xref: /titanic_51/usr/src/uts/common/fs/zfs/dbuf.c (revision a95756817154e8f4dfc0fd0963ce91ea4f5e34fb)
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 	return (0);
74 }
75 
76 /* ARGSUSED */
77 static void
78 dbuf_dest(void *vdb, void *unused)
79 {
80 	dmu_buf_impl_t *db = vdb;
81 	mutex_destroy(&db->db_mtx);
82 	cv_destroy(&db->db_changed);
83 	refcount_destroy(&db->db_holds);
84 }
85 
86 /*
87  * dbuf hash table routines
88  */
89 static dbuf_hash_table_t dbuf_hash_table;
90 
91 static uint64_t dbuf_hash_count;
92 
93 static uint64_t
94 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
95 {
96 	uintptr_t osv = (uintptr_t)os;
97 	uint64_t crc = -1ULL;
98 
99 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
100 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
101 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
102 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
103 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
104 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
105 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
106 
107 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
108 
109 	return (crc);
110 }
111 
112 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
113 
114 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
115 	((dbuf)->db.db_object == (obj) &&		\
116 	(dbuf)->db_objset == (os) &&			\
117 	(dbuf)->db_level == (level) &&			\
118 	(dbuf)->db_blkid == (blkid))
119 
120 dmu_buf_impl_t *
121 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
122 {
123 	dbuf_hash_table_t *h = &dbuf_hash_table;
124 	objset_t *os = dn->dn_objset;
125 	uint64_t obj = dn->dn_object;
126 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
127 	uint64_t idx = hv & h->hash_table_mask;
128 	dmu_buf_impl_t *db;
129 
130 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
131 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
132 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
133 			mutex_enter(&db->db_mtx);
134 			if (db->db_state != DB_EVICTING) {
135 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
136 				return (db);
137 			}
138 			mutex_exit(&db->db_mtx);
139 		}
140 	}
141 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
142 	return (NULL);
143 }
144 
145 /*
146  * Insert an entry into the hash table.  If there is already an element
147  * equal to elem in the hash table, then the already existing element
148  * will be returned and the new element will not be inserted.
149  * Otherwise returns NULL.
150  */
151 static dmu_buf_impl_t *
152 dbuf_hash_insert(dmu_buf_impl_t *db)
153 {
154 	dbuf_hash_table_t *h = &dbuf_hash_table;
155 	objset_t *os = db->db_objset;
156 	uint64_t obj = db->db.db_object;
157 	int level = db->db_level;
158 	uint64_t blkid = db->db_blkid;
159 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
160 	uint64_t idx = hv & h->hash_table_mask;
161 	dmu_buf_impl_t *dbf;
162 
163 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
164 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
165 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
166 			mutex_enter(&dbf->db_mtx);
167 			if (dbf->db_state != DB_EVICTING) {
168 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
169 				return (dbf);
170 			}
171 			mutex_exit(&dbf->db_mtx);
172 		}
173 	}
174 
175 	mutex_enter(&db->db_mtx);
176 	db->db_hash_next = h->hash_table[idx];
177 	h->hash_table[idx] = db;
178 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
179 	atomic_inc_64(&dbuf_hash_count);
180 
181 	return (NULL);
182 }
183 
184 /*
185  * Remove an entry from the hash table.  It must be in the EVICTING state.
186  */
187 static void
188 dbuf_hash_remove(dmu_buf_impl_t *db)
189 {
190 	dbuf_hash_table_t *h = &dbuf_hash_table;
191 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
192 	    db->db_level, db->db_blkid);
193 	uint64_t idx = hv & h->hash_table_mask;
194 	dmu_buf_impl_t *dbf, **dbp;
195 
196 	/*
197 	 * We musn't hold db_mtx to maintain lock ordering:
198 	 * DBUF_HASH_MUTEX > db_mtx.
199 	 */
200 	ASSERT(refcount_is_zero(&db->db_holds));
201 	ASSERT(db->db_state == DB_EVICTING);
202 	ASSERT(!MUTEX_HELD(&db->db_mtx));
203 
204 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
205 	dbp = &h->hash_table[idx];
206 	while ((dbf = *dbp) != db) {
207 		dbp = &dbf->db_hash_next;
208 		ASSERT(dbf != NULL);
209 	}
210 	*dbp = db->db_hash_next;
211 	db->db_hash_next = NULL;
212 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
213 	atomic_dec_64(&dbuf_hash_count);
214 }
215 
216 static arc_evict_func_t dbuf_do_evict;
217 
218 static void
219 dbuf_evict_user(dmu_buf_impl_t *db)
220 {
221 	ASSERT(MUTEX_HELD(&db->db_mtx));
222 
223 	if (db->db_level != 0 || db->db_evict_func == NULL)
224 		return;
225 
226 	db->db_evict_func(&db->db, db->db_user_ptr);
227 	db->db_user_ptr = NULL;
228 	db->db_evict_func = NULL;
229 }
230 
231 boolean_t
232 dbuf_is_metadata(dmu_buf_impl_t *db)
233 {
234 	if (db->db_level > 0) {
235 		return (B_TRUE);
236 	} else {
237 		boolean_t is_metadata;
238 
239 		DB_DNODE_ENTER(db);
240 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
241 		DB_DNODE_EXIT(db);
242 
243 		return (is_metadata);
244 	}
245 }
246 
247 void
248 dbuf_evict(dmu_buf_impl_t *db)
249 {
250 	ASSERT(MUTEX_HELD(&db->db_mtx));
251 	ASSERT(db->db_buf == NULL);
252 	ASSERT(db->db_data_pending == NULL);
253 
254 	dbuf_clear(db);
255 	dbuf_destroy(db);
256 }
257 
258 void
259 dbuf_init(void)
260 {
261 	uint64_t hsize = 1ULL << 16;
262 	dbuf_hash_table_t *h = &dbuf_hash_table;
263 	int i;
264 
265 	/*
266 	 * The hash table is big enough to fill all of physical memory
267 	 * with an average 4K block size.  The table will take up
268 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
269 	 */
270 	while (hsize * 4096 < physmem * PAGESIZE)
271 		hsize <<= 1;
272 
273 retry:
274 	h->hash_table_mask = hsize - 1;
275 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
276 	if (h->hash_table == NULL) {
277 		/* XXX - we should really return an error instead of assert */
278 		ASSERT(hsize > (1ULL << 10));
279 		hsize >>= 1;
280 		goto retry;
281 	}
282 
283 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
284 	    sizeof (dmu_buf_impl_t),
285 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
286 
287 	for (i = 0; i < DBUF_MUTEXES; i++)
288 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
289 }
290 
291 void
292 dbuf_fini(void)
293 {
294 	dbuf_hash_table_t *h = &dbuf_hash_table;
295 	int i;
296 
297 	for (i = 0; i < DBUF_MUTEXES; i++)
298 		mutex_destroy(&h->hash_mutexes[i]);
299 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
300 	kmem_cache_destroy(dbuf_cache);
301 }
302 
303 /*
304  * Other stuff.
305  */
306 
307 #ifdef ZFS_DEBUG
308 static void
309 dbuf_verify(dmu_buf_impl_t *db)
310 {
311 	dnode_t *dn;
312 	dbuf_dirty_record_t *dr;
313 
314 	ASSERT(MUTEX_HELD(&db->db_mtx));
315 
316 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
317 		return;
318 
319 	ASSERT(db->db_objset != NULL);
320 	DB_DNODE_ENTER(db);
321 	dn = DB_DNODE(db);
322 	if (dn == NULL) {
323 		ASSERT(db->db_parent == NULL);
324 		ASSERT(db->db_blkptr == NULL);
325 	} else {
326 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
327 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
328 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
329 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
330 		    db->db_blkid == DMU_SPILL_BLKID ||
331 		    !avl_is_empty(&dn->dn_dbufs));
332 	}
333 	if (db->db_blkid == DMU_BONUS_BLKID) {
334 		ASSERT(dn != NULL);
335 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
336 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
337 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
338 		ASSERT(dn != NULL);
339 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
340 		ASSERT0(db->db.db_offset);
341 	} else {
342 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
343 	}
344 
345 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
346 		ASSERT(dr->dr_dbuf == db);
347 
348 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
349 		ASSERT(dr->dr_dbuf == db);
350 
351 	/*
352 	 * We can't assert that db_size matches dn_datablksz because it
353 	 * can be momentarily different when another thread is doing
354 	 * dnode_set_blksz().
355 	 */
356 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
357 		dr = db->db_data_pending;
358 		/*
359 		 * It should only be modified in syncing context, so
360 		 * make sure we only have one copy of the data.
361 		 */
362 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
363 	}
364 
365 	/* verify db->db_blkptr */
366 	if (db->db_blkptr) {
367 		if (db->db_parent == dn->dn_dbuf) {
368 			/* db is pointed to by the dnode */
369 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
370 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
371 				ASSERT(db->db_parent == NULL);
372 			else
373 				ASSERT(db->db_parent != NULL);
374 			if (db->db_blkid != DMU_SPILL_BLKID)
375 				ASSERT3P(db->db_blkptr, ==,
376 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
377 		} else {
378 			/* db is pointed to by an indirect block */
379 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
380 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
381 			ASSERT3U(db->db_parent->db.db_object, ==,
382 			    db->db.db_object);
383 			/*
384 			 * dnode_grow_indblksz() can make this fail if we don't
385 			 * have the struct_rwlock.  XXX indblksz no longer
386 			 * grows.  safe to do this now?
387 			 */
388 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
389 				ASSERT3P(db->db_blkptr, ==,
390 				    ((blkptr_t *)db->db_parent->db.db_data +
391 				    db->db_blkid % epb));
392 			}
393 		}
394 	}
395 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
396 	    (db->db_buf == NULL || db->db_buf->b_data) &&
397 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
398 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
399 		/*
400 		 * If the blkptr isn't set but they have nonzero data,
401 		 * it had better be dirty, otherwise we'll lose that
402 		 * data when we evict this buffer.
403 		 */
404 		if (db->db_dirtycnt == 0) {
405 			uint64_t *buf = db->db.db_data;
406 			int i;
407 
408 			for (i = 0; i < db->db.db_size >> 3; i++) {
409 				ASSERT(buf[i] == 0);
410 			}
411 		}
412 	}
413 	DB_DNODE_EXIT(db);
414 }
415 #endif
416 
417 static void
418 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
419 {
420 	ASSERT(MUTEX_HELD(&db->db_mtx));
421 	db->db_buf = buf;
422 	if (buf != NULL) {
423 		ASSERT(buf->b_data != NULL);
424 		db->db.db_data = buf->b_data;
425 		if (!arc_released(buf))
426 			arc_set_callback(buf, dbuf_do_evict, db);
427 	} else {
428 		dbuf_evict_user(db);
429 		db->db.db_data = NULL;
430 		if (db->db_state != DB_NOFILL)
431 			db->db_state = DB_UNCACHED;
432 	}
433 }
434 
435 /*
436  * Loan out an arc_buf for read.  Return the loaned arc_buf.
437  */
438 arc_buf_t *
439 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
440 {
441 	arc_buf_t *abuf;
442 
443 	mutex_enter(&db->db_mtx);
444 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
445 		int blksz = db->db.db_size;
446 		spa_t *spa = db->db_objset->os_spa;
447 
448 		mutex_exit(&db->db_mtx);
449 		abuf = arc_loan_buf(spa, blksz);
450 		bcopy(db->db.db_data, abuf->b_data, blksz);
451 	} else {
452 		abuf = db->db_buf;
453 		arc_loan_inuse_buf(abuf, db);
454 		dbuf_set_data(db, NULL);
455 		mutex_exit(&db->db_mtx);
456 	}
457 	return (abuf);
458 }
459 
460 uint64_t
461 dbuf_whichblock(dnode_t *dn, uint64_t offset)
462 {
463 	if (dn->dn_datablkshift) {
464 		return (offset >> dn->dn_datablkshift);
465 	} else {
466 		ASSERT3U(offset, <, dn->dn_datablksz);
467 		return (0);
468 	}
469 }
470 
471 static void
472 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
473 {
474 	dmu_buf_impl_t *db = vdb;
475 
476 	mutex_enter(&db->db_mtx);
477 	ASSERT3U(db->db_state, ==, DB_READ);
478 	/*
479 	 * All reads are synchronous, so we must have a hold on the dbuf
480 	 */
481 	ASSERT(refcount_count(&db->db_holds) > 0);
482 	ASSERT(db->db_buf == NULL);
483 	ASSERT(db->db.db_data == NULL);
484 	if (db->db_level == 0 && db->db_freed_in_flight) {
485 		/* we were freed in flight; disregard any error */
486 		arc_release(buf, db);
487 		bzero(buf->b_data, db->db.db_size);
488 		arc_buf_freeze(buf);
489 		db->db_freed_in_flight = FALSE;
490 		dbuf_set_data(db, buf);
491 		db->db_state = DB_CACHED;
492 	} else if (zio == NULL || zio->io_error == 0) {
493 		dbuf_set_data(db, buf);
494 		db->db_state = DB_CACHED;
495 	} else {
496 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
497 		ASSERT3P(db->db_buf, ==, NULL);
498 		VERIFY(arc_buf_remove_ref(buf, db));
499 		db->db_state = DB_UNCACHED;
500 	}
501 	cv_broadcast(&db->db_changed);
502 	dbuf_rele_and_unlock(db, NULL);
503 }
504 
505 static void
506 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
507 {
508 	dnode_t *dn;
509 	zbookmark_phys_t zb;
510 	uint32_t aflags = ARC_NOWAIT;
511 
512 	DB_DNODE_ENTER(db);
513 	dn = DB_DNODE(db);
514 	ASSERT(!refcount_is_zero(&db->db_holds));
515 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
516 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
517 	ASSERT(MUTEX_HELD(&db->db_mtx));
518 	ASSERT(db->db_state == DB_UNCACHED);
519 	ASSERT(db->db_buf == NULL);
520 
521 	if (db->db_blkid == DMU_BONUS_BLKID) {
522 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
523 
524 		ASSERT3U(bonuslen, <=, db->db.db_size);
525 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
526 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
527 		if (bonuslen < DN_MAX_BONUSLEN)
528 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
529 		if (bonuslen)
530 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
531 		DB_DNODE_EXIT(db);
532 		db->db_state = DB_CACHED;
533 		mutex_exit(&db->db_mtx);
534 		return;
535 	}
536 
537 	/*
538 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
539 	 * processes the delete record and clears the bp while we are waiting
540 	 * for the dn_mtx (resulting in a "no" from block_freed).
541 	 */
542 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
543 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
544 	    BP_IS_HOLE(db->db_blkptr)))) {
545 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
546 
547 		DB_DNODE_EXIT(db);
548 		dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
549 		    db->db.db_size, db, type));
550 		bzero(db->db.db_data, db->db.db_size);
551 		db->db_state = DB_CACHED;
552 		*flags |= DB_RF_CACHED;
553 		mutex_exit(&db->db_mtx);
554 		return;
555 	}
556 
557 	DB_DNODE_EXIT(db);
558 
559 	db->db_state = DB_READ;
560 	mutex_exit(&db->db_mtx);
561 
562 	if (DBUF_IS_L2CACHEABLE(db))
563 		aflags |= ARC_L2CACHE;
564 	if (DBUF_IS_L2COMPRESSIBLE(db))
565 		aflags |= ARC_L2COMPRESS;
566 
567 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
568 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
569 	    db->db.db_object, db->db_level, db->db_blkid);
570 
571 	dbuf_add_ref(db, NULL);
572 
573 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
574 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
575 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
576 	    &aflags, &zb);
577 	if (aflags & ARC_CACHED)
578 		*flags |= DB_RF_CACHED;
579 }
580 
581 int
582 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
583 {
584 	int err = 0;
585 	boolean_t havepzio = (zio != NULL);
586 	boolean_t prefetch;
587 	dnode_t *dn;
588 
589 	/*
590 	 * We don't have to hold the mutex to check db_state because it
591 	 * can't be freed while we have a hold on the buffer.
592 	 */
593 	ASSERT(!refcount_is_zero(&db->db_holds));
594 
595 	if (db->db_state == DB_NOFILL)
596 		return (SET_ERROR(EIO));
597 
598 	DB_DNODE_ENTER(db);
599 	dn = DB_DNODE(db);
600 	if ((flags & DB_RF_HAVESTRUCT) == 0)
601 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
602 
603 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
604 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
605 	    DBUF_IS_CACHEABLE(db);
606 
607 	mutex_enter(&db->db_mtx);
608 	if (db->db_state == DB_CACHED) {
609 		mutex_exit(&db->db_mtx);
610 		if (prefetch)
611 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
612 			    db->db.db_size, TRUE);
613 		if ((flags & DB_RF_HAVESTRUCT) == 0)
614 			rw_exit(&dn->dn_struct_rwlock);
615 		DB_DNODE_EXIT(db);
616 	} else if (db->db_state == DB_UNCACHED) {
617 		spa_t *spa = dn->dn_objset->os_spa;
618 
619 		if (zio == NULL)
620 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
621 		dbuf_read_impl(db, zio, &flags);
622 
623 		/* dbuf_read_impl has dropped db_mtx for us */
624 
625 		if (prefetch)
626 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
627 			    db->db.db_size, flags & DB_RF_CACHED);
628 
629 		if ((flags & DB_RF_HAVESTRUCT) == 0)
630 			rw_exit(&dn->dn_struct_rwlock);
631 		DB_DNODE_EXIT(db);
632 
633 		if (!havepzio)
634 			err = zio_wait(zio);
635 	} else {
636 		/*
637 		 * Another reader came in while the dbuf was in flight
638 		 * between UNCACHED and CACHED.  Either a writer will finish
639 		 * writing the buffer (sending the dbuf to CACHED) or the
640 		 * first reader's request will reach the read_done callback
641 		 * and send the dbuf to CACHED.  Otherwise, a failure
642 		 * occurred and the dbuf went to UNCACHED.
643 		 */
644 		mutex_exit(&db->db_mtx);
645 		if (prefetch)
646 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
647 			    db->db.db_size, TRUE);
648 		if ((flags & DB_RF_HAVESTRUCT) == 0)
649 			rw_exit(&dn->dn_struct_rwlock);
650 		DB_DNODE_EXIT(db);
651 
652 		/* Skip the wait per the caller's request. */
653 		mutex_enter(&db->db_mtx);
654 		if ((flags & DB_RF_NEVERWAIT) == 0) {
655 			while (db->db_state == DB_READ ||
656 			    db->db_state == DB_FILL) {
657 				ASSERT(db->db_state == DB_READ ||
658 				    (flags & DB_RF_HAVESTRUCT) == 0);
659 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
660 				    db, zio_t *, zio);
661 				cv_wait(&db->db_changed, &db->db_mtx);
662 			}
663 			if (db->db_state == DB_UNCACHED)
664 				err = SET_ERROR(EIO);
665 		}
666 		mutex_exit(&db->db_mtx);
667 	}
668 
669 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
670 	return (err);
671 }
672 
673 static void
674 dbuf_noread(dmu_buf_impl_t *db)
675 {
676 	ASSERT(!refcount_is_zero(&db->db_holds));
677 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
678 	mutex_enter(&db->db_mtx);
679 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
680 		cv_wait(&db->db_changed, &db->db_mtx);
681 	if (db->db_state == DB_UNCACHED) {
682 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
683 		spa_t *spa = db->db_objset->os_spa;
684 
685 		ASSERT(db->db_buf == NULL);
686 		ASSERT(db->db.db_data == NULL);
687 		dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
688 		db->db_state = DB_FILL;
689 	} else if (db->db_state == DB_NOFILL) {
690 		dbuf_set_data(db, NULL);
691 	} else {
692 		ASSERT3U(db->db_state, ==, DB_CACHED);
693 	}
694 	mutex_exit(&db->db_mtx);
695 }
696 
697 /*
698  * This is our just-in-time copy function.  It makes a copy of
699  * buffers, that have been modified in a previous transaction
700  * group, before we modify them in the current active group.
701  *
702  * This function is used in two places: when we are dirtying a
703  * buffer for the first time in a txg, and when we are freeing
704  * a range in a dnode that includes this buffer.
705  *
706  * Note that when we are called from dbuf_free_range() we do
707  * not put a hold on the buffer, we just traverse the active
708  * dbuf list for the dnode.
709  */
710 static void
711 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
712 {
713 	dbuf_dirty_record_t *dr = db->db_last_dirty;
714 
715 	ASSERT(MUTEX_HELD(&db->db_mtx));
716 	ASSERT(db->db.db_data != NULL);
717 	ASSERT(db->db_level == 0);
718 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
719 
720 	if (dr == NULL ||
721 	    (dr->dt.dl.dr_data !=
722 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
723 		return;
724 
725 	/*
726 	 * If the last dirty record for this dbuf has not yet synced
727 	 * and its referencing the dbuf data, either:
728 	 *	reset the reference to point to a new copy,
729 	 * or (if there a no active holders)
730 	 *	just null out the current db_data pointer.
731 	 */
732 	ASSERT(dr->dr_txg >= txg - 2);
733 	if (db->db_blkid == DMU_BONUS_BLKID) {
734 		/* Note that the data bufs here are zio_bufs */
735 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
736 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
737 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
738 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
739 		int size = db->db.db_size;
740 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
741 		spa_t *spa = db->db_objset->os_spa;
742 
743 		dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
744 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
745 	} else {
746 		dbuf_set_data(db, NULL);
747 	}
748 }
749 
750 void
751 dbuf_unoverride(dbuf_dirty_record_t *dr)
752 {
753 	dmu_buf_impl_t *db = dr->dr_dbuf;
754 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
755 	uint64_t txg = dr->dr_txg;
756 
757 	ASSERT(MUTEX_HELD(&db->db_mtx));
758 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
759 	ASSERT(db->db_level == 0);
760 
761 	if (db->db_blkid == DMU_BONUS_BLKID ||
762 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
763 		return;
764 
765 	ASSERT(db->db_data_pending != dr);
766 
767 	/* free this block */
768 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
769 		zio_free(db->db_objset->os_spa, txg, bp);
770 
771 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
772 	dr->dt.dl.dr_nopwrite = B_FALSE;
773 
774 	/*
775 	 * Release the already-written buffer, so we leave it in
776 	 * a consistent dirty state.  Note that all callers are
777 	 * modifying the buffer, so they will immediately do
778 	 * another (redundant) arc_release().  Therefore, leave
779 	 * the buf thawed to save the effort of freezing &
780 	 * immediately re-thawing it.
781 	 */
782 	arc_release(dr->dt.dl.dr_data, db);
783 }
784 
785 /*
786  * Evict (if its unreferenced) or clear (if its referenced) any level-0
787  * data blocks in the free range, so that any future readers will find
788  * empty blocks.
789  *
790  * This is a no-op if the dataset is in the middle of an incremental
791  * receive; see comment below for details.
792  */
793 void
794 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
795     dmu_tx_t *tx)
796 {
797 	dmu_buf_impl_t *db, *db_next, db_search;
798 	uint64_t txg = tx->tx_txg;
799 	avl_index_t where;
800 
801 	if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
802 		end_blkid = dn->dn_maxblkid;
803 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
804 
805 	db_search.db_level = 0;
806 	db_search.db_blkid = start_blkid;
807 	db_search.db_state = DB_SEARCH;
808 
809 	mutex_enter(&dn->dn_dbufs_mtx);
810 	if (start_blkid >= dn->dn_unlisted_l0_blkid) {
811 		/* There can't be any dbufs in this range; no need to search. */
812 #ifdef DEBUG
813 		db = avl_find(&dn->dn_dbufs, &db_search, &where);
814 		ASSERT3P(db, ==, NULL);
815 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
816 		ASSERT(db == NULL || db->db_level > 0);
817 #endif
818 		mutex_exit(&dn->dn_dbufs_mtx);
819 		return;
820 	} else if (dmu_objset_is_receiving(dn->dn_objset)) {
821 		/*
822 		 * If we are receiving, we expect there to be no dbufs in
823 		 * the range to be freed, because receive modifies each
824 		 * block at most once, and in offset order.  If this is
825 		 * not the case, it can lead to performance problems,
826 		 * so note that we unexpectedly took the slow path.
827 		 */
828 		atomic_inc_64(&zfs_free_range_recv_miss);
829 	}
830 
831 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
832 	ASSERT3P(db, ==, NULL);
833 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
834 
835 	for (; db != NULL; db = db_next) {
836 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
837 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
838 
839 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
840 			break;
841 		}
842 		ASSERT3U(db->db_blkid, >=, start_blkid);
843 
844 		/* found a level 0 buffer in the range */
845 		mutex_enter(&db->db_mtx);
846 		if (dbuf_undirty(db, tx)) {
847 			/* mutex has been dropped and dbuf destroyed */
848 			continue;
849 		}
850 
851 		if (db->db_state == DB_UNCACHED ||
852 		    db->db_state == DB_NOFILL ||
853 		    db->db_state == DB_EVICTING) {
854 			ASSERT(db->db.db_data == NULL);
855 			mutex_exit(&db->db_mtx);
856 			continue;
857 		}
858 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
859 			/* will be handled in dbuf_read_done or dbuf_rele */
860 			db->db_freed_in_flight = TRUE;
861 			mutex_exit(&db->db_mtx);
862 			continue;
863 		}
864 		if (refcount_count(&db->db_holds) == 0) {
865 			ASSERT(db->db_buf);
866 			dbuf_clear(db);
867 			continue;
868 		}
869 		/* The dbuf is referenced */
870 
871 		if (db->db_last_dirty != NULL) {
872 			dbuf_dirty_record_t *dr = db->db_last_dirty;
873 
874 			if (dr->dr_txg == txg) {
875 				/*
876 				 * This buffer is "in-use", re-adjust the file
877 				 * size to reflect that this buffer may
878 				 * contain new data when we sync.
879 				 */
880 				if (db->db_blkid != DMU_SPILL_BLKID &&
881 				    db->db_blkid > dn->dn_maxblkid)
882 					dn->dn_maxblkid = db->db_blkid;
883 				dbuf_unoverride(dr);
884 			} else {
885 				/*
886 				 * This dbuf is not dirty in the open context.
887 				 * Either uncache it (if its not referenced in
888 				 * the open context) or reset its contents to
889 				 * empty.
890 				 */
891 				dbuf_fix_old_data(db, txg);
892 			}
893 		}
894 		/* clear the contents if its cached */
895 		if (db->db_state == DB_CACHED) {
896 			ASSERT(db->db.db_data != NULL);
897 			arc_release(db->db_buf, db);
898 			bzero(db->db.db_data, db->db.db_size);
899 			arc_buf_freeze(db->db_buf);
900 		}
901 
902 		mutex_exit(&db->db_mtx);
903 	}
904 	mutex_exit(&dn->dn_dbufs_mtx);
905 }
906 
907 static int
908 dbuf_block_freeable(dmu_buf_impl_t *db)
909 {
910 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
911 	uint64_t birth_txg = 0;
912 
913 	/*
914 	 * We don't need any locking to protect db_blkptr:
915 	 * If it's syncing, then db_last_dirty will be set
916 	 * so we'll ignore db_blkptr.
917 	 *
918 	 * This logic ensures that only block births for
919 	 * filled blocks are considered.
920 	 */
921 	ASSERT(MUTEX_HELD(&db->db_mtx));
922 	if (db->db_last_dirty && (db->db_blkptr == NULL ||
923 	    !BP_IS_HOLE(db->db_blkptr))) {
924 		birth_txg = db->db_last_dirty->dr_txg;
925 	} else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
926 		birth_txg = db->db_blkptr->blk_birth;
927 	}
928 
929 	/*
930 	 * If this block don't exist or is in a snapshot, it can't be freed.
931 	 * Don't pass the bp to dsl_dataset_block_freeable() since we
932 	 * are holding the db_mtx lock and might deadlock if we are
933 	 * prefetching a dedup-ed block.
934 	 */
935 	if (birth_txg != 0)
936 		return (ds == NULL ||
937 		    dsl_dataset_block_freeable(ds, NULL, birth_txg));
938 	else
939 		return (B_FALSE);
940 }
941 
942 void
943 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
944 {
945 	arc_buf_t *buf, *obuf;
946 	int osize = db->db.db_size;
947 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
948 	dnode_t *dn;
949 
950 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
951 
952 	DB_DNODE_ENTER(db);
953 	dn = DB_DNODE(db);
954 
955 	/* XXX does *this* func really need the lock? */
956 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
957 
958 	/*
959 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
960 	 * is OK, because there can be no other references to the db
961 	 * when we are changing its size, so no concurrent DB_FILL can
962 	 * be happening.
963 	 */
964 	/*
965 	 * XXX we should be doing a dbuf_read, checking the return
966 	 * value and returning that up to our callers
967 	 */
968 	dmu_buf_will_dirty(&db->db, tx);
969 
970 	/* create the data buffer for the new block */
971 	buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
972 
973 	/* copy old block data to the new block */
974 	obuf = db->db_buf;
975 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
976 	/* zero the remainder */
977 	if (size > osize)
978 		bzero((uint8_t *)buf->b_data + osize, size - osize);
979 
980 	mutex_enter(&db->db_mtx);
981 	dbuf_set_data(db, buf);
982 	VERIFY(arc_buf_remove_ref(obuf, db));
983 	db->db.db_size = size;
984 
985 	if (db->db_level == 0) {
986 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
987 		db->db_last_dirty->dt.dl.dr_data = buf;
988 	}
989 	mutex_exit(&db->db_mtx);
990 
991 	dnode_willuse_space(dn, size-osize, tx);
992 	DB_DNODE_EXIT(db);
993 }
994 
995 void
996 dbuf_release_bp(dmu_buf_impl_t *db)
997 {
998 	objset_t *os = db->db_objset;
999 
1000 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1001 	ASSERT(arc_released(os->os_phys_buf) ||
1002 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
1003 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1004 
1005 	(void) arc_release(db->db_buf, db);
1006 }
1007 
1008 dbuf_dirty_record_t *
1009 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1010 {
1011 	dnode_t *dn;
1012 	objset_t *os;
1013 	dbuf_dirty_record_t **drp, *dr;
1014 	int drop_struct_lock = FALSE;
1015 	boolean_t do_free_accounting = B_FALSE;
1016 	int txgoff = tx->tx_txg & TXG_MASK;
1017 
1018 	ASSERT(tx->tx_txg != 0);
1019 	ASSERT(!refcount_is_zero(&db->db_holds));
1020 	DMU_TX_DIRTY_BUF(tx, db);
1021 
1022 	DB_DNODE_ENTER(db);
1023 	dn = DB_DNODE(db);
1024 	/*
1025 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1026 	 * objects may be dirtied in syncing context, but only if they
1027 	 * were already pre-dirtied in open context.
1028 	 */
1029 	ASSERT(!dmu_tx_is_syncing(tx) ||
1030 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1031 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1032 	    dn->dn_objset->os_dsl_dataset == NULL);
1033 	/*
1034 	 * We make this assert for private objects as well, but after we
1035 	 * check if we're already dirty.  They are allowed to re-dirty
1036 	 * in syncing context.
1037 	 */
1038 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1039 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1040 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1041 
1042 	mutex_enter(&db->db_mtx);
1043 	/*
1044 	 * XXX make this true for indirects too?  The problem is that
1045 	 * transactions created with dmu_tx_create_assigned() from
1046 	 * syncing context don't bother holding ahead.
1047 	 */
1048 	ASSERT(db->db_level != 0 ||
1049 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1050 	    db->db_state == DB_NOFILL);
1051 
1052 	mutex_enter(&dn->dn_mtx);
1053 	/*
1054 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1055 	 * initialize the objset.
1056 	 */
1057 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1058 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1059 		dn->dn_dirtyctx =
1060 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1061 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
1062 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1063 	}
1064 	mutex_exit(&dn->dn_mtx);
1065 
1066 	if (db->db_blkid == DMU_SPILL_BLKID)
1067 		dn->dn_have_spill = B_TRUE;
1068 
1069 	/*
1070 	 * If this buffer is already dirty, we're done.
1071 	 */
1072 	drp = &db->db_last_dirty;
1073 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1074 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1075 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1076 		drp = &dr->dr_next;
1077 	if (dr && dr->dr_txg == tx->tx_txg) {
1078 		DB_DNODE_EXIT(db);
1079 
1080 		if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1081 			/*
1082 			 * If this buffer has already been written out,
1083 			 * we now need to reset its state.
1084 			 */
1085 			dbuf_unoverride(dr);
1086 			if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1087 			    db->db_state != DB_NOFILL)
1088 				arc_buf_thaw(db->db_buf);
1089 		}
1090 		mutex_exit(&db->db_mtx);
1091 		return (dr);
1092 	}
1093 
1094 	/*
1095 	 * Only valid if not already dirty.
1096 	 */
1097 	ASSERT(dn->dn_object == 0 ||
1098 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1099 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1100 
1101 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1102 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1103 	    dn->dn_phys->dn_nlevels > db->db_level ||
1104 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1105 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1106 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1107 
1108 	/*
1109 	 * We should only be dirtying in syncing context if it's the
1110 	 * mos or we're initializing the os or it's a special object.
1111 	 * However, we are allowed to dirty in syncing context provided
1112 	 * we already dirtied it in open context.  Hence we must make
1113 	 * this assertion only if we're not already dirty.
1114 	 */
1115 	os = dn->dn_objset;
1116 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1117 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1118 	ASSERT(db->db.db_size != 0);
1119 
1120 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1121 
1122 	if (db->db_blkid != DMU_BONUS_BLKID) {
1123 		/*
1124 		 * Update the accounting.
1125 		 * Note: we delay "free accounting" until after we drop
1126 		 * the db_mtx.  This keeps us from grabbing other locks
1127 		 * (and possibly deadlocking) in bp_get_dsize() while
1128 		 * also holding the db_mtx.
1129 		 */
1130 		dnode_willuse_space(dn, db->db.db_size, tx);
1131 		do_free_accounting = dbuf_block_freeable(db);
1132 	}
1133 
1134 	/*
1135 	 * If this buffer is dirty in an old transaction group we need
1136 	 * to make a copy of it so that the changes we make in this
1137 	 * transaction group won't leak out when we sync the older txg.
1138 	 */
1139 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1140 	if (db->db_level == 0) {
1141 		void *data_old = db->db_buf;
1142 
1143 		if (db->db_state != DB_NOFILL) {
1144 			if (db->db_blkid == DMU_BONUS_BLKID) {
1145 				dbuf_fix_old_data(db, tx->tx_txg);
1146 				data_old = db->db.db_data;
1147 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1148 				/*
1149 				 * Release the data buffer from the cache so
1150 				 * that we can modify it without impacting
1151 				 * possible other users of this cached data
1152 				 * block.  Note that indirect blocks and
1153 				 * private objects are not released until the
1154 				 * syncing state (since they are only modified
1155 				 * then).
1156 				 */
1157 				arc_release(db->db_buf, db);
1158 				dbuf_fix_old_data(db, tx->tx_txg);
1159 				data_old = db->db_buf;
1160 			}
1161 			ASSERT(data_old != NULL);
1162 		}
1163 		dr->dt.dl.dr_data = data_old;
1164 	} else {
1165 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1166 		list_create(&dr->dt.di.dr_children,
1167 		    sizeof (dbuf_dirty_record_t),
1168 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1169 	}
1170 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1171 		dr->dr_accounted = db->db.db_size;
1172 	dr->dr_dbuf = db;
1173 	dr->dr_txg = tx->tx_txg;
1174 	dr->dr_next = *drp;
1175 	*drp = dr;
1176 
1177 	/*
1178 	 * We could have been freed_in_flight between the dbuf_noread
1179 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1180 	 * happened after the free.
1181 	 */
1182 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1183 	    db->db_blkid != DMU_SPILL_BLKID) {
1184 		mutex_enter(&dn->dn_mtx);
1185 		if (dn->dn_free_ranges[txgoff] != NULL) {
1186 			range_tree_clear(dn->dn_free_ranges[txgoff],
1187 			    db->db_blkid, 1);
1188 		}
1189 		mutex_exit(&dn->dn_mtx);
1190 		db->db_freed_in_flight = FALSE;
1191 	}
1192 
1193 	/*
1194 	 * This buffer is now part of this txg
1195 	 */
1196 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1197 	db->db_dirtycnt += 1;
1198 	ASSERT3U(db->db_dirtycnt, <=, 3);
1199 
1200 	mutex_exit(&db->db_mtx);
1201 
1202 	if (db->db_blkid == DMU_BONUS_BLKID ||
1203 	    db->db_blkid == DMU_SPILL_BLKID) {
1204 		mutex_enter(&dn->dn_mtx);
1205 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1206 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1207 		mutex_exit(&dn->dn_mtx);
1208 		dnode_setdirty(dn, tx);
1209 		DB_DNODE_EXIT(db);
1210 		return (dr);
1211 	} else if (do_free_accounting) {
1212 		blkptr_t *bp = db->db_blkptr;
1213 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1214 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1215 		/*
1216 		 * This is only a guess -- if the dbuf is dirty
1217 		 * in a previous txg, we don't know how much
1218 		 * space it will use on disk yet.  We should
1219 		 * really have the struct_rwlock to access
1220 		 * db_blkptr, but since this is just a guess,
1221 		 * it's OK if we get an odd answer.
1222 		 */
1223 		ddt_prefetch(os->os_spa, bp);
1224 		dnode_willuse_space(dn, -willfree, tx);
1225 	}
1226 
1227 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1228 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1229 		drop_struct_lock = TRUE;
1230 	}
1231 
1232 	if (db->db_level == 0) {
1233 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1234 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1235 	}
1236 
1237 	if (db->db_level+1 < dn->dn_nlevels) {
1238 		dmu_buf_impl_t *parent = db->db_parent;
1239 		dbuf_dirty_record_t *di;
1240 		int parent_held = FALSE;
1241 
1242 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1243 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1244 
1245 			parent = dbuf_hold_level(dn, db->db_level+1,
1246 			    db->db_blkid >> epbs, FTAG);
1247 			ASSERT(parent != NULL);
1248 			parent_held = TRUE;
1249 		}
1250 		if (drop_struct_lock)
1251 			rw_exit(&dn->dn_struct_rwlock);
1252 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1253 		di = dbuf_dirty(parent, tx);
1254 		if (parent_held)
1255 			dbuf_rele(parent, FTAG);
1256 
1257 		mutex_enter(&db->db_mtx);
1258 		/*
1259 		 * Since we've dropped the mutex, it's possible that
1260 		 * dbuf_undirty() might have changed this out from under us.
1261 		 */
1262 		if (db->db_last_dirty == dr ||
1263 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1264 			mutex_enter(&di->dt.di.dr_mtx);
1265 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1266 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1267 			list_insert_tail(&di->dt.di.dr_children, dr);
1268 			mutex_exit(&di->dt.di.dr_mtx);
1269 			dr->dr_parent = di;
1270 		}
1271 		mutex_exit(&db->db_mtx);
1272 	} else {
1273 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1274 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1275 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1276 		mutex_enter(&dn->dn_mtx);
1277 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1278 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1279 		mutex_exit(&dn->dn_mtx);
1280 		if (drop_struct_lock)
1281 			rw_exit(&dn->dn_struct_rwlock);
1282 	}
1283 
1284 	dnode_setdirty(dn, tx);
1285 	DB_DNODE_EXIT(db);
1286 	return (dr);
1287 }
1288 
1289 /*
1290  * Undirty a buffer in the transaction group referenced by the given
1291  * transaction.  Return whether this evicted the dbuf.
1292  */
1293 static boolean_t
1294 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1295 {
1296 	dnode_t *dn;
1297 	uint64_t txg = tx->tx_txg;
1298 	dbuf_dirty_record_t *dr, **drp;
1299 
1300 	ASSERT(txg != 0);
1301 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1302 	ASSERT0(db->db_level);
1303 	ASSERT(MUTEX_HELD(&db->db_mtx));
1304 
1305 	/*
1306 	 * If this buffer is not dirty, we're done.
1307 	 */
1308 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1309 		if (dr->dr_txg <= txg)
1310 			break;
1311 	if (dr == NULL || dr->dr_txg < txg)
1312 		return (B_FALSE);
1313 	ASSERT(dr->dr_txg == txg);
1314 	ASSERT(dr->dr_dbuf == db);
1315 
1316 	DB_DNODE_ENTER(db);
1317 	dn = DB_DNODE(db);
1318 
1319 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1320 
1321 	ASSERT(db->db.db_size != 0);
1322 
1323 	/*
1324 	 * Any space we accounted for in dp_dirty_* will be cleaned up by
1325 	 * dsl_pool_sync().  This is relatively rare so the discrepancy
1326 	 * is not a big deal.
1327 	 */
1328 
1329 	*drp = dr->dr_next;
1330 
1331 	/*
1332 	 * Note that there are three places in dbuf_dirty()
1333 	 * where this dirty record may be put on a list.
1334 	 * Make sure to do a list_remove corresponding to
1335 	 * every one of those list_insert calls.
1336 	 */
1337 	if (dr->dr_parent) {
1338 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1339 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1340 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1341 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1342 	    db->db_level+1 == dn->dn_nlevels) {
1343 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1344 		mutex_enter(&dn->dn_mtx);
1345 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1346 		mutex_exit(&dn->dn_mtx);
1347 	}
1348 	DB_DNODE_EXIT(db);
1349 
1350 	if (db->db_state != DB_NOFILL) {
1351 		dbuf_unoverride(dr);
1352 
1353 		ASSERT(db->db_buf != NULL);
1354 		ASSERT(dr->dt.dl.dr_data != NULL);
1355 		if (dr->dt.dl.dr_data != db->db_buf)
1356 			VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1357 	}
1358 
1359 	if (db->db_level != 0) {
1360 		mutex_destroy(&dr->dt.di.dr_mtx);
1361 		list_destroy(&dr->dt.di.dr_children);
1362 	}
1363 
1364 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1365 
1366 	ASSERT(db->db_dirtycnt > 0);
1367 	db->db_dirtycnt -= 1;
1368 
1369 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1370 		arc_buf_t *buf = db->db_buf;
1371 
1372 		ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1373 		dbuf_set_data(db, NULL);
1374 		VERIFY(arc_buf_remove_ref(buf, db));
1375 		dbuf_evict(db);
1376 		return (B_TRUE);
1377 	}
1378 
1379 	return (B_FALSE);
1380 }
1381 
1382 void
1383 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1384 {
1385 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1386 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1387 
1388 	ASSERT(tx->tx_txg != 0);
1389 	ASSERT(!refcount_is_zero(&db->db_holds));
1390 
1391 	DB_DNODE_ENTER(db);
1392 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1393 		rf |= DB_RF_HAVESTRUCT;
1394 	DB_DNODE_EXIT(db);
1395 	(void) dbuf_read(db, NULL, rf);
1396 	(void) dbuf_dirty(db, tx);
1397 }
1398 
1399 void
1400 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1401 {
1402 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1403 
1404 	db->db_state = DB_NOFILL;
1405 
1406 	dmu_buf_will_fill(db_fake, tx);
1407 }
1408 
1409 void
1410 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1411 {
1412 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1413 
1414 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1415 	ASSERT(tx->tx_txg != 0);
1416 	ASSERT(db->db_level == 0);
1417 	ASSERT(!refcount_is_zero(&db->db_holds));
1418 
1419 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1420 	    dmu_tx_private_ok(tx));
1421 
1422 	dbuf_noread(db);
1423 	(void) dbuf_dirty(db, tx);
1424 }
1425 
1426 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1427 /* ARGSUSED */
1428 void
1429 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1430 {
1431 	mutex_enter(&db->db_mtx);
1432 	DBUF_VERIFY(db);
1433 
1434 	if (db->db_state == DB_FILL) {
1435 		if (db->db_level == 0 && db->db_freed_in_flight) {
1436 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1437 			/* we were freed while filling */
1438 			/* XXX dbuf_undirty? */
1439 			bzero(db->db.db_data, db->db.db_size);
1440 			db->db_freed_in_flight = FALSE;
1441 		}
1442 		db->db_state = DB_CACHED;
1443 		cv_broadcast(&db->db_changed);
1444 	}
1445 	mutex_exit(&db->db_mtx);
1446 }
1447 
1448 void
1449 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1450     bp_embedded_type_t etype, enum zio_compress comp,
1451     int uncompressed_size, int compressed_size, int byteorder,
1452     dmu_tx_t *tx)
1453 {
1454 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1455 	struct dirty_leaf *dl;
1456 	dmu_object_type_t type;
1457 
1458 	DB_DNODE_ENTER(db);
1459 	type = DB_DNODE(db)->dn_type;
1460 	DB_DNODE_EXIT(db);
1461 
1462 	ASSERT0(db->db_level);
1463 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1464 
1465 	dmu_buf_will_not_fill(dbuf, tx);
1466 
1467 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1468 	dl = &db->db_last_dirty->dt.dl;
1469 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
1470 	    data, comp, uncompressed_size, compressed_size);
1471 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1472 	BP_SET_TYPE(&dl->dr_overridden_by, type);
1473 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1474 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1475 
1476 	dl->dr_override_state = DR_OVERRIDDEN;
1477 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1478 }
1479 
1480 /*
1481  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1482  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1483  */
1484 void
1485 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1486 {
1487 	ASSERT(!refcount_is_zero(&db->db_holds));
1488 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1489 	ASSERT(db->db_level == 0);
1490 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1491 	ASSERT(buf != NULL);
1492 	ASSERT(arc_buf_size(buf) == db->db.db_size);
1493 	ASSERT(tx->tx_txg != 0);
1494 
1495 	arc_return_buf(buf, db);
1496 	ASSERT(arc_released(buf));
1497 
1498 	mutex_enter(&db->db_mtx);
1499 
1500 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1501 		cv_wait(&db->db_changed, &db->db_mtx);
1502 
1503 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1504 
1505 	if (db->db_state == DB_CACHED &&
1506 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1507 		mutex_exit(&db->db_mtx);
1508 		(void) dbuf_dirty(db, tx);
1509 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1510 		VERIFY(arc_buf_remove_ref(buf, db));
1511 		xuio_stat_wbuf_copied();
1512 		return;
1513 	}
1514 
1515 	xuio_stat_wbuf_nocopy();
1516 	if (db->db_state == DB_CACHED) {
1517 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1518 
1519 		ASSERT(db->db_buf != NULL);
1520 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1521 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1522 			if (!arc_released(db->db_buf)) {
1523 				ASSERT(dr->dt.dl.dr_override_state ==
1524 				    DR_OVERRIDDEN);
1525 				arc_release(db->db_buf, db);
1526 			}
1527 			dr->dt.dl.dr_data = buf;
1528 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1529 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1530 			arc_release(db->db_buf, db);
1531 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1532 		}
1533 		db->db_buf = NULL;
1534 	}
1535 	ASSERT(db->db_buf == NULL);
1536 	dbuf_set_data(db, buf);
1537 	db->db_state = DB_FILL;
1538 	mutex_exit(&db->db_mtx);
1539 	(void) dbuf_dirty(db, tx);
1540 	dmu_buf_fill_done(&db->db, tx);
1541 }
1542 
1543 /*
1544  * "Clear" the contents of this dbuf.  This will mark the dbuf
1545  * EVICTING and clear *most* of its references.  Unfortunately,
1546  * when we are not holding the dn_dbufs_mtx, we can't clear the
1547  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1548  * in this case.  For callers from the DMU we will usually see:
1549  *	dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1550  * For the arc callback, we will usually see:
1551  *	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1552  * Sometimes, though, we will get a mix of these two:
1553  *	DMU: dbuf_clear()->arc_clear_callback()
1554  *	ARC: dbuf_do_evict()->dbuf_destroy()
1555  *
1556  * This routine will dissociate the dbuf from the arc, by calling
1557  * arc_clear_callback(), but will not evict the data from the ARC.
1558  */
1559 void
1560 dbuf_clear(dmu_buf_impl_t *db)
1561 {
1562 	dnode_t *dn;
1563 	dmu_buf_impl_t *parent = db->db_parent;
1564 	dmu_buf_impl_t *dndb;
1565 	boolean_t dbuf_gone = B_FALSE;
1566 
1567 	ASSERT(MUTEX_HELD(&db->db_mtx));
1568 	ASSERT(refcount_is_zero(&db->db_holds));
1569 
1570 	dbuf_evict_user(db);
1571 
1572 	if (db->db_state == DB_CACHED) {
1573 		ASSERT(db->db.db_data != NULL);
1574 		if (db->db_blkid == DMU_BONUS_BLKID) {
1575 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1576 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1577 		}
1578 		db->db.db_data = NULL;
1579 		db->db_state = DB_UNCACHED;
1580 	}
1581 
1582 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1583 	ASSERT(db->db_data_pending == NULL);
1584 
1585 	db->db_state = DB_EVICTING;
1586 	db->db_blkptr = NULL;
1587 
1588 	DB_DNODE_ENTER(db);
1589 	dn = DB_DNODE(db);
1590 	dndb = dn->dn_dbuf;
1591 	if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1592 		avl_remove(&dn->dn_dbufs, db);
1593 		atomic_dec_32(&dn->dn_dbufs_count);
1594 		membar_producer();
1595 		DB_DNODE_EXIT(db);
1596 		/*
1597 		 * Decrementing the dbuf count means that the hold corresponding
1598 		 * to the removed dbuf is no longer discounted in dnode_move(),
1599 		 * so the dnode cannot be moved until after we release the hold.
1600 		 * The membar_producer() ensures visibility of the decremented
1601 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1602 		 * release any lock.
1603 		 */
1604 		dnode_rele(dn, db);
1605 		db->db_dnode_handle = NULL;
1606 	} else {
1607 		DB_DNODE_EXIT(db);
1608 	}
1609 
1610 	if (db->db_buf)
1611 		dbuf_gone = arc_clear_callback(db->db_buf);
1612 
1613 	if (!dbuf_gone)
1614 		mutex_exit(&db->db_mtx);
1615 
1616 	/*
1617 	 * If this dbuf is referenced from an indirect dbuf,
1618 	 * decrement the ref count on the indirect dbuf.
1619 	 */
1620 	if (parent && parent != dndb)
1621 		dbuf_rele(parent, db);
1622 }
1623 
1624 static int
1625 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1626     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1627 {
1628 	int nlevels, epbs;
1629 
1630 	*parentp = NULL;
1631 	*bpp = NULL;
1632 
1633 	ASSERT(blkid != DMU_BONUS_BLKID);
1634 
1635 	if (blkid == DMU_SPILL_BLKID) {
1636 		mutex_enter(&dn->dn_mtx);
1637 		if (dn->dn_have_spill &&
1638 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1639 			*bpp = &dn->dn_phys->dn_spill;
1640 		else
1641 			*bpp = NULL;
1642 		dbuf_add_ref(dn->dn_dbuf, NULL);
1643 		*parentp = dn->dn_dbuf;
1644 		mutex_exit(&dn->dn_mtx);
1645 		return (0);
1646 	}
1647 
1648 	if (dn->dn_phys->dn_nlevels == 0)
1649 		nlevels = 1;
1650 	else
1651 		nlevels = dn->dn_phys->dn_nlevels;
1652 
1653 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1654 
1655 	ASSERT3U(level * epbs, <, 64);
1656 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1657 	if (level >= nlevels ||
1658 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1659 		/* the buffer has no parent yet */
1660 		return (SET_ERROR(ENOENT));
1661 	} else if (level < nlevels-1) {
1662 		/* this block is referenced from an indirect block */
1663 		int err = dbuf_hold_impl(dn, level+1,
1664 		    blkid >> epbs, fail_sparse, NULL, parentp);
1665 		if (err)
1666 			return (err);
1667 		err = dbuf_read(*parentp, NULL,
1668 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1669 		if (err) {
1670 			dbuf_rele(*parentp, NULL);
1671 			*parentp = NULL;
1672 			return (err);
1673 		}
1674 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1675 		    (blkid & ((1ULL << epbs) - 1));
1676 		return (0);
1677 	} else {
1678 		/* the block is referenced from the dnode */
1679 		ASSERT3U(level, ==, nlevels-1);
1680 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1681 		    blkid < dn->dn_phys->dn_nblkptr);
1682 		if (dn->dn_dbuf) {
1683 			dbuf_add_ref(dn->dn_dbuf, NULL);
1684 			*parentp = dn->dn_dbuf;
1685 		}
1686 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1687 		return (0);
1688 	}
1689 }
1690 
1691 static dmu_buf_impl_t *
1692 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1693     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1694 {
1695 	objset_t *os = dn->dn_objset;
1696 	dmu_buf_impl_t *db, *odb;
1697 
1698 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1699 	ASSERT(dn->dn_type != DMU_OT_NONE);
1700 
1701 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1702 
1703 	db->db_objset = os;
1704 	db->db.db_object = dn->dn_object;
1705 	db->db_level = level;
1706 	db->db_blkid = blkid;
1707 	db->db_last_dirty = NULL;
1708 	db->db_dirtycnt = 0;
1709 	db->db_dnode_handle = dn->dn_handle;
1710 	db->db_parent = parent;
1711 	db->db_blkptr = blkptr;
1712 
1713 	db->db_user_ptr = NULL;
1714 	db->db_evict_func = NULL;
1715 	db->db_immediate_evict = 0;
1716 	db->db_freed_in_flight = 0;
1717 
1718 	if (blkid == DMU_BONUS_BLKID) {
1719 		ASSERT3P(parent, ==, dn->dn_dbuf);
1720 		db->db.db_size = DN_MAX_BONUSLEN -
1721 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1722 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1723 		db->db.db_offset = DMU_BONUS_BLKID;
1724 		db->db_state = DB_UNCACHED;
1725 		/* the bonus dbuf is not placed in the hash table */
1726 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1727 		return (db);
1728 	} else if (blkid == DMU_SPILL_BLKID) {
1729 		db->db.db_size = (blkptr != NULL) ?
1730 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1731 		db->db.db_offset = 0;
1732 	} else {
1733 		int blocksize =
1734 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1735 		db->db.db_size = blocksize;
1736 		db->db.db_offset = db->db_blkid * blocksize;
1737 	}
1738 
1739 	/*
1740 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1741 	 * in the hash table *and* added to the dbufs list.
1742 	 * This prevents a possible deadlock with someone
1743 	 * trying to look up this dbuf before its added to the
1744 	 * dn_dbufs list.
1745 	 */
1746 	mutex_enter(&dn->dn_dbufs_mtx);
1747 	db->db_state = DB_EVICTING;
1748 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1749 		/* someone else inserted it first */
1750 		kmem_cache_free(dbuf_cache, db);
1751 		mutex_exit(&dn->dn_dbufs_mtx);
1752 		return (odb);
1753 	}
1754 	avl_add(&dn->dn_dbufs, db);
1755 	if (db->db_level == 0 && db->db_blkid >=
1756 	    dn->dn_unlisted_l0_blkid)
1757 		dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1758 	db->db_state = DB_UNCACHED;
1759 	mutex_exit(&dn->dn_dbufs_mtx);
1760 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1761 
1762 	if (parent && parent != dn->dn_dbuf)
1763 		dbuf_add_ref(parent, db);
1764 
1765 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1766 	    refcount_count(&dn->dn_holds) > 0);
1767 	(void) refcount_add(&dn->dn_holds, db);
1768 	atomic_inc_32(&dn->dn_dbufs_count);
1769 
1770 	dprintf_dbuf(db, "db=%p\n", db);
1771 
1772 	return (db);
1773 }
1774 
1775 static int
1776 dbuf_do_evict(void *private)
1777 {
1778 	dmu_buf_impl_t *db = private;
1779 
1780 	if (!MUTEX_HELD(&db->db_mtx))
1781 		mutex_enter(&db->db_mtx);
1782 
1783 	ASSERT(refcount_is_zero(&db->db_holds));
1784 
1785 	if (db->db_state != DB_EVICTING) {
1786 		ASSERT(db->db_state == DB_CACHED);
1787 		DBUF_VERIFY(db);
1788 		db->db_buf = NULL;
1789 		dbuf_evict(db);
1790 	} else {
1791 		mutex_exit(&db->db_mtx);
1792 		dbuf_destroy(db);
1793 	}
1794 	return (0);
1795 }
1796 
1797 static void
1798 dbuf_destroy(dmu_buf_impl_t *db)
1799 {
1800 	ASSERT(refcount_is_zero(&db->db_holds));
1801 
1802 	if (db->db_blkid != DMU_BONUS_BLKID) {
1803 		/*
1804 		 * If this dbuf is still on the dn_dbufs list,
1805 		 * remove it from that list.
1806 		 */
1807 		if (db->db_dnode_handle != NULL) {
1808 			dnode_t *dn;
1809 
1810 			DB_DNODE_ENTER(db);
1811 			dn = DB_DNODE(db);
1812 			mutex_enter(&dn->dn_dbufs_mtx);
1813 			avl_remove(&dn->dn_dbufs, db);
1814 			atomic_dec_32(&dn->dn_dbufs_count);
1815 			mutex_exit(&dn->dn_dbufs_mtx);
1816 			DB_DNODE_EXIT(db);
1817 			/*
1818 			 * Decrementing the dbuf count means that the hold
1819 			 * corresponding to the removed dbuf is no longer
1820 			 * discounted in dnode_move(), so the dnode cannot be
1821 			 * moved until after we release the hold.
1822 			 */
1823 			dnode_rele(dn, db);
1824 			db->db_dnode_handle = NULL;
1825 		}
1826 		dbuf_hash_remove(db);
1827 	}
1828 	db->db_parent = NULL;
1829 	db->db_buf = NULL;
1830 
1831 	ASSERT(db->db.db_data == NULL);
1832 	ASSERT(db->db_hash_next == NULL);
1833 	ASSERT(db->db_blkptr == NULL);
1834 	ASSERT(db->db_data_pending == NULL);
1835 
1836 	kmem_cache_free(dbuf_cache, db);
1837 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1838 }
1839 
1840 void
1841 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1842 {
1843 	dmu_buf_impl_t *db = NULL;
1844 	blkptr_t *bp = NULL;
1845 
1846 	ASSERT(blkid != DMU_BONUS_BLKID);
1847 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1848 
1849 	if (dnode_block_freed(dn, blkid))
1850 		return;
1851 
1852 	/* dbuf_find() returns with db_mtx held */
1853 	if (db = dbuf_find(dn, 0, blkid)) {
1854 		/*
1855 		 * This dbuf is already in the cache.  We assume that
1856 		 * it is already CACHED, or else about to be either
1857 		 * read or filled.
1858 		 */
1859 		mutex_exit(&db->db_mtx);
1860 		return;
1861 	}
1862 
1863 	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1864 		if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
1865 			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1866 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1867 			zbookmark_phys_t zb;
1868 
1869 			SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1870 			    dn->dn_object, 0, blkid);
1871 
1872 			(void) arc_read(NULL, dn->dn_objset->os_spa,
1873 			    bp, NULL, NULL, prio,
1874 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1875 			    &aflags, &zb);
1876 		}
1877 		if (db)
1878 			dbuf_rele(db, NULL);
1879 	}
1880 }
1881 
1882 /*
1883  * Returns with db_holds incremented, and db_mtx not held.
1884  * Note: dn_struct_rwlock must be held.
1885  */
1886 int
1887 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1888     void *tag, dmu_buf_impl_t **dbp)
1889 {
1890 	dmu_buf_impl_t *db, *parent = NULL;
1891 
1892 	ASSERT(blkid != DMU_BONUS_BLKID);
1893 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1894 	ASSERT3U(dn->dn_nlevels, >, level);
1895 
1896 	*dbp = NULL;
1897 top:
1898 	/* dbuf_find() returns with db_mtx held */
1899 	db = dbuf_find(dn, level, blkid);
1900 
1901 	if (db == NULL) {
1902 		blkptr_t *bp = NULL;
1903 		int err;
1904 
1905 		ASSERT3P(parent, ==, NULL);
1906 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1907 		if (fail_sparse) {
1908 			if (err == 0 && bp && BP_IS_HOLE(bp))
1909 				err = SET_ERROR(ENOENT);
1910 			if (err) {
1911 				if (parent)
1912 					dbuf_rele(parent, NULL);
1913 				return (err);
1914 			}
1915 		}
1916 		if (err && err != ENOENT)
1917 			return (err);
1918 		db = dbuf_create(dn, level, blkid, parent, bp);
1919 	}
1920 
1921 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1922 		arc_buf_add_ref(db->db_buf, db);
1923 		if (db->db_buf->b_data == NULL) {
1924 			dbuf_clear(db);
1925 			if (parent) {
1926 				dbuf_rele(parent, NULL);
1927 				parent = NULL;
1928 			}
1929 			goto top;
1930 		}
1931 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1932 	}
1933 
1934 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1935 
1936 	/*
1937 	 * If this buffer is currently syncing out, and we are are
1938 	 * still referencing it from db_data, we need to make a copy
1939 	 * of it in case we decide we want to dirty it again in this txg.
1940 	 */
1941 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1942 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
1943 	    db->db_state == DB_CACHED && db->db_data_pending) {
1944 		dbuf_dirty_record_t *dr = db->db_data_pending;
1945 
1946 		if (dr->dt.dl.dr_data == db->db_buf) {
1947 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1948 
1949 			dbuf_set_data(db,
1950 			    arc_buf_alloc(dn->dn_objset->os_spa,
1951 			    db->db.db_size, db, type));
1952 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1953 			    db->db.db_size);
1954 		}
1955 	}
1956 
1957 	(void) refcount_add(&db->db_holds, tag);
1958 	DBUF_VERIFY(db);
1959 	mutex_exit(&db->db_mtx);
1960 
1961 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
1962 	if (parent)
1963 		dbuf_rele(parent, NULL);
1964 
1965 	ASSERT3P(DB_DNODE(db), ==, dn);
1966 	ASSERT3U(db->db_blkid, ==, blkid);
1967 	ASSERT3U(db->db_level, ==, level);
1968 	*dbp = db;
1969 
1970 	return (0);
1971 }
1972 
1973 dmu_buf_impl_t *
1974 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1975 {
1976 	dmu_buf_impl_t *db;
1977 	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1978 	return (err ? NULL : db);
1979 }
1980 
1981 dmu_buf_impl_t *
1982 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1983 {
1984 	dmu_buf_impl_t *db;
1985 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1986 	return (err ? NULL : db);
1987 }
1988 
1989 void
1990 dbuf_create_bonus(dnode_t *dn)
1991 {
1992 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1993 
1994 	ASSERT(dn->dn_bonus == NULL);
1995 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1996 }
1997 
1998 int
1999 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2000 {
2001 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2002 	dnode_t *dn;
2003 
2004 	if (db->db_blkid != DMU_SPILL_BLKID)
2005 		return (SET_ERROR(ENOTSUP));
2006 	if (blksz == 0)
2007 		blksz = SPA_MINBLOCKSIZE;
2008 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2009 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2010 
2011 	DB_DNODE_ENTER(db);
2012 	dn = DB_DNODE(db);
2013 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2014 	dbuf_new_size(db, blksz, tx);
2015 	rw_exit(&dn->dn_struct_rwlock);
2016 	DB_DNODE_EXIT(db);
2017 
2018 	return (0);
2019 }
2020 
2021 void
2022 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2023 {
2024 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2025 }
2026 
2027 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2028 void
2029 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2030 {
2031 	int64_t holds = refcount_add(&db->db_holds, tag);
2032 	ASSERT(holds > 1);
2033 }
2034 
2035 /*
2036  * If you call dbuf_rele() you had better not be referencing the dnode handle
2037  * unless you have some other direct or indirect hold on the dnode. (An indirect
2038  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2039  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2040  * dnode's parent dbuf evicting its dnode handles.
2041  */
2042 void
2043 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2044 {
2045 	mutex_enter(&db->db_mtx);
2046 	dbuf_rele_and_unlock(db, tag);
2047 }
2048 
2049 void
2050 dmu_buf_rele(dmu_buf_t *db, void *tag)
2051 {
2052 	dbuf_rele((dmu_buf_impl_t *)db, tag);
2053 }
2054 
2055 /*
2056  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2057  * db_dirtycnt and db_holds to be updated atomically.
2058  */
2059 void
2060 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2061 {
2062 	int64_t holds;
2063 
2064 	ASSERT(MUTEX_HELD(&db->db_mtx));
2065 	DBUF_VERIFY(db);
2066 
2067 	/*
2068 	 * Remove the reference to the dbuf before removing its hold on the
2069 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2070 	 * buffer has a corresponding dnode hold.
2071 	 */
2072 	holds = refcount_remove(&db->db_holds, tag);
2073 	ASSERT(holds >= 0);
2074 
2075 	/*
2076 	 * We can't freeze indirects if there is a possibility that they
2077 	 * may be modified in the current syncing context.
2078 	 */
2079 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2080 		arc_buf_freeze(db->db_buf);
2081 
2082 	if (holds == db->db_dirtycnt &&
2083 	    db->db_level == 0 && db->db_immediate_evict)
2084 		dbuf_evict_user(db);
2085 
2086 	if (holds == 0) {
2087 		if (db->db_blkid == DMU_BONUS_BLKID) {
2088 			mutex_exit(&db->db_mtx);
2089 
2090 			/*
2091 			 * If the dnode moves here, we cannot cross this barrier
2092 			 * until the move completes.
2093 			 */
2094 			DB_DNODE_ENTER(db);
2095 			atomic_dec_32(&DB_DNODE(db)->dn_dbufs_count);
2096 			DB_DNODE_EXIT(db);
2097 			/*
2098 			 * The bonus buffer's dnode hold is no longer discounted
2099 			 * in dnode_move(). The dnode cannot move until after
2100 			 * the dnode_rele().
2101 			 */
2102 			dnode_rele(DB_DNODE(db), db);
2103 		} else if (db->db_buf == NULL) {
2104 			/*
2105 			 * This is a special case: we never associated this
2106 			 * dbuf with any data allocated from the ARC.
2107 			 */
2108 			ASSERT(db->db_state == DB_UNCACHED ||
2109 			    db->db_state == DB_NOFILL);
2110 			dbuf_evict(db);
2111 		} else if (arc_released(db->db_buf)) {
2112 			arc_buf_t *buf = db->db_buf;
2113 			/*
2114 			 * This dbuf has anonymous data associated with it.
2115 			 */
2116 			dbuf_set_data(db, NULL);
2117 			VERIFY(arc_buf_remove_ref(buf, db));
2118 			dbuf_evict(db);
2119 		} else {
2120 			VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2121 
2122 			/*
2123 			 * A dbuf will be eligible for eviction if either the
2124 			 * 'primarycache' property is set or a duplicate
2125 			 * copy of this buffer is already cached in the arc.
2126 			 *
2127 			 * In the case of the 'primarycache' a buffer
2128 			 * is considered for eviction if it matches the
2129 			 * criteria set in the property.
2130 			 *
2131 			 * To decide if our buffer is considered a
2132 			 * duplicate, we must call into the arc to determine
2133 			 * if multiple buffers are referencing the same
2134 			 * block on-disk. If so, then we simply evict
2135 			 * ourselves.
2136 			 */
2137 			if (!DBUF_IS_CACHEABLE(db)) {
2138 				if (db->db_blkptr != NULL &&
2139 				    !BP_IS_HOLE(db->db_blkptr) &&
2140 				    !BP_IS_EMBEDDED(db->db_blkptr)) {
2141 					spa_t *spa =
2142 					    dmu_objset_spa(db->db_objset);
2143 					blkptr_t bp = *db->db_blkptr;
2144 					dbuf_clear(db);
2145 					arc_freed(spa, &bp);
2146 				} else {
2147 					dbuf_clear(db);
2148 				}
2149 			} else if (arc_buf_eviction_needed(db->db_buf)) {
2150 				dbuf_clear(db);
2151 			} else {
2152 				mutex_exit(&db->db_mtx);
2153 			}
2154 		}
2155 	} else {
2156 		mutex_exit(&db->db_mtx);
2157 	}
2158 }
2159 
2160 #pragma weak dmu_buf_refcount = dbuf_refcount
2161 uint64_t
2162 dbuf_refcount(dmu_buf_impl_t *db)
2163 {
2164 	return (refcount_count(&db->db_holds));
2165 }
2166 
2167 void *
2168 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr,
2169     dmu_buf_evict_func_t *evict_func)
2170 {
2171 	return (dmu_buf_update_user(db_fake, NULL, user_ptr, evict_func));
2172 }
2173 
2174 void *
2175 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr,
2176     dmu_buf_evict_func_t *evict_func)
2177 {
2178 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2179 
2180 	db->db_immediate_evict = TRUE;
2181 	return (dmu_buf_update_user(db_fake, NULL, user_ptr, evict_func));
2182 }
2183 
2184 void *
2185 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2186     dmu_buf_evict_func_t *evict_func)
2187 {
2188 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2189 	ASSERT(db->db_level == 0);
2190 
2191 	ASSERT((user_ptr == NULL) == (evict_func == NULL));
2192 
2193 	mutex_enter(&db->db_mtx);
2194 
2195 	if (db->db_user_ptr == old_user_ptr) {
2196 		db->db_user_ptr = user_ptr;
2197 		db->db_evict_func = evict_func;
2198 	} else {
2199 		old_user_ptr = db->db_user_ptr;
2200 	}
2201 
2202 	mutex_exit(&db->db_mtx);
2203 	return (old_user_ptr);
2204 }
2205 
2206 void *
2207 dmu_buf_get_user(dmu_buf_t *db_fake)
2208 {
2209 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2210 	ASSERT(!refcount_is_zero(&db->db_holds));
2211 
2212 	return (db->db_user_ptr);
2213 }
2214 
2215 boolean_t
2216 dmu_buf_freeable(dmu_buf_t *dbuf)
2217 {
2218 	boolean_t res = B_FALSE;
2219 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2220 
2221 	if (db->db_blkptr)
2222 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2223 		    db->db_blkptr, db->db_blkptr->blk_birth);
2224 
2225 	return (res);
2226 }
2227 
2228 blkptr_t *
2229 dmu_buf_get_blkptr(dmu_buf_t *db)
2230 {
2231 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2232 	return (dbi->db_blkptr);
2233 }
2234 
2235 static void
2236 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2237 {
2238 	/* ASSERT(dmu_tx_is_syncing(tx) */
2239 	ASSERT(MUTEX_HELD(&db->db_mtx));
2240 
2241 	if (db->db_blkptr != NULL)
2242 		return;
2243 
2244 	if (db->db_blkid == DMU_SPILL_BLKID) {
2245 		db->db_blkptr = &dn->dn_phys->dn_spill;
2246 		BP_ZERO(db->db_blkptr);
2247 		return;
2248 	}
2249 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2250 		/*
2251 		 * This buffer was allocated at a time when there was
2252 		 * no available blkptrs from the dnode, or it was
2253 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2254 		 */
2255 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2256 		ASSERT(db->db_parent == NULL);
2257 		db->db_parent = dn->dn_dbuf;
2258 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2259 		DBUF_VERIFY(db);
2260 	} else {
2261 		dmu_buf_impl_t *parent = db->db_parent;
2262 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2263 
2264 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2265 		if (parent == NULL) {
2266 			mutex_exit(&db->db_mtx);
2267 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2268 			(void) dbuf_hold_impl(dn, db->db_level+1,
2269 			    db->db_blkid >> epbs, FALSE, db, &parent);
2270 			rw_exit(&dn->dn_struct_rwlock);
2271 			mutex_enter(&db->db_mtx);
2272 			db->db_parent = parent;
2273 		}
2274 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2275 		    (db->db_blkid & ((1ULL << epbs) - 1));
2276 		DBUF_VERIFY(db);
2277 	}
2278 }
2279 
2280 static void
2281 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2282 {
2283 	dmu_buf_impl_t *db = dr->dr_dbuf;
2284 	dnode_t *dn;
2285 	zio_t *zio;
2286 
2287 	ASSERT(dmu_tx_is_syncing(tx));
2288 
2289 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2290 
2291 	mutex_enter(&db->db_mtx);
2292 
2293 	ASSERT(db->db_level > 0);
2294 	DBUF_VERIFY(db);
2295 
2296 	/* Read the block if it hasn't been read yet. */
2297 	if (db->db_buf == NULL) {
2298 		mutex_exit(&db->db_mtx);
2299 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2300 		mutex_enter(&db->db_mtx);
2301 	}
2302 	ASSERT3U(db->db_state, ==, DB_CACHED);
2303 	ASSERT(db->db_buf != NULL);
2304 
2305 	DB_DNODE_ENTER(db);
2306 	dn = DB_DNODE(db);
2307 	/* Indirect block size must match what the dnode thinks it is. */
2308 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2309 	dbuf_check_blkptr(dn, db);
2310 	DB_DNODE_EXIT(db);
2311 
2312 	/* Provide the pending dirty record to child dbufs */
2313 	db->db_data_pending = dr;
2314 
2315 	mutex_exit(&db->db_mtx);
2316 	dbuf_write(dr, db->db_buf, tx);
2317 
2318 	zio = dr->dr_zio;
2319 	mutex_enter(&dr->dt.di.dr_mtx);
2320 	dbuf_sync_list(&dr->dt.di.dr_children, tx);
2321 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2322 	mutex_exit(&dr->dt.di.dr_mtx);
2323 	zio_nowait(zio);
2324 }
2325 
2326 static void
2327 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2328 {
2329 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2330 	dmu_buf_impl_t *db = dr->dr_dbuf;
2331 	dnode_t *dn;
2332 	objset_t *os;
2333 	uint64_t txg = tx->tx_txg;
2334 
2335 	ASSERT(dmu_tx_is_syncing(tx));
2336 
2337 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2338 
2339 	mutex_enter(&db->db_mtx);
2340 	/*
2341 	 * To be synced, we must be dirtied.  But we
2342 	 * might have been freed after the dirty.
2343 	 */
2344 	if (db->db_state == DB_UNCACHED) {
2345 		/* This buffer has been freed since it was dirtied */
2346 		ASSERT(db->db.db_data == NULL);
2347 	} else if (db->db_state == DB_FILL) {
2348 		/* This buffer was freed and is now being re-filled */
2349 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2350 	} else {
2351 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2352 	}
2353 	DBUF_VERIFY(db);
2354 
2355 	DB_DNODE_ENTER(db);
2356 	dn = DB_DNODE(db);
2357 
2358 	if (db->db_blkid == DMU_SPILL_BLKID) {
2359 		mutex_enter(&dn->dn_mtx);
2360 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2361 		mutex_exit(&dn->dn_mtx);
2362 	}
2363 
2364 	/*
2365 	 * If this is a bonus buffer, simply copy the bonus data into the
2366 	 * dnode.  It will be written out when the dnode is synced (and it
2367 	 * will be synced, since it must have been dirty for dbuf_sync to
2368 	 * be called).
2369 	 */
2370 	if (db->db_blkid == DMU_BONUS_BLKID) {
2371 		dbuf_dirty_record_t **drp;
2372 
2373 		ASSERT(*datap != NULL);
2374 		ASSERT0(db->db_level);
2375 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2376 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2377 		DB_DNODE_EXIT(db);
2378 
2379 		if (*datap != db->db.db_data) {
2380 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
2381 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2382 		}
2383 		db->db_data_pending = NULL;
2384 		drp = &db->db_last_dirty;
2385 		while (*drp != dr)
2386 			drp = &(*drp)->dr_next;
2387 		ASSERT(dr->dr_next == NULL);
2388 		ASSERT(dr->dr_dbuf == db);
2389 		*drp = dr->dr_next;
2390 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
2391 		ASSERT(db->db_dirtycnt > 0);
2392 		db->db_dirtycnt -= 1;
2393 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2394 		return;
2395 	}
2396 
2397 	os = dn->dn_objset;
2398 
2399 	/*
2400 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
2401 	 * operation to sneak in. As a result, we need to ensure that we
2402 	 * don't check the dr_override_state until we have returned from
2403 	 * dbuf_check_blkptr.
2404 	 */
2405 	dbuf_check_blkptr(dn, db);
2406 
2407 	/*
2408 	 * If this buffer is in the middle of an immediate write,
2409 	 * wait for the synchronous IO to complete.
2410 	 */
2411 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2412 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2413 		cv_wait(&db->db_changed, &db->db_mtx);
2414 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2415 	}
2416 
2417 	if (db->db_state != DB_NOFILL &&
2418 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2419 	    refcount_count(&db->db_holds) > 1 &&
2420 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2421 	    *datap == db->db_buf) {
2422 		/*
2423 		 * If this buffer is currently "in use" (i.e., there
2424 		 * are active holds and db_data still references it),
2425 		 * then make a copy before we start the write so that
2426 		 * any modifications from the open txg will not leak
2427 		 * into this write.
2428 		 *
2429 		 * NOTE: this copy does not need to be made for
2430 		 * objects only modified in the syncing context (e.g.
2431 		 * DNONE_DNODE blocks).
2432 		 */
2433 		int blksz = arc_buf_size(*datap);
2434 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2435 		*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2436 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
2437 	}
2438 	db->db_data_pending = dr;
2439 
2440 	mutex_exit(&db->db_mtx);
2441 
2442 	dbuf_write(dr, *datap, tx);
2443 
2444 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2445 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2446 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2447 		DB_DNODE_EXIT(db);
2448 	} else {
2449 		/*
2450 		 * Although zio_nowait() does not "wait for an IO", it does
2451 		 * initiate the IO. If this is an empty write it seems plausible
2452 		 * that the IO could actually be completed before the nowait
2453 		 * returns. We need to DB_DNODE_EXIT() first in case
2454 		 * zio_nowait() invalidates the dbuf.
2455 		 */
2456 		DB_DNODE_EXIT(db);
2457 		zio_nowait(dr->dr_zio);
2458 	}
2459 }
2460 
2461 void
2462 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2463 {
2464 	dbuf_dirty_record_t *dr;
2465 
2466 	while (dr = list_head(list)) {
2467 		if (dr->dr_zio != NULL) {
2468 			/*
2469 			 * If we find an already initialized zio then we
2470 			 * are processing the meta-dnode, and we have finished.
2471 			 * The dbufs for all dnodes are put back on the list
2472 			 * during processing, so that we can zio_wait()
2473 			 * these IOs after initiating all child IOs.
2474 			 */
2475 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2476 			    DMU_META_DNODE_OBJECT);
2477 			break;
2478 		}
2479 		list_remove(list, dr);
2480 		if (dr->dr_dbuf->db_level > 0)
2481 			dbuf_sync_indirect(dr, tx);
2482 		else
2483 			dbuf_sync_leaf(dr, tx);
2484 	}
2485 }
2486 
2487 /* ARGSUSED */
2488 static void
2489 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2490 {
2491 	dmu_buf_impl_t *db = vdb;
2492 	dnode_t *dn;
2493 	blkptr_t *bp = zio->io_bp;
2494 	blkptr_t *bp_orig = &zio->io_bp_orig;
2495 	spa_t *spa = zio->io_spa;
2496 	int64_t delta;
2497 	uint64_t fill = 0;
2498 	int i;
2499 
2500 	ASSERT3P(db->db_blkptr, ==, bp);
2501 
2502 	DB_DNODE_ENTER(db);
2503 	dn = DB_DNODE(db);
2504 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2505 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2506 	zio->io_prev_space_delta = delta;
2507 
2508 	if (bp->blk_birth != 0) {
2509 		ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2510 		    BP_GET_TYPE(bp) == dn->dn_type) ||
2511 		    (db->db_blkid == DMU_SPILL_BLKID &&
2512 		    BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2513 		    BP_IS_EMBEDDED(bp));
2514 		ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2515 	}
2516 
2517 	mutex_enter(&db->db_mtx);
2518 
2519 #ifdef ZFS_DEBUG
2520 	if (db->db_blkid == DMU_SPILL_BLKID) {
2521 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2522 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2523 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2524 	}
2525 #endif
2526 
2527 	if (db->db_level == 0) {
2528 		mutex_enter(&dn->dn_mtx);
2529 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2530 		    db->db_blkid != DMU_SPILL_BLKID)
2531 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2532 		mutex_exit(&dn->dn_mtx);
2533 
2534 		if (dn->dn_type == DMU_OT_DNODE) {
2535 			dnode_phys_t *dnp = db->db.db_data;
2536 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2537 			    i--, dnp++) {
2538 				if (dnp->dn_type != DMU_OT_NONE)
2539 					fill++;
2540 			}
2541 		} else {
2542 			if (BP_IS_HOLE(bp)) {
2543 				fill = 0;
2544 			} else {
2545 				fill = 1;
2546 			}
2547 		}
2548 	} else {
2549 		blkptr_t *ibp = db->db.db_data;
2550 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2551 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2552 			if (BP_IS_HOLE(ibp))
2553 				continue;
2554 			fill += BP_GET_FILL(ibp);
2555 		}
2556 	}
2557 	DB_DNODE_EXIT(db);
2558 
2559 	if (!BP_IS_EMBEDDED(bp))
2560 		bp->blk_fill = fill;
2561 
2562 	mutex_exit(&db->db_mtx);
2563 }
2564 
2565 /*
2566  * The SPA will call this callback several times for each zio - once
2567  * for every physical child i/o (zio->io_phys_children times).  This
2568  * allows the DMU to monitor the progress of each logical i/o.  For example,
2569  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2570  * block.  There may be a long delay before all copies/fragments are completed,
2571  * so this callback allows us to retire dirty space gradually, as the physical
2572  * i/os complete.
2573  */
2574 /* ARGSUSED */
2575 static void
2576 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2577 {
2578 	dmu_buf_impl_t *db = arg;
2579 	objset_t *os = db->db_objset;
2580 	dsl_pool_t *dp = dmu_objset_pool(os);
2581 	dbuf_dirty_record_t *dr;
2582 	int delta = 0;
2583 
2584 	dr = db->db_data_pending;
2585 	ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2586 
2587 	/*
2588 	 * The callback will be called io_phys_children times.  Retire one
2589 	 * portion of our dirty space each time we are called.  Any rounding
2590 	 * error will be cleaned up by dsl_pool_sync()'s call to
2591 	 * dsl_pool_undirty_space().
2592 	 */
2593 	delta = dr->dr_accounted / zio->io_phys_children;
2594 	dsl_pool_undirty_space(dp, delta, zio->io_txg);
2595 }
2596 
2597 /* ARGSUSED */
2598 static void
2599 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2600 {
2601 	dmu_buf_impl_t *db = vdb;
2602 	blkptr_t *bp_orig = &zio->io_bp_orig;
2603 	blkptr_t *bp = db->db_blkptr;
2604 	objset_t *os = db->db_objset;
2605 	dmu_tx_t *tx = os->os_synctx;
2606 	dbuf_dirty_record_t **drp, *dr;
2607 
2608 	ASSERT0(zio->io_error);
2609 	ASSERT(db->db_blkptr == bp);
2610 
2611 	/*
2612 	 * For nopwrites and rewrites we ensure that the bp matches our
2613 	 * original and bypass all the accounting.
2614 	 */
2615 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2616 		ASSERT(BP_EQUAL(bp, bp_orig));
2617 	} else {
2618 		dsl_dataset_t *ds = os->os_dsl_dataset;
2619 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2620 		dsl_dataset_block_born(ds, bp, tx);
2621 	}
2622 
2623 	mutex_enter(&db->db_mtx);
2624 
2625 	DBUF_VERIFY(db);
2626 
2627 	drp = &db->db_last_dirty;
2628 	while ((dr = *drp) != db->db_data_pending)
2629 		drp = &dr->dr_next;
2630 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2631 	ASSERT(dr->dr_dbuf == db);
2632 	ASSERT(dr->dr_next == NULL);
2633 	*drp = dr->dr_next;
2634 
2635 #ifdef ZFS_DEBUG
2636 	if (db->db_blkid == DMU_SPILL_BLKID) {
2637 		dnode_t *dn;
2638 
2639 		DB_DNODE_ENTER(db);
2640 		dn = DB_DNODE(db);
2641 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2642 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2643 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2644 		DB_DNODE_EXIT(db);
2645 	}
2646 #endif
2647 
2648 	if (db->db_level == 0) {
2649 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2650 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2651 		if (db->db_state != DB_NOFILL) {
2652 			if (dr->dt.dl.dr_data != db->db_buf)
2653 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2654 				    db));
2655 			else if (!arc_released(db->db_buf))
2656 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2657 		}
2658 	} else {
2659 		dnode_t *dn;
2660 
2661 		DB_DNODE_ENTER(db);
2662 		dn = DB_DNODE(db);
2663 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2664 		ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2665 		if (!BP_IS_HOLE(db->db_blkptr)) {
2666 			int epbs =
2667 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2668 			ASSERT3U(db->db_blkid, <=,
2669 			    dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2670 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2671 			    db->db.db_size);
2672 			if (!arc_released(db->db_buf))
2673 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2674 		}
2675 		DB_DNODE_EXIT(db);
2676 		mutex_destroy(&dr->dt.di.dr_mtx);
2677 		list_destroy(&dr->dt.di.dr_children);
2678 	}
2679 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
2680 
2681 	cv_broadcast(&db->db_changed);
2682 	ASSERT(db->db_dirtycnt > 0);
2683 	db->db_dirtycnt -= 1;
2684 	db->db_data_pending = NULL;
2685 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2686 }
2687 
2688 static void
2689 dbuf_write_nofill_ready(zio_t *zio)
2690 {
2691 	dbuf_write_ready(zio, NULL, zio->io_private);
2692 }
2693 
2694 static void
2695 dbuf_write_nofill_done(zio_t *zio)
2696 {
2697 	dbuf_write_done(zio, NULL, zio->io_private);
2698 }
2699 
2700 static void
2701 dbuf_write_override_ready(zio_t *zio)
2702 {
2703 	dbuf_dirty_record_t *dr = zio->io_private;
2704 	dmu_buf_impl_t *db = dr->dr_dbuf;
2705 
2706 	dbuf_write_ready(zio, NULL, db);
2707 }
2708 
2709 static void
2710 dbuf_write_override_done(zio_t *zio)
2711 {
2712 	dbuf_dirty_record_t *dr = zio->io_private;
2713 	dmu_buf_impl_t *db = dr->dr_dbuf;
2714 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2715 
2716 	mutex_enter(&db->db_mtx);
2717 	if (!BP_EQUAL(zio->io_bp, obp)) {
2718 		if (!BP_IS_HOLE(obp))
2719 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2720 		arc_release(dr->dt.dl.dr_data, db);
2721 	}
2722 	mutex_exit(&db->db_mtx);
2723 
2724 	dbuf_write_done(zio, NULL, db);
2725 }
2726 
2727 /* Issue I/O to commit a dirty buffer to disk. */
2728 static void
2729 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2730 {
2731 	dmu_buf_impl_t *db = dr->dr_dbuf;
2732 	dnode_t *dn;
2733 	objset_t *os;
2734 	dmu_buf_impl_t *parent = db->db_parent;
2735 	uint64_t txg = tx->tx_txg;
2736 	zbookmark_phys_t zb;
2737 	zio_prop_t zp;
2738 	zio_t *zio;
2739 	int wp_flag = 0;
2740 
2741 	DB_DNODE_ENTER(db);
2742 	dn = DB_DNODE(db);
2743 	os = dn->dn_objset;
2744 
2745 	if (db->db_state != DB_NOFILL) {
2746 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2747 			/*
2748 			 * Private object buffers are released here rather
2749 			 * than in dbuf_dirty() since they are only modified
2750 			 * in the syncing context and we don't want the
2751 			 * overhead of making multiple copies of the data.
2752 			 */
2753 			if (BP_IS_HOLE(db->db_blkptr)) {
2754 				arc_buf_thaw(data);
2755 			} else {
2756 				dbuf_release_bp(db);
2757 			}
2758 		}
2759 	}
2760 
2761 	if (parent != dn->dn_dbuf) {
2762 		/* Our parent is an indirect block. */
2763 		/* We have a dirty parent that has been scheduled for write. */
2764 		ASSERT(parent && parent->db_data_pending);
2765 		/* Our parent's buffer is one level closer to the dnode. */
2766 		ASSERT(db->db_level == parent->db_level-1);
2767 		/*
2768 		 * We're about to modify our parent's db_data by modifying
2769 		 * our block pointer, so the parent must be released.
2770 		 */
2771 		ASSERT(arc_released(parent->db_buf));
2772 		zio = parent->db_data_pending->dr_zio;
2773 	} else {
2774 		/* Our parent is the dnode itself. */
2775 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2776 		    db->db_blkid != DMU_SPILL_BLKID) ||
2777 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2778 		if (db->db_blkid != DMU_SPILL_BLKID)
2779 			ASSERT3P(db->db_blkptr, ==,
2780 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
2781 		zio = dn->dn_zio;
2782 	}
2783 
2784 	ASSERT(db->db_level == 0 || data == db->db_buf);
2785 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2786 	ASSERT(zio);
2787 
2788 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2789 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2790 	    db->db.db_object, db->db_level, db->db_blkid);
2791 
2792 	if (db->db_blkid == DMU_SPILL_BLKID)
2793 		wp_flag = WP_SPILL;
2794 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2795 
2796 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2797 	DB_DNODE_EXIT(db);
2798 
2799 	if (db->db_level == 0 &&
2800 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2801 		/*
2802 		 * The BP for this block has been provided by open context
2803 		 * (by dmu_sync() or dmu_buf_write_embedded()).
2804 		 */
2805 		void *contents = (data != NULL) ? data->b_data : NULL;
2806 
2807 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2808 		    db->db_blkptr, contents, db->db.db_size, &zp,
2809 		    dbuf_write_override_ready, NULL, dbuf_write_override_done,
2810 		    dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2811 		mutex_enter(&db->db_mtx);
2812 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2813 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2814 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2815 		mutex_exit(&db->db_mtx);
2816 	} else if (db->db_state == DB_NOFILL) {
2817 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2818 		    zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2819 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2820 		    db->db_blkptr, NULL, db->db.db_size, &zp,
2821 		    dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2822 		    ZIO_PRIORITY_ASYNC_WRITE,
2823 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2824 	} else {
2825 		ASSERT(arc_released(data));
2826 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
2827 		    db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2828 		    DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2829 		    dbuf_write_physdone, dbuf_write_done, db,
2830 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2831 	}
2832 }
2833