NULL;
2830 
2831 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2832 		    db->db_blkptr, contents, db->db.db_size, &zp,
2833 		    dbuf_write_override_ready, NULL, dbuf_write_override_done,
2834 		    dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2835 		mutex_enter(&db->db_mtx);
2836 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2837 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2838 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2839 		mutex_exit(&db->db_mtx);
2840 	} else if (db->db_state == DB_NOFILL) {
2841 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2842 		    zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2843 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2844 		    db->db_blkptr, NULL, db->db.db_size, &zp,
2845 		    dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2846 		    ZIO_PRIORITY_ASYNC_WRITE,
2847 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2848 	} else {
2849 		ASSERT(arc_released(data));
2850 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
2851 		    db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2852 		    DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2853 		    dbuf_write_physdone, dbuf_write_done, db,
2854 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2855 	}
2856 }
2857