xref: /titanic_41/usr/src/uts/common/fs/zfs/dbuf.c (revision 416a91e24418fb283022309d06e7d7ce0094d70e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
25  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28  */
29 
30 #include <sys/zfs_context.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dbuf.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/spa.h>
40 #include <sys/zio.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/sa.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 
48 /*
49  * Number of times that zfs_free_range() took the slow path while doing
50  * a zfs receive.  A nonzero value indicates a potential performance problem.
51  */
52 uint64_t zfs_free_range_recv_miss;
53 
54 static void dbuf_destroy(dmu_buf_impl_t *db);
55 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
56 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
57 
58 #ifndef __lint
59 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
60     dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp);
61 #endif /* ! __lint */
62 
63 /*
64  * Global data structures and functions for the dbuf cache.
65  */
66 static kmem_cache_t *dbuf_cache;
67 static taskq_t *dbu_evict_taskq;
68 
69 /* ARGSUSED */
70 static int
71 dbuf_cons(void *vdb, void *unused, int kmflag)
72 {
73 	dmu_buf_impl_t *db = vdb;
74 	bzero(db, sizeof (dmu_buf_impl_t));
75 
76 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
77 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
78 	refcount_create(&db->db_holds);
79 
80 	return (0);
81 }
82 
83 /* ARGSUSED */
84 static void
85 dbuf_dest(void *vdb, void *unused)
86 {
87 	dmu_buf_impl_t *db = vdb;
88 	mutex_destroy(&db->db_mtx);
89 	cv_destroy(&db->db_changed);
90 	refcount_destroy(&db->db_holds);
91 }
92 
93 /*
94  * dbuf hash table routines
95  */
96 static dbuf_hash_table_t dbuf_hash_table;
97 
98 static uint64_t dbuf_hash_count;
99 
100 static uint64_t
101 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
102 {
103 	uintptr_t osv = (uintptr_t)os;
104 	uint64_t crc = -1ULL;
105 
106 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
107 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
108 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
109 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
110 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
111 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
112 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
113 
114 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
115 
116 	return (crc);
117 }
118 
119 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
120 
121 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
122 	((dbuf)->db.db_object == (obj) &&		\
123 	(dbuf)->db_objset == (os) &&			\
124 	(dbuf)->db_level == (level) &&			\
125 	(dbuf)->db_blkid == (blkid))
126 
127 dmu_buf_impl_t *
128 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
129 {
130 	dbuf_hash_table_t *h = &dbuf_hash_table;
131 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
132 	uint64_t idx = hv & h->hash_table_mask;
133 	dmu_buf_impl_t *db;
134 
135 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
136 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
137 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
138 			mutex_enter(&db->db_mtx);
139 			if (db->db_state != DB_EVICTING) {
140 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
141 				return (db);
142 			}
143 			mutex_exit(&db->db_mtx);
144 		}
145 	}
146 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
147 	return (NULL);
148 }
149 
150 static dmu_buf_impl_t *
151 dbuf_find_bonus(objset_t *os, uint64_t object)
152 {
153 	dnode_t *dn;
154 	dmu_buf_impl_t *db = NULL;
155 
156 	if (dnode_hold(os, object, FTAG, &dn) == 0) {
157 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
158 		if (dn->dn_bonus != NULL) {
159 			db = dn->dn_bonus;
160 			mutex_enter(&db->db_mtx);
161 		}
162 		rw_exit(&dn->dn_struct_rwlock);
163 		dnode_rele(dn, FTAG);
164 	}
165 	return (db);
166 }
167 
168 /*
169  * Insert an entry into the hash table.  If there is already an element
170  * equal to elem in the hash table, then the already existing element
171  * will be returned and the new element will not be inserted.
172  * Otherwise returns NULL.
173  */
174 static dmu_buf_impl_t *
175 dbuf_hash_insert(dmu_buf_impl_t *db)
176 {
177 	dbuf_hash_table_t *h = &dbuf_hash_table;
178 	objset_t *os = db->db_objset;
179 	uint64_t obj = db->db.db_object;
180 	int level = db->db_level;
181 	uint64_t blkid = db->db_blkid;
182 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
183 	uint64_t idx = hv & h->hash_table_mask;
184 	dmu_buf_impl_t *dbf;
185 
186 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
187 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
188 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
189 			mutex_enter(&dbf->db_mtx);
190 			if (dbf->db_state != DB_EVICTING) {
191 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
192 				return (dbf);
193 			}
194 			mutex_exit(&dbf->db_mtx);
195 		}
196 	}
197 
198 	mutex_enter(&db->db_mtx);
199 	db->db_hash_next = h->hash_table[idx];
200 	h->hash_table[idx] = db;
201 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
202 	atomic_inc_64(&dbuf_hash_count);
203 
204 	return (NULL);
205 }
206 
207 /*
208  * Remove an entry from the hash table.  It must be in the EVICTING state.
209  */
210 static void
211 dbuf_hash_remove(dmu_buf_impl_t *db)
212 {
213 	dbuf_hash_table_t *h = &dbuf_hash_table;
214 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
215 	    db->db_level, db->db_blkid);
216 	uint64_t idx = hv & h->hash_table_mask;
217 	dmu_buf_impl_t *dbf, **dbp;
218 
219 	/*
220 	 * We musn't hold db_mtx to maintain lock ordering:
221 	 * DBUF_HASH_MUTEX > db_mtx.
222 	 */
223 	ASSERT(refcount_is_zero(&db->db_holds));
224 	ASSERT(db->db_state == DB_EVICTING);
225 	ASSERT(!MUTEX_HELD(&db->db_mtx));
226 
227 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
228 	dbp = &h->hash_table[idx];
229 	while ((dbf = *dbp) != db) {
230 		dbp = &dbf->db_hash_next;
231 		ASSERT(dbf != NULL);
232 	}
233 	*dbp = db->db_hash_next;
234 	db->db_hash_next = NULL;
235 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
236 	atomic_dec_64(&dbuf_hash_count);
237 }
238 
239 static arc_evict_func_t dbuf_do_evict;
240 
241 typedef enum {
242 	DBVU_EVICTING,
243 	DBVU_NOT_EVICTING
244 } dbvu_verify_type_t;
245 
246 static void
247 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
248 {
249 #ifdef ZFS_DEBUG
250 	int64_t holds;
251 
252 	if (db->db_user == NULL)
253 		return;
254 
255 	/* Only data blocks support the attachment of user data. */
256 	ASSERT(db->db_level == 0);
257 
258 	/* Clients must resolve a dbuf before attaching user data. */
259 	ASSERT(db->db.db_data != NULL);
260 	ASSERT3U(db->db_state, ==, DB_CACHED);
261 
262 	holds = refcount_count(&db->db_holds);
263 	if (verify_type == DBVU_EVICTING) {
264 		/*
265 		 * Immediate eviction occurs when holds == dirtycnt.
266 		 * For normal eviction buffers, holds is zero on
267 		 * eviction, except when dbuf_fix_old_data() calls
268 		 * dbuf_clear_data().  However, the hold count can grow
269 		 * during eviction even though db_mtx is held (see
270 		 * dmu_bonus_hold() for an example), so we can only
271 		 * test the generic invariant that holds >= dirtycnt.
272 		 */
273 		ASSERT3U(holds, >=, db->db_dirtycnt);
274 	} else {
275 		if (db->db_immediate_evict == TRUE)
276 			ASSERT3U(holds, >=, db->db_dirtycnt);
277 		else
278 			ASSERT3U(holds, >, 0);
279 	}
280 #endif
281 }
282 
283 static void
284 dbuf_evict_user(dmu_buf_impl_t *db)
285 {
286 	dmu_buf_user_t *dbu = db->db_user;
287 
288 	ASSERT(MUTEX_HELD(&db->db_mtx));
289 
290 	if (dbu == NULL)
291 		return;
292 
293 	dbuf_verify_user(db, DBVU_EVICTING);
294 	db->db_user = NULL;
295 
296 #ifdef ZFS_DEBUG
297 	if (dbu->dbu_clear_on_evict_dbufp != NULL)
298 		*dbu->dbu_clear_on_evict_dbufp = NULL;
299 #endif
300 
301 	/*
302 	 * Invoke the callback from a taskq to avoid lock order reversals
303 	 * and limit stack depth.
304 	 */
305 	taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
306 	    &dbu->dbu_tqent);
307 }
308 
309 boolean_t
310 dbuf_is_metadata(dmu_buf_impl_t *db)
311 {
312 	if (db->db_level > 0) {
313 		return (B_TRUE);
314 	} else {
315 		boolean_t is_metadata;
316 
317 		DB_DNODE_ENTER(db);
318 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
319 		DB_DNODE_EXIT(db);
320 
321 		return (is_metadata);
322 	}
323 }
324 
325 void
326 dbuf_evict(dmu_buf_impl_t *db)
327 {
328 	ASSERT(MUTEX_HELD(&db->db_mtx));
329 	ASSERT(db->db_buf == NULL);
330 	ASSERT(db->db_data_pending == NULL);
331 
332 	dbuf_clear(db);
333 	dbuf_destroy(db);
334 }
335 
336 void
337 dbuf_init(void)
338 {
339 	uint64_t hsize = 1ULL << 16;
340 	dbuf_hash_table_t *h = &dbuf_hash_table;
341 	int i;
342 
343 	/*
344 	 * The hash table is big enough to fill all of physical memory
345 	 * with an average 4K block size.  The table will take up
346 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
347 	 */
348 	while (hsize * 4096 < physmem * PAGESIZE)
349 		hsize <<= 1;
350 
351 retry:
352 	h->hash_table_mask = hsize - 1;
353 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
354 	if (h->hash_table == NULL) {
355 		/* XXX - we should really return an error instead of assert */
356 		ASSERT(hsize > (1ULL << 10));
357 		hsize >>= 1;
358 		goto retry;
359 	}
360 
361 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
362 	    sizeof (dmu_buf_impl_t),
363 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
364 
365 	for (i = 0; i < DBUF_MUTEXES; i++)
366 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
367 
368 	/*
369 	 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
370 	 * configuration is not required.
371 	 */
372 	dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
373 }
374 
375 void
376 dbuf_fini(void)
377 {
378 	dbuf_hash_table_t *h = &dbuf_hash_table;
379 	int i;
380 
381 	for (i = 0; i < DBUF_MUTEXES; i++)
382 		mutex_destroy(&h->hash_mutexes[i]);
383 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
384 	kmem_cache_destroy(dbuf_cache);
385 	taskq_destroy(dbu_evict_taskq);
386 }
387 
388 /*
389  * Other stuff.
390  */
391 
392 #ifdef ZFS_DEBUG
393 static void
394 dbuf_verify(dmu_buf_impl_t *db)
395 {
396 	dnode_t *dn;
397 	dbuf_dirty_record_t *dr;
398 
399 	ASSERT(MUTEX_HELD(&db->db_mtx));
400 
401 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
402 		return;
403 
404 	ASSERT(db->db_objset != NULL);
405 	DB_DNODE_ENTER(db);
406 	dn = DB_DNODE(db);
407 	if (dn == NULL) {
408 		ASSERT(db->db_parent == NULL);
409 		ASSERT(db->db_blkptr == NULL);
410 	} else {
411 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
412 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
413 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
414 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
415 		    db->db_blkid == DMU_SPILL_BLKID ||
416 		    !avl_is_empty(&dn->dn_dbufs));
417 	}
418 	if (db->db_blkid == DMU_BONUS_BLKID) {
419 		ASSERT(dn != NULL);
420 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
421 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
422 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
423 		ASSERT(dn != NULL);
424 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
425 		ASSERT0(db->db.db_offset);
426 	} else {
427 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
428 	}
429 
430 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
431 		ASSERT(dr->dr_dbuf == db);
432 
433 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
434 		ASSERT(dr->dr_dbuf == db);
435 
436 	/*
437 	 * We can't assert that db_size matches dn_datablksz because it
438 	 * can be momentarily different when another thread is doing
439 	 * dnode_set_blksz().
440 	 */
441 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
442 		dr = db->db_data_pending;
443 		/*
444 		 * It should only be modified in syncing context, so
445 		 * make sure we only have one copy of the data.
446 		 */
447 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
448 	}
449 
450 	/* verify db->db_blkptr */
451 	if (db->db_blkptr) {
452 		if (db->db_parent == dn->dn_dbuf) {
453 			/* db is pointed to by the dnode */
454 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
455 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
456 				ASSERT(db->db_parent == NULL);
457 			else
458 				ASSERT(db->db_parent != NULL);
459 			if (db->db_blkid != DMU_SPILL_BLKID)
460 				ASSERT3P(db->db_blkptr, ==,
461 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
462 		} else {
463 			/* db is pointed to by an indirect block */
464 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
465 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
466 			ASSERT3U(db->db_parent->db.db_object, ==,
467 			    db->db.db_object);
468 			/*
469 			 * dnode_grow_indblksz() can make this fail if we don't
470 			 * have the struct_rwlock.  XXX indblksz no longer
471 			 * grows.  safe to do this now?
472 			 */
473 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
474 				ASSERT3P(db->db_blkptr, ==,
475 				    ((blkptr_t *)db->db_parent->db.db_data +
476 				    db->db_blkid % epb));
477 			}
478 		}
479 	}
480 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
481 	    (db->db_buf == NULL || db->db_buf->b_data) &&
482 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
483 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
484 		/*
485 		 * If the blkptr isn't set but they have nonzero data,
486 		 * it had better be dirty, otherwise we'll lose that
487 		 * data when we evict this buffer.
488 		 */
489 		if (db->db_dirtycnt == 0) {
490 			uint64_t *buf = db->db.db_data;
491 			int i;
492 
493 			for (i = 0; i < db->db.db_size >> 3; i++) {
494 				ASSERT(buf[i] == 0);
495 			}
496 		}
497 	}
498 	DB_DNODE_EXIT(db);
499 }
500 #endif
501 
502 static void
503 dbuf_clear_data(dmu_buf_impl_t *db)
504 {
505 	ASSERT(MUTEX_HELD(&db->db_mtx));
506 	dbuf_evict_user(db);
507 	db->db_buf = NULL;
508 	db->db.db_data = NULL;
509 	if (db->db_state != DB_NOFILL)
510 		db->db_state = DB_UNCACHED;
511 }
512 
513 static void
514 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
515 {
516 	ASSERT(MUTEX_HELD(&db->db_mtx));
517 	ASSERT(buf != NULL);
518 
519 	db->db_buf = buf;
520 	ASSERT(buf->b_data != NULL);
521 	db->db.db_data = buf->b_data;
522 	if (!arc_released(buf))
523 		arc_set_callback(buf, dbuf_do_evict, db);
524 }
525 
526 /*
527  * Loan out an arc_buf for read.  Return the loaned arc_buf.
528  */
529 arc_buf_t *
530 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
531 {
532 	arc_buf_t *abuf;
533 
534 	mutex_enter(&db->db_mtx);
535 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
536 		int blksz = db->db.db_size;
537 		spa_t *spa = db->db_objset->os_spa;
538 
539 		mutex_exit(&db->db_mtx);
540 		abuf = arc_loan_buf(spa, blksz);
541 		bcopy(db->db.db_data, abuf->b_data, blksz);
542 	} else {
543 		abuf = db->db_buf;
544 		arc_loan_inuse_buf(abuf, db);
545 		dbuf_clear_data(db);
546 		mutex_exit(&db->db_mtx);
547 	}
548 	return (abuf);
549 }
550 
551 /*
552  * Calculate which level n block references the data at the level 0 offset
553  * provided.
554  */
555 uint64_t
556 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
557 {
558 	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
559 		/*
560 		 * The level n blkid is equal to the level 0 blkid divided by
561 		 * the number of level 0s in a level n block.
562 		 *
563 		 * The level 0 blkid is offset >> datablkshift =
564 		 * offset / 2^datablkshift.
565 		 *
566 		 * The number of level 0s in a level n is the number of block
567 		 * pointers in an indirect block, raised to the power of level.
568 		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
569 		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
570 		 *
571 		 * Thus, the level n blkid is: offset /
572 		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
573 		 * = offset / 2^(datablkshift + level *
574 		 *   (indblkshift - SPA_BLKPTRSHIFT))
575 		 * = offset >> (datablkshift + level *
576 		 *   (indblkshift - SPA_BLKPTRSHIFT))
577 		 */
578 		return (offset >> (dn->dn_datablkshift + level *
579 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
580 	} else {
581 		ASSERT3U(offset, <, dn->dn_datablksz);
582 		return (0);
583 	}
584 }
585 
586 static void
587 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
588 {
589 	dmu_buf_impl_t *db = vdb;
590 
591 	mutex_enter(&db->db_mtx);
592 	ASSERT3U(db->db_state, ==, DB_READ);
593 	/*
594 	 * All reads are synchronous, so we must have a hold on the dbuf
595 	 */
596 	ASSERT(refcount_count(&db->db_holds) > 0);
597 	ASSERT(db->db_buf == NULL);
598 	ASSERT(db->db.db_data == NULL);
599 	if (db->db_level == 0 && db->db_freed_in_flight) {
600 		/* we were freed in flight; disregard any error */
601 		arc_release(buf, db);
602 		bzero(buf->b_data, db->db.db_size);
603 		arc_buf_freeze(buf);
604 		db->db_freed_in_flight = FALSE;
605 		dbuf_set_data(db, buf);
606 		db->db_state = DB_CACHED;
607 	} else if (zio == NULL || zio->io_error == 0) {
608 		dbuf_set_data(db, buf);
609 		db->db_state = DB_CACHED;
610 	} else {
611 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
612 		ASSERT3P(db->db_buf, ==, NULL);
613 		VERIFY(arc_buf_remove_ref(buf, db));
614 		db->db_state = DB_UNCACHED;
615 	}
616 	cv_broadcast(&db->db_changed);
617 	dbuf_rele_and_unlock(db, NULL);
618 }
619 
620 static void
621 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
622 {
623 	dnode_t *dn;
624 	zbookmark_phys_t zb;
625 	arc_flags_t aflags = ARC_FLAG_NOWAIT;
626 
627 	DB_DNODE_ENTER(db);
628 	dn = DB_DNODE(db);
629 	ASSERT(!refcount_is_zero(&db->db_holds));
630 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
631 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
632 	ASSERT(MUTEX_HELD(&db->db_mtx));
633 	ASSERT(db->db_state == DB_UNCACHED);
634 	ASSERT(db->db_buf == NULL);
635 
636 	if (db->db_blkid == DMU_BONUS_BLKID) {
637 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
638 
639 		ASSERT3U(bonuslen, <=, db->db.db_size);
640 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
641 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
642 		if (bonuslen < DN_MAX_BONUSLEN)
643 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
644 		if (bonuslen)
645 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
646 		DB_DNODE_EXIT(db);
647 		db->db_state = DB_CACHED;
648 		mutex_exit(&db->db_mtx);
649 		return;
650 	}
651 
652 	/*
653 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
654 	 * processes the delete record and clears the bp while we are waiting
655 	 * for the dn_mtx (resulting in a "no" from block_freed).
656 	 */
657 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
658 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
659 	    BP_IS_HOLE(db->db_blkptr)))) {
660 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
661 
662 		DB_DNODE_EXIT(db);
663 		dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
664 		    db->db.db_size, db, type));
665 		bzero(db->db.db_data, db->db.db_size);
666 		db->db_state = DB_CACHED;
667 		*flags |= DB_RF_CACHED;
668 		mutex_exit(&db->db_mtx);
669 		return;
670 	}
671 
672 	DB_DNODE_EXIT(db);
673 
674 	db->db_state = DB_READ;
675 	mutex_exit(&db->db_mtx);
676 
677 	if (DBUF_IS_L2CACHEABLE(db))
678 		aflags |= ARC_FLAG_L2CACHE;
679 	if (DBUF_IS_L2COMPRESSIBLE(db))
680 		aflags |= ARC_FLAG_L2COMPRESS;
681 
682 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
683 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
684 	    db->db.db_object, db->db_level, db->db_blkid);
685 
686 	dbuf_add_ref(db, NULL);
687 
688 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
689 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
690 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
691 	    &aflags, &zb);
692 	if (aflags & ARC_FLAG_CACHED)
693 		*flags |= DB_RF_CACHED;
694 }
695 
696 int
697 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
698 {
699 	int err = 0;
700 	boolean_t havepzio = (zio != NULL);
701 	boolean_t prefetch;
702 	dnode_t *dn;
703 
704 	/*
705 	 * We don't have to hold the mutex to check db_state because it
706 	 * can't be freed while we have a hold on the buffer.
707 	 */
708 	ASSERT(!refcount_is_zero(&db->db_holds));
709 
710 	if (db->db_state == DB_NOFILL)
711 		return (SET_ERROR(EIO));
712 
713 	DB_DNODE_ENTER(db);
714 	dn = DB_DNODE(db);
715 	if ((flags & DB_RF_HAVESTRUCT) == 0)
716 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
717 
718 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
719 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
720 	    DBUF_IS_CACHEABLE(db);
721 
722 	mutex_enter(&db->db_mtx);
723 	if (db->db_state == DB_CACHED) {
724 		mutex_exit(&db->db_mtx);
725 		if (prefetch)
726 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
727 			    db->db.db_size, TRUE);
728 		if ((flags & DB_RF_HAVESTRUCT) == 0)
729 			rw_exit(&dn->dn_struct_rwlock);
730 		DB_DNODE_EXIT(db);
731 	} else if (db->db_state == DB_UNCACHED) {
732 		spa_t *spa = dn->dn_objset->os_spa;
733 
734 		if (zio == NULL)
735 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
736 		dbuf_read_impl(db, zio, &flags);
737 
738 		/* dbuf_read_impl has dropped db_mtx for us */
739 
740 		if (prefetch)
741 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
742 			    db->db.db_size, flags & DB_RF_CACHED);
743 
744 		if ((flags & DB_RF_HAVESTRUCT) == 0)
745 			rw_exit(&dn->dn_struct_rwlock);
746 		DB_DNODE_EXIT(db);
747 
748 		if (!havepzio)
749 			err = zio_wait(zio);
750 	} else {
751 		/*
752 		 * Another reader came in while the dbuf was in flight
753 		 * between UNCACHED and CACHED.  Either a writer will finish
754 		 * writing the buffer (sending the dbuf to CACHED) or the
755 		 * first reader's request will reach the read_done callback
756 		 * and send the dbuf to CACHED.  Otherwise, a failure
757 		 * occurred and the dbuf went to UNCACHED.
758 		 */
759 		mutex_exit(&db->db_mtx);
760 		if (prefetch)
761 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
762 			    db->db.db_size, TRUE);
763 		if ((flags & DB_RF_HAVESTRUCT) == 0)
764 			rw_exit(&dn->dn_struct_rwlock);
765 		DB_DNODE_EXIT(db);
766 
767 		/* Skip the wait per the caller's request. */
768 		mutex_enter(&db->db_mtx);
769 		if ((flags & DB_RF_NEVERWAIT) == 0) {
770 			while (db->db_state == DB_READ ||
771 			    db->db_state == DB_FILL) {
772 				ASSERT(db->db_state == DB_READ ||
773 				    (flags & DB_RF_HAVESTRUCT) == 0);
774 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
775 				    db, zio_t *, zio);
776 				cv_wait(&db->db_changed, &db->db_mtx);
777 			}
778 			if (db->db_state == DB_UNCACHED)
779 				err = SET_ERROR(EIO);
780 		}
781 		mutex_exit(&db->db_mtx);
782 	}
783 
784 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
785 	return (err);
786 }
787 
788 static void
789 dbuf_noread(dmu_buf_impl_t *db)
790 {
791 	ASSERT(!refcount_is_zero(&db->db_holds));
792 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
793 	mutex_enter(&db->db_mtx);
794 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
795 		cv_wait(&db->db_changed, &db->db_mtx);
796 	if (db->db_state == DB_UNCACHED) {
797 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
798 		spa_t *spa = db->db_objset->os_spa;
799 
800 		ASSERT(db->db_buf == NULL);
801 		ASSERT(db->db.db_data == NULL);
802 		dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
803 		db->db_state = DB_FILL;
804 	} else if (db->db_state == DB_NOFILL) {
805 		dbuf_clear_data(db);
806 	} else {
807 		ASSERT3U(db->db_state, ==, DB_CACHED);
808 	}
809 	mutex_exit(&db->db_mtx);
810 }
811 
812 /*
813  * This is our just-in-time copy function.  It makes a copy of
814  * buffers, that have been modified in a previous transaction
815  * group, before we modify them in the current active group.
816  *
817  * This function is used in two places: when we are dirtying a
818  * buffer for the first time in a txg, and when we are freeing
819  * a range in a dnode that includes this buffer.
820  *
821  * Note that when we are called from dbuf_free_range() we do
822  * not put a hold on the buffer, we just traverse the active
823  * dbuf list for the dnode.
824  */
825 static void
826 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
827 {
828 	dbuf_dirty_record_t *dr = db->db_last_dirty;
829 
830 	ASSERT(MUTEX_HELD(&db->db_mtx));
831 	ASSERT(db->db.db_data != NULL);
832 	ASSERT(db->db_level == 0);
833 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
834 
835 	if (dr == NULL ||
836 	    (dr->dt.dl.dr_data !=
837 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
838 		return;
839 
840 	/*
841 	 * If the last dirty record for this dbuf has not yet synced
842 	 * and its referencing the dbuf data, either:
843 	 *	reset the reference to point to a new copy,
844 	 * or (if there a no active holders)
845 	 *	just null out the current db_data pointer.
846 	 */
847 	ASSERT(dr->dr_txg >= txg - 2);
848 	if (db->db_blkid == DMU_BONUS_BLKID) {
849 		/* Note that the data bufs here are zio_bufs */
850 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
851 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
852 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
853 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
854 		int size = db->db.db_size;
855 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
856 		spa_t *spa = db->db_objset->os_spa;
857 
858 		dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
859 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
860 	} else {
861 		dbuf_clear_data(db);
862 	}
863 }
864 
865 void
866 dbuf_unoverride(dbuf_dirty_record_t *dr)
867 {
868 	dmu_buf_impl_t *db = dr->dr_dbuf;
869 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
870 	uint64_t txg = dr->dr_txg;
871 
872 	ASSERT(MUTEX_HELD(&db->db_mtx));
873 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
874 	ASSERT(db->db_level == 0);
875 
876 	if (db->db_blkid == DMU_BONUS_BLKID ||
877 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
878 		return;
879 
880 	ASSERT(db->db_data_pending != dr);
881 
882 	/* free this block */
883 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
884 		zio_free(db->db_objset->os_spa, txg, bp);
885 
886 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
887 	dr->dt.dl.dr_nopwrite = B_FALSE;
888 
889 	/*
890 	 * Release the already-written buffer, so we leave it in
891 	 * a consistent dirty state.  Note that all callers are
892 	 * modifying the buffer, so they will immediately do
893 	 * another (redundant) arc_release().  Therefore, leave
894 	 * the buf thawed to save the effort of freezing &
895 	 * immediately re-thawing it.
896 	 */
897 	arc_release(dr->dt.dl.dr_data, db);
898 }
899 
900 /*
901  * Evict (if its unreferenced) or clear (if its referenced) any level-0
902  * data blocks in the free range, so that any future readers will find
903  * empty blocks.
904  *
905  * This is a no-op if the dataset is in the middle of an incremental
906  * receive; see comment below for details.
907  */
908 void
909 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
910     dmu_tx_t *tx)
911 {
912 	dmu_buf_impl_t db_search;
913 	dmu_buf_impl_t *db, *db_next;
914 	uint64_t txg = tx->tx_txg;
915 	avl_index_t where;
916 
917 	if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
918 		end_blkid = dn->dn_maxblkid;
919 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
920 
921 	db_search.db_level = 0;
922 	db_search.db_blkid = start_blkid;
923 	db_search.db_state = DB_SEARCH;
924 
925 	mutex_enter(&dn->dn_dbufs_mtx);
926 	if (start_blkid >= dn->dn_unlisted_l0_blkid) {
927 		/* There can't be any dbufs in this range; no need to search. */
928 #ifdef DEBUG
929 		db = avl_find(&dn->dn_dbufs, &db_search, &where);
930 		ASSERT3P(db, ==, NULL);
931 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
932 		ASSERT(db == NULL || db->db_level > 0);
933 #endif
934 		mutex_exit(&dn->dn_dbufs_mtx);
935 		return;
936 	} else if (dmu_objset_is_receiving(dn->dn_objset)) {
937 		/*
938 		 * If we are receiving, we expect there to be no dbufs in
939 		 * the range to be freed, because receive modifies each
940 		 * block at most once, and in offset order.  If this is
941 		 * not the case, it can lead to performance problems,
942 		 * so note that we unexpectedly took the slow path.
943 		 */
944 		atomic_inc_64(&zfs_free_range_recv_miss);
945 	}
946 
947 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
948 	ASSERT3P(db, ==, NULL);
949 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
950 
951 	for (; db != NULL; db = db_next) {
952 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
953 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
954 
955 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
956 			break;
957 		}
958 		ASSERT3U(db->db_blkid, >=, start_blkid);
959 
960 		/* found a level 0 buffer in the range */
961 		mutex_enter(&db->db_mtx);
962 		if (dbuf_undirty(db, tx)) {
963 			/* mutex has been dropped and dbuf destroyed */
964 			continue;
965 		}
966 
967 		if (db->db_state == DB_UNCACHED ||
968 		    db->db_state == DB_NOFILL ||
969 		    db->db_state == DB_EVICTING) {
970 			ASSERT(db->db.db_data == NULL);
971 			mutex_exit(&db->db_mtx);
972 			continue;
973 		}
974 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
975 			/* will be handled in dbuf_read_done or dbuf_rele */
976 			db->db_freed_in_flight = TRUE;
977 			mutex_exit(&db->db_mtx);
978 			continue;
979 		}
980 		if (refcount_count(&db->db_holds) == 0) {
981 			ASSERT(db->db_buf);
982 			dbuf_clear(db);
983 			continue;
984 		}
985 		/* The dbuf is referenced */
986 
987 		if (db->db_last_dirty != NULL) {
988 			dbuf_dirty_record_t *dr = db->db_last_dirty;
989 
990 			if (dr->dr_txg == txg) {
991 				/*
992 				 * This buffer is "in-use", re-adjust the file
993 				 * size to reflect that this buffer may
994 				 * contain new data when we sync.
995 				 */
996 				if (db->db_blkid != DMU_SPILL_BLKID &&
997 				    db->db_blkid > dn->dn_maxblkid)
998 					dn->dn_maxblkid = db->db_blkid;
999 				dbuf_unoverride(dr);
1000 			} else {
1001 				/*
1002 				 * This dbuf is not dirty in the open context.
1003 				 * Either uncache it (if its not referenced in
1004 				 * the open context) or reset its contents to
1005 				 * empty.
1006 				 */
1007 				dbuf_fix_old_data(db, txg);
1008 			}
1009 		}
1010 		/* clear the contents if its cached */
1011 		if (db->db_state == DB_CACHED) {
1012 			ASSERT(db->db.db_data != NULL);
1013 			arc_release(db->db_buf, db);
1014 			bzero(db->db.db_data, db->db.db_size);
1015 			arc_buf_freeze(db->db_buf);
1016 		}
1017 
1018 		mutex_exit(&db->db_mtx);
1019 	}
1020 	mutex_exit(&dn->dn_dbufs_mtx);
1021 }
1022 
1023 static int
1024 dbuf_block_freeable(dmu_buf_impl_t *db)
1025 {
1026 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1027 	uint64_t birth_txg = 0;
1028 
1029 	/*
1030 	 * We don't need any locking to protect db_blkptr:
1031 	 * If it's syncing, then db_last_dirty will be set
1032 	 * so we'll ignore db_blkptr.
1033 	 *
1034 	 * This logic ensures that only block births for
1035 	 * filled blocks are considered.
1036 	 */
1037 	ASSERT(MUTEX_HELD(&db->db_mtx));
1038 	if (db->db_last_dirty && (db->db_blkptr == NULL ||
1039 	    !BP_IS_HOLE(db->db_blkptr))) {
1040 		birth_txg = db->db_last_dirty->dr_txg;
1041 	} else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1042 		birth_txg = db->db_blkptr->blk_birth;
1043 	}
1044 
1045 	/*
1046 	 * If this block don't exist or is in a snapshot, it can't be freed.
1047 	 * Don't pass the bp to dsl_dataset_block_freeable() since we
1048 	 * are holding the db_mtx lock and might deadlock if we are
1049 	 * prefetching a dedup-ed block.
1050 	 */
1051 	if (birth_txg != 0)
1052 		return (ds == NULL ||
1053 		    dsl_dataset_block_freeable(ds, NULL, birth_txg));
1054 	else
1055 		return (B_FALSE);
1056 }
1057 
1058 void
1059 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1060 {
1061 	arc_buf_t *buf, *obuf;
1062 	int osize = db->db.db_size;
1063 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1064 	dnode_t *dn;
1065 
1066 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1067 
1068 	DB_DNODE_ENTER(db);
1069 	dn = DB_DNODE(db);
1070 
1071 	/* XXX does *this* func really need the lock? */
1072 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1073 
1074 	/*
1075 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1076 	 * is OK, because there can be no other references to the db
1077 	 * when we are changing its size, so no concurrent DB_FILL can
1078 	 * be happening.
1079 	 */
1080 	/*
1081 	 * XXX we should be doing a dbuf_read, checking the return
1082 	 * value and returning that up to our callers
1083 	 */
1084 	dmu_buf_will_dirty(&db->db, tx);
1085 
1086 	/* create the data buffer for the new block */
1087 	buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1088 
1089 	/* copy old block data to the new block */
1090 	obuf = db->db_buf;
1091 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1092 	/* zero the remainder */
1093 	if (size > osize)
1094 		bzero((uint8_t *)buf->b_data + osize, size - osize);
1095 
1096 	mutex_enter(&db->db_mtx);
1097 	dbuf_set_data(db, buf);
1098 	VERIFY(arc_buf_remove_ref(obuf, db));
1099 	db->db.db_size = size;
1100 
1101 	if (db->db_level == 0) {
1102 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1103 		db->db_last_dirty->dt.dl.dr_data = buf;
1104 	}
1105 	mutex_exit(&db->db_mtx);
1106 
1107 	dnode_willuse_space(dn, size-osize, tx);
1108 	DB_DNODE_EXIT(db);
1109 }
1110 
1111 void
1112 dbuf_release_bp(dmu_buf_impl_t *db)
1113 {
1114 	objset_t *os = db->db_objset;
1115 
1116 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1117 	ASSERT(arc_released(os->os_phys_buf) ||
1118 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
1119 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1120 
1121 	(void) arc_release(db->db_buf, db);
1122 }
1123 
1124 dbuf_dirty_record_t *
1125 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1126 {
1127 	dnode_t *dn;
1128 	objset_t *os;
1129 	dbuf_dirty_record_t **drp, *dr;
1130 	int drop_struct_lock = FALSE;
1131 	boolean_t do_free_accounting = B_FALSE;
1132 	int txgoff = tx->tx_txg & TXG_MASK;
1133 
1134 	ASSERT(tx->tx_txg != 0);
1135 	ASSERT(!refcount_is_zero(&db->db_holds));
1136 	DMU_TX_DIRTY_BUF(tx, db);
1137 
1138 	DB_DNODE_ENTER(db);
1139 	dn = DB_DNODE(db);
1140 	/*
1141 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1142 	 * objects may be dirtied in syncing context, but only if they
1143 	 * were already pre-dirtied in open context.
1144 	 */
1145 	ASSERT(!dmu_tx_is_syncing(tx) ||
1146 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1147 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1148 	    dn->dn_objset->os_dsl_dataset == NULL);
1149 	/*
1150 	 * We make this assert for private objects as well, but after we
1151 	 * check if we're already dirty.  They are allowed to re-dirty
1152 	 * in syncing context.
1153 	 */
1154 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1155 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1156 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1157 
1158 	mutex_enter(&db->db_mtx);
1159 	/*
1160 	 * XXX make this true for indirects too?  The problem is that
1161 	 * transactions created with dmu_tx_create_assigned() from
1162 	 * syncing context don't bother holding ahead.
1163 	 */
1164 	ASSERT(db->db_level != 0 ||
1165 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1166 	    db->db_state == DB_NOFILL);
1167 
1168 	mutex_enter(&dn->dn_mtx);
1169 	/*
1170 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1171 	 * initialize the objset.
1172 	 */
1173 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1174 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1175 		dn->dn_dirtyctx =
1176 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1177 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
1178 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1179 	}
1180 	mutex_exit(&dn->dn_mtx);
1181 
1182 	if (db->db_blkid == DMU_SPILL_BLKID)
1183 		dn->dn_have_spill = B_TRUE;
1184 
1185 	/*
1186 	 * If this buffer is already dirty, we're done.
1187 	 */
1188 	drp = &db->db_last_dirty;
1189 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1190 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1191 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1192 		drp = &dr->dr_next;
1193 	if (dr && dr->dr_txg == tx->tx_txg) {
1194 		DB_DNODE_EXIT(db);
1195 
1196 		if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1197 			/*
1198 			 * If this buffer has already been written out,
1199 			 * we now need to reset its state.
1200 			 */
1201 			dbuf_unoverride(dr);
1202 			if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1203 			    db->db_state != DB_NOFILL)
1204 				arc_buf_thaw(db->db_buf);
1205 		}
1206 		mutex_exit(&db->db_mtx);
1207 		return (dr);
1208 	}
1209 
1210 	/*
1211 	 * Only valid if not already dirty.
1212 	 */
1213 	ASSERT(dn->dn_object == 0 ||
1214 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1215 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1216 
1217 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1218 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1219 	    dn->dn_phys->dn_nlevels > db->db_level ||
1220 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1221 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1222 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1223 
1224 	/*
1225 	 * We should only be dirtying in syncing context if it's the
1226 	 * mos or we're initializing the os or it's a special object.
1227 	 * However, we are allowed to dirty in syncing context provided
1228 	 * we already dirtied it in open context.  Hence we must make
1229 	 * this assertion only if we're not already dirty.
1230 	 */
1231 	os = dn->dn_objset;
1232 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1233 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1234 	ASSERT(db->db.db_size != 0);
1235 
1236 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1237 
1238 	if (db->db_blkid != DMU_BONUS_BLKID) {
1239 		/*
1240 		 * Update the accounting.
1241 		 * Note: we delay "free accounting" until after we drop
1242 		 * the db_mtx.  This keeps us from grabbing other locks
1243 		 * (and possibly deadlocking) in bp_get_dsize() while
1244 		 * also holding the db_mtx.
1245 		 */
1246 		dnode_willuse_space(dn, db->db.db_size, tx);
1247 		do_free_accounting = dbuf_block_freeable(db);
1248 	}
1249 
1250 	/*
1251 	 * If this buffer is dirty in an old transaction group we need
1252 	 * to make a copy of it so that the changes we make in this
1253 	 * transaction group won't leak out when we sync the older txg.
1254 	 */
1255 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1256 	if (db->db_level == 0) {
1257 		void *data_old = db->db_buf;
1258 
1259 		if (db->db_state != DB_NOFILL) {
1260 			if (db->db_blkid == DMU_BONUS_BLKID) {
1261 				dbuf_fix_old_data(db, tx->tx_txg);
1262 				data_old = db->db.db_data;
1263 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1264 				/*
1265 				 * Release the data buffer from the cache so
1266 				 * that we can modify it without impacting
1267 				 * possible other users of this cached data
1268 				 * block.  Note that indirect blocks and
1269 				 * private objects are not released until the
1270 				 * syncing state (since they are only modified
1271 				 * then).
1272 				 */
1273 				arc_release(db->db_buf, db);
1274 				dbuf_fix_old_data(db, tx->tx_txg);
1275 				data_old = db->db_buf;
1276 			}
1277 			ASSERT(data_old != NULL);
1278 		}
1279 		dr->dt.dl.dr_data = data_old;
1280 	} else {
1281 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1282 		list_create(&dr->dt.di.dr_children,
1283 		    sizeof (dbuf_dirty_record_t),
1284 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1285 	}
1286 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1287 		dr->dr_accounted = db->db.db_size;
1288 	dr->dr_dbuf = db;
1289 	dr->dr_txg = tx->tx_txg;
1290 	dr->dr_next = *drp;
1291 	*drp = dr;
1292 
1293 	/*
1294 	 * We could have been freed_in_flight between the dbuf_noread
1295 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1296 	 * happened after the free.
1297 	 */
1298 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1299 	    db->db_blkid != DMU_SPILL_BLKID) {
1300 		mutex_enter(&dn->dn_mtx);
1301 		if (dn->dn_free_ranges[txgoff] != NULL) {
1302 			range_tree_clear(dn->dn_free_ranges[txgoff],
1303 			    db->db_blkid, 1);
1304 		}
1305 		mutex_exit(&dn->dn_mtx);
1306 		db->db_freed_in_flight = FALSE;
1307 	}
1308 
1309 	/*
1310 	 * This buffer is now part of this txg
1311 	 */
1312 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1313 	db->db_dirtycnt += 1;
1314 	ASSERT3U(db->db_dirtycnt, <=, 3);
1315 
1316 	mutex_exit(&db->db_mtx);
1317 
1318 	if (db->db_blkid == DMU_BONUS_BLKID ||
1319 	    db->db_blkid == DMU_SPILL_BLKID) {
1320 		mutex_enter(&dn->dn_mtx);
1321 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1322 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1323 		mutex_exit(&dn->dn_mtx);
1324 		dnode_setdirty(dn, tx);
1325 		DB_DNODE_EXIT(db);
1326 		return (dr);
1327 	} else if (do_free_accounting) {
1328 		blkptr_t *bp = db->db_blkptr;
1329 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1330 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1331 		/*
1332 		 * This is only a guess -- if the dbuf is dirty
1333 		 * in a previous txg, we don't know how much
1334 		 * space it will use on disk yet.  We should
1335 		 * really have the struct_rwlock to access
1336 		 * db_blkptr, but since this is just a guess,
1337 		 * it's OK if we get an odd answer.
1338 		 */
1339 		ddt_prefetch(os->os_spa, bp);
1340 		dnode_willuse_space(dn, -willfree, tx);
1341 	}
1342 
1343 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1344 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1345 		drop_struct_lock = TRUE;
1346 	}
1347 
1348 	if (db->db_level == 0) {
1349 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1350 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1351 	}
1352 
1353 	if (db->db_level+1 < dn->dn_nlevels) {
1354 		dmu_buf_impl_t *parent = db->db_parent;
1355 		dbuf_dirty_record_t *di;
1356 		int parent_held = FALSE;
1357 
1358 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1359 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1360 
1361 			parent = dbuf_hold_level(dn, db->db_level+1,
1362 			    db->db_blkid >> epbs, FTAG);
1363 			ASSERT(parent != NULL);
1364 			parent_held = TRUE;
1365 		}
1366 		if (drop_struct_lock)
1367 			rw_exit(&dn->dn_struct_rwlock);
1368 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1369 		di = dbuf_dirty(parent, tx);
1370 		if (parent_held)
1371 			dbuf_rele(parent, FTAG);
1372 
1373 		mutex_enter(&db->db_mtx);
1374 		/*
1375 		 * Since we've dropped the mutex, it's possible that
1376 		 * dbuf_undirty() might have changed this out from under us.
1377 		 */
1378 		if (db->db_last_dirty == dr ||
1379 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1380 			mutex_enter(&di->dt.di.dr_mtx);
1381 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1382 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1383 			list_insert_tail(&di->dt.di.dr_children, dr);
1384 			mutex_exit(&di->dt.di.dr_mtx);
1385 			dr->dr_parent = di;
1386 		}
1387 		mutex_exit(&db->db_mtx);
1388 	} else {
1389 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1390 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1391 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1392 		mutex_enter(&dn->dn_mtx);
1393 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1394 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1395 		mutex_exit(&dn->dn_mtx);
1396 		if (drop_struct_lock)
1397 			rw_exit(&dn->dn_struct_rwlock);
1398 	}
1399 
1400 	dnode_setdirty(dn, tx);
1401 	DB_DNODE_EXIT(db);
1402 	return (dr);
1403 }
1404 
1405 /*
1406  * Undirty a buffer in the transaction group referenced by the given
1407  * transaction.  Return whether this evicted the dbuf.
1408  */
1409 static boolean_t
1410 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1411 {
1412 	dnode_t *dn;
1413 	uint64_t txg = tx->tx_txg;
1414 	dbuf_dirty_record_t *dr, **drp;
1415 
1416 	ASSERT(txg != 0);
1417 
1418 	/*
1419 	 * Due to our use of dn_nlevels below, this can only be called
1420 	 * in open context, unless we are operating on the MOS.
1421 	 * From syncing context, dn_nlevels may be different from the
1422 	 * dn_nlevels used when dbuf was dirtied.
1423 	 */
1424 	ASSERT(db->db_objset ==
1425 	    dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1426 	    txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1427 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1428 	ASSERT0(db->db_level);
1429 	ASSERT(MUTEX_HELD(&db->db_mtx));
1430 
1431 	/*
1432 	 * If this buffer is not dirty, we're done.
1433 	 */
1434 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1435 		if (dr->dr_txg <= txg)
1436 			break;
1437 	if (dr == NULL || dr->dr_txg < txg)
1438 		return (B_FALSE);
1439 	ASSERT(dr->dr_txg == txg);
1440 	ASSERT(dr->dr_dbuf == db);
1441 
1442 	DB_DNODE_ENTER(db);
1443 	dn = DB_DNODE(db);
1444 
1445 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1446 
1447 	ASSERT(db->db.db_size != 0);
1448 
1449 	dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1450 	    dr->dr_accounted, txg);
1451 
1452 	*drp = dr->dr_next;
1453 
1454 	/*
1455 	 * Note that there are three places in dbuf_dirty()
1456 	 * where this dirty record may be put on a list.
1457 	 * Make sure to do a list_remove corresponding to
1458 	 * every one of those list_insert calls.
1459 	 */
1460 	if (dr->dr_parent) {
1461 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1462 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1463 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1464 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1465 	    db->db_level + 1 == dn->dn_nlevels) {
1466 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1467 		mutex_enter(&dn->dn_mtx);
1468 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1469 		mutex_exit(&dn->dn_mtx);
1470 	}
1471 	DB_DNODE_EXIT(db);
1472 
1473 	if (db->db_state != DB_NOFILL) {
1474 		dbuf_unoverride(dr);
1475 
1476 		ASSERT(db->db_buf != NULL);
1477 		ASSERT(dr->dt.dl.dr_data != NULL);
1478 		if (dr->dt.dl.dr_data != db->db_buf)
1479 			VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1480 	}
1481 
1482 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1483 
1484 	ASSERT(db->db_dirtycnt > 0);
1485 	db->db_dirtycnt -= 1;
1486 
1487 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1488 		arc_buf_t *buf = db->db_buf;
1489 
1490 		ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1491 		dbuf_clear_data(db);
1492 		VERIFY(arc_buf_remove_ref(buf, db));
1493 		dbuf_evict(db);
1494 		return (B_TRUE);
1495 	}
1496 
1497 	return (B_FALSE);
1498 }
1499 
1500 void
1501 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1502 {
1503 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1504 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1505 
1506 	ASSERT(tx->tx_txg != 0);
1507 	ASSERT(!refcount_is_zero(&db->db_holds));
1508 
1509 	DB_DNODE_ENTER(db);
1510 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1511 		rf |= DB_RF_HAVESTRUCT;
1512 	DB_DNODE_EXIT(db);
1513 	(void) dbuf_read(db, NULL, rf);
1514 	(void) dbuf_dirty(db, tx);
1515 }
1516 
1517 void
1518 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1519 {
1520 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1521 
1522 	db->db_state = DB_NOFILL;
1523 
1524 	dmu_buf_will_fill(db_fake, tx);
1525 }
1526 
1527 void
1528 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1529 {
1530 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1531 
1532 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1533 	ASSERT(tx->tx_txg != 0);
1534 	ASSERT(db->db_level == 0);
1535 	ASSERT(!refcount_is_zero(&db->db_holds));
1536 
1537 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1538 	    dmu_tx_private_ok(tx));
1539 
1540 	dbuf_noread(db);
1541 	(void) dbuf_dirty(db, tx);
1542 }
1543 
1544 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1545 /* ARGSUSED */
1546 void
1547 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1548 {
1549 	mutex_enter(&db->db_mtx);
1550 	DBUF_VERIFY(db);
1551 
1552 	if (db->db_state == DB_FILL) {
1553 		if (db->db_level == 0 && db->db_freed_in_flight) {
1554 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1555 			/* we were freed while filling */
1556 			/* XXX dbuf_undirty? */
1557 			bzero(db->db.db_data, db->db.db_size);
1558 			db->db_freed_in_flight = FALSE;
1559 		}
1560 		db->db_state = DB_CACHED;
1561 		cv_broadcast(&db->db_changed);
1562 	}
1563 	mutex_exit(&db->db_mtx);
1564 }
1565 
1566 void
1567 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1568     bp_embedded_type_t etype, enum zio_compress comp,
1569     int uncompressed_size, int compressed_size, int byteorder,
1570     dmu_tx_t *tx)
1571 {
1572 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1573 	struct dirty_leaf *dl;
1574 	dmu_object_type_t type;
1575 
1576 	DB_DNODE_ENTER(db);
1577 	type = DB_DNODE(db)->dn_type;
1578 	DB_DNODE_EXIT(db);
1579 
1580 	ASSERT0(db->db_level);
1581 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1582 
1583 	dmu_buf_will_not_fill(dbuf, tx);
1584 
1585 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1586 	dl = &db->db_last_dirty->dt.dl;
1587 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
1588 	    data, comp, uncompressed_size, compressed_size);
1589 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1590 	BP_SET_TYPE(&dl->dr_overridden_by, type);
1591 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1592 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1593 
1594 	dl->dr_override_state = DR_OVERRIDDEN;
1595 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1596 }
1597 
1598 /*
1599  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1600  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1601  */
1602 void
1603 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1604 {
1605 	ASSERT(!refcount_is_zero(&db->db_holds));
1606 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1607 	ASSERT(db->db_level == 0);
1608 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1609 	ASSERT(buf != NULL);
1610 	ASSERT(arc_buf_size(buf) == db->db.db_size);
1611 	ASSERT(tx->tx_txg != 0);
1612 
1613 	arc_return_buf(buf, db);
1614 	ASSERT(arc_released(buf));
1615 
1616 	mutex_enter(&db->db_mtx);
1617 
1618 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1619 		cv_wait(&db->db_changed, &db->db_mtx);
1620 
1621 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1622 
1623 	if (db->db_state == DB_CACHED &&
1624 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1625 		mutex_exit(&db->db_mtx);
1626 		(void) dbuf_dirty(db, tx);
1627 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1628 		VERIFY(arc_buf_remove_ref(buf, db));
1629 		xuio_stat_wbuf_copied();
1630 		return;
1631 	}
1632 
1633 	xuio_stat_wbuf_nocopy();
1634 	if (db->db_state == DB_CACHED) {
1635 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1636 
1637 		ASSERT(db->db_buf != NULL);
1638 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1639 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1640 			if (!arc_released(db->db_buf)) {
1641 				ASSERT(dr->dt.dl.dr_override_state ==
1642 				    DR_OVERRIDDEN);
1643 				arc_release(db->db_buf, db);
1644 			}
1645 			dr->dt.dl.dr_data = buf;
1646 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1647 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1648 			arc_release(db->db_buf, db);
1649 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1650 		}
1651 		db->db_buf = NULL;
1652 	}
1653 	ASSERT(db->db_buf == NULL);
1654 	dbuf_set_data(db, buf);
1655 	db->db_state = DB_FILL;
1656 	mutex_exit(&db->db_mtx);
1657 	(void) dbuf_dirty(db, tx);
1658 	dmu_buf_fill_done(&db->db, tx);
1659 }
1660 
1661 /*
1662  * "Clear" the contents of this dbuf.  This will mark the dbuf
1663  * EVICTING and clear *most* of its references.  Unfortunately,
1664  * when we are not holding the dn_dbufs_mtx, we can't clear the
1665  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1666  * in this case.  For callers from the DMU we will usually see:
1667  *	dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1668  * For the arc callback, we will usually see:
1669  *	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1670  * Sometimes, though, we will get a mix of these two:
1671  *	DMU: dbuf_clear()->arc_clear_callback()
1672  *	ARC: dbuf_do_evict()->dbuf_destroy()
1673  *
1674  * This routine will dissociate the dbuf from the arc, by calling
1675  * arc_clear_callback(), but will not evict the data from the ARC.
1676  */
1677 void
1678 dbuf_clear(dmu_buf_impl_t *db)
1679 {
1680 	dnode_t *dn;
1681 	dmu_buf_impl_t *parent = db->db_parent;
1682 	dmu_buf_impl_t *dndb;
1683 	boolean_t dbuf_gone = B_FALSE;
1684 
1685 	ASSERT(MUTEX_HELD(&db->db_mtx));
1686 	ASSERT(refcount_is_zero(&db->db_holds));
1687 
1688 	dbuf_evict_user(db);
1689 
1690 	if (db->db_state == DB_CACHED) {
1691 		ASSERT(db->db.db_data != NULL);
1692 		if (db->db_blkid == DMU_BONUS_BLKID) {
1693 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1694 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1695 		}
1696 		db->db.db_data = NULL;
1697 		db->db_state = DB_UNCACHED;
1698 	}
1699 
1700 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1701 	ASSERT(db->db_data_pending == NULL);
1702 
1703 	db->db_state = DB_EVICTING;
1704 	db->db_blkptr = NULL;
1705 
1706 	DB_DNODE_ENTER(db);
1707 	dn = DB_DNODE(db);
1708 	dndb = dn->dn_dbuf;
1709 	if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1710 		avl_remove(&dn->dn_dbufs, db);
1711 		atomic_dec_32(&dn->dn_dbufs_count);
1712 		membar_producer();
1713 		DB_DNODE_EXIT(db);
1714 		/*
1715 		 * Decrementing the dbuf count means that the hold corresponding
1716 		 * to the removed dbuf is no longer discounted in dnode_move(),
1717 		 * so the dnode cannot be moved until after we release the hold.
1718 		 * The membar_producer() ensures visibility of the decremented
1719 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1720 		 * release any lock.
1721 		 */
1722 		dnode_rele(dn, db);
1723 		db->db_dnode_handle = NULL;
1724 	} else {
1725 		DB_DNODE_EXIT(db);
1726 	}
1727 
1728 	if (db->db_buf)
1729 		dbuf_gone = arc_clear_callback(db->db_buf);
1730 
1731 	if (!dbuf_gone)
1732 		mutex_exit(&db->db_mtx);
1733 
1734 	/*
1735 	 * If this dbuf is referenced from an indirect dbuf,
1736 	 * decrement the ref count on the indirect dbuf.
1737 	 */
1738 	if (parent && parent != dndb)
1739 		dbuf_rele(parent, db);
1740 }
1741 
1742 /*
1743  * Note: While bpp will always be updated if the function returns success,
1744  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1745  * this happens when the dnode is the meta-dnode, or a userused or groupused
1746  * object.
1747  */
1748 static int
1749 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1750     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1751 {
1752 	int nlevels, epbs;
1753 
1754 	*parentp = NULL;
1755 	*bpp = NULL;
1756 
1757 	ASSERT(blkid != DMU_BONUS_BLKID);
1758 
1759 	if (blkid == DMU_SPILL_BLKID) {
1760 		mutex_enter(&dn->dn_mtx);
1761 		if (dn->dn_have_spill &&
1762 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1763 			*bpp = &dn->dn_phys->dn_spill;
1764 		else
1765 			*bpp = NULL;
1766 		dbuf_add_ref(dn->dn_dbuf, NULL);
1767 		*parentp = dn->dn_dbuf;
1768 		mutex_exit(&dn->dn_mtx);
1769 		return (0);
1770 	}
1771 
1772 	if (dn->dn_phys->dn_nlevels == 0)
1773 		nlevels = 1;
1774 	else
1775 		nlevels = dn->dn_phys->dn_nlevels;
1776 
1777 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1778 
1779 	ASSERT3U(level * epbs, <, 64);
1780 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1781 	if (level >= nlevels ||
1782 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1783 		/* the buffer has no parent yet */
1784 		return (SET_ERROR(ENOENT));
1785 	} else if (level < nlevels-1) {
1786 		/* this block is referenced from an indirect block */
1787 		int err = dbuf_hold_impl(dn, level+1,
1788 		    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
1789 		if (err)
1790 			return (err);
1791 		err = dbuf_read(*parentp, NULL,
1792 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1793 		if (err) {
1794 			dbuf_rele(*parentp, NULL);
1795 			*parentp = NULL;
1796 			return (err);
1797 		}
1798 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1799 		    (blkid & ((1ULL << epbs) - 1));
1800 		return (0);
1801 	} else {
1802 		/* the block is referenced from the dnode */
1803 		ASSERT3U(level, ==, nlevels-1);
1804 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1805 		    blkid < dn->dn_phys->dn_nblkptr);
1806 		if (dn->dn_dbuf) {
1807 			dbuf_add_ref(dn->dn_dbuf, NULL);
1808 			*parentp = dn->dn_dbuf;
1809 		}
1810 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1811 		return (0);
1812 	}
1813 }
1814 
1815 static dmu_buf_impl_t *
1816 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1817     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1818 {
1819 	objset_t *os = dn->dn_objset;
1820 	dmu_buf_impl_t *db, *odb;
1821 
1822 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1823 	ASSERT(dn->dn_type != DMU_OT_NONE);
1824 
1825 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1826 
1827 	db->db_objset = os;
1828 	db->db.db_object = dn->dn_object;
1829 	db->db_level = level;
1830 	db->db_blkid = blkid;
1831 	db->db_last_dirty = NULL;
1832 	db->db_dirtycnt = 0;
1833 	db->db_dnode_handle = dn->dn_handle;
1834 	db->db_parent = parent;
1835 	db->db_blkptr = blkptr;
1836 
1837 	db->db_user = NULL;
1838 	db->db_immediate_evict = 0;
1839 	db->db_freed_in_flight = 0;
1840 
1841 	if (blkid == DMU_BONUS_BLKID) {
1842 		ASSERT3P(parent, ==, dn->dn_dbuf);
1843 		db->db.db_size = DN_MAX_BONUSLEN -
1844 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1845 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1846 		db->db.db_offset = DMU_BONUS_BLKID;
1847 		db->db_state = DB_UNCACHED;
1848 		/* the bonus dbuf is not placed in the hash table */
1849 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1850 		return (db);
1851 	} else if (blkid == DMU_SPILL_BLKID) {
1852 		db->db.db_size = (blkptr != NULL) ?
1853 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1854 		db->db.db_offset = 0;
1855 	} else {
1856 		int blocksize =
1857 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1858 		db->db.db_size = blocksize;
1859 		db->db.db_offset = db->db_blkid * blocksize;
1860 	}
1861 
1862 	/*
1863 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1864 	 * in the hash table *and* added to the dbufs list.
1865 	 * This prevents a possible deadlock with someone
1866 	 * trying to look up this dbuf before its added to the
1867 	 * dn_dbufs list.
1868 	 */
1869 	mutex_enter(&dn->dn_dbufs_mtx);
1870 	db->db_state = DB_EVICTING;
1871 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1872 		/* someone else inserted it first */
1873 		kmem_cache_free(dbuf_cache, db);
1874 		mutex_exit(&dn->dn_dbufs_mtx);
1875 		return (odb);
1876 	}
1877 	avl_add(&dn->dn_dbufs, db);
1878 	if (db->db_level == 0 && db->db_blkid >=
1879 	    dn->dn_unlisted_l0_blkid)
1880 		dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1881 	db->db_state = DB_UNCACHED;
1882 	mutex_exit(&dn->dn_dbufs_mtx);
1883 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1884 
1885 	if (parent && parent != dn->dn_dbuf)
1886 		dbuf_add_ref(parent, db);
1887 
1888 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1889 	    refcount_count(&dn->dn_holds) > 0);
1890 	(void) refcount_add(&dn->dn_holds, db);
1891 	atomic_inc_32(&dn->dn_dbufs_count);
1892 
1893 	dprintf_dbuf(db, "db=%p\n", db);
1894 
1895 	return (db);
1896 }
1897 
1898 static int
1899 dbuf_do_evict(void *private)
1900 {
1901 	dmu_buf_impl_t *db = private;
1902 
1903 	if (!MUTEX_HELD(&db->db_mtx))
1904 		mutex_enter(&db->db_mtx);
1905 
1906 	ASSERT(refcount_is_zero(&db->db_holds));
1907 
1908 	if (db->db_state != DB_EVICTING) {
1909 		ASSERT(db->db_state == DB_CACHED);
1910 		DBUF_VERIFY(db);
1911 		db->db_buf = NULL;
1912 		dbuf_evict(db);
1913 	} else {
1914 		mutex_exit(&db->db_mtx);
1915 		dbuf_destroy(db);
1916 	}
1917 	return (0);
1918 }
1919 
1920 static void
1921 dbuf_destroy(dmu_buf_impl_t *db)
1922 {
1923 	ASSERT(refcount_is_zero(&db->db_holds));
1924 
1925 	if (db->db_blkid != DMU_BONUS_BLKID) {
1926 		/*
1927 		 * If this dbuf is still on the dn_dbufs list,
1928 		 * remove it from that list.
1929 		 */
1930 		if (db->db_dnode_handle != NULL) {
1931 			dnode_t *dn;
1932 
1933 			DB_DNODE_ENTER(db);
1934 			dn = DB_DNODE(db);
1935 			mutex_enter(&dn->dn_dbufs_mtx);
1936 			avl_remove(&dn->dn_dbufs, db);
1937 			atomic_dec_32(&dn->dn_dbufs_count);
1938 			mutex_exit(&dn->dn_dbufs_mtx);
1939 			DB_DNODE_EXIT(db);
1940 			/*
1941 			 * Decrementing the dbuf count means that the hold
1942 			 * corresponding to the removed dbuf is no longer
1943 			 * discounted in dnode_move(), so the dnode cannot be
1944 			 * moved until after we release the hold.
1945 			 */
1946 			dnode_rele(dn, db);
1947 			db->db_dnode_handle = NULL;
1948 		}
1949 		dbuf_hash_remove(db);
1950 	}
1951 	db->db_parent = NULL;
1952 	db->db_buf = NULL;
1953 
1954 	ASSERT(db->db.db_data == NULL);
1955 	ASSERT(db->db_hash_next == NULL);
1956 	ASSERT(db->db_blkptr == NULL);
1957 	ASSERT(db->db_data_pending == NULL);
1958 
1959 	kmem_cache_free(dbuf_cache, db);
1960 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1961 }
1962 
1963 typedef struct dbuf_prefetch_arg {
1964 	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
1965 	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
1966 	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
1967 	int dpa_curlevel; /* The current level that we're reading */
1968 	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
1969 	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
1970 	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
1971 } dbuf_prefetch_arg_t;
1972 
1973 /*
1974  * Actually issue the prefetch read for the block given.
1975  */
1976 static void
1977 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
1978 {
1979 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
1980 		return;
1981 
1982 	arc_flags_t aflags =
1983 	    dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
1984 
1985 	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
1986 	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
1987 	ASSERT(dpa->dpa_zio != NULL);
1988 	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
1989 	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1990 	    &aflags, &dpa->dpa_zb);
1991 }
1992 
1993 /*
1994  * Called when an indirect block above our prefetch target is read in.  This
1995  * will either read in the next indirect block down the tree or issue the actual
1996  * prefetch if the next block down is our target.
1997  */
1998 static void
1999 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2000 {
2001 	dbuf_prefetch_arg_t *dpa = private;
2002 
2003 	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2004 	ASSERT3S(dpa->dpa_curlevel, >, 0);
2005 	if (zio != NULL) {
2006 		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2007 		ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2008 		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2009 	}
2010 
2011 	dpa->dpa_curlevel--;
2012 
2013 	uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2014 	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2015 	blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2016 	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2017 	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2018 		kmem_free(dpa, sizeof (*dpa));
2019 	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2020 		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2021 		dbuf_issue_final_prefetch(dpa, bp);
2022 		kmem_free(dpa, sizeof (*dpa));
2023 	} else {
2024 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2025 		zbookmark_phys_t zb;
2026 
2027 		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2028 
2029 		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2030 		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2031 
2032 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2033 		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2034 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2035 		    &iter_aflags, &zb);
2036 	}
2037 	(void) arc_buf_remove_ref(abuf, private);
2038 }
2039 
2040 /*
2041  * Issue prefetch reads for the given block on the given level.  If the indirect
2042  * blocks above that block are not in memory, we will read them in
2043  * asynchronously.  As a result, this call never blocks waiting for a read to
2044  * complete.
2045  */
2046 void
2047 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2048     arc_flags_t aflags)
2049 {
2050 	blkptr_t bp;
2051 	int epbs, nlevels, curlevel;
2052 	uint64_t curblkid;
2053 
2054 	ASSERT(blkid != DMU_BONUS_BLKID);
2055 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2056 
2057 	if (dnode_block_freed(dn, blkid))
2058 		return;
2059 
2060 	/*
2061 	 * This dnode hasn't been written to disk yet, so there's nothing to
2062 	 * prefetch.
2063 	 */
2064 	nlevels = dn->dn_phys->dn_nlevels;
2065 	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2066 		return;
2067 
2068 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2069 	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2070 		return;
2071 
2072 	dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2073 	    level, blkid);
2074 	if (db != NULL) {
2075 		mutex_exit(&db->db_mtx);
2076 		/*
2077 		 * This dbuf already exists.  It is either CACHED, or
2078 		 * (we assume) about to be read or filled.
2079 		 */
2080 		return;
2081 	}
2082 
2083 	/*
2084 	 * Find the closest ancestor (indirect block) of the target block
2085 	 * that is present in the cache.  In this indirect block, we will
2086 	 * find the bp that is at curlevel, curblkid.
2087 	 */
2088 	curlevel = level;
2089 	curblkid = blkid;
2090 	while (curlevel < nlevels - 1) {
2091 		int parent_level = curlevel + 1;
2092 		uint64_t parent_blkid = curblkid >> epbs;
2093 		dmu_buf_impl_t *db;
2094 
2095 		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2096 		    FALSE, TRUE, FTAG, &db) == 0) {
2097 			blkptr_t *bpp = db->db_buf->b_data;
2098 			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2099 			dbuf_rele(db, FTAG);
2100 			break;
2101 		}
2102 
2103 		curlevel = parent_level;
2104 		curblkid = parent_blkid;
2105 	}
2106 
2107 	if (curlevel == nlevels - 1) {
2108 		/* No cached indirect blocks found. */
2109 		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2110 		bp = dn->dn_phys->dn_blkptr[curblkid];
2111 	}
2112 	if (BP_IS_HOLE(&bp))
2113 		return;
2114 
2115 	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2116 
2117 	zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2118 	    ZIO_FLAG_CANFAIL);
2119 
2120 	dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2121 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2122 	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2123 	    dn->dn_object, level, blkid);
2124 	dpa->dpa_curlevel = curlevel;
2125 	dpa->dpa_prio = prio;
2126 	dpa->dpa_aflags = aflags;
2127 	dpa->dpa_spa = dn->dn_objset->os_spa;
2128 	dpa->dpa_epbs = epbs;
2129 	dpa->dpa_zio = pio;
2130 
2131 	/*
2132 	 * If we have the indirect just above us, no need to do the asynchronous
2133 	 * prefetch chain; we'll just run the last step ourselves.  If we're at
2134 	 * a higher level, though, we want to issue the prefetches for all the
2135 	 * indirect blocks asynchronously, so we can go on with whatever we were
2136 	 * doing.
2137 	 */
2138 	if (curlevel == level) {
2139 		ASSERT3U(curblkid, ==, blkid);
2140 		dbuf_issue_final_prefetch(dpa, &bp);
2141 		kmem_free(dpa, sizeof (*dpa));
2142 	} else {
2143 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2144 		zbookmark_phys_t zb;
2145 
2146 		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2147 		    dn->dn_object, curlevel, curblkid);
2148 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2149 		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
2150 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2151 		    &iter_aflags, &zb);
2152 	}
2153 	/*
2154 	 * We use pio here instead of dpa_zio since it's possible that
2155 	 * dpa may have already been freed.
2156 	 */
2157 	zio_nowait(pio);
2158 }
2159 
2160 /*
2161  * Returns with db_holds incremented, and db_mtx not held.
2162  * Note: dn_struct_rwlock must be held.
2163  */
2164 int
2165 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2166     boolean_t fail_sparse, boolean_t fail_uncached,
2167     void *tag, dmu_buf_impl_t **dbp)
2168 {
2169 	dmu_buf_impl_t *db, *parent = NULL;
2170 
2171 	ASSERT(blkid != DMU_BONUS_BLKID);
2172 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2173 	ASSERT3U(dn->dn_nlevels, >, level);
2174 
2175 	*dbp = NULL;
2176 top:
2177 	/* dbuf_find() returns with db_mtx held */
2178 	db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2179 
2180 	if (db == NULL) {
2181 		blkptr_t *bp = NULL;
2182 		int err;
2183 
2184 		if (fail_uncached)
2185 			return (SET_ERROR(ENOENT));
2186 
2187 		ASSERT3P(parent, ==, NULL);
2188 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2189 		if (fail_sparse) {
2190 			if (err == 0 && bp && BP_IS_HOLE(bp))
2191 				err = SET_ERROR(ENOENT);
2192 			if (err) {
2193 				if (parent)
2194 					dbuf_rele(parent, NULL);
2195 				return (err);
2196 			}
2197 		}
2198 		if (err && err != ENOENT)
2199 			return (err);
2200 		db = dbuf_create(dn, level, blkid, parent, bp);
2201 	}
2202 
2203 	if (fail_uncached && db->db_state != DB_CACHED) {
2204 		mutex_exit(&db->db_mtx);
2205 		return (SET_ERROR(ENOENT));
2206 	}
2207 
2208 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
2209 		arc_buf_add_ref(db->db_buf, db);
2210 		if (db->db_buf->b_data == NULL) {
2211 			dbuf_clear(db);
2212 			if (parent) {
2213 				dbuf_rele(parent, NULL);
2214 				parent = NULL;
2215 			}
2216 			goto top;
2217 		}
2218 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2219 	}
2220 
2221 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2222 
2223 	/*
2224 	 * If this buffer is currently syncing out, and we are are
2225 	 * still referencing it from db_data, we need to make a copy
2226 	 * of it in case we decide we want to dirty it again in this txg.
2227 	 */
2228 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2229 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2230 	    db->db_state == DB_CACHED && db->db_data_pending) {
2231 		dbuf_dirty_record_t *dr = db->db_data_pending;
2232 
2233 		if (dr->dt.dl.dr_data == db->db_buf) {
2234 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2235 
2236 			dbuf_set_data(db,
2237 			    arc_buf_alloc(dn->dn_objset->os_spa,
2238 			    db->db.db_size, db, type));
2239 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2240 			    db->db.db_size);
2241 		}
2242 	}
2243 
2244 	(void) refcount_add(&db->db_holds, tag);
2245 	DBUF_VERIFY(db);
2246 	mutex_exit(&db->db_mtx);
2247 
2248 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
2249 	if (parent)
2250 		dbuf_rele(parent, NULL);
2251 
2252 	ASSERT3P(DB_DNODE(db), ==, dn);
2253 	ASSERT3U(db->db_blkid, ==, blkid);
2254 	ASSERT3U(db->db_level, ==, level);
2255 	*dbp = db;
2256 
2257 	return (0);
2258 }
2259 
2260 dmu_buf_impl_t *
2261 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2262 {
2263 	return (dbuf_hold_level(dn, 0, blkid, tag));
2264 }
2265 
2266 dmu_buf_impl_t *
2267 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2268 {
2269 	dmu_buf_impl_t *db;
2270 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2271 	return (err ? NULL : db);
2272 }
2273 
2274 void
2275 dbuf_create_bonus(dnode_t *dn)
2276 {
2277 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2278 
2279 	ASSERT(dn->dn_bonus == NULL);
2280 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2281 }
2282 
2283 int
2284 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2285 {
2286 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2287 	dnode_t *dn;
2288 
2289 	if (db->db_blkid != DMU_SPILL_BLKID)
2290 		return (SET_ERROR(ENOTSUP));
2291 	if (blksz == 0)
2292 		blksz = SPA_MINBLOCKSIZE;
2293 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2294 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2295 
2296 	DB_DNODE_ENTER(db);
2297 	dn = DB_DNODE(db);
2298 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2299 	dbuf_new_size(db, blksz, tx);
2300 	rw_exit(&dn->dn_struct_rwlock);
2301 	DB_DNODE_EXIT(db);
2302 
2303 	return (0);
2304 }
2305 
2306 void
2307 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2308 {
2309 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2310 }
2311 
2312 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2313 void
2314 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2315 {
2316 	int64_t holds = refcount_add(&db->db_holds, tag);
2317 	ASSERT(holds > 1);
2318 }
2319 
2320 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2321 boolean_t
2322 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2323     void *tag)
2324 {
2325 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2326 	dmu_buf_impl_t *found_db;
2327 	boolean_t result = B_FALSE;
2328 
2329 	if (db->db_blkid == DMU_BONUS_BLKID)
2330 		found_db = dbuf_find_bonus(os, obj);
2331 	else
2332 		found_db = dbuf_find(os, obj, 0, blkid);
2333 
2334 	if (found_db != NULL) {
2335 		if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2336 			(void) refcount_add(&db->db_holds, tag);
2337 			result = B_TRUE;
2338 		}
2339 		mutex_exit(&db->db_mtx);
2340 	}
2341 	return (result);
2342 }
2343 
2344 /*
2345  * If you call dbuf_rele() you had better not be referencing the dnode handle
2346  * unless you have some other direct or indirect hold on the dnode. (An indirect
2347  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2348  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2349  * dnode's parent dbuf evicting its dnode handles.
2350  */
2351 void
2352 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2353 {
2354 	mutex_enter(&db->db_mtx);
2355 	dbuf_rele_and_unlock(db, tag);
2356 }
2357 
2358 void
2359 dmu_buf_rele(dmu_buf_t *db, void *tag)
2360 {
2361 	dbuf_rele((dmu_buf_impl_t *)db, tag);
2362 }
2363 
2364 /*
2365  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2366  * db_dirtycnt and db_holds to be updated atomically.
2367  */
2368 void
2369 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2370 {
2371 	int64_t holds;
2372 
2373 	ASSERT(MUTEX_HELD(&db->db_mtx));
2374 	DBUF_VERIFY(db);
2375 
2376 	/*
2377 	 * Remove the reference to the dbuf before removing its hold on the
2378 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2379 	 * buffer has a corresponding dnode hold.
2380 	 */
2381 	holds = refcount_remove(&db->db_holds, tag);
2382 	ASSERT(holds >= 0);
2383 
2384 	/*
2385 	 * We can't freeze indirects if there is a possibility that they
2386 	 * may be modified in the current syncing context.
2387 	 */
2388 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2389 		arc_buf_freeze(db->db_buf);
2390 
2391 	if (holds == db->db_dirtycnt &&
2392 	    db->db_level == 0 && db->db_immediate_evict)
2393 		dbuf_evict_user(db);
2394 
2395 	if (holds == 0) {
2396 		if (db->db_blkid == DMU_BONUS_BLKID) {
2397 			dnode_t *dn;
2398 
2399 			/*
2400 			 * If the dnode moves here, we cannot cross this
2401 			 * barrier until the move completes.
2402 			 */
2403 			DB_DNODE_ENTER(db);
2404 
2405 			dn = DB_DNODE(db);
2406 			atomic_dec_32(&dn->dn_dbufs_count);
2407 
2408 			/*
2409 			 * Decrementing the dbuf count means that the bonus
2410 			 * buffer's dnode hold is no longer discounted in
2411 			 * dnode_move(). The dnode cannot move until after
2412 			 * the dnode_rele_and_unlock() below.
2413 			 */
2414 			DB_DNODE_EXIT(db);
2415 
2416 			/*
2417 			 * Do not reference db after its lock is dropped.
2418 			 * Another thread may evict it.
2419 			 */
2420 			mutex_exit(&db->db_mtx);
2421 
2422 			/*
2423 			 * If the dnode has been freed, evict the bonus
2424 			 * buffer immediately.	The data in the bonus
2425 			 * buffer is no longer relevant and this prevents
2426 			 * a stale bonus buffer from being associated
2427 			 * with this dnode_t should the dnode_t be reused
2428 			 * prior to being destroyed.
2429 			 */
2430 			mutex_enter(&dn->dn_mtx);
2431 			if (dn->dn_type == DMU_OT_NONE ||
2432 			    dn->dn_free_txg != 0) {
2433 				/*
2434 				 * Drop dn_mtx.  It is a leaf lock and
2435 				 * cannot be held when dnode_evict_bonus()
2436 				 * acquires other locks in order to
2437 				 * perform the eviction.
2438 				 *
2439 				 * Freed dnodes cannot be reused until the
2440 				 * last hold is released.  Since this bonus
2441 				 * buffer has a hold, the dnode will remain
2442 				 * in the free state, even without dn_mtx
2443 				 * held, until the dnode_rele_and_unlock()
2444 				 * below.
2445 				 */
2446 				mutex_exit(&dn->dn_mtx);
2447 				dnode_evict_bonus(dn);
2448 				mutex_enter(&dn->dn_mtx);
2449 			}
2450 			dnode_rele_and_unlock(dn, db);
2451 		} else if (db->db_buf == NULL) {
2452 			/*
2453 			 * This is a special case: we never associated this
2454 			 * dbuf with any data allocated from the ARC.
2455 			 */
2456 			ASSERT(db->db_state == DB_UNCACHED ||
2457 			    db->db_state == DB_NOFILL);
2458 			dbuf_evict(db);
2459 		} else if (arc_released(db->db_buf)) {
2460 			arc_buf_t *buf = db->db_buf;
2461 			/*
2462 			 * This dbuf has anonymous data associated with it.
2463 			 */
2464 			dbuf_clear_data(db);
2465 			VERIFY(arc_buf_remove_ref(buf, db));
2466 			dbuf_evict(db);
2467 		} else {
2468 			VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2469 
2470 			/*
2471 			 * A dbuf will be eligible for eviction if either the
2472 			 * 'primarycache' property is set or a duplicate
2473 			 * copy of this buffer is already cached in the arc.
2474 			 *
2475 			 * In the case of the 'primarycache' a buffer
2476 			 * is considered for eviction if it matches the
2477 			 * criteria set in the property.
2478 			 *
2479 			 * To decide if our buffer is considered a
2480 			 * duplicate, we must call into the arc to determine
2481 			 * if multiple buffers are referencing the same
2482 			 * block on-disk. If so, then we simply evict
2483 			 * ourselves.
2484 			 */
2485 			if (!DBUF_IS_CACHEABLE(db)) {
2486 				if (db->db_blkptr != NULL &&
2487 				    !BP_IS_HOLE(db->db_blkptr) &&
2488 				    !BP_IS_EMBEDDED(db->db_blkptr)) {
2489 					spa_t *spa =
2490 					    dmu_objset_spa(db->db_objset);
2491 					blkptr_t bp = *db->db_blkptr;
2492 					dbuf_clear(db);
2493 					arc_freed(spa, &bp);
2494 				} else {
2495 					dbuf_clear(db);
2496 				}
2497 			} else if (db->db_objset->os_evicting ||
2498 			    arc_buf_eviction_needed(db->db_buf)) {
2499 				dbuf_clear(db);
2500 			} else {
2501 				mutex_exit(&db->db_mtx);
2502 			}
2503 		}
2504 	} else {
2505 		mutex_exit(&db->db_mtx);
2506 	}
2507 }
2508 
2509 #pragma weak dmu_buf_refcount = dbuf_refcount
2510 uint64_t
2511 dbuf_refcount(dmu_buf_impl_t *db)
2512 {
2513 	return (refcount_count(&db->db_holds));
2514 }
2515 
2516 void *
2517 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2518     dmu_buf_user_t *new_user)
2519 {
2520 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2521 
2522 	mutex_enter(&db->db_mtx);
2523 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2524 	if (db->db_user == old_user)
2525 		db->db_user = new_user;
2526 	else
2527 		old_user = db->db_user;
2528 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2529 	mutex_exit(&db->db_mtx);
2530 
2531 	return (old_user);
2532 }
2533 
2534 void *
2535 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2536 {
2537 	return (dmu_buf_replace_user(db_fake, NULL, user));
2538 }
2539 
2540 void *
2541 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2542 {
2543 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2544 
2545 	db->db_immediate_evict = TRUE;
2546 	return (dmu_buf_set_user(db_fake, user));
2547 }
2548 
2549 void *
2550 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2551 {
2552 	return (dmu_buf_replace_user(db_fake, user, NULL));
2553 }
2554 
2555 void *
2556 dmu_buf_get_user(dmu_buf_t *db_fake)
2557 {
2558 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2559 
2560 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2561 	return (db->db_user);
2562 }
2563 
2564 void
2565 dmu_buf_user_evict_wait()
2566 {
2567 	taskq_wait(dbu_evict_taskq);
2568 }
2569 
2570 boolean_t
2571 dmu_buf_freeable(dmu_buf_t *dbuf)
2572 {
2573 	boolean_t res = B_FALSE;
2574 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2575 
2576 	if (db->db_blkptr)
2577 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2578 		    db->db_blkptr, db->db_blkptr->blk_birth);
2579 
2580 	return (res);
2581 }
2582 
2583 blkptr_t *
2584 dmu_buf_get_blkptr(dmu_buf_t *db)
2585 {
2586 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2587 	return (dbi->db_blkptr);
2588 }
2589 
2590 static void
2591 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2592 {
2593 	/* ASSERT(dmu_tx_is_syncing(tx) */
2594 	ASSERT(MUTEX_HELD(&db->db_mtx));
2595 
2596 	if (db->db_blkptr != NULL)
2597 		return;
2598 
2599 	if (db->db_blkid == DMU_SPILL_BLKID) {
2600 		db->db_blkptr = &dn->dn_phys->dn_spill;
2601 		BP_ZERO(db->db_blkptr);
2602 		return;
2603 	}
2604 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2605 		/*
2606 		 * This buffer was allocated at a time when there was
2607 		 * no available blkptrs from the dnode, or it was
2608 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2609 		 */
2610 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2611 		ASSERT(db->db_parent == NULL);
2612 		db->db_parent = dn->dn_dbuf;
2613 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2614 		DBUF_VERIFY(db);
2615 	} else {
2616 		dmu_buf_impl_t *parent = db->db_parent;
2617 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2618 
2619 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2620 		if (parent == NULL) {
2621 			mutex_exit(&db->db_mtx);
2622 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2623 			parent = dbuf_hold_level(dn, db->db_level + 1,
2624 			    db->db_blkid >> epbs, db);
2625 			rw_exit(&dn->dn_struct_rwlock);
2626 			mutex_enter(&db->db_mtx);
2627 			db->db_parent = parent;
2628 		}
2629 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2630 		    (db->db_blkid & ((1ULL << epbs) - 1));
2631 		DBUF_VERIFY(db);
2632 	}
2633 }
2634 
2635 static void
2636 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2637 {
2638 	dmu_buf_impl_t *db = dr->dr_dbuf;
2639 	dnode_t *dn;
2640 	zio_t *zio;
2641 
2642 	ASSERT(dmu_tx_is_syncing(tx));
2643 
2644 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2645 
2646 	mutex_enter(&db->db_mtx);
2647 
2648 	ASSERT(db->db_level > 0);
2649 	DBUF_VERIFY(db);
2650 
2651 	/* Read the block if it hasn't been read yet. */
2652 	if (db->db_buf == NULL) {
2653 		mutex_exit(&db->db_mtx);
2654 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2655 		mutex_enter(&db->db_mtx);
2656 	}
2657 	ASSERT3U(db->db_state, ==, DB_CACHED);
2658 	ASSERT(db->db_buf != NULL);
2659 
2660 	DB_DNODE_ENTER(db);
2661 	dn = DB_DNODE(db);
2662 	/* Indirect block size must match what the dnode thinks it is. */
2663 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2664 	dbuf_check_blkptr(dn, db);
2665 	DB_DNODE_EXIT(db);
2666 
2667 	/* Provide the pending dirty record to child dbufs */
2668 	db->db_data_pending = dr;
2669 
2670 	mutex_exit(&db->db_mtx);
2671 	dbuf_write(dr, db->db_buf, tx);
2672 
2673 	zio = dr->dr_zio;
2674 	mutex_enter(&dr->dt.di.dr_mtx);
2675 	dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
2676 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2677 	mutex_exit(&dr->dt.di.dr_mtx);
2678 	zio_nowait(zio);
2679 }
2680 
2681 static void
2682 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2683 {
2684 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2685 	dmu_buf_impl_t *db = dr->dr_dbuf;
2686 	dnode_t *dn;
2687 	objset_t *os;
2688 	uint64_t txg = tx->tx_txg;
2689 
2690 	ASSERT(dmu_tx_is_syncing(tx));
2691 
2692 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2693 
2694 	mutex_enter(&db->db_mtx);
2695 	/*
2696 	 * To be synced, we must be dirtied.  But we
2697 	 * might have been freed after the dirty.
2698 	 */
2699 	if (db->db_state == DB_UNCACHED) {
2700 		/* This buffer has been freed since it was dirtied */
2701 		ASSERT(db->db.db_data == NULL);
2702 	} else if (db->db_state == DB_FILL) {
2703 		/* This buffer was freed and is now being re-filled */
2704 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2705 	} else {
2706 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2707 	}
2708 	DBUF_VERIFY(db);
2709 
2710 	DB_DNODE_ENTER(db);
2711 	dn = DB_DNODE(db);
2712 
2713 	if (db->db_blkid == DMU_SPILL_BLKID) {
2714 		mutex_enter(&dn->dn_mtx);
2715 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2716 		mutex_exit(&dn->dn_mtx);
2717 	}
2718 
2719 	/*
2720 	 * If this is a bonus buffer, simply copy the bonus data into the
2721 	 * dnode.  It will be written out when the dnode is synced (and it
2722 	 * will be synced, since it must have been dirty for dbuf_sync to
2723 	 * be called).
2724 	 */
2725 	if (db->db_blkid == DMU_BONUS_BLKID) {
2726 		dbuf_dirty_record_t **drp;
2727 
2728 		ASSERT(*datap != NULL);
2729 		ASSERT0(db->db_level);
2730 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2731 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2732 		DB_DNODE_EXIT(db);
2733 
2734 		if (*datap != db->db.db_data) {
2735 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
2736 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2737 		}
2738 		db->db_data_pending = NULL;
2739 		drp = &db->db_last_dirty;
2740 		while (*drp != dr)
2741 			drp = &(*drp)->dr_next;
2742 		ASSERT(dr->dr_next == NULL);
2743 		ASSERT(dr->dr_dbuf == db);
2744 		*drp = dr->dr_next;
2745 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
2746 		ASSERT(db->db_dirtycnt > 0);
2747 		db->db_dirtycnt -= 1;
2748 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2749 		return;
2750 	}
2751 
2752 	os = dn->dn_objset;
2753 
2754 	/*
2755 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
2756 	 * operation to sneak in. As a result, we need to ensure that we
2757 	 * don't check the dr_override_state until we have returned from
2758 	 * dbuf_check_blkptr.
2759 	 */
2760 	dbuf_check_blkptr(dn, db);
2761 
2762 	/*
2763 	 * If this buffer is in the middle of an immediate write,
2764 	 * wait for the synchronous IO to complete.
2765 	 */
2766 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2767 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2768 		cv_wait(&db->db_changed, &db->db_mtx);
2769 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2770 	}
2771 
2772 	if (db->db_state != DB_NOFILL &&
2773 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2774 	    refcount_count(&db->db_holds) > 1 &&
2775 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2776 	    *datap == db->db_buf) {
2777 		/*
2778 		 * If this buffer is currently "in use" (i.e., there
2779 		 * are active holds and db_data still references it),
2780 		 * then make a copy before we start the write so that
2781 		 * any modifications from the open txg will not leak
2782 		 * into this write.
2783 		 *
2784 		 * NOTE: this copy does not need to be made for
2785 		 * objects only modified in the syncing context (e.g.
2786 		 * DNONE_DNODE blocks).
2787 		 */
2788 		int blksz = arc_buf_size(*datap);
2789 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2790 		*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2791 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
2792 	}
2793 	db->db_data_pending = dr;
2794 
2795 	mutex_exit(&db->db_mtx);
2796 
2797 	dbuf_write(dr, *datap, tx);
2798 
2799 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2800 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2801 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2802 		DB_DNODE_EXIT(db);
2803 	} else {
2804 		/*
2805 		 * Although zio_nowait() does not "wait for an IO", it does
2806 		 * initiate the IO. If this is an empty write it seems plausible
2807 		 * that the IO could actually be completed before the nowait
2808 		 * returns. We need to DB_DNODE_EXIT() first in case
2809 		 * zio_nowait() invalidates the dbuf.
2810 		 */
2811 		DB_DNODE_EXIT(db);
2812 		zio_nowait(dr->dr_zio);
2813 	}
2814 }
2815 
2816 void
2817 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
2818 {
2819 	dbuf_dirty_record_t *dr;
2820 
2821 	while (dr = list_head(list)) {
2822 		if (dr->dr_zio != NULL) {
2823 			/*
2824 			 * If we find an already initialized zio then we
2825 			 * are processing the meta-dnode, and we have finished.
2826 			 * The dbufs for all dnodes are put back on the list
2827 			 * during processing, so that we can zio_wait()
2828 			 * these IOs after initiating all child IOs.
2829 			 */
2830 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2831 			    DMU_META_DNODE_OBJECT);
2832 			break;
2833 		}
2834 		if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
2835 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
2836 			VERIFY3U(dr->dr_dbuf->db_level, ==, level);
2837 		}
2838 		list_remove(list, dr);
2839 		if (dr->dr_dbuf->db_level > 0)
2840 			dbuf_sync_indirect(dr, tx);
2841 		else
2842 			dbuf_sync_leaf(dr, tx);
2843 	}
2844 }
2845 
2846 /* ARGSUSED */
2847 static void
2848 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2849 {
2850 	dmu_buf_impl_t *db = vdb;
2851 	dnode_t *dn;
2852 	blkptr_t *bp = zio->io_bp;
2853 	blkptr_t *bp_orig = &zio->io_bp_orig;
2854 	spa_t *spa = zio->io_spa;
2855 	int64_t delta;
2856 	uint64_t fill = 0;
2857 	int i;
2858 
2859 	ASSERT3P(db->db_blkptr, ==, bp);
2860 
2861 	DB_DNODE_ENTER(db);
2862 	dn = DB_DNODE(db);
2863 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2864 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2865 	zio->io_prev_space_delta = delta;
2866 
2867 	if (bp->blk_birth != 0) {
2868 		ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2869 		    BP_GET_TYPE(bp) == dn->dn_type) ||
2870 		    (db->db_blkid == DMU_SPILL_BLKID &&
2871 		    BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2872 		    BP_IS_EMBEDDED(bp));
2873 		ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2874 	}
2875 
2876 	mutex_enter(&db->db_mtx);
2877 
2878 #ifdef ZFS_DEBUG
2879 	if (db->db_blkid == DMU_SPILL_BLKID) {
2880 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2881 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2882 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2883 	}
2884 #endif
2885 
2886 	if (db->db_level == 0) {
2887 		mutex_enter(&dn->dn_mtx);
2888 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2889 		    db->db_blkid != DMU_SPILL_BLKID)
2890 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2891 		mutex_exit(&dn->dn_mtx);
2892 
2893 		if (dn->dn_type == DMU_OT_DNODE) {
2894 			dnode_phys_t *dnp = db->db.db_data;
2895 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2896 			    i--, dnp++) {
2897 				if (dnp->dn_type != DMU_OT_NONE)
2898 					fill++;
2899 			}
2900 		} else {
2901 			if (BP_IS_HOLE(bp)) {
2902 				fill = 0;
2903 			} else {
2904 				fill = 1;
2905 			}
2906 		}
2907 	} else {
2908 		blkptr_t *ibp = db->db.db_data;
2909 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2910 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2911 			if (BP_IS_HOLE(ibp))
2912 				continue;
2913 			fill += BP_GET_FILL(ibp);
2914 		}
2915 	}
2916 	DB_DNODE_EXIT(db);
2917 
2918 	if (!BP_IS_EMBEDDED(bp))
2919 		bp->blk_fill = fill;
2920 
2921 	mutex_exit(&db->db_mtx);
2922 }
2923 
2924 /*
2925  * The SPA will call this callback several times for each zio - once
2926  * for every physical child i/o (zio->io_phys_children times).  This
2927  * allows the DMU to monitor the progress of each logical i/o.  For example,
2928  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2929  * block.  There may be a long delay before all copies/fragments are completed,
2930  * so this callback allows us to retire dirty space gradually, as the physical
2931  * i/os complete.
2932  */
2933 /* ARGSUSED */
2934 static void
2935 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2936 {
2937 	dmu_buf_impl_t *db = arg;
2938 	objset_t *os = db->db_objset;
2939 	dsl_pool_t *dp = dmu_objset_pool(os);
2940 	dbuf_dirty_record_t *dr;
2941 	int delta = 0;
2942 
2943 	dr = db->db_data_pending;
2944 	ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2945 
2946 	/*
2947 	 * The callback will be called io_phys_children times.  Retire one
2948 	 * portion of our dirty space each time we are called.  Any rounding
2949 	 * error will be cleaned up by dsl_pool_sync()'s call to
2950 	 * dsl_pool_undirty_space().
2951 	 */
2952 	delta = dr->dr_accounted / zio->io_phys_children;
2953 	dsl_pool_undirty_space(dp, delta, zio->io_txg);
2954 }
2955 
2956 /* ARGSUSED */
2957 static void
2958 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2959 {
2960 	dmu_buf_impl_t *db = vdb;
2961 	blkptr_t *bp_orig = &zio->io_bp_orig;
2962 	blkptr_t *bp = db->db_blkptr;
2963 	objset_t *os = db->db_objset;
2964 	dmu_tx_t *tx = os->os_synctx;
2965 	dbuf_dirty_record_t **drp, *dr;
2966 
2967 	ASSERT0(zio->io_error);
2968 	ASSERT(db->db_blkptr == bp);
2969 
2970 	/*
2971 	 * For nopwrites and rewrites we ensure that the bp matches our
2972 	 * original and bypass all the accounting.
2973 	 */
2974 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2975 		ASSERT(BP_EQUAL(bp, bp_orig));
2976 	} else {
2977 		dsl_dataset_t *ds = os->os_dsl_dataset;
2978 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2979 		dsl_dataset_block_born(ds, bp, tx);
2980 	}
2981 
2982 	mutex_enter(&db->db_mtx);
2983 
2984 	DBUF_VERIFY(db);
2985 
2986 	drp = &db->db_last_dirty;
2987 	while ((dr = *drp) != db->db_data_pending)
2988 		drp = &dr->dr_next;
2989 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2990 	ASSERT(dr->dr_dbuf == db);
2991 	ASSERT(dr->dr_next == NULL);
2992 	*drp = dr->dr_next;
2993 
2994 #ifdef ZFS_DEBUG
2995 	if (db->db_blkid == DMU_SPILL_BLKID) {
2996 		dnode_t *dn;
2997 
2998 		DB_DNODE_ENTER(db);
2999 		dn = DB_DNODE(db);
3000 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3001 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3002 		    db->db_blkptr == &dn->dn_phys->dn_spill);
3003 		DB_DNODE_EXIT(db);
3004 	}
3005 #endif
3006 
3007 	if (db->db_level == 0) {
3008 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3009 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3010 		if (db->db_state != DB_NOFILL) {
3011 			if (dr->dt.dl.dr_data != db->db_buf)
3012 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
3013 				    db));
3014 			else if (!arc_released(db->db_buf))
3015 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
3016 		}
3017 	} else {
3018 		dnode_t *dn;
3019 
3020 		DB_DNODE_ENTER(db);
3021 		dn = DB_DNODE(db);
3022 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3023 		ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3024 		if (!BP_IS_HOLE(db->db_blkptr)) {
3025 			int epbs =
3026 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3027 			ASSERT3U(db->db_blkid, <=,
3028 			    dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3029 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3030 			    db->db.db_size);
3031 			if (!arc_released(db->db_buf))
3032 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
3033 		}
3034 		DB_DNODE_EXIT(db);
3035 		mutex_destroy(&dr->dt.di.dr_mtx);
3036 		list_destroy(&dr->dt.di.dr_children);
3037 	}
3038 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
3039 
3040 	cv_broadcast(&db->db_changed);
3041 	ASSERT(db->db_dirtycnt > 0);
3042 	db->db_dirtycnt -= 1;
3043 	db->db_data_pending = NULL;
3044 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3045 }
3046 
3047 static void
3048 dbuf_write_nofill_ready(zio_t *zio)
3049 {
3050 	dbuf_write_ready(zio, NULL, zio->io_private);
3051 }
3052 
3053 static void
3054 dbuf_write_nofill_done(zio_t *zio)
3055 {
3056 	dbuf_write_done(zio, NULL, zio->io_private);
3057 }
3058 
3059 static void
3060 dbuf_write_override_ready(zio_t *zio)
3061 {
3062 	dbuf_dirty_record_t *dr = zio->io_private;
3063 	dmu_buf_impl_t *db = dr->dr_dbuf;
3064 
3065 	dbuf_write_ready(zio, NULL, db);
3066 }
3067 
3068 static void
3069 dbuf_write_override_done(zio_t *zio)
3070 {
3071 	dbuf_dirty_record_t *dr = zio->io_private;
3072 	dmu_buf_impl_t *db = dr->dr_dbuf;
3073 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3074 
3075 	mutex_enter(&db->db_mtx);
3076 	if (!BP_EQUAL(zio->io_bp, obp)) {
3077 		if (!BP_IS_HOLE(obp))
3078 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3079 		arc_release(dr->dt.dl.dr_data, db);
3080 	}
3081 	mutex_exit(&db->db_mtx);
3082 
3083 	dbuf_write_done(zio, NULL, db);
3084 }
3085 
3086 /* Issue I/O to commit a dirty buffer to disk. */
3087 static void
3088 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3089 {
3090 	dmu_buf_impl_t *db = dr->dr_dbuf;
3091 	dnode_t *dn;
3092 	objset_t *os;
3093 	dmu_buf_impl_t *parent = db->db_parent;
3094 	uint64_t txg = tx->tx_txg;
3095 	zbookmark_phys_t zb;
3096 	zio_prop_t zp;
3097 	zio_t *zio;
3098 	int wp_flag = 0;
3099 
3100 	DB_DNODE_ENTER(db);
3101 	dn = DB_DNODE(db);
3102 	os = dn->dn_objset;
3103 
3104 	if (db->db_state != DB_NOFILL) {
3105 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3106 			/*
3107 			 * Private object buffers are released here rather
3108 			 * than in dbuf_dirty() since they are only modified
3109 			 * in the syncing context and we don't want the
3110 			 * overhead of making multiple copies of the data.
3111 			 */
3112 			if (BP_IS_HOLE(db->db_blkptr)) {
3113 				arc_buf_thaw(data);
3114 			} else {
3115 				dbuf_release_bp(db);
3116 			}
3117 		}
3118 	}
3119 
3120 	if (parent != dn->dn_dbuf) {
3121 		/* Our parent is an indirect block. */
3122 		/* We have a dirty parent that has been scheduled for write. */
3123 		ASSERT(parent && parent->db_data_pending);
3124 		/* Our parent's buffer is one level closer to the dnode. */
3125 		ASSERT(db->db_level == parent->db_level-1);
3126 		/*
3127 		 * We're about to modify our parent's db_data by modifying
3128 		 * our block pointer, so the parent must be released.
3129 		 */
3130 		ASSERT(arc_released(parent->db_buf));
3131 		zio = parent->db_data_pending->dr_zio;
3132 	} else {
3133 		/* Our parent is the dnode itself. */
3134 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3135 		    db->db_blkid != DMU_SPILL_BLKID) ||
3136 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3137 		if (db->db_blkid != DMU_SPILL_BLKID)
3138 			ASSERT3P(db->db_blkptr, ==,
3139 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
3140 		zio = dn->dn_zio;
3141 	}
3142 
3143 	ASSERT(db->db_level == 0 || data == db->db_buf);
3144 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3145 	ASSERT(zio);
3146 
3147 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3148 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3149 	    db->db.db_object, db->db_level, db->db_blkid);
3150 
3151 	if (db->db_blkid == DMU_SPILL_BLKID)
3152 		wp_flag = WP_SPILL;
3153 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3154 
3155 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3156 	DB_DNODE_EXIT(db);
3157 
3158 	if (db->db_level == 0 &&
3159 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3160 		/*
3161 		 * The BP for this block has been provided by open context
3162 		 * (by dmu_sync() or dmu_buf_write_embedded()).
3163 		 */
3164 		void *contents = (data != NULL) ? data->b_data : NULL;
3165 
3166 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
3167 		    db->db_blkptr, contents, db->db.db_size, &zp,
3168 		    dbuf_write_override_ready, NULL, dbuf_write_override_done,
3169 		    dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3170 		mutex_enter(&db->db_mtx);
3171 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3172 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3173 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3174 		mutex_exit(&db->db_mtx);
3175 	} else if (db->db_state == DB_NOFILL) {
3176 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3177 		    zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3178 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
3179 		    db->db_blkptr, NULL, db->db.db_size, &zp,
3180 		    dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
3181 		    ZIO_PRIORITY_ASYNC_WRITE,
3182 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3183 	} else {
3184 		ASSERT(arc_released(data));
3185 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
3186 		    db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
3187 		    DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
3188 		    dbuf_write_physdone, dbuf_write_done, db,
3189 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3190 	}
3191 }
3192