xref: /titanic_41/usr/src/uts/common/fs/zfs/dbuf.c (revision 186d582bd9dbcd38e0aeea49036d47d3426a3536)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/zfs_context.h>
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/spa.h>
35 #include <sys/zio.h>
36 #include <sys/dmu_zfetch.h>
37 
38 static void dbuf_destroy(dmu_buf_impl_t *db);
39 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
40 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
41 
42 /*
43  * Global data structures and functions for the dbuf cache.
44  */
45 static kmem_cache_t *dbuf_cache;
46 
47 /* ARGSUSED */
48 static int
49 dbuf_cons(void *vdb, void *unused, int kmflag)
50 {
51 	dmu_buf_impl_t *db = vdb;
52 	bzero(db, sizeof (dmu_buf_impl_t));
53 
54 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
55 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
56 	refcount_create(&db->db_holds);
57 	return (0);
58 }
59 
60 /* ARGSUSED */
61 static void
62 dbuf_dest(void *vdb, void *unused)
63 {
64 	dmu_buf_impl_t *db = vdb;
65 	mutex_destroy(&db->db_mtx);
66 	cv_destroy(&db->db_changed);
67 	refcount_destroy(&db->db_holds);
68 }
69 
70 /*
71  * dbuf hash table routines
72  */
73 static dbuf_hash_table_t dbuf_hash_table;
74 
75 static uint64_t dbuf_hash_count;
76 
77 static uint64_t
78 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
79 {
80 	uintptr_t osv = (uintptr_t)os;
81 	uint64_t crc = -1ULL;
82 
83 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
84 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
85 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
86 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
87 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
88 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
89 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
90 
91 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
92 
93 	return (crc);
94 }
95 
96 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
97 
98 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
99 	((dbuf)->db.db_object == (obj) &&		\
100 	(dbuf)->db_objset == (os) &&			\
101 	(dbuf)->db_level == (level) &&			\
102 	(dbuf)->db_blkid == (blkid))
103 
104 dmu_buf_impl_t *
105 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
106 {
107 	dbuf_hash_table_t *h = &dbuf_hash_table;
108 	objset_t *os = dn->dn_objset;
109 	uint64_t obj = dn->dn_object;
110 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
111 	uint64_t idx = hv & h->hash_table_mask;
112 	dmu_buf_impl_t *db;
113 
114 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
115 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
116 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
117 			mutex_enter(&db->db_mtx);
118 			if (db->db_state != DB_EVICTING) {
119 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
120 				return (db);
121 			}
122 			mutex_exit(&db->db_mtx);
123 		}
124 	}
125 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
126 	return (NULL);
127 }
128 
129 /*
130  * Insert an entry into the hash table.  If there is already an element
131  * equal to elem in the hash table, then the already existing element
132  * will be returned and the new element will not be inserted.
133  * Otherwise returns NULL.
134  */
135 static dmu_buf_impl_t *
136 dbuf_hash_insert(dmu_buf_impl_t *db)
137 {
138 	dbuf_hash_table_t *h = &dbuf_hash_table;
139 	objset_t *os = db->db_objset;
140 	uint64_t obj = db->db.db_object;
141 	int level = db->db_level;
142 	uint64_t blkid = db->db_blkid;
143 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
144 	uint64_t idx = hv & h->hash_table_mask;
145 	dmu_buf_impl_t *dbf;
146 
147 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
148 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
149 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
150 			mutex_enter(&dbf->db_mtx);
151 			if (dbf->db_state != DB_EVICTING) {
152 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
153 				return (dbf);
154 			}
155 			mutex_exit(&dbf->db_mtx);
156 		}
157 	}
158 
159 	mutex_enter(&db->db_mtx);
160 	db->db_hash_next = h->hash_table[idx];
161 	h->hash_table[idx] = db;
162 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
163 	atomic_add_64(&dbuf_hash_count, 1);
164 
165 	return (NULL);
166 }
167 
168 /*
169  * Remove an entry from the hash table.  This operation will
170  * fail if there are any existing holds on the db.
171  */
172 static void
173 dbuf_hash_remove(dmu_buf_impl_t *db)
174 {
175 	dbuf_hash_table_t *h = &dbuf_hash_table;
176 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
177 	    db->db_level, db->db_blkid);
178 	uint64_t idx = hv & h->hash_table_mask;
179 	dmu_buf_impl_t *dbf, **dbp;
180 
181 	/*
182 	 * We musn't hold db_mtx to maintin lock ordering:
183 	 * DBUF_HASH_MUTEX > db_mtx.
184 	 */
185 	ASSERT(refcount_is_zero(&db->db_holds));
186 	ASSERT(db->db_state == DB_EVICTING);
187 	ASSERT(!MUTEX_HELD(&db->db_mtx));
188 
189 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
190 	dbp = &h->hash_table[idx];
191 	while ((dbf = *dbp) != db) {
192 		dbp = &dbf->db_hash_next;
193 		ASSERT(dbf != NULL);
194 	}
195 	*dbp = db->db_hash_next;
196 	db->db_hash_next = NULL;
197 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
198 	atomic_add_64(&dbuf_hash_count, -1);
199 }
200 
201 static arc_evict_func_t dbuf_do_evict;
202 
203 static void
204 dbuf_evict_user(dmu_buf_impl_t *db)
205 {
206 	ASSERT(MUTEX_HELD(&db->db_mtx));
207 
208 	if (db->db_level != 0 || db->db_evict_func == NULL)
209 		return;
210 
211 	if (db->db_user_data_ptr_ptr)
212 		*db->db_user_data_ptr_ptr = db->db.db_data;
213 	db->db_evict_func(&db->db, db->db_user_ptr);
214 	db->db_user_ptr = NULL;
215 	db->db_user_data_ptr_ptr = NULL;
216 	db->db_evict_func = NULL;
217 }
218 
219 void
220 dbuf_evict(dmu_buf_impl_t *db)
221 {
222 	ASSERT(MUTEX_HELD(&db->db_mtx));
223 	ASSERT(db->db_buf == NULL);
224 	ASSERT(db->db_data_pending == NULL);
225 
226 	dbuf_clear(db);
227 	dbuf_destroy(db);
228 }
229 
230 void
231 dbuf_init(void)
232 {
233 	uint64_t hsize = 1ULL << 16;
234 	dbuf_hash_table_t *h = &dbuf_hash_table;
235 	int i;
236 
237 	/*
238 	 * The hash table is big enough to fill all of physical memory
239 	 * with an average 4K block size.  The table will take up
240 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
241 	 */
242 	while (hsize * 4096 < physmem * PAGESIZE)
243 		hsize <<= 1;
244 
245 retry:
246 	h->hash_table_mask = hsize - 1;
247 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
248 	if (h->hash_table == NULL) {
249 		/* XXX - we should really return an error instead of assert */
250 		ASSERT(hsize > (1ULL << 10));
251 		hsize >>= 1;
252 		goto retry;
253 	}
254 
255 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
256 	    sizeof (dmu_buf_impl_t),
257 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
258 
259 	for (i = 0; i < DBUF_MUTEXES; i++)
260 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
261 }
262 
263 void
264 dbuf_fini(void)
265 {
266 	dbuf_hash_table_t *h = &dbuf_hash_table;
267 	int i;
268 
269 	for (i = 0; i < DBUF_MUTEXES; i++)
270 		mutex_destroy(&h->hash_mutexes[i]);
271 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
272 	kmem_cache_destroy(dbuf_cache);
273 }
274 
275 /*
276  * Other stuff.
277  */
278 
279 #ifdef ZFS_DEBUG
280 static void
281 dbuf_verify(dmu_buf_impl_t *db)
282 {
283 	dnode_t *dn = db->db_dnode;
284 	dbuf_dirty_record_t *dr;
285 
286 	ASSERT(MUTEX_HELD(&db->db_mtx));
287 
288 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
289 		return;
290 
291 	ASSERT(db->db_objset != NULL);
292 	if (dn == NULL) {
293 		ASSERT(db->db_parent == NULL);
294 		ASSERT(db->db_blkptr == NULL);
295 	} else {
296 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
297 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
298 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
299 		ASSERT(db->db_blkid == DB_BONUS_BLKID ||
300 		    list_head(&dn->dn_dbufs));
301 	}
302 	if (db->db_blkid == DB_BONUS_BLKID) {
303 		ASSERT(dn != NULL);
304 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
305 		ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID);
306 	} else {
307 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
308 	}
309 
310 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
311 		ASSERT(dr->dr_dbuf == db);
312 
313 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
314 		ASSERT(dr->dr_dbuf == db);
315 
316 	/*
317 	 * We can't assert that db_size matches dn_datablksz because it
318 	 * can be momentarily different when another thread is doing
319 	 * dnode_set_blksz().
320 	 */
321 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
322 		dr = db->db_data_pending;
323 		/*
324 		 * It should only be modified in syncing context, so
325 		 * make sure we only have one copy of the data.
326 		 */
327 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
328 	}
329 
330 	/* verify db->db_blkptr */
331 	if (db->db_blkptr) {
332 		if (db->db_parent == dn->dn_dbuf) {
333 			/* db is pointed to by the dnode */
334 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
335 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
336 				ASSERT(db->db_parent == NULL);
337 			else
338 				ASSERT(db->db_parent != NULL);
339 			ASSERT3P(db->db_blkptr, ==,
340 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
341 		} else {
342 			/* db is pointed to by an indirect block */
343 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
344 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
345 			ASSERT3U(db->db_parent->db.db_object, ==,
346 			    db->db.db_object);
347 			/*
348 			 * dnode_grow_indblksz() can make this fail if we don't
349 			 * have the struct_rwlock.  XXX indblksz no longer
350 			 * grows.  safe to do this now?
351 			 */
352 			if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) {
353 				ASSERT3P(db->db_blkptr, ==,
354 				    ((blkptr_t *)db->db_parent->db.db_data +
355 				    db->db_blkid % epb));
356 			}
357 		}
358 	}
359 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
360 	    db->db.db_data && db->db_blkid != DB_BONUS_BLKID &&
361 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
362 		/*
363 		 * If the blkptr isn't set but they have nonzero data,
364 		 * it had better be dirty, otherwise we'll lose that
365 		 * data when we evict this buffer.
366 		 */
367 		if (db->db_dirtycnt == 0) {
368 			uint64_t *buf = db->db.db_data;
369 			int i;
370 
371 			for (i = 0; i < db->db.db_size >> 3; i++) {
372 				ASSERT(buf[i] == 0);
373 			}
374 		}
375 	}
376 }
377 #endif
378 
379 static void
380 dbuf_update_data(dmu_buf_impl_t *db)
381 {
382 	ASSERT(MUTEX_HELD(&db->db_mtx));
383 	if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
384 		ASSERT(!refcount_is_zero(&db->db_holds));
385 		*db->db_user_data_ptr_ptr = db->db.db_data;
386 	}
387 }
388 
389 static void
390 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
391 {
392 	ASSERT(MUTEX_HELD(&db->db_mtx));
393 	ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
394 	db->db_buf = buf;
395 	if (buf != NULL) {
396 		ASSERT(buf->b_data != NULL);
397 		db->db.db_data = buf->b_data;
398 		if (!arc_released(buf))
399 			arc_set_callback(buf, dbuf_do_evict, db);
400 		dbuf_update_data(db);
401 	} else {
402 		dbuf_evict_user(db);
403 		db->db.db_data = NULL;
404 		if (db->db_state != DB_NOFILL)
405 			db->db_state = DB_UNCACHED;
406 	}
407 }
408 
409 /*
410  * Loan out an arc_buf for read.  Return the loaned arc_buf.
411  */
412 arc_buf_t *
413 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
414 {
415 	arc_buf_t *abuf;
416 
417 	mutex_enter(&db->db_mtx);
418 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
419 		int blksz = db->db.db_size;
420 		mutex_exit(&db->db_mtx);
421 		abuf = arc_loan_buf(db->db_dnode->dn_objset->os_spa, blksz);
422 		bcopy(db->db.db_data, abuf->b_data, blksz);
423 	} else {
424 		abuf = db->db_buf;
425 		arc_loan_inuse_buf(abuf, db);
426 		dbuf_set_data(db, NULL);
427 		mutex_exit(&db->db_mtx);
428 	}
429 	return (abuf);
430 }
431 
432 uint64_t
433 dbuf_whichblock(dnode_t *dn, uint64_t offset)
434 {
435 	if (dn->dn_datablkshift) {
436 		return (offset >> dn->dn_datablkshift);
437 	} else {
438 		ASSERT3U(offset, <, dn->dn_datablksz);
439 		return (0);
440 	}
441 }
442 
443 static void
444 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
445 {
446 	dmu_buf_impl_t *db = vdb;
447 
448 	mutex_enter(&db->db_mtx);
449 	ASSERT3U(db->db_state, ==, DB_READ);
450 	/*
451 	 * All reads are synchronous, so we must have a hold on the dbuf
452 	 */
453 	ASSERT(refcount_count(&db->db_holds) > 0);
454 	ASSERT(db->db_buf == NULL);
455 	ASSERT(db->db.db_data == NULL);
456 	if (db->db_level == 0 && db->db_freed_in_flight) {
457 		/* we were freed in flight; disregard any error */
458 		arc_release(buf, db);
459 		bzero(buf->b_data, db->db.db_size);
460 		arc_buf_freeze(buf);
461 		db->db_freed_in_flight = FALSE;
462 		dbuf_set_data(db, buf);
463 		db->db_state = DB_CACHED;
464 	} else if (zio == NULL || zio->io_error == 0) {
465 		dbuf_set_data(db, buf);
466 		db->db_state = DB_CACHED;
467 	} else {
468 		ASSERT(db->db_blkid != DB_BONUS_BLKID);
469 		ASSERT3P(db->db_buf, ==, NULL);
470 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
471 		db->db_state = DB_UNCACHED;
472 	}
473 	cv_broadcast(&db->db_changed);
474 	mutex_exit(&db->db_mtx);
475 	dbuf_rele(db, NULL);
476 }
477 
478 static void
479 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
480 {
481 	dnode_t *dn = db->db_dnode;
482 	zbookmark_t zb;
483 	uint32_t aflags = ARC_NOWAIT;
484 	arc_buf_t *pbuf;
485 
486 	ASSERT(!refcount_is_zero(&db->db_holds));
487 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
488 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
489 	ASSERT(MUTEX_HELD(&db->db_mtx));
490 	ASSERT(db->db_state == DB_UNCACHED);
491 	ASSERT(db->db_buf == NULL);
492 
493 	if (db->db_blkid == DB_BONUS_BLKID) {
494 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
495 
496 		ASSERT3U(bonuslen, <=, db->db.db_size);
497 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
498 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
499 		if (bonuslen < DN_MAX_BONUSLEN)
500 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
501 		if (bonuslen)
502 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
503 		dbuf_update_data(db);
504 		db->db_state = DB_CACHED;
505 		mutex_exit(&db->db_mtx);
506 		return;
507 	}
508 
509 	/*
510 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
511 	 * processes the delete record and clears the bp while we are waiting
512 	 * for the dn_mtx (resulting in a "no" from block_freed).
513 	 */
514 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
515 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
516 	    BP_IS_HOLE(db->db_blkptr)))) {
517 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
518 
519 		dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
520 		    db->db.db_size, db, type));
521 		bzero(db->db.db_data, db->db.db_size);
522 		db->db_state = DB_CACHED;
523 		*flags |= DB_RF_CACHED;
524 		mutex_exit(&db->db_mtx);
525 		return;
526 	}
527 
528 	db->db_state = DB_READ;
529 	mutex_exit(&db->db_mtx);
530 
531 	if (DBUF_IS_L2CACHEABLE(db))
532 		aflags |= ARC_L2CACHE;
533 
534 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
535 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
536 	    db->db.db_object, db->db_level, db->db_blkid);
537 
538 	dbuf_add_ref(db, NULL);
539 	/* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
540 
541 	if (db->db_parent)
542 		pbuf = db->db_parent->db_buf;
543 	else
544 		pbuf = db->db_objset->os_phys_buf;
545 
546 	(void) arc_read(zio, dn->dn_objset->os_spa, db->db_blkptr, pbuf,
547 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
548 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
549 	    &aflags, &zb);
550 	if (aflags & ARC_CACHED)
551 		*flags |= DB_RF_CACHED;
552 }
553 
554 int
555 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
556 {
557 	int err = 0;
558 	int havepzio = (zio != NULL);
559 	int prefetch;
560 
561 	/*
562 	 * We don't have to hold the mutex to check db_state because it
563 	 * can't be freed while we have a hold on the buffer.
564 	 */
565 	ASSERT(!refcount_is_zero(&db->db_holds));
566 
567 	if (db->db_state == DB_NOFILL)
568 		return (EIO);
569 
570 	if ((flags & DB_RF_HAVESTRUCT) == 0)
571 		rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER);
572 
573 	prefetch = db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID &&
574 	    (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL &&
575 	    DBUF_IS_CACHEABLE(db);
576 
577 	mutex_enter(&db->db_mtx);
578 	if (db->db_state == DB_CACHED) {
579 		mutex_exit(&db->db_mtx);
580 		if (prefetch)
581 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
582 			    db->db.db_size, TRUE);
583 		if ((flags & DB_RF_HAVESTRUCT) == 0)
584 			rw_exit(&db->db_dnode->dn_struct_rwlock);
585 	} else if (db->db_state == DB_UNCACHED) {
586 		if (zio == NULL) {
587 			zio = zio_root(db->db_dnode->dn_objset->os_spa,
588 			    NULL, NULL, ZIO_FLAG_CANFAIL);
589 		}
590 		dbuf_read_impl(db, zio, &flags);
591 
592 		/* dbuf_read_impl has dropped db_mtx for us */
593 
594 		if (prefetch)
595 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
596 			    db->db.db_size, flags & DB_RF_CACHED);
597 
598 		if ((flags & DB_RF_HAVESTRUCT) == 0)
599 			rw_exit(&db->db_dnode->dn_struct_rwlock);
600 
601 		if (!havepzio)
602 			err = zio_wait(zio);
603 	} else {
604 		mutex_exit(&db->db_mtx);
605 		if (prefetch)
606 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
607 			    db->db.db_size, TRUE);
608 		if ((flags & DB_RF_HAVESTRUCT) == 0)
609 			rw_exit(&db->db_dnode->dn_struct_rwlock);
610 
611 		mutex_enter(&db->db_mtx);
612 		if ((flags & DB_RF_NEVERWAIT) == 0) {
613 			while (db->db_state == DB_READ ||
614 			    db->db_state == DB_FILL) {
615 				ASSERT(db->db_state == DB_READ ||
616 				    (flags & DB_RF_HAVESTRUCT) == 0);
617 				cv_wait(&db->db_changed, &db->db_mtx);
618 			}
619 			if (db->db_state == DB_UNCACHED)
620 				err = EIO;
621 		}
622 		mutex_exit(&db->db_mtx);
623 	}
624 
625 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
626 	return (err);
627 }
628 
629 static void
630 dbuf_noread(dmu_buf_impl_t *db)
631 {
632 	ASSERT(!refcount_is_zero(&db->db_holds));
633 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
634 	mutex_enter(&db->db_mtx);
635 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
636 		cv_wait(&db->db_changed, &db->db_mtx);
637 	if (db->db_state == DB_UNCACHED) {
638 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
639 
640 		ASSERT(db->db_buf == NULL);
641 		ASSERT(db->db.db_data == NULL);
642 		dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
643 		    db->db.db_size, db, type));
644 		db->db_state = DB_FILL;
645 	} else if (db->db_state == DB_NOFILL) {
646 		dbuf_set_data(db, NULL);
647 	} else {
648 		ASSERT3U(db->db_state, ==, DB_CACHED);
649 	}
650 	mutex_exit(&db->db_mtx);
651 }
652 
653 /*
654  * This is our just-in-time copy function.  It makes a copy of
655  * buffers, that have been modified in a previous transaction
656  * group, before we modify them in the current active group.
657  *
658  * This function is used in two places: when we are dirtying a
659  * buffer for the first time in a txg, and when we are freeing
660  * a range in a dnode that includes this buffer.
661  *
662  * Note that when we are called from dbuf_free_range() we do
663  * not put a hold on the buffer, we just traverse the active
664  * dbuf list for the dnode.
665  */
666 static void
667 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
668 {
669 	dbuf_dirty_record_t *dr = db->db_last_dirty;
670 
671 	ASSERT(MUTEX_HELD(&db->db_mtx));
672 	ASSERT(db->db.db_data != NULL);
673 	ASSERT(db->db_level == 0);
674 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
675 
676 	if (dr == NULL ||
677 	    (dr->dt.dl.dr_data !=
678 	    ((db->db_blkid  == DB_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
679 		return;
680 
681 	/*
682 	 * If the last dirty record for this dbuf has not yet synced
683 	 * and its referencing the dbuf data, either:
684 	 * 	reset the reference to point to a new copy,
685 	 * or (if there a no active holders)
686 	 *	just null out the current db_data pointer.
687 	 */
688 	ASSERT(dr->dr_txg >= txg - 2);
689 	if (db->db_blkid == DB_BONUS_BLKID) {
690 		/* Note that the data bufs here are zio_bufs */
691 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
692 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
693 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
694 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
695 		int size = db->db.db_size;
696 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
697 		dr->dt.dl.dr_data = arc_buf_alloc(
698 		    db->db_dnode->dn_objset->os_spa, size, db, type);
699 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
700 	} else {
701 		dbuf_set_data(db, NULL);
702 	}
703 }
704 
705 void
706 dbuf_unoverride(dbuf_dirty_record_t *dr)
707 {
708 	dmu_buf_impl_t *db = dr->dr_dbuf;
709 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
710 	uint64_t txg = dr->dr_txg;
711 
712 	ASSERT(MUTEX_HELD(&db->db_mtx));
713 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
714 	ASSERT(db->db_level == 0);
715 
716 	if (db->db_blkid == DB_BONUS_BLKID ||
717 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
718 		return;
719 
720 	ASSERT(db->db_data_pending != dr);
721 
722 	/* free this block */
723 	if (!BP_IS_HOLE(bp))
724 		dsl_free(spa_get_dsl(db->db_dnode->dn_objset->os_spa), txg, bp);
725 
726 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
727 	/*
728 	 * Release the already-written buffer, so we leave it in
729 	 * a consistent dirty state.  Note that all callers are
730 	 * modifying the buffer, so they will immediately do
731 	 * another (redundant) arc_release().  Therefore, leave
732 	 * the buf thawed to save the effort of freezing &
733 	 * immediately re-thawing it.
734 	 */
735 	arc_release(dr->dt.dl.dr_data, db);
736 }
737 
738 /*
739  * Evict (if its unreferenced) or clear (if its referenced) any level-0
740  * data blocks in the free range, so that any future readers will find
741  * empty blocks.  Also, if we happen accross any level-1 dbufs in the
742  * range that have not already been marked dirty, mark them dirty so
743  * they stay in memory.
744  */
745 void
746 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
747 {
748 	dmu_buf_impl_t *db, *db_next;
749 	uint64_t txg = tx->tx_txg;
750 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
751 	uint64_t first_l1 = start >> epbs;
752 	uint64_t last_l1 = end >> epbs;
753 
754 	if (end > dn->dn_maxblkid) {
755 		end = dn->dn_maxblkid;
756 		last_l1 = end >> epbs;
757 	}
758 	dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
759 	mutex_enter(&dn->dn_dbufs_mtx);
760 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
761 		db_next = list_next(&dn->dn_dbufs, db);
762 		ASSERT(db->db_blkid != DB_BONUS_BLKID);
763 
764 		if (db->db_level == 1 &&
765 		    db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
766 			mutex_enter(&db->db_mtx);
767 			if (db->db_last_dirty &&
768 			    db->db_last_dirty->dr_txg < txg) {
769 				dbuf_add_ref(db, FTAG);
770 				mutex_exit(&db->db_mtx);
771 				dbuf_will_dirty(db, tx);
772 				dbuf_rele(db, FTAG);
773 			} else {
774 				mutex_exit(&db->db_mtx);
775 			}
776 		}
777 
778 		if (db->db_level != 0)
779 			continue;
780 		dprintf_dbuf(db, "found buf %s\n", "");
781 		if (db->db_blkid < start || db->db_blkid > end)
782 			continue;
783 
784 		/* found a level 0 buffer in the range */
785 		if (dbuf_undirty(db, tx))
786 			continue;
787 
788 		mutex_enter(&db->db_mtx);
789 		if (db->db_state == DB_UNCACHED ||
790 		    db->db_state == DB_NOFILL ||
791 		    db->db_state == DB_EVICTING) {
792 			ASSERT(db->db.db_data == NULL);
793 			mutex_exit(&db->db_mtx);
794 			continue;
795 		}
796 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
797 			/* will be handled in dbuf_read_done or dbuf_rele */
798 			db->db_freed_in_flight = TRUE;
799 			mutex_exit(&db->db_mtx);
800 			continue;
801 		}
802 		if (refcount_count(&db->db_holds) == 0) {
803 			ASSERT(db->db_buf);
804 			dbuf_clear(db);
805 			continue;
806 		}
807 		/* The dbuf is referenced */
808 
809 		if (db->db_last_dirty != NULL) {
810 			dbuf_dirty_record_t *dr = db->db_last_dirty;
811 
812 			if (dr->dr_txg == txg) {
813 				/*
814 				 * This buffer is "in-use", re-adjust the file
815 				 * size to reflect that this buffer may
816 				 * contain new data when we sync.
817 				 */
818 				if (db->db_blkid > dn->dn_maxblkid)
819 					dn->dn_maxblkid = db->db_blkid;
820 				dbuf_unoverride(dr);
821 			} else {
822 				/*
823 				 * This dbuf is not dirty in the open context.
824 				 * Either uncache it (if its not referenced in
825 				 * the open context) or reset its contents to
826 				 * empty.
827 				 */
828 				dbuf_fix_old_data(db, txg);
829 			}
830 		}
831 		/* clear the contents if its cached */
832 		if (db->db_state == DB_CACHED) {
833 			ASSERT(db->db.db_data != NULL);
834 			arc_release(db->db_buf, db);
835 			bzero(db->db.db_data, db->db.db_size);
836 			arc_buf_freeze(db->db_buf);
837 		}
838 
839 		mutex_exit(&db->db_mtx);
840 	}
841 	mutex_exit(&dn->dn_dbufs_mtx);
842 }
843 
844 static int
845 dbuf_block_freeable(dmu_buf_impl_t *db)
846 {
847 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
848 	uint64_t birth_txg = 0;
849 
850 	/*
851 	 * We don't need any locking to protect db_blkptr:
852 	 * If it's syncing, then db_last_dirty will be set
853 	 * so we'll ignore db_blkptr.
854 	 */
855 	ASSERT(MUTEX_HELD(&db->db_mtx));
856 	if (db->db_last_dirty)
857 		birth_txg = db->db_last_dirty->dr_txg;
858 	else if (db->db_blkptr)
859 		birth_txg = db->db_blkptr->blk_birth;
860 
861 	/* If we don't exist or are in a snapshot, we can't be freed */
862 	if (birth_txg)
863 		return (ds == NULL ||
864 		    dsl_dataset_block_freeable(ds, birth_txg));
865 	else
866 		return (FALSE);
867 }
868 
869 void
870 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
871 {
872 	arc_buf_t *buf, *obuf;
873 	int osize = db->db.db_size;
874 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
875 
876 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
877 
878 	/* XXX does *this* func really need the lock? */
879 	ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock));
880 
881 	/*
882 	 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
883 	 * is OK, because there can be no other references to the db
884 	 * when we are changing its size, so no concurrent DB_FILL can
885 	 * be happening.
886 	 */
887 	/*
888 	 * XXX we should be doing a dbuf_read, checking the return
889 	 * value and returning that up to our callers
890 	 */
891 	dbuf_will_dirty(db, tx);
892 
893 	/* create the data buffer for the new block */
894 	buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type);
895 
896 	/* copy old block data to the new block */
897 	obuf = db->db_buf;
898 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
899 	/* zero the remainder */
900 	if (size > osize)
901 		bzero((uint8_t *)buf->b_data + osize, size - osize);
902 
903 	mutex_enter(&db->db_mtx);
904 	dbuf_set_data(db, buf);
905 	VERIFY(arc_buf_remove_ref(obuf, db) == 1);
906 	db->db.db_size = size;
907 
908 	if (db->db_level == 0) {
909 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
910 		db->db_last_dirty->dt.dl.dr_data = buf;
911 	}
912 	mutex_exit(&db->db_mtx);
913 
914 	dnode_willuse_space(db->db_dnode, size-osize, tx);
915 }
916 
917 dbuf_dirty_record_t *
918 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
919 {
920 	dnode_t *dn = db->db_dnode;
921 	objset_t *os = dn->dn_objset;
922 	dbuf_dirty_record_t **drp, *dr;
923 	int drop_struct_lock = FALSE;
924 	boolean_t do_free_accounting = B_FALSE;
925 	int txgoff = tx->tx_txg & TXG_MASK;
926 
927 	ASSERT(tx->tx_txg != 0);
928 	ASSERT(!refcount_is_zero(&db->db_holds));
929 	DMU_TX_DIRTY_BUF(tx, db);
930 
931 	/*
932 	 * Shouldn't dirty a regular buffer in syncing context.  Private
933 	 * objects may be dirtied in syncing context, but only if they
934 	 * were already pre-dirtied in open context.
935 	 */
936 	ASSERT(!dmu_tx_is_syncing(tx) ||
937 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
938 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
939 	    dn->dn_objset->os_dsl_dataset == NULL);
940 	/*
941 	 * We make this assert for private objects as well, but after we
942 	 * check if we're already dirty.  They are allowed to re-dirty
943 	 * in syncing context.
944 	 */
945 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
946 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
947 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
948 
949 	mutex_enter(&db->db_mtx);
950 	/*
951 	 * XXX make this true for indirects too?  The problem is that
952 	 * transactions created with dmu_tx_create_assigned() from
953 	 * syncing context don't bother holding ahead.
954 	 */
955 	ASSERT(db->db_level != 0 ||
956 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
957 	    db->db_state == DB_NOFILL);
958 
959 	mutex_enter(&dn->dn_mtx);
960 	/*
961 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
962 	 * initialize the objset.
963 	 */
964 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
965 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
966 		dn->dn_dirtyctx =
967 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
968 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
969 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
970 	}
971 	mutex_exit(&dn->dn_mtx);
972 
973 	/*
974 	 * If this buffer is already dirty, we're done.
975 	 */
976 	drp = &db->db_last_dirty;
977 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
978 	    db->db.db_object == DMU_META_DNODE_OBJECT);
979 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
980 		drp = &dr->dr_next;
981 	if (dr && dr->dr_txg == tx->tx_txg) {
982 		if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) {
983 			/*
984 			 * If this buffer has already been written out,
985 			 * we now need to reset its state.
986 			 */
987 			dbuf_unoverride(dr);
988 			if (db->db.db_object != DMU_META_DNODE_OBJECT &&
989 			    db->db_state != DB_NOFILL)
990 				arc_buf_thaw(db->db_buf);
991 		}
992 		mutex_exit(&db->db_mtx);
993 		return (dr);
994 	}
995 
996 	/*
997 	 * Only valid if not already dirty.
998 	 */
999 	ASSERT(dn->dn_object == 0 ||
1000 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1001 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1002 
1003 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1004 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1005 	    dn->dn_phys->dn_nlevels > db->db_level ||
1006 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1007 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1008 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1009 
1010 	/*
1011 	 * We should only be dirtying in syncing context if it's the
1012 	 * mos or we're initializing the os or it's a special object.
1013 	 * However, we are allowed to dirty in syncing context provided
1014 	 * we already dirtied it in open context.  Hence we must make
1015 	 * this assertion only if we're not already dirty.
1016 	 */
1017 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1018 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1019 	ASSERT(db->db.db_size != 0);
1020 
1021 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1022 
1023 	if (db->db_blkid != DB_BONUS_BLKID) {
1024 		/*
1025 		 * Update the accounting.
1026 		 * Note: we delay "free accounting" until after we drop
1027 		 * the db_mtx.  This keeps us from grabbing other locks
1028 		 * (and possibly deadlocking) in bp_get_dsize() while
1029 		 * also holding the db_mtx.
1030 		 */
1031 		dnode_willuse_space(dn, db->db.db_size, tx);
1032 		do_free_accounting = dbuf_block_freeable(db);
1033 	}
1034 
1035 	/*
1036 	 * If this buffer is dirty in an old transaction group we need
1037 	 * to make a copy of it so that the changes we make in this
1038 	 * transaction group won't leak out when we sync the older txg.
1039 	 */
1040 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1041 	if (db->db_level == 0) {
1042 		void *data_old = db->db_buf;
1043 
1044 		if (db->db_state != DB_NOFILL) {
1045 			if (db->db_blkid == DB_BONUS_BLKID) {
1046 				dbuf_fix_old_data(db, tx->tx_txg);
1047 				data_old = db->db.db_data;
1048 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1049 				/*
1050 				 * Release the data buffer from the cache so
1051 				 * that we can modify it without impacting
1052 				 * possible other users of this cached data
1053 				 * block.  Note that indirect blocks and
1054 				 * private objects are not released until the
1055 				 * syncing state (since they are only modified
1056 				 * then).
1057 				 */
1058 				arc_release(db->db_buf, db);
1059 				dbuf_fix_old_data(db, tx->tx_txg);
1060 				data_old = db->db_buf;
1061 			}
1062 			ASSERT(data_old != NULL);
1063 		}
1064 		dr->dt.dl.dr_data = data_old;
1065 	} else {
1066 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1067 		list_create(&dr->dt.di.dr_children,
1068 		    sizeof (dbuf_dirty_record_t),
1069 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1070 	}
1071 	dr->dr_dbuf = db;
1072 	dr->dr_txg = tx->tx_txg;
1073 	dr->dr_next = *drp;
1074 	*drp = dr;
1075 
1076 	/*
1077 	 * We could have been freed_in_flight between the dbuf_noread
1078 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1079 	 * happened after the free.
1080 	 */
1081 	if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) {
1082 		mutex_enter(&dn->dn_mtx);
1083 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1084 		mutex_exit(&dn->dn_mtx);
1085 		db->db_freed_in_flight = FALSE;
1086 	}
1087 
1088 	/*
1089 	 * This buffer is now part of this txg
1090 	 */
1091 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1092 	db->db_dirtycnt += 1;
1093 	ASSERT3U(db->db_dirtycnt, <=, 3);
1094 
1095 	mutex_exit(&db->db_mtx);
1096 
1097 	if (db->db_blkid == DB_BONUS_BLKID) {
1098 		mutex_enter(&dn->dn_mtx);
1099 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1100 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1101 		mutex_exit(&dn->dn_mtx);
1102 		dnode_setdirty(dn, tx);
1103 		return (dr);
1104 	} else if (do_free_accounting) {
1105 		blkptr_t *bp = db->db_blkptr;
1106 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1107 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1108 		/*
1109 		 * This is only a guess -- if the dbuf is dirty
1110 		 * in a previous txg, we don't know how much
1111 		 * space it will use on disk yet.  We should
1112 		 * really have the struct_rwlock to access
1113 		 * db_blkptr, but since this is just a guess,
1114 		 * it's OK if we get an odd answer.
1115 		 */
1116 		dnode_willuse_space(dn, -willfree, tx);
1117 	}
1118 
1119 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1120 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1121 		drop_struct_lock = TRUE;
1122 	}
1123 
1124 	if (db->db_level == 0) {
1125 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1126 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1127 	}
1128 
1129 	if (db->db_level+1 < dn->dn_nlevels) {
1130 		dmu_buf_impl_t *parent = db->db_parent;
1131 		dbuf_dirty_record_t *di;
1132 		int parent_held = FALSE;
1133 
1134 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1135 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1136 
1137 			parent = dbuf_hold_level(dn, db->db_level+1,
1138 			    db->db_blkid >> epbs, FTAG);
1139 			parent_held = TRUE;
1140 		}
1141 		if (drop_struct_lock)
1142 			rw_exit(&dn->dn_struct_rwlock);
1143 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1144 		di = dbuf_dirty(parent, tx);
1145 		if (parent_held)
1146 			dbuf_rele(parent, FTAG);
1147 
1148 		mutex_enter(&db->db_mtx);
1149 		/*  possible race with dbuf_undirty() */
1150 		if (db->db_last_dirty == dr ||
1151 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1152 			mutex_enter(&di->dt.di.dr_mtx);
1153 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1154 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1155 			list_insert_tail(&di->dt.di.dr_children, dr);
1156 			mutex_exit(&di->dt.di.dr_mtx);
1157 			dr->dr_parent = di;
1158 		}
1159 		mutex_exit(&db->db_mtx);
1160 	} else {
1161 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1162 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1163 		ASSERT(db->db_parent == NULL ||
1164 		    db->db_parent == db->db_dnode->dn_dbuf);
1165 		mutex_enter(&dn->dn_mtx);
1166 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1167 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1168 		mutex_exit(&dn->dn_mtx);
1169 		if (drop_struct_lock)
1170 			rw_exit(&dn->dn_struct_rwlock);
1171 	}
1172 
1173 	dnode_setdirty(dn, tx);
1174 	return (dr);
1175 }
1176 
1177 static int
1178 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1179 {
1180 	dnode_t *dn = db->db_dnode;
1181 	uint64_t txg = tx->tx_txg;
1182 	dbuf_dirty_record_t *dr, **drp;
1183 
1184 	ASSERT(txg != 0);
1185 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
1186 
1187 	mutex_enter(&db->db_mtx);
1188 	/*
1189 	 * If this buffer is not dirty, we're done.
1190 	 */
1191 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1192 		if (dr->dr_txg <= txg)
1193 			break;
1194 	if (dr == NULL || dr->dr_txg < txg) {
1195 		mutex_exit(&db->db_mtx);
1196 		return (0);
1197 	}
1198 	ASSERT(dr->dr_txg == txg);
1199 	ASSERT(dr->dr_dbuf == db);
1200 
1201 	/*
1202 	 * If this buffer is currently held, we cannot undirty
1203 	 * it, since one of the current holders may be in the
1204 	 * middle of an update.  Note that users of dbuf_undirty()
1205 	 * should not place a hold on the dbuf before the call.
1206 	 */
1207 	if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1208 		mutex_exit(&db->db_mtx);
1209 		/* Make sure we don't toss this buffer at sync phase */
1210 		mutex_enter(&dn->dn_mtx);
1211 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1212 		mutex_exit(&dn->dn_mtx);
1213 		return (0);
1214 	}
1215 
1216 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1217 
1218 	ASSERT(db->db.db_size != 0);
1219 
1220 	/* XXX would be nice to fix up dn_towrite_space[] */
1221 
1222 	*drp = dr->dr_next;
1223 
1224 	if (dr->dr_parent) {
1225 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1226 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1227 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1228 	} else if (db->db_level+1 == dn->dn_nlevels) {
1229 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1230 		mutex_enter(&dn->dn_mtx);
1231 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1232 		mutex_exit(&dn->dn_mtx);
1233 	}
1234 
1235 	if (db->db_level == 0) {
1236 		if (db->db_state != DB_NOFILL) {
1237 			dbuf_unoverride(dr);
1238 
1239 			ASSERT(db->db_buf != NULL);
1240 			ASSERT(dr->dt.dl.dr_data != NULL);
1241 			if (dr->dt.dl.dr_data != db->db_buf)
1242 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
1243 				    db) == 1);
1244 		}
1245 	} else {
1246 		ASSERT(db->db_buf != NULL);
1247 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1248 		mutex_destroy(&dr->dt.di.dr_mtx);
1249 		list_destroy(&dr->dt.di.dr_children);
1250 	}
1251 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1252 
1253 	ASSERT(db->db_dirtycnt > 0);
1254 	db->db_dirtycnt -= 1;
1255 
1256 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1257 		arc_buf_t *buf = db->db_buf;
1258 
1259 		ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1260 		dbuf_set_data(db, NULL);
1261 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
1262 		dbuf_evict(db);
1263 		return (1);
1264 	}
1265 
1266 	mutex_exit(&db->db_mtx);
1267 	return (0);
1268 }
1269 
1270 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1271 void
1272 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1273 {
1274 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1275 
1276 	ASSERT(tx->tx_txg != 0);
1277 	ASSERT(!refcount_is_zero(&db->db_holds));
1278 
1279 	if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock))
1280 		rf |= DB_RF_HAVESTRUCT;
1281 	(void) dbuf_read(db, NULL, rf);
1282 	(void) dbuf_dirty(db, tx);
1283 }
1284 
1285 void
1286 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1287 {
1288 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1289 
1290 	db->db_state = DB_NOFILL;
1291 
1292 	dmu_buf_will_fill(db_fake, tx);
1293 }
1294 
1295 void
1296 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1297 {
1298 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1299 
1300 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
1301 	ASSERT(tx->tx_txg != 0);
1302 	ASSERT(db->db_level == 0);
1303 	ASSERT(!refcount_is_zero(&db->db_holds));
1304 
1305 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1306 	    dmu_tx_private_ok(tx));
1307 
1308 	dbuf_noread(db);
1309 	(void) dbuf_dirty(db, tx);
1310 }
1311 
1312 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1313 /* ARGSUSED */
1314 void
1315 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1316 {
1317 	mutex_enter(&db->db_mtx);
1318 	DBUF_VERIFY(db);
1319 
1320 	if (db->db_state == DB_FILL) {
1321 		if (db->db_level == 0 && db->db_freed_in_flight) {
1322 			ASSERT(db->db_blkid != DB_BONUS_BLKID);
1323 			/* we were freed while filling */
1324 			/* XXX dbuf_undirty? */
1325 			bzero(db->db.db_data, db->db.db_size);
1326 			db->db_freed_in_flight = FALSE;
1327 		}
1328 		db->db_state = DB_CACHED;
1329 		cv_broadcast(&db->db_changed);
1330 	}
1331 	mutex_exit(&db->db_mtx);
1332 }
1333 
1334 /*
1335  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1336  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1337  */
1338 void
1339 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1340 {
1341 	ASSERT(!refcount_is_zero(&db->db_holds));
1342 	ASSERT(db->db_dnode->dn_object != DMU_META_DNODE_OBJECT);
1343 	ASSERT(db->db_blkid != DB_BONUS_BLKID);
1344 	ASSERT(db->db_level == 0);
1345 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1346 	ASSERT(buf != NULL);
1347 	ASSERT(arc_buf_size(buf) == db->db.db_size);
1348 	ASSERT(tx->tx_txg != 0);
1349 
1350 	arc_return_buf(buf, db);
1351 	ASSERT(arc_released(buf));
1352 
1353 	mutex_enter(&db->db_mtx);
1354 
1355 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1356 		cv_wait(&db->db_changed, &db->db_mtx);
1357 
1358 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1359 
1360 	if (db->db_state == DB_CACHED &&
1361 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1362 		mutex_exit(&db->db_mtx);
1363 		(void) dbuf_dirty(db, tx);
1364 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1365 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
1366 		xuio_stat_wbuf_copied();
1367 		return;
1368 	}
1369 
1370 	xuio_stat_wbuf_nocopy();
1371 	if (db->db_state == DB_CACHED) {
1372 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1373 
1374 		ASSERT(db->db_buf != NULL);
1375 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1376 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1377 			if (!arc_released(db->db_buf)) {
1378 				ASSERT(dr->dt.dl.dr_override_state ==
1379 				    DR_OVERRIDDEN);
1380 				arc_release(db->db_buf, db);
1381 			}
1382 			dr->dt.dl.dr_data = buf;
1383 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1384 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1385 			arc_release(db->db_buf, db);
1386 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1387 		}
1388 		db->db_buf = NULL;
1389 	}
1390 	ASSERT(db->db_buf == NULL);
1391 	dbuf_set_data(db, buf);
1392 	db->db_state = DB_FILL;
1393 	mutex_exit(&db->db_mtx);
1394 	(void) dbuf_dirty(db, tx);
1395 	dbuf_fill_done(db, tx);
1396 }
1397 
1398 /*
1399  * "Clear" the contents of this dbuf.  This will mark the dbuf
1400  * EVICTING and clear *most* of its references.  Unfortunetely,
1401  * when we are not holding the dn_dbufs_mtx, we can't clear the
1402  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1403  * in this case.  For callers from the DMU we will usually see:
1404  *	dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1405  * For the arc callback, we will usually see:
1406  * 	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1407  * Sometimes, though, we will get a mix of these two:
1408  *	DMU: dbuf_clear()->arc_buf_evict()
1409  *	ARC: dbuf_do_evict()->dbuf_destroy()
1410  */
1411 void
1412 dbuf_clear(dmu_buf_impl_t *db)
1413 {
1414 	dnode_t *dn = db->db_dnode;
1415 	dmu_buf_impl_t *parent = db->db_parent;
1416 	dmu_buf_impl_t *dndb = dn->dn_dbuf;
1417 	int dbuf_gone = FALSE;
1418 
1419 	ASSERT(MUTEX_HELD(&db->db_mtx));
1420 	ASSERT(refcount_is_zero(&db->db_holds));
1421 
1422 	dbuf_evict_user(db);
1423 
1424 	if (db->db_state == DB_CACHED) {
1425 		ASSERT(db->db.db_data != NULL);
1426 		if (db->db_blkid == DB_BONUS_BLKID) {
1427 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1428 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1429 		}
1430 		db->db.db_data = NULL;
1431 		db->db_state = DB_UNCACHED;
1432 	}
1433 
1434 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1435 	ASSERT(db->db_data_pending == NULL);
1436 
1437 	db->db_state = DB_EVICTING;
1438 	db->db_blkptr = NULL;
1439 
1440 	if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1441 		list_remove(&dn->dn_dbufs, db);
1442 		dnode_rele(dn, db);
1443 		db->db_dnode = NULL;
1444 	}
1445 
1446 	if (db->db_buf)
1447 		dbuf_gone = arc_buf_evict(db->db_buf);
1448 
1449 	if (!dbuf_gone)
1450 		mutex_exit(&db->db_mtx);
1451 
1452 	/*
1453 	 * If this dbuf is referened from an indirect dbuf,
1454 	 * decrement the ref count on the indirect dbuf.
1455 	 */
1456 	if (parent && parent != dndb)
1457 		dbuf_rele(parent, db);
1458 }
1459 
1460 static int
1461 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1462     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1463 {
1464 	int nlevels, epbs;
1465 
1466 	*parentp = NULL;
1467 	*bpp = NULL;
1468 
1469 	ASSERT(blkid != DB_BONUS_BLKID);
1470 
1471 	if (dn->dn_phys->dn_nlevels == 0)
1472 		nlevels = 1;
1473 	else
1474 		nlevels = dn->dn_phys->dn_nlevels;
1475 
1476 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1477 
1478 	ASSERT3U(level * epbs, <, 64);
1479 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1480 	if (level >= nlevels ||
1481 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1482 		/* the buffer has no parent yet */
1483 		return (ENOENT);
1484 	} else if (level < nlevels-1) {
1485 		/* this block is referenced from an indirect block */
1486 		int err = dbuf_hold_impl(dn, level+1,
1487 		    blkid >> epbs, fail_sparse, NULL, parentp);
1488 		if (err)
1489 			return (err);
1490 		err = dbuf_read(*parentp, NULL,
1491 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1492 		if (err) {
1493 			dbuf_rele(*parentp, NULL);
1494 			*parentp = NULL;
1495 			return (err);
1496 		}
1497 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1498 		    (blkid & ((1ULL << epbs) - 1));
1499 		return (0);
1500 	} else {
1501 		/* the block is referenced from the dnode */
1502 		ASSERT3U(level, ==, nlevels-1);
1503 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1504 		    blkid < dn->dn_phys->dn_nblkptr);
1505 		if (dn->dn_dbuf) {
1506 			dbuf_add_ref(dn->dn_dbuf, NULL);
1507 			*parentp = dn->dn_dbuf;
1508 		}
1509 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1510 		return (0);
1511 	}
1512 }
1513 
1514 static dmu_buf_impl_t *
1515 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1516     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1517 {
1518 	objset_t *os = dn->dn_objset;
1519 	dmu_buf_impl_t *db, *odb;
1520 
1521 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1522 	ASSERT(dn->dn_type != DMU_OT_NONE);
1523 
1524 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1525 
1526 	db->db_objset = os;
1527 	db->db.db_object = dn->dn_object;
1528 	db->db_level = level;
1529 	db->db_blkid = blkid;
1530 	db->db_last_dirty = NULL;
1531 	db->db_dirtycnt = 0;
1532 	db->db_dnode = dn;
1533 	db->db_parent = parent;
1534 	db->db_blkptr = blkptr;
1535 
1536 	db->db_user_ptr = NULL;
1537 	db->db_user_data_ptr_ptr = NULL;
1538 	db->db_evict_func = NULL;
1539 	db->db_immediate_evict = 0;
1540 	db->db_freed_in_flight = 0;
1541 
1542 	if (blkid == DB_BONUS_BLKID) {
1543 		ASSERT3P(parent, ==, dn->dn_dbuf);
1544 		db->db.db_size = DN_MAX_BONUSLEN -
1545 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1546 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1547 		db->db.db_offset = DB_BONUS_BLKID;
1548 		db->db_state = DB_UNCACHED;
1549 		/* the bonus dbuf is not placed in the hash table */
1550 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1551 		return (db);
1552 	} else {
1553 		int blocksize =
1554 		    db->db_level ? 1<<dn->dn_indblkshift :  dn->dn_datablksz;
1555 		db->db.db_size = blocksize;
1556 		db->db.db_offset = db->db_blkid * blocksize;
1557 	}
1558 
1559 	/*
1560 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1561 	 * in the hash table *and* added to the dbufs list.
1562 	 * This prevents a possible deadlock with someone
1563 	 * trying to look up this dbuf before its added to the
1564 	 * dn_dbufs list.
1565 	 */
1566 	mutex_enter(&dn->dn_dbufs_mtx);
1567 	db->db_state = DB_EVICTING;
1568 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1569 		/* someone else inserted it first */
1570 		kmem_cache_free(dbuf_cache, db);
1571 		mutex_exit(&dn->dn_dbufs_mtx);
1572 		return (odb);
1573 	}
1574 	list_insert_head(&dn->dn_dbufs, db);
1575 	db->db_state = DB_UNCACHED;
1576 	mutex_exit(&dn->dn_dbufs_mtx);
1577 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1578 
1579 	if (parent && parent != dn->dn_dbuf)
1580 		dbuf_add_ref(parent, db);
1581 
1582 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1583 	    refcount_count(&dn->dn_holds) > 0);
1584 	(void) refcount_add(&dn->dn_holds, db);
1585 
1586 	dprintf_dbuf(db, "db=%p\n", db);
1587 
1588 	return (db);
1589 }
1590 
1591 static int
1592 dbuf_do_evict(void *private)
1593 {
1594 	arc_buf_t *buf = private;
1595 	dmu_buf_impl_t *db = buf->b_private;
1596 
1597 	if (!MUTEX_HELD(&db->db_mtx))
1598 		mutex_enter(&db->db_mtx);
1599 
1600 	ASSERT(refcount_is_zero(&db->db_holds));
1601 
1602 	if (db->db_state != DB_EVICTING) {
1603 		ASSERT(db->db_state == DB_CACHED);
1604 		DBUF_VERIFY(db);
1605 		db->db_buf = NULL;
1606 		dbuf_evict(db);
1607 	} else {
1608 		mutex_exit(&db->db_mtx);
1609 		dbuf_destroy(db);
1610 	}
1611 	return (0);
1612 }
1613 
1614 static void
1615 dbuf_destroy(dmu_buf_impl_t *db)
1616 {
1617 	ASSERT(refcount_is_zero(&db->db_holds));
1618 
1619 	if (db->db_blkid != DB_BONUS_BLKID) {
1620 		/*
1621 		 * If this dbuf is still on the dn_dbufs list,
1622 		 * remove it from that list.
1623 		 */
1624 		if (db->db_dnode) {
1625 			dnode_t *dn = db->db_dnode;
1626 
1627 			mutex_enter(&dn->dn_dbufs_mtx);
1628 			list_remove(&dn->dn_dbufs, db);
1629 			mutex_exit(&dn->dn_dbufs_mtx);
1630 
1631 			dnode_rele(dn, db);
1632 			db->db_dnode = NULL;
1633 		}
1634 		dbuf_hash_remove(db);
1635 	}
1636 	db->db_parent = NULL;
1637 	db->db_buf = NULL;
1638 
1639 	ASSERT(!list_link_active(&db->db_link));
1640 	ASSERT(db->db.db_data == NULL);
1641 	ASSERT(db->db_hash_next == NULL);
1642 	ASSERT(db->db_blkptr == NULL);
1643 	ASSERT(db->db_data_pending == NULL);
1644 
1645 	kmem_cache_free(dbuf_cache, db);
1646 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1647 }
1648 
1649 void
1650 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1651 {
1652 	dmu_buf_impl_t *db = NULL;
1653 	blkptr_t *bp = NULL;
1654 
1655 	ASSERT(blkid != DB_BONUS_BLKID);
1656 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1657 
1658 	if (dnode_block_freed(dn, blkid))
1659 		return;
1660 
1661 	/* dbuf_find() returns with db_mtx held */
1662 	if (db = dbuf_find(dn, 0, blkid)) {
1663 		if (refcount_count(&db->db_holds) > 0) {
1664 			/*
1665 			 * This dbuf is active.  We assume that it is
1666 			 * already CACHED, or else about to be either
1667 			 * read or filled.
1668 			 */
1669 			mutex_exit(&db->db_mtx);
1670 			return;
1671 		}
1672 		mutex_exit(&db->db_mtx);
1673 		db = NULL;
1674 	}
1675 
1676 	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1677 		if (bp && !BP_IS_HOLE(bp)) {
1678 			arc_buf_t *pbuf;
1679 			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1680 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1681 			zbookmark_t zb;
1682 
1683 			SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1684 			    dn->dn_object, 0, blkid);
1685 
1686 			if (db)
1687 				pbuf = db->db_buf;
1688 			else
1689 				pbuf = dn->dn_objset->os_phys_buf;
1690 
1691 			(void) arc_read(NULL, dn->dn_objset->os_spa,
1692 			    bp, pbuf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
1693 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1694 			    &aflags, &zb);
1695 		}
1696 		if (db)
1697 			dbuf_rele(db, NULL);
1698 	}
1699 }
1700 
1701 /*
1702  * Returns with db_holds incremented, and db_mtx not held.
1703  * Note: dn_struct_rwlock must be held.
1704  */
1705 int
1706 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1707     void *tag, dmu_buf_impl_t **dbp)
1708 {
1709 	dmu_buf_impl_t *db, *parent = NULL;
1710 
1711 	ASSERT(blkid != DB_BONUS_BLKID);
1712 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1713 	ASSERT3U(dn->dn_nlevels, >, level);
1714 
1715 	*dbp = NULL;
1716 top:
1717 	/* dbuf_find() returns with db_mtx held */
1718 	db = dbuf_find(dn, level, blkid);
1719 
1720 	if (db == NULL) {
1721 		blkptr_t *bp = NULL;
1722 		int err;
1723 
1724 		ASSERT3P(parent, ==, NULL);
1725 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1726 		if (fail_sparse) {
1727 			if (err == 0 && bp && BP_IS_HOLE(bp))
1728 				err = ENOENT;
1729 			if (err) {
1730 				if (parent)
1731 					dbuf_rele(parent, NULL);
1732 				return (err);
1733 			}
1734 		}
1735 		if (err && err != ENOENT)
1736 			return (err);
1737 		db = dbuf_create(dn, level, blkid, parent, bp);
1738 	}
1739 
1740 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1741 		arc_buf_add_ref(db->db_buf, db);
1742 		if (db->db_buf->b_data == NULL) {
1743 			dbuf_clear(db);
1744 			if (parent) {
1745 				dbuf_rele(parent, NULL);
1746 				parent = NULL;
1747 			}
1748 			goto top;
1749 		}
1750 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1751 	}
1752 
1753 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1754 
1755 	/*
1756 	 * If this buffer is currently syncing out, and we are are
1757 	 * still referencing it from db_data, we need to make a copy
1758 	 * of it in case we decide we want to dirty it again in this txg.
1759 	 */
1760 	if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID &&
1761 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
1762 	    db->db_state == DB_CACHED && db->db_data_pending) {
1763 		dbuf_dirty_record_t *dr = db->db_data_pending;
1764 
1765 		if (dr->dt.dl.dr_data == db->db_buf) {
1766 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1767 
1768 			dbuf_set_data(db,
1769 			    arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
1770 			    db->db.db_size, db, type));
1771 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1772 			    db->db.db_size);
1773 		}
1774 	}
1775 
1776 	(void) refcount_add(&db->db_holds, tag);
1777 	dbuf_update_data(db);
1778 	DBUF_VERIFY(db);
1779 	mutex_exit(&db->db_mtx);
1780 
1781 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
1782 	if (parent)
1783 		dbuf_rele(parent, NULL);
1784 
1785 	ASSERT3P(db->db_dnode, ==, dn);
1786 	ASSERT3U(db->db_blkid, ==, blkid);
1787 	ASSERT3U(db->db_level, ==, level);
1788 	*dbp = db;
1789 
1790 	return (0);
1791 }
1792 
1793 dmu_buf_impl_t *
1794 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1795 {
1796 	dmu_buf_impl_t *db;
1797 	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1798 	return (err ? NULL : db);
1799 }
1800 
1801 dmu_buf_impl_t *
1802 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1803 {
1804 	dmu_buf_impl_t *db;
1805 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1806 	return (err ? NULL : db);
1807 }
1808 
1809 void
1810 dbuf_create_bonus(dnode_t *dn)
1811 {
1812 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1813 
1814 	ASSERT(dn->dn_bonus == NULL);
1815 	dn->dn_bonus = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL);
1816 }
1817 
1818 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1819 void
1820 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1821 {
1822 	int64_t holds = refcount_add(&db->db_holds, tag);
1823 	ASSERT(holds > 1);
1824 }
1825 
1826 #pragma weak dmu_buf_rele = dbuf_rele
1827 void
1828 dbuf_rele(dmu_buf_impl_t *db, void *tag)
1829 {
1830 	mutex_enter(&db->db_mtx);
1831 	dbuf_rele_and_unlock(db, tag);
1832 }
1833 
1834 /*
1835  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
1836  * db_dirtycnt and db_holds to be updated atomically.
1837  */
1838 void
1839 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
1840 {
1841 	int64_t holds;
1842 
1843 	ASSERT(MUTEX_HELD(&db->db_mtx));
1844 	DBUF_VERIFY(db);
1845 
1846 	holds = refcount_remove(&db->db_holds, tag);
1847 	ASSERT(holds >= 0);
1848 
1849 	/*
1850 	 * We can't freeze indirects if there is a possibility that they
1851 	 * may be modified in the current syncing context.
1852 	 */
1853 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
1854 		arc_buf_freeze(db->db_buf);
1855 
1856 	if (holds == db->db_dirtycnt &&
1857 	    db->db_level == 0 && db->db_immediate_evict)
1858 		dbuf_evict_user(db);
1859 
1860 	if (holds == 0) {
1861 		if (db->db_blkid == DB_BONUS_BLKID) {
1862 			mutex_exit(&db->db_mtx);
1863 			dnode_rele(db->db_dnode, db);
1864 		} else if (db->db_buf == NULL) {
1865 			/*
1866 			 * This is a special case: we never associated this
1867 			 * dbuf with any data allocated from the ARC.
1868 			 */
1869 			ASSERT(db->db_state == DB_UNCACHED ||
1870 			    db->db_state == DB_NOFILL);
1871 			dbuf_evict(db);
1872 		} else if (arc_released(db->db_buf)) {
1873 			arc_buf_t *buf = db->db_buf;
1874 			/*
1875 			 * This dbuf has anonymous data associated with it.
1876 			 */
1877 			dbuf_set_data(db, NULL);
1878 			VERIFY(arc_buf_remove_ref(buf, db) == 1);
1879 			dbuf_evict(db);
1880 		} else {
1881 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0);
1882 			if (!DBUF_IS_CACHEABLE(db))
1883 				dbuf_clear(db);
1884 			else
1885 				mutex_exit(&db->db_mtx);
1886 		}
1887 	} else {
1888 		mutex_exit(&db->db_mtx);
1889 	}
1890 }
1891 
1892 #pragma weak dmu_buf_refcount = dbuf_refcount
1893 uint64_t
1894 dbuf_refcount(dmu_buf_impl_t *db)
1895 {
1896 	return (refcount_count(&db->db_holds));
1897 }
1898 
1899 void *
1900 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1901     dmu_buf_evict_func_t *evict_func)
1902 {
1903 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1904 	    user_data_ptr_ptr, evict_func));
1905 }
1906 
1907 void *
1908 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1909     dmu_buf_evict_func_t *evict_func)
1910 {
1911 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1912 
1913 	db->db_immediate_evict = TRUE;
1914 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1915 	    user_data_ptr_ptr, evict_func));
1916 }
1917 
1918 void *
1919 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
1920     void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
1921 {
1922 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1923 	ASSERT(db->db_level == 0);
1924 
1925 	ASSERT((user_ptr == NULL) == (evict_func == NULL));
1926 
1927 	mutex_enter(&db->db_mtx);
1928 
1929 	if (db->db_user_ptr == old_user_ptr) {
1930 		db->db_user_ptr = user_ptr;
1931 		db->db_user_data_ptr_ptr = user_data_ptr_ptr;
1932 		db->db_evict_func = evict_func;
1933 
1934 		dbuf_update_data(db);
1935 	} else {
1936 		old_user_ptr = db->db_user_ptr;
1937 	}
1938 
1939 	mutex_exit(&db->db_mtx);
1940 	return (old_user_ptr);
1941 }
1942 
1943 void *
1944 dmu_buf_get_user(dmu_buf_t *db_fake)
1945 {
1946 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1947 	ASSERT(!refcount_is_zero(&db->db_holds));
1948 
1949 	return (db->db_user_ptr);
1950 }
1951 
1952 boolean_t
1953 dmu_buf_freeable(dmu_buf_t *dbuf)
1954 {
1955 	boolean_t res = B_FALSE;
1956 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1957 
1958 	if (db->db_blkptr)
1959 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
1960 		    db->db_blkptr->blk_birth);
1961 
1962 	return (res);
1963 }
1964 
1965 static void
1966 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
1967 {
1968 	/* ASSERT(dmu_tx_is_syncing(tx) */
1969 	ASSERT(MUTEX_HELD(&db->db_mtx));
1970 
1971 	if (db->db_blkptr != NULL)
1972 		return;
1973 
1974 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
1975 		/*
1976 		 * This buffer was allocated at a time when there was
1977 		 * no available blkptrs from the dnode, or it was
1978 		 * inappropriate to hook it in (i.e., nlevels mis-match).
1979 		 */
1980 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
1981 		ASSERT(db->db_parent == NULL);
1982 		db->db_parent = dn->dn_dbuf;
1983 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
1984 		DBUF_VERIFY(db);
1985 	} else {
1986 		dmu_buf_impl_t *parent = db->db_parent;
1987 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1988 
1989 		ASSERT(dn->dn_phys->dn_nlevels > 1);
1990 		if (parent == NULL) {
1991 			mutex_exit(&db->db_mtx);
1992 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
1993 			(void) dbuf_hold_impl(dn, db->db_level+1,
1994 			    db->db_blkid >> epbs, FALSE, db, &parent);
1995 			rw_exit(&dn->dn_struct_rwlock);
1996 			mutex_enter(&db->db_mtx);
1997 			db->db_parent = parent;
1998 		}
1999 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2000 		    (db->db_blkid & ((1ULL << epbs) - 1));
2001 		DBUF_VERIFY(db);
2002 	}
2003 }
2004 
2005 static void
2006 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2007 {
2008 	dmu_buf_impl_t *db = dr->dr_dbuf;
2009 	dnode_t *dn = db->db_dnode;
2010 	zio_t *zio;
2011 
2012 	ASSERT(dmu_tx_is_syncing(tx));
2013 
2014 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2015 
2016 	mutex_enter(&db->db_mtx);
2017 
2018 	ASSERT(db->db_level > 0);
2019 	DBUF_VERIFY(db);
2020 
2021 	if (db->db_buf == NULL) {
2022 		mutex_exit(&db->db_mtx);
2023 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2024 		mutex_enter(&db->db_mtx);
2025 	}
2026 	ASSERT3U(db->db_state, ==, DB_CACHED);
2027 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2028 	ASSERT(db->db_buf != NULL);
2029 
2030 	dbuf_check_blkptr(dn, db);
2031 
2032 	db->db_data_pending = dr;
2033 
2034 	mutex_exit(&db->db_mtx);
2035 	dbuf_write(dr, db->db_buf, tx);
2036 
2037 	zio = dr->dr_zio;
2038 	mutex_enter(&dr->dt.di.dr_mtx);
2039 	dbuf_sync_list(&dr->dt.di.dr_children, tx);
2040 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2041 	mutex_exit(&dr->dt.di.dr_mtx);
2042 	zio_nowait(zio);
2043 }
2044 
2045 static void
2046 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2047 {
2048 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2049 	dmu_buf_impl_t *db = dr->dr_dbuf;
2050 	dnode_t *dn = db->db_dnode;
2051 	objset_t *os = dn->dn_objset;
2052 	uint64_t txg = tx->tx_txg;
2053 
2054 	ASSERT(dmu_tx_is_syncing(tx));
2055 
2056 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2057 
2058 	mutex_enter(&db->db_mtx);
2059 	/*
2060 	 * To be synced, we must be dirtied.  But we
2061 	 * might have been freed after the dirty.
2062 	 */
2063 	if (db->db_state == DB_UNCACHED) {
2064 		/* This buffer has been freed since it was dirtied */
2065 		ASSERT(db->db.db_data == NULL);
2066 	} else if (db->db_state == DB_FILL) {
2067 		/* This buffer was freed and is now being re-filled */
2068 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2069 	} else {
2070 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2071 	}
2072 	DBUF_VERIFY(db);
2073 
2074 	/*
2075 	 * If this is a bonus buffer, simply copy the bonus data into the
2076 	 * dnode.  It will be written out when the dnode is synced (and it
2077 	 * will be synced, since it must have been dirty for dbuf_sync to
2078 	 * be called).
2079 	 */
2080 	if (db->db_blkid == DB_BONUS_BLKID) {
2081 		dbuf_dirty_record_t **drp;
2082 
2083 		ASSERT(*datap != NULL);
2084 		ASSERT3U(db->db_level, ==, 0);
2085 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2086 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2087 		if (*datap != db->db.db_data) {
2088 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
2089 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2090 		}
2091 		db->db_data_pending = NULL;
2092 		drp = &db->db_last_dirty;
2093 		while (*drp != dr)
2094 			drp = &(*drp)->dr_next;
2095 		ASSERT(dr->dr_next == NULL);
2096 		ASSERT(dr->dr_dbuf == db);
2097 		*drp = dr->dr_next;
2098 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
2099 		ASSERT(db->db_dirtycnt > 0);
2100 		db->db_dirtycnt -= 1;
2101 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2102 		return;
2103 	}
2104 
2105 	/*
2106 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
2107 	 * operation to sneak in. As a result, we need to ensure that we
2108 	 * don't check the dr_override_state until we have returned from
2109 	 * dbuf_check_blkptr.
2110 	 */
2111 	dbuf_check_blkptr(dn, db);
2112 
2113 	/*
2114 	 * If this buffer is in the middle of an immdiate write,
2115 	 * wait for the synchronous IO to complete.
2116 	 */
2117 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2118 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2119 		cv_wait(&db->db_changed, &db->db_mtx);
2120 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2121 	}
2122 
2123 	if (db->db_state != DB_NOFILL &&
2124 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2125 	    refcount_count(&db->db_holds) > 1 &&
2126 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2127 	    *datap == db->db_buf) {
2128 		/*
2129 		 * If this buffer is currently "in use" (i.e., there
2130 		 * are active holds and db_data still references it),
2131 		 * then make a copy before we start the write so that
2132 		 * any modifications from the open txg will not leak
2133 		 * into this write.
2134 		 *
2135 		 * NOTE: this copy does not need to be made for
2136 		 * objects only modified in the syncing context (e.g.
2137 		 * DNONE_DNODE blocks).
2138 		 */
2139 		int blksz = arc_buf_size(*datap);
2140 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2141 		*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2142 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
2143 	}
2144 	db->db_data_pending = dr;
2145 
2146 	mutex_exit(&db->db_mtx);
2147 
2148 	dbuf_write(dr, *datap, tx);
2149 
2150 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2151 	if (dn->dn_object == DMU_META_DNODE_OBJECT)
2152 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2153 	else
2154 		zio_nowait(dr->dr_zio);
2155 }
2156 
2157 void
2158 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2159 {
2160 	dbuf_dirty_record_t *dr;
2161 
2162 	while (dr = list_head(list)) {
2163 		if (dr->dr_zio != NULL) {
2164 			/*
2165 			 * If we find an already initialized zio then we
2166 			 * are processing the meta-dnode, and we have finished.
2167 			 * The dbufs for all dnodes are put back on the list
2168 			 * during processing, so that we can zio_wait()
2169 			 * these IOs after initiating all child IOs.
2170 			 */
2171 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2172 			    DMU_META_DNODE_OBJECT);
2173 			break;
2174 		}
2175 		list_remove(list, dr);
2176 		if (dr->dr_dbuf->db_level > 0)
2177 			dbuf_sync_indirect(dr, tx);
2178 		else
2179 			dbuf_sync_leaf(dr, tx);
2180 	}
2181 }
2182 
2183 /* ARGSUSED */
2184 static void
2185 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2186 {
2187 	dmu_buf_impl_t *db = vdb;
2188 	blkptr_t *bp = zio->io_bp;
2189 	blkptr_t *bp_orig = &zio->io_bp_orig;
2190 	dnode_t *dn = db->db_dnode;
2191 	spa_t *spa = zio->io_spa;
2192 	int64_t delta;
2193 	uint64_t fill = 0;
2194 	int i;
2195 
2196 	ASSERT(db->db_blkptr == bp);
2197 
2198 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2199 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2200 	zio->io_prev_space_delta = delta;
2201 
2202 	if (BP_IS_HOLE(bp)) {
2203 		ASSERT(bp->blk_fill == 0);
2204 		return;
2205 	}
2206 
2207 	ASSERT(BP_GET_TYPE(bp) == dn->dn_type);
2208 	ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2209 
2210 	mutex_enter(&db->db_mtx);
2211 
2212 	if (db->db_level == 0) {
2213 		mutex_enter(&dn->dn_mtx);
2214 		if (db->db_blkid > dn->dn_phys->dn_maxblkid)
2215 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2216 		mutex_exit(&dn->dn_mtx);
2217 
2218 		if (dn->dn_type == DMU_OT_DNODE) {
2219 			dnode_phys_t *dnp = db->db.db_data;
2220 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2221 			    i--, dnp++) {
2222 				if (dnp->dn_type != DMU_OT_NONE)
2223 					fill++;
2224 			}
2225 		} else {
2226 			fill = 1;
2227 		}
2228 	} else {
2229 		blkptr_t *ibp = db->db.db_data;
2230 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2231 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2232 			if (BP_IS_HOLE(ibp))
2233 				continue;
2234 			ASSERT3U(BP_GET_LSIZE(ibp), ==,
2235 			    db->db_level == 1 ? dn->dn_datablksz :
2236 			    (1<<dn->dn_phys->dn_indblkshift));
2237 			fill += ibp->blk_fill;
2238 		}
2239 	}
2240 
2241 	bp->blk_fill = fill;
2242 
2243 	mutex_exit(&db->db_mtx);
2244 }
2245 
2246 /* ARGSUSED */
2247 static void
2248 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2249 {
2250 	dmu_buf_impl_t *db = vdb;
2251 	blkptr_t *bp = zio->io_bp;
2252 	blkptr_t *bp_orig = &zio->io_bp_orig;
2253 	dnode_t *dn = db->db_dnode;
2254 	objset_t *os = dn->dn_objset;
2255 	uint64_t txg = zio->io_txg;
2256 	dbuf_dirty_record_t **drp, *dr;
2257 
2258 	ASSERT3U(zio->io_error, ==, 0);
2259 	ASSERT(db->db_blkptr == bp);
2260 
2261 	if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
2262 		ASSERT(BP_EQUAL(bp, bp_orig));
2263 	} else {
2264 		dsl_dataset_t *ds = os->os_dsl_dataset;
2265 		dmu_tx_t *tx = os->os_synctx;
2266 
2267 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2268 		dsl_dataset_block_born(ds, bp, tx);
2269 	}
2270 
2271 	mutex_enter(&db->db_mtx);
2272 
2273 	DBUF_VERIFY(db);
2274 
2275 	drp = &db->db_last_dirty;
2276 	while ((dr = *drp) != db->db_data_pending)
2277 		drp = &dr->dr_next;
2278 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2279 	ASSERT(dr->dr_txg == txg);
2280 	ASSERT(dr->dr_dbuf == db);
2281 	ASSERT(dr->dr_next == NULL);
2282 	*drp = dr->dr_next;
2283 
2284 	if (db->db_level == 0) {
2285 		ASSERT(db->db_blkid != DB_BONUS_BLKID);
2286 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2287 		if (db->db_state != DB_NOFILL) {
2288 			if (dr->dt.dl.dr_data != db->db_buf)
2289 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2290 				    db) == 1);
2291 			else if (!arc_released(db->db_buf))
2292 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2293 		}
2294 	} else {
2295 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2296 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2297 		if (!BP_IS_HOLE(db->db_blkptr)) {
2298 			int epbs =
2299 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2300 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2301 			    db->db.db_size);
2302 			ASSERT3U(dn->dn_phys->dn_maxblkid
2303 			    >> (db->db_level * epbs), >=, db->db_blkid);
2304 			arc_set_callback(db->db_buf, dbuf_do_evict, db);
2305 		}
2306 		mutex_destroy(&dr->dt.di.dr_mtx);
2307 		list_destroy(&dr->dt.di.dr_children);
2308 	}
2309 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
2310 
2311 	cv_broadcast(&db->db_changed);
2312 	ASSERT(db->db_dirtycnt > 0);
2313 	db->db_dirtycnt -= 1;
2314 	db->db_data_pending = NULL;
2315 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2316 }
2317 
2318 static void
2319 dbuf_write_nofill_ready(zio_t *zio)
2320 {
2321 	dbuf_write_ready(zio, NULL, zio->io_private);
2322 }
2323 
2324 static void
2325 dbuf_write_nofill_done(zio_t *zio)
2326 {
2327 	dbuf_write_done(zio, NULL, zio->io_private);
2328 }
2329 
2330 static void
2331 dbuf_write_override_ready(zio_t *zio)
2332 {
2333 	dbuf_dirty_record_t *dr = zio->io_private;
2334 	dmu_buf_impl_t *db = dr->dr_dbuf;
2335 
2336 	dbuf_write_ready(zio, NULL, db);
2337 }
2338 
2339 static void
2340 dbuf_write_override_done(zio_t *zio)
2341 {
2342 	dbuf_dirty_record_t *dr = zio->io_private;
2343 	dmu_buf_impl_t *db = dr->dr_dbuf;
2344 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2345 
2346 	mutex_enter(&db->db_mtx);
2347 	if (!BP_EQUAL(zio->io_bp, obp)) {
2348 		if (!BP_IS_HOLE(obp))
2349 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2350 		arc_release(dr->dt.dl.dr_data, db);
2351 	}
2352 	mutex_exit(&db->db_mtx);
2353 
2354 	dbuf_write_done(zio, NULL, db);
2355 }
2356 
2357 static void
2358 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2359 {
2360 	dmu_buf_impl_t *db = dr->dr_dbuf;
2361 	dnode_t *dn = db->db_dnode;
2362 	objset_t *os = dn->dn_objset;
2363 	dmu_buf_impl_t *parent = db->db_parent;
2364 	uint64_t txg = tx->tx_txg;
2365 	zbookmark_t zb;
2366 	zio_prop_t zp;
2367 	zio_t *zio;
2368 
2369 	if (db->db_state != DB_NOFILL) {
2370 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2371 			/*
2372 			 * Private object buffers are released here rather
2373 			 * than in dbuf_dirty() since they are only modified
2374 			 * in the syncing context and we don't want the
2375 			 * overhead of making multiple copies of the data.
2376 			 */
2377 			if (BP_IS_HOLE(db->db_blkptr)) {
2378 				arc_buf_thaw(data);
2379 			} else {
2380 				arc_release(data, db);
2381 			}
2382 		}
2383 	}
2384 
2385 	if (parent != dn->dn_dbuf) {
2386 		ASSERT(parent && parent->db_data_pending);
2387 		ASSERT(db->db_level == parent->db_level-1);
2388 		ASSERT(arc_released(parent->db_buf));
2389 		zio = parent->db_data_pending->dr_zio;
2390 	} else {
2391 		ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1);
2392 		ASSERT3P(db->db_blkptr, ==,
2393 		    &dn->dn_phys->dn_blkptr[db->db_blkid]);
2394 		zio = dn->dn_zio;
2395 	}
2396 
2397 	ASSERT(db->db_level == 0 || data == db->db_buf);
2398 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2399 	ASSERT(zio);
2400 
2401 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2402 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2403 	    db->db.db_object, db->db_level, db->db_blkid);
2404 
2405 	dmu_write_policy(os, dn, db->db_level,
2406 	    db->db_state == DB_NOFILL ? WP_NOFILL : 0, &zp);
2407 
2408 	if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2409 		ASSERT(db->db_state != DB_NOFILL);
2410 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2411 		    db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2412 		    dbuf_write_override_ready, dbuf_write_override_done, dr,
2413 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2414 		mutex_enter(&db->db_mtx);
2415 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2416 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2417 		    dr->dt.dl.dr_copies);
2418 		mutex_exit(&db->db_mtx);
2419 	} else if (db->db_state == DB_NOFILL) {
2420 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2421 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2422 		    db->db_blkptr, NULL, db->db.db_size, &zp,
2423 		    dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2424 		    ZIO_PRIORITY_ASYNC_WRITE,
2425 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2426 	} else {
2427 		ASSERT(arc_released(data));
2428 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
2429 		    db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp,
2430 		    dbuf_write_ready, dbuf_write_done, db,
2431 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2432 	}
2433 }
2434