xref: /titanic_52/usr/src/uts/common/fs/zfs/dbuf.c (revision 2ac302890e472bf0c11db192dd18f12ded6043f6)
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) 2013 by Delphix. All rights reserved.
25  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26  */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dbuf.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/sa.h>
40 #include <sys/sa_impl.h>
41 
42 static void dbuf_destroy(dmu_buf_impl_t *db);
43 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
44 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
45 
46 /*
47  * Global data structures and functions for the dbuf cache.
48  */
49 static kmem_cache_t *dbuf_cache;
50 
51 /* ARGSUSED */
52 static int
53 dbuf_cons(void *vdb, void *unused, int kmflag)
54 {
55 	dmu_buf_impl_t *db = vdb;
56 	bzero(db, sizeof (dmu_buf_impl_t));
57 
58 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
59 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
60 	refcount_create(&db->db_holds);
61 	return (0);
62 }
63 
64 /* ARGSUSED */
65 static void
66 dbuf_dest(void *vdb, void *unused)
67 {
68 	dmu_buf_impl_t *db = vdb;
69 	mutex_destroy(&db->db_mtx);
70 	cv_destroy(&db->db_changed);
71 	refcount_destroy(&db->db_holds);
72 }
73 
74 /*
75  * dbuf hash table routines
76  */
77 static dbuf_hash_table_t dbuf_hash_table;
78 
79 static uint64_t dbuf_hash_count;
80 
81 static uint64_t
82 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
83 {
84 	uintptr_t osv = (uintptr_t)os;
85 	uint64_t crc = -1ULL;
86 
87 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
88 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
89 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
90 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
91 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
92 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
93 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
94 
95 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
96 
97 	return (crc);
98 }
99 
100 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
101 
102 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
103 	((dbuf)->db.db_object == (obj) &&		\
104 	(dbuf)->db_objset == (os) &&			\
105 	(dbuf)->db_level == (level) &&			\
106 	(dbuf)->db_blkid == (blkid))
107 
108 dmu_buf_impl_t *
109 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
110 {
111 	dbuf_hash_table_t *h = &dbuf_hash_table;
112 	objset_t *os = dn->dn_objset;
113 	uint64_t obj = dn->dn_object;
114 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
115 	uint64_t idx = hv & h->hash_table_mask;
116 	dmu_buf_impl_t *db;
117 
118 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
119 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
120 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
121 			mutex_enter(&db->db_mtx);
122 			if (db->db_state != DB_EVICTING) {
123 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
124 				return (db);
125 			}
126 			mutex_exit(&db->db_mtx);
127 		}
128 	}
129 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
130 	return (NULL);
131 }
132 
133 /*
134  * Insert an entry into the hash table.  If there is already an element
135  * equal to elem in the hash table, then the already existing element
136  * will be returned and the new element will not be inserted.
137  * Otherwise returns NULL.
138  */
139 static dmu_buf_impl_t *
140 dbuf_hash_insert(dmu_buf_impl_t *db)
141 {
142 	dbuf_hash_table_t *h = &dbuf_hash_table;
143 	objset_t *os = db->db_objset;
144 	uint64_t obj = db->db.db_object;
145 	int level = db->db_level;
146 	uint64_t blkid = db->db_blkid;
147 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
148 	uint64_t idx = hv & h->hash_table_mask;
149 	dmu_buf_impl_t *dbf;
150 
151 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
152 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
153 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
154 			mutex_enter(&dbf->db_mtx);
155 			if (dbf->db_state != DB_EVICTING) {
156 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
157 				return (dbf);
158 			}
159 			mutex_exit(&dbf->db_mtx);
160 		}
161 	}
162 
163 	mutex_enter(&db->db_mtx);
164 	db->db_hash_next = h->hash_table[idx];
165 	h->hash_table[idx] = db;
166 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
167 	atomic_add_64(&dbuf_hash_count, 1);
168 
169 	return (NULL);
170 }
171 
172 /*
173  * Remove an entry from the hash table.  This operation will
174  * fail if there are any existing holds on the db.
175  */
176 static void
177 dbuf_hash_remove(dmu_buf_impl_t *db)
178 {
179 	dbuf_hash_table_t *h = &dbuf_hash_table;
180 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
181 	    db->db_level, db->db_blkid);
182 	uint64_t idx = hv & h->hash_table_mask;
183 	dmu_buf_impl_t *dbf, **dbp;
184 
185 	/*
186 	 * We musn't hold db_mtx to maintin lock ordering:
187 	 * DBUF_HASH_MUTEX > db_mtx.
188 	 */
189 	ASSERT(refcount_is_zero(&db->db_holds));
190 	ASSERT(db->db_state == DB_EVICTING);
191 	ASSERT(!MUTEX_HELD(&db->db_mtx));
192 
193 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
194 	dbp = &h->hash_table[idx];
195 	while ((dbf = *dbp) != db) {
196 		dbp = &dbf->db_hash_next;
197 		ASSERT(dbf != NULL);
198 	}
199 	*dbp = db->db_hash_next;
200 	db->db_hash_next = NULL;
201 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
202 	atomic_add_64(&dbuf_hash_count, -1);
203 }
204 
205 static arc_evict_func_t dbuf_do_evict;
206 
207 static void
208 dbuf_evict_user(dmu_buf_impl_t *db)
209 {
210 	ASSERT(MUTEX_HELD(&db->db_mtx));
211 
212 	if (db->db_level != 0 || db->db_evict_func == NULL)
213 		return;
214 
215 	if (db->db_user_data_ptr_ptr)
216 		*db->db_user_data_ptr_ptr = db->db.db_data;
217 	db->db_evict_func(&db->db, db->db_user_ptr);
218 	db->db_user_ptr = NULL;
219 	db->db_user_data_ptr_ptr = NULL;
220 	db->db_evict_func = NULL;
221 }
222 
223 boolean_t
224 dbuf_is_metadata(dmu_buf_impl_t *db)
225 {
226 	if (db->db_level > 0) {
227 		return (B_TRUE);
228 	} else {
229 		boolean_t is_metadata;
230 
231 		DB_DNODE_ENTER(db);
232 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
233 		DB_DNODE_EXIT(db);
234 
235 		return (is_metadata);
236 	}
237 }
238 
239 void
240 dbuf_evict(dmu_buf_impl_t *db)
241 {
242 	ASSERT(MUTEX_HELD(&db->db_mtx));
243 	ASSERT(db->db_buf == NULL);
244 	ASSERT(db->db_data_pending == NULL);
245 
246 	dbuf_clear(db);
247 	dbuf_destroy(db);
248 }
249 
250 void
251 dbuf_init(void)
252 {
253 	uint64_t hsize = 1ULL << 16;
254 	dbuf_hash_table_t *h = &dbuf_hash_table;
255 	int i;
256 
257 	/*
258 	 * The hash table is big enough to fill all of physical memory
259 	 * with an average 4K block size.  The table will take up
260 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
261 	 */
262 	while (hsize * 4096 < physmem * PAGESIZE)
263 		hsize <<= 1;
264 
265 retry:
266 	h->hash_table_mask = hsize - 1;
267 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
268 	if (h->hash_table == NULL) {
269 		/* XXX - we should really return an error instead of assert */
270 		ASSERT(hsize > (1ULL << 10));
271 		hsize >>= 1;
272 		goto retry;
273 	}
274 
275 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
276 	    sizeof (dmu_buf_impl_t),
277 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
278 
279 	for (i = 0; i < DBUF_MUTEXES; i++)
280 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
281 }
282 
283 void
284 dbuf_fini(void)
285 {
286 	dbuf_hash_table_t *h = &dbuf_hash_table;
287 	int i;
288 
289 	for (i = 0; i < DBUF_MUTEXES; i++)
290 		mutex_destroy(&h->hash_mutexes[i]);
291 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
292 	kmem_cache_destroy(dbuf_cache);
293 }
294 
295 /*
296  * Other stuff.
297  */
298 
299 #ifdef ZFS_DEBUG
300 static void
301 dbuf_verify(dmu_buf_impl_t *db)
302 {
303 	dnode_t *dn;
304 	dbuf_dirty_record_t *dr;
305 
306 	ASSERT(MUTEX_HELD(&db->db_mtx));
307 
308 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
309 		return;
310 
311 	ASSERT(db->db_objset != NULL);
312 	DB_DNODE_ENTER(db);
313 	dn = DB_DNODE(db);
314 	if (dn == NULL) {
315 		ASSERT(db->db_parent == NULL);
316 		ASSERT(db->db_blkptr == NULL);
317 	} else {
318 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
319 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
320 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
321 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
322 		    db->db_blkid == DMU_SPILL_BLKID ||
323 		    !list_is_empty(&dn->dn_dbufs));
324 	}
325 	if (db->db_blkid == DMU_BONUS_BLKID) {
326 		ASSERT(dn != NULL);
327 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
328 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
329 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
330 		ASSERT(dn != NULL);
331 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
332 		ASSERT0(db->db.db_offset);
333 	} else {
334 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
335 	}
336 
337 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
338 		ASSERT(dr->dr_dbuf == db);
339 
340 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
341 		ASSERT(dr->dr_dbuf == db);
342 
343 	/*
344 	 * We can't assert that db_size matches dn_datablksz because it
345 	 * can be momentarily different when another thread is doing
346 	 * dnode_set_blksz().
347 	 */
348 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
349 		dr = db->db_data_pending;
350 		/*
351 		 * It should only be modified in syncing context, so
352 		 * make sure we only have one copy of the data.
353 		 */
354 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
355 	}
356 
357 	/* verify db->db_blkptr */
358 	if (db->db_blkptr) {
359 		if (db->db_parent == dn->dn_dbuf) {
360 			/* db is pointed to by the dnode */
361 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
362 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
363 				ASSERT(db->db_parent == NULL);
364 			else
365 				ASSERT(db->db_parent != NULL);
366 			if (db->db_blkid != DMU_SPILL_BLKID)
367 				ASSERT3P(db->db_blkptr, ==,
368 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
369 		} else {
370 			/* db is pointed to by an indirect block */
371 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
372 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
373 			ASSERT3U(db->db_parent->db.db_object, ==,
374 			    db->db.db_object);
375 			/*
376 			 * dnode_grow_indblksz() can make this fail if we don't
377 			 * have the struct_rwlock.  XXX indblksz no longer
378 			 * grows.  safe to do this now?
379 			 */
380 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
381 				ASSERT3P(db->db_blkptr, ==,
382 				    ((blkptr_t *)db->db_parent->db.db_data +
383 				    db->db_blkid % epb));
384 			}
385 		}
386 	}
387 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
388 	    (db->db_buf == NULL || db->db_buf->b_data) &&
389 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
390 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
391 		/*
392 		 * If the blkptr isn't set but they have nonzero data,
393 		 * it had better be dirty, otherwise we'll lose that
394 		 * data when we evict this buffer.
395 		 */
396 		if (db->db_dirtycnt == 0) {
397 			uint64_t *buf = db->db.db_data;
398 			int i;
399 
400 			for (i = 0; i < db->db.db_size >> 3; i++) {
401 				ASSERT(buf[i] == 0);
402 			}
403 		}
404 	}
405 	DB_DNODE_EXIT(db);
406 }
407 #endif
408 
409 static void
410 dbuf_update_data(dmu_buf_impl_t *db)
411 {
412 	ASSERT(MUTEX_HELD(&db->db_mtx));
413 	if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
414 		ASSERT(!refcount_is_zero(&db->db_holds));
415 		*db->db_user_data_ptr_ptr = db->db.db_data;
416 	}
417 }
418 
419 static void
420 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
421 {
422 	ASSERT(MUTEX_HELD(&db->db_mtx));
423 	ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
424 	db->db_buf = buf;
425 	if (buf != NULL) {
426 		ASSERT(buf->b_data != NULL);
427 		db->db.db_data = buf->b_data;
428 		if (!arc_released(buf))
429 			arc_set_callback(buf, dbuf_do_evict, db);
430 		dbuf_update_data(db);
431 	} else {
432 		dbuf_evict_user(db);
433 		db->db.db_data = NULL;
434 		if (db->db_state != DB_NOFILL)
435 			db->db_state = DB_UNCACHED;
436 	}
437 }
438 
439 /*
440  * Loan out an arc_buf for read.  Return the loaned arc_buf.
441  */
442 arc_buf_t *
443 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
444 {
445 	arc_buf_t *abuf;
446 
447 	mutex_enter(&db->db_mtx);
448 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
449 		int blksz = db->db.db_size;
450 		spa_t *spa;
451 
452 		mutex_exit(&db->db_mtx);
453 		DB_GET_SPA(&spa, db);
454 		abuf = arc_loan_buf(spa, blksz);
455 		bcopy(db->db.db_data, abuf->b_data, blksz);
456 	} else {
457 		abuf = db->db_buf;
458 		arc_loan_inuse_buf(abuf, db);
459 		dbuf_set_data(db, NULL);
460 		mutex_exit(&db->db_mtx);
461 	}
462 	return (abuf);
463 }
464 
465 uint64_t
466 dbuf_whichblock(dnode_t *dn, uint64_t offset)
467 {
468 	if (dn->dn_datablkshift) {
469 		return (offset >> dn->dn_datablkshift);
470 	} else {
471 		ASSERT3U(offset, <, dn->dn_datablksz);
472 		return (0);
473 	}
474 }
475 
476 static void
477 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
478 {
479 	dmu_buf_impl_t *db = vdb;
480 
481 	mutex_enter(&db->db_mtx);
482 	ASSERT3U(db->db_state, ==, DB_READ);
483 	/*
484 	 * All reads are synchronous, so we must have a hold on the dbuf
485 	 */
486 	ASSERT(refcount_count(&db->db_holds) > 0);
487 	ASSERT(db->db_buf == NULL);
488 	ASSERT(db->db.db_data == NULL);
489 	if (db->db_level == 0 && db->db_freed_in_flight) {
490 		/* we were freed in flight; disregard any error */
491 		arc_release(buf, db);
492 		bzero(buf->b_data, db->db.db_size);
493 		arc_buf_freeze(buf);
494 		db->db_freed_in_flight = FALSE;
495 		dbuf_set_data(db, buf);
496 		db->db_state = DB_CACHED;
497 	} else if (zio == NULL || zio->io_error == 0) {
498 		dbuf_set_data(db, buf);
499 		db->db_state = DB_CACHED;
500 	} else {
501 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
502 		ASSERT3P(db->db_buf, ==, NULL);
503 		VERIFY(arc_buf_remove_ref(buf, db));
504 		db->db_state = DB_UNCACHED;
505 	}
506 	cv_broadcast(&db->db_changed);
507 	dbuf_rele_and_unlock(db, NULL);
508 }
509 
510 static void
511 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
512 {
513 	dnode_t *dn;
514 	spa_t *spa;
515 	zbookmark_t zb;
516 	uint32_t aflags = ARC_NOWAIT;
517 
518 	DB_DNODE_ENTER(db);
519 	dn = DB_DNODE(db);
520 	ASSERT(!refcount_is_zero(&db->db_holds));
521 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
522 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
523 	ASSERT(MUTEX_HELD(&db->db_mtx));
524 	ASSERT(db->db_state == DB_UNCACHED);
525 	ASSERT(db->db_buf == NULL);
526 
527 	if (db->db_blkid == DMU_BONUS_BLKID) {
528 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
529 
530 		ASSERT3U(bonuslen, <=, db->db.db_size);
531 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
532 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
533 		if (bonuslen < DN_MAX_BONUSLEN)
534 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
535 		if (bonuslen)
536 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
537 		DB_DNODE_EXIT(db);
538 		dbuf_update_data(db);
539 		db->db_state = DB_CACHED;
540 		mutex_exit(&db->db_mtx);
541 		return;
542 	}
543 
544 	/*
545 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
546 	 * processes the delete record and clears the bp while we are waiting
547 	 * for the dn_mtx (resulting in a "no" from block_freed).
548 	 */
549 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
550 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
551 	    BP_IS_HOLE(db->db_blkptr)))) {
552 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
553 
554 		dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
555 		    db->db.db_size, db, type));
556 		DB_DNODE_EXIT(db);
557 		bzero(db->db.db_data, db->db.db_size);
558 		db->db_state = DB_CACHED;
559 		*flags |= DB_RF_CACHED;
560 		mutex_exit(&db->db_mtx);
561 		return;
562 	}
563 
564 	spa = dn->dn_objset->os_spa;
565 	DB_DNODE_EXIT(db);
566 
567 	db->db_state = DB_READ;
568 	mutex_exit(&db->db_mtx);
569 
570 	if (DBUF_IS_L2CACHEABLE(db))
571 		aflags |= ARC_L2CACHE;
572 	if (DBUF_IS_L2COMPRESSIBLE(db))
573 		aflags |= ARC_L2COMPRESS;
574 
575 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
576 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
577 	    db->db.db_object, db->db_level, db->db_blkid);
578 
579 	dbuf_add_ref(db, NULL);
580 
581 	(void) arc_read(zio, spa, db->db_blkptr,
582 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
583 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
584 	    &aflags, &zb);
585 	if (aflags & ARC_CACHED)
586 		*flags |= DB_RF_CACHED;
587 }
588 
589 int
590 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
591 {
592 	int err = 0;
593 	int havepzio = (zio != NULL);
594 	int prefetch;
595 	dnode_t *dn;
596 
597 	/*
598 	 * We don't have to hold the mutex to check db_state because it
599 	 * can't be freed while we have a hold on the buffer.
600 	 */
601 	ASSERT(!refcount_is_zero(&db->db_holds));
602 
603 	if (db->db_state == DB_NOFILL)
604 		return (SET_ERROR(EIO));
605 
606 	DB_DNODE_ENTER(db);
607 	dn = DB_DNODE(db);
608 	if ((flags & DB_RF_HAVESTRUCT) == 0)
609 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
610 
611 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
612 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
613 	    DBUF_IS_CACHEABLE(db);
614 
615 	mutex_enter(&db->db_mtx);
616 	if (db->db_state == DB_CACHED) {
617 		mutex_exit(&db->db_mtx);
618 		if (prefetch)
619 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
620 			    db->db.db_size, TRUE);
621 		if ((flags & DB_RF_HAVESTRUCT) == 0)
622 			rw_exit(&dn->dn_struct_rwlock);
623 		DB_DNODE_EXIT(db);
624 	} else if (db->db_state == DB_UNCACHED) {
625 		spa_t *spa = dn->dn_objset->os_spa;
626 
627 		if (zio == NULL)
628 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
629 		dbuf_read_impl(db, zio, &flags);
630 
631 		/* dbuf_read_impl has dropped db_mtx for us */
632 
633 		if (prefetch)
634 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
635 			    db->db.db_size, flags & DB_RF_CACHED);
636 
637 		if ((flags & DB_RF_HAVESTRUCT) == 0)
638 			rw_exit(&dn->dn_struct_rwlock);
639 		DB_DNODE_EXIT(db);
640 
641 		if (!havepzio)
642 			err = zio_wait(zio);
643 	} else {
644 		mutex_exit(&db->db_mtx);
645 		if (prefetch)
646 			dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
647 			    db->db.db_size, TRUE);
648 		if ((flags & DB_RF_HAVESTRUCT) == 0)
649 			rw_exit(&dn->dn_struct_rwlock);
650 		DB_DNODE_EXIT(db);
651 
652 		mutex_enter(&db->db_mtx);
653 		if ((flags & DB_RF_NEVERWAIT) == 0) {
654 			while (db->db_state == DB_READ ||
655 			    db->db_state == DB_FILL) {
656 				ASSERT(db->db_state == DB_READ ||
657 				    (flags & DB_RF_HAVESTRUCT) == 0);
658 				cv_wait(&db->db_changed, &db->db_mtx);
659 			}
660 			if (db->db_state == DB_UNCACHED)
661 				err = SET_ERROR(EIO);
662 		}
663 		mutex_exit(&db->db_mtx);
664 	}
665 
666 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
667 	return (err);
668 }
669 
670 static void
671 dbuf_noread(dmu_buf_impl_t *db)
672 {
673 	ASSERT(!refcount_is_zero(&db->db_holds));
674 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
675 	mutex_enter(&db->db_mtx);
676 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
677 		cv_wait(&db->db_changed, &db->db_mtx);
678 	if (db->db_state == DB_UNCACHED) {
679 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
680 		spa_t *spa;
681 
682 		ASSERT(db->db_buf == NULL);
683 		ASSERT(db->db.db_data == NULL);
684 		DB_GET_SPA(&spa, db);
685 		dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
686 		db->db_state = DB_FILL;
687 	} else if (db->db_state == DB_NOFILL) {
688 		dbuf_set_data(db, NULL);
689 	} else {
690 		ASSERT3U(db->db_state, ==, DB_CACHED);
691 	}
692 	mutex_exit(&db->db_mtx);
693 }
694 
695 /*
696  * This is our just-in-time copy function.  It makes a copy of
697  * buffers, that have been modified in a previous transaction
698  * group, before we modify them in the current active group.
699  *
700  * This function is used in two places: when we are dirtying a
701  * buffer for the first time in a txg, and when we are freeing
702  * a range in a dnode that includes this buffer.
703  *
704  * Note that when we are called from dbuf_free_range() we do
705  * not put a hold on the buffer, we just traverse the active
706  * dbuf list for the dnode.
707  */
708 static void
709 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
710 {
711 	dbuf_dirty_record_t *dr = db->db_last_dirty;
712 
713 	ASSERT(MUTEX_HELD(&db->db_mtx));
714 	ASSERT(db->db.db_data != NULL);
715 	ASSERT(db->db_level == 0);
716 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
717 
718 	if (dr == NULL ||
719 	    (dr->dt.dl.dr_data !=
720 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
721 		return;
722 
723 	/*
724 	 * If the last dirty record for this dbuf has not yet synced
725 	 * and its referencing the dbuf data, either:
726 	 *	reset the reference to point to a new copy,
727 	 * or (if there a no active holders)
728 	 *	just null out the current db_data pointer.
729 	 */
730 	ASSERT(dr->dr_txg >= txg - 2);
731 	if (db->db_blkid == DMU_BONUS_BLKID) {
732 		/* Note that the data bufs here are zio_bufs */
733 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
734 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
735 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
736 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
737 		int size = db->db.db_size;
738 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
739 		spa_t *spa;
740 
741 		DB_GET_SPA(&spa, db);
742 		dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
743 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
744 	} else {
745 		dbuf_set_data(db, NULL);
746 	}
747 }
748 
749 void
750 dbuf_unoverride(dbuf_dirty_record_t *dr)
751 {
752 	dmu_buf_impl_t *db = dr->dr_dbuf;
753 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
754 	uint64_t txg = dr->dr_txg;
755 
756 	ASSERT(MUTEX_HELD(&db->db_mtx));
757 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
758 	ASSERT(db->db_level == 0);
759 
760 	if (db->db_blkid == DMU_BONUS_BLKID ||
761 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
762 		return;
763 
764 	ASSERT(db->db_data_pending != dr);
765 
766 	/* free this block */
767 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) {
768 		spa_t *spa;
769 
770 		DB_GET_SPA(&spa, db);
771 		zio_free(spa, txg, bp);
772 	}
773 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
774 	dr->dt.dl.dr_nopwrite = B_FALSE;
775 
776 	/*
777 	 * Release the already-written buffer, so we leave it in
778 	 * a consistent dirty state.  Note that all callers are
779 	 * modifying the buffer, so they will immediately do
780 	 * another (redundant) arc_release().  Therefore, leave
781 	 * the buf thawed to save the effort of freezing &
782 	 * immediately re-thawing it.
783 	 */
784 	arc_release(dr->dt.dl.dr_data, db);
785 }
786 
787 /*
788  * Evict (if its unreferenced) or clear (if its referenced) any level-0
789  * data blocks in the free range, so that any future readers will find
790  * empty blocks.  Also, if we happen accross any level-1 dbufs in the
791  * range that have not already been marked dirty, mark them dirty so
792  * they stay in memory.
793  */
794 void
795 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
796 {
797 	dmu_buf_impl_t *db, *db_next;
798 	uint64_t txg = tx->tx_txg;
799 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
800 	uint64_t first_l1 = start >> epbs;
801 	uint64_t last_l1 = end >> epbs;
802 
803 	if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
804 		end = dn->dn_maxblkid;
805 		last_l1 = end >> epbs;
806 	}
807 	dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
808 	mutex_enter(&dn->dn_dbufs_mtx);
809 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
810 		db_next = list_next(&dn->dn_dbufs, db);
811 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
812 
813 		if (db->db_level == 1 &&
814 		    db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
815 			mutex_enter(&db->db_mtx);
816 			if (db->db_last_dirty &&
817 			    db->db_last_dirty->dr_txg < txg) {
818 				dbuf_add_ref(db, FTAG);
819 				mutex_exit(&db->db_mtx);
820 				dbuf_will_dirty(db, tx);
821 				dbuf_rele(db, FTAG);
822 			} else {
823 				mutex_exit(&db->db_mtx);
824 			}
825 		}
826 
827 		if (db->db_level != 0)
828 			continue;
829 		dprintf_dbuf(db, "found buf %s\n", "");
830 		if (db->db_blkid < start || db->db_blkid > end)
831 			continue;
832 
833 		/* found a level 0 buffer in the range */
834 		mutex_enter(&db->db_mtx);
835 		if (dbuf_undirty(db, tx)) {
836 			/* mutex has been dropped and dbuf destroyed */
837 			continue;
838 		}
839 
840 		if (db->db_state == DB_UNCACHED ||
841 		    db->db_state == DB_NOFILL ||
842 		    db->db_state == DB_EVICTING) {
843 			ASSERT(db->db.db_data == NULL);
844 			mutex_exit(&db->db_mtx);
845 			continue;
846 		}
847 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
848 			/* will be handled in dbuf_read_done or dbuf_rele */
849 			db->db_freed_in_flight = TRUE;
850 			mutex_exit(&db->db_mtx);
851 			continue;
852 		}
853 		if (refcount_count(&db->db_holds) == 0) {
854 			ASSERT(db->db_buf);
855 			dbuf_clear(db);
856 			continue;
857 		}
858 		/* The dbuf is referenced */
859 
860 		if (db->db_last_dirty != NULL) {
861 			dbuf_dirty_record_t *dr = db->db_last_dirty;
862 
863 			if (dr->dr_txg == txg) {
864 				/*
865 				 * This buffer is "in-use", re-adjust the file
866 				 * size to reflect that this buffer may
867 				 * contain new data when we sync.
868 				 */
869 				if (db->db_blkid != DMU_SPILL_BLKID &&
870 				    db->db_blkid > dn->dn_maxblkid)
871 					dn->dn_maxblkid = db->db_blkid;
872 				dbuf_unoverride(dr);
873 			} else {
874 				/*
875 				 * This dbuf is not dirty in the open context.
876 				 * Either uncache it (if its not referenced in
877 				 * the open context) or reset its contents to
878 				 * empty.
879 				 */
880 				dbuf_fix_old_data(db, txg);
881 			}
882 		}
883 		/* clear the contents if its cached */
884 		if (db->db_state == DB_CACHED) {
885 			ASSERT(db->db.db_data != NULL);
886 			arc_release(db->db_buf, db);
887 			bzero(db->db.db_data, db->db.db_size);
888 			arc_buf_freeze(db->db_buf);
889 		}
890 
891 		mutex_exit(&db->db_mtx);
892 	}
893 	mutex_exit(&dn->dn_dbufs_mtx);
894 }
895 
896 static int
897 dbuf_block_freeable(dmu_buf_impl_t *db)
898 {
899 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
900 	uint64_t birth_txg = 0;
901 
902 	/*
903 	 * We don't need any locking to protect db_blkptr:
904 	 * If it's syncing, then db_last_dirty will be set
905 	 * so we'll ignore db_blkptr.
906 	 */
907 	ASSERT(MUTEX_HELD(&db->db_mtx));
908 	if (db->db_last_dirty)
909 		birth_txg = db->db_last_dirty->dr_txg;
910 	else if (db->db_blkptr)
911 		birth_txg = db->db_blkptr->blk_birth;
912 
913 	/*
914 	 * If we don't exist or are in a snapshot, we can't be freed.
915 	 * Don't pass the bp to dsl_dataset_block_freeable() since we
916 	 * are holding the db_mtx lock and might deadlock if we are
917 	 * prefetching a dedup-ed block.
918 	 */
919 	if (birth_txg)
920 		return (ds == NULL ||
921 		    dsl_dataset_block_freeable(ds, NULL, birth_txg));
922 	else
923 		return (FALSE);
924 }
925 
926 void
927 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
928 {
929 	arc_buf_t *buf, *obuf;
930 	int osize = db->db.db_size;
931 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
932 	dnode_t *dn;
933 
934 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
935 
936 	DB_DNODE_ENTER(db);
937 	dn = DB_DNODE(db);
938 
939 	/* XXX does *this* func really need the lock? */
940 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
941 
942 	/*
943 	 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
944 	 * is OK, because there can be no other references to the db
945 	 * when we are changing its size, so no concurrent DB_FILL can
946 	 * be happening.
947 	 */
948 	/*
949 	 * XXX we should be doing a dbuf_read, checking the return
950 	 * value and returning that up to our callers
951 	 */
952 	dbuf_will_dirty(db, tx);
953 
954 	/* create the data buffer for the new block */
955 	buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
956 
957 	/* copy old block data to the new block */
958 	obuf = db->db_buf;
959 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
960 	/* zero the remainder */
961 	if (size > osize)
962 		bzero((uint8_t *)buf->b_data + osize, size - osize);
963 
964 	mutex_enter(&db->db_mtx);
965 	dbuf_set_data(db, buf);
966 	VERIFY(arc_buf_remove_ref(obuf, db));
967 	db->db.db_size = size;
968 
969 	if (db->db_level == 0) {
970 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
971 		db->db_last_dirty->dt.dl.dr_data = buf;
972 	}
973 	mutex_exit(&db->db_mtx);
974 
975 	dnode_willuse_space(dn, size-osize, tx);
976 	DB_DNODE_EXIT(db);
977 }
978 
979 void
980 dbuf_release_bp(dmu_buf_impl_t *db)
981 {
982 	objset_t *os;
983 
984 	DB_GET_OBJSET(&os, db);
985 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
986 	ASSERT(arc_released(os->os_phys_buf) ||
987 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
988 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
989 
990 	(void) arc_release(db->db_buf, db);
991 }
992 
993 dbuf_dirty_record_t *
994 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
995 {
996 	dnode_t *dn;
997 	objset_t *os;
998 	dbuf_dirty_record_t **drp, *dr;
999 	int drop_struct_lock = FALSE;
1000 	boolean_t do_free_accounting = B_FALSE;
1001 	int txgoff = tx->tx_txg & TXG_MASK;
1002 
1003 	ASSERT(tx->tx_txg != 0);
1004 	ASSERT(!refcount_is_zero(&db->db_holds));
1005 	DMU_TX_DIRTY_BUF(tx, db);
1006 
1007 	DB_DNODE_ENTER(db);
1008 	dn = DB_DNODE(db);
1009 	/*
1010 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1011 	 * objects may be dirtied in syncing context, but only if they
1012 	 * were already pre-dirtied in open context.
1013 	 */
1014 	ASSERT(!dmu_tx_is_syncing(tx) ||
1015 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1016 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1017 	    dn->dn_objset->os_dsl_dataset == NULL);
1018 	/*
1019 	 * We make this assert for private objects as well, but after we
1020 	 * check if we're already dirty.  They are allowed to re-dirty
1021 	 * in syncing context.
1022 	 */
1023 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1024 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1025 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1026 
1027 	mutex_enter(&db->db_mtx);
1028 	/*
1029 	 * XXX make this true for indirects too?  The problem is that
1030 	 * transactions created with dmu_tx_create_assigned() from
1031 	 * syncing context don't bother holding ahead.
1032 	 */
1033 	ASSERT(db->db_level != 0 ||
1034 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1035 	    db->db_state == DB_NOFILL);
1036 
1037 	mutex_enter(&dn->dn_mtx);
1038 	/*
1039 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1040 	 * initialize the objset.
1041 	 */
1042 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1043 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1044 		dn->dn_dirtyctx =
1045 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1046 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
1047 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1048 	}
1049 	mutex_exit(&dn->dn_mtx);
1050 
1051 	if (db->db_blkid == DMU_SPILL_BLKID)
1052 		dn->dn_have_spill = B_TRUE;
1053 
1054 	/*
1055 	 * If this buffer is already dirty, we're done.
1056 	 */
1057 	drp = &db->db_last_dirty;
1058 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1059 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1060 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1061 		drp = &dr->dr_next;
1062 	if (dr && dr->dr_txg == tx->tx_txg) {
1063 		DB_DNODE_EXIT(db);
1064 
1065 		if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1066 			/*
1067 			 * If this buffer has already been written out,
1068 			 * we now need to reset its state.
1069 			 */
1070 			dbuf_unoverride(dr);
1071 			if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1072 			    db->db_state != DB_NOFILL)
1073 				arc_buf_thaw(db->db_buf);
1074 		}
1075 		mutex_exit(&db->db_mtx);
1076 		return (dr);
1077 	}
1078 
1079 	/*
1080 	 * Only valid if not already dirty.
1081 	 */
1082 	ASSERT(dn->dn_object == 0 ||
1083 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1084 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1085 
1086 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1087 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1088 	    dn->dn_phys->dn_nlevels > db->db_level ||
1089 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1090 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1091 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1092 
1093 	/*
1094 	 * We should only be dirtying in syncing context if it's the
1095 	 * mos or we're initializing the os or it's a special object.
1096 	 * However, we are allowed to dirty in syncing context provided
1097 	 * we already dirtied it in open context.  Hence we must make
1098 	 * this assertion only if we're not already dirty.
1099 	 */
1100 	os = dn->dn_objset;
1101 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1102 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1103 	ASSERT(db->db.db_size != 0);
1104 
1105 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1106 
1107 	if (db->db_blkid != DMU_BONUS_BLKID) {
1108 		/*
1109 		 * Update the accounting.
1110 		 * Note: we delay "free accounting" until after we drop
1111 		 * the db_mtx.  This keeps us from grabbing other locks
1112 		 * (and possibly deadlocking) in bp_get_dsize() while
1113 		 * also holding the db_mtx.
1114 		 */
1115 		dnode_willuse_space(dn, db->db.db_size, tx);
1116 		do_free_accounting = dbuf_block_freeable(db);
1117 	}
1118 
1119 	/*
1120 	 * If this buffer is dirty in an old transaction group we need
1121 	 * to make a copy of it so that the changes we make in this
1122 	 * transaction group won't leak out when we sync the older txg.
1123 	 */
1124 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1125 	if (db->db_level == 0) {
1126 		void *data_old = db->db_buf;
1127 
1128 		if (db->db_state != DB_NOFILL) {
1129 			if (db->db_blkid == DMU_BONUS_BLKID) {
1130 				dbuf_fix_old_data(db, tx->tx_txg);
1131 				data_old = db->db.db_data;
1132 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1133 				/*
1134 				 * Release the data buffer from the cache so
1135 				 * that we can modify it without impacting
1136 				 * possible other users of this cached data
1137 				 * block.  Note that indirect blocks and
1138 				 * private objects are not released until the
1139 				 * syncing state (since they are only modified
1140 				 * then).
1141 				 */
1142 				arc_release(db->db_buf, db);
1143 				dbuf_fix_old_data(db, tx->tx_txg);
1144 				data_old = db->db_buf;
1145 			}
1146 			ASSERT(data_old != NULL);
1147 		}
1148 		dr->dt.dl.dr_data = data_old;
1149 	} else {
1150 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1151 		list_create(&dr->dt.di.dr_children,
1152 		    sizeof (dbuf_dirty_record_t),
1153 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1154 	}
1155 	dr->dr_dbuf = db;
1156 	dr->dr_txg = tx->tx_txg;
1157 	dr->dr_next = *drp;
1158 	*drp = dr;
1159 
1160 	/*
1161 	 * We could have been freed_in_flight between the dbuf_noread
1162 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1163 	 * happened after the free.
1164 	 */
1165 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1166 	    db->db_blkid != DMU_SPILL_BLKID) {
1167 		mutex_enter(&dn->dn_mtx);
1168 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1169 		mutex_exit(&dn->dn_mtx);
1170 		db->db_freed_in_flight = FALSE;
1171 	}
1172 
1173 	/*
1174 	 * This buffer is now part of this txg
1175 	 */
1176 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1177 	db->db_dirtycnt += 1;
1178 	ASSERT3U(db->db_dirtycnt, <=, 3);
1179 
1180 	mutex_exit(&db->db_mtx);
1181 
1182 	if (db->db_blkid == DMU_BONUS_BLKID ||
1183 	    db->db_blkid == DMU_SPILL_BLKID) {
1184 		mutex_enter(&dn->dn_mtx);
1185 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1186 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1187 		mutex_exit(&dn->dn_mtx);
1188 		dnode_setdirty(dn, tx);
1189 		DB_DNODE_EXIT(db);
1190 		return (dr);
1191 	} else if (do_free_accounting) {
1192 		blkptr_t *bp = db->db_blkptr;
1193 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1194 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1195 		/*
1196 		 * This is only a guess -- if the dbuf is dirty
1197 		 * in a previous txg, we don't know how much
1198 		 * space it will use on disk yet.  We should
1199 		 * really have the struct_rwlock to access
1200 		 * db_blkptr, but since this is just a guess,
1201 		 * it's OK if we get an odd answer.
1202 		 */
1203 		ddt_prefetch(os->os_spa, bp);
1204 		dnode_willuse_space(dn, -willfree, tx);
1205 	}
1206 
1207 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1208 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1209 		drop_struct_lock = TRUE;
1210 	}
1211 
1212 	if (db->db_level == 0) {
1213 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1214 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1215 	}
1216 
1217 	if (db->db_level+1 < dn->dn_nlevels) {
1218 		dmu_buf_impl_t *parent = db->db_parent;
1219 		dbuf_dirty_record_t *di;
1220 		int parent_held = FALSE;
1221 
1222 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1223 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1224 
1225 			parent = dbuf_hold_level(dn, db->db_level+1,
1226 			    db->db_blkid >> epbs, FTAG);
1227 			ASSERT(parent != NULL);
1228 			parent_held = TRUE;
1229 		}
1230 		if (drop_struct_lock)
1231 			rw_exit(&dn->dn_struct_rwlock);
1232 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1233 		di = dbuf_dirty(parent, tx);
1234 		if (parent_held)
1235 			dbuf_rele(parent, FTAG);
1236 
1237 		mutex_enter(&db->db_mtx);
1238 		/*  possible race with dbuf_undirty() */
1239 		if (db->db_last_dirty == dr ||
1240 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1241 			mutex_enter(&di->dt.di.dr_mtx);
1242 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1243 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1244 			list_insert_tail(&di->dt.di.dr_children, dr);
1245 			mutex_exit(&di->dt.di.dr_mtx);
1246 			dr->dr_parent = di;
1247 		}
1248 		mutex_exit(&db->db_mtx);
1249 	} else {
1250 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1251 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1252 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1253 		mutex_enter(&dn->dn_mtx);
1254 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1255 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1256 		mutex_exit(&dn->dn_mtx);
1257 		if (drop_struct_lock)
1258 			rw_exit(&dn->dn_struct_rwlock);
1259 	}
1260 
1261 	dnode_setdirty(dn, tx);
1262 	DB_DNODE_EXIT(db);
1263 	return (dr);
1264 }
1265 
1266 /*
1267  * Return TRUE if this evicted the dbuf.
1268  */
1269 static boolean_t
1270 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1271 {
1272 	dnode_t *dn;
1273 	uint64_t txg = tx->tx_txg;
1274 	dbuf_dirty_record_t *dr, **drp;
1275 
1276 	ASSERT(txg != 0);
1277 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1278 	ASSERT0(db->db_level);
1279 	ASSERT(MUTEX_HELD(&db->db_mtx));
1280 
1281 	/*
1282 	 * If this buffer is not dirty, we're done.
1283 	 */
1284 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1285 		if (dr->dr_txg <= txg)
1286 			break;
1287 	if (dr == NULL || dr->dr_txg < txg)
1288 		return (B_FALSE);
1289 	ASSERT(dr->dr_txg == txg);
1290 	ASSERT(dr->dr_dbuf == db);
1291 
1292 	DB_DNODE_ENTER(db);
1293 	dn = DB_DNODE(db);
1294 
1295 	/*
1296 	 * Note:  This code will probably work even if there are concurrent
1297 	 * holders, but it is untested in that scenerio, as the ZPL and
1298 	 * ztest have additional locking (the range locks) that prevents
1299 	 * that type of concurrent access.
1300 	 */
1301 	ASSERT3U(refcount_count(&db->db_holds), ==, db->db_dirtycnt);
1302 
1303 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1304 
1305 	ASSERT(db->db.db_size != 0);
1306 
1307 	/* XXX would be nice to fix up dn_towrite_space[] */
1308 
1309 	*drp = dr->dr_next;
1310 
1311 	/*
1312 	 * Note that there are three places in dbuf_dirty()
1313 	 * where this dirty record may be put on a list.
1314 	 * Make sure to do a list_remove corresponding to
1315 	 * every one of those list_insert calls.
1316 	 */
1317 	if (dr->dr_parent) {
1318 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1319 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1320 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1321 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1322 	    db->db_level+1 == dn->dn_nlevels) {
1323 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1324 		mutex_enter(&dn->dn_mtx);
1325 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1326 		mutex_exit(&dn->dn_mtx);
1327 	}
1328 	DB_DNODE_EXIT(db);
1329 
1330 	if (db->db_state != DB_NOFILL) {
1331 		dbuf_unoverride(dr);
1332 
1333 		ASSERT(db->db_buf != NULL);
1334 		ASSERT(dr->dt.dl.dr_data != NULL);
1335 		if (dr->dt.dl.dr_data != db->db_buf)
1336 			VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1337 	}
1338 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1339 
1340 	ASSERT(db->db_dirtycnt > 0);
1341 	db->db_dirtycnt -= 1;
1342 
1343 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1344 		arc_buf_t *buf = db->db_buf;
1345 
1346 		ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1347 		dbuf_set_data(db, NULL);
1348 		VERIFY(arc_buf_remove_ref(buf, db));
1349 		dbuf_evict(db);
1350 		return (B_TRUE);
1351 	}
1352 
1353 	return (B_FALSE);
1354 }
1355 
1356 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1357 void
1358 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1359 {
1360 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1361 
1362 	ASSERT(tx->tx_txg != 0);
1363 	ASSERT(!refcount_is_zero(&db->db_holds));
1364 
1365 	DB_DNODE_ENTER(db);
1366 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1367 		rf |= DB_RF_HAVESTRUCT;
1368 	DB_DNODE_EXIT(db);
1369 	(void) dbuf_read(db, NULL, rf);
1370 	(void) dbuf_dirty(db, tx);
1371 }
1372 
1373 void
1374 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1375 {
1376 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1377 
1378 	db->db_state = DB_NOFILL;
1379 
1380 	dmu_buf_will_fill(db_fake, tx);
1381 }
1382 
1383 void
1384 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1385 {
1386 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1387 
1388 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1389 	ASSERT(tx->tx_txg != 0);
1390 	ASSERT(db->db_level == 0);
1391 	ASSERT(!refcount_is_zero(&db->db_holds));
1392 
1393 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1394 	    dmu_tx_private_ok(tx));
1395 
1396 	dbuf_noread(db);
1397 	(void) dbuf_dirty(db, tx);
1398 }
1399 
1400 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1401 /* ARGSUSED */
1402 void
1403 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1404 {
1405 	mutex_enter(&db->db_mtx);
1406 	DBUF_VERIFY(db);
1407 
1408 	if (db->db_state == DB_FILL) {
1409 		if (db->db_level == 0 && db->db_freed_in_flight) {
1410 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1411 			/* we were freed while filling */
1412 			/* XXX dbuf_undirty? */
1413 			bzero(db->db.db_data, db->db.db_size);
1414 			db->db_freed_in_flight = FALSE;
1415 		}
1416 		db->db_state = DB_CACHED;
1417 		cv_broadcast(&db->db_changed);
1418 	}
1419 	mutex_exit(&db->db_mtx);
1420 }
1421 
1422 /*
1423  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1424  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1425  */
1426 void
1427 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1428 {
1429 	ASSERT(!refcount_is_zero(&db->db_holds));
1430 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1431 	ASSERT(db->db_level == 0);
1432 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1433 	ASSERT(buf != NULL);
1434 	ASSERT(arc_buf_size(buf) == db->db.db_size);
1435 	ASSERT(tx->tx_txg != 0);
1436 
1437 	arc_return_buf(buf, db);
1438 	ASSERT(arc_released(buf));
1439 
1440 	mutex_enter(&db->db_mtx);
1441 
1442 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1443 		cv_wait(&db->db_changed, &db->db_mtx);
1444 
1445 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1446 
1447 	if (db->db_state == DB_CACHED &&
1448 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1449 		mutex_exit(&db->db_mtx);
1450 		(void) dbuf_dirty(db, tx);
1451 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1452 		VERIFY(arc_buf_remove_ref(buf, db));
1453 		xuio_stat_wbuf_copied();
1454 		return;
1455 	}
1456 
1457 	xuio_stat_wbuf_nocopy();
1458 	if (db->db_state == DB_CACHED) {
1459 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1460 
1461 		ASSERT(db->db_buf != NULL);
1462 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1463 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1464 			if (!arc_released(db->db_buf)) {
1465 				ASSERT(dr->dt.dl.dr_override_state ==
1466 				    DR_OVERRIDDEN);
1467 				arc_release(db->db_buf, db);
1468 			}
1469 			dr->dt.dl.dr_data = buf;
1470 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1471 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1472 			arc_release(db->db_buf, db);
1473 			VERIFY(arc_buf_remove_ref(db->db_buf, db));
1474 		}
1475 		db->db_buf = NULL;
1476 	}
1477 	ASSERT(db->db_buf == NULL);
1478 	dbuf_set_data(db, buf);
1479 	db->db_state = DB_FILL;
1480 	mutex_exit(&db->db_mtx);
1481 	(void) dbuf_dirty(db, tx);
1482 	dbuf_fill_done(db, tx);
1483 }
1484 
1485 /*
1486  * "Clear" the contents of this dbuf.  This will mark the dbuf
1487  * EVICTING and clear *most* of its references.  Unfortunetely,
1488  * when we are not holding the dn_dbufs_mtx, we can't clear the
1489  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1490  * in this case.  For callers from the DMU we will usually see:
1491  *	dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1492  * For the arc callback, we will usually see:
1493  *	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1494  * Sometimes, though, we will get a mix of these two:
1495  *	DMU: dbuf_clear()->arc_buf_evict()
1496  *	ARC: dbuf_do_evict()->dbuf_destroy()
1497  */
1498 void
1499 dbuf_clear(dmu_buf_impl_t *db)
1500 {
1501 	dnode_t *dn;
1502 	dmu_buf_impl_t *parent = db->db_parent;
1503 	dmu_buf_impl_t *dndb;
1504 	int dbuf_gone = FALSE;
1505 
1506 	ASSERT(MUTEX_HELD(&db->db_mtx));
1507 	ASSERT(refcount_is_zero(&db->db_holds));
1508 
1509 	dbuf_evict_user(db);
1510 
1511 	if (db->db_state == DB_CACHED) {
1512 		ASSERT(db->db.db_data != NULL);
1513 		if (db->db_blkid == DMU_BONUS_BLKID) {
1514 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1515 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1516 		}
1517 		db->db.db_data = NULL;
1518 		db->db_state = DB_UNCACHED;
1519 	}
1520 
1521 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1522 	ASSERT(db->db_data_pending == NULL);
1523 
1524 	db->db_state = DB_EVICTING;
1525 	db->db_blkptr = NULL;
1526 
1527 	DB_DNODE_ENTER(db);
1528 	dn = DB_DNODE(db);
1529 	dndb = dn->dn_dbuf;
1530 	if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1531 		list_remove(&dn->dn_dbufs, db);
1532 		(void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1533 		membar_producer();
1534 		DB_DNODE_EXIT(db);
1535 		/*
1536 		 * Decrementing the dbuf count means that the hold corresponding
1537 		 * to the removed dbuf is no longer discounted in dnode_move(),
1538 		 * so the dnode cannot be moved until after we release the hold.
1539 		 * The membar_producer() ensures visibility of the decremented
1540 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1541 		 * release any lock.
1542 		 */
1543 		dnode_rele(dn, db);
1544 		db->db_dnode_handle = NULL;
1545 	} else {
1546 		DB_DNODE_EXIT(db);
1547 	}
1548 
1549 	if (db->db_buf)
1550 		dbuf_gone = arc_buf_evict(db->db_buf);
1551 
1552 	if (!dbuf_gone)
1553 		mutex_exit(&db->db_mtx);
1554 
1555 	/*
1556 	 * If this dbuf is referenced from an indirect dbuf,
1557 	 * decrement the ref count on the indirect dbuf.
1558 	 */
1559 	if (parent && parent != dndb)
1560 		dbuf_rele(parent, db);
1561 }
1562 
1563 static int
1564 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1565     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1566 {
1567 	int nlevels, epbs;
1568 
1569 	*parentp = NULL;
1570 	*bpp = NULL;
1571 
1572 	ASSERT(blkid != DMU_BONUS_BLKID);
1573 
1574 	if (blkid == DMU_SPILL_BLKID) {
1575 		mutex_enter(&dn->dn_mtx);
1576 		if (dn->dn_have_spill &&
1577 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1578 			*bpp = &dn->dn_phys->dn_spill;
1579 		else
1580 			*bpp = NULL;
1581 		dbuf_add_ref(dn->dn_dbuf, NULL);
1582 		*parentp = dn->dn_dbuf;
1583 		mutex_exit(&dn->dn_mtx);
1584 		return (0);
1585 	}
1586 
1587 	if (dn->dn_phys->dn_nlevels == 0)
1588 		nlevels = 1;
1589 	else
1590 		nlevels = dn->dn_phys->dn_nlevels;
1591 
1592 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1593 
1594 	ASSERT3U(level * epbs, <, 64);
1595 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1596 	if (level >= nlevels ||
1597 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1598 		/* the buffer has no parent yet */
1599 		return (SET_ERROR(ENOENT));
1600 	} else if (level < nlevels-1) {
1601 		/* this block is referenced from an indirect block */
1602 		int err = dbuf_hold_impl(dn, level+1,
1603 		    blkid >> epbs, fail_sparse, NULL, parentp);
1604 		if (err)
1605 			return (err);
1606 		err = dbuf_read(*parentp, NULL,
1607 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1608 		if (err) {
1609 			dbuf_rele(*parentp, NULL);
1610 			*parentp = NULL;
1611 			return (err);
1612 		}
1613 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1614 		    (blkid & ((1ULL << epbs) - 1));
1615 		return (0);
1616 	} else {
1617 		/* the block is referenced from the dnode */
1618 		ASSERT3U(level, ==, nlevels-1);
1619 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1620 		    blkid < dn->dn_phys->dn_nblkptr);
1621 		if (dn->dn_dbuf) {
1622 			dbuf_add_ref(dn->dn_dbuf, NULL);
1623 			*parentp = dn->dn_dbuf;
1624 		}
1625 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1626 		return (0);
1627 	}
1628 }
1629 
1630 static dmu_buf_impl_t *
1631 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1632     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1633 {
1634 	objset_t *os = dn->dn_objset;
1635 	dmu_buf_impl_t *db, *odb;
1636 
1637 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1638 	ASSERT(dn->dn_type != DMU_OT_NONE);
1639 
1640 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1641 
1642 	db->db_objset = os;
1643 	db->db.db_object = dn->dn_object;
1644 	db->db_level = level;
1645 	db->db_blkid = blkid;
1646 	db->db_last_dirty = NULL;
1647 	db->db_dirtycnt = 0;
1648 	db->db_dnode_handle = dn->dn_handle;
1649 	db->db_parent = parent;
1650 	db->db_blkptr = blkptr;
1651 
1652 	db->db_user_ptr = NULL;
1653 	db->db_user_data_ptr_ptr = NULL;
1654 	db->db_evict_func = NULL;
1655 	db->db_immediate_evict = 0;
1656 	db->db_freed_in_flight = 0;
1657 
1658 	if (blkid == DMU_BONUS_BLKID) {
1659 		ASSERT3P(parent, ==, dn->dn_dbuf);
1660 		db->db.db_size = DN_MAX_BONUSLEN -
1661 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1662 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1663 		db->db.db_offset = DMU_BONUS_BLKID;
1664 		db->db_state = DB_UNCACHED;
1665 		/* the bonus dbuf is not placed in the hash table */
1666 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1667 		return (db);
1668 	} else if (blkid == DMU_SPILL_BLKID) {
1669 		db->db.db_size = (blkptr != NULL) ?
1670 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1671 		db->db.db_offset = 0;
1672 	} else {
1673 		int blocksize =
1674 		    db->db_level ? 1<<dn->dn_indblkshift :  dn->dn_datablksz;
1675 		db->db.db_size = blocksize;
1676 		db->db.db_offset = db->db_blkid * blocksize;
1677 	}
1678 
1679 	/*
1680 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1681 	 * in the hash table *and* added to the dbufs list.
1682 	 * This prevents a possible deadlock with someone
1683 	 * trying to look up this dbuf before its added to the
1684 	 * dn_dbufs list.
1685 	 */
1686 	mutex_enter(&dn->dn_dbufs_mtx);
1687 	db->db_state = DB_EVICTING;
1688 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1689 		/* someone else inserted it first */
1690 		kmem_cache_free(dbuf_cache, db);
1691 		mutex_exit(&dn->dn_dbufs_mtx);
1692 		return (odb);
1693 	}
1694 	list_insert_head(&dn->dn_dbufs, db);
1695 	db->db_state = DB_UNCACHED;
1696 	mutex_exit(&dn->dn_dbufs_mtx);
1697 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1698 
1699 	if (parent && parent != dn->dn_dbuf)
1700 		dbuf_add_ref(parent, db);
1701 
1702 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1703 	    refcount_count(&dn->dn_holds) > 0);
1704 	(void) refcount_add(&dn->dn_holds, db);
1705 	(void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1706 
1707 	dprintf_dbuf(db, "db=%p\n", db);
1708 
1709 	return (db);
1710 }
1711 
1712 static int
1713 dbuf_do_evict(void *private)
1714 {
1715 	arc_buf_t *buf = private;
1716 	dmu_buf_impl_t *db = buf->b_private;
1717 
1718 	if (!MUTEX_HELD(&db->db_mtx))
1719 		mutex_enter(&db->db_mtx);
1720 
1721 	ASSERT(refcount_is_zero(&db->db_holds));
1722 
1723 	if (db->db_state != DB_EVICTING) {
1724 		ASSERT(db->db_state == DB_CACHED);
1725 		DBUF_VERIFY(db);
1726 		db->db_buf = NULL;
1727 		dbuf_evict(db);
1728 	} else {
1729 		mutex_exit(&db->db_mtx);
1730 		dbuf_destroy(db);
1731 	}
1732 	return (0);
1733 }
1734 
1735 static void
1736 dbuf_destroy(dmu_buf_impl_t *db)
1737 {
1738 	ASSERT(refcount_is_zero(&db->db_holds));
1739 
1740 	if (db->db_blkid != DMU_BONUS_BLKID) {
1741 		/*
1742 		 * If this dbuf is still on the dn_dbufs list,
1743 		 * remove it from that list.
1744 		 */
1745 		if (db->db_dnode_handle != NULL) {
1746 			dnode_t *dn;
1747 
1748 			DB_DNODE_ENTER(db);
1749 			dn = DB_DNODE(db);
1750 			mutex_enter(&dn->dn_dbufs_mtx);
1751 			list_remove(&dn->dn_dbufs, db);
1752 			(void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1753 			mutex_exit(&dn->dn_dbufs_mtx);
1754 			DB_DNODE_EXIT(db);
1755 			/*
1756 			 * Decrementing the dbuf count means that the hold
1757 			 * corresponding to the removed dbuf is no longer
1758 			 * discounted in dnode_move(), so the dnode cannot be
1759 			 * moved until after we release the hold.
1760 			 */
1761 			dnode_rele(dn, db);
1762 			db->db_dnode_handle = NULL;
1763 		}
1764 		dbuf_hash_remove(db);
1765 	}
1766 	db->db_parent = NULL;
1767 	db->db_buf = NULL;
1768 
1769 	ASSERT(!list_link_active(&db->db_link));
1770 	ASSERT(db->db.db_data == NULL);
1771 	ASSERT(db->db_hash_next == NULL);
1772 	ASSERT(db->db_blkptr == NULL);
1773 	ASSERT(db->db_data_pending == NULL);
1774 
1775 	kmem_cache_free(dbuf_cache, db);
1776 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1777 }
1778 
1779 void
1780 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1781 {
1782 	dmu_buf_impl_t *db = NULL;
1783 	blkptr_t *bp = NULL;
1784 
1785 	ASSERT(blkid != DMU_BONUS_BLKID);
1786 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1787 
1788 	if (dnode_block_freed(dn, blkid))
1789 		return;
1790 
1791 	/* dbuf_find() returns with db_mtx held */
1792 	if (db = dbuf_find(dn, 0, blkid)) {
1793 		/*
1794 		 * This dbuf is already in the cache.  We assume that
1795 		 * it is already CACHED, or else about to be either
1796 		 * read or filled.
1797 		 */
1798 		mutex_exit(&db->db_mtx);
1799 		return;
1800 	}
1801 
1802 	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1803 		if (bp && !BP_IS_HOLE(bp)) {
1804 			int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1805 			    ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1806 			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1807 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1808 			zbookmark_t zb;
1809 
1810 			SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1811 			    dn->dn_object, 0, blkid);
1812 
1813 			(void) arc_read(NULL, dn->dn_objset->os_spa,
1814 			    bp, NULL, NULL, priority,
1815 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1816 			    &aflags, &zb);
1817 		}
1818 		if (db)
1819 			dbuf_rele(db, NULL);
1820 	}
1821 }
1822 
1823 /*
1824  * Returns with db_holds incremented, and db_mtx not held.
1825  * Note: dn_struct_rwlock must be held.
1826  */
1827 int
1828 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1829     void *tag, dmu_buf_impl_t **dbp)
1830 {
1831 	dmu_buf_impl_t *db, *parent = NULL;
1832 
1833 	ASSERT(blkid != DMU_BONUS_BLKID);
1834 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1835 	ASSERT3U(dn->dn_nlevels, >, level);
1836 
1837 	*dbp = NULL;
1838 top:
1839 	/* dbuf_find() returns with db_mtx held */
1840 	db = dbuf_find(dn, level, blkid);
1841 
1842 	if (db == NULL) {
1843 		blkptr_t *bp = NULL;
1844 		int err;
1845 
1846 		ASSERT3P(parent, ==, NULL);
1847 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1848 		if (fail_sparse) {
1849 			if (err == 0 && bp && BP_IS_HOLE(bp))
1850 				err = SET_ERROR(ENOENT);
1851 			if (err) {
1852 				if (parent)
1853 					dbuf_rele(parent, NULL);
1854 				return (err);
1855 			}
1856 		}
1857 		if (err && err != ENOENT)
1858 			return (err);
1859 		db = dbuf_create(dn, level, blkid, parent, bp);
1860 	}
1861 
1862 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1863 		arc_buf_add_ref(db->db_buf, db);
1864 		if (db->db_buf->b_data == NULL) {
1865 			dbuf_clear(db);
1866 			if (parent) {
1867 				dbuf_rele(parent, NULL);
1868 				parent = NULL;
1869 			}
1870 			goto top;
1871 		}
1872 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1873 	}
1874 
1875 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1876 
1877 	/*
1878 	 * If this buffer is currently syncing out, and we are are
1879 	 * still referencing it from db_data, we need to make a copy
1880 	 * of it in case we decide we want to dirty it again in this txg.
1881 	 */
1882 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1883 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
1884 	    db->db_state == DB_CACHED && db->db_data_pending) {
1885 		dbuf_dirty_record_t *dr = db->db_data_pending;
1886 
1887 		if (dr->dt.dl.dr_data == db->db_buf) {
1888 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1889 
1890 			dbuf_set_data(db,
1891 			    arc_buf_alloc(dn->dn_objset->os_spa,
1892 			    db->db.db_size, db, type));
1893 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1894 			    db->db.db_size);
1895 		}
1896 	}
1897 
1898 	(void) refcount_add(&db->db_holds, tag);
1899 	dbuf_update_data(db);
1900 	DBUF_VERIFY(db);
1901 	mutex_exit(&db->db_mtx);
1902 
1903 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
1904 	if (parent)
1905 		dbuf_rele(parent, NULL);
1906 
1907 	ASSERT3P(DB_DNODE(db), ==, dn);
1908 	ASSERT3U(db->db_blkid, ==, blkid);
1909 	ASSERT3U(db->db_level, ==, level);
1910 	*dbp = db;
1911 
1912 	return (0);
1913 }
1914 
1915 dmu_buf_impl_t *
1916 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1917 {
1918 	dmu_buf_impl_t *db;
1919 	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1920 	return (err ? NULL : db);
1921 }
1922 
1923 dmu_buf_impl_t *
1924 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1925 {
1926 	dmu_buf_impl_t *db;
1927 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1928 	return (err ? NULL : db);
1929 }
1930 
1931 void
1932 dbuf_create_bonus(dnode_t *dn)
1933 {
1934 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1935 
1936 	ASSERT(dn->dn_bonus == NULL);
1937 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1938 }
1939 
1940 int
1941 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1942 {
1943 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1944 	dnode_t *dn;
1945 
1946 	if (db->db_blkid != DMU_SPILL_BLKID)
1947 		return (SET_ERROR(ENOTSUP));
1948 	if (blksz == 0)
1949 		blksz = SPA_MINBLOCKSIZE;
1950 	if (blksz > SPA_MAXBLOCKSIZE)
1951 		blksz = SPA_MAXBLOCKSIZE;
1952 	else
1953 		blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1954 
1955 	DB_DNODE_ENTER(db);
1956 	dn = DB_DNODE(db);
1957 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1958 	dbuf_new_size(db, blksz, tx);
1959 	rw_exit(&dn->dn_struct_rwlock);
1960 	DB_DNODE_EXIT(db);
1961 
1962 	return (0);
1963 }
1964 
1965 void
1966 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
1967 {
1968 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
1969 }
1970 
1971 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1972 void
1973 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1974 {
1975 	int64_t holds = refcount_add(&db->db_holds, tag);
1976 	ASSERT(holds > 1);
1977 }
1978 
1979 /*
1980  * If you call dbuf_rele() you had better not be referencing the dnode handle
1981  * unless you have some other direct or indirect hold on the dnode. (An indirect
1982  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
1983  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
1984  * dnode's parent dbuf evicting its dnode handles.
1985  */
1986 #pragma weak dmu_buf_rele = dbuf_rele
1987 void
1988 dbuf_rele(dmu_buf_impl_t *db, void *tag)
1989 {
1990 	mutex_enter(&db->db_mtx);
1991 	dbuf_rele_and_unlock(db, tag);
1992 }
1993 
1994 /*
1995  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
1996  * db_dirtycnt and db_holds to be updated atomically.
1997  */
1998 void
1999 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2000 {
2001 	int64_t holds;
2002 
2003 	ASSERT(MUTEX_HELD(&db->db_mtx));
2004 	DBUF_VERIFY(db);
2005 
2006 	/*
2007 	 * Remove the reference to the dbuf before removing its hold on the
2008 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2009 	 * buffer has a corresponding dnode hold.
2010 	 */
2011 	holds = refcount_remove(&db->db_holds, tag);
2012 	ASSERT(holds >= 0);
2013 
2014 	/*
2015 	 * We can't freeze indirects if there is a possibility that they
2016 	 * may be modified in the current syncing context.
2017 	 */
2018 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2019 		arc_buf_freeze(db->db_buf);
2020 
2021 	if (holds == db->db_dirtycnt &&
2022 	    db->db_level == 0 && db->db_immediate_evict)
2023 		dbuf_evict_user(db);
2024 
2025 	if (holds == 0) {
2026 		if (db->db_blkid == DMU_BONUS_BLKID) {
2027 			mutex_exit(&db->db_mtx);
2028 
2029 			/*
2030 			 * If the dnode moves here, we cannot cross this barrier
2031 			 * until the move completes.
2032 			 */
2033 			DB_DNODE_ENTER(db);
2034 			(void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2035 			DB_DNODE_EXIT(db);
2036 			/*
2037 			 * The bonus buffer's dnode hold is no longer discounted
2038 			 * in dnode_move(). The dnode cannot move until after
2039 			 * the dnode_rele().
2040 			 */
2041 			dnode_rele(DB_DNODE(db), db);
2042 		} else if (db->db_buf == NULL) {
2043 			/*
2044 			 * This is a special case: we never associated this
2045 			 * dbuf with any data allocated from the ARC.
2046 			 */
2047 			ASSERT(db->db_state == DB_UNCACHED ||
2048 			    db->db_state == DB_NOFILL);
2049 			dbuf_evict(db);
2050 		} else if (arc_released(db->db_buf)) {
2051 			arc_buf_t *buf = db->db_buf;
2052 			/*
2053 			 * This dbuf has anonymous data associated with it.
2054 			 */
2055 			dbuf_set_data(db, NULL);
2056 			VERIFY(arc_buf_remove_ref(buf, db));
2057 			dbuf_evict(db);
2058 		} else {
2059 			VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2060 
2061 			/*
2062 			 * A dbuf will be eligible for eviction if either the
2063 			 * 'primarycache' property is set or a duplicate
2064 			 * copy of this buffer is already cached in the arc.
2065 			 *
2066 			 * In the case of the 'primarycache' a buffer
2067 			 * is considered for eviction if it matches the
2068 			 * criteria set in the property.
2069 			 *
2070 			 * To decide if our buffer is considered a
2071 			 * duplicate, we must call into the arc to determine
2072 			 * if multiple buffers are referencing the same
2073 			 * block on-disk. If so, then we simply evict
2074 			 * ourselves.
2075 			 */
2076 			if (!DBUF_IS_CACHEABLE(db) ||
2077 			    arc_buf_eviction_needed(db->db_buf))
2078 				dbuf_clear(db);
2079 			else
2080 				mutex_exit(&db->db_mtx);
2081 		}
2082 	} else {
2083 		mutex_exit(&db->db_mtx);
2084 	}
2085 }
2086 
2087 #pragma weak dmu_buf_refcount = dbuf_refcount
2088 uint64_t
2089 dbuf_refcount(dmu_buf_impl_t *db)
2090 {
2091 	return (refcount_count(&db->db_holds));
2092 }
2093 
2094 void *
2095 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2096     dmu_buf_evict_func_t *evict_func)
2097 {
2098 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2099 	    user_data_ptr_ptr, evict_func));
2100 }
2101 
2102 void *
2103 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2104     dmu_buf_evict_func_t *evict_func)
2105 {
2106 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2107 
2108 	db->db_immediate_evict = TRUE;
2109 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2110 	    user_data_ptr_ptr, evict_func));
2111 }
2112 
2113 void *
2114 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2115     void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2116 {
2117 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2118 	ASSERT(db->db_level == 0);
2119 
2120 	ASSERT((user_ptr == NULL) == (evict_func == NULL));
2121 
2122 	mutex_enter(&db->db_mtx);
2123 
2124 	if (db->db_user_ptr == old_user_ptr) {
2125 		db->db_user_ptr = user_ptr;
2126 		db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2127 		db->db_evict_func = evict_func;
2128 
2129 		dbuf_update_data(db);
2130 	} else {
2131 		old_user_ptr = db->db_user_ptr;
2132 	}
2133 
2134 	mutex_exit(&db->db_mtx);
2135 	return (old_user_ptr);
2136 }
2137 
2138 void *
2139 dmu_buf_get_user(dmu_buf_t *db_fake)
2140 {
2141 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2142 	ASSERT(!refcount_is_zero(&db->db_holds));
2143 
2144 	return (db->db_user_ptr);
2145 }
2146 
2147 boolean_t
2148 dmu_buf_freeable(dmu_buf_t *dbuf)
2149 {
2150 	boolean_t res = B_FALSE;
2151 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2152 
2153 	if (db->db_blkptr)
2154 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2155 		    db->db_blkptr, db->db_blkptr->blk_birth);
2156 
2157 	return (res);
2158 }
2159 
2160 blkptr_t *
2161 dmu_buf_get_blkptr(dmu_buf_t *db)
2162 {
2163 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2164 	return (dbi->db_blkptr);
2165 }
2166 
2167 static void
2168 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2169 {
2170 	/* ASSERT(dmu_tx_is_syncing(tx) */
2171 	ASSERT(MUTEX_HELD(&db->db_mtx));
2172 
2173 	if (db->db_blkptr != NULL)
2174 		return;
2175 
2176 	if (db->db_blkid == DMU_SPILL_BLKID) {
2177 		db->db_blkptr = &dn->dn_phys->dn_spill;
2178 		BP_ZERO(db->db_blkptr);
2179 		return;
2180 	}
2181 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2182 		/*
2183 		 * This buffer was allocated at a time when there was
2184 		 * no available blkptrs from the dnode, or it was
2185 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2186 		 */
2187 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2188 		ASSERT(db->db_parent == NULL);
2189 		db->db_parent = dn->dn_dbuf;
2190 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2191 		DBUF_VERIFY(db);
2192 	} else {
2193 		dmu_buf_impl_t *parent = db->db_parent;
2194 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2195 
2196 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2197 		if (parent == NULL) {
2198 			mutex_exit(&db->db_mtx);
2199 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2200 			(void) dbuf_hold_impl(dn, db->db_level+1,
2201 			    db->db_blkid >> epbs, FALSE, db, &parent);
2202 			rw_exit(&dn->dn_struct_rwlock);
2203 			mutex_enter(&db->db_mtx);
2204 			db->db_parent = parent;
2205 		}
2206 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2207 		    (db->db_blkid & ((1ULL << epbs) - 1));
2208 		DBUF_VERIFY(db);
2209 	}
2210 }
2211 
2212 static void
2213 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2214 {
2215 	dmu_buf_impl_t *db = dr->dr_dbuf;
2216 	dnode_t *dn;
2217 	zio_t *zio;
2218 
2219 	ASSERT(dmu_tx_is_syncing(tx));
2220 
2221 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2222 
2223 	mutex_enter(&db->db_mtx);
2224 
2225 	ASSERT(db->db_level > 0);
2226 	DBUF_VERIFY(db);
2227 
2228 	if (db->db_buf == NULL) {
2229 		mutex_exit(&db->db_mtx);
2230 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2231 		mutex_enter(&db->db_mtx);
2232 	}
2233 	ASSERT3U(db->db_state, ==, DB_CACHED);
2234 	ASSERT(db->db_buf != NULL);
2235 
2236 	DB_DNODE_ENTER(db);
2237 	dn = DB_DNODE(db);
2238 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2239 	dbuf_check_blkptr(dn, db);
2240 	DB_DNODE_EXIT(db);
2241 
2242 	db->db_data_pending = dr;
2243 
2244 	mutex_exit(&db->db_mtx);
2245 	dbuf_write(dr, db->db_buf, tx);
2246 
2247 	zio = dr->dr_zio;
2248 	mutex_enter(&dr->dt.di.dr_mtx);
2249 	dbuf_sync_list(&dr->dt.di.dr_children, tx);
2250 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2251 	mutex_exit(&dr->dt.di.dr_mtx);
2252 	zio_nowait(zio);
2253 }
2254 
2255 static void
2256 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2257 {
2258 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2259 	dmu_buf_impl_t *db = dr->dr_dbuf;
2260 	dnode_t *dn;
2261 	objset_t *os;
2262 	uint64_t txg = tx->tx_txg;
2263 
2264 	ASSERT(dmu_tx_is_syncing(tx));
2265 
2266 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2267 
2268 	mutex_enter(&db->db_mtx);
2269 	/*
2270 	 * To be synced, we must be dirtied.  But we
2271 	 * might have been freed after the dirty.
2272 	 */
2273 	if (db->db_state == DB_UNCACHED) {
2274 		/* This buffer has been freed since it was dirtied */
2275 		ASSERT(db->db.db_data == NULL);
2276 	} else if (db->db_state == DB_FILL) {
2277 		/* This buffer was freed and is now being re-filled */
2278 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2279 	} else {
2280 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2281 	}
2282 	DBUF_VERIFY(db);
2283 
2284 	DB_DNODE_ENTER(db);
2285 	dn = DB_DNODE(db);
2286 
2287 	if (db->db_blkid == DMU_SPILL_BLKID) {
2288 		mutex_enter(&dn->dn_mtx);
2289 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2290 		mutex_exit(&dn->dn_mtx);
2291 	}
2292 
2293 	/*
2294 	 * If this is a bonus buffer, simply copy the bonus data into the
2295 	 * dnode.  It will be written out when the dnode is synced (and it
2296 	 * will be synced, since it must have been dirty for dbuf_sync to
2297 	 * be called).
2298 	 */
2299 	if (db->db_blkid == DMU_BONUS_BLKID) {
2300 		dbuf_dirty_record_t **drp;
2301 
2302 		ASSERT(*datap != NULL);
2303 		ASSERT0(db->db_level);
2304 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2305 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2306 		DB_DNODE_EXIT(db);
2307 
2308 		if (*datap != db->db.db_data) {
2309 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
2310 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2311 		}
2312 		db->db_data_pending = NULL;
2313 		drp = &db->db_last_dirty;
2314 		while (*drp != dr)
2315 			drp = &(*drp)->dr_next;
2316 		ASSERT(dr->dr_next == NULL);
2317 		ASSERT(dr->dr_dbuf == db);
2318 		*drp = dr->dr_next;
2319 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
2320 		ASSERT(db->db_dirtycnt > 0);
2321 		db->db_dirtycnt -= 1;
2322 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2323 		return;
2324 	}
2325 
2326 	os = dn->dn_objset;
2327 
2328 	/*
2329 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
2330 	 * operation to sneak in. As a result, we need to ensure that we
2331 	 * don't check the dr_override_state until we have returned from
2332 	 * dbuf_check_blkptr.
2333 	 */
2334 	dbuf_check_blkptr(dn, db);
2335 
2336 	/*
2337 	 * If this buffer is in the middle of an immediate write,
2338 	 * wait for the synchronous IO to complete.
2339 	 */
2340 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2341 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2342 		cv_wait(&db->db_changed, &db->db_mtx);
2343 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2344 	}
2345 
2346 	if (db->db_state != DB_NOFILL &&
2347 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2348 	    refcount_count(&db->db_holds) > 1 &&
2349 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2350 	    *datap == db->db_buf) {
2351 		/*
2352 		 * If this buffer is currently "in use" (i.e., there
2353 		 * are active holds and db_data still references it),
2354 		 * then make a copy before we start the write so that
2355 		 * any modifications from the open txg will not leak
2356 		 * into this write.
2357 		 *
2358 		 * NOTE: this copy does not need to be made for
2359 		 * objects only modified in the syncing context (e.g.
2360 		 * DNONE_DNODE blocks).
2361 		 */
2362 		int blksz = arc_buf_size(*datap);
2363 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2364 		*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2365 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
2366 	}
2367 	db->db_data_pending = dr;
2368 
2369 	mutex_exit(&db->db_mtx);
2370 
2371 	dbuf_write(dr, *datap, tx);
2372 
2373 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2374 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2375 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2376 		DB_DNODE_EXIT(db);
2377 	} else {
2378 		/*
2379 		 * Although zio_nowait() does not "wait for an IO", it does
2380 		 * initiate the IO. If this is an empty write it seems plausible
2381 		 * that the IO could actually be completed before the nowait
2382 		 * returns. We need to DB_DNODE_EXIT() first in case
2383 		 * zio_nowait() invalidates the dbuf.
2384 		 */
2385 		DB_DNODE_EXIT(db);
2386 		zio_nowait(dr->dr_zio);
2387 	}
2388 }
2389 
2390 void
2391 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2392 {
2393 	dbuf_dirty_record_t *dr;
2394 
2395 	while (dr = list_head(list)) {
2396 		if (dr->dr_zio != NULL) {
2397 			/*
2398 			 * If we find an already initialized zio then we
2399 			 * are processing the meta-dnode, and we have finished.
2400 			 * The dbufs for all dnodes are put back on the list
2401 			 * during processing, so that we can zio_wait()
2402 			 * these IOs after initiating all child IOs.
2403 			 */
2404 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2405 			    DMU_META_DNODE_OBJECT);
2406 			break;
2407 		}
2408 		list_remove(list, dr);
2409 		if (dr->dr_dbuf->db_level > 0)
2410 			dbuf_sync_indirect(dr, tx);
2411 		else
2412 			dbuf_sync_leaf(dr, tx);
2413 	}
2414 }
2415 
2416 /* ARGSUSED */
2417 static void
2418 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2419 {
2420 	dmu_buf_impl_t *db = vdb;
2421 	dnode_t *dn;
2422 	blkptr_t *bp = zio->io_bp;
2423 	blkptr_t *bp_orig = &zio->io_bp_orig;
2424 	spa_t *spa = zio->io_spa;
2425 	int64_t delta;
2426 	uint64_t fill = 0;
2427 	int i;
2428 
2429 	ASSERT(db->db_blkptr == bp);
2430 
2431 	DB_DNODE_ENTER(db);
2432 	dn = DB_DNODE(db);
2433 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2434 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2435 	zio->io_prev_space_delta = delta;
2436 
2437 	if (BP_IS_HOLE(bp)) {
2438 		ASSERT(bp->blk_fill == 0);
2439 		DB_DNODE_EXIT(db);
2440 		return;
2441 	}
2442 
2443 	ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2444 	    BP_GET_TYPE(bp) == dn->dn_type) ||
2445 	    (db->db_blkid == DMU_SPILL_BLKID &&
2446 	    BP_GET_TYPE(bp) == dn->dn_bonustype));
2447 	ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2448 
2449 	mutex_enter(&db->db_mtx);
2450 
2451 #ifdef ZFS_DEBUG
2452 	if (db->db_blkid == DMU_SPILL_BLKID) {
2453 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2454 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2455 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2456 	}
2457 #endif
2458 
2459 	if (db->db_level == 0) {
2460 		mutex_enter(&dn->dn_mtx);
2461 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2462 		    db->db_blkid != DMU_SPILL_BLKID)
2463 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2464 		mutex_exit(&dn->dn_mtx);
2465 
2466 		if (dn->dn_type == DMU_OT_DNODE) {
2467 			dnode_phys_t *dnp = db->db.db_data;
2468 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2469 			    i--, dnp++) {
2470 				if (dnp->dn_type != DMU_OT_NONE)
2471 					fill++;
2472 			}
2473 		} else {
2474 			fill = 1;
2475 		}
2476 	} else {
2477 		blkptr_t *ibp = db->db.db_data;
2478 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2479 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2480 			if (BP_IS_HOLE(ibp))
2481 				continue;
2482 			fill += ibp->blk_fill;
2483 		}
2484 	}
2485 	DB_DNODE_EXIT(db);
2486 
2487 	bp->blk_fill = fill;
2488 
2489 	mutex_exit(&db->db_mtx);
2490 }
2491 
2492 /* ARGSUSED */
2493 static void
2494 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2495 {
2496 	dmu_buf_impl_t *db = vdb;
2497 	blkptr_t *bp = zio->io_bp;
2498 	blkptr_t *bp_orig = &zio->io_bp_orig;
2499 	uint64_t txg = zio->io_txg;
2500 	dbuf_dirty_record_t **drp, *dr;
2501 
2502 	ASSERT0(zio->io_error);
2503 	ASSERT(db->db_blkptr == bp);
2504 
2505 	/*
2506 	 * For nopwrites and rewrites we ensure that the bp matches our
2507 	 * original and bypass all the accounting.
2508 	 */
2509 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2510 		ASSERT(BP_EQUAL(bp, bp_orig));
2511 	} else {
2512 		objset_t *os;
2513 		dsl_dataset_t *ds;
2514 		dmu_tx_t *tx;
2515 
2516 		DB_GET_OBJSET(&os, db);
2517 		ds = os->os_dsl_dataset;
2518 		tx = os->os_synctx;
2519 
2520 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2521 		dsl_dataset_block_born(ds, bp, tx);
2522 	}
2523 
2524 	mutex_enter(&db->db_mtx);
2525 
2526 	DBUF_VERIFY(db);
2527 
2528 	drp = &db->db_last_dirty;
2529 	while ((dr = *drp) != db->db_data_pending)
2530 		drp = &dr->dr_next;
2531 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2532 	ASSERT(dr->dr_txg == txg);
2533 	ASSERT(dr->dr_dbuf == db);
2534 	ASSERT(dr->dr_next == NULL);
2535 	*drp = dr->dr_next;
2536 
2537 #ifdef ZFS_DEBUG
2538 	if (db->db_blkid == DMU_SPILL_BLKID) {
2539 		dnode_t *dn;
2540 
2541 		DB_DNODE_ENTER(db);
2542 		dn = DB_DNODE(db);
2543 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2544 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2545 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2546 		DB_DNODE_EXIT(db);
2547 	}
2548 #endif
2549 
2550 	if (db->db_level == 0) {
2551 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2552 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2553 		if (db->db_state != DB_NOFILL) {
2554 			if (dr->dt.dl.dr_data != db->db_buf)
2555 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2556 				    db));
2557 			else if (!arc_released(db->db_buf))
2558 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2559 		}
2560 	} else {
2561 		dnode_t *dn;
2562 
2563 		DB_DNODE_ENTER(db);
2564 		dn = DB_DNODE(db);
2565 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2566 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2567 		if (!BP_IS_HOLE(db->db_blkptr)) {
2568 			int epbs =
2569 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2570 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2571 			    db->db.db_size);
2572 			ASSERT3U(dn->dn_phys->dn_maxblkid
2573 			    >> (db->db_level * epbs), >=, db->db_blkid);
2574 			arc_set_callback(db->db_buf, dbuf_do_evict, db);
2575 		}
2576 		DB_DNODE_EXIT(db);
2577 		mutex_destroy(&dr->dt.di.dr_mtx);
2578 		list_destroy(&dr->dt.di.dr_children);
2579 	}
2580 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
2581 
2582 	cv_broadcast(&db->db_changed);
2583 	ASSERT(db->db_dirtycnt > 0);
2584 	db->db_dirtycnt -= 1;
2585 	db->db_data_pending = NULL;
2586 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2587 }
2588 
2589 static void
2590 dbuf_write_nofill_ready(zio_t *zio)
2591 {
2592 	dbuf_write_ready(zio, NULL, zio->io_private);
2593 }
2594 
2595 static void
2596 dbuf_write_nofill_done(zio_t *zio)
2597 {
2598 	dbuf_write_done(zio, NULL, zio->io_private);
2599 }
2600 
2601 static void
2602 dbuf_write_override_ready(zio_t *zio)
2603 {
2604 	dbuf_dirty_record_t *dr = zio->io_private;
2605 	dmu_buf_impl_t *db = dr->dr_dbuf;
2606 
2607 	dbuf_write_ready(zio, NULL, db);
2608 }
2609 
2610 static void
2611 dbuf_write_override_done(zio_t *zio)
2612 {
2613 	dbuf_dirty_record_t *dr = zio->io_private;
2614 	dmu_buf_impl_t *db = dr->dr_dbuf;
2615 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2616 
2617 	mutex_enter(&db->db_mtx);
2618 	if (!BP_EQUAL(zio->io_bp, obp)) {
2619 		if (!BP_IS_HOLE(obp))
2620 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2621 		arc_release(dr->dt.dl.dr_data, db);
2622 	}
2623 	mutex_exit(&db->db_mtx);
2624 
2625 	dbuf_write_done(zio, NULL, db);
2626 }
2627 
2628 static void
2629 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2630 {
2631 	dmu_buf_impl_t *db = dr->dr_dbuf;
2632 	dnode_t *dn;
2633 	objset_t *os;
2634 	dmu_buf_impl_t *parent = db->db_parent;
2635 	uint64_t txg = tx->tx_txg;
2636 	zbookmark_t zb;
2637 	zio_prop_t zp;
2638 	zio_t *zio;
2639 	int wp_flag = 0;
2640 
2641 	DB_DNODE_ENTER(db);
2642 	dn = DB_DNODE(db);
2643 	os = dn->dn_objset;
2644 
2645 	if (db->db_state != DB_NOFILL) {
2646 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2647 			/*
2648 			 * Private object buffers are released here rather
2649 			 * than in dbuf_dirty() since they are only modified
2650 			 * in the syncing context and we don't want the
2651 			 * overhead of making multiple copies of the data.
2652 			 */
2653 			if (BP_IS_HOLE(db->db_blkptr)) {
2654 				arc_buf_thaw(data);
2655 			} else {
2656 				dbuf_release_bp(db);
2657 			}
2658 		}
2659 	}
2660 
2661 	if (parent != dn->dn_dbuf) {
2662 		ASSERT(parent && parent->db_data_pending);
2663 		ASSERT(db->db_level == parent->db_level-1);
2664 		ASSERT(arc_released(parent->db_buf));
2665 		zio = parent->db_data_pending->dr_zio;
2666 	} else {
2667 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2668 		    db->db_blkid != DMU_SPILL_BLKID) ||
2669 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2670 		if (db->db_blkid != DMU_SPILL_BLKID)
2671 			ASSERT3P(db->db_blkptr, ==,
2672 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
2673 		zio = dn->dn_zio;
2674 	}
2675 
2676 	ASSERT(db->db_level == 0 || data == db->db_buf);
2677 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2678 	ASSERT(zio);
2679 
2680 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2681 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2682 	    db->db.db_object, db->db_level, db->db_blkid);
2683 
2684 	if (db->db_blkid == DMU_SPILL_BLKID)
2685 		wp_flag = WP_SPILL;
2686 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2687 
2688 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2689 	DB_DNODE_EXIT(db);
2690 
2691 	if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2692 		ASSERT(db->db_state != DB_NOFILL);
2693 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2694 		    db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2695 		    dbuf_write_override_ready, dbuf_write_override_done, dr,
2696 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2697 		mutex_enter(&db->db_mtx);
2698 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2699 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2700 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2701 		mutex_exit(&db->db_mtx);
2702 	} else if (db->db_state == DB_NOFILL) {
2703 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2704 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2705 		    db->db_blkptr, NULL, db->db.db_size, &zp,
2706 		    dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2707 		    ZIO_PRIORITY_ASYNC_WRITE,
2708 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2709 	} else {
2710 		ASSERT(arc_released(data));
2711 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
2712 		    db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2713 		    DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2714 		    dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
2715 		    ZIO_FLAG_MUSTSUCCEED, &zb);
2716 	}
2717 }
2718